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/rootnode.hpp"
83 #include "opto/runtime.hpp"
84 #include "opto/stringopts.hpp"
85 #include "opto/type.hpp"
86 #include "opto/vector.hpp"
87 #include "opto/vectornode.hpp"
88 #include "runtime/arguments.hpp"
89 #include "runtime/globals_extension.hpp"
90 #include "runtime/sharedRuntime.hpp"
91 #include "runtime/signature.hpp"
92 #include "runtime/stubRoutines.hpp"
93 #include "runtime/timer.hpp"
94 #include "utilities/align.hpp"
95 #include "utilities/copy.hpp"
96 #include "utilities/hashTable.hpp"
97 #include "utilities/macros.hpp"
98
99 // -------------------- Compile::mach_constant_base_node -----------------------
100 // Constant table base node singleton.
101 MachConstantBaseNode* Compile::mach_constant_base_node() {
102 if (_mach_constant_base_node == nullptr) {
103 _mach_constant_base_node = new MachConstantBaseNode();
104 _mach_constant_base_node->add_req(C->root());
105 }
106 return _mach_constant_base_node;
107 }
108
109
110 /// Support for intrinsics.
111
112 // Return the index at which m must be inserted (or already exists).
113 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
114 class IntrinsicDescPair {
115 private:
116 ciMethod* _m;
117 bool _is_virtual;
118 public:
119 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
120 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
121 ciMethod* m= elt->method();
122 ciMethod* key_m = key->_m;
123 if (key_m < m) return -1;
124 else if (key_m > m) return 1;
125 else {
126 bool is_virtual = elt->is_virtual();
127 bool key_virtual = key->_is_virtual;
128 if (key_virtual < is_virtual) return -1;
129 else if (key_virtual > is_virtual) return 1;
130 else return 0;
131 }
132 }
133 };
134 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
135 #ifdef ASSERT
136 for (int i = 1; i < _intrinsics.length(); i++) {
137 CallGenerator* cg1 = _intrinsics.at(i-1);
138 CallGenerator* cg2 = _intrinsics.at(i);
139 assert(cg1->method() != cg2->method()
140 ? cg1->method() < cg2->method()
141 : cg1->is_virtual() < cg2->is_virtual(),
142 "compiler intrinsics list must stay sorted");
143 }
144 #endif
145 IntrinsicDescPair pair(m, is_virtual);
146 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
147 }
148
149 void Compile::register_intrinsic(CallGenerator* cg) {
150 bool found = false;
151 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
152 assert(!found, "registering twice");
153 _intrinsics.insert_before(index, cg);
154 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
155 }
156
157 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
158 assert(m->is_loaded(), "don't try this on unloaded methods");
159 if (_intrinsics.length() > 0) {
160 bool found = false;
161 int index = intrinsic_insertion_index(m, is_virtual, found);
162 if (found) {
163 return _intrinsics.at(index);
164 }
165 }
166 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
167 if (m->intrinsic_id() != vmIntrinsics::_none &&
168 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
169 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
170 if (cg != nullptr) {
171 // Save it for next time:
172 register_intrinsic(cg);
173 return cg;
174 } else {
175 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
176 }
177 }
178 return nullptr;
179 }
180
181 // Compile::make_vm_intrinsic is defined in library_call.cpp.
182
183 #ifndef PRODUCT
184 // statistics gathering...
185
186 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
187 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
188
189 inline int as_int(vmIntrinsics::ID id) {
190 return vmIntrinsics::as_int(id);
191 }
192
193 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
194 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
195 int oflags = _intrinsic_hist_flags[as_int(id)];
196 assert(flags != 0, "what happened?");
197 if (is_virtual) {
198 flags |= _intrinsic_virtual;
199 }
200 bool changed = (flags != oflags);
201 if ((flags & _intrinsic_worked) != 0) {
202 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
203 if (count == 1) {
204 changed = true; // first time
205 }
206 // increment the overall count also:
207 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
208 }
209 if (changed) {
210 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
211 // Something changed about the intrinsic's virtuality.
212 if ((flags & _intrinsic_virtual) != 0) {
213 // This is the first use of this intrinsic as a virtual call.
214 if (oflags != 0) {
215 // We already saw it as a non-virtual, so note both cases.
216 flags |= _intrinsic_both;
217 }
218 } else if ((oflags & _intrinsic_both) == 0) {
219 // This is the first use of this intrinsic as a non-virtual
220 flags |= _intrinsic_both;
221 }
222 }
223 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
224 }
225 // update the overall flags also:
226 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
227 return changed;
228 }
229
230 static char* format_flags(int flags, char* buf) {
231 buf[0] = 0;
232 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
233 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
234 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
235 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
236 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
237 if (buf[0] == 0) strcat(buf, ",");
238 assert(buf[0] == ',', "must be");
239 return &buf[1];
240 }
241
242 void Compile::print_intrinsic_statistics() {
243 char flagsbuf[100];
244 ttyLocker ttyl;
245 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
246 tty->print_cr("Compiler intrinsic usage:");
247 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
248 if (total == 0) total = 1; // avoid div0 in case of no successes
249 #define PRINT_STAT_LINE(name, c, f) \
250 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
251 for (auto id : EnumRange<vmIntrinsicID>{}) {
252 int flags = _intrinsic_hist_flags[as_int(id)];
253 juint count = _intrinsic_hist_count[as_int(id)];
254 if ((flags | count) != 0) {
255 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
256 }
257 }
258 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
259 if (xtty != nullptr) xtty->tail("statistics");
260 }
261
262 void Compile::print_statistics() {
263 { ttyLocker ttyl;
264 if (xtty != nullptr) xtty->head("statistics type='opto'");
265 Parse::print_statistics();
266 PhaseStringOpts::print_statistics();
267 PhaseCCP::print_statistics();
268 PhaseRegAlloc::print_statistics();
269 PhaseOutput::print_statistics();
270 PhasePeephole::print_statistics();
271 PhaseIdealLoop::print_statistics();
272 ConnectionGraph::print_statistics();
273 PhaseMacroExpand::print_statistics();
274 if (xtty != nullptr) xtty->tail("statistics");
275 }
276 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
277 // put this under its own <statistics> element.
278 print_intrinsic_statistics();
279 }
280 }
281 #endif //PRODUCT
282
283 void Compile::gvn_replace_by(Node* n, Node* nn) {
284 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
285 Node* use = n->last_out(i);
286 bool is_in_table = initial_gvn()->hash_delete(use);
287 uint uses_found = 0;
288 for (uint j = 0; j < use->len(); j++) {
289 if (use->in(j) == n) {
290 if (j < use->req())
291 use->set_req(j, nn);
292 else
293 use->set_prec(j, nn);
294 uses_found++;
295 }
296 }
297 if (is_in_table) {
298 // reinsert into table
299 initial_gvn()->hash_find_insert(use);
300 }
301 record_for_igvn(use);
302 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
303 i -= uses_found; // we deleted 1 or more copies of this edge
304 }
305 }
306
307
308 // Identify all nodes that are reachable from below, useful.
309 // Use breadth-first pass that records state in a Unique_Node_List,
310 // recursive traversal is slower.
311 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
312 int estimated_worklist_size = live_nodes();
313 useful.map( estimated_worklist_size, nullptr ); // preallocate space
314
315 // Initialize worklist
316 if (root() != nullptr) { useful.push(root()); }
317 // If 'top' is cached, declare it useful to preserve cached node
318 if (cached_top_node()) { useful.push(cached_top_node()); }
319
320 // Push all useful nodes onto the list, breadthfirst
321 for( uint next = 0; next < useful.size(); ++next ) {
322 assert( next < unique(), "Unique useful nodes < total nodes");
323 Node *n = useful.at(next);
324 uint max = n->len();
325 for( uint i = 0; i < max; ++i ) {
326 Node *m = n->in(i);
327 if (not_a_node(m)) continue;
328 useful.push(m);
329 }
330 }
331 }
332
333 // Update dead_node_list with any missing dead nodes using useful
334 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
335 void Compile::update_dead_node_list(Unique_Node_List &useful) {
336 uint max_idx = unique();
337 VectorSet& useful_node_set = useful.member_set();
338
339 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
340 // If node with index node_idx is not in useful set,
341 // mark it as dead in dead node list.
342 if (!useful_node_set.test(node_idx)) {
343 record_dead_node(node_idx);
344 }
345 }
346 }
347
348 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
349 int shift = 0;
350 for (int i = 0; i < inlines->length(); i++) {
351 CallGenerator* cg = inlines->at(i);
352 if (useful.member(cg->call_node())) {
353 if (shift > 0) {
354 inlines->at_put(i - shift, cg);
355 }
356 } else {
357 shift++; // skip over the dead element
358 }
359 }
360 if (shift > 0) {
361 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
362 }
363 }
364
365 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
366 assert(dead != nullptr && dead->is_Call(), "sanity");
367 int found = 0;
368 for (int i = 0; i < inlines->length(); i++) {
369 if (inlines->at(i)->call_node() == dead) {
370 inlines->remove_at(i);
371 found++;
372 NOT_DEBUG( break; ) // elements are unique, so exit early
373 }
374 }
375 assert(found <= 1, "not unique");
376 }
377
378 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
379 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
380 for (int i = node_list.length() - 1; i >= 0; i--) {
381 N* node = node_list.at(i);
382 if (!useful.member(node)) {
383 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
384 }
385 }
386 }
387
388 void Compile::remove_useless_node(Node* dead) {
389 remove_modified_node(dead);
390
391 // Constant node that has no out-edges and has only one in-edge from
392 // root is usually dead. However, sometimes reshaping walk makes
393 // it reachable by adding use edges. So, we will NOT count Con nodes
394 // as dead to be conservative about the dead node count at any
395 // given time.
396 if (!dead->is_Con()) {
397 record_dead_node(dead->_idx);
398 }
399 if (dead->is_macro()) {
400 remove_macro_node(dead);
401 }
402 if (dead->is_expensive()) {
403 remove_expensive_node(dead);
404 }
405 if (dead->is_OpaqueTemplateAssertionPredicate()) {
406 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
407 }
408 if (dead->is_ParsePredicate()) {
409 remove_parse_predicate(dead->as_ParsePredicate());
410 }
411 if (dead->for_post_loop_opts_igvn()) {
412 remove_from_post_loop_opts_igvn(dead);
413 }
414 if (dead->is_InlineType()) {
415 remove_inline_type(dead);
416 }
417 if (dead->is_LoadFlat() || dead->is_StoreFlat()) {
418 remove_flat_access(dead);
419 }
420 if (dead->for_merge_stores_igvn()) {
421 remove_from_merge_stores_igvn(dead);
422 }
423 if (dead->is_Call()) {
424 remove_useless_late_inlines( &_late_inlines, dead);
425 remove_useless_late_inlines( &_string_late_inlines, dead);
426 remove_useless_late_inlines( &_boxing_late_inlines, dead);
427 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
428
429 if (dead->is_CallStaticJava()) {
430 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
431 }
432 }
433 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
434 bs->unregister_potential_barrier_node(dead);
435 }
436
437 // Disconnect all useless nodes by disconnecting those at the boundary.
438 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
439 uint next = 0;
440 while (next < useful.size()) {
441 Node *n = useful.at(next++);
442 if (n->is_SafePoint()) {
443 // We're done with a parsing phase. Replaced nodes are not valid
444 // beyond that point.
445 n->as_SafePoint()->delete_replaced_nodes();
446 }
447 // Use raw traversal of out edges since this code removes out edges
448 int max = n->outcnt();
449 for (int j = 0; j < max; ++j) {
450 Node* child = n->raw_out(j);
451 if (!useful.member(child)) {
452 assert(!child->is_top() || child != top(),
453 "If top is cached in Compile object it is in useful list");
454 // Only need to remove this out-edge to the useless node
455 n->raw_del_out(j);
456 --j;
457 --max;
458 if (child->is_data_proj_of_pure_function(n)) {
459 worklist.push(n);
460 }
461 }
462 }
463 if (n->outcnt() == 1 && n->has_special_unique_user()) {
464 assert(useful.member(n->unique_out()), "do not push a useless node");
465 worklist.push(n->unique_out());
466 }
467 if (n->outcnt() == 0) {
468 worklist.push(n);
469 }
470 }
471
472 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
473 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
474 // Remove useless Template Assertion Predicate opaque nodes
475 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
476 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
477 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
478 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
479 remove_useless_nodes(_flat_access_nodes, useful); // remove useless flat access nodes
480 #ifdef ASSERT
481 if (_modified_nodes != nullptr) {
482 _modified_nodes->remove_useless_nodes(useful.member_set());
483 }
484 #endif
485 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
486 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
487 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
488 #ifdef ASSERT
489 if (_modified_nodes != nullptr) {
490 _modified_nodes->remove_useless_nodes(useful.member_set());
491 }
492 #endif
493
494 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
495 bs->eliminate_useless_gc_barriers(useful, this);
496 // clean up the late inline lists
497 remove_useless_late_inlines( &_late_inlines, useful);
498 remove_useless_late_inlines( &_string_late_inlines, useful);
499 remove_useless_late_inlines( &_boxing_late_inlines, useful);
500 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
501 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
502 }
503
504 // ============================================================================
505 //------------------------------CompileWrapper---------------------------------
506 class CompileWrapper : public StackObj {
507 Compile *const _compile;
508 public:
509 CompileWrapper(Compile* compile);
510
511 ~CompileWrapper();
512 };
513
514 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
515 // the Compile* pointer is stored in the current ciEnv:
516 ciEnv* env = compile->env();
517 assert(env == ciEnv::current(), "must already be a ciEnv active");
518 assert(env->compiler_data() == nullptr, "compile already active?");
519 env->set_compiler_data(compile);
520 assert(compile == Compile::current(), "sanity");
521
522 compile->set_type_dict(nullptr);
523 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
524 compile->clone_map().set_clone_idx(0);
525 compile->set_type_last_size(0);
526 compile->set_last_tf(nullptr, nullptr);
527 compile->set_indexSet_arena(nullptr);
528 compile->set_indexSet_free_block_list(nullptr);
529 compile->init_type_arena();
530 Type::Initialize(compile);
531 _compile->begin_method();
532 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
533 }
534 CompileWrapper::~CompileWrapper() {
535 // simulate crash during compilation
536 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
537
538 _compile->end_method();
539 _compile->env()->set_compiler_data(nullptr);
540 }
541
542
543 //----------------------------print_compile_messages---------------------------
544 void Compile::print_compile_messages() {
545 #ifndef PRODUCT
546 // Check if recompiling
547 if (!subsume_loads() && PrintOpto) {
548 // Recompiling without allowing machine instructions to subsume loads
549 tty->print_cr("*********************************************************");
550 tty->print_cr("** Bailout: Recompile without subsuming loads **");
551 tty->print_cr("*********************************************************");
552 }
553 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
554 // Recompiling without escape analysis
555 tty->print_cr("*********************************************************");
556 tty->print_cr("** Bailout: Recompile without escape analysis **");
557 tty->print_cr("*********************************************************");
558 }
559 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
560 // Recompiling without iterative escape analysis
561 tty->print_cr("*********************************************************");
562 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
563 tty->print_cr("*********************************************************");
564 }
565 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
566 // Recompiling without reducing allocation merges
567 tty->print_cr("*********************************************************");
568 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
569 tty->print_cr("*********************************************************");
570 }
571 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
572 // Recompiling without boxing elimination
573 tty->print_cr("*********************************************************");
574 tty->print_cr("** Bailout: Recompile without boxing elimination **");
575 tty->print_cr("*********************************************************");
576 }
577 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
578 // Recompiling without locks coarsening
579 tty->print_cr("*********************************************************");
580 tty->print_cr("** Bailout: Recompile without locks coarsening **");
581 tty->print_cr("*********************************************************");
582 }
583 if (env()->break_at_compile()) {
584 // Open the debugger when compiling this method.
585 tty->print("### Breaking when compiling: ");
586 method()->print_short_name();
587 tty->cr();
588 BREAKPOINT;
589 }
590
591 if( PrintOpto ) {
592 if (is_osr_compilation()) {
593 tty->print("[OSR]%3d", _compile_id);
594 } else {
595 tty->print("%3d", _compile_id);
596 }
597 }
598 #endif
599 }
600
601 #ifndef PRODUCT
602 void Compile::print_phase(const char* phase_name) {
603 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name);
604 }
605
606 void Compile::print_ideal_ir(const char* compile_phase_name) const {
607 // keep the following output all in one block
608 // This output goes directly to the tty, not the compiler log.
609 // To enable tools to match it up with the compilation activity,
610 // be sure to tag this tty output with the compile ID.
611
612 // Node dumping can cause a safepoint, which can break the tty lock.
613 // Buffer all node dumps, so that all safepoints happen before we lock.
614 ResourceMark rm;
615 stringStream ss;
616
617 if (_output == nullptr) {
618 ss.print_cr("AFTER: %s", compile_phase_name);
619 // Print out all nodes in ascending order of index.
620 // It is important that we traverse both inputs and outputs of nodes,
621 // so that we reach all nodes that are connected to Root.
622 root()->dump_bfs(MaxNodeLimit, nullptr, "-+S$", &ss);
623 } else {
624 // Dump the node blockwise if we have a scheduling
625 _output->print_scheduling(&ss);
626 }
627
628 // Check that the lock is not broken by a safepoint.
629 NoSafepointVerifier nsv;
630 ttyLocker ttyl;
631 if (xtty != nullptr) {
632 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
633 compile_id(),
634 is_osr_compilation() ? " compile_kind='osr'" : "",
635 compile_phase_name);
636 }
637
638 tty->print("%s", ss.as_string());
639
640 if (xtty != nullptr) {
641 xtty->tail("ideal");
642 }
643 }
644 #endif
645
646 // ============================================================================
647 //------------------------------Compile standard-------------------------------
648
649 // Compile a method. entry_bci is -1 for normal compilations and indicates
650 // the continuation bci for on stack replacement.
651
652
653 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
654 Options options, DirectiveSet* directive)
655 : Phase(Compiler),
656 _compile_id(ci_env->compile_id()),
657 _options(options),
658 _method(target),
659 _entry_bci(osr_bci),
660 _ilt(nullptr),
661 _stub_function(nullptr),
662 _stub_name(nullptr),
663 _stub_id(StubId::NO_STUBID),
664 _stub_entry_point(nullptr),
665 _max_node_limit(MaxNodeLimit),
666 _post_loop_opts_phase(false),
667 _merge_stores_phase(false),
668 _allow_macro_nodes(true),
669 _inlining_progress(false),
670 _inlining_incrementally(false),
671 _strength_reduction(false),
672 _do_cleanup(false),
673 _has_reserved_stack_access(target->has_reserved_stack_access()),
674 _has_circular_inline_type(false),
675 #ifndef PRODUCT
676 _igv_idx(0),
677 _trace_opto_output(directive->TraceOptoOutputOption),
678 #endif
679 _clinit_barrier_on_entry(false),
680 _stress_seed(0),
681 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
682 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
683 _env(ci_env),
684 _directive(directive),
685 _log(ci_env->log()),
686 _first_failure_details(nullptr),
687 _intrinsics(comp_arena(), 0, 0, nullptr),
688 _macro_nodes(comp_arena(), 8, 0, nullptr),
689 _parse_predicates(comp_arena(), 8, 0, nullptr),
690 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
691 _expensive_nodes(comp_arena(), 8, 0, nullptr),
692 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
693 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
694 _flat_access_nodes(comp_arena(), 8, 0, nullptr),
695 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
696 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
697 _coarsened_locks(comp_arena(), 8, 0, nullptr),
698 _congraph(nullptr),
699 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
700 _unique(0),
701 _dead_node_count(0),
702 _dead_node_list(comp_arena()),
703 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
704 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
705 _node_arena(&_node_arena_one),
706 _mach_constant_base_node(nullptr),
707 _Compile_types(mtCompiler, Arena::Tag::tag_type),
708 _initial_gvn(nullptr),
709 _igvn_worklist(nullptr),
710 _types(nullptr),
711 _node_hash(nullptr),
712 _late_inlines(comp_arena(), 2, 0, nullptr),
713 _string_late_inlines(comp_arena(), 2, 0, nullptr),
714 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
715 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
716 _late_inlines_pos(0),
717 _has_mh_late_inlines(false),
718 _oom(false),
719 _replay_inline_data(nullptr),
720 _inline_printer(this),
721 _java_calls(0),
722 _inner_loops(0),
723 _FIRST_STACK_mask(comp_arena()),
724 _interpreter_frame_size(0),
725 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
726 _output(nullptr)
727 #ifndef PRODUCT
728 ,
729 _in_dump_cnt(0)
730 #endif
731 {
732 C = this;
733 CompileWrapper cw(this);
734
735 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
736 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
737
738 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
739 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
740 // We can always print a disassembly, either abstract (hex dump) or
741 // with the help of a suitable hsdis library. Thus, we should not
742 // couple print_assembly and print_opto_assembly controls.
743 // But: always print opto and regular assembly on compile command 'print'.
744 bool print_assembly = directive->PrintAssemblyOption;
745 set_print_assembly(print_opto_assembly || print_assembly);
746 #else
747 set_print_assembly(false); // must initialize.
748 #endif
749
750 #ifndef PRODUCT
751 set_parsed_irreducible_loop(false);
752 #endif
753
754 if (directive->ReplayInlineOption) {
755 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
756 }
757 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
758 set_print_intrinsics(directive->PrintIntrinsicsOption);
759 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
760
761 if (ProfileTraps) {
762 // Make sure the method being compiled gets its own MDO,
763 // so we can at least track the decompile_count().
764 method()->ensure_method_data();
765 }
766
767 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
768 StressIncrementalInlining || StressMacroExpansion ||
769 StressMacroElimination || StressUnstableIfTraps ||
770 StressBailout || StressLoopPeeling) {
771 initialize_stress_seed(directive);
772 }
773
774 Init(/*do_aliasing=*/ true);
775
776 print_compile_messages();
777
778 _ilt = InlineTree::build_inline_tree_root();
779
780 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
781 assert(num_alias_types() >= AliasIdxRaw, "");
782
783 #define MINIMUM_NODE_HASH 1023
784
785 // GVN that will be run immediately on new nodes
786 uint estimated_size = method()->code_size()*4+64;
787 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
788 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
789 _types = new (comp_arena()) Type_Array(comp_arena());
790 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
791 PhaseGVN gvn;
792 set_initial_gvn(&gvn);
793
794 { // Scope for timing the parser
795 TracePhase tp(_t_parser);
796
797 // Put top into the hash table ASAP.
798 initial_gvn()->transform(top());
799
800 // Set up tf(), start(), and find a CallGenerator.
801 CallGenerator* cg = nullptr;
802 if (is_osr_compilation()) {
803 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
804 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
805 initial_gvn()->set_type_bottom(s);
806 verify_start(s);
807 cg = CallGenerator::for_osr(method(), entry_bci());
808 } else {
809 // Normal case.
810 init_tf(TypeFunc::make(method()));
811 StartNode* s = new StartNode(root(), tf()->domain_cc());
812 initial_gvn()->set_type_bottom(s);
813 verify_start(s);
814 float past_uses = method()->interpreter_invocation_count();
815 float expected_uses = past_uses;
816 cg = CallGenerator::for_inline(method(), expected_uses);
817 }
818 if (failing()) return;
819 if (cg == nullptr) {
820 const char* reason = InlineTree::check_can_parse(method());
821 assert(reason != nullptr, "expect reason for parse failure");
822 stringStream ss;
823 ss.print("cannot parse method: %s", reason);
824 record_method_not_compilable(ss.as_string());
825 return;
826 }
827
828 gvn.set_type(root(), root()->bottom_type());
829
830 JVMState* jvms = build_start_state(start(), tf());
831 if ((jvms = cg->generate(jvms)) == nullptr) {
832 assert(failure_reason() != nullptr, "expect reason for parse failure");
833 stringStream ss;
834 ss.print("method parse failed: %s", failure_reason());
835 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
836 return;
837 }
838 GraphKit kit(jvms);
839
840 if (!kit.stopped()) {
841 // Accept return values, and transfer control we know not where.
842 // This is done by a special, unique ReturnNode bound to root.
843 return_values(kit.jvms());
844 }
845
846 if (kit.has_exceptions()) {
847 // Any exceptions that escape from this call must be rethrown
848 // to whatever caller is dynamically above us on the stack.
849 // This is done by a special, unique RethrowNode bound to root.
850 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
851 }
852
853 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
854
855 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
856 inline_string_calls(true);
857 }
858
859 if (failing()) return;
860
861 // Remove clutter produced by parsing.
862 if (!failing()) {
863 ResourceMark rm;
864 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
865 }
866 }
867
868 // Note: Large methods are capped off in do_one_bytecode().
