1 /*
2 * Copyright (c) 1997, 2026, 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/ciFlatArray.hpp"
28 #include "ci/ciInlineKlass.hpp"
29 #include "ci/ciReplay.hpp"
30 #include "classfile/javaClasses.hpp"
31 #include "code/aotCodeCache.hpp"
32 #include "code/exceptionHandlerTable.hpp"
33 #include "code/nmethod.hpp"
34 #include "compiler/compilationFailureInfo.hpp"
35 #include "compiler/compilationMemoryStatistic.hpp"
36 #include "compiler/compileBroker.hpp"
37 #include "compiler/compileLog.hpp"
38 #include "compiler/compiler_globals.hpp"
39 #include "compiler/compilerDefinitions.hpp"
40 #include "compiler/compilerOracle.hpp"
41 #include "compiler/disassembler.hpp"
42 #include "compiler/oopMap.hpp"
43 #include "gc/shared/barrierSet.hpp"
44 #include "gc/shared/c2/barrierSetC2.hpp"
45 #include "jfr/jfrEvents.hpp"
46 #include "jvm_io.h"
47 #include "memory/allocation.hpp"
48 #include "memory/arena.hpp"
49 #include "memory/resourceArea.hpp"
50 #include "opto/addnode.hpp"
51 #include "opto/block.hpp"
52 #include "opto/c2compiler.hpp"
53 #include "opto/callGenerator.hpp"
54 #include "opto/callnode.hpp"
55 #include "opto/castnode.hpp"
56 #include "opto/cfgnode.hpp"
57 #include "opto/chaitin.hpp"
58 #include "opto/compile.hpp"
59 #include "opto/connode.hpp"
60 #include "opto/convertnode.hpp"
61 #include "opto/divnode.hpp"
62 #include "opto/escape.hpp"
63 #include "opto/idealGraphPrinter.hpp"
64 #include "opto/inlinetypenode.hpp"
65 #include "opto/locknode.hpp"
66 #include "opto/loopnode.hpp"
67 #include "opto/machnode.hpp"
68 #include "opto/macro.hpp"
69 #include "opto/matcher.hpp"
70 #include "opto/mathexactnode.hpp"
71 #include "opto/memnode.hpp"
72 #include "opto/movenode.hpp"
73 #include "opto/mulnode.hpp"
74 #include "opto/multnode.hpp"
75 #include "opto/narrowptrnode.hpp"
76 #include "opto/node.hpp"
77 #include "opto/opaquenode.hpp"
78 #include "opto/opcodes.hpp"
79 #include "opto/output.hpp"
80 #include "opto/parse.hpp"
81 #include "opto/phaseX.hpp"
82 #include "opto/reachability.hpp"
83 #include "opto/rootnode.hpp"
84 #include "opto/runtime.hpp"
85 #include "opto/stringopts.hpp"
86 #include "opto/type.hpp"
87 #include "opto/vector.hpp"
88 #include "opto/vectornode.hpp"
89 #include "runtime/arguments.hpp"
90 #include "runtime/globals_extension.hpp"
91 #include "runtime/sharedRuntime.hpp"
92 #include "runtime/signature.hpp"
93 #include "runtime/stubRoutines.hpp"
94 #include "runtime/timer.hpp"
95 #include "utilities/align.hpp"
96 #include "utilities/copy.hpp"
97 #include "utilities/hashTable.hpp"
98 #include "utilities/macros.hpp"
99
100 // -------------------- Compile::mach_constant_base_node -----------------------
101 // Constant table base node singleton.
102 MachConstantBaseNode* Compile::mach_constant_base_node() {
103 if (_mach_constant_base_node == nullptr) {
104 _mach_constant_base_node = new MachConstantBaseNode();
105 _mach_constant_base_node->add_req(C->root());
106 }
107 return _mach_constant_base_node;
108 }
109
110
111 /// Support for intrinsics.
112
113 // Return the index at which m must be inserted (or already exists).
114 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
115 class IntrinsicDescPair {
116 private:
117 ciMethod* _m;
118 bool _is_virtual;
119 public:
120 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
121 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
122 ciMethod* m= elt->method();
123 ciMethod* key_m = key->_m;
124 if (key_m < m) return -1;
125 else if (key_m > m) return 1;
126 else {
127 bool is_virtual = elt->is_virtual();
128 bool key_virtual = key->_is_virtual;
129 if (key_virtual < is_virtual) return -1;
130 else if (key_virtual > is_virtual) return 1;
131 else return 0;
132 }
133 }
134 };
135 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
136 #ifdef ASSERT
137 for (int i = 1; i < _intrinsics.length(); i++) {
138 CallGenerator* cg1 = _intrinsics.at(i-1);
139 CallGenerator* cg2 = _intrinsics.at(i);
140 assert(cg1->method() != cg2->method()
141 ? cg1->method() < cg2->method()
142 : cg1->is_virtual() < cg2->is_virtual(),
143 "compiler intrinsics list must stay sorted");
144 }
145 #endif
146 IntrinsicDescPair pair(m, is_virtual);
147 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
148 }
149
150 void Compile::register_intrinsic(CallGenerator* cg) {
151 bool found = false;
152 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
153 assert(!found, "registering twice");
154 _intrinsics.insert_before(index, cg);
155 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
156 }
157
158 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
159 assert(m->is_loaded(), "don't try this on unloaded methods");
160 if (_intrinsics.length() > 0) {
161 bool found = false;
162 int index = intrinsic_insertion_index(m, is_virtual, found);
163 if (found) {
164 return _intrinsics.at(index);
165 }
166 }
167 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
168 if (m->intrinsic_id() != vmIntrinsics::_none &&
169 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
170 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
171 if (cg != nullptr) {
172 // Save it for next time:
173 register_intrinsic(cg);
174 return cg;
175 } else {
176 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
177 }
178 }
179 return nullptr;
180 }
181
182 // Compile::make_vm_intrinsic is defined in library_call.cpp.
183
184 #ifndef PRODUCT
185 // statistics gathering...
186
187 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
188 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
189
190 inline int as_int(vmIntrinsics::ID id) {
191 return vmIntrinsics::as_int(id);
192 }
193
194 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
195 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
196 int oflags = _intrinsic_hist_flags[as_int(id)];
197 assert(flags != 0, "what happened?");
198 if (is_virtual) {
199 flags |= _intrinsic_virtual;
200 }
201 bool changed = (flags != oflags);
202 if ((flags & _intrinsic_worked) != 0) {
203 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
204 if (count == 1) {
205 changed = true; // first time
206 }
207 // increment the overall count also:
208 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
209 }
210 if (changed) {
211 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
212 // Something changed about the intrinsic's virtuality.
213 if ((flags & _intrinsic_virtual) != 0) {
214 // This is the first use of this intrinsic as a virtual call.
215 if (oflags != 0) {
216 // We already saw it as a non-virtual, so note both cases.
217 flags |= _intrinsic_both;
218 }
219 } else if ((oflags & _intrinsic_both) == 0) {
220 // This is the first use of this intrinsic as a non-virtual
221 flags |= _intrinsic_both;
222 }
223 }
224 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
225 }
226 // update the overall flags also:
227 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
228 return changed;
229 }
230
231 static char* format_flags(int flags, char* buf) {
232 buf[0] = 0;
233 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
234 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
235 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
236 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
237 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
238 if (buf[0] == 0) strcat(buf, ",");
239 assert(buf[0] == ',', "must be");
240 return &buf[1];
241 }
242
243 void Compile::print_intrinsic_statistics() {
244 char flagsbuf[100];
245 ttyLocker ttyl;
246 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
247 tty->print_cr("Compiler intrinsic usage:");
248 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
249 if (total == 0) total = 1; // avoid div0 in case of no successes
250 #define PRINT_STAT_LINE(name, c, f) \
251 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
252 for (auto id : EnumRange<vmIntrinsicID>{}) {
253 int flags = _intrinsic_hist_flags[as_int(id)];
254 juint count = _intrinsic_hist_count[as_int(id)];
255 if ((flags | count) != 0) {
256 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
257 }
258 }
259 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
260 if (xtty != nullptr) xtty->tail("statistics");
261 }
262
263 void Compile::print_statistics() {
264 { ttyLocker ttyl;
265 if (xtty != nullptr) xtty->head("statistics type='opto'");
266 Parse::print_statistics();
267 PhaseStringOpts::print_statistics();
268 PhaseCCP::print_statistics();
269 PhaseRegAlloc::print_statistics();
270 PhaseOutput::print_statistics();
271 PhasePeephole::print_statistics();
272 PhaseIdealLoop::print_statistics();
273 ConnectionGraph::print_statistics();
274 PhaseMacroExpand::print_statistics();
275 if (xtty != nullptr) xtty->tail("statistics");
276 }
277 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
278 // put this under its own <statistics> element.
279 print_intrinsic_statistics();
280 }
281 }
282 #endif //PRODUCT
283
284 void Compile::gvn_replace_by(Node* n, Node* nn) {
285 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
286 Node* use = n->last_out(i);
287 bool is_in_table = initial_gvn()->hash_delete(use);
288 uint uses_found = 0;
289 for (uint j = 0; j < use->len(); j++) {
290 if (use->in(j) == n) {
291 if (j < use->req())
292 use->set_req(j, nn);
293 else
294 use->set_prec(j, nn);
295 uses_found++;
296 }
297 }
298 if (is_in_table) {
299 // reinsert into table
300 initial_gvn()->hash_find_insert(use);
301 }
302 record_for_igvn(use);
303 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
304 i -= uses_found; // we deleted 1 or more copies of this edge
305 }
306 }
307
308
309 // Identify all nodes that are reachable from below, useful.
310 // Use breadth-first pass that records state in a Unique_Node_List,
311 // recursive traversal is slower.
312 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
313 int estimated_worklist_size = live_nodes();
314 useful.map( estimated_worklist_size, nullptr ); // preallocate space
315
316 // Initialize worklist
317 if (root() != nullptr) { useful.push(root()); }
318 // If 'top' is cached, declare it useful to preserve cached node
319 if (cached_top_node()) { useful.push(cached_top_node()); }
320
321 // Push all useful nodes onto the list, breadthfirst
322 for( uint next = 0; next < useful.size(); ++next ) {
323 assert( next < unique(), "Unique useful nodes < total nodes");
324 Node *n = useful.at(next);
325 uint max = n->len();
326 for( uint i = 0; i < max; ++i ) {
327 Node *m = n->in(i);
328 if (not_a_node(m)) continue;
329 useful.push(m);
330 }
331 }
332 }
333
334 // Update dead_node_list with any missing dead nodes using useful
335 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
336 void Compile::update_dead_node_list(Unique_Node_List &useful) {
337 uint max_idx = unique();
338 VectorSet& useful_node_set = useful.member_set();
339
340 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
341 // If node with index node_idx is not in useful set,
342 // mark it as dead in dead node list.
343 if (!useful_node_set.test(node_idx)) {
344 record_dead_node(node_idx);
345 }
346 }
347 }
348
349 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
350 int shift = 0;
351 for (int i = 0; i < inlines->length(); i++) {
352 CallGenerator* cg = inlines->at(i);
353 if (useful.member(cg->call_node())) {
354 if (shift > 0) {
355 inlines->at_put(i - shift, cg);
356 }
357 } else {
358 shift++; // skip over the dead element
359 }
360 }
361 if (shift > 0) {
362 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
363 }
364 }
365
366 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
367 assert(dead != nullptr && dead->is_Call(), "sanity");
368 int found = 0;
369 for (int i = 0; i < inlines->length(); i++) {
370 if (inlines->at(i)->call_node() == dead) {
371 inlines->remove_at(i);
372 found++;
373 NOT_DEBUG( break; ) // elements are unique, so exit early
374 }
375 }
376 assert(found <= 1, "not unique");
377 }
378
379 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
380 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
381 for (int i = node_list.length() - 1; i >= 0; i--) {
382 N* node = node_list.at(i);
383 if (!useful.member(node)) {
384 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
385 }
386 }
387 }
388
389 void Compile::remove_useless_node(Node* dead) {
390 remove_modified_node(dead);
391
392 // Constant node that has no out-edges and has only one in-edge from
393 // root is usually dead. However, sometimes reshaping walk makes
394 // it reachable by adding use edges. So, we will NOT count Con nodes
395 // as dead to be conservative about the dead node count at any
396 // given time.
397 if (!dead->is_Con()) {
398 record_dead_node(dead->_idx);
399 }
400 if (dead->is_macro()) {
401 remove_macro_node(dead);
402 }
403 if (dead->is_expensive()) {
404 remove_expensive_node(dead);
405 }
406 if (dead->is_ReachabilityFence()) {
407 remove_reachability_fence(dead->as_ReachabilityFence());
408 }
409 if (dead->is_OpaqueTemplateAssertionPredicate()) {
410 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
411 }
412 if (dead->is_ParsePredicate()) {
413 remove_parse_predicate(dead->as_ParsePredicate());
414 }
415 if (dead->for_post_loop_opts_igvn()) {
416 remove_from_post_loop_opts_igvn(dead);
417 }
418 if (dead->is_InlineType()) {
419 remove_inline_type(dead);
420 }
421 if (dead->is_LoadFlat() || dead->is_StoreFlat()) {
422 remove_flat_access(dead);
423 }
424 if (dead->for_merge_stores_igvn()) {
425 remove_from_merge_stores_igvn(dead);
426 }
427 if (dead->is_Call()) {
428 remove_useless_late_inlines( &_late_inlines, dead);
429 remove_useless_late_inlines( &_string_late_inlines, dead);
430 remove_useless_late_inlines( &_boxing_late_inlines, dead);
431 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
432
433 if (dead->is_CallStaticJava()) {
434 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
435 }
436 }
437 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
438 bs->unregister_potential_barrier_node(dead);
439 }
440
441 // Disconnect all useless nodes by disconnecting those at the boundary.
442 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
443 uint next = 0;
444 while (next < useful.size()) {
445 Node *n = useful.at(next++);
446 if (n->is_SafePoint()) {
447 // We're done with a parsing phase. Replaced nodes are not valid
448 // beyond that point.
449 n->as_SafePoint()->delete_replaced_nodes();
450 }
451 // Use raw traversal of out edges since this code removes out edges
452 int max = n->outcnt();
453 for (int j = 0; j < max; ++j) {
454 Node* child = n->raw_out(j);
455 if (!useful.member(child)) {
456 assert(!child->is_top() || child != top(),
457 "If top is cached in Compile object it is in useful list");
458 // Only need to remove this out-edge to the useless node
459 n->raw_del_out(j);
460 --j;
461 --max;
462 if (child->is_data_proj_of_pure_function(n)) {
463 worklist.push(n);
464 }
465 }
466 }
467 if (n->outcnt() == 1 && n->has_special_unique_user()) {
468 assert(useful.member(n->unique_out()), "do not push a useless node");
469 worklist.push(n->unique_out());
470 }
471 if (n->outcnt() == 0) {
472 worklist.push(n);
473 }
474 }
475
476 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
477 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
478 // Remove useless Template Assertion Predicate opaque nodes
479 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
480 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
481 remove_useless_nodes(_reachability_fences, useful); // remove useless node recorded for post loop opts IGVN pass
482 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
483 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
484 remove_useless_nodes(_flat_access_nodes, useful); // remove useless flat access nodes
485 #ifdef ASSERT
486 if (_modified_nodes != nullptr) {
487 _modified_nodes->remove_useless_nodes(useful.member_set());
488 }
489 #endif
490 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
491 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
492 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
493 #ifdef ASSERT
494 if (_modified_nodes != nullptr) {
495 _modified_nodes->remove_useless_nodes(useful.member_set());
496 }
497 #endif
498
499 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
500 bs->eliminate_useless_gc_barriers(useful, this);
501 // clean up the late inline lists
502 remove_useless_late_inlines( &_late_inlines, useful);
503 remove_useless_late_inlines( &_string_late_inlines, useful);
504 remove_useless_late_inlines( &_boxing_late_inlines, useful);
505 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
506 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
507 }
508
509 // ============================================================================
510 //------------------------------CompileWrapper---------------------------------
511 class CompileWrapper : public StackObj {
512 Compile *const _compile;
513 public:
514 CompileWrapper(Compile* compile);
515
516 ~CompileWrapper();
517 };
518
519 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
520 // the Compile* pointer is stored in the current ciEnv:
521 ciEnv* env = compile->env();
522 assert(env == ciEnv::current(), "must already be a ciEnv active");
523 assert(env->compiler_data() == nullptr, "compile already active?");
524 env->set_compiler_data(compile);
525 assert(compile == Compile::current(), "sanity");
526
527 compile->set_type_dict(nullptr);
528 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
529 compile->clone_map().set_clone_idx(0);
530 compile->set_type_last_size(0);
531 compile->set_last_tf(nullptr, nullptr);
532 compile->set_indexSet_arena(nullptr);
533 compile->set_indexSet_free_block_list(nullptr);
534 compile->init_type_arena();
535 Type::Initialize(compile);
536 _compile->begin_method();
537 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
538 }
539 CompileWrapper::~CompileWrapper() {
540 // simulate crash during compilation
541 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
542
543 _compile->end_method();
544 _compile->env()->set_compiler_data(nullptr);
545 }
546
547
548 //----------------------------print_compile_messages---------------------------
549 void Compile::print_compile_messages() {
550 #ifndef PRODUCT
551 // Check if recompiling
552 if (!subsume_loads() && PrintOpto) {
553 // Recompiling without allowing machine instructions to subsume loads
554 tty->print_cr("*********************************************************");
555 tty->print_cr("** Bailout: Recompile without subsuming loads **");
556 tty->print_cr("*********************************************************");
557 }
558 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
559 // Recompiling without escape analysis
560 tty->print_cr("*********************************************************");
561 tty->print_cr("** Bailout: Recompile without escape analysis **");
562 tty->print_cr("*********************************************************");
563 }
564 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
565 // Recompiling without iterative escape analysis
566 tty->print_cr("*********************************************************");
567 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
568 tty->print_cr("*********************************************************");
569 }
570 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
571 // Recompiling without reducing allocation merges
572 tty->print_cr("*********************************************************");
573 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
574 tty->print_cr("*********************************************************");
575 }
576 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
577 // Recompiling without boxing elimination
578 tty->print_cr("*********************************************************");
579 tty->print_cr("** Bailout: Recompile without boxing elimination **");
580 tty->print_cr("*********************************************************");
581 }
582 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
583 // Recompiling without locks coarsening
584 tty->print_cr("*********************************************************");
585 tty->print_cr("** Bailout: Recompile without locks coarsening **");
586 tty->print_cr("*********************************************************");
587 }
588 if (env()->break_at_compile()) {
589 // Open the debugger when compiling this method.
590 tty->print("### Breaking when compiling: ");
591 method()->print_short_name();
592 tty->cr();
593 BREAKPOINT;
594 }
595
596 if( PrintOpto ) {
597 if (is_osr_compilation()) {
598 tty->print("[OSR]%3d", _compile_id);
599 } else {
600 tty->print("%3d", _compile_id);
601 }
602 }
603 #endif
604 }
605
606 #ifndef PRODUCT
607 void Compile::print_phase(const char* phase_name) {
608 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name);
609 }
610
611 void Compile::print_ideal_ir(const char* compile_phase_name) const {
612 // keep the following output all in one block
613 // This output goes directly to the tty, not the compiler log.
614 // To enable tools to match it up with the compilation activity,
615 // be sure to tag this tty output with the compile ID.
616
617 // Node dumping can cause a safepoint, which can break the tty lock.
618 // Buffer all node dumps, so that all safepoints happen before we lock.
619 ResourceMark rm;
620 stringStream ss;
621
622 if (_output == nullptr) {
623 ss.print_cr("AFTER: %s", compile_phase_name);
624 // Print out all nodes in ascending order of index.
625 // It is important that we traverse both inputs and outputs of nodes,
626 // so that we reach all nodes that are connected to Root.
627 root()->dump_bfs(MaxNodeLimit, nullptr, "-+S$", &ss);
628 } else {
629 // Dump the node blockwise if we have a scheduling
630 _output->print_scheduling(&ss);
631 }
632
633 // Check that the lock is not broken by a safepoint.
634 NoSafepointVerifier nsv;
635 ttyLocker ttyl;
636 if (xtty != nullptr) {
637 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
638 compile_id(),
639 is_osr_compilation() ? " compile_kind='osr'" : "",
640 compile_phase_name);
641 }
642
643 tty->print("%s", ss.as_string());
644
645 if (xtty != nullptr) {
646 xtty->tail("ideal");
647 }
648 }
649 #endif
650
651 // ============================================================================
652 //------------------------------Compile standard-------------------------------
653
654 // Compile a method. entry_bci is -1 for normal compilations and indicates
655 // the continuation bci for on stack replacement.
656
657
658 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
659 Options options, DirectiveSet* directive)
660 : Phase(Compiler),
661 _compile_id(ci_env->compile_id()),
662 _options(options),
663 _method(target),
664 _entry_bci(osr_bci),
665 _ilt(nullptr),
666 _stub_function(nullptr),
667 _stub_name(nullptr),
668 _stub_id(StubId::NO_STUBID),
669 _stub_entry_point(nullptr),
670 _max_node_limit(MaxNodeLimit),
671 _node_count_inlining_cutoff(NodeCountInliningCutoff),
672 _post_loop_opts_phase(false),
673 _merge_stores_phase(false),
674 _allow_macro_nodes(true),
675 _inlining_progress(false),
676 _inlining_incrementally(false),
677 _strength_reduction(false),
678 _do_cleanup(false),
679 _has_reserved_stack_access(target->has_reserved_stack_access()),
680 _has_circular_inline_type(false),
681 #ifndef PRODUCT
682 _igv_idx(0),
683 _trace_opto_output(directive->TraceOptoOutputOption),
684 #endif
685 _clinit_barrier_on_entry(false),
686 _stress_seed(0),
687 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
688 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
689 _env(ci_env),
690 _directive(directive),
691 _log(ci_env->log()),
692 _first_failure_details(nullptr),
693 _intrinsics(comp_arena(), 0, 0, nullptr),
694 _macro_nodes(comp_arena(), 8, 0, nullptr),
695 _parse_predicates(comp_arena(), 8, 0, nullptr),
696 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
697 _expensive_nodes(comp_arena(), 8, 0, nullptr),
698 _reachability_fences(comp_arena(), 8, 0, nullptr),
699 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
700 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
701 _flat_access_nodes(comp_arena(), 8, 0, nullptr),
702 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
703 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
704 _coarsened_locks(comp_arena(), 8, 0, nullptr),
705 _congraph(nullptr),
706 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
707 _unique(0),
708 _dead_node_count(0),
709 _dead_node_list(comp_arena()),
710 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
711 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
712 _node_arena(&_node_arena_one),
713 _mach_constant_base_node(nullptr),
714 _Compile_types(mtCompiler, Arena::Tag::tag_type),
715 _initial_gvn(nullptr),
716 _igvn_worklist(nullptr),
717 _types(nullptr),
718 _node_hash(nullptr),
719 _late_inlines(comp_arena(), 2, 0, nullptr),
720 _string_late_inlines(comp_arena(), 2, 0, nullptr),
721 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
722 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
723 _late_inlines_pos(0),
724 _has_mh_late_inlines(false),
725 _oom(false),
726 _replay_inline_data(nullptr),
727 _inline_printer(this),
728 _java_calls(0),
729 _inner_loops(0),
730 _FIRST_STACK_mask(comp_arena()),
731 _interpreter_frame_size(0),
732 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
733 _output(nullptr)
734 #ifndef PRODUCT
735 ,
736 _in_dump_cnt(0)
737 #endif
738 {
739 C = this;
740 CompileWrapper cw(this);
741
742 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
743 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
744
745 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
746 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
747 // We can always print a disassembly, either abstract (hex dump) or
748 // with the help of a suitable hsdis library. Thus, we should not
749 // couple print_assembly and print_opto_assembly controls.
750 // But: always print opto and regular assembly on compile command 'print'.
751 bool print_assembly = directive->PrintAssemblyOption;
752 set_print_assembly(print_opto_assembly || print_assembly);
753 #else
754 set_print_assembly(false); // must initialize.
755 #endif
756
757 #ifndef PRODUCT
758 set_parsed_irreducible_loop(false);
759 #endif
760
761 if (directive->ReplayInlineOption) {
762 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
763 }
764 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
765 set_print_intrinsics(directive->PrintIntrinsicsOption);
766 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
767
768 if (ProfileTraps) {
769 // Make sure the method being compiled gets its own MDO,
770 // so we can at least track the decompile_count().
771 method()->ensure_method_data();
772 }
773
774 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
775 StressIncrementalInlining || StressMacroExpansion ||
776 StressMacroElimination || StressUnstableIfTraps ||
777 StressBailout || StressLoopPeeling || StressCountedLoop ||
778 StressEliminateAllocations) {
779 initialize_stress_seed(directive);
780 }
781
782 Init(/*do_aliasing=*/ true);
783
784 print_compile_messages();
785
786 _ilt = InlineTree::build_inline_tree_root();
787
788 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
789 assert(num_alias_types() >= AliasIdxRaw, "");
790
791 #define MINIMUM_NODE_HASH 1023
792
793 // GVN that will be run immediately on new nodes
794 uint estimated_size = method()->code_size()*4+64;
795 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
796 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
797 _types = new (comp_arena()) Type_Array(comp_arena());
798 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
799 PhaseGVN gvn;
800 set_initial_gvn(&gvn);
801
802 { // Scope for timing the parser
803 TracePhase tp(_t_parser);
804
805 // Put top into the hash table ASAP.
806 initial_gvn()->transform(top());
807
808 // Set up tf(), start(), and find a CallGenerator.
809 CallGenerator* cg = nullptr;
810 if (is_osr_compilation()) {
811 init_tf(TypeFunc::make(method(), false, /* is_osr_compilation = */ true));
812 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
813 initial_gvn()->set_type_bottom(s);
814 verify_start(s);
815 cg = CallGenerator::for_osr(method(), entry_bci());
816 } else {
817 // Normal case.
