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 initialize_stress_seed(directive);
779 }
780
781 Init(/*do_aliasing=*/ true);
782
783 print_compile_messages();
784
785 _ilt = InlineTree::build_inline_tree_root();
786
787 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
788 assert(num_alias_types() >= AliasIdxRaw, "");
789
790 #define MINIMUM_NODE_HASH 1023
791
792 // GVN that will be run immediately on new nodes
793 uint estimated_size = method()->code_size()*4+64;
794 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
795 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
796 _types = new (comp_arena()) Type_Array(comp_arena());
797 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
798 PhaseGVN gvn;
799 set_initial_gvn(&gvn);
800
801 { // Scope for timing the parser
802 TracePhase tp(_t_parser);
803
804 // Put top into the hash table ASAP.
805 initial_gvn()->transform(top());
806
807 // Set up tf(), start(), and find a CallGenerator.
808 CallGenerator* cg = nullptr;
809 if (is_osr_compilation()) {
810 init_tf(TypeFunc::make(method(), false, /* is_osr_compilation = */ true));
811 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
812 initial_gvn()->set_type_bottom(s);
813 verify_start(s);
814 cg = CallGenerator::for_osr(method(), entry_bci());
815 } else {
816 // Normal case.
817 init_tf(TypeFunc::make(method(), false));
818 StartNode* s = new StartNode(root(), tf()->domain_cc());
819 initial_gvn()->set_type_bottom(s);
820 verify_start(s);
821 float past_uses = method()->interpreter_invocation_count();
822 float expected_uses = past_uses;
823 cg = CallGenerator::for_inline(method(), expected_uses);
824 }
825 if (failing()) return;
826 if (cg == nullptr) {
827 const char* reason = InlineTree::check_can_parse(method());
828 assert(reason != nullptr, "expect reason for parse failure");
829 stringStream ss;
830 ss.print("cannot parse method: %s", reason);
831 record_method_not_compilable(ss.as_string());
832 return;
833 }
834
835 gvn.set_type(root(), root()->bottom_type());
836
837 JVMState* jvms = build_start_state(start(), tf());
838 if ((jvms = cg->generate(jvms)) == nullptr) {
839 assert(failure_reason() != nullptr, "expect reason for parse failure");
840 stringStream ss;
841 ss.print("method parse failed: %s", failure_reason());
842 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
843 return;
844 }
845 GraphKit kit(jvms);
846
847 if (!kit.stopped()) {
848 // Accept return values, and transfer control we know not where.
849 // This is done by a special, unique ReturnNode bound to root.
850 return_values(kit.jvms());
851 }
852
853 if (kit.has_exceptions()) {
854 // Any exceptions that escape from this call must be rethrown
855 // to whatever caller is dynamically above us on the stack.
856 // This is done by a special, unique RethrowNode bound to root.
857 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
858 }
859
860 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
861
862 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
863 inline_string_calls(true);
864 }
865
866 if (failing()) return;
867
868 // Remove clutter produced by parsing.
869 if (!failing()) {
870 ResourceMark rm;
871 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
872 }
873 }
874
875 // Note: Large methods are capped off in do_one_bytecode().
876 if (failing()) return;
877
878 // After parsing, node notes are no longer automagic.
879 // They must be propagated by register_new_node_with_optimizer(),
880 // clone(), or the like.
881 set_default_node_notes(nullptr);
882
883 #ifndef PRODUCT
884 if (should_print_igv(1)) {
885 _igv_printer->print_inlining();
886 }
887 #endif
888
889 if (failing()) return;
890 NOT_PRODUCT( verify_graph_edges(); )
891
892 // Now optimize
893 Optimize();
894 if (failing()) return;
895 NOT_PRODUCT( verify_graph_edges(); )
896
897 #ifndef PRODUCT
898 if (should_print_ideal()) {
899 print_ideal_ir("PrintIdeal");
900 }
901 #endif
902
903 #ifdef ASSERT
904 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
905 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
906 #endif
907
908 // Dump compilation data to replay it.
909 if (directive->DumpReplayOption) {
910 env()->dump_replay_data(_compile_id);
911 }
912 if (directive->DumpInlineOption && (ilt() != nullptr)) {
913 env()->dump_inline_data(_compile_id);
914 }
915
916 // Now that we know the size of all the monitors we can add fixed slots:
917 // [...]
918 // rsp+80: saved fp register
919 // rsp+76: Fixed slot 7
920 // rsp+72: Fixed slot 6 (stack increment)
921 // rsp+68: Fixed slot 5
922 // rsp+64: Fixed slot 4 (null marker)
923 // rsp+60: Fixed slot 3
924 // rsp+56: Fixed slot 2 (original deopt pc)
925 // rsp+52: Fixed slot 1
926 // rsp+48: Fixed slot 0 (monitors)
927 // rsp+44: spill
928 // [...]
929
930 // One extra slot for the original deopt pc.
931 int next_slot = fixed_slots();
932 next_slot += VMRegImpl::slots_per_word;
933
934 // One extra slot for the special stack increment value.
935 if (needs_stack_repair()) {
936 next_slot += VMRegImpl::slots_per_word;
937 }
938
939 // One extra slot to hold the null marker at scalarized returns.
940 if (needs_nm_slot()) {
941 next_slot += VMRegImpl::slots_per_word;
942 }
943 set_fixed_slots(next_slot);
944
945 // Compute when to use implicit null checks. Used by matching trap based
946 // nodes and NullCheck optimization.
947 set_allowed_deopt_reasons();
948
949 // Now generate code
950 Code_Gen();
951 }
952
953 // C2 uses runtime stubs serialized generation to initialize its static tables
954 // shared by all compilations, like Type::_shared_type_dict.
955 // At least one stub have to be completely generated to execute intialization
956 // before we can skip the rest stubs generation by loading AOT cached stubs.
957
958 static bool c2_do_stub_init_complete = false;
959
960 //------------------------------Compile----------------------------------------
961 // Compile a runtime stub
962 Compile::Compile(ciEnv* ci_env,
963 TypeFunc_generator generator,
964 address stub_function,
965 const char* stub_name,
966 StubId stub_id,
967 int is_fancy_jump,
968 bool pass_tls,
969 bool return_pc,
970 DirectiveSet* directive)
971 : Phase(Compiler),
972 _compile_id(0),
973 _options(Options::for_runtime_stub()),
974 _method(nullptr),
975 _entry_bci(InvocationEntryBci),
976 _stub_function(stub_function),
977 _stub_name(stub_name),
978 _stub_id(stub_id),
979 _stub_entry_point(nullptr),
980 _max_node_limit(MaxNodeLimit),
981 _node_count_inlining_cutoff(NodeCountInliningCutoff),
982 _post_loop_opts_phase(false),
983 _merge_stores_phase(false),
984 _allow_macro_nodes(true),
985 _inlining_progress(false),
986 _inlining_incrementally(false),
987 _has_reserved_stack_access(false),
988 _has_circular_inline_type(false),
989 #ifndef PRODUCT
990 _igv_idx(0),
991 _trace_opto_output(directive->TraceOptoOutputOption),
992 #endif
993 _clinit_barrier_on_entry(false),
994 _stress_seed(0),
995 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
996 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
997 _env(ci_env),
998 _directive(directive),
999 _log(ci_env->log()),
1000 _first_failure_details(nullptr),
1001 _reachability_fences(comp_arena(), 8, 0, nullptr),
1002 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
1003 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
1004 _congraph(nullptr),
1005 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
1006 _unique(0),
1007 _dead_node_count(0),
1008 _dead_node_list(comp_arena()),
1009 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
1010 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
1011 _node_arena(&_node_arena_one),
1012 _mach_constant_base_node(nullptr),
1013 _Compile_types(mtCompiler, Arena::Tag::tag_type),
1014 _initial_gvn(nullptr),
1015 _igvn_worklist(nullptr),
1016 _types(nullptr),
1017 _node_hash(nullptr),
1018 _has_mh_late_inlines(false),
1019 _oom(false),
1020 _replay_inline_data(nullptr),
1021 _inline_printer(this),
1022 _java_calls(0),
1023 _inner_loops(0),
1024 _FIRST_STACK_mask(comp_arena()),
1025 _interpreter_frame_size(0),
1026 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
1027 _output(nullptr),
1028 #ifndef PRODUCT
1029 _in_dump_cnt(0),
1030 #endif
1031 _allowed_reasons(0) {
1032 C = this;
1033
1034 // try to reuse an existing stub
1035 if (c2_do_stub_init_complete) {
1036 BlobId blob_id = StubInfo::blob(_stub_id);
1037 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
1038 if (blob != nullptr) {
1039 RuntimeStub* rs = blob->as_runtime_stub();
1040 _stub_entry_point = rs->entry_point();
1041 return;
1042 }
1043 }
1044
1045 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
1046 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
1047
1048 #ifndef PRODUCT
1049 set_print_assembly(PrintFrameConverterAssembly);
1050 set_parsed_irreducible_loop(false);
1051 #else
1052 set_print_assembly(false); // Must initialize.
1053 #endif
1054 set_has_irreducible_loop(false); // no loops
1055
1056 CompileWrapper cw(this);
1057 Init(/*do_aliasing=*/ false);
1058 init_tf((*generator)());
1059
1060 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1061 _types = new (comp_arena()) Type_Array(comp_arena());
1062 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1063
1064 if (StressLCM || StressGCM || StressBailout) {
1065 initialize_stress_seed(directive);
1066 }
1067
1068 {
1069 PhaseGVN gvn;
1070 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1071 gvn.transform(top());
1072
1073 GraphKit kit;
1074 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1075 }
1076
1077 NOT_PRODUCT( verify_graph_edges(); )
1078
1079 Code_Gen();
1080
1081 // First successful stub generation will set it to `true`
1082 // and it will stay `true` after that.
1083 c2_do_stub_init_complete = c2_do_stub_init_complete || (_stub_entry_point != nullptr);
1084 }
1085
1086 Compile::~Compile() {
1087 delete _first_failure_details;
1088 };
1089
1090 //------------------------------Init-------------------------------------------
1091 // Prepare for a single compilation
1092 void Compile::Init(bool aliasing) {
1093 _do_aliasing = aliasing;
1094 _unique = 0;
1095 _regalloc = nullptr;
1096
1097 _tf = nullptr; // filled in later
1098 _top = nullptr; // cached later
1099 _matcher = nullptr; // filled in later
1100 _cfg = nullptr; // filled in later
1101
1102 _node_note_array = nullptr;
1103 _default_node_notes = nullptr;
1104 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1105
1106 _immutable_memory = nullptr; // filled in at first inquiry
1107
1108 #ifdef ASSERT
1109 _phase_optimize_finished = false;
1110 _phase_verify_ideal_loop = false;
1111 _exception_backedge = false;
1112 _type_verify = nullptr;
1113 #endif
1114
1115 // Globally visible Nodes
1116 // First set TOP to null to give safe behavior during creation of RootNode
1117 set_cached_top_node(nullptr);
1118 set_root(new RootNode());
1119 // Now that you have a Root to point to, create the real TOP
1120 set_cached_top_node( new ConNode(Type::TOP) );
1121 set_recent_alloc(nullptr, nullptr);
1122
1123 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1124 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1125 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1126 env()->set_dependencies(new Dependencies(env()));
1127
1128 _fixed_slots = 0;
1129 set_has_split_ifs(false);
1130 set_has_loops(false); // first approximation
1131 set_has_stringbuilder(false);
1132 set_has_boxed_value(false);
1133 _trap_can_recompile = false; // no traps emitted yet
1134 _major_progress = true; // start out assuming good things will happen
1135 set_has_unsafe_access(false);
1136 set_max_vector_size(0);
1137 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1138 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1139 set_decompile_count(0);
1140
1141 #ifndef PRODUCT
1142 _phase_counter = 0;
1143 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1144 #endif
1145
1146 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1147 _loop_opts_cnt = LoopOptsCount;
1148 _has_flat_accesses = false;
1149 _flat_accesses_share_alias = true;
1150 _scalarize_in_safepoints = false;
1151 _needs_nm_slot = false;
1152
1153 set_do_inlining(Inline);
1154 set_max_inline_size(MaxInlineSize);
1155 set_freq_inline_size(FreqInlineSize);
1156 set_do_scheduling(OptoScheduling);
1157
1158 set_do_vector_loop(false);
1159 set_has_monitors(false);
1160 set_has_scoped_access(false);
1161
1162 if (AllowVectorizeOnDemand) {
1163 if (has_method() && _directive->VectorizeOption) {
1164 set_do_vector_loop(true);
1165 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());})
1166 } else if (has_method() && method()->name() != nullptr &&
1167 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1168 set_do_vector_loop(true);
1169 }
1170 }
1171 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1172 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());})
1173
1174 _max_node_limit = _directive->MaxNodeLimitOption;
1175
1176 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1177 set_clinit_barrier_on_entry(true);
1178 }
1179 if (debug_info()->recording_non_safepoints()) {
1180 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1181 (comp_arena(), 8, 0, nullptr));
1182 set_default_node_notes(Node_Notes::make(this));
1183 }
1184
1185 const int grow_ats = 16;
1186 _max_alias_types = grow_ats;
1187 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1188 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1189 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1190 {
1191 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1192 }
1193 // Initialize the first few types.
1194 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1195 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1196 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1197 _num_alias_types = AliasIdxRaw+1;
1198 // Zero out the alias type cache.
1199 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1200 // A null adr_type hits in the cache right away. Preload the right answer.
1201 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1202 }
1203
1204 #ifdef ASSERT
1205 // Verify that the current StartNode is valid.
1206 void Compile::verify_start(StartNode* s) const {
1207 assert(failing_internal() || s == start(), "should be StartNode");
1208 }
1209 #endif
1210
1211 /**
1212 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1213 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1214 * the ideal graph.
1215 */
1216 StartNode* Compile::start() const {
1217 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1218 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1219 Node* start = root()->fast_out(i);
1220 if (start->is_Start()) {
1221 return start->as_Start();
1222 }
1223 }
1224 fatal("Did not find Start node!");
1225 return nullptr;
1226 }
1227
1228 //-------------------------------immutable_memory-------------------------------------
1229 // Access immutable memory
1230 Node* Compile::immutable_memory() {
1231 if (_immutable_memory != nullptr) {
1232 return _immutable_memory;
1233 }
1234 StartNode* s = start();
1235 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1236 Node *p = s->fast_out(i);
1237 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1238 _immutable_memory = p;
1239 return _immutable_memory;
1240 }
1241 }
1242 ShouldNotReachHere();
1243 return nullptr;
1244 }
1245
1246 //----------------------set_cached_top_node------------------------------------
1247 // Install the cached top node, and make sure Node::is_top works correctly.
1248 void Compile::set_cached_top_node(Node* tn) {
1249 if (tn != nullptr) verify_top(tn);
1250 Node* old_top = _top;
1251 _top = tn;
1252 // Calling Node::setup_is_top allows the nodes the chance to adjust
1253 // their _out arrays.
1254 if (_top != nullptr) _top->setup_is_top();
1255 if (old_top != nullptr) old_top->setup_is_top();
1256 assert(_top == nullptr || top()->is_top(), "");
1257 }
1258
1259 #ifdef ASSERT
1260 uint Compile::count_live_nodes_by_graph_walk() {
1261 Unique_Node_List useful(comp_arena());
1262 // Get useful node list by walking the graph.
1263 identify_useful_nodes(useful);
1264 return useful.size();
1265 }
1266
1267 void Compile::print_missing_nodes() {
1268
1269 // Return if CompileLog is null and PrintIdealNodeCount is false.
1270 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1271 return;
1272 }
1273
1274 // This is an expensive function. It is executed only when the user
1275 // specifies VerifyIdealNodeCount option or otherwise knows the
1276 // additional work that needs to be done to identify reachable nodes
1277 // by walking the flow graph and find the missing ones using
1278 // _dead_node_list.
1279
1280 Unique_Node_List useful(comp_arena());
1281 // Get useful node list by walking the graph.
1282 identify_useful_nodes(useful);
1283
1284 uint l_nodes = C->live_nodes();
1285 uint l_nodes_by_walk = useful.size();
1286
1287 if (l_nodes != l_nodes_by_walk) {
1288 if (_log != nullptr) {
1289 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1290 _log->stamp();
1291 _log->end_head();
1292 }
1293 VectorSet& useful_member_set = useful.member_set();
1294 int last_idx = l_nodes_by_walk;
1295 for (int i = 0; i < last_idx; i++) {
1296 if (useful_member_set.test(i)) {
1297 if (_dead_node_list.test(i)) {
1298 if (_log != nullptr) {
1299 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1300 }
1301 if (PrintIdealNodeCount) {
1302 // Print the log message to tty
1303 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1304 useful.at(i)->dump();
1305 }
1306 }
1307 }
1308 else if (! _dead_node_list.test(i)) {
1309 if (_log != nullptr) {
1310 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1311 }
1312 if (PrintIdealNodeCount) {
1313 // Print the log message to tty
1314 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1315 }
1316 }
1317 }
1318 if (_log != nullptr) {
1319 _log->tail("mismatched_nodes");
1320 }
1321 }
1322 }
1323 void Compile::record_modified_node(Node* n) {
1324 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1325 _modified_nodes->push(n);
1326 }
1327 }
1328
1329 void Compile::remove_modified_node(Node* n) {
1330 if (_modified_nodes != nullptr) {
1331 _modified_nodes->remove(n);
1332 }
1333 }
1334 #endif
1335
1336 #ifndef PRODUCT
1337 void Compile::verify_top(Node* tn) const {
1338 if (tn != nullptr) {
1339 assert(tn->is_Con(), "top node must be a constant");
1340 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1341 assert(tn->in(0) != nullptr, "must have live top node");
1342 }
1343 }
1344 #endif
1345
1346
1347 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1348
1349 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1350 guarantee(arr != nullptr, "");
1351 int num_blocks = arr->length();
1352 if (grow_by < num_blocks) grow_by = num_blocks;
1353 int num_notes = grow_by * _node_notes_block_size;
1354 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1355 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1356 while (num_notes > 0) {
1357 arr->append(notes);
1358 notes += _node_notes_block_size;
1359 num_notes -= _node_notes_block_size;
1360 }
1361 assert(num_notes == 0, "exact multiple, please");
1362 }
1363
1364 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1365 if (source == nullptr || dest == nullptr) return false;
1366
1367 if (dest->is_Con())
1368 return false; // Do not push debug info onto constants.
1369
1370 #ifdef ASSERT
1371 // Leave a bread crumb trail pointing to the original node:
1372 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1373 dest->set_debug_orig(source);
1374 }
1375 #endif
1376
1377 if (node_note_array() == nullptr)
1378 return false; // Not collecting any notes now.
1379
1380 // This is a copy onto a pre-existing node, which may already have notes.
1381 // If both nodes have notes, do not overwrite any pre-existing notes.
1382 Node_Notes* source_notes = node_notes_at(source->_idx);
1383 if (source_notes == nullptr || source_notes->is_clear()) return false;
1384 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1385 if (dest_notes == nullptr || dest_notes->is_clear()) {
1386 return set_node_notes_at(dest->_idx, source_notes);
1387 }
1388
1389 Node_Notes merged_notes = (*source_notes);
1390 // The order of operations here ensures that dest notes will win...
1391 merged_notes.update_from(dest_notes);
1392 return set_node_notes_at(dest->_idx, &merged_notes);
1393 }
1394
1395
1396 //--------------------------allow_range_check_smearing-------------------------
1397 // Gating condition for coalescing similar range checks.
1398 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1399 // single covering check that is at least as strong as any of them.
1400 // If the optimization succeeds, the simplified (strengthened) range check
1401 // will always succeed. If it fails, we will deopt, and then give up
1402 // on the optimization.
1403 bool Compile::allow_range_check_smearing() const {
1404 // If this method has already thrown a range-check,
1405 // assume it was because we already tried range smearing
1406 // and it failed.
1407 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1408 return !already_trapped;
1409 }
1410
1411
1412 //------------------------------flatten_alias_type-----------------------------
1413 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1414 assert(do_aliasing(), "Aliasing should be enabled");
1415 int offset = tj->offset();
1416 TypePtr::PTR ptr = tj->ptr();
1417
1418 // Known instance (scalarizable allocation) alias only with itself.
1419 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1420 tj->is_oopptr()->is_known_instance();
1421
1422 // Process weird unsafe references.
1423 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1424 assert(InlineUnsafeOps || StressReflectiveCode || UseAcmpFastPath, "indeterminate pointers come only from unsafe ops");
1425 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1426 tj = TypeOopPtr::BOTTOM;
1427 ptr = tj->ptr();
1428 offset = tj->offset();
1429 }
1430
1431 // Array pointers need some flattening
1432 const TypeAryPtr* ta = tj->isa_aryptr();
1433 if( ta && is_known_inst ) {
1434 if ( offset != Type::OffsetBot &&
1435 offset > arrayOopDesc::length_offset_in_bytes() ) {
1436 offset = Type::OffsetBot; // Flatten constant access into array body only
1437 tj = ta = ta->
1438 remove_speculative()->
1439 cast_to_ptr_type(ptr)->
1440 with_offset(offset);
1441 }
1442 } else if (ta != nullptr) {
1443 // Common slices
1444 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1445 return TypeAryPtr::RANGE;
1446 } else if (offset == oopDesc::klass_offset_in_bytes()) {
1447 return TypeInstPtr::KLASS;
1448 } else if (offset == oopDesc::mark_offset_in_bytes()) {
1449 return TypeInstPtr::MARK;
1450 }
1451
1452 // Remove size and stability
1453 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());
1454 // Remove ptr, const_oop, and offset
1455 if (ta->elem() == Type::BOTTOM) {
1456 // Bottom array (meet of int[] and byte[] for example), accesses to it will be done with
1457 // Unsafe. This should alias with all arrays. For now just leave it as it is (this is
1458 // incorrect, see JDK-8331133).
1459 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, false, Type::Offset::bottom);
1460 } else if (ta->elem()->make_oopptr() != nullptr) {
1461 // Object arrays, keep field_offset
1462 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, ta->klass_is_exact(), Type::Offset::bottom, Type::Offset(ta->field_offset()));
1463 } else {
1464 // Primitive arrays
1465 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, ta->exact_klass(), true, Type::Offset::bottom);
1466 }
1467
1468 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1469 // cannot be distinguished by bytecode alone.
