1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/macroAssembler.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "ci/ciReplay.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "code/aotCodeCache.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compiler_globals.hpp"
37 #include "compiler/compilerDefinitions.hpp"
38 #include "compiler/compilerOracle.hpp"
39 #include "compiler/disassembler.hpp"
40 #include "compiler/oopMap.hpp"
41 #include "gc/shared/barrierSet.hpp"
42 #include "gc/shared/c2/barrierSetC2.hpp"
43 #include "jfr/jfrEvents.hpp"
44 #include "jvm_io.h"
45 #include "memory/allocation.hpp"
46 #include "memory/arena.hpp"
47 #include "memory/resourceArea.hpp"
48 #include "opto/addnode.hpp"
49 #include "opto/block.hpp"
50 #include "opto/c2compiler.hpp"
51 #include "opto/callGenerator.hpp"
52 #include "opto/callnode.hpp"
53 #include "opto/castnode.hpp"
54 #include "opto/cfgnode.hpp"
55 #include "opto/chaitin.hpp"
56 #include "opto/compile.hpp"
57 #include "opto/connode.hpp"
58 #include "opto/convertnode.hpp"
59 #include "opto/divnode.hpp"
60 #include "opto/escape.hpp"
61 #include "opto/idealGraphPrinter.hpp"
62 #include "opto/inlinetypenode.hpp"
63 #include "opto/locknode.hpp"
64 #include "opto/loopnode.hpp"
65 #include "opto/machnode.hpp"
66 #include "opto/macro.hpp"
67 #include "opto/matcher.hpp"
68 #include "opto/mathexactnode.hpp"
69 #include "opto/memnode.hpp"
70 #include "opto/movenode.hpp"
71 #include "opto/mulnode.hpp"
72 #include "opto/narrowptrnode.hpp"
73 #include "opto/node.hpp"
74 #include "opto/opaquenode.hpp"
75 #include "opto/opcodes.hpp"
76 #include "opto/output.hpp"
77 #include "opto/parse.hpp"
78 #include "opto/phaseX.hpp"
79 #include "opto/rootnode.hpp"
80 #include "opto/runtime.hpp"
81 #include "opto/stringopts.hpp"
82 #include "opto/type.hpp"
83 #include "opto/vector.hpp"
84 #include "opto/vectornode.hpp"
85 #include "runtime/globals_extension.hpp"
86 #include "runtime/sharedRuntime.hpp"
87 #include "runtime/signature.hpp"
88 #include "runtime/stubRoutines.hpp"
89 #include "runtime/timer.hpp"
90 #include "utilities/align.hpp"
91 #include "utilities/copy.hpp"
92 #include "utilities/macros.hpp"
93 #include "utilities/resourceHash.hpp"
94
95 // -------------------- Compile::mach_constant_base_node -----------------------
96 // Constant table base node singleton.
97 MachConstantBaseNode* Compile::mach_constant_base_node() {
98 if (_mach_constant_base_node == nullptr) {
99 _mach_constant_base_node = new MachConstantBaseNode();
100 _mach_constant_base_node->add_req(C->root());
101 }
102 return _mach_constant_base_node;
103 }
104
105
106 /// Support for intrinsics.
107
108 // Return the index at which m must be inserted (or already exists).
109 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
110 class IntrinsicDescPair {
111 private:
112 ciMethod* _m;
113 bool _is_virtual;
114 public:
115 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
116 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
117 ciMethod* m= elt->method();
118 ciMethod* key_m = key->_m;
119 if (key_m < m) return -1;
120 else if (key_m > m) return 1;
121 else {
122 bool is_virtual = elt->is_virtual();
123 bool key_virtual = key->_is_virtual;
124 if (key_virtual < is_virtual) return -1;
125 else if (key_virtual > is_virtual) return 1;
126 else return 0;
127 }
128 }
129 };
130 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
131 #ifdef ASSERT
132 for (int i = 1; i < _intrinsics.length(); i++) {
133 CallGenerator* cg1 = _intrinsics.at(i-1);
134 CallGenerator* cg2 = _intrinsics.at(i);
135 assert(cg1->method() != cg2->method()
136 ? cg1->method() < cg2->method()
137 : cg1->is_virtual() < cg2->is_virtual(),
138 "compiler intrinsics list must stay sorted");
139 }
140 #endif
141 IntrinsicDescPair pair(m, is_virtual);
142 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
143 }
144
145 void Compile::register_intrinsic(CallGenerator* cg) {
146 bool found = false;
147 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
148 assert(!found, "registering twice");
149 _intrinsics.insert_before(index, cg);
150 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
151 }
152
153 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
154 assert(m->is_loaded(), "don't try this on unloaded methods");
155 if (_intrinsics.length() > 0) {
156 bool found = false;
157 int index = intrinsic_insertion_index(m, is_virtual, found);
158 if (found) {
159 return _intrinsics.at(index);
160 }
161 }
162 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
163 if (m->intrinsic_id() != vmIntrinsics::_none &&
164 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
165 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
166 if (cg != nullptr) {
167 // Save it for next time:
168 register_intrinsic(cg);
169 return cg;
170 } else {
171 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
172 }
173 }
174 return nullptr;
175 }
176
177 // Compile::make_vm_intrinsic is defined in library_call.cpp.
178
179 #ifndef PRODUCT
180 // statistics gathering...
181
182 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
183 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
184
185 inline int as_int(vmIntrinsics::ID id) {
186 return vmIntrinsics::as_int(id);
187 }
188
189 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
190 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
191 int oflags = _intrinsic_hist_flags[as_int(id)];
192 assert(flags != 0, "what happened?");
193 if (is_virtual) {
194 flags |= _intrinsic_virtual;
195 }
196 bool changed = (flags != oflags);
197 if ((flags & _intrinsic_worked) != 0) {
198 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
199 if (count == 1) {
200 changed = true; // first time
201 }
202 // increment the overall count also:
203 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
204 }
205 if (changed) {
206 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
207 // Something changed about the intrinsic's virtuality.
208 if ((flags & _intrinsic_virtual) != 0) {
209 // This is the first use of this intrinsic as a virtual call.
210 if (oflags != 0) {
211 // We already saw it as a non-virtual, so note both cases.
212 flags |= _intrinsic_both;
213 }
214 } else if ((oflags & _intrinsic_both) == 0) {
215 // This is the first use of this intrinsic as a non-virtual
216 flags |= _intrinsic_both;
217 }
218 }
219 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
220 }
221 // update the overall flags also:
222 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
223 return changed;
224 }
225
226 static char* format_flags(int flags, char* buf) {
227 buf[0] = 0;
228 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
229 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
230 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
231 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
232 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
233 if (buf[0] == 0) strcat(buf, ",");
234 assert(buf[0] == ',', "must be");
235 return &buf[1];
236 }
237
238 void Compile::print_intrinsic_statistics() {
239 char flagsbuf[100];
240 ttyLocker ttyl;
241 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
242 tty->print_cr("Compiler intrinsic usage:");
243 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
244 if (total == 0) total = 1; // avoid div0 in case of no successes
245 #define PRINT_STAT_LINE(name, c, f) \
246 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
247 for (auto id : EnumRange<vmIntrinsicID>{}) {
248 int flags = _intrinsic_hist_flags[as_int(id)];
249 juint count = _intrinsic_hist_count[as_int(id)];
250 if ((flags | count) != 0) {
251 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
252 }
253 }
254 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
255 if (xtty != nullptr) xtty->tail("statistics");
256 }
257
258 void Compile::print_statistics() {
259 { ttyLocker ttyl;
260 if (xtty != nullptr) xtty->head("statistics type='opto'");
261 Parse::print_statistics();
262 PhaseStringOpts::print_statistics();
263 PhaseCCP::print_statistics();
264 PhaseRegAlloc::print_statistics();
265 PhaseOutput::print_statistics();
266 PhasePeephole::print_statistics();
267 PhaseIdealLoop::print_statistics();
268 ConnectionGraph::print_statistics();
269 PhaseMacroExpand::print_statistics();
270 if (xtty != nullptr) xtty->tail("statistics");
271 }
272 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
273 // put this under its own <statistics> element.
274 print_intrinsic_statistics();
275 }
276 }
277 #endif //PRODUCT
278
279 void Compile::gvn_replace_by(Node* n, Node* nn) {
280 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
281 Node* use = n->last_out(i);
282 bool is_in_table = initial_gvn()->hash_delete(use);
283 uint uses_found = 0;
284 for (uint j = 0; j < use->len(); j++) {
285 if (use->in(j) == n) {
286 if (j < use->req())
287 use->set_req(j, nn);
288 else
289 use->set_prec(j, nn);
290 uses_found++;
291 }
292 }
293 if (is_in_table) {
294 // reinsert into table
295 initial_gvn()->hash_find_insert(use);
296 }
297 record_for_igvn(use);
298 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
299 i -= uses_found; // we deleted 1 or more copies of this edge
300 }
301 }
302
303
304 // Identify all nodes that are reachable from below, useful.
305 // Use breadth-first pass that records state in a Unique_Node_List,
306 // recursive traversal is slower.
307 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
308 int estimated_worklist_size = live_nodes();
309 useful.map( estimated_worklist_size, nullptr ); // preallocate space
310
311 // Initialize worklist
312 if (root() != nullptr) { useful.push(root()); }
313 // If 'top' is cached, declare it useful to preserve cached node
314 if (cached_top_node()) { useful.push(cached_top_node()); }
315
316 // Push all useful nodes onto the list, breadthfirst
317 for( uint next = 0; next < useful.size(); ++next ) {
318 assert( next < unique(), "Unique useful nodes < total nodes");
319 Node *n = useful.at(next);
320 uint max = n->len();
321 for( uint i = 0; i < max; ++i ) {
322 Node *m = n->in(i);
323 if (not_a_node(m)) continue;
324 useful.push(m);
325 }
326 }
327 }
328
329 // Update dead_node_list with any missing dead nodes using useful
330 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
331 void Compile::update_dead_node_list(Unique_Node_List &useful) {
332 uint max_idx = unique();
333 VectorSet& useful_node_set = useful.member_set();
334
335 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
336 // If node with index node_idx is not in useful set,
337 // mark it as dead in dead node list.
338 if (!useful_node_set.test(node_idx)) {
339 record_dead_node(node_idx);
340 }
341 }
342 }
343
344 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
345 int shift = 0;
346 for (int i = 0; i < inlines->length(); i++) {
347 CallGenerator* cg = inlines->at(i);
348 if (useful.member(cg->call_node())) {
349 if (shift > 0) {
350 inlines->at_put(i - shift, cg);
351 }
352 } else {
353 shift++; // skip over the dead element
354 }
355 }
356 if (shift > 0) {
357 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
358 }
359 }
360
361 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
362 assert(dead != nullptr && dead->is_Call(), "sanity");
363 int found = 0;
364 for (int i = 0; i < inlines->length(); i++) {
365 if (inlines->at(i)->call_node() == dead) {
366 inlines->remove_at(i);
367 found++;
368 NOT_DEBUG( break; ) // elements are unique, so exit early
369 }
370 }
371 assert(found <= 1, "not unique");
372 }
373
374 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
375 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
376 for (int i = node_list.length() - 1; i >= 0; i--) {
377 N* node = node_list.at(i);
378 if (!useful.member(node)) {
379 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
380 }
381 }
382 }
383
384 void Compile::remove_useless_node(Node* dead) {
385 remove_modified_node(dead);
386
387 // Constant node that has no out-edges and has only one in-edge from
388 // root is usually dead. However, sometimes reshaping walk makes
389 // it reachable by adding use edges. So, we will NOT count Con nodes
390 // as dead to be conservative about the dead node count at any
391 // given time.
392 if (!dead->is_Con()) {
393 record_dead_node(dead->_idx);
394 }
395 if (dead->is_macro()) {
396 remove_macro_node(dead);
397 }
398 if (dead->is_expensive()) {
399 remove_expensive_node(dead);
400 }
401 if (dead->is_OpaqueTemplateAssertionPredicate()) {
402 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
403 }
404 if (dead->is_ParsePredicate()) {
405 remove_parse_predicate(dead->as_ParsePredicate());
406 }
407 if (dead->for_post_loop_opts_igvn()) {
408 remove_from_post_loop_opts_igvn(dead);
409 }
410 if (dead->is_InlineType()) {
411 remove_inline_type(dead);
412 }
413 if (dead->for_merge_stores_igvn()) {
414 remove_from_merge_stores_igvn(dead);
415 }
416 if (dead->is_Call()) {
417 remove_useless_late_inlines( &_late_inlines, dead);
418 remove_useless_late_inlines( &_string_late_inlines, dead);
419 remove_useless_late_inlines( &_boxing_late_inlines, dead);
420 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
421
422 if (dead->is_CallStaticJava()) {
423 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
424 }
425 }
426 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
427 bs->unregister_potential_barrier_node(dead);
428 }
429
430 // Disconnect all useless nodes by disconnecting those at the boundary.
431 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
432 uint next = 0;
433 while (next < useful.size()) {
434 Node *n = useful.at(next++);
435 if (n->is_SafePoint()) {
436 // We're done with a parsing phase. Replaced nodes are not valid
437 // beyond that point.
438 n->as_SafePoint()->delete_replaced_nodes();
439 }
440 // Use raw traversal of out edges since this code removes out edges
441 int max = n->outcnt();
442 for (int j = 0; j < max; ++j) {
443 Node* child = n->raw_out(j);
444 if (!useful.member(child)) {
445 assert(!child->is_top() || child != top(),
446 "If top is cached in Compile object it is in useful list");
447 // Only need to remove this out-edge to the useless node
448 n->raw_del_out(j);
449 --j;
450 --max;
451 if (child->is_data_proj_of_pure_function(n)) {
452 worklist.push(n);
453 }
454 }
455 }
456 if (n->outcnt() == 1 && n->has_special_unique_user()) {
457 assert(useful.member(n->unique_out()), "do not push a useless node");
458 worklist.push(n->unique_out());
459 }
460 if (n->outcnt() == 0) {
461 worklist.push(n);
462 }
463 }
464
465 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
466 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
467 // Remove useless Template Assertion Predicate opaque nodes
468 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
469 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
470 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
471 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
472 #ifdef ASSERT
473 if (_modified_nodes != nullptr) {
474 _modified_nodes->remove_useless_nodes(useful.member_set());
475 }
476 #endif
477 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
478 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
479 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
480 #ifdef ASSERT
481 if (_modified_nodes != nullptr) {
482 _modified_nodes->remove_useless_nodes(useful.member_set());
483 }
484 #endif
485
486 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
487 bs->eliminate_useless_gc_barriers(useful, this);
488 // clean up the late inline lists
489 remove_useless_late_inlines( &_late_inlines, useful);
490 remove_useless_late_inlines( &_string_late_inlines, useful);
491 remove_useless_late_inlines( &_boxing_late_inlines, useful);
492 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
493 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
494 }
495
496 // ============================================================================
497 //------------------------------CompileWrapper---------------------------------
498 class CompileWrapper : public StackObj {
499 Compile *const _compile;
500 public:
501 CompileWrapper(Compile* compile);
502
503 ~CompileWrapper();
504 };
505
506 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
507 // the Compile* pointer is stored in the current ciEnv:
508 ciEnv* env = compile->env();
509 assert(env == ciEnv::current(), "must already be a ciEnv active");
510 assert(env->compiler_data() == nullptr, "compile already active?");
511 env->set_compiler_data(compile);
512 assert(compile == Compile::current(), "sanity");
513
514 compile->set_type_dict(nullptr);
515 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
516 compile->clone_map().set_clone_idx(0);
517 compile->set_type_last_size(0);
518 compile->set_last_tf(nullptr, nullptr);
519 compile->set_indexSet_arena(nullptr);
520 compile->set_indexSet_free_block_list(nullptr);
521 compile->init_type_arena();
522 Type::Initialize(compile);
523 _compile->begin_method();
524 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
525 }
526 CompileWrapper::~CompileWrapper() {
527 // simulate crash during compilation
528 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
529
530 _compile->end_method();
531 _compile->env()->set_compiler_data(nullptr);
532 }
533
534
535 //----------------------------print_compile_messages---------------------------
536 void Compile::print_compile_messages() {
537 #ifndef PRODUCT
538 // Check if recompiling
539 if (!subsume_loads() && PrintOpto) {
540 // Recompiling without allowing machine instructions to subsume loads
541 tty->print_cr("*********************************************************");
542 tty->print_cr("** Bailout: Recompile without subsuming loads **");
543 tty->print_cr("*********************************************************");
544 }
545 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
546 // Recompiling without escape analysis
547 tty->print_cr("*********************************************************");
548 tty->print_cr("** Bailout: Recompile without escape analysis **");
549 tty->print_cr("*********************************************************");
550 }
551 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
552 // Recompiling without iterative escape analysis
553 tty->print_cr("*********************************************************");
554 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
555 tty->print_cr("*********************************************************");
556 }
557 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
558 // Recompiling without reducing allocation merges
559 tty->print_cr("*********************************************************");
560 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
561 tty->print_cr("*********************************************************");
562 }
563 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
564 // Recompiling without boxing elimination
565 tty->print_cr("*********************************************************");
566 tty->print_cr("** Bailout: Recompile without boxing elimination **");
567 tty->print_cr("*********************************************************");
568 }
569 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
570 // Recompiling without locks coarsening
571 tty->print_cr("*********************************************************");
572 tty->print_cr("** Bailout: Recompile without locks coarsening **");
573 tty->print_cr("*********************************************************");
574 }
575 if (env()->break_at_compile()) {
576 // Open the debugger when compiling this method.
577 tty->print("### Breaking when compiling: ");
578 method()->print_short_name();
579 tty->cr();
580 BREAKPOINT;
581 }
582
583 if( PrintOpto ) {
584 if (is_osr_compilation()) {
585 tty->print("[OSR]%3d", _compile_id);
586 } else {
587 tty->print("%3d", _compile_id);
588 }
589 }
590 #endif
591 }
592
593 #ifndef PRODUCT
594 void Compile::print_phase(const char* phase_name) {
595 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name);
596 }
597
598 void Compile::print_ideal_ir(const char* phase_name) {
599 // keep the following output all in one block
600 // This output goes directly to the tty, not the compiler log.
601 // To enable tools to match it up with the compilation activity,
602 // be sure to tag this tty output with the compile ID.
603
604 // Node dumping can cause a safepoint, which can break the tty lock.
605 // Buffer all node dumps, so that all safepoints happen before we lock.
606 ResourceMark rm;
607 stringStream ss;
608
609 if (_output == nullptr) {
610 ss.print_cr("AFTER: %s", phase_name);
611 // Print out all nodes in ascending order of index.
612 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
613 } else {
614 // Dump the node blockwise if we have a scheduling
615 _output->print_scheduling(&ss);
616 }
617
618 // Check that the lock is not broken by a safepoint.
619 NoSafepointVerifier nsv;
620 ttyLocker ttyl;
621 if (xtty != nullptr) {
622 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
623 compile_id(),
624 is_osr_compilation() ? " compile_kind='osr'" : "",
625 phase_name);
626 }
627
628 tty->print("%s", ss.as_string());
629
630 if (xtty != nullptr) {
631 xtty->tail("ideal");
632 }
633 }
634 #endif
635
636 // ============================================================================
637 //------------------------------Compile standard-------------------------------
638
639 // Compile a method. entry_bci is -1 for normal compilations and indicates
640 // the continuation bci for on stack replacement.
641
642
643 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
644 Options options, DirectiveSet* directive)
645 : Phase(Compiler),
646 _compile_id(ci_env->compile_id()),
647 _options(options),
648 _method(target),
649 _entry_bci(osr_bci),
650 _ilt(nullptr),
651 _stub_function(nullptr),
652 _stub_name(nullptr),
653 _stub_id(-1),
654 _stub_entry_point(nullptr),
655 _max_node_limit(MaxNodeLimit),
656 _post_loop_opts_phase(false),
657 _merge_stores_phase(false),
658 _allow_macro_nodes(true),
659 _inlining_progress(false),
660 _inlining_incrementally(false),
661 _do_cleanup(false),
662 _has_reserved_stack_access(target->has_reserved_stack_access()),
663 _has_circular_inline_type(false),
664 #ifndef PRODUCT
665 _igv_idx(0),
666 _trace_opto_output(directive->TraceOptoOutputOption),
667 #endif
668 _has_method_handle_invokes(false),
669 _clinit_barrier_on_entry(false),
670 _stress_seed(0),
671 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
672 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
673 _env(ci_env),
674 _directive(directive),
675 _log(ci_env->log()),
676 _first_failure_details(nullptr),
677 _intrinsics(comp_arena(), 0, 0, nullptr),
678 _macro_nodes(comp_arena(), 8, 0, nullptr),
679 _parse_predicates(comp_arena(), 8, 0, nullptr),
680 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
681 _expensive_nodes(comp_arena(), 8, 0, nullptr),
682 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
683 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
684 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
685 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
686 _coarsened_locks(comp_arena(), 8, 0, nullptr),
687 _congraph(nullptr),
688 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
689 _unique(0),
690 _dead_node_count(0),
691 _dead_node_list(comp_arena()),
692 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
693 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
694 _node_arena(&_node_arena_one),
695 _mach_constant_base_node(nullptr),
696 _Compile_types(mtCompiler, Arena::Tag::tag_type),
697 _initial_gvn(nullptr),
698 _igvn_worklist(nullptr),
699 _types(nullptr),
700 _node_hash(nullptr),
701 _late_inlines(comp_arena(), 2, 0, nullptr),
702 _string_late_inlines(comp_arena(), 2, 0, nullptr),
703 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
704 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
705 _late_inlines_pos(0),
706 _number_of_mh_late_inlines(0),
707 _oom(false),
708 _replay_inline_data(nullptr),
709 _inline_printer(this),
710 _java_calls(0),
711 _inner_loops(0),
712 _interpreter_frame_size(0),
713 _output(nullptr)
714 #ifndef PRODUCT
715 ,
716 _in_dump_cnt(0)
717 #endif
718 {
719 C = this;
720 CompileWrapper cw(this);
721
722 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
723 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
724
725 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
726 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
727 // We can always print a disassembly, either abstract (hex dump) or
728 // with the help of a suitable hsdis library. Thus, we should not
729 // couple print_assembly and print_opto_assembly controls.
730 // But: always print opto and regular assembly on compile command 'print'.
731 bool print_assembly = directive->PrintAssemblyOption;
732 set_print_assembly(print_opto_assembly || print_assembly);
733 #else
734 set_print_assembly(false); // must initialize.
735 #endif
736
737 #ifndef PRODUCT
738 set_parsed_irreducible_loop(false);
739 #endif
740
741 if (directive->ReplayInlineOption) {
742 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
743 }
744 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
745 set_print_intrinsics(directive->PrintIntrinsicsOption);
746 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
747
748 if (ProfileTraps) {
749 // Make sure the method being compiled gets its own MDO,
750 // so we can at least track the decompile_count().