869 if (failing()) return;
870
871 // After parsing, node notes are no longer automagic.
872 // They must be propagated by register_new_node_with_optimizer(),
873 // clone(), or the like.
874 set_default_node_notes(nullptr);
875
876 #ifndef PRODUCT
877 if (should_print_igv(1)) {
878 _igv_printer->print_inlining();
879 }
880 #endif
881
882 if (failing()) return;
883 NOT_PRODUCT( verify_graph_edges(); )
884
885 // Now optimize
886 Optimize();
887 if (failing()) return;
888 NOT_PRODUCT( verify_graph_edges(); )
889
890 #ifndef PRODUCT
891 if (should_print_ideal()) {
892 print_ideal_ir("PrintIdeal");
893 }
894 #endif
895
896 #ifdef ASSERT
897 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
898 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
899 #endif
900
901 // Dump compilation data to replay it.
902 if (directive->DumpReplayOption) {
903 env()->dump_replay_data(_compile_id);
904 }
905 if (directive->DumpInlineOption && (ilt() != nullptr)) {
906 env()->dump_inline_data(_compile_id);
907 }
908
909 // Now that we know the size of all the monitors we can add a fixed slot
910 // for the original deopt pc.
911 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
912 if (needs_stack_repair()) {
913 // One extra slot for the special stack increment value
914 next_slot += 2;
915 }
916 // TODO 8284443 Only reserve extra slot if needed
917 if (InlineTypeReturnedAsFields) {
918 // One extra slot to hold the null marker for a nullable
919 // inline type return if we run out of registers.
920 next_slot += 2;
921 }
922 set_fixed_slots(next_slot);
923
924 // Compute when to use implicit null checks. Used by matching trap based
925 // nodes and NullCheck optimization.
926 set_allowed_deopt_reasons();
927
928 // Now generate code
929 Code_Gen();
930 }
931
932 //------------------------------Compile----------------------------------------
933 // Compile a runtime stub
934 Compile::Compile(ciEnv* ci_env,
935 TypeFunc_generator generator,
936 address stub_function,
937 const char* stub_name,
938 StubId stub_id,
939 int is_fancy_jump,
940 bool pass_tls,
941 bool return_pc,
942 DirectiveSet* directive)
943 : Phase(Compiler),
944 _compile_id(0),
945 _options(Options::for_runtime_stub()),
946 _method(nullptr),
947 _entry_bci(InvocationEntryBci),
948 _stub_function(stub_function),
949 _stub_name(stub_name),
950 _stub_id(stub_id),
951 _stub_entry_point(nullptr),
952 _max_node_limit(MaxNodeLimit),
953 _post_loop_opts_phase(false),
954 _merge_stores_phase(false),
955 _allow_macro_nodes(true),
956 _inlining_progress(false),
957 _inlining_incrementally(false),
958 _has_reserved_stack_access(false),
959 _has_circular_inline_type(false),
960 #ifndef PRODUCT
961 _igv_idx(0),
962 _trace_opto_output(directive->TraceOptoOutputOption),
963 #endif
964 _clinit_barrier_on_entry(false),
965 _stress_seed(0),
966 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
967 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
968 _env(ci_env),
969 _directive(directive),
970 _log(ci_env->log()),
971 _first_failure_details(nullptr),
972 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
973 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
974 _congraph(nullptr),
975 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
976 _unique(0),
977 _dead_node_count(0),
978 _dead_node_list(comp_arena()),
979 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
980 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
981 _node_arena(&_node_arena_one),
982 _mach_constant_base_node(nullptr),
983 _Compile_types(mtCompiler, Arena::Tag::tag_type),
984 _initial_gvn(nullptr),
985 _igvn_worklist(nullptr),
986 _types(nullptr),
987 _node_hash(nullptr),
988 _has_mh_late_inlines(false),
989 _oom(false),
990 _replay_inline_data(nullptr),
991 _inline_printer(this),
992 _java_calls(0),
993 _inner_loops(0),
994 _FIRST_STACK_mask(comp_arena()),
995 _interpreter_frame_size(0),
996 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
997 _output(nullptr),
998 #ifndef PRODUCT
999 _in_dump_cnt(0),
1000 #endif
1001 _allowed_reasons(0) {
1002 C = this;
1003
1004 // try to reuse an existing stub
1005 {
1006 BlobId blob_id = StubInfo::blob(_stub_id);
1007 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
1008 if (blob != nullptr) {
1009 RuntimeStub* rs = blob->as_runtime_stub();
1010 _stub_entry_point = rs->entry_point();
1011 return;
1012 }
1013 }
1014
1015 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
1016 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
1017
1018 #ifndef PRODUCT
1019 set_print_assembly(PrintFrameConverterAssembly);
1020 set_parsed_irreducible_loop(false);
1021 #else
1022 set_print_assembly(false); // Must initialize.
1023 #endif
1024 set_has_irreducible_loop(false); // no loops
1025
1026 CompileWrapper cw(this);
1027 Init(/*do_aliasing=*/ false);
1028 init_tf((*generator)());
1029
1030 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1031 _types = new (comp_arena()) Type_Array(comp_arena());
1032 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1033
1034 if (StressLCM || StressGCM || StressBailout) {
1035 initialize_stress_seed(directive);
1036 }
1037
1038 {
1039 PhaseGVN gvn;
1040 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1041 gvn.transform(top());
1042
1043 GraphKit kit;
1044 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1045 }
1046
1047 NOT_PRODUCT( verify_graph_edges(); )
1048
1049 Code_Gen();
1050 }
1051
1052 Compile::~Compile() {
1053 delete _first_failure_details;
1054 };
1055
1056 //------------------------------Init-------------------------------------------
1057 // Prepare for a single compilation
1058 void Compile::Init(bool aliasing) {
1059 _do_aliasing = aliasing;
1060 _unique = 0;
1061 _regalloc = nullptr;
1062
1063 _tf = nullptr; // filled in later
1064 _top = nullptr; // cached later
1065 _matcher = nullptr; // filled in later
1066 _cfg = nullptr; // filled in later
1067
1068 _node_note_array = nullptr;
1069 _default_node_notes = nullptr;
1070 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1071
1072 _immutable_memory = nullptr; // filled in at first inquiry
1073
1074 #ifdef ASSERT
1075 _phase_optimize_finished = false;
1076 _phase_verify_ideal_loop = false;
1077 _exception_backedge = false;
1078 _type_verify = nullptr;
1079 #endif
1080
1081 // Globally visible Nodes
1082 // First set TOP to null to give safe behavior during creation of RootNode
1083 set_cached_top_node(nullptr);
1084 set_root(new RootNode());
1085 // Now that you have a Root to point to, create the real TOP
1086 set_cached_top_node( new ConNode(Type::TOP) );
1087 set_recent_alloc(nullptr, nullptr);
1088
1089 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1090 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1091 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1092 env()->set_dependencies(new Dependencies(env()));
1093
1094 _fixed_slots = 0;
1095 set_has_split_ifs(false);
1096 set_has_loops(false); // first approximation
1097 set_has_stringbuilder(false);
1098 set_has_boxed_value(false);
1099 _trap_can_recompile = false; // no traps emitted yet
1100 _major_progress = true; // start out assuming good things will happen
1101 set_has_unsafe_access(false);
1102 set_max_vector_size(0);
1103 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1104 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1105 set_decompile_count(0);
1106
1107 #ifndef PRODUCT
1108 _phase_counter = 0;
1109 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1110 #endif
1111
1112 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1113 _loop_opts_cnt = LoopOptsCount;
1114 _has_flat_accesses = false;
1115 _flat_accesses_share_alias = true;
1116 _scalarize_in_safepoints = false;
1117
1118 set_do_inlining(Inline);
1119 set_max_inline_size(MaxInlineSize);
1120 set_freq_inline_size(FreqInlineSize);
1121 set_do_scheduling(OptoScheduling);
1122
1123 set_do_vector_loop(false);
1124 set_has_monitors(false);
1125 set_has_scoped_access(false);
1126
1127 if (AllowVectorizeOnDemand) {
1128 if (has_method() && _directive->VectorizeOption) {
1129 set_do_vector_loop(true);
1130 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());})
1131 } else if (has_method() && method()->name() != nullptr &&
1132 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1133 set_do_vector_loop(true);
1134 }
1135 }
1136 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1137 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());})
1138
1139 _max_node_limit = _directive->MaxNodeLimitOption;
1140
1141 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1142 set_clinit_barrier_on_entry(true);
1143 }
1144 if (debug_info()->recording_non_safepoints()) {
1145 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1146 (comp_arena(), 8, 0, nullptr));
1147 set_default_node_notes(Node_Notes::make(this));
1148 }
1149
1150 const int grow_ats = 16;
1151 _max_alias_types = grow_ats;
1152 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1153 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1154 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1155 {
1156 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1157 }
1158 // Initialize the first few types.
1159 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1160 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1161 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1162 _num_alias_types = AliasIdxRaw+1;
1163 // Zero out the alias type cache.
1164 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1165 // A null adr_type hits in the cache right away. Preload the right answer.
1166 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1167 }
1168
1169 #ifdef ASSERT
1170 // Verify that the current StartNode is valid.
1171 void Compile::verify_start(StartNode* s) const {
1172 assert(failing_internal() || s == start(), "should be StartNode");
1173 }
1174 #endif
1175
1176 /**
1177 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1178 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1179 * the ideal graph.
1180 */
1181 StartNode* Compile::start() const {
1182 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1183 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1184 Node* start = root()->fast_out(i);
1185 if (start->is_Start()) {
1186 return start->as_Start();
1187 }
1188 }
1189 fatal("Did not find Start node!");
1190 return nullptr;
1191 }
1192
1193 //-------------------------------immutable_memory-------------------------------------
1194 // Access immutable memory
1195 Node* Compile::immutable_memory() {
1196 if (_immutable_memory != nullptr) {
1197 return _immutable_memory;
1198 }
1199 StartNode* s = start();
1200 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1201 Node *p = s->fast_out(i);
1202 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1203 _immutable_memory = p;
1204 return _immutable_memory;
1205 }
1206 }
1207 ShouldNotReachHere();
1208 return nullptr;
1209 }
1210
1211 //----------------------set_cached_top_node------------------------------------
1212 // Install the cached top node, and make sure Node::is_top works correctly.
1213 void Compile::set_cached_top_node(Node* tn) {
1214 if (tn != nullptr) verify_top(tn);
1215 Node* old_top = _top;
1216 _top = tn;
1217 // Calling Node::setup_is_top allows the nodes the chance to adjust
1218 // their _out arrays.
1219 if (_top != nullptr) _top->setup_is_top();
1220 if (old_top != nullptr) old_top->setup_is_top();
1221 assert(_top == nullptr || top()->is_top(), "");
1222 }
1223
1224 #ifdef ASSERT
1225 uint Compile::count_live_nodes_by_graph_walk() {
1226 Unique_Node_List useful(comp_arena());
1227 // Get useful node list by walking the graph.
1228 identify_useful_nodes(useful);
1229 return useful.size();
1230 }
1231
1232 void Compile::print_missing_nodes() {
1233
1234 // Return if CompileLog is null and PrintIdealNodeCount is false.
1235 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1236 return;
1237 }
1238
1239 // This is an expensive function. It is executed only when the user
1240 // specifies VerifyIdealNodeCount option or otherwise knows the
1241 // additional work that needs to be done to identify reachable nodes
1242 // by walking the flow graph and find the missing ones using
1243 // _dead_node_list.
1244
1245 Unique_Node_List useful(comp_arena());
1246 // Get useful node list by walking the graph.
1247 identify_useful_nodes(useful);
1248
1249 uint l_nodes = C->live_nodes();
1250 uint l_nodes_by_walk = useful.size();
1251
1252 if (l_nodes != l_nodes_by_walk) {
1253 if (_log != nullptr) {
1254 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1255 _log->stamp();
1256 _log->end_head();
1257 }
1258 VectorSet& useful_member_set = useful.member_set();
1259 int last_idx = l_nodes_by_walk;
1260 for (int i = 0; i < last_idx; i++) {
1261 if (useful_member_set.test(i)) {
1262 if (_dead_node_list.test(i)) {
1263 if (_log != nullptr) {
1264 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1265 }
1266 if (PrintIdealNodeCount) {
1267 // Print the log message to tty
1268 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1269 useful.at(i)->dump();
1270 }
1271 }
1272 }
1273 else if (! _dead_node_list.test(i)) {
1274 if (_log != nullptr) {
1275 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1276 }
1277 if (PrintIdealNodeCount) {
1278 // Print the log message to tty
1279 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1280 }
1281 }
1282 }
1283 if (_log != nullptr) {
1284 _log->tail("mismatched_nodes");
1285 }
1286 }
1287 }
1288 void Compile::record_modified_node(Node* n) {
1289 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1290 _modified_nodes->push(n);
1291 }
1292 }
1293
1294 void Compile::remove_modified_node(Node* n) {
1295 if (_modified_nodes != nullptr) {
1296 _modified_nodes->remove(n);
1297 }
1298 }
1299 #endif
1300
1301 #ifndef PRODUCT
1302 void Compile::verify_top(Node* tn) const {
1303 if (tn != nullptr) {
1304 assert(tn->is_Con(), "top node must be a constant");
1305 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1306 assert(tn->in(0) != nullptr, "must have live top node");
1307 }
1308 }
1309 #endif
1310
1311
1312 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1313
1314 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1315 guarantee(arr != nullptr, "");
1316 int num_blocks = arr->length();
1317 if (grow_by < num_blocks) grow_by = num_blocks;
1318 int num_notes = grow_by * _node_notes_block_size;
1319 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1320 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1321 while (num_notes > 0) {
1322 arr->append(notes);
1323 notes += _node_notes_block_size;
1324 num_notes -= _node_notes_block_size;
1325 }
1326 assert(num_notes == 0, "exact multiple, please");
1327 }
1328
1329 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1330 if (source == nullptr || dest == nullptr) return false;
1331
1332 if (dest->is_Con())
1333 return false; // Do not push debug info onto constants.
1334
1335 #ifdef ASSERT
1336 // Leave a bread crumb trail pointing to the original node:
1337 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1338 dest->set_debug_orig(source);
1339 }
1340 #endif
1341
1342 if (node_note_array() == nullptr)
1343 return false; // Not collecting any notes now.
1344
1345 // This is a copy onto a pre-existing node, which may already have notes.
1346 // If both nodes have notes, do not overwrite any pre-existing notes.
1347 Node_Notes* source_notes = node_notes_at(source->_idx);
1348 if (source_notes == nullptr || source_notes->is_clear()) return false;
1349 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1350 if (dest_notes == nullptr || dest_notes->is_clear()) {
1351 return set_node_notes_at(dest->_idx, source_notes);
1352 }
1353
1354 Node_Notes merged_notes = (*source_notes);
1355 // The order of operations here ensures that dest notes will win...
1356 merged_notes.update_from(dest_notes);
1357 return set_node_notes_at(dest->_idx, &merged_notes);
1358 }
1359
1360
1361 //--------------------------allow_range_check_smearing-------------------------
1362 // Gating condition for coalescing similar range checks.
1363 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1364 // single covering check that is at least as strong as any of them.
1365 // If the optimization succeeds, the simplified (strengthened) range check
1366 // will always succeed. If it fails, we will deopt, and then give up
1367 // on the optimization.
1368 bool Compile::allow_range_check_smearing() const {
1369 // If this method has already thrown a range-check,
1370 // assume it was because we already tried range smearing
1371 // and it failed.
1372 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1373 return !already_trapped;
1374 }
1375
1376
1377 //------------------------------flatten_alias_type-----------------------------
1378 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1379 assert(do_aliasing(), "Aliasing should be enabled");
1380 int offset = tj->offset();
1381 TypePtr::PTR ptr = tj->ptr();
1382
1383 // Known instance (scalarizable allocation) alias only with itself.
1384 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1385 tj->is_oopptr()->is_known_instance();
1386
1387 // Process weird unsafe references.
1388 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1389 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1390 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1391 tj = TypeOopPtr::BOTTOM;
1392 ptr = tj->ptr();
1393 offset = tj->offset();
1394 }
1395
1396 // Array pointers need some flattening
1397 const TypeAryPtr* ta = tj->isa_aryptr();
1398 if( ta && is_known_inst ) {
1399 if ( offset != Type::OffsetBot &&
1400 offset > arrayOopDesc::length_offset_in_bytes() ) {
1401 offset = Type::OffsetBot; // Flatten constant access into array body only
1402 tj = ta = ta->
1403 remove_speculative()->
1404 cast_to_ptr_type(ptr)->
1405 with_offset(offset);
1406 }
1407 } else if (ta) {
1408 // Common slices
1409 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1410 return TypeAryPtr::RANGE;
1411 } else if (offset == oopDesc::klass_offset_in_bytes()) {
1412 return TypeInstPtr::KLASS;
1413 } else if (offset == oopDesc::mark_offset_in_bytes()) {
1414 return TypeInstPtr::MARK;
1415 }
1416
1417 // Remove size and stability
1418 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());
1419 // Remove ptr, const_oop, and offset
1420 if (ta->elem() == Type::BOTTOM) {
1421 // Bottom array (meet of int[] and byte[] for example), accesses to it will be done with
1422 // Unsafe. This should alias with all arrays. For now just leave it as it is (this is
1423 // incorrect, see JDK-8331133).
1424 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, false, Type::Offset::bottom);
1425 } else if (ta->elem()->make_oopptr() != nullptr) {
1426 // Object arrays, keep field_offset
1427 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, ta->klass_is_exact(), Type::Offset::bottom, Type::Offset(ta->field_offset()));
1428 } else {
1429 // Primitive arrays
1430 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, ta->exact_klass(), true, Type::Offset::bottom);
1431 }
1432
1433 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1434 // cannot be distinguished by bytecode alone.
1435 if (ta->elem() == TypeInt::BOOL) {
1436 tj = ta = TypeAryPtr::BYTES;
1437 }
1438
1439 // All arrays of references share the same slice
1440 if (!ta->is_flat() && ta->elem()->make_oopptr() != nullptr) {
1441 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, TypeInt::POS, false, false, true, true, true);
1442 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, tary, nullptr, false, Type::Offset::bottom);
1443 }
1444
1445 if (ta->is_flat()) {
1446 if (_flat_accesses_share_alias) {
1447 // Initially all flattened array accesses share a single slice
1448 tj = ta = TypeAryPtr::INLINES;
1449 } else {
1450 // Flat accesses are always exact
1451 tj = ta = ta->cast_to_exactness(true);
1452 }
1453 }
1454 }
1455
1456 // Oop pointers need some flattening
1457 const TypeInstPtr *to = tj->isa_instptr();
1458 if (to && to != TypeOopPtr::BOTTOM) {
1459 ciInstanceKlass* ik = to->instance_klass();
1460 tj = to = to->cast_to_maybe_flat_in_array(); // flatten to maybe flat in array
1461 if( ptr == TypePtr::Constant ) {
1462 if (ik != ciEnv::current()->Class_klass() ||
1463 offset < ik->layout_helper_size_in_bytes()) {
1464 // No constant oop pointers (such as Strings); they alias with
1465 // unknown strings.
1466 assert(!is_known_inst, "not scalarizable allocation");
1467 tj = to = to->
1468 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1469 remove_speculative()->
1470 cast_to_ptr_type(TypePtr::BotPTR)->
1471 cast_to_exactness(false);
1472 }
1473 } else if( is_known_inst ) {
1474 tj = to; // Keep NotNull and klass_is_exact for instance type
1475 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1476 // During the 2nd round of IterGVN, NotNull castings are removed.
1477 // Make sure the Bottom and NotNull variants alias the same.
1478 // Also, make sure exact and non-exact variants alias the same.
1479 tj = to = to->
1480 remove_speculative()->
1481 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1482 cast_to_ptr_type(TypePtr::BotPTR)->
1483 cast_to_exactness(false);
1484 }
1485 if (to->speculative() != nullptr) {
1486 tj = to = to->remove_speculative();
1487 }
1488 // Canonicalize the holder of this field
1489 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1490 // First handle header references such as a LoadKlassNode, even if the
1491 // object's klass is unloaded at compile time (4965979).
1492 if (!is_known_inst) { // Do it only for non-instance types
1493 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1494 }
1495 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1496 // Static fields are in the space above the normal instance
1497 // fields in the java.lang.Class instance.
1498 if (ik != ciEnv::current()->Class_klass()) {
1499 to = nullptr;
1500 tj = TypeOopPtr::BOTTOM;
1501 offset = tj->offset();
1502 }
1503 } else {
1504 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1505 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1506 assert(tj->offset() == offset, "no change to offset expected");
1507 bool xk = to->klass_is_exact();
1508 int instance_id = to->instance_id();
1509
1510 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1511 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1512 // its interfaces are included.
1513 if (xk && ik->equals(canonical_holder)) {
1514 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id,
1515 TypePtr::MaybeFlat), "exact type should be canonical type");
1516 } else {
1517 assert(xk || !is_known_inst, "Known instance should be exact type");
1518 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id,
1519 TypePtr::MaybeFlat);
1520 }
1521 }
1522 }
1523
1524 // Klass pointers to object array klasses need some flattening
1525 const TypeKlassPtr *tk = tj->isa_klassptr();
1526 if( tk ) {
1527 // If we are referencing a field within a Klass, we need
1528 // to assume the worst case of an Object. Both exact and
1529 // inexact types must flatten to the same alias class so
1530 // use NotNull as the PTR.
1531 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1532 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1533 env()->Object_klass(),
1534 Type::Offset(offset),
1535 TypePtr::MaybeFlat);
1536 }
1537
1538 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1539 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1540 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1541 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset), TypePtr::MaybeFlat);
1542 } else {
1543 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());
1544 }
1545 }
1546 // Check for precise loads from the primary supertype array and force them
1547 // to the supertype cache alias index. Check for generic array loads from
1548 // the primary supertype array and also force them to the supertype cache
1549 // alias index. Since the same load can reach both, we need to merge
1550 // these 2 disparate memories into the same alias class. Since the
1551 // primary supertype array is read-only, there's no chance of confusion
1552 // where we bypass an array load and an array store.
1553 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1554 if (offset == Type::OffsetBot ||
1555 (offset >= primary_supers_offset &&
1556 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1557 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1558 offset = in_bytes(Klass::secondary_super_cache_offset());
1559 tj = tk = tk->with_offset(offset);
1560 }
1561 }
1562
1563 // Flatten all Raw pointers together.
1564 if (tj->base() == Type::RawPtr)
1565 tj = TypeRawPtr::BOTTOM;
1566
1567 if (tj->base() == Type::AnyPtr)
1568 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1569
1570 offset = tj->offset();
1571 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1572
1573 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1574 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1575 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1576 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1577 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1578 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1579 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1580 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1581 assert( tj->ptr() != TypePtr::TopPTR &&
1582 tj->ptr() != TypePtr::AnyNull &&
1583 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1584 // assert( tj->ptr() != TypePtr::Constant ||
1585 // tj->base() == Type::RawPtr ||
1586 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1587
1588 return tj;
1589 }
1590
1591 void Compile::AliasType::Init(int i, const TypePtr* at) {
1592 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1593 _index = i;
1594 _adr_type = at;
1595 _field = nullptr;
1596 _element = nullptr;
1597 _is_rewritable = true; // default
1598 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1599 if (atoop != nullptr && atoop->is_known_instance()) {
1600 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1601 _general_index = Compile::current()->get_alias_index(gt);
1602 } else {
1603 _general_index = 0;
1604 }
1605 }
1606
1607 BasicType Compile::AliasType::basic_type() const {
1608 if (element() != nullptr) {
1609 const Type* element = adr_type()->is_aryptr()->elem();
1610 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1611 } if (field() != nullptr) {
1612 return field()->layout_type();
1613 } else {
1614 return T_ILLEGAL; // unknown
1615 }
1616 }
1617
1618 //---------------------------------print_on------------------------------------
1619 #ifndef PRODUCT
1620 void Compile::AliasType::print_on(outputStream* st) {
1621 if (index() < 10)
1622 st->print("@ <%d> ", index());
1623 else st->print("@ <%d>", index());
1624 st->print(is_rewritable() ? " " : " RO");
1625 int offset = adr_type()->offset();
1626 if (offset == Type::OffsetBot)
1627 st->print(" +any");
1628 else st->print(" +%-3d", offset);
1629 st->print(" in ");
1630 adr_type()->dump_on(st);
1631 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1632 if (field() != nullptr && tjp) {
1633 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1634 tjp->offset() != field()->offset_in_bytes()) {
1635 st->print(" != ");
1636 field()->print();
1637 st->print(" ***");
1638 }
1639 }
1640 }
1641
1642 void print_alias_types() {
1643 Compile* C = Compile::current();
1644 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1645 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1646 C->alias_type(idx)->print_on(tty);
1647 tty->cr();
1648 }
1649 }
1650 #endif
1651
1652
1653 //----------------------------probe_alias_cache--------------------------------
1654 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1655 intptr_t key = (intptr_t) adr_type;
1656 key ^= key >> logAliasCacheSize;
1657 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1658 }
1659
1660
1661 //-----------------------------grow_alias_types--------------------------------
1662 void Compile::grow_alias_types() {
1663 const int old_ats = _max_alias_types; // how many before?
