1 /*
2 * Copyright (c) 1997, 2024, 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 "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/javaClasses.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/compilerOracle.hpp"
37 #include "compiler/compiler_globals.hpp"
38 #include "compiler/disassembler.hpp"
39 #include "compiler/oopMap.hpp"
40 #include "gc/shared/barrierSet.hpp"
41 #include "gc/shared/c2/barrierSetC2.hpp"
42 #include "jfr/jfrEvents.hpp"
43 #include "jvm_io.h"
44 #include "memory/allocation.hpp"
45 #include "memory/resourceArea.hpp"
46 #include "opto/addnode.hpp"
47 #include "opto/block.hpp"
48 #include "opto/c2compiler.hpp"
49 #include "opto/callGenerator.hpp"
50 #include "opto/callnode.hpp"
51 #include "opto/castnode.hpp"
52 #include "opto/cfgnode.hpp"
53 #include "opto/chaitin.hpp"
54 #include "opto/compile.hpp"
55 #include "opto/connode.hpp"
56 #include "opto/convertnode.hpp"
57 #include "opto/divnode.hpp"
58 #include "opto/escape.hpp"
59 #include "opto/idealGraphPrinter.hpp"
60 #include "opto/locknode.hpp"
61 #include "opto/loopnode.hpp"
62 #include "opto/machnode.hpp"
63 #include "opto/macro.hpp"
64 #include "opto/matcher.hpp"
65 #include "opto/mathexactnode.hpp"
66 #include "opto/memnode.hpp"
67 #include "opto/mulnode.hpp"
68 #include "opto/narrowptrnode.hpp"
69 #include "opto/node.hpp"
70 #include "opto/opcodes.hpp"
71 #include "opto/output.hpp"
72 #include "opto/parse.hpp"
73 #include "opto/phaseX.hpp"
74 #include "opto/rootnode.hpp"
75 #include "opto/runtime.hpp"
76 #include "opto/stringopts.hpp"
77 #include "opto/type.hpp"
78 #include "opto/vector.hpp"
79 #include "opto/vectornode.hpp"
80 #include "runtime/globals_extension.hpp"
81 #include "runtime/sharedRuntime.hpp"
82 #include "runtime/signature.hpp"
83 #include "runtime/stubRoutines.hpp"
84 #include "runtime/timer.hpp"
85 #include "utilities/align.hpp"
86 #include "utilities/copy.hpp"
87 #include "utilities/macros.hpp"
88 #include "utilities/resourceHash.hpp"
89
90 // -------------------- Compile::mach_constant_base_node -----------------------
91 // Constant table base node singleton.
92 MachConstantBaseNode* Compile::mach_constant_base_node() {
93 if (_mach_constant_base_node == nullptr) {
94 _mach_constant_base_node = new MachConstantBaseNode();
95 _mach_constant_base_node->add_req(C->root());
96 }
97 return _mach_constant_base_node;
98 }
99
100
101 /// Support for intrinsics.
102
103 // Return the index at which m must be inserted (or already exists).
104 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
105 class IntrinsicDescPair {
106 private:
107 ciMethod* _m;
108 bool _is_virtual;
109 public:
110 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
111 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
112 ciMethod* m= elt->method();
113 ciMethod* key_m = key->_m;
114 if (key_m < m) return -1;
115 else if (key_m > m) return 1;
116 else {
117 bool is_virtual = elt->is_virtual();
118 bool key_virtual = key->_is_virtual;
119 if (key_virtual < is_virtual) return -1;
120 else if (key_virtual > is_virtual) return 1;
121 else return 0;
122 }
123 }
124 };
125 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
126 #ifdef ASSERT
127 for (int i = 1; i < _intrinsics.length(); i++) {
128 CallGenerator* cg1 = _intrinsics.at(i-1);
129 CallGenerator* cg2 = _intrinsics.at(i);
130 assert(cg1->method() != cg2->method()
131 ? cg1->method() < cg2->method()
132 : cg1->is_virtual() < cg2->is_virtual(),
133 "compiler intrinsics list must stay sorted");
134 }
135 #endif
136 IntrinsicDescPair pair(m, is_virtual);
137 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
138 }
139
140 void Compile::register_intrinsic(CallGenerator* cg) {
141 bool found = false;
142 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
143 assert(!found, "registering twice");
144 _intrinsics.insert_before(index, cg);
145 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
146 }
147
148 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
149 assert(m->is_loaded(), "don't try this on unloaded methods");
150 if (_intrinsics.length() > 0) {
151 bool found = false;
152 int index = intrinsic_insertion_index(m, is_virtual, found);
153 if (found) {
154 return _intrinsics.at(index);
155 }
156 }
157 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
158 if (m->intrinsic_id() != vmIntrinsics::_none &&
159 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
160 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
161 if (cg != nullptr) {
162 // Save it for next time:
163 register_intrinsic(cg);
164 return cg;
165 } else {
166 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
167 }
168 }
169 return nullptr;
170 }
171
172 // Compile::make_vm_intrinsic is defined in library_call.cpp.
173
174 #ifndef PRODUCT
175 // statistics gathering...
176
177 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
179
180 inline int as_int(vmIntrinsics::ID id) {
181 return vmIntrinsics::as_int(id);
182 }
183
184 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
185 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
186 int oflags = _intrinsic_hist_flags[as_int(id)];
187 assert(flags != 0, "what happened?");
188 if (is_virtual) {
189 flags |= _intrinsic_virtual;
190 }
191 bool changed = (flags != oflags);
192 if ((flags & _intrinsic_worked) != 0) {
193 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
194 if (count == 1) {
195 changed = true; // first time
196 }
197 // increment the overall count also:
198 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
199 }
200 if (changed) {
201 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
202 // Something changed about the intrinsic's virtuality.
203 if ((flags & _intrinsic_virtual) != 0) {
204 // This is the first use of this intrinsic as a virtual call.
205 if (oflags != 0) {
206 // We already saw it as a non-virtual, so note both cases.
207 flags |= _intrinsic_both;
208 }
209 } else if ((oflags & _intrinsic_both) == 0) {
210 // This is the first use of this intrinsic as a non-virtual
211 flags |= _intrinsic_both;
212 }
213 }
214 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
215 }
216 // update the overall flags also:
217 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
218 return changed;
219 }
220
221 static char* format_flags(int flags, char* buf) {
222 buf[0] = 0;
223 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
224 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
225 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
226 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
227 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
228 if (buf[0] == 0) strcat(buf, ",");
229 assert(buf[0] == ',', "must be");
230 return &buf[1];
231 }
232
233 void Compile::print_intrinsic_statistics() {
234 char flagsbuf[100];
235 ttyLocker ttyl;
236 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
237 tty->print_cr("Compiler intrinsic usage:");
238 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
239 if (total == 0) total = 1; // avoid div0 in case of no successes
240 #define PRINT_STAT_LINE(name, c, f) \
241 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
242 for (auto id : EnumRange<vmIntrinsicID>{}) {
243 int flags = _intrinsic_hist_flags[as_int(id)];
244 juint count = _intrinsic_hist_count[as_int(id)];
245 if ((flags | count) != 0) {
246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
247 }
248 }
249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
250 if (xtty != nullptr) xtty->tail("statistics");
251 }
252
253 void Compile::print_statistics() {
254 { ttyLocker ttyl;
255 if (xtty != nullptr) xtty->head("statistics type='opto'");
256 Parse::print_statistics();
257 PhaseStringOpts::print_statistics();
258 PhaseCCP::print_statistics();
259 PhaseRegAlloc::print_statistics();
260 PhaseOutput::print_statistics();
261 PhasePeephole::print_statistics();
262 PhaseIdealLoop::print_statistics();
263 ConnectionGraph::print_statistics();
264 PhaseMacroExpand::print_statistics();
265 if (xtty != nullptr) xtty->tail("statistics");
266 }
267 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
268 // put this under its own <statistics> element.
269 print_intrinsic_statistics();
270 }
271 }
272 #endif //PRODUCT
273
274 void Compile::gvn_replace_by(Node* n, Node* nn) {
275 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
276 Node* use = n->last_out(i);
277 bool is_in_table = initial_gvn()->hash_delete(use);
278 uint uses_found = 0;
279 for (uint j = 0; j < use->len(); j++) {
280 if (use->in(j) == n) {
281 if (j < use->req())
282 use->set_req(j, nn);
283 else
284 use->set_prec(j, nn);
285 uses_found++;
286 }
287 }
288 if (is_in_table) {
289 // reinsert into table
290 initial_gvn()->hash_find_insert(use);
291 }
292 record_for_igvn(use);
293 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
294 i -= uses_found; // we deleted 1 or more copies of this edge
295 }
296 }
297
298
299 // Identify all nodes that are reachable from below, useful.
300 // Use breadth-first pass that records state in a Unique_Node_List,
301 // recursive traversal is slower.
302 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
303 int estimated_worklist_size = live_nodes();
304 useful.map( estimated_worklist_size, nullptr ); // preallocate space
305
306 // Initialize worklist
307 if (root() != nullptr) { useful.push(root()); }
308 // If 'top' is cached, declare it useful to preserve cached node
309 if (cached_top_node()) { useful.push(cached_top_node()); }
310
311 // Push all useful nodes onto the list, breadthfirst
312 for( uint next = 0; next < useful.size(); ++next ) {
313 assert( next < unique(), "Unique useful nodes < total nodes");
314 Node *n = useful.at(next);
315 uint max = n->len();
316 for( uint i = 0; i < max; ++i ) {
317 Node *m = n->in(i);
318 if (not_a_node(m)) continue;
319 useful.push(m);
320 }
321 }
322 }
323
324 // Update dead_node_list with any missing dead nodes using useful
325 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
326 void Compile::update_dead_node_list(Unique_Node_List &useful) {
327 uint max_idx = unique();
328 VectorSet& useful_node_set = useful.member_set();
329
330 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
331 // If node with index node_idx is not in useful set,
332 // mark it as dead in dead node list.
333 if (!useful_node_set.test(node_idx)) {
334 record_dead_node(node_idx);
335 }
336 }
337 }
338
339 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
340 int shift = 0;
341 for (int i = 0; i < inlines->length(); i++) {
342 CallGenerator* cg = inlines->at(i);
343 if (useful.member(cg->call_node())) {
344 if (shift > 0) {
345 inlines->at_put(i - shift, cg);
346 }
347 } else {
348 shift++; // skip over the dead element
349 }
350 }
351 if (shift > 0) {
352 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
353 }
354 }
355
356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
357 assert(dead != nullptr && dead->is_Call(), "sanity");
358 int found = 0;
359 for (int i = 0; i < inlines->length(); i++) {
360 if (inlines->at(i)->call_node() == dead) {
361 inlines->remove_at(i);
362 found++;
363 NOT_DEBUG( break; ) // elements are unique, so exit early
364 }
365 }
366 assert(found <= 1, "not unique");
367 }
368
369 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
370 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
371 for (int i = node_list.length() - 1; i >= 0; i--) {
372 N* node = node_list.at(i);
373 if (!useful.member(node)) {
374 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
375 }
376 }
377 }
378
379 void Compile::remove_useless_node(Node* dead) {
380 remove_modified_node(dead);
381
382 // Constant node that has no out-edges and has only one in-edge from
383 // root is usually dead. However, sometimes reshaping walk makes
384 // it reachable by adding use edges. So, we will NOT count Con nodes
385 // as dead to be conservative about the dead node count at any
386 // given time.
387 if (!dead->is_Con()) {
388 record_dead_node(dead->_idx);
389 }
390 if (dead->is_macro()) {
391 remove_macro_node(dead);
392 }
393 if (dead->is_expensive()) {
394 remove_expensive_node(dead);
395 }
396 if (dead->is_OpaqueTemplateAssertionPredicate()) {
397 remove_template_assertion_predicate_opaq(dead);
398 }
399 if (dead->is_ParsePredicate()) {
400 remove_parse_predicate(dead->as_ParsePredicate());
401 }
402 if (dead->for_post_loop_opts_igvn()) {
403 remove_from_post_loop_opts_igvn(dead);
404 }
405 if (dead->is_Call()) {
406 remove_useless_late_inlines( &_late_inlines, dead);
407 remove_useless_late_inlines( &_string_late_inlines, dead);
408 remove_useless_late_inlines( &_boxing_late_inlines, dead);
409 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
410
411 if (dead->is_CallStaticJava()) {
412 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
413 }
414 }
415 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
416 bs->unregister_potential_barrier_node(dead);
417 }
418
419 // Disconnect all useless nodes by disconnecting those at the boundary.
420 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
421 uint next = 0;
422 while (next < useful.size()) {
423 Node *n = useful.at(next++);
424 if (n->is_SafePoint()) {
425 // We're done with a parsing phase. Replaced nodes are not valid
426 // beyond that point.
427 n->as_SafePoint()->delete_replaced_nodes();
428 }
429 // Use raw traversal of out edges since this code removes out edges
430 int max = n->outcnt();
431 for (int j = 0; j < max; ++j) {
432 Node* child = n->raw_out(j);
433 if (!useful.member(child)) {
434 assert(!child->is_top() || child != top(),
435 "If top is cached in Compile object it is in useful list");
436 // Only need to remove this out-edge to the useless node
437 n->raw_del_out(j);
438 --j;
439 --max;
440 }
441 }
442 if (n->outcnt() == 1 && n->has_special_unique_user()) {
443 assert(useful.member(n->unique_out()), "do not push a useless node");
444 worklist.push(n->unique_out());
445 }
446 }
447
448 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
449 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
450 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
451 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
452 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
453 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
454 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
455 #ifdef ASSERT
456 if (_modified_nodes != nullptr) {
457 _modified_nodes->remove_useless_nodes(useful.member_set());
458 }
459 #endif
460
461 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
462 bs->eliminate_useless_gc_barriers(useful, this);
463 // clean up the late inline lists
464 remove_useless_late_inlines( &_late_inlines, useful);
465 remove_useless_late_inlines( &_string_late_inlines, useful);
466 remove_useless_late_inlines( &_boxing_late_inlines, useful);
467 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
468 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
469 }
470
471 // ============================================================================
472 //------------------------------CompileWrapper---------------------------------
473 class CompileWrapper : public StackObj {
474 Compile *const _compile;
475 public:
476 CompileWrapper(Compile* compile);
477
478 ~CompileWrapper();
479 };
480
481 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
482 // the Compile* pointer is stored in the current ciEnv:
483 ciEnv* env = compile->env();
484 assert(env == ciEnv::current(), "must already be a ciEnv active");
485 assert(env->compiler_data() == nullptr, "compile already active?");
486 env->set_compiler_data(compile);
487 assert(compile == Compile::current(), "sanity");
488
489 compile->set_type_dict(nullptr);
490 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
491 compile->clone_map().set_clone_idx(0);
492 compile->set_type_last_size(0);
493 compile->set_last_tf(nullptr, nullptr);
494 compile->set_indexSet_arena(nullptr);
495 compile->set_indexSet_free_block_list(nullptr);
496 compile->init_type_arena();
497 Type::Initialize(compile);
498 _compile->begin_method();
499 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
500 }
501 CompileWrapper::~CompileWrapper() {
502 // simulate crash during compilation
503 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
504
505 _compile->end_method();
506 _compile->env()->set_compiler_data(nullptr);
507 }
508
509
510 //----------------------------print_compile_messages---------------------------
511 void Compile::print_compile_messages() {
512 #ifndef PRODUCT
513 // Check if recompiling
514 if (!subsume_loads() && PrintOpto) {
515 // Recompiling without allowing machine instructions to subsume loads
516 tty->print_cr("*********************************************************");
517 tty->print_cr("** Bailout: Recompile without subsuming loads **");
518 tty->print_cr("*********************************************************");
519 }
520 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
521 // Recompiling without escape analysis
522 tty->print_cr("*********************************************************");
523 tty->print_cr("** Bailout: Recompile without escape analysis **");
524 tty->print_cr("*********************************************************");
525 }
526 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
527 // Recompiling without iterative escape analysis
528 tty->print_cr("*********************************************************");
529 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
530 tty->print_cr("*********************************************************");
531 }
532 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
533 // Recompiling without reducing allocation merges
534 tty->print_cr("*********************************************************");
535 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
536 tty->print_cr("*********************************************************");
537 }
538 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
539 // Recompiling without boxing elimination
540 tty->print_cr("*********************************************************");
541 tty->print_cr("** Bailout: Recompile without boxing elimination **");
542 tty->print_cr("*********************************************************");
543 }
544 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
545 // Recompiling without locks coarsening
546 tty->print_cr("*********************************************************");
547 tty->print_cr("** Bailout: Recompile without locks coarsening **");
548 tty->print_cr("*********************************************************");
549 }
550 if (env()->break_at_compile()) {
551 // Open the debugger when compiling this method.
552 tty->print("### Breaking when compiling: ");
553 method()->print_short_name();
554 tty->cr();
555 BREAKPOINT;
556 }
557
558 if( PrintOpto ) {
559 if (is_osr_compilation()) {
560 tty->print("[OSR]%3d", _compile_id);
561 } else {
562 tty->print("%3d", _compile_id);
563 }
564 }
565 #endif
566 }
567
568 #ifndef PRODUCT
569 void Compile::print_ideal_ir(const char* phase_name) {
570 // keep the following output all in one block
571 // This output goes directly to the tty, not the compiler log.
572 // To enable tools to match it up with the compilation activity,
573 // be sure to tag this tty output with the compile ID.
574
575 // Node dumping can cause a safepoint, which can break the tty lock.
576 // Buffer all node dumps, so that all safepoints happen before we lock.
577 ResourceMark rm;
578 stringStream ss;
579
580 if (_output == nullptr) {
581 ss.print_cr("AFTER: %s", phase_name);
582 // Print out all nodes in ascending order of index.
583 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
584 } else {
585 // Dump the node blockwise if we have a scheduling
586 _output->print_scheduling(&ss);
587 }
588
589 // Check that the lock is not broken by a safepoint.
590 NoSafepointVerifier nsv;
591 ttyLocker ttyl;
592 if (xtty != nullptr) {
593 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
594 compile_id(),
595 is_osr_compilation() ? " compile_kind='osr'" : "",
596 phase_name);
597 }
598
599 tty->print("%s", ss.as_string());
600
601 if (xtty != nullptr) {
602 xtty->tail("ideal");
603 }
604 }
605 #endif
606
607 // ============================================================================
608 //------------------------------Compile standard-------------------------------
609
610 // Compile a method. entry_bci is -1 for normal compilations and indicates
611 // the continuation bci for on stack replacement.
612
613
614 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
615 Options options, DirectiveSet* directive)
616 : Phase(Compiler),
617 _compile_id(ci_env->compile_id()),
618 _options(options),
619 _method(target),
620 _entry_bci(osr_bci),
621 _ilt(nullptr),
622 _stub_function(nullptr),
623 _stub_name(nullptr),
624 _stub_entry_point(nullptr),
625 _max_node_limit(MaxNodeLimit),
626 _post_loop_opts_phase(false),
627 _allow_macro_nodes(true),
628 _inlining_progress(false),
629 _inlining_incrementally(false),
630 _do_cleanup(false),
631 _has_reserved_stack_access(target->has_reserved_stack_access()),
632 #ifndef PRODUCT
633 _igv_idx(0),
634 _trace_opto_output(directive->TraceOptoOutputOption),
635 #endif
636 _has_method_handle_invokes(false),
637 _clinit_barrier_on_entry(false),
638 _has_clinit_barriers(false),
639 _stress_seed(0),
640 _comp_arena(mtCompiler),
641 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
642 _env(ci_env),
643 _directive(directive),
644 _log(ci_env->log()),
645 _first_failure_details(nullptr),
646 _intrinsics (comp_arena(), 0, 0, nullptr),
647 _macro_nodes (comp_arena(), 8, 0, nullptr),
648 _parse_predicates (comp_arena(), 8, 0, nullptr),
649 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
650 _expensive_nodes (comp_arena(), 8, 0, nullptr),
651 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
652 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
653 _coarsened_locks (comp_arena(), 8, 0, nullptr),
654 _congraph(nullptr),
655 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
656 _unique(0),
657 _dead_node_count(0),
658 _dead_node_list(comp_arena()),
659 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
660 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
661 _node_arena(&_node_arena_one),
662 _mach_constant_base_node(nullptr),
663 _Compile_types(mtCompiler),
664 _initial_gvn(nullptr),
665 _igvn_worklist(nullptr),
666 _types(nullptr),
667 _node_hash(nullptr),
668 _late_inlines(comp_arena(), 2, 0, nullptr),
669 _string_late_inlines(comp_arena(), 2, 0, nullptr),
670 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
671 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
672 _late_inlines_pos(0),
673 _number_of_mh_late_inlines(0),
674 _oom(false),
675 _print_inlining_stream(new (mtCompiler) stringStream()),
676 _print_inlining_list(nullptr),
677 _print_inlining_idx(0),
678 _print_inlining_output(nullptr),
679 _replay_inline_data(nullptr),
680 _java_calls(0),
681 _inner_loops(0),
682 _interpreter_frame_size(0),
683 _output(nullptr)
684 #ifndef PRODUCT
685 , _in_dump_cnt(0)
686 #endif
687 {
688 C = this;
689 CompileWrapper cw(this);
690
691 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
692 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
693
694 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
695 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
696 // We can always print a disassembly, either abstract (hex dump) or
697 // with the help of a suitable hsdis library. Thus, we should not
698 // couple print_assembly and print_opto_assembly controls.
