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 _stress_seed(0),
639 _comp_arena(mtCompiler),
640 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
641 _env(ci_env),
642 _directive(directive),
643 _log(ci_env->log()),
644 _first_failure_details(nullptr),
645 _intrinsics (comp_arena(), 0, 0, nullptr),
646 _macro_nodes (comp_arena(), 8, 0, nullptr),
647 _parse_predicates (comp_arena(), 8, 0, nullptr),
648 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
649 _expensive_nodes (comp_arena(), 8, 0, nullptr),
650 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
651 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
652 _coarsened_locks (comp_arena(), 8, 0, nullptr),
653 _congraph(nullptr),
654 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
655 _unique(0),
656 _dead_node_count(0),
657 _dead_node_list(comp_arena()),
658 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
659 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
660 _node_arena(&_node_arena_one),
661 _mach_constant_base_node(nullptr),
662 _Compile_types(mtCompiler),
663 _initial_gvn(nullptr),
664 _igvn_worklist(nullptr),
665 _types(nullptr),
666 _node_hash(nullptr),
667 _late_inlines(comp_arena(), 2, 0, nullptr),
668 _string_late_inlines(comp_arena(), 2, 0, nullptr),
669 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
670 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
671 _late_inlines_pos(0),
672 _number_of_mh_late_inlines(0),
673 _oom(false),
674 _print_inlining_stream(new (mtCompiler) stringStream()),
675 _print_inlining_list(nullptr),
676 _print_inlining_idx(0),
677 _print_inlining_output(nullptr),
678 _replay_inline_data(nullptr),
679 _java_calls(0),
680 _inner_loops(0),
681 _interpreter_frame_size(0),
682 _output(nullptr)
683 #ifndef PRODUCT
684 , _in_dump_cnt(0)
685 #endif
686 {
687 C = this;
688 CompileWrapper cw(this);
689
690 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
691 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
692
693 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
694 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
695 // We can always print a disassembly, either abstract (hex dump) or
696 // with the help of a suitable hsdis library. Thus, we should not
697 // couple print_assembly and print_opto_assembly controls.
698 // But: always print opto and regular assembly on compile command 'print'.
699 bool print_assembly = directive->PrintAssemblyOption;
700 set_print_assembly(print_opto_assembly || print_assembly);
701 #else
702 set_print_assembly(false); // must initialize.
703 #endif
704
705 #ifndef PRODUCT
706 set_parsed_irreducible_loop(false);
707 #endif
708
709 if (directive->ReplayInlineOption) {
710 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
711 }
712 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
713 set_print_intrinsics(directive->PrintIntrinsicsOption);
714 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
715
716 if (ProfileTraps) {
717 // Make sure the method being compiled gets its own MDO,
718 // so we can at least track the decompile_count().
719 method()->ensure_method_data();
720 }
721
722 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
723 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout) {
724 initialize_stress_seed(directive);
725 }
726
727 Init(/*do_aliasing=*/ true);
728
729 print_compile_messages();
730
731 _ilt = InlineTree::build_inline_tree_root();
732
733 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
734 assert(num_alias_types() >= AliasIdxRaw, "");
735
736 #define MINIMUM_NODE_HASH 1023
737
738 // GVN that will be run immediately on new nodes
739 uint estimated_size = method()->code_size()*4+64;
740 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
741 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
742 _types = new (comp_arena()) Type_Array(comp_arena());
743 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
744 PhaseGVN gvn;
745 set_initial_gvn(&gvn);
746
747 print_inlining_init();
748 { // Scope for timing the parser
749 TracePhase tp(_t_parser);
750
751 // Put top into the hash table ASAP.
752 initial_gvn()->transform(top());
753
754 // Set up tf(), start(), and find a CallGenerator.
755 CallGenerator* cg = nullptr;
756 if (is_osr_compilation()) {
757 const TypeTuple *domain = StartOSRNode::osr_domain();
758 const TypeTuple *range = TypeTuple::make_range(method()->signature());
759 init_tf(TypeFunc::make(domain, range));
760 StartNode* s = new StartOSRNode(root(), domain);
761 initial_gvn()->set_type_bottom(s);
762 verify_start(s);
763 cg = CallGenerator::for_osr(method(), entry_bci());
764 } else {
765 // Normal case.
766 init_tf(TypeFunc::make(method()));
767 StartNode* s = new StartNode(root(), tf()->domain());
768 initial_gvn()->set_type_bottom(s);
769 verify_start(s);
770 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
771 // With java.lang.ref.reference.get() we must go through the
772 // intrinsic - even when get() is the root
773 // method of the compile - so that, if necessary, the value in
774 // the referent field of the reference object gets recorded by
775 // the pre-barrier code.
776 cg = find_intrinsic(method(), false);
777 }
778 if (cg == nullptr) {
779 float past_uses = method()->interpreter_invocation_count();
780 float expected_uses = past_uses;
781 cg = CallGenerator::for_inline(method(), expected_uses);
782 }
783 }
784 if (failing()) return;
785 if (cg == nullptr) {
786 const char* reason = InlineTree::check_can_parse(method());
787 assert(reason != nullptr, "expect reason for parse failure");
788 stringStream ss;
789 ss.print("cannot parse method: %s", reason);
790 record_method_not_compilable(ss.as_string());
791 return;
792 }
793
794 gvn.set_type(root(), root()->bottom_type());
795
796 JVMState* jvms = build_start_state(start(), tf());
797 if ((jvms = cg->generate(jvms)) == nullptr) {
798 assert(failure_reason() != nullptr, "expect reason for parse failure");
799 stringStream ss;
800 ss.print("method parse failed: %s", failure_reason());
801 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
802 return;
803 }
804 GraphKit kit(jvms);
805
806 if (!kit.stopped()) {
807 // Accept return values, and transfer control we know not where.
808 // This is done by a special, unique ReturnNode bound to root.
809 return_values(kit.jvms());
810 }
811
812 if (kit.has_exceptions()) {
813 // Any exceptions that escape from this call must be rethrown
814 // to whatever caller is dynamically above us on the stack.
815 // This is done by a special, unique RethrowNode bound to root.
816 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
817 }
818
819 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
820
821 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
822 inline_string_calls(true);
823 }
824
825 if (failing()) return;
826
827 // Remove clutter produced by parsing.
828 if (!failing()) {
829 ResourceMark rm;
830 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
831 }
832 }
833
834 // Note: Large methods are capped off in do_one_bytecode().
835 if (failing()) return;
836
837 // After parsing, node notes are no longer automagic.
838 // They must be propagated by register_new_node_with_optimizer(),
839 // clone(), or the like.
840 set_default_node_notes(nullptr);
841
842 #ifndef PRODUCT
843 if (should_print_igv(1)) {
844 _igv_printer->print_inlining();
845 }
846 #endif
847
848 if (failing()) return;
849 NOT_PRODUCT( verify_graph_edges(); )
850
851 // Now optimize
852 Optimize();
853 if (failing()) return;
854 NOT_PRODUCT( verify_graph_edges(); )
855
856 #ifndef PRODUCT
857 if (should_print_ideal()) {
858 print_ideal_ir("print_ideal");
859 }
860 #endif
861
862 #ifdef ASSERT
863 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
864 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
865 #endif
866
867 // Dump compilation data to replay it.
868 if (directive->DumpReplayOption) {
869 env()->dump_replay_data(_compile_id);
870 }
871 if (directive->DumpInlineOption && (ilt() != nullptr)) {
872 env()->dump_inline_data(_compile_id);
873 }
874
875 // Now that we know the size of all the monitors we can add a fixed slot
876 // for the original deopt pc.
877 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
878 set_fixed_slots(next_slot);
879
880 // Compute when to use implicit null checks. Used by matching trap based
881 // nodes and NullCheck optimization.
882 set_allowed_deopt_reasons();
883
884 // Now generate code
885 Code_Gen();
886 }
887
888 //------------------------------Compile----------------------------------------
889 // Compile a runtime stub
890 Compile::Compile( ciEnv* ci_env,
891 TypeFunc_generator generator,
892 address stub_function,
893 const char *stub_name,
894 int is_fancy_jump,
895 bool pass_tls,
896 bool return_pc,
897 DirectiveSet* directive)
898 : Phase(Compiler),
899 _compile_id(0),
900 _options(Options::for_runtime_stub()),
901 _method(nullptr),
902 _entry_bci(InvocationEntryBci),
903 _stub_function(stub_function),
904 _stub_name(stub_name),
905 _stub_entry_point(nullptr),
906 _max_node_limit(MaxNodeLimit),
907 _post_loop_opts_phase(false),
908 _allow_macro_nodes(true),
909 _inlining_progress(false),
910 _inlining_incrementally(false),
911 _has_reserved_stack_access(false),
912 #ifndef PRODUCT
913 _igv_idx(0),
914 _trace_opto_output(directive->TraceOptoOutputOption),
915 #endif
916 _has_method_handle_invokes(false),
917 _clinit_barrier_on_entry(false),
918 _stress_seed(0),
919 _comp_arena(mtCompiler),
920 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
921 _env(ci_env),
922 _directive(directive),
923 _log(ci_env->log()),
924 _first_failure_details(nullptr),
925 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
926 _congraph(nullptr),
927 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
928 _unique(0),
929 _dead_node_count(0),
930 _dead_node_list(comp_arena()),
931 _node_arena_one(mtCompiler),
932 _node_arena_two(mtCompiler),
933 _node_arena(&_node_arena_one),
934 _mach_constant_base_node(nullptr),
935 _Compile_types(mtCompiler),
936 _initial_gvn(nullptr),
937 _igvn_worklist(nullptr),
938 _types(nullptr),
939 _node_hash(nullptr),
940 _number_of_mh_late_inlines(0),
941 _oom(false),
942 _print_inlining_stream(new (mtCompiler) stringStream()),
943 _print_inlining_list(nullptr),
944 _print_inlining_idx(0),
945 _print_inlining_output(nullptr),
946 _replay_inline_data(nullptr),
947 _java_calls(0),
948 _inner_loops(0),
949 _interpreter_frame_size(0),
950 _output(nullptr),
951 #ifndef PRODUCT
952 _in_dump_cnt(0),
953 #endif
954 _allowed_reasons(0) {
955 C = this;
956
957 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
958 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
959
960 #ifndef PRODUCT
961 set_print_assembly(PrintFrameConverterAssembly);
962 set_parsed_irreducible_loop(false);
963 #else
964 set_print_assembly(false); // Must initialize.
965 #endif
966 set_has_irreducible_loop(false); // no loops
967
968 CompileWrapper cw(this);
969 Init(/*do_aliasing=*/ false);
970 init_tf((*generator)());
971
972 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
973 _types = new (comp_arena()) Type_Array(comp_arena());
974 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
975
976 if (StressLCM || StressGCM || StressBailout) {
977 initialize_stress_seed(directive);
978 }
979
980 {
981 PhaseGVN gvn;
982 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
983 gvn.transform(top());
984
985 GraphKit kit;
986 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
987 }
988
989 NOT_PRODUCT( verify_graph_edges(); )
990
991 Code_Gen();
992 }
993
994 Compile::~Compile() {
995 delete _print_inlining_stream;
996 delete _first_failure_details;
997 };
998
999 //------------------------------Init-------------------------------------------
1000 // Prepare for a single compilation
1001 void Compile::Init(bool aliasing) {
1002 _do_aliasing = aliasing;
1003 _unique = 0;
1004 _regalloc = nullptr;
1005
1006 _tf = nullptr; // filled in later
1007 _top = nullptr; // cached later
1008 _matcher = nullptr; // filled in later
1009 _cfg = nullptr; // filled in later
1010
1011 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1012
1013 _node_note_array = nullptr;
1014 _default_node_notes = nullptr;
1015 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1016
1017 _immutable_memory = nullptr; // filled in at first inquiry
1018
1019 #ifdef ASSERT
1020 _phase_optimize_finished = false;
1021 _phase_verify_ideal_loop = false;
1022 _exception_backedge = false;
1023 _type_verify = nullptr;
1024 #endif
1025
1026 // Globally visible Nodes
1027 // First set TOP to null to give safe behavior during creation of RootNode
1028 set_cached_top_node(nullptr);
1029 set_root(new RootNode());
1030 // Now that you have a Root to point to, create the real TOP
1031 set_cached_top_node( new ConNode(Type::TOP) );
1032 set_recent_alloc(nullptr, nullptr);
1033
1034 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1035 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1036 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1037 env()->set_dependencies(new Dependencies(env()));
1038
1039 _fixed_slots = 0;
1040 set_has_split_ifs(false);
1041 set_has_loops(false); // first approximation
1042 set_has_stringbuilder(false);
1043 set_has_boxed_value(false);
1044 _trap_can_recompile = false; // no traps emitted yet
1045 _major_progress = true; // start out assuming good things will happen
1046 set_has_unsafe_access(false);
1047 set_max_vector_size(0);
1048 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1049 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1050 set_decompile_count(0);
1051
1052 #ifndef PRODUCT
1053 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1054 #endif
1055
1056 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1057 _loop_opts_cnt = LoopOptsCount;
1058 set_do_inlining(Inline);
1059 set_max_inline_size(MaxInlineSize);
1060 set_freq_inline_size(FreqInlineSize);
1061 set_do_scheduling(OptoScheduling);
1062
1063 set_do_vector_loop(false);
1064 set_has_monitors(false);
1065 set_has_scoped_access(false);
1066
1067 if (AllowVectorizeOnDemand) {
1068 if (has_method() && _directive->VectorizeOption) {
1069 set_do_vector_loop(true);
1070 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());})
1071 } else if (has_method() && method()->name() != nullptr &&
1072 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1073 set_do_vector_loop(true);
1074 }
1075 }
1076 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1077 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());})
1078
1079 _max_node_limit = _directive->MaxNodeLimitOption;
1080
1081 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1082 set_clinit_barrier_on_entry(true);
1083 }
1084 if (debug_info()->recording_non_safepoints()) {
1085 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1086 (comp_arena(), 8, 0, nullptr));
1087 set_default_node_notes(Node_Notes::make(this));
1088 }
1089
1090 const int grow_ats = 16;
1091 _max_alias_types = grow_ats;
1092 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1093 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1094 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1095 {
1096 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1097 }
1098 // Initialize the first few types.
1099 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1100 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1101 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1102 _num_alias_types = AliasIdxRaw+1;
1103 // Zero out the alias type cache.
1104 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1105 // A null adr_type hits in the cache right away. Preload the right answer.
1106 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1107 }
1108
1109 #ifdef ASSERT
1110 // Verify that the current StartNode is valid.
1111 void Compile::verify_start(StartNode* s) const {
1112 assert(failing_internal() || s == start(), "should be StartNode");
1113 }
1114 #endif
1115
1116 /**
1117 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1118 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1119 * the ideal graph.
1120 */
1121 StartNode* Compile::start() const {
1122 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1123 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1124 Node* start = root()->fast_out(i);
1125 if (start->is_Start()) {
1126 return start->as_Start();
1127 }
1128 }
1129 fatal("Did not find Start node!");
1130 return nullptr;
1131 }
1132
1133 //-------------------------------immutable_memory-------------------------------------
1134 // Access immutable memory
1135 Node* Compile::immutable_memory() {
1136 if (_immutable_memory != nullptr) {
1137 return _immutable_memory;
1138 }
1139 StartNode* s = start();
1140 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1141 Node *p = s->fast_out(i);
1142 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1143 _immutable_memory = p;
1144 return _immutable_memory;
1145 }
1146 }
1147 ShouldNotReachHere();
1148 return nullptr;
1149 }
1150
1151 //----------------------set_cached_top_node------------------------------------
1152 // Install the cached top node, and make sure Node::is_top works correctly.
1153 void Compile::set_cached_top_node(Node* tn) {
1154 if (tn != nullptr) verify_top(tn);
1155 Node* old_top = _top;
1156 _top = tn;
1157 // Calling Node::setup_is_top allows the nodes the chance to adjust
1158 // their _out arrays.
1159 if (_top != nullptr) _top->setup_is_top();
1160 if (old_top != nullptr) old_top->setup_is_top();
1161 assert(_top == nullptr || top()->is_top(), "");
1162 }
1163
1164 #ifdef ASSERT
1165 uint Compile::count_live_nodes_by_graph_walk() {
1166 Unique_Node_List useful(comp_arena());
1167 // Get useful node list by walking the graph.
1168 identify_useful_nodes(useful);
1169 return useful.size();
1170 }
1171
1172 void Compile::print_missing_nodes() {
1173
1174 // Return if CompileLog is null and PrintIdealNodeCount is false.
1175 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1176 return;
1177 }
1178
1179 // This is an expensive function. It is executed only when the user
1180 // specifies VerifyIdealNodeCount option or otherwise knows the
1181 // additional work that needs to be done to identify reachable nodes
1182 // by walking the flow graph and find the missing ones using
1183 // _dead_node_list.
1184
1185 Unique_Node_List useful(comp_arena());
1186 // Get useful node list by walking the graph.
1187 identify_useful_nodes(useful);
1188
1189 uint l_nodes = C->live_nodes();
1190 uint l_nodes_by_walk = useful.size();
1191
1192 if (l_nodes != l_nodes_by_walk) {
1193 if (_log != nullptr) {
1194 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1195 _log->stamp();
1196 _log->end_head();
1197 }
1198 VectorSet& useful_member_set = useful.member_set();
1199 int last_idx = l_nodes_by_walk;
1200 for (int i = 0; i < last_idx; i++) {
1201 if (useful_member_set.test(i)) {
1202 if (_dead_node_list.test(i)) {
1203 if (_log != nullptr) {
1204 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1205 }
1206 if (PrintIdealNodeCount) {
1207 // Print the log message to tty
1208 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1209 useful.at(i)->dump();
1210 }
1211 }
1212 }
1213 else if (! _dead_node_list.test(i)) {
1214 if (_log != nullptr) {
1215 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1216 }
1217 if (PrintIdealNodeCount) {
1218 // Print the log message to tty
1219 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1220 }
1221 }
1222 }
1223 if (_log != nullptr) {
1224 _log->tail("mismatched_nodes");
1225 }
1226 }
1227 }
1228 void Compile::record_modified_node(Node* n) {
1229 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1230 _modified_nodes->push(n);
1231 }
1232 }
1233
1234 void Compile::remove_modified_node(Node* n) {
1235 if (_modified_nodes != nullptr) {
1236 _modified_nodes->remove(n);
1237 }
1238 }
1239 #endif
1240
1241 #ifndef PRODUCT
1242 void Compile::verify_top(Node* tn) const {
1243 if (tn != nullptr) {
1244 assert(tn->is_Con(), "top node must be a constant");
1245 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1246 assert(tn->in(0) != nullptr, "must have live top node");
1247 }
1248 }
1249 #endif
1250
1251
1252 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1253
1254 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1255 guarantee(arr != nullptr, "");
1256 int num_blocks = arr->length();
1257 if (grow_by < num_blocks) grow_by = num_blocks;
1258 int num_notes = grow_by * _node_notes_block_size;
1259 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1260 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1261 while (num_notes > 0) {
1262 arr->append(notes);
1263 notes += _node_notes_block_size;
1264 num_notes -= _node_notes_block_size;
1265 }
1266 assert(num_notes == 0, "exact multiple, please");
1267 }
1268
1269 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1270 if (source == nullptr || dest == nullptr) return false;
1271
1272 if (dest->is_Con())
1273 return false; // Do not push debug info onto constants.