1664 const int new_ats = old_ats; // how many more?
1665 const int grow_ats = old_ats+new_ats; // how many now?
1666 _max_alias_types = grow_ats;
1667 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1668 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1669 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1670 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1671 }
1672
1673
1674 //--------------------------------find_alias_type------------------------------
1675 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1676 if (!do_aliasing()) {
1677 return alias_type(AliasIdxBot);
1678 }
1679
1680 AliasCacheEntry* ace = nullptr;
1681 if (!uncached) {
1682 ace = probe_alias_cache(adr_type);
1683 if (ace->_adr_type == adr_type) {
1684 return alias_type(ace->_index);
1685 }
1686 }
1687
1688 // Handle special cases.
1689 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1690 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1691
1692 // Do it the slow way.
1693 const TypePtr* flat = flatten_alias_type(adr_type);
1694
1695 #ifdef ASSERT
1696 {
1697 ResourceMark rm;
1698 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1699 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1700 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1701 Type::str(adr_type));
1702 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1703 const TypeOopPtr* foop = flat->is_oopptr();
1704 // Scalarizable allocations have exact klass always.
1705 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1706 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1707 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1708 Type::str(foop), Type::str(xoop));
1709 }
1710 }
1711 #endif
1712
1713 int idx = AliasIdxTop;
1714 for (int i = 0; i < num_alias_types(); i++) {
1715 if (alias_type(i)->adr_type() == flat) {
1716 idx = i;
1717 break;
1718 }
1719 }
1720
1721 if (idx == AliasIdxTop) {
1722 if (no_create) return nullptr;
1723 // Grow the array if necessary.
1724 if (_num_alias_types == _max_alias_types) grow_alias_types();
1725 // Add a new alias type.
1726 idx = _num_alias_types++;
1727 _alias_types[idx]->Init(idx, flat);
1728 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1729 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1730 if (flat->isa_instptr()) {
1731 if (flat->offset() == java_lang_Class::klass_offset()
1732 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1733 alias_type(idx)->set_rewritable(false);
1734 }
1735 ciField* field = nullptr;
1736 if (flat->isa_aryptr()) {
1737 #ifdef ASSERT
1738 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1739 // (T_BYTE has the weakest alignment and size restrictions...)
1740 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1741 #endif
1742 const Type* elemtype = flat->is_aryptr()->elem();
1743 if (flat->offset() == TypePtr::OffsetBot) {
1744 alias_type(idx)->set_element(elemtype);
1745 }
1746 int field_offset = flat->is_aryptr()->field_offset().get();
1747 if (flat->is_flat() &&
1748 field_offset != Type::OffsetBot) {
1749 ciInlineKlass* vk = elemtype->inline_klass();
1750 field_offset += vk->payload_offset();
1751 field = vk->get_field_by_offset(field_offset, false);
1752 }
1753 }
1754 if (flat->isa_klassptr()) {
1755 if (UseCompactObjectHeaders) {
1756 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1757 alias_type(idx)->set_rewritable(false);
1758 }
1759 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1760 alias_type(idx)->set_rewritable(false);
1761 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1762 alias_type(idx)->set_rewritable(false);
1763 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1764 alias_type(idx)->set_rewritable(false);
1765 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1766 alias_type(idx)->set_rewritable(false);
1767 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1768 alias_type(idx)->set_rewritable(false);
1769 }
1770
1771 if (flat->isa_instklassptr()) {
1772 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1773 alias_type(idx)->set_rewritable(false);
1774 }
1775 }
1776 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1777 // but the base pointer type is not distinctive enough to identify
1778 // references into JavaThread.)
1779
1780 // Check for final fields.
1781 const TypeInstPtr* tinst = flat->isa_instptr();
1782 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1783 if (tinst->const_oop() != nullptr &&
1784 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1785 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1786 // static field
1787 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1788 field = k->get_field_by_offset(tinst->offset(), true);
1789 } else if (tinst->is_inlinetypeptr()) {
1790 // Inline type field
1791 ciInlineKlass* vk = tinst->inline_klass();
1792 field = vk->get_field_by_offset(tinst->offset(), false);
1793 } else {
1794 ciInstanceKlass *k = tinst->instance_klass();
1795 field = k->get_field_by_offset(tinst->offset(), false);
1796 }
1797 }
1798 assert(field == nullptr ||
1799 original_field == nullptr ||
1800 (field->holder() == original_field->holder() &&
1801 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1802 field->is_static() == original_field->is_static()), "wrong field?");
1803 // Set field() and is_rewritable() attributes.
1804 if (field != nullptr) {
1805 alias_type(idx)->set_field(field);
1806 if (flat->isa_aryptr()) {
1807 // Fields of flat arrays are rewritable although they are declared final
1808 assert(flat->is_flat(), "must be a flat array");
1809 alias_type(idx)->set_rewritable(true);
1810 }
1811 }
1812 }
1813
1814 // Fill the cache for next time.
1815 if (!uncached) {
1816 ace->_adr_type = adr_type;
1817 ace->_index = idx;
1818 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1819
1820 // Might as well try to fill the cache for the flattened version, too.
1821 AliasCacheEntry* face = probe_alias_cache(flat);
1822 if (face->_adr_type == nullptr) {
1823 face->_adr_type = flat;
1824 face->_index = idx;
1825 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1826 }
1827 }
1828
1829 return alias_type(idx);
1830 }
1831
1832
1833 Compile::AliasType* Compile::alias_type(ciField* field) {
1834 const TypeOopPtr* t;
1835 if (field->is_static())
1836 t = TypeInstPtr::make(field->holder()->java_mirror());
1837 else
1838 t = TypeOopPtr::make_from_klass_raw(field->holder());
1839 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1840 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1841 return atp;
1842 }
1843
1844
1845 //------------------------------have_alias_type--------------------------------
1846 bool Compile::have_alias_type(const TypePtr* adr_type) {
1847 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1848 if (ace->_adr_type == adr_type) {
1849 return true;
1850 }
1851
1852 // Handle special cases.
1853 if (adr_type == nullptr) return true;
1854 if (adr_type == TypePtr::BOTTOM) return true;
1855
1856 return find_alias_type(adr_type, true, nullptr) != nullptr;
1857 }
1858
1859 //-----------------------------must_alias--------------------------------------
1860 // True if all values of the given address type are in the given alias category.
1861 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1862 if (alias_idx == AliasIdxBot) return true; // the universal category
1863 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1864 if (alias_idx == AliasIdxTop) return false; // the empty category
1865 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1866
1867 // the only remaining possible overlap is identity
1868 int adr_idx = get_alias_index(adr_type);
1869 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1870 assert(adr_idx == alias_idx ||
1871 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1872 && adr_type != TypeOopPtr::BOTTOM),
1873 "should not be testing for overlap with an unsafe pointer");
1874 return adr_idx == alias_idx;
1875 }
1876
1877 //------------------------------can_alias--------------------------------------
1878 // True if any values of the given address type are in the given alias category.
1879 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1880 if (alias_idx == AliasIdxTop) return false; // the empty category
1881 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1882 // Known instance doesn't alias with bottom memory
1883 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1884 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1885
1886 // the only remaining possible overlap is identity
1887 int adr_idx = get_alias_index(adr_type);
1888 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1889 return adr_idx == alias_idx;
1890 }
1891
1892 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1893 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1894 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1895 if (parse_predicate_count() == 0) {
1896 return;
1897 }
1898 for (int i = 0; i < parse_predicate_count(); i++) {
1899 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1900 parse_predicate->mark_useless(igvn);
1901 }
1902 _parse_predicates.clear();
1903 }
1904
1905 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1906 if (!n->for_post_loop_opts_igvn()) {
1907 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1908 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1909 _for_post_loop_igvn.append(n);
1910 }
1911 }
1912
1913 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1914 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1915 _for_post_loop_igvn.remove(n);
1916 }
1917
1918 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1919 // Verify that all previous optimizations produced a valid graph
1920 // at least to this point, even if no loop optimizations were done.
1921 PhaseIdealLoop::verify(igvn);
1922
1923 if (_print_phase_loop_opts) {
1924 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1925 }
1926 C->set_post_loop_opts_phase(); // no more loop opts allowed
1927
1928 assert(!C->major_progress(), "not cleared");
1929
1930 if (_for_post_loop_igvn.length() > 0) {
1931 while (_for_post_loop_igvn.length() > 0) {
1932 Node* n = _for_post_loop_igvn.pop();
1933 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1934 igvn._worklist.push(n);
1935 }
1936 igvn.optimize();
1937 if (failing()) return;
1938 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1939 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1940
1941 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1942 if (C->major_progress()) {
1943 C->clear_major_progress(); // ensure that major progress is now clear
1944 }
1945 }
1946 }
1947
1948 void Compile::add_inline_type(Node* n) {
1949 assert(n->is_InlineType(), "unexpected node");
1950 _inline_type_nodes.push(n);
1951 }
1952
1953 void Compile::remove_inline_type(Node* n) {
1954 assert(n->is_InlineType(), "unexpected node");
1955 if (_inline_type_nodes.contains(n)) {
1956 _inline_type_nodes.remove(n);
1957 }
1958 }
1959
1960 // Does the return value keep otherwise useless inline type allocations alive?
1961 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1962 ResourceMark rm;
1963 Unique_Node_List wq;
1964 wq.push(ret_val);
1965 bool some_allocations = false;
1966 for (uint i = 0; i < wq.size(); i++) {
1967 Node* n = wq.at(i);
1968 if (n->outcnt() > 1) {
1969 // Some other use for the allocation
1970 return false;
1971 } else if (n->is_InlineType()) {
1972 wq.push(n->in(1));
1973 } else if (n->is_Phi()) {
1974 for (uint j = 1; j < n->req(); j++) {
1975 wq.push(n->in(j));
1976 }
1977 } else if (n->is_CheckCastPP() &&
1978 n->in(1)->is_Proj() &&
1979 n->in(1)->in(0)->is_Allocate()) {
1980 some_allocations = true;
1981 } else if (n->is_CheckCastPP()) {
1982 wq.push(n->in(1));
1983 }
1984 }
1985 return some_allocations;
1986 }
1987
1988 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1989 // Make sure that the return value does not keep an otherwise unused allocation alive
1990 if (tf()->returns_inline_type_as_fields()) {
1991 Node* ret = nullptr;
1992 for (uint i = 1; i < root()->req(); i++) {
1993 Node* in = root()->in(i);
1994 if (in->Opcode() == Op_Return) {
1995 assert(ret == nullptr, "only one return");
1996 ret = in;
1997 }
1998 }
1999 if (ret != nullptr) {
2000 Node* ret_val = ret->in(TypeFunc::Parms);
2001 if (igvn.type(ret_val)->isa_oopptr() &&
2002 return_val_keeps_allocations_alive(ret_val)) {
2003 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
2004 assert(ret_val->outcnt() == 0, "should be dead now");
2005 igvn.remove_dead_node(ret_val);
2006 }
2007 }
2008 }
2009 if (_inline_type_nodes.length() == 0) {
2010 // keep the graph canonical
2011 igvn.optimize();
2012 return;
2013 }
2014 // Scalarize inline types in safepoint debug info.
2015 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2016 set_scalarize_in_safepoints(true);
2017 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2018 InlineTypeNode* vt = _inline_type_nodes.at(i)->as_InlineType();
2019 vt->make_scalar_in_safepoints(&igvn);
2020 igvn.record_for_igvn(vt);
2021 }
2022 if (remove) {
2023 // Remove inline type nodes by replacing them with their oop input
2024 while (_inline_type_nodes.length() > 0) {
2025 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2026 if (vt->outcnt() == 0) {
2027 igvn.remove_dead_node(vt);
2028 continue;
2029 }
2030 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2031 DEBUG_ONLY(bool must_be_buffered = false);
2032 Node* u = vt->out(i);
2033 // Check if any users are blackholes. If so, rewrite them to use either the
2034 // allocated buffer, or individual components, instead of the inline type node
2035 // that goes away.
2036 if (u->is_Blackhole()) {
2037 BlackholeNode* bh = u->as_Blackhole();
2038
2039 // Unlink the old input
2040 int idx = bh->find_edge(vt);
2041 assert(idx != -1, "The edge should be there");
2042 bh->del_req(idx);
2043 --i;
2044
2045 if (vt->is_allocated(&igvn)) {
2046 // Already has the allocated instance, blackhole that
2047 bh->add_req(vt->get_oop());
2048 } else {
2049 // Not allocated yet, blackhole the components
2050 for (uint c = 0; c < vt->field_count(); c++) {
2051 bh->add_req(vt->field_value(c));
2052 }
2053 }
2054
2055 // Node modified, record for IGVN
2056 igvn.record_for_igvn(bh);
2057 }
2058 #ifdef ASSERT
2059 // Verify that inline type is buffered when replacing by oop
2060 else if (u->is_InlineType()) {
2061 // InlineType uses don't need buffering because they are about to be replaced as well
2062 } else if (u->is_Phi()) {
2063 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2064 } else {
2065 must_be_buffered = true;
2066 }
2067 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2068 vt->dump(0);
2069 u->dump(0);
2070 assert(false, "Should have been buffered");
2071 }
2072 #endif
2073 }
2074 igvn.replace_node(vt, vt->get_oop());
2075 }
2076 }
2077 igvn.optimize();
2078 }
2079
2080 void Compile::add_flat_access(Node* n) {
2081 assert(n != nullptr && (n->Opcode() == Op_LoadFlat || n->Opcode() == Op_StoreFlat), "unexpected node %s", n == nullptr ? "nullptr" : n->Name());
2082 assert(!_flat_access_nodes.contains(n), "duplicate insertion");
2083 _flat_access_nodes.push(n);
2084 }
2085
2086 void Compile::remove_flat_access(Node* n) {
2087 assert(n != nullptr && (n->Opcode() == Op_LoadFlat || n->Opcode() == Op_StoreFlat), "unexpected node %s", n == nullptr ? "nullptr" : n->Name());
2088 _flat_access_nodes.remove_if_existing(n);
2089 }
2090
2091 void Compile::process_flat_accesses(PhaseIterGVN& igvn) {
2092 assert(igvn._worklist.size() == 0, "should be empty");
2093 igvn.set_delay_transform(true);
2094 for (int i = _flat_access_nodes.length() - 1; i >= 0; i--) {
2095 Node* n = _flat_access_nodes.at(i);
2096 assert(n != nullptr, "unexpected nullptr");
2097 if (n->is_LoadFlat()) {
2098 LoadFlatNode* loadn = n->as_LoadFlat();
2099 // Expending a flat load atomically means that we get a chunk of memory spanning multiple fields
2100 // that we chop with bitwise operations. That is too subtle for some optimizations, especially
2101 // constant folding when fields are constant. If we can get a constant object from which we are
2102 // flat-loading, we can simply replace the loads at compilation-time by the field of the constant
2103 // object.
2104 ciInstance* loaded_from = nullptr;
2105 if (FoldStableValues) {
2106 const TypeOopPtr* base_type = igvn.type(loadn->base())->is_oopptr();
2107 ciObject* oop = base_type->const_oop();
2108 int off = igvn.type(loadn->ptr())->isa_ptr()->offset();
2109
2110 if (oop != nullptr && oop->is_instance()) {
2111 ciInstance* holder = oop->as_instance();
2112 ciKlass* klass = holder->klass();
2113 ciInstanceKlass* iklass = klass->as_instance_klass();
2114 ciField* field = iklass->get_non_flat_field_by_offset(off);
2115
2116 if (field->is_stable()) {
2117 ciConstant fv = holder->field_value(field);
2118 if (is_reference_type(fv.basic_type()) && fv.as_object()->is_instance()) {
2119 // The field value is an object, not null. We can use stability.
2120 loaded_from = fv.as_object()->as_instance();
2121 }
2122 }
2123 } else if (oop != nullptr && oop->is_array() && off != Type::OffsetBot) {
2124 ciArray* array = oop->as_array();
2125 ciConstant elt = array->element_value_by_offset(off);
2126 const TypeAryPtr* aryptr = base_type->is_aryptr();
2127 if (aryptr->is_stable() && aryptr->is_atomic() && is_reference_type(elt.basic_type()) && elt.as_object()->is_instance()) {
2128 loaded_from = elt.as_object()->as_instance();
2129 }
2130 }
2131 }
2132
2133 if (loaded_from != nullptr) {
2134 loadn->expand_constant(igvn, loaded_from);
2135 } else {
2136 loadn->expand_atomic(igvn);
2137 }
2138 } else {
2139 n->as_StoreFlat()->expand_atomic(igvn);
2140 }
2141 }
2142 _flat_access_nodes.clear_and_deallocate();
2143 igvn.set_delay_transform(false);
2144 igvn.optimize();
2145 }
2146
2147 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2148 DEBUG_ONLY(igvn.verify_empty_worklist(nullptr));
2149 if (!_has_flat_accesses) {
2150 return;
2151 }
2152 // Initially, all flat array accesses share the same slice to
2153 // keep dependencies with Object[] array accesses (that could be
2154 // to a flat array) correct. We're done with parsing so we
2155 // now know all flat array accesses in this compile
2156 // unit. Let's move flat array accesses to their own slice,
2157 // one per element field. This should help memory access
2158 // optimizations.
2159 ResourceMark rm;
2160 Unique_Node_List wq;
2161 wq.push(root());
2162
2163 Node_List mergememnodes;
2164 Node_List memnodes;
2165
2166 // Alias index currently shared by all flat memory accesses
2167 int index = get_alias_index(TypeAryPtr::INLINES);
2168
2169 // Find MergeMem nodes and flat array accesses
2170 for (uint i = 0; i < wq.size(); i++) {
2171 Node* n = wq.at(i);
2172 if (n->is_Mem()) {
2173 const TypePtr* adr_type = nullptr;
2174 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2175 if (adr_type == TypeAryPtr::INLINES) {
2176 memnodes.push(n);
2177 }
2178 } else if (n->is_MergeMem()) {
2179 MergeMemNode* mm = n->as_MergeMem();
2180 if (mm->memory_at(index) != mm->base_memory()) {
2181 mergememnodes.push(n);
2182 }
2183 }
2184 for (uint j = 0; j < n->req(); j++) {
2185 Node* m = n->in(j);
2186 if (m != nullptr) {
2187 wq.push(m);
2188 }
2189 }
2190 }
2191
2192 _flat_accesses_share_alias = false;
2193
2194 // We are going to change the slice for the flat array
2195 // accesses so we need to clear the cache entries that refer to
2196 // them.
2197 for (uint i = 0; i < AliasCacheSize; i++) {
2198 AliasCacheEntry* ace = &_alias_cache[i];
2199 if (ace->_adr_type != nullptr &&
2200 ace->_adr_type->is_flat()) {
2201 ace->_adr_type = nullptr;
2202 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2203 }
2204 }
2205
2206 #ifdef ASSERT
2207 for (uint i = 0; i < memnodes.size(); i++) {
2208 Node* m = memnodes.at(i);
2209 const TypePtr* adr_type = m->adr_type();
2210 m->as_Mem()->set_adr_type(adr_type);
2211 }
2212 #endif // ASSERT
2213
2214 int start_alias = num_alias_types(); // Start of new aliases
2215 Node_Stack stack(0);
2216 #ifdef ASSERT
2217 VectorSet seen(Thread::current()->resource_area());
2218 #endif
2219 // Now let's fix the memory graph so each flat array access
2220 // is moved to the right slice. Start from the MergeMem nodes.
2221 uint last = unique();
2222 for (uint i = 0; i < mergememnodes.size(); i++) {
2223 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2224 if (current->outcnt() == 0) {
2225 // This node is killed by a previous iteration
2226 continue;
2227 }
2228
2229 Node* n = current->memory_at(index);
2230 MergeMemNode* mm = nullptr;
2231 do {
2232 // Follow memory edges through memory accesses, phis and
2233 // narrow membars and push nodes on the stack. Once we hit
2234 // bottom memory, we pop element off the stack one at a
2235 // time, in reverse order, and move them to the right slice
2236 // by changing their memory edges.
2237 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() ||
2238 (n->adr_type() == TypeAryPtr::INLINES && !n->is_NarrowMemProj())) {
2239 assert(!seen.test_set(n->_idx), "");
2240 // Uses (a load for instance) will need to be moved to the
2241 // right slice as well and will get a new memory state
2242 // that we don't know yet. The use could also be the
2243 // backedge of a loop. We put a place holder node between
2244 // the memory node and its uses. We replace that place
2245 // holder with the correct memory state once we know it,
2246 // i.e. when nodes are popped off the stack. Using the
2247 // place holder make the logic work in the presence of
2248 // loops.
2249 if (n->outcnt() > 1) {
2250 Node* place_holder = nullptr;
2251 assert(!n->has_out_with(Op_Node), "");
2252 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2253 Node* u = n->out(k);
2254 if (u != current && u->_idx < last) {
2255 bool success = false;
2256 for (uint l = 0; l < u->req(); l++) {
2257 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2258 continue;
2259 }
2260 Node* in = u->in(l);
2261 if (in == n) {
2262 if (place_holder == nullptr) {
2263 place_holder = new Node(1);
2264 place_holder->init_req(0, n);
2265 }
2266 igvn.replace_input_of(u, l, place_holder);
2267 success = true;
2268 }
2269 }
2270 if (success) {
2271 --k;
2272 }
2273 }
2274 }
2275 }
2276 if (n->is_Phi()) {
2277 stack.push(n, 1);
2278 n = n->in(1);
2279 } else if (n->is_Mem()) {
2280 stack.push(n, n->req());
2281 n = n->in(MemNode::Memory);
2282 } else {
2283 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2284 stack.push(n, n->req());
2285 n = n->in(0)->in(TypeFunc::Memory);
2286 }
2287 } else {
2288 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || n->is_NarrowMemProj(), "");
2289 // Build a new MergeMem node to carry the new memory state
2290 // as we build it. IGVN should fold extraneous MergeMem
2291 // nodes.