818 init_tf(TypeFunc::make(method(), false));
819 StartNode* s = new StartNode(root(), tf()->domain_cc());
820 initial_gvn()->set_type_bottom(s);
821 verify_start(s);
822 float past_uses = method()->interpreter_invocation_count();
823 float expected_uses = past_uses;
824 cg = CallGenerator::for_inline(method(), expected_uses);
825 }
826 if (failing()) return;
827 if (cg == nullptr) {
828 const char* reason = InlineTree::check_can_parse(method());
829 assert(reason != nullptr, "expect reason for parse failure");
830 stringStream ss;
831 ss.print("cannot parse method: %s", reason);
832 record_method_not_compilable(ss.as_string());
833 return;
834 }
835
836 gvn.set_type(root(), root()->bottom_type());
837
838 JVMState* jvms = build_start_state(start(), tf());
839 if ((jvms = cg->generate(jvms)) == nullptr) {
840 assert(failure_reason() != nullptr, "expect reason for parse failure");
841 stringStream ss;
842 ss.print("method parse failed: %s", failure_reason());
843 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
844 return;
845 }
846 GraphKit kit(jvms);
847
848 if (!kit.stopped()) {
849 // Accept return values, and transfer control we know not where.
850 // This is done by a special, unique ReturnNode bound to root.
851 return_values(kit.jvms());
852 }
853
854 if (kit.has_exceptions()) {
855 // Any exceptions that escape from this call must be rethrown
856 // to whatever caller is dynamically above us on the stack.
857 // This is done by a special, unique RethrowNode bound to root.
858 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
859 }
860
861 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
862
863 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
864 inline_string_calls(true);
865 }
866
867 if (failing()) return;
868
869 // Remove clutter produced by parsing.
870 if (!failing()) {
871 ResourceMark rm;
872 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
873 }
874 }
875
876 // Note: Large methods are capped off in do_one_bytecode().
877 if (failing()) return;
878
879 // After parsing, node notes are no longer automagic.
880 // They must be propagated by register_new_node_with_optimizer(),
881 // clone(), or the like.
882 set_default_node_notes(nullptr);
883
884 #ifndef PRODUCT
885 if (should_print_igv(1)) {
886 _igv_printer->print_inlining();
887 }
888 #endif
889
890 if (failing()) return;
891 NOT_PRODUCT( verify_graph_edges(); )
892
893 // Now optimize
894 Optimize();
895 if (failing()) return;
896 NOT_PRODUCT( verify_graph_edges(); )
897
898 #ifndef PRODUCT
899 if (should_print_ideal()) {
900 print_ideal_ir("PrintIdeal");
901 }
902 #endif
903
904 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
905 bs->final_refinement(this);
906
907 #ifdef ASSERT
908 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
909 #endif
910
911 // Dump compilation data to replay it.
912 if (directive->DumpReplayOption) {
913 env()->dump_replay_data(_compile_id);
914 }
915 if (directive->DumpInlineOption && (ilt() != nullptr)) {
916 env()->dump_inline_data(_compile_id);
917 }
918
919 // Now that we know the size of all the monitors we can add fixed slots:
920 // [...]
921 // rsp+80: saved fp register
922 // rsp+76: Fixed slot 7
923 // rsp+72: Fixed slot 6 (stack increment)
924 // rsp+68: Fixed slot 5
925 // rsp+64: Fixed slot 4 (null marker)
926 // rsp+60: Fixed slot 3
927 // rsp+56: Fixed slot 2 (original deopt pc)
928 // rsp+52: Fixed slot 1
929 // rsp+48: Fixed slot 0 (monitors)
930 // rsp+44: spill
931 // [...]
932
933 // One extra slot for the original deopt pc.
934 int next_slot = fixed_slots();
935 next_slot += VMRegImpl::slots_per_word;
936
937 // One extra slot for the special stack increment value.
938 if (needs_stack_repair()) {
939 next_slot += VMRegImpl::slots_per_word;
940 }
941
942 // One extra slot to hold the null marker at scalarized returns.
943 if (needs_nm_slot()) {
944 next_slot += VMRegImpl::slots_per_word;
945 }
946 set_fixed_slots(next_slot);
947
948 // Compute when to use implicit null checks. Used by matching trap based
949 // nodes and NullCheck optimization.
950 set_allowed_deopt_reasons();
951
952 // Now generate code
953 Code_Gen();
954 }
955
956 // C2 uses runtime stubs serialized generation to initialize its static tables
957 // shared by all compilations, like Type::_shared_type_dict.
958 // At least one stub have to be completely generated to execute intialization
959 // before we can skip the rest stubs generation by loading AOT cached stubs.
960
961 static bool c2_do_stub_init_complete = false;
962
963 //------------------------------Compile----------------------------------------
964 // Compile a runtime stub
965 Compile::Compile(ciEnv* ci_env,
966 TypeFunc_generator generator,
967 address stub_function,
968 const char* stub_name,
969 StubId stub_id,
970 int is_fancy_jump,
971 bool pass_tls,
972 bool return_pc,
973 DirectiveSet* directive)
974 : Phase(Compiler),
975 _compile_id(0),
976 _options(Options::for_runtime_stub()),
977 _method(nullptr),
978 _entry_bci(InvocationEntryBci),
979 _stub_function(stub_function),
980 _stub_name(stub_name),
981 _stub_id(stub_id),
982 _stub_entry_point(nullptr),
983 _max_node_limit(MaxNodeLimit),
984 _node_count_inlining_cutoff(NodeCountInliningCutoff),
985 _post_loop_opts_phase(false),
986 _merge_stores_phase(false),
987 _allow_macro_nodes(true),
988 _inlining_progress(false),
989 _inlining_incrementally(false),
990 _has_reserved_stack_access(false),
991 _has_circular_inline_type(false),
992 #ifndef PRODUCT
993 _igv_idx(0),
994 _trace_opto_output(directive->TraceOptoOutputOption),
995 #endif
996 _clinit_barrier_on_entry(false),
997 _stress_seed(0),
998 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
999 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
1000 _env(ci_env),
1001 _directive(directive),
1002 _log(ci_env->log()),
1003 _first_failure_details(nullptr),
1004 _reachability_fences(comp_arena(), 8, 0, nullptr),
1005 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
1006 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
1007 _congraph(nullptr),
1008 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
1009 _unique(0),
1010 _dead_node_count(0),
1011 _dead_node_list(comp_arena()),
1012 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
1013 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
1014 _node_arena(&_node_arena_one),
1015 _mach_constant_base_node(nullptr),
1016 _Compile_types(mtCompiler, Arena::Tag::tag_type),
1017 _initial_gvn(nullptr),
1018 _igvn_worklist(nullptr),
1019 _types(nullptr),
1020 _node_hash(nullptr),
1021 _has_mh_late_inlines(false),
1022 _oom(false),
1023 _replay_inline_data(nullptr),
1024 _inline_printer(this),
1025 _java_calls(0),
1026 _inner_loops(0),
1027 _FIRST_STACK_mask(comp_arena()),
1028 _interpreter_frame_size(0),
1029 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
1030 _output(nullptr),
1031 #ifndef PRODUCT
1032 _in_dump_cnt(0),
1033 #endif
1034 _allowed_reasons(0) {
1035 C = this;
1036
1037 // try to reuse an existing stub
1038 if (c2_do_stub_init_complete) {
1039 BlobId blob_id = StubInfo::blob(_stub_id);
1040 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
1041 if (blob != nullptr) {
1042 RuntimeStub* rs = blob->as_runtime_stub();
1043 _stub_entry_point = rs->entry_point();
1044 return;
1045 }
1046 }
1047
1048 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
1049 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
1050
1051 #ifndef PRODUCT
1052 set_print_assembly(PrintFrameConverterAssembly);
1053 set_parsed_irreducible_loop(false);
1054 #else
1055 set_print_assembly(false); // Must initialize.
1056 #endif
1057 set_has_irreducible_loop(false); // no loops
1058
1059 CompileWrapper cw(this);
1060 Init(/*do_aliasing=*/ false);
1061 init_tf((*generator)());
1062
1063 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1064 _types = new (comp_arena()) Type_Array(comp_arena());
1065 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1066
1067 if (StressLCM || StressGCM || StressBailout) {
1068 initialize_stress_seed(directive);
1069 }
1070
1071 {
1072 PhaseGVN gvn;
1073 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1074 gvn.transform(top());
1075
1076 GraphKit kit;
1077 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1078 }
1079
1080 NOT_PRODUCT( verify_graph_edges(); )
1081
1082 Code_Gen();
1083
1084 // First successful stub generation will set it to `true`
1085 // and it will stay `true` after that.
1086 c2_do_stub_init_complete = c2_do_stub_init_complete || (_stub_entry_point != nullptr);
1087 }
1088
1089 Compile::~Compile() {
1090 delete _first_failure_details;
1091 };
1092
1093 //------------------------------Init-------------------------------------------
1094 // Prepare for a single compilation
1095 void Compile::Init(bool aliasing) {
1096 _do_aliasing = aliasing;
1097 _unique = 0;
1098 _regalloc = nullptr;
1099
1100 _tf = nullptr; // filled in later
1101 _top = nullptr; // cached later
1102 _matcher = nullptr; // filled in later
1103 _cfg = nullptr; // filled in later
1104
1105 _node_note_array = nullptr;
1106 _default_node_notes = nullptr;
1107 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1108
1109 _immutable_memory = nullptr; // filled in at first inquiry
1110
1111 #ifdef ASSERT
1112 _phase_optimize_finished = false;
1113 _phase_verify_ideal_loop = false;
1114 _exception_backedge = false;
1115 _type_verify = nullptr;
1116 #endif
1117
1118 // Globally visible Nodes
1119 // First set TOP to null to give safe behavior during creation of RootNode
1120 set_cached_top_node(nullptr);
1121 set_root(new RootNode());
1122 // Now that you have a Root to point to, create the real TOP
1123 set_cached_top_node( new ConNode(Type::TOP) );
1124 set_recent_alloc(nullptr, nullptr);
1125
1126 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1127 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1128 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1129 env()->set_dependencies(new Dependencies(env()));
1130
1131 _fixed_slots = 0;
1132 set_has_split_ifs(false);
1133 set_has_loops(false); // first approximation
1134 set_has_stringbuilder(false);
1135 set_has_boxed_value(false);
1136 _trap_can_recompile = false; // no traps emitted yet
1137 _major_progress = true; // start out assuming good things will happen
1138 set_has_unsafe_access(false);
1139 set_max_vector_size(0);
1140 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1141 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1142 set_decompile_count(0);
1143
1144 #ifndef PRODUCT
1145 _phase_counter = 0;
1146 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1147 #endif
1148
1149 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1150 _loop_opts_cnt = LoopOptsCount;
1151 _has_flat_accesses = false;
1152 _flat_accesses_share_alias = true;
1153 _scalarize_in_safepoints = false;
1154 _needs_nm_slot = false;
1155
1156 set_do_inlining(Inline);
1157 set_max_inline_size(MaxInlineSize);
1158 set_freq_inline_size(FreqInlineSize);
1159 set_do_scheduling(OptoScheduling);
1160
1161 set_do_vector_loop(false);
1162 set_has_monitors(false);
1163 set_has_scoped_access(false);
1164
1165 if (AllowVectorizeOnDemand) {
1166 if (has_method() && _directive->VectorizeOption) {
1167 set_do_vector_loop(true);
1168 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());})
1169 } else if (has_method() && method()->name() != nullptr &&
1170 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1171 set_do_vector_loop(true);
1172 }
1173 }
1174 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1175 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());})
1176
1177 _max_node_limit = _directive->MaxNodeLimitOption;
1178
1179 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1180 set_clinit_barrier_on_entry(true);
1181 }
1182 if (debug_info()->recording_non_safepoints()) {
1183 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1184 (comp_arena(), 8, 0, nullptr));
1185 set_default_node_notes(Node_Notes::make(this));
1186 }
1187
1188 const int grow_ats = 16;
1189 _max_alias_types = grow_ats;
1190 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1191 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1192 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1193 {
1194 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1195 }
1196 // Initialize the first few types.
1197 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1198 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1199 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1200 _num_alias_types = AliasIdxRaw+1;
1201 // Zero out the alias type cache.
1202 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1203 // A null adr_type hits in the cache right away. Preload the right answer.
1204 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1205 }
1206
1207 #ifdef ASSERT
1208 // Verify that the current StartNode is valid.
1209 void Compile::verify_start(StartNode* s) const {
1210 assert(failing_internal() || s == start(), "should be StartNode");
1211 }
1212 #endif
1213
1214 /**
1215 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1216 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1217 * the ideal graph.
1218 */
1219 StartNode* Compile::start() const {
1220 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1221 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1222 Node* start = root()->fast_out(i);
1223 if (start->is_Start()) {
1224 return start->as_Start();
1225 }
1226 }
1227 fatal("Did not find Start node!");
1228 return nullptr;
1229 }
1230
1231 //-------------------------------immutable_memory-------------------------------------
1232 // Access immutable memory
1233 Node* Compile::immutable_memory() {
1234 if (_immutable_memory != nullptr) {
1235 return _immutable_memory;
1236 }
1237 StartNode* s = start();
1238 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1239 Node *p = s->fast_out(i);
1240 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1241 _immutable_memory = p;
1242 return _immutable_memory;
1243 }
1244 }
1245 ShouldNotReachHere();
1246 return nullptr;
1247 }
1248
1249 //----------------------set_cached_top_node------------------------------------
1250 // Install the cached top node, and make sure Node::is_top works correctly.
1251 void Compile::set_cached_top_node(Node* tn) {
1252 if (tn != nullptr) verify_top(tn);
1253 Node* old_top = _top;
1254 _top = tn;
1255 // Calling Node::setup_is_top allows the nodes the chance to adjust
1256 // their _out arrays.
1257 if (_top != nullptr) _top->setup_is_top();
1258 if (old_top != nullptr) old_top->setup_is_top();
1259 assert(_top == nullptr || top()->is_top(), "");
1260 }
1261
1262 #ifdef ASSERT
1263 uint Compile::count_live_nodes_by_graph_walk() {
1264 Unique_Node_List useful(comp_arena());
1265 // Get useful node list by walking the graph.
1266 identify_useful_nodes(useful);
1267 return useful.size();
1268 }
1269
1270 void Compile::print_missing_nodes() {
1271
1272 // Return if CompileLog is null and PrintIdealNodeCount is false.
1273 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1274 return;
1275 }
1276
1277 // This is an expensive function. It is executed only when the user
1278 // specifies VerifyIdealNodeCount option or otherwise knows the
1279 // additional work that needs to be done to identify reachable nodes
1280 // by walking the flow graph and find the missing ones using
1281 // _dead_node_list.
1282
1283 Unique_Node_List useful(comp_arena());
1284 // Get useful node list by walking the graph.
1285 identify_useful_nodes(useful);
1286
1287 uint l_nodes = C->live_nodes();
1288 uint l_nodes_by_walk = useful.size();
1289
1290 if (l_nodes != l_nodes_by_walk) {
1291 if (_log != nullptr) {
1292 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1293 _log->stamp();
1294 _log->end_head();
1295 }
1296 VectorSet& useful_member_set = useful.member_set();
1297 int last_idx = l_nodes_by_walk;
1298 for (int i = 0; i < last_idx; i++) {
1299 if (useful_member_set.test(i)) {
1300 if (_dead_node_list.test(i)) {
1301 if (_log != nullptr) {
1302 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1303 }
1304 if (PrintIdealNodeCount) {
1305 // Print the log message to tty
1306 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1307 useful.at(i)->dump();
1308 }
1309 }
1310 }
1311 else if (! _dead_node_list.test(i)) {
1312 if (_log != nullptr) {
1313 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1314 }
1315 if (PrintIdealNodeCount) {
1316 // Print the log message to tty
1317 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1318 }
1319 }
1320 }
1321 if (_log != nullptr) {
1322 _log->tail("mismatched_nodes");
1323 }
1324 }
1325 }
1326 void Compile::record_modified_node(Node* n) {
1327 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1328 _modified_nodes->push(n);
1329 }
1330 }
1331
1332 void Compile::remove_modified_node(Node* n) {
1333 if (_modified_nodes != nullptr) {
1334 _modified_nodes->remove(n);
1335 }
1336 }
1337 #endif
1338
1339 #ifndef PRODUCT
1340 void Compile::verify_top(Node* tn) const {
1341 if (tn != nullptr) {
1342 assert(tn->is_Con(), "top node must be a constant");
1343 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1344 assert(tn->in(0) != nullptr, "must have live top node");
1345 }
1346 }
1347 #endif
1348
1349
1350 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1351
1352 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1353 guarantee(arr != nullptr, "");
1354 int num_blocks = arr->length();
1355 if (grow_by < num_blocks) grow_by = num_blocks;
1356 int num_notes = grow_by * _node_notes_block_size;
1357 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1358 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1359 while (num_notes > 0) {
1360 arr->append(notes);
1361 notes += _node_notes_block_size;
1362 num_notes -= _node_notes_block_size;
1363 }
1364 assert(num_notes == 0, "exact multiple, please");
1365 }
1366
1367 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1368 if (source == nullptr || dest == nullptr) return false;
1369
1370 if (dest->is_Con())
1371 return false; // Do not push debug info onto constants.
1372
1373 #ifdef ASSERT
1374 // Leave a bread crumb trail pointing to the original node:
1375 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1376 dest->set_debug_orig(source);
1377 }
1378 #endif
1379
1380 if (node_note_array() == nullptr)
1381 return false; // Not collecting any notes now.
1382
1383 // This is a copy onto a pre-existing node, which may already have notes.
1384 // If both nodes have notes, do not overwrite any pre-existing notes.
1385 Node_Notes* source_notes = node_notes_at(source->_idx);
1386 if (source_notes == nullptr || source_notes->is_clear()) return false;
1387 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1388 if (dest_notes == nullptr || dest_notes->is_clear()) {
1389 return set_node_notes_at(dest->_idx, source_notes);
1390 }
1391
1392 Node_Notes merged_notes = (*source_notes);
1393 // The order of operations here ensures that dest notes will win...
1394 merged_notes.update_from(dest_notes);
1395 return set_node_notes_at(dest->_idx, &merged_notes);
1396 }
1397
1398
1399 //--------------------------allow_range_check_smearing-------------------------
1400 // Gating condition for coalescing similar range checks.
1401 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1402 // single covering check that is at least as strong as any of them.
1403 // If the optimization succeeds, the simplified (strengthened) range check
1404 // will always succeed. If it fails, we will deopt, and then give up
1405 // on the optimization.
1406 bool Compile::allow_range_check_smearing() const {
1407 // If this method has already thrown a range-check,
1408 // assume it was because we already tried range smearing
1409 // and it failed.
1410 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1411 return !already_trapped;
1412 }
1413
1414
1415 //------------------------------flatten_alias_type-----------------------------
1416 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1417 assert(do_aliasing(), "Aliasing should be enabled");
1418 int offset = tj->offset();
1419 TypePtr::PTR ptr = tj->ptr();
1420
1421 // Known instance (scalarizable allocation) alias only with itself.
1422 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1423 tj->is_oopptr()->is_known_instance();
1424
1425 // Process weird unsafe references.
1426 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1427 assert(InlineUnsafeOps || StressReflectiveCode || UseAcmpFastPath, "indeterminate pointers come only from unsafe ops");
1428 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1429 tj = TypeOopPtr::BOTTOM;
1430 ptr = tj->ptr();
1431 offset = tj->offset();
1432 }
1433
1434 // Array pointers need some flattening
1435 const TypeAryPtr* ta = tj->isa_aryptr();
1436 if( ta && is_known_inst ) {
1437 if ( offset != Type::OffsetBot &&
1438 offset > arrayOopDesc::length_offset_in_bytes() ) {
1439 offset = Type::OffsetBot; // Flatten constant access into array body only
1440 tj = ta = ta->
1441 remove_speculative()->
1442 cast_to_ptr_type(ptr)->
1443 with_offset(offset);
1444 }
1445 } else if (ta != nullptr) {
1446 // Common slices
1447 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1448 return TypeAryPtr::RANGE;
1449 } else if (offset == oopDesc::klass_offset_in_bytes()) {
1450 return TypeInstPtr::KLASS;
1451 } else if (offset == oopDesc::mark_offset_in_bytes()) {
1452 return TypeInstPtr::MARK;
1453 }
1454
1455 // Remove size and stability
1456 const TypeAry* normalized_ary = TypeAry::make(ta->elem(), TypeInt::POS, false, ta->is_flat(), ta->is_not_flat(), ta->is_not_null_free(), ta->is_atomic());
1457 // Remove ptr, const_oop, and offset
1458 if (ta->elem() == Type::BOTTOM) {
1459 // Bottom array (meet of int[] and byte[] for example), accesses to it will be done with
1460 // Unsafe. This should alias with all arrays. For now just leave it as it is (this is
1461 // incorrect, see JDK-8331133).
1462 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, false, Type::Offset::bottom);
1463 } else if (ta->elem()->make_oopptr() != nullptr) {
1464 // Object arrays, keep field_offset
1465 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, ta->klass_is_exact(), Type::Offset::bottom, Type::Offset(ta->field_offset()));
1466 } else {
1467 // Primitive arrays
1468 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, ta->exact_klass(), true, Type::Offset::bottom);
1469 }
1470
1471 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1472 // cannot be distinguished by bytecode alone.
1473 if (ta->elem() == TypeInt::BOOL) {
1474 tj = ta = TypeAryPtr::BYTES;
1475 }
1476
1477 // All arrays of references share the same slice
1478 if (!ta->is_flat() && ta->elem()->make_oopptr() != nullptr) {
1479 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, TypeInt::POS, false, false, true, true, true);
1480 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, tary, nullptr, false, Type::Offset::bottom);
1481 }
1482
1483 if (ta->is_flat()) {
1484 if (_flat_accesses_share_alias) {
1485 // Initially all flattened array accesses share a single slice
1486 tj = ta = TypeAryPtr::INLINES;
1487 } else {
1488 // Flat accesses are always exact
1489 tj = ta = ta->cast_to_exactness(true);
1490 }
1491 }
1492 }
1493
1494 // Oop pointers need some flattening
1495 const TypeInstPtr *to = tj->isa_instptr();
1496 if (to && to != TypeOopPtr::BOTTOM) {
1497 ciInstanceKlass* ik = to->instance_klass();
1498 tj = to = to->cast_to_maybe_flat_in_array(); // flatten to maybe flat in array
1499 if( ptr == TypePtr::Constant ) {
1500 if (ik != ciEnv::current()->Class_klass() ||
1501 offset < ik->layout_helper_size_in_bytes()) {
1502 // No constant oop pointers (such as Strings); they alias with
1503 // unknown strings.
1504 assert(!is_known_inst, "not scalarizable allocation");
1505 tj = to = to->
1506 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1507 remove_speculative()->
1508 cast_to_ptr_type(TypePtr::BotPTR)->
1509 cast_to_exactness(false);
1510 }
1511 } else if( is_known_inst ) {
1512 tj = to; // Keep NotNull and klass_is_exact for instance type
1513 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1514 // During the 2nd round of IterGVN, NotNull castings are removed.
1515 // Make sure the Bottom and NotNull variants alias the same.
1516 // Also, make sure exact and non-exact variants alias the same.
1517 tj = to = to->
1518 remove_speculative()->
1519 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1520 cast_to_ptr_type(TypePtr::BotPTR)->
1521 cast_to_exactness(false);
1522 }
1523 if (to->speculative() != nullptr) {
1524 tj = to = to->remove_speculative();
1525 }
1526 // Canonicalize the holder of this field
1527 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1528 // First handle header references such as a LoadKlassNode, even if the
1529 // object's klass is unloaded at compile time (4965979).
1530 if (!is_known_inst) { // Do it only for non-instance types
1531 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1532 }
1533 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1534 // Static fields are in the space above the normal instance
1535 // fields in the java.lang.Class instance.
1536 if (ik != ciEnv::current()->Class_klass()) {
1537 to = nullptr;
1538 tj = TypeOopPtr::BOTTOM;
1539 offset = tj->offset();
1540 }
1541 } else {
1542 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1543 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1544 assert(tj->offset() == offset, "no change to offset expected");
1545 bool xk = to->klass_is_exact();
1546 int instance_id = to->instance_id();
1547
1548 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1549 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1550 // its interfaces are included.
1551 if (xk && ik->equals(canonical_holder)) {
1552 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id,
1553 TypePtr::MaybeFlat), "exact type should be canonical type");
1554 } else {
1555 assert(xk || !is_known_inst, "Known instance should be exact type");
1556 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id,
1557 TypePtr::MaybeFlat);
1558 }
1559 }
1560 }
1561
1562 // Klass pointers to object array klasses need some flattening
1563 const TypeKlassPtr *tk = tj->isa_klassptr();
1564 if( tk ) {
1565 // If we are referencing a field within a Klass, we need
1566 // to assume the worst case of an Object. Both exact and
1567 // inexact types must flatten to the same alias class so
1568 // use NotNull as the PTR.
1569 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1570 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1571 env()->Object_klass(),
1572 Type::Offset(offset),
1573 TypePtr::MaybeFlat);
1574 }
1575
1576 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1577 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1578 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1579 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset), TypePtr::MaybeFlat);
1580 } else {
1581 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(), tk->is_atomic(), tk->is_aryklassptr()->is_refined_type());
1582 }
1583 }
1584 // Check for precise loads from the primary supertype array and force them
1585 // to the supertype cache alias index. Check for generic array loads from
1586 // the primary supertype array and also force them to the supertype cache
1587 // alias index. Since the same load can reach both, we need to merge
1588 // these 2 disparate memories into the same alias class. Since the
1589 // primary supertype array is read-only, there's no chance of confusion
1590 // where we bypass an array load and an array store.
1591 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1592 if (offset == Type::OffsetBot ||
1593 (offset >= primary_supers_offset &&
1594 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1595 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1596 offset = in_bytes(Klass::secondary_super_cache_offset());
1597 tj = tk = tk->with_offset(offset);
1598 }
1599 }
1600
1601 // Flatten all Raw pointers together.
