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