2292 if (n->is_NarrowMemProj()) {
2293 // We need 1 NarrowMemProj for each slice of this array
2294 InitializeNode* init = n->in(0)->as_Initialize();
2295 AllocateNode* alloc = init->allocation();
2296 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2297 const TypeAryKlassPtr* klass_type = klass_node->bottom_type()->isa_aryklassptr();
2298 assert(klass_type != nullptr, "must be an array");
2299 assert(klass_type->klass_is_exact(), "must be an exact klass");
2300 ciArrayKlass* klass = klass_type->exact_klass()->as_array_klass();
2301 assert(klass->is_flat_array_klass(), "must be a flat array");
2302 ciInlineKlass* elem_klass = klass->element_klass()->as_inline_klass();
2303 const TypeAryPtr* oop_type = klass_type->as_instance_type()->is_aryptr();
2304 assert(oop_type->klass_is_exact(), "must be an exact klass");
2305
2306 Node* base = alloc->in(TypeFunc::Memory);
2307 assert(base->bottom_type() == Type::MEMORY, "the memory input of AllocateNode must be a memory");
2308 assert(base->adr_type() == TypePtr::BOTTOM, "the memory input of AllocateNode must be a bottom memory");
2309 // Must create a MergeMem with base as the base memory, do not clone if base is a
2310 // MergeMem because it may not be processed yet
2311 mm = MergeMemNode::make(nullptr);
2312 mm->set_base_memory(base);
2313 for (int j = 0; j < elem_klass->nof_nonstatic_fields(); j++) {
2314 int field_offset = elem_klass->nonstatic_field_at(j)->offset_in_bytes() - elem_klass->payload_offset();
2315 const TypeAryPtr* field_ptr = oop_type->with_offset(Type::OffsetBot)->with_field_offset(field_offset);
2316 int field_alias_idx = get_alias_index(field_ptr);
2317 assert(field_ptr == get_adr_type(field_alias_idx), "must match");
2318 Node* new_proj = new NarrowMemProjNode(init, field_ptr);
2319 igvn.register_new_node_with_optimizer(new_proj);
2320 mm->set_memory_at(field_alias_idx, new_proj);
2321 }
2322 if (!klass->is_elem_null_free()) {
2323 int nm_offset = elem_klass->null_marker_offset_in_payload();
2324 const TypeAryPtr* nm_ptr = oop_type->with_offset(Type::OffsetBot)->with_field_offset(nm_offset);
2325 int nm_alias_idx = get_alias_index(nm_ptr);
2326 assert(nm_ptr == get_adr_type(nm_alias_idx), "must match");
2327 Node* new_proj = new NarrowMemProjNode(init, nm_ptr);
2328 igvn.register_new_node_with_optimizer(new_proj);
2329 mm->set_memory_at(nm_alias_idx, new_proj);
2330 }
2331
2332 // Replace all uses of the old NarrowMemProj with the correct state
2333 MergeMemNode* new_n = MergeMemNode::make(mm);
2334 igvn.register_new_node_with_optimizer(new_n);
2335 igvn.replace_node(n, new_n);
2336 } else {
2337 // Must create a MergeMem with n as the base memory, do not clone if n is a MergeMem
2338 // because it may not be processed yet
2339 mm = MergeMemNode::make(nullptr);
2340 mm->set_base_memory(n);
2341 }
2342
2343 igvn.register_new_node_with_optimizer(mm);
2344 while (stack.size() > 0) {
2345 Node* m = stack.node();
2346 uint idx = stack.index();
2347 if (m->is_Mem()) {
2348 // Move memory node to its new slice
2349 const TypePtr* adr_type = m->adr_type();
2350 int alias = get_alias_index(adr_type);
2351 Node* prev = mm->memory_at(alias);
2352 igvn.replace_input_of(m, MemNode::Memory, prev);
2353 mm->set_memory_at(alias, m);
2354 } else if (m->is_Phi()) {
2355 // We need as many new phis as there are new aliases
2356 Node* new_phi_in = MergeMemNode::make(mm);
2357 igvn.register_new_node_with_optimizer(new_phi_in);
2358 igvn.replace_input_of(m, idx, new_phi_in);
2359 if (idx == m->req()-1) {
2360 Node* r = m->in(0);
2361 for (int j = start_alias; j < num_alias_types(); j++) {
2362 const TypePtr* adr_type = get_adr_type(j);
2363 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2364 continue;
2365 }
2366 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2367 igvn.register_new_node_with_optimizer(phi);
2368 for (uint k = 1; k < m->req(); k++) {
2369 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2370 }
2371 mm->set_memory_at(j, phi);
2372 }
2373 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2374 igvn.register_new_node_with_optimizer(base_phi);
2375 for (uint k = 1; k < m->req(); k++) {
2376 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2377 }
2378 mm->set_base_memory(base_phi);
2379 }
2380 } else {
2381 // This is a MemBarCPUOrder node from
2382 // Parse::array_load()/Parse::array_store(), in the
2383 // branch that handles flat arrays hidden under
2384 // an Object[] array. We also need one new membar per
2385 // new alias to keep the unknown access that the
2386 // membars protect properly ordered with accesses to
2387 // known flat array.
2388 assert(m->is_Proj(), "projection expected");
2389 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2390 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2391 for (int j = start_alias; j < num_alias_types(); j++) {
2392 const TypePtr* adr_type = get_adr_type(j);
2393 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2394 continue;
2395 }
2396 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2397 igvn.register_new_node_with_optimizer(mb);
2398 Node* mem = mm->memory_at(j);
2399 mb->init_req(TypeFunc::Control, ctrl);
2400 mb->init_req(TypeFunc::Memory, mem);
2401 ctrl = new ProjNode(mb, TypeFunc::Control);
2402 igvn.register_new_node_with_optimizer(ctrl);
2403 mem = new ProjNode(mb, TypeFunc::Memory);
2404 igvn.register_new_node_with_optimizer(mem);
2405 mm->set_memory_at(j, mem);
2406 }
2407 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2408 }
2409 if (idx < m->req()-1) {
2410 idx += 1;
2411 stack.set_index(idx);
2412 n = m->in(idx);
2413 break;
2414 }
2415 // Take care of place holder nodes
2416 if (m->has_out_with(Op_Node)) {
2417 Node* place_holder = m->find_out_with(Op_Node);
2418 if (place_holder != nullptr) {
2419 Node* mm_clone = mm->clone();
2420 igvn.register_new_node_with_optimizer(mm_clone);
2421 Node* hook = new Node(1);
2422 hook->init_req(0, mm);
2423 igvn.replace_node(place_holder, mm_clone);
2424 hook->destruct(&igvn);
2425 }
2426 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2427 }
2428 stack.pop();
2429 }
2430 }
2431 } while(stack.size() > 0);
2432 // Fix the memory state at the MergeMem we started from
2433 igvn.rehash_node_delayed(current);
2434 for (int j = start_alias; j < num_alias_types(); j++) {
2435 const TypePtr* adr_type = get_adr_type(j);
2436 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2437 continue;
2438 }
2439 current->set_memory_at(j, mm);
2440 }
2441 current->set_memory_at(index, current->base_memory());
2442 }
2443 igvn.optimize();
2444
2445 #ifdef ASSERT
2446 wq.clear();
2447 wq.push(root());
2448 for (uint i = 0; i < wq.size(); i++) {
2449 Node* n = wq.at(i);
2450 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2451 for (uint j = 0; j < n->req(); j++) {
2452 Node* m = n->in(j);
2453 if (m != nullptr) {
2454 wq.push(m);
2455 }
2456 }
2457 }
2458 #endif
2459
2460 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2461 }
2462
2463 void Compile::record_for_merge_stores_igvn(Node* n) {
2464 if (!n->for_merge_stores_igvn()) {
2465 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
2466 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2467 _for_merge_stores_igvn.append(n);
2468 }
2469 }
2470
2471 void Compile::remove_from_merge_stores_igvn(Node* n) {
2472 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2473 _for_merge_stores_igvn.remove(n);
2474 }
2475
2476 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
2477 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
2478 // the stores, and we merge the wrong sequence of stores.
2479 // Example:
2480 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
2481 // Apply MergeStores:
2482 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
2483 // Remove more RangeChecks:
2484 // StoreI [ StoreL ] StoreI
2485 // But now it would have been better to do this instead:
2486 // [ StoreL ] [ StoreL ]
2487 //
2488 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
2489 // since we never unset _merge_stores_phase.
2490 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
2491 C->set_merge_stores_phase();
2492
2493 if (_for_merge_stores_igvn.length() > 0) {
2494 while (_for_merge_stores_igvn.length() > 0) {
2495 Node* n = _for_merge_stores_igvn.pop();
2496 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2497 igvn._worklist.push(n);
2498 }
2499 igvn.optimize();
2500 if (failing()) return;
2501 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
2502 print_method(PHASE_AFTER_MERGE_STORES, 3);
2503 }
2504 }
2505
2506 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2507 if (OptimizeUnstableIf) {
2508 _unstable_if_traps.append(trap);
2509 }
2510 }
2511
2512 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2513 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2514 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2515 Node* n = trap->uncommon_trap();
2516 if (!useful.member(n)) {
2517 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2518 }
2519 }
2520 }
2521
2522 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2523 // or fold-compares case. Return true if succeed or not found.
2524 //
2525 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2526 // false and ask fold-compares to yield.
2527 //
2528 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2529 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2530 // when deoptimization does happen.
2531 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2532 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2533 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2534 if (trap->uncommon_trap() == unc) {
2535 if (yield && trap->modified()) {
2536 return false;
2537 }
2538 _unstable_if_traps.delete_at(i);
2539 break;
2540 }
2541 }
2542 return true;
2543 }
2544
2545 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2546 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2547 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2548 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2549 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2550 CallStaticJavaNode* unc = trap->uncommon_trap();
2551 int next_bci = trap->next_bci();
2552 bool modified = trap->modified();
2553
2554 if (next_bci != -1 && !modified) {
2555 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2556 JVMState* jvms = unc->jvms();
2557 ciMethod* method = jvms->method();
2558 ciBytecodeStream iter(method);
2559
2560 iter.force_bci(jvms->bci());
2561 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2562 Bytecodes::Code c = iter.cur_bc();
2563 Node* lhs = nullptr;
2564 Node* rhs = nullptr;
2565 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2566 lhs = unc->peek_operand(0);
2567 rhs = unc->peek_operand(1);
2568 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2569 lhs = unc->peek_operand(0);
2570 }
2571
2572 ResourceMark rm;
2573 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2574 assert(live_locals.is_valid(), "broken liveness info");
2575 int len = (int)live_locals.size();
2576
2577 for (int i = 0; i < len; i++) {
2578 Node* local = unc->local(jvms, i);
2579 // kill local using the liveness of next_bci.
2580 // give up when the local looks like an operand to secure reexecution.
2581 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2582 uint idx = jvms->locoff() + i;
2583 #ifdef ASSERT
2584 if (PrintOpto && Verbose) {
2585 tty->print("[unstable_if] kill local#%d: ", idx);
2586 local->dump();
2587 tty->cr();
2588 }
2589 #endif
2590 igvn.replace_input_of(unc, idx, top());
2591 modified = true;
2592 }
2593 }
2594 }
2595
2596 // keep the modified trap for late query
2597 if (modified) {
2598 trap->set_modified();
2599 } else {
2600 _unstable_if_traps.delete_at(i);
2601 }
2602 }
2603 igvn.optimize();
2604 }
2605
2606 // StringOpts and late inlining of string methods
2607 void Compile::inline_string_calls(bool parse_time) {
2608 {
2609 // remove useless nodes to make the usage analysis simpler
2610 ResourceMark rm;
2611 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2612 }
2613
2614 {
2615 ResourceMark rm;
2616 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2617 PhaseStringOpts pso(initial_gvn());
2618 print_method(PHASE_AFTER_STRINGOPTS, 3);
2619 }
2620
2621 // now inline anything that we skipped the first time around
2622 if (!parse_time) {
2623 _late_inlines_pos = _late_inlines.length();
2624 }
2625
2626 while (_string_late_inlines.length() > 0) {
2627 CallGenerator* cg = _string_late_inlines.pop();
2628 cg->do_late_inline();
2629 if (failing()) return;
2630 }
2631 _string_late_inlines.trunc_to(0);
2632 }
2633
2634 // Late inlining of boxing methods
2635 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2636 if (_boxing_late_inlines.length() > 0) {
2637 assert(has_boxed_value(), "inconsistent");
2638
2639 set_inlining_incrementally(true);
2640
2641 igvn_worklist()->ensure_empty(); // should be done with igvn
2642
2643 _late_inlines_pos = _late_inlines.length();
2644
2645 while (_boxing_late_inlines.length() > 0) {
2646 CallGenerator* cg = _boxing_late_inlines.pop();
2647 cg->do_late_inline();
2648 if (failing()) return;
2649 }
2650 _boxing_late_inlines.trunc_to(0);
2651
2652 inline_incrementally_cleanup(igvn);
2653
2654 set_inlining_incrementally(false);
2655 }
2656 }
2657
2658 bool Compile::inline_incrementally_one() {
2659 assert(IncrementalInline, "incremental inlining should be on");
2660 assert(_late_inlines.length() > 0, "should have been checked by caller");
2661
2662 TracePhase tp(_t_incrInline_inline);
2663
2664 set_inlining_progress(false);
2665 set_do_cleanup(false);
2666
2667 for (int i = 0; i < _late_inlines.length(); i++) {
2668 _late_inlines_pos = i+1;
2669 CallGenerator* cg = _late_inlines.at(i);
2670 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2671 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2672 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2673 if (should_stress_inlining()) {
2674 // randomly add repeated inline attempt if stress-inlining
2675 cg->call_node()->set_generator(cg);
2676 C->igvn_worklist()->push(cg->call_node());
2677 continue;
2678 }
2679 cg->do_late_inline();
2680 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2681 if (failing()) {
2682 return false;
2683 } else if (inlining_progress()) {
2684 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2685 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2686 break; // process one call site at a time
2687 } else {
2688 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2689 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2690 // Avoid potential infinite loop if node already in the IGVN list
2691 assert(false, "scheduled for IGVN during inlining attempt");
2692 } else {
2693 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2694 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2695 cg->call_node()->set_generator(cg);
2696 }
2697 }
2698 } else {
2699 // Ignore late inline direct calls when inlining is not allowed.
2700 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2701 }
2702 }
2703 // Remove processed elements.
2704 _late_inlines.remove_till(_late_inlines_pos);
2705 _late_inlines_pos = 0;
2706
2707 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2708
2709 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2710
2711 set_inlining_progress(false);
2712 set_do_cleanup(false);
2713
2714 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2715 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2716 }
2717
2718 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2719 {
2720 TracePhase tp(_t_incrInline_pru);
2721 ResourceMark rm;
2722 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2723 }
2724 {
2725 TracePhase tp(_t_incrInline_igvn);
2726 igvn.reset();
2727 igvn.optimize();
2728 if (failing()) return;
2729 }
2730 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2731 }
2732
2733 template<typename E>
2734 static void shuffle_array(Compile& C, GrowableArray<E>& array) {
2735 if (array.length() < 2) {
2736 return;
2737 }
2738 for (uint i = array.length() - 1; i >= 1; i--) {
2739 uint j = C.random() % (i + 1);
2740 swap(array.at(i), array.at(j));
2741 }
2742 }
2743
2744 void Compile::shuffle_late_inlines() {
2745 shuffle_array(*C, _late_inlines);
2746 }
2747
2748 // Perform incremental inlining until bound on number of live nodes is reached
2749 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2750 TracePhase tp(_t_incrInline);
2751
2752 set_inlining_incrementally(true);
2753 uint low_live_nodes = 0;
2754
2755 if (StressIncrementalInlining) {
2756 shuffle_late_inlines();
2757 }
2758
2759 while (_late_inlines.length() > 0) {
2760 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2761 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2762 TracePhase tp(_t_incrInline_ideal);
2763 // PhaseIdealLoop is expensive so we only try it once we are
2764 // out of live nodes and we only try it again if the previous
2765 // helped got the number of nodes down significantly
2766 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2767 if (failing()) return;
2768 low_live_nodes = live_nodes();
2769 _major_progress = true;
2770 }
2771
2772 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2773 bool do_print_inlining = print_inlining() || print_intrinsics();
2774 if (do_print_inlining || log() != nullptr) {
2775 // Print inlining message for candidates that we couldn't inline for lack of space.
2776 for (int i = 0; i < _late_inlines.length(); i++) {
2777 CallGenerator* cg = _late_inlines.at(i);
2778 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2779 if (do_print_inlining) {
2780 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2781 }
2782 log_late_inline_failure(cg, msg);
2783 }
2784 }
2785 break; // finish
2786 }
2787 }
2788
2789 igvn_worklist()->ensure_empty(); // should be done with igvn
2790
2791 if (_late_inlines.length() == 0) {
2792 break; // no more progress
2793 }
2794
2795 while (inline_incrementally_one()) {
2796 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2797 }
2798 if (failing()) return;
2799
2800 inline_incrementally_cleanup(igvn);
2801
2802 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2803
2804 if (failing()) return;
2805 }
2806
2807 igvn_worklist()->ensure_empty(); // should be done with igvn
2808
2809 if (_string_late_inlines.length() > 0) {
2810 assert(has_stringbuilder(), "inconsistent");
2811
2812 inline_string_calls(false);
2813
2814 if (failing()) return;
2815
2816 inline_incrementally_cleanup(igvn);
2817 }
2818
2819 set_inlining_incrementally(false);
2820 }
2821
2822 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2823 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2824 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2825 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2826 // as if "inlining_incrementally() == true" were set.
2827 assert(inlining_incrementally() == false, "not allowed");
2828 set_strength_reduction(true);
2829 #ifdef ASSERT
2830 Unique_Node_List* modified_nodes = _modified_nodes;
2831 _modified_nodes = nullptr;
2832 #endif
2833 assert(_late_inlines.length() > 0, "sanity");
2834
2835 if (StressIncrementalInlining) {
2836 shuffle_late_inlines();
2837 }
2838
2839 while (_late_inlines.length() > 0) {
2840 igvn_worklist()->ensure_empty(); // should be done with igvn
2841
2842 while (inline_incrementally_one()) {
2843 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2844 }
2845 if (failing()) return;
2846
2847 inline_incrementally_cleanup(igvn);
2848 }
2849 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2850 set_strength_reduction(false);
2851 }
2852
2853 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2854 if (_loop_opts_cnt > 0) {
2855 while (major_progress() && (_loop_opts_cnt > 0)) {
2856 TracePhase tp(_t_idealLoop);
2857 PhaseIdealLoop::optimize(igvn, mode);
2858 _loop_opts_cnt--;
2859 if (failing()) return false;
2860 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2861 }
2862 }
2863 return true;
2864 }
2865
2866 // Remove edges from "root" to each SafePoint at a backward branch.
2867 // They were inserted during parsing (see add_safepoint()) to make
2868 // infinite loops without calls or exceptions visible to root, i.e.,
2869 // useful.
2870 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2871 Node *r = root();
2872 if (r != nullptr) {
2873 for (uint i = r->req(); i < r->len(); ++i) {
2874 Node *n = r->in(i);
2875 if (n != nullptr && n->is_SafePoint()) {
2876 r->rm_prec(i);
2877 if (n->outcnt() == 0) {
2878 igvn.remove_dead_node(n);
2879 }
2880 --i;
2881 }
2882 }
2883 // Parsing may have added top inputs to the root node (Path
2884 // leading to the Halt node proven dead). Make sure we get a
2885 // chance to clean them up.
2886 igvn._worklist.push(r);
2887 igvn.optimize();
2888 }
2889 }
2890
2891 //------------------------------Optimize---------------------------------------
2892 // Given a graph, optimize it.
2893 void Compile::Optimize() {
2894 TracePhase tp(_t_optimizer);
2895
2896 #ifndef PRODUCT
2897 if (env()->break_at_compile()) {
2898 BREAKPOINT;
2899 }
2900
2901 #endif
2902
2903 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2904 #ifdef ASSERT
2905 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2906 #endif
2907
2908 ResourceMark rm;
2909
2910 NOT_PRODUCT( verify_graph_edges(); )
2911
2912 print_method(PHASE_AFTER_PARSING, 1);
2913
2914 {
2915 // Iterative Global Value Numbering, including ideal transforms
2916 PhaseIterGVN igvn;
2917 #ifdef ASSERT
2918 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2919 #endif
2920 {
2921 TracePhase tp(_t_iterGVN);
2922 igvn.optimize();
2923 }
2924
2925 if (failing()) return;
2926
2927 print_method(PHASE_ITER_GVN1, 2);
2928
2929 process_for_unstable_if_traps(igvn);
2930
2931 if (failing()) return;
2932
2933 inline_incrementally(igvn);
2934
2935 print_method(PHASE_INCREMENTAL_INLINE, 2);
2936
2937 if (failing()) return;
2938
2939 if (eliminate_boxing()) {
2940 // Inline valueOf() methods now.
2941 inline_boxing_calls(igvn);
2942
2943 if (failing()) return;
2944
2945 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2946 inline_incrementally(igvn);
2947 }
2948
2949 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2950
2951 if (failing()) return;
2952 }
2953
2954 // Remove the speculative part of types and clean up the graph from
2955 // the extra CastPP nodes whose only purpose is to carry them. Do
2956 // that early so that optimizations are not disrupted by the extra
2957 // CastPP nodes.
2958 remove_speculative_types(igvn);
2959
2960 if (failing()) return;
2961
2962 // No more new expensive nodes will be added to the list from here
2963 // so keep only the actual candidates for optimizations.
2964 cleanup_expensive_nodes(igvn);
2965
2966 if (failing()) return;
2967
2968 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2969 if (EnableVectorSupport && has_vbox_nodes()) {
2970 TracePhase tp(_t_vector);
2971 PhaseVector pv(igvn);
2972 pv.optimize_vector_boxes();
2973 if (failing()) return;
2974 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2975 }
2976 assert(!has_vbox_nodes(), "sanity");
2977
2978 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2979 Compile::TracePhase tp(_t_renumberLive);
2980 igvn_worklist()->ensure_empty(); // should be done with igvn
2981 {
2982 ResourceMark rm;
2983 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2984 }
2985 igvn.reset();
2986 igvn.optimize();
2987 if (failing()) return;
2988 }
2989
2990 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2991 // safepoints
2992 remove_root_to_sfpts_edges(igvn);
2993
2994 // Process inline type nodes now that all inlining is over
2995 process_inline_types(igvn);
2996
2997 adjust_flat_array_access_aliases(igvn);
2998
2999 if (failing()) return;
3000
3001 if (C->macro_count() > 0) {
3002 // Eliminate some macro nodes before EA to reduce analysis pressure
3003 PhaseMacroExpand mexp(igvn);
3004 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
3005 if (failing()) {
3006 return;
3007 }
3008 igvn.set_delay_transform(false);
3009 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3010 }
3011
3012 _print_phase_loop_opts = has_loops();
3013 if (_print_phase_loop_opts) {
3014 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
3015 }
3016
3017 // Perform escape analysis
3018 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
3019 if (has_loops()) {
3020 // Cleanup graph (remove dead nodes).
3021 TracePhase tp(_t_idealLoop);
3022 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
3023 if (failing()) {
3024 return;
3025 }
3026 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
3027 if (C->macro_count() > 0) {
3028 // Eliminate some macro nodes before EA to reduce analysis pressure
3029 PhaseMacroExpand mexp(igvn);
3030 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
3031 if (failing()) {
3032 return;
3033 }
3034 igvn.set_delay_transform(false);
3035 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3036 }
3037 }
3038
3039 bool progress;
3040 do {
3041 ConnectionGraph::do_analysis(this, &igvn);
3042
3043 if (failing()) return;
3044
3045 int mcount = macro_count(); // Record number of allocations and locks before IGVN
3046
3047 // Optimize out fields loads from scalar replaceable allocations.
3048 igvn.optimize();
3049 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
3050
3051 if (failing()) return;
3052
3053 if (congraph() != nullptr && macro_count() > 0) {
3054 TracePhase tp(_t_macroEliminate);
3055 PhaseMacroExpand mexp(igvn);
3056 mexp.eliminate_macro_nodes();
3057 if (failing()) {
3058 return;
3059 }
3060 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
3061
3062 igvn.set_delay_transform(false);
3063 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3064 }
3065
3066 ConnectionGraph::verify_ram_nodes(this, root());
3067 if (failing()) return;
3068
3069 progress = do_iterative_escape_analysis() &&
3070 (macro_count() < mcount) &&
3071 ConnectionGraph::has_candidates(this);
3072 // Try again if candidates exist and made progress
3073 // by removing some allocations and/or locks.
3074 } while (progress);
3075 }
3076
3077 process_flat_accesses(igvn);
3078 if (failing()) {
3079 return;
3080 }
3081
3082 // Loop transforms on the ideal graph. Range Check Elimination,
3083 // peeling, unrolling, etc.
3084
3085 // Set loop opts counter
3086 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
3087 {
3088 TracePhase tp(_t_idealLoop);
3089 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
3090 _loop_opts_cnt--;
3091 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
3092 if (failing()) return;
3093 }
3094 // Loop opts pass if partial peeling occurred in previous pass
3095 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
3096 TracePhase tp(_t_idealLoop);
3097 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
3098 _loop_opts_cnt--;
3099 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
3100 if (failing()) return;
3101 }
3102 // Loop opts pass for loop-unrolling before CCP
3103 if(major_progress() && (_loop_opts_cnt > 0)) {
3104 TracePhase tp(_t_idealLoop);
3105 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
3106 _loop_opts_cnt--;
3107 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
3108 }
3109 if (!failing()) {
3110 // Verify that last round of loop opts produced a valid graph
3111 PhaseIdealLoop::verify(igvn);
3112 }
3113 }
3114 if (failing()) return;
3115
3116 // Conditional Constant Propagation;
3117 print_method(PHASE_BEFORE_CCP1, 2);
3118 PhaseCCP ccp( &igvn );
3119 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
3120 {
3121 TracePhase tp(_t_ccp);
3122 ccp.do_transform();
3123 }
3124 print_method(PHASE_CCP1, 2);
3125
3126 assert( true, "Break here to ccp.dump_old2new_map()");
3127
3128 // Iterative Global Value Numbering, including ideal transforms
3129 {
3130 TracePhase tp(_t_iterGVN2);
3131 igvn.reset_from_igvn(&ccp);
3132 igvn.optimize();
3133 }
3134 print_method(PHASE_ITER_GVN2, 2);
3135
3136 if (failing()) return;
3137
3138 // Loop transforms on the ideal graph. Range Check Elimination,
3139 // peeling, unrolling, etc.