1602 if (tj->base() == Type::RawPtr)
1603 tj = TypeRawPtr::BOTTOM;
1604
1605 if (tj->base() == Type::AnyPtr)
1606 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1607
1608 offset = tj->offset();
1609 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1610
1611 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1612 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1613 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1614 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1615 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1616 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1617 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1618 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1619 assert( tj->ptr() != TypePtr::TopPTR &&
1620 tj->ptr() != TypePtr::AnyNull &&
1621 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1622 // assert( tj->ptr() != TypePtr::Constant ||
1623 // tj->base() == Type::RawPtr ||
1624 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1625
1626 return tj;
1627 }
1628
1629 void Compile::AliasType::Init(int i, const TypePtr* at) {
1630 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1631 _index = i;
1632 _adr_type = at;
1633 _field = nullptr;
1634 _element = nullptr;
1635 _is_rewritable = true; // default
1636 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1637 if (atoop != nullptr && atoop->is_known_instance()) {
1638 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1639 _general_index = Compile::current()->get_alias_index(gt);
1640 } else {
1641 _general_index = 0;
1642 }
1643 }
1644
1645 BasicType Compile::AliasType::basic_type() const {
1646 if (element() != nullptr) {
1647 const Type* element = adr_type()->is_aryptr()->elem();
1648 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1649 } if (field() != nullptr) {
1650 return field()->layout_type();
1651 } else {
1652 return T_ILLEGAL; // unknown
1653 }
1654 }
1655
1656 //---------------------------------print_on------------------------------------
1657 #ifndef PRODUCT
1658 void Compile::AliasType::print_on(outputStream* st) {
1659 if (index() < 10)
1660 st->print("@ <%d> ", index());
1661 else st->print("@ <%d>", index());
1662 st->print(is_rewritable() ? " " : " RO");
1663 int offset = adr_type()->offset();
1664 if (offset == Type::OffsetBot)
1665 st->print(" +any");
1666 else st->print(" +%-3d", offset);
1667 st->print(" in ");
1668 adr_type()->dump_on(st);
1669 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1670 if (field() != nullptr && tjp) {
1671 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1672 tjp->offset() != field()->offset_in_bytes()) {
1673 st->print(" != ");
1674 field()->print();
1675 st->print(" ***");
1676 }
1677 }
1678 }
1679
1680 void print_alias_types() {
1681 Compile* C = Compile::current();
1682 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1683 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1684 C->alias_type(idx)->print_on(tty);
1685 tty->cr();
1686 }
1687 }
1688 #endif
1689
1690
1691 //----------------------------probe_alias_cache--------------------------------
1692 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1693 intptr_t key = (intptr_t) adr_type;
1694 key ^= key >> logAliasCacheSize;
1695 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1696 }
1697
1698
1699 //-----------------------------grow_alias_types--------------------------------
1700 void Compile::grow_alias_types() {
1701 const int old_ats = _max_alias_types; // how many before?
1702 const int new_ats = old_ats; // how many more?
1703 const int grow_ats = old_ats+new_ats; // how many now?
1704 _max_alias_types = grow_ats;
1705 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), _alias_types, old_ats, grow_ats);
1706 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1707 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1708 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1709 }
1710
1711
1712 //--------------------------------find_alias_type------------------------------
1713 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1714 if (!do_aliasing()) {
1715 return alias_type(AliasIdxBot);
1716 }
1717
1718 AliasCacheEntry* ace = nullptr;
1719 if (!uncached) {
1720 ace = probe_alias_cache(adr_type);
1721 if (ace->_adr_type == adr_type) {
1722 return alias_type(ace->_index);
1723 }
1724 }
1725
1726 // Handle special cases.
1727 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1728 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1729
1730 // Do it the slow way.
1731 const TypePtr* flat = flatten_alias_type(adr_type);
1732
1733 #ifdef ASSERT
1734 {
1735 ResourceMark rm;
1736 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1737 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1738 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1739 Type::str(adr_type));
1740 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1741 const TypeOopPtr* foop = flat->is_oopptr();
1742 // Scalarizable allocations have exact klass always.
1743 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1744 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1745 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1746 Type::str(foop), Type::str(xoop));
1747 }
1748 }
1749 #endif
1750
1751 int idx = AliasIdxTop;
1752 for (int i = 0; i < num_alias_types(); i++) {
1753 if (alias_type(i)->adr_type() == flat) {
1754 idx = i;
1755 break;
1756 }
1757 }
1758
1759 if (idx == AliasIdxTop) {
1760 if (no_create) return nullptr;
1761 // Grow the array if necessary.
1762 if (_num_alias_types == _max_alias_types) grow_alias_types();
1763 // Add a new alias type.
1764 idx = _num_alias_types++;
1765 _alias_types[idx]->Init(idx, flat);
1766 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1767 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1768 if (flat->isa_instptr()) {
1769 if (flat->offset() == java_lang_Class::klass_offset()
1770 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1771 alias_type(idx)->set_rewritable(false);
1772 }
1773 ciField* field = nullptr;
1774 if (flat->isa_aryptr()) {
1775 #ifdef ASSERT
1776 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1777 // (T_BYTE has the weakest alignment and size restrictions...)
1778 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1779 #endif
1780 const Type* elemtype = flat->is_aryptr()->elem();
1781 if (flat->offset() == TypePtr::OffsetBot) {
1782 alias_type(idx)->set_element(elemtype);
1783 }
1784 int field_offset = flat->is_aryptr()->field_offset().get();
1785 if (flat->is_flat() &&
1786 field_offset != Type::OffsetBot) {
1787 ciInlineKlass* vk = elemtype->inline_klass();
1788 field_offset += vk->payload_offset();
1789 field = vk->get_field_by_offset(field_offset, false);
1790 }
1791 }
1792 if (flat->isa_klassptr()) {
1793 if (UseCompactObjectHeaders) {
1794 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1795 alias_type(idx)->set_rewritable(false);
1796 }
1797 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1798 alias_type(idx)->set_rewritable(false);
1799 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1800 alias_type(idx)->set_rewritable(false);
1801 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1802 alias_type(idx)->set_rewritable(false);
1803 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1804 alias_type(idx)->set_rewritable(false);
1805 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1806 alias_type(idx)->set_rewritable(false);
1807 }
1808
1809 if (flat->isa_instklassptr()) {
1810 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1811 alias_type(idx)->set_rewritable(false);
1812 }
1813 }
1814 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1815 // but the base pointer type is not distinctive enough to identify
1816 // references into JavaThread.)
1817
1818 // Check for final fields.
1819 const TypeInstPtr* tinst = flat->isa_instptr();
1820 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1821 if (tinst->const_oop() != nullptr &&
1822 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1823 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1824 // static field
1825 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1826 field = k->get_field_by_offset(tinst->offset(), true);
1827 } else if (tinst->is_inlinetypeptr()) {
1828 // Inline type field
1829 ciInlineKlass* vk = tinst->inline_klass();
1830 field = vk->get_field_by_offset(tinst->offset(), false);
1831 } else {
1832 ciInstanceKlass *k = tinst->instance_klass();
1833 field = k->get_field_by_offset(tinst->offset(), false);
1834 }
1835 }
1836 assert(field == nullptr ||
1837 original_field == nullptr ||
1838 (field->holder() == original_field->holder() &&
1839 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1840 field->is_static() == original_field->is_static()), "wrong field?");
1841 // Set field() and is_rewritable() attributes.
1842 if (field != nullptr) {
1843 alias_type(idx)->set_field(field);
1844 if (flat->isa_aryptr()) {
1845 // Fields of flat arrays are rewritable although they are declared final
1846 assert(flat->is_flat(), "must be a flat array");
1847 alias_type(idx)->set_rewritable(true);
1848 }
1849 }
1850 }
1851
1852 // Fill the cache for next time.
1853 if (!uncached) {
1854 ace->_adr_type = adr_type;
1855 ace->_index = idx;
1856 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1857
1858 // Might as well try to fill the cache for the flattened version, too.
1859 AliasCacheEntry* face = probe_alias_cache(flat);
1860 if (face->_adr_type == nullptr) {
1861 face->_adr_type = flat;
1862 face->_index = idx;
1863 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1864 }
1865 }
1866
1867 return alias_type(idx);
1868 }
1869
1870
1871 Compile::AliasType* Compile::alias_type(ciField* field) {
1872 const TypeOopPtr* t;
1873 if (field->is_static())
1874 t = TypeInstPtr::make(field->holder()->java_mirror());
1875 else
1876 t = TypeOopPtr::make_from_klass_raw(field->holder());
1877 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1878 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1879 return atp;
1880 }
1881
1882
1883 //------------------------------have_alias_type--------------------------------
1884 bool Compile::have_alias_type(const TypePtr* adr_type) {
1885 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1886 if (ace->_adr_type == adr_type) {
1887 return true;
1888 }
1889
1890 // Handle special cases.
1891 if (adr_type == nullptr) return true;
1892 if (adr_type == TypePtr::BOTTOM) return true;
1893
1894 return find_alias_type(adr_type, true, nullptr) != nullptr;
1895 }
1896
1897 //-----------------------------must_alias--------------------------------------
1898 // True if all values of the given address type are in the given alias category.
1899 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1900 if (alias_idx == AliasIdxBot) return true; // the universal category
1901 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1902 if (alias_idx == AliasIdxTop) return false; // the empty category
1903 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1904
1905 // the only remaining possible overlap is identity
1906 int adr_idx = get_alias_index(adr_type);
1907 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1908 assert(adr_idx == alias_idx ||
1909 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1910 && adr_type != TypeOopPtr::BOTTOM),
1911 "should not be testing for overlap with an unsafe pointer");
1912 return adr_idx == alias_idx;
1913 }
1914
1915 //------------------------------can_alias--------------------------------------
1916 // True if any values of the given address type are in the given alias category.
1917 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1918 if (alias_idx == AliasIdxTop) return false; // the empty category
1919 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1920 // Known instance doesn't alias with bottom memory
1921 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1922 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1923
1924 // the only remaining possible overlap is identity
1925 int adr_idx = get_alias_index(adr_type);
1926 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1927 return adr_idx == alias_idx;
1928 }
1929
1930 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1931 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1932 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1933 if (parse_predicate_count() == 0) {
1934 return;
1935 }
1936 for (int i = 0; i < parse_predicate_count(); i++) {
1937 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1938 parse_predicate->mark_useless(igvn);
1939 }
1940 _parse_predicates.clear();
1941 }
1942
1943 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1944 if (!n->for_post_loop_opts_igvn()) {
1945 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1946 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1947 _for_post_loop_igvn.append(n);
1948 }
1949 }
1950
1951 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1952 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1953 _for_post_loop_igvn.remove(n);
1954 }
1955
1956 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1957 // Verify that all previous optimizations produced a valid graph
1958 // at least to this point, even if no loop optimizations were done.
1959 PhaseIdealLoop::verify(igvn);
1960
1961 if (_print_phase_loop_opts) {
1962 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1963 }
1964 C->set_post_loop_opts_phase(); // no more loop opts allowed
1965
1966 assert(!C->major_progress(), "not cleared");
1967
1968 if (_for_post_loop_igvn.length() > 0) {
1969 while (_for_post_loop_igvn.length() > 0) {
1970 Node* n = _for_post_loop_igvn.pop();
1971 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1972 igvn._worklist.push(n);
1973 }
1974 igvn.optimize();
1975 if (failing()) return;
1976 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1977 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1978
1979 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1980 if (C->major_progress()) {
1981 C->clear_major_progress(); // ensure that major progress is now clear
1982 }
1983 }
1984 }
1985
1986 void Compile::add_inline_type(Node* n) {
1987 assert(n->is_InlineType(), "unexpected node");
1988 _inline_type_nodes.push(n);
1989 }
1990
1991 void Compile::remove_inline_type(Node* n) {
1992 assert(n->is_InlineType(), "unexpected node");
1993 if (_inline_type_nodes.contains(n)) {
1994 _inline_type_nodes.remove(n);
1995 }
1996 }
1997
1998 // Does the return value keep otherwise useless inline type allocations alive?
1999 static bool return_val_keeps_allocations_alive(Node* ret_val) {
2000 ResourceMark rm;
2001 Unique_Node_List wq;
2002 wq.push(ret_val);
2003 bool some_allocations = false;
2004 for (uint i = 0; i < wq.size(); i++) {
2005 Node* n = wq.at(i);
2006 if (n->outcnt() > 1) {
2007 // Some other use for the allocation
2008 return false;
2009 } else if (n->is_InlineType()) {
2010 wq.push(n->in(1));
2011 } else if (n->is_Phi()) {
2012 for (uint j = 1; j < n->req(); j++) {
2013 wq.push(n->in(j));
2014 }
2015 } else if (n->is_CheckCastPP() &&
2016 n->in(1)->is_Proj() &&
2017 n->in(1)->in(0)->is_Allocate()) {
2018 some_allocations = true;
2019 } else if (n->is_CheckCastPP() || n->is_CastPP()) {
2020 wq.push(n->in(1));
2021 }
2022 }
2023 return some_allocations;
2024 }
2025
2026 bool Compile::clear_argument_if_only_used_as_buffer_at_calls(Node* result_cast, PhaseIterGVN& igvn) {
2027 ResourceMark rm;
2028 Unique_Node_List wq;
2029 wq.push(result_cast);
2030 Node_List calls;
2031 for (uint i = 0; i < wq.size(); ++i) {
2032 Node* n = wq.at(i);
2033 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2034 Node* u = n->fast_out(j);
2035 if (u->is_Phi()) {
2036 wq.push(u);
2037 } else if (u->is_InlineType() && u->as_InlineType()->get_oop() == n) {
2038 wq.push(u);
2039 } else if (u->is_CallJava()) {
2040 CallJavaNode* call = u->as_CallJava();
2041 if (call->method() != nullptr && call->method()->mismatch()) {
2042 return false;
2043 }
2044 uint nargs = call->tf()->domain_cc()->cnt();
2045 for (uint k = TypeFunc::Parms; k < nargs; k++) {
2046 Node* in = call->in(k);
2047 if (in == n && (call->method() == nullptr || !call->method()->is_scalarized_buffer_arg(k - TypeFunc::Parms))) {
2048 return false;
2049 }
2050 }
2051 calls.push(call);
2052 } else if (u->Opcode() == Op_EncodeP) {
2053 wq.push(u);
2054 } else if (u->is_AddP()) {
2055 wq.push(u);
2056 } else if (u->is_Store() && u->in(MemNode::Address) == n) {
2057 // storing to the buffer is fine
2058 } else if (u->is_SafePoint()) {
2059 SafePointNode* sfpt = u->as_SafePoint();
2060 int input = u->find_edge(n);
2061 JVMState* jvms = sfpt->jvms();
2062 if (jvms != nullptr) {
2063 if (input < (int)jvms->debug_start()) {
2064 return false;
2065 }
2066 }
2067 } else {
2068 return false;
2069 }
2070 }
2071 }
2072 for (uint i = 0; i < calls.size(); ++i) {
2073 CallJavaNode* call = calls.at(i)->as_CallJava();
2074 uint nargs = call->tf()->domain_cc()->cnt();
2075 for (uint k = TypeFunc::Parms; k < nargs; k++) {
2076 Node* in = call->in(k);
2077 if (wq.member(in)) {
2078 assert(call->method()->is_scalarized_buffer_arg(k - TypeFunc::Parms), "only buffer argument removed here");
2079 igvn.replace_input_of(call, k, igvn.zerocon(T_OBJECT));
2080 }
2081 }
2082 }
2083 return true;
2084 }
2085
2086 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
2087 // Make sure that the return value does not keep an otherwise unused allocation alive
2088 if (tf()->returns_inline_type_as_fields()) {
2089 Node* ret = nullptr;
2090 for (uint i = 1; i < root()->req(); i++) {
2091 Node* in = root()->in(i);
2092 if (in->Opcode() == Op_Return) {
2093 assert(ret == nullptr, "only one return");
2094 ret = in;
2095 }
2096 }
2097 if (ret != nullptr) {
2098 Node* ret_val = ret->in(TypeFunc::Parms);
2099 if (igvn.type(ret_val)->isa_oopptr() &&
2100 return_val_keeps_allocations_alive(ret_val)) {
2101 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
2102 assert(ret_val->outcnt() == 0, "should be dead now");
2103 igvn.remove_dead_node(ret_val, PhaseIterGVN::NodeOrigin::Graph);
2104 }
2105 }
2106 }
2107 // if a newly allocated object is a value that's only passed as argument to calls as (possibly null) buffers, then
2108 // clear the call argument inputs so the allocation node can be removed
2109 for (int i = 0; i < C->macro_count(); ++i) {
2110 Node* macro_node = C->macro_node(i);
2111 if (macro_node->Opcode() == Op_Allocate) {
2112 AllocateNode* allocate = macro_node->as_Allocate();
2113 Node* result_cast = allocate->result_cast();
2114 if (result_cast != nullptr) {
2115 const Type* result_type = igvn.type(result_cast);
2116 if (result_type->is_inlinetypeptr()) {
2117 clear_argument_if_only_used_as_buffer_at_calls(result_cast, igvn);
2118 }
2119 }
2120 }
2121 }
2122
2123 if (_inline_type_nodes.length() == 0) {
2124 // keep the graph canonical
2125 igvn.optimize();
2126 return;
2127 }
2128 // Scalarize inline types in safepoint debug info.
2129 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2130 set_scalarize_in_safepoints(true);
2131 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2132 InlineTypeNode* vt = _inline_type_nodes.at(i)->as_InlineType();
2133 vt->make_scalar_in_safepoints(&igvn);
2134 igvn.record_for_igvn(vt);
2135 }
2136 if (remove) {
2137 // Remove inline type nodes by replacing them with their oop input
2138 while (_inline_type_nodes.length() > 0) {
2139 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2140 if (vt->outcnt() == 0) {
2141 igvn.remove_dead_node(vt, PhaseIterGVN::NodeOrigin::Graph);
2142 continue;
2143 }
2144 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2145 DEBUG_ONLY(bool must_be_buffered = false);
2146 Node* u = vt->out(i);
2147 // Check if any users are blackholes. If so, rewrite them to use either the
2148 // allocated buffer, or individual components, instead of the inline type node
2149 // that goes away.
2150 if (u->is_Blackhole()) {
2151 BlackholeNode* bh = u->as_Blackhole();
2152
2153 // Unlink the old input
2154 int idx = bh->find_edge(vt);
2155 assert(idx != -1, "The edge should be there");
2156 bh->del_req(idx);
2157 --i;
2158
2159 if (vt->is_allocated(&igvn)) {
2160 // Already has the allocated instance, blackhole that
2161 bh->add_req(vt->get_oop());
2162 } else {
2163 // Not allocated yet, blackhole the components
2164 for (uint c = 0; c < vt->field_count(); c++) {
2165 bh->add_req(vt->field_value(c));
2166 }
2167 }
2168
2169 // Node modified, record for IGVN
2170 igvn.record_for_igvn(bh);
2171 }
2172 #ifdef ASSERT
2173 // Verify that inline type is buffered when replacing by oop
2174 else if (u->is_InlineType()) {
2175 // InlineType uses don't need buffering because they are about to be replaced as well
2176 } else {
2177 must_be_buffered = true;
2178 }
2179 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2180 vt->dump(0);
2181 u->dump(0);
2182 assert(false, "Should have been buffered");
2183 }
2184 #endif
2185 }
2186 igvn.replace_node(vt, vt->get_oop());
2187 }
2188 }
2189 igvn.optimize();
2190 }
2191
2192 void Compile::add_flat_access(Node* n) {
2193 assert(n != nullptr && (n->Opcode() == Op_LoadFlat || n->Opcode() == Op_StoreFlat), "unexpected node %s", n == nullptr ? "nullptr" : n->Name());
2194 assert(!_flat_access_nodes.contains(n), "duplicate insertion");
2195 _flat_access_nodes.push(n);
2196 }
2197
2198 void Compile::remove_flat_access(Node* n) {
2199 assert(n != nullptr && (n->Opcode() == Op_LoadFlat || n->Opcode() == Op_StoreFlat), "unexpected node %s", n == nullptr ? "nullptr" : n->Name());
2200 _flat_access_nodes.remove_if_existing(n);
2201 }
2202
2203 void Compile::process_flat_accesses(PhaseIterGVN& igvn) {
2204 assert(igvn._worklist.size() == 0, "should be empty");
2205 igvn.set_delay_transform(true);
2206 for (int i = _flat_access_nodes.length() - 1; i >= 0; i--) {
2207 Node* n = _flat_access_nodes.at(i);
2208 assert(n != nullptr, "unexpected nullptr");
2209 if (n->is_LoadFlat()) {
2210 LoadFlatNode* loadn = n->as_LoadFlat();
2211 // Expending a flat load atomically means that we get a chunk of memory spanning multiple fields
2212 // that we chop with bitwise operations. That is too subtle for some optimizations, especially
2213 // constant folding when fields are constant. If we can get a constant object from which we are
2214 // flat-loading, we can simply replace the loads at compilation-time by the field of the constant
2215 // object.
2216 ciInstance* loaded_from = nullptr;
2217 if (FoldStableValues) {
2218 const TypeOopPtr* base_type = igvn.type(loadn->base())->is_oopptr();
2219 ciObject* oop = base_type->const_oop();
2220 int off = igvn.type(loadn->ptr())->isa_ptr()->offset();
2221
2222 if (oop != nullptr && oop->is_instance()) {
2223 ciInstance* holder = oop->as_instance();
2224 ciKlass* klass = holder->klass();
2225 ciInstanceKlass* iklass = klass->as_instance_klass();
2226 ciField* field = iklass->get_non_flat_field_by_offset(off);
2227
2228 if (field->is_stable()) {
2229 ciConstant fv = holder->field_value(field);
2230 if (is_reference_type(fv.basic_type()) && fv.as_object()->is_instance()) {
2231 // The field value is an object, not null. We can use stability.
2232 loaded_from = fv.as_object()->as_instance();
2233 }
2234 }
2235 } else if (oop != nullptr && oop->is_array() && off != Type::OffsetBot) {
2236 ciArray* array = oop->as_array();
2237 ciConstant elt = array->element_value_by_offset(off);
2238 const TypeAryPtr* aryptr = base_type->is_aryptr();
2239 if (aryptr->is_stable() && aryptr->is_atomic() && is_reference_type(elt.basic_type()) && elt.as_object()->is_instance()) {
2240 loaded_from = elt.as_object()->as_instance();
2241 }
2242 }
2243 }
2244
2245 if (loaded_from != nullptr) {
2246 loadn->expand_constant(igvn, loaded_from);
2247 } else {
2248 loadn->expand_atomic(igvn);
2249 }
2250 } else {
2251 n->as_StoreFlat()->expand_atomic(igvn);
2252 }
2253 }
2254 _flat_access_nodes.clear_and_deallocate();
2255 igvn.set_delay_transform(false);
2256 igvn.optimize();
2257 }
2258
2259 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2260 DEBUG_ONLY(igvn.verify_empty_worklist(nullptr));
2261 if (!_has_flat_accesses) {
2262 return;
2263 }
2264 // Initially, all flat array accesses share the same slice to
2265 // keep dependencies with Object[] array accesses (that could be
2266 // to a flat array) correct. We're done with parsing so we
2267 // now know all flat array accesses in this compile
2268 // unit. Let's move flat array accesses to their own slice,
2269 // one per element field. This should help memory access
2270 // optimizations.
2271 ResourceMark rm;
2272 Unique_Node_List wq;
2273 wq.push(root());
2274
2275 Node_List mergememnodes;
2276 Node_List memnodes;
2277
2278 // Alias index currently shared by all flat memory accesses
2279 int index = get_alias_index(TypeAryPtr::INLINES);
2280
2281 // Find MergeMem nodes and flat array accesses
2282 for (uint i = 0; i < wq.size(); i++) {
2283 Node* n = wq.at(i);
2284 if (n->is_Mem()) {
2285 const TypePtr* adr_type = nullptr;
2286 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2287 if (adr_type == TypeAryPtr::INLINES) {
2288 memnodes.push(n);
2289 }
2290 } else if (n->is_MergeMem()) {
2291 MergeMemNode* mm = n->as_MergeMem();
2292 if (mm->memory_at(index) != mm->base_memory()) {
2293 mergememnodes.push(n);
2294 }
2295 }
2296 for (uint j = 0; j < n->req(); j++) {
2297 Node* m = n->in(j);
2298 if (m != nullptr) {
2299 wq.push(m);
2300 }
2301 }
2302 }
2303
2304 _flat_accesses_share_alias = false;
2305
2306 // We are going to change the slice for the flat array
2307 // accesses so we need to clear the cache entries that refer to
2308 // them.
2309 for (uint i = 0; i < AliasCacheSize; i++) {
2310 AliasCacheEntry* ace = &_alias_cache[i];
2311 if (ace->_adr_type != nullptr &&
2312 ace->_adr_type->is_flat()) {
2313 ace->_adr_type = nullptr;
2314 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2315 }
2316 }
2317
2318 #ifdef ASSERT
2319 for (uint i = 0; i < memnodes.size(); i++) {
2320 Node* m = memnodes.at(i);
2321 const TypePtr* adr_type = m->adr_type();
2322 m->as_Mem()->set_adr_type(adr_type);
2323 }
2324 #endif // ASSERT
2325
2326 int start_alias = num_alias_types(); // Start of new aliases
2327 Node_Stack stack(0);
2328 #ifdef ASSERT
2329 VectorSet seen(Thread::current()->resource_area());
2330 #endif
2331 // Now let's fix the memory graph so each flat array access
2332 // is moved to the right slice. Start from the MergeMem nodes.
2333 uint last = unique();
2334 for (uint i = 0; i < mergememnodes.size(); i++) {
2335 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2336 if (current->outcnt() == 0) {
2337 // This node is killed by a previous iteration
2338 continue;
2339 }
2340
2341 Node* n = current->memory_at(index);
2342 MergeMemNode* mm = nullptr;
2343 do {
2344 // Follow memory edges through memory accesses, phis and
2345 // narrow membars and push nodes on the stack. Once we hit
2346 // bottom memory, we pop element off the stack one at a
2347 // time, in reverse order, and move them to the right slice
2348 // by changing their memory edges.