1274
1275 #ifdef ASSERT
1276 // Leave a bread crumb trail pointing to the original node:
1277 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1278 dest->set_debug_orig(source);
1279 }
1280 #endif
1281
1282 if (node_note_array() == nullptr)
1283 return false; // Not collecting any notes now.
1284
1285 // This is a copy onto a pre-existing node, which may already have notes.
1286 // If both nodes have notes, do not overwrite any pre-existing notes.
1287 Node_Notes* source_notes = node_notes_at(source->_idx);
1288 if (source_notes == nullptr || source_notes->is_clear()) return false;
1289 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1290 if (dest_notes == nullptr || dest_notes->is_clear()) {
1291 return set_node_notes_at(dest->_idx, source_notes);
1292 }
1293
1294 Node_Notes merged_notes = (*source_notes);
1295 // The order of operations here ensures that dest notes will win...
1296 merged_notes.update_from(dest_notes);
1297 return set_node_notes_at(dest->_idx, &merged_notes);
1298 }
1299
1300
1301 //--------------------------allow_range_check_smearing-------------------------
1302 // Gating condition for coalescing similar range checks.
1303 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1304 // single covering check that is at least as strong as any of them.
1305 // If the optimization succeeds, the simplified (strengthened) range check
1306 // will always succeed. If it fails, we will deopt, and then give up
1307 // on the optimization.
1308 bool Compile::allow_range_check_smearing() const {
1309 // If this method has already thrown a range-check,
1310 // assume it was because we already tried range smearing
1311 // and it failed.
1312 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1313 return !already_trapped;
1314 }
1315
1316
1317 //------------------------------flatten_alias_type-----------------------------
1318 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1319 assert(do_aliasing(), "Aliasing should be enabled");
1320 int offset = tj->offset();
1321 TypePtr::PTR ptr = tj->ptr();
1322
1323 // Known instance (scalarizable allocation) alias only with itself.
1324 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1325 tj->is_oopptr()->is_known_instance();
1326
1327 // Process weird unsafe references.
1328 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1329 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1330 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1331 tj = TypeOopPtr::BOTTOM;
1332 ptr = tj->ptr();
1333 offset = tj->offset();
1334 }
1335
1336 // Array pointers need some flattening
1337 const TypeAryPtr* ta = tj->isa_aryptr();
1338 if (ta && ta->is_stable()) {
1339 // Erase stability property for alias analysis.
1340 tj = ta = ta->cast_to_stable(false);
1341 }
1342 if( ta && is_known_inst ) {
1343 if ( offset != Type::OffsetBot &&
1344 offset > arrayOopDesc::length_offset_in_bytes() ) {
1345 offset = Type::OffsetBot; // Flatten constant access into array body only
1346 tj = ta = ta->
1347 remove_speculative()->
1348 cast_to_ptr_type(ptr)->
1349 with_offset(offset);
1350 }
1351 } else if (ta) {
1352 // For arrays indexed by constant indices, we flatten the alias
1353 // space to include all of the array body. Only the header, klass
1354 // and array length can be accessed un-aliased.
1355 if( offset != Type::OffsetBot ) {
1356 if( ta->const_oop() ) { // MethodData* or Method*
1357 offset = Type::OffsetBot; // Flatten constant access into array body
1358 tj = ta = ta->
1359 remove_speculative()->
1360 cast_to_ptr_type(ptr)->
1361 cast_to_exactness(false)->
1362 with_offset(offset);
1363 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1364 // range is OK as-is.
1365 tj = ta = TypeAryPtr::RANGE;
1366 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1367 tj = TypeInstPtr::KLASS; // all klass loads look alike
1368 ta = TypeAryPtr::RANGE; // generic ignored junk
1369 ptr = TypePtr::BotPTR;
1370 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1371 tj = TypeInstPtr::MARK;
1372 ta = TypeAryPtr::RANGE; // generic ignored junk
1373 ptr = TypePtr::BotPTR;
1374 } else { // Random constant offset into array body
1375 offset = Type::OffsetBot; // Flatten constant access into array body
1376 tj = ta = ta->
1377 remove_speculative()->
1378 cast_to_ptr_type(ptr)->
1379 cast_to_exactness(false)->
1380 with_offset(offset);
1381 }
1382 }
1383 // Arrays of fixed size alias with arrays of unknown size.
1384 if (ta->size() != TypeInt::POS) {
1385 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1386 tj = ta = ta->
1387 remove_speculative()->
1388 cast_to_ptr_type(ptr)->
1389 with_ary(tary)->
1390 cast_to_exactness(false);
1391 }
1392 // Arrays of known objects become arrays of unknown objects.
1393 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1394 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1395 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1396 }
1397 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1398 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1399 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1400 }
1401 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1402 // cannot be distinguished by bytecode alone.
1403 if (ta->elem() == TypeInt::BOOL) {
1404 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1405 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1406 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1407 }
1408 // During the 2nd round of IterGVN, NotNull castings are removed.
1409 // Make sure the Bottom and NotNull variants alias the same.
1410 // Also, make sure exact and non-exact variants alias the same.
1411 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1412 tj = ta = ta->
1413 remove_speculative()->
1414 cast_to_ptr_type(TypePtr::BotPTR)->
1415 cast_to_exactness(false)->
1416 with_offset(offset);
1417 }
1418 }
1419
1420 // Oop pointers need some flattening
1421 const TypeInstPtr *to = tj->isa_instptr();
1422 if (to && to != TypeOopPtr::BOTTOM) {
1423 ciInstanceKlass* ik = to->instance_klass();
1424 if( ptr == TypePtr::Constant ) {
1425 if (ik != ciEnv::current()->Class_klass() ||
1426 offset < ik->layout_helper_size_in_bytes()) {
1427 // No constant oop pointers (such as Strings); they alias with
1428 // unknown strings.
1429 assert(!is_known_inst, "not scalarizable allocation");
1430 tj = to = to->
1431 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1432 remove_speculative()->
1433 cast_to_ptr_type(TypePtr::BotPTR)->
1434 cast_to_exactness(false);
1435 }
1436 } else if( is_known_inst ) {
1437 tj = to; // Keep NotNull and klass_is_exact for instance type
1438 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1439 // During the 2nd round of IterGVN, NotNull castings are removed.
1440 // Make sure the Bottom and NotNull variants alias the same.
1441 // Also, make sure exact and non-exact variants alias the same.
1442 tj = to = to->
1443 remove_speculative()->
1444 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1445 cast_to_ptr_type(TypePtr::BotPTR)->
1446 cast_to_exactness(false);
1447 }
1448 if (to->speculative() != nullptr) {
1449 tj = to = to->remove_speculative();
1450 }
1451 // Canonicalize the holder of this field
1452 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1453 // First handle header references such as a LoadKlassNode, even if the
1454 // object's klass is unloaded at compile time (4965979).
1455 if (!is_known_inst) { // Do it only for non-instance types
1456 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1457 }
1458 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1459 // Static fields are in the space above the normal instance
1460 // fields in the java.lang.Class instance.
1461 if (ik != ciEnv::current()->Class_klass()) {
1462 to = nullptr;
1463 tj = TypeOopPtr::BOTTOM;
1464 offset = tj->offset();
1465 }
1466 } else {
1467 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1468 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1469 assert(tj->offset() == offset, "no change to offset expected");
1470 bool xk = to->klass_is_exact();
1471 int instance_id = to->instance_id();
1472
1473 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1474 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1475 // its interfaces are included.
1476 if (xk && ik->equals(canonical_holder)) {
1477 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1478 } else {
1479 assert(xk || !is_known_inst, "Known instance should be exact type");
1480 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1481 }
1482 }
1483 }
1484
1485 // Klass pointers to object array klasses need some flattening
1486 const TypeKlassPtr *tk = tj->isa_klassptr();
1487 if( tk ) {
1488 // If we are referencing a field within a Klass, we need
1489 // to assume the worst case of an Object. Both exact and
1490 // inexact types must flatten to the same alias class so
1491 // use NotNull as the PTR.
1492 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1493 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1494 env()->Object_klass(),
1495 offset);
1496 }
1497
1498 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1499 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1500 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1501 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1502 } else {
1503 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1504 }
1505 }
1506
1507 // Check for precise loads from the primary supertype array and force them
1508 // to the supertype cache alias index. Check for generic array loads from
1509 // the primary supertype array and also force them to the supertype cache
1510 // alias index. Since the same load can reach both, we need to merge
1511 // these 2 disparate memories into the same alias class. Since the
1512 // primary supertype array is read-only, there's no chance of confusion
1513 // where we bypass an array load and an array store.
1514 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1515 if (offset == Type::OffsetBot ||
1516 (offset >= primary_supers_offset &&
1517 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1518 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1519 offset = in_bytes(Klass::secondary_super_cache_offset());
1520 tj = tk = tk->with_offset(offset);
1521 }
1522 }
1523
1524 // Flatten all Raw pointers together.
1525 if (tj->base() == Type::RawPtr)
1526 tj = TypeRawPtr::BOTTOM;
1527
1528 if (tj->base() == Type::AnyPtr)
1529 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1530
1531 offset = tj->offset();
1532 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1533
1534 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1535 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1536 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1537 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1538 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1539 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1540 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1541 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1542 assert( tj->ptr() != TypePtr::TopPTR &&
1543 tj->ptr() != TypePtr::AnyNull &&
1544 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1545 // assert( tj->ptr() != TypePtr::Constant ||
1546 // tj->base() == Type::RawPtr ||
1547 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1548
1549 return tj;
1550 }
1551
1552 void Compile::AliasType::Init(int i, const TypePtr* at) {
1553 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1554 _index = i;
1555 _adr_type = at;
1556 _field = nullptr;
1557 _element = nullptr;
1558 _is_rewritable = true; // default
1559 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1560 if (atoop != nullptr && atoop->is_known_instance()) {
1561 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1562 _general_index = Compile::current()->get_alias_index(gt);
1563 } else {
1564 _general_index = 0;
1565 }
1566 }
1567
1568 BasicType Compile::AliasType::basic_type() const {
1569 if (element() != nullptr) {
1570 const Type* element = adr_type()->is_aryptr()->elem();
1571 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1572 } if (field() != nullptr) {
1573 return field()->layout_type();
1574 } else {
1575 return T_ILLEGAL; // unknown
1576 }
1577 }
1578
1579 //---------------------------------print_on------------------------------------
1580 #ifndef PRODUCT
1581 void Compile::AliasType::print_on(outputStream* st) {
1582 if (index() < 10)
1583 st->print("@ <%d> ", index());
1584 else st->print("@ <%d>", index());
1585 st->print(is_rewritable() ? " " : " RO");
1586 int offset = adr_type()->offset();
1587 if (offset == Type::OffsetBot)
1588 st->print(" +any");
1589 else st->print(" +%-3d", offset);
1590 st->print(" in ");
1591 adr_type()->dump_on(st);
1592 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1593 if (field() != nullptr && tjp) {
1594 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1595 tjp->offset() != field()->offset_in_bytes()) {
1596 st->print(" != ");
1597 field()->print();
1598 st->print(" ***");
1599 }
1600 }
1601 }
1602
1603 void print_alias_types() {
1604 Compile* C = Compile::current();
1605 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1606 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1607 C->alias_type(idx)->print_on(tty);
1608 tty->cr();
1609 }
1610 }
1611 #endif
1612
1613
1614 //----------------------------probe_alias_cache--------------------------------
1615 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1616 intptr_t key = (intptr_t) adr_type;
1617 key ^= key >> logAliasCacheSize;
1618 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1619 }
1620
1621
1622 //-----------------------------grow_alias_types--------------------------------
1623 void Compile::grow_alias_types() {
1624 const int old_ats = _max_alias_types; // how many before?
1625 const int new_ats = old_ats; // how many more?
1626 const int grow_ats = old_ats+new_ats; // how many now?
1627 _max_alias_types = grow_ats;
1628 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1629 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1630 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1631 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1632 }
1633
1634
1635 //--------------------------------find_alias_type------------------------------
1636 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1637 if (!do_aliasing()) {
1638 return alias_type(AliasIdxBot);
1639 }
1640
1641 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1642 if (ace->_adr_type == adr_type) {
1643 return alias_type(ace->_index);
1644 }
1645
1646 // Handle special cases.
1647 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1648 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1649
1650 // Do it the slow way.
1651 const TypePtr* flat = flatten_alias_type(adr_type);
1652
1653 #ifdef ASSERT
1654 {
1655 ResourceMark rm;
1656 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1657 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1658 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1659 Type::str(adr_type));
1660 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1661 const TypeOopPtr* foop = flat->is_oopptr();
1662 // Scalarizable allocations have exact klass always.
1663 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1664 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1665 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1666 Type::str(foop), Type::str(xoop));
1667 }
1668 }
1669 #endif
1670
1671 int idx = AliasIdxTop;
1672 for (int i = 0; i < num_alias_types(); i++) {
1673 if (alias_type(i)->adr_type() == flat) {
1674 idx = i;
1675 break;
1676 }
1677 }
1678
1679 if (idx == AliasIdxTop) {
1680 if (no_create) return nullptr;
1681 // Grow the array if necessary.
1682 if (_num_alias_types == _max_alias_types) grow_alias_types();
1683 // Add a new alias type.
1684 idx = _num_alias_types++;
1685 _alias_types[idx]->Init(idx, flat);
1686 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1687 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1688 if (flat->isa_instptr()) {
1689 if (flat->offset() == java_lang_Class::klass_offset()
1690 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1691 alias_type(idx)->set_rewritable(false);
1692 }
1693 if (flat->isa_aryptr()) {
1694 #ifdef ASSERT
1695 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1696 // (T_BYTE has the weakest alignment and size restrictions...)
1697 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1698 #endif
1699 if (flat->offset() == TypePtr::OffsetBot) {
1700 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1701 }
1702 }
1703 if (flat->isa_klassptr()) {
1704 if (UseCompactObjectHeaders) {
1705 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1706 alias_type(idx)->set_rewritable(false);
1707 }
1708 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1709 alias_type(idx)->set_rewritable(false);
1710 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1711 alias_type(idx)->set_rewritable(false);
1712 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1713 alias_type(idx)->set_rewritable(false);
1714 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1715 alias_type(idx)->set_rewritable(false);
1716 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1717 alias_type(idx)->set_rewritable(false);
1718 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1719 alias_type(idx)->set_rewritable(false);
1720 }
1721 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1722 // but the base pointer type is not distinctive enough to identify
1723 // references into JavaThread.)
1724
1725 // Check for final fields.
1726 const TypeInstPtr* tinst = flat->isa_instptr();
1727 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1728 ciField* field;
1729 if (tinst->const_oop() != nullptr &&
1730 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1731 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1732 // static field
1733 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1734 field = k->get_field_by_offset(tinst->offset(), true);
1735 } else {
1736 ciInstanceKlass *k = tinst->instance_klass();
1737 field = k->get_field_by_offset(tinst->offset(), false);
1738 }
1739 assert(field == nullptr ||
1740 original_field == nullptr ||
1741 (field->holder() == original_field->holder() &&
1742 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1743 field->is_static() == original_field->is_static()), "wrong field?");
1744 // Set field() and is_rewritable() attributes.
1745 if (field != nullptr) alias_type(idx)->set_field(field);
1746 }
1747 }
1748
1749 // Fill the cache for next time.
1750 ace->_adr_type = adr_type;
1751 ace->_index = idx;
1752 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1753
1754 // Might as well try to fill the cache for the flattened version, too.
1755 AliasCacheEntry* face = probe_alias_cache(flat);
1756 if (face->_adr_type == nullptr) {
1757 face->_adr_type = flat;
1758 face->_index = idx;
1759 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1760 }
1761
1762 return alias_type(idx);
1763 }
1764
1765
1766 Compile::AliasType* Compile::alias_type(ciField* field) {
1767 const TypeOopPtr* t;
1768 if (field->is_static())
1769 t = TypeInstPtr::make(field->holder()->java_mirror());
1770 else
1771 t = TypeOopPtr::make_from_klass_raw(field->holder());
1772 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1773 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1774 return atp;
1775 }
1776
1777
1778 //------------------------------have_alias_type--------------------------------
1779 bool Compile::have_alias_type(const TypePtr* adr_type) {
1780 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1781 if (ace->_adr_type == adr_type) {
1782 return true;
1783 }
1784
1785 // Handle special cases.
1786 if (adr_type == nullptr) return true;
1787 if (adr_type == TypePtr::BOTTOM) return true;
1788
1789 return find_alias_type(adr_type, true, nullptr) != nullptr;
1790 }
1791
1792 //-----------------------------must_alias--------------------------------------
1793 // True if all values of the given address type are in the given alias category.
1794 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1795 if (alias_idx == AliasIdxBot) return true; // the universal category
1796 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1797 if (alias_idx == AliasIdxTop) return false; // the empty category
1798 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1799
1800 // the only remaining possible overlap is identity
1801 int adr_idx = get_alias_index(adr_type);
1802 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1803 assert(adr_idx == alias_idx ||
1804 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1805 && adr_type != TypeOopPtr::BOTTOM),
1806 "should not be testing for overlap with an unsafe pointer");
1807 return adr_idx == alias_idx;
1808 }
1809
1810 //------------------------------can_alias--------------------------------------
1811 // True if any values of the given address type are in the given alias category.