3140 if (!optimize_loops(igvn, LoopOptsDefault)) {
3141 return;
3142 }
3143
3144 if (failing()) return;
3145
3146 C->clear_major_progress(); // ensure that major progress is now clear
3147
3148 process_for_post_loop_opts_igvn(igvn);
3149
3150 process_for_merge_stores_igvn(igvn);
3151
3152 if (failing()) return;
3153
3154 #ifdef ASSERT
3155 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
3156 #endif
3157
3158 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
3159
3160 if (_late_inlines.length() > 0) {
3161 // More opportunities to optimize virtual and MH calls.
3162 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
3163 process_late_inline_calls_no_inline(igvn);
3164 }
3165
3166 {
3167 TracePhase tp(_t_macroExpand);
3168 PhaseMacroExpand mex(igvn);
3169 // Last attempt to eliminate macro nodes.
3170 mex.eliminate_macro_nodes();
3171 if (failing()) {
3172 return;
3173 }
3174
3175 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
3176 // Do not allow new macro nodes once we start to eliminate and expand
3177 C->reset_allow_macro_nodes();
3178 // Last attempt to eliminate macro nodes before expand
3179 mex.eliminate_macro_nodes();
3180 if (failing()) {
3181 return;
3182 }
3183 mex.eliminate_opaque_looplimit_macro_nodes();
3184 if (failing()) {
3185 return;
3186 }
3187 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
3188 if (mex.expand_macro_nodes()) {
3189 assert(failing(), "must bail out w/ explicit message");
3190 return;
3191 }
3192 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
3193 }
3194
3195 // Process inline type nodes again and remove them. From here
3196 // on we don't need to keep track of field values anymore.
3197 process_inline_types(igvn, /* remove= */ true);
3198
3199 {
3200 TracePhase tp(_t_barrierExpand);
3201 if (bs->expand_barriers(this, igvn)) {
3202 assert(failing(), "must bail out w/ explicit message");
3203 return;
3204 }
3205 print_method(PHASE_BARRIER_EXPANSION, 2);
3206 }
3207
3208 if (C->max_vector_size() > 0) {
3209 C->optimize_logic_cones(igvn);
3210 igvn.optimize();
3211 if (failing()) return;
3212 }
3213
3214 DEBUG_ONLY( _modified_nodes = nullptr; )
3215 DEBUG_ONLY( _late_inlines.clear(); )
3216
3217 assert(igvn._worklist.size() == 0, "not empty");
3218 } // (End scope of igvn; run destructor if necessary for asserts.)
3219
3220 check_no_dead_use();
3221
3222 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
3223 // to remove hashes to unlock nodes for modifications.
3224 C->node_hash()->clear();
3225
3226 // A method with only infinite loops has no edges entering loops from root
3227 {
3228 TracePhase tp(_t_graphReshaping);
3229 if (final_graph_reshaping()) {
3230 assert(failing(), "must bail out w/ explicit message");
3231 return;
3232 }
3233 }
3234
3235 print_method(PHASE_OPTIMIZE_FINISHED, 2);
3236 DEBUG_ONLY(set_phase_optimize_finished();)
3237 }
3238
3239 #ifdef ASSERT
3240 void Compile::check_no_dead_use() const {
3241 ResourceMark rm;
3242 Unique_Node_List wq;
3243 wq.push(root());
3244 for (uint i = 0; i < wq.size(); ++i) {
3245 Node* n = wq.at(i);
3246 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3247 Node* u = n->fast_out(j);
3248 if (u->outcnt() == 0 && !u->is_Con()) {
3249 u->dump();
3250 fatal("no reachable node should have no use");
3251 }
3252 wq.push(u);
3253 }
3254 }
3255 }
3256 #endif
3257
3258 void Compile::inline_vector_reboxing_calls() {
3259 if (C->_vector_reboxing_late_inlines.length() > 0) {
3260 _late_inlines_pos = C->_late_inlines.length();
3261 while (_vector_reboxing_late_inlines.length() > 0) {
3262 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
3263 cg->do_late_inline();
3264 if (failing()) return;
3265 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
3266 }
3267 _vector_reboxing_late_inlines.trunc_to(0);
3268 }
3269 }
3270
3271 bool Compile::has_vbox_nodes() {
3272 if (C->_vector_reboxing_late_inlines.length() > 0) {
3273 return true;
3274 }
3275 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3276 Node * n = C->macro_node(macro_idx);
3277 assert(n->is_macro(), "only macro nodes expected here");
3278 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3279 return true;
3280 }
3281 }
3282 return false;
3283 }
3284
3285 //---------------------------- Bitwise operation packing optimization ---------------------------
3286
3287 static bool is_vector_unary_bitwise_op(Node* n) {
3288 return n->Opcode() == Op_XorV &&
3289 VectorNode::is_vector_bitwise_not_pattern(n);
3290 }
3291
3292 static bool is_vector_binary_bitwise_op(Node* n) {
3293 switch (n->Opcode()) {
3294 case Op_AndV:
3295 case Op_OrV:
3296 return true;
3297
3298 case Op_XorV:
3299 return !is_vector_unary_bitwise_op(n);
3300
3301 default:
3302 return false;
3303 }
3304 }
3305
3306 static bool is_vector_ternary_bitwise_op(Node* n) {
3307 return n->Opcode() == Op_MacroLogicV;
3308 }
3309
3310 static bool is_vector_bitwise_op(Node* n) {
3311 return is_vector_unary_bitwise_op(n) ||
3312 is_vector_binary_bitwise_op(n) ||
3313 is_vector_ternary_bitwise_op(n);
3314 }
3315
3316 static bool is_vector_bitwise_cone_root(Node* n) {
3317 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
3318 return false;
3319 }
3320 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3321 if (is_vector_bitwise_op(n->fast_out(i))) {
3322 return false;
3323 }
3324 }
3325 return true;
3326 }
3327
3328 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3329 uint cnt = 0;
3330 if (is_vector_bitwise_op(n)) {
3331 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3332 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3333 for (uint i = 1; i < inp_cnt; i++) {
3334 Node* in = n->in(i);
3335 bool skip = VectorNode::is_all_ones_vector(in);
3336 if (!skip && !inputs.member(in)) {
3337 inputs.push(in);
3338 cnt++;
3339 }
3340 }
3341 assert(cnt <= 1, "not unary");
3342 } else {
3343 uint last_req = inp_cnt;
3344 if (is_vector_ternary_bitwise_op(n)) {
3345 last_req = inp_cnt - 1; // skip last input
3346 }
3347 for (uint i = 1; i < last_req; i++) {
3348 Node* def = n->in(i);
3349 if (!inputs.member(def)) {
3350 inputs.push(def);
3351 cnt++;
3352 }
3353 }
3354 }
3355 } else { // not a bitwise operations
3356 if (!inputs.member(n)) {
3357 inputs.push(n);
3358 cnt++;
3359 }
3360 }
3361 return cnt;
3362 }
3363
3364 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3365 Unique_Node_List useful_nodes;
3366 C->identify_useful_nodes(useful_nodes);
3367
3368 for (uint i = 0; i < useful_nodes.size(); i++) {
3369 Node* n = useful_nodes.at(i);
3370 if (is_vector_bitwise_cone_root(n)) {
3371 list.push(n);
3372 }
3373 }
3374 }
3375
3376 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3377 const TypeVect* vt,
3378 Unique_Node_List& partition,
3379 Unique_Node_List& inputs) {
3380 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3381 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3382 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3383
3384 Node* in1 = inputs.at(0);
3385 Node* in2 = inputs.at(1);
3386 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3387
3388 uint func = compute_truth_table(partition, inputs);
3389
3390 Node* pn = partition.at(partition.size() - 1);
3391 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3392 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3393 }
3394
3395 static uint extract_bit(uint func, uint pos) {
3396 return (func & (1 << pos)) >> pos;
3397 }
3398
3399 //
3400 // A macro logic node represents a truth table. It has 4 inputs,
3401 // First three inputs corresponds to 3 columns of a truth table
3402 // and fourth input captures the logic function.
3403 //
3404 // eg. fn = (in1 AND in2) OR in3;
3405 //
3406 // MacroNode(in1,in2,in3,fn)
3407 //
3408 // -----------------
3409 // in1 in2 in3 fn
3410 // -----------------
3411 // 0 0 0 0
3412 // 0 0 1 1
3413 // 0 1 0 0
3414 // 0 1 1 1
3415 // 1 0 0 0
3416 // 1 0 1 1
3417 // 1 1 0 1
3418 // 1 1 1 1
3419 //
3420
3421 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3422 int res = 0;
3423 for (int i = 0; i < 8; i++) {
3424 int bit1 = extract_bit(in1, i);
3425 int bit2 = extract_bit(in2, i);
3426 int bit3 = extract_bit(in3, i);
3427
3428 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3429 int func_bit = extract_bit(func, func_bit_pos);
3430
3431 res |= func_bit << i;
3432 }
3433 return res;
3434 }
3435
3436 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
3437 assert(n != nullptr, "");
3438 assert(eval_map.contains(n), "absent");
3439 return *(eval_map.get(n));
3440 }
3441
3442 static void eval_operands(Node* n,
3443 uint& func1, uint& func2, uint& func3,
3444 HashTable<Node*,uint>& eval_map) {
3445 assert(is_vector_bitwise_op(n), "");
3446
3447 if (is_vector_unary_bitwise_op(n)) {
3448 Node* opnd = n->in(1);
3449 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3450 opnd = n->in(2);
3451 }
3452 func1 = eval_operand(opnd, eval_map);
3453 } else if (is_vector_binary_bitwise_op(n)) {
3454 func1 = eval_operand(n->in(1), eval_map);
3455 func2 = eval_operand(n->in(2), eval_map);
3456 } else {
3457 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3458 func1 = eval_operand(n->in(1), eval_map);
3459 func2 = eval_operand(n->in(2), eval_map);
3460 func3 = eval_operand(n->in(3), eval_map);
3461 }
3462 }
3463
3464 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3465 assert(inputs.size() <= 3, "sanity");
3466 ResourceMark rm;
3467 uint res = 0;
3468 HashTable<Node*,uint> eval_map;
3469
3470 // Populate precomputed functions for inputs.
3471 // Each input corresponds to one column of 3 input truth-table.
3472 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3473 0xCC, // (_, b, _) -> b
3474 0xF0 }; // (a, _, _) -> a
3475 for (uint i = 0; i < inputs.size(); i++) {
3476 eval_map.put(inputs.at(i), input_funcs[2-i]);
3477 }
3478
3479 for (uint i = 0; i < partition.size(); i++) {
3480 Node* n = partition.at(i);
3481
3482 uint func1 = 0, func2 = 0, func3 = 0;
3483 eval_operands(n, func1, func2, func3, eval_map);
3484
3485 switch (n->Opcode()) {
3486 case Op_OrV:
3487 assert(func3 == 0, "not binary");
3488 res = func1 | func2;
3489 break;
3490 case Op_AndV:
3491 assert(func3 == 0, "not binary");
3492 res = func1 & func2;
3493 break;
3494 case Op_XorV:
3495 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3496 assert(func2 == 0 && func3 == 0, "not unary");
3497 res = (~func1) & 0xFF;
3498 } else {
3499 assert(func3 == 0, "not binary");
3500 res = func1 ^ func2;
3501 }
3502 break;
3503 case Op_MacroLogicV:
3504 // Ordering of inputs may change during evaluation of sub-tree
3505 // containing MacroLogic node as a child node, thus a re-evaluation
3506 // makes sure that function is evaluated in context of current
3507 // inputs.
3508 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3509 break;
3510
3511 default: assert(false, "not supported: %s", n->Name());
3512 }
3513 assert(res <= 0xFF, "invalid");
3514 eval_map.put(n, res);
3515 }
3516 return res;
3517 }
3518
3519 // Criteria under which nodes gets packed into a macro logic node:-
3520 // 1) Parent and both child nodes are all unmasked or masked with
3521 // same predicates.
3522 // 2) Masked parent can be packed with left child if it is predicated
3523 // and both have same predicates.
3524 // 3) Masked parent can be packed with right child if its un-predicated
3525 // or has matching predication condition.
3526 // 4) An unmasked parent can be packed with an unmasked child.
3527 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3528 assert(partition.size() == 0, "not empty");
3529 assert(inputs.size() == 0, "not empty");
3530 if (is_vector_ternary_bitwise_op(n)) {
3531 return false;
3532 }
3533
3534 bool is_unary_op = is_vector_unary_bitwise_op(n);
3535 if (is_unary_op) {
3536 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3537 return false; // too few inputs
3538 }
3539
3540 bool pack_left_child = true;
3541 bool pack_right_child = true;
3542
3543 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3544 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3545
3546 int left_child_input_cnt = 0;
3547 int right_child_input_cnt = 0;
3548
3549 bool parent_is_predicated = n->is_predicated_vector();
3550 bool left_child_predicated = n->in(1)->is_predicated_vector();
3551 bool right_child_predicated = n->in(2)->is_predicated_vector();
3552
3553 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
3554 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3555 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3556
3557 do {
3558 if (pack_left_child && left_child_LOP &&
3559 ((!parent_is_predicated && !left_child_predicated) ||
3560 ((parent_is_predicated && left_child_predicated &&
3561 parent_pred == left_child_pred)))) {
3562 partition.push(n->in(1));
3563 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3564 } else {
3565 inputs.push(n->in(1));
3566 left_child_input_cnt = 1;
3567 }
3568
3569 if (pack_right_child && right_child_LOP &&
3570 (!right_child_predicated ||
3571 (right_child_predicated && parent_is_predicated &&
3572 parent_pred == right_child_pred))) {
3573 partition.push(n->in(2));
3574 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3575 } else {
3576 inputs.push(n->in(2));
3577 right_child_input_cnt = 1;
3578 }
3579
3580 if (inputs.size() > 3) {
3581 assert(partition.size() > 0, "");
3582 inputs.clear();
3583 partition.clear();
3584 if (left_child_input_cnt > right_child_input_cnt) {
3585 pack_left_child = false;
3586 } else {
3587 pack_right_child = false;
3588 }
3589 } else {
3590 break;
3591 }
3592 } while(true);
3593
3594 if(partition.size()) {
3595 partition.push(n);
3596 }
3597
3598 return (partition.size() == 2 || partition.size() == 3) &&
3599 (inputs.size() == 2 || inputs.size() == 3);
3600 }
3601
3602 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3603 assert(is_vector_bitwise_op(n), "not a root");
3604
3605 visited.set(n->_idx);
3606
3607 // 1) Do a DFS walk over the logic cone.
3608 for (uint i = 1; i < n->req(); i++) {
3609 Node* in = n->in(i);
3610 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3611 process_logic_cone_root(igvn, in, visited);
3612 }
3613 }
3614
3615 // 2) Bottom up traversal: Merge node[s] with
3616 // the parent to form macro logic node.
3617 Unique_Node_List partition;
3618 Unique_Node_List inputs;
3619 if (compute_logic_cone(n, partition, inputs)) {
3620 const TypeVect* vt = n->bottom_type()->is_vect();
3621 Node* pn = partition.at(partition.size() - 1);
3622 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3623 if (mask == nullptr ||
3624 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3625 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3626 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3627 igvn.replace_node(n, macro_logic);
3628 }
3629 }
3630 }
3631
3632 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3633 ResourceMark rm;
3634 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3635 Unique_Node_List list;
3636 collect_logic_cone_roots(list);
3637
3638 while (list.size() > 0) {
3639 Node* n = list.pop();
3640 const TypeVect* vt = n->bottom_type()->is_vect();
3641 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3642 if (supported) {
3643 VectorSet visited(comp_arena());
3644 process_logic_cone_root(igvn, n, visited);
3645 }
3646 }
3647 }
3648 }
3649
3650 //------------------------------Code_Gen---------------------------------------
3651 // Given a graph, generate code for it
3652 void Compile::Code_Gen() {
3653 if (failing()) {
3654 return;
3655 }
3656
3657 // Perform instruction selection. You might think we could reclaim Matcher
3658 // memory PDQ, but actually the Matcher is used in generating spill code.
3659 // Internals of the Matcher (including some VectorSets) must remain live
3660 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3661 // set a bit in reclaimed memory.
3662
3663 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3664 // nodes. Mapping is only valid at the root of each matched subtree.
3665 NOT_PRODUCT( verify_graph_edges(); )
3666
3667 Matcher matcher;
3668 _matcher = &matcher;
3669 {
3670 TracePhase tp(_t_matcher);
3671 matcher.match();
3672 if (failing()) {
3673 return;
3674 }
3675 }
3676 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3677 // nodes. Mapping is only valid at the root of each matched subtree.
3678 NOT_PRODUCT( verify_graph_edges(); )
3679
3680 // If you have too many nodes, or if matching has failed, bail out
3681 check_node_count(0, "out of nodes matching instructions");
3682 if (failing()) {
3683 return;
3684 }
3685
3686 print_method(PHASE_MATCHING, 2);
3687
3688 // Build a proper-looking CFG
3689 PhaseCFG cfg(node_arena(), root(), matcher);
3690 if (failing()) {
3691 return;
3692 }
3693 _cfg = &cfg;
3694 {
3695 TracePhase tp(_t_scheduler);
3696 bool success = cfg.do_global_code_motion();
3697 if (!success) {
3698 return;
3699 }
3700
3701 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3702 NOT_PRODUCT( verify_graph_edges(); )
3703 cfg.verify();
3704 if (failing()) {
3705 return;
3706 }
3707 }
3708
3709 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3710 _regalloc = ®alloc;
3711 {
3712 TracePhase tp(_t_registerAllocation);
3713 // Perform register allocation. After Chaitin, use-def chains are
3714 // no longer accurate (at spill code) and so must be ignored.
3715 // Node->LRG->reg mappings are still accurate.
3716 _regalloc->Register_Allocate();
3717
3718 // Bail out if the allocator builds too many nodes
3719 if (failing()) {
3720 return;
3721 }
3722
3723 print_method(PHASE_REGISTER_ALLOCATION, 2);
3724 }
3725
3726 // Prior to register allocation we kept empty basic blocks in case the
3727 // the allocator needed a place to spill. After register allocation we
3728 // are not adding any new instructions. If any basic block is empty, we
3729 // can now safely remove it.
3730 {
3731 TracePhase tp(_t_blockOrdering);
3732 cfg.remove_empty_blocks();
3733 if (do_freq_based_layout()) {
3734 PhaseBlockLayout layout(cfg);
3735 } else {
3736 cfg.set_loop_alignment();
3737 }
3738 cfg.fixup_flow();
3739 cfg.remove_unreachable_blocks();
3740 cfg.verify_dominator_tree();
3741 print_method(PHASE_BLOCK_ORDERING, 3);
3742 }
3743
3744 // Apply peephole optimizations
3745 if( OptoPeephole ) {
3746 TracePhase tp(_t_peephole);
3747 PhasePeephole peep( _regalloc, cfg);
3748 peep.do_transform();
3749 print_method(PHASE_PEEPHOLE, 3);
3750 }
3751
3752 // Do late expand if CPU requires this.
3753 if (Matcher::require_postalloc_expand) {
3754 TracePhase tp(_t_postalloc_expand);
3755 cfg.postalloc_expand(_regalloc);
3756 print_method(PHASE_POSTALLOC_EXPAND, 3);
3757 }
3758
3759 #ifdef ASSERT
3760 {
3761 CompilationMemoryStatistic::do_test_allocations();
3762 if (failing()) return;
3763 }
3764 #endif
3765
3766 // Convert Nodes to instruction bits in a buffer
3767 {
3768 TracePhase tp(_t_output);
3769 PhaseOutput output;
3770 output.Output();
3771 if (failing()) return;
3772 output.install();
3773 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3774 }
3775
3776 // He's dead, Jim.
3777 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3778 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3779 }
3780
3781 //------------------------------Final_Reshape_Counts---------------------------
3782 // This class defines counters to help identify when a method
3783 // may/must be executed using hardware with only 24-bit precision.
3784 struct Final_Reshape_Counts : public StackObj {
3785 int _call_count; // count non-inlined 'common' calls
3786 int _float_count; // count float ops requiring 24-bit precision
3787 int _double_count; // count double ops requiring more precision
3788 int _java_call_count; // count non-inlined 'java' calls
3789 int _inner_loop_count; // count loops which need alignment
3790 VectorSet _visited; // Visitation flags
3791 Node_List _tests; // Set of IfNodes & PCTableNodes
3792
3793 Final_Reshape_Counts() :
3794 _call_count(0), _float_count(0), _double_count(0),
3795 _java_call_count(0), _inner_loop_count(0) { }
3796
3797 void inc_call_count () { _call_count ++; }
3798 void inc_float_count () { _float_count ++; }
3799 void inc_double_count() { _double_count++; }
3800 void inc_java_call_count() { _java_call_count++; }
3801 void inc_inner_loop_count() { _inner_loop_count++; }
3802
3803 int get_call_count () const { return _call_count ; }
3804 int get_float_count () const { return _float_count ; }
3805 int get_double_count() const { return _double_count; }
3806 int get_java_call_count() const { return _java_call_count; }
3807 int get_inner_loop_count() const { return _inner_loop_count; }
3808 };
3809
3810 //------------------------------final_graph_reshaping_impl----------------------
3811 // Implement items 1-5 from final_graph_reshaping below.
3812 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3813
3814 if ( n->outcnt() == 0 ) return; // dead node
3815 uint nop = n->Opcode();
3816
3817 // Check for 2-input instruction with "last use" on right input.
3818 // Swap to left input. Implements item (2).
3819 if( n->req() == 3 && // two-input instruction
3820 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3821 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3822 n->in(2)->outcnt() == 1 &&// right use IS a last use
3823 !n->in(2)->is_Con() ) { // right use is not a constant
3824 // Check for commutative opcode
3825 switch( nop ) {
3826 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3827 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3828 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3829 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3830 case Op_AndL: case Op_XorL: case Op_OrL:
3831 case Op_AndI: case Op_XorI: case Op_OrI: {
3832 // Move "last use" input to left by swapping inputs
3833 n->swap_edges(1, 2);
3834 break;
3835 }
3836 default:
3837 break;
3838 }
3839 }
3840
3841 #ifdef ASSERT
3842 if( n->is_Mem() ) {
3843 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3844 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3845 // oop will be recorded in oop map if load crosses safepoint
3846 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3847 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3848 "raw memory operations should have control edge");
3849 }
3850 if (n->is_MemBar()) {
3851 MemBarNode* mb = n->as_MemBar();
3852 if (mb->trailing_store() || mb->trailing_load_store()) {
3853 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3854 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3855 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3856 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3857 } else if (mb->leading()) {
3858 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3859 }
3860 }
3861 #endif
3862 // Count FPU ops and common calls, implements item (3)
3863 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3864 if (!gc_handled) {
3865 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3866 }
3867
3868 // Collect CFG split points
3869 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3870 frc._tests.push(n);
3871 }
3872 }
3873
3874 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3875 if (!UseDivMod) {
3876 return;
3877 }
3878
3879 // Check if "a % b" and "a / b" both exist
3880 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3881 if (d == nullptr) {
3882 return;
3883 }
3884
3885 // Replace them with a fused divmod if supported
3886 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3887 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3888 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3889 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3890 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3891 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3892 // DivMod node so the dependency is not lost.
3893 divmod->add_prec_from(n);
3894 divmod->add_prec_from(d);
3895 d->subsume_by(divmod->div_proj(), this);
3896 n->subsume_by(divmod->mod_proj(), this);
3897 } else {
3898 // Replace "a % b" with "a - ((a / b) * b)"
3899 Node* mult = MulNode::make(d, d->in(2), bt);
3900 Node* sub = SubNode::make(d->in(1), mult, bt);
3901 n->subsume_by(sub, this);
3902 }
3903 }
3904
3905 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3906 switch( nop ) {
3907 // Count all float operations that may use FPU
3908 case Op_AddF:
3909 case Op_SubF:
3910 case Op_MulF:
3911 case Op_DivF:
3912 case Op_NegF:
3913 case Op_ModF:
3914 case Op_ConvI2F:
3915 case Op_ConF:
3916 case Op_CmpF:
3917 case Op_CmpF3:
3918 case Op_StoreF:
3919 case Op_LoadF:
3920 // case Op_ConvL2F: // longs are split into 32-bit halves
3921 frc.inc_float_count();
3922 break;
3923
3924 case Op_ConvF2D:
3925 case Op_ConvD2F:
3926 frc.inc_float_count();
3927 frc.inc_double_count();
3928 break;
3929
3930 // Count all double operations that may use FPU
3931 case Op_AddD:
3932 case Op_SubD:
3933 case Op_MulD:
3934 case Op_DivD:
3935 case Op_NegD:
3936 case Op_ModD:
3937 case Op_ConvI2D:
3938 case Op_ConvD2I:
3939 // case Op_ConvL2D: // handled by leaf call
3940 // case Op_ConvD2L: // handled by leaf call
3941 case Op_ConD:
3942 case Op_CmpD:
3943 case Op_CmpD3:
3944 case Op_StoreD:
3945 case Op_LoadD:
3946 case Op_LoadD_unaligned:
3947 frc.inc_double_count();
3948 break;
3949 case Op_Opaque1: // Remove Opaque Nodes before matching
3950 n->subsume_by(n->in(1), this);
3951 break;
3952 case Op_CallLeafPure: {
3953 // If the pure call is not supported, then lower to a CallLeaf.