751 method()->ensure_method_data();
752 }
753
754 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
755 StressIncrementalInlining || StressMacroExpansion ||
756 StressMacroElimination || StressUnstableIfTraps ||
757 StressBailout || StressLoopPeeling) {
758 initialize_stress_seed(directive);
759 }
760
761 Init(/*do_aliasing=*/ true);
762
763 print_compile_messages();
764
765 _ilt = InlineTree::build_inline_tree_root();
766
767 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
768 assert(num_alias_types() >= AliasIdxRaw, "");
769
770 #define MINIMUM_NODE_HASH 1023
771
772 // GVN that will be run immediately on new nodes
773 uint estimated_size = method()->code_size()*4+64;
774 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
775 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
776 _types = new (comp_arena()) Type_Array(comp_arena());
777 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
778 PhaseGVN gvn;
779 set_initial_gvn(&gvn);
780
781 { // Scope for timing the parser
782 TracePhase tp(_t_parser);
783
784 // Put top into the hash table ASAP.
785 initial_gvn()->transform(top());
786
787 // Set up tf(), start(), and find a CallGenerator.
788 CallGenerator* cg = nullptr;
789 if (is_osr_compilation()) {
790 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
791 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
792 initial_gvn()->set_type_bottom(s);
793 verify_start(s);
794 cg = CallGenerator::for_osr(method(), entry_bci());
795 } else {
796 // Normal case.
797 init_tf(TypeFunc::make(method()));
798 StartNode* s = new StartNode(root(), tf()->domain_cc());
799 initial_gvn()->set_type_bottom(s);
800 verify_start(s);
801 float past_uses = method()->interpreter_invocation_count();
802 float expected_uses = past_uses;
803 cg = CallGenerator::for_inline(method(), expected_uses);
804 }
805 if (failing()) return;
806 if (cg == nullptr) {
807 const char* reason = InlineTree::check_can_parse(method());
808 assert(reason != nullptr, "expect reason for parse failure");
809 stringStream ss;
810 ss.print("cannot parse method: %s", reason);
811 record_method_not_compilable(ss.as_string());
812 return;
813 }
814
815 gvn.set_type(root(), root()->bottom_type());
816
817 JVMState* jvms = build_start_state(start(), tf());
818 if ((jvms = cg->generate(jvms)) == nullptr) {
819 assert(failure_reason() != nullptr, "expect reason for parse failure");
820 stringStream ss;
821 ss.print("method parse failed: %s", failure_reason());
822 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
823 return;
824 }
825 GraphKit kit(jvms);
826
827 if (!kit.stopped()) {
828 // Accept return values, and transfer control we know not where.
829 // This is done by a special, unique ReturnNode bound to root.
830 return_values(kit.jvms());
831 }
832
833 if (kit.has_exceptions()) {
834 // Any exceptions that escape from this call must be rethrown
835 // to whatever caller is dynamically above us on the stack.
836 // This is done by a special, unique RethrowNode bound to root.
837 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
838 }
839
840 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
841
842 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
843 inline_string_calls(true);
844 }
845
846 if (failing()) return;
847
848 // Remove clutter produced by parsing.
849 if (!failing()) {
850 ResourceMark rm;
851 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
852 }
853 }
854
855 // Note: Large methods are capped off in do_one_bytecode().
856 if (failing()) return;
857
858 // After parsing, node notes are no longer automagic.
859 // They must be propagated by register_new_node_with_optimizer(),
860 // clone(), or the like.
861 set_default_node_notes(nullptr);
862
863 #ifndef PRODUCT
864 if (should_print_igv(1)) {
865 _igv_printer->print_inlining();
866 }
867 #endif
868
869 if (failing()) return;
870 NOT_PRODUCT( verify_graph_edges(); )
871
872 // Now optimize
873 Optimize();
874 if (failing()) return;
875 NOT_PRODUCT( verify_graph_edges(); )
876
877 #ifndef PRODUCT
878 if (should_print_ideal()) {
879 print_ideal_ir("print_ideal");
880 }
881 #endif
882
883 #ifdef ASSERT
884 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
885 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
886 #endif
887
888 // Dump compilation data to replay it.
889 if (directive->DumpReplayOption) {
890 env()->dump_replay_data(_compile_id);
891 }
892 if (directive->DumpInlineOption && (ilt() != nullptr)) {
893 env()->dump_inline_data(_compile_id);
894 }
895
896 // Now that we know the size of all the monitors we can add a fixed slot
897 // for the original deopt pc.
898 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
899 if (needs_stack_repair()) {
900 // One extra slot for the special stack increment value
901 next_slot += 2;
902 }
903 // TODO 8284443 Only reserve extra slot if needed
904 if (InlineTypeReturnedAsFields) {
905 // One extra slot to hold the null marker for a nullable
906 // inline type return if we run out of registers.
907 next_slot += 2;
908 }
909 set_fixed_slots(next_slot);
910
911 // Compute when to use implicit null checks. Used by matching trap based
912 // nodes and NullCheck optimization.
913 set_allowed_deopt_reasons();
914
915 // Now generate code
916 Code_Gen();
917 }
918
919 //------------------------------Compile----------------------------------------
920 // Compile a runtime stub
921 Compile::Compile(ciEnv* ci_env,
922 TypeFunc_generator generator,
923 address stub_function,
924 const char* stub_name,
925 int stub_id,
926 int is_fancy_jump,
927 bool pass_tls,
928 bool return_pc,
929 DirectiveSet* directive)
930 : Phase(Compiler),
931 _compile_id(0),
932 _options(Options::for_runtime_stub()),
933 _method(nullptr),
934 _entry_bci(InvocationEntryBci),
935 _stub_function(stub_function),
936 _stub_name(stub_name),
937 _stub_id(stub_id),
938 _stub_entry_point(nullptr),
939 _max_node_limit(MaxNodeLimit),
940 _post_loop_opts_phase(false),
941 _merge_stores_phase(false),
942 _allow_macro_nodes(true),
943 _inlining_progress(false),
944 _inlining_incrementally(false),
945 _has_reserved_stack_access(false),
946 _has_circular_inline_type(false),
947 #ifndef PRODUCT
948 _igv_idx(0),
949 _trace_opto_output(directive->TraceOptoOutputOption),
950 #endif
951 _has_method_handle_invokes(false),
952 _clinit_barrier_on_entry(false),
953 _stress_seed(0),
954 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
955 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
956 _env(ci_env),
957 _directive(directive),
958 _log(ci_env->log()),
959 _first_failure_details(nullptr),
960 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
961 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
962 _congraph(nullptr),
963 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
964 _unique(0),
965 _dead_node_count(0),
966 _dead_node_list(comp_arena()),
967 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
968 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
969 _node_arena(&_node_arena_one),
970 _mach_constant_base_node(nullptr),
971 _Compile_types(mtCompiler, Arena::Tag::tag_type),
972 _initial_gvn(nullptr),
973 _igvn_worklist(nullptr),
974 _types(nullptr),
975 _node_hash(nullptr),
976 _number_of_mh_late_inlines(0),
977 _oom(false),
978 _replay_inline_data(nullptr),
979 _inline_printer(this),
980 _java_calls(0),
981 _inner_loops(0),
982 _interpreter_frame_size(0),
983 _output(nullptr),
984 #ifndef PRODUCT
985 _in_dump_cnt(0),
986 #endif
987 _allowed_reasons(0) {
988 C = this;
989
990 // try to reuse an existing stub
991 {
992 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, _stub_id, stub_name);
993 if (blob != nullptr) {
994 RuntimeStub* rs = blob->as_runtime_stub();
995 _stub_entry_point = rs->entry_point();
996 return;
997 }
998 }
999
1000 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
1001 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
1002
1003 #ifndef PRODUCT
1004 set_print_assembly(PrintFrameConverterAssembly);
1005 set_parsed_irreducible_loop(false);
1006 #else
1007 set_print_assembly(false); // Must initialize.
1008 #endif
1009 set_has_irreducible_loop(false); // no loops
1010
1011 CompileWrapper cw(this);
1012 Init(/*do_aliasing=*/ false);
1013 init_tf((*generator)());
1014
1015 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1016 _types = new (comp_arena()) Type_Array(comp_arena());
1017 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1018
1019 if (StressLCM || StressGCM || StressBailout) {
1020 initialize_stress_seed(directive);
1021 }
1022
1023 {
1024 PhaseGVN gvn;
1025 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1026 gvn.transform(top());
1027
1028 GraphKit kit;
1029 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1030 }
1031
1032 NOT_PRODUCT( verify_graph_edges(); )
1033
1034 Code_Gen();
1035 }
1036
1037 Compile::~Compile() {
1038 delete _first_failure_details;
1039 };
1040
1041 //------------------------------Init-------------------------------------------
1042 // Prepare for a single compilation
1043 void Compile::Init(bool aliasing) {
1044 _do_aliasing = aliasing;
1045 _unique = 0;
1046 _regalloc = nullptr;
1047
1048 _tf = nullptr; // filled in later
1049 _top = nullptr; // cached later
1050 _matcher = nullptr; // filled in later
1051 _cfg = nullptr; // filled in later
1052
1053 _node_note_array = nullptr;
1054 _default_node_notes = nullptr;
1055 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1056
1057 _immutable_memory = nullptr; // filled in at first inquiry
1058
1059 #ifdef ASSERT
1060 _phase_optimize_finished = false;
1061 _phase_verify_ideal_loop = false;
1062 _exception_backedge = false;
1063 _type_verify = nullptr;
1064 #endif
1065
1066 // Globally visible Nodes
1067 // First set TOP to null to give safe behavior during creation of RootNode
1068 set_cached_top_node(nullptr);
1069 set_root(new RootNode());
1070 // Now that you have a Root to point to, create the real TOP
1071 set_cached_top_node( new ConNode(Type::TOP) );
1072 set_recent_alloc(nullptr, nullptr);
1073
1074 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1075 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1076 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1077 env()->set_dependencies(new Dependencies(env()));
1078
1079 _fixed_slots = 0;
1080 set_has_split_ifs(false);
1081 set_has_loops(false); // first approximation
1082 set_has_stringbuilder(false);
1083 set_has_boxed_value(false);
1084 _trap_can_recompile = false; // no traps emitted yet
1085 _major_progress = true; // start out assuming good things will happen
1086 set_has_unsafe_access(false);
1087 set_max_vector_size(0);
1088 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1089 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1090 set_decompile_count(0);
1091
1092 #ifndef PRODUCT
1093 _phase_counter = 0;
1094 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1095 #endif
1096
1097 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1098 _loop_opts_cnt = LoopOptsCount;
1099 _has_flat_accesses = false;
1100 _flat_accesses_share_alias = true;
1101 _scalarize_in_safepoints = false;
1102
1103 set_do_inlining(Inline);
1104 set_max_inline_size(MaxInlineSize);
1105 set_freq_inline_size(FreqInlineSize);
1106 set_do_scheduling(OptoScheduling);
1107
1108 set_do_vector_loop(false);
1109 set_has_monitors(false);
1110 set_has_scoped_access(false);
1111
1112 if (AllowVectorizeOnDemand) {
1113 if (has_method() && _directive->VectorizeOption) {
1114 set_do_vector_loop(true);
1115 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());})
1116 } else if (has_method() && method()->name() != nullptr &&
1117 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1118 set_do_vector_loop(true);
1119 }
1120 }
1121 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1122 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());})
1123
1124 _max_node_limit = _directive->MaxNodeLimitOption;
1125
1126 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1127 set_clinit_barrier_on_entry(true);
1128 }
1129 if (debug_info()->recording_non_safepoints()) {
1130 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1131 (comp_arena(), 8, 0, nullptr));
1132 set_default_node_notes(Node_Notes::make(this));
1133 }
1134
1135 const int grow_ats = 16;
1136 _max_alias_types = grow_ats;
1137 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1138 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1139 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1140 {
1141 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1142 }
1143 // Initialize the first few types.
1144 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1145 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1146 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1147 _num_alias_types = AliasIdxRaw+1;
1148 // Zero out the alias type cache.
1149 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1150 // A null adr_type hits in the cache right away. Preload the right answer.
1151 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1152 }
1153
1154 #ifdef ASSERT
1155 // Verify that the current StartNode is valid.
1156 void Compile::verify_start(StartNode* s) const {
1157 assert(failing_internal() || s == start(), "should be StartNode");
1158 }
1159 #endif
1160
1161 /**
1162 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1163 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1164 * the ideal graph.
1165 */
1166 StartNode* Compile::start() const {
1167 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1168 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1169 Node* start = root()->fast_out(i);
1170 if (start->is_Start()) {
1171 return start->as_Start();
1172 }
1173 }
1174 fatal("Did not find Start node!");
1175 return nullptr;
1176 }
1177
1178 //-------------------------------immutable_memory-------------------------------------
1179 // Access immutable memory
1180 Node* Compile::immutable_memory() {
1181 if (_immutable_memory != nullptr) {
1182 return _immutable_memory;
1183 }
1184 StartNode* s = start();
1185 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1186 Node *p = s->fast_out(i);
1187 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1188 _immutable_memory = p;
1189 return _immutable_memory;
1190 }
1191 }
1192 ShouldNotReachHere();
1193 return nullptr;
1194 }
1195
1196 //----------------------set_cached_top_node------------------------------------
1197 // Install the cached top node, and make sure Node::is_top works correctly.
1198 void Compile::set_cached_top_node(Node* tn) {
1199 if (tn != nullptr) verify_top(tn);
1200 Node* old_top = _top;
1201 _top = tn;
1202 // Calling Node::setup_is_top allows the nodes the chance to adjust
1203 // their _out arrays.
1204 if (_top != nullptr) _top->setup_is_top();
1205 if (old_top != nullptr) old_top->setup_is_top();
1206 assert(_top == nullptr || top()->is_top(), "");
1207 }
1208
1209 #ifdef ASSERT
1210 uint Compile::count_live_nodes_by_graph_walk() {
1211 Unique_Node_List useful(comp_arena());
1212 // Get useful node list by walking the graph.
1213 identify_useful_nodes(useful);
1214 return useful.size();
1215 }
1216
1217 void Compile::print_missing_nodes() {
1218
1219 // Return if CompileLog is null and PrintIdealNodeCount is false.
1220 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1221 return;
1222 }
1223
1224 // This is an expensive function. It is executed only when the user
1225 // specifies VerifyIdealNodeCount option or otherwise knows the
1226 // additional work that needs to be done to identify reachable nodes
1227 // by walking the flow graph and find the missing ones using
1228 // _dead_node_list.
1229
1230 Unique_Node_List useful(comp_arena());
1231 // Get useful node list by walking the graph.
1232 identify_useful_nodes(useful);
1233
1234 uint l_nodes = C->live_nodes();
1235 uint l_nodes_by_walk = useful.size();
1236
1237 if (l_nodes != l_nodes_by_walk) {
1238 if (_log != nullptr) {
1239 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1240 _log->stamp();
1241 _log->end_head();
1242 }
1243 VectorSet& useful_member_set = useful.member_set();
1244 int last_idx = l_nodes_by_walk;
1245 for (int i = 0; i < last_idx; i++) {
1246 if (useful_member_set.test(i)) {
1247 if (_dead_node_list.test(i)) {
1248 if (_log != nullptr) {
1249 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1250 }
1251 if (PrintIdealNodeCount) {
1252 // Print the log message to tty
1253 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1254 useful.at(i)->dump();
1255 }
1256 }
1257 }
1258 else if (! _dead_node_list.test(i)) {
1259 if (_log != nullptr) {
1260 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1261 }
1262 if (PrintIdealNodeCount) {
1263 // Print the log message to tty
1264 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1265 }
1266 }
1267 }
1268 if (_log != nullptr) {
1269 _log->tail("mismatched_nodes");
1270 }
1271 }
1272 }
1273 void Compile::record_modified_node(Node* n) {
1274 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1275 _modified_nodes->push(n);
1276 }
1277 }
1278
1279 void Compile::remove_modified_node(Node* n) {
1280 if (_modified_nodes != nullptr) {
1281 _modified_nodes->remove(n);
1282 }
1283 }
1284 #endif
1285
1286 #ifndef PRODUCT
1287 void Compile::verify_top(Node* tn) const {
1288 if (tn != nullptr) {
1289 assert(tn->is_Con(), "top node must be a constant");
1290 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1291 assert(tn->in(0) != nullptr, "must have live top node");
1292 }
1293 }
1294 #endif
1295
1296
1297 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1298
1299 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1300 guarantee(arr != nullptr, "");
1301 int num_blocks = arr->length();
1302 if (grow_by < num_blocks) grow_by = num_blocks;
1303 int num_notes = grow_by * _node_notes_block_size;
1304 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1305 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1306 while (num_notes > 0) {
1307 arr->append(notes);
1308 notes += _node_notes_block_size;
1309 num_notes -= _node_notes_block_size;
1310 }
1311 assert(num_notes == 0, "exact multiple, please");
1312 }
1313
1314 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1315 if (source == nullptr || dest == nullptr) return false;
1316
1317 if (dest->is_Con())
1318 return false; // Do not push debug info onto constants.
1319
1320 #ifdef ASSERT
1321 // Leave a bread crumb trail pointing to the original node:
1322 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1323 dest->set_debug_orig(source);
1324 }
1325 #endif
1326
1327 if (node_note_array() == nullptr)
1328 return false; // Not collecting any notes now.
1329
1330 // This is a copy onto a pre-existing node, which may already have notes.
1331 // If both nodes have notes, do not overwrite any pre-existing notes.
1332 Node_Notes* source_notes = node_notes_at(source->_idx);
1333 if (source_notes == nullptr || source_notes->is_clear()) return false;
1334 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1335 if (dest_notes == nullptr || dest_notes->is_clear()) {
1336 return set_node_notes_at(dest->_idx, source_notes);
1337 }
1338
1339 Node_Notes merged_notes = (*source_notes);
1340 // The order of operations here ensures that dest notes will win...
1341 merged_notes.update_from(dest_notes);
1342 return set_node_notes_at(dest->_idx, &merged_notes);
1343 }
1344
1345
1346 //--------------------------allow_range_check_smearing-------------------------
1347 // Gating condition for coalescing similar range checks.
1348 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1349 // single covering check that is at least as strong as any of them.
1350 // If the optimization succeeds, the simplified (strengthened) range check
1351 // will always succeed. If it fails, we will deopt, and then give up
1352 // on the optimization.
1353 bool Compile::allow_range_check_smearing() const {
1354 // If this method has already thrown a range-check,
1355 // assume it was because we already tried range smearing
1356 // and it failed.
1357 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1358 return !already_trapped;
1359 }
1360
1361
1362 //------------------------------flatten_alias_type-----------------------------
1363 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1364 assert(do_aliasing(), "Aliasing should be enabled");
1365 int offset = tj->offset();
1366 TypePtr::PTR ptr = tj->ptr();
1367
1368 // Known instance (scalarizable allocation) alias only with itself.
1369 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1370 tj->is_oopptr()->is_known_instance();
1371
1372 // Process weird unsafe references.
1373 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1374 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1375 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1376 tj = TypeOopPtr::BOTTOM;
1377 ptr = tj->ptr();
1378 offset = tj->offset();
1379 }
1380
1381 // Array pointers need some flattening
1382 const TypeAryPtr* ta = tj->isa_aryptr();
1383 if (ta && ta->is_stable()) {
1384 // Erase stability property for alias analysis.
1385 tj = ta = ta->cast_to_stable(false);
1386 }
1387 if (ta && ta->is_not_flat()) {
1388 // Erase not flat property for alias analysis.
1389 tj = ta = ta->cast_to_not_flat(false);
1390 }
1391 if (ta && ta->is_not_null_free()) {
1392 // Erase not null free property for alias analysis.
1393 tj = ta = ta->cast_to_not_null_free(false);
1394 }
1395
1396 if( ta && is_known_inst ) {
1397 if ( offset != Type::OffsetBot &&
1398 offset > arrayOopDesc::length_offset_in_bytes() ) {
1399 offset = Type::OffsetBot; // Flatten constant access into array body only
1400 tj = ta = ta->
1401 remove_speculative()->
1402 cast_to_ptr_type(ptr)->
1403 with_offset(offset);
1404 }
1405 } else if (ta) {
1406 // For arrays indexed by constant indices, we flatten the alias
1407 // space to include all of the array body. Only the header, klass
1408 // and array length can be accessed un-aliased.
1409 // For flat inline type array, each field has its own slice so
1410 // we must include the field offset.
1411 if( offset != Type::OffsetBot ) {
1412 if( ta->const_oop() ) { // MethodData* or Method*
1413 offset = Type::OffsetBot; // Flatten constant access into array body
1414 tj = ta = ta->
1415 remove_speculative()->
1416 cast_to_ptr_type(ptr)->
1417 cast_to_exactness(false)->
1418 with_offset(offset);
1419 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1420 // range is OK as-is.
1421 tj = ta = TypeAryPtr::RANGE;
1422 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1423 tj = TypeInstPtr::KLASS; // all klass loads look alike
1424 ta = TypeAryPtr::RANGE; // generic ignored junk
1425 ptr = TypePtr::BotPTR;
1426 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1427 tj = TypeInstPtr::MARK;
1428 ta = TypeAryPtr::RANGE; // generic ignored junk
1429 ptr = TypePtr::BotPTR;
1430 } else { // Random constant offset into array body
1431 offset = Type::OffsetBot; // Flatten constant access into array body
1432 tj = ta = ta->
1433 remove_speculative()->
1434 cast_to_ptr_type(ptr)->
1435 cast_to_exactness(false)->
1436 with_offset(offset);
1437 }
1438 }
1439 // Arrays of fixed size alias with arrays of unknown size.
1440 if (ta->size() != TypeInt::POS) {
1441 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1442 tj = ta = ta->
1443 remove_speculative()->
1444 cast_to_ptr_type(ptr)->
1445 with_ary(tary)->
1446 cast_to_exactness(false);
1447 }
1448 // Arrays of known objects become arrays of unknown objects.
1449 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1450 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1451 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1452 }
1453 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1454 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1455 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1456 }
1457 // Initially all flattened array accesses share a single slice
1458 if (ta->is_flat() && ta->elem() != TypeInstPtr::BOTTOM && _flat_accesses_share_alias) {
1459 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size(), /* stable= */ false, /* flat= */ true);
1460 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1461 }
1462 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1463 // cannot be distinguished by bytecode alone.
1464 if (ta->elem() == TypeInt::BOOL) {
1465 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1466 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1467 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1468 }
1469 // During the 2nd round of IterGVN, NotNull castings are removed.
1470 // Make sure the Bottom and NotNull variants alias the same.
1471 // Also, make sure exact and non-exact variants alias the same.
1472 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1473 tj = ta = ta->
1474 remove_speculative()->
1475 cast_to_ptr_type(TypePtr::BotPTR)->
1476 cast_to_exactness(false)->
1477 with_offset(offset);
1478 }
1479 }
1480
1481 // Oop pointers need some flattening
1482 const TypeInstPtr *to = tj->isa_instptr();
1483 if (to && to != TypeOopPtr::BOTTOM) {
1484 ciInstanceKlass* ik = to->instance_klass();
1485 if( ptr == TypePtr::Constant ) {
1486 if (ik != ciEnv::current()->Class_klass() ||
1487 offset < ik->layout_helper_size_in_bytes()) {
1488 // No constant oop pointers (such as Strings); they alias with
1489 // unknown strings.