1812 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1813 if (alias_idx == AliasIdxTop) return false; // the empty category
1814 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1815 // Known instance doesn't alias with bottom memory
1816 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1817 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1818
1819 // the only remaining possible overlap is identity
1820 int adr_idx = get_alias_index(adr_type);
1821 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1822 return adr_idx == alias_idx;
1823 }
1824
1825 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1826 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1827 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1828 if (parse_predicate_count() == 0) {
1829 return;
1830 }
1831 for (int i = 0; i < parse_predicate_count(); i++) {
1832 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1833 parse_predicate->mark_useless();
1834 igvn._worklist.push(parse_predicate);
1835 }
1836 _parse_predicates.clear();
1837 }
1838
1839 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1840 if (!n->for_post_loop_opts_igvn()) {
1841 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1842 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1843 _for_post_loop_igvn.append(n);
1844 }
1845 }
1846
1847 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1848 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1849 _for_post_loop_igvn.remove(n);
1850 }
1851
1852 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1853 // Verify that all previous optimizations produced a valid graph
1854 // at least to this point, even if no loop optimizations were done.
1855 PhaseIdealLoop::verify(igvn);
1856
1857 C->set_post_loop_opts_phase(); // no more loop opts allowed
1858
1859 assert(!C->major_progress(), "not cleared");
1860
1861 if (_for_post_loop_igvn.length() > 0) {
1862 while (_for_post_loop_igvn.length() > 0) {
1863 Node* n = _for_post_loop_igvn.pop();
1864 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1865 igvn._worklist.push(n);
1866 }
1867 igvn.optimize();
1868 if (failing()) return;
1869 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1870 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1871
1872 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1873 if (C->major_progress()) {
1874 C->clear_major_progress(); // ensure that major progress is now clear
1875 }
1876 }
1877 }
1878
1879 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1880 if (OptimizeUnstableIf) {
1881 _unstable_if_traps.append(trap);
1882 }
1883 }
1884
1885 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1886 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1887 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1888 Node* n = trap->uncommon_trap();
1889 if (!useful.member(n)) {
1890 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1891 }
1892 }
1893 }
1894
1895 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1896 // or fold-compares case. Return true if succeed or not found.
1897 //
1898 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1899 // false and ask fold-compares to yield.
1900 //
1901 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1902 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1903 // when deoptimization does happen.
1904 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1905 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1906 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1907 if (trap->uncommon_trap() == unc) {
1908 if (yield && trap->modified()) {
1909 return false;
1910 }
1911 _unstable_if_traps.delete_at(i);
1912 break;
1913 }
1914 }
1915 return true;
1916 }
1917
1918 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1919 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1920 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1921 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1922 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1923 CallStaticJavaNode* unc = trap->uncommon_trap();
1924 int next_bci = trap->next_bci();
1925 bool modified = trap->modified();
1926
1927 if (next_bci != -1 && !modified) {
1928 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1929 JVMState* jvms = unc->jvms();
1930 ciMethod* method = jvms->method();
1931 ciBytecodeStream iter(method);
1932
1933 iter.force_bci(jvms->bci());
1934 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1935 Bytecodes::Code c = iter.cur_bc();
1936 Node* lhs = nullptr;
1937 Node* rhs = nullptr;
1938 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1939 lhs = unc->peek_operand(0);
1940 rhs = unc->peek_operand(1);
1941 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1942 lhs = unc->peek_operand(0);
1943 }
1944
1945 ResourceMark rm;
1946 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1947 assert(live_locals.is_valid(), "broken liveness info");
1948 int len = (int)live_locals.size();
1949
1950 for (int i = 0; i < len; i++) {
1951 Node* local = unc->local(jvms, i);
1952 // kill local using the liveness of next_bci.
1953 // give up when the local looks like an operand to secure reexecution.
1954 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1955 uint idx = jvms->locoff() + i;
1956 #ifdef ASSERT
1957 if (PrintOpto && Verbose) {
1958 tty->print("[unstable_if] kill local#%d: ", idx);
1959 local->dump();
1960 tty->cr();
1961 }
1962 #endif
1963 igvn.replace_input_of(unc, idx, top());
1964 modified = true;
1965 }
1966 }
1967 }
1968
1969 // keep the mondified trap for late query
1970 if (modified) {
1971 trap->set_modified();
1972 } else {
1973 _unstable_if_traps.delete_at(i);
1974 }
1975 }
1976 igvn.optimize();
1977 }
1978
1979 // StringOpts and late inlining of string methods
1980 void Compile::inline_string_calls(bool parse_time) {
1981 {
1982 // remove useless nodes to make the usage analysis simpler
1983 ResourceMark rm;
1984 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1985 }
1986
1987 {
1988 ResourceMark rm;
1989 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1990 PhaseStringOpts pso(initial_gvn());
1991 print_method(PHASE_AFTER_STRINGOPTS, 3);
1992 }
1993
1994 // now inline anything that we skipped the first time around
1995 if (!parse_time) {
1996 _late_inlines_pos = _late_inlines.length();
1997 }
1998
1999 while (_string_late_inlines.length() > 0) {
2000 CallGenerator* cg = _string_late_inlines.pop();
2001 cg->do_late_inline();
2002 if (failing()) return;
2003 }
2004 _string_late_inlines.trunc_to(0);
2005 }
2006
2007 // Late inlining of boxing methods
2008 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2009 if (_boxing_late_inlines.length() > 0) {
2010 assert(has_boxed_value(), "inconsistent");
2011
2012 set_inlining_incrementally(true);
2013
2014 igvn_worklist()->ensure_empty(); // should be done with igvn
2015
2016 _late_inlines_pos = _late_inlines.length();
2017
2018 while (_boxing_late_inlines.length() > 0) {
2019 CallGenerator* cg = _boxing_late_inlines.pop();
2020 cg->do_late_inline();
2021 if (failing()) return;
2022 }
2023 _boxing_late_inlines.trunc_to(0);
2024
2025 inline_incrementally_cleanup(igvn);
2026
2027 set_inlining_incrementally(false);
2028 }
2029 }
2030
2031 bool Compile::inline_incrementally_one() {
2032 assert(IncrementalInline, "incremental inlining should be on");
2033
2034 TracePhase tp(_t_incrInline_inline);
2035
2036 set_inlining_progress(false);
2037 set_do_cleanup(false);
2038
2039 for (int i = 0; i < _late_inlines.length(); i++) {
2040 _late_inlines_pos = i+1;
2041 CallGenerator* cg = _late_inlines.at(i);
2042 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2043 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2044 cg->do_late_inline();
2045 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2046 if (failing()) {
2047 return false;
2048 } else if (inlining_progress()) {
2049 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2050 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2051 break; // process one call site at a time
2052 }
2053 } else {
2054 // Ignore late inline direct calls when inlining is not allowed.
2055 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2056 }
2057 }
2058 // Remove processed elements.
2059 _late_inlines.remove_till(_late_inlines_pos);
2060 _late_inlines_pos = 0;
2061
2062 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2063
2064 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2065
2066 set_inlining_progress(false);
2067 set_do_cleanup(false);
2068
2069 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2070 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2071 }
2072
2073 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2074 {
2075 TracePhase tp(_t_incrInline_pru);
2076 ResourceMark rm;
2077 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2078 }
2079 {
2080 TracePhase tp(_t_incrInline_igvn);
2081 igvn.reset_from_gvn(initial_gvn());
2082 igvn.optimize();
2083 if (failing()) return;
2084 }
2085 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2086 }
2087
2088 // Perform incremental inlining until bound on number of live nodes is reached
2089 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2090 TracePhase tp(_t_incrInline);
2091
2092 set_inlining_incrementally(true);
2093 uint low_live_nodes = 0;
2094
2095 while (_late_inlines.length() > 0) {
2096 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2097 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2098 TracePhase tp(_t_incrInline_ideal);
2099 // PhaseIdealLoop is expensive so we only try it once we are
2100 // out of live nodes and we only try it again if the previous
2101 // helped got the number of nodes down significantly
2102 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2103 if (failing()) return;
2104 low_live_nodes = live_nodes();
2105 _major_progress = true;
2106 }
2107
2108 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2109 bool do_print_inlining = print_inlining() || print_intrinsics();
2110 if (do_print_inlining || log() != nullptr) {
2111 // Print inlining message for candidates that we couldn't inline for lack of space.
2112 for (int i = 0; i < _late_inlines.length(); i++) {
2113 CallGenerator* cg = _late_inlines.at(i);
2114 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2115 if (do_print_inlining) {
2116 cg->print_inlining_late(InliningResult::FAILURE, msg);
2117 }
2118 log_late_inline_failure(cg, msg);
2119 }
2120 }
2121 break; // finish
2122 }
2123 }
2124
2125 igvn_worklist()->ensure_empty(); // should be done with igvn
2126
2127 while (inline_incrementally_one()) {
2128 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2129 }
2130 if (failing()) return;
2131
2132 inline_incrementally_cleanup(igvn);
2133
2134 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2135
2136 if (failing()) return;
2137
2138 if (_late_inlines.length() == 0) {
2139 break; // no more progress
2140 }
2141 }
2142
2143 igvn_worklist()->ensure_empty(); // should be done with igvn
2144
2145 if (_string_late_inlines.length() > 0) {
2146 assert(has_stringbuilder(), "inconsistent");
2147
2148 inline_string_calls(false);
2149
2150 if (failing()) return;
2151
2152 inline_incrementally_cleanup(igvn);
2153 }
2154
2155 set_inlining_incrementally(false);
2156 }
2157
2158 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2159 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2160 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2161 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2162 // as if "inlining_incrementally() == true" were set.
2163 assert(inlining_incrementally() == false, "not allowed");
2164 assert(_modified_nodes == nullptr, "not allowed");
2165 assert(_late_inlines.length() > 0, "sanity");
2166
2167 while (_late_inlines.length() > 0) {
2168 igvn_worklist()->ensure_empty(); // should be done with igvn
2169
2170 while (inline_incrementally_one()) {
2171 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2172 }
2173 if (failing()) return;
2174
2175 inline_incrementally_cleanup(igvn);
2176 }
2177 }
2178
2179 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2180 if (_loop_opts_cnt > 0) {
2181 while (major_progress() && (_loop_opts_cnt > 0)) {
2182 TracePhase tp(_t_idealLoop);
2183 PhaseIdealLoop::optimize(igvn, mode);
2184 _loop_opts_cnt--;
2185 if (failing()) return false;
2186 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2187 }
2188 }
2189 return true;
2190 }
2191
2192 // Remove edges from "root" to each SafePoint at a backward branch.
2193 // They were inserted during parsing (see add_safepoint()) to make
2194 // infinite loops without calls or exceptions visible to root, i.e.,
2195 // useful.
2196 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2197 Node *r = root();
2198 if (r != nullptr) {
2199 for (uint i = r->req(); i < r->len(); ++i) {
2200 Node *n = r->in(i);
2201 if (n != nullptr && n->is_SafePoint()) {
2202 r->rm_prec(i);
2203 if (n->outcnt() == 0) {
2204 igvn.remove_dead_node(n);
2205 }
2206 --i;
2207 }
2208 }
2209 // Parsing may have added top inputs to the root node (Path
2210 // leading to the Halt node proven dead). Make sure we get a
2211 // chance to clean them up.
2212 igvn._worklist.push(r);
2213 igvn.optimize();
2214 }
2215 }
2216
2217 //------------------------------Optimize---------------------------------------
2218 // Given a graph, optimize it.
2219 void Compile::Optimize() {
2220 TracePhase tp(_t_optimizer);
2221
2222 #ifndef PRODUCT
2223 if (env()->break_at_compile()) {
2224 BREAKPOINT;
2225 }
2226
2227 #endif
2228
2229 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2230 #ifdef ASSERT
2231 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2232 #endif
2233
2234 ResourceMark rm;
2235
2236 print_inlining_reinit();
2237
2238 NOT_PRODUCT( verify_graph_edges(); )
2239
2240 print_method(PHASE_AFTER_PARSING, 1);
2241
2242 {
2243 // Iterative Global Value Numbering, including ideal transforms
2244 // Initialize IterGVN with types and values from parse-time GVN
2245 PhaseIterGVN igvn(initial_gvn());
2246 #ifdef ASSERT
2247 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2248 #endif
2249 {
2250 TracePhase tp(_t_iterGVN);
2251 igvn.optimize();
2252 }
2253
2254 if (failing()) return;
2255
2256 print_method(PHASE_ITER_GVN1, 2);
2257
2258 process_for_unstable_if_traps(igvn);
2259
2260 if (failing()) return;
2261
2262 inline_incrementally(igvn);
2263
2264 print_method(PHASE_INCREMENTAL_INLINE, 2);
2265
2266 if (failing()) return;
2267
2268 if (eliminate_boxing()) {
2269 // Inline valueOf() methods now.
2270 inline_boxing_calls(igvn);
2271
2272 if (failing()) return;
2273
2274 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2275 inline_incrementally(igvn);
2276 }
2277
2278 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2279
2280 if (failing()) return;
2281 }
2282
2283 // Remove the speculative part of types and clean up the graph from
2284 // the extra CastPP nodes whose only purpose is to carry them. Do
2285 // that early so that optimizations are not disrupted by the extra
2286 // CastPP nodes.
2287 remove_speculative_types(igvn);
2288
2289 if (failing()) return;
2290
2291 // No more new expensive nodes will be added to the list from here
2292 // so keep only the actual candidates for optimizations.
2293 cleanup_expensive_nodes(igvn);
2294
2295 if (failing()) return;
2296
2297 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2298 if (EnableVectorSupport && has_vbox_nodes()) {
2299 TracePhase tp(_t_vector);
2300 PhaseVector pv(igvn);
2301 pv.optimize_vector_boxes();
2302 if (failing()) return;
2303 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2304 }
2305 assert(!has_vbox_nodes(), "sanity");
2306
2307 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2308 Compile::TracePhase tp(_t_renumberLive);
2309 igvn_worklist()->ensure_empty(); // should be done with igvn
2310 {
2311 ResourceMark rm;
2312 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2313 }
2314 igvn.reset_from_gvn(initial_gvn());
2315 igvn.optimize();
2316 if (failing()) return;
2317 }
2318
2319 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2320 // safepoints
2321 remove_root_to_sfpts_edges(igvn);
2322
2323 if (failing()) return;
2324
2325 // Perform escape analysis
2326 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2327 if (has_loops()) {
2328 // Cleanup graph (remove dead nodes).
2329 TracePhase tp(_t_idealLoop);
2330 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2331 if (failing()) return;
2332 }
2333 bool progress;
2334 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2335 do {
2336 ConnectionGraph::do_analysis(this, &igvn);
2337
2338 if (failing()) return;
2339
2340 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2341
2342 // Optimize out fields loads from scalar replaceable allocations.
2343 igvn.optimize();
2344 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2345
2346 if (failing()) return;
2347
2348 if (congraph() != nullptr && macro_count() > 0) {
2349 TracePhase tp(_t_macroEliminate);
2350 PhaseMacroExpand mexp(igvn);
2351 mexp.eliminate_macro_nodes();
2352 if (failing()) return;
2353
2354 igvn.set_delay_transform(false);
2355 igvn.optimize();
2356 if (failing()) return;
2357
2358 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2359 }
2360
2361 ConnectionGraph::verify_ram_nodes(this, root());
2362 if (failing()) return;
2363
2364 progress = do_iterative_escape_analysis() &&
2365 (macro_count() < mcount) &&
2366 ConnectionGraph::has_candidates(this);
2367 // Try again if candidates exist and made progress
2368 // by removing some allocations and/or locks.
2369 } while (progress);
2370 }
2371
2372 // Loop transforms on the ideal graph. Range Check Elimination,
2373 // peeling, unrolling, etc.
2374
2375 // Set loop opts counter
2376 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2377 {
2378 TracePhase tp(_t_idealLoop);
2379 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2380 _loop_opts_cnt--;
2381 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2382 if (failing()) return;
2383 }
2384 // Loop opts pass if partial peeling occurred in previous pass
2385 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2386 TracePhase tp(_t_idealLoop);
2387 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2388 _loop_opts_cnt--;
2389 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2390 if (failing()) return;
2391 }
2392 // Loop opts pass for loop-unrolling before CCP
2393 if(major_progress() && (_loop_opts_cnt > 0)) {
2394 TracePhase tp(_t_idealLoop);
2395 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2396 _loop_opts_cnt--;
2397 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2398 }
2399 if (!failing()) {
2400 // Verify that last round of loop opts produced a valid graph
2401 PhaseIdealLoop::verify(igvn);
2402 }
2403 }
2404 if (failing()) return;
2405
2406 // Conditional Constant Propagation;
2407 print_method(PHASE_BEFORE_CCP1, 2);
2408 PhaseCCP ccp( &igvn );
2409 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2410 {
2411 TracePhase tp(_t_ccp);
2412 ccp.do_transform();
2413 }
2414 print_method(PHASE_CCP1, 2);
2415
2416 assert( true, "Break here to ccp.dump_old2new_map()");
2417
2418 // Iterative Global Value Numbering, including ideal transforms
2419 {
2420 TracePhase tp(_t_iterGVN2);
2421 igvn.reset_from_igvn(&ccp);
2422 igvn.optimize();
2423 }
2424 print_method(PHASE_ITER_GVN2, 2);
2425
2426 if (failing()) return;
2427
2428 // Loop transforms on the ideal graph. Range Check Elimination,
2429 // peeling, unrolling, etc.
2430 if (!optimize_loops(igvn, LoopOptsDefault)) {
2431 return;
2432 }
2433
2434 if (failing()) return;
2435
2436 C->clear_major_progress(); // ensure that major progress is now clear
2437
2438 process_for_post_loop_opts_igvn(igvn);
2439
2440 if (failing()) return;
2441
2442 #ifdef ASSERT
2443 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2444 #endif
2445
2446 {
2447 TracePhase tp(_t_macroExpand);
2448 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2449 PhaseMacroExpand mex(igvn);
2450 if (mex.expand_macro_nodes()) {
2451 assert(failing(), "must bail out w/ explicit message");
2452 return;
2453 }
2454 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2455 }
2456
2457 {
2458 TracePhase tp(_t_barrierExpand);
2459 if (bs->expand_barriers(this, igvn)) {
2460 assert(failing(), "must bail out w/ explicit message");
2461 return;
2462 }
2463 print_method(PHASE_BARRIER_EXPANSION, 2);
2464 }
2465
2466 if (C->max_vector_size() > 0) {
2467 C->optimize_logic_cones(igvn);
2468 igvn.optimize();
2469 if (failing()) return;
2470 }
2471
2472 DEBUG_ONLY( _modified_nodes = nullptr; )
2473
2474 assert(igvn._worklist.size() == 0, "not empty");
2475
2476 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2477
2478 if (_late_inlines.length() > 0) {
2479 // More opportunities to optimize virtual and MH calls.
2480 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2481 process_late_inline_calls_no_inline(igvn);
2482 if (failing()) return;
2483 }
2484 } // (End scope of igvn; run destructor if necessary for asserts.)