699 // But: always print opto and regular assembly on compile command 'print'.
700 bool print_assembly = directive->PrintAssemblyOption;
701 set_print_assembly(print_opto_assembly || print_assembly);
702 #else
703 set_print_assembly(false); // must initialize.
704 #endif
705
706 #ifndef PRODUCT
707 set_parsed_irreducible_loop(false);
708 #endif
709
710 if (directive->ReplayInlineOption) {
711 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
712 }
713 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
714 set_print_intrinsics(directive->PrintIntrinsicsOption);
715 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
716
717 if (ProfileTraps) {
718 // Make sure the method being compiled gets its own MDO,
719 // so we can at least track the decompile_count().
720 method()->ensure_method_data();
721 }
722
723 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
724 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout) {
725 initialize_stress_seed(directive);
726 }
727
728 Init(/*do_aliasing=*/ true);
729
730 print_compile_messages();
731
732 _ilt = InlineTree::build_inline_tree_root();
733
734 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
735 assert(num_alias_types() >= AliasIdxRaw, "");
736
737 #define MINIMUM_NODE_HASH 1023
738
739 // GVN that will be run immediately on new nodes
740 uint estimated_size = method()->code_size()*4+64;
741 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
742 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
743 _types = new (comp_arena()) Type_Array(comp_arena());
744 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
745 PhaseGVN gvn;
746 set_initial_gvn(&gvn);
747
748 print_inlining_init();
749 { // Scope for timing the parser
750 TracePhase tp(_t_parser);
751
752 // Put top into the hash table ASAP.
753 initial_gvn()->transform(top());
754
755 // Set up tf(), start(), and find a CallGenerator.
756 CallGenerator* cg = nullptr;
757 if (is_osr_compilation()) {
758 const TypeTuple *domain = StartOSRNode::osr_domain();
759 const TypeTuple *range = TypeTuple::make_range(method()->signature());
760 init_tf(TypeFunc::make(domain, range));
761 StartNode* s = new StartOSRNode(root(), domain);
762 initial_gvn()->set_type_bottom(s);
763 verify_start(s);
764 cg = CallGenerator::for_osr(method(), entry_bci());
765 } else {
766 // Normal case.
767 init_tf(TypeFunc::make(method()));
768 StartNode* s = new StartNode(root(), tf()->domain());
769 initial_gvn()->set_type_bottom(s);
770 verify_start(s);
771 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
772 // With java.lang.ref.reference.get() we must go through the
773 // intrinsic - even when get() is the root
774 // method of the compile - so that, if necessary, the value in
775 // the referent field of the reference object gets recorded by
776 // the pre-barrier code.
777 cg = find_intrinsic(method(), false);
778 }
779 if (cg == nullptr) {
780 float past_uses = method()->interpreter_invocation_count();
781 float expected_uses = past_uses;
782 cg = CallGenerator::for_inline(method(), expected_uses);
783 }
784 }
785 if (failing()) return;
786 if (cg == nullptr) {
787 const char* reason = InlineTree::check_can_parse(method());
788 assert(reason != nullptr, "expect reason for parse failure");
789 stringStream ss;
790 ss.print("cannot parse method: %s", reason);
791 record_method_not_compilable(ss.as_string());
792 return;
793 }
794
795 gvn.set_type(root(), root()->bottom_type());
796
797 JVMState* jvms = build_start_state(start(), tf());
798 if ((jvms = cg->generate(jvms)) == nullptr) {
799 assert(failure_reason() != nullptr, "expect reason for parse failure");
800 stringStream ss;
801 ss.print("method parse failed: %s", failure_reason());
802 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
803 return;
804 }
805 GraphKit kit(jvms);
806
807 if (!kit.stopped()) {
808 // Accept return values, and transfer control we know not where.
809 // This is done by a special, unique ReturnNode bound to root.
810 return_values(kit.jvms());
811 }
812
813 if (kit.has_exceptions()) {
814 // Any exceptions that escape from this call must be rethrown
815 // to whatever caller is dynamically above us on the stack.
816 // This is done by a special, unique RethrowNode bound to root.
817 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
818 }
819
820 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
821
822 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
823 inline_string_calls(true);
824 }
825
826 if (failing()) return;
827
828 // Remove clutter produced by parsing.
829 if (!failing()) {
830 ResourceMark rm;
831 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
832 }
833 }
834
835 // Note: Large methods are capped off in do_one_bytecode().
836 if (failing()) return;
837
838 // After parsing, node notes are no longer automagic.
839 // They must be propagated by register_new_node_with_optimizer(),
840 // clone(), or the like.
841 set_default_node_notes(nullptr);
842
843 #ifndef PRODUCT
844 if (should_print_igv(1)) {
845 _igv_printer->print_inlining();
846 }
847 #endif
848
849 if (failing()) return;
850 NOT_PRODUCT( verify_graph_edges(); )
851
852 // Now optimize
853 Optimize();
854 if (failing()) return;
855 NOT_PRODUCT( verify_graph_edges(); )
856
857 #ifndef PRODUCT
858 if (should_print_ideal()) {
859 print_ideal_ir("print_ideal");
860 }
861 #endif
862
863 #ifdef ASSERT
864 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
865 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
866 #endif
867
868 // Dump compilation data to replay it.
869 if (directive->DumpReplayOption) {
870 env()->dump_replay_data(_compile_id);
871 }
872 if (directive->DumpInlineOption && (ilt() != nullptr)) {
873 env()->dump_inline_data(_compile_id);
874 }
875
876 // Now that we know the size of all the monitors we can add a fixed slot
877 // for the original deopt pc.
878 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
879 set_fixed_slots(next_slot);
880
881 // Compute when to use implicit null checks. Used by matching trap based
882 // nodes and NullCheck optimization.
883 set_allowed_deopt_reasons();
884
885 // Now generate code
886 Code_Gen();
887 }
888
889 //------------------------------Compile----------------------------------------
890 // Compile a runtime stub
891 Compile::Compile( ciEnv* ci_env,
892 TypeFunc_generator generator,
893 address stub_function,
894 const char *stub_name,
895 int is_fancy_jump,
896 bool pass_tls,
897 bool return_pc,
898 DirectiveSet* directive)
899 : Phase(Compiler),
900 _compile_id(0),
901 _options(Options::for_runtime_stub()),
902 _method(nullptr),
903 _entry_bci(InvocationEntryBci),
904 _stub_function(stub_function),
905 _stub_name(stub_name),
906 _stub_entry_point(nullptr),
907 _max_node_limit(MaxNodeLimit),
908 _post_loop_opts_phase(false),
909 _allow_macro_nodes(true),
910 _inlining_progress(false),
911 _inlining_incrementally(false),
912 _has_reserved_stack_access(false),
913 #ifndef PRODUCT
914 _igv_idx(0),
915 _trace_opto_output(directive->TraceOptoOutputOption),
916 #endif
917 _has_method_handle_invokes(false),
918 _clinit_barrier_on_entry(false),
919 _has_clinit_barriers(false),
920 _stress_seed(0),
921 _comp_arena(mtCompiler),
922 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
923 _env(ci_env),
924 _directive(directive),
925 _log(ci_env->log()),
926 _first_failure_details(nullptr),
927 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
928 _congraph(nullptr),
929 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
930 _unique(0),
931 _dead_node_count(0),
932 _dead_node_list(comp_arena()),
933 _node_arena_one(mtCompiler),
934 _node_arena_two(mtCompiler),
935 _node_arena(&_node_arena_one),
936 _mach_constant_base_node(nullptr),
937 _Compile_types(mtCompiler),
938 _initial_gvn(nullptr),
939 _igvn_worklist(nullptr),
940 _types(nullptr),
941 _node_hash(nullptr),
942 _number_of_mh_late_inlines(0),
943 _oom(false),
944 _print_inlining_stream(new (mtCompiler) stringStream()),
945 _print_inlining_list(nullptr),
946 _print_inlining_idx(0),
947 _print_inlining_output(nullptr),
948 _replay_inline_data(nullptr),
949 _java_calls(0),
950 _inner_loops(0),
951 _interpreter_frame_size(0),
952 _output(nullptr),
953 #ifndef PRODUCT
954 _in_dump_cnt(0),
955 #endif
956 _allowed_reasons(0) {
957 C = this;
958
959 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
960 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
961
962 #ifndef PRODUCT
963 set_print_assembly(PrintFrameConverterAssembly);
964 set_parsed_irreducible_loop(false);
965 #else
966 set_print_assembly(false); // Must initialize.
967 #endif
968 set_has_irreducible_loop(false); // no loops
969
970 CompileWrapper cw(this);
971 Init(/*do_aliasing=*/ false);
972 init_tf((*generator)());
973
974 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
975 _types = new (comp_arena()) Type_Array(comp_arena());
976 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
977
978 if (StressLCM || StressGCM || StressBailout) {
979 initialize_stress_seed(directive);
980 }
981
982 {
983 PhaseGVN gvn;
984 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
985 gvn.transform(top());
986
987 GraphKit kit;
988 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
989 }
990
991 NOT_PRODUCT( verify_graph_edges(); )
992
993 Code_Gen();
994 }
995
996 Compile::~Compile() {
997 delete _print_inlining_stream;
998 delete _first_failure_details;
999 };
1000
1001 //------------------------------Init-------------------------------------------
1002 // Prepare for a single compilation
1003 void Compile::Init(bool aliasing) {
1004 _do_aliasing = aliasing;
1005 _unique = 0;
1006 _regalloc = nullptr;
1007
1008 _tf = nullptr; // filled in later
1009 _top = nullptr; // cached later
1010 _matcher = nullptr; // filled in later
1011 _cfg = nullptr; // filled in later
1012
1013 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1014
1015 _node_note_array = nullptr;
1016 _default_node_notes = nullptr;
1017 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1018
1019 _immutable_memory = nullptr; // filled in at first inquiry
1020
1021 #ifdef ASSERT
1022 _phase_optimize_finished = false;
1023 _phase_verify_ideal_loop = false;
1024 _exception_backedge = false;
1025 _type_verify = nullptr;
1026 #endif
1027
1028 // Globally visible Nodes
1029 // First set TOP to null to give safe behavior during creation of RootNode
1030 set_cached_top_node(nullptr);
1031 set_root(new RootNode());
1032 // Now that you have a Root to point to, create the real TOP
1033 set_cached_top_node( new ConNode(Type::TOP) );
1034 set_recent_alloc(nullptr, nullptr);
1035
1036 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1037 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1038 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1039 env()->set_dependencies(new Dependencies(env()));
1040
1041 _fixed_slots = 0;
1042 set_has_split_ifs(false);
1043 set_has_loops(false); // first approximation
1044 set_has_stringbuilder(false);
1045 set_has_boxed_value(false);
1046 _trap_can_recompile = false; // no traps emitted yet
1047 _major_progress = true; // start out assuming good things will happen
1048 set_has_unsafe_access(false);
1049 set_max_vector_size(0);
1050 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1051 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1052 set_decompile_count(0);
1053
1054 #ifndef PRODUCT
1055 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1056 #endif
1057
1058 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1059 _loop_opts_cnt = LoopOptsCount;
1060 set_do_inlining(Inline);
1061 set_max_inline_size(MaxInlineSize);
1062 set_freq_inline_size(FreqInlineSize);
1063 set_do_scheduling(OptoScheduling);
1064
1065 set_do_vector_loop(false);
1066 set_has_monitors(false);
1067 set_has_scoped_access(false);
1068
1069 if (AllowVectorizeOnDemand) {
1070 if (has_method() && _directive->VectorizeOption) {
1071 set_do_vector_loop(true);
1072 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());})
1073 } else if (has_method() && method()->name() != nullptr &&
1074 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1075 set_do_vector_loop(true);
1076 }
1077 }
1078 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1079 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());})
1080
1081 _max_node_limit = _directive->MaxNodeLimitOption;
1082
1083 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() &&
1084 (method()->needs_clinit_barrier() || (do_clinit_barriers() && method()->is_static()))) {
1085 set_clinit_barrier_on_entry(true);
1086 if (do_clinit_barriers()) {
1087 set_has_clinit_barriers(true); // Entry clinit barrier is in prolog code.
1088 }
1089 }
1090 if (debug_info()->recording_non_safepoints()) {
1091 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1092 (comp_arena(), 8, 0, nullptr));
1093 set_default_node_notes(Node_Notes::make(this));
1094 }
1095
1096 const int grow_ats = 16;
1097 _max_alias_types = grow_ats;
1098 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1099 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1100 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1101 {
1102 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1103 }
1104 // Initialize the first few types.
1105 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1106 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1107 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1108 _num_alias_types = AliasIdxRaw+1;
1109 // Zero out the alias type cache.
1110 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1111 // A null adr_type hits in the cache right away. Preload the right answer.
1112 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1113 }
1114
1115 #ifdef ASSERT
1116 // Verify that the current StartNode is valid.
1117 void Compile::verify_start(StartNode* s) const {
1118 assert(failing_internal() || s == start(), "should be StartNode");
1119 }
1120 #endif
1121
1122 /**
1123 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1124 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1125 * the ideal graph.
1126 */
1127 StartNode* Compile::start() const {
1128 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1129 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1130 Node* start = root()->fast_out(i);
1131 if (start->is_Start()) {
1132 return start->as_Start();
1133 }
1134 }
1135 fatal("Did not find Start node!");
1136 return nullptr;
1137 }
1138
1139 //-------------------------------immutable_memory-------------------------------------
1140 // Access immutable memory
1141 Node* Compile::immutable_memory() {
1142 if (_immutable_memory != nullptr) {
1143 return _immutable_memory;
1144 }
1145 StartNode* s = start();
1146 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1147 Node *p = s->fast_out(i);
1148 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1149 _immutable_memory = p;
1150 return _immutable_memory;
1151 }
1152 }
1153 ShouldNotReachHere();
1154 return nullptr;
1155 }
1156
1157 //----------------------set_cached_top_node------------------------------------
1158 // Install the cached top node, and make sure Node::is_top works correctly.
1159 void Compile::set_cached_top_node(Node* tn) {
1160 if (tn != nullptr) verify_top(tn);
1161 Node* old_top = _top;
1162 _top = tn;
1163 // Calling Node::setup_is_top allows the nodes the chance to adjust
1164 // their _out arrays.
1165 if (_top != nullptr) _top->setup_is_top();
1166 if (old_top != nullptr) old_top->setup_is_top();
1167 assert(_top == nullptr || top()->is_top(), "");
1168 }
1169
1170 #ifdef ASSERT
1171 uint Compile::count_live_nodes_by_graph_walk() {
1172 Unique_Node_List useful(comp_arena());
1173 // Get useful node list by walking the graph.
1174 identify_useful_nodes(useful);
1175 return useful.size();
1176 }
1177
1178 void Compile::print_missing_nodes() {
1179
1180 // Return if CompileLog is null and PrintIdealNodeCount is false.
1181 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1182 return;
1183 }
1184
1185 // This is an expensive function. It is executed only when the user
1186 // specifies VerifyIdealNodeCount option or otherwise knows the
1187 // additional work that needs to be done to identify reachable nodes
1188 // by walking the flow graph and find the missing ones using
1189 // _dead_node_list.
1190
1191 Unique_Node_List useful(comp_arena());
1192 // Get useful node list by walking the graph.
1193 identify_useful_nodes(useful);
1194
1195 uint l_nodes = C->live_nodes();
1196 uint l_nodes_by_walk = useful.size();
1197
1198 if (l_nodes != l_nodes_by_walk) {
1199 if (_log != nullptr) {
1200 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1201 _log->stamp();
1202 _log->end_head();
1203 }
1204 VectorSet& useful_member_set = useful.member_set();
1205 int last_idx = l_nodes_by_walk;
1206 for (int i = 0; i < last_idx; i++) {
1207 if (useful_member_set.test(i)) {
1208 if (_dead_node_list.test(i)) {
1209 if (_log != nullptr) {
1210 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1211 }
1212 if (PrintIdealNodeCount) {
1213 // Print the log message to tty
1214 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1215 useful.at(i)->dump();
1216 }
1217 }
1218 }
1219 else if (! _dead_node_list.test(i)) {
1220 if (_log != nullptr) {
1221 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1222 }
1223 if (PrintIdealNodeCount) {
1224 // Print the log message to tty
1225 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1226 }
1227 }
1228 }
1229 if (_log != nullptr) {
1230 _log->tail("mismatched_nodes");
1231 }
1232 }
1233 }
1234 void Compile::record_modified_node(Node* n) {
1235 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1236 _modified_nodes->push(n);
1237 }
1238 }
1239
1240 void Compile::remove_modified_node(Node* n) {
1241 if (_modified_nodes != nullptr) {
1242 _modified_nodes->remove(n);
1243 }
1244 }
1245 #endif
1246
1247 #ifndef PRODUCT
1248 void Compile::verify_top(Node* tn) const {
1249 if (tn != nullptr) {
1250 assert(tn->is_Con(), "top node must be a constant");
1251 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1252 assert(tn->in(0) != nullptr, "must have live top node");
1253 }
1254 }
1255 #endif
1256
1257
1258 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1259
1260 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1261 guarantee(arr != nullptr, "");
1262 int num_blocks = arr->length();
1263 if (grow_by < num_blocks) grow_by = num_blocks;
1264 int num_notes = grow_by * _node_notes_block_size;
1265 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1266 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1267 while (num_notes > 0) {
1268 arr->append(notes);
1269 notes += _node_notes_block_size;
1270 num_notes -= _node_notes_block_size;
1271 }
1272 assert(num_notes == 0, "exact multiple, please");
1273 }
1274
1275 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1276 if (source == nullptr || dest == nullptr) return false;
1277
1278 if (dest->is_Con())
1279 return false; // Do not push debug info onto constants.
1280
1281 #ifdef ASSERT
1282 // Leave a bread crumb trail pointing to the original node:
1283 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1284 dest->set_debug_orig(source);
1285 }
1286 #endif
1287
1288 if (node_note_array() == nullptr)
1289 return false; // Not collecting any notes now.
1290
1291 // This is a copy onto a pre-existing node, which may already have notes.
1292 // If both nodes have notes, do not overwrite any pre-existing notes.
1293 Node_Notes* source_notes = node_notes_at(source->_idx);
1294 if (source_notes == nullptr || source_notes->is_clear()) return false;
1295 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1296 if (dest_notes == nullptr || dest_notes->is_clear()) {
1297 return set_node_notes_at(dest->_idx, source_notes);
1298 }
1299
1300 Node_Notes merged_notes = (*source_notes);
1301 // The order of operations here ensures that dest notes will win...
1302 merged_notes.update_from(dest_notes);
1303 return set_node_notes_at(dest->_idx, &merged_notes);
1304 }
1305
1306
1307 //--------------------------allow_range_check_smearing-------------------------
1308 // Gating condition for coalescing similar range checks.
1309 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1310 // single covering check that is at least as strong as any of them.
1311 // If the optimization succeeds, the simplified (strengthened) range check
1312 // will always succeed. If it fails, we will deopt, and then give up
1313 // on the optimization.
1314 bool Compile::allow_range_check_smearing() const {
1315 // If this method has already thrown a range-check,
1316 // assume it was because we already tried range smearing
1317 // and it failed.
1318 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1319 return !already_trapped;
1320 }
1321
1322
1323 //------------------------------flatten_alias_type-----------------------------
1324 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1325 assert(do_aliasing(), "Aliasing should be enabled");
1326 int offset = tj->offset();
1327 TypePtr::PTR ptr = tj->ptr();
1328
1329 // Known instance (scalarizable allocation) alias only with itself.
1330 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1331 tj->is_oopptr()->is_known_instance();
1332
1333 // Process weird unsafe references.
1334 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1335 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1336 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1337 tj = TypeOopPtr::BOTTOM;
1338 ptr = tj->ptr();
1339 offset = tj->offset();
1340 }
1341
1342 // Array pointers need some flattening
1343 const TypeAryPtr* ta = tj->isa_aryptr();
1344 if (ta && ta->is_stable()) {
1345 // Erase stability property for alias analysis.
1346 tj = ta = ta->cast_to_stable(false);
1347 }
1348 if( ta && is_known_inst ) {
1349 if ( offset != Type::OffsetBot &&
1350 offset > arrayOopDesc::length_offset_in_bytes() ) {
1351 offset = Type::OffsetBot; // Flatten constant access into array body only
1352 tj = ta = ta->
1353 remove_speculative()->
1354 cast_to_ptr_type(ptr)->
1355 with_offset(offset);
1356 }
1357 } else if (ta) {
1358 // For arrays indexed by constant indices, we flatten the alias
1359 // space to include all of the array body. Only the header, klass
1360 // and array length can be accessed un-aliased.