1490 assert(!is_known_inst, "not scalarizable allocation");
1491 tj = to = to->
1492 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1493 remove_speculative()->
1494 cast_to_ptr_type(TypePtr::BotPTR)->
1495 cast_to_exactness(false);
1496 }
1497 } else if( is_known_inst ) {
1498 tj = to; // Keep NotNull and klass_is_exact for instance type
1499 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1500 // During the 2nd round of IterGVN, NotNull castings are removed.
1501 // Make sure the Bottom and NotNull variants alias the same.
1502 // Also, make sure exact and non-exact variants alias the same.
1503 tj = to = to->
1504 remove_speculative()->
1505 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1506 cast_to_ptr_type(TypePtr::BotPTR)->
1507 cast_to_exactness(false);
1508 }
1509 if (to->speculative() != nullptr) {
1510 tj = to = to->remove_speculative();
1511 }
1512 // Canonicalize the holder of this field
1513 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1514 // First handle header references such as a LoadKlassNode, even if the
1515 // object's klass is unloaded at compile time (4965979).
1516 if (!is_known_inst) { // Do it only for non-instance types
1517 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1518 }
1519 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1520 // Static fields are in the space above the normal instance
1521 // fields in the java.lang.Class instance.
1522 if (ik != ciEnv::current()->Class_klass()) {
1523 to = nullptr;
1524 tj = TypeOopPtr::BOTTOM;
1525 offset = tj->offset();
1526 }
1527 } else {
1528 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1529 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1530 assert(tj->offset() == offset, "no change to offset expected");
1531 bool xk = to->klass_is_exact();
1532 int instance_id = to->instance_id();
1533
1534 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1535 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1536 // its interfaces are included.
1537 if (xk && ik->equals(canonical_holder)) {
1538 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id), "exact type should be canonical type");
1539 } else {
1540 assert(xk || !is_known_inst, "Known instance should be exact type");
1541 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id);
1542 }
1543 }
1544 }
1545
1546 // Klass pointers to object array klasses need some flattening
1547 const TypeKlassPtr *tk = tj->isa_klassptr();
1548 if( tk ) {
1549 // If we are referencing a field within a Klass, we need
1550 // to assume the worst case of an Object. Both exact and
1551 // inexact types must flatten to the same alias class so
1552 // use NotNull as the PTR.
1553 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1554 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1555 env()->Object_klass(),
1556 Type::Offset(offset));
1557 }
1558
1559 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1560 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1561 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1562 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset));
1563 } else {
1564 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_vm_type());
1565 }
1566 }
1567 // Check for precise loads from the primary supertype array and force them
1568 // to the supertype cache alias index. Check for generic array loads from
1569 // the primary supertype array and also force them to the supertype cache
1570 // alias index. Since the same load can reach both, we need to merge
1571 // these 2 disparate memories into the same alias class. Since the
1572 // primary supertype array is read-only, there's no chance of confusion
1573 // where we bypass an array load and an array store.
1574 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1575 if (offset == Type::OffsetBot ||
1576 (offset >= primary_supers_offset &&
1577 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1578 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1579 offset = in_bytes(Klass::secondary_super_cache_offset());
1580 tj = tk = tk->with_offset(offset);
1581 }
1582 }
1583
1584 // Flatten all Raw pointers together.
1585 if (tj->base() == Type::RawPtr)
1586 tj = TypeRawPtr::BOTTOM;
1587
1588 if (tj->base() == Type::AnyPtr)
1589 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1590
1591 offset = tj->offset();
1592 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1593
1594 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1595 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1596 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1597 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1598 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1599 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1600 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1601 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1602 assert( tj->ptr() != TypePtr::TopPTR &&
1603 tj->ptr() != TypePtr::AnyNull &&
1604 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1605 // assert( tj->ptr() != TypePtr::Constant ||
1606 // tj->base() == Type::RawPtr ||
1607 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1608
1609 return tj;
1610 }
1611
1612 void Compile::AliasType::Init(int i, const TypePtr* at) {
1613 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1614 _index = i;
1615 _adr_type = at;
1616 _field = nullptr;
1617 _element = nullptr;
1618 _is_rewritable = true; // default
1619 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1620 if (atoop != nullptr && atoop->is_known_instance()) {
1621 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1622 _general_index = Compile::current()->get_alias_index(gt);
1623 } else {
1624 _general_index = 0;
1625 }
1626 }
1627
1628 BasicType Compile::AliasType::basic_type() const {
1629 if (element() != nullptr) {
1630 const Type* element = adr_type()->is_aryptr()->elem();
1631 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1632 } if (field() != nullptr) {
1633 return field()->layout_type();
1634 } else {
1635 return T_ILLEGAL; // unknown
1636 }
1637 }
1638
1639 //---------------------------------print_on------------------------------------
1640 #ifndef PRODUCT
1641 void Compile::AliasType::print_on(outputStream* st) {
1642 if (index() < 10)
1643 st->print("@ <%d> ", index());
1644 else st->print("@ <%d>", index());
1645 st->print(is_rewritable() ? " " : " RO");
1646 int offset = adr_type()->offset();
1647 if (offset == Type::OffsetBot)
1648 st->print(" +any");
1649 else st->print(" +%-3d", offset);
1650 st->print(" in ");
1651 adr_type()->dump_on(st);
1652 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1653 if (field() != nullptr && tjp) {
1654 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1655 tjp->offset() != field()->offset_in_bytes()) {
1656 st->print(" != ");
1657 field()->print();
1658 st->print(" ***");
1659 }
1660 }
1661 }
1662
1663 void print_alias_types() {
1664 Compile* C = Compile::current();
1665 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1666 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1667 C->alias_type(idx)->print_on(tty);
1668 tty->cr();
1669 }
1670 }
1671 #endif
1672
1673
1674 //----------------------------probe_alias_cache--------------------------------
1675 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1676 intptr_t key = (intptr_t) adr_type;
1677 key ^= key >> logAliasCacheSize;
1678 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1679 }
1680
1681
1682 //-----------------------------grow_alias_types--------------------------------
1683 void Compile::grow_alias_types() {
1684 const int old_ats = _max_alias_types; // how many before?
1685 const int new_ats = old_ats; // how many more?
1686 const int grow_ats = old_ats+new_ats; // how many now?
1687 _max_alias_types = grow_ats;
1688 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1689 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1690 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1691 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1692 }
1693
1694
1695 //--------------------------------find_alias_type------------------------------
1696 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1697 if (!do_aliasing()) {
1698 return alias_type(AliasIdxBot);
1699 }
1700
1701 AliasCacheEntry* ace = nullptr;
1702 if (!uncached) {
1703 ace = probe_alias_cache(adr_type);
1704 if (ace->_adr_type == adr_type) {
1705 return alias_type(ace->_index);
1706 }
1707 }
1708
1709 // Handle special cases.
1710 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1711 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1712
1713 // Do it the slow way.
1714 const TypePtr* flat = flatten_alias_type(adr_type);
1715
1716 #ifdef ASSERT
1717 {
1718 ResourceMark rm;
1719 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1720 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1721 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1722 Type::str(adr_type));
1723 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1724 const TypeOopPtr* foop = flat->is_oopptr();
1725 // Scalarizable allocations have exact klass always.
1726 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1727 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1728 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1729 Type::str(foop), Type::str(xoop));
1730 }
1731 }
1732 #endif
1733
1734 int idx = AliasIdxTop;
1735 for (int i = 0; i < num_alias_types(); i++) {
1736 if (alias_type(i)->adr_type() == flat) {
1737 idx = i;
1738 break;
1739 }
1740 }
1741
1742 if (idx == AliasIdxTop) {
1743 if (no_create) return nullptr;
1744 // Grow the array if necessary.
1745 if (_num_alias_types == _max_alias_types) grow_alias_types();
1746 // Add a new alias type.
1747 idx = _num_alias_types++;
1748 _alias_types[idx]->Init(idx, flat);
1749 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1750 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1751 if (flat->isa_instptr()) {
1752 if (flat->offset() == java_lang_Class::klass_offset()
1753 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1754 alias_type(idx)->set_rewritable(false);
1755 }
1756 ciField* field = nullptr;
1757 if (flat->isa_aryptr()) {
1758 #ifdef ASSERT
1759 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1760 // (T_BYTE has the weakest alignment and size restrictions...)
1761 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1762 #endif
1763 const Type* elemtype = flat->is_aryptr()->elem();
1764 if (flat->offset() == TypePtr::OffsetBot) {
1765 alias_type(idx)->set_element(elemtype);
1766 }
1767 int field_offset = flat->is_aryptr()->field_offset().get();
1768 if (flat->is_flat() &&
1769 field_offset != Type::OffsetBot) {
1770 ciInlineKlass* vk = elemtype->inline_klass();
1771 field_offset += vk->payload_offset();
1772 field = vk->get_field_by_offset(field_offset, false);
1773 }
1774 }
1775 if (flat->isa_klassptr()) {
1776 if (UseCompactObjectHeaders) {
1777 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1778 alias_type(idx)->set_rewritable(false);
1779 }
1780 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1781 alias_type(idx)->set_rewritable(false);
1782 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1783 alias_type(idx)->set_rewritable(false);
1784 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1785 alias_type(idx)->set_rewritable(false);
1786 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1787 alias_type(idx)->set_rewritable(false);
1788 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1789 alias_type(idx)->set_rewritable(false);
1790 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1791 alias_type(idx)->set_rewritable(false);
1792 }
1793 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1794 // but the base pointer type is not distinctive enough to identify
1795 // references into JavaThread.)
1796
1797 // Check for final fields.
1798 const TypeInstPtr* tinst = flat->isa_instptr();
1799 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1800 if (tinst->const_oop() != nullptr &&
1801 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1802 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1803 // static field
1804 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1805 field = k->get_field_by_offset(tinst->offset(), true);
1806 } else if (tinst->is_inlinetypeptr()) {
1807 // Inline type field
1808 ciInlineKlass* vk = tinst->inline_klass();
1809 field = vk->get_field_by_offset(tinst->offset(), false);
1810 } else {
1811 ciInstanceKlass *k = tinst->instance_klass();
1812 field = k->get_field_by_offset(tinst->offset(), false);
1813 }
1814 }
1815 assert(field == nullptr ||
1816 original_field == nullptr ||
1817 (field->holder() == original_field->holder() &&
1818 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1819 field->is_static() == original_field->is_static()), "wrong field?");
1820 // Set field() and is_rewritable() attributes.
1821 if (field != nullptr) {
1822 alias_type(idx)->set_field(field);
1823 if (flat->isa_aryptr()) {
1824 // Fields of flat arrays are rewritable although they are declared final
1825 assert(flat->is_flat(), "must be a flat array");
1826 alias_type(idx)->set_rewritable(true);
1827 }
1828 }
1829 }
1830
1831 // Fill the cache for next time.
1832 if (!uncached) {
1833 ace->_adr_type = adr_type;
1834 ace->_index = idx;
1835 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1836
1837 // Might as well try to fill the cache for the flattened version, too.
1838 AliasCacheEntry* face = probe_alias_cache(flat);
1839 if (face->_adr_type == nullptr) {
1840 face->_adr_type = flat;
1841 face->_index = idx;
1842 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1843 }
1844 }
1845
1846 return alias_type(idx);
1847 }
1848
1849
1850 Compile::AliasType* Compile::alias_type(ciField* field) {
1851 const TypeOopPtr* t;
1852 if (field->is_static())
1853 t = TypeInstPtr::make(field->holder()->java_mirror());
1854 else
1855 t = TypeOopPtr::make_from_klass_raw(field->holder());
1856 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1857 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1858 return atp;
1859 }
1860
1861
1862 //------------------------------have_alias_type--------------------------------
1863 bool Compile::have_alias_type(const TypePtr* adr_type) {
1864 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1865 if (ace->_adr_type == adr_type) {
1866 return true;
1867 }
1868
1869 // Handle special cases.
1870 if (adr_type == nullptr) return true;
1871 if (adr_type == TypePtr::BOTTOM) return true;
1872
1873 return find_alias_type(adr_type, true, nullptr) != nullptr;
1874 }
1875
1876 //-----------------------------must_alias--------------------------------------
1877 // True if all values of the given address type are in the given alias category.
1878 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1879 if (alias_idx == AliasIdxBot) return true; // the universal category
1880 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1881 if (alias_idx == AliasIdxTop) return false; // the empty category
1882 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1883
1884 // the only remaining possible overlap is identity
1885 int adr_idx = get_alias_index(adr_type);
1886 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1887 assert(adr_idx == alias_idx ||
1888 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1889 && adr_type != TypeOopPtr::BOTTOM),
1890 "should not be testing for overlap with an unsafe pointer");
1891 return adr_idx == alias_idx;
1892 }
1893
1894 //------------------------------can_alias--------------------------------------
1895 // True if any values of the given address type are in the given alias category.
1896 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1897 if (alias_idx == AliasIdxTop) return false; // the empty category
1898 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1899 // Known instance doesn't alias with bottom memory
1900 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1901 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1902
1903 // the only remaining possible overlap is identity
1904 int adr_idx = get_alias_index(adr_type);
1905 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1906 return adr_idx == alias_idx;
1907 }
1908
1909 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1910 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1911 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1912 if (parse_predicate_count() == 0) {
1913 return;
1914 }
1915 for (int i = 0; i < parse_predicate_count(); i++) {
1916 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1917 parse_predicate->mark_useless(igvn);
1918 }
1919 _parse_predicates.clear();
1920 }
1921
1922 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1923 if (!n->for_post_loop_opts_igvn()) {
1924 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1925 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1926 _for_post_loop_igvn.append(n);
1927 }
1928 }
1929
1930 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1931 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1932 _for_post_loop_igvn.remove(n);
1933 }
1934
1935 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1936 // Verify that all previous optimizations produced a valid graph
1937 // at least to this point, even if no loop optimizations were done.
1938 PhaseIdealLoop::verify(igvn);
1939
1940 if (has_loops() || _loop_opts_cnt > 0) {
1941 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1942 }
1943 C->set_post_loop_opts_phase(); // no more loop opts allowed
1944
1945 assert(!C->major_progress(), "not cleared");
1946
1947 if (_for_post_loop_igvn.length() > 0) {
1948 while (_for_post_loop_igvn.length() > 0) {
1949 Node* n = _for_post_loop_igvn.pop();
1950 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1951 igvn._worklist.push(n);
1952 }
1953 igvn.optimize();
1954 if (failing()) return;
1955 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1956 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1957
1958 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1959 if (C->major_progress()) {
1960 C->clear_major_progress(); // ensure that major progress is now clear
1961 }
1962 }
1963 }
1964
1965 void Compile::add_inline_type(Node* n) {
1966 assert(n->is_InlineType(), "unexpected node");
1967 _inline_type_nodes.push(n);
1968 }
1969
1970 void Compile::remove_inline_type(Node* n) {
1971 assert(n->is_InlineType(), "unexpected node");
1972 if (_inline_type_nodes.contains(n)) {
1973 _inline_type_nodes.remove(n);
1974 }
1975 }
1976
1977 // Does the return value keep otherwise useless inline type allocations alive?
1978 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1979 ResourceMark rm;
1980 Unique_Node_List wq;
1981 wq.push(ret_val);
1982 bool some_allocations = false;
1983 for (uint i = 0; i < wq.size(); i++) {
1984 Node* n = wq.at(i);
1985 if (n->outcnt() > 1) {
1986 // Some other use for the allocation
1987 return false;
1988 } else if (n->is_InlineType()) {
1989 wq.push(n->in(1));
1990 } else if (n->is_Phi()) {
1991 for (uint j = 1; j < n->req(); j++) {
1992 wq.push(n->in(j));
1993 }
1994 } else if (n->is_CheckCastPP() &&
1995 n->in(1)->is_Proj() &&
1996 n->in(1)->in(0)->is_Allocate()) {
1997 some_allocations = true;
1998 } else if (n->is_CheckCastPP()) {
1999 wq.push(n->in(1));
2000 }
2001 }
2002 return some_allocations;
2003 }
2004
2005 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
2006 // Make sure that the return value does not keep an otherwise unused allocation alive
2007 if (tf()->returns_inline_type_as_fields()) {
2008 Node* ret = nullptr;
2009 for (uint i = 1; i < root()->req(); i++) {
2010 Node* in = root()->in(i);
2011 if (in->Opcode() == Op_Return) {
2012 assert(ret == nullptr, "only one return");
2013 ret = in;
2014 }
2015 }
2016 if (ret != nullptr) {
2017 Node* ret_val = ret->in(TypeFunc::Parms);
2018 if (igvn.type(ret_val)->isa_oopptr() &&
2019 return_val_keeps_allocations_alive(ret_val)) {
2020 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
2021 assert(ret_val->outcnt() == 0, "should be dead now");
2022 igvn.remove_dead_node(ret_val);
2023 }
2024 }
2025 }
2026 if (_inline_type_nodes.length() == 0) {
2027 return;
2028 }
2029 // Scalarize inline types in safepoint debug info.
2030 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2031 set_scalarize_in_safepoints(true);
2032 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2033 InlineTypeNode* vt = _inline_type_nodes.at(i)->as_InlineType();
2034 vt->make_scalar_in_safepoints(&igvn);
2035 igvn.record_for_igvn(vt);
2036 }
2037 if (remove) {
2038 // Remove inline type nodes by replacing them with their oop input
2039 while (_inline_type_nodes.length() > 0) {
2040 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2041 if (vt->outcnt() == 0) {
2042 igvn.remove_dead_node(vt);
2043 continue;
2044 }
2045 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2046 DEBUG_ONLY(bool must_be_buffered = false);
2047 Node* u = vt->out(i);
2048 // Check if any users are blackholes. If so, rewrite them to use either the
2049 // allocated buffer, or individual components, instead of the inline type node
2050 // that goes away.
2051 if (u->is_Blackhole()) {
2052 BlackholeNode* bh = u->as_Blackhole();
2053
2054 // Unlink the old input
2055 int idx = bh->find_edge(vt);
2056 assert(idx != -1, "The edge should be there");
2057 bh->del_req(idx);
2058 --i;
2059
2060 if (vt->is_allocated(&igvn)) {
2061 // Already has the allocated instance, blackhole that
2062 bh->add_req(vt->get_oop());
2063 } else {
2064 // Not allocated yet, blackhole the components
2065 for (uint c = 0; c < vt->field_count(); c++) {
2066 bh->add_req(vt->field_value(c));
2067 }
2068 }
2069
2070 // Node modified, record for IGVN
2071 igvn.record_for_igvn(bh);
2072 }
2073 #ifdef ASSERT
2074 // Verify that inline type is buffered when replacing by oop
2075 else if (u->is_InlineType()) {
2076 // InlineType uses don't need buffering because they are about to be replaced as well
2077 } else if (u->is_Phi()) {
2078 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2079 } else {
2080 must_be_buffered = true;
2081 }
2082 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2083 vt->dump(0);
2084 u->dump(0);
2085 assert(false, "Should have been buffered");
2086 }
2087 #endif
2088 }
2089 igvn.replace_node(vt, vt->get_oop());
2090 }
2091 }
2092 igvn.optimize();
2093 }
2094
2095 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2096 if (!_has_flat_accesses) {
2097 return;
2098 }
2099 // Initially, all flat array accesses share the same slice to
2100 // keep dependencies with Object[] array accesses (that could be
2101 // to a flat array) correct. We're done with parsing so we
2102 // now know all flat array accesses in this compile
2103 // unit. Let's move flat array accesses to their own slice,
2104 // one per element field. This should help memory access
2105 // optimizations.
2106 ResourceMark rm;
2107 Unique_Node_List wq;
2108 wq.push(root());
2109
2110 Node_List mergememnodes;
2111 Node_List memnodes;
2112
2113 // Alias index currently shared by all flat memory accesses
2114 int index = get_alias_index(TypeAryPtr::INLINES);
2115
2116 // Find MergeMem nodes and flat array accesses
2117 for (uint i = 0; i < wq.size(); i++) {
2118 Node* n = wq.at(i);
2119 if (n->is_Mem()) {
2120 const TypePtr* adr_type = nullptr;
2121 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2122 if (adr_type == TypeAryPtr::INLINES) {
2123 memnodes.push(n);
2124 }
2125 } else if (n->is_MergeMem()) {
2126 MergeMemNode* mm = n->as_MergeMem();
2127 if (mm->memory_at(index) != mm->base_memory()) {
2128 mergememnodes.push(n);
2129 }
2130 }
2131 for (uint j = 0; j < n->req(); j++) {
2132 Node* m = n->in(j);
2133 if (m != nullptr) {
2134 wq.push(m);
2135 }
2136 }
2137 }
2138
2139 if (memnodes.size() > 0) {
2140 _flat_accesses_share_alias = false;
2141
2142 // We are going to change the slice for the flat array
2143 // accesses so we need to clear the cache entries that refer to
2144 // them.
2145 for (uint i = 0; i < AliasCacheSize; i++) {
2146 AliasCacheEntry* ace = &_alias_cache[i];
2147 if (ace->_adr_type != nullptr &&
2148 ace->_adr_type->is_flat()) {
2149 ace->_adr_type = nullptr;
2150 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2151 }
2152 }
2153
2154 // Find what aliases we are going to add
2155 int start_alias = num_alias_types()-1;
2156 int stop_alias = 0;
2157
2158 for (uint i = 0; i < memnodes.size(); i++) {
2159 Node* m = memnodes.at(i);
2160 const TypePtr* adr_type = nullptr;
2161 adr_type = m->adr_type();
2162 #ifdef ASSERT
2163 m->as_Mem()->set_adr_type(adr_type);
2164 #endif
2165 int idx = get_alias_index(adr_type);
2166 start_alias = MIN2(start_alias, idx);
2167 stop_alias = MAX2(stop_alias, idx);
2168 }
2169
2170 assert(stop_alias >= start_alias, "should have expanded aliases");
2171
2172 Node_Stack stack(0);
2173 #ifdef ASSERT
2174 VectorSet seen(Thread::current()->resource_area());
2175 #endif
2176 // Now let's fix the memory graph so each flat array access
2177 // is moved to the right slice. Start from the MergeMem nodes.
2178 uint last = unique();
2179 for (uint i = 0; i < mergememnodes.size(); i++) {
2180 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2181 Node* n = current->memory_at(index);
2182 MergeMemNode* mm = nullptr;
2183 do {
2184 // Follow memory edges through memory accesses, phis and
2185 // narrow membars and push nodes on the stack. Once we hit
2186 // bottom memory, we pop element off the stack one at a
2187 // time, in reverse order, and move them to the right slice
2188 // by changing their memory edges.
2189 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2190 assert(!seen.test_set(n->_idx), "");
2191 // Uses (a load for instance) will need to be moved to the
2192 // right slice as well and will get a new memory state
2193 // that we don't know yet. The use could also be the
2194 // backedge of a loop. We put a place holder node between
2195 // the memory node and its uses. We replace that place
2196 // holder with the correct memory state once we know it,
2197 // i.e. when nodes are popped off the stack. Using the
2198 // place holder make the logic work in the presence of
2199 // loops.