1470 if (ta->elem() == TypeInt::BOOL) {
1471 tj = ta = TypeAryPtr::BYTES;
1472 }
1473
1474 // All arrays of references share the same slice
1475 if (!ta->is_flat() && ta->elem()->make_oopptr() != nullptr) {
1476 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, TypeInt::POS, false, false, true, true, true);
1477 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, tary, nullptr, false, Type::Offset::bottom);
1478 }
1479
1480 if (ta->is_flat()) {
1481 if (_flat_accesses_share_alias) {
1482 // Initially all flattened array accesses share a single slice
1483 tj = ta = TypeAryPtr::INLINES;
1484 } else {
1485 // Flat accesses are always exact
1486 tj = ta = ta->cast_to_exactness(true);
1487 }
1488 }
1489 }
1490
1491 // Oop pointers need some flattening
1492 const TypeInstPtr *to = tj->isa_instptr();
1493 if (to && to != TypeOopPtr::BOTTOM) {
1494 ciInstanceKlass* ik = to->instance_klass();
1495 tj = to = to->cast_to_maybe_flat_in_array(); // flatten to maybe flat in array
1496 if( ptr == TypePtr::Constant ) {
1497 if (ik != ciEnv::current()->Class_klass() ||
1498 offset < ik->layout_helper_size_in_bytes()) {
1499 // No constant oop pointers (such as Strings); they alias with
1500 // unknown strings.
1501 assert(!is_known_inst, "not scalarizable allocation");
1502 tj = to = to->
1503 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1504 remove_speculative()->
1505 cast_to_ptr_type(TypePtr::BotPTR)->
1506 cast_to_exactness(false);
1507 }
1508 } else if( is_known_inst ) {
1509 tj = to; // Keep NotNull and klass_is_exact for instance type
1510 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1511 // During the 2nd round of IterGVN, NotNull castings are removed.
1512 // Make sure the Bottom and NotNull variants alias the same.
1513 // Also, make sure exact and non-exact variants alias the same.
1514 tj = to = to->
1515 remove_speculative()->
1516 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1517 cast_to_ptr_type(TypePtr::BotPTR)->
1518 cast_to_exactness(false);
1519 }
1520 if (to->speculative() != nullptr) {
1521 tj = to = to->remove_speculative();
1522 }
1523 // Canonicalize the holder of this field
1524 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1525 // First handle header references such as a LoadKlassNode, even if the
1526 // object's klass is unloaded at compile time (4965979).
1527 if (!is_known_inst) { // Do it only for non-instance types
1528 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1529 }
1530 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1531 // Static fields are in the space above the normal instance
1532 // fields in the java.lang.Class instance.
1533 if (ik != ciEnv::current()->Class_klass()) {
1534 to = nullptr;
1535 tj = TypeOopPtr::BOTTOM;
1536 offset = tj->offset();
1537 }
1538 } else {
1539 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1540 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1541 assert(tj->offset() == offset, "no change to offset expected");
1542 bool xk = to->klass_is_exact();
1543 int instance_id = to->instance_id();
1544
1545 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1546 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1547 // its interfaces are included.
1548 if (xk && ik->equals(canonical_holder)) {
1549 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id,
1550 TypePtr::MaybeFlat), "exact type should be canonical type");
1551 } else {
1552 assert(xk || !is_known_inst, "Known instance should be exact type");
1553 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id,
1554 TypePtr::MaybeFlat);
1555 }
1556 }
1557 }
1558
1559 // Klass pointers to object array klasses need some flattening
1560 const TypeKlassPtr *tk = tj->isa_klassptr();
1561 if( tk ) {
1562 // If we are referencing a field within a Klass, we need
1563 // to assume the worst case of an Object. Both exact and
1564 // inexact types must flatten to the same alias class so
1565 // use NotNull as the PTR.
1566 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1567 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1568 env()->Object_klass(),
1569 Type::Offset(offset),
1570 TypePtr::MaybeFlat);
1571 }
1572
1573 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1574 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1575 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1576 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset), TypePtr::MaybeFlat);
1577 } else {
1578 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());
1579 }
1580 }
1581 // Check for precise loads from the primary supertype array and force them
1582 // to the supertype cache alias index. Check for generic array loads from
1583 // the primary supertype array and also force them to the supertype cache
1584 // alias index. Since the same load can reach both, we need to merge
1585 // these 2 disparate memories into the same alias class. Since the
1586 // primary supertype array is read-only, there's no chance of confusion
1587 // where we bypass an array load and an array store.
1588 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1589 if (offset == Type::OffsetBot ||
1590 (offset >= primary_supers_offset &&
1591 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1592 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1593 offset = in_bytes(Klass::secondary_super_cache_offset());
1594 tj = tk = tk->with_offset(offset);
1595 }
1596 }
1597
1598 // Flatten all Raw pointers together.
1599 if (tj->base() == Type::RawPtr)
1600 tj = TypeRawPtr::BOTTOM;
1601
1602 if (tj->base() == Type::AnyPtr)
1603 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1604
1605 offset = tj->offset();
1606 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1607
1608 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1609 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1610 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1611 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1612 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1613 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1614 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1615 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1616 assert( tj->ptr() != TypePtr::TopPTR &&
1617 tj->ptr() != TypePtr::AnyNull &&
1618 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1619 // assert( tj->ptr() != TypePtr::Constant ||
1620 // tj->base() == Type::RawPtr ||
1621 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1622
1623 return tj;
1624 }
1625
1626 void Compile::AliasType::Init(int i, const TypePtr* at) {
1627 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1628 _index = i;
1629 _adr_type = at;
1630 _field = nullptr;
1631 _element = nullptr;
1632 _is_rewritable = true; // default
1633 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1634 if (atoop != nullptr && atoop->is_known_instance()) {
1635 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1636 _general_index = Compile::current()->get_alias_index(gt);
1637 } else {
1638 _general_index = 0;
1639 }
1640 }
1641
1642 BasicType Compile::AliasType::basic_type() const {
1643 if (element() != nullptr) {
1644 const Type* element = adr_type()->is_aryptr()->elem();
1645 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1646 } if (field() != nullptr) {
1647 return field()->layout_type();
1648 } else {
1649 return T_ILLEGAL; // unknown
1650 }
1651 }
1652
1653 //---------------------------------print_on------------------------------------
1654 #ifndef PRODUCT
1655 void Compile::AliasType::print_on(outputStream* st) {
1656 if (index() < 10)
1657 st->print("@ <%d> ", index());
1658 else st->print("@ <%d>", index());
1659 st->print(is_rewritable() ? " " : " RO");
1660 int offset = adr_type()->offset();
1661 if (offset == Type::OffsetBot)
1662 st->print(" +any");
1663 else st->print(" +%-3d", offset);
1664 st->print(" in ");
1665 adr_type()->dump_on(st);
1666 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1667 if (field() != nullptr && tjp) {
1668 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1669 tjp->offset() != field()->offset_in_bytes()) {
1670 st->print(" != ");
1671 field()->print();
1672 st->print(" ***");
1673 }
1674 }
1675 }
1676
1677 void print_alias_types() {
1678 Compile* C = Compile::current();
1679 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1680 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1681 C->alias_type(idx)->print_on(tty);
1682 tty->cr();
1683 }
1684 }
1685 #endif
1686
1687
1688 //----------------------------probe_alias_cache--------------------------------
1689 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1690 intptr_t key = (intptr_t) adr_type;
1691 key ^= key >> logAliasCacheSize;
1692 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1693 }
1694
1695
1696 //-----------------------------grow_alias_types--------------------------------
1697 void Compile::grow_alias_types() {
1698 const int old_ats = _max_alias_types; // how many before?
1699 const int new_ats = old_ats; // how many more?
1700 const int grow_ats = old_ats+new_ats; // how many now?
1701 _max_alias_types = grow_ats;
1702 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), _alias_types, old_ats, grow_ats);
1703 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1704 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1705 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1706 }
1707
1708
1709 //--------------------------------find_alias_type------------------------------
1710 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1711 if (!do_aliasing()) {
1712 return alias_type(AliasIdxBot);
1713 }
1714
1715 AliasCacheEntry* ace = nullptr;
1716 if (!uncached) {
1717 ace = probe_alias_cache(adr_type);
1718 if (ace->_adr_type == adr_type) {
1719 return alias_type(ace->_index);
1720 }
1721 }
1722
1723 // Handle special cases.
1724 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1725 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1726
1727 // Do it the slow way.
1728 const TypePtr* flat = flatten_alias_type(adr_type);
1729
1730 #ifdef ASSERT
1731 {
1732 ResourceMark rm;
1733 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1734 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1735 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1736 Type::str(adr_type));
1737 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1738 const TypeOopPtr* foop = flat->is_oopptr();
1739 // Scalarizable allocations have exact klass always.
1740 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1741 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1742 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1743 Type::str(foop), Type::str(xoop));
1744 }
1745 }
1746 #endif
1747
1748 int idx = AliasIdxTop;
1749 for (int i = 0; i < num_alias_types(); i++) {
1750 if (alias_type(i)->adr_type() == flat) {
1751 idx = i;
1752 break;
1753 }
1754 }
1755
1756 if (idx == AliasIdxTop) {
1757 if (no_create) return nullptr;
1758 // Grow the array if necessary.
1759 if (_num_alias_types == _max_alias_types) grow_alias_types();
1760 // Add a new alias type.
1761 idx = _num_alias_types++;
1762 _alias_types[idx]->Init(idx, flat);
1763 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1764 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1765 if (flat->isa_instptr()) {
1766 if (flat->offset() == java_lang_Class::klass_offset()
1767 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1768 alias_type(idx)->set_rewritable(false);
1769 }
1770 ciField* field = nullptr;
1771 if (flat->isa_aryptr()) {
1772 #ifdef ASSERT
1773 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1774 // (T_BYTE has the weakest alignment and size restrictions...)
1775 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1776 #endif
1777 const Type* elemtype = flat->is_aryptr()->elem();
1778 if (flat->offset() == TypePtr::OffsetBot) {
1779 alias_type(idx)->set_element(elemtype);
1780 }
1781 int field_offset = flat->is_aryptr()->field_offset().get();
1782 if (flat->is_flat() &&
1783 field_offset != Type::OffsetBot) {
1784 ciInlineKlass* vk = elemtype->inline_klass();
1785 field_offset += vk->payload_offset();
1786 field = vk->get_field_by_offset(field_offset, false);
1787 }
1788 }
1789 if (flat->isa_klassptr()) {
1790 if (UseCompactObjectHeaders) {
1791 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1792 alias_type(idx)->set_rewritable(false);
1793 }
1794 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1795 alias_type(idx)->set_rewritable(false);
1796 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1797 alias_type(idx)->set_rewritable(false);
1798 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1799 alias_type(idx)->set_rewritable(false);
1800 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1801 alias_type(idx)->set_rewritable(false);
1802 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1803 alias_type(idx)->set_rewritable(false);
1804 }
1805
1806 if (flat->isa_instklassptr()) {
1807 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1808 alias_type(idx)->set_rewritable(false);
1809 }
1810 }
1811 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1812 // but the base pointer type is not distinctive enough to identify
1813 // references into JavaThread.)
1814
1815 // Check for final fields.
1816 const TypeInstPtr* tinst = flat->isa_instptr();
1817 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1818 if (tinst->const_oop() != nullptr &&
1819 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1820 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1821 // static field
1822 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1823 field = k->get_field_by_offset(tinst->offset(), true);
1824 } else if (tinst->is_inlinetypeptr()) {
1825 // Inline type field
1826 ciInlineKlass* vk = tinst->inline_klass();
1827 field = vk->get_field_by_offset(tinst->offset(), false);
1828 } else {
1829 ciInstanceKlass *k = tinst->instance_klass();
1830 field = k->get_field_by_offset(tinst->offset(), false);
1831 }
1832 }
1833 assert(field == nullptr ||
1834 original_field == nullptr ||
1835 (field->holder() == original_field->holder() &&
1836 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1837 field->is_static() == original_field->is_static()), "wrong field?");
1838 // Set field() and is_rewritable() attributes.
1839 if (field != nullptr) {
1840 alias_type(idx)->set_field(field);
1841 if (flat->isa_aryptr()) {
1842 // Fields of flat arrays are rewritable although they are declared final
1843 assert(flat->is_flat(), "must be a flat array");
1844 alias_type(idx)->set_rewritable(true);
1845 }
1846 }
1847 }
1848
1849 // Fill the cache for next time.
1850 if (!uncached) {
1851 ace->_adr_type = adr_type;
1852 ace->_index = idx;
1853 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1854
1855 // Might as well try to fill the cache for the flattened version, too.
1856 AliasCacheEntry* face = probe_alias_cache(flat);
1857 if (face->_adr_type == nullptr) {
1858 face->_adr_type = flat;
1859 face->_index = idx;
1860 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1861 }
1862 }
1863
1864 return alias_type(idx);
1865 }
1866
1867
1868 Compile::AliasType* Compile::alias_type(ciField* field) {
1869 const TypeOopPtr* t;
1870 if (field->is_static())
1871 t = TypeInstPtr::make(field->holder()->java_mirror());
1872 else
1873 t = TypeOopPtr::make_from_klass_raw(field->holder());
1874 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1875 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1876 return atp;
1877 }
1878
1879
1880 //------------------------------have_alias_type--------------------------------
1881 bool Compile::have_alias_type(const TypePtr* adr_type) {
1882 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1883 if (ace->_adr_type == adr_type) {
1884 return true;
1885 }
1886
1887 // Handle special cases.
1888 if (adr_type == nullptr) return true;
1889 if (adr_type == TypePtr::BOTTOM) return true;
1890
1891 return find_alias_type(adr_type, true, nullptr) != nullptr;
1892 }
1893
1894 //-----------------------------must_alias--------------------------------------
1895 // True if all values of the given address type are in the given alias category.
1896 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1897 if (alias_idx == AliasIdxBot) return true; // the universal category
1898 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1899 if (alias_idx == AliasIdxTop) return false; // the empty category
1900 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1901
1902 // the only remaining possible overlap is identity
1903 int adr_idx = get_alias_index(adr_type);
1904 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1905 assert(adr_idx == alias_idx ||
1906 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1907 && adr_type != TypeOopPtr::BOTTOM),
1908 "should not be testing for overlap with an unsafe pointer");
1909 return adr_idx == alias_idx;
1910 }
1911
1912 //------------------------------can_alias--------------------------------------
1913 // True if any values of the given address type are in the given alias category.
1914 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1915 if (alias_idx == AliasIdxTop) return false; // the empty category
1916 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1917 // Known instance doesn't alias with bottom memory
1918 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1919 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1920
1921 // the only remaining possible overlap is identity
1922 int adr_idx = get_alias_index(adr_type);
1923 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1924 return adr_idx == alias_idx;
1925 }
1926
1927 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1928 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1929 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1930 if (parse_predicate_count() == 0) {
1931 return;
1932 }
1933 for (int i = 0; i < parse_predicate_count(); i++) {
1934 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1935 parse_predicate->mark_useless(igvn);
1936 }
1937 _parse_predicates.clear();
1938 }
1939
1940 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1941 if (!n->for_post_loop_opts_igvn()) {
1942 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1943 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1944 _for_post_loop_igvn.append(n);
1945 }
1946 }
1947
1948 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1949 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1950 _for_post_loop_igvn.remove(n);
1951 }
1952
1953 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1954 // Verify that all previous optimizations produced a valid graph
1955 // at least to this point, even if no loop optimizations were done.
1956 PhaseIdealLoop::verify(igvn);
1957
1958 if (_print_phase_loop_opts) {
1959 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1960 }
1961 C->set_post_loop_opts_phase(); // no more loop opts allowed
1962
1963 assert(!C->major_progress(), "not cleared");
1964
1965 if (_for_post_loop_igvn.length() > 0) {
1966 while (_for_post_loop_igvn.length() > 0) {
1967 Node* n = _for_post_loop_igvn.pop();
1968 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1969 igvn._worklist.push(n);
1970 }
1971 igvn.optimize();
1972 if (failing()) return;
1973 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1974 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1975
1976 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1977 if (C->major_progress()) {
1978 C->clear_major_progress(); // ensure that major progress is now clear
1979 }
1980 }
1981 }
1982
1983 void Compile::add_inline_type(Node* n) {
1984 assert(n->is_InlineType(), "unexpected node");
1985 _inline_type_nodes.push(n);
1986 }
1987
1988 void Compile::remove_inline_type(Node* n) {
1989 assert(n->is_InlineType(), "unexpected node");
1990 if (_inline_type_nodes.contains(n)) {
1991 _inline_type_nodes.remove(n);
1992 }
1993 }
1994
1995 // Does the return value keep otherwise useless inline type allocations alive?
1996 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1997 ResourceMark rm;
1998 Unique_Node_List wq;
1999 wq.push(ret_val);
2000 bool some_allocations = false;
2001 for (uint i = 0; i < wq.size(); i++) {
2002 Node* n = wq.at(i);
2003 if (n->outcnt() > 1) {
2004 // Some other use for the allocation
2005 return false;
2006 } else if (n->is_InlineType()) {
2007 wq.push(n->in(1));
2008 } else if (n->is_Phi()) {
2009 for (uint j = 1; j < n->req(); j++) {
2010 wq.push(n->in(j));
2011 }
2012 } else if (n->is_CheckCastPP() &&
2013 n->in(1)->is_Proj() &&
2014 n->in(1)->in(0)->is_Allocate()) {
2015 some_allocations = true;
2016 } else if (n->is_CheckCastPP() || n->is_CastPP()) {
2017 wq.push(n->in(1));
2018 }
2019 }
2020 return some_allocations;
2021 }
2022
2023 bool Compile::clear_argument_if_only_used_as_buffer_at_calls(Node* result_cast, PhaseIterGVN& igvn) {
2024 ResourceMark rm;
2025 Unique_Node_List wq;
2026 wq.push(result_cast);
2027 Node_List calls;
2028 for (uint i = 0; i < wq.size(); ++i) {
2029 Node* n = wq.at(i);
2030 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2031 Node* u = n->fast_out(j);
2032 if (u->is_Phi()) {
2033 wq.push(u);
2034 } else if (u->is_InlineType() && u->as_InlineType()->get_oop() == n) {
2035 wq.push(u);
2036 } else if (u->is_CallJava()) {
2037 CallJavaNode* call = u->as_CallJava();
2038 if (call->method() != nullptr && call->method()->mismatch()) {
2039 return false;
2040 }
2041 uint nargs = call->tf()->domain_cc()->cnt();
2042 for (uint k = TypeFunc::Parms; k < nargs; k++) {
2043 Node* in = call->in(k);
2044 if (in == n && (call->method() == nullptr || !call->method()->is_scalarized_buffer_arg(k - TypeFunc::Parms))) {
2045 return false;
2046 }
2047 }
2048 calls.push(call);
2049 } else if (u->Opcode() == Op_EncodeP) {
2050 wq.push(u);
2051 } else if (u->is_AddP()) {
2052 wq.push(u);
2053 } else if (u->is_Store() && u->in(MemNode::Address) == n) {
2054 // storing to the buffer is fine
2055 } else if (u->is_SafePoint()) {
2056 SafePointNode* sfpt = u->as_SafePoint();
2057 int input = u->find_edge(n);
2058 JVMState* jvms = sfpt->jvms();
2059 if (jvms != nullptr) {
2060 if (input < (int)jvms->debug_start()) {
2061 return false;
2062 }
2063 }
2064 } else {
2065 return false;
2066 }
2067 }
2068 }
2069 for (uint i = 0; i < calls.size(); ++i) {
2070 CallJavaNode* call = calls.at(i)->as_CallJava();
2071 uint nargs = call->tf()->domain_cc()->cnt();
2072 for (uint k = TypeFunc::Parms; k < nargs; k++) {
2073 Node* in = call->in(k);
2074 if (wq.member(in)) {
2075 assert(call->method()->is_scalarized_buffer_arg(k - TypeFunc::Parms), "only buffer argument removed here");
2076 igvn.replace_input_of(call, k, igvn.zerocon(T_OBJECT));
2077 }
2078 }
2079 }
2080 return true;
2081 }
2082
2083 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
2084 // Make sure that the return value does not keep an otherwise unused allocation alive
2085 if (tf()->returns_inline_type_as_fields()) {
2086 Node* ret = nullptr;
2087 for (uint i = 1; i < root()->req(); i++) {
2088 Node* in = root()->in(i);
2089 if (in->Opcode() == Op_Return) {
2090 assert(ret == nullptr, "only one return");
2091 ret = in;
2092 }
2093 }
2094 if (ret != nullptr) {
2095 Node* ret_val = ret->in(TypeFunc::Parms);
2096 if (igvn.type(ret_val)->isa_oopptr() &&
2097 return_val_keeps_allocations_alive(ret_val)) {
2098 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
2099 assert(ret_val->outcnt() == 0, "should be dead now");
2100 igvn.remove_dead_node(ret_val, PhaseIterGVN::NodeOrigin::Graph);
2101 }
2102 }
2103 }
2104 // if a newly allocated object is a value that's only passed as argument to calls as (possibly null) buffers, then
2105 // clear the call argument inputs so the allocation node can be removed
2106 for (int i = 0; i < C->macro_count(); ++i) {
2107 Node* macro_node = C->macro_node(i);
2108 if (macro_node->Opcode() == Op_Allocate) {
2109 AllocateNode* allocate = macro_node->as_Allocate();
2110 Node* result_cast = allocate->result_cast();
2111 if (result_cast != nullptr) {
2112 const Type* result_type = igvn.type(result_cast);
2113 if (result_type->is_inlinetypeptr()) {
2114 clear_argument_if_only_used_as_buffer_at_calls(result_cast, igvn);
2115 }
2116 }
2117 }
2118 }
2119
2120 if (_inline_type_nodes.length() == 0) {
2121 // keep the graph canonical
2122 igvn.optimize();
2123 return;
2124 }
2125 // Scalarize inline types in safepoint debug info.
2126 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2127 set_scalarize_in_safepoints(true);
2128 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2129 InlineTypeNode* vt = _inline_type_nodes.at(i)->as_InlineType();
2130 vt->make_scalar_in_safepoints(&igvn);
2131 igvn.record_for_igvn(vt);
2132 }
2133 if (remove) {
2134 // Remove inline type nodes by replacing them with their oop input
2135 while (_inline_type_nodes.length() > 0) {
2136 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2137 if (vt->outcnt() == 0) {
2138 igvn.remove_dead_node(vt, PhaseIterGVN::NodeOrigin::Graph);
2139 continue;
2140 }
2141 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2142 DEBUG_ONLY(bool must_be_buffered = false);
2143 Node* u = vt->out(i);
2144 // Check if any users are blackholes. If so, rewrite them to use either the
2145 // allocated buffer, or individual components, instead of the inline type node
2146 // that goes away.