1361 if( offset != Type::OffsetBot ) {
1362 if( ta->const_oop() ) { // MethodData* or Method*
1363 offset = Type::OffsetBot; // Flatten constant access into array body
1364 tj = ta = ta->
1365 remove_speculative()->
1366 cast_to_ptr_type(ptr)->
1367 cast_to_exactness(false)->
1368 with_offset(offset);
1369 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1370 // range is OK as-is.
1371 tj = ta = TypeAryPtr::RANGE;
1372 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1373 tj = TypeInstPtr::KLASS; // all klass loads look alike
1374 ta = TypeAryPtr::RANGE; // generic ignored junk
1375 ptr = TypePtr::BotPTR;
1376 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1377 tj = TypeInstPtr::MARK;
1378 ta = TypeAryPtr::RANGE; // generic ignored junk
1379 ptr = TypePtr::BotPTR;
1380 } else { // Random constant offset into array body
1381 offset = Type::OffsetBot; // Flatten constant access into array body
1382 tj = ta = ta->
1383 remove_speculative()->
1384 cast_to_ptr_type(ptr)->
1385 cast_to_exactness(false)->
1386 with_offset(offset);
1387 }
1388 }
1389 // Arrays of fixed size alias with arrays of unknown size.
1390 if (ta->size() != TypeInt::POS) {
1391 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1392 tj = ta = ta->
1393 remove_speculative()->
1394 cast_to_ptr_type(ptr)->
1395 with_ary(tary)->
1396 cast_to_exactness(false);
1397 }
1398 // Arrays of known objects become arrays of unknown objects.
1399 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1400 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1401 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1402 }
1403 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1404 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1405 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1406 }
1407 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1408 // cannot be distinguished by bytecode alone.
1409 if (ta->elem() == TypeInt::BOOL) {
1410 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1411 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1412 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1413 }
1414 // During the 2nd round of IterGVN, NotNull castings are removed.
1415 // Make sure the Bottom and NotNull variants alias the same.
1416 // Also, make sure exact and non-exact variants alias the same.
1417 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1418 tj = ta = ta->
1419 remove_speculative()->
1420 cast_to_ptr_type(TypePtr::BotPTR)->
1421 cast_to_exactness(false)->
1422 with_offset(offset);
1423 }
1424 }
1425
1426 // Oop pointers need some flattening
1427 const TypeInstPtr *to = tj->isa_instptr();
1428 if (to && to != TypeOopPtr::BOTTOM) {
1429 ciInstanceKlass* ik = to->instance_klass();
1430 if( ptr == TypePtr::Constant ) {
1431 if (ik != ciEnv::current()->Class_klass() ||
1432 offset < ik->layout_helper_size_in_bytes()) {
1433 // No constant oop pointers (such as Strings); they alias with
1434 // unknown strings.
1435 assert(!is_known_inst, "not scalarizable allocation");
1436 tj = to = to->
1437 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1438 remove_speculative()->
1439 cast_to_ptr_type(TypePtr::BotPTR)->
1440 cast_to_exactness(false);
1441 }
1442 } else if( is_known_inst ) {
1443 tj = to; // Keep NotNull and klass_is_exact for instance type
1444 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1445 // During the 2nd round of IterGVN, NotNull castings are removed.
1446 // Make sure the Bottom and NotNull variants alias the same.
1447 // Also, make sure exact and non-exact variants alias the same.
1448 tj = to = to->
1449 remove_speculative()->
1450 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1451 cast_to_ptr_type(TypePtr::BotPTR)->
1452 cast_to_exactness(false);
1453 }
1454 if (to->speculative() != nullptr) {
1455 tj = to = to->remove_speculative();
1456 }
1457 // Canonicalize the holder of this field
1458 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1459 // First handle header references such as a LoadKlassNode, even if the
1460 // object's klass is unloaded at compile time (4965979).
1461 if (!is_known_inst) { // Do it only for non-instance types
1462 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1463 }
1464 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1465 // Static fields are in the space above the normal instance
1466 // fields in the java.lang.Class instance.
1467 if (ik != ciEnv::current()->Class_klass()) {
1468 to = nullptr;
1469 tj = TypeOopPtr::BOTTOM;
1470 offset = tj->offset();
1471 }
1472 } else {
1473 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1474 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1475 assert(tj->offset() == offset, "no change to offset expected");
1476 bool xk = to->klass_is_exact();
1477 int instance_id = to->instance_id();
1478
1479 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1480 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1481 // its interfaces are included.
1482 if (xk && ik->equals(canonical_holder)) {
1483 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1484 } else {
1485 assert(xk || !is_known_inst, "Known instance should be exact type");
1486 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1487 }
1488 }
1489 }
1490
1491 // Klass pointers to object array klasses need some flattening
1492 const TypeKlassPtr *tk = tj->isa_klassptr();
1493 if( tk ) {
1494 // If we are referencing a field within a Klass, we need
1495 // to assume the worst case of an Object. Both exact and
1496 // inexact types must flatten to the same alias class so
1497 // use NotNull as the PTR.
1498 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1499 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1500 env()->Object_klass(),
1501 offset);
1502 }
1503
1504 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1505 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1506 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1507 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1508 } else {
1509 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1510 }
1511 }
1512
1513 // Check for precise loads from the primary supertype array and force them
1514 // to the supertype cache alias index. Check for generic array loads from
1515 // the primary supertype array and also force them to the supertype cache
1516 // alias index. Since the same load can reach both, we need to merge
1517 // these 2 disparate memories into the same alias class. Since the
1518 // primary supertype array is read-only, there's no chance of confusion
1519 // where we bypass an array load and an array store.
1520 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1521 if (offset == Type::OffsetBot ||
1522 (offset >= primary_supers_offset &&
1523 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1524 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1525 offset = in_bytes(Klass::secondary_super_cache_offset());
1526 tj = tk = tk->with_offset(offset);
1527 }
1528 }
1529
1530 // Flatten all Raw pointers together.
1531 if (tj->base() == Type::RawPtr)
1532 tj = TypeRawPtr::BOTTOM;
1533
1534 if (tj->base() == Type::AnyPtr)
1535 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1536
1537 offset = tj->offset();
1538 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1539
1540 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1541 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1542 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1543 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1544 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1545 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1546 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1547 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1548 assert( tj->ptr() != TypePtr::TopPTR &&
1549 tj->ptr() != TypePtr::AnyNull &&
1550 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1551 // assert( tj->ptr() != TypePtr::Constant ||
1552 // tj->base() == Type::RawPtr ||
1553 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1554
1555 return tj;
1556 }
1557
1558 void Compile::AliasType::Init(int i, const TypePtr* at) {
1559 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1560 _index = i;
1561 _adr_type = at;
1562 _field = nullptr;
1563 _element = nullptr;
1564 _is_rewritable = true; // default
1565 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1566 if (atoop != nullptr && atoop->is_known_instance()) {
1567 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1568 _general_index = Compile::current()->get_alias_index(gt);
1569 } else {
1570 _general_index = 0;
1571 }
1572 }
1573
1574 BasicType Compile::AliasType::basic_type() const {
1575 if (element() != nullptr) {
1576 const Type* element = adr_type()->is_aryptr()->elem();
1577 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1578 } if (field() != nullptr) {
1579 return field()->layout_type();
1580 } else {
1581 return T_ILLEGAL; // unknown
1582 }
1583 }
1584
1585 //---------------------------------print_on------------------------------------
1586 #ifndef PRODUCT
1587 void Compile::AliasType::print_on(outputStream* st) {
1588 if (index() < 10)
1589 st->print("@ <%d> ", index());
1590 else st->print("@ <%d>", index());
1591 st->print(is_rewritable() ? " " : " RO");
1592 int offset = adr_type()->offset();
1593 if (offset == Type::OffsetBot)
1594 st->print(" +any");
1595 else st->print(" +%-3d", offset);
1596 st->print(" in ");
1597 adr_type()->dump_on(st);
1598 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1599 if (field() != nullptr && tjp) {
1600 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1601 tjp->offset() != field()->offset_in_bytes()) {
1602 st->print(" != ");
1603 field()->print();
1604 st->print(" ***");
1605 }
1606 }
1607 }
1608
1609 void print_alias_types() {
1610 Compile* C = Compile::current();
1611 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1612 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1613 C->alias_type(idx)->print_on(tty);
1614 tty->cr();
1615 }
1616 }
1617 #endif
1618
1619
1620 //----------------------------probe_alias_cache--------------------------------
1621 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1622 intptr_t key = (intptr_t) adr_type;
1623 key ^= key >> logAliasCacheSize;
1624 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1625 }
1626
1627
1628 //-----------------------------grow_alias_types--------------------------------
1629 void Compile::grow_alias_types() {
1630 const int old_ats = _max_alias_types; // how many before?
1631 const int new_ats = old_ats; // how many more?
1632 const int grow_ats = old_ats+new_ats; // how many now?
1633 _max_alias_types = grow_ats;
1634 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1635 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1636 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1637 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1638 }
1639
1640
1641 //--------------------------------find_alias_type------------------------------
1642 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1643 if (!do_aliasing()) {
1644 return alias_type(AliasIdxBot);
1645 }
1646
1647 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1648 if (ace->_adr_type == adr_type) {
1649 return alias_type(ace->_index);
1650 }
1651
1652 // Handle special cases.
1653 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1654 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1655
1656 // Do it the slow way.
1657 const TypePtr* flat = flatten_alias_type(adr_type);
1658
1659 #ifdef ASSERT
1660 {
1661 ResourceMark rm;
1662 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1663 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1664 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1665 Type::str(adr_type));
1666 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1667 const TypeOopPtr* foop = flat->is_oopptr();
1668 // Scalarizable allocations have exact klass always.
1669 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1670 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1671 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1672 Type::str(foop), Type::str(xoop));
1673 }
1674 }
1675 #endif
1676
1677 int idx = AliasIdxTop;
1678 for (int i = 0; i < num_alias_types(); i++) {
1679 if (alias_type(i)->adr_type() == flat) {
1680 idx = i;
1681 break;
1682 }
1683 }
1684
1685 if (idx == AliasIdxTop) {
1686 if (no_create) return nullptr;
1687 // Grow the array if necessary.
1688 if (_num_alias_types == _max_alias_types) grow_alias_types();
1689 // Add a new alias type.
1690 idx = _num_alias_types++;
1691 _alias_types[idx]->Init(idx, flat);
1692 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1693 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1694 if (flat->isa_instptr()) {
1695 if (flat->offset() == java_lang_Class::klass_offset()
1696 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1697 alias_type(idx)->set_rewritable(false);
1698 }
1699 if (flat->isa_aryptr()) {
1700 #ifdef ASSERT
1701 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1702 // (T_BYTE has the weakest alignment and size restrictions...)
1703 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1704 #endif
1705 if (flat->offset() == TypePtr::OffsetBot) {
1706 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1707 }
1708 }
1709 if (flat->isa_klassptr()) {
1710 if (UseCompactObjectHeaders) {
1711 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1712 alias_type(idx)->set_rewritable(false);
1713 }
1714 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1715 alias_type(idx)->set_rewritable(false);
1716 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1717 alias_type(idx)->set_rewritable(false);
1718 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1719 alias_type(idx)->set_rewritable(false);
1720 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1721 alias_type(idx)->set_rewritable(false);
1722 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1723 alias_type(idx)->set_rewritable(false);
1724 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1725 alias_type(idx)->set_rewritable(false);
1726 }
1727 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1728 // but the base pointer type is not distinctive enough to identify
1729 // references into JavaThread.)
1730
1731 // Check for final fields.
1732 const TypeInstPtr* tinst = flat->isa_instptr();
1733 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1734 ciField* field;
1735 if (tinst->const_oop() != nullptr &&
1736 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1737 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1738 // static field
1739 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1740 field = k->get_field_by_offset(tinst->offset(), true);
1741 } else {
1742 ciInstanceKlass *k = tinst->instance_klass();
1743 field = k->get_field_by_offset(tinst->offset(), false);
1744 }
1745 assert(field == nullptr ||
1746 original_field == nullptr ||
1747 (field->holder() == original_field->holder() &&
1748 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1749 field->is_static() == original_field->is_static()), "wrong field?");
1750 // Set field() and is_rewritable() attributes.
1751 if (field != nullptr) alias_type(idx)->set_field(field);
1752 }
1753 }
1754
1755 // Fill the cache for next time.
1756 ace->_adr_type = adr_type;
1757 ace->_index = idx;
1758 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1759
1760 // Might as well try to fill the cache for the flattened version, too.
1761 AliasCacheEntry* face = probe_alias_cache(flat);
1762 if (face->_adr_type == nullptr) {
1763 face->_adr_type = flat;
1764 face->_index = idx;
1765 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1766 }
1767
1768 return alias_type(idx);
1769 }
1770
1771
1772 Compile::AliasType* Compile::alias_type(ciField* field) {
1773 const TypeOopPtr* t;
1774 if (field->is_static())
1775 t = TypeInstPtr::make(field->holder()->java_mirror());
1776 else
1777 t = TypeOopPtr::make_from_klass_raw(field->holder());
1778 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1779 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1780 return atp;
1781 }
1782
1783
1784 //------------------------------have_alias_type--------------------------------
1785 bool Compile::have_alias_type(const TypePtr* adr_type) {
1786 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1787 if (ace->_adr_type == adr_type) {
1788 return true;
1789 }
1790
1791 // Handle special cases.
1792 if (adr_type == nullptr) return true;
1793 if (adr_type == TypePtr::BOTTOM) return true;
1794
1795 return find_alias_type(adr_type, true, nullptr) != nullptr;
1796 }
1797
1798 //-----------------------------must_alias--------------------------------------
1799 // True if all values of the given address type are in the given alias category.
1800 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1801 if (alias_idx == AliasIdxBot) return true; // the universal category
1802 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1803 if (alias_idx == AliasIdxTop) return false; // the empty category
1804 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1805
1806 // the only remaining possible overlap is identity
1807 int adr_idx = get_alias_index(adr_type);
1808 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1809 assert(adr_idx == alias_idx ||
1810 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1811 && adr_type != TypeOopPtr::BOTTOM),
1812 "should not be testing for overlap with an unsafe pointer");
1813 return adr_idx == alias_idx;
1814 }
1815
1816 //------------------------------can_alias--------------------------------------
1817 // True if any values of the given address type are in the given alias category.
1818 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1819 if (alias_idx == AliasIdxTop) return false; // the empty category
1820 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1821 // Known instance doesn't alias with bottom memory
1822 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1823 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1824
1825 // the only remaining possible overlap is identity
1826 int adr_idx = get_alias_index(adr_type);
1827 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1828 return adr_idx == alias_idx;
1829 }
1830
1831 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1832 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1833 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1834 if (parse_predicate_count() == 0) {
1835 return;
1836 }
1837 for (int i = 0; i < parse_predicate_count(); i++) {
1838 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1839 parse_predicate->mark_useless();
1840 igvn._worklist.push(parse_predicate);
1841 }
1842 _parse_predicates.clear();
1843 }
1844
1845 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1846 if (!n->for_post_loop_opts_igvn()) {
1847 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1848 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1849 _for_post_loop_igvn.append(n);
1850 }
1851 }
1852
1853 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1854 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1855 _for_post_loop_igvn.remove(n);
1856 }
1857
1858 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1859 // Verify that all previous optimizations produced a valid graph
1860 // at least to this point, even if no loop optimizations were done.
1861 PhaseIdealLoop::verify(igvn);
1862
1863 C->set_post_loop_opts_phase(); // no more loop opts allowed
1864
1865 assert(!C->major_progress(), "not cleared");
1866
1867 if (_for_post_loop_igvn.length() > 0) {
1868 while (_for_post_loop_igvn.length() > 0) {
1869 Node* n = _for_post_loop_igvn.pop();
1870 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1871 igvn._worklist.push(n);
1872 }
1873 igvn.optimize();
1874 if (failing()) return;
1875 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1876 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1877
1878 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1879 if (C->major_progress()) {
1880 C->clear_major_progress(); // ensure that major progress is now clear
1881 }
1882 }
1883 }
1884
1885 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1886 if (OptimizeUnstableIf) {
1887 _unstable_if_traps.append(trap);
1888 }
1889 }
1890
1891 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1892 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1893 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1894 Node* n = trap->uncommon_trap();
1895 if (!useful.member(n)) {
1896 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1897 }
1898 }
1899 }
1900
1901 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1902 // or fold-compares case. Return true if succeed or not found.
1903 //
1904 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1905 // false and ask fold-compares to yield.
1906 //
1907 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1908 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1909 // when deoptimization does happen.
1910 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1911 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1912 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1913 if (trap->uncommon_trap() == unc) {
1914 if (yield && trap->modified()) {
1915 return false;
1916 }
1917 _unstable_if_traps.delete_at(i);
1918 break;
1919 }
1920 }
1921 return true;
1922 }
1923
1924 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1925 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1926 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1927 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1928 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1929 CallStaticJavaNode* unc = trap->uncommon_trap();
1930 int next_bci = trap->next_bci();
1931 bool modified = trap->modified();
1932
1933 if (next_bci != -1 && !modified) {
1934 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1935 JVMState* jvms = unc->jvms();
1936 ciMethod* method = jvms->method();
1937 ciBytecodeStream iter(method);
1938
1939 iter.force_bci(jvms->bci());
1940 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1941 Bytecodes::Code c = iter.cur_bc();
1942 Node* lhs = nullptr;
1943 Node* rhs = nullptr;
1944 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1945 lhs = unc->peek_operand(0);
1946 rhs = unc->peek_operand(1);
1947 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1948 lhs = unc->peek_operand(0);
1949 }
1950
1951 ResourceMark rm;
1952 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1953 assert(live_locals.is_valid(), "broken liveness info");
1954 int len = (int)live_locals.size();
1955
1956 for (int i = 0; i < len; i++) {
1957 Node* local = unc->local(jvms, i);
1958 // kill local using the liveness of next_bci.
1959 // give up when the local looks like an operand to secure reexecution.
1960 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1961 uint idx = jvms->locoff() + i;
1962 #ifdef ASSERT
1963 if (PrintOpto && Verbose) {
1964 tty->print("[unstable_if] kill local#%d: ", idx);
1965 local->dump();
1966 tty->cr();
1967 }
1968 #endif
1969 igvn.replace_input_of(unc, idx, top());
1970 modified = true;
1971 }
1972 }
1973 }
1974
1975 // keep the mondified trap for late query
1976 if (modified) {
1977 trap->set_modified();
1978 } else {
1979 _unstable_if_traps.delete_at(i);
1980 }
1981 }
1982 igvn.optimize();
1983 }
1984
1985 // StringOpts and late inlining of string methods
1986 void Compile::inline_string_calls(bool parse_time) {
1987 {
1988 // remove useless nodes to make the usage analysis simpler
1989 ResourceMark rm;
1990 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1991 }
1992
1993 {
1994 ResourceMark rm;
1995 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1996 PhaseStringOpts pso(initial_gvn());
1997 print_method(PHASE_AFTER_STRINGOPTS, 3);
1998 }
1999
2000 // now inline anything that we skipped the first time around
2001 if (!parse_time) {
2002 _late_inlines_pos = _late_inlines.length();
2003 }
2004
2005 while (_string_late_inlines.length() > 0) {
2006 CallGenerator* cg = _string_late_inlines.pop();
2007 cg->do_late_inline();
2008 if (failing()) return;
2009 }
2010 _string_late_inlines.trunc_to(0);
2011 }
2012
2013 // Late inlining of boxing methods
2014 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2015 if (_boxing_late_inlines.length() > 0) {
2016 assert(has_boxed_value(), "inconsistent");
2017
2018 set_inlining_incrementally(true);
2019
2020 igvn_worklist()->ensure_empty(); // should be done with igvn
2021
2022 _late_inlines_pos = _late_inlines.length();
2023
2024 while (_boxing_late_inlines.length() > 0) {
2025 CallGenerator* cg = _boxing_late_inlines.pop();
2026 cg->do_late_inline();
2027 if (failing()) return;
2028 }
2029 _boxing_late_inlines.trunc_to(0);
2030
2031 inline_incrementally_cleanup(igvn);
2032
2033 set_inlining_incrementally(false);
2034 }
2035 }
2036
2037 bool Compile::inline_incrementally_one() {
2038 assert(IncrementalInline, "incremental inlining should be on");
2039
2040 TracePhase tp(_t_incrInline_inline);
2041
2042 set_inlining_progress(false);
2043 set_do_cleanup(false);
2044
2045 for (int i = 0; i < _late_inlines.length(); i++) {
2046 _late_inlines_pos = i+1;
2047 CallGenerator* cg = _late_inlines.at(i);
2048 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2049 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2050 cg->do_late_inline();
2051 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2052 if (failing()) {
2053 return false;
2054 } else if (inlining_progress()) {
2055 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2056 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2057 break; // process one call site at a time
2058 }
2059 } else {
2060 // Ignore late inline direct calls when inlining is not allowed.