2349 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() ||
2350 (n->adr_type() == TypeAryPtr::INLINES && !n->is_NarrowMemProj())) {
2351 assert(!seen.test_set(n->_idx), "");
2352 // Uses (a load for instance) will need to be moved to the
2353 // right slice as well and will get a new memory state
2354 // that we don't know yet. The use could also be the
2355 // backedge of a loop. We put a place holder node between
2356 // the memory node and its uses. We replace that place
2357 // holder with the correct memory state once we know it,
2358 // i.e. when nodes are popped off the stack. Using the
2359 // place holder make the logic work in the presence of
2360 // loops.
2361 if (n->outcnt() > 1) {
2362 Node* place_holder = nullptr;
2363 assert(!n->has_out_with(Op_Node), "");
2364 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2365 Node* u = n->out(k);
2366 if (u != current && u->_idx < last) {
2367 bool success = false;
2368 for (uint l = 0; l < u->req(); l++) {
2369 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2370 continue;
2371 }
2372 Node* in = u->in(l);
2373 if (in == n) {
2374 if (place_holder == nullptr) {
2375 place_holder = new Node(1);
2376 place_holder->init_req(0, n);
2377 }
2378 igvn.replace_input_of(u, l, place_holder);
2379 success = true;
2380 }
2381 }
2382 if (success) {
2383 --k;
2384 }
2385 }
2386 }
2387 }
2388 if (n->is_Phi()) {
2389 stack.push(n, 1);
2390 n = n->in(1);
2391 } else if (n->is_Mem()) {
2392 stack.push(n, n->req());
2393 n = n->in(MemNode::Memory);
2394 } else {
2395 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2396 stack.push(n, n->req());
2397 n = n->in(0)->in(TypeFunc::Memory);
2398 }
2399 } else {
2400 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || n->is_NarrowMemProj(), "");
2401 // Build a new MergeMem node to carry the new memory state
2402 // as we build it. IGVN should fold extraneous MergeMem
2403 // nodes.
2404 if (n->is_NarrowMemProj()) {
2405 // We need 1 NarrowMemProj for each slice of this array
2406 InitializeNode* init = n->in(0)->as_Initialize();
2407 AllocateNode* alloc = init->allocation();
2408 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2409 const TypeAryKlassPtr* klass_type = klass_node->bottom_type()->isa_aryklassptr();
2410 assert(klass_type != nullptr, "must be an array");
2411 assert(klass_type->klass_is_exact(), "must be an exact klass");
2412 ciArrayKlass* klass = klass_type->exact_klass()->as_array_klass();
2413 assert(klass->is_flat_array_klass(), "must be a flat array");
2414 ciInlineKlass* elem_klass = klass->element_klass()->as_inline_klass();
2415 const TypeAryPtr* oop_type = klass_type->as_exact_instance_type()->is_aryptr();
2416 assert(oop_type->klass_is_exact(), "must be an exact klass");
2417
2418 Node* base = alloc->in(TypeFunc::Memory);
2419 assert(base->bottom_type() == Type::MEMORY, "the memory input of AllocateNode must be a memory");
2420 assert(base->adr_type() == TypePtr::BOTTOM, "the memory input of AllocateNode must be a bottom memory");
2421 // Must create a MergeMem with base as the base memory, do not clone if base is a
2422 // MergeMem because it may not be processed yet
2423 mm = MergeMemNode::make(nullptr);
2424 mm->set_base_memory(base);
2425 for (int j = 0; j < elem_klass->nof_nonstatic_fields(); j++) {
2426 int field_offset = elem_klass->nonstatic_field_at(j)->offset_in_bytes() - elem_klass->payload_offset();
2427 const TypeAryPtr* field_ptr = oop_type->with_offset(Type::OffsetBot)->with_field_offset(field_offset);
2428 int field_alias_idx = get_alias_index(field_ptr);
2429 assert(field_ptr == get_adr_type(field_alias_idx), "must match");
2430 Node* new_proj = new NarrowMemProjNode(init, field_ptr);
2431 igvn.register_new_node_with_optimizer(new_proj);
2432 mm->set_memory_at(field_alias_idx, new_proj);
2433 }
2434 if (!klass->is_elem_null_free()) {
2435 int nm_offset = elem_klass->null_marker_offset_in_payload();
2436 const TypeAryPtr* nm_ptr = oop_type->with_offset(Type::OffsetBot)->with_field_offset(nm_offset);
2437 int nm_alias_idx = get_alias_index(nm_ptr);
2438 assert(nm_ptr == get_adr_type(nm_alias_idx), "must match");
2439 Node* new_proj = new NarrowMemProjNode(init, nm_ptr);
2440 igvn.register_new_node_with_optimizer(new_proj);
2441 mm->set_memory_at(nm_alias_idx, new_proj);
2442 }
2443
2444 // Replace all uses of the old NarrowMemProj with the correct state
2445 MergeMemNode* new_n = MergeMemNode::make(mm);
2446 igvn.register_new_node_with_optimizer(new_n);
2447 igvn.replace_node(n, new_n);
2448 } else {
2449 // Must create a MergeMem with n as the base memory, do not clone if n is a MergeMem
2450 // because it may not be processed yet
2451 mm = MergeMemNode::make(nullptr);
2452 mm->set_base_memory(n);
2453 }
2454
2455 igvn.register_new_node_with_optimizer(mm);
2456 while (stack.size() > 0) {
2457 Node* m = stack.node();
2458 uint idx = stack.index();
2459 if (m->is_Mem()) {
2460 // Move memory node to its new slice
2461 const TypePtr* adr_type = m->adr_type();
2462 int alias = get_alias_index(adr_type);
2463 Node* prev = mm->memory_at(alias);
2464 igvn.replace_input_of(m, MemNode::Memory, prev);
2465 mm->set_memory_at(alias, m);
2466 } else if (m->is_Phi()) {
2467 // We need as many new phis as there are new aliases
2468 Node* new_phi_in = MergeMemNode::make(mm);
2469 igvn.register_new_node_with_optimizer(new_phi_in);
2470 igvn.replace_input_of(m, idx, new_phi_in);
2471 if (idx == m->req()-1) {
2472 Node* r = m->in(0);
2473 for (int j = start_alias; j < num_alias_types(); j++) {
2474 const TypePtr* adr_type = get_adr_type(j);
2475 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2476 continue;
2477 }
2478 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2479 igvn.register_new_node_with_optimizer(phi);
2480 for (uint k = 1; k < m->req(); k++) {
2481 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2482 }
2483 mm->set_memory_at(j, phi);
2484 }
2485 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2486 igvn.register_new_node_with_optimizer(base_phi);
2487 for (uint k = 1; k < m->req(); k++) {
2488 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2489 }
2490 mm->set_base_memory(base_phi);
2491 }
2492 } else {
2493 // This is a MemBarCPUOrder node from
2494 // Parse::array_load()/Parse::array_store(), in the
2495 // branch that handles flat arrays hidden under
2496 // an Object[] array. We also need one new membar per
2497 // new alias to keep the unknown access that the
2498 // membars protect properly ordered with accesses to
2499 // known flat array.
2500 assert(m->is_Proj(), "projection expected");
2501 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2502 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2503 for (int j = start_alias; j < num_alias_types(); j++) {
2504 const TypePtr* adr_type = get_adr_type(j);
2505 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2506 continue;
2507 }
2508 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2509 igvn.register_new_node_with_optimizer(mb);
2510 Node* mem = mm->memory_at(j);
2511 mb->init_req(TypeFunc::Control, ctrl);
2512 mb->init_req(TypeFunc::Memory, mem);
2513 ctrl = new ProjNode(mb, TypeFunc::Control);
2514 igvn.register_new_node_with_optimizer(ctrl);
2515 mem = new ProjNode(mb, TypeFunc::Memory);
2516 igvn.register_new_node_with_optimizer(mem);
2517 mm->set_memory_at(j, mem);
2518 }
2519 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2520 }
2521 if (idx < m->req()-1) {
2522 idx += 1;
2523 stack.set_index(idx);
2524 n = m->in(idx);
2525 break;
2526 }
2527 // Take care of place holder nodes
2528 if (m->has_out_with(Op_Node)) {
2529 Node* place_holder = m->find_out_with(Op_Node);
2530 if (place_holder != nullptr) {
2531 Node* mm_clone = mm->clone();
2532 igvn.register_new_node_with_optimizer(mm_clone);
2533 Node* hook = new Node(1);
2534 hook->init_req(0, mm);
2535 igvn.replace_node(place_holder, mm_clone);
2536 hook->destruct(&igvn);
2537 }
2538 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2539 }
2540 stack.pop();
2541 }
2542 }
2543 } while(stack.size() > 0);
2544 // Fix the memory state at the MergeMem we started from
2545 igvn.rehash_node_delayed(current);
2546 for (int j = start_alias; j < num_alias_types(); j++) {
2547 const TypePtr* adr_type = get_adr_type(j);
2548 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2549 continue;
2550 }
2551 current->set_memory_at(j, mm);
2552 }
2553 current->set_memory_at(index, current->base_memory());
2554 }
2555 igvn.optimize();
2556
2557 #ifdef ASSERT
2558 wq.clear();
2559 wq.push(root());
2560 for (uint i = 0; i < wq.size(); i++) {
2561 Node* n = wq.at(i);
2562 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2563 for (uint j = 0; j < n->req(); j++) {
2564 Node* m = n->in(j);
2565 if (m != nullptr) {
2566 wq.push(m);
2567 }
2568 }
2569 }
2570 #endif
2571
2572 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2573 }
2574
2575 void Compile::record_for_merge_stores_igvn(Node* n) {
2576 if (!n->for_merge_stores_igvn()) {
2577 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
2578 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2579 _for_merge_stores_igvn.append(n);
2580 }
2581 }
2582
2583 void Compile::remove_from_merge_stores_igvn(Node* n) {
2584 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2585 _for_merge_stores_igvn.remove(n);
2586 }
2587
2588 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
2589 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
2590 // the stores, and we merge the wrong sequence of stores.
2591 // Example:
2592 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
2593 // Apply MergeStores:
2594 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
2595 // Remove more RangeChecks:
2596 // StoreI [ StoreL ] StoreI
2597 // But now it would have been better to do this instead:
2598 // [ StoreL ] [ StoreL ]
2599 //
2600 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
2601 // since we never unset _merge_stores_phase.
2602 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
2603 C->set_merge_stores_phase();
2604
2605 if (_for_merge_stores_igvn.length() > 0) {
2606 while (_for_merge_stores_igvn.length() > 0) {
2607 Node* n = _for_merge_stores_igvn.pop();
2608 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2609 igvn._worklist.push(n);
2610 }
2611 igvn.optimize();
2612 if (failing()) return;
2613 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
2614 print_method(PHASE_AFTER_MERGE_STORES, 3);
2615 }
2616 }
2617
2618 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2619 if (OptimizeUnstableIf) {
2620 _unstable_if_traps.append(trap);
2621 }
2622 }
2623
2624 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2625 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2626 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2627 Node* n = trap->uncommon_trap();
2628 if (!useful.member(n)) {
2629 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2630 }
2631 }
2632 }
2633
2634 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2635 // or fold-compares case. Return true if succeed or not found.
2636 //
2637 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2638 // false and ask fold-compares to yield.
2639 //
2640 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2641 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2642 // when deoptimization does happen.
2643 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2644 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2645 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2646 if (trap->uncommon_trap() == unc) {
2647 if (yield && trap->modified()) {
2648 return false;
2649 }
2650 _unstable_if_traps.delete_at(i);
2651 break;
2652 }
2653 }
2654 return true;
2655 }
2656
2657 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2658 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2659 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2660 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2661 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2662 CallStaticJavaNode* unc = trap->uncommon_trap();
2663 int next_bci = trap->next_bci();
2664 bool modified = trap->modified();
2665
2666 if (next_bci != -1 && !modified) {
2667 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2668 JVMState* jvms = unc->jvms();
2669 ciMethod* method = jvms->method();
2670 ciBytecodeStream iter(method);
2671
2672 iter.force_bci(jvms->bci());
2673 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2674 Bytecodes::Code c = iter.cur_bc();
2675 Node* lhs = nullptr;
2676 Node* rhs = nullptr;
2677 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2678 lhs = unc->peek_operand(0);
2679 rhs = unc->peek_operand(1);
2680 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2681 lhs = unc->peek_operand(0);
2682 }
2683
2684 ResourceMark rm;
2685 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2686 assert(live_locals.is_valid(), "broken liveness info");
2687 int len = (int)live_locals.size();
2688
2689 for (int i = 0; i < len; i++) {
2690 Node* local = unc->local(jvms, i);
2691 // kill local using the liveness of next_bci.
2692 // give up when the local looks like an operand to secure reexecution.
2693 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2694 uint idx = jvms->locoff() + i;
2695 #ifdef ASSERT
2696 if (PrintOpto && Verbose) {
2697 tty->print("[unstable_if] kill local#%d: ", idx);
2698 local->dump();
2699 tty->cr();
2700 }
2701 #endif
2702 igvn.replace_input_of(unc, idx, top());
2703 modified = true;
2704 }
2705 }
2706 }
2707
2708 // keep the modified trap for late query
2709 if (modified) {
2710 trap->set_modified();
2711 } else {
2712 _unstable_if_traps.delete_at(i);
2713 }
2714 }
2715 igvn.optimize();
2716 }
2717
2718 // StringOpts and late inlining of string methods
2719 void Compile::inline_string_calls(bool parse_time) {
2720 {
2721 // remove useless nodes to make the usage analysis simpler
2722 ResourceMark rm;
2723 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2724 }
2725
2726 {
2727 ResourceMark rm;
2728 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2729 PhaseStringOpts pso(initial_gvn());
2730 print_method(PHASE_AFTER_STRINGOPTS, 3);
2731 }
2732
2733 // now inline anything that we skipped the first time around
2734 if (!parse_time) {
2735 _late_inlines_pos = _late_inlines.length();
2736 }
2737
2738 assert(!do_cleanup(), "already set");
2739
2740 while (_string_late_inlines.length() > 0) {
2741 CallGenerator* cg = _string_late_inlines.pop();
2742 cg->do_late_inline();
2743 if (failing()) return;
2744 set_do_cleanup(false); // ignore and reset
2745 }
2746 _string_late_inlines.trunc_to(0);
2747 }
2748
2749 // Late inlining of boxing methods
2750 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2751 if (_boxing_late_inlines.length() > 0) {
2752 assert(has_boxed_value(), "inconsistent");
2753
2754 set_inlining_incrementally(true);
2755
2756 igvn_worklist()->ensure_empty(); // should be done with igvn
2757
2758 _late_inlines_pos = _late_inlines.length();
2759
2760 assert(!do_cleanup(), "already set");
2761
2762 while (_boxing_late_inlines.length() > 0) {
2763 CallGenerator* cg = _boxing_late_inlines.pop();
2764 cg->do_late_inline();
2765 if (failing()) return;
2766 set_do_cleanup(false); // ignore and reset
2767 }
2768 _boxing_late_inlines.trunc_to(0);
2769
2770 inline_incrementally_cleanup(igvn);
2771
2772 set_inlining_incrementally(false);
2773 }
2774 }
2775
2776 bool Compile::inline_incrementally_one() {
2777 assert(IncrementalInline, "incremental inlining should be on");
2778 assert(_late_inlines.length() > 0, "should have been checked by caller");
2779
2780 TracePhase tp(_t_incrInline_inline);
2781
2782 set_inlining_progress(false);
2783 set_do_cleanup(false);
2784
2785 for (int i = 0; i < _late_inlines.length(); i++) {
2786 _late_inlines_pos = i+1;
2787 CallGenerator* cg = _late_inlines.at(i);
2788 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2789 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2790 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2791 if (should_stress_inlining()) {
2792 // randomly add repeated inline attempt if stress-inlining
2793 cg->call_node()->set_generator(cg);
2794 C->igvn_worklist()->push(cg->call_node());
2795 continue;
2796 }
2797 cg->do_late_inline();
2798 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2799 if (failing()) {
2800 return false;
2801 } else if (inlining_progress()) {
2802 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2803 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2804 break; // process one call site at a time
2805 } else {
2806 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2807 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2808 // Avoid potential infinite loop if node already in the IGVN list
2809 assert(false, "scheduled for IGVN during inlining attempt");
2810 } else {
2811 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2812 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2813 cg->call_node()->set_generator(cg);
2814 }
2815 }
2816 } else {
2817 // Ignore late inline direct calls when inlining is not allowed.
2818 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2819 }
2820 }
2821 // Remove processed elements.
2822 _late_inlines.remove_till(_late_inlines_pos);
2823 _late_inlines_pos = 0;
2824
2825 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2826
2827 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2828
2829 set_inlining_progress(false);
2830 set_do_cleanup(false);
2831
2832 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2833 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2834 }
2835
2836 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2837 {
2838 TracePhase tp(_t_incrInline_pru);
2839 ResourceMark rm;
2840 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2841 }
2842 {
2843 TracePhase tp(_t_incrInline_igvn);
2844 igvn.reset();
2845 igvn.optimize();
2846 if (failing()) return;
2847 }
2848 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2849 }
2850
2851 template<typename E>
2852 static void shuffle_array(Compile& C, GrowableArray<E>& array) {
2853 if (array.length() < 2) {
2854 return;
2855 }
2856 for (uint i = array.length() - 1; i >= 1; i--) {
2857 uint j = C.random() % (i + 1);
2858 swap(array.at(i), array.at(j));
2859 }
2860 }
2861
2862 void Compile::shuffle_late_inlines() {
2863 shuffle_array(*C, _late_inlines);
2864 }
2865
2866 // Perform incremental inlining until bound on number of live nodes is reached
2867 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2868 TracePhase tp(_t_incrInline);
2869
2870 set_inlining_incrementally(true);
2871 uint low_live_nodes = 0;
2872
2873 if (StressIncrementalInlining) {
2874 shuffle_late_inlines();
2875 }
2876
2877 while (_late_inlines.length() > 0) {
2878 if (live_nodes() > node_count_inlining_cutoff()) {
2879 if (low_live_nodes < node_count_inlining_cutoff() * 8 / 10) {
2880 TracePhase tp(_t_incrInline_ideal);
2881 // PhaseIdealLoop is expensive so we only try it once we are
2882 // out of live nodes and we only try it again if the previous
2883 // helped got the number of nodes down significantly
2884 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2885 if (failing()) return;
2886 low_live_nodes = live_nodes();
2887 _major_progress = true;
2888 }
2889
2890 if (live_nodes() > node_count_inlining_cutoff()) {
2891 bool do_print_inlining = print_inlining() || print_intrinsics();
2892 if (do_print_inlining || log() != nullptr) {
2893 // Print inlining message for candidates that we couldn't inline for lack of space.
2894 for (int i = 0; i < _late_inlines.length(); i++) {
2895 CallGenerator* cg = _late_inlines.at(i);
2896 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2897 if (do_print_inlining) {
2898 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2899 }
2900 log_late_inline_failure(cg, msg);
2901 }
2902 }
2903 break; // finish
2904 }
2905 }
2906
2907 igvn_worklist()->ensure_empty(); // should be done with igvn
2908
2909 if (_late_inlines.length() == 0) {
2910 break; // no more progress
2911 }
2912
2913 while (inline_incrementally_one()) {
2914 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2915 }
2916 if (failing()) return;
2917
2918 inline_incrementally_cleanup(igvn);
2919
2920 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2921
2922 if (failing()) return;
2923 }
2924
2925 igvn_worklist()->ensure_empty(); // should be done with igvn
2926
2927 if (_string_late_inlines.length() > 0) {
2928 assert(has_stringbuilder(), "inconsistent");
2929
2930 inline_string_calls(false);
2931
2932 if (failing()) return;
2933
2934 inline_incrementally_cleanup(igvn);
2935 }
2936
2937 set_inlining_incrementally(false);
2938 }
2939
2940 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2941 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2942 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2943 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2944 // as if "inlining_incrementally() == true" were set.
2945 assert(inlining_incrementally() == false, "not allowed");
2946 set_strength_reduction(true);
2947 #ifdef ASSERT
2948 Unique_Node_List* modified_nodes = _modified_nodes;
2949 _modified_nodes = nullptr;
2950 #endif
2951 assert(_late_inlines.length() > 0, "sanity");
2952
2953 if (StressIncrementalInlining) {
2954 shuffle_late_inlines();
2955 }
2956
2957 while (_late_inlines.length() > 0) {
2958 igvn_worklist()->ensure_empty(); // should be done with igvn
2959
2960 while (inline_incrementally_one()) {
2961 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2962 }
2963 if (failing()) return;
2964
2965 inline_incrementally_cleanup(igvn);
2966 }
2967 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2968 set_strength_reduction(false);
2969 }
2970
2971 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2972 if (_loop_opts_cnt > 0) {
2973 while (major_progress() && (_loop_opts_cnt > 0)) {
2974 TracePhase tp(_t_idealLoop);
2975 PhaseIdealLoop::optimize(igvn, mode);
2976 _loop_opts_cnt--;
2977 if (failing()) return false;
2978 if (major_progress()) {
2979 print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2980 }
2981 }
2982 }
2983 return true;
2984 }
2985
2986 // Remove edges from "root" to each SafePoint at a backward branch.
2987 // They were inserted during parsing (see add_safepoint()) to make
2988 // infinite loops without calls or exceptions visible to root, i.e.,
2989 // useful.
2990 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2991 Node *r = root();
2992 if (r != nullptr) {
2993 for (uint i = r->req(); i < r->len(); ++i) {
2994 Node *n = r->in(i);
2995 if (n != nullptr && n->is_SafePoint()) {
2996 r->rm_prec(i);
2997 if (n->outcnt() == 0) {
2998 igvn.remove_dead_node(n, PhaseIterGVN::NodeOrigin::Graph);
2999 }
3000 --i;
3001 }
3002 }
3003 // Parsing may have added top inputs to the root node (Path
3004 // leading to the Halt node proven dead). Make sure we get a
3005 // chance to clean them up.
3006 igvn._worklist.push(r);
3007 igvn.optimize();
3008 }
3009 }
3010
3011 //------------------------------Optimize---------------------------------------
3012 // Given a graph, optimize it.
3013 void Compile::Optimize() {
3014 TracePhase tp(_t_optimizer);
3015
3016 #ifndef PRODUCT
3017 if (env()->break_at_compile()) {
3018 BREAKPOINT;
3019 }
3020
3021 #endif
3022
3023 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
3024 #ifdef ASSERT
3025 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
3026 #endif
3027
3028 ResourceMark rm;
3029
3030 NOT_PRODUCT( verify_graph_edges(); )
3031
3032 print_method(PHASE_AFTER_PARSING, 1);
3033
3034 {
3035 // Iterative Global Value Numbering, including ideal transforms
3036 PhaseIterGVN igvn;
3037 #ifdef ASSERT
3038 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
3039 #endif
3040 {
3041 TracePhase tp(_t_iterGVN);
3042 igvn.optimize(true);
3043 }
3044
3045 if (failing()) return;
3046
3047 print_method(PHASE_ITER_GVN1, 2);
3048
3049 process_for_unstable_if_traps(igvn);
3050
3051 if (failing()) return;
3052
3053 inline_incrementally(igvn);
3054
3055 print_method(PHASE_INCREMENTAL_INLINE, 2);
3056
3057 if (failing()) return;
3058
3059 if (eliminate_boxing()) {
3060 // Inline valueOf() methods now.
3061 inline_boxing_calls(igvn);
3062
3063 if (failing()) return;
3064
3065 if (AlwaysIncrementalInline || StressIncrementalInlining) {
3066 inline_incrementally(igvn);
3067 }
3068
3069 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
3070
3071 if (failing()) return;
3072 }
3073
3074 // Remove the speculative part of types and clean up the graph from
3075 // the extra CastPP nodes whose only purpose is to carry them. Do
3076 // that early so that optimizations are not disrupted by the extra
3077 // CastPP nodes.
3078 remove_speculative_types(igvn);
3079
3080 if (failing()) return;
3081
3082 // No more new expensive nodes will be added to the list from here
3083 // so keep only the actual candidates for optimizations.
3084 cleanup_expensive_nodes(igvn);
3085
3086 if (failing()) return;
3087
3088 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
3089 if (EnableVectorSupport && has_vbox_nodes()) {
3090 TracePhase tp(_t_vector);
3091 PhaseVector pv(igvn);
3092 pv.optimize_vector_boxes();
3093 if (failing()) return;
3094 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
3095 }
3096 assert(!has_vbox_nodes(), "sanity");
3097
3098 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
3099 Compile::TracePhase tp(_t_renumberLive);
3100 igvn_worklist()->ensure_empty(); // should be done with igvn
3101 {
3102 ResourceMark rm;
3103 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
3104 }
3105 igvn.reset();
3106 igvn.optimize(true);
3107 if (failing()) return;
3108 }
3109
3110 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
3111 // safepoints
3112 remove_root_to_sfpts_edges(igvn);
3113
3114 // Process inline type nodes now that all inlining is over
3115 process_inline_types(igvn);
3116
3117 adjust_flat_array_access_aliases(igvn);
3118
3119 if (failing()) return;
3120
3121 if (C->macro_count() > 0) {
3122 // Eliminate some macro nodes before EA to reduce analysis pressure
3123 PhaseMacroExpand mexp(igvn);
3124 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
3125 if (failing()) {
3126 return;
3127 }
3128 igvn.set_delay_transform(false);
3129 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3130 }
3131
3132 _print_phase_loop_opts = has_loops();
3133 if (_print_phase_loop_opts) {
3134 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
3135 }
3136
3137 // Perform escape analysis
3138 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
3139 if (has_loops()) {
3140 // Cleanup graph (remove dead nodes).