2485
2486 check_no_dead_use();
2487
2488 process_print_inlining();
2489
2490 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2491 // to remove hashes to unlock nodes for modifications.
2492 C->node_hash()->clear();
2493
2494 // A method with only infinite loops has no edges entering loops from root
2495 {
2496 TracePhase tp(_t_graphReshaping);
2497 if (final_graph_reshaping()) {
2498 assert(failing(), "must bail out w/ explicit message");
2499 return;
2500 }
2501 }
2502
2503 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2504 DEBUG_ONLY(set_phase_optimize_finished();)
2505 }
2506
2507 #ifdef ASSERT
2508 void Compile::check_no_dead_use() const {
2509 ResourceMark rm;
2510 Unique_Node_List wq;
2511 wq.push(root());
2512 for (uint i = 0; i < wq.size(); ++i) {
2513 Node* n = wq.at(i);
2514 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2515 Node* u = n->fast_out(j);
2516 if (u->outcnt() == 0 && !u->is_Con()) {
2517 u->dump();
2518 fatal("no reachable node should have no use");
2519 }
2520 wq.push(u);
2521 }
2522 }
2523 }
2524 #endif
2525
2526 void Compile::inline_vector_reboxing_calls() {
2527 if (C->_vector_reboxing_late_inlines.length() > 0) {
2528 _late_inlines_pos = C->_late_inlines.length();
2529 while (_vector_reboxing_late_inlines.length() > 0) {
2530 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2531 cg->do_late_inline();
2532 if (failing()) return;
2533 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2534 }
2535 _vector_reboxing_late_inlines.trunc_to(0);
2536 }
2537 }
2538
2539 bool Compile::has_vbox_nodes() {
2540 if (C->_vector_reboxing_late_inlines.length() > 0) {
2541 return true;
2542 }
2543 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2544 Node * n = C->macro_node(macro_idx);
2545 assert(n->is_macro(), "only macro nodes expected here");
2546 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2547 return true;
2548 }
2549 }
2550 return false;
2551 }
2552
2553 //---------------------------- Bitwise operation packing optimization ---------------------------
2554
2555 static bool is_vector_unary_bitwise_op(Node* n) {
2556 return n->Opcode() == Op_XorV &&
2557 VectorNode::is_vector_bitwise_not_pattern(n);
2558 }
2559
2560 static bool is_vector_binary_bitwise_op(Node* n) {
2561 switch (n->Opcode()) {
2562 case Op_AndV:
2563 case Op_OrV:
2564 return true;
2565
2566 case Op_XorV:
2567 return !is_vector_unary_bitwise_op(n);
2568
2569 default:
2570 return false;
2571 }
2572 }
2573
2574 static bool is_vector_ternary_bitwise_op(Node* n) {
2575 return n->Opcode() == Op_MacroLogicV;
2576 }
2577
2578 static bool is_vector_bitwise_op(Node* n) {
2579 return is_vector_unary_bitwise_op(n) ||
2580 is_vector_binary_bitwise_op(n) ||
2581 is_vector_ternary_bitwise_op(n);
2582 }
2583
2584 static bool is_vector_bitwise_cone_root(Node* n) {
2585 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2586 return false;
2587 }
2588 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2589 if (is_vector_bitwise_op(n->fast_out(i))) {
2590 return false;
2591 }
2592 }
2593 return true;
2594 }
2595
2596 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2597 uint cnt = 0;
2598 if (is_vector_bitwise_op(n)) {
2599 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2600 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2601 for (uint i = 1; i < inp_cnt; i++) {
2602 Node* in = n->in(i);
2603 bool skip = VectorNode::is_all_ones_vector(in);
2604 if (!skip && !inputs.member(in)) {
2605 inputs.push(in);
2606 cnt++;
2607 }
2608 }
2609 assert(cnt <= 1, "not unary");
2610 } else {
2611 uint last_req = inp_cnt;
2612 if (is_vector_ternary_bitwise_op(n)) {
2613 last_req = inp_cnt - 1; // skip last input
2614 }
2615 for (uint i = 1; i < last_req; i++) {
2616 Node* def = n->in(i);
2617 if (!inputs.member(def)) {
2618 inputs.push(def);
2619 cnt++;
2620 }
2621 }
2622 }
2623 } else { // not a bitwise operations
2624 if (!inputs.member(n)) {
2625 inputs.push(n);
2626 cnt++;
2627 }
2628 }
2629 return cnt;
2630 }
2631
2632 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2633 Unique_Node_List useful_nodes;
2634 C->identify_useful_nodes(useful_nodes);
2635
2636 for (uint i = 0; i < useful_nodes.size(); i++) {
2637 Node* n = useful_nodes.at(i);
2638 if (is_vector_bitwise_cone_root(n)) {
2639 list.push(n);
2640 }
2641 }
2642 }
2643
2644 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2645 const TypeVect* vt,
2646 Unique_Node_List& partition,
2647 Unique_Node_List& inputs) {
2648 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2649 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2650 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2651
2652 Node* in1 = inputs.at(0);
2653 Node* in2 = inputs.at(1);
2654 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2655
2656 uint func = compute_truth_table(partition, inputs);
2657
2658 Node* pn = partition.at(partition.size() - 1);
2659 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2660 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2661 }
2662
2663 static uint extract_bit(uint func, uint pos) {
2664 return (func & (1 << pos)) >> pos;
2665 }
2666
2667 //
2668 // A macro logic node represents a truth table. It has 4 inputs,
2669 // First three inputs corresponds to 3 columns of a truth table
2670 // and fourth input captures the logic function.
2671 //
2672 // eg. fn = (in1 AND in2) OR in3;
2673 //
2674 // MacroNode(in1,in2,in3,fn)
2675 //
2676 // -----------------
2677 // in1 in2 in3 fn
2678 // -----------------
2679 // 0 0 0 0
2680 // 0 0 1 1
2681 // 0 1 0 0
2682 // 0 1 1 1
2683 // 1 0 0 0
2684 // 1 0 1 1
2685 // 1 1 0 1
2686 // 1 1 1 1
2687 //
2688
2689 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2690 int res = 0;
2691 for (int i = 0; i < 8; i++) {
2692 int bit1 = extract_bit(in1, i);
2693 int bit2 = extract_bit(in2, i);
2694 int bit3 = extract_bit(in3, i);
2695
2696 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2697 int func_bit = extract_bit(func, func_bit_pos);
2698
2699 res |= func_bit << i;
2700 }
2701 return res;
2702 }
2703
2704 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2705 assert(n != nullptr, "");
2706 assert(eval_map.contains(n), "absent");
2707 return *(eval_map.get(n));
2708 }
2709
2710 static void eval_operands(Node* n,
2711 uint& func1, uint& func2, uint& func3,
2712 ResourceHashtable<Node*,uint>& eval_map) {
2713 assert(is_vector_bitwise_op(n), "");
2714
2715 if (is_vector_unary_bitwise_op(n)) {
2716 Node* opnd = n->in(1);
2717 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2718 opnd = n->in(2);
2719 }
2720 func1 = eval_operand(opnd, eval_map);
2721 } else if (is_vector_binary_bitwise_op(n)) {
2722 func1 = eval_operand(n->in(1), eval_map);
2723 func2 = eval_operand(n->in(2), eval_map);
2724 } else {
2725 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2726 func1 = eval_operand(n->in(1), eval_map);
2727 func2 = eval_operand(n->in(2), eval_map);
2728 func3 = eval_operand(n->in(3), eval_map);
2729 }
2730 }
2731
2732 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2733 assert(inputs.size() <= 3, "sanity");
2734 ResourceMark rm;
2735 uint res = 0;
2736 ResourceHashtable<Node*,uint> eval_map;
2737
2738 // Populate precomputed functions for inputs.
2739 // Each input corresponds to one column of 3 input truth-table.
2740 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2741 0xCC, // (_, b, _) -> b
2742 0xF0 }; // (a, _, _) -> a
2743 for (uint i = 0; i < inputs.size(); i++) {
2744 eval_map.put(inputs.at(i), input_funcs[2-i]);
2745 }
2746
2747 for (uint i = 0; i < partition.size(); i++) {
2748 Node* n = partition.at(i);
2749
2750 uint func1 = 0, func2 = 0, func3 = 0;
2751 eval_operands(n, func1, func2, func3, eval_map);
2752
2753 switch (n->Opcode()) {
2754 case Op_OrV:
2755 assert(func3 == 0, "not binary");
2756 res = func1 | func2;
2757 break;
2758 case Op_AndV:
2759 assert(func3 == 0, "not binary");
2760 res = func1 & func2;
2761 break;
2762 case Op_XorV:
2763 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2764 assert(func2 == 0 && func3 == 0, "not unary");
2765 res = (~func1) & 0xFF;
2766 } else {
2767 assert(func3 == 0, "not binary");
2768 res = func1 ^ func2;
2769 }
2770 break;
2771 case Op_MacroLogicV:
2772 // Ordering of inputs may change during evaluation of sub-tree
2773 // containing MacroLogic node as a child node, thus a re-evaluation
2774 // makes sure that function is evaluated in context of current
2775 // inputs.
2776 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2777 break;
2778
2779 default: assert(false, "not supported: %s", n->Name());
2780 }
2781 assert(res <= 0xFF, "invalid");
2782 eval_map.put(n, res);
2783 }
2784 return res;
2785 }
2786
2787 // Criteria under which nodes gets packed into a macro logic node:-
2788 // 1) Parent and both child nodes are all unmasked or masked with
2789 // same predicates.
2790 // 2) Masked parent can be packed with left child if it is predicated
2791 // and both have same predicates.
2792 // 3) Masked parent can be packed with right child if its un-predicated
2793 // or has matching predication condition.
2794 // 4) An unmasked parent can be packed with an unmasked child.
2795 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2796 assert(partition.size() == 0, "not empty");
2797 assert(inputs.size() == 0, "not empty");
2798 if (is_vector_ternary_bitwise_op(n)) {
2799 return false;
2800 }
2801
2802 bool is_unary_op = is_vector_unary_bitwise_op(n);
2803 if (is_unary_op) {
2804 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2805 return false; // too few inputs
2806 }
2807
2808 bool pack_left_child = true;
2809 bool pack_right_child = true;
2810
2811 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2812 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2813
2814 int left_child_input_cnt = 0;
2815 int right_child_input_cnt = 0;
2816
2817 bool parent_is_predicated = n->is_predicated_vector();
2818 bool left_child_predicated = n->in(1)->is_predicated_vector();
2819 bool right_child_predicated = n->in(2)->is_predicated_vector();
2820
2821 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2822 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2823 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2824
2825 do {
2826 if (pack_left_child && left_child_LOP &&
2827 ((!parent_is_predicated && !left_child_predicated) ||
2828 ((parent_is_predicated && left_child_predicated &&
2829 parent_pred == left_child_pred)))) {
2830 partition.push(n->in(1));
2831 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2832 } else {
2833 inputs.push(n->in(1));
2834 left_child_input_cnt = 1;
2835 }
2836
2837 if (pack_right_child && right_child_LOP &&
2838 (!right_child_predicated ||
2839 (right_child_predicated && parent_is_predicated &&
2840 parent_pred == right_child_pred))) {
2841 partition.push(n->in(2));
2842 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2843 } else {
2844 inputs.push(n->in(2));
2845 right_child_input_cnt = 1;
2846 }
2847
2848 if (inputs.size() > 3) {
2849 assert(partition.size() > 0, "");
2850 inputs.clear();
2851 partition.clear();
2852 if (left_child_input_cnt > right_child_input_cnt) {
2853 pack_left_child = false;
2854 } else {
2855 pack_right_child = false;
2856 }
2857 } else {
2858 break;
2859 }
2860 } while(true);
2861
2862 if(partition.size()) {
2863 partition.push(n);
2864 }
2865
2866 return (partition.size() == 2 || partition.size() == 3) &&
2867 (inputs.size() == 2 || inputs.size() == 3);
2868 }
2869
2870 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2871 assert(is_vector_bitwise_op(n), "not a root");
2872
2873 visited.set(n->_idx);
2874
2875 // 1) Do a DFS walk over the logic cone.
2876 for (uint i = 1; i < n->req(); i++) {
2877 Node* in = n->in(i);
2878 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2879 process_logic_cone_root(igvn, in, visited);
2880 }
2881 }
2882
2883 // 2) Bottom up traversal: Merge node[s] with
2884 // the parent to form macro logic node.
2885 Unique_Node_List partition;
2886 Unique_Node_List inputs;
2887 if (compute_logic_cone(n, partition, inputs)) {
2888 const TypeVect* vt = n->bottom_type()->is_vect();
2889 Node* pn = partition.at(partition.size() - 1);
2890 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2891 if (mask == nullptr ||
2892 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2893 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2894 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2895 igvn.replace_node(n, macro_logic);
2896 }
2897 }
2898 }
2899
2900 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2901 ResourceMark rm;
2902 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2903 Unique_Node_List list;
2904 collect_logic_cone_roots(list);
2905
2906 while (list.size() > 0) {
2907 Node* n = list.pop();
2908 const TypeVect* vt = n->bottom_type()->is_vect();
2909 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2910 if (supported) {
2911 VectorSet visited(comp_arena());
2912 process_logic_cone_root(igvn, n, visited);
2913 }
2914 }
2915 }
2916 }
2917
2918 //------------------------------Code_Gen---------------------------------------
2919 // Given a graph, generate code for it
2920 void Compile::Code_Gen() {
2921 if (failing()) {
2922 return;
2923 }
2924
2925 // Perform instruction selection. You might think we could reclaim Matcher
2926 // memory PDQ, but actually the Matcher is used in generating spill code.
2927 // Internals of the Matcher (including some VectorSets) must remain live
2928 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2929 // set a bit in reclaimed memory.
2930
2931 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2932 // nodes. Mapping is only valid at the root of each matched subtree.
2933 NOT_PRODUCT( verify_graph_edges(); )
2934
2935 Matcher matcher;
2936 _matcher = &matcher;
2937 {
2938 TracePhase tp(_t_matcher);
2939 matcher.match();
2940 if (failing()) {
2941 return;
2942 }
2943 }
2944 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2945 // nodes. Mapping is only valid at the root of each matched subtree.
2946 NOT_PRODUCT( verify_graph_edges(); )
2947
2948 // If you have too many nodes, or if matching has failed, bail out
2949 check_node_count(0, "out of nodes matching instructions");
2950 if (failing()) {
2951 return;
2952 }
2953
2954 print_method(PHASE_MATCHING, 2);
2955
2956 // Build a proper-looking CFG
2957 PhaseCFG cfg(node_arena(), root(), matcher);
2958 if (failing()) {
2959 return;
2960 }
2961 _cfg = &cfg;
2962 {
2963 TracePhase tp(_t_scheduler);
2964 bool success = cfg.do_global_code_motion();
2965 if (!success) {
2966 return;
2967 }
2968
2969 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2970 NOT_PRODUCT( verify_graph_edges(); )
2971 cfg.verify();
2972 if (failing()) {
2973 return;
2974 }
2975 }
2976
2977 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2978 _regalloc = ®alloc;
2979 {
2980 TracePhase tp(_t_registerAllocation);
2981 // Perform register allocation. After Chaitin, use-def chains are
2982 // no longer accurate (at spill code) and so must be ignored.
2983 // Node->LRG->reg mappings are still accurate.
2984 _regalloc->Register_Allocate();
2985
2986 // Bail out if the allocator builds too many nodes
2987 if (failing()) {
2988 return;
2989 }
2990
2991 print_method(PHASE_REGISTER_ALLOCATION, 2);
2992 }
2993
2994 // Prior to register allocation we kept empty basic blocks in case the
2995 // the allocator needed a place to spill. After register allocation we
2996 // are not adding any new instructions. If any basic block is empty, we
2997 // can now safely remove it.
2998 {
2999 TracePhase tp(_t_blockOrdering);
3000 cfg.remove_empty_blocks();
3001 if (do_freq_based_layout()) {
3002 PhaseBlockLayout layout(cfg);
3003 } else {
3004 cfg.set_loop_alignment();
3005 }
3006 cfg.fixup_flow();
3007 cfg.remove_unreachable_blocks();
3008 cfg.verify_dominator_tree();
3009 print_method(PHASE_BLOCK_ORDERING, 3);
3010 }
3011
3012 // Apply peephole optimizations
3013 if( OptoPeephole ) {
3014 TracePhase tp(_t_peephole);
3015 PhasePeephole peep( _regalloc, cfg);
3016 peep.do_transform();
3017 print_method(PHASE_PEEPHOLE, 3);
3018 }
3019
3020 // Do late expand if CPU requires this.
3021 if (Matcher::require_postalloc_expand) {
3022 TracePhase tp(_t_postalloc_expand);
3023 cfg.postalloc_expand(_regalloc);
3024 print_method(PHASE_POSTALLOC_EXPAND, 3);
3025 }
3026
3027 // Convert Nodes to instruction bits in a buffer
3028 {
3029 TracePhase tp(_t_output);
3030 PhaseOutput output;
3031 output.Output();
3032 if (failing()) return;
3033 output.install();
3034 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3035 }
3036
3037 // He's dead, Jim.
3038 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3039 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3040 }
3041
3042 //------------------------------Final_Reshape_Counts---------------------------
3043 // This class defines counters to help identify when a method
3044 // may/must be executed using hardware with only 24-bit precision.
3045 struct Final_Reshape_Counts : public StackObj {
3046 int _call_count; // count non-inlined 'common' calls
3047 int _float_count; // count float ops requiring 24-bit precision
3048 int _double_count; // count double ops requiring more precision
3049 int _java_call_count; // count non-inlined 'java' calls
3050 int _inner_loop_count; // count loops which need alignment
3051 VectorSet _visited; // Visitation flags
3052 Node_List _tests; // Set of IfNodes & PCTableNodes
3053
3054 Final_Reshape_Counts() :
3055 _call_count(0), _float_count(0), _double_count(0),
3056 _java_call_count(0), _inner_loop_count(0) { }
3057
3058 void inc_call_count () { _call_count ++; }
3059 void inc_float_count () { _float_count ++; }
3060 void inc_double_count() { _double_count++; }
3061 void inc_java_call_count() { _java_call_count++; }
3062 void inc_inner_loop_count() { _inner_loop_count++; }
3063
3064 int get_call_count () const { return _call_count ; }
3065 int get_float_count () const { return _float_count ; }
3066 int get_double_count() const { return _double_count; }
3067 int get_java_call_count() const { return _java_call_count; }
3068 int get_inner_loop_count() const { return _inner_loop_count; }
3069 };
3070
3071 //------------------------------final_graph_reshaping_impl----------------------
3072 // Implement items 1-5 from final_graph_reshaping below.