2061 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2062 }
2063 }
2064 // Remove processed elements.
2065 _late_inlines.remove_till(_late_inlines_pos);
2066 _late_inlines_pos = 0;
2067
2068 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2069
2070 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2071
2072 set_inlining_progress(false);
2073 set_do_cleanup(false);
2074
2075 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2076 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2077 }
2078
2079 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2080 {
2081 TracePhase tp(_t_incrInline_pru);
2082 ResourceMark rm;
2083 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2084 }
2085 {
2086 TracePhase tp(_t_incrInline_igvn);
2087 igvn.reset_from_gvn(initial_gvn());
2088 igvn.optimize();
2089 if (failing()) return;
2090 }
2091 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2092 }
2093
2094 // Perform incremental inlining until bound on number of live nodes is reached
2095 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2096 TracePhase tp(_t_incrInline);
2097
2098 set_inlining_incrementally(true);
2099 uint low_live_nodes = 0;
2100
2101 while (_late_inlines.length() > 0) {
2102 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2103 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2104 TracePhase tp(_t_incrInline_ideal);
2105 // PhaseIdealLoop is expensive so we only try it once we are
2106 // out of live nodes and we only try it again if the previous
2107 // helped got the number of nodes down significantly
2108 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2109 if (failing()) return;
2110 low_live_nodes = live_nodes();
2111 _major_progress = true;
2112 }
2113
2114 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2115 bool do_print_inlining = print_inlining() || print_intrinsics();
2116 if (do_print_inlining || log() != nullptr) {
2117 // Print inlining message for candidates that we couldn't inline for lack of space.
2118 for (int i = 0; i < _late_inlines.length(); i++) {
2119 CallGenerator* cg = _late_inlines.at(i);
2120 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2121 if (do_print_inlining) {
2122 cg->print_inlining_late(InliningResult::FAILURE, msg);
2123 }
2124 log_late_inline_failure(cg, msg);
2125 }
2126 }
2127 break; // finish
2128 }
2129 }
2130
2131 igvn_worklist()->ensure_empty(); // should be done with igvn
2132
2133 while (inline_incrementally_one()) {
2134 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2135 }
2136 if (failing()) return;
2137
2138 inline_incrementally_cleanup(igvn);
2139
2140 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2141
2142 if (failing()) return;
2143
2144 if (_late_inlines.length() == 0) {
2145 break; // no more progress
2146 }
2147 }
2148
2149 igvn_worklist()->ensure_empty(); // should be done with igvn
2150
2151 if (_string_late_inlines.length() > 0) {
2152 assert(has_stringbuilder(), "inconsistent");
2153
2154 inline_string_calls(false);
2155
2156 if (failing()) return;
2157
2158 inline_incrementally_cleanup(igvn);
2159 }
2160
2161 set_inlining_incrementally(false);
2162 }
2163
2164 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2165 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2166 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2167 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2168 // as if "inlining_incrementally() == true" were set.
2169 assert(inlining_incrementally() == false, "not allowed");
2170 assert(_modified_nodes == nullptr, "not allowed");
2171 assert(_late_inlines.length() > 0, "sanity");
2172
2173 while (_late_inlines.length() > 0) {
2174 igvn_worklist()->ensure_empty(); // should be done with igvn
2175
2176 while (inline_incrementally_one()) {
2177 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2178 }
2179 if (failing()) return;
2180
2181 inline_incrementally_cleanup(igvn);
2182 }
2183 }
2184
2185 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2186 if (_loop_opts_cnt > 0) {
2187 while (major_progress() && (_loop_opts_cnt > 0)) {
2188 TracePhase tp(_t_idealLoop);
2189 PhaseIdealLoop::optimize(igvn, mode);
2190 _loop_opts_cnt--;
2191 if (failing()) return false;
2192 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2193 }
2194 }
2195 return true;
2196 }
2197
2198 // Remove edges from "root" to each SafePoint at a backward branch.
2199 // They were inserted during parsing (see add_safepoint()) to make
2200 // infinite loops without calls or exceptions visible to root, i.e.,
2201 // useful.
2202 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2203 Node *r = root();
2204 if (r != nullptr) {
2205 for (uint i = r->req(); i < r->len(); ++i) {
2206 Node *n = r->in(i);
2207 if (n != nullptr && n->is_SafePoint()) {
2208 r->rm_prec(i);
2209 if (n->outcnt() == 0) {
2210 igvn.remove_dead_node(n);
2211 }
2212 --i;
2213 }
2214 }
2215 // Parsing may have added top inputs to the root node (Path
2216 // leading to the Halt node proven dead). Make sure we get a
2217 // chance to clean them up.
2218 igvn._worklist.push(r);
2219 igvn.optimize();
2220 }
2221 }
2222
2223 //------------------------------Optimize---------------------------------------
2224 // Given a graph, optimize it.
2225 void Compile::Optimize() {
2226 TracePhase tp(_t_optimizer);
2227
2228 #ifndef PRODUCT
2229 if (env()->break_at_compile()) {
2230 BREAKPOINT;
2231 }
2232
2233 #endif
2234
2235 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2236 #ifdef ASSERT
2237 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2238 #endif
2239
2240 ResourceMark rm;
2241
2242 print_inlining_reinit();
2243
2244 NOT_PRODUCT( verify_graph_edges(); )
2245
2246 print_method(PHASE_AFTER_PARSING, 1);
2247
2248 {
2249 // Iterative Global Value Numbering, including ideal transforms
2250 // Initialize IterGVN with types and values from parse-time GVN
2251 PhaseIterGVN igvn(initial_gvn());
2252 #ifdef ASSERT
2253 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2254 #endif
2255 {
2256 TracePhase tp(_t_iterGVN);
2257 igvn.optimize();
2258 }
2259
2260 if (failing()) return;
2261
2262 print_method(PHASE_ITER_GVN1, 2);
2263
2264 process_for_unstable_if_traps(igvn);
2265
2266 if (failing()) return;
2267
2268 inline_incrementally(igvn);
2269
2270 print_method(PHASE_INCREMENTAL_INLINE, 2);
2271
2272 if (failing()) return;
2273
2274 if (eliminate_boxing()) {
2275 // Inline valueOf() methods now.
2276 inline_boxing_calls(igvn);
2277
2278 if (failing()) return;
2279
2280 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2281 inline_incrementally(igvn);
2282 }
2283
2284 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2285
2286 if (failing()) return;
2287 }
2288
2289 // Remove the speculative part of types and clean up the graph from
2290 // the extra CastPP nodes whose only purpose is to carry them. Do
2291 // that early so that optimizations are not disrupted by the extra
2292 // CastPP nodes.
2293 remove_speculative_types(igvn);
2294
2295 if (failing()) return;
2296
2297 // No more new expensive nodes will be added to the list from here
2298 // so keep only the actual candidates for optimizations.
2299 cleanup_expensive_nodes(igvn);
2300
2301 if (failing()) return;
2302
2303 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2304 if (EnableVectorSupport && has_vbox_nodes()) {
2305 TracePhase tp(_t_vector);
2306 PhaseVector pv(igvn);
2307 pv.optimize_vector_boxes();
2308 if (failing()) return;
2309 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2310 }
2311 assert(!has_vbox_nodes(), "sanity");
2312
2313 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2314 Compile::TracePhase tp(_t_renumberLive);
2315 igvn_worklist()->ensure_empty(); // should be done with igvn
2316 {
2317 ResourceMark rm;
2318 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2319 }
2320 igvn.reset_from_gvn(initial_gvn());
2321 igvn.optimize();
2322 if (failing()) return;
2323 }
2324
2325 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2326 // safepoints
2327 remove_root_to_sfpts_edges(igvn);
2328
2329 if (failing()) return;
2330
2331 // Perform escape analysis
2332 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2333 if (has_loops()) {
2334 // Cleanup graph (remove dead nodes).
2335 TracePhase tp(_t_idealLoop);
2336 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2337 if (failing()) return;
2338 }
2339 bool progress;
2340 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2341 do {
2342 ConnectionGraph::do_analysis(this, &igvn);
2343
2344 if (failing()) return;
2345
2346 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2347
2348 // Optimize out fields loads from scalar replaceable allocations.
2349 igvn.optimize();
2350 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2351
2352 if (failing()) return;
2353
2354 if (congraph() != nullptr && macro_count() > 0) {
2355 TracePhase tp(_t_macroEliminate);
2356 PhaseMacroExpand mexp(igvn);
2357 mexp.eliminate_macro_nodes();
2358 if (failing()) return;
2359
2360 igvn.set_delay_transform(false);
2361 igvn.optimize();
2362 if (failing()) return;
2363
2364 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2365 }
2366
2367 ConnectionGraph::verify_ram_nodes(this, root());
2368 if (failing()) return;
2369
2370 progress = do_iterative_escape_analysis() &&
2371 (macro_count() < mcount) &&
2372 ConnectionGraph::has_candidates(this);
2373 // Try again if candidates exist and made progress
2374 // by removing some allocations and/or locks.
2375 } while (progress);
2376 }
2377
2378 // Loop transforms on the ideal graph. Range Check Elimination,
2379 // peeling, unrolling, etc.
2380
2381 // Set loop opts counter
2382 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2383 {
2384 TracePhase tp(_t_idealLoop);
2385 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2386 _loop_opts_cnt--;
2387 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2388 if (failing()) return;
2389 }
2390 // Loop opts pass if partial peeling occurred in previous pass
2391 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2392 TracePhase tp(_t_idealLoop);
2393 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2394 _loop_opts_cnt--;
2395 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2396 if (failing()) return;
2397 }
2398 // Loop opts pass for loop-unrolling before CCP
2399 if(major_progress() && (_loop_opts_cnt > 0)) {
2400 TracePhase tp(_t_idealLoop);
2401 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2402 _loop_opts_cnt--;
2403 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2404 }
2405 if (!failing()) {
2406 // Verify that last round of loop opts produced a valid graph
2407 PhaseIdealLoop::verify(igvn);
2408 }
2409 }
2410 if (failing()) return;
2411
2412 // Conditional Constant Propagation;
2413 print_method(PHASE_BEFORE_CCP1, 2);
2414 PhaseCCP ccp( &igvn );
2415 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2416 {
2417 TracePhase tp(_t_ccp);
2418 ccp.do_transform();
2419 }
2420 print_method(PHASE_CCP1, 2);
2421
2422 assert( true, "Break here to ccp.dump_old2new_map()");
2423
2424 // Iterative Global Value Numbering, including ideal transforms
2425 {
2426 TracePhase tp(_t_iterGVN2);
2427 igvn.reset_from_igvn(&ccp);
2428 igvn.optimize();
2429 }
2430 print_method(PHASE_ITER_GVN2, 2);
2431
2432 if (failing()) return;
2433
2434 // Loop transforms on the ideal graph. Range Check Elimination,
2435 // peeling, unrolling, etc.
2436 if (!optimize_loops(igvn, LoopOptsDefault)) {
2437 return;
2438 }
2439
2440 if (failing()) return;
2441
2442 C->clear_major_progress(); // ensure that major progress is now clear
2443
2444 process_for_post_loop_opts_igvn(igvn);
2445
2446 if (failing()) return;
2447
2448 #ifdef ASSERT
2449 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2450 #endif
2451
2452 {
2453 TracePhase tp(_t_macroExpand);
2454 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2455 PhaseMacroExpand mex(igvn);
2456 if (mex.expand_macro_nodes()) {
2457 assert(failing(), "must bail out w/ explicit message");
2458 return;
2459 }
2460 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2461 }
2462
2463 {
2464 TracePhase tp(_t_barrierExpand);
2465 if (bs->expand_barriers(this, igvn)) {
2466 assert(failing(), "must bail out w/ explicit message");
2467 return;
2468 }
2469 print_method(PHASE_BARRIER_EXPANSION, 2);
2470 }
2471
2472 if (C->max_vector_size() > 0) {
2473 C->optimize_logic_cones(igvn);
2474 igvn.optimize();
2475 if (failing()) return;
2476 }
2477
2478 DEBUG_ONLY( _modified_nodes = nullptr; )
2479
2480 assert(igvn._worklist.size() == 0, "not empty");
2481
2482 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2483
2484 if (_late_inlines.length() > 0) {
2485 // More opportunities to optimize virtual and MH calls.
2486 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2487 process_late_inline_calls_no_inline(igvn);
2488 if (failing()) return;
2489 }
2490 } // (End scope of igvn; run destructor if necessary for asserts.)
2491
2492 check_no_dead_use();
2493
2494 process_print_inlining();
2495
2496 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2497 // to remove hashes to unlock nodes for modifications.
2498 C->node_hash()->clear();
2499
2500 // A method with only infinite loops has no edges entering loops from root
2501 {
2502 TracePhase tp(_t_graphReshaping);
2503 if (final_graph_reshaping()) {
2504 assert(failing(), "must bail out w/ explicit message");
2505 return;
2506 }
2507 }
2508
2509 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2510 DEBUG_ONLY(set_phase_optimize_finished();)
2511 }
2512
2513 #ifdef ASSERT
2514 void Compile::check_no_dead_use() const {
2515 ResourceMark rm;
2516 Unique_Node_List wq;
2517 wq.push(root());
2518 for (uint i = 0; i < wq.size(); ++i) {
2519 Node* n = wq.at(i);
2520 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2521 Node* u = n->fast_out(j);
2522 if (u->outcnt() == 0 && !u->is_Con()) {
2523 u->dump();
2524 fatal("no reachable node should have no use");
2525 }
2526 wq.push(u);
2527 }
2528 }
2529 }
2530 #endif
2531
2532 void Compile::inline_vector_reboxing_calls() {
2533 if (C->_vector_reboxing_late_inlines.length() > 0) {
2534 _late_inlines_pos = C->_late_inlines.length();
2535 while (_vector_reboxing_late_inlines.length() > 0) {
2536 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2537 cg->do_late_inline();
2538 if (failing()) return;
2539 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2540 }
2541 _vector_reboxing_late_inlines.trunc_to(0);
2542 }
2543 }
2544
2545 bool Compile::has_vbox_nodes() {
2546 if (C->_vector_reboxing_late_inlines.length() > 0) {
2547 return true;
2548 }
2549 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2550 Node * n = C->macro_node(macro_idx);
2551 assert(n->is_macro(), "only macro nodes expected here");
2552 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2553 return true;
2554 }
2555 }
2556 return false;
2557 }
2558
2559 //---------------------------- Bitwise operation packing optimization ---------------------------
2560
2561 static bool is_vector_unary_bitwise_op(Node* n) {
2562 return n->Opcode() == Op_XorV &&
2563 VectorNode::is_vector_bitwise_not_pattern(n);
2564 }
2565
2566 static bool is_vector_binary_bitwise_op(Node* n) {
2567 switch (n->Opcode()) {
2568 case Op_AndV:
2569 case Op_OrV:
2570 return true;
2571
2572 case Op_XorV:
2573 return !is_vector_unary_bitwise_op(n);
2574
2575 default:
2576 return false;
2577 }
2578 }
2579
2580 static bool is_vector_ternary_bitwise_op(Node* n) {
2581 return n->Opcode() == Op_MacroLogicV;
2582 }
2583
2584 static bool is_vector_bitwise_op(Node* n) {
2585 return is_vector_unary_bitwise_op(n) ||
2586 is_vector_binary_bitwise_op(n) ||
2587 is_vector_ternary_bitwise_op(n);
2588 }
2589
2590 static bool is_vector_bitwise_cone_root(Node* n) {
2591 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2592 return false;
2593 }
2594 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2595 if (is_vector_bitwise_op(n->fast_out(i))) {
2596 return false;
2597 }
2598 }
2599 return true;
2600 }
2601
2602 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2603 uint cnt = 0;
2604 if (is_vector_bitwise_op(n)) {
2605 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2606 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2607 for (uint i = 1; i < inp_cnt; i++) {
2608 Node* in = n->in(i);
2609 bool skip = VectorNode::is_all_ones_vector(in);
2610 if (!skip && !inputs.member(in)) {
2611 inputs.push(in);
2612 cnt++;
2613 }
2614 }
2615 assert(cnt <= 1, "not unary");
2616 } else {
2617 uint last_req = inp_cnt;
2618 if (is_vector_ternary_bitwise_op(n)) {
2619 last_req = inp_cnt - 1; // skip last input
2620 }
2621 for (uint i = 1; i < last_req; i++) {
2622 Node* def = n->in(i);
2623 if (!inputs.member(def)) {
2624 inputs.push(def);
2625 cnt++;
2626 }
2627 }
2628 }
2629 } else { // not a bitwise operations
2630 if (!inputs.member(n)) {
2631 inputs.push(n);
2632 cnt++;
2633 }
2634 }
2635 return cnt;
2636 }
2637
2638 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2639 Unique_Node_List useful_nodes;
2640 C->identify_useful_nodes(useful_nodes);
2641
2642 for (uint i = 0; i < useful_nodes.size(); i++) {
2643 Node* n = useful_nodes.at(i);
2644 if (is_vector_bitwise_cone_root(n)) {
2645 list.push(n);
2646 }
2647 }
2648 }
2649
2650 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2651 const TypeVect* vt,
2652 Unique_Node_List& partition,
2653 Unique_Node_List& inputs) {
2654 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2655 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2656 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2657
2658 Node* in1 = inputs.at(0);
2659 Node* in2 = inputs.at(1);
2660 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2661
2662 uint func = compute_truth_table(partition, inputs);
2663
2664 Node* pn = partition.at(partition.size() - 1);
2665 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2666 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2667 }
2668
2669 static uint extract_bit(uint func, uint pos) {
2670 return (func & (1 << pos)) >> pos;
2671 }
2672
2673 //
2674 // A macro logic node represents a truth table. It has 4 inputs,
2675 // First three inputs corresponds to 3 columns of a truth table
2676 // and fourth input captures the logic function.
2677 //
2678 // eg. fn = (in1 AND in2) OR in3;
2679 //
2680 // MacroNode(in1,in2,in3,fn)
2681 //
2682 // -----------------
2683 // in1 in2 in3 fn
2684 // -----------------
2685 // 0 0 0 0
2686 // 0 0 1 1
2687 // 0 1 0 0
2688 // 0 1 1 1
2689 // 1 0 0 0
2690 // 1 0 1 1
2691 // 1 1 0 1
2692 // 1 1 1 1
2693 //
2694
2695 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2696 int res = 0;
2697 for (int i = 0; i < 8; i++) {
2698 int bit1 = extract_bit(in1, i);
2699 int bit2 = extract_bit(in2, i);
2700 int bit3 = extract_bit(in3, i);
2701
2702 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2703 int func_bit = extract_bit(func, func_bit_pos);
2704
2705 res |= func_bit << i;
2706 }
2707 return res;
2708 }
2709
2710 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2711 assert(n != nullptr, "");
2712 assert(eval_map.contains(n), "absent");
2713 return *(eval_map.get(n));
2714 }
2715
2716 static void eval_operands(Node* n,
2717 uint& func1, uint& func2, uint& func3,
2718 ResourceHashtable<Node*,uint>& eval_map) {
2719 assert(is_vector_bitwise_op(n), "");
2720
2721 if (is_vector_unary_bitwise_op(n)) {
2722 Node* opnd = n->in(1);
2723 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2724 opnd = n->in(2);
2725 }
2726 func1 = eval_operand(opnd, eval_map);
2727 } else if (is_vector_binary_bitwise_op(n)) {
2728 func1 = eval_operand(n->in(1), eval_map);
2729 func2 = eval_operand(n->in(2), eval_map);
2730 } else {
2731 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2732 func1 = eval_operand(n->in(1), eval_map);
2733 func2 = eval_operand(n->in(2), eval_map);
2734 func3 = eval_operand(n->in(3), eval_map);
2735 }
2736 }
2737
2738 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2739 assert(inputs.size() <= 3, "sanity");
2740 ResourceMark rm;
2741 uint res = 0;
2742 ResourceHashtable<Node*,uint> eval_map;
2743
2744 // Populate precomputed functions for inputs.
2745 // Each input corresponds to one column of 3 input truth-table.
2746 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2747 0xCC, // (_, b, _) -> b
2748 0xF0 }; // (a, _, _) -> a
2749 for (uint i = 0; i < inputs.size(); i++) {
2750 eval_map.put(inputs.at(i), input_funcs[2-i]);
2751 }
2752
2753 for (uint i = 0; i < partition.size(); i++) {
2754 Node* n = partition.at(i);
2755
2756 uint func1 = 0, func2 = 0, func3 = 0;
2757 eval_operands(n, func1, func2, func3, eval_map);
2758
2759 switch (n->Opcode()) {
2760 case Op_OrV:
2761 assert(func3 == 0, "not binary");
2762 res = func1 | func2;
2763 break;
2764 case Op_AndV:
2765 assert(func3 == 0, "not binary");
2766 res = func1 & func2;
2767 break;
2768 case Op_XorV:
2769 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2770 assert(func2 == 0 && func3 == 0, "not unary");
2771 res = (~func1) & 0xFF;
2772 } else {
2773 assert(func3 == 0, "not binary");
2774 res = func1 ^ func2;
2775 }
2776 break;
2777 case Op_MacroLogicV:
2778 // Ordering of inputs may change during evaluation of sub-tree
2779 // containing MacroLogic node as a child node, thus a re-evaluation
2780 // makes sure that function is evaluated in context of current
2781 // inputs.
