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->Opcode() == Op_Opaque4) {
397 remove_template_assertion_predicate_opaq(dead);
398 }
399 if (dead->is_ParsePredicate()) {
400 remove_parse_predicate(dead->as_ParsePredicate());
401 }
402 if (dead->for_post_loop_opts_igvn()) {
403 remove_from_post_loop_opts_igvn(dead);
404 }
405 if (dead->is_Call()) {
406 remove_useless_late_inlines( &_late_inlines, dead);
407 remove_useless_late_inlines( &_string_late_inlines, dead);
408 remove_useless_late_inlines( &_boxing_late_inlines, dead);
409 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
410
411 if (dead->is_CallStaticJava()) {
412 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
413 }
414 }
415 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
416 bs->unregister_potential_barrier_node(dead);
417 }
418
419 // Disconnect all useless nodes by disconnecting those at the boundary.
420 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
421 uint next = 0;
422 while (next < useful.size()) {
423 Node *n = useful.at(next++);
424 if (n->is_SafePoint()) {
425 // We're done with a parsing phase. Replaced nodes are not valid
426 // beyond that point.
427 n->as_SafePoint()->delete_replaced_nodes();
428 }
429 // Use raw traversal of out edges since this code removes out edges
430 int max = n->outcnt();
431 for (int j = 0; j < max; ++j) {
432 Node* child = n->raw_out(j);
433 if (!useful.member(child)) {
434 assert(!child->is_top() || child != top(),
435 "If top is cached in Compile object it is in useful list");
436 // Only need to remove this out-edge to the useless node
437 n->raw_del_out(j);
438 --j;
439 --max;
440 }
441 }
442 if (n->outcnt() == 1 && n->has_special_unique_user()) {
443 assert(useful.member(n->unique_out()), "do not push a useless node");
444 worklist.push(n->unique_out());
445 }
446 }
447
448 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
449 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
450 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
451 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
452 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
453 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
454 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
455 #ifdef ASSERT
456 if (_modified_nodes != nullptr) {
457 _modified_nodes->remove_useless_nodes(useful.member_set());
458 }
459 #endif
460
461 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
462 bs->eliminate_useless_gc_barriers(useful, this);
463 // clean up the late inline lists
464 remove_useless_late_inlines( &_late_inlines, useful);
465 remove_useless_late_inlines( &_string_late_inlines, useful);
466 remove_useless_late_inlines( &_boxing_late_inlines, useful);
467 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
468 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
469 }
470
471 // ============================================================================
472 //------------------------------CompileWrapper---------------------------------
473 class CompileWrapper : public StackObj {
474 Compile *const _compile;
475 public:
476 CompileWrapper(Compile* compile);
477
478 ~CompileWrapper();
479 };
480
481 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
482 // the Compile* pointer is stored in the current ciEnv:
483 ciEnv* env = compile->env();
484 assert(env == ciEnv::current(), "must already be a ciEnv active");
485 assert(env->compiler_data() == nullptr, "compile already active?");
486 env->set_compiler_data(compile);
487 assert(compile == Compile::current(), "sanity");
488
489 compile->set_type_dict(nullptr);
490 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
491 compile->clone_map().set_clone_idx(0);
492 compile->set_type_last_size(0);
493 compile->set_last_tf(nullptr, nullptr);
494 compile->set_indexSet_arena(nullptr);
495 compile->set_indexSet_free_block_list(nullptr);
496 compile->init_type_arena();
497 Type::Initialize(compile);
498 _compile->begin_method();
499 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
500 }
501 CompileWrapper::~CompileWrapper() {
502 // simulate crash during compilation
503 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
504
505 _compile->end_method();
506 _compile->env()->set_compiler_data(nullptr);
507 }
508
509
510 //----------------------------print_compile_messages---------------------------
511 void Compile::print_compile_messages() {
512 #ifndef PRODUCT
513 // Check if recompiling
514 if (!subsume_loads() && PrintOpto) {
515 // Recompiling without allowing machine instructions to subsume loads
516 tty->print_cr("*********************************************************");
517 tty->print_cr("** Bailout: Recompile without subsuming loads **");
518 tty->print_cr("*********************************************************");
519 }
520 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
521 // Recompiling without escape analysis
522 tty->print_cr("*********************************************************");
523 tty->print_cr("** Bailout: Recompile without escape analysis **");
524 tty->print_cr("*********************************************************");
525 }
526 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
527 // Recompiling without iterative escape analysis
528 tty->print_cr("*********************************************************");
529 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
530 tty->print_cr("*********************************************************");
531 }
532 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
533 // Recompiling without reducing allocation merges
534 tty->print_cr("*********************************************************");
535 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
536 tty->print_cr("*********************************************************");
537 }
538 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
539 // Recompiling without boxing elimination
540 tty->print_cr("*********************************************************");
541 tty->print_cr("** Bailout: Recompile without boxing elimination **");
542 tty->print_cr("*********************************************************");
543 }
544 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
545 // Recompiling without locks coarsening
546 tty->print_cr("*********************************************************");
547 tty->print_cr("** Bailout: Recompile without locks coarsening **");
548 tty->print_cr("*********************************************************");
549 }
550 if (env()->break_at_compile()) {
551 // Open the debugger when compiling this method.
552 tty->print("### Breaking when compiling: ");
553 method()->print_short_name();
554 tty->cr();
555 BREAKPOINT;
556 }
557
558 if( PrintOpto ) {
559 if (is_osr_compilation()) {
560 tty->print("[OSR]%3d", _compile_id);
561 } else {
562 tty->print("%3d", _compile_id);
563 }
564 }
565 #endif
566 }
567
568 #ifndef PRODUCT
569 void Compile::print_ideal_ir(const char* phase_name) {
570 // keep the following output all in one block
571 // This output goes directly to the tty, not the compiler log.
572 // To enable tools to match it up with the compilation activity,
573 // be sure to tag this tty output with the compile ID.
574
575 // Node dumping can cause a safepoint, which can break the tty lock.
576 // Buffer all node dumps, so that all safepoints happen before we lock.
577 ResourceMark rm;
578 stringStream ss;
579
580 if (_output == nullptr) {
581 ss.print_cr("AFTER: %s", phase_name);
582 // Print out all nodes in ascending order of index.
583 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
584 } else {
585 // Dump the node blockwise if we have a scheduling
586 _output->print_scheduling(&ss);
587 }
588
589 // Check that the lock is not broken by a safepoint.
590 NoSafepointVerifier nsv;
591 ttyLocker ttyl;
592 if (xtty != nullptr) {
593 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
594 compile_id(),
595 is_osr_compilation() ? " compile_kind='osr'" : "",
596 phase_name);
597 }
598
599 tty->print("%s", ss.as_string());
600
601 if (xtty != nullptr) {
602 xtty->tail("ideal");
603 }
604 }
605 #endif
606
607 // ============================================================================
608 //------------------------------Compile standard-------------------------------
609
610 // Compile a method. entry_bci is -1 for normal compilations and indicates
611 // the continuation bci for on stack replacement.
612
613
614 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
615 Options options, DirectiveSet* directive)
616 : Phase(Compiler),
617 _compile_id(ci_env->compile_id()),
618 _options(options),
619 _method(target),
620 _entry_bci(osr_bci),
621 _ilt(nullptr),
622 _stub_function(nullptr),
623 _stub_name(nullptr),
624 _stub_entry_point(nullptr),
625 _max_node_limit(MaxNodeLimit),
626 _post_loop_opts_phase(false),
627 _allow_macro_nodes(true),
628 _inlining_progress(false),
629 _inlining_incrementally(false),
630 _do_cleanup(false),
631 _has_reserved_stack_access(target->has_reserved_stack_access()),
632 #ifndef PRODUCT
633 _igv_idx(0),
634 _trace_opto_output(directive->TraceOptoOutputOption),
635 #endif
636 _has_method_handle_invokes(false),
637 _clinit_barrier_on_entry(false),
638 _has_clinit_barriers(false),
639 _stress_seed(0),
640 _comp_arena(mtCompiler),
641 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
642 _env(ci_env),
643 _directive(directive),
644 _log(ci_env->log()),
645 _first_failure_details(nullptr),
646 _intrinsics (comp_arena(), 0, 0, nullptr),
647 _macro_nodes (comp_arena(), 8, 0, nullptr),
648 _parse_predicates (comp_arena(), 8, 0, nullptr),
649 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
650 _expensive_nodes (comp_arena(), 8, 0, nullptr),
651 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
652 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
653 _coarsened_locks (comp_arena(), 8, 0, nullptr),
654 _congraph(nullptr),
655 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
656 _unique(0),
657 _dead_node_count(0),
658 _dead_node_list(comp_arena()),
659 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
660 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
661 _node_arena(&_node_arena_one),
662 _mach_constant_base_node(nullptr),
663 _Compile_types(mtCompiler),
664 _initial_gvn(nullptr),
665 _igvn_worklist(nullptr),
666 _types(nullptr),
667 _node_hash(nullptr),
668 _late_inlines(comp_arena(), 2, 0, nullptr),
669 _string_late_inlines(comp_arena(), 2, 0, nullptr),
670 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
671 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
672 _late_inlines_pos(0),
673 _number_of_mh_late_inlines(0),
674 _oom(false),
675 _print_inlining_stream(new (mtCompiler) stringStream()),
676 _print_inlining_list(nullptr),
677 _print_inlining_idx(0),
678 _print_inlining_output(nullptr),
679 _replay_inline_data(nullptr),
680 _java_calls(0),
681 _inner_loops(0),
682 _interpreter_frame_size(0),
683 _output(nullptr)
684 #ifndef PRODUCT
685 , _in_dump_cnt(0)
686 #endif
687 {
688 C = this;
689 CompileWrapper cw(this);
690
691 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
692 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
693
694 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
695 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
696 // We can always print a disassembly, either abstract (hex dump) or
697 // with the help of a suitable hsdis library. Thus, we should not
698 // couple print_assembly and print_opto_assembly controls.
699 // But: always print opto and regular assembly on compile command 'print'.
700 bool print_assembly = directive->PrintAssemblyOption;
701 set_print_assembly(print_opto_assembly || print_assembly);
702 #else
703 set_print_assembly(false); // must initialize.
704 #endif
705
706 #ifndef PRODUCT
707 set_parsed_irreducible_loop(false);
708 #endif
709
710 if (directive->ReplayInlineOption) {
711 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
712 }
713 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
714 set_print_intrinsics(directive->PrintIntrinsicsOption);
715 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
716
717 if (ProfileTraps) {
718 // Make sure the method being compiled gets its own MDO,
719 // so we can at least track the decompile_count().
720 method()->ensure_method_data();
721 }
722
723 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
724 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout) {
725 initialize_stress_seed(directive);
726 }
727
728 Init(/*do_aliasing=*/ true);
729
730 print_compile_messages();
731
732 _ilt = InlineTree::build_inline_tree_root();
733
734 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
735 assert(num_alias_types() >= AliasIdxRaw, "");
736
737 #define MINIMUM_NODE_HASH 1023
738
739 // GVN that will be run immediately on new nodes
740 uint estimated_size = method()->code_size()*4+64;
741 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
742 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
743 _types = new (comp_arena()) Type_Array(comp_arena());
744 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
745 PhaseGVN gvn;
746 set_initial_gvn(&gvn);
747
748 print_inlining_init();
749 { // Scope for timing the parser
750 TracePhase tp("parse", &timers[_t_parser]);
751
752 // Put top into the hash table ASAP.
753 initial_gvn()->transform(top());
754
755 // Set up tf(), start(), and find a CallGenerator.
756 CallGenerator* cg = nullptr;
757 if (is_osr_compilation()) {
758 const TypeTuple *domain = StartOSRNode::osr_domain();
759 const TypeTuple *range = TypeTuple::make_range(method()->signature());
760 init_tf(TypeFunc::make(domain, range));
761 StartNode* s = new StartOSRNode(root(), domain);
762 initial_gvn()->set_type_bottom(s);
763 verify_start(s);
764 cg = CallGenerator::for_osr(method(), entry_bci());
765 } else {
766 // Normal case.
767 init_tf(TypeFunc::make(method()));
768 StartNode* s = new StartNode(root(), tf()->domain());
769 initial_gvn()->set_type_bottom(s);
770 verify_start(s);
771 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
772 // With java.lang.ref.reference.get() we must go through the
773 // intrinsic - even when get() is the root
774 // method of the compile - so that, if necessary, the value in
775 // the referent field of the reference object gets recorded by
776 // the pre-barrier code.
777 cg = find_intrinsic(method(), false);
778 }
779 if (cg == nullptr) {
780 float past_uses = method()->interpreter_invocation_count();
781 float expected_uses = past_uses;
782 cg = CallGenerator::for_inline(method(), expected_uses);
783 }
784 }
785 if (failing()) return;
786 if (cg == nullptr) {
787 const char* reason = InlineTree::check_can_parse(method());
788 assert(reason != nullptr, "expect reason for parse failure");
789 stringStream ss;
790 ss.print("cannot parse method: %s", reason);
791 record_method_not_compilable(ss.as_string());
792 return;
793 }
794
795 gvn.set_type(root(), root()->bottom_type());
796
797 JVMState* jvms = build_start_state(start(), tf());
798 if ((jvms = cg->generate(jvms)) == nullptr) {
799 assert(failure_reason() != nullptr, "expect reason for parse failure");
800 stringStream ss;
801 ss.print("method parse failed: %s", failure_reason());
802 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
803 return;
804 }
805 GraphKit kit(jvms);
806
807 if (!kit.stopped()) {
808 // Accept return values, and transfer control we know not where.