2147 if (u->is_Blackhole()) {
2148 BlackholeNode* bh = u->as_Blackhole();
2149
2150 // Unlink the old input
2151 int idx = bh->find_edge(vt);
2152 assert(idx != -1, "The edge should be there");
2153 bh->del_req(idx);
2154 --i;
2155
2156 if (vt->is_allocated(&igvn)) {
2157 // Already has the allocated instance, blackhole that
2158 bh->add_req(vt->get_oop());
2159 } else {
2160 // Not allocated yet, blackhole the components
2161 for (uint c = 0; c < vt->field_count(); c++) {
2162 bh->add_req(vt->field_value(c));
2163 }
2164 }
2165
2166 // Node modified, record for IGVN
2167 igvn.record_for_igvn(bh);
2168 }
2169 #ifdef ASSERT
2170 // Verify that inline type is buffered when replacing by oop
2171 else if (u->is_InlineType()) {
2172 // InlineType uses don't need buffering because they are about to be replaced as well
2173 } else if (u->is_Phi()) {
2174 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2175 } else {
2176 must_be_buffered = true;
2177 }
2178 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2179 vt->dump(0);
2180 u->dump(0);
2181 assert(false, "Should have been buffered");
2182 }
2183 #endif
2184 }
2185 igvn.replace_node(vt, vt->get_oop());
2186 }
2187 }
2188 igvn.optimize();
2189 }
2190
2191 void Compile::add_flat_access(Node* n) {
2192 assert(n != nullptr && (n->Opcode() == Op_LoadFlat || n->Opcode() == Op_StoreFlat), "unexpected node %s", n == nullptr ? "nullptr" : n->Name());
2193 assert(!_flat_access_nodes.contains(n), "duplicate insertion");
2194 _flat_access_nodes.push(n);
2195 }
2196
2197 void Compile::remove_flat_access(Node* n) {
2198 assert(n != nullptr && (n->Opcode() == Op_LoadFlat || n->Opcode() == Op_StoreFlat), "unexpected node %s", n == nullptr ? "nullptr" : n->Name());
2199 _flat_access_nodes.remove_if_existing(n);
2200 }
2201
2202 void Compile::process_flat_accesses(PhaseIterGVN& igvn) {
2203 assert(igvn._worklist.size() == 0, "should be empty");
2204 igvn.set_delay_transform(true);
2205 for (int i = _flat_access_nodes.length() - 1; i >= 0; i--) {
2206 Node* n = _flat_access_nodes.at(i);
2207 assert(n != nullptr, "unexpected nullptr");
2208 if (n->is_LoadFlat()) {
2209 LoadFlatNode* loadn = n->as_LoadFlat();
2210 // Expending a flat load atomically means that we get a chunk of memory spanning multiple fields
2211 // that we chop with bitwise operations. That is too subtle for some optimizations, especially
2212 // constant folding when fields are constant. If we can get a constant object from which we are
2213 // flat-loading, we can simply replace the loads at compilation-time by the field of the constant
2214 // object.
2215 ciInstance* loaded_from = nullptr;
2216 if (FoldStableValues) {
2217 const TypeOopPtr* base_type = igvn.type(loadn->base())->is_oopptr();
2218 ciObject* oop = base_type->const_oop();
2219 int off = igvn.type(loadn->ptr())->isa_ptr()->offset();
2220
2221 if (oop != nullptr && oop->is_instance()) {
2222 ciInstance* holder = oop->as_instance();
2223 ciKlass* klass = holder->klass();
2224 ciInstanceKlass* iklass = klass->as_instance_klass();
2225 ciField* field = iklass->get_non_flat_field_by_offset(off);
2226
2227 if (field->is_stable()) {
2228 ciConstant fv = holder->field_value(field);
2229 if (is_reference_type(fv.basic_type()) && fv.as_object()->is_instance()) {
2230 // The field value is an object, not null. We can use stability.
2231 loaded_from = fv.as_object()->as_instance();
2232 }
2233 }
2234 } else if (oop != nullptr && oop->is_array() && off != Type::OffsetBot) {
2235 ciArray* array = oop->as_array();
2236 ciConstant elt = array->element_value_by_offset(off);
2237 const TypeAryPtr* aryptr = base_type->is_aryptr();
2238 if (aryptr->is_stable() && aryptr->is_atomic() && is_reference_type(elt.basic_type()) && elt.as_object()->is_instance()) {
2239 loaded_from = elt.as_object()->as_instance();
2240 }
2241 }
2242 }
2243
2244 if (loaded_from != nullptr) {
2245 loadn->expand_constant(igvn, loaded_from);
2246 } else {
2247 loadn->expand_atomic(igvn);
2248 }
2249 } else {
2250 n->as_StoreFlat()->expand_atomic(igvn);
2251 }
2252 }
2253 _flat_access_nodes.clear_and_deallocate();
2254 igvn.set_delay_transform(false);
2255 igvn.optimize();
2256 }
2257
2258 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2259 DEBUG_ONLY(igvn.verify_empty_worklist(nullptr));
2260 if (!_has_flat_accesses) {
2261 return;
2262 }
2263 // Initially, all flat array accesses share the same slice to
2264 // keep dependencies with Object[] array accesses (that could be
2265 // to a flat array) correct. We're done with parsing so we
2266 // now know all flat array accesses in this compile
2267 // unit. Let's move flat array accesses to their own slice,
2268 // one per element field. This should help memory access
2269 // optimizations.
2270 ResourceMark rm;
2271 Unique_Node_List wq;
2272 wq.push(root());
2273
2274 Node_List mergememnodes;
2275 Node_List memnodes;
2276
2277 // Alias index currently shared by all flat memory accesses
2278 int index = get_alias_index(TypeAryPtr::INLINES);
2279
2280 // Find MergeMem nodes and flat array accesses
2281 for (uint i = 0; i < wq.size(); i++) {
2282 Node* n = wq.at(i);
2283 if (n->is_Mem()) {
2284 const TypePtr* adr_type = nullptr;
2285 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2286 if (adr_type == TypeAryPtr::INLINES) {
2287 memnodes.push(n);
2288 }
2289 } else if (n->is_MergeMem()) {
2290 MergeMemNode* mm = n->as_MergeMem();
2291 if (mm->memory_at(index) != mm->base_memory()) {
2292 mergememnodes.push(n);
2293 }
2294 }
2295 for (uint j = 0; j < n->req(); j++) {
2296 Node* m = n->in(j);
2297 if (m != nullptr) {
2298 wq.push(m);
2299 }
2300 }
2301 }
2302
2303 _flat_accesses_share_alias = false;
2304
2305 // We are going to change the slice for the flat array
2306 // accesses so we need to clear the cache entries that refer to
2307 // them.
2308 for (uint i = 0; i < AliasCacheSize; i++) {
2309 AliasCacheEntry* ace = &_alias_cache[i];
2310 if (ace->_adr_type != nullptr &&
2311 ace->_adr_type->is_flat()) {
2312 ace->_adr_type = nullptr;
2313 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2314 }
2315 }
2316
2317 #ifdef ASSERT
2318 for (uint i = 0; i < memnodes.size(); i++) {
2319 Node* m = memnodes.at(i);
2320 const TypePtr* adr_type = m->adr_type();
2321 m->as_Mem()->set_adr_type(adr_type);
2322 }
2323 #endif // ASSERT
2324
2325 int start_alias = num_alias_types(); // Start of new aliases
2326 Node_Stack stack(0);
2327 #ifdef ASSERT
2328 VectorSet seen(Thread::current()->resource_area());
2329 #endif
2330 // Now let's fix the memory graph so each flat array access
2331 // is moved to the right slice. Start from the MergeMem nodes.
2332 uint last = unique();
2333 for (uint i = 0; i < mergememnodes.size(); i++) {
2334 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2335 if (current->outcnt() == 0) {
2336 // This node is killed by a previous iteration
2337 continue;
2338 }
2339
2340 Node* n = current->memory_at(index);
2341 MergeMemNode* mm = nullptr;
2342 do {
2343 // Follow memory edges through memory accesses, phis and
2344 // narrow membars and push nodes on the stack. Once we hit
2345 // bottom memory, we pop element off the stack one at a
2346 // time, in reverse order, and move them to the right slice
2347 // by changing their memory edges.
2348 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() ||
2349 (n->adr_type() == TypeAryPtr::INLINES && !n->is_NarrowMemProj())) {
2350 assert(!seen.test_set(n->_idx), "");
2351 // Uses (a load for instance) will need to be moved to the
2352 // right slice as well and will get a new memory state
2353 // that we don't know yet. The use could also be the
2354 // backedge of a loop. We put a place holder node between
2355 // the memory node and its uses. We replace that place
2356 // holder with the correct memory state once we know it,
2357 // i.e. when nodes are popped off the stack. Using the
2358 // place holder make the logic work in the presence of
2359 // loops.
2360 if (n->outcnt() > 1) {
2361 Node* place_holder = nullptr;
2362 assert(!n->has_out_with(Op_Node), "");
2363 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2364 Node* u = n->out(k);
2365 if (u != current && u->_idx < last) {
2366 bool success = false;
2367 for (uint l = 0; l < u->req(); l++) {
2368 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2369 continue;
2370 }
2371 Node* in = u->in(l);
2372 if (in == n) {
2373 if (place_holder == nullptr) {
2374 place_holder = new Node(1);
2375 place_holder->init_req(0, n);
2376 }
2377 igvn.replace_input_of(u, l, place_holder);
2378 success = true;
2379 }
2380 }
2381 if (success) {
2382 --k;
2383 }
2384 }
2385 }
2386 }
2387 if (n->is_Phi()) {
2388 stack.push(n, 1);
2389 n = n->in(1);
2390 } else if (n->is_Mem()) {
2391 stack.push(n, n->req());
2392 n = n->in(MemNode::Memory);
2393 } else {
2394 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2395 stack.push(n, n->req());
2396 n = n->in(0)->in(TypeFunc::Memory);
2397 }
2398 } else {
2399 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || n->is_NarrowMemProj(), "");
2400 // Build a new MergeMem node to carry the new memory state
2401 // as we build it. IGVN should fold extraneous MergeMem
2402 // nodes.
2403 if (n->is_NarrowMemProj()) {
2404 // We need 1 NarrowMemProj for each slice of this array
2405 InitializeNode* init = n->in(0)->as_Initialize();
2406 AllocateNode* alloc = init->allocation();
2407 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2408 const TypeAryKlassPtr* klass_type = klass_node->bottom_type()->isa_aryklassptr();
2409 assert(klass_type != nullptr, "must be an array");
2410 assert(klass_type->klass_is_exact(), "must be an exact klass");
2411 ciArrayKlass* klass = klass_type->exact_klass()->as_array_klass();
2412 assert(klass->is_flat_array_klass(), "must be a flat array");
2413 ciInlineKlass* elem_klass = klass->element_klass()->as_inline_klass();
2414 const TypeAryPtr* oop_type = klass_type->as_instance_type()->is_aryptr();
2415 assert(oop_type->klass_is_exact(), "must be an exact klass");
2416
2417 Node* base = alloc->in(TypeFunc::Memory);
2418 assert(base->bottom_type() == Type::MEMORY, "the memory input of AllocateNode must be a memory");
2419 assert(base->adr_type() == TypePtr::BOTTOM, "the memory input of AllocateNode must be a bottom memory");
2420 // Must create a MergeMem with base as the base memory, do not clone if base is a
2421 // MergeMem because it may not be processed yet
2422 mm = MergeMemNode::make(nullptr);
2423 mm->set_base_memory(base);
2424 for (int j = 0; j < elem_klass->nof_nonstatic_fields(); j++) {
2425 int field_offset = elem_klass->nonstatic_field_at(j)->offset_in_bytes() - elem_klass->payload_offset();
2426 const TypeAryPtr* field_ptr = oop_type->with_offset(Type::OffsetBot)->with_field_offset(field_offset);
2427 int field_alias_idx = get_alias_index(field_ptr);
2428 assert(field_ptr == get_adr_type(field_alias_idx), "must match");
2429 Node* new_proj = new NarrowMemProjNode(init, field_ptr);
2430 igvn.register_new_node_with_optimizer(new_proj);
2431 mm->set_memory_at(field_alias_idx, new_proj);
2432 }
2433 if (!klass->is_elem_null_free()) {
2434 int nm_offset = elem_klass->null_marker_offset_in_payload();
2435 const TypeAryPtr* nm_ptr = oop_type->with_offset(Type::OffsetBot)->with_field_offset(nm_offset);
2436 int nm_alias_idx = get_alias_index(nm_ptr);
2437 assert(nm_ptr == get_adr_type(nm_alias_idx), "must match");
2438 Node* new_proj = new NarrowMemProjNode(init, nm_ptr);
2439 igvn.register_new_node_with_optimizer(new_proj);
2440 mm->set_memory_at(nm_alias_idx, new_proj);
2441 }
2442
2443 // Replace all uses of the old NarrowMemProj with the correct state
2444 MergeMemNode* new_n = MergeMemNode::make(mm);
2445 igvn.register_new_node_with_optimizer(new_n);
2446 igvn.replace_node(n, new_n);
2447 } else {
2448 // Must create a MergeMem with n as the base memory, do not clone if n is a MergeMem
2449 // because it may not be processed yet
2450 mm = MergeMemNode::make(nullptr);
2451 mm->set_base_memory(n);
2452 }
2453
2454 igvn.register_new_node_with_optimizer(mm);
2455 while (stack.size() > 0) {
2456 Node* m = stack.node();
2457 uint idx = stack.index();
2458 if (m->is_Mem()) {
2459 // Move memory node to its new slice
2460 const TypePtr* adr_type = m->adr_type();
2461 int alias = get_alias_index(adr_type);
2462 Node* prev = mm->memory_at(alias);
2463 igvn.replace_input_of(m, MemNode::Memory, prev);
2464 mm->set_memory_at(alias, m);
2465 } else if (m->is_Phi()) {
2466 // We need as many new phis as there are new aliases
2467 Node* new_phi_in = MergeMemNode::make(mm);
2468 igvn.register_new_node_with_optimizer(new_phi_in);
2469 igvn.replace_input_of(m, idx, new_phi_in);
2470 if (idx == m->req()-1) {
2471 Node* r = m->in(0);
2472 for (int j = start_alias; j < num_alias_types(); j++) {
2473 const TypePtr* adr_type = get_adr_type(j);
2474 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2475 continue;
2476 }
2477 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2478 igvn.register_new_node_with_optimizer(phi);
2479 for (uint k = 1; k < m->req(); k++) {
2480 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2481 }
2482 mm->set_memory_at(j, phi);
2483 }
2484 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2485 igvn.register_new_node_with_optimizer(base_phi);
2486 for (uint k = 1; k < m->req(); k++) {
2487 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2488 }
2489 mm->set_base_memory(base_phi);
2490 }
2491 } else {
2492 // This is a MemBarCPUOrder node from
2493 // Parse::array_load()/Parse::array_store(), in the
2494 // branch that handles flat arrays hidden under
2495 // an Object[] array. We also need one new membar per
2496 // new alias to keep the unknown access that the
2497 // membars protect properly ordered with accesses to
2498 // known flat array.
2499 assert(m->is_Proj(), "projection expected");
2500 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2501 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2502 for (int j = start_alias; j < num_alias_types(); j++) {
2503 const TypePtr* adr_type = get_adr_type(j);
2504 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2505 continue;
2506 }
2507 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2508 igvn.register_new_node_with_optimizer(mb);
2509 Node* mem = mm->memory_at(j);
2510 mb->init_req(TypeFunc::Control, ctrl);
2511 mb->init_req(TypeFunc::Memory, mem);
2512 ctrl = new ProjNode(mb, TypeFunc::Control);
2513 igvn.register_new_node_with_optimizer(ctrl);
2514 mem = new ProjNode(mb, TypeFunc::Memory);
2515 igvn.register_new_node_with_optimizer(mem);
2516 mm->set_memory_at(j, mem);
2517 }
2518 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2519 }
2520 if (idx < m->req()-1) {
2521 idx += 1;
2522 stack.set_index(idx);
2523 n = m->in(idx);
2524 break;
2525 }
2526 // Take care of place holder nodes
2527 if (m->has_out_with(Op_Node)) {
2528 Node* place_holder = m->find_out_with(Op_Node);
2529 if (place_holder != nullptr) {
2530 Node* mm_clone = mm->clone();
2531 igvn.register_new_node_with_optimizer(mm_clone);
2532 Node* hook = new Node(1);
2533 hook->init_req(0, mm);
2534 igvn.replace_node(place_holder, mm_clone);
2535 hook->destruct(&igvn);
2536 }
2537 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2538 }
2539 stack.pop();
2540 }
2541 }
2542 } while(stack.size() > 0);
2543 // Fix the memory state at the MergeMem we started from
2544 igvn.rehash_node_delayed(current);
2545 for (int j = start_alias; j < num_alias_types(); j++) {
2546 const TypePtr* adr_type = get_adr_type(j);
2547 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2548 continue;
2549 }
2550 current->set_memory_at(j, mm);
2551 }
2552 current->set_memory_at(index, current->base_memory());
2553 }
2554 igvn.optimize();
2555
2556 #ifdef ASSERT
2557 wq.clear();
2558 wq.push(root());
2559 for (uint i = 0; i < wq.size(); i++) {
2560 Node* n = wq.at(i);
2561 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2562 for (uint j = 0; j < n->req(); j++) {
2563 Node* m = n->in(j);
2564 if (m != nullptr) {
2565 wq.push(m);
2566 }
2567 }
2568 }
2569 #endif
2570
2571 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2572 }
2573
2574 void Compile::record_for_merge_stores_igvn(Node* n) {
2575 if (!n->for_merge_stores_igvn()) {
2576 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
2577 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2578 _for_merge_stores_igvn.append(n);
2579 }
2580 }
2581
2582 void Compile::remove_from_merge_stores_igvn(Node* n) {
2583 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2584 _for_merge_stores_igvn.remove(n);
2585 }
2586
2587 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
2588 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
2589 // the stores, and we merge the wrong sequence of stores.
2590 // Example:
2591 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
2592 // Apply MergeStores:
2593 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
2594 // Remove more RangeChecks:
2595 // StoreI [ StoreL ] StoreI
2596 // But now it would have been better to do this instead:
2597 // [ StoreL ] [ StoreL ]
2598 //
2599 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
2600 // since we never unset _merge_stores_phase.
2601 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
2602 C->set_merge_stores_phase();
2603
2604 if (_for_merge_stores_igvn.length() > 0) {
2605 while (_for_merge_stores_igvn.length() > 0) {
2606 Node* n = _for_merge_stores_igvn.pop();
2607 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2608 igvn._worklist.push(n);
2609 }
2610 igvn.optimize();
2611 if (failing()) return;
2612 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
2613 print_method(PHASE_AFTER_MERGE_STORES, 3);
2614 }
2615 }
2616
2617 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2618 if (OptimizeUnstableIf) {
2619 _unstable_if_traps.append(trap);
2620 }
2621 }
2622
2623 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2624 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2625 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2626 Node* n = trap->uncommon_trap();
2627 if (!useful.member(n)) {
2628 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2629 }
2630 }
2631 }
2632
2633 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2634 // or fold-compares case. Return true if succeed or not found.
2635 //
2636 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2637 // false and ask fold-compares to yield.
2638 //
2639 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2640 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2641 // when deoptimization does happen.
2642 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2643 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2644 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2645 if (trap->uncommon_trap() == unc) {
2646 if (yield && trap->modified()) {
2647 return false;
2648 }
2649 _unstable_if_traps.delete_at(i);
2650 break;
2651 }
2652 }
2653 return true;
2654 }
2655
2656 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2657 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2658 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2659 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2660 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2661 CallStaticJavaNode* unc = trap->uncommon_trap();
2662 int next_bci = trap->next_bci();
2663 bool modified = trap->modified();
2664
2665 if (next_bci != -1 && !modified) {
2666 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2667 JVMState* jvms = unc->jvms();
2668 ciMethod* method = jvms->method();
2669 ciBytecodeStream iter(method);
2670
2671 iter.force_bci(jvms->bci());
2672 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2673 Bytecodes::Code c = iter.cur_bc();
2674 Node* lhs = nullptr;
2675 Node* rhs = nullptr;
2676 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2677 lhs = unc->peek_operand(0);
2678 rhs = unc->peek_operand(1);
2679 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2680 lhs = unc->peek_operand(0);
2681 }
2682
2683 ResourceMark rm;
2684 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2685 assert(live_locals.is_valid(), "broken liveness info");
2686 int len = (int)live_locals.size();
2687
2688 for (int i = 0; i < len; i++) {
2689 Node* local = unc->local(jvms, i);
2690 // kill local using the liveness of next_bci.
2691 // give up when the local looks like an operand to secure reexecution.