2782 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2783 break;
2784
2785 default: assert(false, "not supported: %s", n->Name());
2786 }
2787 assert(res <= 0xFF, "invalid");
2788 eval_map.put(n, res);
2789 }
2790 return res;
2791 }
2792
2793 // Criteria under which nodes gets packed into a macro logic node:-
2794 // 1) Parent and both child nodes are all unmasked or masked with
2795 // same predicates.
2796 // 2) Masked parent can be packed with left child if it is predicated
2797 // and both have same predicates.
2798 // 3) Masked parent can be packed with right child if its un-predicated
2799 // or has matching predication condition.
2800 // 4) An unmasked parent can be packed with an unmasked child.
2801 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2802 assert(partition.size() == 0, "not empty");
2803 assert(inputs.size() == 0, "not empty");
2804 if (is_vector_ternary_bitwise_op(n)) {
2805 return false;
2806 }
2807
2808 bool is_unary_op = is_vector_unary_bitwise_op(n);
2809 if (is_unary_op) {
2810 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2811 return false; // too few inputs
2812 }
2813
2814 bool pack_left_child = true;
2815 bool pack_right_child = true;
2816
2817 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2818 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2819
2820 int left_child_input_cnt = 0;
2821 int right_child_input_cnt = 0;
2822
2823 bool parent_is_predicated = n->is_predicated_vector();
2824 bool left_child_predicated = n->in(1)->is_predicated_vector();
2825 bool right_child_predicated = n->in(2)->is_predicated_vector();
2826
2827 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2828 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2829 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2830
2831 do {
2832 if (pack_left_child && left_child_LOP &&
2833 ((!parent_is_predicated && !left_child_predicated) ||
2834 ((parent_is_predicated && left_child_predicated &&
2835 parent_pred == left_child_pred)))) {
2836 partition.push(n->in(1));
2837 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2838 } else {
2839 inputs.push(n->in(1));
2840 left_child_input_cnt = 1;
2841 }
2842
2843 if (pack_right_child && right_child_LOP &&
2844 (!right_child_predicated ||
2845 (right_child_predicated && parent_is_predicated &&
2846 parent_pred == right_child_pred))) {
2847 partition.push(n->in(2));
2848 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2849 } else {
2850 inputs.push(n->in(2));
2851 right_child_input_cnt = 1;
2852 }
2853
2854 if (inputs.size() > 3) {
2855 assert(partition.size() > 0, "");
2856 inputs.clear();
2857 partition.clear();
2858 if (left_child_input_cnt > right_child_input_cnt) {
2859 pack_left_child = false;
2860 } else {
2861 pack_right_child = false;
2862 }
2863 } else {
2864 break;
2865 }
2866 } while(true);
2867
2868 if(partition.size()) {
2869 partition.push(n);
2870 }
2871
2872 return (partition.size() == 2 || partition.size() == 3) &&
2873 (inputs.size() == 2 || inputs.size() == 3);
2874 }
2875
2876 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2877 assert(is_vector_bitwise_op(n), "not a root");
2878
2879 visited.set(n->_idx);
2880
2881 // 1) Do a DFS walk over the logic cone.
2882 for (uint i = 1; i < n->req(); i++) {
2883 Node* in = n->in(i);
2884 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2885 process_logic_cone_root(igvn, in, visited);
2886 }
2887 }
2888
2889 // 2) Bottom up traversal: Merge node[s] with
2890 // the parent to form macro logic node.
2891 Unique_Node_List partition;
2892 Unique_Node_List inputs;
2893 if (compute_logic_cone(n, partition, inputs)) {
2894 const TypeVect* vt = n->bottom_type()->is_vect();
2895 Node* pn = partition.at(partition.size() - 1);
2896 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2897 if (mask == nullptr ||
2898 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2899 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2900 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2901 igvn.replace_node(n, macro_logic);
2902 }
2903 }
2904 }
2905
2906 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2907 ResourceMark rm;
2908 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2909 Unique_Node_List list;
2910 collect_logic_cone_roots(list);
2911
2912 while (list.size() > 0) {
2913 Node* n = list.pop();
2914 const TypeVect* vt = n->bottom_type()->is_vect();
2915 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2916 if (supported) {
2917 VectorSet visited(comp_arena());
2918 process_logic_cone_root(igvn, n, visited);
2919 }
2920 }
2921 }
2922 }
2923
2924 //------------------------------Code_Gen---------------------------------------
2925 // Given a graph, generate code for it
2926 void Compile::Code_Gen() {
2927 if (failing()) {
2928 return;
2929 }
2930
2931 // Perform instruction selection. You might think we could reclaim Matcher
2932 // memory PDQ, but actually the Matcher is used in generating spill code.
2933 // Internals of the Matcher (including some VectorSets) must remain live
2934 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2935 // set a bit in reclaimed memory.
2936
2937 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2938 // nodes. Mapping is only valid at the root of each matched subtree.
2939 NOT_PRODUCT( verify_graph_edges(); )
2940
2941 Matcher matcher;
2942 _matcher = &matcher;
2943 {
2944 TracePhase tp(_t_matcher);
2945 matcher.match();
2946 if (failing()) {
2947 return;
2948 }
2949 }
2950 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2951 // nodes. Mapping is only valid at the root of each matched subtree.
2952 NOT_PRODUCT( verify_graph_edges(); )
2953
2954 // If you have too many nodes, or if matching has failed, bail out
2955 check_node_count(0, "out of nodes matching instructions");
2956 if (failing()) {
2957 return;
2958 }
2959
2960 print_method(PHASE_MATCHING, 2);
2961
2962 // Build a proper-looking CFG
2963 PhaseCFG cfg(node_arena(), root(), matcher);
2964 if (failing()) {
2965 return;
2966 }
2967 _cfg = &cfg;
2968 {
2969 TracePhase tp(_t_scheduler);
2970 bool success = cfg.do_global_code_motion();
2971 if (!success) {
2972 return;
2973 }
2974
2975 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2976 NOT_PRODUCT( verify_graph_edges(); )
2977 cfg.verify();
2978 }
2979
2980 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2981 _regalloc = ®alloc;
2982 {
2983 TracePhase tp(_t_registerAllocation);
2984 // Perform register allocation. After Chaitin, use-def chains are
2985 // no longer accurate (at spill code) and so must be ignored.
2986 // Node->LRG->reg mappings are still accurate.
2987 _regalloc->Register_Allocate();
2988
2989 // Bail out if the allocator builds too many nodes
2990 if (failing()) {
2991 return;
2992 }
2993
2994 print_method(PHASE_REGISTER_ALLOCATION, 2);
2995 }
2996
2997 // Prior to register allocation we kept empty basic blocks in case the
2998 // the allocator needed a place to spill. After register allocation we
2999 // are not adding any new instructions. If any basic block is empty, we
3000 // can now safely remove it.
3001 {
3002 TracePhase tp(_t_blockOrdering);
3003 cfg.remove_empty_blocks();
3004 if (do_freq_based_layout()) {
3005 PhaseBlockLayout layout(cfg);
3006 } else {
3007 cfg.set_loop_alignment();
3008 }
3009 cfg.fixup_flow();
3010 cfg.remove_unreachable_blocks();
3011 cfg.verify_dominator_tree();
3012 print_method(PHASE_BLOCK_ORDERING, 3);
3013 }
3014
3015 // Apply peephole optimizations
3016 if( OptoPeephole ) {
3017 TracePhase tp(_t_peephole);
3018 PhasePeephole peep( _regalloc, cfg);
3019 peep.do_transform();
3020 print_method(PHASE_PEEPHOLE, 3);
3021 }
3022
3023 // Do late expand if CPU requires this.
3024 if (Matcher::require_postalloc_expand) {
3025 TracePhase tp(_t_postalloc_expand);
3026 cfg.postalloc_expand(_regalloc);
3027 print_method(PHASE_POSTALLOC_EXPAND, 3);
3028 }
3029
3030 // Convert Nodes to instruction bits in a buffer
3031 {
3032 TracePhase tp(_t_output);
3033 PhaseOutput output;
3034 output.Output();
3035 if (failing()) return;
3036 output.install();
3037 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3038 }
3039
3040 // He's dead, Jim.
3041 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3042 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3043 }
3044
3045 //------------------------------Final_Reshape_Counts---------------------------
3046 // This class defines counters to help identify when a method
3047 // may/must be executed using hardware with only 24-bit precision.
3048 struct Final_Reshape_Counts : public StackObj {
3049 int _call_count; // count non-inlined 'common' calls
3050 int _float_count; // count float ops requiring 24-bit precision
3051 int _double_count; // count double ops requiring more precision
3052 int _java_call_count; // count non-inlined 'java' calls
3053 int _inner_loop_count; // count loops which need alignment
3054 VectorSet _visited; // Visitation flags
3055 Node_List _tests; // Set of IfNodes & PCTableNodes
3056
3057 Final_Reshape_Counts() :
3058 _call_count(0), _float_count(0), _double_count(0),
3059 _java_call_count(0), _inner_loop_count(0) { }
3060
3061 void inc_call_count () { _call_count ++; }
3062 void inc_float_count () { _float_count ++; }
3063 void inc_double_count() { _double_count++; }
3064 void inc_java_call_count() { _java_call_count++; }
3065 void inc_inner_loop_count() { _inner_loop_count++; }
3066
3067 int get_call_count () const { return _call_count ; }
3068 int get_float_count () const { return _float_count ; }
3069 int get_double_count() const { return _double_count; }
3070 int get_java_call_count() const { return _java_call_count; }
3071 int get_inner_loop_count() const { return _inner_loop_count; }
3072 };
3073
3074 //------------------------------final_graph_reshaping_impl----------------------
3075 // Implement items 1-5 from final_graph_reshaping below.
3076 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3077
3078 if ( n->outcnt() == 0 ) return; // dead node
3079 uint nop = n->Opcode();
3080
3081 // Check for 2-input instruction with "last use" on right input.
3082 // Swap to left input. Implements item (2).
3083 if( n->req() == 3 && // two-input instruction
3084 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3085 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3086 n->in(2)->outcnt() == 1 &&// right use IS a last use
3087 !n->in(2)->is_Con() ) { // right use is not a constant
3088 // Check for commutative opcode
3089 switch( nop ) {
3090 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3091 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3092 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3093 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3094 case Op_AndL: case Op_XorL: case Op_OrL:
3095 case Op_AndI: case Op_XorI: case Op_OrI: {
3096 // Move "last use" input to left by swapping inputs
3097 n->swap_edges(1, 2);
3098 break;
3099 }
3100 default:
3101 break;
3102 }
3103 }
3104
3105 #ifdef ASSERT
3106 if( n->is_Mem() ) {
3107 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3108 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3109 // oop will be recorded in oop map if load crosses safepoint
3110 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3111 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3112 "raw memory operations should have control edge");
3113 }
3114 if (n->is_MemBar()) {
3115 MemBarNode* mb = n->as_MemBar();
3116 if (mb->trailing_store() || mb->trailing_load_store()) {
3117 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3118 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3119 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3120 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3121 } else if (mb->leading()) {
3122 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3123 }
3124 }
3125 #endif
3126 // Count FPU ops and common calls, implements item (3)
3127 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3128 if (!gc_handled) {
3129 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3130 }
3131
3132 // Collect CFG split points
3133 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3134 frc._tests.push(n);
3135 }
3136 }
3137
3138 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3139 if (!UseDivMod) {
3140 return;
3141 }
3142
3143 // Check if "a % b" and "a / b" both exist
3144 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3145 if (d == nullptr) {
3146 return;
3147 }
3148
3149 // Replace them with a fused divmod if supported
3150 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3151 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3152 d->subsume_by(divmod->div_proj(), this);
3153 n->subsume_by(divmod->mod_proj(), this);
3154 } else {
3155 // Replace "a % b" with "a - ((a / b) * b)"
3156 Node* mult = MulNode::make(d, d->in(2), bt);
3157 Node* sub = SubNode::make(d->in(1), mult, bt);
3158 n->subsume_by(sub, this);
3159 }
3160 }
3161
3162 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3163 switch( nop ) {
3164 // Count all float operations that may use FPU
3165 case Op_AddF:
3166 case Op_SubF:
3167 case Op_MulF:
3168 case Op_DivF:
3169 case Op_NegF:
3170 case Op_ModF:
3171 case Op_ConvI2F:
3172 case Op_ConF:
3173 case Op_CmpF:
3174 case Op_CmpF3:
3175 case Op_StoreF:
3176 case Op_LoadF:
3177 // case Op_ConvL2F: // longs are split into 32-bit halves
3178 frc.inc_float_count();
3179 break;
3180
3181 case Op_ConvF2D:
3182 case Op_ConvD2F:
3183 frc.inc_float_count();
3184 frc.inc_double_count();
3185 break;
3186
3187 // Count all double operations that may use FPU
3188 case Op_AddD:
3189 case Op_SubD:
3190 case Op_MulD:
3191 case Op_DivD:
3192 case Op_NegD:
3193 case Op_ModD:
3194 case Op_ConvI2D:
3195 case Op_ConvD2I:
3196 // case Op_ConvL2D: // handled by leaf call
3197 // case Op_ConvD2L: // handled by leaf call
3198 case Op_ConD:
3199 case Op_CmpD:
3200 case Op_CmpD3:
3201 case Op_StoreD:
3202 case Op_LoadD:
3203 case Op_LoadD_unaligned:
3204 frc.inc_double_count();
3205 break;
3206 case Op_Opaque1: // Remove Opaque Nodes before matching
3207 n->subsume_by(n->in(1), this);
3208 break;
3209 case Op_CallStaticJava:
3210 case Op_CallJava:
3211 case Op_CallDynamicJava:
3212 frc.inc_java_call_count(); // Count java call site;
3213 case Op_CallRuntime:
3214 case Op_CallLeaf:
3215 case Op_CallLeafVector:
3216 case Op_CallLeafNoFP: {
3217 assert (n->is_Call(), "");
3218 CallNode *call = n->as_Call();
3219 // Count call sites where the FP mode bit would have to be flipped.
3220 // Do not count uncommon runtime calls:
3221 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3222 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3223 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3224 frc.inc_call_count(); // Count the call site
3225 } else { // See if uncommon argument is shared
3226 Node *n = call->in(TypeFunc::Parms);
3227 int nop = n->Opcode();
3228 // Clone shared simple arguments to uncommon calls, item (1).
3229 if (n->outcnt() > 1 &&
3230 !n->is_Proj() &&
3231 nop != Op_CreateEx &&
3232 nop != Op_CheckCastPP &&
3233 nop != Op_DecodeN &&
3234 nop != Op_DecodeNKlass &&
3235 !n->is_Mem() &&
3236 !n->is_Phi()) {
3237 Node *x = n->clone();
3238 call->set_req(TypeFunc::Parms, x);
3239 }
3240 }
3241 break;
3242 }
3243 case Op_StoreB:
3244 case Op_StoreC:
3245 case Op_StoreI:
3246 case Op_StoreL:
3247 case Op_CompareAndSwapB:
3248 case Op_CompareAndSwapS:
3249 case Op_CompareAndSwapI:
3250 case Op_CompareAndSwapL:
3251 case Op_CompareAndSwapP:
3252 case Op_CompareAndSwapN:
3253 case Op_WeakCompareAndSwapB:
3254 case Op_WeakCompareAndSwapS:
3255 case Op_WeakCompareAndSwapI:
3256 case Op_WeakCompareAndSwapL:
3257 case Op_WeakCompareAndSwapP:
3258 case Op_WeakCompareAndSwapN:
3259 case Op_CompareAndExchangeB:
3260 case Op_CompareAndExchangeS:
3261 case Op_CompareAndExchangeI:
3262 case Op_CompareAndExchangeL:
3263 case Op_CompareAndExchangeP:
3264 case Op_CompareAndExchangeN:
3265 case Op_GetAndAddS:
3266 case Op_GetAndAddB:
3267 case Op_GetAndAddI:
3268 case Op_GetAndAddL:
3269 case Op_GetAndSetS:
3270 case Op_GetAndSetB:
3271 case Op_GetAndSetI:
3272 case Op_GetAndSetL:
3273 case Op_GetAndSetP:
3274 case Op_GetAndSetN:
3275 case Op_StoreP:
3276 case Op_StoreN:
3277 case Op_StoreNKlass:
3278 case Op_LoadB:
3279 case Op_LoadUB:
3280 case Op_LoadUS:
3281 case Op_LoadI:
3282 case Op_LoadKlass:
3283 case Op_LoadNKlass:
3284 case Op_LoadL:
3285 case Op_LoadL_unaligned:
3286 case Op_LoadP:
3287 case Op_LoadN:
3288 case Op_LoadRange:
3289 case Op_LoadS:
3290 break;
3291
3292 case Op_AddP: { // Assert sane base pointers
3293 Node *addp = n->in(AddPNode::Address);
3294 assert( !addp->is_AddP() ||
3295 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3296 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3297 "Base pointers must match (addp %u)", addp->_idx );
3298 #ifdef _LP64
3299 if ((UseCompressedOops || UseCompressedClassPointers) &&
3300 addp->Opcode() == Op_ConP &&
3301 addp == n->in(AddPNode::Base) &&
3302 n->in(AddPNode::Offset)->is_Con()) {
3303 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3304 // on the platform and on the compressed oops mode.
3305 // Use addressing with narrow klass to load with offset on x86.
3306 // Some platforms can use the constant pool to load ConP.
3307 // Do this transformation here since IGVN will convert ConN back to ConP.
3308 const Type* t = addp->bottom_type();
3309 bool is_oop = t->isa_oopptr() != nullptr;
3310 bool is_klass = t->isa_klassptr() != nullptr;
3311
3312 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3313 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3314 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3315 Node* nn = nullptr;
3316
3317 int op = is_oop ? Op_ConN : Op_ConNKlass;
3318
3319 // Look for existing ConN node of the same exact type.
3320 Node* r = root();
3321 uint cnt = r->outcnt();
3322 for (uint i = 0; i < cnt; i++) {
3323 Node* m = r->raw_out(i);
3324 if (m!= nullptr && m->Opcode() == op &&
3325 m->bottom_type()->make_ptr() == t) {
3326 nn = m;
3327 break;
3328 }
3329 }
3330 if (nn != nullptr) {
3331 // Decode a narrow oop to match address
3332 // [R12 + narrow_oop_reg<<3 + offset]
3333 if (is_oop) {
3334 nn = new DecodeNNode(nn, t);
3335 } else {
3336 nn = new DecodeNKlassNode(nn, t);
3337 }
3338 // Check for succeeding AddP which uses the same Base.
3339 // Otherwise we will run into the assertion above when visiting that guy.
3340 for (uint i = 0; i < n->outcnt(); ++i) {
3341 Node *out_i = n->raw_out(i);
3342 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3343 out_i->set_req(AddPNode::Base, nn);
3344 #ifdef ASSERT
3345 for (uint j = 0; j < out_i->outcnt(); ++j) {
3346 Node *out_j = out_i->raw_out(j);
3347 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3348 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3349 }
3350 #endif
3351 }
3352 }
3353 n->set_req(AddPNode::Base, nn);
3354 n->set_req(AddPNode::Address, nn);
3355 if (addp->outcnt() == 0) {
3356 addp->disconnect_inputs(this);
3357 }
3358 }
3359 }
3360 }
3361 #endif
3362 break;
3363 }
3364
3365 case Op_CastPP: {
3366 // Remove CastPP nodes to gain more freedom during scheduling but
3367 // keep the dependency they encode as control or precedence edges
3368 // (if control is set already) on memory operations. Some CastPP
3369 // nodes don't have a control (don't carry a dependency): skip
3370 // those.