809 // This is done by a special, unique ReturnNode bound to root.
810 return_values(kit.jvms());
811 }
812
813 if (kit.has_exceptions()) {
814 // Any exceptions that escape from this call must be rethrown
815 // to whatever caller is dynamically above us on the stack.
816 // This is done by a special, unique RethrowNode bound to root.
817 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
818 }
819
820 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
821
822 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
823 inline_string_calls(true);
824 }
825
826 if (failing()) return;
827
828 // Remove clutter produced by parsing.
829 if (!failing()) {
830 ResourceMark rm;
831 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
832 }
833 }
834
835 // Note: Large methods are capped off in do_one_bytecode().
836 if (failing()) return;
837
838 // After parsing, node notes are no longer automagic.
839 // They must be propagated by register_new_node_with_optimizer(),
840 // clone(), or the like.
841 set_default_node_notes(nullptr);
842
843 #ifndef PRODUCT
844 if (should_print_igv(1)) {
845 _igv_printer->print_inlining();
846 }
847 #endif
848
849 if (failing()) return;
850 NOT_PRODUCT( verify_graph_edges(); )
851
852 // Now optimize
853 Optimize();
854 if (failing()) return;
855 NOT_PRODUCT( verify_graph_edges(); )
856
857 #ifndef PRODUCT
858 if (should_print_ideal()) {
859 print_ideal_ir("print_ideal");
860 }
861 #endif
862
863 #ifdef ASSERT
864 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
865 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
866 #endif
867
868 // Dump compilation data to replay it.
869 if (directive->DumpReplayOption) {
870 env()->dump_replay_data(_compile_id);
871 }
872 if (directive->DumpInlineOption && (ilt() != nullptr)) {
873 env()->dump_inline_data(_compile_id);
874 }
875
876 // Now that we know the size of all the monitors we can add a fixed slot
877 // for the original deopt pc.
878 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
879 set_fixed_slots(next_slot);
880
881 // Compute when to use implicit null checks. Used by matching trap based
882 // nodes and NullCheck optimization.
883 set_allowed_deopt_reasons();
884
885 // Now generate code
886 Code_Gen();
887 }
888
889 //------------------------------Compile----------------------------------------
890 // Compile a runtime stub
891 Compile::Compile( ciEnv* ci_env,
892 TypeFunc_generator generator,
893 address stub_function,
894 const char *stub_name,
895 int is_fancy_jump,
896 bool pass_tls,
897 bool return_pc,
898 DirectiveSet* directive)
899 : Phase(Compiler),
900 _compile_id(0),
901 _options(Options::for_runtime_stub()),
902 _method(nullptr),
903 _entry_bci(InvocationEntryBci),
904 _stub_function(stub_function),
905 _stub_name(stub_name),
906 _stub_entry_point(nullptr),
907 _max_node_limit(MaxNodeLimit),
908 _post_loop_opts_phase(false),
909 _allow_macro_nodes(true),
910 _inlining_progress(false),
911 _inlining_incrementally(false),
912 _has_reserved_stack_access(false),
913 #ifndef PRODUCT
914 _igv_idx(0),
915 _trace_opto_output(directive->TraceOptoOutputOption),
916 #endif
917 _has_method_handle_invokes(false),
918 _clinit_barrier_on_entry(false),
919 _has_clinit_barriers(false),
920 _stress_seed(0),
921 _comp_arena(mtCompiler),
922 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
923 _env(ci_env),
924 _directive(directive),
925 _log(ci_env->log()),
926 _first_failure_details(nullptr),
927 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
928 _congraph(nullptr),
929 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
930 _unique(0),
931 _dead_node_count(0),
932 _dead_node_list(comp_arena()),
933 _node_arena_one(mtCompiler),
934 _node_arena_two(mtCompiler),
935 _node_arena(&_node_arena_one),
936 _mach_constant_base_node(nullptr),
937 _Compile_types(mtCompiler),
938 _initial_gvn(nullptr),
939 _igvn_worklist(nullptr),
940 _types(nullptr),
941 _node_hash(nullptr),
942 _number_of_mh_late_inlines(0),
943 _oom(false),
944 _print_inlining_stream(new (mtCompiler) stringStream()),
945 _print_inlining_list(nullptr),
946 _print_inlining_idx(0),
947 _print_inlining_output(nullptr),
948 _replay_inline_data(nullptr),
949 _java_calls(0),
950 _inner_loops(0),
951 _interpreter_frame_size(0),
952 _output(nullptr),
953 #ifndef PRODUCT
954 _in_dump_cnt(0),
955 #endif
956 _allowed_reasons(0) {
957 C = this;
958
959 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
960 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
961
962 #ifndef PRODUCT
963 set_print_assembly(PrintFrameConverterAssembly);
964 set_parsed_irreducible_loop(false);
965 #else
966 set_print_assembly(false); // Must initialize.
967 #endif
968 set_has_irreducible_loop(false); // no loops
969
970 CompileWrapper cw(this);
971 Init(/*do_aliasing=*/ false);
972 init_tf((*generator)());
973
974 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
975 _types = new (comp_arena()) Type_Array(comp_arena());
976 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
977
978 if (StressLCM || StressGCM || StressBailout) {
979 initialize_stress_seed(directive);
980 }
981
982 {
983 PhaseGVN gvn;
984 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
985 gvn.transform(top());
986
987 GraphKit kit;
988 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
989 }
990
991 NOT_PRODUCT( verify_graph_edges(); )
992
993 Code_Gen();
994 }
995
996 Compile::~Compile() {
997 delete _print_inlining_stream;
998 delete _first_failure_details;
999 };
1000
1001 //------------------------------Init-------------------------------------------
1002 // Prepare for a single compilation
1003 void Compile::Init(bool aliasing) {
1004 _do_aliasing = aliasing;
1005 _unique = 0;
1006 _regalloc = nullptr;
1007
1008 _tf = nullptr; // filled in later
1009 _top = nullptr; // cached later
1010 _matcher = nullptr; // filled in later
1011 _cfg = nullptr; // filled in later
1012
1013 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1014
1015 _node_note_array = nullptr;
1016 _default_node_notes = nullptr;
1017 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1018
1019 _immutable_memory = nullptr; // filled in at first inquiry
1020
1021 #ifdef ASSERT
1022 _phase_optimize_finished = false;
1023 _phase_verify_ideal_loop = false;
1024 _exception_backedge = false;
1025 _type_verify = nullptr;
1026 #endif
1027
1028 // Globally visible Nodes
1029 // First set TOP to null to give safe behavior during creation of RootNode
1030 set_cached_top_node(nullptr);
1031 set_root(new RootNode());
1032 // Now that you have a Root to point to, create the real TOP
1033 set_cached_top_node( new ConNode(Type::TOP) );
1034 set_recent_alloc(nullptr, nullptr);
1035
1036 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1037 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1038 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1039 env()->set_dependencies(new Dependencies(env()));
1040
1041 _fixed_slots = 0;
1042 set_has_split_ifs(false);
1043 set_has_loops(false); // first approximation
1044 set_has_stringbuilder(false);
1045 set_has_boxed_value(false);
1046 _trap_can_recompile = false; // no traps emitted yet
1047 _major_progress = true; // start out assuming good things will happen
1048 set_has_unsafe_access(false);
1049 set_max_vector_size(0);
1050 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1051 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1052 set_decompile_count(0);
1053
1054 #ifndef PRODUCT
1055 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1056 #endif
1057
1058 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1059 _loop_opts_cnt = LoopOptsCount;
1060 set_do_inlining(Inline);
1061 set_max_inline_size(MaxInlineSize);
1062 set_freq_inline_size(FreqInlineSize);
1063 set_do_scheduling(OptoScheduling);
1064
1065 set_do_vector_loop(false);
1066 set_has_monitors(false);
1067 set_has_scoped_access(false);
1068
1069 if (AllowVectorizeOnDemand) {
1070 if (has_method() && _directive->VectorizeOption) {
1071 set_do_vector_loop(true);
1072 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());})
1073 } else if (has_method() && method()->name() != nullptr &&
1074 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1075 set_do_vector_loop(true);
1076 }
1077 }
1078 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1079 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());})
1080
1081 _max_node_limit = _directive->MaxNodeLimitOption;
1082
1083 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() &&
1084 (method()->needs_clinit_barrier() || (do_clinit_barriers() && method()->is_static()))) {
1085 set_clinit_barrier_on_entry(true);
1086 if (do_clinit_barriers()) {
1087 set_has_clinit_barriers(true); // Entry clinit barrier is in prolog code.
1088 }
1089 }
1090 if (debug_info()->recording_non_safepoints()) {
1091 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1092 (comp_arena(), 8, 0, nullptr));
1093 set_default_node_notes(Node_Notes::make(this));
1094 }
1095
1096 const int grow_ats = 16;
1097 _max_alias_types = grow_ats;
1098 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1099 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1100 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1101 {
1102 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1103 }
1104 // Initialize the first few types.
1105 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1106 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1107 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1108 _num_alias_types = AliasIdxRaw+1;
1109 // Zero out the alias type cache.
1110 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1111 // A null adr_type hits in the cache right away. Preload the right answer.
1112 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1113 }
1114
1115 #ifdef ASSERT
1116 // Verify that the current StartNode is valid.
1117 void Compile::verify_start(StartNode* s) const {
1118 assert(failing_internal() || s == start(), "should be StartNode");
1119 }
1120 #endif
1121
1122 /**
1123 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1124 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1125 * the ideal graph.
1126 */
1127 StartNode* Compile::start() const {
1128 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1129 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1130 Node* start = root()->fast_out(i);
1131 if (start->is_Start()) {
1132 return start->as_Start();
1133 }
1134 }
1135 fatal("Did not find Start node!");
1136 return nullptr;
1137 }
1138
1139 //-------------------------------immutable_memory-------------------------------------
1140 // Access immutable memory
1141 Node* Compile::immutable_memory() {
1142 if (_immutable_memory != nullptr) {
1143 return _immutable_memory;
1144 }
1145 StartNode* s = start();
1146 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1147 Node *p = s->fast_out(i);
1148 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1149 _immutable_memory = p;
1150 return _immutable_memory;
1151 }
1152 }
1153 ShouldNotReachHere();
1154 return nullptr;
1155 }
1156
1157 //----------------------set_cached_top_node------------------------------------
1158 // Install the cached top node, and make sure Node::is_top works correctly.
1159 void Compile::set_cached_top_node(Node* tn) {
1160 if (tn != nullptr) verify_top(tn);
1161 Node* old_top = _top;
1162 _top = tn;
1163 // Calling Node::setup_is_top allows the nodes the chance to adjust
1164 // their _out arrays.
1165 if (_top != nullptr) _top->setup_is_top();
1166 if (old_top != nullptr) old_top->setup_is_top();
1167 assert(_top == nullptr || top()->is_top(), "");
1168 }
1169
1170 #ifdef ASSERT
1171 uint Compile::count_live_nodes_by_graph_walk() {
1172 Unique_Node_List useful(comp_arena());
1173 // Get useful node list by walking the graph.
1174 identify_useful_nodes(useful);
1175 return useful.size();
1176 }
1177
1178 void Compile::print_missing_nodes() {
1179
1180 // Return if CompileLog is null and PrintIdealNodeCount is false.
1181 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1182 return;
1183 }
1184
1185 // This is an expensive function. It is executed only when the user
1186 // specifies VerifyIdealNodeCount option or otherwise knows the
1187 // additional work that needs to be done to identify reachable nodes
1188 // by walking the flow graph and find the missing ones using
1189 // _dead_node_list.
1190
1191 Unique_Node_List useful(comp_arena());
1192 // Get useful node list by walking the graph.
1193 identify_useful_nodes(useful);
1194
1195 uint l_nodes = C->live_nodes();
1196 uint l_nodes_by_walk = useful.size();
1197
1198 if (l_nodes != l_nodes_by_walk) {
1199 if (_log != nullptr) {
1200 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1201 _log->stamp();
1202 _log->end_head();
1203 }
1204 VectorSet& useful_member_set = useful.member_set();
1205 int last_idx = l_nodes_by_walk;
1206 for (int i = 0; i < last_idx; i++) {
1207 if (useful_member_set.test(i)) {
1208 if (_dead_node_list.test(i)) {
1209 if (_log != nullptr) {
1210 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1211 }
1212 if (PrintIdealNodeCount) {
1213 // Print the log message to tty
1214 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1215 useful.at(i)->dump();
1216 }
1217 }
1218 }
1219 else if (! _dead_node_list.test(i)) {
1220 if (_log != nullptr) {
1221 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1222 }
1223 if (PrintIdealNodeCount) {
1224 // Print the log message to tty
1225 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1226 }
1227 }
1228 }
1229 if (_log != nullptr) {
1230 _log->tail("mismatched_nodes");
1231 }
1232 }
1233 }
1234 void Compile::record_modified_node(Node* n) {
1235 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1236 _modified_nodes->push(n);
1237 }
1238 }
1239
1240 void Compile::remove_modified_node(Node* n) {
1241 if (_modified_nodes != nullptr) {
1242 _modified_nodes->remove(n);
1243 }
1244 }
1245 #endif
1246
1247 #ifndef PRODUCT
1248 void Compile::verify_top(Node* tn) const {
1249 if (tn != nullptr) {
1250 assert(tn->is_Con(), "top node must be a constant");
1251 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1252 assert(tn->in(0) != nullptr, "must have live top node");
1253 }
1254 }
1255 #endif
1256
1257
1258 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1259
1260 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1261 guarantee(arr != nullptr, "");
1262 int num_blocks = arr->length();
1263 if (grow_by < num_blocks) grow_by = num_blocks;
1264 int num_notes = grow_by * _node_notes_block_size;
1265 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1266 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1267 while (num_notes > 0) {
1268 arr->append(notes);
1269 notes += _node_notes_block_size;
1270 num_notes -= _node_notes_block_size;
1271 }
1272 assert(num_notes == 0, "exact multiple, please");
1273 }
1274
1275 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1276 if (source == nullptr || dest == nullptr) return false;
1277
1278 if (dest->is_Con())
1279 return false; // Do not push debug info onto constants.