2200 if (n->outcnt() > 1) {
2201 Node* place_holder = nullptr;
2202 assert(!n->has_out_with(Op_Node), "");
2203 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2204 Node* u = n->out(k);
2205 if (u != current && u->_idx < last) {
2206 bool success = false;
2207 for (uint l = 0; l < u->req(); l++) {
2208 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2209 continue;
2210 }
2211 Node* in = u->in(l);
2212 if (in == n) {
2213 if (place_holder == nullptr) {
2214 place_holder = new Node(1);
2215 place_holder->init_req(0, n);
2216 }
2217 igvn.replace_input_of(u, l, place_holder);
2218 success = true;
2219 }
2220 }
2221 if (success) {
2222 --k;
2223 }
2224 }
2225 }
2226 }
2227 if (n->is_Phi()) {
2228 stack.push(n, 1);
2229 n = n->in(1);
2230 } else if (n->is_Mem()) {
2231 stack.push(n, n->req());
2232 n = n->in(MemNode::Memory);
2233 } else {
2234 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2235 stack.push(n, n->req());
2236 n = n->in(0)->in(TypeFunc::Memory);
2237 }
2238 } else {
2239 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2240 // Build a new MergeMem node to carry the new memory state
2241 // as we build it. IGVN should fold extraneous MergeMem
2242 // nodes.
2243 mm = MergeMemNode::make(n);
2244 igvn.register_new_node_with_optimizer(mm);
2245 while (stack.size() > 0) {
2246 Node* m = stack.node();
2247 uint idx = stack.index();
2248 if (m->is_Mem()) {
2249 // Move memory node to its new slice
2250 const TypePtr* adr_type = m->adr_type();
2251 int alias = get_alias_index(adr_type);
2252 Node* prev = mm->memory_at(alias);
2253 igvn.replace_input_of(m, MemNode::Memory, prev);
2254 mm->set_memory_at(alias, m);
2255 } else if (m->is_Phi()) {
2256 // We need as many new phis as there are new aliases
2257 igvn.replace_input_of(m, idx, mm);
2258 if (idx == m->req()-1) {
2259 Node* r = m->in(0);
2260 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2261 const TypePtr* adr_type = get_adr_type(j);
2262 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2263 continue;
2264 }
2265 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2266 igvn.register_new_node_with_optimizer(phi);
2267 for (uint k = 1; k < m->req(); k++) {
2268 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2269 }
2270 mm->set_memory_at(j, phi);
2271 }
2272 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2273 igvn.register_new_node_with_optimizer(base_phi);
2274 for (uint k = 1; k < m->req(); k++) {
2275 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2276 }
2277 mm->set_base_memory(base_phi);
2278 }
2279 } else {
2280 // This is a MemBarCPUOrder node from
2281 // Parse::array_load()/Parse::array_store(), in the
2282 // branch that handles flat arrays hidden under
2283 // an Object[] array. We also need one new membar per
2284 // new alias to keep the unknown access that the
2285 // membars protect properly ordered with accesses to
2286 // known flat array.
2287 assert(m->is_Proj(), "projection expected");
2288 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2289 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2290 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2291 const TypePtr* adr_type = get_adr_type(j);
2292 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2293 continue;
2294 }
2295 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2296 igvn.register_new_node_with_optimizer(mb);
2297 Node* mem = mm->memory_at(j);
2298 mb->init_req(TypeFunc::Control, ctrl);
2299 mb->init_req(TypeFunc::Memory, mem);
2300 ctrl = new ProjNode(mb, TypeFunc::Control);
2301 igvn.register_new_node_with_optimizer(ctrl);
2302 mem = new ProjNode(mb, TypeFunc::Memory);
2303 igvn.register_new_node_with_optimizer(mem);
2304 mm->set_memory_at(j, mem);
2305 }
2306 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2307 }
2308 if (idx < m->req()-1) {
2309 idx += 1;
2310 stack.set_index(idx);
2311 n = m->in(idx);
2312 break;
2313 }
2314 // Take care of place holder nodes
2315 if (m->has_out_with(Op_Node)) {
2316 Node* place_holder = m->find_out_with(Op_Node);
2317 if (place_holder != nullptr) {
2318 Node* mm_clone = mm->clone();
2319 igvn.register_new_node_with_optimizer(mm_clone);
2320 Node* hook = new Node(1);
2321 hook->init_req(0, mm);
2322 igvn.replace_node(place_holder, mm_clone);
2323 hook->destruct(&igvn);
2324 }
2325 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2326 }
2327 stack.pop();
2328 }
2329 }
2330 } while(stack.size() > 0);
2331 // Fix the memory state at the MergeMem we started from
2332 igvn.rehash_node_delayed(current);
2333 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2334 const TypePtr* adr_type = get_adr_type(j);
2335 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2336 continue;
2337 }
2338 current->set_memory_at(j, mm);
2339 }
2340 current->set_memory_at(index, current->base_memory());
2341 }
2342 igvn.optimize();
2343 }
2344 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2345 #ifdef ASSERT
2346 if (!_flat_accesses_share_alias) {
2347 wq.clear();
2348 wq.push(root());
2349 for (uint i = 0; i < wq.size(); i++) {
2350 Node* n = wq.at(i);
2351 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2352 for (uint j = 0; j < n->req(); j++) {
2353 Node* m = n->in(j);
2354 if (m != nullptr) {
2355 wq.push(m);
2356 }
2357 }
2358 }
2359 }
2360 #endif
2361 }
2362
2363 void Compile::record_for_merge_stores_igvn(Node* n) {
2364 if (!n->for_merge_stores_igvn()) {
2365 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
2366 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2367 _for_merge_stores_igvn.append(n);
2368 }
2369 }
2370
2371 void Compile::remove_from_merge_stores_igvn(Node* n) {
2372 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2373 _for_merge_stores_igvn.remove(n);
2374 }
2375
2376 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
2377 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
2378 // the stores, and we merge the wrong sequence of stores.
2379 // Example:
2380 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
2381 // Apply MergeStores:
2382 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
2383 // Remove more RangeChecks:
2384 // StoreI [ StoreL ] StoreI
2385 // But now it would have been better to do this instead:
2386 // [ StoreL ] [ StoreL ]
2387 //
2388 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
2389 // since we never unset _merge_stores_phase.
2390 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
2391 C->set_merge_stores_phase();
2392
2393 if (_for_merge_stores_igvn.length() > 0) {
2394 while (_for_merge_stores_igvn.length() > 0) {
2395 Node* n = _for_merge_stores_igvn.pop();
2396 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2397 igvn._worklist.push(n);
2398 }
2399 igvn.optimize();
2400 if (failing()) return;
2401 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
2402 print_method(PHASE_AFTER_MERGE_STORES, 3);
2403 }
2404 }
2405
2406 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2407 if (OptimizeUnstableIf) {
2408 _unstable_if_traps.append(trap);
2409 }
2410 }
2411
2412 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2413 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2414 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2415 Node* n = trap->uncommon_trap();
2416 if (!useful.member(n)) {
2417 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2418 }
2419 }
2420 }
2421
2422 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2423 // or fold-compares case. Return true if succeed or not found.
2424 //
2425 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2426 // false and ask fold-compares to yield.
2427 //
2428 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2429 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2430 // when deoptimization does happen.
2431 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2432 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2433 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2434 if (trap->uncommon_trap() == unc) {
2435 if (yield && trap->modified()) {
2436 return false;
2437 }
2438 _unstable_if_traps.delete_at(i);
2439 break;
2440 }
2441 }
2442 return true;
2443 }
2444
2445 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2446 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2447 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2448 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2449 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2450 CallStaticJavaNode* unc = trap->uncommon_trap();
2451 int next_bci = trap->next_bci();
2452 bool modified = trap->modified();
2453
2454 if (next_bci != -1 && !modified) {
2455 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2456 JVMState* jvms = unc->jvms();
2457 ciMethod* method = jvms->method();
2458 ciBytecodeStream iter(method);
2459
2460 iter.force_bci(jvms->bci());
2461 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2462 Bytecodes::Code c = iter.cur_bc();
2463 Node* lhs = nullptr;
2464 Node* rhs = nullptr;
2465 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2466 lhs = unc->peek_operand(0);
2467 rhs = unc->peek_operand(1);
2468 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2469 lhs = unc->peek_operand(0);
2470 }
2471
2472 ResourceMark rm;
2473 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2474 assert(live_locals.is_valid(), "broken liveness info");
2475 int len = (int)live_locals.size();
2476
2477 for (int i = 0; i < len; i++) {
2478 Node* local = unc->local(jvms, i);
2479 // kill local using the liveness of next_bci.
2480 // give up when the local looks like an operand to secure reexecution.
2481 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2482 uint idx = jvms->locoff() + i;
2483 #ifdef ASSERT
2484 if (PrintOpto && Verbose) {
2485 tty->print("[unstable_if] kill local#%d: ", idx);
2486 local->dump();
2487 tty->cr();
2488 }
2489 #endif
2490 igvn.replace_input_of(unc, idx, top());
2491 modified = true;
2492 }
2493 }
2494 }
2495
2496 // keep the modified trap for late query
2497 if (modified) {
2498 trap->set_modified();
2499 } else {
2500 _unstable_if_traps.delete_at(i);
2501 }
2502 }
2503 igvn.optimize();
2504 }
2505
2506 // StringOpts and late inlining of string methods
2507 void Compile::inline_string_calls(bool parse_time) {
2508 {
2509 // remove useless nodes to make the usage analysis simpler
2510 ResourceMark rm;
2511 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2512 }
2513
2514 {
2515 ResourceMark rm;
2516 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2517 PhaseStringOpts pso(initial_gvn());
2518 print_method(PHASE_AFTER_STRINGOPTS, 3);
2519 }
2520
2521 // now inline anything that we skipped the first time around
2522 if (!parse_time) {
2523 _late_inlines_pos = _late_inlines.length();
2524 }
2525
2526 while (_string_late_inlines.length() > 0) {
2527 CallGenerator* cg = _string_late_inlines.pop();
2528 cg->do_late_inline();
2529 if (failing()) return;
2530 }
2531 _string_late_inlines.trunc_to(0);
2532 }
2533
2534 // Late inlining of boxing methods
2535 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2536 if (_boxing_late_inlines.length() > 0) {
2537 assert(has_boxed_value(), "inconsistent");
2538
2539 set_inlining_incrementally(true);
2540
2541 igvn_worklist()->ensure_empty(); // should be done with igvn
2542
2543 _late_inlines_pos = _late_inlines.length();
2544
2545 while (_boxing_late_inlines.length() > 0) {
2546 CallGenerator* cg = _boxing_late_inlines.pop();
2547 cg->do_late_inline();
2548 if (failing()) return;
2549 }
2550 _boxing_late_inlines.trunc_to(0);
2551
2552 inline_incrementally_cleanup(igvn);
2553
2554 set_inlining_incrementally(false);
2555 }
2556 }
2557
2558 bool Compile::inline_incrementally_one() {
2559 assert(IncrementalInline, "incremental inlining should be on");
2560
2561 TracePhase tp(_t_incrInline_inline);
2562
2563 set_inlining_progress(false);
2564 set_do_cleanup(false);
2565
2566 for (int i = 0; i < _late_inlines.length(); i++) {
2567 _late_inlines_pos = i+1;
2568 CallGenerator* cg = _late_inlines.at(i);
2569 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2570 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2571 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2572 cg->do_late_inline();
2573 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2574 if (failing()) {
2575 return false;
2576 } else if (inlining_progress()) {
2577 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2578 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2579 break; // process one call site at a time
2580 } else {
2581 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2582 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2583 // Avoid potential infinite loop if node already in the IGVN list
2584 assert(false, "scheduled for IGVN during inlining attempt");
2585 } else {
2586 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2587 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2588 cg->call_node()->set_generator(cg);
2589 }
2590 }
2591 } else {
2592 // Ignore late inline direct calls when inlining is not allowed.
2593 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2594 }
2595 }
2596 // Remove processed elements.
2597 _late_inlines.remove_till(_late_inlines_pos);
2598 _late_inlines_pos = 0;
2599
2600 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2601
2602 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2603
2604 set_inlining_progress(false);
2605 set_do_cleanup(false);
2606
2607 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2608 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2609 }
2610
2611 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2612 {
2613 TracePhase tp(_t_incrInline_pru);
2614 ResourceMark rm;
2615 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2616 }
2617 {
2618 TracePhase tp(_t_incrInline_igvn);
2619 igvn.reset_from_gvn(initial_gvn());
2620 igvn.optimize();
2621 if (failing()) return;
2622 }
2623 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2624 }
2625
2626 // Perform incremental inlining until bound on number of live nodes is reached
2627 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2628 TracePhase tp(_t_incrInline);
2629
2630 set_inlining_incrementally(true);
2631 uint low_live_nodes = 0;
2632
2633 while (_late_inlines.length() > 0) {
2634 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2635 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2636 TracePhase tp(_t_incrInline_ideal);
2637 // PhaseIdealLoop is expensive so we only try it once we are
2638 // out of live nodes and we only try it again if the previous
2639 // helped got the number of nodes down significantly
2640 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2641 if (failing()) return;
2642 low_live_nodes = live_nodes();
2643 _major_progress = true;
2644 }
2645
2646 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2647 bool do_print_inlining = print_inlining() || print_intrinsics();
2648 if (do_print_inlining || log() != nullptr) {
2649 // Print inlining message for candidates that we couldn't inline for lack of space.
2650 for (int i = 0; i < _late_inlines.length(); i++) {
2651 CallGenerator* cg = _late_inlines.at(i);
2652 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2653 if (do_print_inlining) {
2654 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2655 }
2656 log_late_inline_failure(cg, msg);
2657 }
2658 }
2659 break; // finish
2660 }
2661 }
2662
2663 igvn_worklist()->ensure_empty(); // should be done with igvn
2664
2665 while (inline_incrementally_one()) {
2666 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2667 }
2668 if (failing()) return;
2669
2670 inline_incrementally_cleanup(igvn);
2671
2672 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2673
2674 if (failing()) return;
2675
2676 if (_late_inlines.length() == 0) {
2677 break; // no more progress
2678 }
2679 }
2680
2681 igvn_worklist()->ensure_empty(); // should be done with igvn
2682
2683 if (_string_late_inlines.length() > 0) {
2684 assert(has_stringbuilder(), "inconsistent");
2685
2686 inline_string_calls(false);
2687
2688 if (failing()) return;
2689
2690 inline_incrementally_cleanup(igvn);
2691 }
2692
2693 set_inlining_incrementally(false);
2694 }
2695
2696 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2697 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2698 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2699 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2700 // as if "inlining_incrementally() == true" were set.
2701 assert(inlining_incrementally() == false, "not allowed");
2702 #ifdef ASSERT
2703 Unique_Node_List* modified_nodes = _modified_nodes;
2704 _modified_nodes = nullptr;
2705 #endif
2706 assert(_late_inlines.length() > 0, "sanity");
2707
2708 while (_late_inlines.length() > 0) {
2709 igvn_worklist()->ensure_empty(); // should be done with igvn
2710
2711 while (inline_incrementally_one()) {
2712 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2713 }
2714 if (failing()) return;
2715
2716 inline_incrementally_cleanup(igvn);
2717 }
2718 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2719 }
2720
2721 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2722 if (_loop_opts_cnt > 0) {
2723 while (major_progress() && (_loop_opts_cnt > 0)) {
2724 TracePhase tp(_t_idealLoop);
2725 PhaseIdealLoop::optimize(igvn, mode);
2726 _loop_opts_cnt--;
2727 if (failing()) return false;
2728 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2729 }
2730 }
2731 return true;
2732 }
2733
2734 // Remove edges from "root" to each SafePoint at a backward branch.
2735 // They were inserted during parsing (see add_safepoint()) to make
2736 // infinite loops without calls or exceptions visible to root, i.e.,
2737 // useful.
2738 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2739 Node *r = root();
2740 if (r != nullptr) {
2741 for (uint i = r->req(); i < r->len(); ++i) {
2742 Node *n = r->in(i);
2743 if (n != nullptr && n->is_SafePoint()) {
2744 r->rm_prec(i);
2745 if (n->outcnt() == 0) {
2746 igvn.remove_dead_node(n);
2747 }
2748 --i;
2749 }
2750 }
2751 // Parsing may have added top inputs to the root node (Path
2752 // leading to the Halt node proven dead). Make sure we get a
2753 // chance to clean them up.
2754 igvn._worklist.push(r);
2755 igvn.optimize();
2756 }
2757 }
2758
2759 //------------------------------Optimize---------------------------------------
2760 // Given a graph, optimize it.
2761 void Compile::Optimize() {
2762 TracePhase tp(_t_optimizer);
2763
2764 #ifndef PRODUCT
2765 if (env()->break_at_compile()) {
2766 BREAKPOINT;
2767 }
2768
2769 #endif
2770
2771 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2772 #ifdef ASSERT
2773 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2774 #endif
2775
2776 ResourceMark rm;
2777
2778 NOT_PRODUCT( verify_graph_edges(); )
2779
2780 print_method(PHASE_AFTER_PARSING, 1);
2781
2782 {
2783 // Iterative Global Value Numbering, including ideal transforms
2784 // Initialize IterGVN with types and values from parse-time GVN
2785 PhaseIterGVN igvn(initial_gvn());
2786 #ifdef ASSERT
2787 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2788 #endif
2789 {
2790 TracePhase tp(_t_iterGVN);
2791 igvn.optimize();
2792 }
2793
2794 if (failing()) return;
2795
2796 print_method(PHASE_ITER_GVN1, 2);
2797
2798 process_for_unstable_if_traps(igvn);
2799
2800 if (failing()) return;
2801
2802 inline_incrementally(igvn);
2803
2804 print_method(PHASE_INCREMENTAL_INLINE, 2);
2805
2806 if (failing()) return;
2807
2808 if (eliminate_boxing()) {
2809 // Inline valueOf() methods now.
2810 inline_boxing_calls(igvn);
2811
2812 if (failing()) return;
2813
2814 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2815 inline_incrementally(igvn);
2816 }
2817
2818 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2819
2820 if (failing()) return;
2821 }
2822
2823 // Remove the speculative part of types and clean up the graph from
2824 // the extra CastPP nodes whose only purpose is to carry them. Do
2825 // that early so that optimizations are not disrupted by the extra
2826 // CastPP nodes.
2827 remove_speculative_types(igvn);
2828
2829 if (failing()) return;
2830
2831 // No more new expensive nodes will be added to the list from here
2832 // so keep only the actual candidates for optimizations.
2833 cleanup_expensive_nodes(igvn);
2834
2835 if (failing()) return;
2836
2837 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2838 if (EnableVectorSupport && has_vbox_nodes()) {
2839 TracePhase tp(_t_vector);
2840 PhaseVector pv(igvn);
2841 pv.optimize_vector_boxes();
2842 if (failing()) return;
2843 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2844 }
2845 assert(!has_vbox_nodes(), "sanity");
2846
2847 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2848 Compile::TracePhase tp(_t_renumberLive);
2849 igvn_worklist()->ensure_empty(); // should be done with igvn
2850 {
2851 ResourceMark rm;
2852 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2853 }
2854 igvn.reset_from_gvn(initial_gvn());
2855 igvn.optimize();
2856 if (failing()) return;
2857 }
2858
2859 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2860 // safepoints
2861 remove_root_to_sfpts_edges(igvn);
2862
2863 // Process inline type nodes now that all inlining is over
2864 process_inline_types(igvn);
2865
2866 adjust_flat_array_access_aliases(igvn);
2867
2868 if (failing()) return;
2869
2870 if (C->macro_count() > 0) {
2871 // Eliminate some macro nodes before EA to reduce analysis pressure
2872 PhaseMacroExpand mexp(igvn);
2873 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
2874 if (failing()) {
2875 return;
2876 }
2877 igvn.set_delay_transform(false);
2878 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2879 }
2880
2881 if (has_loops()) {
2882 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2883 }
2884
2885 // Perform escape analysis
2886 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2887 if (has_loops()) {
2888 // Cleanup graph (remove dead nodes).
2889 TracePhase tp(_t_idealLoop);
2890 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2891 if (failing()) {
2892 return;
2893 }
2894 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2895 if (C->macro_count() > 0) {
2896 // Eliminate some macro nodes before EA to reduce analysis pressure
2897 PhaseMacroExpand mexp(igvn);
2898 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
2899 if (failing()) {
2900 return;
2901 }
2902 igvn.set_delay_transform(false);
2903 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2904 }
2905 }
2906
2907 bool progress;
2908 do {
2909 ConnectionGraph::do_analysis(this, &igvn);
2910
2911 if (failing()) return;
2912
2913 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2914
2915 // Optimize out fields loads from scalar replaceable allocations.
2916 igvn.optimize();
2917 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2918
2919 if (failing()) return;
2920
2921 if (congraph() != nullptr && macro_count() > 0) {
2922 TracePhase tp(_t_macroEliminate);
2923 PhaseMacroExpand mexp(igvn);
2924 mexp.eliminate_macro_nodes();
2925 if (failing()) {
2926 return;
2927 }
2928 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2929
2930 igvn.set_delay_transform(false);
2931 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2932 }
2933
2934 ConnectionGraph::verify_ram_nodes(this, root());
2935 if (failing()) return;
2936
2937 progress = do_iterative_escape_analysis() &&
2938 (macro_count() < mcount) &&
2939 ConnectionGraph::has_candidates(this);
2940 // Try again if candidates exist and made progress
2941 // by removing some allocations and/or locks.
2942 } while (progress);
2943 }
2944
2945 // Loop transforms on the ideal graph. Range Check Elimination,
2946 // peeling, unrolling, etc.
2947
2948 // Set loop opts counter
2949 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2950 {
2951 TracePhase tp(_t_idealLoop);
2952 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2953 _loop_opts_cnt--;
2954 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2955 if (failing()) return;
2956 }
2957 // Loop opts pass if partial peeling occurred in previous pass
2958 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2959 TracePhase tp(_t_idealLoop);
2960 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2961 _loop_opts_cnt--;
2962 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2963 if (failing()) return;
2964 }
2965 // Loop opts pass for loop-unrolling before CCP
2966 if(major_progress() && (_loop_opts_cnt > 0)) {
2967 TracePhase tp(_t_idealLoop);
2968 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2969 _loop_opts_cnt--;
2970 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2971 }
2972 if (!failing()) {
2973 // Verify that last round of loop opts produced a valid graph
2974 PhaseIdealLoop::verify(igvn);
2975 }
2976 }
2977 if (failing()) return;
2978
2979 // Conditional Constant Propagation;
2980 print_method(PHASE_BEFORE_CCP1, 2);
2981 PhaseCCP ccp( &igvn );
2982 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2983 {
2984 TracePhase tp(_t_ccp);
2985 ccp.do_transform();
2986 }
2987 print_method(PHASE_CCP1, 2);
2988
2989 assert( true, "Break here to ccp.dump_old2new_map()");
2990
2991 // Iterative Global Value Numbering, including ideal transforms
2992 {
2993 TracePhase tp(_t_iterGVN2);
2994 igvn.reset_from_igvn(&ccp);
2995 igvn.optimize();
2996 }
2997 print_method(PHASE_ITER_GVN2, 2);
2998
2999 if (failing()) return;
3000
3001 // Loop transforms on the ideal graph. Range Check Elimination,
3002 // peeling, unrolling, etc.
3003 if (!optimize_loops(igvn, LoopOptsDefault)) {
3004 return;
3005 }
3006
3007 if (failing()) return;
3008
3009 C->clear_major_progress(); // ensure that major progress is now clear
3010
3011 process_for_post_loop_opts_igvn(igvn);
3012
3013 process_for_merge_stores_igvn(igvn);
3014
3015 if (failing()) return;
3016
3017 #ifdef ASSERT
3018 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
3019 #endif
3020
3021 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
3022
3023 if (_late_inlines.length() > 0) {
3024 // More opportunities to optimize virtual and MH calls.