3073 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3074
3075 if ( n->outcnt() == 0 ) return; // dead node
3076 uint nop = n->Opcode();
3077
3078 // Check for 2-input instruction with "last use" on right input.
3079 // Swap to left input. Implements item (2).
3080 if( n->req() == 3 && // two-input instruction
3081 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3082 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3083 n->in(2)->outcnt() == 1 &&// right use IS a last use
3084 !n->in(2)->is_Con() ) { // right use is not a constant
3085 // Check for commutative opcode
3086 switch( nop ) {
3087 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3088 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3089 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3090 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3091 case Op_AndL: case Op_XorL: case Op_OrL:
3092 case Op_AndI: case Op_XorI: case Op_OrI: {
3093 // Move "last use" input to left by swapping inputs
3094 n->swap_edges(1, 2);
3095 break;
3096 }
3097 default:
3098 break;
3099 }
3100 }
3101
3102 #ifdef ASSERT
3103 if( n->is_Mem() ) {
3104 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3105 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3106 // oop will be recorded in oop map if load crosses safepoint
3107 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3108 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3109 "raw memory operations should have control edge");
3110 }
3111 if (n->is_MemBar()) {
3112 MemBarNode* mb = n->as_MemBar();
3113 if (mb->trailing_store() || mb->trailing_load_store()) {
3114 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3115 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3116 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3117 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3118 } else if (mb->leading()) {
3119 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3120 }
3121 }
3122 #endif
3123 // Count FPU ops and common calls, implements item (3)
3124 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3125 if (!gc_handled) {
3126 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3127 }
3128
3129 // Collect CFG split points
3130 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3131 frc._tests.push(n);
3132 }
3133 }
3134
3135 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3136 if (!UseDivMod) {
3137 return;
3138 }
3139
3140 // Check if "a % b" and "a / b" both exist
3141 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3142 if (d == nullptr) {
3143 return;
3144 }
3145
3146 // Replace them with a fused divmod if supported
3147 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3148 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3149 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3150 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3151 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3152 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3153 // DivMod node so the dependency is not lost.
3154 divmod->add_prec_from(n);
3155 divmod->add_prec_from(d);
3156 d->subsume_by(divmod->div_proj(), this);
3157 n->subsume_by(divmod->mod_proj(), this);
3158 } else {
3159 // Replace "a % b" with "a - ((a / b) * b)"
3160 Node* mult = MulNode::make(d, d->in(2), bt);
3161 Node* sub = SubNode::make(d->in(1), mult, bt);
3162 n->subsume_by(sub, this);
3163 }
3164 }
3165
3166 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3167 switch( nop ) {
3168 // Count all float operations that may use FPU
3169 case Op_AddF:
3170 case Op_SubF:
3171 case Op_MulF:
3172 case Op_DivF:
3173 case Op_NegF:
3174 case Op_ModF:
3175 case Op_ConvI2F:
3176 case Op_ConF:
3177 case Op_CmpF:
3178 case Op_CmpF3:
3179 case Op_StoreF:
3180 case Op_LoadF:
3181 // case Op_ConvL2F: // longs are split into 32-bit halves
3182 frc.inc_float_count();
3183 break;
3184
3185 case Op_ConvF2D:
3186 case Op_ConvD2F:
3187 frc.inc_float_count();
3188 frc.inc_double_count();
3189 break;
3190
3191 // Count all double operations that may use FPU
3192 case Op_AddD:
3193 case Op_SubD:
3194 case Op_MulD:
3195 case Op_DivD:
3196 case Op_NegD:
3197 case Op_ModD:
3198 case Op_ConvI2D:
3199 case Op_ConvD2I:
3200 // case Op_ConvL2D: // handled by leaf call
3201 // case Op_ConvD2L: // handled by leaf call
3202 case Op_ConD:
3203 case Op_CmpD:
3204 case Op_CmpD3:
3205 case Op_StoreD:
3206 case Op_LoadD:
3207 case Op_LoadD_unaligned:
3208 frc.inc_double_count();
3209 break;
3210 case Op_Opaque1: // Remove Opaque Nodes before matching
3211 n->subsume_by(n->in(1), this);
3212 break;
3213 case Op_CallStaticJava:
3214 case Op_CallJava:
3215 case Op_CallDynamicJava:
3216 frc.inc_java_call_count(); // Count java call site;
3217 case Op_CallRuntime:
3218 case Op_CallLeaf:
3219 case Op_CallLeafVector:
3220 case Op_CallLeafNoFP: {
3221 assert (n->is_Call(), "");
3222 CallNode *call = n->as_Call();
3223 // Count call sites where the FP mode bit would have to be flipped.
3224 // Do not count uncommon runtime calls:
3225 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3226 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3227 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3228 frc.inc_call_count(); // Count the call site
3229 } else { // See if uncommon argument is shared
3230 Node *n = call->in(TypeFunc::Parms);
3231 int nop = n->Opcode();
3232 // Clone shared simple arguments to uncommon calls, item (1).
3233 if (n->outcnt() > 1 &&
3234 !n->is_Proj() &&
3235 nop != Op_CreateEx &&
3236 nop != Op_CheckCastPP &&
3237 nop != Op_DecodeN &&
3238 nop != Op_DecodeNKlass &&
3239 !n->is_Mem() &&
3240 !n->is_Phi()) {
3241 Node *x = n->clone();
3242 call->set_req(TypeFunc::Parms, x);
3243 }
3244 }
3245 break;
3246 }
3247 case Op_StoreB:
3248 case Op_StoreC:
3249 case Op_StoreI:
3250 case Op_StoreL:
3251 case Op_CompareAndSwapB:
3252 case Op_CompareAndSwapS:
3253 case Op_CompareAndSwapI:
3254 case Op_CompareAndSwapL:
3255 case Op_CompareAndSwapP:
3256 case Op_CompareAndSwapN:
3257 case Op_WeakCompareAndSwapB:
3258 case Op_WeakCompareAndSwapS:
3259 case Op_WeakCompareAndSwapI:
3260 case Op_WeakCompareAndSwapL:
3261 case Op_WeakCompareAndSwapP:
3262 case Op_WeakCompareAndSwapN:
3263 case Op_CompareAndExchangeB:
3264 case Op_CompareAndExchangeS:
3265 case Op_CompareAndExchangeI:
3266 case Op_CompareAndExchangeL:
3267 case Op_CompareAndExchangeP:
3268 case Op_CompareAndExchangeN:
3269 case Op_GetAndAddS:
3270 case Op_GetAndAddB:
3271 case Op_GetAndAddI:
3272 case Op_GetAndAddL:
3273 case Op_GetAndSetS:
3274 case Op_GetAndSetB:
3275 case Op_GetAndSetI:
3276 case Op_GetAndSetL:
3277 case Op_GetAndSetP:
3278 case Op_GetAndSetN:
3279 case Op_StoreP:
3280 case Op_StoreN:
3281 case Op_StoreNKlass:
3282 case Op_LoadB:
3283 case Op_LoadUB:
3284 case Op_LoadUS:
3285 case Op_LoadI:
3286 case Op_LoadKlass:
3287 case Op_LoadNKlass:
3288 case Op_LoadL:
3289 case Op_LoadL_unaligned:
3290 case Op_LoadP:
3291 case Op_LoadN:
3292 case Op_LoadRange:
3293 case Op_LoadS:
3294 break;
3295
3296 case Op_AddP: { // Assert sane base pointers
3297 Node *addp = n->in(AddPNode::Address);
3298 assert( !addp->is_AddP() ||
3299 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3300 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3301 "Base pointers must match (addp %u)", addp->_idx );
3302 #ifdef _LP64
3303 if ((UseCompressedOops || UseCompressedClassPointers) &&
3304 addp->Opcode() == Op_ConP &&
3305 addp == n->in(AddPNode::Base) &&
3306 n->in(AddPNode::Offset)->is_Con()) {
3307 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3308 // on the platform and on the compressed oops mode.
3309 // Use addressing with narrow klass to load with offset on x86.
3310 // Some platforms can use the constant pool to load ConP.
3311 // Do this transformation here since IGVN will convert ConN back to ConP.
3312 const Type* t = addp->bottom_type();
3313 bool is_oop = t->isa_oopptr() != nullptr;
3314 bool is_klass = t->isa_klassptr() != nullptr;
3315
3316 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3317 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3318 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3319 Node* nn = nullptr;
3320
3321 int op = is_oop ? Op_ConN : Op_ConNKlass;
3322
3323 // Look for existing ConN node of the same exact type.
3324 Node* r = root();
3325 uint cnt = r->outcnt();
3326 for (uint i = 0; i < cnt; i++) {
3327 Node* m = r->raw_out(i);
3328 if (m!= nullptr && m->Opcode() == op &&
3329 m->bottom_type()->make_ptr() == t) {
3330 nn = m;
3331 break;
3332 }
3333 }
3334 if (nn != nullptr) {
3335 // Decode a narrow oop to match address
3336 // [R12 + narrow_oop_reg<<3 + offset]
3337 if (is_oop) {
3338 nn = new DecodeNNode(nn, t);
3339 } else {
3340 nn = new DecodeNKlassNode(nn, t);
3341 }
3342 // Check for succeeding AddP which uses the same Base.
3343 // Otherwise we will run into the assertion above when visiting that guy.
3344 for (uint i = 0; i < n->outcnt(); ++i) {
3345 Node *out_i = n->raw_out(i);
3346 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3347 out_i->set_req(AddPNode::Base, nn);
3348 #ifdef ASSERT
3349 for (uint j = 0; j < out_i->outcnt(); ++j) {
3350 Node *out_j = out_i->raw_out(j);
3351 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3352 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3353 }
3354 #endif
3355 }
3356 }
3357 n->set_req(AddPNode::Base, nn);
3358 n->set_req(AddPNode::Address, nn);
3359 if (addp->outcnt() == 0) {
3360 addp->disconnect_inputs(this);
3361 }
3362 }
3363 }
3364 }
3365 #endif
3366 break;
3367 }
3368
3369 case Op_CastPP: {
3370 // Remove CastPP nodes to gain more freedom during scheduling but
3371 // keep the dependency they encode as control or precedence edges
3372 // (if control is set already) on memory operations. Some CastPP
3373 // nodes don't have a control (don't carry a dependency): skip
3374 // those.
3375 if (n->in(0) != nullptr) {
3376 ResourceMark rm;
3377 Unique_Node_List wq;
3378 wq.push(n);
3379 for (uint next = 0; next < wq.size(); ++next) {
3380 Node *m = wq.at(next);
3381 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3382 Node* use = m->fast_out(i);
3383 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3384 use->ensure_control_or_add_prec(n->in(0));
3385 } else {
3386 switch(use->Opcode()) {
3387 case Op_AddP:
3388 case Op_DecodeN:
3389 case Op_DecodeNKlass:
3390 case Op_CheckCastPP:
3391 case Op_CastPP:
3392 wq.push(use);
3393 break;
3394 }
3395 }
3396 }
3397 }
3398 }
3399 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3400 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3401 Node* in1 = n->in(1);
3402 const Type* t = n->bottom_type();
3403 Node* new_in1 = in1->clone();
3404 new_in1->as_DecodeN()->set_type(t);
3405
3406 if (!Matcher::narrow_oop_use_complex_address()) {
3407 //
3408 // x86, ARM and friends can handle 2 adds in addressing mode
3409 // and Matcher can fold a DecodeN node into address by using
3410 // a narrow oop directly and do implicit null check in address:
3411 //
3412 // [R12 + narrow_oop_reg<<3 + offset]
3413 // NullCheck narrow_oop_reg
3414 //
3415 // On other platforms (Sparc) we have to keep new DecodeN node and
3416 // use it to do implicit null check in address:
3417 //
3418 // decode_not_null narrow_oop_reg, base_reg
3419 // [base_reg + offset]
3420 // NullCheck base_reg
3421 //
3422 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3423 // to keep the information to which null check the new DecodeN node
3424 // corresponds to use it as value in implicit_null_check().
3425 //
3426 new_in1->set_req(0, n->in(0));
3427 }
3428
3429 n->subsume_by(new_in1, this);
3430 if (in1->outcnt() == 0) {
3431 in1->disconnect_inputs(this);
3432 }
3433 } else {
3434 n->subsume_by(n->in(1), this);
3435 if (n->outcnt() == 0) {
3436 n->disconnect_inputs(this);
3437 }
3438 }
3439 break;
3440 }
3441 case Op_CastII: {
3442 n->as_CastII()->remove_range_check_cast(this);
3443 break;
3444 }
3445 #ifdef _LP64
3446 case Op_CmpP:
3447 // Do this transformation here to preserve CmpPNode::sub() and
3448 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3449 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3450 Node* in1 = n->in(1);
3451 Node* in2 = n->in(2);
3452 if (!in1->is_DecodeNarrowPtr()) {
3453 in2 = in1;
3454 in1 = n->in(2);
3455 }
3456 assert(in1->is_DecodeNarrowPtr(), "sanity");
3457
3458 Node* new_in2 = nullptr;
3459 if (in2->is_DecodeNarrowPtr()) {
3460 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3461 new_in2 = in2->in(1);
3462 } else if (in2->Opcode() == Op_ConP) {
3463 const Type* t = in2->bottom_type();
3464 if (t == TypePtr::NULL_PTR) {
3465 assert(in1->is_DecodeN(), "compare klass to null?");
3466 // Don't convert CmpP null check into CmpN if compressed
3467 // oops implicit null check is not generated.
3468 // This will allow to generate normal oop implicit null check.
3469 if (Matcher::gen_narrow_oop_implicit_null_checks())
3470 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3471 //
3472 // This transformation together with CastPP transformation above
3473 // will generated code for implicit null checks for compressed oops.
3474 //
3475 // The original code after Optimize()
3476 //
3477 // LoadN memory, narrow_oop_reg
3478 // decode narrow_oop_reg, base_reg
3479 // CmpP base_reg, nullptr
3480 // CastPP base_reg // NotNull
3481 // Load [base_reg + offset], val_reg
3482 //
3483 // after these transformations will be
3484 //
3485 // LoadN memory, narrow_oop_reg
3486 // CmpN narrow_oop_reg, nullptr
3487 // decode_not_null narrow_oop_reg, base_reg
3488 // Load [base_reg + offset], val_reg
3489 //
3490 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3491 // since narrow oops can be used in debug info now (see the code in
3492 // final_graph_reshaping_walk()).
3493 //
3494 // At the end the code will be matched to
3495 // on x86:
3496 //
3497 // Load_narrow_oop memory, narrow_oop_reg
3498 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3499 // NullCheck narrow_oop_reg
3500 //
3501 // and on sparc:
3502 //
3503 // Load_narrow_oop memory, narrow_oop_reg
3504 // decode_not_null narrow_oop_reg, base_reg
3505 // Load [base_reg + offset], val_reg
3506 // NullCheck base_reg
3507 //
3508 } else if (t->isa_oopptr()) {
3509 new_in2 = ConNode::make(t->make_narrowoop());
3510 } else if (t->isa_klassptr()) {
3511 new_in2 = ConNode::make(t->make_narrowklass());
3512 }
3513 }
3514 if (new_in2 != nullptr) {
3515 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3516 n->subsume_by(cmpN, this);
3517 if (in1->outcnt() == 0) {
3518 in1->disconnect_inputs(this);
3519 }
3520 if (in2->outcnt() == 0) {
3521 in2->disconnect_inputs(this);
3522 }
3523 }
3524 }
3525 break;
3526
3527 case Op_DecodeN:
3528 case Op_DecodeNKlass:
3529 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3530 // DecodeN could be pinned when it can't be fold into
3531 // an address expression, see the code for Op_CastPP above.
3532 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3533 break;
3534
3535 case Op_EncodeP:
3536 case Op_EncodePKlass: {
3537 Node* in1 = n->in(1);
3538 if (in1->is_DecodeNarrowPtr()) {
3539 n->subsume_by(in1->in(1), this);
3540 } else if (in1->Opcode() == Op_ConP) {
3541 const Type* t = in1->bottom_type();
3542 if (t == TypePtr::NULL_PTR) {
3543 assert(t->isa_oopptr(), "null klass?");
3544 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3545 } else if (t->isa_oopptr()) {
3546 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3547 } else if (t->isa_klassptr()) {
3548 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3549 }
3550 }
3551 if (in1->outcnt() == 0) {
3552 in1->disconnect_inputs(this);
3553 }
3554 break;
3555 }
3556
3557 case Op_Proj: {
3558 if (OptimizeStringConcat || IncrementalInline) {
3559 ProjNode* proj = n->as_Proj();
3560 if (proj->_is_io_use) {
3561 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3562 // Separate projections were used for the exception path which
3563 // are normally removed by a late inline. If it wasn't inlined
3564 // then they will hang around and should just be replaced with
3565 // the original one. Merge them.
3566 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3567 if (non_io_proj != nullptr) {
3568 proj->subsume_by(non_io_proj , this);
3569 }
3570 }
3571 }
3572 break;
3573 }
3574
3575 case Op_Phi:
3576 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3577 // The EncodeP optimization may create Phi with the same edges
3578 // for all paths. It is not handled well by Register Allocator.
3579 Node* unique_in = n->in(1);
3580 assert(unique_in != nullptr, "");
3581 uint cnt = n->req();
3582 for (uint i = 2; i < cnt; i++) {
3583 Node* m = n->in(i);
3584 assert(m != nullptr, "");
3585 if (unique_in != m)
3586 unique_in = nullptr;
3587 }
3588 if (unique_in != nullptr) {
3589 n->subsume_by(unique_in, this);
3590 }
3591 }
3592 break;
3593
3594 #endif
3595
3596 case Op_ModI:
3597 handle_div_mod_op(n, T_INT, false);
3598 break;
3599
3600 case Op_ModL:
3601 handle_div_mod_op(n, T_LONG, false);
3602 break;
3603
3604 case Op_UModI:
3605 handle_div_mod_op(n, T_INT, true);
3606 break;
3607
3608 case Op_UModL:
3609 handle_div_mod_op(n, T_LONG, true);
3610 break;
3611
3612 case Op_LoadVector:
3613 case Op_StoreVector:
3614 #ifdef ASSERT
3615 // Add VerifyVectorAlignment node between adr and load / store.
3616 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3617 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3618 n->as_StoreVector()->must_verify_alignment();
3619 if (must_verify_alignment) {
3620 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3621 n->as_StoreVector()->memory_size();
3622 // The memory access should be aligned to the vector width in bytes.