3141 TracePhase tp(_t_idealLoop);
3142 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
3143 if (failing()) {
3144 return;
3145 }
3146 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
3147 if (C->macro_count() > 0) {
3148 // Eliminate some macro nodes before EA to reduce analysis pressure
3149 PhaseMacroExpand mexp(igvn);
3150 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
3151 if (failing()) {
3152 return;
3153 }
3154 igvn.set_delay_transform(false);
3155 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3156 }
3157 }
3158
3159 bool progress;
3160 do {
3161 ConnectionGraph::do_analysis(this, &igvn);
3162
3163 if (failing()) return;
3164
3165 int mcount = macro_count(); // Record number of allocations and locks before IGVN
3166
3167 // Optimize out fields loads from scalar replaceable allocations.
3168 igvn.optimize(true);
3169 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
3170
3171 if (failing()) return;
3172
3173 if (congraph() != nullptr && macro_count() > 0) {
3174 TracePhase tp(_t_macroEliminate);
3175 PhaseMacroExpand mexp(igvn);
3176 mexp.eliminate_macro_nodes();
3177 if (failing()) {
3178 return;
3179 }
3180 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
3181
3182 igvn.set_delay_transform(false);
3183 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3184 }
3185
3186 ConnectionGraph::verify_ram_nodes(this, root());
3187 if (failing()) return;
3188
3189 progress = do_iterative_escape_analysis() &&
3190 (macro_count() < mcount) &&
3191 ConnectionGraph::has_candidates(this);
3192 // Try again if candidates exist and made progress
3193 // by removing some allocations and/or locks.
3194 } while (progress);
3195 }
3196
3197 process_flat_accesses(igvn);
3198 if (failing()) {
3199 return;
3200 }
3201
3202 // Loop transforms on the ideal graph. Range Check Elimination,
3203 // peeling, unrolling, etc.
3204
3205 // Set loop opts counter
3206 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
3207 {
3208 TracePhase tp(_t_idealLoop);
3209 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
3210 _loop_opts_cnt--;
3211 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
3212 if (failing()) return;
3213 }
3214 // Loop opts pass if partial peeling occurred in previous pass
3215 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
3216 TracePhase tp(_t_idealLoop);
3217 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
3218 _loop_opts_cnt--;
3219 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
3220 if (failing()) return;
3221 }
3222 // Loop opts pass for loop-unrolling before CCP
3223 if(major_progress() && (_loop_opts_cnt > 0)) {
3224 TracePhase tp(_t_idealLoop);
3225 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
3226 _loop_opts_cnt--;
3227 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
3228 }
3229 if (!failing()) {
3230 // Verify that last round of loop opts produced a valid graph
3231 PhaseIdealLoop::verify(igvn);
3232 }
3233 }
3234 if (failing()) return;
3235
3236 // Conditional Constant Propagation;
3237 print_method(PHASE_BEFORE_CCP1, 2);
3238 PhaseCCP ccp( &igvn );
3239 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
3240 {
3241 TracePhase tp(_t_ccp);
3242 ccp.do_transform();
3243 }
3244 print_method(PHASE_CCP1, 2);
3245
3246 assert( true, "Break here to ccp.dump_old2new_map()");
3247
3248 // Iterative Global Value Numbering, including ideal transforms
3249 {
3250 TracePhase tp(_t_iterGVN2);
3251 igvn.reset_from_igvn(&ccp);
3252 igvn.optimize(true);
3253 }
3254 print_method(PHASE_ITER_GVN2, 2);
3255
3256 if (failing()) return;
3257
3258 // Loop transforms on the ideal graph. Range Check Elimination,
3259 // peeling, unrolling, etc.
3260 if (!optimize_loops(igvn, LoopOptsDefault)) {
3261 return;
3262 }
3263
3264 C->clear_major_progress(); // ensure that major progress is now clear
3265
3266 process_for_post_loop_opts_igvn(igvn);
3267
3268 if (failing()) return;
3269
3270 // Once loop optimizations are over, it is safe to get rid of all reachability fence nodes and
3271 // migrate reachability edges to safepoints.
3272 if (OptimizeReachabilityFences && _reachability_fences.length() > 0) {
3273 TracePhase tp1(_t_idealLoop);
3274 TracePhase tp2(_t_reachability);
3275 PhaseIdealLoop::optimize(igvn, PostLoopOptsExpandReachabilityFences);
3276 print_method(PHASE_EXPAND_REACHABILITY_FENCES, 2);
3277 if (failing()) return;
3278 assert(_reachability_fences.length() == 0 || PreserveReachabilityFencesOnConstants, "no RF nodes allowed");
3279 }
3280
3281 process_for_merge_stores_igvn(igvn);
3282
3283 if (failing()) return;
3284
3285 #ifdef ASSERT
3286 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
3287 #endif
3288
3289 if (_late_inlines.length() > 0) {
3290 // More opportunities to optimize virtual and MH calls.
3291 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
3292 process_late_inline_calls_no_inline(igvn);
3293 if (failing()) {
3294 return;
3295 }
3296 }
3297 assert(_late_inlines.length() == 0, "late inline queue must be drained");
3298
3299 // Process inline types before macro expansion. Otherwise, we will not be able to
3300 // remove unused allocations because it cannot match the expanded allocation.
3301 process_inline_types(igvn);
3302
3303 {
3304 TracePhase tp(_t_macroExpand);
3305 PhaseMacroExpand mex(igvn);
3306 // Last attempt to eliminate macro nodes.
3307 mex.eliminate_macro_nodes();
3308 if (failing()) {
3309 return;
3310 }
3311
3312 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
3313 // Do not allow new macro nodes once we start to eliminate and expand
3314 C->reset_allow_macro_nodes();
3315 // Last attempt to eliminate macro nodes before expand
3316 mex.eliminate_macro_nodes();
3317 if (failing()) {
3318 return;
3319 }
3320 mex.eliminate_opaque_looplimit_macro_nodes();
3321 if (failing()) {
3322 return;
3323 }
3324 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
3325 if (mex.expand_macro_nodes()) {
3326 assert(failing(), "must bail out w/ explicit message");
3327 return;
3328 }
3329 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
3330 }
3331
3332 // Process inline type nodes again and remove them. From here
3333 // on we don't need to keep track of field values anymore.
3334 process_inline_types(igvn, /* remove= */ true);
3335
3336 {
3337 TracePhase tp(_t_barrierExpand);
3338 if (bs->expand_barriers(this, igvn)) {
3339 assert(failing(), "must bail out w/ explicit message");
3340 return;
3341 }
3342 print_method(PHASE_BARRIER_EXPANSION, 2);
3343 }
3344
3345 if (C->max_vector_size() > 0) {
3346 C->optimize_logic_cones(igvn);
3347 igvn.optimize();
3348 if (failing()) return;
3349 }
3350
3351 DEBUG_ONLY( _modified_nodes = nullptr; )
3352 DEBUG_ONLY( _late_inlines.clear(); )
3353
3354 assert(igvn._worklist.size() == 0, "not empty");
3355 } // (End scope of igvn; run destructor if necessary for asserts.)
3356
3357 check_no_dead_use();
3358
3359 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
3360 // to remove hashes to unlock nodes for modifications.
3361 C->node_hash()->clear();
3362
3363 // A method with only infinite loops has no edges entering loops from root
3364 {
3365 TracePhase tp(_t_graphReshaping);
3366 if (final_graph_reshaping()) {
3367 assert(failing(), "must bail out w/ explicit message");
3368 return;
3369 }
3370 }
3371
3372 print_method(PHASE_OPTIMIZE_FINISHED, 2);
3373 DEBUG_ONLY(set_phase_optimize_finished();)
3374 }
3375
3376 #ifdef ASSERT
3377 void Compile::check_no_dead_use() const {
3378 ResourceMark rm;
3379 Unique_Node_List wq;
3380 wq.push(root());
3381 for (uint i = 0; i < wq.size(); ++i) {
3382 Node* n = wq.at(i);
3383 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3384 Node* u = n->fast_out(j);
3385 if (u->outcnt() == 0 && !u->is_Con()) {
3386 u->dump();
3387 fatal("no reachable node should have no use");
3388 }
3389 wq.push(u);
3390 }
3391 }
3392 }
3393 #endif
3394
3395 void Compile::inline_vector_reboxing_calls() {
3396 assert(!do_cleanup(), "already set");
3397 if (C->_vector_reboxing_late_inlines.length() > 0) {
3398 _late_inlines_pos = C->_late_inlines.length();
3399 while (_vector_reboxing_late_inlines.length() > 0) {
3400 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
3401 cg->do_late_inline();
3402 if (failing()) return;
3403 assert(!do_cleanup(), "should not be set");
3404 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
3405 }
3406 _vector_reboxing_late_inlines.trunc_to(0);
3407 }
3408 }
3409
3410 bool Compile::has_vbox_nodes() {
3411 if (C->_vector_reboxing_late_inlines.length() > 0) {
3412 return true;
3413 }
3414 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3415 Node * n = C->macro_node(macro_idx);
3416 assert(n->is_macro(), "only macro nodes expected here");
3417 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3418 return true;
3419 }
3420 }
3421 return false;
3422 }
3423
3424 //---------------------------- Bitwise operation packing optimization ---------------------------
3425
3426 static bool is_vector_unary_bitwise_op(Node* n) {
3427 return n->Opcode() == Op_XorV &&
3428 VectorNode::is_vector_bitwise_not_pattern(n);
3429 }
3430
3431 static bool is_vector_binary_bitwise_op(Node* n) {
3432 switch (n->Opcode()) {
3433 case Op_AndV:
3434 case Op_OrV:
3435 return true;
3436
3437 case Op_XorV:
3438 return !is_vector_unary_bitwise_op(n);
3439
3440 default:
3441 return false;
3442 }
3443 }
3444
3445 static bool is_vector_ternary_bitwise_op(Node* n) {
3446 return n->Opcode() == Op_MacroLogicV;
3447 }
3448
3449 static bool is_vector_bitwise_op(Node* n) {
3450 return is_vector_unary_bitwise_op(n) ||
3451 is_vector_binary_bitwise_op(n) ||
3452 is_vector_ternary_bitwise_op(n);
3453 }
3454
3455 static bool is_vector_bitwise_cone_root(Node* n) {
3456 if (n->bottom_type()->isa_pvectmask() || !is_vector_bitwise_op(n)) {
3457 return false;
3458 }
3459 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3460 if (is_vector_bitwise_op(n->fast_out(i))) {
3461 return false;
3462 }
3463 }
3464 return true;
3465 }
3466
3467 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3468 uint cnt = 0;
3469 if (is_vector_bitwise_op(n)) {
3470 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3471 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3472 assert(n->req() == (n->is_predicated_vector() ? 4 : 3), "must have 2 data inputs");
3473 Node* opnd = VectorNode::is_all_ones_vector(n->in(1)) ? n->in(2) : n->in(1);
3474 if (!inputs.member(opnd)) {
3475 inputs.push(opnd);
3476 cnt++;
3477 }
3478 assert(cnt <= 1, "not unary");
3479 } else {
3480 uint last_req = inp_cnt;
3481 if (is_vector_ternary_bitwise_op(n)) {
3482 last_req = inp_cnt - 1; // skip last input
3483 }
3484 for (uint i = 1; i < last_req; i++) {
3485 Node* def = n->in(i);
3486 if (!inputs.member(def)) {
3487 inputs.push(def);
3488 cnt++;
3489 }
3490 }
3491 }
3492 } else { // not a bitwise operations
3493 if (!inputs.member(n)) {
3494 inputs.push(n);
3495 cnt++;
3496 }
3497 }
3498 return cnt;
3499 }
3500
3501 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3502 Unique_Node_List useful_nodes;
3503 C->identify_useful_nodes(useful_nodes);
3504
3505 for (uint i = 0; i < useful_nodes.size(); i++) {
3506 Node* n = useful_nodes.at(i);
3507 if (is_vector_bitwise_cone_root(n)) {
3508 list.push(n);
3509 }
3510 }
3511 }
3512
3513 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3514 const TypeVect* vt,
3515 Unique_Node_List& partition,
3516 Unique_Node_List& inputs) {
3517 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3518 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3519 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3520
3521 Node* in1 = inputs.at(0);
3522 Node* in2 = inputs.at(1);
3523 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3524
3525 uint func = compute_truth_table(partition, inputs);
3526
3527 Node* pn = partition.at(partition.size() - 1);
3528 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3529 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3530 }
3531
3532 static uint extract_bit(uint func, uint pos) {
3533 return (func & (1 << pos)) >> pos;
3534 }
3535
3536 //
3537 // A macro logic node represents a truth table. It has 4 inputs,
3538 // First three inputs corresponds to 3 columns of a truth table
3539 // and fourth input captures the logic function.
3540 //
3541 // eg. fn = (in1 AND in2) OR in3;
3542 //
3543 // MacroNode(in1,in2,in3,fn)
3544 //
3545 // -----------------
3546 // in1 in2 in3 fn
3547 // -----------------
3548 // 0 0 0 0
3549 // 0 0 1 1
3550 // 0 1 0 0
3551 // 0 1 1 1
3552 // 1 0 0 0
3553 // 1 0 1 1
3554 // 1 1 0 1
3555 // 1 1 1 1
3556 //
3557
3558 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3559 int res = 0;
3560 for (int i = 0; i < 8; i++) {
3561 int bit1 = extract_bit(in1, i);
3562 int bit2 = extract_bit(in2, i);
3563 int bit3 = extract_bit(in3, i);
3564
3565 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3566 int func_bit = extract_bit(func, func_bit_pos);
3567
3568 res |= func_bit << i;
3569 }
3570 return res;
3571 }
3572
3573 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
3574 assert(n != nullptr, "");
3575 assert(eval_map.contains(n), "absent");
3576 return *(eval_map.get(n));
3577 }
3578
3579 static void eval_operands(Node* n,
3580 uint& func1, uint& func2, uint& func3,
3581 HashTable<Node*,uint>& eval_map) {
3582 assert(is_vector_bitwise_op(n), "");
3583
3584 if (is_vector_unary_bitwise_op(n)) {
3585 Node* opnd = n->in(1);
3586 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3587 opnd = n->in(2);
3588 }
3589 func1 = eval_operand(opnd, eval_map);
3590 } else if (is_vector_binary_bitwise_op(n)) {
3591 func1 = eval_operand(n->in(1), eval_map);
3592 func2 = eval_operand(n->in(2), eval_map);
3593 } else {
3594 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3595 func1 = eval_operand(n->in(1), eval_map);
3596 func2 = eval_operand(n->in(2), eval_map);
3597 func3 = eval_operand(n->in(3), eval_map);
3598 }
3599 }
3600
3601 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3602 assert(inputs.size() <= 3, "sanity");
3603 ResourceMark rm;
3604 uint res = 0;
3605 HashTable<Node*,uint> eval_map;
3606
3607 // Populate precomputed functions for inputs.
3608 // Each input corresponds to one column of 3 input truth-table.
3609 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3610 0xCC, // (_, b, _) -> b
3611 0xF0 }; // (a, _, _) -> a
3612 for (uint i = 0; i < inputs.size(); i++) {
3613 eval_map.put(inputs.at(i), input_funcs[2-i]);
3614 }
3615
3616 for (uint i = 0; i < partition.size(); i++) {
3617 Node* n = partition.at(i);
3618
3619 uint func1 = 0, func2 = 0, func3 = 0;
3620 eval_operands(n, func1, func2, func3, eval_map);
3621
3622 switch (n->Opcode()) {
3623 case Op_OrV:
3624 assert(func3 == 0, "not binary");
3625 res = func1 | func2;
3626 break;
3627 case Op_AndV:
3628 assert(func3 == 0, "not binary");
3629 res = func1 & func2;
3630 break;
3631 case Op_XorV:
3632 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3633 assert(func2 == 0 && func3 == 0, "not unary");
3634 res = (~func1) & 0xFF;
3635 } else {
3636 assert(func3 == 0, "not binary");
3637 res = func1 ^ func2;
3638 }
3639 break;
3640 case Op_MacroLogicV:
3641 // Ordering of inputs may change during evaluation of sub-tree
3642 // containing MacroLogic node as a child node, thus a re-evaluation
3643 // makes sure that function is evaluated in context of current
3644 // inputs.
3645 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3646 break;
3647
3648 default: assert(false, "not supported: %s", n->Name());
3649 }
3650 assert(res <= 0xFF, "invalid");
3651 eval_map.put(n, res);
3652 }
3653 return res;
3654 }
3655
3656 // Criteria under which nodes gets packed into a macro logic node:-
3657 // 1) Parent and both child nodes are all unmasked or masked with
3658 // same predicates.
3659 // 2) Masked parent can be packed with left child if it is predicated
3660 // and both have same predicates.
3661 // 3) Masked parent can be packed with right child if its un-predicated
3662 // or has matching predication condition.
3663 // 4) An unmasked parent can be packed with an unmasked child.
3664 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3665 assert(partition.size() == 0, "not empty");
3666 assert(inputs.size() == 0, "not empty");
3667 if (is_vector_ternary_bitwise_op(n)) {
3668 return false;
3669 }
3670
3671 bool is_unary_op = is_vector_unary_bitwise_op(n);
3672 if (is_unary_op) {
3673 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3674 return false; // too few inputs
3675 }
3676
3677 bool pack_left_child = true;
3678 bool pack_right_child = true;
3679
3680 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3681 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3682
3683 int left_child_input_cnt = 0;
3684 int right_child_input_cnt = 0;
3685
3686 bool parent_is_predicated = n->is_predicated_vector();
3687 bool left_child_predicated = n->in(1)->is_predicated_vector();
3688 bool right_child_predicated = n->in(2)->is_predicated_vector();
3689
3690 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
3691 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3692 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3693
3694 do {
3695 if (pack_left_child && left_child_LOP &&
3696 ((!parent_is_predicated && !left_child_predicated) ||
3697 ((parent_is_predicated && left_child_predicated &&
3698 parent_pred == left_child_pred)))) {
3699 partition.push(n->in(1));
3700 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3701 } else {
3702 inputs.push(n->in(1));
3703 left_child_input_cnt = 1;
3704 }
3705
3706 if (pack_right_child && right_child_LOP &&
3707 (!right_child_predicated ||
3708 (right_child_predicated && parent_is_predicated &&
3709 parent_pred == right_child_pred))) {
3710 partition.push(n->in(2));
3711 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3712 } else {
3713 inputs.push(n->in(2));
3714 right_child_input_cnt = 1;
3715 }
3716
3717 if (inputs.size() > 3) {
3718 assert(partition.size() > 0, "");
3719 inputs.clear();
3720 partition.clear();
3721 if (left_child_input_cnt > right_child_input_cnt) {
3722 pack_left_child = false;
3723 } else {
3724 pack_right_child = false;
3725 }
3726 } else {
3727 break;
3728 }
3729 } while(true);
3730
3731 if(partition.size()) {
3732 partition.push(n);
3733 }
3734
3735 return (partition.size() == 2 || partition.size() == 3) &&
3736 (inputs.size() == 2 || inputs.size() == 3);
3737 }
3738
3739 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3740 assert(is_vector_bitwise_op(n), "not a root");
3741
3742 visited.set(n->_idx);
3743
3744 // 1) Do a DFS walk over the logic cone.
3745 for (uint i = 1; i < n->req(); i++) {
3746 Node* in = n->in(i);
3747 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3748 process_logic_cone_root(igvn, in, visited);
3749 }
3750 }
3751
3752 // 2) Bottom up traversal: Merge node[s] with
3753 // the parent to form macro logic node.
3754 Unique_Node_List partition;
3755 Unique_Node_List inputs;
3756 if (compute_logic_cone(n, partition, inputs)) {
3757 const TypeVect* vt = n->bottom_type()->is_vect();
3758 Node* pn = partition.at(partition.size() - 1);
3759 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3760 if (mask == nullptr ||
3761 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3762 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3763 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3764 igvn.replace_node(n, macro_logic);
3765 }
3766 }
3767 }
3768
3769 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3770 ResourceMark rm;
3771 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3772 Unique_Node_List list;
3773 collect_logic_cone_roots(list);
3774
3775 while (list.size() > 0) {
3776 Node* n = list.pop();
3777 const TypeVect* vt = n->bottom_type()->is_vect();
3778 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3779 if (supported) {
3780 VectorSet visited(comp_arena());
3781 process_logic_cone_root(igvn, n, visited);
3782 }
3783 }
3784 }
3785 }
3786
3787 //------------------------------Code_Gen---------------------------------------
3788 // Given a graph, generate code for it
3789 void Compile::Code_Gen() {
3790 if (failing()) {
3791 return;
3792 }
3793
3794 // Perform instruction selection. You might think we could reclaim Matcher
3795 // memory PDQ, but actually the Matcher is used in generating spill code.
3796 // Internals of the Matcher (including some VectorSets) must remain live
3797 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3798 // set a bit in reclaimed memory.
3799
3800 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3801 // nodes. Mapping is only valid at the root of each matched subtree.
3802 NOT_PRODUCT( verify_graph_edges(); )
3803
3804 Matcher matcher;
3805 _matcher = &matcher;
3806 {
3807 TracePhase tp(_t_matcher);
3808 matcher.match();
3809 if (failing()) {
3810 return;
3811 }
3812 }
3813 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3814 // nodes. Mapping is only valid at the root of each matched subtree.
3815 NOT_PRODUCT( verify_graph_edges(); )
3816
3817 // If you have too many nodes, or if matching has failed, bail out
3818 check_node_count(0, "out of nodes matching instructions");
3819 if (failing()) {
3820 return;
3821 }
3822
3823 print_method(PHASE_MATCHING, 2);
3824
3825 // Build a proper-looking CFG
3826 PhaseCFG cfg(node_arena(), root(), matcher);
3827 if (failing()) {
3828 return;
3829 }
3830 _cfg = &cfg;
3831 {
3832 TracePhase tp(_t_scheduler);
3833 bool success = cfg.do_global_code_motion();
3834 if (!success) {
3835 return;
3836 }
3837
3838 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3839 NOT_PRODUCT( verify_graph_edges(); )
3840 cfg.verify();
3841 if (failing()) {
3842 return;
3843 }
3844 }
3845
3846 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3847 _regalloc = ®alloc;
3848 {
3849 TracePhase tp(_t_registerAllocation);
3850 // Perform register allocation. After Chaitin, use-def chains are
3851 // no longer accurate (at spill code) and so must be ignored.
3852 // Node->LRG->reg mappings are still accurate.
3853 _regalloc->Register_Allocate();
3854
3855 // Bail out if the allocator builds too many nodes
3856 if (failing()) {
3857 return;
3858 }
3859
3860 print_method(PHASE_REGISTER_ALLOCATION, 2);
3861 }
3862
3863 // Prior to register allocation we kept empty basic blocks in case the
3864 // the allocator needed a place to spill. After register allocation we
3865 // are not adding any new instructions. If any basic block is empty, we
3866 // can now safely remove it.
3867 {
3868 TracePhase tp(_t_blockOrdering);
3869 cfg.remove_empty_blocks();
3870 if (do_freq_based_layout()) {
3871 PhaseBlockLayout layout(cfg);
3872 } else {
3873 cfg.set_loop_alignment();
3874 }
3875 cfg.fixup_flow();
3876 cfg.remove_unreachable_blocks();
3877 cfg.verify_dominator_tree();
3878 print_method(PHASE_BLOCK_ORDERING, 3);
3879 }
3880
3881 // Apply peephole optimizations
3882 if( OptoPeephole ) {
3883 TracePhase tp(_t_peephole);
3884 PhasePeephole peep( _regalloc, cfg);
3885 peep.do_transform();
3886 print_method(PHASE_PEEPHOLE, 3);
3887 }
3888
3889 // Do late expand if CPU requires this.
3890 if (Matcher::require_postalloc_expand) {
3891 TracePhase tp(_t_postalloc_expand);
3892 cfg.postalloc_expand(_regalloc);
3893 print_method(PHASE_POSTALLOC_EXPAND, 3);
3894 }
3895
3896 #ifdef ASSERT
3897 {
3898 CompilationMemoryStatistic::do_test_allocations();
3899 if (failing()) return;
3900 }
3901 #endif
3902
3903 // Convert Nodes to instruction bits in a buffer
3904 {
3905 TracePhase tp(_t_output);
3906 PhaseOutput output;
3907 output.Output();
3908 if (failing()) return;
3909 output.install();
3910 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3911 }
3912
3913 // He's dead, Jim.
3914 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3915 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3916 }
3917
3918 //------------------------------Final_Reshape_Counts---------------------------
3919 // This class defines counters and node lists collected during
3920 // the final graph reshaping.
3921 struct Final_Reshape_Counts : public StackObj {
3922 int _java_call_count; // count non-inlined 'java' calls
3923 int _inner_loop_count; // count loops which need alignment
3924 VectorSet _visited; // Visitation flags
3925 Node_List _tests; // Set of IfNodes & PCTableNodes
3926
3927 Final_Reshape_Counts() :
3928 _java_call_count(0), _inner_loop_count(0) { }
3929
3930 void inc_java_call_count() { _java_call_count++; }
3931 void inc_inner_loop_count() { _inner_loop_count++; }
3932
3933 int get_java_call_count() const { return _java_call_count; }
3934 int get_inner_loop_count() const { return _inner_loop_count; }
3935 };
3936
3937 //------------------------------final_graph_reshaping_impl----------------------
3938 // Implement items 1-5 from final_graph_reshaping below.
3939 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3940
3941 if ( n->outcnt() == 0 ) return; // dead node
3942 uint nop = n->Opcode();
3943
3944 // Check for 2-input instruction with "last use" on right input.
3945 // Swap to left input. Implements item (2).
3946 if( n->req() == 3 && // two-input instruction
3947 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3948 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3949 n->in(2)->outcnt() == 1 &&// right use IS a last use
3950 !n->in(2)->is_Con() ) { // right use is not a constant
3951 // Check for commutative opcode
3952 switch( nop ) {
3953 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddHF: case Op_AddL:
3954 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3955 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3956 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulHF: case Op_MulL:
3957 case Op_AndL: case Op_XorL: case Op_OrL:
3958 case Op_AndI: case Op_XorI: case Op_OrI: {
3959 // Move "last use" input to left by swapping inputs
3960 n->swap_edges(1, 2);
3961 break;
3962 }
3963 default:
3964 break;
3965 }
3966 }
3967
3968 #ifdef ASSERT
3969 if( n->is_Mem() ) {
3970 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3971 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3972 // oop will be recorded in oop map if load crosses safepoint
3973 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3974 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3975 "raw memory operations should have control edge");
3976 }
3977 if (n->is_MemBar()) {
3978 MemBarNode* mb = n->as_MemBar();
3979 if (mb->trailing_store() || mb->trailing_load_store()) {
3980 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3981 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3982 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3983 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3984 } else if (mb->leading()) {
3985 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3986 }
3987 }
3988 if (n->is_CallLeafPure()) {
3989 // A pure call whose result projection is unused should have been
3990 // eliminated by CallLeafPureNode::Ideal during IGVN.