1280
1281 #ifdef ASSERT
1282 // Leave a bread crumb trail pointing to the original node:
1283 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1284 dest->set_debug_orig(source);
1285 }
1286 #endif
1287
1288 if (node_note_array() == nullptr)
1289 return false; // Not collecting any notes now.
1290
1291 // This is a copy onto a pre-existing node, which may already have notes.
1292 // If both nodes have notes, do not overwrite any pre-existing notes.
1293 Node_Notes* source_notes = node_notes_at(source->_idx);
1294 if (source_notes == nullptr || source_notes->is_clear()) return false;
1295 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1296 if (dest_notes == nullptr || dest_notes->is_clear()) {
1297 return set_node_notes_at(dest->_idx, source_notes);
1298 }
1299
1300 Node_Notes merged_notes = (*source_notes);
1301 // The order of operations here ensures that dest notes will win...
1302 merged_notes.update_from(dest_notes);
1303 return set_node_notes_at(dest->_idx, &merged_notes);
1304 }
1305
1306
1307 //--------------------------allow_range_check_smearing-------------------------
1308 // Gating condition for coalescing similar range checks.
1309 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1310 // single covering check that is at least as strong as any of them.
1311 // If the optimization succeeds, the simplified (strengthened) range check
1312 // will always succeed. If it fails, we will deopt, and then give up
1313 // on the optimization.
1314 bool Compile::allow_range_check_smearing() const {
1315 // If this method has already thrown a range-check,
1316 // assume it was because we already tried range smearing
1317 // and it failed.
1318 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1319 return !already_trapped;
1320 }
1321
1322
1323 //------------------------------flatten_alias_type-----------------------------
1324 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1325 assert(do_aliasing(), "Aliasing should be enabled");
1326 int offset = tj->offset();
1327 TypePtr::PTR ptr = tj->ptr();
1328
1329 // Known instance (scalarizable allocation) alias only with itself.
1330 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1331 tj->is_oopptr()->is_known_instance();
1332
1333 // Process weird unsafe references.
1334 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1335 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1336 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1337 tj = TypeOopPtr::BOTTOM;
1338 ptr = tj->ptr();
1339 offset = tj->offset();
1340 }
1341
1342 // Array pointers need some flattening
1343 const TypeAryPtr* ta = tj->isa_aryptr();
1344 if (ta && ta->is_stable()) {
1345 // Erase stability property for alias analysis.
1346 tj = ta = ta->cast_to_stable(false);
1347 }
1348 if( ta && is_known_inst ) {
1349 if ( offset != Type::OffsetBot &&
1350 offset > arrayOopDesc::length_offset_in_bytes() ) {
1351 offset = Type::OffsetBot; // Flatten constant access into array body only
1352 tj = ta = ta->
1353 remove_speculative()->
1354 cast_to_ptr_type(ptr)->
1355 with_offset(offset);
1356 }
1357 } else if (ta) {
1358 // For arrays indexed by constant indices, we flatten the alias
1359 // space to include all of the array body. Only the header, klass
1360 // and array length can be accessed un-aliased.
1361 if( offset != Type::OffsetBot ) {
1362 if( ta->const_oop() ) { // MethodData* or Method*
1363 offset = Type::OffsetBot; // Flatten constant access into array body
1364 tj = ta = ta->
1365 remove_speculative()->
1366 cast_to_ptr_type(ptr)->
1367 cast_to_exactness(false)->
1368 with_offset(offset);
1369 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1370 // range is OK as-is.
1371 tj = ta = TypeAryPtr::RANGE;
1372 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1373 tj = TypeInstPtr::KLASS; // all klass loads look alike
1374 ta = TypeAryPtr::RANGE; // generic ignored junk
1375 ptr = TypePtr::BotPTR;
1376 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1377 tj = TypeInstPtr::MARK;
1378 ta = TypeAryPtr::RANGE; // generic ignored junk
1379 ptr = TypePtr::BotPTR;
1380 } else { // Random constant offset into array body
1381 offset = Type::OffsetBot; // Flatten constant access into array body
1382 tj = ta = ta->
1383 remove_speculative()->
1384 cast_to_ptr_type(ptr)->
1385 cast_to_exactness(false)->
1386 with_offset(offset);
1387 }
1388 }
1389 // Arrays of fixed size alias with arrays of unknown size.
1390 if (ta->size() != TypeInt::POS) {
1391 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1392 tj = ta = ta->
1393 remove_speculative()->
1394 cast_to_ptr_type(ptr)->
1395 with_ary(tary)->
1396 cast_to_exactness(false);
1397 }
1398 // Arrays of known objects become arrays of unknown objects.
1399 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1400 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1401 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1402 }
1403 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1404 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1405 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1406 }
1407 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1408 // cannot be distinguished by bytecode alone.
1409 if (ta->elem() == TypeInt::BOOL) {
1410 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1411 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1412 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1413 }
1414 // During the 2nd round of IterGVN, NotNull castings are removed.
1415 // Make sure the Bottom and NotNull variants alias the same.
1416 // Also, make sure exact and non-exact variants alias the same.
1417 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1418 tj = ta = ta->
1419 remove_speculative()->
1420 cast_to_ptr_type(TypePtr::BotPTR)->
1421 cast_to_exactness(false)->
1422 with_offset(offset);
1423 }
1424 }
1425
1426 // Oop pointers need some flattening
1427 const TypeInstPtr *to = tj->isa_instptr();
1428 if (to && to != TypeOopPtr::BOTTOM) {
1429 ciInstanceKlass* ik = to->instance_klass();
1430 if( ptr == TypePtr::Constant ) {
1431 if (ik != ciEnv::current()->Class_klass() ||
1432 offset < ik->layout_helper_size_in_bytes()) {
1433 // No constant oop pointers (such as Strings); they alias with
1434 // unknown strings.
1435 assert(!is_known_inst, "not scalarizable allocation");
1436 tj = to = to->
1437 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1438 remove_speculative()->
1439 cast_to_ptr_type(TypePtr::BotPTR)->
1440 cast_to_exactness(false);
1441 }
1442 } else if( is_known_inst ) {
1443 tj = to; // Keep NotNull and klass_is_exact for instance type
1444 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1445 // During the 2nd round of IterGVN, NotNull castings are removed.
1446 // Make sure the Bottom and NotNull variants alias the same.
1447 // Also, make sure exact and non-exact variants alias the same.
1448 tj = to = to->
1449 remove_speculative()->
1450 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1451 cast_to_ptr_type(TypePtr::BotPTR)->
1452 cast_to_exactness(false);
1453 }
1454 if (to->speculative() != nullptr) {
1455 tj = to = to->remove_speculative();
1456 }
1457 // Canonicalize the holder of this field
1458 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1459 // First handle header references such as a LoadKlassNode, even if the
1460 // object's klass is unloaded at compile time (4965979).
1461 if (!is_known_inst) { // Do it only for non-instance types
1462 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1463 }
1464 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1465 // Static fields are in the space above the normal instance
1466 // fields in the java.lang.Class instance.
1467 if (ik != ciEnv::current()->Class_klass()) {
1468 to = nullptr;
1469 tj = TypeOopPtr::BOTTOM;
1470 offset = tj->offset();
1471 }
1472 } else {
1473 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1474 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1475 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1476 if( is_known_inst ) {
1477 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id());
1478 } else {
1479 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset);
1480 }
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 (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1705 alias_type(idx)->set_rewritable(false);
1706 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1707 alias_type(idx)->set_rewritable(false);
1708 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1709 alias_type(idx)->set_rewritable(false);
1710 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1711 alias_type(idx)->set_rewritable(false);
1712 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1713 alias_type(idx)->set_rewritable(false);
1714 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1715 alias_type(idx)->set_rewritable(false);
1716 }
1717 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1718 // but the base pointer type is not distinctive enough to identify
1719 // references into JavaThread.)
1720
1721 // Check for final fields.
1722 const TypeInstPtr* tinst = flat->isa_instptr();
1723 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1724 ciField* field;
1725 if (tinst->const_oop() != nullptr &&
1726 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1727 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1728 // static field
1729 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1730 field = k->get_field_by_offset(tinst->offset(), true);
1731 } else {
1732 ciInstanceKlass *k = tinst->instance_klass();
1733 field = k->get_field_by_offset(tinst->offset(), false);
1734 }
1735 assert(field == nullptr ||
1736 original_field == nullptr ||
1737 (field->holder() == original_field->holder() &&
1738 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1739 field->is_static() == original_field->is_static()), "wrong field?");
1740 // Set field() and is_rewritable() attributes.
1741 if (field != nullptr) alias_type(idx)->set_field(field);
1742 }
1743 }
1744
1745 // Fill the cache for next time.
1746 ace->_adr_type = adr_type;
1747 ace->_index = idx;
1748 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1749
1750 // Might as well try to fill the cache for the flattened version, too.
1751 AliasCacheEntry* face = probe_alias_cache(flat);
1752 if (face->_adr_type == nullptr) {
1753 face->_adr_type = flat;
1754 face->_index = idx;
1755 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1756 }
1757
1758 return alias_type(idx);
1759 }
1760
1761
1762 Compile::AliasType* Compile::alias_type(ciField* field) {
1763 const TypeOopPtr* t;
1764 if (field->is_static())
1765 t = TypeInstPtr::make(field->holder()->java_mirror());
1766 else
1767 t = TypeOopPtr::make_from_klass_raw(field->holder());
1768 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1769 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1770 return atp;
1771 }
1772
1773
1774 //------------------------------have_alias_type--------------------------------
1775 bool Compile::have_alias_type(const TypePtr* adr_type) {
1776 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1777 if (ace->_adr_type == adr_type) {
1778 return true;
1779 }
1780
1781 // Handle special cases.
1782 if (adr_type == nullptr) return true;
1783 if (adr_type == TypePtr::BOTTOM) return true;
1784
1785 return find_alias_type(adr_type, true, nullptr) != nullptr;
1786 }
1787
1788 //-----------------------------must_alias--------------------------------------
1789 // True if all values of the given address type are in the given alias category.
1790 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1791 if (alias_idx == AliasIdxBot) return true; // the universal category
1792 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1793 if (alias_idx == AliasIdxTop) return false; // the empty category
1794 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1795
1796 // the only remaining possible overlap is identity
1797 int adr_idx = get_alias_index(adr_type);
1798 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1799 assert(adr_idx == alias_idx ||
1800 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1801 && adr_type != TypeOopPtr::BOTTOM),
1802 "should not be testing for overlap with an unsafe pointer");
1803 return adr_idx == alias_idx;
1804 }
1805
1806 //------------------------------can_alias--------------------------------------
1807 // True if any values of the given address type are in the given alias category.
1808 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1809 if (alias_idx == AliasIdxTop) return false; // the empty category
1810 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1811 // Known instance doesn't alias with bottom memory
1812 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1813 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1814
1815 // the only remaining possible overlap is identity
1816 int adr_idx = get_alias_index(adr_type);
1817 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1818 return adr_idx == alias_idx;
1819 }
1820
1821 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1822 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1823 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1824 if (parse_predicate_count() == 0) {
1825 return;
1826 }
1827 for (int i = 0; i < parse_predicate_count(); i++) {
1828 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1829 parse_predicate->mark_useless();
1830 igvn._worklist.push(parse_predicate);
1831 }
1832 _parse_predicates.clear();
1833 }
1834
1835 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1836 if (!n->for_post_loop_opts_igvn()) {
1837 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1838 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1839 _for_post_loop_igvn.append(n);
1840 }
1841 }
1842
1843 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1844 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1845 _for_post_loop_igvn.remove(n);
1846 }
1847
1848 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1849 // Verify that all previous optimizations produced a valid graph
1850 // at least to this point, even if no loop optimizations were done.
1851 PhaseIdealLoop::verify(igvn);
1852
1853 C->set_post_loop_opts_phase(); // no more loop opts allowed
1854
1855 assert(!C->major_progress(), "not cleared");
1856
1857 if (_for_post_loop_igvn.length() > 0) {
1858 while (_for_post_loop_igvn.length() > 0) {
1859 Node* n = _for_post_loop_igvn.pop();
1860 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1861 igvn._worklist.push(n);
1862 }
1863 igvn.optimize();
1864 if (failing()) return;
1865 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1866 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1867
1868 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1869 if (C->major_progress()) {
1870 C->clear_major_progress(); // ensure that major progress is now clear
1871 }
1872 }
1873 }
1874
1875 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1876 if (OptimizeUnstableIf) {
1877 _unstable_if_traps.append(trap);
1878 }
1879 }
1880
1881 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1882 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1883 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1884 Node* n = trap->uncommon_trap();
1885 if (!useful.member(n)) {
1886 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1887 }
1888 }
1889 }
1890
1891 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1892 // or fold-compares case. Return true if succeed or not found.
1893 //
1894 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1895 // false and ask fold-compares to yield.
1896 //
1897 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1898 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1899 // when deoptimization does happen.