3025 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
3026 process_late_inline_calls_no_inline(igvn);
3027 }
3028
3029 {
3030 TracePhase tp(_t_macroExpand);
3031 PhaseMacroExpand mex(igvn);
3032 // Last attempt to eliminate macro nodes.
3033 mex.eliminate_macro_nodes();
3034 if (failing()) {
3035 return;
3036 }
3037
3038 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
3039 // Do not allow new macro nodes once we start to eliminate and expand
3040 C->reset_allow_macro_nodes();
3041 // Last attempt to eliminate macro nodes before expand
3042 mex.eliminate_macro_nodes();
3043 if (failing()) {
3044 return;
3045 }
3046 mex.eliminate_opaque_looplimit_macro_nodes();
3047 if (failing()) {
3048 return;
3049 }
3050 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
3051 if (mex.expand_macro_nodes()) {
3052 assert(failing(), "must bail out w/ explicit message");
3053 return;
3054 }
3055 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
3056 }
3057
3058 // Process inline type nodes again and remove them. From here
3059 // on we don't need to keep track of field values anymore.
3060 process_inline_types(igvn, /* remove= */ true);
3061
3062 {
3063 TracePhase tp(_t_barrierExpand);
3064 if (bs->expand_barriers(this, igvn)) {
3065 assert(failing(), "must bail out w/ explicit message");
3066 return;
3067 }
3068 print_method(PHASE_BARRIER_EXPANSION, 2);
3069 }
3070
3071 if (C->max_vector_size() > 0) {
3072 C->optimize_logic_cones(igvn);
3073 igvn.optimize();
3074 if (failing()) return;
3075 }
3076
3077 DEBUG_ONLY( _modified_nodes = nullptr; )
3078 DEBUG_ONLY( _late_inlines.clear(); )
3079
3080 assert(igvn._worklist.size() == 0, "not empty");
3081 } // (End scope of igvn; run destructor if necessary for asserts.)
3082
3083 check_no_dead_use();
3084
3085 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
3086 // to remove hashes to unlock nodes for modifications.
3087 C->node_hash()->clear();
3088
3089 // A method with only infinite loops has no edges entering loops from root
3090 {
3091 TracePhase tp(_t_graphReshaping);
3092 if (final_graph_reshaping()) {
3093 assert(failing(), "must bail out w/ explicit message");
3094 return;
3095 }
3096 }
3097
3098 print_method(PHASE_OPTIMIZE_FINISHED, 2);
3099 DEBUG_ONLY(set_phase_optimize_finished();)
3100 }
3101
3102 #ifdef ASSERT
3103 void Compile::check_no_dead_use() const {
3104 ResourceMark rm;
3105 Unique_Node_List wq;
3106 wq.push(root());
3107 for (uint i = 0; i < wq.size(); ++i) {
3108 Node* n = wq.at(i);
3109 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3110 Node* u = n->fast_out(j);
3111 if (u->outcnt() == 0 && !u->is_Con()) {
3112 u->dump();
3113 fatal("no reachable node should have no use");
3114 }
3115 wq.push(u);
3116 }
3117 }
3118 }
3119 #endif
3120
3121 void Compile::inline_vector_reboxing_calls() {
3122 if (C->_vector_reboxing_late_inlines.length() > 0) {
3123 _late_inlines_pos = C->_late_inlines.length();
3124 while (_vector_reboxing_late_inlines.length() > 0) {
3125 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
3126 cg->do_late_inline();
3127 if (failing()) return;
3128 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
3129 }
3130 _vector_reboxing_late_inlines.trunc_to(0);
3131 }
3132 }
3133
3134 bool Compile::has_vbox_nodes() {
3135 if (C->_vector_reboxing_late_inlines.length() > 0) {
3136 return true;
3137 }
3138 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3139 Node * n = C->macro_node(macro_idx);
3140 assert(n->is_macro(), "only macro nodes expected here");
3141 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3142 return true;
3143 }
3144 }
3145 return false;
3146 }
3147
3148 //---------------------------- Bitwise operation packing optimization ---------------------------
3149
3150 static bool is_vector_unary_bitwise_op(Node* n) {
3151 return n->Opcode() == Op_XorV &&
3152 VectorNode::is_vector_bitwise_not_pattern(n);
3153 }
3154
3155 static bool is_vector_binary_bitwise_op(Node* n) {
3156 switch (n->Opcode()) {
3157 case Op_AndV:
3158 case Op_OrV:
3159 return true;
3160
3161 case Op_XorV:
3162 return !is_vector_unary_bitwise_op(n);
3163
3164 default:
3165 return false;
3166 }
3167 }
3168
3169 static bool is_vector_ternary_bitwise_op(Node* n) {
3170 return n->Opcode() == Op_MacroLogicV;
3171 }
3172
3173 static bool is_vector_bitwise_op(Node* n) {
3174 return is_vector_unary_bitwise_op(n) ||
3175 is_vector_binary_bitwise_op(n) ||
3176 is_vector_ternary_bitwise_op(n);
3177 }
3178
3179 static bool is_vector_bitwise_cone_root(Node* n) {
3180 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
3181 return false;
3182 }
3183 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3184 if (is_vector_bitwise_op(n->fast_out(i))) {
3185 return false;
3186 }
3187 }
3188 return true;
3189 }
3190
3191 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3192 uint cnt = 0;
3193 if (is_vector_bitwise_op(n)) {
3194 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3195 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3196 for (uint i = 1; i < inp_cnt; i++) {
3197 Node* in = n->in(i);
3198 bool skip = VectorNode::is_all_ones_vector(in);
3199 if (!skip && !inputs.member(in)) {
3200 inputs.push(in);
3201 cnt++;
3202 }
3203 }
3204 assert(cnt <= 1, "not unary");
3205 } else {
3206 uint last_req = inp_cnt;
3207 if (is_vector_ternary_bitwise_op(n)) {
3208 last_req = inp_cnt - 1; // skip last input
3209 }
3210 for (uint i = 1; i < last_req; i++) {
3211 Node* def = n->in(i);
3212 if (!inputs.member(def)) {
3213 inputs.push(def);
3214 cnt++;
3215 }
3216 }
3217 }
3218 } else { // not a bitwise operations
3219 if (!inputs.member(n)) {
3220 inputs.push(n);
3221 cnt++;
3222 }
3223 }
3224 return cnt;
3225 }
3226
3227 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3228 Unique_Node_List useful_nodes;
3229 C->identify_useful_nodes(useful_nodes);
3230
3231 for (uint i = 0; i < useful_nodes.size(); i++) {
3232 Node* n = useful_nodes.at(i);
3233 if (is_vector_bitwise_cone_root(n)) {
3234 list.push(n);
3235 }
3236 }
3237 }
3238
3239 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3240 const TypeVect* vt,
3241 Unique_Node_List& partition,
3242 Unique_Node_List& inputs) {
3243 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3244 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3245 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3246
3247 Node* in1 = inputs.at(0);
3248 Node* in2 = inputs.at(1);
3249 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3250
3251 uint func = compute_truth_table(partition, inputs);
3252
3253 Node* pn = partition.at(partition.size() - 1);
3254 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3255 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3256 }
3257
3258 static uint extract_bit(uint func, uint pos) {
3259 return (func & (1 << pos)) >> pos;
3260 }
3261
3262 //
3263 // A macro logic node represents a truth table. It has 4 inputs,
3264 // First three inputs corresponds to 3 columns of a truth table
3265 // and fourth input captures the logic function.
3266 //
3267 // eg. fn = (in1 AND in2) OR in3;
3268 //
3269 // MacroNode(in1,in2,in3,fn)
3270 //
3271 // -----------------
3272 // in1 in2 in3 fn
3273 // -----------------
3274 // 0 0 0 0
3275 // 0 0 1 1
3276 // 0 1 0 0
3277 // 0 1 1 1
3278 // 1 0 0 0
3279 // 1 0 1 1
3280 // 1 1 0 1
3281 // 1 1 1 1
3282 //
3283
3284 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3285 int res = 0;
3286 for (int i = 0; i < 8; i++) {
3287 int bit1 = extract_bit(in1, i);
3288 int bit2 = extract_bit(in2, i);
3289 int bit3 = extract_bit(in3, i);
3290
3291 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3292 int func_bit = extract_bit(func, func_bit_pos);
3293
3294 res |= func_bit << i;
3295 }
3296 return res;
3297 }
3298
3299 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
3300 assert(n != nullptr, "");
3301 assert(eval_map.contains(n), "absent");
3302 return *(eval_map.get(n));
3303 }
3304
3305 static void eval_operands(Node* n,
3306 uint& func1, uint& func2, uint& func3,
3307 ResourceHashtable<Node*,uint>& eval_map) {
3308 assert(is_vector_bitwise_op(n), "");
3309
3310 if (is_vector_unary_bitwise_op(n)) {
3311 Node* opnd = n->in(1);
3312 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3313 opnd = n->in(2);
3314 }
3315 func1 = eval_operand(opnd, eval_map);
3316 } else if (is_vector_binary_bitwise_op(n)) {
3317 func1 = eval_operand(n->in(1), eval_map);
3318 func2 = eval_operand(n->in(2), eval_map);
3319 } else {
3320 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3321 func1 = eval_operand(n->in(1), eval_map);
3322 func2 = eval_operand(n->in(2), eval_map);
3323 func3 = eval_operand(n->in(3), eval_map);
3324 }
3325 }
3326
3327 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3328 assert(inputs.size() <= 3, "sanity");
3329 ResourceMark rm;
3330 uint res = 0;
3331 ResourceHashtable<Node*,uint> eval_map;
3332
3333 // Populate precomputed functions for inputs.
3334 // Each input corresponds to one column of 3 input truth-table.
3335 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3336 0xCC, // (_, b, _) -> b
3337 0xF0 }; // (a, _, _) -> a
3338 for (uint i = 0; i < inputs.size(); i++) {
3339 eval_map.put(inputs.at(i), input_funcs[2-i]);
3340 }
3341
3342 for (uint i = 0; i < partition.size(); i++) {
3343 Node* n = partition.at(i);
3344
3345 uint func1 = 0, func2 = 0, func3 = 0;
3346 eval_operands(n, func1, func2, func3, eval_map);
3347
3348 switch (n->Opcode()) {
3349 case Op_OrV:
3350 assert(func3 == 0, "not binary");
3351 res = func1 | func2;
3352 break;
3353 case Op_AndV:
3354 assert(func3 == 0, "not binary");
3355 res = func1 & func2;
3356 break;
3357 case Op_XorV:
3358 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3359 assert(func2 == 0 && func3 == 0, "not unary");
3360 res = (~func1) & 0xFF;
3361 } else {
3362 assert(func3 == 0, "not binary");
3363 res = func1 ^ func2;
3364 }
3365 break;
3366 case Op_MacroLogicV:
3367 // Ordering of inputs may change during evaluation of sub-tree
3368 // containing MacroLogic node as a child node, thus a re-evaluation
3369 // makes sure that function is evaluated in context of current
3370 // inputs.
3371 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3372 break;
3373
3374 default: assert(false, "not supported: %s", n->Name());
3375 }
3376 assert(res <= 0xFF, "invalid");
3377 eval_map.put(n, res);
3378 }
3379 return res;
3380 }
3381
3382 // Criteria under which nodes gets packed into a macro logic node:-
3383 // 1) Parent and both child nodes are all unmasked or masked with
3384 // same predicates.
3385 // 2) Masked parent can be packed with left child if it is predicated
3386 // and both have same predicates.
3387 // 3) Masked parent can be packed with right child if its un-predicated
3388 // or has matching predication condition.
3389 // 4) An unmasked parent can be packed with an unmasked child.
3390 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3391 assert(partition.size() == 0, "not empty");
3392 assert(inputs.size() == 0, "not empty");
3393 if (is_vector_ternary_bitwise_op(n)) {
3394 return false;
3395 }
3396
3397 bool is_unary_op = is_vector_unary_bitwise_op(n);
3398 if (is_unary_op) {
3399 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3400 return false; // too few inputs
3401 }
3402
3403 bool pack_left_child = true;
3404 bool pack_right_child = true;
3405
3406 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3407 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3408
3409 int left_child_input_cnt = 0;
3410 int right_child_input_cnt = 0;
3411
3412 bool parent_is_predicated = n->is_predicated_vector();
3413 bool left_child_predicated = n->in(1)->is_predicated_vector();
3414 bool right_child_predicated = n->in(2)->is_predicated_vector();
3415
3416 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
3417 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3418 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3419
3420 do {
3421 if (pack_left_child && left_child_LOP &&
3422 ((!parent_is_predicated && !left_child_predicated) ||
3423 ((parent_is_predicated && left_child_predicated &&
3424 parent_pred == left_child_pred)))) {
3425 partition.push(n->in(1));
3426 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3427 } else {
3428 inputs.push(n->in(1));
3429 left_child_input_cnt = 1;
3430 }
3431
3432 if (pack_right_child && right_child_LOP &&
3433 (!right_child_predicated ||
3434 (right_child_predicated && parent_is_predicated &&
3435 parent_pred == right_child_pred))) {
3436 partition.push(n->in(2));
3437 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3438 } else {
3439 inputs.push(n->in(2));
3440 right_child_input_cnt = 1;
3441 }
3442
3443 if (inputs.size() > 3) {
3444 assert(partition.size() > 0, "");
3445 inputs.clear();
3446 partition.clear();
3447 if (left_child_input_cnt > right_child_input_cnt) {
3448 pack_left_child = false;
3449 } else {
3450 pack_right_child = false;
3451 }
3452 } else {
3453 break;
3454 }
3455 } while(true);
3456
3457 if(partition.size()) {
3458 partition.push(n);
3459 }
3460
3461 return (partition.size() == 2 || partition.size() == 3) &&
3462 (inputs.size() == 2 || inputs.size() == 3);
3463 }
3464
3465 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3466 assert(is_vector_bitwise_op(n), "not a root");
3467
3468 visited.set(n->_idx);
3469
3470 // 1) Do a DFS walk over the logic cone.
3471 for (uint i = 1; i < n->req(); i++) {
3472 Node* in = n->in(i);
3473 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3474 process_logic_cone_root(igvn, in, visited);
3475 }
3476 }
3477
3478 // 2) Bottom up traversal: Merge node[s] with
3479 // the parent to form macro logic node.
3480 Unique_Node_List partition;
3481 Unique_Node_List inputs;
3482 if (compute_logic_cone(n, partition, inputs)) {
3483 const TypeVect* vt = n->bottom_type()->is_vect();
3484 Node* pn = partition.at(partition.size() - 1);
3485 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3486 if (mask == nullptr ||
3487 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3488 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3489 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3490 igvn.replace_node(n, macro_logic);
3491 }
3492 }
3493 }
3494
3495 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3496 ResourceMark rm;
3497 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3498 Unique_Node_List list;
3499 collect_logic_cone_roots(list);
3500
3501 while (list.size() > 0) {
3502 Node* n = list.pop();
3503 const TypeVect* vt = n->bottom_type()->is_vect();
3504 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3505 if (supported) {
3506 VectorSet visited(comp_arena());
3507 process_logic_cone_root(igvn, n, visited);
3508 }
3509 }
3510 }
3511 }
3512
3513 //------------------------------Code_Gen---------------------------------------
3514 // Given a graph, generate code for it
3515 void Compile::Code_Gen() {
3516 if (failing()) {
3517 return;
3518 }
3519
3520 // Perform instruction selection. You might think we could reclaim Matcher
3521 // memory PDQ, but actually the Matcher is used in generating spill code.
3522 // Internals of the Matcher (including some VectorSets) must remain live
3523 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3524 // set a bit in reclaimed memory.
3525
3526 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3527 // nodes. Mapping is only valid at the root of each matched subtree.
3528 NOT_PRODUCT( verify_graph_edges(); )
3529
3530 Matcher matcher;
3531 _matcher = &matcher;
3532 {
3533 TracePhase tp(_t_matcher);
3534 matcher.match();
3535 if (failing()) {
3536 return;
3537 }
3538 }
3539 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3540 // nodes. Mapping is only valid at the root of each matched subtree.
3541 NOT_PRODUCT( verify_graph_edges(); )
3542
3543 // If you have too many nodes, or if matching has failed, bail out
3544 check_node_count(0, "out of nodes matching instructions");
3545 if (failing()) {
3546 return;
3547 }
3548
3549 print_method(PHASE_MATCHING, 2);
3550
3551 // Build a proper-looking CFG
3552 PhaseCFG cfg(node_arena(), root(), matcher);
3553 if (failing()) {
3554 return;
3555 }
3556 _cfg = &cfg;
3557 {
3558 TracePhase tp(_t_scheduler);
3559 bool success = cfg.do_global_code_motion();
3560 if (!success) {
3561 return;
3562 }
3563
3564 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3565 NOT_PRODUCT( verify_graph_edges(); )
3566 cfg.verify();
3567 if (failing()) {
3568 return;
3569 }
3570 }
3571
3572 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3573 _regalloc = ®alloc;
3574 {
3575 TracePhase tp(_t_registerAllocation);
3576 // Perform register allocation. After Chaitin, use-def chains are
3577 // no longer accurate (at spill code) and so must be ignored.
3578 // Node->LRG->reg mappings are still accurate.
3579 _regalloc->Register_Allocate();
3580
3581 // Bail out if the allocator builds too many nodes
3582 if (failing()) {
3583 return;
3584 }
3585
3586 print_method(PHASE_REGISTER_ALLOCATION, 2);
3587 }
3588
3589 // Prior to register allocation we kept empty basic blocks in case the
3590 // the allocator needed a place to spill. After register allocation we
3591 // are not adding any new instructions. If any basic block is empty, we
3592 // can now safely remove it.
3593 {
3594 TracePhase tp(_t_blockOrdering);
3595 cfg.remove_empty_blocks();
3596 if (do_freq_based_layout()) {
3597 PhaseBlockLayout layout(cfg);
3598 } else {
3599 cfg.set_loop_alignment();
3600 }
3601 cfg.fixup_flow();
3602 cfg.remove_unreachable_blocks();
3603 cfg.verify_dominator_tree();
3604 print_method(PHASE_BLOCK_ORDERING, 3);
3605 }
3606
3607 // Apply peephole optimizations
3608 if( OptoPeephole ) {
3609 TracePhase tp(_t_peephole);
3610 PhasePeephole peep( _regalloc, cfg);
3611 peep.do_transform();
3612 print_method(PHASE_PEEPHOLE, 3);
3613 }
3614
3615 // Do late expand if CPU requires this.
3616 if (Matcher::require_postalloc_expand) {
3617 TracePhase tp(_t_postalloc_expand);
3618 cfg.postalloc_expand(_regalloc);
3619 print_method(PHASE_POSTALLOC_EXPAND, 3);
3620 }
3621
3622 #ifdef ASSERT
3623 {
3624 CompilationMemoryStatistic::do_test_allocations();
3625 if (failing()) return;
3626 }
3627 #endif
3628
3629 // Convert Nodes to instruction bits in a buffer
3630 {
3631 TracePhase tp(_t_output);
3632 PhaseOutput output;
3633 output.Output();
3634 if (failing()) return;
3635 output.install();
3636 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3637 }
3638
3639 // He's dead, Jim.
3640 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3641 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3642 }
3643
3644 //------------------------------Final_Reshape_Counts---------------------------
3645 // This class defines counters to help identify when a method
3646 // may/must be executed using hardware with only 24-bit precision.
3647 struct Final_Reshape_Counts : public StackObj {
3648 int _call_count; // count non-inlined 'common' calls
3649 int _float_count; // count float ops requiring 24-bit precision
3650 int _double_count; // count double ops requiring more precision
3651 int _java_call_count; // count non-inlined 'java' calls
3652 int _inner_loop_count; // count loops which need alignment
3653 VectorSet _visited; // Visitation flags
3654 Node_List _tests; // Set of IfNodes & PCTableNodes
3655
3656 Final_Reshape_Counts() :
3657 _call_count(0), _float_count(0), _double_count(0),
3658 _java_call_count(0), _inner_loop_count(0) { }
3659
3660 void inc_call_count () { _call_count ++; }
3661 void inc_float_count () { _float_count ++; }
3662 void inc_double_count() { _double_count++; }
3663 void inc_java_call_count() { _java_call_count++; }
3664 void inc_inner_loop_count() { _inner_loop_count++; }
3665
3666 int get_call_count () const { return _call_count ; }
3667 int get_float_count () const { return _float_count ; }
3668 int get_double_count() const { return _double_count; }
3669 int get_java_call_count() const { return _java_call_count; }
3670 int get_inner_loop_count() const { return _inner_loop_count; }
3671 };
3672
3673 //------------------------------final_graph_reshaping_impl----------------------
3674 // Implement items 1-5 from final_graph_reshaping below.
3675 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3676
3677 if ( n->outcnt() == 0 ) return; // dead node
3678 uint nop = n->Opcode();
3679
3680 // Check for 2-input instruction with "last use" on right input.
3681 // Swap to left input. Implements item (2).
3682 if( n->req() == 3 && // two-input instruction
3683 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3684 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3685 n->in(2)->outcnt() == 1 &&// right use IS a last use
3686 !n->in(2)->is_Con() ) { // right use is not a constant
3687 // Check for commutative opcode
3688 switch( nop ) {
3689 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3690 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3691 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3692 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3693 case Op_AndL: case Op_XorL: case Op_OrL:
3694 case Op_AndI: case Op_XorI: case Op_OrI: {
3695 // Move "last use" input to left by swapping inputs
3696 n->swap_edges(1, 2);
3697 break;
3698 }
3699 default:
3700 break;
3701 }
3702 }
3703
3704 #ifdef ASSERT
3705 if( n->is_Mem() ) {
3706 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3707 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3708 // oop will be recorded in oop map if load crosses safepoint
3709 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3710 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3711 "raw memory operations should have control edge");
3712 }
3713 if (n->is_MemBar()) {
3714 MemBarNode* mb = n->as_MemBar();
3715 if (mb->trailing_store() || mb->trailing_load_store()) {
3716 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3717 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3718 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3719 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3720 } else if (mb->leading()) {
3721 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3722 }
3723 }
3724 #endif
3725 // Count FPU ops and common calls, implements item (3)
3726 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3727 if (!gc_handled) {
3728 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3729 }
3730
3731 // Collect CFG split points
3732 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3733 frc._tests.push(n);
3734 }
3735 }
3736
3737 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3738 if (!UseDivMod) {
3739 return;
3740 }
3741
3742 // Check if "a % b" and "a / b" both exist
3743 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3744 if (d == nullptr) {
3745 return;
3746 }
3747
3748 // Replace them with a fused divmod if supported
3749 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3750 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3751 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3752 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3753 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3754 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3755 // DivMod node so the dependency is not lost.