3991 assert(n->as_CallLeafPure()->proj_out_or_null(TypeFunc::Parms) != nullptr,
3992 "unused CallLeafPureNode should have been removed before final graph reshaping");
3993 }
3994 #endif
3995 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3996 if (!gc_handled) {
3997 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3998 }
3999
4000 // Collect CFG split points
4001 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
4002 frc._tests.push(n);
4003 }
4004 }
4005
4006 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
4007 if (!UseDivMod) {
4008 return;
4009 }
4010
4011 // Check if "a % b" and "a / b" both exist
4012 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
4013 if (d == nullptr) {
4014 return;
4015 }
4016
4017 // Replace them with a fused divmod if supported
4018 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
4019 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
4020 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
4021 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
4022 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
4023 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
4024 // DivMod node so the dependency is not lost.
4025 divmod->add_prec_from(n);
4026 divmod->add_prec_from(d);
4027 d->subsume_by(divmod->first_proj(), this);
4028 n->subsume_by(divmod->second_proj(), this);
4029 } else {
4030 // Replace "a % b" with "a - ((a / b) * b)"
4031 Node* mult = MulNode::make(d, d->in(2), bt);
4032 Node* sub = SubNode::make(d->in(1), mult, bt);
4033 n->subsume_by(sub, this);
4034 }
4035 }
4036
4037 void Compile::handle_mulhi_mul_op(Node* n, bool is_unsigned) {
4038 const int fused_opcode = is_unsigned ? Op_UMulHiLoL : Op_MulHiLoL;
4039 if (!Matcher::has_match_rule(fused_opcode)) {
4040 return;
4041 }
4042
4043 Node* mul = n->find_similar(Op_MulL, true);
4044
4045 if (mul == nullptr) {
4046 return;
4047 }
4048
4049 MulHiLoLNode* mul_hi_lo = is_unsigned ? static_cast<MulHiLoLNode*>(UMulHiLoLNode::make(n))
4050 : MulHiLoLNode::make(n);
4051 mul->subsume_by(mul_hi_lo->first_proj(), this);
4052 n->subsume_by(mul_hi_lo->second_proj(), this);
4053 }
4054
4055 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
4056 switch( nop ) {
4057 case Op_Opaque1: // Remove Opaque Nodes before matching
4058 n->subsume_by(n->in(1), this);
4059 break;
4060 case Op_CallLeafPure: {
4061 // If the pure call is not supported, then lower to a CallLeaf.
4062 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
4063 CallNode* call = n->as_Call();
4064 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
4065 call->_name, TypeRawPtr::BOTTOM);
4066 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
4067 new_call->init_req(TypeFunc::I_O, C->top());
4068 new_call->init_req(TypeFunc::Memory, C->top());
4069 new_call->init_req(TypeFunc::ReturnAdr, C->top());
4070 new_call->init_req(TypeFunc::FramePtr, C->top());
4071 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain_sig()->cnt(); i++) {
4072 new_call->init_req(i, call->in(i));
4073 }
4074 n->subsume_by(new_call, this);
4075 }
4076 break;
4077 }
4078 case Op_CallStaticJava:
4079 case Op_CallJava:
4080 case Op_CallDynamicJava:
4081 frc.inc_java_call_count(); // Count java call site;
4082 case Op_CallRuntime:
4083 case Op_CallLeaf:
4084 case Op_CallLeafVector:
4085 case Op_CallLeafNoFP: {
4086 assert (n->is_Call(), "");
4087 CallNode *call = n->as_Call();
4088 // See if uncommon argument is shared
4089 if (call->is_CallStaticJava() && call->as_CallStaticJava()->_name) {
4090 Node *n = call->in(TypeFunc::Parms);
4091 int nop = n->Opcode();
4092 // Clone shared simple arguments to uncommon calls, item (1).
4093 if (n->outcnt() > 1 &&
4094 !n->is_Proj() &&
4095 nop != Op_CreateEx &&
4096 nop != Op_CheckCastPP &&
4097 nop != Op_DecodeN &&
4098 nop != Op_DecodeNKlass &&
4099 !n->is_Mem() &&
4100 !n->is_Phi()) {
4101 Node *x = n->clone();
4102 call->set_req(TypeFunc::Parms, x);
4103 }
4104 }
4105 break;
4106 }
4107
4108 // Mem nodes need explicit cases to satisfy assert(!n->is_Mem()) in default.
4109 case Op_StoreF:
4110 case Op_LoadF:
4111 case Op_StoreD:
4112 case Op_LoadD:
4113 case Op_LoadD_unaligned:
4114 case Op_StoreB:
4115 case Op_StoreC:
4116 case Op_StoreI:
4117 case Op_StoreL:
4118 case Op_StoreLSpecial:
4119 case Op_CompareAndSwapB:
4120 case Op_CompareAndSwapS:
4121 case Op_CompareAndSwapI:
4122 case Op_CompareAndSwapL:
4123 case Op_CompareAndSwapP:
4124 case Op_CompareAndSwapN:
4125 case Op_WeakCompareAndSwapB:
4126 case Op_WeakCompareAndSwapS:
4127 case Op_WeakCompareAndSwapI:
4128 case Op_WeakCompareAndSwapL:
4129 case Op_WeakCompareAndSwapP:
4130 case Op_WeakCompareAndSwapN:
4131 case Op_CompareAndExchangeB:
4132 case Op_CompareAndExchangeS:
4133 case Op_CompareAndExchangeI:
4134 case Op_CompareAndExchangeL:
4135 case Op_CompareAndExchangeP:
4136 case Op_CompareAndExchangeN:
4137 case Op_GetAndAddS:
4138 case Op_GetAndAddB:
4139 case Op_GetAndAddI:
4140 case Op_GetAndAddL:
4141 case Op_GetAndSetS:
4142 case Op_GetAndSetB:
4143 case Op_GetAndSetI:
4144 case Op_GetAndSetL:
4145 case Op_GetAndSetP:
4146 case Op_GetAndSetN:
4147 case Op_StoreP:
4148 case Op_StoreN:
4149 case Op_StoreNKlass:
4150 case Op_LoadB:
4151 case Op_LoadUB:
4152 case Op_LoadUS:
4153 case Op_LoadI:
4154 case Op_LoadKlass:
4155 case Op_LoadNKlass:
4156 case Op_LoadL:
4157 case Op_LoadL_unaligned:
4158 case Op_LoadP:
4159 case Op_LoadN:
4160 case Op_LoadRange:
4161 case Op_LoadS:
4162 case Op_LoadVectorGather:
4163 case Op_StoreVectorScatter:
4164 case Op_LoadVectorGatherMasked:
4165 case Op_StoreVectorScatterMasked:
4166 case Op_LoadVectorMasked:
4167 case Op_StoreVectorMasked:
4168 break;
4169
4170 case Op_AddP: { // Assert sane base pointers
4171 Node *addp = n->in(AddPNode::Address);
4172 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
4173 #ifdef _LP64
4174 if (addp->Opcode() == Op_ConP &&
4175 addp == n->in(AddPNode::Base) &&
4176 n->in(AddPNode::Offset)->is_Con()) {
4177 // If the transformation of ConP to ConN+DecodeN is beneficial depends
4178 // on the platform and on the compressed oops mode.
4179 // Use addressing with narrow klass to load with offset on x86.
4180 // Some platforms can use the constant pool to load ConP.
4181 // Do this transformation here since IGVN will convert ConN back to ConP.
4182 const Type* t = addp->bottom_type();
4183 bool is_oop = t->isa_oopptr() != nullptr;
4184 bool is_klass = t->isa_klassptr() != nullptr;
4185
4186 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
4187 (is_klass && Matcher::const_klass_prefer_decode() &&
4188 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
4189 Node* nn = nullptr;
4190
4191 int op = is_oop ? Op_ConN : Op_ConNKlass;
4192
4193 // Look for existing ConN node of the same exact type.
4194 Node* r = root();
4195 uint cnt = r->outcnt();
4196 for (uint i = 0; i < cnt; i++) {
4197 Node* m = r->raw_out(i);
4198 if (m!= nullptr && m->Opcode() == op &&
4199 m->bottom_type()->make_ptr() == t) {
4200 nn = m;
4201 break;
4202 }
4203 }
4204 if (nn != nullptr) {
4205 // Decode a narrow oop to match address
4206 // [R12 + narrow_oop_reg<<3 + offset]
4207 if (is_oop) {
4208 nn = new DecodeNNode(nn, t);
4209 } else {
4210 nn = new DecodeNKlassNode(nn, t);
4211 }
4212 // Check for succeeding AddP which uses the same Base.
4213 // Otherwise we will run into the assertion above when visiting that guy.
4214 for (uint i = 0; i < n->outcnt(); ++i) {
4215 Node *out_i = n->raw_out(i);
4216 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
4217 out_i->set_req(AddPNode::Base, nn);
4218 #ifdef ASSERT
4219 for (uint j = 0; j < out_i->outcnt(); ++j) {
4220 Node *out_j = out_i->raw_out(j);
4221 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
4222 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
4223 }
4224 #endif
4225 }
4226 }
4227 n->set_req(AddPNode::Base, nn);
4228 n->set_req(AddPNode::Address, nn);
4229 if (addp->outcnt() == 0) {
4230 addp->disconnect_inputs(this);
4231 }
4232 }
4233 }
4234 }
4235 #endif
4236 break;
4237 }
4238
4239 case Op_CastPP: {
4240 // Remove CastPP nodes to gain more freedom during scheduling but
4241 // keep the dependency they encode as control or precedence edges
4242 // (if control is set already) on memory operations. Some CastPP
4243 // nodes don't have a control (don't carry a dependency): skip
4244 // those.
4245 if (n->in(0) != nullptr) {
4246 ResourceMark rm;
4247 Unique_Node_List wq;
4248 wq.push(n);
4249
4250
4251 // When we remove a CastPP, we need to pin all of its transitive users under the control of
4252 // the removed node. The simplest approach is to pin all of the uses of the removed CastPP,
4253 // but it is overly conservative, as an AddP does not really need pinning. As a result, we
4254 // look through those nodes that do not need pinning and only pin memory access nodes under
4255 // n->in(0).
4256 for (uint next = 0; next < wq.size(); ++next) {
4257 Node *m = wq.at(next);
4258 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
4259 Node* use = m->fast_out(i);
4260 int use_op = use->Opcode();
4261 if (use->is_CFG() || use->pinned() || // already pinned at the exact control
4262 use->is_Cmp() || use_op == Op_CastP2X || use_op == Op_Conv2B) { // pure computations
4263 continue;
4264 } else if (use->is_EncodeNarrowPtr() || // EncodeP remembers whether its input is nullable, so it must be pinned
4265 use_op == Op_PartialSubtypeCheck || // This accesses its pointer inputs, so it must depend on them being not-null
4266 use->is_Mem() || use->is_memory_access_intrinsic()) {
4267 use->ensure_control_or_add_prec(n->in(0));
4268 } else if (use_op == Op_AddP ||
4269 use_op == Op_CastPP || use_op == Op_CheckCastPP ||
4270 use_op == Op_CMoveP || use_op == Op_CMoveN ||
4271 use_op == Op_DecodeN || use_op == Op_DecodeNKlass ||
4272 use_op == Op_VerifyVectorAlignment) {
4273 // Look through use to find memory accesses if use does not need pinning
4274 wq.push(use);
4275 } else {
4276 // Should have handled all kinds of nodes, verify that we do not unexpectedly arrive
4277 // here
4278 assert(false, "unexpected node %s", use->Name());
4279 // Be conservative in product and pin the unexpected use
4280 use->ensure_control_or_add_prec(n->in(0));
4281 }
4282 }
4283 }
4284 }
4285 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
4286 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
4287 Node* in1 = n->in(1);
4288 const Type* t = n->bottom_type();
4289 Node* new_in1 = in1->clone();
4290 new_in1->as_DecodeN()->set_type(t);
4291
4292 if (!Matcher::narrow_oop_use_complex_address()) {
4293 //
4294 // x86, ARM and friends can handle 2 adds in addressing mode
4295 // and Matcher can fold a DecodeN node into address by using
4296 // a narrow oop directly and do implicit null check in address:
4297 //
4298 // [R12 + narrow_oop_reg<<3 + offset]
4299 // NullCheck narrow_oop_reg
4300 //
4301 // On other platforms (Sparc) we have to keep new DecodeN node and
4302 // use it to do implicit null check in address:
4303 //
4304 // decode_not_null narrow_oop_reg, base_reg
4305 // [base_reg + offset]
4306 // NullCheck base_reg
4307 //
4308 // Pin the new DecodeN node to non-null path on these platform (Sparc)
4309 // to keep the information to which null check the new DecodeN node
4310 // corresponds to use it as value in implicit_null_check().
4311 //
4312 new_in1->set_req(0, n->in(0));
4313 }
4314
4315 n->subsume_by(new_in1, this);
4316 if (in1->outcnt() == 0) {
4317 in1->disconnect_inputs(this);
4318 }
4319 } else {
4320 n->subsume_by(n->in(1), this);
4321 if (n->outcnt() == 0) {
4322 n->disconnect_inputs(this);
4323 }
4324 }
4325 break;
4326 }
4327 case Op_CastII: {
4328 n->as_CastII()->remove_range_check_cast(this);
4329 break;
4330 }
4331 #ifdef _LP64
4332 case Op_CmpP:
4333 // Do this transformation here to preserve CmpPNode::sub() and
4334 // other TypePtr related Ideal optimizations (for example, ptr nullness).
4335 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
4336 Node* in1 = n->in(1);
4337 Node* in2 = n->in(2);
4338 if (!in1->is_DecodeNarrowPtr()) {
4339 in2 = in1;
4340 in1 = n->in(2);
4341 }
4342 assert(in1->is_DecodeNarrowPtr(), "sanity");
4343
4344 Node* new_in2 = nullptr;
4345 if (in2->is_DecodeNarrowPtr()) {
4346 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
4347 new_in2 = in2->in(1);
4348 } else if (in2->Opcode() == Op_ConP) {
4349 const Type* t = in2->bottom_type();
4350 if (t == TypePtr::NULL_PTR) {
4351 assert(in1->is_DecodeN(), "compare klass to null?");
4352 // Don't convert CmpP null check into CmpN if compressed
4353 // oops implicit null check is not generated.
4354 // This will allow to generate normal oop implicit null check.
4355 if (Matcher::gen_narrow_oop_implicit_null_checks())
4356 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
4357 //
4358 // This transformation together with CastPP transformation above
4359 // will generated code for implicit null checks for compressed oops.
4360 //
4361 // The original code after Optimize()
4362 //
4363 // LoadN memory, narrow_oop_reg
4364 // decode narrow_oop_reg, base_reg
4365 // CmpP base_reg, nullptr
4366 // CastPP base_reg // NotNull
4367 // Load [base_reg + offset], val_reg
4368 //
4369 // after these transformations will be
4370 //
4371 // LoadN memory, narrow_oop_reg
4372 // CmpN narrow_oop_reg, nullptr
4373 // decode_not_null narrow_oop_reg, base_reg
4374 // Load [base_reg + offset], val_reg
4375 //
4376 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
4377 // since narrow oops can be used in debug info now (see the code in
4378 // final_graph_reshaping_walk()).
4379 //
4380 // At the end the code will be matched to
4381 // on x86:
4382 //
4383 // Load_narrow_oop memory, narrow_oop_reg
4384 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
4385 // NullCheck narrow_oop_reg
4386 //
4387 // and on sparc:
4388 //
4389 // Load_narrow_oop memory, narrow_oop_reg
4390 // decode_not_null narrow_oop_reg, base_reg
4391 // Load [base_reg + offset], val_reg
4392 // NullCheck base_reg
4393 //
4394 } else if (t->isa_oopptr()) {
4395 new_in2 = ConNode::make(t->make_narrowoop());
4396 } else if (t->isa_klassptr()) {
4397 ciKlass* klass = t->is_klassptr()->exact_klass();
4398 if (klass->is_in_encoding_range()) {
4399 new_in2 = ConNode::make(t->make_narrowklass());
4400 }
4401 }
4402 }
4403 if (new_in2 != nullptr) {
4404 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
4405 n->subsume_by(cmpN, this);
4406 if (in1->outcnt() == 0) {
4407 in1->disconnect_inputs(this);
4408 }
4409 if (in2->outcnt() == 0) {
4410 in2->disconnect_inputs(this);
4411 }
4412 }
4413 }
4414 break;
4415
4416 case Op_DecodeN:
4417 case Op_DecodeNKlass:
4418 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4419 // DecodeN could be pinned when it can't be fold into
4420 // an address expression, see the code for Op_CastPP above.
4421 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4422 break;
4423
4424 case Op_EncodeP:
4425 case Op_EncodePKlass: {
4426 Node* in1 = n->in(1);
4427 if (in1->is_DecodeNarrowPtr()) {
4428 n->subsume_by(in1->in(1), this);
4429 } else if (in1->Opcode() == Op_ConP) {
4430 const Type* t = in1->bottom_type();
4431 if (t == TypePtr::NULL_PTR) {
4432 assert(t->isa_oopptr(), "null klass?");
4433 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4434 } else if (t->isa_oopptr()) {
4435 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4436 } else if (t->isa_klassptr()) {
4437 ciKlass* klass = t->is_klassptr()->exact_klass();
4438 if (klass->is_in_encoding_range()) {
4439 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4440 } else {
4441 assert(false, "unencodable klass in ConP -> EncodeP");
4442 C->record_failure("unencodable klass in ConP -> EncodeP");
4443 }
4444 }
4445 }
4446 if (in1->outcnt() == 0) {
4447 in1->disconnect_inputs(this);
4448 }
4449 break;
4450 }
4451
4452 case Op_Proj: {
4453 if (OptimizeStringConcat || IncrementalInline) {
4454 ProjNode* proj = n->as_Proj();
4455 if (proj->_is_io_use) {
4456 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4457 // Separate projections were used for the exception path which
4458 // are normally removed by a late inline. If it wasn't inlined
4459 // then they will hang around and should just be replaced with
4460 // the original one. Merge them.
4461 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4462 if (non_io_proj != nullptr) {
4463 proj->subsume_by(non_io_proj , this);
4464 }
4465 }
4466 }
4467 break;
4468 }
4469
4470 case Op_Phi:
4471 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4472 // The EncodeP optimization may create Phi with the same edges
4473 // for all paths. It is not handled well by Register Allocator.
4474 Node* unique_in = n->in(1);
4475 assert(unique_in != nullptr, "");
4476 uint cnt = n->req();
4477 for (uint i = 2; i < cnt; i++) {
4478 Node* m = n->in(i);
4479 assert(m != nullptr, "");
4480 if (unique_in != m)
4481 unique_in = nullptr;
4482 }
4483 if (unique_in != nullptr) {
4484 n->subsume_by(unique_in, this);
4485 }
4486 }
4487 break;
4488
4489 #endif
4490
4491 case Op_ModI:
4492 handle_div_mod_op(n, T_INT, false);
4493 break;
4494
4495 case Op_ModL:
4496 handle_div_mod_op(n, T_LONG, false);
4497 break;
4498
4499 case Op_UModI:
4500 handle_div_mod_op(n, T_INT, true);
4501 break;
4502
4503 case Op_UModL:
4504 handle_div_mod_op(n, T_LONG, true);
4505 break;
4506
4507 case Op_MulHiL:
4508 handle_mulhi_mul_op(n, false);
4509 break;
4510
4511 case Op_UMulHiL:
4512 handle_mulhi_mul_op(n, true);
4513 break;
4514
4515 case Op_LoadVector:
4516 case Op_StoreVector:
4517 #ifdef ASSERT
4518 // Add VerifyVectorAlignment node between adr and load / store.
4519 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
4520 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
4521 n->as_StoreVector()->must_verify_alignment();
4522 if (must_verify_alignment) {
4523 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
4524 n->as_StoreVector()->memory_size();
4525 // The memory access should be aligned to the vector width in bytes.
4526 // However, the underlying array is possibly less well aligned, but at least
4527 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
4528 // a loop we can expect at least the following alignment:
4529 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
4530 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
4531 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
4532 // Create mask from alignment. e.g. 0b1000 -> 0b0111
4533 jlong mask = guaranteed_alignment - 1;
4534 Node* mask_con = ConLNode::make(mask);
4535 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
4536 n->set_req(MemNode::Address, va);
4537 }
4538 }
4539 #endif
4540 break;
4541
4542 case Op_PackB:
4543 case Op_PackS:
4544 case Op_PackI:
4545 case Op_PackF:
4546 case Op_PackL:
4547 case Op_PackD:
4548 if (n->req()-1 > 2) {
4549 // Replace many operand PackNodes with a binary tree for matching
4550 PackNode* p = (PackNode*) n;
4551 Node* btp = p->binary_tree_pack(1, n->req());
4552 n->subsume_by(btp, this);
4553 }
4554 break;
4555 case Op_Loop:
4556 // When StressCountedLoop is enabled, this loop may intentionally avoid a counted loop conversion.
4557 // This is expected behavior for the stress mode, which exercises alternative compilation paths.
4558 if (!StressCountedLoop) {
4559 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4560 }
4561 case Op_CountedLoop:
4562 case Op_LongCountedLoop:
4563 case Op_OuterStripMinedLoop:
4564 if (n->as_Loop()->is_inner_loop()) {
4565 frc.inc_inner_loop_count();
4566 }
4567 n->as_Loop()->verify_strip_mined(0);
4568 break;
4569 case Op_LShiftI:
4570 case Op_RShiftI:
4571 case Op_URShiftI:
4572 case Op_LShiftL:
4573 case Op_RShiftL:
4574 case Op_URShiftL:
4575 if (Matcher::need_masked_shift_count) {
4576 // The cpu's shift instructions don't restrict the count to the
4577 // lower 5/6 bits. We need to do the masking ourselves.
4578 Node* in2 = n->in(2);
4579 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4580 const TypeInt* t = in2->find_int_type();
4581 if (t != nullptr && t->is_con()) {
4582 juint shift = t->get_con();
4583 if (shift > mask) { // Unsigned cmp
4584 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4585 }
4586 } else {
4587 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
4588 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4589 n->set_req(2, shift);
4590 }
4591 }
4592 if (in2->outcnt() == 0) { // Remove dead node
4593 in2->disconnect_inputs(this);
4594 }
4595 }
4596 break;
4597 case Op_MemBarStoreStore:
4598 case Op_MemBarRelease:
4599 // Break the link with AllocateNode: it is no longer useful and
4600 // confuses register allocation.
4601 if (n->req() > MemBarNode::Precedent) {
4602 n->set_req(MemBarNode::Precedent, top());
4603 }
4604 break;
4605 case Op_MemBarAcquire: {
4606 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4607 // At parse time, the trailing MemBarAcquire for a volatile load
4608 // is created with an edge to the load. After optimizations,
4609 // that input may be a chain of Phis. If those phis have no
4610 // other use, then the MemBarAcquire keeps them alive and
4611 // register allocation can be confused.
4612 dead_nodes.push(n->in(MemBarNode::Precedent));
4613 n->set_req(MemBarNode::Precedent, top());
4614 }
4615 break;
4616 }
4617 case Op_RangeCheck: {
4618 RangeCheckNode* rc = n->as_RangeCheck();
4619 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4620 n->subsume_by(iff, this);
4621 frc._tests.push(iff);
4622 break;
4623 }
4624 case Op_ConvI2L: {
4625 if (!Matcher::convi2l_type_required) {
4626 // Code generation on some platforms doesn't need accurate
4627 // ConvI2L types. Widening the type can help remove redundant
4628 // address computations.
4629 n->as_Type()->set_type(TypeLong::INT);
4630 ResourceMark rm;
4631 Unique_Node_List wq;
4632 wq.push(n);
4633 for (uint next = 0; next < wq.size(); next++) {
4634 Node *m = wq.at(next);
4635
4636 for(;;) {
4637 // Loop over all nodes with identical inputs edges as m
4638 Node* k = m->find_similar(m->Opcode());
4639 if (k == nullptr) {
4640 break;
4641 }
4642 // Push their uses so we get a chance to remove node made
4643 // redundant
4644 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4645 Node* u = k->fast_out(i);
4646 if (u->Opcode() == Op_LShiftL ||
4647 u->Opcode() == Op_AddL ||
4648 u->Opcode() == Op_SubL ||
4649 u->Opcode() == Op_AddP) {
4650 wq.push(u);
4651 }
4652 }
4653 // Replace all nodes with identical edges as m with m
4654 k->subsume_by(m, this);
4655 }
4656 }
4657 }
4658 break;
4659 }
4660 case Op_CmpUL: {
4661 if (!Matcher::has_match_rule(Op_CmpUL)) {
4662 // No support for unsigned long comparisons
4663 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4664 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4665 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4666 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4667 Node* andl = new AndLNode(orl, remove_sign_mask);
4668 Node* cmp = new CmpLNode(andl, n->in(2));
4669 n->subsume_by(cmp, this);
4670 }
4671 break;
4672 }
4673 #ifdef ASSERT
4674 case Op_InlineType: {
4675 n->dump(-1);
4676 assert(false, "inline type node was not removed");
4677 break;
4678 }
4679 case Op_ConNKlass: {
4680 const TypePtr* tp = n->as_Type()->type()->make_ptr();
4681 ciKlass* klass = tp->is_klassptr()->exact_klass();
4682 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
4683 break;
4684 }
4685 #endif
4686 default:
4687 assert(!n->is_Call(), "");
4688 assert(!n->is_Mem(), "");
4689 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4690 break;
4691 }
4692 }
4693
4694 //------------------------------final_graph_reshaping_walk---------------------
4695 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4696 // requires that the walk visits a node's inputs before visiting the node.