2692 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2693 uint idx = jvms->locoff() + i;
2694 #ifdef ASSERT
2695 if (PrintOpto && Verbose) {
2696 tty->print("[unstable_if] kill local#%d: ", idx);
2697 local->dump();
2698 tty->cr();
2699 }
2700 #endif
2701 igvn.replace_input_of(unc, idx, top());
2702 modified = true;
2703 }
2704 }
2705 }
2706
2707 // keep the modified trap for late query
2708 if (modified) {
2709 trap->set_modified();
2710 } else {
2711 _unstable_if_traps.delete_at(i);
2712 }
2713 }
2714 igvn.optimize();
2715 }
2716
2717 // StringOpts and late inlining of string methods
2718 void Compile::inline_string_calls(bool parse_time) {
2719 {
2720 // remove useless nodes to make the usage analysis simpler
2721 ResourceMark rm;
2722 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2723 }
2724
2725 {
2726 ResourceMark rm;
2727 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2728 PhaseStringOpts pso(initial_gvn());
2729 print_method(PHASE_AFTER_STRINGOPTS, 3);
2730 }
2731
2732 // now inline anything that we skipped the first time around
2733 if (!parse_time) {
2734 _late_inlines_pos = _late_inlines.length();
2735 }
2736
2737 while (_string_late_inlines.length() > 0) {
2738 CallGenerator* cg = _string_late_inlines.pop();
2739 cg->do_late_inline();
2740 if (failing()) return;
2741 }
2742 _string_late_inlines.trunc_to(0);
2743 }
2744
2745 // Late inlining of boxing methods
2746 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2747 if (_boxing_late_inlines.length() > 0) {
2748 assert(has_boxed_value(), "inconsistent");
2749
2750 set_inlining_incrementally(true);
2751
2752 igvn_worklist()->ensure_empty(); // should be done with igvn
2753
2754 _late_inlines_pos = _late_inlines.length();
2755
2756 while (_boxing_late_inlines.length() > 0) {
2757 CallGenerator* cg = _boxing_late_inlines.pop();
2758 cg->do_late_inline();
2759 if (failing()) return;
2760 }
2761 _boxing_late_inlines.trunc_to(0);
2762
2763 inline_incrementally_cleanup(igvn);
2764
2765 set_inlining_incrementally(false);
2766 }
2767 }
2768
2769 bool Compile::inline_incrementally_one() {
2770 assert(IncrementalInline, "incremental inlining should be on");
2771 assert(_late_inlines.length() > 0, "should have been checked by caller");
2772
2773 TracePhase tp(_t_incrInline_inline);
2774
2775 set_inlining_progress(false);
2776 set_do_cleanup(false);
2777
2778 for (int i = 0; i < _late_inlines.length(); i++) {
2779 _late_inlines_pos = i+1;
2780 CallGenerator* cg = _late_inlines.at(i);
2781 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2782 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2783 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2784 if (should_stress_inlining()) {
2785 // randomly add repeated inline attempt if stress-inlining
2786 cg->call_node()->set_generator(cg);
2787 C->igvn_worklist()->push(cg->call_node());
2788 continue;
2789 }
2790 cg->do_late_inline();
2791 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2792 if (failing()) {
2793 return false;
2794 } else if (inlining_progress()) {
2795 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2796 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2797 break; // process one call site at a time
2798 } else {
2799 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2800 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2801 // Avoid potential infinite loop if node already in the IGVN list
2802 assert(false, "scheduled for IGVN during inlining attempt");
2803 } else {
2804 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2805 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2806 cg->call_node()->set_generator(cg);
2807 }
2808 }
2809 } else {
2810 // Ignore late inline direct calls when inlining is not allowed.
2811 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2812 }
2813 }
2814 // Remove processed elements.
2815 _late_inlines.remove_till(_late_inlines_pos);
2816 _late_inlines_pos = 0;
2817
2818 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2819
2820 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2821
2822 set_inlining_progress(false);
2823 set_do_cleanup(false);
2824
2825 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2826 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2827 }
2828
2829 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2830 {
2831 TracePhase tp(_t_incrInline_pru);
2832 ResourceMark rm;
2833 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2834 }
2835 {
2836 TracePhase tp(_t_incrInline_igvn);
2837 igvn.reset();
2838 igvn.optimize();
2839 if (failing()) return;
2840 }
2841 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2842 }
2843
2844 template<typename E>
2845 static void shuffle_array(Compile& C, GrowableArray<E>& array) {
2846 if (array.length() < 2) {
2847 return;
2848 }
2849 for (uint i = array.length() - 1; i >= 1; i--) {
2850 uint j = C.random() % (i + 1);
2851 swap(array.at(i), array.at(j));
2852 }
2853 }
2854
2855 void Compile::shuffle_late_inlines() {
2856 shuffle_array(*C, _late_inlines);
2857 }
2858
2859 // Perform incremental inlining until bound on number of live nodes is reached
2860 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2861 TracePhase tp(_t_incrInline);
2862
2863 set_inlining_incrementally(true);
2864 uint low_live_nodes = 0;
2865
2866 if (StressIncrementalInlining) {
2867 shuffle_late_inlines();
2868 }
2869
2870 while (_late_inlines.length() > 0) {
2871 if (live_nodes() > node_count_inlining_cutoff()) {
2872 if (low_live_nodes < node_count_inlining_cutoff() * 8 / 10) {
2873 TracePhase tp(_t_incrInline_ideal);
2874 // PhaseIdealLoop is expensive so we only try it once we are
2875 // out of live nodes and we only try it again if the previous
2876 // helped got the number of nodes down significantly
2877 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2878 if (failing()) return;
2879 low_live_nodes = live_nodes();
2880 _major_progress = true;
2881 }
2882
2883 if (live_nodes() > node_count_inlining_cutoff()) {
2884 bool do_print_inlining = print_inlining() || print_intrinsics();
2885 if (do_print_inlining || log() != nullptr) {
2886 // Print inlining message for candidates that we couldn't inline for lack of space.
2887 for (int i = 0; i < _late_inlines.length(); i++) {
2888 CallGenerator* cg = _late_inlines.at(i);
2889 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2890 if (do_print_inlining) {
2891 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2892 }
2893 log_late_inline_failure(cg, msg);
2894 }
2895 }
2896 break; // finish
2897 }
2898 }
2899
2900 igvn_worklist()->ensure_empty(); // should be done with igvn
2901
2902 if (_late_inlines.length() == 0) {
2903 break; // no more progress
2904 }
2905
2906 while (inline_incrementally_one()) {
2907 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2908 }
2909 if (failing()) return;
2910
2911 inline_incrementally_cleanup(igvn);
2912
2913 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2914
2915 if (failing()) return;
2916 }
2917
2918 igvn_worklist()->ensure_empty(); // should be done with igvn
2919
2920 if (_string_late_inlines.length() > 0) {
2921 assert(has_stringbuilder(), "inconsistent");
2922
2923 inline_string_calls(false);
2924
2925 if (failing()) return;
2926
2927 inline_incrementally_cleanup(igvn);
2928 }
2929
2930 set_inlining_incrementally(false);
2931 }
2932
2933 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2934 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2935 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2936 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2937 // as if "inlining_incrementally() == true" were set.
2938 assert(inlining_incrementally() == false, "not allowed");
2939 set_strength_reduction(true);
2940 #ifdef ASSERT
2941 Unique_Node_List* modified_nodes = _modified_nodes;
2942 _modified_nodes = nullptr;
2943 #endif
2944 assert(_late_inlines.length() > 0, "sanity");
2945
2946 if (StressIncrementalInlining) {
2947 shuffle_late_inlines();
2948 }
2949
2950 while (_late_inlines.length() > 0) {
2951 igvn_worklist()->ensure_empty(); // should be done with igvn
2952
2953 while (inline_incrementally_one()) {
2954 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2955 }
2956 if (failing()) return;
2957
2958 inline_incrementally_cleanup(igvn);
2959 }
2960 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2961 set_strength_reduction(false);
2962 }
2963
2964 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2965 if (_loop_opts_cnt > 0) {
2966 while (major_progress() && (_loop_opts_cnt > 0)) {
2967 TracePhase tp(_t_idealLoop);
2968 PhaseIdealLoop::optimize(igvn, mode);
2969 _loop_opts_cnt--;
2970 if (failing()) return false;
2971 if (major_progress()) {
2972 print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2973 }
2974 }
2975 }
2976 return true;
2977 }
2978
2979 // Remove edges from "root" to each SafePoint at a backward branch.
2980 // They were inserted during parsing (see add_safepoint()) to make
2981 // infinite loops without calls or exceptions visible to root, i.e.,
2982 // useful.
2983 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2984 Node *r = root();
2985 if (r != nullptr) {
2986 for (uint i = r->req(); i < r->len(); ++i) {
2987 Node *n = r->in(i);
2988 if (n != nullptr && n->is_SafePoint()) {
2989 r->rm_prec(i);
2990 if (n->outcnt() == 0) {
2991 igvn.remove_dead_node(n, PhaseIterGVN::NodeOrigin::Graph);
2992 }
2993 --i;
2994 }
2995 }
2996 // Parsing may have added top inputs to the root node (Path
2997 // leading to the Halt node proven dead). Make sure we get a
2998 // chance to clean them up.
2999 igvn._worklist.push(r);
3000 igvn.optimize();
3001 }
3002 }
3003
3004 //------------------------------Optimize---------------------------------------
3005 // Given a graph, optimize it.
3006 void Compile::Optimize() {
3007 TracePhase tp(_t_optimizer);
3008
3009 #ifndef PRODUCT
3010 if (env()->break_at_compile()) {
3011 BREAKPOINT;
3012 }
3013
3014 #endif
3015
3016 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
3017 #ifdef ASSERT
3018 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
3019 #endif
3020
3021 ResourceMark rm;
3022
3023 NOT_PRODUCT( verify_graph_edges(); )
3024
3025 print_method(PHASE_AFTER_PARSING, 1);
3026
3027 {
3028 // Iterative Global Value Numbering, including ideal transforms
3029 PhaseIterGVN igvn;
3030 #ifdef ASSERT
3031 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
3032 #endif
3033 {
3034 TracePhase tp(_t_iterGVN);
3035 igvn.optimize(true);
3036 }
3037
3038 if (failing()) return;
3039
3040 print_method(PHASE_ITER_GVN1, 2);
3041
3042 process_for_unstable_if_traps(igvn);
3043
3044 if (failing()) return;
3045
3046 inline_incrementally(igvn);
3047
3048 print_method(PHASE_INCREMENTAL_INLINE, 2);
3049
3050 if (failing()) return;
3051
3052 if (eliminate_boxing()) {
3053 // Inline valueOf() methods now.
3054 inline_boxing_calls(igvn);
3055
3056 if (failing()) return;
3057
3058 if (AlwaysIncrementalInline || StressIncrementalInlining) {
3059 inline_incrementally(igvn);
3060 }
3061
3062 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
3063
3064 if (failing()) return;
3065 }
3066
3067 // Remove the speculative part of types and clean up the graph from
3068 // the extra CastPP nodes whose only purpose is to carry them. Do
3069 // that early so that optimizations are not disrupted by the extra
3070 // CastPP nodes.
3071 remove_speculative_types(igvn);
3072
3073 if (failing()) return;
3074
3075 // No more new expensive nodes will be added to the list from here
3076 // so keep only the actual candidates for optimizations.
3077 cleanup_expensive_nodes(igvn);
3078
3079 if (failing()) return;
3080
3081 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
3082 if (EnableVectorSupport && has_vbox_nodes()) {
3083 TracePhase tp(_t_vector);
3084 PhaseVector pv(igvn);
3085 pv.optimize_vector_boxes();
3086 if (failing()) return;
3087 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
3088 }
3089 assert(!has_vbox_nodes(), "sanity");
3090
3091 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
3092 Compile::TracePhase tp(_t_renumberLive);
3093 igvn_worklist()->ensure_empty(); // should be done with igvn
3094 {
3095 ResourceMark rm;
3096 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
3097 }
3098 igvn.reset();
3099 igvn.optimize(true);
3100 if (failing()) return;
3101 }
3102
3103 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
3104 // safepoints
3105 remove_root_to_sfpts_edges(igvn);
3106
3107 // Process inline type nodes now that all inlining is over
3108 process_inline_types(igvn);
3109
3110 adjust_flat_array_access_aliases(igvn);
3111
3112 if (failing()) return;
3113
3114 if (C->macro_count() > 0) {
3115 // Eliminate some macro nodes before EA to reduce analysis pressure
3116 PhaseMacroExpand mexp(igvn);
3117 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
3118 if (failing()) {
3119 return;
3120 }
3121 igvn.set_delay_transform(false);
3122 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3123 }
3124
3125 _print_phase_loop_opts = has_loops();
3126 if (_print_phase_loop_opts) {
3127 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
3128 }
3129
3130 // Perform escape analysis
3131 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
3132 if (has_loops()) {
3133 // Cleanup graph (remove dead nodes).
3134 TracePhase tp(_t_idealLoop);
3135 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
3136 if (failing()) {
3137 return;
3138 }
3139 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
3140 if (C->macro_count() > 0) {
3141 // Eliminate some macro nodes before EA to reduce analysis pressure
3142 PhaseMacroExpand mexp(igvn);
3143 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
3144 if (failing()) {
3145 return;
3146 }
3147 igvn.set_delay_transform(false);
3148 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3149 }
3150 }
3151
3152 bool progress;
3153 do {
3154 ConnectionGraph::do_analysis(this, &igvn);
3155
3156 if (failing()) return;
3157
3158 int mcount = macro_count(); // Record number of allocations and locks before IGVN
3159
3160 // Optimize out fields loads from scalar replaceable allocations.
3161 igvn.optimize(true);
3162 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
3163
3164 if (failing()) return;
3165
3166 if (congraph() != nullptr && macro_count() > 0) {
3167 TracePhase tp(_t_macroEliminate);
3168 PhaseMacroExpand mexp(igvn);
3169 mexp.eliminate_macro_nodes();
3170 if (failing()) {
3171 return;
3172 }
3173 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
3174
3175 igvn.set_delay_transform(false);
3176 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3177 }
3178
3179 ConnectionGraph::verify_ram_nodes(this, root());
3180 if (failing()) return;
3181
3182 progress = do_iterative_escape_analysis() &&
3183 (macro_count() < mcount) &&
3184 ConnectionGraph::has_candidates(this);
3185 // Try again if candidates exist and made progress
3186 // by removing some allocations and/or locks.
3187 } while (progress);
3188 }
3189
3190 process_flat_accesses(igvn);
3191 if (failing()) {
3192 return;
3193 }
3194
3195 // Loop transforms on the ideal graph. Range Check Elimination,
3196 // peeling, unrolling, etc.
3197
3198 // Set loop opts counter
3199 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
3200 {
3201 TracePhase tp(_t_idealLoop);
3202 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
3203 _loop_opts_cnt--;
3204 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
3205 if (failing()) return;
3206 }
3207 // Loop opts pass if partial peeling occurred in previous pass
3208 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
3209 TracePhase tp(_t_idealLoop);
3210 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
3211 _loop_opts_cnt--;
3212 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
3213 if (failing()) return;
3214 }
3215 // Loop opts pass for loop-unrolling before CCP
3216 if(major_progress() && (_loop_opts_cnt > 0)) {
3217 TracePhase tp(_t_idealLoop);
3218 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
3219 _loop_opts_cnt--;
3220 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
3221 }
3222 if (!failing()) {
3223 // Verify that last round of loop opts produced a valid graph
3224 PhaseIdealLoop::verify(igvn);
3225 }
3226 }
3227 if (failing()) return;
3228
3229 // Conditional Constant Propagation;
3230 print_method(PHASE_BEFORE_CCP1, 2);
3231 PhaseCCP ccp( &igvn );
3232 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
3233 {
3234 TracePhase tp(_t_ccp);
3235 ccp.do_transform();
3236 }
3237 print_method(PHASE_CCP1, 2);
3238
3239 assert( true, "Break here to ccp.dump_old2new_map()");
3240
3241 // Iterative Global Value Numbering, including ideal transforms
3242 {
3243 TracePhase tp(_t_iterGVN2);
3244 igvn.reset_from_igvn(&ccp);
3245 igvn.optimize(true);
3246 }
3247 print_method(PHASE_ITER_GVN2, 2);
3248
3249 if (failing()) return;
3250
3251 // Loop transforms on the ideal graph. Range Check Elimination,
3252 // peeling, unrolling, etc.
3253 if (!optimize_loops(igvn, LoopOptsDefault)) {
3254 return;
3255 }
3256
3257 C->clear_major_progress(); // ensure that major progress is now clear
3258
3259 process_for_post_loop_opts_igvn(igvn);
3260
3261 if (failing()) return;
3262
3263 // Once loop optimizations are over, it is safe to get rid of all reachability fence nodes and
3264 // migrate reachability edges to safepoints.
3265 if (OptimizeReachabilityFences && _reachability_fences.length() > 0) {
3266 TracePhase tp1(_t_idealLoop);
3267 TracePhase tp2(_t_reachability);
3268 PhaseIdealLoop::optimize(igvn, PostLoopOptsExpandReachabilityFences);
3269 print_method(PHASE_EXPAND_REACHABILITY_FENCES, 2);
3270 if (failing()) return;
3271 assert(_reachability_fences.length() == 0 || PreserveReachabilityFencesOnConstants, "no RF nodes allowed");
3272 }
3273
3274 process_for_merge_stores_igvn(igvn);
3275
3276 if (failing()) return;
3277
3278 #ifdef ASSERT
3279 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
3280 #endif
3281
3282 if (_late_inlines.length() > 0) {
3283 // More opportunities to optimize virtual and MH calls.
3284 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
3285 process_late_inline_calls_no_inline(igvn);
3286 if (failing()) {
3287 return;
3288 }
3289 }
3290 assert(_late_inlines.length() == 0, "late inline queue must be drained");
3291
3292 // Process inline types before macro expansion. Otherwise, we will not be able to
3293 // remove unused allocations because it cannot match the expanded allocation.
3294 process_inline_types(igvn);
3295
3296 {
3297 TracePhase tp(_t_macroExpand);
3298 PhaseMacroExpand mex(igvn);
3299 // Last attempt to eliminate macro nodes.
3300 mex.eliminate_macro_nodes();
3301 if (failing()) {
3302 return;
3303 }
3304
3305 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
3306 // Do not allow new macro nodes once we start to eliminate and expand
3307 C->reset_allow_macro_nodes();
3308 // Last attempt to eliminate macro nodes before expand
3309 mex.eliminate_macro_nodes();
3310 if (failing()) {
3311 return;
3312 }
3313 mex.eliminate_opaque_looplimit_macro_nodes();
3314 if (failing()) {
3315 return;
3316 }
3317 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
3318 if (mex.expand_macro_nodes()) {
3319 assert(failing(), "must bail out w/ explicit message");
3320 return;
3321 }
3322 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
3323 }
3324
3325 // Process inline type nodes again and remove them. From here
3326 // on we don't need to keep track of field values anymore.
3327 process_inline_types(igvn, /* remove= */ true);
3328
3329 {
3330 TracePhase tp(_t_barrierExpand);
3331 if (bs->expand_barriers(this, igvn)) {
3332 assert(failing(), "must bail out w/ explicit message");
3333 return;
3334 }
3335 print_method(PHASE_BARRIER_EXPANSION, 2);
3336 }
3337
3338 if (C->max_vector_size() > 0) {
3339 C->optimize_logic_cones(igvn);
3340 igvn.optimize();
3341 if (failing()) return;
3342 }
3343
3344 DEBUG_ONLY( _modified_nodes = nullptr; )
3345 DEBUG_ONLY( _late_inlines.clear(); )
3346
3347 assert(igvn._worklist.size() == 0, "not empty");
3348 } // (End scope of igvn; run destructor if necessary for asserts.)
3349
3350 check_no_dead_use();
3351
3352 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
3353 // to remove hashes to unlock nodes for modifications.
3354 C->node_hash()->clear();
3355
3356 // A method with only infinite loops has no edges entering loops from root
3357 {
3358 TracePhase tp(_t_graphReshaping);
3359 if (final_graph_reshaping()) {
3360 assert(failing(), "must bail out w/ explicit message");
3361 return;
3362 }
3363 }
3364
3365 print_method(PHASE_OPTIMIZE_FINISHED, 2);
3366 DEBUG_ONLY(set_phase_optimize_finished();)
3367 }
3368
3369 #ifdef ASSERT
3370 void Compile::check_no_dead_use() const {
3371 ResourceMark rm;
3372 Unique_Node_List wq;
3373 wq.push(root());
3374 for (uint i = 0; i < wq.size(); ++i) {
3375 Node* n = wq.at(i);
3376 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3377 Node* u = n->fast_out(j);
3378 if (u->outcnt() == 0 && !u->is_Con()) {
3379 u->dump();
3380 fatal("no reachable node should have no use");
3381 }
3382 wq.push(u);
3383 }
3384 }
3385 }
3386 #endif
3387
3388 void Compile::inline_vector_reboxing_calls() {
3389 if (C->_vector_reboxing_late_inlines.length() > 0) {
3390 _late_inlines_pos = C->_late_inlines.length();
3391 while (_vector_reboxing_late_inlines.length() > 0) {
3392 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
3393 cg->do_late_inline();
3394 if (failing()) return;
3395 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
3396 }
3397 _vector_reboxing_late_inlines.trunc_to(0);
3398 }
3399 }
3400
3401 bool Compile::has_vbox_nodes() {
3402 if (C->_vector_reboxing_late_inlines.length() > 0) {
3403 return true;
3404 }
3405 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3406 Node * n = C->macro_node(macro_idx);
3407 assert(n->is_macro(), "only macro nodes expected here");
3408 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3409 return true;
3410 }
3411 }
3412 return false;
3413 }
3414
3415 //---------------------------- Bitwise operation packing optimization ---------------------------
3416
3417 static bool is_vector_unary_bitwise_op(Node* n) {
3418 return n->Opcode() == Op_XorV &&
3419 VectorNode::is_vector_bitwise_not_pattern(n);
3420 }
3421
3422 static bool is_vector_binary_bitwise_op(Node* n) {
3423 switch (n->Opcode()) {
3424 case Op_AndV:
3425 case Op_OrV:
3426 return true;
3427
3428 case Op_XorV:
3429 return !is_vector_unary_bitwise_op(n);
3430
3431 default:
3432 return false;
3433 }
3434 }
3435
3436 static bool is_vector_ternary_bitwise_op(Node* n) {
3437 return n->Opcode() == Op_MacroLogicV;
3438 }
3439
3440 static bool is_vector_bitwise_op(Node* n) {
3441 return is_vector_unary_bitwise_op(n) ||
3442 is_vector_binary_bitwise_op(n) ||
3443 is_vector_ternary_bitwise_op(n);
3444 }
3445
3446 static bool is_vector_bitwise_cone_root(Node* n) {
3447 if (n->bottom_type()->isa_pvectmask() || !is_vector_bitwise_op(n)) {
3448 return false;
3449 }
3450 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3451 if (is_vector_bitwise_op(n->fast_out(i))) {
3452 return false;
3453 }
3454 }
3455 return true;
3456 }
3457
3458 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3459 uint cnt = 0;
3460 if (is_vector_bitwise_op(n)) {
3461 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3462 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3463 for (uint i = 1; i < inp_cnt; i++) {
3464 Node* in = n->in(i);
3465 bool skip = VectorNode::is_all_ones_vector(in);
3466 if (!skip && !inputs.member(in)) {
3467 inputs.push(in);
3468 cnt++;
3469 }
3470 }
3471 assert(cnt <= 1, "not unary");
3472 } else {
3473 uint last_req = inp_cnt;
3474 if (is_vector_ternary_bitwise_op(n)) {
3475 last_req = inp_cnt - 1; // skip last input
3476 }
3477 for (uint i = 1; i < last_req; i++) {
3478 Node* def = n->in(i);
3479 if (!inputs.member(def)) {
3480 inputs.push(def);
3481 cnt++;
3482 }
3483 }
3484 }
3485 } else { // not a bitwise operations
3486 if (!inputs.member(n)) {
3487 inputs.push(n);
3488 cnt++;
3489 }
3490 }
3491 return cnt;
3492 }
3493
3494 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3495 Unique_Node_List useful_nodes;
3496 C->identify_useful_nodes(useful_nodes);
3497
3498 for (uint i = 0; i < useful_nodes.size(); i++) {
3499 Node* n = useful_nodes.at(i);
3500 if (is_vector_bitwise_cone_root(n)) {
3501 list.push(n);
3502 }
3503 }
3504 }
3505
3506 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3507 const TypeVect* vt,
3508 Unique_Node_List& partition,
3509 Unique_Node_List& inputs) {
3510 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3511 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3512 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3513
3514 Node* in1 = inputs.at(0);
3515 Node* in2 = inputs.at(1);
3516 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3517
3518 uint func = compute_truth_table(partition, inputs);
3519
3520 Node* pn = partition.at(partition.size() - 1);
3521 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3522 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3523 }
3524
3525 static uint extract_bit(uint func, uint pos) {
3526 return (func & (1 << pos)) >> pos;
3527 }
3528
3529 //
3530 // A macro logic node represents a truth table. It has 4 inputs,
3531 // First three inputs corresponds to 3 columns of a truth table
3532 // and fourth input captures the logic function.