1900 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1901 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1902 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1903 if (trap->uncommon_trap() == unc) {
1904 if (yield && trap->modified()) {
1905 return false;
1906 }
1907 _unstable_if_traps.delete_at(i);
1908 break;
1909 }
1910 }
1911 return true;
1912 }
1913
1914 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1915 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1916 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1917 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1918 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1919 CallStaticJavaNode* unc = trap->uncommon_trap();
1920 int next_bci = trap->next_bci();
1921 bool modified = trap->modified();
1922
1923 if (next_bci != -1 && !modified) {
1924 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1925 JVMState* jvms = unc->jvms();
1926 ciMethod* method = jvms->method();
1927 ciBytecodeStream iter(method);
1928
1929 iter.force_bci(jvms->bci());
1930 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1931 Bytecodes::Code c = iter.cur_bc();
1932 Node* lhs = nullptr;
1933 Node* rhs = nullptr;
1934 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1935 lhs = unc->peek_operand(0);
1936 rhs = unc->peek_operand(1);
1937 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1938 lhs = unc->peek_operand(0);
1939 }
1940
1941 ResourceMark rm;
1942 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1943 assert(live_locals.is_valid(), "broken liveness info");
1944 int len = (int)live_locals.size();
1945
1946 for (int i = 0; i < len; i++) {
1947 Node* local = unc->local(jvms, i);
1948 // kill local using the liveness of next_bci.
1949 // give up when the local looks like an operand to secure reexecution.
1950 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1951 uint idx = jvms->locoff() + i;
1952 #ifdef ASSERT
1953 if (PrintOpto && Verbose) {
1954 tty->print("[unstable_if] kill local#%d: ", idx);
1955 local->dump();
1956 tty->cr();
1957 }
1958 #endif
1959 igvn.replace_input_of(unc, idx, top());
1960 modified = true;
1961 }
1962 }
1963 }
1964
1965 // keep the mondified trap for late query
1966 if (modified) {
1967 trap->set_modified();
1968 } else {
1969 _unstable_if_traps.delete_at(i);
1970 }
1971 }
1972 igvn.optimize();
1973 }
1974
1975 // StringOpts and late inlining of string methods
1976 void Compile::inline_string_calls(bool parse_time) {
1977 {
1978 // remove useless nodes to make the usage analysis simpler
1979 ResourceMark rm;
1980 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1981 }
1982
1983 {
1984 ResourceMark rm;
1985 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1986 PhaseStringOpts pso(initial_gvn());
1987 print_method(PHASE_AFTER_STRINGOPTS, 3);
1988 }
1989
1990 // now inline anything that we skipped the first time around
1991 if (!parse_time) {
1992 _late_inlines_pos = _late_inlines.length();
1993 }
1994
1995 while (_string_late_inlines.length() > 0) {
1996 CallGenerator* cg = _string_late_inlines.pop();
1997 cg->do_late_inline();
1998 if (failing()) return;
1999 }
2000 _string_late_inlines.trunc_to(0);
2001 }
2002
2003 // Late inlining of boxing methods
2004 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2005 if (_boxing_late_inlines.length() > 0) {
2006 assert(has_boxed_value(), "inconsistent");
2007
2008 set_inlining_incrementally(true);
2009
2010 igvn_worklist()->ensure_empty(); // should be done with igvn
2011
2012 _late_inlines_pos = _late_inlines.length();
2013
2014 while (_boxing_late_inlines.length() > 0) {
2015 CallGenerator* cg = _boxing_late_inlines.pop();
2016 cg->do_late_inline();
2017 if (failing()) return;
2018 }
2019 _boxing_late_inlines.trunc_to(0);
2020
2021 inline_incrementally_cleanup(igvn);
2022
2023 set_inlining_incrementally(false);
2024 }
2025 }
2026
2027 bool Compile::inline_incrementally_one() {
2028 assert(IncrementalInline, "incremental inlining should be on");
2029
2030 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2031
2032 set_inlining_progress(false);
2033 set_do_cleanup(false);
2034
2035 for (int i = 0; i < _late_inlines.length(); i++) {
2036 _late_inlines_pos = i+1;
2037 CallGenerator* cg = _late_inlines.at(i);
2038 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2039 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2040 cg->do_late_inline();
2041 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2042 if (failing()) {
2043 return false;
2044 } else if (inlining_progress()) {
2045 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2046 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2047 break; // process one call site at a time
2048 }
2049 } else {
2050 // Ignore late inline direct calls when inlining is not allowed.
2051 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2052 }
2053 }
2054 // Remove processed elements.
2055 _late_inlines.remove_till(_late_inlines_pos);
2056 _late_inlines_pos = 0;
2057
2058 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2059
2060 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2061
2062 set_inlining_progress(false);
2063 set_do_cleanup(false);
2064
2065 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2066 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2067 }
2068
2069 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2070 {
2071 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2072 ResourceMark rm;
2073 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2074 }
2075 {
2076 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2077 igvn.reset_from_gvn(initial_gvn());
2078 igvn.optimize();
2079 if (failing()) return;
2080 }
2081 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2082 }
2083
2084 // Perform incremental inlining until bound on number of live nodes is reached
2085 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2086 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2087
2088 set_inlining_incrementally(true);
2089 uint low_live_nodes = 0;
2090
2091 while (_late_inlines.length() > 0) {
2092 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2093 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2094 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2095 // PhaseIdealLoop is expensive so we only try it once we are
2096 // out of live nodes and we only try it again if the previous
2097 // helped got the number of nodes down significantly
2098 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2099 if (failing()) return;
2100 low_live_nodes = live_nodes();
2101 _major_progress = true;
2102 }
2103
2104 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2105 bool do_print_inlining = print_inlining() || print_intrinsics();
2106 if (do_print_inlining || log() != nullptr) {
2107 // Print inlining message for candidates that we couldn't inline for lack of space.
2108 for (int i = 0; i < _late_inlines.length(); i++) {
2109 CallGenerator* cg = _late_inlines.at(i);
2110 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2111 if (do_print_inlining) {
2112 cg->print_inlining_late(InliningResult::FAILURE, msg);
2113 }
2114 log_late_inline_failure(cg, msg);
2115 }
2116 }
2117 break; // finish
2118 }
2119 }
2120
2121 igvn_worklist()->ensure_empty(); // should be done with igvn
2122
2123 while (inline_incrementally_one()) {
2124 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2125 }
2126 if (failing()) return;
2127
2128 inline_incrementally_cleanup(igvn);
2129
2130 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2131
2132 if (failing()) return;
2133
2134 if (_late_inlines.length() == 0) {
2135 break; // no more progress
2136 }
2137 }
2138
2139 igvn_worklist()->ensure_empty(); // should be done with igvn
2140
2141 if (_string_late_inlines.length() > 0) {
2142 assert(has_stringbuilder(), "inconsistent");
2143
2144 inline_string_calls(false);
2145
2146 if (failing()) return;
2147
2148 inline_incrementally_cleanup(igvn);
2149 }
2150
2151 set_inlining_incrementally(false);
2152 }
2153
2154 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2155 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2156 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2157 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2158 // as if "inlining_incrementally() == true" were set.
2159 assert(inlining_incrementally() == false, "not allowed");
2160 assert(_modified_nodes == nullptr, "not allowed");
2161 assert(_late_inlines.length() > 0, "sanity");
2162
2163 while (_late_inlines.length() > 0) {
2164 igvn_worklist()->ensure_empty(); // should be done with igvn
2165
2166 while (inline_incrementally_one()) {
2167 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2168 }
2169 if (failing()) return;
2170
2171 inline_incrementally_cleanup(igvn);
2172 }
2173 }
2174
2175 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2176 if (_loop_opts_cnt > 0) {
2177 while (major_progress() && (_loop_opts_cnt > 0)) {
2178 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2179 PhaseIdealLoop::optimize(igvn, mode);
2180 _loop_opts_cnt--;
2181 if (failing()) return false;
2182 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2183 }
2184 }
2185 return true;
2186 }
2187
2188 // Remove edges from "root" to each SafePoint at a backward branch.
2189 // They were inserted during parsing (see add_safepoint()) to make
2190 // infinite loops without calls or exceptions visible to root, i.e.,
2191 // useful.
2192 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2193 Node *r = root();
2194 if (r != nullptr) {
2195 for (uint i = r->req(); i < r->len(); ++i) {
2196 Node *n = r->in(i);
2197 if (n != nullptr && n->is_SafePoint()) {
2198 r->rm_prec(i);
2199 if (n->outcnt() == 0) {
2200 igvn.remove_dead_node(n);
2201 }
2202 --i;
2203 }
2204 }
2205 // Parsing may have added top inputs to the root node (Path
2206 // leading to the Halt node proven dead). Make sure we get a
2207 // chance to clean them up.
2208 igvn._worklist.push(r);
2209 igvn.optimize();
2210 }
2211 }
2212
2213 //------------------------------Optimize---------------------------------------
2214 // Given a graph, optimize it.
2215 void Compile::Optimize() {
2216 TracePhase tp("optimizer", &timers[_t_optimizer]);
2217
2218 #ifndef PRODUCT
2219 if (env()->break_at_compile()) {
2220 BREAKPOINT;
2221 }
2222
2223 #endif
2224
2225 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2226 #ifdef ASSERT
2227 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2228 #endif
2229
2230 ResourceMark rm;
2231
2232 print_inlining_reinit();
2233
2234 NOT_PRODUCT( verify_graph_edges(); )
2235
2236 print_method(PHASE_AFTER_PARSING, 1);
2237
2238 {
2239 // Iterative Global Value Numbering, including ideal transforms
2240 // Initialize IterGVN with types and values from parse-time GVN
2241 PhaseIterGVN igvn(initial_gvn());
2242 #ifdef ASSERT
2243 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2244 #endif
2245 {
2246 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2247 igvn.optimize();
2248 }
2249
2250 if (failing()) return;
2251
2252 print_method(PHASE_ITER_GVN1, 2);
2253
2254 process_for_unstable_if_traps(igvn);
2255
2256 if (failing()) return;
2257
2258 inline_incrementally(igvn);
2259
2260 print_method(PHASE_INCREMENTAL_INLINE, 2);
2261
2262 if (failing()) return;
2263
2264 if (eliminate_boxing()) {
2265 // Inline valueOf() methods now.
2266 inline_boxing_calls(igvn);
2267
2268 if (failing()) return;
2269
2270 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2271 inline_incrementally(igvn);
2272 }
2273
2274 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2275
2276 if (failing()) return;
2277 }
2278
2279 // Remove the speculative part of types and clean up the graph from
2280 // the extra CastPP nodes whose only purpose is to carry them. Do
2281 // that early so that optimizations are not disrupted by the extra
2282 // CastPP nodes.
2283 remove_speculative_types(igvn);
2284
2285 if (failing()) return;
2286
2287 // No more new expensive nodes will be added to the list from here
2288 // so keep only the actual candidates for optimizations.
2289 cleanup_expensive_nodes(igvn);
2290
2291 if (failing()) return;
2292
2293 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2294 if (EnableVectorSupport && has_vbox_nodes()) {
2295 TracePhase tp("", &timers[_t_vector]);
2296 PhaseVector pv(igvn);
2297 pv.optimize_vector_boxes();
2298 if (failing()) return;
2299 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2300 }
2301 assert(!has_vbox_nodes(), "sanity");
2302
2303 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2304 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2305 igvn_worklist()->ensure_empty(); // should be done with igvn
2306 {
2307 ResourceMark rm;
2308 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2309 }
2310 igvn.reset_from_gvn(initial_gvn());
2311 igvn.optimize();
2312 if (failing()) return;
2313 }
2314
2315 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2316 // safepoints
2317 remove_root_to_sfpts_edges(igvn);
2318
2319 if (failing()) return;
2320
2321 // Perform escape analysis
2322 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2323 if (has_loops()) {
2324 // Cleanup graph (remove dead nodes).
2325 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2326 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2327 if (failing()) return;
2328 }
2329 bool progress;
2330 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2331 do {
2332 ConnectionGraph::do_analysis(this, &igvn);
2333
2334 if (failing()) return;
2335
2336 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2337
2338 // Optimize out fields loads from scalar replaceable allocations.
2339 igvn.optimize();
2340 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2341
2342 if (failing()) return;
2343
2344 if (congraph() != nullptr && macro_count() > 0) {
2345 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2346 PhaseMacroExpand mexp(igvn);
2347 mexp.eliminate_macro_nodes();
2348 if (failing()) return;
2349
2350 igvn.set_delay_transform(false);
2351 igvn.optimize();
2352 if (failing()) return;
2353
2354 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2355 }
2356
2357 ConnectionGraph::verify_ram_nodes(this, root());
2358 if (failing()) return;
2359
2360 progress = do_iterative_escape_analysis() &&
2361 (macro_count() < mcount) &&
2362 ConnectionGraph::has_candidates(this);
2363 // Try again if candidates exist and made progress
2364 // by removing some allocations and/or locks.
2365 } while (progress);
2366 }
2367
2368 // Loop transforms on the ideal graph. Range Check Elimination,
2369 // peeling, unrolling, etc.