3954 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3955 CallNode* call = n->as_Call();
3956 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3957 call->_name, TypeRawPtr::BOTTOM);
3958 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3959 new_call->init_req(TypeFunc::I_O, C->top());
3960 new_call->init_req(TypeFunc::Memory, C->top());
3961 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3962 new_call->init_req(TypeFunc::FramePtr, C->top());
3963 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain_sig()->cnt(); i++) {
3964 new_call->init_req(i, call->in(i));
3965 }
3966 n->subsume_by(new_call, this);
3967 }
3968 frc.inc_call_count();
3969 break;
3970 }
3971 case Op_CallStaticJava:
3972 case Op_CallJava:
3973 case Op_CallDynamicJava:
3974 frc.inc_java_call_count(); // Count java call site;
3975 case Op_CallRuntime:
3976 case Op_CallLeaf:
3977 case Op_CallLeafVector:
3978 case Op_CallLeafNoFP: {
3979 assert (n->is_Call(), "");
3980 CallNode *call = n->as_Call();
3981 // Count call sites where the FP mode bit would have to be flipped.
3982 // Do not count uncommon runtime calls:
3983 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3984 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3985 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3986 frc.inc_call_count(); // Count the call site
3987 } else { // See if uncommon argument is shared
3988 Node *n = call->in(TypeFunc::Parms);
3989 int nop = n->Opcode();
3990 // Clone shared simple arguments to uncommon calls, item (1).
3991 if (n->outcnt() > 1 &&
3992 !n->is_Proj() &&
3993 nop != Op_CreateEx &&
3994 nop != Op_CheckCastPP &&
3995 nop != Op_DecodeN &&
3996 nop != Op_DecodeNKlass &&
3997 !n->is_Mem() &&
3998 !n->is_Phi()) {
3999 Node *x = n->clone();
4000 call->set_req(TypeFunc::Parms, x);
4001 }
4002 }
4003 break;
4004 }
4005 case Op_StoreB:
4006 case Op_StoreC:
4007 case Op_StoreI:
4008 case Op_StoreL:
4009 case Op_StoreLSpecial:
4010 case Op_CompareAndSwapB:
4011 case Op_CompareAndSwapS:
4012 case Op_CompareAndSwapI:
4013 case Op_CompareAndSwapL:
4014 case Op_CompareAndSwapP:
4015 case Op_CompareAndSwapN:
4016 case Op_WeakCompareAndSwapB:
4017 case Op_WeakCompareAndSwapS:
4018 case Op_WeakCompareAndSwapI:
4019 case Op_WeakCompareAndSwapL:
4020 case Op_WeakCompareAndSwapP:
4021 case Op_WeakCompareAndSwapN:
4022 case Op_CompareAndExchangeB:
4023 case Op_CompareAndExchangeS:
4024 case Op_CompareAndExchangeI:
4025 case Op_CompareAndExchangeL:
4026 case Op_CompareAndExchangeP:
4027 case Op_CompareAndExchangeN:
4028 case Op_GetAndAddS:
4029 case Op_GetAndAddB:
4030 case Op_GetAndAddI:
4031 case Op_GetAndAddL:
4032 case Op_GetAndSetS:
4033 case Op_GetAndSetB:
4034 case Op_GetAndSetI:
4035 case Op_GetAndSetL:
4036 case Op_GetAndSetP:
4037 case Op_GetAndSetN:
4038 case Op_StoreP:
4039 case Op_StoreN:
4040 case Op_StoreNKlass:
4041 case Op_LoadB:
4042 case Op_LoadUB:
4043 case Op_LoadUS:
4044 case Op_LoadI:
4045 case Op_LoadKlass:
4046 case Op_LoadNKlass:
4047 case Op_LoadL:
4048 case Op_LoadL_unaligned:
4049 case Op_LoadP:
4050 case Op_LoadN:
4051 case Op_LoadRange:
4052 case Op_LoadS:
4053 break;
4054
4055 case Op_AddP: { // Assert sane base pointers
4056 Node *addp = n->in(AddPNode::Address);
4057 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
4058 #ifdef _LP64
4059 if ((UseCompressedOops || UseCompressedClassPointers) &&
4060 addp->Opcode() == Op_ConP &&
4061 addp == n->in(AddPNode::Base) &&
4062 n->in(AddPNode::Offset)->is_Con()) {
4063 // If the transformation of ConP to ConN+DecodeN is beneficial depends
4064 // on the platform and on the compressed oops mode.
4065 // Use addressing with narrow klass to load with offset on x86.
4066 // Some platforms can use the constant pool to load ConP.
4067 // Do this transformation here since IGVN will convert ConN back to ConP.
4068 const Type* t = addp->bottom_type();
4069 bool is_oop = t->isa_oopptr() != nullptr;
4070 bool is_klass = t->isa_klassptr() != nullptr;
4071
4072 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
4073 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
4074 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
4075 Node* nn = nullptr;
4076
4077 int op = is_oop ? Op_ConN : Op_ConNKlass;
4078
4079 // Look for existing ConN node of the same exact type.
4080 Node* r = root();
4081 uint cnt = r->outcnt();
4082 for (uint i = 0; i < cnt; i++) {
4083 Node* m = r->raw_out(i);
4084 if (m!= nullptr && m->Opcode() == op &&
4085 m->bottom_type()->make_ptr() == t) {
4086 nn = m;
4087 break;
4088 }
4089 }
4090 if (nn != nullptr) {
4091 // Decode a narrow oop to match address
4092 // [R12 + narrow_oop_reg<<3 + offset]
4093 if (is_oop) {
4094 nn = new DecodeNNode(nn, t);
4095 } else {
4096 nn = new DecodeNKlassNode(nn, t);
4097 }
4098 // Check for succeeding AddP which uses the same Base.
4099 // Otherwise we will run into the assertion above when visiting that guy.
4100 for (uint i = 0; i < n->outcnt(); ++i) {
4101 Node *out_i = n->raw_out(i);
4102 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
4103 out_i->set_req(AddPNode::Base, nn);
4104 #ifdef ASSERT
4105 for (uint j = 0; j < out_i->outcnt(); ++j) {
4106 Node *out_j = out_i->raw_out(j);
4107 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
4108 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
4109 }
4110 #endif
4111 }
4112 }
4113 n->set_req(AddPNode::Base, nn);
4114 n->set_req(AddPNode::Address, nn);
4115 if (addp->outcnt() == 0) {
4116 addp->disconnect_inputs(this);
4117 }
4118 }
4119 }
4120 }
4121 #endif
4122 break;
4123 }
4124
4125 case Op_CastPP: {
4126 // Remove CastPP nodes to gain more freedom during scheduling but
4127 // keep the dependency they encode as control or precedence edges
4128 // (if control is set already) on memory operations. Some CastPP
4129 // nodes don't have a control (don't carry a dependency): skip
4130 // those.
4131 if (n->in(0) != nullptr) {
4132 ResourceMark rm;
4133 Unique_Node_List wq;
4134 wq.push(n);
4135 for (uint next = 0; next < wq.size(); ++next) {
4136 Node *m = wq.at(next);
4137 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
4138 Node* use = m->fast_out(i);
4139 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
4140 use->ensure_control_or_add_prec(n->in(0));
4141 } else {
4142 switch(use->Opcode()) {
4143 case Op_AddP:
4144 case Op_DecodeN:
4145 case Op_DecodeNKlass:
4146 case Op_CheckCastPP:
4147 case Op_CastPP:
4148 wq.push(use);
4149 break;
4150 }
4151 }
4152 }
4153 }
4154 }
4155 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
4156 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
4157 Node* in1 = n->in(1);
4158 const Type* t = n->bottom_type();
4159 Node* new_in1 = in1->clone();
4160 new_in1->as_DecodeN()->set_type(t);
4161
4162 if (!Matcher::narrow_oop_use_complex_address()) {
4163 //
4164 // x86, ARM and friends can handle 2 adds in addressing mode
4165 // and Matcher can fold a DecodeN node into address by using
4166 // a narrow oop directly and do implicit null check in address:
4167 //
4168 // [R12 + narrow_oop_reg<<3 + offset]
4169 // NullCheck narrow_oop_reg
4170 //
4171 // On other platforms (Sparc) we have to keep new DecodeN node and
4172 // use it to do implicit null check in address:
4173 //
4174 // decode_not_null narrow_oop_reg, base_reg
4175 // [base_reg + offset]
4176 // NullCheck base_reg
4177 //
4178 // Pin the new DecodeN node to non-null path on these platform (Sparc)
4179 // to keep the information to which null check the new DecodeN node
4180 // corresponds to use it as value in implicit_null_check().
4181 //
4182 new_in1->set_req(0, n->in(0));
4183 }
4184
4185 n->subsume_by(new_in1, this);
4186 if (in1->outcnt() == 0) {
4187 in1->disconnect_inputs(this);
4188 }
4189 } else {
4190 n->subsume_by(n->in(1), this);
4191 if (n->outcnt() == 0) {
4192 n->disconnect_inputs(this);
4193 }
4194 }
4195 break;
4196 }
4197 case Op_CastII: {
4198 n->as_CastII()->remove_range_check_cast(this);
4199 break;
4200 }
4201 #ifdef _LP64
4202 case Op_CmpP:
4203 // Do this transformation here to preserve CmpPNode::sub() and
4204 // other TypePtr related Ideal optimizations (for example, ptr nullness).
4205 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
4206 Node* in1 = n->in(1);
4207 Node* in2 = n->in(2);
4208 if (!in1->is_DecodeNarrowPtr()) {
4209 in2 = in1;
4210 in1 = n->in(2);
4211 }
4212 assert(in1->is_DecodeNarrowPtr(), "sanity");
4213
4214 Node* new_in2 = nullptr;
4215 if (in2->is_DecodeNarrowPtr()) {
4216 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
4217 new_in2 = in2->in(1);
4218 } else if (in2->Opcode() == Op_ConP) {
4219 const Type* t = in2->bottom_type();
4220 if (t == TypePtr::NULL_PTR) {
4221 assert(in1->is_DecodeN(), "compare klass to null?");
4222 // Don't convert CmpP null check into CmpN if compressed
4223 // oops implicit null check is not generated.
4224 // This will allow to generate normal oop implicit null check.
4225 if (Matcher::gen_narrow_oop_implicit_null_checks())
4226 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
4227 //
4228 // This transformation together with CastPP transformation above
4229 // will generated code for implicit null checks for compressed oops.
4230 //
4231 // The original code after Optimize()
4232 //
4233 // LoadN memory, narrow_oop_reg
4234 // decode narrow_oop_reg, base_reg
4235 // CmpP base_reg, nullptr
4236 // CastPP base_reg // NotNull
4237 // Load [base_reg + offset], val_reg
4238 //
4239 // after these transformations will be
4240 //
4241 // LoadN memory, narrow_oop_reg
4242 // CmpN narrow_oop_reg, nullptr
4243 // decode_not_null narrow_oop_reg, base_reg
4244 // Load [base_reg + offset], val_reg
4245 //
4246 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
4247 // since narrow oops can be used in debug info now (see the code in
4248 // final_graph_reshaping_walk()).
4249 //
4250 // At the end the code will be matched to
4251 // on x86:
4252 //
4253 // Load_narrow_oop memory, narrow_oop_reg
4254 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
4255 // NullCheck narrow_oop_reg
4256 //
4257 // and on sparc:
4258 //
4259 // Load_narrow_oop memory, narrow_oop_reg
4260 // decode_not_null narrow_oop_reg, base_reg
4261 // Load [base_reg + offset], val_reg
4262 // NullCheck base_reg
4263 //
4264 } else if (t->isa_oopptr()) {
4265 new_in2 = ConNode::make(t->make_narrowoop());
4266 } else if (t->isa_klassptr()) {
4267 ciKlass* klass = t->is_klassptr()->exact_klass();
4268 if (klass->is_in_encoding_range()) {
4269 new_in2 = ConNode::make(t->make_narrowklass());
4270 }
4271 }
4272 }
4273 if (new_in2 != nullptr) {
4274 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
4275 n->subsume_by(cmpN, this);
4276 if (in1->outcnt() == 0) {
4277 in1->disconnect_inputs(this);
4278 }
4279 if (in2->outcnt() == 0) {
4280 in2->disconnect_inputs(this);
4281 }
4282 }
4283 }
4284 break;
4285
4286 case Op_DecodeN:
4287 case Op_DecodeNKlass:
4288 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4289 // DecodeN could be pinned when it can't be fold into
4290 // an address expression, see the code for Op_CastPP above.
4291 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4292 break;
4293
4294 case Op_EncodeP:
4295 case Op_EncodePKlass: {
4296 Node* in1 = n->in(1);
4297 if (in1->is_DecodeNarrowPtr()) {
4298 n->subsume_by(in1->in(1), this);
4299 } else if (in1->Opcode() == Op_ConP) {
4300 const Type* t = in1->bottom_type();
4301 if (t == TypePtr::NULL_PTR) {
4302 assert(t->isa_oopptr(), "null klass?");
4303 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4304 } else if (t->isa_oopptr()) {
4305 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4306 } else if (t->isa_klassptr()) {
4307 ciKlass* klass = t->is_klassptr()->exact_klass();
4308 if (klass->is_in_encoding_range()) {
4309 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4310 } else {
4311 assert(false, "unencodable klass in ConP -> EncodeP");
4312 C->record_failure("unencodable klass in ConP -> EncodeP");
4313 }
4314 }
4315 }
4316 if (in1->outcnt() == 0) {
4317 in1->disconnect_inputs(this);
4318 }
4319 break;
4320 }
4321
4322 case Op_Proj: {
4323 if (OptimizeStringConcat || IncrementalInline) {
4324 ProjNode* proj = n->as_Proj();
4325 if (proj->_is_io_use) {
4326 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4327 // Separate projections were used for the exception path which
4328 // are normally removed by a late inline. If it wasn't inlined
4329 // then they will hang around and should just be replaced with
4330 // the original one. Merge them.
4331 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4332 if (non_io_proj != nullptr) {
4333 proj->subsume_by(non_io_proj , this);
4334 }
4335 }
4336 }
4337 break;
4338 }
4339
4340 case Op_Phi:
4341 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4342 // The EncodeP optimization may create Phi with the same edges
4343 // for all paths. It is not handled well by Register Allocator.
4344 Node* unique_in = n->in(1);
4345 assert(unique_in != nullptr, "");
4346 uint cnt = n->req();
4347 for (uint i = 2; i < cnt; i++) {
4348 Node* m = n->in(i);
4349 assert(m != nullptr, "");
4350 if (unique_in != m)
4351 unique_in = nullptr;
4352 }
4353 if (unique_in != nullptr) {
4354 n->subsume_by(unique_in, this);
4355 }
4356 }
4357 break;
4358
4359 #endif
4360
4361 case Op_ModI:
4362 handle_div_mod_op(n, T_INT, false);
4363 break;
4364
4365 case Op_ModL:
4366 handle_div_mod_op(n, T_LONG, false);
4367 break;
4368
4369 case Op_UModI:
4370 handle_div_mod_op(n, T_INT, true);
4371 break;
4372
4373 case Op_UModL:
4374 handle_div_mod_op(n, T_LONG, true);
4375 break;
4376
4377 case Op_LoadVector:
4378 case Op_StoreVector:
4379 #ifdef ASSERT
4380 // Add VerifyVectorAlignment node between adr and load / store.
4381 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
4382 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
4383 n->as_StoreVector()->must_verify_alignment();
4384 if (must_verify_alignment) {
4385 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
4386 n->as_StoreVector()->memory_size();
4387 // The memory access should be aligned to the vector width in bytes.
4388 // However, the underlying array is possibly less well aligned, but at least
4389 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
4390 // a loop we can expect at least the following alignment:
4391 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
4392 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
4393 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
4394 // Create mask from alignment. e.g. 0b1000 -> 0b0111
4395 jlong mask = guaranteed_alignment - 1;
4396 Node* mask_con = ConLNode::make(mask);
4397 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
4398 n->set_req(MemNode::Address, va);
4399 }
4400 }
4401 #endif
4402 break;
4403
4404 case Op_LoadVectorGather:
4405 case Op_StoreVectorScatter:
4406 case Op_LoadVectorGatherMasked:
4407 case Op_StoreVectorScatterMasked:
4408 case Op_VectorCmpMasked:
4409 case Op_VectorMaskGen:
4410 case Op_LoadVectorMasked:
4411 case Op_StoreVectorMasked:
4412 break;
4413
4414 case Op_AddReductionVI:
4415 case Op_AddReductionVL:
4416 case Op_AddReductionVF:
4417 case Op_AddReductionVD:
4418 case Op_MulReductionVI:
4419 case Op_MulReductionVL:
4420 case Op_MulReductionVF:
4421 case Op_MulReductionVD:
4422 case Op_MinReductionV:
4423 case Op_MaxReductionV:
4424 case Op_UMinReductionV:
4425 case Op_UMaxReductionV:
4426 case Op_AndReductionV:
4427 case Op_OrReductionV:
4428 case Op_XorReductionV:
4429 break;
4430
4431 case Op_PackB:
4432 case Op_PackS:
4433 case Op_PackI:
4434 case Op_PackF:
4435 case Op_PackL:
4436 case Op_PackD:
4437 if (n->req()-1 > 2) {
4438 // Replace many operand PackNodes with a binary tree for matching
4439 PackNode* p = (PackNode*) n;
4440 Node* btp = p->binary_tree_pack(1, n->req());
4441 n->subsume_by(btp, this);
4442 }
4443 break;
4444 case Op_Loop:
4445 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4446 case Op_CountedLoop:
4447 case Op_LongCountedLoop:
4448 case Op_OuterStripMinedLoop:
4449 if (n->as_Loop()->is_inner_loop()) {
4450 frc.inc_inner_loop_count();
4451 }
4452 n->as_Loop()->verify_strip_mined(0);
4453 break;
4454 case Op_LShiftI:
4455 case Op_RShiftI:
4456 case Op_URShiftI:
4457 case Op_LShiftL:
4458 case Op_RShiftL:
4459 case Op_URShiftL:
4460 if (Matcher::need_masked_shift_count) {
4461 // The cpu's shift instructions don't restrict the count to the
4462 // lower 5/6 bits. We need to do the masking ourselves.
4463 Node* in2 = n->in(2);
4464 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4465 const TypeInt* t = in2->find_int_type();
4466 if (t != nullptr && t->is_con()) {
4467 juint shift = t->get_con();
4468 if (shift > mask) { // Unsigned cmp
4469 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4470 }
4471 } else {
4472 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
4473 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4474 n->set_req(2, shift);
4475 }
4476 }
4477 if (in2->outcnt() == 0) { // Remove dead node
4478 in2->disconnect_inputs(this);
4479 }
4480 }
4481 break;
4482 case Op_MemBarStoreStore:
4483 case Op_MemBarRelease:
4484 // Break the link with AllocateNode: it is no longer useful and
4485 // confuses register allocation.
4486 if (n->req() > MemBarNode::Precedent) {
4487 n->set_req(MemBarNode::Precedent, top());
4488 }
4489 break;
4490 case Op_MemBarAcquire: {
4491 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4492 // At parse time, the trailing MemBarAcquire for a volatile load
4493 // is created with an edge to the load. After optimizations,
4494 // that input may be a chain of Phis. If those phis have no
4495 // other use, then the MemBarAcquire keeps them alive and
4496 // register allocation can be confused.
4497 dead_nodes.push(n->in(MemBarNode::Precedent));
4498 n->set_req(MemBarNode::Precedent, top());
4499 }
4500 break;
4501 }
4502 case Op_Blackhole:
4503 break;
4504 case Op_RangeCheck: {
4505 RangeCheckNode* rc = n->as_RangeCheck();
4506 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4507 n->subsume_by(iff, this);
4508 frc._tests.push(iff);
4509 break;
4510 }
4511 case Op_ConvI2L: {
4512 if (!Matcher::convi2l_type_required) {
4513 // Code generation on some platforms doesn't need accurate
4514 // ConvI2L types. Widening the type can help remove redundant
4515 // address computations.
4516 n->as_Type()->set_type(TypeLong::INT);
4517 ResourceMark rm;
4518 Unique_Node_List wq;
4519 wq.push(n);
4520 for (uint next = 0; next < wq.size(); next++) {
4521 Node *m = wq.at(next);
4522
4523 for(;;) {
4524 // Loop over all nodes with identical inputs edges as m
4525 Node* k = m->find_similar(m->Opcode());
4526 if (k == nullptr) {
4527 break;
4528 }
4529 // Push their uses so we get a chance to remove node made
4530 // redundant
4531 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4532 Node* u = k->fast_out(i);
4533 if (u->Opcode() == Op_LShiftL ||
4534 u->Opcode() == Op_AddL ||
4535 u->Opcode() == Op_SubL ||
4536 u->Opcode() == Op_AddP) {
4537 wq.push(u);
4538 }
4539 }
4540 // Replace all nodes with identical edges as m with m
4541 k->subsume_by(m, this);
4542 }
4543 }
4544 }
4545 break;
4546 }
4547 case Op_CmpUL: {
4548 if (!Matcher::has_match_rule(Op_CmpUL)) {
4549 // No support for unsigned long comparisons
4550 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4551 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4552 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4553 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4554 Node* andl = new AndLNode(orl, remove_sign_mask);
4555 Node* cmp = new CmpLNode(andl, n->in(2));
4556 n->subsume_by(cmp, this);
4557 }
4558 break;
4559 }
4560 #ifdef ASSERT
4561 case Op_InlineType: {
4562 n->dump(-1);
4563 assert(false, "inline type node was not removed");
4564 break;
4565 }
4566 case Op_ConNKlass: {
4567 const TypePtr* tp = n->as_Type()->type()->make_ptr();
4568 ciKlass* klass = tp->is_klassptr()->exact_klass();
4569 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
4570 break;
4571 }
4572 #endif
4573 default:
4574 assert(!n->is_Call(), "");
4575 assert(!n->is_Mem(), "");
4576 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4577 break;
4578 }
4579 }
4580
4581 //------------------------------final_graph_reshaping_walk---------------------
4582 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4583 // requires that the walk visits a node's inputs before visiting the node.
4584 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4585 Unique_Node_List sfpt;
4586
4587 frc._visited.set(root->_idx); // first, mark node as visited
4588 uint cnt = root->req();
4589 Node *n = root;
4590 uint i = 0;
4591 while (true) {
4592 if (i < cnt) {
4593 // Place all non-visited non-null inputs onto stack
4594 Node* m = n->in(i);
4595 ++i;
4596 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
4597 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
4598 // compute worst case interpreter size in case of a deoptimization
4599 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4600
4601 sfpt.push(m);
4602 }
4603 cnt = m->req();
4604 nstack.push(n, i); // put on stack parent and next input's index
4605 n = m;
4606 i = 0;
4607 }
4608 } else {
4609 // Now do post-visit work
4610 final_graph_reshaping_impl(n, frc, dead_nodes);
4611 if (nstack.is_empty())
4612 break; // finished
4613 n = nstack.node(); // Get node from stack
4614 cnt = n->req();
4615 i = nstack.index();
4616 nstack.pop(); // Shift to the next node on stack
4617 }
4618 }
4619
4620 // Skip next transformation if compressed oops are not used.
4621 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4622 (!UseCompressedOops && !UseCompressedClassPointers))
4623 return;
4624
4625 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4626 // It could be done for an uncommon traps or any safepoints/calls
4627 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4628 while (sfpt.size() > 0) {
4629 n = sfpt.pop();
4630 JVMState *jvms = n->as_SafePoint()->jvms();
4631 assert(jvms != nullptr, "sanity");
4632 int start = jvms->debug_start();
4633 int end = n->req();
4634 bool is_uncommon = (n->is_CallStaticJava() &&
4635 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4636 for (int j = start; j < end; j++) {
4637 Node* in = n->in(j);
4638 if (in->is_DecodeNarrowPtr()) {
4639 bool safe_to_skip = true;
4640 if (!is_uncommon ) {
4641 // Is it safe to skip?