3756 divmod->add_prec_from(n);
3757 divmod->add_prec_from(d);
3758 d->subsume_by(divmod->div_proj(), this);
3759 n->subsume_by(divmod->mod_proj(), this);
3760 } else {
3761 // Replace "a % b" with "a - ((a / b) * b)"
3762 Node* mult = MulNode::make(d, d->in(2), bt);
3763 Node* sub = SubNode::make(d->in(1), mult, bt);
3764 n->subsume_by(sub, this);
3765 }
3766 }
3767
3768 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3769 switch( nop ) {
3770 // Count all float operations that may use FPU
3771 case Op_AddF:
3772 case Op_SubF:
3773 case Op_MulF:
3774 case Op_DivF:
3775 case Op_NegF:
3776 case Op_ModF:
3777 case Op_ConvI2F:
3778 case Op_ConF:
3779 case Op_CmpF:
3780 case Op_CmpF3:
3781 case Op_StoreF:
3782 case Op_LoadF:
3783 // case Op_ConvL2F: // longs are split into 32-bit halves
3784 frc.inc_float_count();
3785 break;
3786
3787 case Op_ConvF2D:
3788 case Op_ConvD2F:
3789 frc.inc_float_count();
3790 frc.inc_double_count();
3791 break;
3792
3793 // Count all double operations that may use FPU
3794 case Op_AddD:
3795 case Op_SubD:
3796 case Op_MulD:
3797 case Op_DivD:
3798 case Op_NegD:
3799 case Op_ModD:
3800 case Op_ConvI2D:
3801 case Op_ConvD2I:
3802 // case Op_ConvL2D: // handled by leaf call
3803 // case Op_ConvD2L: // handled by leaf call
3804 case Op_ConD:
3805 case Op_CmpD:
3806 case Op_CmpD3:
3807 case Op_StoreD:
3808 case Op_LoadD:
3809 case Op_LoadD_unaligned:
3810 frc.inc_double_count();
3811 break;
3812 case Op_Opaque1: // Remove Opaque Nodes before matching
3813 n->subsume_by(n->in(1), this);
3814 break;
3815 case Op_CallStaticJava:
3816 case Op_CallJava:
3817 case Op_CallDynamicJava:
3818 frc.inc_java_call_count(); // Count java call site;
3819 case Op_CallRuntime:
3820 case Op_CallLeaf:
3821 case Op_CallLeafVector:
3822 case Op_CallLeafNoFP: {
3823 assert (n->is_Call(), "");
3824 CallNode *call = n->as_Call();
3825 // Count call sites where the FP mode bit would have to be flipped.
3826 // Do not count uncommon runtime calls:
3827 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3828 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3829 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3830 frc.inc_call_count(); // Count the call site
3831 } else { // See if uncommon argument is shared
3832 Node *n = call->in(TypeFunc::Parms);
3833 int nop = n->Opcode();
3834 // Clone shared simple arguments to uncommon calls, item (1).
3835 if (n->outcnt() > 1 &&
3836 !n->is_Proj() &&
3837 nop != Op_CreateEx &&
3838 nop != Op_CheckCastPP &&
3839 nop != Op_DecodeN &&
3840 nop != Op_DecodeNKlass &&
3841 !n->is_Mem() &&
3842 !n->is_Phi()) {
3843 Node *x = n->clone();
3844 call->set_req(TypeFunc::Parms, x);
3845 }
3846 }
3847 break;
3848 }
3849 case Op_StoreB:
3850 case Op_StoreC:
3851 case Op_StoreI:
3852 case Op_StoreL:
3853 case Op_StoreLSpecial:
3854 case Op_CompareAndSwapB:
3855 case Op_CompareAndSwapS:
3856 case Op_CompareAndSwapI:
3857 case Op_CompareAndSwapL:
3858 case Op_CompareAndSwapP:
3859 case Op_CompareAndSwapN:
3860 case Op_WeakCompareAndSwapB:
3861 case Op_WeakCompareAndSwapS:
3862 case Op_WeakCompareAndSwapI:
3863 case Op_WeakCompareAndSwapL:
3864 case Op_WeakCompareAndSwapP:
3865 case Op_WeakCompareAndSwapN:
3866 case Op_CompareAndExchangeB:
3867 case Op_CompareAndExchangeS:
3868 case Op_CompareAndExchangeI:
3869 case Op_CompareAndExchangeL:
3870 case Op_CompareAndExchangeP:
3871 case Op_CompareAndExchangeN:
3872 case Op_GetAndAddS:
3873 case Op_GetAndAddB:
3874 case Op_GetAndAddI:
3875 case Op_GetAndAddL:
3876 case Op_GetAndSetS:
3877 case Op_GetAndSetB:
3878 case Op_GetAndSetI:
3879 case Op_GetAndSetL:
3880 case Op_GetAndSetP:
3881 case Op_GetAndSetN:
3882 case Op_StoreP:
3883 case Op_StoreN:
3884 case Op_StoreNKlass:
3885 case Op_LoadB:
3886 case Op_LoadUB:
3887 case Op_LoadUS:
3888 case Op_LoadI:
3889 case Op_LoadKlass:
3890 case Op_LoadNKlass:
3891 case Op_LoadL:
3892 case Op_LoadL_unaligned:
3893 case Op_LoadP:
3894 case Op_LoadN:
3895 case Op_LoadRange:
3896 case Op_LoadS:
3897 break;
3898
3899 case Op_AddP: { // Assert sane base pointers
3900 Node *addp = n->in(AddPNode::Address);
3901 assert( !addp->is_AddP() ||
3902 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3903 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3904 "Base pointers must match (addp %u)", addp->_idx );
3905 #ifdef _LP64
3906 if ((UseCompressedOops || UseCompressedClassPointers) &&
3907 addp->Opcode() == Op_ConP &&
3908 addp == n->in(AddPNode::Base) &&
3909 n->in(AddPNode::Offset)->is_Con()) {
3910 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3911 // on the platform and on the compressed oops mode.
3912 // Use addressing with narrow klass to load with offset on x86.
3913 // Some platforms can use the constant pool to load ConP.
3914 // Do this transformation here since IGVN will convert ConN back to ConP.
3915 const Type* t = addp->bottom_type();
3916 bool is_oop = t->isa_oopptr() != nullptr;
3917 bool is_klass = t->isa_klassptr() != nullptr;
3918
3919 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3920 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3921 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3922 Node* nn = nullptr;
3923
3924 int op = is_oop ? Op_ConN : Op_ConNKlass;
3925
3926 // Look for existing ConN node of the same exact type.
3927 Node* r = root();
3928 uint cnt = r->outcnt();
3929 for (uint i = 0; i < cnt; i++) {
3930 Node* m = r->raw_out(i);
3931 if (m!= nullptr && m->Opcode() == op &&
3932 m->bottom_type()->make_ptr() == t) {
3933 nn = m;
3934 break;
3935 }
3936 }
3937 if (nn != nullptr) {
3938 // Decode a narrow oop to match address
3939 // [R12 + narrow_oop_reg<<3 + offset]
3940 if (is_oop) {
3941 nn = new DecodeNNode(nn, t);
3942 } else {
3943 nn = new DecodeNKlassNode(nn, t);
3944 }
3945 // Check for succeeding AddP which uses the same Base.
3946 // Otherwise we will run into the assertion above when visiting that guy.
3947 for (uint i = 0; i < n->outcnt(); ++i) {
3948 Node *out_i = n->raw_out(i);
3949 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3950 out_i->set_req(AddPNode::Base, nn);
3951 #ifdef ASSERT
3952 for (uint j = 0; j < out_i->outcnt(); ++j) {
3953 Node *out_j = out_i->raw_out(j);
3954 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3955 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3956 }
3957 #endif
3958 }
3959 }
3960 n->set_req(AddPNode::Base, nn);
3961 n->set_req(AddPNode::Address, nn);
3962 if (addp->outcnt() == 0) {
3963 addp->disconnect_inputs(this);
3964 }
3965 }
3966 }
3967 }
3968 #endif
3969 break;
3970 }
3971
3972 case Op_CastPP: {
3973 // Remove CastPP nodes to gain more freedom during scheduling but
3974 // keep the dependency they encode as control or precedence edges
3975 // (if control is set already) on memory operations. Some CastPP
3976 // nodes don't have a control (don't carry a dependency): skip
3977 // those.
3978 if (n->in(0) != nullptr) {
3979 ResourceMark rm;
3980 Unique_Node_List wq;
3981 wq.push(n);
3982 for (uint next = 0; next < wq.size(); ++next) {
3983 Node *m = wq.at(next);
3984 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3985 Node* use = m->fast_out(i);
3986 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3987 use->ensure_control_or_add_prec(n->in(0));
3988 } else {
3989 switch(use->Opcode()) {
3990 case Op_AddP:
3991 case Op_DecodeN:
3992 case Op_DecodeNKlass:
3993 case Op_CheckCastPP:
3994 case Op_CastPP:
3995 wq.push(use);
3996 break;
3997 }
3998 }
3999 }
4000 }
4001 }
4002 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
4003 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
4004 Node* in1 = n->in(1);
4005 const Type* t = n->bottom_type();
4006 Node* new_in1 = in1->clone();
4007 new_in1->as_DecodeN()->set_type(t);
4008
4009 if (!Matcher::narrow_oop_use_complex_address()) {
4010 //
4011 // x86, ARM and friends can handle 2 adds in addressing mode
4012 // and Matcher can fold a DecodeN node into address by using
4013 // a narrow oop directly and do implicit null check in address:
4014 //
4015 // [R12 + narrow_oop_reg<<3 + offset]
4016 // NullCheck narrow_oop_reg
4017 //
4018 // On other platforms (Sparc) we have to keep new DecodeN node and
4019 // use it to do implicit null check in address:
4020 //
4021 // decode_not_null narrow_oop_reg, base_reg
4022 // [base_reg + offset]
4023 // NullCheck base_reg
4024 //
4025 // Pin the new DecodeN node to non-null path on these platform (Sparc)
4026 // to keep the information to which null check the new DecodeN node
4027 // corresponds to use it as value in implicit_null_check().
4028 //
4029 new_in1->set_req(0, n->in(0));
4030 }
4031
4032 n->subsume_by(new_in1, this);
4033 if (in1->outcnt() == 0) {
4034 in1->disconnect_inputs(this);
4035 }
4036 } else {
4037 n->subsume_by(n->in(1), this);
4038 if (n->outcnt() == 0) {
4039 n->disconnect_inputs(this);
4040 }
4041 }
4042 break;
4043 }
4044 case Op_CastII: {
4045 n->as_CastII()->remove_range_check_cast(this);
4046 break;
4047 }
4048 #ifdef _LP64
4049 case Op_CmpP:
4050 // Do this transformation here to preserve CmpPNode::sub() and
4051 // other TypePtr related Ideal optimizations (for example, ptr nullness).
4052 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
4053 Node* in1 = n->in(1);
4054 Node* in2 = n->in(2);
4055 if (!in1->is_DecodeNarrowPtr()) {
4056 in2 = in1;
4057 in1 = n->in(2);
4058 }
4059 assert(in1->is_DecodeNarrowPtr(), "sanity");
4060
4061 Node* new_in2 = nullptr;
4062 if (in2->is_DecodeNarrowPtr()) {
4063 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
4064 new_in2 = in2->in(1);
4065 } else if (in2->Opcode() == Op_ConP) {
4066 const Type* t = in2->bottom_type();
4067 if (t == TypePtr::NULL_PTR) {
4068 assert(in1->is_DecodeN(), "compare klass to null?");
4069 // Don't convert CmpP null check into CmpN if compressed
4070 // oops implicit null check is not generated.
4071 // This will allow to generate normal oop implicit null check.
4072 if (Matcher::gen_narrow_oop_implicit_null_checks())
4073 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
4074 //
4075 // This transformation together with CastPP transformation above
4076 // will generated code for implicit null checks for compressed oops.
4077 //
4078 // The original code after Optimize()
4079 //
4080 // LoadN memory, narrow_oop_reg
4081 // decode narrow_oop_reg, base_reg
4082 // CmpP base_reg, nullptr
4083 // CastPP base_reg // NotNull
4084 // Load [base_reg + offset], val_reg
4085 //
4086 // after these transformations will be
4087 //
4088 // LoadN memory, narrow_oop_reg
4089 // CmpN narrow_oop_reg, nullptr
4090 // decode_not_null narrow_oop_reg, base_reg
4091 // Load [base_reg + offset], val_reg
4092 //
4093 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
4094 // since narrow oops can be used in debug info now (see the code in
4095 // final_graph_reshaping_walk()).
4096 //
4097 // At the end the code will be matched to
4098 // on x86:
4099 //
4100 // Load_narrow_oop memory, narrow_oop_reg
4101 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
4102 // NullCheck narrow_oop_reg
4103 //
4104 // and on sparc:
4105 //
4106 // Load_narrow_oop memory, narrow_oop_reg
4107 // decode_not_null narrow_oop_reg, base_reg
4108 // Load [base_reg + offset], val_reg
4109 // NullCheck base_reg
4110 //
4111 } else if (t->isa_oopptr()) {
4112 new_in2 = ConNode::make(t->make_narrowoop());
4113 } else if (t->isa_klassptr()) {
4114 new_in2 = ConNode::make(t->make_narrowklass());
4115 }
4116 }
4117 if (new_in2 != nullptr) {
4118 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
4119 n->subsume_by(cmpN, this);
4120 if (in1->outcnt() == 0) {
4121 in1->disconnect_inputs(this);
4122 }
4123 if (in2->outcnt() == 0) {
4124 in2->disconnect_inputs(this);
4125 }
4126 }
4127 }
4128 break;
4129
4130 case Op_DecodeN:
4131 case Op_DecodeNKlass:
4132 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4133 // DecodeN could be pinned when it can't be fold into
4134 // an address expression, see the code for Op_CastPP above.
4135 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4136 break;
4137
4138 case Op_EncodeP:
4139 case Op_EncodePKlass: {
4140 Node* in1 = n->in(1);
4141 if (in1->is_DecodeNarrowPtr()) {
4142 n->subsume_by(in1->in(1), this);
4143 } else if (in1->Opcode() == Op_ConP) {
4144 const Type* t = in1->bottom_type();
4145 if (t == TypePtr::NULL_PTR) {
4146 assert(t->isa_oopptr(), "null klass?");
4147 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4148 } else if (t->isa_oopptr()) {
4149 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4150 } else if (t->isa_klassptr()) {
4151 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4152 }
4153 }
4154 if (in1->outcnt() == 0) {
4155 in1->disconnect_inputs(this);
4156 }
4157 break;
4158 }
4159
4160 case Op_Proj: {
4161 if (OptimizeStringConcat || IncrementalInline) {
4162 ProjNode* proj = n->as_Proj();
4163 if (proj->_is_io_use) {
4164 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4165 // Separate projections were used for the exception path which
4166 // are normally removed by a late inline. If it wasn't inlined
4167 // then they will hang around and should just be replaced with
4168 // the original one. Merge them.
4169 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4170 if (non_io_proj != nullptr) {
4171 proj->subsume_by(non_io_proj , this);
4172 }
4173 }
4174 }
4175 break;
4176 }
4177
4178 case Op_Phi:
4179 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4180 // The EncodeP optimization may create Phi with the same edges
4181 // for all paths. It is not handled well by Register Allocator.
4182 Node* unique_in = n->in(1);
4183 assert(unique_in != nullptr, "");
4184 uint cnt = n->req();
4185 for (uint i = 2; i < cnt; i++) {
4186 Node* m = n->in(i);
4187 assert(m != nullptr, "");
4188 if (unique_in != m)
4189 unique_in = nullptr;
4190 }
4191 if (unique_in != nullptr) {
4192 n->subsume_by(unique_in, this);
4193 }
4194 }
4195 break;
4196
4197 #endif
4198
4199 case Op_ModI:
4200 handle_div_mod_op(n, T_INT, false);
4201 break;
4202
4203 case Op_ModL:
4204 handle_div_mod_op(n, T_LONG, false);
4205 break;
4206
4207 case Op_UModI:
4208 handle_div_mod_op(n, T_INT, true);
4209 break;
4210
4211 case Op_UModL:
4212 handle_div_mod_op(n, T_LONG, true);
4213 break;
4214
4215 case Op_LoadVector:
4216 case Op_StoreVector:
4217 #ifdef ASSERT
4218 // Add VerifyVectorAlignment node between adr and load / store.
4219 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
4220 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
4221 n->as_StoreVector()->must_verify_alignment();
4222 if (must_verify_alignment) {
4223 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
4224 n->as_StoreVector()->memory_size();
4225 // The memory access should be aligned to the vector width in bytes.
4226 // However, the underlying array is possibly less well aligned, but at least
4227 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
4228 // a loop we can expect at least the following alignment:
4229 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
4230 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
4231 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
4232 // Create mask from alignment. e.g. 0b1000 -> 0b0111
4233 jlong mask = guaranteed_alignment - 1;
4234 Node* mask_con = ConLNode::make(mask);
4235 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
4236 n->set_req(MemNode::Address, va);
4237 }
4238 }
4239 #endif
4240 break;
4241
4242 case Op_LoadVectorGather:
4243 case Op_StoreVectorScatter:
4244 case Op_LoadVectorGatherMasked:
4245 case Op_StoreVectorScatterMasked:
4246 case Op_VectorCmpMasked:
4247 case Op_VectorMaskGen:
4248 case Op_LoadVectorMasked:
4249 case Op_StoreVectorMasked:
4250 break;
4251
4252 case Op_AddReductionVI:
4253 case Op_AddReductionVL:
4254 case Op_AddReductionVF:
4255 case Op_AddReductionVD:
4256 case Op_MulReductionVI:
4257 case Op_MulReductionVL:
4258 case Op_MulReductionVF:
4259 case Op_MulReductionVD:
4260 case Op_MinReductionV:
4261 case Op_MaxReductionV:
4262 case Op_AndReductionV:
4263 case Op_OrReductionV:
4264 case Op_XorReductionV:
4265 break;
4266
4267 case Op_PackB:
4268 case Op_PackS:
4269 case Op_PackI:
4270 case Op_PackF:
4271 case Op_PackL:
4272 case Op_PackD:
4273 if (n->req()-1 > 2) {
4274 // Replace many operand PackNodes with a binary tree for matching
4275 PackNode* p = (PackNode*) n;
4276 Node* btp = p->binary_tree_pack(1, n->req());
4277 n->subsume_by(btp, this);
4278 }
4279 break;
4280 case Op_Loop:
4281 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4282 case Op_CountedLoop:
4283 case Op_LongCountedLoop:
4284 case Op_OuterStripMinedLoop:
4285 if (n->as_Loop()->is_inner_loop()) {
4286 frc.inc_inner_loop_count();
4287 }
4288 n->as_Loop()->verify_strip_mined(0);
4289 break;
4290 case Op_LShiftI:
4291 case Op_RShiftI:
4292 case Op_URShiftI:
4293 case Op_LShiftL:
4294 case Op_RShiftL:
4295 case Op_URShiftL:
4296 if (Matcher::need_masked_shift_count) {
4297 // The cpu's shift instructions don't restrict the count to the
4298 // lower 5/6 bits. We need to do the masking ourselves.
4299 Node* in2 = n->in(2);
4300 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4301 const TypeInt* t = in2->find_int_type();
4302 if (t != nullptr && t->is_con()) {
4303 juint shift = t->get_con();
4304 if (shift > mask) { // Unsigned cmp
4305 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4306 }
4307 } else {
4308 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
4309 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4310 n->set_req(2, shift);
4311 }
4312 }
4313 if (in2->outcnt() == 0) { // Remove dead node
4314 in2->disconnect_inputs(this);
4315 }
4316 }
4317 break;
4318 case Op_MemBarStoreStore:
4319 case Op_MemBarRelease:
4320 // Break the link with AllocateNode: it is no longer useful and
4321 // confuses register allocation.
4322 if (n->req() > MemBarNode::Precedent) {
4323 n->set_req(MemBarNode::Precedent, top());
4324 }
4325 break;
4326 case Op_MemBarAcquire: {
4327 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4328 // At parse time, the trailing MemBarAcquire for a volatile load
4329 // is created with an edge to the load. After optimizations,
4330 // that input may be a chain of Phis. If those phis have no
4331 // other use, then the MemBarAcquire keeps them alive and
4332 // register allocation can be confused.
4333 dead_nodes.push(n->in(MemBarNode::Precedent));
4334 n->set_req(MemBarNode::Precedent, top());
4335 }
4336 break;
4337 }
4338 case Op_Blackhole:
4339 break;
4340 case Op_RangeCheck: {
4341 RangeCheckNode* rc = n->as_RangeCheck();
4342 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4343 n->subsume_by(iff, this);
4344 frc._tests.push(iff);
4345 break;
4346 }
4347 case Op_ConvI2L: {
4348 if (!Matcher::convi2l_type_required) {
4349 // Code generation on some platforms doesn't need accurate
4350 // ConvI2L types. Widening the type can help remove redundant
4351 // address computations.
4352 n->as_Type()->set_type(TypeLong::INT);
4353 ResourceMark rm;
4354 Unique_Node_List wq;
4355 wq.push(n);
4356 for (uint next = 0; next < wq.size(); next++) {
4357 Node *m = wq.at(next);
4358
4359 for(;;) {
4360 // Loop over all nodes with identical inputs edges as m
4361 Node* k = m->find_similar(m->Opcode());
4362 if (k == nullptr) {
4363 break;
4364 }
4365 // Push their uses so we get a chance to remove node made
4366 // redundant
4367 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4368 Node* u = k->fast_out(i);
4369 if (u->Opcode() == Op_LShiftL ||
4370 u->Opcode() == Op_AddL ||
4371 u->Opcode() == Op_SubL ||
4372 u->Opcode() == Op_AddP) {
4373 wq.push(u);
4374 }
4375 }
4376 // Replace all nodes with identical edges as m with m
4377 k->subsume_by(m, this);
4378 }
4379 }
4380 }
4381 break;
4382 }
4383 case Op_CmpUL: {
4384 if (!Matcher::has_match_rule(Op_CmpUL)) {
4385 // No support for unsigned long comparisons
4386 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4387 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4388 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4389 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4390 Node* andl = new AndLNode(orl, remove_sign_mask);
4391 Node* cmp = new CmpLNode(andl, n->in(2));
4392 n->subsume_by(cmp, this);
4393 }
4394 break;
4395 }
4396 #ifdef ASSERT
4397 case Op_InlineType: {
4398 n->dump(-1);
4399 assert(false, "inline type node was not removed");
4400 break;
4401 }
4402 case Op_ConNKlass: {
4403 const TypePtr* tp = n->as_Type()->type()->make_ptr();
4404 ciKlass* klass = tp->is_klassptr()->exact_klass();
4405 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
4406 break;
4407 }
4408 #endif
4409 default:
4410 assert(!n->is_Call(), "");
4411 assert(!n->is_Mem(), "");
4412 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4413 break;
4414 }
4415 }
4416
4417 //------------------------------final_graph_reshaping_walk---------------------
4418 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4419 // requires that the walk visits a node's inputs before visiting the node.