2370
2371 // Set loop opts counter
2372 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2373 {
2374 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2375 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2376 _loop_opts_cnt--;
2377 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2378 if (failing()) return;
2379 }
2380 // Loop opts pass if partial peeling occurred in previous pass
2381 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2382 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2383 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2384 _loop_opts_cnt--;
2385 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2386 if (failing()) return;
2387 }
2388 // Loop opts pass for loop-unrolling before CCP
2389 if(major_progress() && (_loop_opts_cnt > 0)) {
2390 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2391 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2392 _loop_opts_cnt--;
2393 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2394 }
2395 if (!failing()) {
2396 // Verify that last round of loop opts produced a valid graph
2397 PhaseIdealLoop::verify(igvn);
2398 }
2399 }
2400 if (failing()) return;
2401
2402 // Conditional Constant Propagation;
2403 print_method(PHASE_BEFORE_CCP1, 2);
2404 PhaseCCP ccp( &igvn );
2405 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2406 {
2407 TracePhase tp("ccp", &timers[_t_ccp]);
2408 ccp.do_transform();
2409 }
2410 print_method(PHASE_CCP1, 2);
2411
2412 assert( true, "Break here to ccp.dump_old2new_map()");
2413
2414 // Iterative Global Value Numbering, including ideal transforms
2415 {
2416 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2417 igvn.reset_from_igvn(&ccp);
2418 igvn.optimize();
2419 }
2420 print_method(PHASE_ITER_GVN2, 2);
2421
2422 if (failing()) return;
2423
2424 // Loop transforms on the ideal graph. Range Check Elimination,
2425 // peeling, unrolling, etc.
2426 if (!optimize_loops(igvn, LoopOptsDefault)) {
2427 return;
2428 }
2429
2430 if (failing()) return;
2431
2432 C->clear_major_progress(); // ensure that major progress is now clear
2433
2434 process_for_post_loop_opts_igvn(igvn);
2435
2436 if (failing()) return;
2437
2438 #ifdef ASSERT
2439 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2440 #endif
2441
2442 {
2443 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2444 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2445 PhaseMacroExpand mex(igvn);
2446 if (mex.expand_macro_nodes()) {
2447 assert(failing(), "must bail out w/ explicit message");
2448 return;
2449 }
2450 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2451 }
2452
2453 {
2454 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2455 if (bs->expand_barriers(this, igvn)) {
2456 assert(failing(), "must bail out w/ explicit message");
2457 return;
2458 }
2459 print_method(PHASE_BARRIER_EXPANSION, 2);
2460 }
2461
2462 if (C->max_vector_size() > 0) {
2463 C->optimize_logic_cones(igvn);
2464 igvn.optimize();
2465 if (failing()) return;
2466 }
2467
2468 DEBUG_ONLY( _modified_nodes = nullptr; )
2469
2470 assert(igvn._worklist.size() == 0, "not empty");
2471
2472 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2473
2474 if (_late_inlines.length() > 0) {
2475 // More opportunities to optimize virtual and MH calls.
2476 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2477 process_late_inline_calls_no_inline(igvn);
2478 if (failing()) return;
2479 }
2480 } // (End scope of igvn; run destructor if necessary for asserts.)
2481
2482 check_no_dead_use();
2483
2484 process_print_inlining();
2485
2486 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2487 // to remove hashes to unlock nodes for modifications.
2488 C->node_hash()->clear();
2489
2490 // A method with only infinite loops has no edges entering loops from root
2491 {
2492 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2493 if (final_graph_reshaping()) {
2494 assert(failing(), "must bail out w/ explicit message");
2495 return;
2496 }
2497 }
2498
2499 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2500 DEBUG_ONLY(set_phase_optimize_finished();)
2501 }
2502
2503 #ifdef ASSERT
2504 void Compile::check_no_dead_use() const {
2505 ResourceMark rm;
2506 Unique_Node_List wq;
2507 wq.push(root());
2508 for (uint i = 0; i < wq.size(); ++i) {
2509 Node* n = wq.at(i);
2510 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2511 Node* u = n->fast_out(j);
2512 if (u->outcnt() == 0 && !u->is_Con()) {
2513 u->dump();
2514 fatal("no reachable node should have no use");
2515 }
2516 wq.push(u);
2517 }
2518 }
2519 }
2520 #endif
2521
2522 void Compile::inline_vector_reboxing_calls() {
2523 if (C->_vector_reboxing_late_inlines.length() > 0) {
2524 _late_inlines_pos = C->_late_inlines.length();
2525 while (_vector_reboxing_late_inlines.length() > 0) {
2526 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2527 cg->do_late_inline();
2528 if (failing()) return;
2529 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2530 }
2531 _vector_reboxing_late_inlines.trunc_to(0);
2532 }
2533 }
2534
2535 bool Compile::has_vbox_nodes() {
2536 if (C->_vector_reboxing_late_inlines.length() > 0) {
2537 return true;
2538 }
2539 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2540 Node * n = C->macro_node(macro_idx);
2541 assert(n->is_macro(), "only macro nodes expected here");
2542 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2543 return true;
2544 }
2545 }
2546 return false;
2547 }
2548
2549 //---------------------------- Bitwise operation packing optimization ---------------------------
2550
2551 static bool is_vector_unary_bitwise_op(Node* n) {
2552 return n->Opcode() == Op_XorV &&
2553 VectorNode::is_vector_bitwise_not_pattern(n);
2554 }
2555
2556 static bool is_vector_binary_bitwise_op(Node* n) {
2557 switch (n->Opcode()) {
2558 case Op_AndV:
2559 case Op_OrV:
2560 return true;
2561
2562 case Op_XorV:
2563 return !is_vector_unary_bitwise_op(n);
2564
2565 default:
2566 return false;
2567 }
2568 }
2569
2570 static bool is_vector_ternary_bitwise_op(Node* n) {
2571 return n->Opcode() == Op_MacroLogicV;
2572 }
2573
2574 static bool is_vector_bitwise_op(Node* n) {
2575 return is_vector_unary_bitwise_op(n) ||
2576 is_vector_binary_bitwise_op(n) ||
2577 is_vector_ternary_bitwise_op(n);
2578 }
2579
2580 static bool is_vector_bitwise_cone_root(Node* n) {
2581 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2582 return false;
2583 }
2584 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2585 if (is_vector_bitwise_op(n->fast_out(i))) {
2586 return false;
2587 }
2588 }
2589 return true;
2590 }
2591
2592 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2593 uint cnt = 0;
2594 if (is_vector_bitwise_op(n)) {
2595 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2596 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2597 for (uint i = 1; i < inp_cnt; i++) {
2598 Node* in = n->in(i);
2599 bool skip = VectorNode::is_all_ones_vector(in);
2600 if (!skip && !inputs.member(in)) {
2601 inputs.push(in);
2602 cnt++;
2603 }
2604 }
2605 assert(cnt <= 1, "not unary");
2606 } else {
2607 uint last_req = inp_cnt;
2608 if (is_vector_ternary_bitwise_op(n)) {
2609 last_req = inp_cnt - 1; // skip last input
2610 }
2611 for (uint i = 1; i < last_req; i++) {
2612 Node* def = n->in(i);
2613 if (!inputs.member(def)) {
2614 inputs.push(def);
2615 cnt++;
2616 }
2617 }
2618 }
2619 } else { // not a bitwise operations
2620 if (!inputs.member(n)) {
2621 inputs.push(n);
2622 cnt++;
2623 }
2624 }
2625 return cnt;
2626 }
2627
2628 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2629 Unique_Node_List useful_nodes;
2630 C->identify_useful_nodes(useful_nodes);
2631
2632 for (uint i = 0; i < useful_nodes.size(); i++) {
2633 Node* n = useful_nodes.at(i);
2634 if (is_vector_bitwise_cone_root(n)) {
2635 list.push(n);
2636 }
2637 }
2638 }
2639
2640 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2641 const TypeVect* vt,
2642 Unique_Node_List& partition,
2643 Unique_Node_List& inputs) {
2644 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2645 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2646 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2647
2648 Node* in1 = inputs.at(0);
2649 Node* in2 = inputs.at(1);
2650 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2651
2652 uint func = compute_truth_table(partition, inputs);
2653
2654 Node* pn = partition.at(partition.size() - 1);
2655 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2656 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2657 }
2658
2659 static uint extract_bit(uint func, uint pos) {
2660 return (func & (1 << pos)) >> pos;
2661 }
2662
2663 //
2664 // A macro logic node represents a truth table. It has 4 inputs,
2665 // First three inputs corresponds to 3 columns of a truth table
2666 // and fourth input captures the logic function.
2667 //
2668 // eg. fn = (in1 AND in2) OR in3;
2669 //
2670 // MacroNode(in1,in2,in3,fn)
2671 //
2672 // -----------------
2673 // in1 in2 in3 fn
2674 // -----------------
2675 // 0 0 0 0
2676 // 0 0 1 1
2677 // 0 1 0 0
2678 // 0 1 1 1
2679 // 1 0 0 0
2680 // 1 0 1 1
2681 // 1 1 0 1
2682 // 1 1 1 1
2683 //
2684
2685 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2686 int res = 0;
2687 for (int i = 0; i < 8; i++) {
2688 int bit1 = extract_bit(in1, i);
2689 int bit2 = extract_bit(in2, i);
2690 int bit3 = extract_bit(in3, i);
2691
2692 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2693 int func_bit = extract_bit(func, func_bit_pos);
2694
2695 res |= func_bit << i;
2696 }
2697 return res;
2698 }
2699
2700 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2701 assert(n != nullptr, "");
2702 assert(eval_map.contains(n), "absent");
2703 return *(eval_map.get(n));
2704 }
2705
2706 static void eval_operands(Node* n,
2707 uint& func1, uint& func2, uint& func3,
2708 ResourceHashtable<Node*,uint>& eval_map) {
2709 assert(is_vector_bitwise_op(n), "");
2710
2711 if (is_vector_unary_bitwise_op(n)) {
2712 Node* opnd = n->in(1);
2713 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2714 opnd = n->in(2);
2715 }
2716 func1 = eval_operand(opnd, eval_map);
2717 } else if (is_vector_binary_bitwise_op(n)) {
2718 func1 = eval_operand(n->in(1), eval_map);
2719 func2 = eval_operand(n->in(2), eval_map);
2720 } else {
2721 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2722 func1 = eval_operand(n->in(1), eval_map);
2723 func2 = eval_operand(n->in(2), eval_map);
2724 func3 = eval_operand(n->in(3), eval_map);
2725 }
2726 }
2727
2728 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2729 assert(inputs.size() <= 3, "sanity");
2730 ResourceMark rm;
2731 uint res = 0;
2732 ResourceHashtable<Node*,uint> eval_map;
2733
2734 // Populate precomputed functions for inputs.
2735 // Each input corresponds to one column of 3 input truth-table.
2736 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2737 0xCC, // (_, b, _) -> b
2738 0xF0 }; // (a, _, _) -> a
2739 for (uint i = 0; i < inputs.size(); i++) {
2740 eval_map.put(inputs.at(i), input_funcs[2-i]);
2741 }
2742
2743 for (uint i = 0; i < partition.size(); i++) {
2744 Node* n = partition.at(i);
2745
2746 uint func1 = 0, func2 = 0, func3 = 0;
2747 eval_operands(n, func1, func2, func3, eval_map);
2748
2749 switch (n->Opcode()) {
2750 case Op_OrV:
2751 assert(func3 == 0, "not binary");
2752 res = func1 | func2;
2753 break;
2754 case Op_AndV:
2755 assert(func3 == 0, "not binary");
2756 res = func1 & func2;
2757 break;
2758 case Op_XorV:
2759 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2760 assert(func2 == 0 && func3 == 0, "not unary");
2761 res = (~func1) & 0xFF;
2762 } else {
2763 assert(func3 == 0, "not binary");
2764 res = func1 ^ func2;
2765 }
2766 break;
2767 case Op_MacroLogicV:
2768 // Ordering of inputs may change during evaluation of sub-tree
2769 // containing MacroLogic node as a child node, thus a re-evaluation
2770 // makes sure that function is evaluated in context of current
2771 // inputs.
2772 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2773 break;
2774
2775 default: assert(false, "not supported: %s", n->Name());
2776 }
2777 assert(res <= 0xFF, "invalid");
2778 eval_map.put(n, res);
2779 }
2780 return res;
2781 }
2782
2783 // Criteria under which nodes gets packed into a macro logic node:-
2784 // 1) Parent and both child nodes are all unmasked or masked with
2785 // same predicates.
2786 // 2) Masked parent can be packed with left child if it is predicated
2787 // and both have same predicates.
2788 // 3) Masked parent can be packed with right child if its un-predicated
2789 // or has matching predication condition.
2790 // 4) An unmasked parent can be packed with an unmasked child.
2791 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2792 assert(partition.size() == 0, "not empty");
2793 assert(inputs.size() == 0, "not empty");
2794 if (is_vector_ternary_bitwise_op(n)) {
2795 return false;
2796 }
2797
2798 bool is_unary_op = is_vector_unary_bitwise_op(n);
2799 if (is_unary_op) {
2800 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2801 return false; // too few inputs
2802 }
2803
2804 bool pack_left_child = true;
2805 bool pack_right_child = true;
2806
2807 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2808 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2809
2810 int left_child_input_cnt = 0;
2811 int right_child_input_cnt = 0;
2812
2813 bool parent_is_predicated = n->is_predicated_vector();
2814 bool left_child_predicated = n->in(1)->is_predicated_vector();
2815 bool right_child_predicated = n->in(2)->is_predicated_vector();
2816
2817 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2818 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2819 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2820
2821 do {
2822 if (pack_left_child && left_child_LOP &&
2823 ((!parent_is_predicated && !left_child_predicated) ||
2824 ((parent_is_predicated && left_child_predicated &&
2825 parent_pred == left_child_pred)))) {
2826 partition.push(n->in(1));
2827 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2828 } else {
2829 inputs.push(n->in(1));
2830 left_child_input_cnt = 1;
2831 }
2832
2833 if (pack_right_child && right_child_LOP &&
2834 (!right_child_predicated ||
2835 (right_child_predicated && parent_is_predicated &&
2836 parent_pred == right_child_pred))) {
2837 partition.push(n->in(2));
2838 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2839 } else {
2840 inputs.push(n->in(2));
2841 right_child_input_cnt = 1;
2842 }
2843
2844 if (inputs.size() > 3) {
2845 assert(partition.size() > 0, "");
2846 inputs.clear();
2847 partition.clear();
2848 if (left_child_input_cnt > right_child_input_cnt) {
2849 pack_left_child = false;
2850 } else {
2851 pack_right_child = false;
2852 }
2853 } else {
2854 break;
2855 }
2856 } while(true);
2857
2858 if(partition.size()) {
2859 partition.push(n);
2860 }
2861
2862 return (partition.size() == 2 || partition.size() == 3) &&
2863 (inputs.size() == 2 || inputs.size() == 3);
2864 }
2865
2866 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2867 assert(is_vector_bitwise_op(n), "not a root");
2868
2869 visited.set(n->_idx);
2870
2871 // 1) Do a DFS walk over the logic cone.