3371 if (n->in(0) != nullptr) {
3372 ResourceMark rm;
3373 Unique_Node_List wq;
3374 wq.push(n);
3375 for (uint next = 0; next < wq.size(); ++next) {
3376 Node *m = wq.at(next);
3377 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3378 Node* use = m->fast_out(i);
3379 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3380 use->ensure_control_or_add_prec(n->in(0));
3381 } else {
3382 switch(use->Opcode()) {
3383 case Op_AddP:
3384 case Op_DecodeN:
3385 case Op_DecodeNKlass:
3386 case Op_CheckCastPP:
3387 case Op_CastPP:
3388 wq.push(use);
3389 break;
3390 }
3391 }
3392 }
3393 }
3394 }
3395 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3396 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3397 Node* in1 = n->in(1);
3398 const Type* t = n->bottom_type();
3399 Node* new_in1 = in1->clone();
3400 new_in1->as_DecodeN()->set_type(t);
3401
3402 if (!Matcher::narrow_oop_use_complex_address()) {
3403 //
3404 // x86, ARM and friends can handle 2 adds in addressing mode
3405 // and Matcher can fold a DecodeN node into address by using
3406 // a narrow oop directly and do implicit null check in address:
3407 //
3408 // [R12 + narrow_oop_reg<<3 + offset]
3409 // NullCheck narrow_oop_reg
3410 //
3411 // On other platforms (Sparc) we have to keep new DecodeN node and
3412 // use it to do implicit null check in address:
3413 //
3414 // decode_not_null narrow_oop_reg, base_reg
3415 // [base_reg + offset]
3416 // NullCheck base_reg
3417 //
3418 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3419 // to keep the information to which null check the new DecodeN node
3420 // corresponds to use it as value in implicit_null_check().
3421 //
3422 new_in1->set_req(0, n->in(0));
3423 }
3424
3425 n->subsume_by(new_in1, this);
3426 if (in1->outcnt() == 0) {
3427 in1->disconnect_inputs(this);
3428 }
3429 } else {
3430 n->subsume_by(n->in(1), this);
3431 if (n->outcnt() == 0) {
3432 n->disconnect_inputs(this);
3433 }
3434 }
3435 break;
3436 }
3437 #ifdef _LP64
3438 case Op_CmpP:
3439 // Do this transformation here to preserve CmpPNode::sub() and
3440 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3441 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3442 Node* in1 = n->in(1);
3443 Node* in2 = n->in(2);
3444 if (!in1->is_DecodeNarrowPtr()) {
3445 in2 = in1;
3446 in1 = n->in(2);
3447 }
3448 assert(in1->is_DecodeNarrowPtr(), "sanity");
3449
3450 Node* new_in2 = nullptr;
3451 if (in2->is_DecodeNarrowPtr()) {
3452 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3453 new_in2 = in2->in(1);
3454 } else if (in2->Opcode() == Op_ConP) {
3455 const Type* t = in2->bottom_type();
3456 if (t == TypePtr::NULL_PTR) {
3457 assert(in1->is_DecodeN(), "compare klass to null?");
3458 // Don't convert CmpP null check into CmpN if compressed
3459 // oops implicit null check is not generated.
3460 // This will allow to generate normal oop implicit null check.
3461 if (Matcher::gen_narrow_oop_implicit_null_checks())
3462 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3463 //
3464 // This transformation together with CastPP transformation above
3465 // will generated code for implicit null checks for compressed oops.
3466 //
3467 // The original code after Optimize()
3468 //
3469 // LoadN memory, narrow_oop_reg
3470 // decode narrow_oop_reg, base_reg
3471 // CmpP base_reg, nullptr
3472 // CastPP base_reg // NotNull
3473 // Load [base_reg + offset], val_reg
3474 //
3475 // after these transformations will be
3476 //
3477 // LoadN memory, narrow_oop_reg
3478 // CmpN narrow_oop_reg, nullptr
3479 // decode_not_null narrow_oop_reg, base_reg
3480 // Load [base_reg + offset], val_reg
3481 //
3482 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3483 // since narrow oops can be used in debug info now (see the code in
3484 // final_graph_reshaping_walk()).
3485 //
3486 // At the end the code will be matched to
3487 // on x86:
3488 //
3489 // Load_narrow_oop memory, narrow_oop_reg
3490 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3491 // NullCheck narrow_oop_reg
3492 //
3493 // and on sparc:
3494 //
3495 // Load_narrow_oop memory, narrow_oop_reg
3496 // decode_not_null narrow_oop_reg, base_reg
3497 // Load [base_reg + offset], val_reg
3498 // NullCheck base_reg
3499 //
3500 } else if (t->isa_oopptr()) {
3501 new_in2 = ConNode::make(t->make_narrowoop());
3502 } else if (t->isa_klassptr()) {
3503 new_in2 = ConNode::make(t->make_narrowklass());
3504 }
3505 }
3506 if (new_in2 != nullptr) {
3507 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3508 n->subsume_by(cmpN, this);
3509 if (in1->outcnt() == 0) {
3510 in1->disconnect_inputs(this);
3511 }
3512 if (in2->outcnt() == 0) {
3513 in2->disconnect_inputs(this);
3514 }
3515 }
3516 }
3517 break;
3518
3519 case Op_DecodeN:
3520 case Op_DecodeNKlass:
3521 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3522 // DecodeN could be pinned when it can't be fold into
3523 // an address expression, see the code for Op_CastPP above.
3524 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3525 break;
3526
3527 case Op_EncodeP:
3528 case Op_EncodePKlass: {
3529 Node* in1 = n->in(1);
3530 if (in1->is_DecodeNarrowPtr()) {
3531 n->subsume_by(in1->in(1), this);
3532 } else if (in1->Opcode() == Op_ConP) {
3533 const Type* t = in1->bottom_type();
3534 if (t == TypePtr::NULL_PTR) {
3535 assert(t->isa_oopptr(), "null klass?");
3536 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3537 } else if (t->isa_oopptr()) {
3538 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3539 } else if (t->isa_klassptr()) {
3540 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3541 }
3542 }
3543 if (in1->outcnt() == 0) {
3544 in1->disconnect_inputs(this);
3545 }
3546 break;
3547 }
3548
3549 case Op_Proj: {
3550 if (OptimizeStringConcat || IncrementalInline) {
3551 ProjNode* proj = n->as_Proj();
3552 if (proj->_is_io_use) {
3553 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3554 // Separate projections were used for the exception path which
3555 // are normally removed by a late inline. If it wasn't inlined
3556 // then they will hang around and should just be replaced with
3557 // the original one. Merge them.
3558 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3559 if (non_io_proj != nullptr) {
3560 proj->subsume_by(non_io_proj , this);
3561 }
3562 }
3563 }
3564 break;
3565 }
3566
3567 case Op_Phi:
3568 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3569 // The EncodeP optimization may create Phi with the same edges
3570 // for all paths. It is not handled well by Register Allocator.
3571 Node* unique_in = n->in(1);
3572 assert(unique_in != nullptr, "");
3573 uint cnt = n->req();
3574 for (uint i = 2; i < cnt; i++) {
3575 Node* m = n->in(i);
3576 assert(m != nullptr, "");
3577 if (unique_in != m)
3578 unique_in = nullptr;
3579 }
3580 if (unique_in != nullptr) {
3581 n->subsume_by(unique_in, this);
3582 }
3583 }
3584 break;
3585
3586 #endif
3587
3588 #ifdef ASSERT
3589 case Op_CastII:
3590 // Verify that all range check dependent CastII nodes were removed.
3591 if (n->isa_CastII()->has_range_check()) {
3592 n->dump(3);
3593 assert(false, "Range check dependent CastII node was not removed");
3594 }
3595 break;
3596 #endif
3597
3598 case Op_ModI:
3599 handle_div_mod_op(n, T_INT, false);
3600 break;
3601
3602 case Op_ModL:
3603 handle_div_mod_op(n, T_LONG, false);
3604 break;
3605
3606 case Op_UModI:
3607 handle_div_mod_op(n, T_INT, true);
3608 break;
3609
3610 case Op_UModL:
3611 handle_div_mod_op(n, T_LONG, true);
3612 break;
3613
3614 case Op_LoadVector:
3615 case Op_StoreVector:
3616 #ifdef ASSERT
3617 // Add VerifyVectorAlignment node between adr and load / store.
3618 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3619 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3620 n->as_StoreVector()->must_verify_alignment();
3621 if (must_verify_alignment) {
3622 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3623 n->as_StoreVector()->memory_size();
3624 // The memory access should be aligned to the vector width in bytes.
3625 // However, the underlying array is possibly less well aligned, but at least
3626 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3627 // a loop we can expect at least the following alignment:
3628 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3629 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3630 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3631 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3632 jlong mask = guaranteed_alignment - 1;
3633 Node* mask_con = ConLNode::make(mask);
3634 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3635 n->set_req(MemNode::Address, va);
3636 }
3637 }
3638 #endif
3639 break;
3640
3641 case Op_LoadVectorGather:
3642 case Op_StoreVectorScatter:
3643 case Op_LoadVectorGatherMasked:
3644 case Op_StoreVectorScatterMasked:
3645 case Op_VectorCmpMasked:
3646 case Op_VectorMaskGen:
3647 case Op_LoadVectorMasked:
3648 case Op_StoreVectorMasked:
3649 break;
3650
3651 case Op_AddReductionVI:
3652 case Op_AddReductionVL:
3653 case Op_AddReductionVF:
3654 case Op_AddReductionVD:
3655 case Op_MulReductionVI:
3656 case Op_MulReductionVL:
3657 case Op_MulReductionVF:
3658 case Op_MulReductionVD:
3659 case Op_MinReductionV:
3660 case Op_MaxReductionV:
3661 case Op_AndReductionV:
3662 case Op_OrReductionV:
3663 case Op_XorReductionV:
3664 break;
3665
3666 case Op_PackB:
3667 case Op_PackS:
3668 case Op_PackI:
3669 case Op_PackF:
3670 case Op_PackL:
3671 case Op_PackD:
3672 if (n->req()-1 > 2) {
3673 // Replace many operand PackNodes with a binary tree for matching
3674 PackNode* p = (PackNode*) n;
3675 Node* btp = p->binary_tree_pack(1, n->req());
3676 n->subsume_by(btp, this);
3677 }
3678 break;
3679 case Op_Loop:
3680 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3681 case Op_CountedLoop:
3682 case Op_LongCountedLoop:
3683 case Op_OuterStripMinedLoop:
3684 if (n->as_Loop()->is_inner_loop()) {
3685 frc.inc_inner_loop_count();
3686 }
3687 n->as_Loop()->verify_strip_mined(0);
3688 break;
3689 case Op_LShiftI:
3690 case Op_RShiftI:
3691 case Op_URShiftI:
3692 case Op_LShiftL:
3693 case Op_RShiftL:
3694 case Op_URShiftL:
3695 if (Matcher::need_masked_shift_count) {
3696 // The cpu's shift instructions don't restrict the count to the
3697 // lower 5/6 bits. We need to do the masking ourselves.
3698 Node* in2 = n->in(2);
3699 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3700 const TypeInt* t = in2->find_int_type();
3701 if (t != nullptr && t->is_con()) {
3702 juint shift = t->get_con();
3703 if (shift > mask) { // Unsigned cmp
3704 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3705 }
3706 } else {
3707 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3708 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3709 n->set_req(2, shift);
3710 }
3711 }
3712 if (in2->outcnt() == 0) { // Remove dead node
3713 in2->disconnect_inputs(this);
3714 }
3715 }
3716 break;
3717 case Op_MemBarStoreStore:
3718 case Op_MemBarRelease:
3719 // Break the link with AllocateNode: it is no longer useful and
3720 // confuses register allocation.
3721 if (n->req() > MemBarNode::Precedent) {
3722 n->set_req(MemBarNode::Precedent, top());
3723 }
3724 break;
3725 case Op_MemBarAcquire: {
3726 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3727 // At parse time, the trailing MemBarAcquire for a volatile load
3728 // is created with an edge to the load. After optimizations,
3729 // that input may be a chain of Phis. If those phis have no
3730 // other use, then the MemBarAcquire keeps them alive and
3731 // register allocation can be confused.
3732 dead_nodes.push(n->in(MemBarNode::Precedent));
3733 n->set_req(MemBarNode::Precedent, top());
3734 }
3735 break;
3736 }
3737 case Op_Blackhole:
3738 break;
3739 case Op_RangeCheck: {
3740 RangeCheckNode* rc = n->as_RangeCheck();
3741 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3742 n->subsume_by(iff, this);
3743 frc._tests.push(iff);
3744 break;
3745 }
3746 case Op_ConvI2L: {
3747 if (!Matcher::convi2l_type_required) {
3748 // Code generation on some platforms doesn't need accurate
3749 // ConvI2L types. Widening the type can help remove redundant
3750 // address computations.
3751 n->as_Type()->set_type(TypeLong::INT);
3752 ResourceMark rm;
3753 Unique_Node_List wq;
3754 wq.push(n);
3755 for (uint next = 0; next < wq.size(); next++) {
3756 Node *m = wq.at(next);
3757
3758 for(;;) {
3759 // Loop over all nodes with identical inputs edges as m
3760 Node* k = m->find_similar(m->Opcode());
3761 if (k == nullptr) {
3762 break;
3763 }
3764 // Push their uses so we get a chance to remove node made
3765 // redundant
3766 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3767 Node* u = k->fast_out(i);
3768 if (u->Opcode() == Op_LShiftL ||
3769 u->Opcode() == Op_AddL ||
3770 u->Opcode() == Op_SubL ||
3771 u->Opcode() == Op_AddP) {
3772 wq.push(u);
3773 }
3774 }
3775 // Replace all nodes with identical edges as m with m
3776 k->subsume_by(m, this);
3777 }
3778 }
3779 }
3780 break;
3781 }
3782 case Op_CmpUL: {
3783 if (!Matcher::has_match_rule(Op_CmpUL)) {
3784 // No support for unsigned long comparisons
3785 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3786 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3787 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3788 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3789 Node* andl = new AndLNode(orl, remove_sign_mask);
3790 Node* cmp = new CmpLNode(andl, n->in(2));
3791 n->subsume_by(cmp, this);
3792 }
3793 break;
3794 }
3795 #ifdef ASSERT
3796 case Op_ConNKlass: {
3797 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3798 ciKlass* klass = tp->is_klassptr()->exact_klass();
3799 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3800 break;
3801 }
3802 #endif
3803 default:
3804 assert(!n->is_Call(), "");
3805 assert(!n->is_Mem(), "");
3806 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3807 break;
3808 }
3809 }
3810
3811 //------------------------------final_graph_reshaping_walk---------------------
3812 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3813 // requires that the walk visits a node's inputs before visiting the node.
3814 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3815 Unique_Node_List sfpt;
3816
3817 frc._visited.set(root->_idx); // first, mark node as visited
3818 uint cnt = root->req();
3819 Node *n = root;
3820 uint i = 0;
3821 while (true) {
3822 if (i < cnt) {
3823 // Place all non-visited non-null inputs onto stack
3824 Node* m = n->in(i);
3825 ++i;
3826 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3827 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3828 // compute worst case interpreter size in case of a deoptimization
3829 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3830
3831 sfpt.push(m);
3832 }
3833 cnt = m->req();
3834 nstack.push(n, i); // put on stack parent and next input's index
3835 n = m;
3836 i = 0;
3837 }
3838 } else {
3839 // Now do post-visit work
3840 final_graph_reshaping_impl(n, frc, dead_nodes);
3841 if (nstack.is_empty())
3842 break; // finished
3843 n = nstack.node(); // Get node from stack
3844 cnt = n->req();
3845 i = nstack.index();
3846 nstack.pop(); // Shift to the next node on stack
3847 }
3848 }
3849
3850 // Skip next transformation if compressed oops are not used.
3851 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3852 (!UseCompressedOops && !UseCompressedClassPointers))
3853 return;
3854
3855 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3856 // It could be done for an uncommon traps or any safepoints/calls
3857 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3858 while (sfpt.size() > 0) {
3859 n = sfpt.pop();
3860 JVMState *jvms = n->as_SafePoint()->jvms();
3861 assert(jvms != nullptr, "sanity");
3862 int start = jvms->debug_start();
3863 int end = n->req();
3864 bool is_uncommon = (n->is_CallStaticJava() &&
3865 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3866 for (int j = start; j < end; j++) {
3867 Node* in = n->in(j);
3868 if (in->is_DecodeNarrowPtr()) {
3869 bool safe_to_skip = true;
3870 if (!is_uncommon ) {
3871 // Is it safe to skip?
3872 for (uint i = 0; i < in->outcnt(); i++) {
3873 Node* u = in->raw_out(i);
3874 if (!u->is_SafePoint() ||
3875 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3876 safe_to_skip = false;
3877 }
3878 }
3879 }
3880 if (safe_to_skip) {
3881 n->set_req(j, in->in(1));
3882 }
3883 if (in->outcnt() == 0) {
3884 in->disconnect_inputs(this);
3885 }
3886 }
3887 }
3888 }
3889 }
3890
3891 //------------------------------final_graph_reshaping--------------------------
3892 // Final Graph Reshaping.
3893 //
3894 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3895 // and not commoned up and forced early. Must come after regular
3896 // optimizations to avoid GVN undoing the cloning. Clone constant
3897 // inputs to Loop Phis; these will be split by the allocator anyways.
3898 // Remove Opaque nodes.
3899 // (2) Move last-uses by commutative operations to the left input to encourage
3900 // Intel update-in-place two-address operations and better register usage
3901 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3902 // calls canonicalizing them back.
3903 // (3) Count the number of double-precision FP ops, single-precision FP ops
3904 // and call sites. On Intel, we can get correct rounding either by
3905 // forcing singles to memory (requires extra stores and loads after each
3906 // FP bytecode) or we can set a rounding mode bit (requires setting and
3907 // clearing the mode bit around call sites). The mode bit is only used
3908 // if the relative frequency of single FP ops to calls is low enough.
3909 // This is a key transform for SPEC mpeg_audio.
3910 // (4) Detect infinite loops; blobs of code reachable from above but not
3911 // below. Several of the Code_Gen algorithms fail on such code shapes,
3912 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3913 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3914 // Detection is by looking for IfNodes where only 1 projection is
3915 // reachable from below or CatchNodes missing some targets.
3916 // (5) Assert for insane oop offsets in debug mode.
3917
3918 bool Compile::final_graph_reshaping() {
3919 // an infinite loop may have been eliminated by the optimizer,
3920 // in which case the graph will be empty.
3921 if (root()->req() == 1) {
3922 // Do not compile method that is only a trivial infinite loop,
3923 // since the content of the loop may have been eliminated.
3924 record_method_not_compilable("trivial infinite loop");
3925 return true;
3926 }
3927
3928 // Expensive nodes have their control input set to prevent the GVN
3929 // from freely commoning them. There's no GVN beyond this point so
3930 // no need to keep the control input. We want the expensive nodes to
3931 // be freely moved to the least frequent code path by gcm.
3932 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3933 for (int i = 0; i < expensive_count(); i++) {
3934 _expensive_nodes.at(i)->set_req(0, nullptr);
3935 }
3936
3937 Final_Reshape_Counts frc;
3938
3939 // Visit everybody reachable!
3940 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3941 Node_Stack nstack(live_nodes() >> 1);
3942 Unique_Node_List dead_nodes;
3943 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
3944
3945 // Check for unreachable (from below) code (i.e., infinite loops).
3946 for( uint i = 0; i < frc._tests.size(); i++ ) {
3947 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3948 // Get number of CFG targets.
3949 // Note that PCTables include exception targets after calls.
3950 uint required_outcnt = n->required_outcnt();
3951 if (n->outcnt() != required_outcnt) {
3952 // Check for a few special cases. Rethrow Nodes never take the
3953 // 'fall-thru' path, so expected kids is 1 less.
3954 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3955 if (n->in(0)->in(0)->is_Call()) {
3956 CallNode* call = n->in(0)->in(0)->as_Call();
3957 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3958 required_outcnt--; // Rethrow always has 1 less kid
3959 } else if (call->req() > TypeFunc::Parms &&
3960 call->is_CallDynamicJava()) {
3961 // Check for null receiver. In such case, the optimizer has
3962 // detected that the virtual call will always result in a null
3963 // pointer exception. The fall-through projection of this CatchNode
3964 // will not be populated.
3965 Node* arg0 = call->in(TypeFunc::Parms);
3966 if (arg0->is_Type() &&
3967 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3968 required_outcnt--;
3969 }
3970 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
3971 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3972 // Check for illegal array length. In such case, the optimizer has
3973 // detected that the allocation attempt will always result in an
3974 // exception. There is no fall-through projection of this CatchNode .
3975 assert(call->is_CallStaticJava(), "static call expected");
3976 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
3977 uint valid_length_test_input = call->req() - 1;
3978 Node* valid_length_test = call->in(valid_length_test_input);
3979 call->del_req(valid_length_test_input);
3980 if (valid_length_test->find_int_con(1) == 0) {
3981 required_outcnt--;
3982 }
3983 dead_nodes.push(valid_length_test);
3984 assert(n->outcnt() == required_outcnt, "malformed control flow");
3985 continue;
3986 }
3987 }
3988 }
3989
3990 // Recheck with a better notion of 'required_outcnt'
3991 if (n->outcnt() != required_outcnt) {
3992 record_method_not_compilable("malformed control flow");
3993 return true; // Not all targets reachable!