3533 //
3534 // eg. fn = (in1 AND in2) OR in3;
3535 //
3536 // MacroNode(in1,in2,in3,fn)
3537 //
3538 // -----------------
3539 // in1 in2 in3 fn
3540 // -----------------
3541 // 0 0 0 0
3542 // 0 0 1 1
3543 // 0 1 0 0
3544 // 0 1 1 1
3545 // 1 0 0 0
3546 // 1 0 1 1
3547 // 1 1 0 1
3548 // 1 1 1 1
3549 //
3550
3551 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3552 int res = 0;
3553 for (int i = 0; i < 8; i++) {
3554 int bit1 = extract_bit(in1, i);
3555 int bit2 = extract_bit(in2, i);
3556 int bit3 = extract_bit(in3, i);
3557
3558 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3559 int func_bit = extract_bit(func, func_bit_pos);
3560
3561 res |= func_bit << i;
3562 }
3563 return res;
3564 }
3565
3566 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
3567 assert(n != nullptr, "");
3568 assert(eval_map.contains(n), "absent");
3569 return *(eval_map.get(n));
3570 }
3571
3572 static void eval_operands(Node* n,
3573 uint& func1, uint& func2, uint& func3,
3574 HashTable<Node*,uint>& eval_map) {
3575 assert(is_vector_bitwise_op(n), "");
3576
3577 if (is_vector_unary_bitwise_op(n)) {
3578 Node* opnd = n->in(1);
3579 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3580 opnd = n->in(2);
3581 }
3582 func1 = eval_operand(opnd, eval_map);
3583 } else if (is_vector_binary_bitwise_op(n)) {
3584 func1 = eval_operand(n->in(1), eval_map);
3585 func2 = eval_operand(n->in(2), eval_map);
3586 } else {
3587 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3588 func1 = eval_operand(n->in(1), eval_map);
3589 func2 = eval_operand(n->in(2), eval_map);
3590 func3 = eval_operand(n->in(3), eval_map);
3591 }
3592 }
3593
3594 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3595 assert(inputs.size() <= 3, "sanity");
3596 ResourceMark rm;
3597 uint res = 0;
3598 HashTable<Node*,uint> eval_map;
3599
3600 // Populate precomputed functions for inputs.
3601 // Each input corresponds to one column of 3 input truth-table.
3602 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3603 0xCC, // (_, b, _) -> b
3604 0xF0 }; // (a, _, _) -> a
3605 for (uint i = 0; i < inputs.size(); i++) {
3606 eval_map.put(inputs.at(i), input_funcs[2-i]);
3607 }
3608
3609 for (uint i = 0; i < partition.size(); i++) {
3610 Node* n = partition.at(i);
3611
3612 uint func1 = 0, func2 = 0, func3 = 0;
3613 eval_operands(n, func1, func2, func3, eval_map);
3614
3615 switch (n->Opcode()) {
3616 case Op_OrV:
3617 assert(func3 == 0, "not binary");
3618 res = func1 | func2;
3619 break;
3620 case Op_AndV:
3621 assert(func3 == 0, "not binary");
3622 res = func1 & func2;
3623 break;
3624 case Op_XorV:
3625 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3626 assert(func2 == 0 && func3 == 0, "not unary");
3627 res = (~func1) & 0xFF;
3628 } else {
3629 assert(func3 == 0, "not binary");
3630 res = func1 ^ func2;
3631 }
3632 break;
3633 case Op_MacroLogicV:
3634 // Ordering of inputs may change during evaluation of sub-tree
3635 // containing MacroLogic node as a child node, thus a re-evaluation
3636 // makes sure that function is evaluated in context of current
3637 // inputs.
3638 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3639 break;
3640
3641 default: assert(false, "not supported: %s", n->Name());
3642 }
3643 assert(res <= 0xFF, "invalid");
3644 eval_map.put(n, res);
3645 }
3646 return res;
3647 }
3648
3649 // Criteria under which nodes gets packed into a macro logic node:-
3650 // 1) Parent and both child nodes are all unmasked or masked with
3651 // same predicates.
3652 // 2) Masked parent can be packed with left child if it is predicated
3653 // and both have same predicates.
3654 // 3) Masked parent can be packed with right child if its un-predicated
3655 // or has matching predication condition.
3656 // 4) An unmasked parent can be packed with an unmasked child.
3657 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3658 assert(partition.size() == 0, "not empty");
3659 assert(inputs.size() == 0, "not empty");
3660 if (is_vector_ternary_bitwise_op(n)) {
3661 return false;
3662 }
3663
3664 bool is_unary_op = is_vector_unary_bitwise_op(n);
3665 if (is_unary_op) {
3666 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3667 return false; // too few inputs
3668 }
3669
3670 bool pack_left_child = true;
3671 bool pack_right_child = true;
3672
3673 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3674 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3675
3676 int left_child_input_cnt = 0;
3677 int right_child_input_cnt = 0;
3678
3679 bool parent_is_predicated = n->is_predicated_vector();
3680 bool left_child_predicated = n->in(1)->is_predicated_vector();
3681 bool right_child_predicated = n->in(2)->is_predicated_vector();
3682
3683 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
3684 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3685 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3686
3687 do {
3688 if (pack_left_child && left_child_LOP &&
3689 ((!parent_is_predicated && !left_child_predicated) ||
3690 ((parent_is_predicated && left_child_predicated &&
3691 parent_pred == left_child_pred)))) {
3692 partition.push(n->in(1));
3693 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3694 } else {
3695 inputs.push(n->in(1));
3696 left_child_input_cnt = 1;
3697 }
3698
3699 if (pack_right_child && right_child_LOP &&
3700 (!right_child_predicated ||
3701 (right_child_predicated && parent_is_predicated &&
3702 parent_pred == right_child_pred))) {
3703 partition.push(n->in(2));
3704 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3705 } else {
3706 inputs.push(n->in(2));
3707 right_child_input_cnt = 1;
3708 }
3709
3710 if (inputs.size() > 3) {
3711 assert(partition.size() > 0, "");
3712 inputs.clear();
3713 partition.clear();
3714 if (left_child_input_cnt > right_child_input_cnt) {
3715 pack_left_child = false;
3716 } else {
3717 pack_right_child = false;
3718 }
3719 } else {
3720 break;
3721 }
3722 } while(true);
3723
3724 if(partition.size()) {
3725 partition.push(n);
3726 }
3727
3728 return (partition.size() == 2 || partition.size() == 3) &&
3729 (inputs.size() == 2 || inputs.size() == 3);
3730 }
3731
3732 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3733 assert(is_vector_bitwise_op(n), "not a root");
3734
3735 visited.set(n->_idx);
3736
3737 // 1) Do a DFS walk over the logic cone.
3738 for (uint i = 1; i < n->req(); i++) {
3739 Node* in = n->in(i);
3740 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3741 process_logic_cone_root(igvn, in, visited);
3742 }
3743 }
3744
3745 // 2) Bottom up traversal: Merge node[s] with
3746 // the parent to form macro logic node.
3747 Unique_Node_List partition;
3748 Unique_Node_List inputs;
3749 if (compute_logic_cone(n, partition, inputs)) {
3750 const TypeVect* vt = n->bottom_type()->is_vect();
3751 Node* pn = partition.at(partition.size() - 1);
3752 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3753 if (mask == nullptr ||
3754 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3755 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3756 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3757 igvn.replace_node(n, macro_logic);
3758 }
3759 }
3760 }
3761
3762 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3763 ResourceMark rm;
3764 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3765 Unique_Node_List list;
3766 collect_logic_cone_roots(list);
3767
3768 while (list.size() > 0) {
3769 Node* n = list.pop();
3770 const TypeVect* vt = n->bottom_type()->is_vect();
3771 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3772 if (supported) {
3773 VectorSet visited(comp_arena());
3774 process_logic_cone_root(igvn, n, visited);
3775 }
3776 }
3777 }
3778 }
3779
3780 //------------------------------Code_Gen---------------------------------------
3781 // Given a graph, generate code for it
3782 void Compile::Code_Gen() {
3783 if (failing()) {
3784 return;
3785 }
3786
3787 // Perform instruction selection. You might think we could reclaim Matcher
3788 // memory PDQ, but actually the Matcher is used in generating spill code.
3789 // Internals of the Matcher (including some VectorSets) must remain live
3790 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3791 // set a bit in reclaimed memory.
3792
3793 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3794 // nodes. Mapping is only valid at the root of each matched subtree.
3795 NOT_PRODUCT( verify_graph_edges(); )
3796
3797 Matcher matcher;
3798 _matcher = &matcher;
3799 {
3800 TracePhase tp(_t_matcher);
3801 matcher.match();
3802 if (failing()) {
3803 return;
3804 }
3805 }
3806 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3807 // nodes. Mapping is only valid at the root of each matched subtree.
3808 NOT_PRODUCT( verify_graph_edges(); )
3809
3810 // If you have too many nodes, or if matching has failed, bail out
3811 check_node_count(0, "out of nodes matching instructions");
3812 if (failing()) {
3813 return;
3814 }
3815
3816 print_method(PHASE_MATCHING, 2);
3817
3818 // Build a proper-looking CFG
3819 PhaseCFG cfg(node_arena(), root(), matcher);
3820 if (failing()) {
3821 return;
3822 }
3823 _cfg = &cfg;
3824 {
3825 TracePhase tp(_t_scheduler);
3826 bool success = cfg.do_global_code_motion();
3827 if (!success) {
3828 return;
3829 }
3830
3831 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3832 NOT_PRODUCT( verify_graph_edges(); )
3833 cfg.verify();
3834 if (failing()) {
3835 return;
3836 }
3837 }
3838
3839 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3840 _regalloc = ®alloc;
3841 {
3842 TracePhase tp(_t_registerAllocation);
3843 // Perform register allocation. After Chaitin, use-def chains are
3844 // no longer accurate (at spill code) and so must be ignored.
3845 // Node->LRG->reg mappings are still accurate.
3846 _regalloc->Register_Allocate();
3847
3848 // Bail out if the allocator builds too many nodes
3849 if (failing()) {
3850 return;
3851 }
3852
3853 print_method(PHASE_REGISTER_ALLOCATION, 2);
3854 }
3855
3856 // Prior to register allocation we kept empty basic blocks in case the
3857 // the allocator needed a place to spill. After register allocation we
3858 // are not adding any new instructions. If any basic block is empty, we
3859 // can now safely remove it.
3860 {
3861 TracePhase tp(_t_blockOrdering);
3862 cfg.remove_empty_blocks();
3863 if (do_freq_based_layout()) {
3864 PhaseBlockLayout layout(cfg);
3865 } else {
3866 cfg.set_loop_alignment();
3867 }
3868 cfg.fixup_flow();
3869 cfg.remove_unreachable_blocks();
3870 cfg.verify_dominator_tree();
3871 print_method(PHASE_BLOCK_ORDERING, 3);
3872 }
3873
3874 // Apply peephole optimizations
3875 if( OptoPeephole ) {
3876 TracePhase tp(_t_peephole);
3877 PhasePeephole peep( _regalloc, cfg);
3878 peep.do_transform();
3879 print_method(PHASE_PEEPHOLE, 3);
3880 }
3881
3882 // Do late expand if CPU requires this.
3883 if (Matcher::require_postalloc_expand) {
3884 TracePhase tp(_t_postalloc_expand);
3885 cfg.postalloc_expand(_regalloc);
3886 print_method(PHASE_POSTALLOC_EXPAND, 3);
3887 }
3888
3889 #ifdef ASSERT
3890 {
3891 CompilationMemoryStatistic::do_test_allocations();
3892 if (failing()) return;
3893 }
3894 #endif
3895
3896 // Convert Nodes to instruction bits in a buffer
3897 {
3898 TracePhase tp(_t_output);
3899 PhaseOutput output;
3900 output.Output();
3901 if (failing()) return;
3902 output.install();
3903 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3904 }
3905
3906 // He's dead, Jim.
3907 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3908 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3909 }
3910
3911 //------------------------------Final_Reshape_Counts---------------------------
3912 // This class defines counters and node lists collected during
3913 // the final graph reshaping.
3914 struct Final_Reshape_Counts : public StackObj {
3915 int _java_call_count; // count non-inlined 'java' calls
3916 int _inner_loop_count; // count loops which need alignment
3917 VectorSet _visited; // Visitation flags
3918 Node_List _tests; // Set of IfNodes & PCTableNodes
3919
3920 Final_Reshape_Counts() :
3921 _java_call_count(0), _inner_loop_count(0) { }
3922
3923 void inc_java_call_count() { _java_call_count++; }
3924 void inc_inner_loop_count() { _inner_loop_count++; }
3925
3926 int get_java_call_count() const { return _java_call_count; }
3927 int get_inner_loop_count() const { return _inner_loop_count; }
3928 };
3929
3930 //------------------------------final_graph_reshaping_impl----------------------
3931 // Implement items 1-5 from final_graph_reshaping below.
3932 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3933
3934 if ( n->outcnt() == 0 ) return; // dead node
3935 uint nop = n->Opcode();
3936
3937 // Check for 2-input instruction with "last use" on right input.
3938 // Swap to left input. Implements item (2).
3939 if( n->req() == 3 && // two-input instruction
3940 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3941 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3942 n->in(2)->outcnt() == 1 &&// right use IS a last use
3943 !n->in(2)->is_Con() ) { // right use is not a constant
3944 // Check for commutative opcode
3945 switch( nop ) {
3946 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddHF: case Op_AddL:
3947 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3948 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3949 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulHF: case Op_MulL:
3950 case Op_AndL: case Op_XorL: case Op_OrL:
3951 case Op_AndI: case Op_XorI: case Op_OrI: {
3952 // Move "last use" input to left by swapping inputs
3953 n->swap_edges(1, 2);
3954 break;
3955 }
3956 default:
3957 break;
3958 }
3959 }
3960
3961 #ifdef ASSERT
3962 if( n->is_Mem() ) {
3963 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3964 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3965 // oop will be recorded in oop map if load crosses safepoint
3966 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3967 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3968 "raw memory operations should have control edge");
3969 }
3970 if (n->is_MemBar()) {
3971 MemBarNode* mb = n->as_MemBar();
3972 if (mb->trailing_store() || mb->trailing_load_store()) {
3973 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3974 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3975 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3976 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3977 } else if (mb->leading()) {
3978 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3979 }
3980 }
3981 if (n->is_CallLeafPure()) {
3982 // A pure call whose result projection is unused should have been
3983 // eliminated by CallLeafPureNode::Ideal during IGVN.
3984 assert(n->as_CallLeafPure()->proj_out_or_null(TypeFunc::Parms) != nullptr,
3985 "unused CallLeafPureNode should have been removed before final graph reshaping");
3986 }
3987 #endif
3988 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3989 if (!gc_handled) {
3990 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3991 }
3992
3993 // Collect CFG split points
3994 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3995 frc._tests.push(n);
3996 }
3997 }
3998
3999 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
4000 if (!UseDivMod) {
4001 return;
4002 }
4003
4004 // Check if "a % b" and "a / b" both exist
4005 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
4006 if (d == nullptr) {
4007 return;
4008 }
4009
4010 // Replace them with a fused divmod if supported
4011 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
4012 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
4013 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
4014 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
4015 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
4016 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
4017 // DivMod node so the dependency is not lost.
4018 divmod->add_prec_from(n);
4019 divmod->add_prec_from(d);
4020 d->subsume_by(divmod->div_proj(), this);
4021 n->subsume_by(divmod->mod_proj(), this);
4022 } else {
4023 // Replace "a % b" with "a - ((a / b) * b)"
4024 Node* mult = MulNode::make(d, d->in(2), bt);
4025 Node* sub = SubNode::make(d->in(1), mult, bt);
4026 n->subsume_by(sub, this);
4027 }
4028 }
4029
4030 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
4031 switch( nop ) {
4032 case Op_Opaque1: // Remove Opaque Nodes before matching
4033 n->subsume_by(n->in(1), this);
4034 break;
4035 case Op_CallLeafPure: {
4036 // If the pure call is not supported, then lower to a CallLeaf.
4037 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
4038 CallNode* call = n->as_Call();
4039 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
4040 call->_name, TypeRawPtr::BOTTOM);
4041 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
4042 new_call->init_req(TypeFunc::I_O, C->top());
4043 new_call->init_req(TypeFunc::Memory, C->top());
4044 new_call->init_req(TypeFunc::ReturnAdr, C->top());
4045 new_call->init_req(TypeFunc::FramePtr, C->top());
4046 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain_sig()->cnt(); i++) {
4047 new_call->init_req(i, call->in(i));
4048 }
4049 n->subsume_by(new_call, this);
4050 }
4051 break;
4052 }
4053 case Op_CallStaticJava:
4054 case Op_CallJava:
4055 case Op_CallDynamicJava:
4056 frc.inc_java_call_count(); // Count java call site;
4057 case Op_CallRuntime:
4058 case Op_CallLeaf:
4059 case Op_CallLeafVector:
4060 case Op_CallLeafNoFP: {
4061 assert (n->is_Call(), "");
4062 CallNode *call = n->as_Call();
4063 // See if uncommon argument is shared
4064 if (call->is_CallStaticJava() && call->as_CallStaticJava()->_name) {
4065 Node *n = call->in(TypeFunc::Parms);
4066 int nop = n->Opcode();
4067 // Clone shared simple arguments to uncommon calls, item (1).
4068 if (n->outcnt() > 1 &&
4069 !n->is_Proj() &&
4070 nop != Op_CreateEx &&
4071 nop != Op_CheckCastPP &&
4072 nop != Op_DecodeN &&
4073 nop != Op_DecodeNKlass &&
4074 !n->is_Mem() &&
4075 !n->is_Phi()) {
4076 Node *x = n->clone();
4077 call->set_req(TypeFunc::Parms, x);
4078 }
4079 }
4080 break;
4081 }
4082
4083 // Mem nodes need explicit cases to satisfy assert(!n->is_Mem()) in default.
4084 case Op_StoreF:
4085 case Op_LoadF:
4086 case Op_StoreD:
4087 case Op_LoadD:
4088 case Op_LoadD_unaligned:
4089 case Op_StoreB:
4090 case Op_StoreC:
4091 case Op_StoreI:
4092 case Op_StoreL:
4093 case Op_StoreLSpecial:
4094 case Op_CompareAndSwapB:
4095 case Op_CompareAndSwapS:
4096 case Op_CompareAndSwapI:
4097 case Op_CompareAndSwapL:
4098 case Op_CompareAndSwapP:
4099 case Op_CompareAndSwapN:
4100 case Op_WeakCompareAndSwapB:
4101 case Op_WeakCompareAndSwapS:
4102 case Op_WeakCompareAndSwapI:
4103 case Op_WeakCompareAndSwapL:
4104 case Op_WeakCompareAndSwapP:
4105 case Op_WeakCompareAndSwapN:
4106 case Op_CompareAndExchangeB:
4107 case Op_CompareAndExchangeS:
4108 case Op_CompareAndExchangeI:
4109 case Op_CompareAndExchangeL:
4110 case Op_CompareAndExchangeP:
4111 case Op_CompareAndExchangeN:
4112 case Op_GetAndAddS:
4113 case Op_GetAndAddB:
4114 case Op_GetAndAddI:
4115 case Op_GetAndAddL:
4116 case Op_GetAndSetS:
4117 case Op_GetAndSetB:
4118 case Op_GetAndSetI:
4119 case Op_GetAndSetL:
4120 case Op_GetAndSetP:
4121 case Op_GetAndSetN:
4122 case Op_StoreP:
4123 case Op_StoreN:
4124 case Op_StoreNKlass:
4125 case Op_LoadB:
4126 case Op_LoadUB:
4127 case Op_LoadUS:
4128 case Op_LoadI:
4129 case Op_LoadKlass:
4130 case Op_LoadNKlass:
4131 case Op_LoadL:
4132 case Op_LoadL_unaligned:
4133 case Op_LoadP:
4134 case Op_LoadN:
4135 case Op_LoadRange:
4136 case Op_LoadS:
4137 case Op_LoadVectorGather:
4138 case Op_StoreVectorScatter:
4139 case Op_LoadVectorGatherMasked:
4140 case Op_StoreVectorScatterMasked:
4141 case Op_LoadVectorMasked:
4142 case Op_StoreVectorMasked:
4143 break;
4144
4145 case Op_AddP: { // Assert sane base pointers
4146 Node *addp = n->in(AddPNode::Address);
4147 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
4148 #ifdef _LP64
4149 if (addp->Opcode() == Op_ConP &&
4150 addp == n->in(AddPNode::Base) &&
4151 n->in(AddPNode::Offset)->is_Con()) {
4152 // If the transformation of ConP to ConN+DecodeN is beneficial depends
4153 // on the platform and on the compressed oops mode.
4154 // Use addressing with narrow klass to load with offset on x86.
4155 // Some platforms can use the constant pool to load ConP.
4156 // Do this transformation here since IGVN will convert ConN back to ConP.
4157 const Type* t = addp->bottom_type();
4158 bool is_oop = t->isa_oopptr() != nullptr;
4159 bool is_klass = t->isa_klassptr() != nullptr;
4160
4161 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
4162 (is_klass && Matcher::const_klass_prefer_decode() &&
4163 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
4164 Node* nn = nullptr;
4165
4166 int op = is_oop ? Op_ConN : Op_ConNKlass;
4167
4168 // Look for existing ConN node of the same exact type.