4642 for (uint i = 0; i < in->outcnt(); i++) {
4643 Node* u = in->raw_out(i);
4644 if (!u->is_SafePoint() ||
4645 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4646 safe_to_skip = false;
4647 }
4648 }
4649 }
4650 if (safe_to_skip) {
4651 n->set_req(j, in->in(1));
4652 }
4653 if (in->outcnt() == 0) {
4654 in->disconnect_inputs(this);
4655 }
4656 }
4657 }
4658 }
4659 }
4660
4661 //------------------------------final_graph_reshaping--------------------------
4662 // Final Graph Reshaping.
4663 //
4664 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4665 // and not commoned up and forced early. Must come after regular
4666 // optimizations to avoid GVN undoing the cloning. Clone constant
4667 // inputs to Loop Phis; these will be split by the allocator anyways.
4668 // Remove Opaque nodes.
4669 // (2) Move last-uses by commutative operations to the left input to encourage
4670 // Intel update-in-place two-address operations and better register usage
4671 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4672 // calls canonicalizing them back.
4673 // (3) Count the number of double-precision FP ops, single-precision FP ops
4674 // and call sites. On Intel, we can get correct rounding either by
4675 // forcing singles to memory (requires extra stores and loads after each
4676 // FP bytecode) or we can set a rounding mode bit (requires setting and
4677 // clearing the mode bit around call sites). The mode bit is only used
4678 // if the relative frequency of single FP ops to calls is low enough.
4679 // This is a key transform for SPEC mpeg_audio.
4680 // (4) Detect infinite loops; blobs of code reachable from above but not
4681 // below. Several of the Code_Gen algorithms fail on such code shapes,
4682 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4683 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4684 // Detection is by looking for IfNodes where only 1 projection is
4685 // reachable from below or CatchNodes missing some targets.
4686 // (5) Assert for insane oop offsets in debug mode.
4687
4688 bool Compile::final_graph_reshaping() {
4689 // an infinite loop may have been eliminated by the optimizer,
4690 // in which case the graph will be empty.
4691 if (root()->req() == 1) {
4692 // Do not compile method that is only a trivial infinite loop,
4693 // since the content of the loop may have been eliminated.
4694 record_method_not_compilable("trivial infinite loop");
4695 return true;
4696 }
4697
4698 // Expensive nodes have their control input set to prevent the GVN
4699 // from freely commoning them. There's no GVN beyond this point so
4700 // no need to keep the control input. We want the expensive nodes to
4701 // be freely moved to the least frequent code path by gcm.
4702 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4703 for (int i = 0; i < expensive_count(); i++) {
4704 _expensive_nodes.at(i)->set_req(0, nullptr);
4705 }
4706
4707 Final_Reshape_Counts frc;
4708
4709 // Visit everybody reachable!
4710 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4711 Node_Stack nstack(live_nodes() >> 1);
4712 Unique_Node_List dead_nodes;
4713 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4714
4715 // Check for unreachable (from below) code (i.e., infinite loops).
4716 for( uint i = 0; i < frc._tests.size(); i++ ) {
4717 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4718 // Get number of CFG targets.
4719 // Note that PCTables include exception targets after calls.
4720 uint required_outcnt = n->required_outcnt();
4721 if (n->outcnt() != required_outcnt) {
4722 // Check for a few special cases. Rethrow Nodes never take the
4723 // 'fall-thru' path, so expected kids is 1 less.
4724 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4725 if (n->in(0)->in(0)->is_Call()) {
4726 CallNode* call = n->in(0)->in(0)->as_Call();
4727 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4728 required_outcnt--; // Rethrow always has 1 less kid
4729 } else if (call->req() > TypeFunc::Parms &&
4730 call->is_CallDynamicJava()) {
4731 // Check for null receiver. In such case, the optimizer has
4732 // detected that the virtual call will always result in a null
4733 // pointer exception. The fall-through projection of this CatchNode
4734 // will not be populated.
4735 Node* arg0 = call->in(TypeFunc::Parms);
4736 if (arg0->is_Type() &&
4737 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4738 required_outcnt--;
4739 }
4740 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4741 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4742 // Check for illegal array length. In such case, the optimizer has
4743 // detected that the allocation attempt will always result in an
4744 // exception. There is no fall-through projection of this CatchNode .
4745 assert(call->is_CallStaticJava(), "static call expected");
4746 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4747 uint valid_length_test_input = call->req() - 1;
4748 Node* valid_length_test = call->in(valid_length_test_input);
4749 call->del_req(valid_length_test_input);
4750 if (valid_length_test->find_int_con(1) == 0) {
4751 required_outcnt--;
4752 }
4753 dead_nodes.push(valid_length_test);
4754 assert(n->outcnt() == required_outcnt, "malformed control flow");
4755 continue;
4756 }
4757 }
4758 }
4759
4760 // Recheck with a better notion of 'required_outcnt'
4761 if (n->outcnt() != required_outcnt) {
4762 record_method_not_compilable("malformed control flow");
4763 return true; // Not all targets reachable!
4764 }
4765 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4766 CallNode* call = n->in(0)->in(0)->as_Call();
4767 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4768 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4769 assert(call->is_CallStaticJava(), "static call expected");
4770 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4771 uint valid_length_test_input = call->req() - 1;
4772 dead_nodes.push(call->in(valid_length_test_input));
4773 call->del_req(valid_length_test_input); // valid length test useless now
4774 }
4775 }
4776 // Check that I actually visited all kids. Unreached kids
4777 // must be infinite loops.
4778 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4779 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4780 record_method_not_compilable("infinite loop");
4781 return true; // Found unvisited kid; must be unreach
4782 }
4783
4784 // Here so verification code in final_graph_reshaping_walk()
4785 // always see an OuterStripMinedLoopEnd
4786 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4787 IfNode* init_iff = n->as_If();
4788 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4789 n->subsume_by(iff, this);
4790 }
4791 }
4792
4793 while (dead_nodes.size() > 0) {
4794 Node* m = dead_nodes.pop();
4795 if (m->outcnt() == 0 && m != top()) {
4796 for (uint j = 0; j < m->req(); j++) {
4797 Node* in = m->in(j);
4798 if (in != nullptr) {
4799 dead_nodes.push(in);
4800 }
4801 }
4802 m->disconnect_inputs(this);
4803 }
4804 }
4805
4806 set_java_calls(frc.get_java_call_count());
4807 set_inner_loops(frc.get_inner_loop_count());
4808
4809 // No infinite loops, no reason to bail out.
4810 return false;
4811 }
4812
4813 //-----------------------------too_many_traps----------------------------------
4814 // Report if there are too many traps at the current method and bci.
4815 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4816 bool Compile::too_many_traps(ciMethod* method,
4817 int bci,
4818 Deoptimization::DeoptReason reason) {
4819 ciMethodData* md = method->method_data();
4820 if (md->is_empty()) {
4821 // Assume the trap has not occurred, or that it occurred only
4822 // because of a transient condition during start-up in the interpreter.
4823 return false;
4824 }
4825 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4826 if (md->has_trap_at(bci, m, reason) != 0) {
4827 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4828 // Also, if there are multiple reasons, or if there is no per-BCI record,
4829 // assume the worst.
4830 if (log())
4831 log()->elem("observe trap='%s' count='%d'",
4832 Deoptimization::trap_reason_name(reason),
4833 md->trap_count(reason));
4834 return true;
4835 } else {
4836 // Ignore method/bci and see if there have been too many globally.
4837 return too_many_traps(reason, md);
4838 }
4839 }
4840
4841 // Less-accurate variant which does not require a method and bci.
4842 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4843 ciMethodData* logmd) {
4844 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4845 // Too many traps globally.
4846 // Note that we use cumulative trap_count, not just md->trap_count.
4847 if (log()) {
4848 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4849 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4850 Deoptimization::trap_reason_name(reason),
4851 mcount, trap_count(reason));
4852 }
4853 return true;
4854 } else {
4855 // The coast is clear.
4856 return false;
4857 }
4858 }
4859
4860 //--------------------------too_many_recompiles--------------------------------
4861 // Report if there are too many recompiles at the current method and bci.
4862 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4863 // Is not eager to return true, since this will cause the compiler to use
4864 // Action_none for a trap point, to avoid too many recompilations.
4865 bool Compile::too_many_recompiles(ciMethod* method,
4866 int bci,
4867 Deoptimization::DeoptReason reason) {
4868 ciMethodData* md = method->method_data();
4869 if (md->is_empty()) {
4870 // Assume the trap has not occurred, or that it occurred only
4871 // because of a transient condition during start-up in the interpreter.
4872 return false;
4873 }
4874 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4875 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4876 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4877 Deoptimization::DeoptReason per_bc_reason
4878 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4879 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4880 if ((per_bc_reason == Deoptimization::Reason_none
4881 || md->has_trap_at(bci, m, reason) != 0)
4882 // The trap frequency measure we care about is the recompile count:
4883 && md->trap_recompiled_at(bci, m)
4884 && md->overflow_recompile_count() >= bc_cutoff) {
4885 // Do not emit a trap here if it has already caused recompilations.
4886 // Also, if there are multiple reasons, or if there is no per-BCI record,
4887 // assume the worst.
4888 if (log())
4889 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4890 Deoptimization::trap_reason_name(reason),
4891 md->trap_count(reason),
4892 md->overflow_recompile_count());
4893 return true;
4894 } else if (trap_count(reason) != 0
4895 && decompile_count() >= m_cutoff) {
4896 // Too many recompiles globally, and we have seen this sort of trap.
4897 // Use cumulative decompile_count, not just md->decompile_count.
4898 if (log())
4899 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4900 Deoptimization::trap_reason_name(reason),
4901 md->trap_count(reason), trap_count(reason),
4902 md->decompile_count(), decompile_count());
4903 return true;
4904 } else {
4905 // The coast is clear.
4906 return false;
4907 }
4908 }
4909
4910 // Compute when not to trap. Used by matching trap based nodes and
4911 // NullCheck optimization.
4912 void Compile::set_allowed_deopt_reasons() {
4913 _allowed_reasons = 0;
4914 if (is_method_compilation()) {
4915 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4916 assert(rs < BitsPerInt, "recode bit map");
4917 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4918 _allowed_reasons |= nth_bit(rs);
4919 }
4920 }
4921 }
4922 }
4923
4924 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4925 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4926 }
4927
4928 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4929 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4930 }
4931
4932 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4933 if (holder->is_initialized()) {
4934 return false;
4935 }
4936 if (holder->is_being_initialized()) {
4937 if (accessing_method->holder() == holder) {
4938 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4939 // <init>, or a static method. In all those cases, there was an initialization
4940 // barrier on the holder klass passed.
4941 if (accessing_method->is_class_initializer() ||
4942 accessing_method->is_object_constructor() ||
4943 accessing_method->is_static()) {
4944 return false;
4945 }
4946 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4947 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4948 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4949 // child class can become fully initialized while its parent class is still being initialized.
4950 if (accessing_method->is_class_initializer()) {
4951 return false;
4952 }
4953 }
4954 ciMethod* root = method(); // the root method of compilation
4955 if (root != accessing_method) {
4956 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4957 }
4958 }
4959 return true;
4960 }
4961
4962 #ifndef PRODUCT
4963 //------------------------------verify_bidirectional_edges---------------------
4964 // For each input edge to a node (ie - for each Use-Def edge), verify that
4965 // there is a corresponding Def-Use edge.
4966 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4967 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4968 uint stack_size = live_nodes() >> 4;
4969 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4970 if (root_and_safepoints != nullptr) {
4971 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4972 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4973 Node* root_or_safepoint = root_and_safepoints->at(i);
4974 // If the node is a safepoint, let's check if it still has a control input
4975 // Lack of control input signifies that this node was killed by CCP or
4976 // recursively by remove_globally_dead_node and it shouldn't be a starting
4977 // point.
4978 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4979 nstack.push(root_or_safepoint);
4980 }
4981 }
4982 } else {
4983 nstack.push(_root);
4984 }
4985
4986 while (nstack.size() > 0) {
4987 Node* n = nstack.pop();
4988 if (visited.member(n)) {
4989 continue;
4990 }
4991 visited.push(n);
4992
4993 // Walk over all input edges, checking for correspondence
4994 uint length = n->len();
4995 for (uint i = 0; i < length; i++) {
4996 Node* in = n->in(i);
4997 if (in != nullptr && !visited.member(in)) {
4998 nstack.push(in); // Put it on stack
4999 }
5000 if (in != nullptr && !in->is_top()) {
5001 // Count instances of `next`
5002 int cnt = 0;
5003 for (uint idx = 0; idx < in->_outcnt; idx++) {
5004 if (in->_out[idx] == n) {
5005 cnt++;
5006 }
5007 }
5008 assert(cnt > 0, "Failed to find Def-Use edge.");
5009 // Check for duplicate edges
5010 // walk the input array downcounting the input edges to n
5011 for (uint j = 0; j < length; j++) {
5012 if (n->in(j) == in) {
5013 cnt--;
5014 }
5015 }
5016 assert(cnt == 0, "Mismatched edge count.");
5017 } else if (in == nullptr) {
5018 assert(i == 0 || i >= n->req() ||
5019 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
5020 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
5021 (n->is_Unlock() && i == (n->req() - 1)) ||
5022 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
5023 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
5024 } else {
5025 assert(in->is_top(), "sanity");
5026 // Nothing to check.
5027 }
5028 }
5029 }
5030 }
5031
5032 //------------------------------verify_graph_edges---------------------------
5033 // Walk the Graph and verify that there is a one-to-one correspondence
5034 // between Use-Def edges and Def-Use edges in the graph.
5035 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
5036 if (VerifyGraphEdges) {
5037 Unique_Node_List visited;
5038
5039 // Call graph walk to check edges
5040 verify_bidirectional_edges(visited, root_and_safepoints);
5041 if (no_dead_code) {
5042 // Now make sure that no visited node is used by an unvisited node.
5043 bool dead_nodes = false;
5044 Unique_Node_List checked;
5045 while (visited.size() > 0) {
5046 Node* n = visited.pop();
5047 checked.push(n);
5048 for (uint i = 0; i < n->outcnt(); i++) {
5049 Node* use = n->raw_out(i);
5050 if (checked.member(use)) continue; // already checked
5051 if (visited.member(use)) continue; // already in the graph
5052 if (use->is_Con()) continue; // a dead ConNode is OK
5053 // At this point, we have found a dead node which is DU-reachable.
5054 if (!dead_nodes) {
5055 tty->print_cr("*** Dead nodes reachable via DU edges:");
5056 dead_nodes = true;
5057 }
5058 use->dump(2);
5059 tty->print_cr("---");
5060 checked.push(use); // No repeats; pretend it is now checked.
5061 }
5062 }
5063 assert(!dead_nodes, "using nodes must be reachable from root");
5064 }
5065 }
5066 }
5067 #endif
5068
5069 // The Compile object keeps track of failure reasons separately from the ciEnv.
5070 // This is required because there is not quite a 1-1 relation between the
5071 // ciEnv and its compilation task and the Compile object. Note that one
5072 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
5073 // to backtrack and retry without subsuming loads. Other than this backtracking
5074 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
5075 // by the logic in C2Compiler.
5076 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
5077 if (log() != nullptr) {
5078 log()->elem("failure reason='%s' phase='compile'", reason);
5079 }
5080 if (_failure_reason.get() == nullptr) {
5081 // Record the first failure reason.
5082 _failure_reason.set(reason);
5083 if (CaptureBailoutInformation) {
5084 _first_failure_details = new CompilationFailureInfo(reason);
5085 }
5086 } else {
5087 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
5088 }
5089
5090 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
5091 C->print_method(PHASE_FAILURE, 1);
5092 }
5093 _root = nullptr; // flush the graph, too
5094 }
5095
5096 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
5097 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
5098 _compile(Compile::current()),
5099 _log(nullptr),
5100 _dolog(CITimeVerbose)
5101 {
5102 assert(_compile != nullptr, "sanity check");
5103 assert(id != PhaseTraceId::_t_none, "Don't use none");
5104 if (_dolog) {
5105 _log = _compile->log();
5106 }
5107 if (_log != nullptr) {
5108 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
5109 _log->stamp();
5110 _log->end_head();
5111 }
5112
5113 // Inform memory statistic, if enabled
5114 if (CompilationMemoryStatistic::enabled()) {
5115 CompilationMemoryStatistic::on_phase_start((int)id, name);
5116 }
5117 }
5118
5119 Compile::TracePhase::TracePhase(PhaseTraceId id)
5120 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
5121
5122 Compile::TracePhase::~TracePhase() {
5123
5124 // Inform memory statistic, if enabled
5125 if (CompilationMemoryStatistic::enabled()) {
5126 CompilationMemoryStatistic::on_phase_end();
5127 }
5128
5129 if (_compile->failing_internal()) {
5130 if (_log != nullptr) {
5131 _log->done("phase");
5132 }
5133 return; // timing code, not stressing bailouts.
5134 }
5135 #ifdef ASSERT
5136 if (PrintIdealNodeCount) {
5137 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
5138 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
5139 }
5140
5141 if (VerifyIdealNodeCount) {
5142 _compile->print_missing_nodes();
5143 }
5144 #endif
5145
5146 if (_log != nullptr) {
5147 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
5148 }
5149 }
5150
5151 //----------------------------static_subtype_check-----------------------------
5152 // Shortcut important common cases when superklass is exact:
5153 // (0) superklass is java.lang.Object (can occur in reflective code)
5154 // (1) subklass is already limited to a subtype of superklass => always ok
5155 // (2) subklass does not overlap with superklass => always fail
5156 // (3) superklass has NO subtypes and we can check with a simple compare.
5157 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
5158 if (skip) {
5159 return SSC_full_test; // Let caller generate the general case.
5160 }
5161
5162 if (subk->is_java_subtype_of(superk)) {
5163 return SSC_always_true; // (0) and (1) this test cannot fail
5164 }
5165
5166 if (!subk->maybe_java_subtype_of(superk)) {
5167 return SSC_always_false; // (2) true path dead; no dynamic test needed
5168 }
5169
5170 const Type* superelem = superk;
5171 if (superk->isa_aryklassptr()) {
5172 int ignored;
5173 superelem = superk->is_aryklassptr()->base_element_type(ignored);
5174
5175 // Do not fold the subtype check to an array klass pointer comparison for null-able inline type arrays
5176 // because null-free [LMyValue <: null-able [LMyValue but the klasses are different. Perform a full test.
5177 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
5178 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
5179 return SSC_full_test;
5180 }
5181 }
5182
5183 if (superelem->isa_instklassptr()) {
5184 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
5185 if (!ik->has_subklass()) {
5186 if (!ik->is_final()) {
5187 // Add a dependency if there is a chance of a later subclass.
5188 dependencies()->assert_leaf_type(ik);
5189 }
5190 if (!superk->maybe_java_subtype_of(subk)) {
5191 return SSC_always_false;
5192 }
5193 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5194 }
5195 } else {
5196 // A primitive array type has no subtypes.
5197 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5198 }
5199
5200 return SSC_full_test;
5201 }
5202
5203 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
5204 #ifdef _LP64
5205 // The scaled index operand to AddP must be a clean 64-bit value.
5206 // Java allows a 32-bit int to be incremented to a negative
5207 // value, which appears in a 64-bit register as a large
5208 // positive number. Using that large positive number as an
5209 // operand in pointer arithmetic has bad consequences.
5210 // On the other hand, 32-bit overflow is rare, and the possibility
5211 // can often be excluded, if we annotate the ConvI2L node with
5212 // a type assertion that its value is known to be a small positive
5213 // number. (The prior range check has ensured this.)
5214 // This assertion is used by ConvI2LNode::Ideal.
5215 int index_max = max_jint - 1; // array size is max_jint, index is one less
5216 if (sizetype != nullptr && sizetype->_hi > 0) {
5217 index_max = sizetype->_hi - 1;
5218 }
5219 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
5220 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
5221 #endif
5222 return idx;
5223 }
5224
5225 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
5226 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
5227 if (ctrl != nullptr) {
5228 // Express control dependency by a CastII node with a narrow type.
5229 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
5230 // node from floating above the range check during loop optimizations. Otherwise, the
5231 // ConvI2L node may be eliminated independently of the range check, causing the data path
5232 // to become TOP while the control path is still there (although it's unreachable).
5233 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
5234 value = phase->transform(value);
5235 }
5236 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
5237 return phase->transform(new ConvI2LNode(value, ltype));
5238 }
5239
5240 void Compile::dump_print_inlining() {
5241 inline_printer()->print_on(tty);
5242 }
5243
5244 void Compile::log_late_inline(CallGenerator* cg) {
5245 if (log() != nullptr) {
5246 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5247 cg->unique_id());
5248 JVMState* p = cg->call_node()->jvms();
5249 while (p != nullptr) {
5250 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5251 p = p->caller();
5252 }
5253 log()->tail("late_inline");
5254 }
5255 }
5256
5257 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5258 log_late_inline(cg);
5259 if (log() != nullptr) {
5260 log()->inline_fail(msg);
5261 }
5262 }
5263
5264 void Compile::log_inline_id(CallGenerator* cg) {
5265 if (log() != nullptr) {
5266 // The LogCompilation tool needs a unique way to identify late
5267 // inline call sites. This id must be unique for this call site in
5268 // this compilation. Try to have it unique across compilations as
5269 // well because it can be convenient when grepping through the log
5270 // file.
5271 // Distinguish OSR compilations from others in case CICountOSR is
5272 // on.
5273 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5274 cg->set_unique_id(id);
5275 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5276 }
5277 }
5278
5279 void Compile::log_inline_failure(const char* msg) {
5280 if (C->log() != nullptr) {
5281 C->log()->inline_fail(msg);
5282 }
5283 }
5284
5285
5286 // Dump inlining replay data to the stream.
5287 // Don't change thread state and acquire any locks.
5288 void Compile::dump_inline_data(outputStream* out) {
5289 InlineTree* inl_tree = ilt();
5290 if (inl_tree != nullptr) {
5291 out->print(" inline %d", inl_tree->count());
5292 inl_tree->dump_replay_data(out);
5293 }
5294 }
5295
5296 void Compile::dump_inline_data_reduced(outputStream* out) {
5297 assert(ReplayReduce, "");
5298
5299 InlineTree* inl_tree = ilt();
5300 if (inl_tree == nullptr) {
5301 return;
5302 }
5303 // Enable iterative replay file reduction
5304 // Output "compile" lines for depth 1 subtrees,
5305 // simulating that those trees were compiled
5306 // instead of inlined.
5307 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5308 InlineTree* sub = inl_tree->subtrees().at(i);
5309 if (sub->inline_level() != 1) {
5310 continue;
5311 }
5312
5313 ciMethod* method = sub->method();
5314 int entry_bci = -1;
5315 int comp_level = env()->task()->comp_level();
5316 out->print("compile ");
5317 method->dump_name_as_ascii(out);
5318 out->print(" %d %d", entry_bci, comp_level);
5319 out->print(" inline %d", sub->count());
5320 sub->dump_replay_data(out, -1);
5321 out->cr();
5322 }
5323 }
5324
5325 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5326 if (n1->Opcode() < n2->Opcode()) return -1;
5327 else if (n1->Opcode() > n2->Opcode()) return 1;
5328
5329 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5330 for (uint i = 1; i < n1->req(); i++) {
5331 if (n1->in(i) < n2->in(i)) return -1;
5332 else if (n1->in(i) > n2->in(i)) return 1;
5333 }
5334
5335 return 0;
5336 }
5337
5338 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5339 Node* n1 = *n1p;
5340 Node* n2 = *n2p;
5341
5342 return cmp_expensive_nodes(n1, n2);
5343 }
5344
5345 void Compile::sort_expensive_nodes() {
5346 if (!expensive_nodes_sorted()) {
5347 _expensive_nodes.sort(cmp_expensive_nodes);
5348 }
5349 }
5350
5351 bool Compile::expensive_nodes_sorted() const {
5352 for (int i = 1; i < _expensive_nodes.length(); i++) {
5353 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5354 return false;
5355 }
5356 }
5357 return true;
5358 }
5359
5360 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5361 if (_expensive_nodes.length() == 0) {
5362 return false;
5363 }
5364
5365 assert(OptimizeExpensiveOps, "optimization off?");
5366
5367 // Take this opportunity to remove dead nodes from the list
5368 int j = 0;
5369 for (int i = 0; i < _expensive_nodes.length(); i++) {
5370 Node* n = _expensive_nodes.at(i);
5371 if (!n->is_unreachable(igvn)) {
5372 assert(n->is_expensive(), "should be expensive");
5373 _expensive_nodes.at_put(j, n);
5374 j++;
5375 }
5376 }
5377 _expensive_nodes.trunc_to(j);
5378
5379 // Then sort the list so that similar nodes are next to each other
5380 // and check for at least two nodes of identical kind with same data
5381 // inputs.