2872 for (uint i = 1; i < n->req(); i++) {
2873 Node* in = n->in(i);
2874 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2875 process_logic_cone_root(igvn, in, visited);
2876 }
2877 }
2878
2879 // 2) Bottom up traversal: Merge node[s] with
2880 // the parent to form macro logic node.
2881 Unique_Node_List partition;
2882 Unique_Node_List inputs;
2883 if (compute_logic_cone(n, partition, inputs)) {
2884 const TypeVect* vt = n->bottom_type()->is_vect();
2885 Node* pn = partition.at(partition.size() - 1);
2886 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2887 if (mask == nullptr ||
2888 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2889 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2890 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2891 igvn.replace_node(n, macro_logic);
2892 }
2893 }
2894 }
2895
2896 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2897 ResourceMark rm;
2898 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2899 Unique_Node_List list;
2900 collect_logic_cone_roots(list);
2901
2902 while (list.size() > 0) {
2903 Node* n = list.pop();
2904 const TypeVect* vt = n->bottom_type()->is_vect();
2905 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2906 if (supported) {
2907 VectorSet visited(comp_arena());
2908 process_logic_cone_root(igvn, n, visited);
2909 }
2910 }
2911 }
2912 }
2913
2914 //------------------------------Code_Gen---------------------------------------
2915 // Given a graph, generate code for it
2916 void Compile::Code_Gen() {
2917 if (failing()) {
2918 return;
2919 }
2920
2921 // Perform instruction selection. You might think we could reclaim Matcher
2922 // memory PDQ, but actually the Matcher is used in generating spill code.
2923 // Internals of the Matcher (including some VectorSets) must remain live
2924 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2925 // set a bit in reclaimed memory.
2926
2927 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2928 // nodes. Mapping is only valid at the root of each matched subtree.
2929 NOT_PRODUCT( verify_graph_edges(); )
2930
2931 Matcher matcher;
2932 _matcher = &matcher;
2933 {
2934 TracePhase tp("matcher", &timers[_t_matcher]);
2935 matcher.match();
2936 if (failing()) {
2937 return;
2938 }
2939 }
2940 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2941 // nodes. Mapping is only valid at the root of each matched subtree.
2942 NOT_PRODUCT( verify_graph_edges(); )
2943
2944 // If you have too many nodes, or if matching has failed, bail out
2945 check_node_count(0, "out of nodes matching instructions");
2946 if (failing()) {
2947 return;
2948 }
2949
2950 print_method(PHASE_MATCHING, 2);
2951
2952 // Build a proper-looking CFG
2953 PhaseCFG cfg(node_arena(), root(), matcher);
2954 if (failing()) {
2955 return;
2956 }
2957 _cfg = &cfg;
2958 {
2959 TracePhase tp("scheduler", &timers[_t_scheduler]);
2960 bool success = cfg.do_global_code_motion();
2961 if (!success) {
2962 return;
2963 }
2964
2965 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2966 NOT_PRODUCT( verify_graph_edges(); )
2967 cfg.verify();
2968 }
2969
2970 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2971 _regalloc = ®alloc;
2972 {
2973 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2974 // Perform register allocation. After Chaitin, use-def chains are
2975 // no longer accurate (at spill code) and so must be ignored.
2976 // Node->LRG->reg mappings are still accurate.
2977 _regalloc->Register_Allocate();
2978
2979 // Bail out if the allocator builds too many nodes
2980 if (failing()) {
2981 return;
2982 }
2983
2984 print_method(PHASE_REGISTER_ALLOCATION, 2);
2985 }
2986
2987 // Prior to register allocation we kept empty basic blocks in case the
2988 // the allocator needed a place to spill. After register allocation we
2989 // are not adding any new instructions. If any basic block is empty, we
2990 // can now safely remove it.
2991 {
2992 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2993 cfg.remove_empty_blocks();
2994 if (do_freq_based_layout()) {
2995 PhaseBlockLayout layout(cfg);
2996 } else {
2997 cfg.set_loop_alignment();
2998 }
2999 cfg.fixup_flow();
3000 cfg.remove_unreachable_blocks();
3001 cfg.verify_dominator_tree();
3002 print_method(PHASE_BLOCK_ORDERING, 3);
3003 }
3004
3005 // Apply peephole optimizations
3006 if( OptoPeephole ) {
3007 TracePhase tp("peephole", &timers[_t_peephole]);
3008 PhasePeephole peep( _regalloc, cfg);
3009 peep.do_transform();
3010 print_method(PHASE_PEEPHOLE, 3);
3011 }
3012
3013 // Do late expand if CPU requires this.
3014 if (Matcher::require_postalloc_expand) {
3015 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3016 cfg.postalloc_expand(_regalloc);
3017 print_method(PHASE_POSTALLOC_EXPAND, 3);
3018 }
3019
3020 // Convert Nodes to instruction bits in a buffer
3021 {
3022 TracePhase tp("output", &timers[_t_output]);
3023 PhaseOutput output;
3024 output.Output();
3025 if (failing()) return;
3026 output.install();
3027 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3028 }
3029
3030 // He's dead, Jim.
3031 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3032 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3033 }
3034
3035 //------------------------------Final_Reshape_Counts---------------------------
3036 // This class defines counters to help identify when a method
3037 // may/must be executed using hardware with only 24-bit precision.
3038 struct Final_Reshape_Counts : public StackObj {
3039 int _call_count; // count non-inlined 'common' calls
3040 int _float_count; // count float ops requiring 24-bit precision
3041 int _double_count; // count double ops requiring more precision
3042 int _java_call_count; // count non-inlined 'java' calls
3043 int _inner_loop_count; // count loops which need alignment
3044 VectorSet _visited; // Visitation flags
3045 Node_List _tests; // Set of IfNodes & PCTableNodes
3046
3047 Final_Reshape_Counts() :
3048 _call_count(0), _float_count(0), _double_count(0),
3049 _java_call_count(0), _inner_loop_count(0) { }
3050
3051 void inc_call_count () { _call_count ++; }
3052 void inc_float_count () { _float_count ++; }
3053 void inc_double_count() { _double_count++; }
3054 void inc_java_call_count() { _java_call_count++; }
3055 void inc_inner_loop_count() { _inner_loop_count++; }
3056
3057 int get_call_count () const { return _call_count ; }
3058 int get_float_count () const { return _float_count ; }
3059 int get_double_count() const { return _double_count; }
3060 int get_java_call_count() const { return _java_call_count; }
3061 int get_inner_loop_count() const { return _inner_loop_count; }
3062 };
3063
3064 // Eliminate trivially redundant StoreCMs and accumulate their
3065 // precedence edges.
3066 void Compile::eliminate_redundant_card_marks(Node* n) {
3067 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3068 if (n->in(MemNode::Address)->outcnt() > 1) {
3069 // There are multiple users of the same address so it might be
3070 // possible to eliminate some of the StoreCMs
3071 Node* mem = n->in(MemNode::Memory);
3072 Node* adr = n->in(MemNode::Address);
3073 Node* val = n->in(MemNode::ValueIn);
3074 Node* prev = n;
3075 bool done = false;
3076 // Walk the chain of StoreCMs eliminating ones that match. As
3077 // long as it's a chain of single users then the optimization is
3078 // safe. Eliminating partially redundant StoreCMs would require
3079 // cloning copies down the other paths.
3080 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3081 if (adr == mem->in(MemNode::Address) &&
3082 val == mem->in(MemNode::ValueIn)) {
3083 // redundant StoreCM
3084 if (mem->req() > MemNode::OopStore) {
3085 // Hasn't been processed by this code yet.
3086 n->add_prec(mem->in(MemNode::OopStore));
3087 } else {
3088 // Already converted to precedence edge
3089 for (uint i = mem->req(); i < mem->len(); i++) {
3090 // Accumulate any precedence edges
3091 if (mem->in(i) != nullptr) {
3092 n->add_prec(mem->in(i));
3093 }
3094 }
3095 // Everything above this point has been processed.
3096 done = true;
3097 }
3098 // Eliminate the previous StoreCM
3099 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3100 assert(mem->outcnt() == 0, "should be dead");
3101 mem->disconnect_inputs(this);
3102 } else {
3103 prev = mem;
3104 }
3105 mem = prev->in(MemNode::Memory);
3106 }
3107 }
3108 }
3109
3110 //------------------------------final_graph_reshaping_impl----------------------
3111 // Implement items 1-5 from final_graph_reshaping below.
3112 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3113
3114 if ( n->outcnt() == 0 ) return; // dead node
3115 uint nop = n->Opcode();
3116
3117 // Check for 2-input instruction with "last use" on right input.
3118 // Swap to left input. Implements item (2).
3119 if( n->req() == 3 && // two-input instruction
3120 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3121 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3122 n->in(2)->outcnt() == 1 &&// right use IS a last use
3123 !n->in(2)->is_Con() ) { // right use is not a constant
3124 // Check for commutative opcode
3125 switch( nop ) {
3126 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3127 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3128 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3129 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3130 case Op_AndL: case Op_XorL: case Op_OrL:
3131 case Op_AndI: case Op_XorI: case Op_OrI: {
3132 // Move "last use" input to left by swapping inputs
3133 n->swap_edges(1, 2);
3134 break;
3135 }
3136 default:
3137 break;
3138 }
3139 }
3140
3141 #ifdef ASSERT
3142 if( n->is_Mem() ) {
3143 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3144 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3145 // oop will be recorded in oop map if load crosses safepoint
3146 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3147 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3148 "raw memory operations should have control edge");
3149 }
3150 if (n->is_MemBar()) {
3151 MemBarNode* mb = n->as_MemBar();
3152 if (mb->trailing_store() || mb->trailing_load_store()) {
3153 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3154 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3155 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3156 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3157 } else if (mb->leading()) {
3158 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3159 }
3160 }
3161 #endif
3162 // Count FPU ops and common calls, implements item (3)
3163 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3164 if (!gc_handled) {
3165 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3166 }
3167
3168 // Collect CFG split points
3169 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3170 frc._tests.push(n);
3171 }
3172 }
3173
3174 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3175 if (!UseDivMod) {
3176 return;
3177 }
3178
3179 // Check if "a % b" and "a / b" both exist
3180 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3181 if (d == nullptr) {
3182 return;
3183 }
3184
3185 // Replace them with a fused divmod if supported
3186 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3187 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3188 d->subsume_by(divmod->div_proj(), this);
3189 n->subsume_by(divmod->mod_proj(), this);
3190 } else {
3191 // Replace "a % b" with "a - ((a / b) * b)"
3192 Node* mult = MulNode::make(d, d->in(2), bt);
3193 Node* sub = SubNode::make(d->in(1), mult, bt);
3194 n->subsume_by(sub, this);
3195 }
3196 }
3197
3198 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3199 switch( nop ) {
3200 // Count all float operations that may use FPU
3201 case Op_AddF:
3202 case Op_SubF:
3203 case Op_MulF:
3204 case Op_DivF:
3205 case Op_NegF:
3206 case Op_ModF:
3207 case Op_ConvI2F:
3208 case Op_ConF:
3209 case Op_CmpF:
3210 case Op_CmpF3:
3211 case Op_StoreF:
3212 case Op_LoadF:
3213 // case Op_ConvL2F: // longs are split into 32-bit halves
3214 frc.inc_float_count();
3215 break;
3216
3217 case Op_ConvF2D:
3218 case Op_ConvD2F:
3219 frc.inc_float_count();
3220 frc.inc_double_count();
3221 break;
3222
3223 // Count all double operations that may use FPU
3224 case Op_AddD:
3225 case Op_SubD:
3226 case Op_MulD:
3227 case Op_DivD:
3228 case Op_NegD:
3229 case Op_ModD:
3230 case Op_ConvI2D:
3231 case Op_ConvD2I:
3232 // case Op_ConvL2D: // handled by leaf call
3233 // case Op_ConvD2L: // handled by leaf call
3234 case Op_ConD:
3235 case Op_CmpD:
3236 case Op_CmpD3:
3237 case Op_StoreD:
3238 case Op_LoadD:
3239 case Op_LoadD_unaligned:
3240 frc.inc_double_count();
3241 break;
3242 case Op_Opaque1: // Remove Opaque Nodes before matching
3243 n->subsume_by(n->in(1), this);
3244 break;
3245 case Op_CallStaticJava:
3246 case Op_CallJava:
3247 case Op_CallDynamicJava:
3248 frc.inc_java_call_count(); // Count java call site;
3249 case Op_CallRuntime:
3250 case Op_CallLeaf:
3251 case Op_CallLeafVector:
3252 case Op_CallLeafNoFP: {
3253 assert (n->is_Call(), "");
3254 CallNode *call = n->as_Call();
3255 // Count call sites where the FP mode bit would have to be flipped.