4169 Node* r = root();
4170 uint cnt = r->outcnt();
4171 for (uint i = 0; i < cnt; i++) {
4172 Node* m = r->raw_out(i);
4173 if (m!= nullptr && m->Opcode() == op &&
4174 m->bottom_type()->make_ptr() == t) {
4175 nn = m;
4176 break;
4177 }
4178 }
4179 if (nn != nullptr) {
4180 // Decode a narrow oop to match address
4181 // [R12 + narrow_oop_reg<<3 + offset]
4182 if (is_oop) {
4183 nn = new DecodeNNode(nn, t);
4184 } else {
4185 nn = new DecodeNKlassNode(nn, t);
4186 }
4187 // Check for succeeding AddP which uses the same Base.
4188 // Otherwise we will run into the assertion above when visiting that guy.
4189 for (uint i = 0; i < n->outcnt(); ++i) {
4190 Node *out_i = n->raw_out(i);
4191 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
4192 out_i->set_req(AddPNode::Base, nn);
4193 #ifdef ASSERT
4194 for (uint j = 0; j < out_i->outcnt(); ++j) {
4195 Node *out_j = out_i->raw_out(j);
4196 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
4197 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
4198 }
4199 #endif
4200 }
4201 }
4202 n->set_req(AddPNode::Base, nn);
4203 n->set_req(AddPNode::Address, nn);
4204 if (addp->outcnt() == 0) {
4205 addp->disconnect_inputs(this);
4206 }
4207 }
4208 }
4209 }
4210 #endif
4211 break;
4212 }
4213
4214 case Op_CastPP: {
4215 // Remove CastPP nodes to gain more freedom during scheduling but
4216 // keep the dependency they encode as control or precedence edges
4217 // (if control is set already) on memory operations. Some CastPP
4218 // nodes don't have a control (don't carry a dependency): skip
4219 // those.
4220 if (n->in(0) != nullptr) {
4221 ResourceMark rm;
4222 Unique_Node_List wq;
4223 wq.push(n);
4224 for (uint next = 0; next < wq.size(); ++next) {
4225 Node *m = wq.at(next);
4226 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
4227 Node* use = m->fast_out(i);
4228 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
4229 use->ensure_control_or_add_prec(n->in(0));
4230 } else {
4231 switch(use->Opcode()) {
4232 case Op_AddP:
4233 case Op_DecodeN:
4234 case Op_DecodeNKlass:
4235 case Op_CheckCastPP:
4236 case Op_CastPP:
4237 wq.push(use);
4238 break;
4239 }
4240 }
4241 }
4242 }
4243 }
4244 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
4245 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
4246 Node* in1 = n->in(1);
4247 const Type* t = n->bottom_type();
4248 Node* new_in1 = in1->clone();
4249 new_in1->as_DecodeN()->set_type(t);
4250
4251 if (!Matcher::narrow_oop_use_complex_address()) {
4252 //
4253 // x86, ARM and friends can handle 2 adds in addressing mode
4254 // and Matcher can fold a DecodeN node into address by using
4255 // a narrow oop directly and do implicit null check in address:
4256 //
4257 // [R12 + narrow_oop_reg<<3 + offset]
4258 // NullCheck narrow_oop_reg
4259 //
4260 // On other platforms (Sparc) we have to keep new DecodeN node and
4261 // use it to do implicit null check in address:
4262 //
4263 // decode_not_null narrow_oop_reg, base_reg
4264 // [base_reg + offset]
4265 // NullCheck base_reg
4266 //
4267 // Pin the new DecodeN node to non-null path on these platform (Sparc)
4268 // to keep the information to which null check the new DecodeN node
4269 // corresponds to use it as value in implicit_null_check().
4270 //
4271 new_in1->set_req(0, n->in(0));
4272 }
4273
4274 n->subsume_by(new_in1, this);
4275 if (in1->outcnt() == 0) {
4276 in1->disconnect_inputs(this);
4277 }
4278 } else {
4279 n->subsume_by(n->in(1), this);
4280 if (n->outcnt() == 0) {
4281 n->disconnect_inputs(this);
4282 }
4283 }
4284 break;
4285 }
4286 case Op_CastII: {
4287 n->as_CastII()->remove_range_check_cast(this);
4288 break;
4289 }
4290 #ifdef _LP64
4291 case Op_CmpP:
4292 // Do this transformation here to preserve CmpPNode::sub() and
4293 // other TypePtr related Ideal optimizations (for example, ptr nullness).
4294 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
4295 Node* in1 = n->in(1);
4296 Node* in2 = n->in(2);
4297 if (!in1->is_DecodeNarrowPtr()) {
4298 in2 = in1;
4299 in1 = n->in(2);
4300 }
4301 assert(in1->is_DecodeNarrowPtr(), "sanity");
4302
4303 Node* new_in2 = nullptr;
4304 if (in2->is_DecodeNarrowPtr()) {
4305 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
4306 new_in2 = in2->in(1);
4307 } else if (in2->Opcode() == Op_ConP) {
4308 const Type* t = in2->bottom_type();
4309 if (t == TypePtr::NULL_PTR) {
4310 assert(in1->is_DecodeN(), "compare klass to null?");
4311 // Don't convert CmpP null check into CmpN if compressed
4312 // oops implicit null check is not generated.
4313 // This will allow to generate normal oop implicit null check.
4314 if (Matcher::gen_narrow_oop_implicit_null_checks())
4315 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
4316 //
4317 // This transformation together with CastPP transformation above
4318 // will generated code for implicit null checks for compressed oops.
4319 //
4320 // The original code after Optimize()
4321 //
4322 // LoadN memory, narrow_oop_reg
4323 // decode narrow_oop_reg, base_reg
4324 // CmpP base_reg, nullptr
4325 // CastPP base_reg // NotNull
4326 // Load [base_reg + offset], val_reg
4327 //
4328 // after these transformations will be
4329 //
4330 // LoadN memory, narrow_oop_reg
4331 // CmpN narrow_oop_reg, nullptr
4332 // decode_not_null narrow_oop_reg, base_reg
4333 // Load [base_reg + offset], val_reg
4334 //
4335 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
4336 // since narrow oops can be used in debug info now (see the code in
4337 // final_graph_reshaping_walk()).
4338 //
4339 // At the end the code will be matched to
4340 // on x86:
4341 //
4342 // Load_narrow_oop memory, narrow_oop_reg
4343 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
4344 // NullCheck narrow_oop_reg
4345 //
4346 // and on sparc:
4347 //
4348 // Load_narrow_oop memory, narrow_oop_reg
4349 // decode_not_null narrow_oop_reg, base_reg
4350 // Load [base_reg + offset], val_reg
4351 // NullCheck base_reg
4352 //
4353 } else if (t->isa_oopptr()) {
4354 new_in2 = ConNode::make(t->make_narrowoop());
4355 } else if (t->isa_klassptr()) {
4356 ciKlass* klass = t->is_klassptr()->exact_klass();
4357 if (klass->is_in_encoding_range()) {
4358 new_in2 = ConNode::make(t->make_narrowklass());
4359 }
4360 }
4361 }
4362 if (new_in2 != nullptr) {
4363 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
4364 n->subsume_by(cmpN, this);
4365 if (in1->outcnt() == 0) {
4366 in1->disconnect_inputs(this);
4367 }
4368 if (in2->outcnt() == 0) {
4369 in2->disconnect_inputs(this);
4370 }
4371 }
4372 }
4373 break;
4374
4375 case Op_DecodeN:
4376 case Op_DecodeNKlass:
4377 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4378 // DecodeN could be pinned when it can't be fold into
4379 // an address expression, see the code for Op_CastPP above.
4380 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4381 break;
4382
4383 case Op_EncodeP:
4384 case Op_EncodePKlass: {
4385 Node* in1 = n->in(1);
4386 if (in1->is_DecodeNarrowPtr()) {
4387 n->subsume_by(in1->in(1), this);
4388 } else if (in1->Opcode() == Op_ConP) {
4389 const Type* t = in1->bottom_type();
4390 if (t == TypePtr::NULL_PTR) {
4391 assert(t->isa_oopptr(), "null klass?");
4392 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4393 } else if (t->isa_oopptr()) {
4394 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4395 } else if (t->isa_klassptr()) {
4396 ciKlass* klass = t->is_klassptr()->exact_klass();
4397 if (klass->is_in_encoding_range()) {
4398 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4399 } else {
4400 assert(false, "unencodable klass in ConP -> EncodeP");
4401 C->record_failure("unencodable klass in ConP -> EncodeP");
4402 }
4403 }
4404 }
4405 if (in1->outcnt() == 0) {
4406 in1->disconnect_inputs(this);
4407 }
4408 break;
4409 }
4410
4411 case Op_Proj: {
4412 if (OptimizeStringConcat || IncrementalInline) {
4413 ProjNode* proj = n->as_Proj();
4414 if (proj->_is_io_use) {
4415 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4416 // Separate projections were used for the exception path which
4417 // are normally removed by a late inline. If it wasn't inlined
4418 // then they will hang around and should just be replaced with
4419 // the original one. Merge them.
4420 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4421 if (non_io_proj != nullptr) {
4422 proj->subsume_by(non_io_proj , this);
4423 }
4424 }
4425 }
4426 break;
4427 }
4428
4429 case Op_Phi:
4430 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4431 // The EncodeP optimization may create Phi with the same edges
4432 // for all paths. It is not handled well by Register Allocator.
4433 Node* unique_in = n->in(1);
4434 assert(unique_in != nullptr, "");
4435 uint cnt = n->req();
4436 for (uint i = 2; i < cnt; i++) {
4437 Node* m = n->in(i);
4438 assert(m != nullptr, "");
4439 if (unique_in != m)
4440 unique_in = nullptr;
4441 }
4442 if (unique_in != nullptr) {
4443 n->subsume_by(unique_in, this);
4444 }
4445 }
4446 break;
4447
4448 #endif
4449
4450 case Op_ModI:
4451 handle_div_mod_op(n, T_INT, false);
4452 break;
4453
4454 case Op_ModL:
4455 handle_div_mod_op(n, T_LONG, false);
4456 break;
4457
4458 case Op_UModI:
4459 handle_div_mod_op(n, T_INT, true);
4460 break;
4461
4462 case Op_UModL:
4463 handle_div_mod_op(n, T_LONG, true);
4464 break;
4465
4466 case Op_LoadVector:
4467 case Op_StoreVector:
4468 #ifdef ASSERT
4469 // Add VerifyVectorAlignment node between adr and load / store.
4470 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
4471 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
4472 n->as_StoreVector()->must_verify_alignment();
4473 if (must_verify_alignment) {
4474 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
4475 n->as_StoreVector()->memory_size();
4476 // The memory access should be aligned to the vector width in bytes.
4477 // However, the underlying array is possibly less well aligned, but at least
4478 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
4479 // a loop we can expect at least the following alignment:
4480 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
4481 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
4482 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
4483 // Create mask from alignment. e.g. 0b1000 -> 0b0111
4484 jlong mask = guaranteed_alignment - 1;
4485 Node* mask_con = ConLNode::make(mask);
4486 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
4487 n->set_req(MemNode::Address, va);
4488 }
4489 }
4490 #endif
4491 break;
4492
4493 case Op_PackB:
4494 case Op_PackS:
4495 case Op_PackI:
4496 case Op_PackF:
4497 case Op_PackL:
4498 case Op_PackD:
4499 if (n->req()-1 > 2) {
4500 // Replace many operand PackNodes with a binary tree for matching
4501 PackNode* p = (PackNode*) n;
4502 Node* btp = p->binary_tree_pack(1, n->req());
4503 n->subsume_by(btp, this);
4504 }
4505 break;
4506 case Op_Loop:
4507 // When StressCountedLoop is enabled, this loop may intentionally avoid a counted loop conversion.
4508 // This is expected behavior for the stress mode, which exercises alternative compilation paths.
4509 if (!StressCountedLoop) {
4510 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4511 }
4512 case Op_CountedLoop:
4513 case Op_LongCountedLoop:
4514 case Op_OuterStripMinedLoop:
4515 if (n->as_Loop()->is_inner_loop()) {
4516 frc.inc_inner_loop_count();
4517 }
4518 n->as_Loop()->verify_strip_mined(0);
4519 break;
4520 case Op_LShiftI:
4521 case Op_RShiftI:
4522 case Op_URShiftI:
4523 case Op_LShiftL:
4524 case Op_RShiftL:
4525 case Op_URShiftL:
4526 if (Matcher::need_masked_shift_count) {
4527 // The cpu's shift instructions don't restrict the count to the
4528 // lower 5/6 bits. We need to do the masking ourselves.
4529 Node* in2 = n->in(2);
4530 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4531 const TypeInt* t = in2->find_int_type();
4532 if (t != nullptr && t->is_con()) {
4533 juint shift = t->get_con();
4534 if (shift > mask) { // Unsigned cmp
4535 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4536 }
4537 } else {
4538 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
4539 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4540 n->set_req(2, shift);
4541 }
4542 }
4543 if (in2->outcnt() == 0) { // Remove dead node
4544 in2->disconnect_inputs(this);
4545 }
4546 }
4547 break;
4548 case Op_MemBarStoreStore:
4549 case Op_MemBarRelease:
4550 // Break the link with AllocateNode: it is no longer useful and
4551 // confuses register allocation.
4552 if (n->req() > MemBarNode::Precedent) {
4553 n->set_req(MemBarNode::Precedent, top());
4554 }
4555 break;
4556 case Op_MemBarAcquire: {
4557 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4558 // At parse time, the trailing MemBarAcquire for a volatile load
4559 // is created with an edge to the load. After optimizations,
4560 // that input may be a chain of Phis. If those phis have no
4561 // other use, then the MemBarAcquire keeps them alive and
4562 // register allocation can be confused.
4563 dead_nodes.push(n->in(MemBarNode::Precedent));
4564 n->set_req(MemBarNode::Precedent, top());
4565 }
4566 break;
4567 }
4568 case Op_RangeCheck: {
4569 RangeCheckNode* rc = n->as_RangeCheck();
4570 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4571 n->subsume_by(iff, this);
4572 frc._tests.push(iff);
4573 break;
4574 }
4575 case Op_ConvI2L: {
4576 if (!Matcher::convi2l_type_required) {
4577 // Code generation on some platforms doesn't need accurate
4578 // ConvI2L types. Widening the type can help remove redundant
4579 // address computations.
4580 n->as_Type()->set_type(TypeLong::INT);
4581 ResourceMark rm;
4582 Unique_Node_List wq;
4583 wq.push(n);
4584 for (uint next = 0; next < wq.size(); next++) {
4585 Node *m = wq.at(next);
4586
4587 for(;;) {
4588 // Loop over all nodes with identical inputs edges as m
4589 Node* k = m->find_similar(m->Opcode());
4590 if (k == nullptr) {
4591 break;
4592 }
4593 // Push their uses so we get a chance to remove node made
4594 // redundant
4595 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4596 Node* u = k->fast_out(i);
4597 if (u->Opcode() == Op_LShiftL ||
4598 u->Opcode() == Op_AddL ||
4599 u->Opcode() == Op_SubL ||
4600 u->Opcode() == Op_AddP) {
4601 wq.push(u);
4602 }
4603 }
4604 // Replace all nodes with identical edges as m with m
4605 k->subsume_by(m, this);
4606 }
4607 }
4608 }
4609 break;
4610 }
4611 case Op_CmpUL: {
4612 if (!Matcher::has_match_rule(Op_CmpUL)) {
4613 // No support for unsigned long comparisons
4614 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4615 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4616 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4617 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4618 Node* andl = new AndLNode(orl, remove_sign_mask);
4619 Node* cmp = new CmpLNode(andl, n->in(2));
4620 n->subsume_by(cmp, this);
4621 }
4622 break;
4623 }
4624 #ifdef ASSERT
4625 case Op_InlineType: {
4626 n->dump(-1);
4627 assert(false, "inline type node was not removed");
4628 break;
4629 }
4630 case Op_ConNKlass: {
4631 const TypePtr* tp = n->as_Type()->type()->make_ptr();
4632 ciKlass* klass = tp->is_klassptr()->exact_klass();
4633 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
4634 break;
4635 }
4636 #endif
4637 default:
4638 assert(!n->is_Call(), "");
4639 assert(!n->is_Mem(), "");
4640 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4641 break;
4642 }
4643 }
4644
4645 //------------------------------final_graph_reshaping_walk---------------------
4646 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4647 // requires that the walk visits a node's inputs before visiting the node.
4648 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4649 Unique_Node_List sfpt;
4650
4651 frc._visited.set(root->_idx); // first, mark node as visited
4652 uint cnt = root->req();
4653 Node *n = root;
4654 uint i = 0;
4655 while (true) {
4656 if (i < cnt) {
4657 // Place all non-visited non-null inputs onto stack
4658 Node* m = n->in(i);
4659 ++i;
4660 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
4661 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
4662 // compute worst case interpreter size in case of a deoptimization
4663 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4664
4665 sfpt.push(m);
4666 }
4667 cnt = m->req();
4668 nstack.push(n, i); // put on stack parent and next input's index
4669 n = m;
4670 i = 0;
4671 }
4672 } else {
4673 // Now do post-visit work
4674 final_graph_reshaping_impl(n, frc, dead_nodes);
4675 if (nstack.is_empty())
4676 break; // finished
4677 n = nstack.node(); // Get node from stack
4678 cnt = n->req();
4679 i = nstack.index();
4680 nstack.pop(); // Shift to the next node on stack
4681 }
4682 }
4683
4684 expand_reachability_edges(sfpt);
4685
4686 // Skip next transformation if compressed oops are not used.
4687 if (UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks())
4688 return;
4689
4690 // Go over ReachabilityFence nodes to skip DecodeN nodes for referents.
4691 // The sole purpose of RF node is to keep the referent oop alive and
4692 // decoding the oop for that is not needed.
4693 for (int i = 0; i < C->reachability_fences_count(); i++) {
4694 ReachabilityFenceNode* rf = C->reachability_fence(i);
4695 DecodeNNode* dn = rf->in(1)->isa_DecodeN();
4696 if (dn != nullptr) {
4697 if (!dn->has_non_debug_uses() || Matcher::narrow_oop_use_complex_address()) {
4698 rf->set_req(1, dn->in(1));
4699 if (dn->outcnt() == 0) {
4700 dn->disconnect_inputs(this);
4701 }
4702 }
4703 }
4704 }
4705
4706 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4707 // It could be done for an uncommon traps or any safepoints/calls
4708 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4709 while (sfpt.size() > 0) {
4710 n = sfpt.pop();
4711 JVMState *jvms = n->as_SafePoint()->jvms();
4712 assert(jvms != nullptr, "sanity");
4713 int start = jvms->debug_start();
4714 int end = n->req();
4715 bool is_uncommon = (n->is_CallStaticJava() &&
4716 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4717 for (int j = start; j < end; j++) {
4718 Node* in = n->in(j);
4719 if (in->is_DecodeNarrowPtr() && (is_uncommon || !in->has_non_debug_uses())) {
4720 n->set_req(j, in->in(1));
4721 if (in->outcnt() == 0) {
4722 in->disconnect_inputs(this);
4723 }
4724 }
4725 }
4726 }
4727 }
4728
4729 //------------------------------final_graph_reshaping--------------------------
4730 // Final Graph Reshaping.
4731 //
4732 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4733 // and not commoned up and forced early. Must come after regular
4734 // optimizations to avoid GVN undoing the cloning. Clone constant
4735 // inputs to Loop Phis; these will be split by the allocator anyways.
4736 // Remove Opaque nodes.
4737 // (2) Move last-uses by commutative operations to the left input to encourage
4738 // Intel update-in-place two-address operations and better register usage
4739 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4740 // calls canonicalizing them back.
4741 // (3) Detect infinite loops; blobs of code reachable from above but not
4742 // below. Several of the Code_Gen algorithms fail on such code shapes,
4743 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4744 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4745 // Detection is by looking for IfNodes where only 1 projection is
4746 // reachable from below or CatchNodes missing some targets.
4747 // (4) Assert for insane oop offsets in debug mode.
4748
4749 bool Compile::final_graph_reshaping() {
4750 // an infinite loop may have been eliminated by the optimizer,
4751 // in which case the graph will be empty.
4752 if (root()->req() == 1) {
4753 // Do not compile method that is only a trivial infinite loop,
4754 // since the content of the loop may have been eliminated.
4755 record_method_not_compilable("trivial infinite loop");
4756 return true;
4757 }
4758
4759 // Expensive nodes have their control input set to prevent the GVN
4760 // from freely commoning them. There's no GVN beyond this point so
4761 // no need to keep the control input. We want the expensive nodes to
4762 // be freely moved to the least frequent code path by gcm.
4763 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4764 for (int i = 0; i < expensive_count(); i++) {
4765 _expensive_nodes.at(i)->set_req(0, nullptr);
4766 }
4767
4768 Final_Reshape_Counts frc;
4769
4770 // Visit everybody reachable!
4771 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4772 Node_Stack nstack(live_nodes() >> 1);
4773 Unique_Node_List dead_nodes;
4774 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4775
4776 // Check for unreachable (from below) code (i.e., infinite loops).
4777 for( uint i = 0; i < frc._tests.size(); i++ ) {
4778 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4779 // Get number of CFG targets.
4780 // Note that PCTables include exception targets after calls.
4781 uint required_outcnt = n->required_outcnt();
4782 if (n->outcnt() != required_outcnt) {
4783 // Check for a few special cases. Rethrow Nodes never take the
4784 // 'fall-thru' path, so expected kids is 1 less.
4785 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4786 if (n->in(0)->in(0)->is_Call()) {
4787 CallNode* call = n->in(0)->in(0)->as_Call();
4788 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4789 required_outcnt--; // Rethrow always has 1 less kid
4790 } else if (call->req() > TypeFunc::Parms &&
4791 call->is_CallDynamicJava()) {
4792 // Check for null receiver. In such case, the optimizer has
4793 // detected that the virtual call will always result in a null
4794 // pointer exception. The fall-through projection of this CatchNode
4795 // will not be populated.
4796 Node* arg0 = call->in(TypeFunc::Parms);
4797 if (arg0->is_Type() &&
4798 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4799 required_outcnt--;
4800 }
4801 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4802 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4803 // Check for illegal array length. In such case, the optimizer has
4804 // detected that the allocation attempt will always result in an
4805 // exception. There is no fall-through projection of this CatchNode .