4697 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4698 Unique_Node_List sfpt;
4699
4700 frc._visited.set(root->_idx); // first, mark node as visited
4701 uint cnt = root->req();
4702 Node *n = root;
4703 uint i = 0;
4704 while (true) {
4705 if (i < cnt) {
4706 // Place all non-visited non-null inputs onto stack
4707 Node* m = n->in(i);
4708 ++i;
4709 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
4710 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
4711 // compute worst case interpreter size in case of a deoptimization
4712 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4713
4714 sfpt.push(m);
4715 }
4716 cnt = m->req();
4717 nstack.push(n, i); // put on stack parent and next input's index
4718 n = m;
4719 i = 0;
4720 }
4721 } else {
4722 // Now do post-visit work
4723 final_graph_reshaping_impl(n, frc, dead_nodes);
4724 if (nstack.is_empty())
4725 break; // finished
4726 n = nstack.node(); // Get node from stack
4727 cnt = n->req();
4728 i = nstack.index();
4729 nstack.pop(); // Shift to the next node on stack
4730 }
4731 }
4732
4733 expand_reachability_edges(sfpt);
4734
4735 // Skip next transformation if compressed oops are not used.
4736 if (UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks())
4737 return;
4738
4739 // Go over ReachabilityFence nodes to skip DecodeN nodes for referents.
4740 // The sole purpose of RF node is to keep the referent oop alive and
4741 // decoding the oop for that is not needed.
4742 for (int i = 0; i < C->reachability_fences_count(); i++) {
4743 ReachabilityFenceNode* rf = C->reachability_fence(i);
4744 DecodeNNode* dn = rf->in(1)->isa_DecodeN();
4745 if (dn != nullptr) {
4746 if (!dn->has_non_debug_uses() || Matcher::narrow_oop_use_complex_address()) {
4747 rf->set_req(1, dn->in(1));
4748 if (dn->outcnt() == 0) {
4749 dn->disconnect_inputs(this);
4750 }
4751 }
4752 }
4753 }
4754
4755 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4756 // It could be done for an uncommon traps or any safepoints/calls
4757 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4758 while (sfpt.size() > 0) {
4759 n = sfpt.pop();
4760 JVMState *jvms = n->as_SafePoint()->jvms();
4761 assert(jvms != nullptr, "sanity");
4762 int start = jvms->debug_start();
4763 int end = n->req();
4764 bool is_uncommon = (n->is_CallStaticJava() &&
4765 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4766 for (int j = start; j < end; j++) {
4767 Node* in = n->in(j);
4768 if (in->is_DecodeNarrowPtr() && (is_uncommon || !in->has_non_debug_uses())) {
4769 n->set_req(j, in->in(1));
4770 if (in->outcnt() == 0) {
4771 in->disconnect_inputs(this);
4772 }
4773 }
4774 }
4775 }
4776 }
4777
4778 //------------------------------final_graph_reshaping--------------------------
4779 // Final Graph Reshaping.
4780 //
4781 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4782 // and not commoned up and forced early. Must come after regular
4783 // optimizations to avoid GVN undoing the cloning. Clone constant
4784 // inputs to Loop Phis; these will be split by the allocator anyways.
4785 // Remove Opaque nodes.
4786 // (2) Move last-uses by commutative operations to the left input to encourage
4787 // Intel update-in-place two-address operations and better register usage
4788 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4789 // calls canonicalizing them back.
4790 // (3) Detect infinite loops; blobs of code reachable from above but not
4791 // below. Several of the Code_Gen algorithms fail on such code shapes,
4792 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4793 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4794 // Detection is by looking for IfNodes where only 1 projection is
4795 // reachable from below or CatchNodes missing some targets.
4796 // (4) Assert for insane oop offsets in debug mode.
4797
4798 bool Compile::final_graph_reshaping() {
4799 // an infinite loop may have been eliminated by the optimizer,
4800 // in which case the graph will be empty.
4801 if (root()->req() == 1) {
4802 // Do not compile method that is only a trivial infinite loop,
4803 // since the content of the loop may have been eliminated.
4804 record_method_not_compilable("trivial infinite loop");
4805 return true;
4806 }
4807
4808 // Expensive nodes have their control input set to prevent the GVN
4809 // from freely commoning them. There's no GVN beyond this point so
4810 // no need to keep the control input. We want the expensive nodes to
4811 // be freely moved to the least frequent code path by gcm.
4812 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4813 for (int i = 0; i < expensive_count(); i++) {
4814 _expensive_nodes.at(i)->set_req(0, nullptr);
4815 }
4816
4817 Final_Reshape_Counts frc;
4818
4819 // Visit everybody reachable!
4820 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4821 Node_Stack nstack(live_nodes() >> 1);
4822 Unique_Node_List dead_nodes;
4823 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4824
4825 // Check for unreachable (from below) code (i.e., infinite loops).
4826 for( uint i = 0; i < frc._tests.size(); i++ ) {
4827 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4828 // Get number of CFG targets.
4829 // Note that PCTables include exception targets after calls.
4830 uint required_outcnt = n->required_outcnt();
4831 if (n->outcnt() != required_outcnt) {
4832 // Check for a few special cases. Rethrow Nodes never take the
4833 // 'fall-thru' path, so expected kids is 1 less.
4834 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4835 if (n->in(0)->in(0)->is_Call()) {
4836 CallNode* call = n->in(0)->in(0)->as_Call();
4837 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4838 required_outcnt--; // Rethrow always has 1 less kid
4839 } else if (call->req() > TypeFunc::Parms &&
4840 call->is_CallDynamicJava()) {
4841 // Check for null receiver. In such case, the optimizer has
4842 // detected that the virtual call will always result in a null
4843 // pointer exception. The fall-through projection of this CatchNode
4844 // will not be populated.
4845 Node* arg0 = call->in(TypeFunc::Parms);
4846 if (arg0->is_Type() &&
4847 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4848 required_outcnt--;
4849 }
4850 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4851 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4852 // Check for illegal array length. In such case, the optimizer has
4853 // detected that the allocation attempt will always result in an
4854 // exception. There is no fall-through projection of this CatchNode .
4855 assert(call->is_CallStaticJava(), "static call expected");
4856 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4857 uint valid_length_test_input = call->req() - 1;
4858 Node* valid_length_test = call->in(valid_length_test_input);
4859 call->del_req(valid_length_test_input);
4860 if (valid_length_test->find_int_con(1) == 0) {
4861 required_outcnt--;
4862 }
4863 dead_nodes.push(valid_length_test);
4864 assert(n->outcnt() == required_outcnt, "malformed control flow");
4865 continue;
4866 }
4867 }
4868 }
4869
4870 // Recheck with a better notion of 'required_outcnt'
4871 if (n->outcnt() != required_outcnt) {
4872 record_method_not_compilable("malformed control flow");
4873 return true; // Not all targets reachable!
4874 }
4875 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4876 CallNode* call = n->in(0)->in(0)->as_Call();
4877 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4878 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4879 assert(call->is_CallStaticJava(), "static call expected");
4880 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4881 uint valid_length_test_input = call->req() - 1;
4882 dead_nodes.push(call->in(valid_length_test_input));
4883 call->del_req(valid_length_test_input); // valid length test useless now
4884 }
4885 }
4886 // Check that I actually visited all kids. Unreached kids
4887 // must be infinite loops.
4888 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4889 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4890 record_method_not_compilable("infinite loop");
4891 return true; // Found unvisited kid; must be unreach
4892 }
4893
4894 // Here so verification code in final_graph_reshaping_walk()
4895 // always see an OuterStripMinedLoopEnd
4896 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4897 IfNode* init_iff = n->as_If();
4898 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4899 n->subsume_by(iff, this);
4900 }
4901 }
4902
4903 while (dead_nodes.size() > 0) {
4904 Node* m = dead_nodes.pop();
4905 if (m->outcnt() == 0 && m != top()) {
4906 for (uint j = 0; j < m->req(); j++) {
4907 Node* in = m->in(j);
4908 if (in != nullptr) {
4909 dead_nodes.push(in);
4910 }
4911 }
4912 m->disconnect_inputs(this);
4913 }
4914 }
4915
4916 set_java_calls(frc.get_java_call_count());
4917 set_inner_loops(frc.get_inner_loop_count());
4918
4919 // No infinite loops, no reason to bail out.
4920 return false;
4921 }
4922
4923 //-----------------------------too_many_traps----------------------------------
4924 // Report if there are too many traps at the current method and bci.
4925 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4926 bool Compile::too_many_traps(ciMethod* method,
4927 int bci,
4928 Deoptimization::DeoptReason reason) {
4929 ciMethodData* md = method->method_data();
4930 if (md->is_empty()) {
4931 // Assume the trap has not occurred, or that it occurred only
4932 // because of a transient condition during start-up in the interpreter.
4933 return false;
4934 }
4935 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4936 if (md->has_trap_at(bci, m, reason) != 0) {
4937 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4938 // Also, if there are multiple reasons, or if there is no per-BCI record,
4939 // assume the worst.
4940 if (log())
4941 log()->elem("observe trap='%s' count='%d'",
4942 Deoptimization::trap_reason_name(reason),
4943 md->trap_count(reason));
4944 return true;
4945 } else {
4946 // Ignore method/bci and see if there have been too many globally.
4947 return too_many_traps(reason, md);
4948 }
4949 }
4950
4951 // Less-accurate variant which does not require a method and bci.
4952 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4953 ciMethodData* logmd) {
4954 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4955 // Too many traps globally.
4956 // Note that we use cumulative trap_count, not just md->trap_count.
4957 if (log()) {
4958 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4959 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4960 Deoptimization::trap_reason_name(reason),
4961 mcount, trap_count(reason));
4962 }
4963 return true;
4964 } else {
4965 // The coast is clear.
4966 return false;
4967 }
4968 }
4969
4970 //--------------------------too_many_recompiles--------------------------------
4971 // Report if there are too many recompiles at the current method and bci.
4972 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4973 // Is not eager to return true, since this will cause the compiler to use
4974 // Action_none for a trap point, to avoid too many recompilations.
4975 bool Compile::too_many_recompiles(ciMethod* method,
4976 int bci,
4977 Deoptimization::DeoptReason reason) {
4978 ciMethodData* md = method->method_data();
4979 if (md->is_empty()) {
4980 // Assume the trap has not occurred, or that it occurred only
4981 // because of a transient condition during start-up in the interpreter.
4982 return false;
4983 }
4984 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4985 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4986 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4987 Deoptimization::DeoptReason per_bc_reason
4988 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4989 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4990 if ((per_bc_reason == Deoptimization::Reason_none
4991 || md->has_trap_at(bci, m, reason) != 0)
4992 // The trap frequency measure we care about is the recompile count:
4993 && md->trap_recompiled_at(bci, m)
4994 && md->overflow_recompile_count() >= bc_cutoff) {
4995 // Do not emit a trap here if it has already caused recompilations.
4996 // Also, if there are multiple reasons, or if there is no per-BCI record,
4997 // assume the worst.
4998 if (log())
4999 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
5000 Deoptimization::trap_reason_name(reason),
5001 md->trap_count(reason),
5002 md->overflow_recompile_count());
5003 return true;
5004 } else if (trap_count(reason) != 0
5005 && decompile_count() >= m_cutoff) {
5006 // Too many recompiles globally, and we have seen this sort of trap.
5007 // Use cumulative decompile_count, not just md->decompile_count.
5008 if (log())
5009 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
5010 Deoptimization::trap_reason_name(reason),
5011 md->trap_count(reason), trap_count(reason),
5012 md->decompile_count(), decompile_count());
5013 return true;
5014 } else {
5015 // The coast is clear.
5016 return false;
5017 }
5018 }
5019
5020 // Compute when not to trap. Used by matching trap based nodes and
5021 // NullCheck optimization.
5022 void Compile::set_allowed_deopt_reasons() {
5023 _allowed_reasons = 0;
5024 if (is_method_compilation()) {
5025 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
5026 assert(rs < BitsPerInt, "recode bit map");
5027 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
5028 _allowed_reasons |= nth_bit(rs);
5029 }
5030 }
5031 }
5032 }
5033
5034 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
5035 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
5036 }
5037
5038 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
5039 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
5040 }
5041
5042 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
5043 if (holder->is_initialized()) {
5044 return false;
5045 }
5046 if (holder->is_being_initialized()) {
5047 if (accessing_method->holder() == holder) {
5048 // Access inside a class. The barrier can be elided when access happens in <clinit>,
5049 // <init>, or a static method. In all those cases, there was an initialization
5050 // barrier on the holder klass passed.
5051 if (accessing_method->is_class_initializer() ||
5052 accessing_method->is_object_constructor() ||
5053 accessing_method->is_static()) {
5054 return false;
5055 }
5056 } else if (accessing_method->holder()->is_subclass_of(holder)) {
5057 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
5058 // In case of <init> or a static method, the barrier is on the subclass is not enough:
5059 // child class can become fully initialized while its parent class is still being initialized.
5060 if (accessing_method->is_class_initializer()) {
5061 return false;
5062 }
5063 }
5064 ciMethod* root = method(); // the root method of compilation
5065 if (root != accessing_method) {
5066 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
5067 }
5068 }
5069 return true;
5070 }
5071
5072 #ifndef PRODUCT
5073 //------------------------------verify_bidirectional_edges---------------------
5074 // For each input edge to a node (ie - for each Use-Def edge), verify that
5075 // there is a corresponding Def-Use edge.
5076 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
5077 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
5078 uint stack_size = live_nodes() >> 4;
5079 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
5080 if (root_and_safepoints != nullptr) {
5081 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
5082 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
5083 Node* root_or_safepoint = root_and_safepoints->at(i);
5084 // If the node is a safepoint, let's check if it still has a control input
5085 // Lack of control input signifies that this node was killed by CCP or
5086 // recursively by remove_globally_dead_node and it shouldn't be a starting
5087 // point.
5088 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
5089 nstack.push(root_or_safepoint);
5090 }
5091 }
5092 } else {
5093 nstack.push(_root);
5094 }
5095
5096 while (nstack.size() > 0) {
5097 Node* n = nstack.pop();
5098 if (visited.member(n)) {
5099 continue;
5100 }
5101 visited.push(n);
5102
5103 // Walk over all input edges, checking for correspondence
5104 uint length = n->len();
5105 for (uint i = 0; i < length; i++) {
5106 Node* in = n->in(i);
5107 if (in != nullptr && !visited.member(in)) {
5108 nstack.push(in); // Put it on stack
5109 }
5110 if (in != nullptr && !in->is_top()) {
5111 // Count instances of `next`
5112 int cnt = 0;
5113 for (uint idx = 0; idx < in->_outcnt; idx++) {
5114 if (in->_out[idx] == n) {
5115 cnt++;
5116 }
5117 }
5118 assert(cnt > 0, "Failed to find Def-Use edge.");
5119 // Check for duplicate edges
5120 // walk the input array downcounting the input edges to n
5121 for (uint j = 0; j < length; j++) {
5122 if (n->in(j) == in) {
5123 cnt--;
5124 }
5125 }
5126 assert(cnt == 0, "Mismatched edge count.");
5127 } else if (in == nullptr) {
5128 assert(i == 0 || i >= n->req() ||
5129 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
5130 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
5131 (n->is_Unlock() && i == (n->req() - 1)) ||
5132 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
5133 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
5134 } else {
5135 assert(in->is_top(), "sanity");
5136 // Nothing to check.
5137 }
5138 }
5139 }
5140 }
5141
5142 //------------------------------verify_graph_edges---------------------------
5143 // Walk the Graph and verify that there is a one-to-one correspondence
5144 // between Use-Def edges and Def-Use edges in the graph.
5145 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
5146 if (VerifyGraphEdges) {
5147 Unique_Node_List visited;
5148
5149 // Call graph walk to check edges
5150 verify_bidirectional_edges(visited, root_and_safepoints);
5151 if (no_dead_code) {
5152 // Now make sure that no visited node is used by an unvisited node.
5153 bool dead_nodes = false;
5154 Unique_Node_List checked;
5155 while (visited.size() > 0) {
5156 Node* n = visited.pop();
5157 checked.push(n);
5158 for (uint i = 0; i < n->outcnt(); i++) {
5159 Node* use = n->raw_out(i);
5160 if (checked.member(use)) continue; // already checked
5161 if (visited.member(use)) continue; // already in the graph
5162 if (use->is_Con()) continue; // a dead ConNode is OK
5163 // At this point, we have found a dead node which is DU-reachable.
5164 if (!dead_nodes) {
5165 tty->print_cr("*** Dead nodes reachable via DU edges:");
5166 dead_nodes = true;
5167 }
5168 use->dump(2);
5169 tty->print_cr("---");
5170 checked.push(use); // No repeats; pretend it is now checked.
5171 }
5172 }
5173 assert(!dead_nodes, "using nodes must be reachable from root");
5174 }
5175 }
5176 }
5177 #endif
5178
5179 // The Compile object keeps track of failure reasons separately from the ciEnv.
5180 // This is required because there is not quite a 1-1 relation between the
5181 // ciEnv and its compilation task and the Compile object. Note that one
5182 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
5183 // to backtrack and retry without subsuming loads. Other than this backtracking
5184 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
5185 // by the logic in C2Compiler.
5186 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
5187 if (log() != nullptr) {
5188 log()->elem("failure reason='%s' phase='compile'", reason);
5189 }
5190 if (_failure_reason.get() == nullptr) {
5191 // Record the first failure reason.
5192 _failure_reason.set(reason);
5193 if (CaptureBailoutInformation) {
5194 _first_failure_details = new CompilationFailureInfo(reason);
5195 }
5196 } else {
5197 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
5198 }
5199
5200 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
5201 C->print_method(PHASE_FAILURE, 1);
5202 }
5203 _root = nullptr; // flush the graph, too
5204 }
5205
5206 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
5207 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
5208 _compile(Compile::current()),
5209 _log(nullptr),
5210 _dolog(CITimeVerbose)
5211 {
5212 assert(_compile != nullptr, "sanity check");
5213 assert(id != PhaseTraceId::_t_none, "Don't use none");
5214 if (_dolog) {
5215 _log = _compile->log();
5216 }
5217 if (_log != nullptr) {
5218 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
5219 _log->stamp();
5220 _log->end_head();
5221 }
5222
5223 // Inform memory statistic, if enabled
5224 if (CompilationMemoryStatistic::enabled()) {
5225 CompilationMemoryStatistic::on_phase_start((int)id, name);
5226 }
5227 }
5228
5229 Compile::TracePhase::TracePhase(PhaseTraceId id)
5230 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
5231
5232 Compile::TracePhase::~TracePhase() {
5233
5234 // Inform memory statistic, if enabled
5235 if (CompilationMemoryStatistic::enabled()) {
5236 CompilationMemoryStatistic::on_phase_end();
5237 }
5238
5239 if (_compile->failing_internal()) {
5240 if (_log != nullptr) {
5241 _log->done("phase");
5242 }
5243 return; // timing code, not stressing bailouts.
5244 }
5245 #ifdef ASSERT
5246 if (PrintIdealNodeCount) {
5247 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
5248 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
5249 }
5250
5251 if (VerifyIdealNodeCount) {
5252 _compile->print_missing_nodes();
5253 }
5254 #endif
5255
5256 if (_log != nullptr) {
5257 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
5258 }
5259 }
5260
5261 //----------------------------static_subtype_check-----------------------------
5262 // Shortcut important common cases when superklass is exact:
5263 // (0) superklass is java.lang.Object (can occur in reflective code)
5264 // (1) subklass is already limited to a subtype of superklass => always ok
5265 // (2) subklass does not overlap with superklass => always fail
5266 // (3) superklass has NO subtypes and we can check with a simple compare.
5267 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
5268 if (skip) {
5269 return SSC_full_test; // Let caller generate the general case.
5270 }
5271
5272 if (subk->is_java_subtype_of(superk)) {
5273 return SSC_always_true; // (0) and (1) this test cannot fail
5274 }
5275
5276 if (!subk->maybe_java_subtype_of(superk)) {
5277 return SSC_always_false; // (2) true path dead; no dynamic test needed
5278 }
5279
5280 const Type* superelem = superk;
5281 if (superk->isa_aryklassptr()) {
5282 int ignored;
5283 superelem = superk->is_aryklassptr()->base_element_type(ignored);
5284
5285 // Do not fold the subtype check to an array klass pointer comparison for null-able inline type arrays
5286 // because null-free [LMyValue <: null-able [LMyValue but the klasses are different. Perform a full test.
5287 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
5288 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
5289 return SSC_full_test;
5290 }
5291 }
5292
5293 if (superelem->isa_instklassptr()) {
5294 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
5295 if (!ik->has_subklass()) {
5296 if (!ik->is_final()) {
5297 // Add a dependency if there is a chance of a later subclass.
5298 dependencies()->assert_leaf_type(ik);
5299 }
5300 if (!superk->maybe_java_subtype_of(subk)) {
5301 return SSC_always_false;
5302 }
5303 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5304 }
5305 } else {
5306 // A primitive array type has no subtypes.
5307 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5308 }
5309
5310 return SSC_full_test;
5311 }
5312
5313 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
5314 #ifdef _LP64
5315 // The scaled index operand to AddP must be a clean 64-bit value.
5316 // Java allows a 32-bit int to be incremented to a negative
5317 // value, which appears in a 64-bit register as a large
5318 // positive number. Using that large positive number as an
5319 // operand in pointer arithmetic has bad consequences.
5320 // On the other hand, 32-bit overflow is rare, and the possibility
5321 // can often be excluded, if we annotate the ConvI2L node with
5322 // a type assertion that its value is known to be a small positive
5323 // number. (The prior range check has ensured this.)
5324 // This assertion is used by ConvI2LNode::Ideal.
5325 int index_max = max_jint - 1; // array size is max_jint, index is one less
5326 if (sizetype != nullptr && sizetype->_hi > 0) {
5327 index_max = sizetype->_hi - 1;
5328 }
5329 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
5330 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
5331 #endif
5332 return idx;
5333 }
5334
5335 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
5336 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
5337 if (ctrl != nullptr) {
5338 // Express control dependency by a CastII node with a narrow type.
5339 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
5340 // node from floating above the range check during loop optimizations. Otherwise, the
5341 // ConvI2L node may be eliminated independently of the range check, causing the data path
5342 // to become TOP while the control path is still there (although it's unreachable).
5343 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
5344 value = phase->transform(value);
5345 }
5346 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
5347 return phase->transform(new ConvI2LNode(value, ltype));
5348 }
5349
5350 void Compile::dump_print_inlining() {
5351 inline_printer()->print_on(tty);
5352 }
5353
5354 void Compile::log_late_inline(CallGenerator* cg) {
5355 if (log() != nullptr) {
5356 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5357 cg->unique_id());
5358 JVMState* p = cg->call_node()->jvms();
5359 while (p != nullptr) {
5360 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5361 p = p->caller();
5362 }
5363 log()->tail("late_inline");
5364 }
5365 }
5366
5367 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5368 log_late_inline(cg);
5369 if (log() != nullptr) {
5370 log()->inline_fail(msg);
5371 }
5372 }
5373
5374 void Compile::log_inline_id(CallGenerator* cg) {
5375 if (log() != nullptr) {
5376 // The LogCompilation tool needs a unique way to identify late
5377 // inline call sites. This id must be unique for this call site in
5378 // this compilation. Try to have it unique across compilations as
5379 // well because it can be convenient when grepping through the log
5380 // file.
5381 // Distinguish OSR compilations from others in case CICountOSR is
5382 // on.
5383 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5384 cg->set_unique_id(id);
5385 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5386 }
5387 }
5388
5389 void Compile::log_inline_failure(const char* msg) {
5390 if (C->log() != nullptr) {
5391 C->log()->inline_fail(msg);
5392 }
5393 }
5394
5395
5396 // Dump inlining replay data to the stream.
5397 // Don't change thread state and acquire any locks.
5398 void Compile::dump_inline_data(outputStream* out) {
5399 InlineTree* inl_tree = ilt();
5400 if (inl_tree != nullptr) {
5401 out->print(" inline %d", inl_tree->count());
5402 inl_tree->dump_replay_data(out);
5403 }
5404 }
5405
5406 void Compile::dump_inline_data_reduced(outputStream* out) {
5407 assert(ReplayReduce, "");
5408
5409 InlineTree* inl_tree = ilt();
5410 if (inl_tree == nullptr) {
5411 return;
5412 }
5413 // Enable iterative replay file reduction
5414 // Output "compile" lines for depth 1 subtrees,
5415 // simulating that those trees were compiled
5416 // instead of inlined.
5417 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5418 InlineTree* sub = inl_tree->subtrees().at(i);
5419 if (sub->inline_level() != 1) {
5420 continue;
5421 }
5422
5423 ciMethod* method = sub->method();
5424 int entry_bci = -1;
5425 int comp_level = env()->task()->comp_level();
5426 out->print("compile ");
5427 method->dump_name_as_ascii(out);
5428 out->print(" %d %d", entry_bci, comp_level);
5429 out->print(" inline %d", sub->count());
5430 sub->dump_replay_data(out, -1);
5431 out->cr();
5432 }
5433 }
5434
5435 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5436 if (n1->Opcode() < n2->Opcode()) return -1;
5437 else if (n1->Opcode() > n2->Opcode()) return 1;
5438
5439 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5440 for (uint i = 1; i < n1->req(); i++) {
5441 if (n1->in(i) < n2->in(i)) return -1;
5442 else if (n1->in(i) > n2->in(i)) return 1;
5443 }
5444
5445 return 0;
5446 }
5447
5448 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5449 Node* n1 = *n1p;
5450 Node* n2 = *n2p;
5451
5452 return cmp_expensive_nodes(n1, n2);
5453 }
5454
5455 void Compile::sort_expensive_nodes() {
5456 if (!expensive_nodes_sorted()) {
5457 _expensive_nodes.sort(cmp_expensive_nodes);
5458 }
5459 }
5460
5461 bool Compile::expensive_nodes_sorted() const {
5462 for (int i = 1; i < _expensive_nodes.length(); i++) {
5463 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5464 return false;
5465 }
5466 }
5467 return true;
5468 }
5469
5470 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5471 if (_expensive_nodes.length() == 0) {
5472 return false;
5473 }
5474
5475 assert(OptimizeExpensiveOps, "optimization off?");
5476
5477 // Take this opportunity to remove dead nodes from the list
5478 int j = 0;
5479 for (int i = 0; i < _expensive_nodes.length(); i++) {
5480 Node* n = _expensive_nodes.at(i);
5481 if (!n->is_unreachable(igvn)) {
5482 assert(n->is_expensive(), "should be expensive");
5483 _expensive_nodes.at_put(j, n);
5484 j++;
5485 }
5486 }
5487 _expensive_nodes.trunc_to(j);
5488
5489 // Then sort the list so that similar nodes are next to each other
5490 // and check for at least two nodes of identical kind with same data
5491 // inputs.