3256 // Do not count uncommon runtime calls:
3257 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3258 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3259 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3260 frc.inc_call_count(); // Count the call site
3261 } else { // See if uncommon argument is shared
3262 Node *n = call->in(TypeFunc::Parms);
3263 int nop = n->Opcode();
3264 // Clone shared simple arguments to uncommon calls, item (1).
3265 if (n->outcnt() > 1 &&
3266 !n->is_Proj() &&
3267 nop != Op_CreateEx &&
3268 nop != Op_CheckCastPP &&
3269 nop != Op_DecodeN &&
3270 nop != Op_DecodeNKlass &&
3271 !n->is_Mem() &&
3272 !n->is_Phi()) {
3273 Node *x = n->clone();
3274 call->set_req(TypeFunc::Parms, x);
3275 }
3276 }
3277 break;
3278 }
3279
3280 case Op_StoreCM:
3281 {
3282 // Convert OopStore dependence into precedence edge
3283 Node* prec = n->in(MemNode::OopStore);
3284 n->del_req(MemNode::OopStore);
3285 n->add_prec(prec);
3286 eliminate_redundant_card_marks(n);
3287 }
3288
3289 // fall through
3290
3291 case Op_StoreB:
3292 case Op_StoreC:
3293 case Op_StoreI:
3294 case Op_StoreL:
3295 case Op_CompareAndSwapB:
3296 case Op_CompareAndSwapS:
3297 case Op_CompareAndSwapI:
3298 case Op_CompareAndSwapL:
3299 case Op_CompareAndSwapP:
3300 case Op_CompareAndSwapN:
3301 case Op_WeakCompareAndSwapB:
3302 case Op_WeakCompareAndSwapS:
3303 case Op_WeakCompareAndSwapI:
3304 case Op_WeakCompareAndSwapL:
3305 case Op_WeakCompareAndSwapP:
3306 case Op_WeakCompareAndSwapN:
3307 case Op_CompareAndExchangeB:
3308 case Op_CompareAndExchangeS:
3309 case Op_CompareAndExchangeI:
3310 case Op_CompareAndExchangeL:
3311 case Op_CompareAndExchangeP:
3312 case Op_CompareAndExchangeN:
3313 case Op_GetAndAddS:
3314 case Op_GetAndAddB:
3315 case Op_GetAndAddI:
3316 case Op_GetAndAddL:
3317 case Op_GetAndSetS:
3318 case Op_GetAndSetB:
3319 case Op_GetAndSetI:
3320 case Op_GetAndSetL:
3321 case Op_GetAndSetP:
3322 case Op_GetAndSetN:
3323 case Op_StoreP:
3324 case Op_StoreN:
3325 case Op_StoreNKlass:
3326 case Op_LoadB:
3327 case Op_LoadUB:
3328 case Op_LoadUS:
3329 case Op_LoadI:
3330 case Op_LoadKlass:
3331 case Op_LoadNKlass:
3332 case Op_LoadL:
3333 case Op_LoadL_unaligned:
3334 case Op_LoadP:
3335 case Op_LoadN:
3336 case Op_LoadRange:
3337 case Op_LoadS:
3338 break;
3339
3340 case Op_AddP: { // Assert sane base pointers
3341 Node *addp = n->in(AddPNode::Address);
3342 assert( !addp->is_AddP() ||
3343 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3344 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3345 "Base pointers must match (addp %u)", addp->_idx );
3346 #ifdef _LP64
3347 if ((UseCompressedOops || UseCompressedClassPointers) &&
3348 addp->Opcode() == Op_ConP &&
3349 addp == n->in(AddPNode::Base) &&
3350 n->in(AddPNode::Offset)->is_Con()) {
3351 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3352 // on the platform and on the compressed oops mode.
3353 // Use addressing with narrow klass to load with offset on x86.
3354 // Some platforms can use the constant pool to load ConP.
3355 // Do this transformation here since IGVN will convert ConN back to ConP.
3356 const Type* t = addp->bottom_type();
3357 bool is_oop = t->isa_oopptr() != nullptr;
3358 bool is_klass = t->isa_klassptr() != nullptr;
3359
3360 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3361 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3362 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3363 Node* nn = nullptr;
3364
3365 int op = is_oop ? Op_ConN : Op_ConNKlass;
3366
3367 // Look for existing ConN node of the same exact type.
3368 Node* r = root();
3369 uint cnt = r->outcnt();
3370 for (uint i = 0; i < cnt; i++) {
3371 Node* m = r->raw_out(i);
3372 if (m!= nullptr && m->Opcode() == op &&
3373 m->bottom_type()->make_ptr() == t) {
3374 nn = m;
3375 break;
3376 }
3377 }
3378 if (nn != nullptr) {
3379 // Decode a narrow oop to match address
3380 // [R12 + narrow_oop_reg<<3 + offset]
3381 if (is_oop) {
3382 nn = new DecodeNNode(nn, t);
3383 } else {
3384 nn = new DecodeNKlassNode(nn, t);
3385 }
3386 // Check for succeeding AddP which uses the same Base.
3387 // Otherwise we will run into the assertion above when visiting that guy.
3388 for (uint i = 0; i < n->outcnt(); ++i) {
3389 Node *out_i = n->raw_out(i);
3390 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3391 out_i->set_req(AddPNode::Base, nn);
3392 #ifdef ASSERT
3393 for (uint j = 0; j < out_i->outcnt(); ++j) {
3394 Node *out_j = out_i->raw_out(j);
3395 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3396 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3397 }
3398 #endif
3399 }
3400 }
3401 n->set_req(AddPNode::Base, nn);
3402 n->set_req(AddPNode::Address, nn);
3403 if (addp->outcnt() == 0) {
3404 addp->disconnect_inputs(this);
3405 }
3406 }
3407 }
3408 }
3409 #endif
3410 break;
3411 }
3412
3413 case Op_CastPP: {
3414 // Remove CastPP nodes to gain more freedom during scheduling but
3415 // keep the dependency they encode as control or precedence edges
3416 // (if control is set already) on memory operations. Some CastPP
3417 // nodes don't have a control (don't carry a dependency): skip
3418 // those.
3419 if (n->in(0) != nullptr) {
3420 ResourceMark rm;
3421 Unique_Node_List wq;
3422 wq.push(n);
3423 for (uint next = 0; next < wq.size(); ++next) {
3424 Node *m = wq.at(next);
3425 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3426 Node* use = m->fast_out(i);
3427 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3428 use->ensure_control_or_add_prec(n->in(0));
3429 } else {
3430 switch(use->Opcode()) {
3431 case Op_AddP:
3432 case Op_DecodeN:
3433 case Op_DecodeNKlass:
3434 case Op_CheckCastPP:
3435 case Op_CastPP:
3436 wq.push(use);
3437 break;
3438 }
3439 }
3440 }
3441 }
3442 }
3443 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3444 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3445 Node* in1 = n->in(1);
3446 const Type* t = n->bottom_type();
3447 Node* new_in1 = in1->clone();
3448 new_in1->as_DecodeN()->set_type(t);
3449
3450 if (!Matcher::narrow_oop_use_complex_address()) {
3451 //
3452 // x86, ARM and friends can handle 2 adds in addressing mode
3453 // and Matcher can fold a DecodeN node into address by using
3454 // a narrow oop directly and do implicit null check in address:
3455 //
3456 // [R12 + narrow_oop_reg<<3 + offset]
3457 // NullCheck narrow_oop_reg
3458 //
3459 // On other platforms (Sparc) we have to keep new DecodeN node and
3460 // use it to do implicit null check in address:
3461 //
3462 // decode_not_null narrow_oop_reg, base_reg
3463 // [base_reg + offset]
3464 // NullCheck base_reg
3465 //
3466 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3467 // to keep the information to which null check the new DecodeN node
3468 // corresponds to use it as value in implicit_null_check().
3469 //
3470 new_in1->set_req(0, n->in(0));
3471 }
3472
3473 n->subsume_by(new_in1, this);
3474 if (in1->outcnt() == 0) {
3475 in1->disconnect_inputs(this);
3476 }
3477 } else {
3478 n->subsume_by(n->in(1), this);
3479 if (n->outcnt() == 0) {
3480 n->disconnect_inputs(this);
3481 }
3482 }
3483 break;
3484 }
3485 #ifdef _LP64
3486 case Op_CmpP:
3487 // Do this transformation here to preserve CmpPNode::sub() and
3488 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3489 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3490 Node* in1 = n->in(1);
3491 Node* in2 = n->in(2);
3492 if (!in1->is_DecodeNarrowPtr()) {
3493 in2 = in1;
3494 in1 = n->in(2);
3495 }
3496 assert(in1->is_DecodeNarrowPtr(), "sanity");
3497
3498 Node* new_in2 = nullptr;
3499 if (in2->is_DecodeNarrowPtr()) {
3500 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3501 new_in2 = in2->in(1);
3502 } else if (in2->Opcode() == Op_ConP) {
3503 const Type* t = in2->bottom_type();
3504 if (t == TypePtr::NULL_PTR) {
3505 assert(in1->is_DecodeN(), "compare klass to null?");
3506 // Don't convert CmpP null check into CmpN if compressed
3507 // oops implicit null check is not generated.
3508 // This will allow to generate normal oop implicit null check.
3509 if (Matcher::gen_narrow_oop_implicit_null_checks())
3510 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3511 //
3512 // This transformation together with CastPP transformation above
3513 // will generated code for implicit null checks for compressed oops.
3514 //
3515 // The original code after Optimize()
3516 //
3517 // LoadN memory, narrow_oop_reg
3518 // decode narrow_oop_reg, base_reg
3519 // CmpP base_reg, nullptr
3520 // CastPP base_reg // NotNull
3521 // Load [base_reg + offset], val_reg
3522 //
3523 // after these transformations will be
3524 //
3525 // LoadN memory, narrow_oop_reg
3526 // CmpN narrow_oop_reg, nullptr
3527 // decode_not_null narrow_oop_reg, base_reg
3528 // Load [base_reg + offset], val_reg
3529 //
3530 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3531 // since narrow oops can be used in debug info now (see the code in
3532 // final_graph_reshaping_walk()).
3533 //
3534 // At the end the code will be matched to
3535 // on x86:
3536 //
3537 // Load_narrow_oop memory, narrow_oop_reg
3538 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3539 // NullCheck narrow_oop_reg
3540 //
3541 // and on sparc:
3542 //
3543 // Load_narrow_oop memory, narrow_oop_reg
3544 // decode_not_null narrow_oop_reg, base_reg
3545 // Load [base_reg + offset], val_reg
3546 // NullCheck base_reg
3547 //
3548 } else if (t->isa_oopptr()) {
3549 new_in2 = ConNode::make(t->make_narrowoop());
3550 } else if (t->isa_klassptr()) {
3551 new_in2 = ConNode::make(t->make_narrowklass());
3552 }
3553 }
3554 if (new_in2 != nullptr) {
3555 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3556 n->subsume_by(cmpN, this);
3557 if (in1->outcnt() == 0) {
3558 in1->disconnect_inputs(this);
3559 }
3560 if (in2->outcnt() == 0) {
3561 in2->disconnect_inputs(this);
3562 }
3563 }
3564 }
3565 break;
3566
3567 case Op_DecodeN:
3568 case Op_DecodeNKlass:
3569 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3570 // DecodeN could be pinned when it can't be fold into
3571 // an address expression, see the code for Op_CastPP above.
3572 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3573 break;
3574
3575 case Op_EncodeP:
3576 case Op_EncodePKlass: {
3577 Node* in1 = n->in(1);
3578 if (in1->is_DecodeNarrowPtr()) {
3579 n->subsume_by(in1->in(1), this);
3580 } else if (in1->Opcode() == Op_ConP) {
3581 const Type* t = in1->bottom_type();
3582 if (t == TypePtr::NULL_PTR) {
3583 assert(t->isa_oopptr(), "null klass?");
3584 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3585 } else if (t->isa_oopptr()) {
3586 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3587 } else if (t->isa_klassptr()) {
3588 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3589 }
3590 }
3591 if (in1->outcnt() == 0) {
3592 in1->disconnect_inputs(this);
3593 }
3594 break;
3595 }
3596
3597 case Op_Proj: {
3598 if (OptimizeStringConcat || IncrementalInline) {
3599 ProjNode* proj = n->as_Proj();
3600 if (proj->_is_io_use) {
3601 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3602 // Separate projections were used for the exception path which
3603 // are normally removed by a late inline. If it wasn't inlined
3604 // then they will hang around and should just be replaced with
3605 // the original one. Merge them.
3606 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3607 if (non_io_proj != nullptr) {
3608 proj->subsume_by(non_io_proj , this);
3609 }
3610 }
3611 }
3612 break;
3613 }
3614
3615 case Op_Phi:
3616 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3617 // The EncodeP optimization may create Phi with the same edges
3618 // for all paths. It is not handled well by Register Allocator.
3619 Node* unique_in = n->in(1);
3620 assert(unique_in != nullptr, "");
3621 uint cnt = n->req();
3622 for (uint i = 2; i < cnt; i++) {
3623 Node* m = n->in(i);
3624 assert(m != nullptr, "");
3625 if (unique_in != m)
3626 unique_in = nullptr;
3627 }
3628 if (unique_in != nullptr) {
3629 n->subsume_by(unique_in, this);
3630 }
3631 }
3632 break;
3633
3634 #endif
3635
3636 #ifdef ASSERT
3637 case Op_CastII:
3638 // Verify that all range check dependent CastII nodes were removed.