5382 sort_expensive_nodes();
5383
5384 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5385 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5386 return true;
5387 }
5388 }
5389
5390 return false;
5391 }
5392
5393 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5394 if (_expensive_nodes.length() == 0) {
5395 return;
5396 }
5397
5398 assert(OptimizeExpensiveOps, "optimization off?");
5399
5400 // Sort to bring similar nodes next to each other and clear the
5401 // control input of nodes for which there's only a single copy.
5402 sort_expensive_nodes();
5403
5404 int j = 0;
5405 int identical = 0;
5406 int i = 0;
5407 bool modified = false;
5408 for (; i < _expensive_nodes.length()-1; i++) {
5409 assert(j <= i, "can't write beyond current index");
5410 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5411 identical++;
5412 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5413 continue;
5414 }
5415 if (identical > 0) {
5416 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5417 identical = 0;
5418 } else {
5419 Node* n = _expensive_nodes.at(i);
5420 igvn.replace_input_of(n, 0, nullptr);
5421 igvn.hash_insert(n);
5422 modified = true;
5423 }
5424 }
5425 if (identical > 0) {
5426 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5427 } else if (_expensive_nodes.length() >= 1) {
5428 Node* n = _expensive_nodes.at(i);
5429 igvn.replace_input_of(n, 0, nullptr);
5430 igvn.hash_insert(n);
5431 modified = true;
5432 }
5433 _expensive_nodes.trunc_to(j);
5434 if (modified) {
5435 igvn.optimize();
5436 }
5437 }
5438
5439 void Compile::add_expensive_node(Node * n) {
5440 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5441 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5442 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5443 if (OptimizeExpensiveOps) {
5444 _expensive_nodes.append(n);
5445 } else {
5446 // Clear control input and let IGVN optimize expensive nodes if
5447 // OptimizeExpensiveOps is off.
5448 n->set_req(0, nullptr);
5449 }
5450 }
5451
5452 /**
5453 * Track coarsened Lock and Unlock nodes.
5454 */
5455
5456 class Lock_List : public Node_List {
5457 uint _origin_cnt;
5458 public:
5459 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5460 uint origin_cnt() const { return _origin_cnt; }
5461 };
5462
5463 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5464 int length = locks.length();
5465 if (length > 0) {
5466 // Have to keep this list until locks elimination during Macro nodes elimination.
5467 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5468 AbstractLockNode* alock = locks.at(0);
5469 BoxLockNode* box = alock->box_node()->as_BoxLock();
5470 for (int i = 0; i < length; i++) {
5471 AbstractLockNode* lock = locks.at(i);
5472 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5473 locks_list->push(lock);
5474 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
5475 if (this_box != box) {
5476 // Locking regions (BoxLock) could be Unbalanced here:
5477 // - its coarsened locks were eliminated in earlier
5478 // macro nodes elimination followed by loop unroll
5479 // - it is OSR locking region (no Lock node)
5480 // Preserve Unbalanced status in such cases.
5481 if (!this_box->is_unbalanced()) {
5482 this_box->set_coarsened();
5483 }
5484 if (!box->is_unbalanced()) {
5485 box->set_coarsened();
5486 }
5487 }
5488 }
5489 _coarsened_locks.append(locks_list);
5490 }
5491 }
5492
5493 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5494 int count = coarsened_count();
5495 for (int i = 0; i < count; i++) {
5496 Node_List* locks_list = _coarsened_locks.at(i);
5497 for (uint j = 0; j < locks_list->size(); j++) {
5498 Node* lock = locks_list->at(j);
5499 assert(lock->is_AbstractLock(), "sanity");
5500 if (!useful.member(lock)) {
5501 locks_list->yank(lock);
5502 }
5503 }
5504 }
5505 }
5506
5507 void Compile::remove_coarsened_lock(Node* n) {
5508 if (n->is_AbstractLock()) {
5509 int count = coarsened_count();
5510 for (int i = 0; i < count; i++) {
5511 Node_List* locks_list = _coarsened_locks.at(i);
5512 locks_list->yank(n);
5513 }
5514 }
5515 }
5516
5517 bool Compile::coarsened_locks_consistent() {
5518 int count = coarsened_count();
5519 for (int i = 0; i < count; i++) {
5520 bool unbalanced = false;
5521 bool modified = false; // track locks kind modifications
5522 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5523 uint size = locks_list->size();
5524 if (size == 0) {
5525 unbalanced = false; // All locks were eliminated - good
5526 } else if (size != locks_list->origin_cnt()) {
5527 unbalanced = true; // Some locks were removed from list
5528 } else {
5529 for (uint j = 0; j < size; j++) {
5530 Node* lock = locks_list->at(j);
5531 // All nodes in group should have the same state (modified or not)
5532 if (!lock->as_AbstractLock()->is_coarsened()) {
5533 if (j == 0) {
5534 // first on list was modified, the rest should be too for consistency
5535 modified = true;
5536 } else if (!modified) {
5537 // this lock was modified but previous locks on the list were not
5538 unbalanced = true;
5539 break;
5540 }
5541 } else if (modified) {
5542 // previous locks on list were modified but not this lock
5543 unbalanced = true;
5544 break;
5545 }
5546 }
5547 }
5548 if (unbalanced) {
5549 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5550 #ifdef ASSERT
5551 if (PrintEliminateLocks) {
5552 tty->print_cr("=== unbalanced coarsened locks ===");
5553 for (uint l = 0; l < size; l++) {
5554 locks_list->at(l)->dump();
5555 }
5556 }
5557 #endif
5558 record_failure(C2Compiler::retry_no_locks_coarsening());
5559 return false;
5560 }
5561 }
5562 return true;
5563 }
5564
5565 // Mark locking regions (identified by BoxLockNode) as unbalanced if
5566 // locks coarsening optimization removed Lock/Unlock nodes from them.
5567 // Such regions become unbalanced because coarsening only removes part
5568 // of Lock/Unlock nodes in region. As result we can't execute other
5569 // locks elimination optimizations which assume all code paths have
5570 // corresponding pair of Lock/Unlock nodes - they are balanced.
5571 void Compile::mark_unbalanced_boxes() const {
5572 int count = coarsened_count();
5573 for (int i = 0; i < count; i++) {
5574 Node_List* locks_list = _coarsened_locks.at(i);
5575 uint size = locks_list->size();
5576 if (size > 0) {
5577 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
5578 BoxLockNode* box = alock->box_node()->as_BoxLock();
5579 if (alock->is_coarsened()) {
5580 // coarsened_locks_consistent(), which is called before this method, verifies
5581 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
5582 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5583 for (uint j = 1; j < size; j++) {
5584 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
5585 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
5586 if (box != this_box) {
5587 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5588 box->set_unbalanced();
5589 this_box->set_unbalanced();
5590 }
5591 }
5592 }
5593 }
5594 }
5595 }
5596
5597 /**
5598 * Remove the speculative part of types and clean up the graph
5599 */
5600 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5601 if (UseTypeSpeculation) {
5602 Unique_Node_List worklist;
5603 worklist.push(root());
5604 int modified = 0;
5605 // Go over all type nodes that carry a speculative type, drop the
5606 // speculative part of the type and enqueue the node for an igvn
5607 // which may optimize it out.
5608 for (uint next = 0; next < worklist.size(); ++next) {
5609 Node *n = worklist.at(next);
5610 if (n->is_Type()) {
5611 TypeNode* tn = n->as_Type();
5612 const Type* t = tn->type();
5613 const Type* t_no_spec = t->remove_speculative();
5614 if (t_no_spec != t) {
5615 bool in_hash = igvn.hash_delete(n);
5616 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5617 tn->set_type(t_no_spec);
5618 igvn.hash_insert(n);
5619 igvn._worklist.push(n); // give it a chance to go away
5620 modified++;
5621 }
5622 }
5623 // Iterate over outs - endless loops is unreachable from below
5624 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5625 Node *m = n->fast_out(i);
5626 if (not_a_node(m)) {
5627 continue;
5628 }
5629 worklist.push(m);
5630 }
5631 }
5632 // Drop the speculative part of all types in the igvn's type table
5633 igvn.remove_speculative_types();
5634 if (modified > 0) {
5635 igvn.optimize();
5636 if (failing()) return;
5637 }
5638 #ifdef ASSERT
5639 // Verify that after the IGVN is over no speculative type has resurfaced
5640 worklist.clear();
5641 worklist.push(root());
5642 for (uint next = 0; next < worklist.size(); ++next) {
5643 Node *n = worklist.at(next);
5644 const Type* t = igvn.type_or_null(n);
5645 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5646 if (n->is_Type()) {
5647 t = n->as_Type()->type();
5648 assert(t == t->remove_speculative(), "no more speculative types");
5649 }
5650 // Iterate over outs - endless loops is unreachable from below
5651 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5652 Node *m = n->fast_out(i);
5653 if (not_a_node(m)) {
5654 continue;
5655 }
5656 worklist.push(m);
5657 }
5658 }
5659 igvn.check_no_speculative_types();
5660 #endif
5661 }
5662 }
5663
5664 // Auxiliary methods to support randomized stressing/fuzzing.
5665
5666 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5667 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5668 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5669 FLAG_SET_ERGO(StressSeed, _stress_seed);
5670 } else {
5671 _stress_seed = StressSeed;
5672 }
5673 if (_log != nullptr) {
5674 _log->elem("stress_test seed='%u'", _stress_seed);
5675 }
5676 }
5677
5678 int Compile::random() {
5679 _stress_seed = os::next_random(_stress_seed);
5680 return static_cast<int>(_stress_seed);
5681 }
5682
5683 // This method can be called the arbitrary number of times, with current count
5684 // as the argument. The logic allows selecting a single candidate from the
5685 // running list of candidates as follows:
5686 // int count = 0;
5687 // Cand* selected = null;
5688 // while(cand = cand->next()) {
5689 // if (randomized_select(++count)) {
5690 // selected = cand;
5691 // }
5692 // }
5693 //
5694 // Including count equalizes the chances any candidate is "selected".
5695 // This is useful when we don't have the complete list of candidates to choose
5696 // from uniformly. In this case, we need to adjust the randomicity of the
5697 // selection, or else we will end up biasing the selection towards the latter
5698 // candidates.
5699 //
5700 // Quick back-envelope calculation shows that for the list of n candidates
5701 // the equal probability for the candidate to persist as "best" can be
5702 // achieved by replacing it with "next" k-th candidate with the probability
5703 // of 1/k. It can be easily shown that by the end of the run, the
5704 // probability for any candidate is converged to 1/n, thus giving the
5705 // uniform distribution among all the candidates.
5706 //
5707 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5708 #define RANDOMIZED_DOMAIN_POW 29
5709 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5710 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5711 bool Compile::randomized_select(int count) {
5712 assert(count > 0, "only positive");
5713 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5714 }
5715
5716 #ifdef ASSERT
5717 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5718 bool Compile::fail_randomly() {
5719 if ((random() % StressBailoutMean) != 0) {
5720 return false;
5721 }
5722 record_failure("StressBailout");
5723 return true;
5724 }
5725
5726 bool Compile::failure_is_artificial() {
5727 return C->failure_reason_is("StressBailout");
5728 }
5729 #endif
5730
5731 CloneMap& Compile::clone_map() { return _clone_map; }
5732 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5733
5734 void NodeCloneInfo::dump_on(outputStream* st) const {
5735 st->print(" {%d:%d} ", idx(), gen());
5736 }
5737
5738 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5739 uint64_t val = value(old->_idx);
5740 NodeCloneInfo cio(val);
5741 assert(val != 0, "old node should be in the map");
5742 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5743 insert(nnn->_idx, cin.get());
5744 #ifndef PRODUCT
5745 if (is_debug()) {
5746 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5747 }
5748 #endif
5749 }
5750
5751 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5752 NodeCloneInfo cio(value(old->_idx));
5753 if (cio.get() == 0) {
5754 cio.set(old->_idx, 0);
5755 insert(old->_idx, cio.get());
5756 #ifndef PRODUCT
5757 if (is_debug()) {
5758 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5759 }
5760 #endif
5761 }
5762 clone(old, nnn, gen);
5763 }
5764
5765 int CloneMap::max_gen() const {
5766 int g = 0;
5767 DictI di(_dict);
5768 for(; di.test(); ++di) {
5769 int t = gen(di._key);
5770 if (g < t) {
5771 g = t;
5772 #ifndef PRODUCT
5773 if (is_debug()) {
5774 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5775 }
5776 #endif
5777 }
5778 }
5779 return g;
5780 }
5781
5782 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5783 uint64_t val = value(key);
5784 if (val != 0) {
5785 NodeCloneInfo ni(val);
5786 ni.dump_on(st);
5787 }
5788 }
5789
5790 void Compile::shuffle_macro_nodes() {
5791 shuffle_array(*C, _macro_nodes);
5792 }
5793
5794 // Move Allocate nodes to the start of the list
5795 void Compile::sort_macro_nodes() {
5796 int count = macro_count();
5797 int allocates = 0;
5798 for (int i = 0; i < count; i++) {
5799 Node* n = macro_node(i);
5800 if (n->is_Allocate()) {
5801 if (i != allocates) {
5802 Node* tmp = macro_node(allocates);
5803 _macro_nodes.at_put(allocates, n);
5804 _macro_nodes.at_put(i, tmp);
5805 }
5806 allocates++;
5807 }
5808 }
5809 }
5810
5811 void Compile::print_method(CompilerPhaseType compile_phase, int level, Node* n) {
5812 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5813 EventCompilerPhase event(UNTIMED);
5814 if (event.should_commit()) {
5815 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, compile_phase, C->_compile_id, level);
5816 }
5817 #ifndef PRODUCT
5818 ResourceMark rm;
5819 stringStream ss;
5820 ss.print_raw(CompilerPhaseTypeHelper::to_description(compile_phase));
5821 int iter = ++_igv_phase_iter[compile_phase];
5822 if (iter > 1) {
5823 ss.print(" %d", iter);
5824 }
5825 if (n != nullptr) {
5826 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5827 if (n->is_Call()) {
5828 CallNode* call = n->as_Call();
5829 if (call->_name != nullptr) {
5830 // E.g. uncommon traps etc.
5831 ss.print(" - %s", call->_name);
5832 } else if (call->is_CallJava()) {
5833 CallJavaNode* call_java = call->as_CallJava();
5834 if (call_java->method() != nullptr) {
5835 ss.print(" -");
5836 call_java->method()->print_short_name(&ss);
5837 }
5838 }
5839 }
5840 }
5841
5842 const char* name = ss.as_string();
5843 if (should_print_igv(level)) {
5844 _igv_printer->print_graph(name);
5845 }
5846 if (should_print_phase(level)) {
5847 print_phase(name);
5848 }
5849 if (should_print_ideal_phase(compile_phase)) {
5850 print_ideal_ir(CompilerPhaseTypeHelper::to_name(compile_phase));
5851 }
5852 #endif
5853 C->_latest_stage_start_counter.stamp();
5854 }
5855
5856 // Only used from CompileWrapper
5857 void Compile::begin_method() {
5858 #ifndef PRODUCT
5859 if (_method != nullptr && should_print_igv(1)) {
5860 _igv_printer->begin_method();
5861 }
5862 #endif
5863 C->_latest_stage_start_counter.stamp();
5864 }
5865
5866 // Only used from CompileWrapper
5867 void Compile::end_method() {
5868 EventCompilerPhase event(UNTIMED);
5869 if (event.should_commit()) {
5870 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5871 }
5872
5873 #ifndef PRODUCT
5874 if (_method != nullptr && should_print_igv(1)) {
5875 _igv_printer->end_method();
5876 }
5877 #endif
5878 }
5879
5880 #ifndef PRODUCT
5881 bool Compile::should_print_phase(const int level) const {
5882 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5883 _method != nullptr; // Do not print phases for stubs.
5884 }
5885
5886 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5887 return _directive->should_print_ideal_phase(cpt);
5888 }
5889
5890 void Compile::init_igv() {
5891 if (_igv_printer == nullptr) {
5892 _igv_printer = IdealGraphPrinter::printer();
5893 _igv_printer->set_compile(this);
5894 }
5895 }
5896
5897 bool Compile::should_print_igv(const int level) {
5898 PRODUCT_RETURN_(return false;);
5899
5900 if (PrintIdealGraphLevel < 0) { // disabled by the user
5901 return false;
5902 }
5903
5904 bool need = directive()->IGVPrintLevelOption >= level;
5905 if (need) {
5906 Compile::init_igv();
5907 }
5908 return need;
5909 }
5910
5911 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5912 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5913
5914 // Called from debugger. Prints method to the default file with the default phase name.
5915 // This works regardless of any Ideal Graph Visualizer flags set or not.
5916 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5917 void igv_print(void* sp, void* fp, void* pc) {
5918 frame fr(sp, fp, pc);
5919 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5920 }
5921
5922 // Same as igv_print() above but with a specified phase name.
5923 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5924 frame fr(sp, fp, pc);
5925 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5926 }
5927
5928 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5929 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5930 // This works regardless of any Ideal Graph Visualizer flags set or not.
5931 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5932 void igv_print(bool network, void* sp, void* fp, void* pc) {
5933 frame fr(sp, fp, pc);
5934 if (network) {
5935 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5936 } else {
5937 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5938 }
5939 }
5940
5941 // Same as igv_print(bool network, ...) above but with a specified phase name.
5942 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5943 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5944 frame fr(sp, fp, pc);
5945 if (network) {
5946 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5947 } else {
5948 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5949 }
5950 }
5951
5952 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5953 void igv_print_default() {
5954 Compile::current()->print_method(PHASE_DEBUG, 0);
5955 }
5956
5957 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5958 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5959 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5960 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5961 void igv_append(void* sp, void* fp, void* pc) {
5962 frame fr(sp, fp, pc);
5963 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5964 }
5965
5966 // Same as igv_append(...) above but with a specified phase name.
5967 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5968 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5969 frame fr(sp, fp, pc);
5970 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5971 }
5972
5973 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5974 const char* file_name = "custom_debug.xml";
5975 if (_debug_file_printer == nullptr) {
5976 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5977 } else {
5978 _debug_file_printer->update_compiled_method(C->method());
5979 }
5980 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5981 _debug_file_printer->print_graph(phase_name, fr);
5982 }
5983
5984 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5985 ResourceMark rm;
5986 GrowableArray<const Node*> empty_list;
5987 igv_print_graph_to_network(phase_name, empty_list, fr);
5988 }
5989
5990 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5991 if (_debug_network_printer == nullptr) {
5992 _debug_network_printer = new IdealGraphPrinter(C);
5993 } else {
5994 _debug_network_printer->update_compiled_method(C->method());
5995 }
5996 tty->print_cr("Method printed over network stream to IGV");
5997 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5998 }
5999 #endif // !PRODUCT
6000
6001 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
6002 if (type != nullptr && phase->type(value)->higher_equal(type)) {
6003 return value;
6004 }
6005 Node* result = nullptr;
6006 if (bt == T_BYTE) {
6007 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
6008 result = new RShiftINode(result, phase->intcon(24));
6009 } else if (bt == T_BOOLEAN) {
6010 result = new AndINode(value, phase->intcon(0xFF));
6011 } else if (bt == T_CHAR) {
6012 result = new AndINode(value,phase->intcon(0xFFFF));
6013 } else if (bt == T_FLOAT) {
6014 result = new MoveI2FNode(value);
6015 } else {
6016 assert(bt == T_SHORT, "unexpected narrow type");
6017 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
6018 result = new RShiftINode(result, phase->intcon(16));
6019 }
6020 if (transform_res) {
6021 result = phase->transform(result);
6022 }
6023 return result;
6024 }
6025
6026 void Compile::record_method_not_compilable_oom() {
6027 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
6028 }
6029
6030 #ifndef PRODUCT
6031 // Collects all the control inputs from nodes on the worklist and from their data dependencies
6032 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
6033 // Follow non-control edges until we reach CFG nodes
6034 for (uint i = 0; i < worklist.size(); i++) {
6035 const Node* n = worklist.at(i);
6036 for (uint j = 0; j < n->req(); j++) {
6037 Node* in = n->in(j);
6038 if (in == nullptr || in->is_Root()) {
6039 continue;
6040 }
6041 if (in->is_CFG()) {
6042 if (in->is_Call()) {
6043 // The return value of a call is only available if the call did not result in an exception
6044 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
6045 if (control_proj_use->is_Catch()) {
6046 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
6047 candidates.push(fall_through);
6048 continue;
6049 }
6050 }
6051
6052 if (in->is_Multi()) {
6053 // We got here by following data inputs so we should only have one control use
6054 // (no IfNode, etc)
6055 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
6056 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
6057 } else {
6058 candidates.push(in);
6059 }
6060 } else {
6061 worklist.push(in);
6062 }
6063 }
6064 }
6065 }
6066
6067 // Returns the candidate node that is a descendant to all the other candidates
6068 static Node* pick_control(Unique_Node_List& candidates) {
6069 Unique_Node_List worklist;
6070 worklist.copy(candidates);
6071
6072 // Traverse backwards through the CFG
6073 for (uint i = 0; i < worklist.size(); i++) {
6074 const Node* n = worklist.at(i);
6075 if (n->is_Root()) {
6076 continue;
6077 }
6078 for (uint j = 0; j < n->req(); j++) {
6079 // Skip backedge of loops to avoid cycles
6080 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
6081 continue;
6082 }
6083
6084 Node* pred = n->in(j);
6085 if (pred != nullptr && pred != n && pred->is_CFG()) {
6086 worklist.push(pred);
6087 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
6088 candidates.remove(pred);
6089 }
6090 }
6091 }
6092
6093 assert(candidates.size() == 1, "unexpected control flow");
6094 return candidates.at(0);
6095 }
6096
6097 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
6098 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
6099 call->init_req((*pos)++, parm);
6100 const BasicType bt = parm->bottom_type()->basic_type();
6101 if (bt == T_LONG || bt == T_DOUBLE) {
6102 call->init_req((*pos)++, half);
6103 }
6104 }
6105
6106 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
6107 Node* parm0, Node* parm1,
6108 Node* parm2, Node* parm3,
6109 Node* parm4, Node* parm5,
6110 Node* parm6) const {
6111 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
6112 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
6113 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
6114
6115 // find the most suitable control input
6116 Unique_Node_List worklist, candidates;
6117 if (parm0 != nullptr) { worklist.push(parm0);
6118 if (parm1 != nullptr) { worklist.push(parm1);
6119 if (parm2 != nullptr) { worklist.push(parm2);
6120 if (parm3 != nullptr) { worklist.push(parm3);
6121 if (parm4 != nullptr) { worklist.push(parm4);
6122 if (parm5 != nullptr) { worklist.push(parm5);
6123 if (parm6 != nullptr) { worklist.push(parm6);
6124 /* close each nested if ===> */ } } } } } } }
6125 find_candidate_control_inputs(worklist, candidates);
6126 Node* control = nullptr;
6127 if (candidates.size() == 0) {
6128 control = C->start()->proj_out(TypeFunc::Control);
6129 } else {
6130 control = pick_control(candidates);
6131 }
6132
6133 // find all the previous users of the control we picked
6134 GrowableArray<Node*> users_of_control;
6135 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
6136 Node* use = control->fast_out(i);
6137 if (use->is_CFG() && use != control) {
6138 users_of_control.push(use);
6139 }
6140 }
6141
6142 // we do not actually care about IO and memory as it uses neither
6143 call->init_req(TypeFunc::Control, control);
6144 call->init_req(TypeFunc::I_O, top());
6145 call->init_req(TypeFunc::Memory, top());
6146 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
6147 call->init_req(TypeFunc::ReturnAdr, top());
6148
6149 int pos = TypeFunc::Parms;
6150 call->init_req(pos++, str_node);
6151 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
6152 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
6153 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
6154 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
6155 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
6156 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
6157 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
6158 /* close each nested if ===> */ } } } } } } }
6159 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
6160
6161 call = gvn->transform(call);
6162 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
6163
6164 // rewire previous users to have the new call as control instead
6165 PhaseIterGVN* igvn = gvn->is_IterGVN();
6166 for (int i = 0; i < users_of_control.length(); i++) {
6167 Node* use = users_of_control.at(i);
6168 for (uint j = 0; j < use->req(); j++) {
6169 if (use->in(j) == control) {
6170 if (igvn != nullptr) {
6171 igvn->replace_input_of(use, j, call_control_proj);
6172 } else {
6173 gvn->hash_delete(use);
6174 use->set_req(j, call_control_proj);
6175 gvn->hash_insert(use);
6176 }
6177 }
6178 }
6179 }
6180
6181 return call;
6182 }
6183 #endif // !PRODUCT