3994 }
3995 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
3996 CallNode* call = n->in(0)->in(0)->as_Call();
3997 if (call->entry_point() == OptoRuntime::new_array_Java() ||
3998 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3999 assert(call->is_CallStaticJava(), "static call expected");
4000 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4001 uint valid_length_test_input = call->req() - 1;
4002 dead_nodes.push(call->in(valid_length_test_input));
4003 call->del_req(valid_length_test_input); // valid length test useless now
4004 }
4005 }
4006 // Check that I actually visited all kids. Unreached kids
4007 // must be infinite loops.
4008 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4009 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4010 record_method_not_compilable("infinite loop");
4011 return true; // Found unvisited kid; must be unreach
4012 }
4013
4014 // Here so verification code in final_graph_reshaping_walk()
4015 // always see an OuterStripMinedLoopEnd
4016 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4017 IfNode* init_iff = n->as_If();
4018 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4019 n->subsume_by(iff, this);
4020 }
4021 }
4022
4023 while (dead_nodes.size() > 0) {
4024 Node* m = dead_nodes.pop();
4025 if (m->outcnt() == 0 && m != top()) {
4026 for (uint j = 0; j < m->req(); j++) {
4027 Node* in = m->in(j);
4028 if (in != nullptr) {
4029 dead_nodes.push(in);
4030 }
4031 }
4032 m->disconnect_inputs(this);
4033 }
4034 }
4035
4036 #ifdef IA32
4037 // If original bytecodes contained a mixture of floats and doubles
4038 // check if the optimizer has made it homogeneous, item (3).
4039 if (UseSSE == 0 &&
4040 frc.get_float_count() > 32 &&
4041 frc.get_double_count() == 0 &&
4042 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4043 set_24_bit_selection_and_mode(false, true);
4044 }
4045 #endif // IA32
4046
4047 set_java_calls(frc.get_java_call_count());
4048 set_inner_loops(frc.get_inner_loop_count());
4049
4050 // No infinite loops, no reason to bail out.
4051 return false;
4052 }
4053
4054 //-----------------------------too_many_traps----------------------------------
4055 // Report if there are too many traps at the current method and bci.
4056 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4057 bool Compile::too_many_traps(ciMethod* method,
4058 int bci,
4059 Deoptimization::DeoptReason reason) {
4060 if (method->has_trap_at(bci)) {
4061 return true;
4062 }
4063 ciMethodData* md = method->method_data();
4064 if (md->is_empty()) {
4065 // Assume the trap has not occurred, or that it occurred only
4066 // because of a transient condition during start-up in the interpreter.
4067 return false;
4068 }
4069 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4070 if (md->has_trap_at(bci, m, reason) != 0) {
4071 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4072 // Also, if there are multiple reasons, or if there is no per-BCI record,
4073 // assume the worst.
4074 if (log())
4075 log()->elem("observe trap='%s' count='%d'",
4076 Deoptimization::trap_reason_name(reason),
4077 md->trap_count(reason));
4078 return true;
4079 } else {
4080 // Ignore method/bci and see if there have been too many globally.
4081 return too_many_traps(reason, md);
4082 }
4083 }
4084
4085 // Less-accurate variant which does not require a method and bci.
4086 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4087 ciMethodData* logmd) {
4088 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4089 // Too many traps globally.
4090 // Note that we use cumulative trap_count, not just md->trap_count.
4091 if (log()) {
4092 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4093 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4094 Deoptimization::trap_reason_name(reason),
4095 mcount, trap_count(reason));
4096 }
4097 return true;
4098 } else {
4099 // The coast is clear.
4100 return false;
4101 }
4102 }
4103
4104 //--------------------------too_many_recompiles--------------------------------
4105 // Report if there are too many recompiles at the current method and bci.
4106 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4107 // Is not eager to return true, since this will cause the compiler to use
4108 // Action_none for a trap point, to avoid too many recompilations.
4109 bool Compile::too_many_recompiles(ciMethod* method,
4110 int bci,
4111 Deoptimization::DeoptReason reason) {
4112 ciMethodData* md = method->method_data();
4113 if (md->is_empty()) {
4114 // Assume the trap has not occurred, or that it occurred only
4115 // because of a transient condition during start-up in the interpreter.
4116 return false;
4117 }
4118 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4119 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4120 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4121 Deoptimization::DeoptReason per_bc_reason
4122 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4123 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4124 if ((per_bc_reason == Deoptimization::Reason_none
4125 || md->has_trap_at(bci, m, reason) != 0)
4126 // The trap frequency measure we care about is the recompile count:
4127 && md->trap_recompiled_at(bci, m)
4128 && md->overflow_recompile_count() >= bc_cutoff) {
4129 // Do not emit a trap here if it has already caused recompilations.
4130 // Also, if there are multiple reasons, or if there is no per-BCI record,
4131 // assume the worst.
4132 if (log())
4133 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4134 Deoptimization::trap_reason_name(reason),
4135 md->trap_count(reason),
4136 md->overflow_recompile_count());
4137 return true;
4138 } else if (trap_count(reason) != 0
4139 && decompile_count() >= m_cutoff) {
4140 // Too many recompiles globally, and we have seen this sort of trap.
4141 // Use cumulative decompile_count, not just md->decompile_count.
4142 if (log())
4143 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4144 Deoptimization::trap_reason_name(reason),
4145 md->trap_count(reason), trap_count(reason),
4146 md->decompile_count(), decompile_count());
4147 return true;
4148 } else {
4149 // The coast is clear.
4150 return false;
4151 }
4152 }
4153
4154 // Compute when not to trap. Used by matching trap based nodes and
4155 // NullCheck optimization.
4156 void Compile::set_allowed_deopt_reasons() {
4157 _allowed_reasons = 0;
4158 if (is_method_compilation()) {
4159 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4160 assert(rs < BitsPerInt, "recode bit map");
4161 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4162 _allowed_reasons |= nth_bit(rs);
4163 }
4164 }
4165 }
4166 }
4167
4168 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4169 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4170 }
4171
4172 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4173 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4174 }
4175
4176 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4177 if (holder->is_initialized() && !do_clinit_barriers()) {
4178 return false;
4179 }
4180 if (holder->is_being_initialized() || do_clinit_barriers()) {
4181 if (accessing_method->holder() == holder) {
4182 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4183 // <init>, or a static method. In all those cases, there was an initialization
4184 // barrier on the holder klass passed.
4185 if (accessing_method->is_static_initializer() ||
4186 accessing_method->is_object_initializer() ||
4187 accessing_method->is_static()) {
4188 return false;
4189 }
4190 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4191 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4192 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4193 // child class can become fully initialized while its parent class is still being initialized.
4194 if (accessing_method->is_static_initializer()) {
4195 return false;
4196 }
4197 }
4198 ciMethod* root = method(); // the root method of compilation
4199 if (root != accessing_method) {
4200 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4201 }
4202 }
4203 return true;
4204 }
4205
4206 #ifndef PRODUCT
4207 //------------------------------verify_bidirectional_edges---------------------
4208 // For each input edge to a node (ie - for each Use-Def edge), verify that
4209 // there is a corresponding Def-Use edge.
4210 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4211 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4212 uint stack_size = live_nodes() >> 4;
4213 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4214 nstack.push(_root);
4215
4216 while (nstack.size() > 0) {
4217 Node* n = nstack.pop();
4218 if (visited.member(n)) {
4219 continue;
4220 }
4221 visited.push(n);
4222
4223 // Walk over all input edges, checking for correspondence
4224 uint length = n->len();
4225 for (uint i = 0; i < length; i++) {
4226 Node* in = n->in(i);
4227 if (in != nullptr && !visited.member(in)) {
4228 nstack.push(in); // Put it on stack
4229 }
4230 if (in != nullptr && !in->is_top()) {
4231 // Count instances of `next`
4232 int cnt = 0;
4233 for (uint idx = 0; idx < in->_outcnt; idx++) {
4234 if (in->_out[idx] == n) {
4235 cnt++;
4236 }
4237 }
4238 assert(cnt > 0, "Failed to find Def-Use edge.");
4239 // Check for duplicate edges
4240 // walk the input array downcounting the input edges to n
4241 for (uint j = 0; j < length; j++) {
4242 if (n->in(j) == in) {
4243 cnt--;
4244 }
4245 }
4246 assert(cnt == 0, "Mismatched edge count.");
4247 } else if (in == nullptr) {
4248 assert(i == 0 || i >= n->req() ||
4249 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4250 (n->is_Unlock() && i == (n->req() - 1)) ||
4251 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4252 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4253 } else {
4254 assert(in->is_top(), "sanity");
4255 // Nothing to check.
4256 }
4257 }
4258 }
4259 }
4260
4261 //------------------------------verify_graph_edges---------------------------
4262 // Walk the Graph and verify that there is a one-to-one correspondence
4263 // between Use-Def edges and Def-Use edges in the graph.
4264 void Compile::verify_graph_edges(bool no_dead_code) {
4265 if (VerifyGraphEdges) {
4266 Unique_Node_List visited;
4267
4268 // Call graph walk to check edges
4269 verify_bidirectional_edges(visited);
4270 if (no_dead_code) {
4271 // Now make sure that no visited node is used by an unvisited node.
4272 bool dead_nodes = false;
4273 Unique_Node_List checked;
4274 while (visited.size() > 0) {
4275 Node* n = visited.pop();
4276 checked.push(n);
4277 for (uint i = 0; i < n->outcnt(); i++) {
4278 Node* use = n->raw_out(i);
4279 if (checked.member(use)) continue; // already checked
4280 if (visited.member(use)) continue; // already in the graph
4281 if (use->is_Con()) continue; // a dead ConNode is OK
4282 // At this point, we have found a dead node which is DU-reachable.
4283 if (!dead_nodes) {
4284 tty->print_cr("*** Dead nodes reachable via DU edges:");
4285 dead_nodes = true;
4286 }
4287 use->dump(2);
4288 tty->print_cr("---");
4289 checked.push(use); // No repeats; pretend it is now checked.
4290 }
4291 }
4292 assert(!dead_nodes, "using nodes must be reachable from root");
4293 }
4294 }
4295 }
4296 #endif
4297
4298 // The Compile object keeps track of failure reasons separately from the ciEnv.
4299 // This is required because there is not quite a 1-1 relation between the
4300 // ciEnv and its compilation task and the Compile object. Note that one
4301 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4302 // to backtrack and retry without subsuming loads. Other than this backtracking
4303 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4304 // by the logic in C2Compiler.
4305 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4306 if (log() != nullptr) {
4307 log()->elem("failure reason='%s' phase='compile'", reason);
4308 }
4309 if (_failure_reason.get() == nullptr) {
4310 // Record the first failure reason.
4311 _failure_reason.set(reason);
4312 if (CaptureBailoutInformation) {
4313 _first_failure_details = new CompilationFailureInfo(reason);
4314 }
4315 } else {
4316 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4317 }
4318
4319 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4320 C->print_method(PHASE_FAILURE, 1);
4321 }
4322 _root = nullptr; // flush the graph, too
4323 }
4324
4325 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4326 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4327 _compile(Compile::current()),
4328 _log(nullptr),
4329 _dolog(CITimeVerbose)
4330 {
4331 assert(_compile != nullptr, "sanity check");
4332 if (_dolog) {
4333 _log = _compile->log();
4334 }
4335 if (_log != nullptr) {
4336 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4337 _log->stamp();
4338 _log->end_head();
4339 }
4340 }
4341
4342 Compile::TracePhase::TracePhase(PhaseTraceId id)
4343 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4344
4345 Compile::TracePhase::~TracePhase() {
4346 if (_compile->failing_internal()) {
4347 return; // timing code, not stressing bailouts.
4348 }
4349 #ifdef ASSERT
4350 if (PrintIdealNodeCount) {
4351 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4352 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4353 }
4354
4355 if (VerifyIdealNodeCount) {
4356 _compile->print_missing_nodes();
4357 }
4358 #endif
4359
4360 if (_log != nullptr) {
4361 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4362 }
4363 }
4364
4365 //----------------------------static_subtype_check-----------------------------
4366 // Shortcut important common cases when superklass is exact:
4367 // (0) superklass is java.lang.Object (can occur in reflective code)
4368 // (1) subklass is already limited to a subtype of superklass => always ok
4369 // (2) subklass does not overlap with superklass => always fail
4370 // (3) superklass has NO subtypes and we can check with a simple compare.
4371 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4372 if (skip) {
4373 return SSC_full_test; // Let caller generate the general case.
4374 }
4375
4376 if (subk->is_java_subtype_of(superk)) {
4377 return SSC_always_true; // (0) and (1) this test cannot fail
4378 }
4379
4380 if (!subk->maybe_java_subtype_of(superk)) {
4381 return SSC_always_false; // (2) true path dead; no dynamic test needed
4382 }
4383
4384 const Type* superelem = superk;
4385 if (superk->isa_aryklassptr()) {
4386 int ignored;
4387 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4388 }
4389
4390 if (superelem->isa_instklassptr()) {
4391 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4392 if (!ik->has_subklass()) {
4393 if (!ik->is_final()) {
4394 // Add a dependency if there is a chance of a later subclass.
4395 dependencies()->assert_leaf_type(ik);
4396 }
4397 if (!superk->maybe_java_subtype_of(subk)) {
4398 return SSC_always_false;
4399 }
4400 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4401 }
4402 } else {
4403 // A primitive array type has no subtypes.
4404 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4405 }
4406
4407 return SSC_full_test;
4408 }
4409
4410 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4411 #ifdef _LP64
4412 // The scaled index operand to AddP must be a clean 64-bit value.
4413 // Java allows a 32-bit int to be incremented to a negative
4414 // value, which appears in a 64-bit register as a large
4415 // positive number. Using that large positive number as an
4416 // operand in pointer arithmetic has bad consequences.
4417 // On the other hand, 32-bit overflow is rare, and the possibility
4418 // can often be excluded, if we annotate the ConvI2L node with
4419 // a type assertion that its value is known to be a small positive
4420 // number. (The prior range check has ensured this.)
4421 // This assertion is used by ConvI2LNode::Ideal.
4422 int index_max = max_jint - 1; // array size is max_jint, index is one less
4423 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4424 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4425 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4426 #endif
4427 return idx;
4428 }
4429
4430 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4431 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4432 if (ctrl != nullptr) {
4433 // Express control dependency by a CastII node with a narrow type.
4434 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4435 // node from floating above the range check during loop optimizations. Otherwise, the
4436 // ConvI2L node may be eliminated independently of the range check, causing the data path
4437 // to become TOP while the control path is still there (although it's unreachable).
4438 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4439 value = phase->transform(value);
4440 }
4441 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4442 return phase->transform(new ConvI2LNode(value, ltype));
4443 }
4444
4445 // The message about the current inlining is accumulated in
4446 // _print_inlining_stream and transferred into the _print_inlining_list
4447 // once we know whether inlining succeeds or not. For regular
4448 // inlining, messages are appended to the buffer pointed by
4449 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4450 // a new buffer is added after _print_inlining_idx in the list. This
4451 // way we can update the inlining message for late inlining call site
4452 // when the inlining is attempted again.
4453 void Compile::print_inlining_init() {
4454 if (print_inlining() || print_intrinsics()) {
4455 // print_inlining_init is actually called several times.
4456 print_inlining_reset();
4457 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4458 }
4459 }
4460
4461 void Compile::print_inlining_reinit() {
4462 if (print_inlining() || print_intrinsics()) {
4463 print_inlining_reset();
4464 }
4465 }
4466
4467 void Compile::print_inlining_reset() {
4468 _print_inlining_stream->reset();
4469 }
4470
4471 void Compile::print_inlining_commit() {
4472 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4473 // Transfer the message from _print_inlining_stream to the current
4474 // _print_inlining_list buffer and clear _print_inlining_stream.
4475 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4476 print_inlining_reset();
4477 }
4478
4479 void Compile::print_inlining_push() {
4480 // Add new buffer to the _print_inlining_list at current position
4481 _print_inlining_idx++;
4482 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4483 }
4484
4485 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4486 return _print_inlining_list->at(_print_inlining_idx);
4487 }
4488
4489 void Compile::print_inlining_update(CallGenerator* cg) {
4490 if (print_inlining() || print_intrinsics()) {
4491 if (cg->is_late_inline()) {
4492 if (print_inlining_current()->cg() != cg &&
4493 (print_inlining_current()->cg() != nullptr ||
4494 print_inlining_current()->ss()->size() != 0)) {
4495 print_inlining_push();
4496 }
4497 print_inlining_commit();
4498 print_inlining_current()->set_cg(cg);
4499 } else {
4500 if (print_inlining_current()->cg() != nullptr) {
4501 print_inlining_push();
4502 }
4503 print_inlining_commit();
4504 }
4505 }
4506 }
4507
4508 void Compile::print_inlining_move_to(CallGenerator* cg) {
4509 // We resume inlining at a late inlining call site. Locate the
4510 // corresponding inlining buffer so that we can update it.
4511 if (print_inlining() || print_intrinsics()) {
4512 for (int i = 0; i < _print_inlining_list->length(); i++) {
4513 if (_print_inlining_list->at(i)->cg() == cg) {
4514 _print_inlining_idx = i;
4515 return;
4516 }
4517 }
4518 ShouldNotReachHere();
4519 }
4520 }
4521
4522 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4523 if (print_inlining() || print_intrinsics()) {
4524 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4525 assert(print_inlining_current()->cg() == cg, "wrong entry");
4526 // replace message with new message
4527 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4528 print_inlining_commit();
4529 print_inlining_current()->set_cg(cg);
4530 }
4531 }
4532
4533 void Compile::print_inlining_assert_ready() {
4534 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4535 }
4536
4537 void Compile::process_print_inlining() {
4538 assert(_late_inlines.length() == 0, "not drained yet");
4539 if (print_inlining() || print_intrinsics()) {
4540 ResourceMark rm;
4541 stringStream ss;
4542 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4543 for (int i = 0; i < _print_inlining_list->length(); i++) {
4544 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4545 ss.print("%s", pib->ss()->freeze());
4546 delete pib;
4547 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4548 }
4549 // Reset _print_inlining_list, it only contains destructed objects.
4550 // It is on the arena, so it will be freed when the arena is reset.
4551 _print_inlining_list = nullptr;
4552 // _print_inlining_stream won't be used anymore, either.
4553 print_inlining_reset();
4554 size_t end = ss.size();
4555 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4556 strncpy(_print_inlining_output, ss.freeze(), end+1);
4557 _print_inlining_output[end] = 0;
4558 }
4559 }
4560
4561 void Compile::dump_print_inlining() {
4562 if (_print_inlining_output != nullptr) {
4563 tty->print_raw(_print_inlining_output);
4564 }
4565 }
4566
4567 void Compile::log_late_inline(CallGenerator* cg) {
4568 if (log() != nullptr) {
4569 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4570 cg->unique_id());
4571 JVMState* p = cg->call_node()->jvms();
4572 while (p != nullptr) {
4573 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4574 p = p->caller();
4575 }
4576 log()->tail("late_inline");
4577 }
4578 }
4579
4580 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4581 log_late_inline(cg);
4582 if (log() != nullptr) {
4583 log()->inline_fail(msg);
4584 }
4585 }
4586
4587 void Compile::log_inline_id(CallGenerator* cg) {
4588 if (log() != nullptr) {
4589 // The LogCompilation tool needs a unique way to identify late
4590 // inline call sites. This id must be unique for this call site in
4591 // this compilation. Try to have it unique across compilations as
4592 // well because it can be convenient when grepping through the log
4593 // file.
4594 // Distinguish OSR compilations from others in case CICountOSR is
4595 // on.
4596 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4597 cg->set_unique_id(id);
4598 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4599 }
4600 }
4601
4602 void Compile::log_inline_failure(const char* msg) {
4603 if (C->log() != nullptr) {
4604 C->log()->inline_fail(msg);
4605 }
4606 }
4607
4608
4609 // Dump inlining replay data to the stream.
4610 // Don't change thread state and acquire any locks.
4611 void Compile::dump_inline_data(outputStream* out) {
4612 InlineTree* inl_tree = ilt();
4613 if (inl_tree != nullptr) {
4614 out->print(" inline %d", inl_tree->count());
4615 inl_tree->dump_replay_data(out);
4616 }
4617 }
4618
4619 void Compile::dump_inline_data_reduced(outputStream* out) {
4620 assert(ReplayReduce, "");
4621
4622 InlineTree* inl_tree = ilt();
4623 if (inl_tree == nullptr) {
4624 return;
4625 }
4626 // Enable iterative replay file reduction
4627 // Output "compile" lines for depth 1 subtrees,
4628 // simulating that those trees were compiled
4629 // instead of inlined.