3623 // However, the underlying array is possibly less well aligned, but at least
3624 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3625 // a loop we can expect at least the following alignment:
3626 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3627 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3628 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3629 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3630 jlong mask = guaranteed_alignment - 1;
3631 Node* mask_con = ConLNode::make(mask);
3632 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3633 n->set_req(MemNode::Address, va);
3634 }
3635 }
3636 #endif
3637 break;
3638
3639 case Op_LoadVectorGather:
3640 case Op_StoreVectorScatter:
3641 case Op_LoadVectorGatherMasked:
3642 case Op_StoreVectorScatterMasked:
3643 case Op_VectorCmpMasked:
3644 case Op_VectorMaskGen:
3645 case Op_LoadVectorMasked:
3646 case Op_StoreVectorMasked:
3647 break;
3648
3649 case Op_AddReductionVI:
3650 case Op_AddReductionVL:
3651 case Op_AddReductionVF:
3652 case Op_AddReductionVD:
3653 case Op_MulReductionVI:
3654 case Op_MulReductionVL:
3655 case Op_MulReductionVF:
3656 case Op_MulReductionVD:
3657 case Op_MinReductionV:
3658 case Op_MaxReductionV:
3659 case Op_AndReductionV:
3660 case Op_OrReductionV:
3661 case Op_XorReductionV:
3662 break;
3663
3664 case Op_PackB:
3665 case Op_PackS:
3666 case Op_PackI:
3667 case Op_PackF:
3668 case Op_PackL:
3669 case Op_PackD:
3670 if (n->req()-1 > 2) {
3671 // Replace many operand PackNodes with a binary tree for matching
3672 PackNode* p = (PackNode*) n;
3673 Node* btp = p->binary_tree_pack(1, n->req());
3674 n->subsume_by(btp, this);
3675 }
3676 break;
3677 case Op_Loop:
3678 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3679 case Op_CountedLoop:
3680 case Op_LongCountedLoop:
3681 case Op_OuterStripMinedLoop:
3682 if (n->as_Loop()->is_inner_loop()) {
3683 frc.inc_inner_loop_count();
3684 }
3685 n->as_Loop()->verify_strip_mined(0);
3686 break;
3687 case Op_LShiftI:
3688 case Op_RShiftI:
3689 case Op_URShiftI:
3690 case Op_LShiftL:
3691 case Op_RShiftL:
3692 case Op_URShiftL:
3693 if (Matcher::need_masked_shift_count) {
3694 // The cpu's shift instructions don't restrict the count to the
3695 // lower 5/6 bits. We need to do the masking ourselves.
3696 Node* in2 = n->in(2);
3697 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3698 const TypeInt* t = in2->find_int_type();
3699 if (t != nullptr && t->is_con()) {
3700 juint shift = t->get_con();
3701 if (shift > mask) { // Unsigned cmp
3702 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3703 }
3704 } else {
3705 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3706 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3707 n->set_req(2, shift);
3708 }
3709 }
3710 if (in2->outcnt() == 0) { // Remove dead node
3711 in2->disconnect_inputs(this);
3712 }
3713 }
3714 break;
3715 case Op_MemBarStoreStore:
3716 case Op_MemBarRelease:
3717 // Break the link with AllocateNode: it is no longer useful and
3718 // confuses register allocation.
3719 if (n->req() > MemBarNode::Precedent) {
3720 n->set_req(MemBarNode::Precedent, top());
3721 }
3722 break;
3723 case Op_MemBarAcquire: {
3724 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3725 // At parse time, the trailing MemBarAcquire for a volatile load
3726 // is created with an edge to the load. After optimizations,
3727 // that input may be a chain of Phis. If those phis have no
3728 // other use, then the MemBarAcquire keeps them alive and
3729 // register allocation can be confused.
3730 dead_nodes.push(n->in(MemBarNode::Precedent));
3731 n->set_req(MemBarNode::Precedent, top());
3732 }
3733 break;
3734 }
3735 case Op_Blackhole:
3736 break;
3737 case Op_RangeCheck: {
3738 RangeCheckNode* rc = n->as_RangeCheck();
3739 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3740 n->subsume_by(iff, this);
3741 frc._tests.push(iff);
3742 break;
3743 }
3744 case Op_ConvI2L: {
3745 if (!Matcher::convi2l_type_required) {
3746 // Code generation on some platforms doesn't need accurate
3747 // ConvI2L types. Widening the type can help remove redundant
3748 // address computations.
3749 n->as_Type()->set_type(TypeLong::INT);
3750 ResourceMark rm;
3751 Unique_Node_List wq;
3752 wq.push(n);
3753 for (uint next = 0; next < wq.size(); next++) {
3754 Node *m = wq.at(next);
3755
3756 for(;;) {
3757 // Loop over all nodes with identical inputs edges as m
3758 Node* k = m->find_similar(m->Opcode());
3759 if (k == nullptr) {
3760 break;
3761 }
3762 // Push their uses so we get a chance to remove node made
3763 // redundant
3764 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3765 Node* u = k->fast_out(i);
3766 if (u->Opcode() == Op_LShiftL ||
3767 u->Opcode() == Op_AddL ||
3768 u->Opcode() == Op_SubL ||
3769 u->Opcode() == Op_AddP) {
3770 wq.push(u);
3771 }
3772 }
3773 // Replace all nodes with identical edges as m with m
3774 k->subsume_by(m, this);
3775 }
3776 }
3777 }
3778 break;
3779 }
3780 case Op_CmpUL: {
3781 if (!Matcher::has_match_rule(Op_CmpUL)) {
3782 // No support for unsigned long comparisons
3783 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3784 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3785 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3786 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3787 Node* andl = new AndLNode(orl, remove_sign_mask);
3788 Node* cmp = new CmpLNode(andl, n->in(2));
3789 n->subsume_by(cmp, this);
3790 }
3791 break;
3792 }
3793 #ifdef ASSERT
3794 case Op_ConNKlass: {
3795 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3796 ciKlass* klass = tp->is_klassptr()->exact_klass();
3797 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3798 break;
3799 }
3800 #endif
3801 default:
3802 assert(!n->is_Call(), "");
3803 assert(!n->is_Mem(), "");
3804 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3805 break;
3806 }
3807 }
3808
3809 //------------------------------final_graph_reshaping_walk---------------------
3810 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3811 // requires that the walk visits a node's inputs before visiting the node.
3812 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3813 Unique_Node_List sfpt;
3814
3815 frc._visited.set(root->_idx); // first, mark node as visited
3816 uint cnt = root->req();
3817 Node *n = root;
3818 uint i = 0;
3819 while (true) {
3820 if (i < cnt) {
3821 // Place all non-visited non-null inputs onto stack
3822 Node* m = n->in(i);
3823 ++i;
3824 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3825 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3826 // compute worst case interpreter size in case of a deoptimization
3827 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3828
3829 sfpt.push(m);
3830 }
3831 cnt = m->req();
3832 nstack.push(n, i); // put on stack parent and next input's index
3833 n = m;
3834 i = 0;
3835 }
3836 } else {
3837 // Now do post-visit work
3838 final_graph_reshaping_impl(n, frc, dead_nodes);
3839 if (nstack.is_empty())
3840 break; // finished
3841 n = nstack.node(); // Get node from stack
3842 cnt = n->req();
3843 i = nstack.index();
3844 nstack.pop(); // Shift to the next node on stack
3845 }
3846 }
3847
3848 // Skip next transformation if compressed oops are not used.
3849 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3850 (!UseCompressedOops && !UseCompressedClassPointers))
3851 return;
3852
3853 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3854 // It could be done for an uncommon traps or any safepoints/calls
3855 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3856 while (sfpt.size() > 0) {
3857 n = sfpt.pop();
3858 JVMState *jvms = n->as_SafePoint()->jvms();
3859 assert(jvms != nullptr, "sanity");
3860 int start = jvms->debug_start();
3861 int end = n->req();
3862 bool is_uncommon = (n->is_CallStaticJava() &&
3863 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3864 for (int j = start; j < end; j++) {
3865 Node* in = n->in(j);
3866 if (in->is_DecodeNarrowPtr()) {
3867 bool safe_to_skip = true;
3868 if (!is_uncommon ) {
3869 // Is it safe to skip?
3870 for (uint i = 0; i < in->outcnt(); i++) {
3871 Node* u = in->raw_out(i);
3872 if (!u->is_SafePoint() ||
3873 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3874 safe_to_skip = false;
3875 }
3876 }
3877 }
3878 if (safe_to_skip) {
3879 n->set_req(j, in->in(1));
3880 }
3881 if (in->outcnt() == 0) {
3882 in->disconnect_inputs(this);
3883 }
3884 }
3885 }
3886 }
3887 }
3888
3889 //------------------------------final_graph_reshaping--------------------------
3890 // Final Graph Reshaping.
3891 //
3892 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3893 // and not commoned up and forced early. Must come after regular
3894 // optimizations to avoid GVN undoing the cloning. Clone constant
3895 // inputs to Loop Phis; these will be split by the allocator anyways.
3896 // Remove Opaque nodes.
3897 // (2) Move last-uses by commutative operations to the left input to encourage
3898 // Intel update-in-place two-address operations and better register usage
3899 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3900 // calls canonicalizing them back.
3901 // (3) Count the number of double-precision FP ops, single-precision FP ops
3902 // and call sites. On Intel, we can get correct rounding either by
3903 // forcing singles to memory (requires extra stores and loads after each
3904 // FP bytecode) or we can set a rounding mode bit (requires setting and
3905 // clearing the mode bit around call sites). The mode bit is only used
3906 // if the relative frequency of single FP ops to calls is low enough.
3907 // This is a key transform for SPEC mpeg_audio.
3908 // (4) Detect infinite loops; blobs of code reachable from above but not
3909 // below. Several of the Code_Gen algorithms fail on such code shapes,
3910 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3911 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3912 // Detection is by looking for IfNodes where only 1 projection is
3913 // reachable from below or CatchNodes missing some targets.
3914 // (5) Assert for insane oop offsets in debug mode.
3915
3916 bool Compile::final_graph_reshaping() {
3917 // an infinite loop may have been eliminated by the optimizer,
3918 // in which case the graph will be empty.
3919 if (root()->req() == 1) {
3920 // Do not compile method that is only a trivial infinite loop,
3921 // since the content of the loop may have been eliminated.
3922 record_method_not_compilable("trivial infinite loop");
3923 return true;
3924 }
3925
3926 // Expensive nodes have their control input set to prevent the GVN
3927 // from freely commoning them. There's no GVN beyond this point so
3928 // no need to keep the control input. We want the expensive nodes to
3929 // be freely moved to the least frequent code path by gcm.
3930 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3931 for (int i = 0; i < expensive_count(); i++) {
3932 _expensive_nodes.at(i)->set_req(0, nullptr);
3933 }
3934
3935 Final_Reshape_Counts frc;
3936
3937 // Visit everybody reachable!
3938 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3939 Node_Stack nstack(live_nodes() >> 1);
3940 Unique_Node_List dead_nodes;
3941 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
3942
3943 // Check for unreachable (from below) code (i.e., infinite loops).
3944 for( uint i = 0; i < frc._tests.size(); i++ ) {
3945 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3946 // Get number of CFG targets.
3947 // Note that PCTables include exception targets after calls.
3948 uint required_outcnt = n->required_outcnt();
3949 if (n->outcnt() != required_outcnt) {
3950 // Check for a few special cases. Rethrow Nodes never take the
3951 // 'fall-thru' path, so expected kids is 1 less.
3952 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3953 if (n->in(0)->in(0)->is_Call()) {
3954 CallNode* call = n->in(0)->in(0)->as_Call();
3955 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3956 required_outcnt--; // Rethrow always has 1 less kid
3957 } else if (call->req() > TypeFunc::Parms &&
3958 call->is_CallDynamicJava()) {
3959 // Check for null receiver. In such case, the optimizer has
3960 // detected that the virtual call will always result in a null
3961 // pointer exception. The fall-through projection of this CatchNode
3962 // will not be populated.
3963 Node* arg0 = call->in(TypeFunc::Parms);
3964 if (arg0->is_Type() &&
3965 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3966 required_outcnt--;
3967 }
3968 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
3969 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3970 // Check for illegal array length. In such case, the optimizer has
3971 // detected that the allocation attempt will always result in an
3972 // exception. There is no fall-through projection of this CatchNode .
3973 assert(call->is_CallStaticJava(), "static call expected");
3974 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
3975 uint valid_length_test_input = call->req() - 1;
3976 Node* valid_length_test = call->in(valid_length_test_input);
3977 call->del_req(valid_length_test_input);
3978 if (valid_length_test->find_int_con(1) == 0) {
3979 required_outcnt--;
3980 }
3981 dead_nodes.push(valid_length_test);
3982 assert(n->outcnt() == required_outcnt, "malformed control flow");
3983 continue;
3984 }
3985 }
3986 }
3987
3988 // Recheck with a better notion of 'required_outcnt'
3989 if (n->outcnt() != required_outcnt) {
3990 record_method_not_compilable("malformed control flow");
3991 return true; // Not all targets reachable!
3992 }
3993 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
3994 CallNode* call = n->in(0)->in(0)->as_Call();
3995 if (call->entry_point() == OptoRuntime::new_array_Java() ||
3996 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3997 assert(call->is_CallStaticJava(), "static call expected");
3998 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
3999 uint valid_length_test_input = call->req() - 1;
4000 dead_nodes.push(call->in(valid_length_test_input));
4001 call->del_req(valid_length_test_input); // valid length test useless now
4002 }
4003 }
4004 // Check that I actually visited all kids. Unreached kids
4005 // must be infinite loops.
4006 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4007 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4008 record_method_not_compilable("infinite loop");
4009 return true; // Found unvisited kid; must be unreach
4010 }
4011
4012 // Here so verification code in final_graph_reshaping_walk()
4013 // always see an OuterStripMinedLoopEnd
4014 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4015 IfNode* init_iff = n->as_If();
4016 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4017 n->subsume_by(iff, this);
4018 }
4019 }
4020
4021 while (dead_nodes.size() > 0) {
4022 Node* m = dead_nodes.pop();
4023 if (m->outcnt() == 0 && m != top()) {
4024 for (uint j = 0; j < m->req(); j++) {
4025 Node* in = m->in(j);
4026 if (in != nullptr) {
4027 dead_nodes.push(in);
4028 }
4029 }
4030 m->disconnect_inputs(this);
4031 }
4032 }
4033
4034 #ifdef IA32
4035 // If original bytecodes contained a mixture of floats and doubles
4036 // check if the optimizer has made it homogeneous, item (3).
4037 if (UseSSE == 0 &&
4038 frc.get_float_count() > 32 &&
4039 frc.get_double_count() == 0 &&
4040 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4041 set_24_bit_selection_and_mode(false, true);
4042 }
4043 #endif // IA32
4044
4045 set_java_calls(frc.get_java_call_count());
4046 set_inner_loops(frc.get_inner_loop_count());
4047
4048 // No infinite loops, no reason to bail out.
4049 return false;
4050 }
4051
4052 //-----------------------------too_many_traps----------------------------------
4053 // Report if there are too many traps at the current method and bci.
4054 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4055 bool Compile::too_many_traps(ciMethod* method,
4056 int bci,
4057 Deoptimization::DeoptReason reason) {
4058 ciMethodData* md = method->method_data();
4059 if (md->is_empty()) {
4060 // Assume the trap has not occurred, or that it occurred only
4061 // because of a transient condition during start-up in the interpreter.
4062 return false;
4063 }
4064 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4065 if (md->has_trap_at(bci, m, reason) != 0) {
4066 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4067 // Also, if there are multiple reasons, or if there is no per-BCI record,
4068 // assume the worst.
4069 if (log())
4070 log()->elem("observe trap='%s' count='%d'",
4071 Deoptimization::trap_reason_name(reason),
4072 md->trap_count(reason));
4073 return true;
4074 } else {
4075 // Ignore method/bci and see if there have been too many globally.
4076 return too_many_traps(reason, md);
4077 }
4078 }
4079
4080 // Less-accurate variant which does not require a method and bci.
4081 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4082 ciMethodData* logmd) {
4083 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4084 // Too many traps globally.
4085 // Note that we use cumulative trap_count, not just md->trap_count.
4086 if (log()) {
4087 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4088 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4089 Deoptimization::trap_reason_name(reason),
4090 mcount, trap_count(reason));
4091 }
4092 return true;
4093 } else {
4094 // The coast is clear.
4095 return false;
4096 }
4097 }
4098
4099 //--------------------------too_many_recompiles--------------------------------
4100 // Report if there are too many recompiles at the current method and bci.
4101 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4102 // Is not eager to return true, since this will cause the compiler to use
4103 // Action_none for a trap point, to avoid too many recompilations.
4104 bool Compile::too_many_recompiles(ciMethod* method,
4105 int bci,
4106 Deoptimization::DeoptReason reason) {
4107 ciMethodData* md = method->method_data();
4108 if (md->is_empty()) {
4109 // Assume the trap has not occurred, or that it occurred only
4110 // because of a transient condition during start-up in the interpreter.
4111 return false;
4112 }
4113 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4114 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4115 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4116 Deoptimization::DeoptReason per_bc_reason
4117 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4118 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4119 if ((per_bc_reason == Deoptimization::Reason_none
4120 || md->has_trap_at(bci, m, reason) != 0)
4121 // The trap frequency measure we care about is the recompile count:
4122 && md->trap_recompiled_at(bci, m)
4123 && md->overflow_recompile_count() >= bc_cutoff) {
4124 // Do not emit a trap here if it has already caused recompilations.
4125 // Also, if there are multiple reasons, or if there is no per-BCI record,
4126 // assume the worst.
4127 if (log())
4128 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4129 Deoptimization::trap_reason_name(reason),
4130 md->trap_count(reason),
4131 md->overflow_recompile_count());
4132 return true;
4133 } else if (trap_count(reason) != 0
4134 && decompile_count() >= m_cutoff) {
4135 // Too many recompiles globally, and we have seen this sort of trap.
4136 // Use cumulative decompile_count, not just md->decompile_count.
4137 if (log())
4138 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4139 Deoptimization::trap_reason_name(reason),
4140 md->trap_count(reason), trap_count(reason),
4141 md->decompile_count(), decompile_count());
4142 return true;
4143 } else {
4144 // The coast is clear.
4145 return false;
4146 }
4147 }
4148
4149 // Compute when not to trap. Used by matching trap based nodes and
4150 // NullCheck optimization.