4806 assert(call->is_CallStaticJava(), "static call expected");
4807 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4808 uint valid_length_test_input = call->req() - 1;
4809 Node* valid_length_test = call->in(valid_length_test_input);
4810 call->del_req(valid_length_test_input);
4811 if (valid_length_test->find_int_con(1) == 0) {
4812 required_outcnt--;
4813 }
4814 dead_nodes.push(valid_length_test);
4815 assert(n->outcnt() == required_outcnt, "malformed control flow");
4816 continue;
4817 }
4818 }
4819 }
4820
4821 // Recheck with a better notion of 'required_outcnt'
4822 if (n->outcnt() != required_outcnt) {
4823 record_method_not_compilable("malformed control flow");
4824 return true; // Not all targets reachable!
4825 }
4826 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4827 CallNode* call = n->in(0)->in(0)->as_Call();
4828 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4829 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4830 assert(call->is_CallStaticJava(), "static call expected");
4831 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4832 uint valid_length_test_input = call->req() - 1;
4833 dead_nodes.push(call->in(valid_length_test_input));
4834 call->del_req(valid_length_test_input); // valid length test useless now
4835 }
4836 }
4837 // Check that I actually visited all kids. Unreached kids
4838 // must be infinite loops.
4839 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4840 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4841 record_method_not_compilable("infinite loop");
4842 return true; // Found unvisited kid; must be unreach
4843 }
4844
4845 // Here so verification code in final_graph_reshaping_walk()
4846 // always see an OuterStripMinedLoopEnd
4847 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4848 IfNode* init_iff = n->as_If();
4849 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4850 n->subsume_by(iff, this);
4851 }
4852 }
4853
4854 while (dead_nodes.size() > 0) {
4855 Node* m = dead_nodes.pop();
4856 if (m->outcnt() == 0 && m != top()) {
4857 for (uint j = 0; j < m->req(); j++) {
4858 Node* in = m->in(j);
4859 if (in != nullptr) {
4860 dead_nodes.push(in);
4861 }
4862 }
4863 m->disconnect_inputs(this);
4864 }
4865 }
4866
4867 set_java_calls(frc.get_java_call_count());
4868 set_inner_loops(frc.get_inner_loop_count());
4869
4870 // No infinite loops, no reason to bail out.
4871 return false;
4872 }
4873
4874 //-----------------------------too_many_traps----------------------------------
4875 // Report if there are too many traps at the current method and bci.
4876 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4877 bool Compile::too_many_traps(ciMethod* method,
4878 int bci,
4879 Deoptimization::DeoptReason reason) {
4880 ciMethodData* md = method->method_data();
4881 if (md->is_empty()) {
4882 // Assume the trap has not occurred, or that it occurred only
4883 // because of a transient condition during start-up in the interpreter.
4884 return false;
4885 }
4886 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4887 if (md->has_trap_at(bci, m, reason) != 0) {
4888 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4889 // Also, if there are multiple reasons, or if there is no per-BCI record,
4890 // assume the worst.
4891 if (log())
4892 log()->elem("observe trap='%s' count='%d'",
4893 Deoptimization::trap_reason_name(reason),
4894 md->trap_count(reason));
4895 return true;
4896 } else {
4897 // Ignore method/bci and see if there have been too many globally.
4898 return too_many_traps(reason, md);
4899 }
4900 }
4901
4902 // Less-accurate variant which does not require a method and bci.
4903 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4904 ciMethodData* logmd) {
4905 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4906 // Too many traps globally.
4907 // Note that we use cumulative trap_count, not just md->trap_count.
4908 if (log()) {
4909 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4910 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4911 Deoptimization::trap_reason_name(reason),
4912 mcount, trap_count(reason));
4913 }
4914 return true;
4915 } else {
4916 // The coast is clear.
4917 return false;
4918 }
4919 }
4920
4921 //--------------------------too_many_recompiles--------------------------------
4922 // Report if there are too many recompiles at the current method and bci.
4923 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4924 // Is not eager to return true, since this will cause the compiler to use
4925 // Action_none for a trap point, to avoid too many recompilations.
4926 bool Compile::too_many_recompiles(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 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4936 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4937 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4938 Deoptimization::DeoptReason per_bc_reason
4939 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4940 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4941 if ((per_bc_reason == Deoptimization::Reason_none
4942 || md->has_trap_at(bci, m, reason) != 0)
4943 // The trap frequency measure we care about is the recompile count:
4944 && md->trap_recompiled_at(bci, m)
4945 && md->overflow_recompile_count() >= bc_cutoff) {
4946 // Do not emit a trap here if it has already caused recompilations.
4947 // Also, if there are multiple reasons, or if there is no per-BCI record,
4948 // assume the worst.
4949 if (log())
4950 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4951 Deoptimization::trap_reason_name(reason),
4952 md->trap_count(reason),
4953 md->overflow_recompile_count());
4954 return true;
4955 } else if (trap_count(reason) != 0
4956 && decompile_count() >= m_cutoff) {
4957 // Too many recompiles globally, and we have seen this sort of trap.
4958 // Use cumulative decompile_count, not just md->decompile_count.
4959 if (log())
4960 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4961 Deoptimization::trap_reason_name(reason),
4962 md->trap_count(reason), trap_count(reason),
4963 md->decompile_count(), decompile_count());
4964 return true;
4965 } else {
4966 // The coast is clear.
4967 return false;
4968 }
4969 }
4970
4971 // Compute when not to trap. Used by matching trap based nodes and
4972 // NullCheck optimization.
4973 void Compile::set_allowed_deopt_reasons() {
4974 _allowed_reasons = 0;
4975 if (is_method_compilation()) {
4976 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4977 assert(rs < BitsPerInt, "recode bit map");
4978 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4979 _allowed_reasons |= nth_bit(rs);
4980 }
4981 }
4982 }
4983 }
4984
4985 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4986 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4987 }
4988
4989 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4990 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4991 }
4992
4993 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4994 if (holder->is_initialized()) {
4995 return false;
4996 }
4997 if (holder->is_being_initialized()) {
4998 if (accessing_method->holder() == holder) {
4999 // Access inside a class. The barrier can be elided when access happens in <clinit>,
5000 // <init>, or a static method. In all those cases, there was an initialization
5001 // barrier on the holder klass passed.
5002 if (accessing_method->is_class_initializer() ||
5003 accessing_method->is_object_constructor() ||
5004 accessing_method->is_static()) {
5005 return false;
5006 }
5007 } else if (accessing_method->holder()->is_subclass_of(holder)) {
5008 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
5009 // In case of <init> or a static method, the barrier is on the subclass is not enough:
5010 // child class can become fully initialized while its parent class is still being initialized.
5011 if (accessing_method->is_class_initializer()) {
5012 return false;
5013 }
5014 }
5015 ciMethod* root = method(); // the root method of compilation
5016 if (root != accessing_method) {
5017 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
5018 }
5019 }
5020 return true;
5021 }
5022
5023 #ifndef PRODUCT
5024 //------------------------------verify_bidirectional_edges---------------------
5025 // For each input edge to a node (ie - for each Use-Def edge), verify that
5026 // there is a corresponding Def-Use edge.
5027 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
5028 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
5029 uint stack_size = live_nodes() >> 4;
5030 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
5031 if (root_and_safepoints != nullptr) {
5032 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
5033 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
5034 Node* root_or_safepoint = root_and_safepoints->at(i);
5035 // If the node is a safepoint, let's check if it still has a control input
5036 // Lack of control input signifies that this node was killed by CCP or
5037 // recursively by remove_globally_dead_node and it shouldn't be a starting
5038 // point.
5039 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
5040 nstack.push(root_or_safepoint);
5041 }
5042 }
5043 } else {
5044 nstack.push(_root);
5045 }
5046
5047 while (nstack.size() > 0) {
5048 Node* n = nstack.pop();
5049 if (visited.member(n)) {
5050 continue;
5051 }
5052 visited.push(n);
5053
5054 // Walk over all input edges, checking for correspondence
5055 uint length = n->len();
5056 for (uint i = 0; i < length; i++) {
5057 Node* in = n->in(i);
5058 if (in != nullptr && !visited.member(in)) {
5059 nstack.push(in); // Put it on stack
5060 }
5061 if (in != nullptr && !in->is_top()) {
5062 // Count instances of `next`
5063 int cnt = 0;
5064 for (uint idx = 0; idx < in->_outcnt; idx++) {
5065 if (in->_out[idx] == n) {
5066 cnt++;
5067 }
5068 }
5069 assert(cnt > 0, "Failed to find Def-Use edge.");
5070 // Check for duplicate edges
5071 // walk the input array downcounting the input edges to n
5072 for (uint j = 0; j < length; j++) {
5073 if (n->in(j) == in) {
5074 cnt--;
5075 }
5076 }
5077 assert(cnt == 0, "Mismatched edge count.");
5078 } else if (in == nullptr) {
5079 assert(i == 0 || i >= n->req() ||
5080 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
5081 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
5082 (n->is_Unlock() && i == (n->req() - 1)) ||
5083 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
5084 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
5085 } else {
5086 assert(in->is_top(), "sanity");
5087 // Nothing to check.
5088 }
5089 }
5090 }
5091 }
5092
5093 //------------------------------verify_graph_edges---------------------------
5094 // Walk the Graph and verify that there is a one-to-one correspondence
5095 // between Use-Def edges and Def-Use edges in the graph.
5096 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
5097 if (VerifyGraphEdges) {
5098 Unique_Node_List visited;
5099
5100 // Call graph walk to check edges
5101 verify_bidirectional_edges(visited, root_and_safepoints);
5102 if (no_dead_code) {
5103 // Now make sure that no visited node is used by an unvisited node.
5104 bool dead_nodes = false;
5105 Unique_Node_List checked;
5106 while (visited.size() > 0) {
5107 Node* n = visited.pop();
5108 checked.push(n);
5109 for (uint i = 0; i < n->outcnt(); i++) {
5110 Node* use = n->raw_out(i);
5111 if (checked.member(use)) continue; // already checked
5112 if (visited.member(use)) continue; // already in the graph
5113 if (use->is_Con()) continue; // a dead ConNode is OK
5114 // At this point, we have found a dead node which is DU-reachable.
5115 if (!dead_nodes) {
5116 tty->print_cr("*** Dead nodes reachable via DU edges:");
5117 dead_nodes = true;
5118 }
5119 use->dump(2);
5120 tty->print_cr("---");
5121 checked.push(use); // No repeats; pretend it is now checked.
5122 }
5123 }
5124 assert(!dead_nodes, "using nodes must be reachable from root");
5125 }
5126 }
5127 }
5128 #endif
5129
5130 // The Compile object keeps track of failure reasons separately from the ciEnv.
5131 // This is required because there is not quite a 1-1 relation between the
5132 // ciEnv and its compilation task and the Compile object. Note that one
5133 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
5134 // to backtrack and retry without subsuming loads. Other than this backtracking
5135 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
5136 // by the logic in C2Compiler.
5137 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
5138 if (log() != nullptr) {
5139 log()->elem("failure reason='%s' phase='compile'", reason);
5140 }
5141 if (_failure_reason.get() == nullptr) {
5142 // Record the first failure reason.
5143 _failure_reason.set(reason);
5144 if (CaptureBailoutInformation) {
5145 _first_failure_details = new CompilationFailureInfo(reason);
5146 }
5147 } else {
5148 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
5149 }
5150
5151 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
5152 C->print_method(PHASE_FAILURE, 1);
5153 }
5154 _root = nullptr; // flush the graph, too
5155 }
5156
5157 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
5158 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
5159 _compile(Compile::current()),
5160 _log(nullptr),
5161 _dolog(CITimeVerbose)
5162 {
5163 assert(_compile != nullptr, "sanity check");
5164 assert(id != PhaseTraceId::_t_none, "Don't use none");
5165 if (_dolog) {
5166 _log = _compile->log();
5167 }
5168 if (_log != nullptr) {
5169 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
5170 _log->stamp();
5171 _log->end_head();
5172 }
5173
5174 // Inform memory statistic, if enabled
5175 if (CompilationMemoryStatistic::enabled()) {
5176 CompilationMemoryStatistic::on_phase_start((int)id, name);
5177 }
5178 }
5179
5180 Compile::TracePhase::TracePhase(PhaseTraceId id)
5181 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
5182
5183 Compile::TracePhase::~TracePhase() {
5184
5185 // Inform memory statistic, if enabled
5186 if (CompilationMemoryStatistic::enabled()) {
5187 CompilationMemoryStatistic::on_phase_end();
5188 }
5189
5190 if (_compile->failing_internal()) {
5191 if (_log != nullptr) {
5192 _log->done("phase");
5193 }
5194 return; // timing code, not stressing bailouts.
5195 }
5196 #ifdef ASSERT
5197 if (PrintIdealNodeCount) {
5198 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
5199 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
5200 }
5201
5202 if (VerifyIdealNodeCount) {
5203 _compile->print_missing_nodes();
5204 }
5205 #endif
5206
5207 if (_log != nullptr) {
5208 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
5209 }
5210 }
5211
5212 //----------------------------static_subtype_check-----------------------------
5213 // Shortcut important common cases when superklass is exact:
5214 // (0) superklass is java.lang.Object (can occur in reflective code)
5215 // (1) subklass is already limited to a subtype of superklass => always ok
5216 // (2) subklass does not overlap with superklass => always fail
5217 // (3) superklass has NO subtypes and we can check with a simple compare.
5218 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
5219 if (skip) {
5220 return SSC_full_test; // Let caller generate the general case.
5221 }
5222
5223 if (subk->is_java_subtype_of(superk)) {
5224 return SSC_always_true; // (0) and (1) this test cannot fail
5225 }
5226
5227 if (!subk->maybe_java_subtype_of(superk)) {
5228 return SSC_always_false; // (2) true path dead; no dynamic test needed
5229 }
5230
5231 const Type* superelem = superk;
5232 if (superk->isa_aryklassptr()) {
5233 int ignored;
5234 superelem = superk->is_aryklassptr()->base_element_type(ignored);
5235
5236 // Do not fold the subtype check to an array klass pointer comparison for null-able inline type arrays
5237 // because null-free [LMyValue <: null-able [LMyValue but the klasses are different. Perform a full test.
5238 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
5239 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
5240 return SSC_full_test;
5241 }
5242 }
5243
5244 if (superelem->isa_instklassptr()) {
5245 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
5246 if (!ik->has_subklass()) {
5247 if (!ik->is_final()) {
5248 // Add a dependency if there is a chance of a later subclass.
5249 dependencies()->assert_leaf_type(ik);
5250 }
5251 if (!superk->maybe_java_subtype_of(subk)) {
5252 return SSC_always_false;
5253 }
5254 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5255 }
5256 } else {
5257 // A primitive array type has no subtypes.
5258 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5259 }
5260
5261 return SSC_full_test;
5262 }
5263
5264 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
5265 #ifdef _LP64
5266 // The scaled index operand to AddP must be a clean 64-bit value.
5267 // Java allows a 32-bit int to be incremented to a negative
5268 // value, which appears in a 64-bit register as a large
5269 // positive number. Using that large positive number as an
5270 // operand in pointer arithmetic has bad consequences.
5271 // On the other hand, 32-bit overflow is rare, and the possibility
5272 // can often be excluded, if we annotate the ConvI2L node with
5273 // a type assertion that its value is known to be a small positive
5274 // number. (The prior range check has ensured this.)
5275 // This assertion is used by ConvI2LNode::Ideal.
5276 int index_max = max_jint - 1; // array size is max_jint, index is one less
5277 if (sizetype != nullptr && sizetype->_hi > 0) {
5278 index_max = sizetype->_hi - 1;
5279 }
5280 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
5281 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
5282 #endif
5283 return idx;
5284 }
5285
5286 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
5287 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
5288 if (ctrl != nullptr) {
5289 // Express control dependency by a CastII node with a narrow type.
5290 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
5291 // node from floating above the range check during loop optimizations. Otherwise, the
5292 // ConvI2L node may be eliminated independently of the range check, causing the data path
5293 // to become TOP while the control path is still there (although it's unreachable).
5294 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
5295 value = phase->transform(value);
5296 }
5297 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
5298 return phase->transform(new ConvI2LNode(value, ltype));
5299 }
5300
5301 void Compile::dump_print_inlining() {
5302 inline_printer()->print_on(tty);
5303 }
5304
5305 void Compile::log_late_inline(CallGenerator* cg) {
5306 if (log() != nullptr) {
5307 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5308 cg->unique_id());
5309 JVMState* p = cg->call_node()->jvms();
5310 while (p != nullptr) {
5311 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5312 p = p->caller();
5313 }
5314 log()->tail("late_inline");
5315 }
5316 }
5317
5318 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5319 log_late_inline(cg);
5320 if (log() != nullptr) {
5321 log()->inline_fail(msg);
5322 }
5323 }
5324
5325 void Compile::log_inline_id(CallGenerator* cg) {
5326 if (log() != nullptr) {
5327 // The LogCompilation tool needs a unique way to identify late
5328 // inline call sites. This id must be unique for this call site in
5329 // this compilation. Try to have it unique across compilations as
5330 // well because it can be convenient when grepping through the log
5331 // file.
5332 // Distinguish OSR compilations from others in case CICountOSR is
5333 // on.
5334 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5335 cg->set_unique_id(id);
5336 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5337 }
5338 }
5339
5340 void Compile::log_inline_failure(const char* msg) {
5341 if (C->log() != nullptr) {
5342 C->log()->inline_fail(msg);
5343 }
5344 }
5345
5346
5347 // Dump inlining replay data to the stream.
5348 // Don't change thread state and acquire any locks.
5349 void Compile::dump_inline_data(outputStream* out) {
5350 InlineTree* inl_tree = ilt();
5351 if (inl_tree != nullptr) {
5352 out->print(" inline %d", inl_tree->count());
5353 inl_tree->dump_replay_data(out);
5354 }
5355 }
5356
5357 void Compile::dump_inline_data_reduced(outputStream* out) {
5358 assert(ReplayReduce, "");
5359
5360 InlineTree* inl_tree = ilt();
5361 if (inl_tree == nullptr) {
5362 return;
5363 }
5364 // Enable iterative replay file reduction
5365 // Output "compile" lines for depth 1 subtrees,
5366 // simulating that those trees were compiled
5367 // instead of inlined.
5368 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5369 InlineTree* sub = inl_tree->subtrees().at(i);
5370 if (sub->inline_level() != 1) {
5371 continue;
5372 }
5373
5374 ciMethod* method = sub->method();
5375 int entry_bci = -1;
5376 int comp_level = env()->task()->comp_level();
5377 out->print("compile ");
5378 method->dump_name_as_ascii(out);
5379 out->print(" %d %d", entry_bci, comp_level);
5380 out->print(" inline %d", sub->count());
5381 sub->dump_replay_data(out, -1);
5382 out->cr();
5383 }
5384 }
5385
5386 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5387 if (n1->Opcode() < n2->Opcode()) return -1;
5388 else if (n1->Opcode() > n2->Opcode()) return 1;
5389
5390 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5391 for (uint i = 1; i < n1->req(); i++) {
5392 if (n1->in(i) < n2->in(i)) return -1;
5393 else if (n1->in(i) > n2->in(i)) return 1;
5394 }
5395
5396 return 0;
5397 }
5398
5399 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5400 Node* n1 = *n1p;
5401 Node* n2 = *n2p;
5402
5403 return cmp_expensive_nodes(n1, n2);
5404 }
5405
5406 void Compile::sort_expensive_nodes() {
5407 if (!expensive_nodes_sorted()) {
5408 _expensive_nodes.sort(cmp_expensive_nodes);
5409 }
5410 }
5411
5412 bool Compile::expensive_nodes_sorted() const {
5413 for (int i = 1; i < _expensive_nodes.length(); i++) {
5414 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5415 return false;
5416 }
5417 }
5418 return true;
5419 }
5420
5421 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5422 if (_expensive_nodes.length() == 0) {
5423 return false;
5424 }
5425
5426 assert(OptimizeExpensiveOps, "optimization off?");
5427
5428 // Take this opportunity to remove dead nodes from the list
5429 int j = 0;
5430 for (int i = 0; i < _expensive_nodes.length(); i++) {
5431 Node* n = _expensive_nodes.at(i);
5432 if (!n->is_unreachable(igvn)) {
5433 assert(n->is_expensive(), "should be expensive");
5434 _expensive_nodes.at_put(j, n);
5435 j++;
5436 }
5437 }
5438 _expensive_nodes.trunc_to(j);
5439
5440 // Then sort the list so that similar nodes are next to each other
5441 // and check for at least two nodes of identical kind with same data
5442 // inputs.
5443 sort_expensive_nodes();
5444
5445 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5446 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5447 return true;
5448 }
5449 }
5450
5451 return false;
5452 }
5453
5454 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5455 if (_expensive_nodes.length() == 0) {
5456 return;
5457 }
5458
5459 assert(OptimizeExpensiveOps, "optimization off?");
5460
5461 // Sort to bring similar nodes next to each other and clear the
5462 // control input of nodes for which there's only a single copy.
5463 sort_expensive_nodes();
5464
5465 int j = 0;
5466 int identical = 0;
5467 int i = 0;
5468 bool modified = false;
5469 for (; i < _expensive_nodes.length()-1; i++) {
5470 assert(j <= i, "can't write beyond current index");
5471 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5472 identical++;
5473 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5474 continue;
5475 }
5476 if (identical > 0) {
5477 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5478 identical = 0;
5479 } else {
5480 Node* n = _expensive_nodes.at(i);
5481 igvn.replace_input_of(n, 0, nullptr);
5482 igvn.hash_insert(n);
5483 modified = true;
5484 }
5485 }
5486 if (identical > 0) {
5487 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5488 } else if (_expensive_nodes.length() >= 1) {
5489 Node* n = _expensive_nodes.at(i);
5490 igvn.replace_input_of(n, 0, nullptr);
5491 igvn.hash_insert(n);
5492 modified = true;
5493 }
5494 _expensive_nodes.trunc_to(j);
5495 if (modified) {
5496 igvn.optimize();
5497 }
5498 }
5499
5500 void Compile::add_expensive_node(Node * n) {
5501 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5502 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5503 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5504 if (OptimizeExpensiveOps) {
5505 _expensive_nodes.append(n);
5506 } else {
5507 // Clear control input and let IGVN optimize expensive nodes if
5508 // OptimizeExpensiveOps is off.
5509 n->set_req(0, nullptr);
5510 }
5511 }
5512
5513 /**
5514 * Track coarsened Lock and Unlock nodes.
5515 */
5516
5517 class Lock_List : public Node_List {
5518 uint _origin_cnt;
5519 public:
5520 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5521 uint origin_cnt() const { return _origin_cnt; }
5522 };
5523
5524 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5525 int length = locks.length();
5526 if (length > 0) {
5527 // Have to keep this list until locks elimination during Macro nodes elimination.