4630 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4631 InlineTree* sub = inl_tree->subtrees().at(i);
4632 if (sub->inline_level() != 1) {
4633 continue;
4634 }
4635
4636 ciMethod* method = sub->method();
4637 int entry_bci = -1;
4638 int comp_level = env()->task()->comp_level();
4639 out->print("compile ");
4640 method->dump_name_as_ascii(out);
4641 out->print(" %d %d", entry_bci, comp_level);
4642 out->print(" inline %d", sub->count());
4643 sub->dump_replay_data(out, -1);
4644 out->cr();
4645 }
4646 }
4647
4648 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4649 if (n1->Opcode() < n2->Opcode()) return -1;
4650 else if (n1->Opcode() > n2->Opcode()) return 1;
4651
4652 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4653 for (uint i = 1; i < n1->req(); i++) {
4654 if (n1->in(i) < n2->in(i)) return -1;
4655 else if (n1->in(i) > n2->in(i)) return 1;
4656 }
4657
4658 return 0;
4659 }
4660
4661 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4662 Node* n1 = *n1p;
4663 Node* n2 = *n2p;
4664
4665 return cmp_expensive_nodes(n1, n2);
4666 }
4667
4668 void Compile::sort_expensive_nodes() {
4669 if (!expensive_nodes_sorted()) {
4670 _expensive_nodes.sort(cmp_expensive_nodes);
4671 }
4672 }
4673
4674 bool Compile::expensive_nodes_sorted() const {
4675 for (int i = 1; i < _expensive_nodes.length(); i++) {
4676 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4677 return false;
4678 }
4679 }
4680 return true;
4681 }
4682
4683 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4684 if (_expensive_nodes.length() == 0) {
4685 return false;
4686 }
4687
4688 assert(OptimizeExpensiveOps, "optimization off?");
4689
4690 // Take this opportunity to remove dead nodes from the list
4691 int j = 0;
4692 for (int i = 0; i < _expensive_nodes.length(); i++) {
4693 Node* n = _expensive_nodes.at(i);
4694 if (!n->is_unreachable(igvn)) {
4695 assert(n->is_expensive(), "should be expensive");
4696 _expensive_nodes.at_put(j, n);
4697 j++;
4698 }
4699 }
4700 _expensive_nodes.trunc_to(j);
4701
4702 // Then sort the list so that similar nodes are next to each other
4703 // and check for at least two nodes of identical kind with same data
4704 // inputs.
4705 sort_expensive_nodes();
4706
4707 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4708 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4709 return true;
4710 }
4711 }
4712
4713 return false;
4714 }
4715
4716 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4717 if (_expensive_nodes.length() == 0) {
4718 return;
4719 }
4720
4721 assert(OptimizeExpensiveOps, "optimization off?");
4722
4723 // Sort to bring similar nodes next to each other and clear the
4724 // control input of nodes for which there's only a single copy.
4725 sort_expensive_nodes();
4726
4727 int j = 0;
4728 int identical = 0;
4729 int i = 0;
4730 bool modified = false;
4731 for (; i < _expensive_nodes.length()-1; i++) {
4732 assert(j <= i, "can't write beyond current index");
4733 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4734 identical++;
4735 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4736 continue;
4737 }
4738 if (identical > 0) {
4739 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4740 identical = 0;
4741 } else {
4742 Node* n = _expensive_nodes.at(i);
4743 igvn.replace_input_of(n, 0, nullptr);
4744 igvn.hash_insert(n);
4745 modified = true;
4746 }
4747 }
4748 if (identical > 0) {
4749 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4750 } else if (_expensive_nodes.length() >= 1) {
4751 Node* n = _expensive_nodes.at(i);
4752 igvn.replace_input_of(n, 0, nullptr);
4753 igvn.hash_insert(n);
4754 modified = true;
4755 }
4756 _expensive_nodes.trunc_to(j);
4757 if (modified) {
4758 igvn.optimize();
4759 }
4760 }
4761
4762 void Compile::add_expensive_node(Node * n) {
4763 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4764 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4765 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4766 if (OptimizeExpensiveOps) {
4767 _expensive_nodes.append(n);
4768 } else {
4769 // Clear control input and let IGVN optimize expensive nodes if
4770 // OptimizeExpensiveOps is off.
4771 n->set_req(0, nullptr);
4772 }
4773 }
4774
4775 /**
4776 * Track coarsened Lock and Unlock nodes.
4777 */
4778
4779 class Lock_List : public Node_List {
4780 uint _origin_cnt;
4781 public:
4782 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4783 uint origin_cnt() const { return _origin_cnt; }
4784 };
4785
4786 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4787 int length = locks.length();
4788 if (length > 0) {
4789 // Have to keep this list until locks elimination during Macro nodes elimination.
4790 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4791 AbstractLockNode* alock = locks.at(0);
4792 BoxLockNode* box = alock->box_node()->as_BoxLock();
4793 for (int i = 0; i < length; i++) {
4794 AbstractLockNode* lock = locks.at(i);
4795 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4796 locks_list->push(lock);
4797 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4798 if (this_box != box) {
4799 // Locking regions (BoxLock) could be Unbalanced here:
4800 // - its coarsened locks were eliminated in earlier
4801 // macro nodes elimination followed by loop unroll
4802 // - it is OSR locking region (no Lock node)
4803 // Preserve Unbalanced status in such cases.
4804 if (!this_box->is_unbalanced()) {
4805 this_box->set_coarsened();
4806 }
4807 if (!box->is_unbalanced()) {
4808 box->set_coarsened();
4809 }
4810 }
4811 }
4812 _coarsened_locks.append(locks_list);
4813 }
4814 }
4815
4816 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4817 int count = coarsened_count();
4818 for (int i = 0; i < count; i++) {
4819 Node_List* locks_list = _coarsened_locks.at(i);
4820 for (uint j = 0; j < locks_list->size(); j++) {
4821 Node* lock = locks_list->at(j);
4822 assert(lock->is_AbstractLock(), "sanity");
4823 if (!useful.member(lock)) {
4824 locks_list->yank(lock);
4825 }
4826 }
4827 }
4828 }
4829
4830 void Compile::remove_coarsened_lock(Node* n) {
4831 if (n->is_AbstractLock()) {
4832 int count = coarsened_count();
4833 for (int i = 0; i < count; i++) {
4834 Node_List* locks_list = _coarsened_locks.at(i);
4835 locks_list->yank(n);
4836 }
4837 }
4838 }
4839
4840 bool Compile::coarsened_locks_consistent() {
4841 int count = coarsened_count();
4842 for (int i = 0; i < count; i++) {
4843 bool unbalanced = false;
4844 bool modified = false; // track locks kind modifications
4845 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4846 uint size = locks_list->size();
4847 if (size == 0) {
4848 unbalanced = false; // All locks were eliminated - good
4849 } else if (size != locks_list->origin_cnt()) {
4850 unbalanced = true; // Some locks were removed from list
4851 } else {
4852 for (uint j = 0; j < size; j++) {
4853 Node* lock = locks_list->at(j);
4854 // All nodes in group should have the same state (modified or not)
4855 if (!lock->as_AbstractLock()->is_coarsened()) {
4856 if (j == 0) {
4857 // first on list was modified, the rest should be too for consistency
4858 modified = true;
4859 } else if (!modified) {
4860 // this lock was modified but previous locks on the list were not
4861 unbalanced = true;
4862 break;
4863 }
4864 } else if (modified) {
4865 // previous locks on list were modified but not this lock
4866 unbalanced = true;
4867 break;
4868 }
4869 }
4870 }
4871 if (unbalanced) {
4872 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4873 #ifdef ASSERT
4874 if (PrintEliminateLocks) {
4875 tty->print_cr("=== unbalanced coarsened locks ===");
4876 for (uint l = 0; l < size; l++) {
4877 locks_list->at(l)->dump();
4878 }
4879 }
4880 #endif
4881 record_failure(C2Compiler::retry_no_locks_coarsening());
4882 return false;
4883 }
4884 }
4885 return true;
4886 }
4887
4888 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4889 // locks coarsening optimization removed Lock/Unlock nodes from them.
4890 // Such regions become unbalanced because coarsening only removes part
4891 // of Lock/Unlock nodes in region. As result we can't execute other
4892 // locks elimination optimizations which assume all code paths have
4893 // corresponding pair of Lock/Unlock nodes - they are balanced.
4894 void Compile::mark_unbalanced_boxes() const {
4895 int count = coarsened_count();
4896 for (int i = 0; i < count; i++) {
4897 Node_List* locks_list = _coarsened_locks.at(i);
4898 uint size = locks_list->size();
4899 if (size > 0) {
4900 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4901 BoxLockNode* box = alock->box_node()->as_BoxLock();
4902 if (alock->is_coarsened()) {
4903 // coarsened_locks_consistent(), which is called before this method, verifies
4904 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4905 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4906 for (uint j = 1; j < size; j++) {
4907 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4908 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4909 if (box != this_box) {
4910 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4911 box->set_unbalanced();
4912 this_box->set_unbalanced();
4913 }
4914 }
4915 }
4916 }
4917 }
4918 }
4919
4920 /**
4921 * Remove the speculative part of types and clean up the graph
4922 */
4923 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4924 if (UseTypeSpeculation) {
4925 Unique_Node_List worklist;
4926 worklist.push(root());
4927 int modified = 0;
4928 // Go over all type nodes that carry a speculative type, drop the
4929 // speculative part of the type and enqueue the node for an igvn
4930 // which may optimize it out.
4931 for (uint next = 0; next < worklist.size(); ++next) {
4932 Node *n = worklist.at(next);
4933 if (n->is_Type()) {
4934 TypeNode* tn = n->as_Type();
4935 const Type* t = tn->type();
4936 const Type* t_no_spec = t->remove_speculative();
4937 if (t_no_spec != t) {
4938 bool in_hash = igvn.hash_delete(n);
4939 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4940 tn->set_type(t_no_spec);
4941 igvn.hash_insert(n);
4942 igvn._worklist.push(n); // give it a chance to go away
4943 modified++;
4944 }
4945 }
4946 // Iterate over outs - endless loops is unreachable from below
4947 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4948 Node *m = n->fast_out(i);
4949 if (not_a_node(m)) {
4950 continue;
4951 }
4952 worklist.push(m);
4953 }
4954 }
4955 // Drop the speculative part of all types in the igvn's type table
4956 igvn.remove_speculative_types();
4957 if (modified > 0) {
4958 igvn.optimize();
4959 if (failing()) return;
4960 }
4961 #ifdef ASSERT
4962 // Verify that after the IGVN is over no speculative type has resurfaced
4963 worklist.clear();
4964 worklist.push(root());
4965 for (uint next = 0; next < worklist.size(); ++next) {
4966 Node *n = worklist.at(next);
4967 const Type* t = igvn.type_or_null(n);
4968 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4969 if (n->is_Type()) {
4970 t = n->as_Type()->type();
4971 assert(t == t->remove_speculative(), "no more speculative types");
4972 }
4973 // Iterate over outs - endless loops is unreachable from below
4974 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4975 Node *m = n->fast_out(i);
4976 if (not_a_node(m)) {
4977 continue;
4978 }
4979 worklist.push(m);
4980 }
4981 }
4982 igvn.check_no_speculative_types();
4983 #endif
4984 }
4985 }
4986
4987 // Auxiliary methods to support randomized stressing/fuzzing.
4988
4989 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
4990 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
4991 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
4992 FLAG_SET_ERGO(StressSeed, _stress_seed);
4993 } else {
4994 _stress_seed = StressSeed;
4995 }
4996 if (_log != nullptr) {
4997 _log->elem("stress_test seed='%u'", _stress_seed);
4998 }
4999 }
5000
5001 int Compile::random() {
5002 _stress_seed = os::next_random(_stress_seed);
5003 return static_cast<int>(_stress_seed);
5004 }
5005
5006 // This method can be called the arbitrary number of times, with current count
5007 // as the argument. The logic allows selecting a single candidate from the
5008 // running list of candidates as follows:
5009 // int count = 0;
5010 // Cand* selected = null;
5011 // while(cand = cand->next()) {
5012 // if (randomized_select(++count)) {
5013 // selected = cand;
5014 // }
5015 // }
5016 //
5017 // Including count equalizes the chances any candidate is "selected".
5018 // This is useful when we don't have the complete list of candidates to choose
5019 // from uniformly. In this case, we need to adjust the randomicity of the
5020 // selection, or else we will end up biasing the selection towards the latter
5021 // candidates.
5022 //
5023 // Quick back-envelope calculation shows that for the list of n candidates
5024 // the equal probability for the candidate to persist as "best" can be
5025 // achieved by replacing it with "next" k-th candidate with the probability
5026 // of 1/k. It can be easily shown that by the end of the run, the
5027 // probability for any candidate is converged to 1/n, thus giving the
5028 // uniform distribution among all the candidates.
5029 //
5030 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5031 #define RANDOMIZED_DOMAIN_POW 29
5032 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5033 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5034 bool Compile::randomized_select(int count) {
5035 assert(count > 0, "only positive");
5036 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5037 }
5038
5039 #ifdef ASSERT
5040 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5041 bool Compile::fail_randomly() {
5042 if ((random() % StressBailoutMean) != 0) {
5043 return false;
5044 }
5045 record_failure("StressBailout");
5046 return true;
5047 }
5048
5049 bool Compile::failure_is_artificial() {
5050 assert(failing_internal(), "should be failing");
5051 return C->failure_reason_is("StressBailout");
5052 }
5053 #endif
5054
5055 CloneMap& Compile::clone_map() { return _clone_map; }
5056 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5057
5058 void NodeCloneInfo::dump_on(outputStream* st) const {
5059 st->print(" {%d:%d} ", idx(), gen());
5060 }
5061
5062 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5063 uint64_t val = value(old->_idx);
5064 NodeCloneInfo cio(val);
5065 assert(val != 0, "old node should be in the map");
5066 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5067 insert(nnn->_idx, cin.get());
5068 #ifndef PRODUCT
5069 if (is_debug()) {
5070 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5071 }
5072 #endif
5073 }
5074
5075 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5076 NodeCloneInfo cio(value(old->_idx));
5077 if (cio.get() == 0) {
5078 cio.set(old->_idx, 0);
5079 insert(old->_idx, cio.get());
5080 #ifndef PRODUCT
5081 if (is_debug()) {
5082 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5083 }
5084 #endif
5085 }
5086 clone(old, nnn, gen);
5087 }
5088
5089 int CloneMap::max_gen() const {
5090 int g = 0;
5091 DictI di(_dict);
5092 for(; di.test(); ++di) {
5093 int t = gen(di._key);
5094 if (g < t) {
5095 g = t;
5096 #ifndef PRODUCT
5097 if (is_debug()) {
5098 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5099 }
5100 #endif
5101 }
5102 }
5103 return g;
5104 }
5105
5106 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5107 uint64_t val = value(key);
5108 if (val != 0) {
5109 NodeCloneInfo ni(val);
5110 ni.dump_on(st);
5111 }
5112 }
5113
5114 void Compile::shuffle_macro_nodes() {
5115 if (_macro_nodes.length() < 2) {
5116 return;
5117 }
5118 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5119 uint j = C->random() % (i + 1);
5120 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5121 }
5122 }
5123
5124 // Move Allocate nodes to the start of the list
5125 void Compile::sort_macro_nodes() {
5126 int count = macro_count();
5127 int allocates = 0;
5128 for (int i = 0; i < count; i++) {
5129 Node* n = macro_node(i);
5130 if (n->is_Allocate()) {
5131 if (i != allocates) {
5132 Node* tmp = macro_node(allocates);
5133 _macro_nodes.at_put(allocates, n);
5134 _macro_nodes.at_put(i, tmp);
5135 }
5136 allocates++;
5137 }
5138 }
5139 }
5140
5141 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5142 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5143 EventCompilerPhase event(UNTIMED);
5144 if (event.should_commit()) {
5145 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5146 }
5147 #ifndef PRODUCT
5148 ResourceMark rm;
5149 stringStream ss;
5150 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5151 int iter = ++_igv_phase_iter[cpt];
5152 if (iter > 1) {
5153 ss.print(" %d", iter);
5154 }
5155 if (n != nullptr) {
5156 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5157 if (n->is_Call()) {
5158 CallNode* call = n->as_Call();
5159 if (call->_name != nullptr) {
5160 // E.g. uncommon traps etc.
5161 ss.print(" - %s", call->_name);
5162 } else if (call->is_CallJava()) {
5163 CallJavaNode* call_java = call->as_CallJava();
5164 if (call_java->method() != nullptr) {
5165 ss.print(" -");
5166 call_java->method()->print_short_name(&ss);
5167 }
5168 }
5169 }
5170 }
5171
5172 const char* name = ss.as_string();
5173 if (should_print_igv(level)) {
5174 _igv_printer->print_method(name, level);
5175 }
5176 if (should_print_phase(cpt)) {
5177 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5178 }
5179 #endif
5180 C->_latest_stage_start_counter.stamp();
5181 }
5182
5183 // Only used from CompileWrapper
5184 void Compile::begin_method() {
5185 #ifndef PRODUCT
5186 if (_method != nullptr && should_print_igv(1)) {
5187 _igv_printer->begin_method();
5188 }
5189 #endif
5190 C->_latest_stage_start_counter.stamp();
5191 }
5192
5193 // Only used from CompileWrapper
5194 void Compile::end_method() {
5195 EventCompilerPhase event(UNTIMED);
5196 if (event.should_commit()) {
5197 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5198 }
5199
5200 #ifndef PRODUCT
5201 if (_method != nullptr && should_print_igv(1)) {
5202 _igv_printer->end_method();
5203 }
5204 #endif
5205 }
5206
5207 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5208 #ifndef PRODUCT
5209 if (_directive->should_print_phase(cpt)) {
5210 return true;
5211 }
5212 #endif
5213 return false;
5214 }
5215
5216 bool Compile::should_print_igv(const int level) {
5217 #ifndef PRODUCT
5218 if (PrintIdealGraphLevel < 0) { // disabled by the user
5219 return false;
5220 }
5221
5222 bool need = directive()->IGVPrintLevelOption >= level;
5223 if (need && !_igv_printer) {
5224 _igv_printer = IdealGraphPrinter::printer();
5225 _igv_printer->set_compile(this);
5226 }
5227 return need;
5228 #else
5229 return false;
5230 #endif
5231 }
5232
5233 #ifndef PRODUCT
5234 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5235 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5236
5237 // Called from debugger. Prints method to the default file with the default phase name.
5238 // This works regardless of any Ideal Graph Visualizer flags set or not.
5239 void igv_print() {
5240 Compile::current()->igv_print_method_to_file();
5241 }
5242
5243 // Same as igv_print() above but with a specified phase name.
5244 void igv_print(const char* phase_name) {
5245 Compile::current()->igv_print_method_to_file(phase_name);
5246 }
5247
5248 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5249 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5250 // This works regardless of any Ideal Graph Visualizer flags set or not.
5251 void igv_print(bool network) {
5252 if (network) {
5253 Compile::current()->igv_print_method_to_network();
5254 } else {
5255 Compile::current()->igv_print_method_to_file();
5256 }
5257 }
5258
5259 // Same as igv_print(bool network) above but with a specified phase name.
5260 void igv_print(bool network, const char* phase_name) {
5261 if (network) {
5262 Compile::current()->igv_print_method_to_network(phase_name);
5263 } else {
5264 Compile::current()->igv_print_method_to_file(phase_name);
5265 }
5266 }
5267
5268 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5269 void igv_print_default() {
5270 Compile::current()->print_method(PHASE_DEBUG, 0);
5271 }
5272
5273 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5274 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5275 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5276 void igv_append() {
5277 Compile::current()->igv_print_method_to_file("Debug", true);
5278 }
5279
5280 // Same as igv_append() above but with a specified phase name.
5281 void igv_append(const char* phase_name) {
5282 Compile::current()->igv_print_method_to_file(phase_name, true);
5283 }
5284
5285 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5286 const char* file_name = "custom_debug.xml";
5287 if (_debug_file_printer == nullptr) {
5288 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5289 } else {
5290 _debug_file_printer->update_compiled_method(C->method());
5291 }
5292 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5293 _debug_file_printer->print(phase_name, (Node*)C->root());
5294 }
5295
5296 void Compile::igv_print_method_to_network(const char* phase_name) {
5297 if (_debug_network_printer == nullptr) {
5298 _debug_network_printer = new IdealGraphPrinter(C);
5299 } else {
5300 _debug_network_printer->update_compiled_method(C->method());
5301 }
5302 tty->print_cr("Method printed over network stream to IGV");
5303 _debug_network_printer->print(phase_name, (Node*)C->root());
5304 }
5305 #endif
5306
5307 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5308 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5309 return value;
5310 }
5311 Node* result = nullptr;
5312 if (bt == T_BYTE) {
5313 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5314 result = new RShiftINode(result, phase->intcon(24));
5315 } else if (bt == T_BOOLEAN) {
5316 result = new AndINode(value, phase->intcon(0xFF));
5317 } else if (bt == T_CHAR) {
5318 result = new AndINode(value,phase->intcon(0xFFFF));
5319 } else {
5320 assert(bt == T_SHORT, "unexpected narrow type");
5321 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5322 result = new RShiftINode(result, phase->intcon(16));
5323 }
5324 if (transform_res) {
5325 result = phase->transform(result);
5326 }
5327 return result;
5328 }
5329
5330 void Compile::record_method_not_compilable_oom() {
5331 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5332 }