4420 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4421 Unique_Node_List sfpt;
4422
4423 frc._visited.set(root->_idx); // first, mark node as visited
4424 uint cnt = root->req();
4425 Node *n = root;
4426 uint i = 0;
4427 while (true) {
4428 if (i < cnt) {
4429 // Place all non-visited non-null inputs onto stack
4430 Node* m = n->in(i);
4431 ++i;
4432 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
4433 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
4434 // compute worst case interpreter size in case of a deoptimization
4435 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4436
4437 sfpt.push(m);
4438 }
4439 cnt = m->req();
4440 nstack.push(n, i); // put on stack parent and next input's index
4441 n = m;
4442 i = 0;
4443 }
4444 } else {
4445 // Now do post-visit work
4446 final_graph_reshaping_impl(n, frc, dead_nodes);
4447 if (nstack.is_empty())
4448 break; // finished
4449 n = nstack.node(); // Get node from stack
4450 cnt = n->req();
4451 i = nstack.index();
4452 nstack.pop(); // Shift to the next node on stack
4453 }
4454 }
4455
4456 // Skip next transformation if compressed oops are not used.
4457 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4458 (!UseCompressedOops && !UseCompressedClassPointers))
4459 return;
4460
4461 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4462 // It could be done for an uncommon traps or any safepoints/calls
4463 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4464 while (sfpt.size() > 0) {
4465 n = sfpt.pop();
4466 JVMState *jvms = n->as_SafePoint()->jvms();
4467 assert(jvms != nullptr, "sanity");
4468 int start = jvms->debug_start();
4469 int end = n->req();
4470 bool is_uncommon = (n->is_CallStaticJava() &&
4471 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4472 for (int j = start; j < end; j++) {
4473 Node* in = n->in(j);
4474 if (in->is_DecodeNarrowPtr()) {
4475 bool safe_to_skip = true;
4476 if (!is_uncommon ) {
4477 // Is it safe to skip?
4478 for (uint i = 0; i < in->outcnt(); i++) {
4479 Node* u = in->raw_out(i);
4480 if (!u->is_SafePoint() ||
4481 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4482 safe_to_skip = false;
4483 }
4484 }
4485 }
4486 if (safe_to_skip) {
4487 n->set_req(j, in->in(1));
4488 }
4489 if (in->outcnt() == 0) {
4490 in->disconnect_inputs(this);
4491 }
4492 }
4493 }
4494 }
4495 }
4496
4497 //------------------------------final_graph_reshaping--------------------------
4498 // Final Graph Reshaping.
4499 //
4500 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4501 // and not commoned up and forced early. Must come after regular
4502 // optimizations to avoid GVN undoing the cloning. Clone constant
4503 // inputs to Loop Phis; these will be split by the allocator anyways.
4504 // Remove Opaque nodes.
4505 // (2) Move last-uses by commutative operations to the left input to encourage
4506 // Intel update-in-place two-address operations and better register usage
4507 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4508 // calls canonicalizing them back.
4509 // (3) Count the number of double-precision FP ops, single-precision FP ops
4510 // and call sites. On Intel, we can get correct rounding either by
4511 // forcing singles to memory (requires extra stores and loads after each
4512 // FP bytecode) or we can set a rounding mode bit (requires setting and
4513 // clearing the mode bit around call sites). The mode bit is only used
4514 // if the relative frequency of single FP ops to calls is low enough.
4515 // This is a key transform for SPEC mpeg_audio.
4516 // (4) Detect infinite loops; blobs of code reachable from above but not
4517 // below. Several of the Code_Gen algorithms fail on such code shapes,
4518 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4519 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4520 // Detection is by looking for IfNodes where only 1 projection is
4521 // reachable from below or CatchNodes missing some targets.
4522 // (5) Assert for insane oop offsets in debug mode.
4523
4524 bool Compile::final_graph_reshaping() {
4525 // an infinite loop may have been eliminated by the optimizer,
4526 // in which case the graph will be empty.
4527 if (root()->req() == 1) {
4528 // Do not compile method that is only a trivial infinite loop,
4529 // since the content of the loop may have been eliminated.
4530 record_method_not_compilable("trivial infinite loop");
4531 return true;
4532 }
4533
4534 // Expensive nodes have their control input set to prevent the GVN
4535 // from freely commoning them. There's no GVN beyond this point so
4536 // no need to keep the control input. We want the expensive nodes to
4537 // be freely moved to the least frequent code path by gcm.
4538 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4539 for (int i = 0; i < expensive_count(); i++) {
4540 _expensive_nodes.at(i)->set_req(0, nullptr);
4541 }
4542
4543 Final_Reshape_Counts frc;
4544
4545 // Visit everybody reachable!
4546 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4547 Node_Stack nstack(live_nodes() >> 1);
4548 Unique_Node_List dead_nodes;
4549 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4550
4551 // Check for unreachable (from below) code (i.e., infinite loops).
4552 for( uint i = 0; i < frc._tests.size(); i++ ) {
4553 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4554 // Get number of CFG targets.
4555 // Note that PCTables include exception targets after calls.
4556 uint required_outcnt = n->required_outcnt();
4557 if (n->outcnt() != required_outcnt) {
4558 // Check for a few special cases. Rethrow Nodes never take the
4559 // 'fall-thru' path, so expected kids is 1 less.
4560 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4561 if (n->in(0)->in(0)->is_Call()) {
4562 CallNode* call = n->in(0)->in(0)->as_Call();
4563 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4564 required_outcnt--; // Rethrow always has 1 less kid
4565 } else if (call->req() > TypeFunc::Parms &&
4566 call->is_CallDynamicJava()) {
4567 // Check for null receiver. In such case, the optimizer has
4568 // detected that the virtual call will always result in a null
4569 // pointer exception. The fall-through projection of this CatchNode
4570 // will not be populated.
4571 Node* arg0 = call->in(TypeFunc::Parms);
4572 if (arg0->is_Type() &&
4573 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4574 required_outcnt--;
4575 }
4576 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4577 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4578 // Check for illegal array length. In such case, the optimizer has
4579 // detected that the allocation attempt will always result in an
4580 // exception. There is no fall-through projection of this CatchNode .
4581 assert(call->is_CallStaticJava(), "static call expected");
4582 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4583 uint valid_length_test_input = call->req() - 1;
4584 Node* valid_length_test = call->in(valid_length_test_input);
4585 call->del_req(valid_length_test_input);
4586 if (valid_length_test->find_int_con(1) == 0) {
4587 required_outcnt--;
4588 }
4589 dead_nodes.push(valid_length_test);
4590 assert(n->outcnt() == required_outcnt, "malformed control flow");
4591 continue;
4592 }
4593 }
4594 }
4595
4596 // Recheck with a better notion of 'required_outcnt'
4597 if (n->outcnt() != required_outcnt) {
4598 record_method_not_compilable("malformed control flow");
4599 return true; // Not all targets reachable!
4600 }
4601 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4602 CallNode* call = n->in(0)->in(0)->as_Call();
4603 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4604 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4605 assert(call->is_CallStaticJava(), "static call expected");
4606 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4607 uint valid_length_test_input = call->req() - 1;
4608 dead_nodes.push(call->in(valid_length_test_input));
4609 call->del_req(valid_length_test_input); // valid length test useless now
4610 }
4611 }
4612 // Check that I actually visited all kids. Unreached kids
4613 // must be infinite loops.
4614 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4615 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4616 record_method_not_compilable("infinite loop");
4617 return true; // Found unvisited kid; must be unreach
4618 }
4619
4620 // Here so verification code in final_graph_reshaping_walk()
4621 // always see an OuterStripMinedLoopEnd
4622 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4623 IfNode* init_iff = n->as_If();
4624 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4625 n->subsume_by(iff, this);
4626 }
4627 }
4628
4629 while (dead_nodes.size() > 0) {
4630 Node* m = dead_nodes.pop();
4631 if (m->outcnt() == 0 && m != top()) {
4632 for (uint j = 0; j < m->req(); j++) {
4633 Node* in = m->in(j);
4634 if (in != nullptr) {
4635 dead_nodes.push(in);
4636 }
4637 }
4638 m->disconnect_inputs(this);
4639 }
4640 }
4641
4642 set_java_calls(frc.get_java_call_count());
4643 set_inner_loops(frc.get_inner_loop_count());
4644
4645 // No infinite loops, no reason to bail out.
4646 return false;
4647 }
4648
4649 //-----------------------------too_many_traps----------------------------------
4650 // Report if there are too many traps at the current method and bci.
4651 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4652 bool Compile::too_many_traps(ciMethod* method,
4653 int bci,
4654 Deoptimization::DeoptReason reason) {
4655 ciMethodData* md = method->method_data();
4656 if (md->is_empty()) {
4657 // Assume the trap has not occurred, or that it occurred only
4658 // because of a transient condition during start-up in the interpreter.
4659 return false;
4660 }
4661 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4662 if (md->has_trap_at(bci, m, reason) != 0) {
4663 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4664 // Also, if there are multiple reasons, or if there is no per-BCI record,
4665 // assume the worst.
4666 if (log())
4667 log()->elem("observe trap='%s' count='%d'",
4668 Deoptimization::trap_reason_name(reason),
4669 md->trap_count(reason));
4670 return true;
4671 } else {
4672 // Ignore method/bci and see if there have been too many globally.
4673 return too_many_traps(reason, md);
4674 }
4675 }
4676
4677 // Less-accurate variant which does not require a method and bci.
4678 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4679 ciMethodData* logmd) {
4680 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4681 // Too many traps globally.
4682 // Note that we use cumulative trap_count, not just md->trap_count.
4683 if (log()) {
4684 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4685 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4686 Deoptimization::trap_reason_name(reason),
4687 mcount, trap_count(reason));
4688 }
4689 return true;
4690 } else {
4691 // The coast is clear.
4692 return false;
4693 }
4694 }
4695
4696 //--------------------------too_many_recompiles--------------------------------
4697 // Report if there are too many recompiles at the current method and bci.
4698 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4699 // Is not eager to return true, since this will cause the compiler to use
4700 // Action_none for a trap point, to avoid too many recompilations.
4701 bool Compile::too_many_recompiles(ciMethod* method,
4702 int bci,
4703 Deoptimization::DeoptReason reason) {
4704 ciMethodData* md = method->method_data();
4705 if (md->is_empty()) {
4706 // Assume the trap has not occurred, or that it occurred only
4707 // because of a transient condition during start-up in the interpreter.
4708 return false;
4709 }
4710 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4711 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4712 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4713 Deoptimization::DeoptReason per_bc_reason
4714 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4715 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4716 if ((per_bc_reason == Deoptimization::Reason_none
4717 || md->has_trap_at(bci, m, reason) != 0)
4718 // The trap frequency measure we care about is the recompile count:
4719 && md->trap_recompiled_at(bci, m)
4720 && md->overflow_recompile_count() >= bc_cutoff) {
4721 // Do not emit a trap here if it has already caused recompilations.
4722 // Also, if there are multiple reasons, or if there is no per-BCI record,
4723 // assume the worst.
4724 if (log())
4725 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4726 Deoptimization::trap_reason_name(reason),
4727 md->trap_count(reason),
4728 md->overflow_recompile_count());
4729 return true;
4730 } else if (trap_count(reason) != 0
4731 && decompile_count() >= m_cutoff) {
4732 // Too many recompiles globally, and we have seen this sort of trap.
4733 // Use cumulative decompile_count, not just md->decompile_count.
4734 if (log())
4735 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4736 Deoptimization::trap_reason_name(reason),
4737 md->trap_count(reason), trap_count(reason),
4738 md->decompile_count(), decompile_count());
4739 return true;
4740 } else {
4741 // The coast is clear.
4742 return false;
4743 }
4744 }
4745
4746 // Compute when not to trap. Used by matching trap based nodes and
4747 // NullCheck optimization.
4748 void Compile::set_allowed_deopt_reasons() {
4749 _allowed_reasons = 0;
4750 if (is_method_compilation()) {
4751 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4752 assert(rs < BitsPerInt, "recode bit map");
4753 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4754 _allowed_reasons |= nth_bit(rs);
4755 }
4756 }
4757 }
4758 }
4759
4760 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4761 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4762 }
4763
4764 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4765 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4766 }
4767
4768 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4769 if (holder->is_initialized()) {
4770 return false;
4771 }
4772 if (holder->is_being_initialized()) {
4773 if (accessing_method->holder() == holder) {
4774 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4775 // <init>, or a static method. In all those cases, there was an initialization
4776 // barrier on the holder klass passed.
4777 if (accessing_method->is_class_initializer() ||
4778 accessing_method->is_object_constructor() ||
4779 accessing_method->is_static()) {
4780 return false;
4781 }
4782 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4783 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4784 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4785 // child class can become fully initialized while its parent class is still being initialized.
4786 if (accessing_method->is_class_initializer()) {
4787 return false;
4788 }
4789 }
4790 ciMethod* root = method(); // the root method of compilation
4791 if (root != accessing_method) {
4792 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4793 }
4794 }
4795 return true;
4796 }
4797
4798 #ifndef PRODUCT
4799 //------------------------------verify_bidirectional_edges---------------------
4800 // For each input edge to a node (ie - for each Use-Def edge), verify that
4801 // there is a corresponding Def-Use edge.
4802 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4803 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4804 uint stack_size = live_nodes() >> 4;
4805 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4806 if (root_and_safepoints != nullptr) {
4807 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4808 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4809 Node* root_or_safepoint = root_and_safepoints->at(i);
4810 // If the node is a safepoint, let's check if it still has a control input
4811 // Lack of control input signifies that this node was killed by CCP or
4812 // recursively by remove_globally_dead_node and it shouldn't be a starting
4813 // point.
4814 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4815 nstack.push(root_or_safepoint);
4816 }
4817 }
4818 } else {
4819 nstack.push(_root);
4820 }
4821
4822 while (nstack.size() > 0) {
4823 Node* n = nstack.pop();
4824 if (visited.member(n)) {
4825 continue;
4826 }
4827 visited.push(n);
4828
4829 // Walk over all input edges, checking for correspondence
4830 uint length = n->len();
4831 for (uint i = 0; i < length; i++) {
4832 Node* in = n->in(i);
4833 if (in != nullptr && !visited.member(in)) {
4834 nstack.push(in); // Put it on stack
4835 }
4836 if (in != nullptr && !in->is_top()) {
4837 // Count instances of `next`
4838 int cnt = 0;
4839 for (uint idx = 0; idx < in->_outcnt; idx++) {
4840 if (in->_out[idx] == n) {
4841 cnt++;
4842 }
4843 }
4844 assert(cnt > 0, "Failed to find Def-Use edge.");
4845 // Check for duplicate edges
4846 // walk the input array downcounting the input edges to n
4847 for (uint j = 0; j < length; j++) {
4848 if (n->in(j) == in) {
4849 cnt--;
4850 }
4851 }
4852 assert(cnt == 0, "Mismatched edge count.");
4853 } else if (in == nullptr) {
4854 assert(i == 0 || i >= n->req() ||
4855 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4856 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
4857 (n->is_Unlock() && i == (n->req() - 1)) ||
4858 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4859 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
4860 } else {
4861 assert(in->is_top(), "sanity");
4862 // Nothing to check.
4863 }
4864 }
4865 }
4866 }
4867
4868 //------------------------------verify_graph_edges---------------------------
4869 // Walk the Graph and verify that there is a one-to-one correspondence
4870 // between Use-Def edges and Def-Use edges in the graph.
4871 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4872 if (VerifyGraphEdges) {
4873 Unique_Node_List visited;
4874
4875 // Call graph walk to check edges
4876 verify_bidirectional_edges(visited, root_and_safepoints);
4877 if (no_dead_code) {
4878 // Now make sure that no visited node is used by an unvisited node.
4879 bool dead_nodes = false;
4880 Unique_Node_List checked;
4881 while (visited.size() > 0) {
4882 Node* n = visited.pop();
4883 checked.push(n);
4884 for (uint i = 0; i < n->outcnt(); i++) {
4885 Node* use = n->raw_out(i);
4886 if (checked.member(use)) continue; // already checked
4887 if (visited.member(use)) continue; // already in the graph
4888 if (use->is_Con()) continue; // a dead ConNode is OK
4889 // At this point, we have found a dead node which is DU-reachable.
4890 if (!dead_nodes) {
4891 tty->print_cr("*** Dead nodes reachable via DU edges:");
4892 dead_nodes = true;
4893 }
4894 use->dump(2);
4895 tty->print_cr("---");
4896 checked.push(use); // No repeats; pretend it is now checked.
4897 }
4898 }
4899 assert(!dead_nodes, "using nodes must be reachable from root");
4900 }
4901 }
4902 }
4903 #endif
4904
4905 // The Compile object keeps track of failure reasons separately from the ciEnv.
4906 // This is required because there is not quite a 1-1 relation between the
4907 // ciEnv and its compilation task and the Compile object. Note that one
4908 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4909 // to backtrack and retry without subsuming loads. Other than this backtracking
4910 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4911 // by the logic in C2Compiler.
4912 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4913 if (log() != nullptr) {
4914 log()->elem("failure reason='%s' phase='compile'", reason);
4915 }
4916 if (_failure_reason.get() == nullptr) {
4917 // Record the first failure reason.
4918 _failure_reason.set(reason);
4919 if (CaptureBailoutInformation) {
4920 _first_failure_details = new CompilationFailureInfo(reason);
4921 }
4922 } else {
4923 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4924 }
4925
4926 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4927 C->print_method(PHASE_FAILURE, 1);
4928 }
4929 _root = nullptr; // flush the graph, too
4930 }
4931
4932 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4933 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4934 _compile(Compile::current()),
4935 _log(nullptr),
4936 _dolog(CITimeVerbose)
4937 {
4938 assert(_compile != nullptr, "sanity check");
4939 assert(id != PhaseTraceId::_t_none, "Don't use none");
4940 if (_dolog) {
4941 _log = _compile->log();
4942 }
4943 if (_log != nullptr) {
4944 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4945 _log->stamp();
4946 _log->end_head();
4947 }
4948
4949 // Inform memory statistic, if enabled
4950 if (CompilationMemoryStatistic::enabled()) {
4951 CompilationMemoryStatistic::on_phase_start((int)id, name);
4952 }
4953 }
4954
4955 Compile::TracePhase::TracePhase(PhaseTraceId id)
4956 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4957
4958 Compile::TracePhase::~TracePhase() {
4959
4960 // Inform memory statistic, if enabled
4961 if (CompilationMemoryStatistic::enabled()) {
4962 CompilationMemoryStatistic::on_phase_end();
4963 }
4964
4965 if (_compile->failing_internal()) {
4966 if (_log != nullptr) {
4967 _log->done("phase");
4968 }
4969 return; // timing code, not stressing bailouts.
4970 }
4971 #ifdef ASSERT
4972 if (PrintIdealNodeCount) {
4973 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4974 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4975 }
4976
4977 if (VerifyIdealNodeCount) {
4978 _compile->print_missing_nodes();
4979 }
4980 #endif
4981
4982 if (_log != nullptr) {
4983 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4984 }
4985 }
4986
4987 //----------------------------static_subtype_check-----------------------------
4988 // Shortcut important common cases when superklass is exact:
4989 // (0) superklass is java.lang.Object (can occur in reflective code)
4990 // (1) subklass is already limited to a subtype of superklass => always ok
4991 // (2) subklass does not overlap with superklass => always fail
4992 // (3) superklass has NO subtypes and we can check with a simple compare.
4993 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4994 if (skip) {
4995 return SSC_full_test; // Let caller generate the general case.
4996 }
4997
4998 if (subk->is_java_subtype_of(superk)) {
4999 return SSC_always_true; // (0) and (1) this test cannot fail
5000 }
5001
5002 if (!subk->maybe_java_subtype_of(superk)) {
5003 return SSC_always_false; // (2) true path dead; no dynamic test needed
5004 }
5005
5006 const Type* superelem = superk;
5007 if (superk->isa_aryklassptr()) {
5008 int ignored;
5009 superelem = superk->is_aryklassptr()->base_element_type(ignored);
5010
5011 // Do not fold the subtype check to an array klass pointer comparison for null-able inline type arrays
5012 // because null-free [LMyValue <: null-able [LMyValue but the klasses are different. Perform a full test.
5013 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
5014 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
5015 return SSC_full_test;
5016 }
5017 }
5018
5019 if (superelem->isa_instklassptr()) {
5020 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
5021 if (!ik->has_subklass()) {
5022 if (!ik->is_final()) {
5023 // Add a dependency if there is a chance of a later subclass.
5024 dependencies()->assert_leaf_type(ik);
5025 }
5026 if (!superk->maybe_java_subtype_of(subk)) {
5027 return SSC_always_false;
5028 }
5029 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5030 }
5031 } else {
5032 // A primitive array type has no subtypes.
5033 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5034 }
5035
5036 return SSC_full_test;
5037 }
5038
5039 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
5040 #ifdef _LP64
5041 // The scaled index operand to AddP must be a clean 64-bit value.
5042 // Java allows a 32-bit int to be incremented to a negative
5043 // value, which appears in a 64-bit register as a large
5044 // positive number. Using that large positive number as an
5045 // operand in pointer arithmetic has bad consequences.
5046 // On the other hand, 32-bit overflow is rare, and the possibility
5047 // can often be excluded, if we annotate the ConvI2L node with
5048 // a type assertion that its value is known to be a small positive
5049 // number. (The prior range check has ensured this.)
5050 // This assertion is used by ConvI2LNode::Ideal.
5051 int index_max = max_jint - 1; // array size is max_jint, index is one less
5052 if (sizetype != nullptr && sizetype->_hi > 0) {
5053 index_max = sizetype->_hi - 1;
5054 }
5055 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
5056 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
5057 #endif
5058 return idx;
5059 }
5060
5061 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
5062 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
5063 if (ctrl != nullptr) {
5064 // Express control dependency by a CastII node with a narrow type.
5065 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
5066 // node from floating above the range check during loop optimizations. Otherwise, the
5067 // ConvI2L node may be eliminated independently of the range check, causing the data path
5068 // to become TOP while the control path is still there (although it's unreachable).
5069 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
5070 value = phase->transform(value);
5071 }
5072 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
5073 return phase->transform(new ConvI2LNode(value, ltype));
5074 }
5075
5076 void Compile::dump_print_inlining() {
5077 inline_printer()->print_on(tty);
5078 }
5079
5080 void Compile::log_late_inline(CallGenerator* cg) {
5081 if (log() != nullptr) {
5082 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5083 cg->unique_id());
5084 JVMState* p = cg->call_node()->jvms();
5085 while (p != nullptr) {
5086 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5087 p = p->caller();
5088 }
5089 log()->tail("late_inline");
5090 }
5091 }
5092
5093 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5094 log_late_inline(cg);
5095 if (log() != nullptr) {
5096 log()->inline_fail(msg);
5097 }
5098 }
5099
5100 void Compile::log_inline_id(CallGenerator* cg) {
5101 if (log() != nullptr) {
5102 // The LogCompilation tool needs a unique way to identify late
5103 // inline call sites. This id must be unique for this call site in
5104 // this compilation. Try to have it unique across compilations as
5105 // well because it can be convenient when grepping through the log
5106 // file.
5107 // Distinguish OSR compilations from others in case CICountOSR is
5108 // on.
5109 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5110 cg->set_unique_id(id);
5111 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5112 }
5113 }
5114
5115 void Compile::log_inline_failure(const char* msg) {
5116 if (C->log() != nullptr) {
5117 C->log()->inline_fail(msg);
5118 }
5119 }
5120
5121
5122 // Dump inlining replay data to the stream.
5123 // Don't change thread state and acquire any locks.