5528 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5529 AbstractLockNode* alock = locks.at(0);
5530 BoxLockNode* box = alock->box_node()->as_BoxLock();
5531 for (int i = 0; i < length; i++) {
5532 AbstractLockNode* lock = locks.at(i);
5533 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5534 locks_list->push(lock);
5535 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
5536 if (this_box != box) {
5537 // Locking regions (BoxLock) could be Unbalanced here:
5538 // - its coarsened locks were eliminated in earlier
5539 // macro nodes elimination followed by loop unroll
5540 // - it is OSR locking region (no Lock node)
5541 // Preserve Unbalanced status in such cases.
5542 if (!this_box->is_unbalanced()) {
5543 this_box->set_coarsened();
5544 }
5545 if (!box->is_unbalanced()) {
5546 box->set_coarsened();
5547 }
5548 }
5549 }
5550 _coarsened_locks.append(locks_list);
5551 }
5552 }
5553
5554 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5555 int count = coarsened_count();
5556 for (int i = 0; i < count; i++) {
5557 Node_List* locks_list = _coarsened_locks.at(i);
5558 for (uint j = 0; j < locks_list->size(); j++) {
5559 Node* lock = locks_list->at(j);
5560 assert(lock->is_AbstractLock(), "sanity");
5561 if (!useful.member(lock)) {
5562 locks_list->yank(lock);
5563 }
5564 }
5565 }
5566 }
5567
5568 void Compile::remove_coarsened_lock(Node* n) {
5569 if (n->is_AbstractLock()) {
5570 int count = coarsened_count();
5571 for (int i = 0; i < count; i++) {
5572 Node_List* locks_list = _coarsened_locks.at(i);
5573 locks_list->yank(n);
5574 }
5575 }
5576 }
5577
5578 bool Compile::coarsened_locks_consistent() {
5579 int count = coarsened_count();
5580 for (int i = 0; i < count; i++) {
5581 bool unbalanced = false;
5582 bool modified = false; // track locks kind modifications
5583 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5584 uint size = locks_list->size();
5585 if (size == 0) {
5586 unbalanced = false; // All locks were eliminated - good
5587 } else if (size != locks_list->origin_cnt()) {
5588 unbalanced = true; // Some locks were removed from list
5589 } else {
5590 for (uint j = 0; j < size; j++) {
5591 Node* lock = locks_list->at(j);
5592 // All nodes in group should have the same state (modified or not)
5593 if (!lock->as_AbstractLock()->is_coarsened()) {
5594 if (j == 0) {
5595 // first on list was modified, the rest should be too for consistency
5596 modified = true;
5597 } else if (!modified) {
5598 // this lock was modified but previous locks on the list were not
5599 unbalanced = true;
5600 break;
5601 }
5602 } else if (modified) {
5603 // previous locks on list were modified but not this lock
5604 unbalanced = true;
5605 break;
5606 }
5607 }
5608 }
5609 if (unbalanced) {
5610 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5611 #ifdef ASSERT
5612 if (PrintEliminateLocks) {
5613 tty->print_cr("=== unbalanced coarsened locks ===");
5614 for (uint l = 0; l < size; l++) {
5615 locks_list->at(l)->dump();
5616 }
5617 }
5618 #endif
5619 record_failure(C2Compiler::retry_no_locks_coarsening());
5620 return false;
5621 }
5622 }
5623 return true;
5624 }
5625
5626 // Mark locking regions (identified by BoxLockNode) as unbalanced if
5627 // locks coarsening optimization removed Lock/Unlock nodes from them.
5628 // Such regions become unbalanced because coarsening only removes part
5629 // of Lock/Unlock nodes in region. As result we can't execute other
5630 // locks elimination optimizations which assume all code paths have
5631 // corresponding pair of Lock/Unlock nodes - they are balanced.
5632 void Compile::mark_unbalanced_boxes() const {
5633 int count = coarsened_count();
5634 for (int i = 0; i < count; i++) {
5635 Node_List* locks_list = _coarsened_locks.at(i);
5636 uint size = locks_list->size();
5637 if (size > 0) {
5638 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
5639 BoxLockNode* box = alock->box_node()->as_BoxLock();
5640 if (alock->is_coarsened()) {
5641 // coarsened_locks_consistent(), which is called before this method, verifies
5642 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
5643 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5644 for (uint j = 1; j < size; j++) {
5645 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
5646 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
5647 if (box != this_box) {
5648 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5649 box->set_unbalanced();
5650 this_box->set_unbalanced();
5651 }
5652 }
5653 }
5654 }
5655 }
5656 }
5657
5658 /**
5659 * Remove the speculative part of types and clean up the graph
5660 */
5661 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5662 if (UseTypeSpeculation) {
5663 Unique_Node_List worklist;
5664 worklist.push(root());
5665 int modified = 0;
5666 // Go over all type nodes that carry a speculative type, drop the
5667 // speculative part of the type and enqueue the node for an igvn
5668 // which may optimize it out.
5669 for (uint next = 0; next < worklist.size(); ++next) {
5670 Node *n = worklist.at(next);
5671 if (n->is_Type()) {
5672 TypeNode* tn = n->as_Type();
5673 const Type* t = tn->type();
5674 const Type* t_no_spec = t->remove_speculative();
5675 if (t_no_spec != t) {
5676 bool in_hash = igvn.hash_delete(n);
5677 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5678 tn->set_type(t_no_spec);
5679 igvn.hash_insert(n);
5680 igvn._worklist.push(n); // give it a chance to go away
5681 modified++;
5682 }
5683 }
5684 // Iterate over outs - endless loops is unreachable from below
5685 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5686 Node *m = n->fast_out(i);
5687 if (not_a_node(m)) {
5688 continue;
5689 }
5690 worklist.push(m);
5691 }
5692 }
5693 // Drop the speculative part of all types in the igvn's type table
5694 igvn.remove_speculative_types();
5695 if (modified > 0) {
5696 igvn.optimize();
5697 if (failing()) return;
5698 }
5699 #ifdef ASSERT
5700 // Verify that after the IGVN is over no speculative type has resurfaced
5701 worklist.clear();
5702 worklist.push(root());
5703 for (uint next = 0; next < worklist.size(); ++next) {
5704 Node *n = worklist.at(next);
5705 const Type* t = igvn.type_or_null(n);
5706 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5707 if (n->is_Type()) {
5708 t = n->as_Type()->type();
5709 assert(t == t->remove_speculative(), "no more speculative types");
5710 }
5711 // Iterate over outs - endless loops is unreachable from below
5712 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5713 Node *m = n->fast_out(i);
5714 if (not_a_node(m)) {
5715 continue;
5716 }
5717 worklist.push(m);
5718 }
5719 }
5720 igvn.check_no_speculative_types();
5721 #endif
5722 }
5723 }
5724
5725 // Auxiliary methods to support randomized stressing/fuzzing.
5726
5727 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5728 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5729 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5730 FLAG_SET_ERGO(StressSeed, _stress_seed);
5731 } else {
5732 _stress_seed = StressSeed;
5733 }
5734 if (_log != nullptr) {
5735 _log->elem("stress_test seed='%u'", _stress_seed);
5736 }
5737 }
5738
5739 int Compile::random() {
5740 _stress_seed = os::next_random(_stress_seed);
5741 return static_cast<int>(_stress_seed);
5742 }
5743
5744 // This method can be called the arbitrary number of times, with current count
5745 // as the argument. The logic allows selecting a single candidate from the
5746 // running list of candidates as follows:
5747 // int count = 0;
5748 // Cand* selected = null;
5749 // while(cand = cand->next()) {
5750 // if (randomized_select(++count)) {
5751 // selected = cand;
5752 // }
5753 // }
5754 //
5755 // Including count equalizes the chances any candidate is "selected".
5756 // This is useful when we don't have the complete list of candidates to choose
5757 // from uniformly. In this case, we need to adjust the randomicity of the
5758 // selection, or else we will end up biasing the selection towards the latter
5759 // candidates.
5760 //
5761 // Quick back-envelope calculation shows that for the list of n candidates
5762 // the equal probability for the candidate to persist as "best" can be
5763 // achieved by replacing it with "next" k-th candidate with the probability
5764 // of 1/k. It can be easily shown that by the end of the run, the
5765 // probability for any candidate is converged to 1/n, thus giving the
5766 // uniform distribution among all the candidates.
5767 //
5768 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5769 #define RANDOMIZED_DOMAIN_POW 29
5770 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5771 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5772 bool Compile::randomized_select(int count) {
5773 assert(count > 0, "only positive");
5774 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5775 }
5776
5777 #ifdef ASSERT
5778 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5779 bool Compile::fail_randomly() {
5780 if ((random() % StressBailoutMean) != 0) {
5781 return false;
5782 }
5783 record_failure("StressBailout");
5784 return true;
5785 }
5786
5787 bool Compile::failure_is_artificial() {
5788 return C->failure_reason_is("StressBailout");
5789 }
5790 #endif
5791
5792 CloneMap& Compile::clone_map() { return _clone_map; }
5793 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5794
5795 void NodeCloneInfo::dump_on(outputStream* st) const {
5796 st->print(" {%d:%d} ", idx(), gen());
5797 }
5798
5799 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5800 uint64_t val = value(old->_idx);
5801 NodeCloneInfo cio(val);
5802 assert(val != 0, "old node should be in the map");
5803 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5804 insert(nnn->_idx, cin.get());
5805 #ifndef PRODUCT
5806 if (is_debug()) {
5807 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5808 }
5809 #endif
5810 }
5811
5812 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5813 NodeCloneInfo cio(value(old->_idx));
5814 if (cio.get() == 0) {
5815 cio.set(old->_idx, 0);
5816 insert(old->_idx, cio.get());
5817 #ifndef PRODUCT
5818 if (is_debug()) {
5819 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5820 }
5821 #endif
5822 }
5823 clone(old, nnn, gen);
5824 }
5825
5826 int CloneMap::max_gen() const {
5827 int g = 0;
5828 DictI di(_dict);
5829 for(; di.test(); ++di) {
5830 int t = gen(di._key);
5831 if (g < t) {
5832 g = t;
5833 #ifndef PRODUCT
5834 if (is_debug()) {
5835 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5836 }
5837 #endif
5838 }
5839 }
5840 return g;
5841 }
5842
5843 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5844 uint64_t val = value(key);
5845 if (val != 0) {
5846 NodeCloneInfo ni(val);
5847 ni.dump_on(st);
5848 }
5849 }
5850
5851 void Compile::shuffle_macro_nodes() {
5852 shuffle_array(*C, _macro_nodes);
5853 }
5854
5855 // Move Allocate nodes to the start of the list
5856 void Compile::sort_macro_nodes() {
5857 int count = macro_count();
5858 int allocates = 0;
5859 for (int i = 0; i < count; i++) {
5860 Node* n = macro_node(i);
5861 if (n->is_Allocate()) {
5862 if (i != allocates) {
5863 Node* tmp = macro_node(allocates);
5864 _macro_nodes.at_put(allocates, n);
5865 _macro_nodes.at_put(i, tmp);
5866 }
5867 allocates++;
5868 }
5869 }
5870 }
5871
5872 void Compile::print_method(CompilerPhaseType compile_phase, int level, Node* n) {
5873 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5874 EventCompilerPhase event(UNTIMED);
5875 if (event.should_commit()) {
5876 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, compile_phase, C->_compile_id, level);
5877 }
5878 #ifndef PRODUCT
5879 ResourceMark rm;
5880 stringStream ss;
5881 ss.print_raw(CompilerPhaseTypeHelper::to_description(compile_phase));
5882 int iter = ++_igv_phase_iter[compile_phase];
5883 if (iter > 1) {
5884 ss.print(" %d", iter);
5885 }
5886 if (n != nullptr) {
5887 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5888 if (n->is_Call()) {
5889 CallNode* call = n->as_Call();
5890 if (call->_name != nullptr) {
5891 // E.g. uncommon traps etc.
5892 ss.print(" - %s", call->_name);
5893 } else if (call->is_CallJava()) {
5894 CallJavaNode* call_java = call->as_CallJava();
5895 if (call_java->method() != nullptr) {
5896 ss.print(" -");
5897 call_java->method()->print_short_name(&ss);
5898 }
5899 }
5900 }
5901 }
5902
5903 const char* name = ss.as_string();
5904 if (should_print_igv(level)) {
5905 _igv_printer->print_graph(name);
5906 }
5907 if (should_print_phase(level)) {
5908 print_phase(name);
5909 }
5910 if (should_print_ideal_phase(compile_phase)) {
5911 print_ideal_ir(CompilerPhaseTypeHelper::to_name(compile_phase));
5912 }
5913 #endif
5914 C->_latest_stage_start_counter.stamp();
5915 }
5916
5917 // Only used from CompileWrapper
5918 void Compile::begin_method() {
5919 #ifndef PRODUCT
5920 if (_method != nullptr && should_print_igv(1)) {
5921 _igv_printer->begin_method();
5922 }
5923 #endif
5924 C->_latest_stage_start_counter.stamp();
5925 }
5926
5927 // Only used from CompileWrapper
5928 void Compile::end_method() {
5929 EventCompilerPhase event(UNTIMED);
5930 if (event.should_commit()) {
5931 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5932 }
5933
5934 #ifndef PRODUCT
5935 if (_method != nullptr && should_print_igv(1)) {
5936 _igv_printer->end_method();
5937 }
5938 #endif
5939 }
5940
5941 #ifndef PRODUCT
5942 bool Compile::should_print_phase(const int level) const {
5943 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5944 _method != nullptr; // Do not print phases for stubs.
5945 }
5946
5947 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5948 return _directive->should_print_ideal_phase(cpt);
5949 }
5950
5951 void Compile::init_igv() {
5952 if (_igv_printer == nullptr) {
5953 _igv_printer = IdealGraphPrinter::printer();
5954 _igv_printer->set_compile(this);
5955 }
5956 }
5957
5958 bool Compile::should_print_igv(const int level) {
5959 PRODUCT_RETURN_(return false;);
5960
5961 if (PrintIdealGraphLevel < 0) { // disabled by the user
5962 return false;
5963 }
5964
5965 bool need = directive()->IGVPrintLevelOption >= level;
5966 if (need) {
5967 Compile::init_igv();
5968 }
5969 return need;
5970 }
5971
5972 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5973 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5974
5975 // Called from debugger. Prints method to the default file with the default phase name.
5976 // This works regardless of any Ideal Graph Visualizer flags set or not.
5977 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5978 void igv_print(void* sp, void* fp, void* pc) {
5979 frame fr(sp, fp, pc);
5980 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5981 }
5982
5983 // Same as igv_print() above but with a specified phase name.
5984 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5985 frame fr(sp, fp, pc);
5986 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5987 }
5988
5989 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5990 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5991 // This works regardless of any Ideal Graph Visualizer flags set or not.
5992 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5993 void igv_print(bool network, void* sp, void* fp, void* pc) {
5994 frame fr(sp, fp, pc);
5995 if (network) {
5996 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5997 } else {
5998 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5999 }
6000 }
6001
6002 // Same as igv_print(bool network, ...) above but with a specified phase name.
6003 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
6004 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
6005 frame fr(sp, fp, pc);
6006 if (network) {
6007 Compile::current()->igv_print_method_to_network(phase_name, &fr);
6008 } else {
6009 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
6010 }
6011 }
6012
6013 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
6014 void igv_print_default() {
6015 Compile::current()->print_method(PHASE_DEBUG, 0);
6016 }
6017
6018 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
6019 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
6020 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
6021 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
6022 void igv_append(void* sp, void* fp, void* pc) {
6023 frame fr(sp, fp, pc);
6024 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
6025 }
6026
6027 // Same as igv_append(...) above but with a specified phase name.
6028 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
6029 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
6030 frame fr(sp, fp, pc);
6031 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
6032 }
6033
6034 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
6035 const char* file_name = "custom_debug.xml";
6036 if (_debug_file_printer == nullptr) {
6037 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
6038 } else {
6039 _debug_file_printer->update_compiled_method(C->method());
6040 }
6041 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
6042 _debug_file_printer->print_graph(phase_name, fr);
6043 }
6044
6045 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
6046 ResourceMark rm;
6047 GrowableArray<const Node*> empty_list;
6048 igv_print_graph_to_network(phase_name, empty_list, fr);
6049 }
6050
6051 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
6052 if (_debug_network_printer == nullptr) {
6053 _debug_network_printer = new IdealGraphPrinter(C);
6054 } else {
6055 _debug_network_printer->update_compiled_method(C->method());
6056 }
6057 tty->print_cr("Method printed over network stream to IGV");
6058 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
6059 }
6060 #endif // !PRODUCT
6061
6062 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
6063 if (type != nullptr && phase->type(value)->higher_equal(type)) {
6064 return value;
6065 }
6066 Node* result = nullptr;
6067 if (bt == T_BYTE) {
6068 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
6069 result = new RShiftINode(result, phase->intcon(24));
6070 } else if (bt == T_BOOLEAN) {
6071 result = new AndINode(value, phase->intcon(0xFF));
6072 } else if (bt == T_CHAR) {
6073 result = new AndINode(value,phase->intcon(0xFFFF));
6074 } else if (bt == T_FLOAT) {
6075 result = new MoveI2FNode(value);
6076 } else {
6077 assert(bt == T_SHORT, "unexpected narrow type");
6078 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
6079 result = new RShiftINode(result, phase->intcon(16));
6080 }
6081 if (transform_res) {
6082 result = phase->transform(result);
6083 }
6084 return result;
6085 }
6086
6087 void Compile::record_method_not_compilable_oom() {
6088 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
6089 }
6090
6091 #ifndef PRODUCT
6092 // Collects all the control inputs from nodes on the worklist and from their data dependencies
6093 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
6094 // Follow non-control edges until we reach CFG nodes
6095 for (uint i = 0; i < worklist.size(); i++) {
6096 const Node* n = worklist.at(i);
6097 for (uint j = 0; j < n->req(); j++) {
6098 Node* in = n->in(j);
6099 if (in == nullptr || in->is_Root()) {
6100 continue;
6101 }
6102 if (in->is_CFG()) {
6103 if (in->is_Call()) {
6104 // The return value of a call is only available if the call did not result in an exception
6105 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
6106 if (control_proj_use->is_Catch()) {
6107 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
6108 candidates.push(fall_through);
6109 continue;
6110 }
6111 }
6112
6113 if (in->is_Multi()) {
6114 // We got here by following data inputs so we should only have one control use
6115 // (no IfNode, etc)
6116 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
6117 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
6118 } else {
6119 candidates.push(in);
6120 }
6121 } else {
6122 worklist.push(in);
6123 }
6124 }
6125 }
6126 }
6127
6128 // Returns the candidate node that is a descendant to all the other candidates
6129 static Node* pick_control(Unique_Node_List& candidates) {
6130 Unique_Node_List worklist;
6131 worklist.copy(candidates);
6132
6133 // Traverse backwards through the CFG
6134 for (uint i = 0; i < worklist.size(); i++) {
6135 const Node* n = worklist.at(i);
6136 if (n->is_Root()) {
6137 continue;
6138 }
6139 for (uint j = 0; j < n->req(); j++) {
6140 // Skip backedge of loops to avoid cycles
6141 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
6142 continue;
6143 }
6144
6145 Node* pred = n->in(j);
6146 if (pred != nullptr && pred != n && pred->is_CFG()) {
6147 worklist.push(pred);
6148 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
6149 candidates.remove(pred);
6150 }
6151 }
6152 }
6153
6154 assert(candidates.size() == 1, "unexpected control flow");
6155 return candidates.at(0);
6156 }
6157
6158 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
6159 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
6160 call->init_req((*pos)++, parm);
6161 const BasicType bt = parm->bottom_type()->basic_type();
6162 if (bt == T_LONG || bt == T_DOUBLE) {
6163 call->init_req((*pos)++, half);
6164 }
6165 }
6166
6167 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
6168 Node* parm0, Node* parm1,
6169 Node* parm2, Node* parm3,
6170 Node* parm4, Node* parm5,
6171 Node* parm6) const {
6172 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
6173 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
6174 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
6175
6176 // find the most suitable control input
6177 Unique_Node_List worklist, candidates;
6178 if (parm0 != nullptr) { worklist.push(parm0);
6179 if (parm1 != nullptr) { worklist.push(parm1);
6180 if (parm2 != nullptr) { worklist.push(parm2);
6181 if (parm3 != nullptr) { worklist.push(parm3);
6182 if (parm4 != nullptr) { worklist.push(parm4);
6183 if (parm5 != nullptr) { worklist.push(parm5);
6184 if (parm6 != nullptr) { worklist.push(parm6);
6185 /* close each nested if ===> */ } } } } } } }
6186 find_candidate_control_inputs(worklist, candidates);
6187 Node* control = nullptr;
6188 if (candidates.size() == 0) {
6189 control = C->start()->proj_out(TypeFunc::Control);
6190 } else {
6191 control = pick_control(candidates);
6192 }
6193
6194 // find all the previous users of the control we picked
6195 GrowableArray<Node*> users_of_control;
6196 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
6197 Node* use = control->fast_out(i);
6198 if (use->is_CFG() && use != control) {
6199 users_of_control.push(use);
6200 }
6201 }
6202
6203 // we do not actually care about IO and memory as it uses neither
6204 call->init_req(TypeFunc::Control, control);
6205 call->init_req(TypeFunc::I_O, top());
6206 call->init_req(TypeFunc::Memory, top());
6207 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
6208 call->init_req(TypeFunc::ReturnAdr, top());
6209
6210 int pos = TypeFunc::Parms;
6211 call->init_req(pos++, str_node);
6212 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
6213 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
6214 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
6215 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
6216 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
6217 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
6218 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
6219 /* close each nested if ===> */ } } } } } } }
6220 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
6221
6222 call = gvn->transform(call);
6223 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
6224
6225 // rewire previous users to have the new call as control instead
6226 PhaseIterGVN* igvn = gvn->is_IterGVN();
6227 for (int i = 0; i < users_of_control.length(); i++) {
6228 Node* use = users_of_control.at(i);
6229 for (uint j = 0; j < use->req(); j++) {
6230 if (use->in(j) == control) {
6231 if (igvn != nullptr) {
6232 igvn->replace_input_of(use, j, call_control_proj);
6233 } else {
6234 gvn->hash_delete(use);
6235 use->set_req(j, call_control_proj);
6236 gvn->hash_insert(use);
6237 }
6238 }
6239 }
6240 }
6241
6242 return call;
6243 }
6244 #endif // !PRODUCT