4151 void Compile::set_allowed_deopt_reasons() {
4152 _allowed_reasons = 0;
4153 if (is_method_compilation()) {
4154 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4155 assert(rs < BitsPerInt, "recode bit map");
4156 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4157 _allowed_reasons |= nth_bit(rs);
4158 }
4159 }
4160 }
4161 }
4162
4163 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4164 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4165 }
4166
4167 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4168 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4169 }
4170
4171 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4172 if (holder->is_initialized()) {
4173 return false;
4174 }
4175 if (holder->is_being_initialized()) {
4176 if (accessing_method->holder() == holder) {
4177 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4178 // <init>, or a static method. In all those cases, there was an initialization
4179 // barrier on the holder klass passed.
4180 if (accessing_method->is_static_initializer() ||
4181 accessing_method->is_object_initializer() ||
4182 accessing_method->is_static()) {
4183 return false;
4184 }
4185 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4186 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4187 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4188 // child class can become fully initialized while its parent class is still being initialized.
4189 if (accessing_method->is_static_initializer()) {
4190 return false;
4191 }
4192 }
4193 ciMethod* root = method(); // the root method of compilation
4194 if (root != accessing_method) {
4195 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4196 }
4197 }
4198 return true;
4199 }
4200
4201 #ifndef PRODUCT
4202 //------------------------------verify_bidirectional_edges---------------------
4203 // For each input edge to a node (ie - for each Use-Def edge), verify that
4204 // there is a corresponding Def-Use edge.
4205 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4206 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4207 uint stack_size = live_nodes() >> 4;
4208 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4209 nstack.push(_root);
4210
4211 while (nstack.size() > 0) {
4212 Node* n = nstack.pop();
4213 if (visited.member(n)) {
4214 continue;
4215 }
4216 visited.push(n);
4217
4218 // Walk over all input edges, checking for correspondence
4219 uint length = n->len();
4220 for (uint i = 0; i < length; i++) {
4221 Node* in = n->in(i);
4222 if (in != nullptr && !visited.member(in)) {
4223 nstack.push(in); // Put it on stack
4224 }
4225 if (in != nullptr && !in->is_top()) {
4226 // Count instances of `next`
4227 int cnt = 0;
4228 for (uint idx = 0; idx < in->_outcnt; idx++) {
4229 if (in->_out[idx] == n) {
4230 cnt++;
4231 }
4232 }
4233 assert(cnt > 0, "Failed to find Def-Use edge.");
4234 // Check for duplicate edges
4235 // walk the input array downcounting the input edges to n
4236 for (uint j = 0; j < length; j++) {
4237 if (n->in(j) == in) {
4238 cnt--;
4239 }
4240 }
4241 assert(cnt == 0, "Mismatched edge count.");
4242 } else if (in == nullptr) {
4243 assert(i == 0 || i >= n->req() ||
4244 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4245 (n->is_Unlock() && i == (n->req() - 1)) ||
4246 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4247 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4248 } else {
4249 assert(in->is_top(), "sanity");
4250 // Nothing to check.
4251 }
4252 }
4253 }
4254 }
4255
4256 //------------------------------verify_graph_edges---------------------------
4257 // Walk the Graph and verify that there is a one-to-one correspondence
4258 // between Use-Def edges and Def-Use edges in the graph.
4259 void Compile::verify_graph_edges(bool no_dead_code) {
4260 if (VerifyGraphEdges) {
4261 Unique_Node_List visited;
4262
4263 // Call graph walk to check edges
4264 verify_bidirectional_edges(visited);
4265 if (no_dead_code) {
4266 // Now make sure that no visited node is used by an unvisited node.
4267 bool dead_nodes = false;
4268 Unique_Node_List checked;
4269 while (visited.size() > 0) {
4270 Node* n = visited.pop();
4271 checked.push(n);
4272 for (uint i = 0; i < n->outcnt(); i++) {
4273 Node* use = n->raw_out(i);
4274 if (checked.member(use)) continue; // already checked
4275 if (visited.member(use)) continue; // already in the graph
4276 if (use->is_Con()) continue; // a dead ConNode is OK
4277 // At this point, we have found a dead node which is DU-reachable.
4278 if (!dead_nodes) {
4279 tty->print_cr("*** Dead nodes reachable via DU edges:");
4280 dead_nodes = true;
4281 }
4282 use->dump(2);
4283 tty->print_cr("---");
4284 checked.push(use); // No repeats; pretend it is now checked.
4285 }
4286 }
4287 assert(!dead_nodes, "using nodes must be reachable from root");
4288 }
4289 }
4290 }
4291 #endif
4292
4293 // The Compile object keeps track of failure reasons separately from the ciEnv.
4294 // This is required because there is not quite a 1-1 relation between the
4295 // ciEnv and its compilation task and the Compile object. Note that one
4296 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4297 // to backtrack and retry without subsuming loads. Other than this backtracking
4298 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4299 // by the logic in C2Compiler.
4300 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4301 if (log() != nullptr) {
4302 log()->elem("failure reason='%s' phase='compile'", reason);
4303 }
4304 if (_failure_reason.get() == nullptr) {
4305 // Record the first failure reason.
4306 _failure_reason.set(reason);
4307 if (CaptureBailoutInformation) {
4308 _first_failure_details = new CompilationFailureInfo(reason);
4309 }
4310 } else {
4311 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4312 }
4313
4314 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4315 C->print_method(PHASE_FAILURE, 1);
4316 }
4317 _root = nullptr; // flush the graph, too
4318 }
4319
4320 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4321 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4322 _compile(Compile::current()),
4323 _log(nullptr),
4324 _dolog(CITimeVerbose)
4325 {
4326 assert(_compile != nullptr, "sanity check");
4327 if (_dolog) {
4328 _log = _compile->log();
4329 }
4330 if (_log != nullptr) {
4331 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4332 _log->stamp();
4333 _log->end_head();
4334 }
4335 }
4336
4337 Compile::TracePhase::TracePhase(PhaseTraceId id)
4338 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4339
4340 Compile::TracePhase::~TracePhase() {
4341 if (_compile->failing_internal()) {
4342 if (_log != nullptr) {
4343 _log->done("phase");
4344 }
4345 return; // timing code, not stressing bailouts.
4346 }
4347 #ifdef ASSERT
4348 if (PrintIdealNodeCount) {
4349 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4350 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4351 }
4352
4353 if (VerifyIdealNodeCount) {
4354 _compile->print_missing_nodes();
4355 }
4356 #endif
4357
4358 if (_log != nullptr) {
4359 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4360 }
4361 }
4362
4363 //----------------------------static_subtype_check-----------------------------
4364 // Shortcut important common cases when superklass is exact:
4365 // (0) superklass is java.lang.Object (can occur in reflective code)
4366 // (1) subklass is already limited to a subtype of superklass => always ok
4367 // (2) subklass does not overlap with superklass => always fail
4368 // (3) superklass has NO subtypes and we can check with a simple compare.
4369 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4370 if (skip) {
4371 return SSC_full_test; // Let caller generate the general case.
4372 }
4373
4374 if (subk->is_java_subtype_of(superk)) {
4375 return SSC_always_true; // (0) and (1) this test cannot fail
4376 }
4377
4378 if (!subk->maybe_java_subtype_of(superk)) {
4379 return SSC_always_false; // (2) true path dead; no dynamic test needed
4380 }
4381
4382 const Type* superelem = superk;
4383 if (superk->isa_aryklassptr()) {
4384 int ignored;
4385 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4386 }
4387
4388 if (superelem->isa_instklassptr()) {
4389 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4390 if (!ik->has_subklass()) {
4391 if (!ik->is_final()) {
4392 // Add a dependency if there is a chance of a later subclass.
4393 dependencies()->assert_leaf_type(ik);
4394 }
4395 if (!superk->maybe_java_subtype_of(subk)) {
4396 return SSC_always_false;
4397 }
4398 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4399 }
4400 } else {
4401 // A primitive array type has no subtypes.
4402 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4403 }
4404
4405 return SSC_full_test;
4406 }
4407
4408 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4409 #ifdef _LP64
4410 // The scaled index operand to AddP must be a clean 64-bit value.
4411 // Java allows a 32-bit int to be incremented to a negative
4412 // value, which appears in a 64-bit register as a large
4413 // positive number. Using that large positive number as an
4414 // operand in pointer arithmetic has bad consequences.
4415 // On the other hand, 32-bit overflow is rare, and the possibility
4416 // can often be excluded, if we annotate the ConvI2L node with
4417 // a type assertion that its value is known to be a small positive
4418 // number. (The prior range check has ensured this.)
4419 // This assertion is used by ConvI2LNode::Ideal.
4420 int index_max = max_jint - 1; // array size is max_jint, index is one less
4421 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4422 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4423 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4424 #endif
4425 return idx;
4426 }
4427
4428 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4429 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4430 if (ctrl != nullptr) {
4431 // Express control dependency by a CastII node with a narrow type.
4432 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4433 // node from floating above the range check during loop optimizations. Otherwise, the
4434 // ConvI2L node may be eliminated independently of the range check, causing the data path
4435 // to become TOP while the control path is still there (although it's unreachable).
4436 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4437 value = phase->transform(value);
4438 }
4439 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4440 return phase->transform(new ConvI2LNode(value, ltype));
4441 }
4442
4443 // The message about the current inlining is accumulated in
4444 // _print_inlining_stream and transferred into the _print_inlining_list
4445 // once we know whether inlining succeeds or not. For regular
4446 // inlining, messages are appended to the buffer pointed by
4447 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4448 // a new buffer is added after _print_inlining_idx in the list. This
4449 // way we can update the inlining message for late inlining call site
4450 // when the inlining is attempted again.
4451 void Compile::print_inlining_init() {
4452 if (print_inlining() || print_intrinsics()) {
4453 // print_inlining_init is actually called several times.
4454 print_inlining_reset();
4455 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4456 }
4457 }
4458
4459 void Compile::print_inlining_reinit() {
4460 if (print_inlining() || print_intrinsics()) {
4461 print_inlining_reset();
4462 }
4463 }
4464
4465 void Compile::print_inlining_reset() {
4466 _print_inlining_stream->reset();
4467 }
4468
4469 void Compile::print_inlining_commit() {
4470 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4471 // Transfer the message from _print_inlining_stream to the current
4472 // _print_inlining_list buffer and clear _print_inlining_stream.
4473 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4474 print_inlining_reset();
4475 }
4476
4477 void Compile::print_inlining_push() {
4478 // Add new buffer to the _print_inlining_list at current position
4479 _print_inlining_idx++;
4480 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4481 }
4482
4483 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4484 return _print_inlining_list->at(_print_inlining_idx);
4485 }
4486
4487 void Compile::print_inlining_update(CallGenerator* cg) {
4488 if (print_inlining() || print_intrinsics()) {
4489 if (cg->is_late_inline()) {
4490 if (print_inlining_current()->cg() != cg &&
4491 (print_inlining_current()->cg() != nullptr ||
4492 print_inlining_current()->ss()->size() != 0)) {
4493 print_inlining_push();
4494 }
4495 print_inlining_commit();
4496 print_inlining_current()->set_cg(cg);
4497 } else {
4498 if (print_inlining_current()->cg() != nullptr) {
4499 print_inlining_push();
4500 }
4501 print_inlining_commit();
4502 }
4503 }
4504 }
4505
4506 void Compile::print_inlining_move_to(CallGenerator* cg) {
4507 // We resume inlining at a late inlining call site. Locate the
4508 // corresponding inlining buffer so that we can update it.
4509 if (print_inlining() || print_intrinsics()) {
4510 for (int i = 0; i < _print_inlining_list->length(); i++) {
4511 if (_print_inlining_list->at(i)->cg() == cg) {
4512 _print_inlining_idx = i;
4513 return;
4514 }
4515 }
4516 ShouldNotReachHere();
4517 }
4518 }
4519
4520 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4521 if (print_inlining() || print_intrinsics()) {
4522 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4523 assert(print_inlining_current()->cg() == cg, "wrong entry");
4524 // replace message with new message
4525 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4526 print_inlining_commit();
4527 print_inlining_current()->set_cg(cg);
4528 }
4529 }
4530
4531 void Compile::print_inlining_assert_ready() {
4532 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4533 }
4534
4535 void Compile::process_print_inlining() {
4536 assert(_late_inlines.length() == 0, "not drained yet");
4537 if (print_inlining() || print_intrinsics()) {
4538 ResourceMark rm;
4539 stringStream ss;
4540 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4541 for (int i = 0; i < _print_inlining_list->length(); i++) {
4542 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4543 ss.print("%s", pib->ss()->freeze());
4544 delete pib;
4545 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4546 }
4547 // Reset _print_inlining_list, it only contains destructed objects.
4548 // It is on the arena, so it will be freed when the arena is reset.
4549 _print_inlining_list = nullptr;
4550 // _print_inlining_stream won't be used anymore, either.
4551 print_inlining_reset();
4552 size_t end = ss.size();
4553 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4554 strncpy(_print_inlining_output, ss.freeze(), end+1);
4555 _print_inlining_output[end] = 0;
4556 }
4557 }
4558
4559 void Compile::dump_print_inlining() {
4560 if (_print_inlining_output != nullptr) {
4561 tty->print_raw(_print_inlining_output);
4562 }
4563 }
4564
4565 void Compile::log_late_inline(CallGenerator* cg) {
4566 if (log() != nullptr) {
4567 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4568 cg->unique_id());
4569 JVMState* p = cg->call_node()->jvms();
4570 while (p != nullptr) {
4571 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4572 p = p->caller();
4573 }
4574 log()->tail("late_inline");
4575 }
4576 }
4577
4578 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4579 log_late_inline(cg);
4580 if (log() != nullptr) {
4581 log()->inline_fail(msg);
4582 }
4583 }
4584
4585 void Compile::log_inline_id(CallGenerator* cg) {
4586 if (log() != nullptr) {
4587 // The LogCompilation tool needs a unique way to identify late
4588 // inline call sites. This id must be unique for this call site in
4589 // this compilation. Try to have it unique across compilations as
4590 // well because it can be convenient when grepping through the log
4591 // file.
4592 // Distinguish OSR compilations from others in case CICountOSR is
4593 // on.
4594 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4595 cg->set_unique_id(id);
4596 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4597 }
4598 }
4599
4600 void Compile::log_inline_failure(const char* msg) {
4601 if (C->log() != nullptr) {
4602 C->log()->inline_fail(msg);
4603 }
4604 }
4605
4606
4607 // Dump inlining replay data to the stream.
4608 // Don't change thread state and acquire any locks.
4609 void Compile::dump_inline_data(outputStream* out) {
4610 InlineTree* inl_tree = ilt();
4611 if (inl_tree != nullptr) {
4612 out->print(" inline %d", inl_tree->count());
4613 inl_tree->dump_replay_data(out);
4614 }
4615 }
4616
4617 void Compile::dump_inline_data_reduced(outputStream* out) {
4618 assert(ReplayReduce, "");
4619
4620 InlineTree* inl_tree = ilt();
4621 if (inl_tree == nullptr) {
4622 return;
4623 }
4624 // Enable iterative replay file reduction
4625 // Output "compile" lines for depth 1 subtrees,
4626 // simulating that those trees were compiled
4627 // instead of inlined.
4628 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4629 InlineTree* sub = inl_tree->subtrees().at(i);
4630 if (sub->inline_level() != 1) {
4631 continue;
4632 }
4633
4634 ciMethod* method = sub->method();
4635 int entry_bci = -1;
4636 int comp_level = env()->task()->comp_level();
4637 out->print("compile ");
4638 method->dump_name_as_ascii(out);
4639 out->print(" %d %d", entry_bci, comp_level);
4640 out->print(" inline %d", sub->count());
4641 sub->dump_replay_data(out, -1);
4642 out->cr();
4643 }
4644 }
4645
4646 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4647 if (n1->Opcode() < n2->Opcode()) return -1;
4648 else if (n1->Opcode() > n2->Opcode()) return 1;
4649
4650 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4651 for (uint i = 1; i < n1->req(); i++) {
4652 if (n1->in(i) < n2->in(i)) return -1;
4653 else if (n1->in(i) > n2->in(i)) return 1;
4654 }
4655
4656 return 0;
4657 }
4658
4659 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4660 Node* n1 = *n1p;
4661 Node* n2 = *n2p;
4662
4663 return cmp_expensive_nodes(n1, n2);
4664 }
4665
4666 void Compile::sort_expensive_nodes() {
4667 if (!expensive_nodes_sorted()) {
4668 _expensive_nodes.sort(cmp_expensive_nodes);
4669 }
4670 }
4671
4672 bool Compile::expensive_nodes_sorted() const {
4673 for (int i = 1; i < _expensive_nodes.length(); i++) {
4674 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4675 return false;
4676 }
4677 }
4678 return true;
4679 }
4680
4681 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4682 if (_expensive_nodes.length() == 0) {
4683 return false;
4684 }
4685
4686 assert(OptimizeExpensiveOps, "optimization off?");
4687
4688 // Take this opportunity to remove dead nodes from the list
4689 int j = 0;
4690 for (int i = 0; i < _expensive_nodes.length(); i++) {
4691 Node* n = _expensive_nodes.at(i);
4692 if (!n->is_unreachable(igvn)) {
4693 assert(n->is_expensive(), "should be expensive");
4694 _expensive_nodes.at_put(j, n);
4695 j++;
4696 }
4697 }
4698 _expensive_nodes.trunc_to(j);
4699
4700 // Then sort the list so that similar nodes are next to each other
4701 // and check for at least two nodes of identical kind with same data
4702 // inputs.
4703 sort_expensive_nodes();
4704
4705 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4706 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4707 return true;
4708 }
4709 }
4710
4711 return false;
4712 }
4713
4714 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4715 if (_expensive_nodes.length() == 0) {
4716 return;
4717 }
4718
4719 assert(OptimizeExpensiveOps, "optimization off?");
4720
4721 // Sort to bring similar nodes next to each other and clear the
4722 // control input of nodes for which there's only a single copy.