5124 void Compile::dump_inline_data(outputStream* out) {
5125 InlineTree* inl_tree = ilt();
5126 if (inl_tree != nullptr) {
5127 out->print(" inline %d", inl_tree->count());
5128 inl_tree->dump_replay_data(out);
5129 }
5130 }
5131
5132 void Compile::dump_inline_data_reduced(outputStream* out) {
5133 assert(ReplayReduce, "");
5134
5135 InlineTree* inl_tree = ilt();
5136 if (inl_tree == nullptr) {
5137 return;
5138 }
5139 // Enable iterative replay file reduction
5140 // Output "compile" lines for depth 1 subtrees,
5141 // simulating that those trees were compiled
5142 // instead of inlined.
5143 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5144 InlineTree* sub = inl_tree->subtrees().at(i);
5145 if (sub->inline_level() != 1) {
5146 continue;
5147 }
5148
5149 ciMethod* method = sub->method();
5150 int entry_bci = -1;
5151 int comp_level = env()->task()->comp_level();
5152 out->print("compile ");
5153 method->dump_name_as_ascii(out);
5154 out->print(" %d %d", entry_bci, comp_level);
5155 out->print(" inline %d", sub->count());
5156 sub->dump_replay_data(out, -1);
5157 out->cr();
5158 }
5159 }
5160
5161 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5162 if (n1->Opcode() < n2->Opcode()) return -1;
5163 else if (n1->Opcode() > n2->Opcode()) return 1;
5164
5165 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5166 for (uint i = 1; i < n1->req(); i++) {
5167 if (n1->in(i) < n2->in(i)) return -1;
5168 else if (n1->in(i) > n2->in(i)) return 1;
5169 }
5170
5171 return 0;
5172 }
5173
5174 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5175 Node* n1 = *n1p;
5176 Node* n2 = *n2p;
5177
5178 return cmp_expensive_nodes(n1, n2);
5179 }
5180
5181 void Compile::sort_expensive_nodes() {
5182 if (!expensive_nodes_sorted()) {
5183 _expensive_nodes.sort(cmp_expensive_nodes);
5184 }
5185 }
5186
5187 bool Compile::expensive_nodes_sorted() const {
5188 for (int i = 1; i < _expensive_nodes.length(); i++) {
5189 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5190 return false;
5191 }
5192 }
5193 return true;
5194 }
5195
5196 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5197 if (_expensive_nodes.length() == 0) {
5198 return false;
5199 }
5200
5201 assert(OptimizeExpensiveOps, "optimization off?");
5202
5203 // Take this opportunity to remove dead nodes from the list
5204 int j = 0;
5205 for (int i = 0; i < _expensive_nodes.length(); i++) {
5206 Node* n = _expensive_nodes.at(i);
5207 if (!n->is_unreachable(igvn)) {
5208 assert(n->is_expensive(), "should be expensive");
5209 _expensive_nodes.at_put(j, n);
5210 j++;
5211 }
5212 }
5213 _expensive_nodes.trunc_to(j);
5214
5215 // Then sort the list so that similar nodes are next to each other
5216 // and check for at least two nodes of identical kind with same data
5217 // inputs.
5218 sort_expensive_nodes();
5219
5220 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5221 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5222 return true;
5223 }
5224 }
5225
5226 return false;
5227 }
5228
5229 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5230 if (_expensive_nodes.length() == 0) {
5231 return;
5232 }
5233
5234 assert(OptimizeExpensiveOps, "optimization off?");
5235
5236 // Sort to bring similar nodes next to each other and clear the
5237 // control input of nodes for which there's only a single copy.
5238 sort_expensive_nodes();
5239
5240 int j = 0;
5241 int identical = 0;
5242 int i = 0;
5243 bool modified = false;
5244 for (; i < _expensive_nodes.length()-1; i++) {
5245 assert(j <= i, "can't write beyond current index");
5246 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5247 identical++;
5248 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5249 continue;
5250 }
5251 if (identical > 0) {
5252 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5253 identical = 0;
5254 } else {
5255 Node* n = _expensive_nodes.at(i);
5256 igvn.replace_input_of(n, 0, nullptr);
5257 igvn.hash_insert(n);
5258 modified = true;
5259 }
5260 }
5261 if (identical > 0) {
5262 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5263 } else if (_expensive_nodes.length() >= 1) {
5264 Node* n = _expensive_nodes.at(i);
5265 igvn.replace_input_of(n, 0, nullptr);
5266 igvn.hash_insert(n);
5267 modified = true;
5268 }
5269 _expensive_nodes.trunc_to(j);
5270 if (modified) {
5271 igvn.optimize();
5272 }
5273 }
5274
5275 void Compile::add_expensive_node(Node * n) {
5276 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5277 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5278 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5279 if (OptimizeExpensiveOps) {
5280 _expensive_nodes.append(n);
5281 } else {
5282 // Clear control input and let IGVN optimize expensive nodes if
5283 // OptimizeExpensiveOps is off.
5284 n->set_req(0, nullptr);
5285 }
5286 }
5287
5288 /**
5289 * Track coarsened Lock and Unlock nodes.
5290 */
5291
5292 class Lock_List : public Node_List {
5293 uint _origin_cnt;
5294 public:
5295 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5296 uint origin_cnt() const { return _origin_cnt; }
5297 };
5298
5299 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5300 int length = locks.length();
5301 if (length > 0) {
5302 // Have to keep this list until locks elimination during Macro nodes elimination.
5303 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5304 AbstractLockNode* alock = locks.at(0);
5305 BoxLockNode* box = alock->box_node()->as_BoxLock();
5306 for (int i = 0; i < length; i++) {
5307 AbstractLockNode* lock = locks.at(i);
5308 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5309 locks_list->push(lock);
5310 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
5311 if (this_box != box) {
5312 // Locking regions (BoxLock) could be Unbalanced here:
5313 // - its coarsened locks were eliminated in earlier
5314 // macro nodes elimination followed by loop unroll
5315 // - it is OSR locking region (no Lock node)
5316 // Preserve Unbalanced status in such cases.
5317 if (!this_box->is_unbalanced()) {
5318 this_box->set_coarsened();
5319 }
5320 if (!box->is_unbalanced()) {
5321 box->set_coarsened();
5322 }
5323 }
5324 }
5325 _coarsened_locks.append(locks_list);
5326 }
5327 }
5328
5329 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5330 int count = coarsened_count();
5331 for (int i = 0; i < count; i++) {
5332 Node_List* locks_list = _coarsened_locks.at(i);
5333 for (uint j = 0; j < locks_list->size(); j++) {
5334 Node* lock = locks_list->at(j);
5335 assert(lock->is_AbstractLock(), "sanity");
5336 if (!useful.member(lock)) {
5337 locks_list->yank(lock);
5338 }
5339 }
5340 }
5341 }
5342
5343 void Compile::remove_coarsened_lock(Node* n) {
5344 if (n->is_AbstractLock()) {
5345 int count = coarsened_count();
5346 for (int i = 0; i < count; i++) {
5347 Node_List* locks_list = _coarsened_locks.at(i);
5348 locks_list->yank(n);
5349 }
5350 }
5351 }
5352
5353 bool Compile::coarsened_locks_consistent() {
5354 int count = coarsened_count();
5355 for (int i = 0; i < count; i++) {
5356 bool unbalanced = false;
5357 bool modified = false; // track locks kind modifications
5358 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5359 uint size = locks_list->size();
5360 if (size == 0) {
5361 unbalanced = false; // All locks were eliminated - good
5362 } else if (size != locks_list->origin_cnt()) {
5363 unbalanced = true; // Some locks were removed from list
5364 } else {
5365 for (uint j = 0; j < size; j++) {
5366 Node* lock = locks_list->at(j);
5367 // All nodes in group should have the same state (modified or not)
5368 if (!lock->as_AbstractLock()->is_coarsened()) {
5369 if (j == 0) {
5370 // first on list was modified, the rest should be too for consistency
5371 modified = true;
5372 } else if (!modified) {
5373 // this lock was modified but previous locks on the list were not
5374 unbalanced = true;
5375 break;
5376 }
5377 } else if (modified) {
5378 // previous locks on list were modified but not this lock
5379 unbalanced = true;
5380 break;
5381 }
5382 }
5383 }
5384 if (unbalanced) {
5385 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5386 #ifdef ASSERT
5387 if (PrintEliminateLocks) {
5388 tty->print_cr("=== unbalanced coarsened locks ===");
5389 for (uint l = 0; l < size; l++) {
5390 locks_list->at(l)->dump();
5391 }
5392 }
5393 #endif
5394 record_failure(C2Compiler::retry_no_locks_coarsening());
5395 return false;
5396 }
5397 }
5398 return true;
5399 }
5400
5401 // Mark locking regions (identified by BoxLockNode) as unbalanced if
5402 // locks coarsening optimization removed Lock/Unlock nodes from them.
5403 // Such regions become unbalanced because coarsening only removes part
5404 // of Lock/Unlock nodes in region. As result we can't execute other
5405 // locks elimination optimizations which assume all code paths have
5406 // corresponding pair of Lock/Unlock nodes - they are balanced.
5407 void Compile::mark_unbalanced_boxes() const {
5408 int count = coarsened_count();
5409 for (int i = 0; i < count; i++) {
5410 Node_List* locks_list = _coarsened_locks.at(i);
5411 uint size = locks_list->size();
5412 if (size > 0) {
5413 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
5414 BoxLockNode* box = alock->box_node()->as_BoxLock();
5415 if (alock->is_coarsened()) {
5416 // coarsened_locks_consistent(), which is called before this method, verifies
5417 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
5418 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5419 for (uint j = 1; j < size; j++) {
5420 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
5421 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
5422 if (box != this_box) {
5423 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5424 box->set_unbalanced();
5425 this_box->set_unbalanced();
5426 }
5427 }
5428 }
5429 }
5430 }
5431 }
5432
5433 /**
5434 * Remove the speculative part of types and clean up the graph
5435 */
5436 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5437 if (UseTypeSpeculation) {
5438 Unique_Node_List worklist;
5439 worklist.push(root());
5440 int modified = 0;
5441 // Go over all type nodes that carry a speculative type, drop the
5442 // speculative part of the type and enqueue the node for an igvn
5443 // which may optimize it out.
5444 for (uint next = 0; next < worklist.size(); ++next) {
5445 Node *n = worklist.at(next);
5446 if (n->is_Type()) {
5447 TypeNode* tn = n->as_Type();
5448 const Type* t = tn->type();
5449 const Type* t_no_spec = t->remove_speculative();
5450 if (t_no_spec != t) {
5451 bool in_hash = igvn.hash_delete(n);
5452 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5453 tn->set_type(t_no_spec);
5454 igvn.hash_insert(n);
5455 igvn._worklist.push(n); // give it a chance to go away
5456 modified++;
5457 }
5458 }
5459 // Iterate over outs - endless loops is unreachable from below
5460 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5461 Node *m = n->fast_out(i);
5462 if (not_a_node(m)) {
5463 continue;
5464 }
5465 worklist.push(m);
5466 }
5467 }
5468 // Drop the speculative part of all types in the igvn's type table
5469 igvn.remove_speculative_types();
5470 if (modified > 0) {
5471 igvn.optimize();
5472 if (failing()) return;
5473 }
5474 #ifdef ASSERT
5475 // Verify that after the IGVN is over no speculative type has resurfaced
5476 worklist.clear();
5477 worklist.push(root());
5478 for (uint next = 0; next < worklist.size(); ++next) {
5479 Node *n = worklist.at(next);
5480 const Type* t = igvn.type_or_null(n);
5481 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5482 if (n->is_Type()) {
5483 t = n->as_Type()->type();
5484 assert(t == t->remove_speculative(), "no more speculative types");
5485 }
5486 // Iterate over outs - endless loops is unreachable from below
5487 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5488 Node *m = n->fast_out(i);
5489 if (not_a_node(m)) {
5490 continue;
5491 }
5492 worklist.push(m);
5493 }
5494 }
5495 igvn.check_no_speculative_types();
5496 #endif
5497 }
5498 }
5499
5500 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5501 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5502 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5503 if (!EnableValhalla || ta == nullptr || tb == nullptr ||
5504 ta->is_zero_type() || tb->is_zero_type() ||
5505 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5506 // Use old acmp if one operand is null or not an inline type
5507 return new CmpPNode(a, b);
5508 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5509 // We know that one operand is an inline type. Therefore,
5510 // new acmp will only return true if both operands are nullptr.
5511 // Check if both operands are null by or'ing the oops.
5512 a = phase->transform(new CastP2XNode(nullptr, a));
5513 b = phase->transform(new CastP2XNode(nullptr, b));
5514 a = phase->transform(new OrXNode(a, b));
5515 return new CmpXNode(a, phase->MakeConX(0));
5516 }
5517 // Use new acmp
5518 return nullptr;
5519 }
5520
5521 // Auxiliary methods to support randomized stressing/fuzzing.
5522
5523 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5524 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5525 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5526 FLAG_SET_ERGO(StressSeed, _stress_seed);
5527 } else {
5528 _stress_seed = StressSeed;
5529 }
5530 if (_log != nullptr) {
5531 _log->elem("stress_test seed='%u'", _stress_seed);
5532 }
5533 }
5534
5535 int Compile::random() {
5536 _stress_seed = os::next_random(_stress_seed);
5537 return static_cast<int>(_stress_seed);
5538 }
5539
5540 // This method can be called the arbitrary number of times, with current count
5541 // as the argument. The logic allows selecting a single candidate from the
5542 // running list of candidates as follows:
5543 // int count = 0;
5544 // Cand* selected = null;
5545 // while(cand = cand->next()) {
5546 // if (randomized_select(++count)) {
5547 // selected = cand;
5548 // }
5549 // }
5550 //
5551 // Including count equalizes the chances any candidate is "selected".
5552 // This is useful when we don't have the complete list of candidates to choose
5553 // from uniformly. In this case, we need to adjust the randomicity of the
5554 // selection, or else we will end up biasing the selection towards the latter
5555 // candidates.
5556 //
5557 // Quick back-envelope calculation shows that for the list of n candidates
5558 // the equal probability for the candidate to persist as "best" can be
5559 // achieved by replacing it with "next" k-th candidate with the probability
5560 // of 1/k. It can be easily shown that by the end of the run, the
5561 // probability for any candidate is converged to 1/n, thus giving the
5562 // uniform distribution among all the candidates.
5563 //
5564 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5565 #define RANDOMIZED_DOMAIN_POW 29
5566 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5567 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5568 bool Compile::randomized_select(int count) {
5569 assert(count > 0, "only positive");
5570 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5571 }
5572
5573 #ifdef ASSERT
5574 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5575 bool Compile::fail_randomly() {
5576 if ((random() % StressBailoutMean) != 0) {
5577 return false;
5578 }
5579 record_failure("StressBailout");
5580 return true;
5581 }
5582
5583 bool Compile::failure_is_artificial() {
5584 return C->failure_reason_is("StressBailout");
5585 }
5586 #endif
5587
5588 CloneMap& Compile::clone_map() { return _clone_map; }
5589 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5590
5591 void NodeCloneInfo::dump_on(outputStream* st) const {
5592 st->print(" {%d:%d} ", idx(), gen());
5593 }
5594
5595 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5596 uint64_t val = value(old->_idx);
5597 NodeCloneInfo cio(val);
5598 assert(val != 0, "old node should be in the map");
5599 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5600 insert(nnn->_idx, cin.get());
5601 #ifndef PRODUCT
5602 if (is_debug()) {
5603 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5604 }
5605 #endif
5606 }
5607
5608 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5609 NodeCloneInfo cio(value(old->_idx));
5610 if (cio.get() == 0) {
5611 cio.set(old->_idx, 0);
5612 insert(old->_idx, cio.get());
5613 #ifndef PRODUCT
5614 if (is_debug()) {
5615 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5616 }
5617 #endif
5618 }
5619 clone(old, nnn, gen);
5620 }
5621
5622 int CloneMap::max_gen() const {
5623 int g = 0;
5624 DictI di(_dict);
5625 for(; di.test(); ++di) {
5626 int t = gen(di._key);
5627 if (g < t) {
5628 g = t;
5629 #ifndef PRODUCT
5630 if (is_debug()) {
5631 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5632 }
5633 #endif
5634 }
5635 }
5636 return g;
5637 }
5638
5639 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5640 uint64_t val = value(key);
5641 if (val != 0) {
5642 NodeCloneInfo ni(val);
5643 ni.dump_on(st);
5644 }
5645 }
5646
5647 void Compile::shuffle_macro_nodes() {
5648 if (_macro_nodes.length() < 2) {
5649 return;
5650 }
5651 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5652 uint j = C->random() % (i + 1);
5653 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5654 }
5655 }
5656
5657 // Move Allocate nodes to the start of the list
5658 void Compile::sort_macro_nodes() {
5659 int count = macro_count();
5660 int allocates = 0;
5661 for (int i = 0; i < count; i++) {
5662 Node* n = macro_node(i);
5663 if (n->is_Allocate()) {
5664 if (i != allocates) {
5665 Node* tmp = macro_node(allocates);
5666 _macro_nodes.at_put(allocates, n);
5667 _macro_nodes.at_put(i, tmp);
5668 }
5669 allocates++;
5670 }
5671 }
5672 }
5673
5674 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5675 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5676 EventCompilerPhase event(UNTIMED);
5677 if (event.should_commit()) {
5678 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5679 }
5680 #ifndef PRODUCT
5681 ResourceMark rm;
5682 stringStream ss;
5683 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5684 int iter = ++_igv_phase_iter[cpt];
5685 if (iter > 1) {
5686 ss.print(" %d", iter);
5687 }
5688 if (n != nullptr) {
5689 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5690 if (n->is_Call()) {
5691 CallNode* call = n->as_Call();
5692 if (call->_name != nullptr) {
5693 // E.g. uncommon traps etc.
5694 ss.print(" - %s", call->_name);
5695 } else if (call->is_CallJava()) {
5696 CallJavaNode* call_java = call->as_CallJava();
5697 if (call_java->method() != nullptr) {
5698 ss.print(" -");
5699 call_java->method()->print_short_name(&ss);
5700 }
5701 }
5702 }
5703 }
5704
5705 const char* name = ss.as_string();
5706 if (should_print_igv(level)) {
5707 _igv_printer->print_graph(name);
5708 }
5709 if (should_print_phase(level)) {
5710 print_phase(name);
5711 }
5712 if (should_print_ideal_phase(cpt)) {
5713 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5714 }
5715 #endif
5716 C->_latest_stage_start_counter.stamp();
5717 }
5718
5719 // Only used from CompileWrapper
5720 void Compile::begin_method() {
5721 #ifndef PRODUCT
5722 if (_method != nullptr && should_print_igv(1)) {
5723 _igv_printer->begin_method();
5724 }
5725 #endif
5726 C->_latest_stage_start_counter.stamp();
5727 }
5728
5729 // Only used from CompileWrapper
5730 void Compile::end_method() {
5731 EventCompilerPhase event(UNTIMED);
5732 if (event.should_commit()) {
5733 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5734 }
5735
5736 #ifndef PRODUCT
5737 if (_method != nullptr && should_print_igv(1)) {
5738 _igv_printer->end_method();
5739 }
5740 #endif
5741 }
5742
5743 #ifndef PRODUCT
5744 bool Compile::should_print_phase(const int level) const {
5745 return PrintPhaseLevel > 0 && directive()->PhasePrintLevelOption >= level &&
5746 _method != nullptr; // Do not print phases for stubs.
5747 }
5748
5749 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5750 return _directive->should_print_ideal_phase(cpt);
5751 }
5752
5753 void Compile::init_igv() {
5754 if (_igv_printer == nullptr) {
5755 _igv_printer = IdealGraphPrinter::printer();
5756 _igv_printer->set_compile(this);
5757 }
5758 }
5759
5760 bool Compile::should_print_igv(const int level) {
5761 PRODUCT_RETURN_(return false;);
5762
5763 if (PrintIdealGraphLevel < 0) { // disabled by the user
5764 return false;
5765 }
5766
5767 bool need = directive()->IGVPrintLevelOption >= level;
5768 if (need) {
5769 Compile::init_igv();
5770 }
5771 return need;
5772 }
5773
5774 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5775 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5776
5777 // Called from debugger. Prints method to the default file with the default phase name.
5778 // This works regardless of any Ideal Graph Visualizer flags set or not.
5779 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5780 void igv_print(void* sp, void* fp, void* pc) {
5781 frame fr(sp, fp, pc);
5782 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5783 }
5784
5785 // Same as igv_print() above but with a specified phase name.
5786 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5787 frame fr(sp, fp, pc);
5788 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5789 }
5790
5791 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5792 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5793 // This works regardless of any Ideal Graph Visualizer flags set or not.
5794 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5795 void igv_print(bool network, void* sp, void* fp, void* pc) {
5796 frame fr(sp, fp, pc);
5797 if (network) {
5798 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5799 } else {
5800 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5801 }
5802 }
5803
5804 // Same as igv_print(bool network, ...) above but with a specified phase name.
5805 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5806 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5807 frame fr(sp, fp, pc);
5808 if (network) {
5809 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5810 } else {
5811 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5812 }
5813 }
5814
5815 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5816 void igv_print_default() {
5817 Compile::current()->print_method(PHASE_DEBUG, 0);
5818 }
5819
5820 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5821 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5822 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5823 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5824 void igv_append(void* sp, void* fp, void* pc) {
5825 frame fr(sp, fp, pc);
5826 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5827 }
5828
5829 // Same as igv_append(...) above but with a specified phase name.
5830 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5831 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5832 frame fr(sp, fp, pc);
5833 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5834 }
5835
5836 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5837 const char* file_name = "custom_debug.xml";
5838 if (_debug_file_printer == nullptr) {
5839 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5840 } else {
5841 _debug_file_printer->update_compiled_method(C->method());
5842 }
5843 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5844 _debug_file_printer->print_graph(phase_name, fr);
5845 }
5846
5847 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5848 ResourceMark rm;
5849 GrowableArray<const Node*> empty_list;
5850 igv_print_graph_to_network(phase_name, empty_list, fr);
5851 }
5852
5853 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5854 if (_debug_network_printer == nullptr) {
5855 _debug_network_printer = new IdealGraphPrinter(C);
5856 } else {
5857 _debug_network_printer->update_compiled_method(C->method());
5858 }
5859 tty->print_cr("Method printed over network stream to IGV");
5860 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5861 }
5862 #endif // !PRODUCT
5863
5864 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5865 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5866 return value;
5867 }
5868 Node* result = nullptr;
5869 if (bt == T_BYTE) {
5870 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5871 result = new RShiftINode(result, phase->intcon(24));
5872 } else if (bt == T_BOOLEAN) {
5873 result = new AndINode(value, phase->intcon(0xFF));
5874 } else if (bt == T_CHAR) {
5875 result = new AndINode(value,phase->intcon(0xFFFF));
5876 } else if (bt == T_FLOAT) {
5877 result = new MoveI2FNode(value);
5878 } else {
5879 assert(bt == T_SHORT, "unexpected narrow type");
5880 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5881 result = new RShiftINode(result, phase->intcon(16));
5882 }
5883 if (transform_res) {
5884 result = phase->transform(result);
5885 }
5886 return result;
5887 }
5888
5889 void Compile::record_method_not_compilable_oom() {
5890 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5891 }