5492 sort_expensive_nodes();
5493
5494 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5495 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5496 return true;
5497 }
5498 }
5499
5500 return false;
5501 }
5502
5503 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5504 if (_expensive_nodes.length() == 0) {
5505 return;
5506 }
5507
5508 assert(OptimizeExpensiveOps, "optimization off?");
5509
5510 // Sort to bring similar nodes next to each other and clear the
5511 // control input of nodes for which there's only a single copy.
5512 sort_expensive_nodes();
5513
5514 int j = 0;
5515 int identical = 0;
5516 int i = 0;
5517 bool modified = false;
5518 for (; i < _expensive_nodes.length()-1; i++) {
5519 assert(j <= i, "can't write beyond current index");
5520 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5521 identical++;
5522 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5523 continue;
5524 }
5525 if (identical > 0) {
5526 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5527 identical = 0;
5528 } else {
5529 Node* n = _expensive_nodes.at(i);
5530 igvn.replace_input_of(n, 0, nullptr);
5531 igvn.hash_insert(n);
5532 modified = true;
5533 }
5534 }
5535 if (identical > 0) {
5536 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5537 } else if (_expensive_nodes.length() >= 1) {
5538 Node* n = _expensive_nodes.at(i);
5539 igvn.replace_input_of(n, 0, nullptr);
5540 igvn.hash_insert(n);
5541 modified = true;
5542 }
5543 _expensive_nodes.trunc_to(j);
5544 if (modified) {
5545 igvn.optimize();
5546 }
5547 }
5548
5549 void Compile::add_expensive_node(Node * n) {
5550 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5551 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5552 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5553 if (OptimizeExpensiveOps) {
5554 _expensive_nodes.append(n);
5555 } else {
5556 // Clear control input and let IGVN optimize expensive nodes if
5557 // OptimizeExpensiveOps is off.
5558 n->set_req(0, nullptr);
5559 }
5560 }
5561
5562 /**
5563 * Track coarsened Lock and Unlock nodes.
5564 */
5565
5566 class Lock_List : public Node_List {
5567 uint _origin_cnt;
5568 public:
5569 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5570 uint origin_cnt() const { return _origin_cnt; }
5571 };
5572
5573 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5574 int length = locks.length();
5575 if (length > 0) {
5576 // Have to keep this list until locks elimination during Macro nodes elimination.
5577 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5578 AbstractLockNode* alock = locks.at(0);
5579 BoxLockNode* box = alock->box_node()->as_BoxLock();
5580 for (int i = 0; i < length; i++) {
5581 AbstractLockNode* lock = locks.at(i);
5582 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5583 locks_list->push(lock);
5584 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
5585 if (this_box != box) {
5586 // Locking regions (BoxLock) could be Unbalanced here:
5587 // - its coarsened locks were eliminated in earlier
5588 // macro nodes elimination followed by loop unroll
5589 // - it is OSR locking region (no Lock node)
5590 // Preserve Unbalanced status in such cases.
5591 if (!this_box->is_unbalanced()) {
5592 this_box->set_coarsened();
5593 }
5594 if (!box->is_unbalanced()) {
5595 box->set_coarsened();
5596 }
5597 }
5598 }
5599 _coarsened_locks.append(locks_list);
5600 }
5601 }
5602
5603 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5604 int count = coarsened_count();
5605 for (int i = 0; i < count; i++) {
5606 Node_List* locks_list = _coarsened_locks.at(i);
5607 for (uint j = 0; j < locks_list->size(); j++) {
5608 Node* lock = locks_list->at(j);
5609 assert(lock->is_AbstractLock(), "sanity");
5610 if (!useful.member(lock)) {
5611 locks_list->yank(lock);
5612 }
5613 }
5614 }
5615 }
5616
5617 void Compile::remove_coarsened_lock(Node* n) {
5618 if (n->is_AbstractLock()) {
5619 int count = coarsened_count();
5620 for (int i = 0; i < count; i++) {
5621 Node_List* locks_list = _coarsened_locks.at(i);
5622 locks_list->yank(n);
5623 }
5624 }
5625 }
5626
5627 bool Compile::coarsened_locks_consistent() {
5628 int count = coarsened_count();
5629 for (int i = 0; i < count; i++) {
5630 bool unbalanced = false;
5631 bool modified = false; // track locks kind modifications
5632 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5633 uint size = locks_list->size();
5634 if (size == 0) {
5635 unbalanced = false; // All locks were eliminated - good
5636 } else if (size != locks_list->origin_cnt()) {
5637 unbalanced = true; // Some locks were removed from list
5638 } else {
5639 for (uint j = 0; j < size; j++) {
5640 Node* lock = locks_list->at(j);
5641 // All nodes in group should have the same state (modified or not)
5642 if (!lock->as_AbstractLock()->is_coarsened()) {
5643 if (j == 0) {
5644 // first on list was modified, the rest should be too for consistency
5645 modified = true;
5646 } else if (!modified) {
5647 // this lock was modified but previous locks on the list were not
5648 unbalanced = true;
5649 break;
5650 }
5651 } else if (modified) {
5652 // previous locks on list were modified but not this lock
5653 unbalanced = true;
5654 break;
5655 }
5656 }
5657 }
5658 if (unbalanced) {
5659 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5660 #ifdef ASSERT
5661 if (PrintEliminateLocks) {
5662 tty->print_cr("=== unbalanced coarsened locks ===");
5663 for (uint l = 0; l < size; l++) {
5664 locks_list->at(l)->dump();
5665 }
5666 }
5667 #endif
5668 record_failure(C2Compiler::retry_no_locks_coarsening());
5669 return false;
5670 }
5671 }
5672 return true;
5673 }
5674
5675 // Mark locking regions (identified by BoxLockNode) as unbalanced if
5676 // locks coarsening optimization removed Lock/Unlock nodes from them.
5677 // Such regions become unbalanced because coarsening only removes part
5678 // of Lock/Unlock nodes in region. As result we can't execute other
5679 // locks elimination optimizations which assume all code paths have
5680 // corresponding pair of Lock/Unlock nodes - they are balanced.
5681 void Compile::mark_unbalanced_boxes() const {
5682 int count = coarsened_count();
5683 for (int i = 0; i < count; i++) {
5684 Node_List* locks_list = _coarsened_locks.at(i);
5685 uint size = locks_list->size();
5686 if (size > 0) {
5687 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
5688 BoxLockNode* box = alock->box_node()->as_BoxLock();
5689 if (alock->is_coarsened()) {
5690 // coarsened_locks_consistent(), which is called before this method, verifies
5691 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
5692 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5693 for (uint j = 1; j < size; j++) {
5694 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
5695 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
5696 if (box != this_box) {
5697 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5698 box->set_unbalanced();
5699 this_box->set_unbalanced();
5700 }
5701 }
5702 }
5703 }
5704 }
5705 }
5706
5707 /**
5708 * Remove the speculative part of types and clean up the graph
5709 */
5710 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5711 if (UseTypeSpeculation) {
5712 Unique_Node_List worklist;
5713 worklist.push(root());
5714 int modified = 0;
5715 // Go over all type nodes that carry a speculative type, drop the
5716 // speculative part of the type and enqueue the node for an igvn
5717 // which may optimize it out.
5718 for (uint next = 0; next < worklist.size(); ++next) {
5719 Node *n = worklist.at(next);
5720 if (n->is_Type()) {
5721 TypeNode* tn = n->as_Type();
5722 const Type* t = tn->type();
5723 const Type* t_no_spec = t->remove_speculative();
5724 if (t_no_spec != t) {
5725 bool in_hash = igvn.hash_delete(n);
5726 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5727 tn->set_type(t_no_spec);
5728 igvn.hash_insert(n);
5729 igvn._worklist.push(n); // give it a chance to go away
5730 modified++;
5731 }
5732 }
5733 // Iterate over outs - endless loops is unreachable from below
5734 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5735 Node *m = n->fast_out(i);
5736 if (not_a_node(m)) {
5737 continue;
5738 }
5739 worklist.push(m);
5740 }
5741 }
5742 // Drop the speculative part of all types in the igvn's type table
5743 igvn.remove_speculative_types();
5744 if (modified > 0) {
5745 igvn.optimize();
5746 if (failing()) return;
5747 }
5748 #ifdef ASSERT
5749 // Verify that after the IGVN is over no speculative type has resurfaced
5750 worklist.clear();
5751 worklist.push(root());
5752 for (uint next = 0; next < worklist.size(); ++next) {
5753 Node *n = worklist.at(next);
5754 const Type* t = igvn.type_or_null(n);
5755 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5756 if (n->is_Type()) {
5757 t = n->as_Type()->type();
5758 assert(t == t->remove_speculative(), "no more speculative types");
5759 }
5760 // Iterate over outs - endless loops is unreachable from below
5761 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5762 Node *m = n->fast_out(i);
5763 if (not_a_node(m)) {
5764 continue;
5765 }
5766 worklist.push(m);
5767 }
5768 }
5769 igvn.check_no_speculative_types();
5770 #endif
5771 }
5772 }
5773
5774 // Auxiliary methods to support randomized stressing/fuzzing.
5775
5776 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5777 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5778 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5779 FLAG_SET_ERGO(StressSeed, _stress_seed);
5780 } else {
5781 _stress_seed = StressSeed;
5782 }
5783 if (_log != nullptr) {
5784 _log->elem("stress_test seed='%u'", _stress_seed);
5785 }
5786 }
5787
5788 int Compile::random() {
5789 _stress_seed = os::next_random(_stress_seed);
5790 return static_cast<int>(_stress_seed);
5791 }
5792
5793 // This method can be called the arbitrary number of times, with current count
5794 // as the argument. The logic allows selecting a single candidate from the
5795 // running list of candidates as follows:
5796 // int count = 0;
5797 // Cand* selected = null;
5798 // while(cand = cand->next()) {
5799 // if (randomized_select(++count)) {
5800 // selected = cand;
5801 // }
5802 // }
5803 //
5804 // Including count equalizes the chances any candidate is "selected".
5805 // This is useful when we don't have the complete list of candidates to choose
5806 // from uniformly. In this case, we need to adjust the randomicity of the
5807 // selection, or else we will end up biasing the selection towards the latter
5808 // candidates.
5809 //
5810 // Quick back-envelope calculation shows that for the list of n candidates
5811 // the equal probability for the candidate to persist as "best" can be
5812 // achieved by replacing it with "next" k-th candidate with the probability
5813 // of 1/k. It can be easily shown that by the end of the run, the
5814 // probability for any candidate is converged to 1/n, thus giving the
5815 // uniform distribution among all the candidates.
5816 //
5817 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5818 #define RANDOMIZED_DOMAIN_POW 29
5819 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5820 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5821 bool Compile::randomized_select(int count) {
5822 assert(count > 0, "only positive");
5823 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5824 }
5825
5826 #ifdef ASSERT
5827 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5828 bool Compile::fail_randomly() {
5829 if ((random() % StressBailoutMean) != 0) {
5830 return false;
5831 }
5832 record_failure("StressBailout");
5833 return true;
5834 }
5835
5836 bool Compile::failure_is_artificial() {
5837 return C->failure_reason_is("StressBailout");
5838 }
5839 #endif
5840
5841 CloneMap& Compile::clone_map() { return _clone_map; }
5842 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5843
5844 void NodeCloneInfo::dump_on(outputStream* st) const {
5845 st->print(" {%d:%d} ", idx(), gen());
5846 }
5847
5848 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5849 uint64_t val = value(old->_idx);
5850 NodeCloneInfo cio(val);
5851 assert(val != 0, "old node should be in the map");
5852 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5853 insert(nnn->_idx, cin.get());
5854 #ifndef PRODUCT
5855 if (is_debug()) {
5856 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5857 }
5858 #endif
5859 }
5860
5861 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5862 NodeCloneInfo cio(value(old->_idx));
5863 if (cio.get() == 0) {
5864 cio.set(old->_idx, 0);
5865 insert(old->_idx, cio.get());
5866 #ifndef PRODUCT
5867 if (is_debug()) {
5868 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5869 }
5870 #endif
5871 }
5872 clone(old, nnn, gen);
5873 }
5874
5875 int CloneMap::max_gen() const {
5876 int g = 0;
5877 DictI di(_dict);
5878 for(; di.test(); ++di) {
5879 int t = gen(di._key);
5880 if (g < t) {
5881 g = t;
5882 #ifndef PRODUCT
5883 if (is_debug()) {
5884 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5885 }
5886 #endif
5887 }
5888 }
5889 return g;
5890 }
5891
5892 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5893 uint64_t val = value(key);
5894 if (val != 0) {
5895 NodeCloneInfo ni(val);
5896 ni.dump_on(st);
5897 }
5898 }
5899
5900 void Compile::shuffle_macro_nodes() {
5901 shuffle_array(*C, _macro_nodes);
5902 }
5903
5904 // Move Allocate nodes to the start of the list
5905 void Compile::sort_macro_nodes() {
5906 int count = macro_count();
5907 int allocates = 0;
5908 for (int i = 0; i < count; i++) {
5909 Node* n = macro_node(i);
5910 if (n->is_Allocate()) {
5911 if (i != allocates) {
5912 Node* tmp = macro_node(allocates);
5913 _macro_nodes.at_put(allocates, n);
5914 _macro_nodes.at_put(i, tmp);
5915 }
5916 allocates++;
5917 }
5918 }
5919 }
5920
5921 void Compile::print_method(CompilerPhaseType compile_phase, int level, Node* n) {
5922 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5923 EventCompilerPhase event(UNTIMED);
5924 if (event.should_commit()) {
5925 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, compile_phase, C->_compile_id, level);
5926 }
5927 #ifndef PRODUCT
5928 ResourceMark rm;
5929 stringStream ss;
5930 ss.print_raw(CompilerPhaseTypeHelper::to_description(compile_phase));
5931 int iter = ++_igv_phase_iter[compile_phase];
5932 if (iter > 1) {
5933 ss.print(" %d", iter);
5934 }
5935 if (n != nullptr) {
5936 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5937 if (n->is_Call()) {
5938 CallNode* call = n->as_Call();
5939 if (call->_name != nullptr) {
5940 // E.g. uncommon traps etc.
5941 ss.print(" - %s", call->_name);
5942 } else if (call->is_CallJava()) {
5943 CallJavaNode* call_java = call->as_CallJava();
5944 if (call_java->method() != nullptr) {
5945 ss.print(" -");
5946 call_java->method()->print_short_name(&ss);
5947 }
5948 }
5949 }
5950 }
5951
5952 const char* name = ss.as_string();
5953 if (should_print_igv(level)) {
5954 _igv_printer->print_graph(name);
5955 }
5956 if (should_print_phase(level)) {
5957 print_phase(name);
5958 }
5959 if (should_print_ideal_phase(compile_phase)) {
5960 print_ideal_ir(CompilerPhaseTypeHelper::to_name(compile_phase));
5961 }
5962 #endif
5963 C->_latest_stage_start_counter.stamp();
5964 }
5965
5966 // Only used from CompileWrapper
5967 void Compile::begin_method() {
5968 #ifndef PRODUCT
5969 if (_method != nullptr && should_print_igv(1)) {
5970 _igv_printer->begin_method();
5971 }
5972 #endif
5973 C->_latest_stage_start_counter.stamp();
5974 }
5975
5976 // Only used from CompileWrapper
5977 void Compile::end_method() {
5978 EventCompilerPhase event(UNTIMED);
5979 if (event.should_commit()) {
5980 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5981 }
5982
5983 #ifndef PRODUCT
5984 if (_method != nullptr && should_print_igv(1)) {
5985 _igv_printer->end_method();
5986 }
5987 #endif
5988 }
5989
5990 #ifndef PRODUCT
5991 bool Compile::should_print_phase(const int level) const {
5992 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5993 _method != nullptr; // Do not print phases for stubs.
5994 }
5995
5996 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5997 return _directive->should_print_ideal_phase(cpt);
5998 }
5999
6000 void Compile::init_igv() {
6001 if (_igv_printer == nullptr) {
6002 _igv_printer = IdealGraphPrinter::printer();
6003 _igv_printer->set_compile(this);
6004 }
6005 }
6006
6007 bool Compile::should_print_igv(const int level) {
6008 PRODUCT_RETURN_(return false;);
6009
6010 if (PrintIdealGraphLevel < 0) { // disabled by the user
6011 return false;
6012 }
6013
6014 bool need = directive()->IGVPrintLevelOption >= level;
6015 if (need) {
6016 Compile::init_igv();
6017 }
6018 return need;
6019 }
6020
6021 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
6022 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
6023
6024 // Called from debugger. Prints method to the default file with the default phase name.
6025 // This works regardless of any Ideal Graph Visualizer flags set or not.
6026 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
6027 void igv_print(void* sp, void* fp, void* pc) {
6028 frame fr(sp, fp, pc);
6029 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
6030 }
6031
6032 // Same as igv_print() above but with a specified phase name.
6033 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
6034 frame fr(sp, fp, pc);
6035 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
6036 }
6037
6038 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
6039 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
6040 // This works regardless of any Ideal Graph Visualizer flags set or not.
6041 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
6042 void igv_print(bool network, void* sp, void* fp, void* pc) {
6043 frame fr(sp, fp, pc);
6044 if (network) {
6045 Compile::current()->igv_print_method_to_network(nullptr, &fr);
6046 } else {
6047 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
6048 }
6049 }
6050
6051 // Same as igv_print(bool network, ...) above but with a specified phase name.
6052 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
6053 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
6054 frame fr(sp, fp, pc);
6055 if (network) {
6056 Compile::current()->igv_print_method_to_network(phase_name, &fr);
6057 } else {
6058 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
6059 }
6060 }
6061
6062 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
6063 void igv_print_default() {
6064 Compile::current()->print_method(PHASE_DEBUG, 0);
6065 }
6066
6067 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
6068 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
6069 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
6070 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
6071 void igv_append(void* sp, void* fp, void* pc) {
6072 frame fr(sp, fp, pc);
6073 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
6074 }
6075
6076 // Same as igv_append(...) above but with a specified phase name.
6077 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
6078 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
6079 frame fr(sp, fp, pc);
6080 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
6081 }
6082
6083 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
6084 const char* file_name = "custom_debug.xml";
6085 if (_debug_file_printer == nullptr) {
6086 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
6087 } else {
6088 _debug_file_printer->update_compiled_method(C->method());
6089 }
6090 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
6091 _debug_file_printer->print_graph(phase_name, fr);
6092 }
6093
6094 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
6095 ResourceMark rm;
6096 GrowableArray<const Node*> empty_list;
6097 igv_print_graph_to_network(phase_name, empty_list, fr);
6098 }
6099
6100 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
6101 if (_debug_network_printer == nullptr) {
6102 _debug_network_printer = new IdealGraphPrinter(C);
6103 } else {
6104 _debug_network_printer->update_compiled_method(C->method());
6105 }
6106 tty->print_cr("Method printed over network stream to IGV");
6107 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
6108 }
6109 #endif // !PRODUCT
6110
6111 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
6112 precond(type != nullptr);
6113
6114 if (phase->type(value)->higher_equal(type)) {
6115 return value;
6116 }
6117 Node* result = nullptr;
6118 if (bt == T_BYTE) {
6119 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
6120 result = new RShiftINode(result, phase->intcon(24));
6121 } else if (bt == T_BOOLEAN) {
6122 assert(type == TypeInt::BOOL || type == TypeInt::UBYTE, "unexpected boolean type: %s", Type::str(type));
6123 Node* mask = phase->intcon(type == TypeInt::BOOL ? 1 : 0xFF);
6124 result = new AndINode(value, mask);
6125 } else if (bt == T_CHAR) {
6126 result = new AndINode(value,phase->intcon(0xFFFF));
6127 } else if (bt == T_FLOAT) {
6128 result = new MoveI2FNode(value);
6129 } else {
6130 assert(bt == T_SHORT, "unexpected narrow type");
6131 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
6132 result = new RShiftINode(result, phase->intcon(16));
6133 }
6134 if (transform_res) {
6135 result = phase->transform(result);
6136 }
6137 return result;
6138 }
6139
6140 void Compile::record_method_not_compilable_oom() {
6141 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
6142 }
6143
6144 #ifndef PRODUCT
6145 // Collects all the control inputs from nodes on the worklist and from their data dependencies
6146 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
6147 // Follow non-control edges until we reach CFG nodes
6148 for (uint i = 0; i < worklist.size(); i++) {
6149 const Node* n = worklist.at(i);
6150 for (uint j = 0; j < n->req(); j++) {
6151 Node* in = n->in(j);
6152 if (in == nullptr || in->is_Root()) {
6153 continue;
6154 }
6155 if (in->is_CFG()) {
6156 if (in->is_Call()) {
6157 // The return value of a call is only available if the call did not result in an exception
6158 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
6159 if (control_proj_use->is_Catch()) {
6160 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
6161 candidates.push(fall_through);
6162 continue;
6163 }
6164 }
6165
6166 if (in->is_Multi()) {
6167 // We got here by following data inputs so we should only have one control use
6168 // (no IfNode, etc)
6169 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
6170 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
6171 } else {
6172 candidates.push(in);
6173 }
6174 } else {
6175 worklist.push(in);
6176 }
6177 }
6178 }
6179 }
6180
6181 // Returns the candidate node that is a descendant to all the other candidates
6182 static Node* pick_control(Unique_Node_List& candidates) {
6183 Unique_Node_List worklist;
6184 worklist.copy(candidates);
6185
6186 // Traverse backwards through the CFG
6187 for (uint i = 0; i < worklist.size(); i++) {
6188 const Node* n = worklist.at(i);
6189 if (n->is_Root()) {
6190 continue;
6191 }
6192 for (uint j = 0; j < n->req(); j++) {
6193 // Skip backedge of loops to avoid cycles
6194 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
6195 continue;
6196 }
6197
6198 Node* pred = n->in(j);
6199 if (pred != nullptr && pred != n && pred->is_CFG()) {
6200 worklist.push(pred);
6201 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
6202 candidates.remove(pred);
6203 }
6204 }
6205 }
6206
6207 assert(candidates.size() == 1, "unexpected control flow");
6208 return candidates.at(0);
6209 }
6210
6211 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
6212 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
6213 call->init_req((*pos)++, parm);
6214 const BasicType bt = parm->bottom_type()->basic_type();
6215 if (bt == T_LONG || bt == T_DOUBLE) {
6216 call->init_req((*pos)++, half);
6217 }
6218 }
6219
6220 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
6221 Node* parm0, Node* parm1,
6222 Node* parm2, Node* parm3,
6223 Node* parm4, Node* parm5,
6224 Node* parm6) const {
6225 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
6226 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
6227 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
6228
6229 // find the most suitable control input
6230 Unique_Node_List worklist, candidates;
6231 if (parm0 != nullptr) { worklist.push(parm0);
6232 if (parm1 != nullptr) { worklist.push(parm1);
6233 if (parm2 != nullptr) { worklist.push(parm2);
6234 if (parm3 != nullptr) { worklist.push(parm3);
6235 if (parm4 != nullptr) { worklist.push(parm4);
6236 if (parm5 != nullptr) { worklist.push(parm5);
6237 if (parm6 != nullptr) { worklist.push(parm6);
6238 /* close each nested if ===> */ } } } } } } }
6239 find_candidate_control_inputs(worklist, candidates);
6240 Node* control = nullptr;
6241 if (candidates.size() == 0) {
6242 control = C->start()->proj_out(TypeFunc::Control);
6243 } else {
6244 control = pick_control(candidates);
6245 }
6246
6247 // find all the previous users of the control we picked
6248 GrowableArray<Node*> users_of_control;
6249 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
6250 Node* use = control->fast_out(i);
6251 if (use->is_CFG() && use != control) {
6252 users_of_control.push(use);
6253 }
6254 }
6255
6256 // we do not actually care about IO and memory as it uses neither
6257 call->init_req(TypeFunc::Control, control);
6258 call->init_req(TypeFunc::I_O, top());
6259 call->init_req(TypeFunc::Memory, top());
6260 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
6261 call->init_req(TypeFunc::ReturnAdr, top());
6262
6263 int pos = TypeFunc::Parms;
6264 call->init_req(pos++, str_node);
6265 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
6266 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
6267 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
6268 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
6269 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
6270 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
6271 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
6272 /* close each nested if ===> */ } } } } } } }
6273 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
6274
6275 call = gvn->transform(call);
6276 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
6277
6278 // rewire previous users to have the new call as control instead
6279 PhaseIterGVN* igvn = gvn->is_IterGVN();
6280 for (int i = 0; i < users_of_control.length(); i++) {
6281 Node* use = users_of_control.at(i);
6282 for (uint j = 0; j < use->req(); j++) {
6283 if (use->in(j) == control) {
6284 if (igvn != nullptr) {
6285 igvn->replace_input_of(use, j, call_control_proj);
6286 } else {
6287 gvn->hash_delete(use);
6288 use->set_req(j, call_control_proj);
6289 gvn->hash_insert(use);
6290 }
6291 }
6292 }
6293 }
6294
6295 return call;
6296 }
6297 #endif // !PRODUCT