4723 sort_expensive_nodes();
4724
4725 int j = 0;
4726 int identical = 0;
4727 int i = 0;
4728 bool modified = false;
4729 for (; i < _expensive_nodes.length()-1; i++) {
4730 assert(j <= i, "can't write beyond current index");
4731 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4732 identical++;
4733 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4734 continue;
4735 }
4736 if (identical > 0) {
4737 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4738 identical = 0;
4739 } else {
4740 Node* n = _expensive_nodes.at(i);
4741 igvn.replace_input_of(n, 0, nullptr);
4742 igvn.hash_insert(n);
4743 modified = true;
4744 }
4745 }
4746 if (identical > 0) {
4747 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4748 } else if (_expensive_nodes.length() >= 1) {
4749 Node* n = _expensive_nodes.at(i);
4750 igvn.replace_input_of(n, 0, nullptr);
4751 igvn.hash_insert(n);
4752 modified = true;
4753 }
4754 _expensive_nodes.trunc_to(j);
4755 if (modified) {
4756 igvn.optimize();
4757 }
4758 }
4759
4760 void Compile::add_expensive_node(Node * n) {
4761 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4762 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4763 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4764 if (OptimizeExpensiveOps) {
4765 _expensive_nodes.append(n);
4766 } else {
4767 // Clear control input and let IGVN optimize expensive nodes if
4768 // OptimizeExpensiveOps is off.
4769 n->set_req(0, nullptr);
4770 }
4771 }
4772
4773 /**
4774 * Track coarsened Lock and Unlock nodes.
4775 */
4776
4777 class Lock_List : public Node_List {
4778 uint _origin_cnt;
4779 public:
4780 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4781 uint origin_cnt() const { return _origin_cnt; }
4782 };
4783
4784 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4785 int length = locks.length();
4786 if (length > 0) {
4787 // Have to keep this list until locks elimination during Macro nodes elimination.
4788 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4789 AbstractLockNode* alock = locks.at(0);
4790 BoxLockNode* box = alock->box_node()->as_BoxLock();
4791 for (int i = 0; i < length; i++) {
4792 AbstractLockNode* lock = locks.at(i);
4793 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4794 locks_list->push(lock);
4795 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4796 if (this_box != box) {
4797 // Locking regions (BoxLock) could be Unbalanced here:
4798 // - its coarsened locks were eliminated in earlier
4799 // macro nodes elimination followed by loop unroll
4800 // - it is OSR locking region (no Lock node)
4801 // Preserve Unbalanced status in such cases.
4802 if (!this_box->is_unbalanced()) {
4803 this_box->set_coarsened();
4804 }
4805 if (!box->is_unbalanced()) {
4806 box->set_coarsened();
4807 }
4808 }
4809 }
4810 _coarsened_locks.append(locks_list);
4811 }
4812 }
4813
4814 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4815 int count = coarsened_count();
4816 for (int i = 0; i < count; i++) {
4817 Node_List* locks_list = _coarsened_locks.at(i);
4818 for (uint j = 0; j < locks_list->size(); j++) {
4819 Node* lock = locks_list->at(j);
4820 assert(lock->is_AbstractLock(), "sanity");
4821 if (!useful.member(lock)) {
4822 locks_list->yank(lock);
4823 }
4824 }
4825 }
4826 }
4827
4828 void Compile::remove_coarsened_lock(Node* n) {
4829 if (n->is_AbstractLock()) {
4830 int count = coarsened_count();
4831 for (int i = 0; i < count; i++) {
4832 Node_List* locks_list = _coarsened_locks.at(i);
4833 locks_list->yank(n);
4834 }
4835 }
4836 }
4837
4838 bool Compile::coarsened_locks_consistent() {
4839 int count = coarsened_count();
4840 for (int i = 0; i < count; i++) {
4841 bool unbalanced = false;
4842 bool modified = false; // track locks kind modifications
4843 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4844 uint size = locks_list->size();
4845 if (size == 0) {
4846 unbalanced = false; // All locks were eliminated - good
4847 } else if (size != locks_list->origin_cnt()) {
4848 unbalanced = true; // Some locks were removed from list
4849 } else {
4850 for (uint j = 0; j < size; j++) {
4851 Node* lock = locks_list->at(j);
4852 // All nodes in group should have the same state (modified or not)
4853 if (!lock->as_AbstractLock()->is_coarsened()) {
4854 if (j == 0) {
4855 // first on list was modified, the rest should be too for consistency
4856 modified = true;
4857 } else if (!modified) {
4858 // this lock was modified but previous locks on the list were not
4859 unbalanced = true;
4860 break;
4861 }
4862 } else if (modified) {
4863 // previous locks on list were modified but not this lock
4864 unbalanced = true;
4865 break;
4866 }
4867 }
4868 }
4869 if (unbalanced) {
4870 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4871 #ifdef ASSERT
4872 if (PrintEliminateLocks) {
4873 tty->print_cr("=== unbalanced coarsened locks ===");
4874 for (uint l = 0; l < size; l++) {
4875 locks_list->at(l)->dump();
4876 }
4877 }
4878 #endif
4879 record_failure(C2Compiler::retry_no_locks_coarsening());
4880 return false;
4881 }
4882 }
4883 return true;
4884 }
4885
4886 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4887 // locks coarsening optimization removed Lock/Unlock nodes from them.
4888 // Such regions become unbalanced because coarsening only removes part
4889 // of Lock/Unlock nodes in region. As result we can't execute other
4890 // locks elimination optimizations which assume all code paths have
4891 // corresponding pair of Lock/Unlock nodes - they are balanced.
4892 void Compile::mark_unbalanced_boxes() const {
4893 int count = coarsened_count();
4894 for (int i = 0; i < count; i++) {
4895 Node_List* locks_list = _coarsened_locks.at(i);
4896 uint size = locks_list->size();
4897 if (size > 0) {
4898 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4899 BoxLockNode* box = alock->box_node()->as_BoxLock();
4900 if (alock->is_coarsened()) {
4901 // coarsened_locks_consistent(), which is called before this method, verifies
4902 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4903 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4904 for (uint j = 1; j < size; j++) {
4905 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4906 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4907 if (box != this_box) {
4908 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4909 box->set_unbalanced();
4910 this_box->set_unbalanced();
4911 }
4912 }
4913 }
4914 }
4915 }
4916 }
4917
4918 /**
4919 * Remove the speculative part of types and clean up the graph
4920 */
4921 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4922 if (UseTypeSpeculation) {
4923 Unique_Node_List worklist;
4924 worklist.push(root());
4925 int modified = 0;
4926 // Go over all type nodes that carry a speculative type, drop the
4927 // speculative part of the type and enqueue the node for an igvn
4928 // which may optimize it out.
4929 for (uint next = 0; next < worklist.size(); ++next) {
4930 Node *n = worklist.at(next);
4931 if (n->is_Type()) {
4932 TypeNode* tn = n->as_Type();
4933 const Type* t = tn->type();
4934 const Type* t_no_spec = t->remove_speculative();
4935 if (t_no_spec != t) {
4936 bool in_hash = igvn.hash_delete(n);
4937 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4938 tn->set_type(t_no_spec);
4939 igvn.hash_insert(n);
4940 igvn._worklist.push(n); // give it a chance to go away
4941 modified++;
4942 }
4943 }
4944 // Iterate over outs - endless loops is unreachable from below
4945 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4946 Node *m = n->fast_out(i);
4947 if (not_a_node(m)) {
4948 continue;
4949 }
4950 worklist.push(m);
4951 }
4952 }
4953 // Drop the speculative part of all types in the igvn's type table
4954 igvn.remove_speculative_types();
4955 if (modified > 0) {
4956 igvn.optimize();
4957 if (failing()) return;
4958 }
4959 #ifdef ASSERT
4960 // Verify that after the IGVN is over no speculative type has resurfaced
4961 worklist.clear();
4962 worklist.push(root());
4963 for (uint next = 0; next < worklist.size(); ++next) {
4964 Node *n = worklist.at(next);
4965 const Type* t = igvn.type_or_null(n);
4966 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4967 if (n->is_Type()) {
4968 t = n->as_Type()->type();
4969 assert(t == t->remove_speculative(), "no more speculative types");
4970 }
4971 // Iterate over outs - endless loops is unreachable from below
4972 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4973 Node *m = n->fast_out(i);
4974 if (not_a_node(m)) {
4975 continue;
4976 }
4977 worklist.push(m);
4978 }
4979 }
4980 igvn.check_no_speculative_types();
4981 #endif
4982 }
4983 }
4984
4985 // Auxiliary methods to support randomized stressing/fuzzing.
4986
4987 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
4988 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
4989 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
4990 FLAG_SET_ERGO(StressSeed, _stress_seed);
4991 } else {
4992 _stress_seed = StressSeed;
4993 }
4994 if (_log != nullptr) {
4995 _log->elem("stress_test seed='%u'", _stress_seed);
4996 }
4997 }
4998
4999 int Compile::random() {
5000 _stress_seed = os::next_random(_stress_seed);
5001 return static_cast<int>(_stress_seed);
5002 }
5003
5004 // This method can be called the arbitrary number of times, with current count
5005 // as the argument. The logic allows selecting a single candidate from the
5006 // running list of candidates as follows:
5007 // int count = 0;
5008 // Cand* selected = null;
5009 // while(cand = cand->next()) {
5010 // if (randomized_select(++count)) {
5011 // selected = cand;
5012 // }
5013 // }
5014 //
5015 // Including count equalizes the chances any candidate is "selected".
5016 // This is useful when we don't have the complete list of candidates to choose
5017 // from uniformly. In this case, we need to adjust the randomicity of the
5018 // selection, or else we will end up biasing the selection towards the latter
5019 // candidates.
5020 //
5021 // Quick back-envelope calculation shows that for the list of n candidates
5022 // the equal probability for the candidate to persist as "best" can be
5023 // achieved by replacing it with "next" k-th candidate with the probability
5024 // of 1/k. It can be easily shown that by the end of the run, the
5025 // probability for any candidate is converged to 1/n, thus giving the
5026 // uniform distribution among all the candidates.
5027 //
5028 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5029 #define RANDOMIZED_DOMAIN_POW 29
5030 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5031 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5032 bool Compile::randomized_select(int count) {
5033 assert(count > 0, "only positive");
5034 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5035 }
5036
5037 #ifdef ASSERT
5038 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5039 bool Compile::fail_randomly() {
5040 if ((random() % StressBailoutMean) != 0) {
5041 return false;
5042 }
5043 record_failure("StressBailout");
5044 return true;
5045 }
5046
5047 bool Compile::failure_is_artificial() {
5048 return C->failure_reason_is("StressBailout");
5049 }
5050 #endif
5051
5052 CloneMap& Compile::clone_map() { return _clone_map; }
5053 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5054
5055 void NodeCloneInfo::dump_on(outputStream* st) const {
5056 st->print(" {%d:%d} ", idx(), gen());
5057 }
5058
5059 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5060 uint64_t val = value(old->_idx);
5061 NodeCloneInfo cio(val);
5062 assert(val != 0, "old node should be in the map");
5063 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5064 insert(nnn->_idx, cin.get());
5065 #ifndef PRODUCT
5066 if (is_debug()) {
5067 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5068 }
5069 #endif
5070 }
5071
5072 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5073 NodeCloneInfo cio(value(old->_idx));
5074 if (cio.get() == 0) {
5075 cio.set(old->_idx, 0);
5076 insert(old->_idx, cio.get());
5077 #ifndef PRODUCT
5078 if (is_debug()) {
5079 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5080 }
5081 #endif
5082 }
5083 clone(old, nnn, gen);
5084 }
5085
5086 int CloneMap::max_gen() const {
5087 int g = 0;
5088 DictI di(_dict);
5089 for(; di.test(); ++di) {
5090 int t = gen(di._key);
5091 if (g < t) {
5092 g = t;
5093 #ifndef PRODUCT
5094 if (is_debug()) {
5095 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5096 }
5097 #endif
5098 }
5099 }
5100 return g;
5101 }
5102
5103 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5104 uint64_t val = value(key);
5105 if (val != 0) {
5106 NodeCloneInfo ni(val);
5107 ni.dump_on(st);
5108 }
5109 }
5110
5111 void Compile::shuffle_macro_nodes() {
5112 if (_macro_nodes.length() < 2) {
5113 return;
5114 }
5115 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5116 uint j = C->random() % (i + 1);
5117 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5118 }
5119 }
5120
5121 // Move Allocate nodes to the start of the list
5122 void Compile::sort_macro_nodes() {
5123 int count = macro_count();
5124 int allocates = 0;
5125 for (int i = 0; i < count; i++) {
5126 Node* n = macro_node(i);
5127 if (n->is_Allocate()) {
5128 if (i != allocates) {
5129 Node* tmp = macro_node(allocates);
5130 _macro_nodes.at_put(allocates, n);
5131 _macro_nodes.at_put(i, tmp);
5132 }
5133 allocates++;
5134 }
5135 }
5136 }
5137
5138 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5139 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5140 EventCompilerPhase event(UNTIMED);
5141 if (event.should_commit()) {
5142 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5143 }
5144 #ifndef PRODUCT
5145 ResourceMark rm;
5146 stringStream ss;
5147 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5148 int iter = ++_igv_phase_iter[cpt];
5149 if (iter > 1) {
5150 ss.print(" %d", iter);
5151 }
5152 if (n != nullptr) {
5153 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5154 if (n->is_Call()) {
5155 CallNode* call = n->as_Call();
5156 if (call->_name != nullptr) {
5157 // E.g. uncommon traps etc.
5158 ss.print(" - %s", call->_name);
5159 } else if (call->is_CallJava()) {
5160 CallJavaNode* call_java = call->as_CallJava();
5161 if (call_java->method() != nullptr) {
5162 ss.print(" -");
5163 call_java->method()->print_short_name(&ss);
5164 }
5165 }
5166 }
5167 }
5168
5169 const char* name = ss.as_string();
5170 if (should_print_igv(level)) {
5171 _igv_printer->print_graph(name);
5172 }
5173 if (should_print_phase(cpt)) {
5174 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5175 }
5176 #endif
5177 C->_latest_stage_start_counter.stamp();
5178 }
5179
5180 // Only used from CompileWrapper
5181 void Compile::begin_method() {
5182 #ifndef PRODUCT
5183 if (_method != nullptr && should_print_igv(1)) {
5184 _igv_printer->begin_method();
5185 }
5186 #endif
5187 C->_latest_stage_start_counter.stamp();
5188 }
5189
5190 // Only used from CompileWrapper
5191 void Compile::end_method() {
5192 EventCompilerPhase event(UNTIMED);
5193 if (event.should_commit()) {
5194 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5195 }
5196
5197 #ifndef PRODUCT
5198 if (_method != nullptr && should_print_igv(1)) {
5199 _igv_printer->end_method();
5200 }
5201 #endif
5202 }
5203
5204 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5205 #ifndef PRODUCT
5206 if (_directive->should_print_phase(cpt)) {
5207 return true;
5208 }
5209 #endif
5210 return false;
5211 }
5212
5213 #ifndef PRODUCT
5214 void Compile::init_igv() {
5215 if (_igv_printer == nullptr) {
5216 _igv_printer = IdealGraphPrinter::printer();
5217 _igv_printer->set_compile(this);
5218 }
5219 }
5220 #endif
5221
5222 bool Compile::should_print_igv(const int level) {
5223 #ifndef PRODUCT
5224 if (PrintIdealGraphLevel < 0) { // disabled by the user
5225 return false;
5226 }
5227
5228 bool need = directive()->IGVPrintLevelOption >= level;
5229 if (need) {
5230 Compile::init_igv();
5231 }
5232 return need;
5233 #else
5234 return false;
5235 #endif
5236 }
5237
5238 #ifndef PRODUCT
5239 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5240 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5241
5242 // Called from debugger. Prints method to the default file with the default phase name.
5243 // This works regardless of any Ideal Graph Visualizer flags set or not.
5244 void igv_print() {
5245 Compile::current()->igv_print_method_to_file();
5246 }
5247
5248 // Same as igv_print() above but with a specified phase name.
5249 void igv_print(const char* phase_name) {
5250 Compile::current()->igv_print_method_to_file(phase_name);
5251 }
5252
5253 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5254 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5255 // This works regardless of any Ideal Graph Visualizer flags set or not.
5256 void igv_print(bool network) {
5257 if (network) {
5258 Compile::current()->igv_print_method_to_network();
5259 } else {
5260 Compile::current()->igv_print_method_to_file();
5261 }
5262 }
5263
5264 // Same as igv_print(bool network) above but with a specified phase name.
5265 void igv_print(bool network, const char* phase_name) {
5266 if (network) {
5267 Compile::current()->igv_print_method_to_network(phase_name);
5268 } else {
5269 Compile::current()->igv_print_method_to_file(phase_name);
5270 }
5271 }
5272
5273 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5274 void igv_print_default() {
5275 Compile::current()->print_method(PHASE_DEBUG, 0);
5276 }
5277
5278 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5279 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5280 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5281 void igv_append() {
5282 Compile::current()->igv_print_method_to_file("Debug", true);
5283 }
5284
5285 // Same as igv_append() above but with a specified phase name.
5286 void igv_append(const char* phase_name) {
5287 Compile::current()->igv_print_method_to_file(phase_name, true);
5288 }
5289
5290 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5291 const char* file_name = "custom_debug.xml";
5292 if (_debug_file_printer == nullptr) {
5293 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5294 } else {
5295 _debug_file_printer->update_compiled_method(C->method());
5296 }
5297 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5298 _debug_file_printer->print_graph(phase_name);
5299 }
5300
5301 void Compile::igv_print_method_to_network(const char* phase_name) {
5302 ResourceMark rm;
5303 GrowableArray<const Node*> empty_list;
5304 igv_print_graph_to_network(phase_name, (Node*) C->root(), empty_list);
5305 }
5306
5307 void Compile::igv_print_graph_to_network(const char* name, Node* node, GrowableArray<const Node*>& visible_nodes) {
5308 if (_debug_network_printer == nullptr) {
5309 _debug_network_printer = new IdealGraphPrinter(C);
5310 } else {
5311 _debug_network_printer->update_compiled_method(C->method());
5312 }
5313 tty->print_cr("Method printed over network stream to IGV");
5314 _debug_network_printer->print(name, C->root(), visible_nodes);
5315 }
5316 #endif
5317
5318 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5319 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5320 return value;
5321 }
5322 Node* result = nullptr;
5323 if (bt == T_BYTE) {
5324 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5325 result = new RShiftINode(result, phase->intcon(24));
5326 } else if (bt == T_BOOLEAN) {
5327 result = new AndINode(value, phase->intcon(0xFF));
5328 } else if (bt == T_CHAR) {
5329 result = new AndINode(value,phase->intcon(0xFFFF));
5330 } else {
5331 assert(bt == T_SHORT, "unexpected narrow type");
5332 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5333 result = new RShiftINode(result, phase->intcon(16));
5334 }
5335 if (transform_res) {
5336 result = phase->transform(result);
5337 }
5338 return result;
5339 }
5340
5341 void Compile::record_method_not_compilable_oom() {
5342 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5343 }