3639 if (n->isa_CastII()->has_range_check()) {
3640 n->dump(3);
3641 assert(false, "Range check dependent CastII node was not removed");
3642 }
3643 break;
3644 #endif
3645
3646 case Op_ModI:
3647 handle_div_mod_op(n, T_INT, false);
3648 break;
3649
3650 case Op_ModL:
3651 handle_div_mod_op(n, T_LONG, false);
3652 break;
3653
3654 case Op_UModI:
3655 handle_div_mod_op(n, T_INT, true);
3656 break;
3657
3658 case Op_UModL:
3659 handle_div_mod_op(n, T_LONG, true);
3660 break;
3661
3662 case Op_LoadVector:
3663 case Op_StoreVector:
3664 #ifdef ASSERT
3665 // Add VerifyVectorAlignment node between adr and load / store.
3666 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3667 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3668 n->as_StoreVector()->must_verify_alignment();
3669 if (must_verify_alignment) {
3670 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3671 n->as_StoreVector()->memory_size();
3672 // The memory access should be aligned to the vector width in bytes.
3673 // However, the underlying array is possibly less well aligned, but at least
3674 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3675 // a loop we can expect at least the following alignment:
3676 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3677 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3678 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3679 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3680 jlong mask = guaranteed_alignment - 1;
3681 Node* mask_con = ConLNode::make(mask);
3682 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3683 n->set_req(MemNode::Address, va);
3684 }
3685 }
3686 #endif
3687 break;
3688
3689 case Op_LoadVectorGather:
3690 case Op_StoreVectorScatter:
3691 case Op_LoadVectorGatherMasked:
3692 case Op_StoreVectorScatterMasked:
3693 case Op_VectorCmpMasked:
3694 case Op_VectorMaskGen:
3695 case Op_LoadVectorMasked:
3696 case Op_StoreVectorMasked:
3697 break;
3698
3699 case Op_AddReductionVI:
3700 case Op_AddReductionVL:
3701 case Op_AddReductionVF:
3702 case Op_AddReductionVD:
3703 case Op_MulReductionVI:
3704 case Op_MulReductionVL:
3705 case Op_MulReductionVF:
3706 case Op_MulReductionVD:
3707 case Op_MinReductionV:
3708 case Op_MaxReductionV:
3709 case Op_AndReductionV:
3710 case Op_OrReductionV:
3711 case Op_XorReductionV:
3712 break;
3713
3714 case Op_PackB:
3715 case Op_PackS:
3716 case Op_PackI:
3717 case Op_PackF:
3718 case Op_PackL:
3719 case Op_PackD:
3720 if (n->req()-1 > 2) {
3721 // Replace many operand PackNodes with a binary tree for matching
3722 PackNode* p = (PackNode*) n;
3723 Node* btp = p->binary_tree_pack(1, n->req());
3724 n->subsume_by(btp, this);
3725 }
3726 break;
3727 case Op_Loop:
3728 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3729 case Op_CountedLoop:
3730 case Op_LongCountedLoop:
3731 case Op_OuterStripMinedLoop:
3732 if (n->as_Loop()->is_inner_loop()) {
3733 frc.inc_inner_loop_count();
3734 }
3735 n->as_Loop()->verify_strip_mined(0);
3736 break;
3737 case Op_LShiftI:
3738 case Op_RShiftI:
3739 case Op_URShiftI:
3740 case Op_LShiftL:
3741 case Op_RShiftL:
3742 case Op_URShiftL:
3743 if (Matcher::need_masked_shift_count) {
3744 // The cpu's shift instructions don't restrict the count to the
3745 // lower 5/6 bits. We need to do the masking ourselves.
3746 Node* in2 = n->in(2);
3747 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3748 const TypeInt* t = in2->find_int_type();
3749 if (t != nullptr && t->is_con()) {
3750 juint shift = t->get_con();
3751 if (shift > mask) { // Unsigned cmp
3752 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3753 }
3754 } else {
3755 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3756 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3757 n->set_req(2, shift);
3758 }
3759 }
3760 if (in2->outcnt() == 0) { // Remove dead node
3761 in2->disconnect_inputs(this);
3762 }
3763 }
3764 break;
3765 case Op_MemBarStoreStore:
3766 case Op_MemBarRelease:
3767 // Break the link with AllocateNode: it is no longer useful and
3768 // confuses register allocation.
3769 if (n->req() > MemBarNode::Precedent) {
3770 n->set_req(MemBarNode::Precedent, top());
3771 }
3772 break;
3773 case Op_MemBarAcquire: {
3774 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3775 // At parse time, the trailing MemBarAcquire for a volatile load
3776 // is created with an edge to the load. After optimizations,
3777 // that input may be a chain of Phis. If those phis have no
3778 // other use, then the MemBarAcquire keeps them alive and
3779 // register allocation can be confused.
3780 dead_nodes.push(n->in(MemBarNode::Precedent));
3781 n->set_req(MemBarNode::Precedent, top());
3782 }
3783 break;
3784 }
3785 case Op_Blackhole:
3786 break;
3787 case Op_RangeCheck: {
3788 RangeCheckNode* rc = n->as_RangeCheck();
3789 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3790 n->subsume_by(iff, this);
3791 frc._tests.push(iff);
3792 break;
3793 }
3794 case Op_ConvI2L: {
3795 if (!Matcher::convi2l_type_required) {
3796 // Code generation on some platforms doesn't need accurate
3797 // ConvI2L types. Widening the type can help remove redundant
3798 // address computations.
3799 n->as_Type()->set_type(TypeLong::INT);
3800 ResourceMark rm;
3801 Unique_Node_List wq;
3802 wq.push(n);
3803 for (uint next = 0; next < wq.size(); next++) {
3804 Node *m = wq.at(next);
3805
3806 for(;;) {
3807 // Loop over all nodes with identical inputs edges as m
3808 Node* k = m->find_similar(m->Opcode());
3809 if (k == nullptr) {
3810 break;
3811 }
3812 // Push their uses so we get a chance to remove node made
3813 // redundant
3814 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3815 Node* u = k->fast_out(i);
3816 if (u->Opcode() == Op_LShiftL ||
3817 u->Opcode() == Op_AddL ||
3818 u->Opcode() == Op_SubL ||
3819 u->Opcode() == Op_AddP) {
3820 wq.push(u);
3821 }
3822 }
3823 // Replace all nodes with identical edges as m with m
3824 k->subsume_by(m, this);
3825 }
3826 }
3827 }
3828 break;
3829 }
3830 case Op_CmpUL: {
3831 if (!Matcher::has_match_rule(Op_CmpUL)) {
3832 // No support for unsigned long comparisons
3833 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3834 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3835 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3836 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3837 Node* andl = new AndLNode(orl, remove_sign_mask);
3838 Node* cmp = new CmpLNode(andl, n->in(2));
3839 n->subsume_by(cmp, this);
3840 }
3841 break;
3842 }
3843 default:
3844 assert(!n->is_Call(), "");
3845 assert(!n->is_Mem(), "");
3846 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3847 break;
3848 }
3849 }
3850
3851 //------------------------------final_graph_reshaping_walk---------------------
3852 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3853 // requires that the walk visits a node's inputs before visiting the node.
3854 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3855 Unique_Node_List sfpt;
3856
3857 frc._visited.set(root->_idx); // first, mark node as visited
3858 uint cnt = root->req();
3859 Node *n = root;
3860 uint i = 0;
3861 while (true) {
3862 if (i < cnt) {
3863 // Place all non-visited non-null inputs onto stack
3864 Node* m = n->in(i);
3865 ++i;
3866 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3867 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3868 // compute worst case interpreter size in case of a deoptimization
3869 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3870
3871 sfpt.push(m);
3872 }
3873 cnt = m->req();
3874 nstack.push(n, i); // put on stack parent and next input's index
3875 n = m;
3876 i = 0;
3877 }
3878 } else {
3879 // Now do post-visit work
3880 final_graph_reshaping_impl(n, frc, dead_nodes);
3881 if (nstack.is_empty())
3882 break; // finished
3883 n = nstack.node(); // Get node from stack
3884 cnt = n->req();
3885 i = nstack.index();
3886 nstack.pop(); // Shift to the next node on stack
3887 }
3888 }
3889
3890 // Skip next transformation if compressed oops are not used.
3891 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3892 (!UseCompressedOops && !UseCompressedClassPointers))
3893 return;
3894
3895 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3896 // It could be done for an uncommon traps or any safepoints/calls
3897 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3898 while (sfpt.size() > 0) {
3899 n = sfpt.pop();
3900 JVMState *jvms = n->as_SafePoint()->jvms();
3901 assert(jvms != nullptr, "sanity");
3902 int start = jvms->debug_start();
3903 int end = n->req();
3904 bool is_uncommon = (n->is_CallStaticJava() &&
3905 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3906 for (int j = start; j < end; j++) {
3907 Node* in = n->in(j);
3908 if (in->is_DecodeNarrowPtr()) {
3909 bool safe_to_skip = true;
3910 if (!is_uncommon ) {
3911 // Is it safe to skip?
3912 for (uint i = 0; i < in->outcnt(); i++) {
3913 Node* u = in->raw_out(i);
3914 if (!u->is_SafePoint() ||
3915 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3916 safe_to_skip = false;
3917 }
3918 }
3919 }
3920 if (safe_to_skip) {
3921 n->set_req(j, in->in(1));
3922 }
3923 if (in->outcnt() == 0) {
3924 in->disconnect_inputs(this);
3925 }
3926 }
3927 }
3928 }
3929 }
3930
3931 //------------------------------final_graph_reshaping--------------------------
3932 // Final Graph Reshaping.
3933 //
3934 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3935 // and not commoned up and forced early. Must come after regular
3936 // optimizations to avoid GVN undoing the cloning. Clone constant
3937 // inputs to Loop Phis; these will be split by the allocator anyways.
3938 // Remove Opaque nodes.
3939 // (2) Move last-uses by commutative operations to the left input to encourage
3940 // Intel update-in-place two-address operations and better register usage
3941 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3942 // calls canonicalizing them back.
3943 // (3) Count the number of double-precision FP ops, single-precision FP ops
3944 // and call sites. On Intel, we can get correct rounding either by
3945 // forcing singles to memory (requires extra stores and loads after each
3946 // FP bytecode) or we can set a rounding mode bit (requires setting and
3947 // clearing the mode bit around call sites). The mode bit is only used
3948 // if the relative frequency of single FP ops to calls is low enough.
3949 // This is a key transform for SPEC mpeg_audio.
3950 // (4) Detect infinite loops; blobs of code reachable from above but not
3951 // below. Several of the Code_Gen algorithms fail on such code shapes,
3952 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3953 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3954 // Detection is by looking for IfNodes where only 1 projection is
3955 // reachable from below or CatchNodes missing some targets.
3956 // (5) Assert for insane oop offsets in debug mode.
3957
3958 bool Compile::final_graph_reshaping() {
3959 // an infinite loop may have been eliminated by the optimizer,
3960 // in which case the graph will be empty.
3961 if (root()->req() == 1) {
3962 // Do not compile method that is only a trivial infinite loop,
3963 // since the content of the loop may have been eliminated.
3964 record_method_not_compilable("trivial infinite loop");
3965 return true;
3966 }
3967
3968 // Expensive nodes have their control input set to prevent the GVN
3969 // from freely commoning them. There's no GVN beyond this point so
3970 // no need to keep the control input. We want the expensive nodes to
3971 // be freely moved to the least frequent code path by gcm.
3972 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3973 for (int i = 0; i < expensive_count(); i++) {
3974 _expensive_nodes.at(i)->set_req(0, nullptr);
3975 }
3976
3977 Final_Reshape_Counts frc;
3978
3979 // Visit everybody reachable!
3980 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3981 Node_Stack nstack(live_nodes() >> 1);
3982 Unique_Node_List dead_nodes;
3983 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
3984
3985 // Check for unreachable (from below) code (i.e., infinite loops).
3986 for( uint i = 0; i < frc._tests.size(); i++ ) {
3987 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3988 // Get number of CFG targets.
3989 // Note that PCTables include exception targets after calls.
3990 uint required_outcnt = n->required_outcnt();
3991 if (n->outcnt() != required_outcnt) {
3992 // Check for a few special cases. Rethrow Nodes never take the
3993 // 'fall-thru' path, so expected kids is 1 less.
3994 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3995 if (n->in(0)->in(0)->is_Call()) {
3996 CallNode* call = n->in(0)->in(0)->as_Call();
3997 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3998 required_outcnt--; // Rethrow always has 1 less kid
3999 } else if (call->req() > TypeFunc::Parms &&
4000 call->is_CallDynamicJava()) {
4001 // Check for null receiver. In such case, the optimizer has
4002 // detected that the virtual call will always result in a null
4003 // pointer exception. The fall-through projection of this CatchNode
4004 // will not be populated.
4005 Node* arg0 = call->in(TypeFunc::Parms);
4006 if (arg0->is_Type() &&
4007 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4008 required_outcnt--;
4009 }
4010 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4011 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4012 // Check for illegal array length. In such case, the optimizer has
4013 // detected that the allocation attempt will always result in an
4014 // exception. There is no fall-through projection of this CatchNode .
4015 assert(call->is_CallStaticJava(), "static call expected");
4016 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4017 uint valid_length_test_input = call->req() - 1;
4018 Node* valid_length_test = call->in(valid_length_test_input);
4019 call->del_req(valid_length_test_input);
4020 if (valid_length_test->find_int_con(1) == 0) {
4021 required_outcnt--;
4022 }
4023 dead_nodes.push(valid_length_test);
4024 assert(n->outcnt() == required_outcnt, "malformed control flow");
4025 continue;
4026 }
4027 }
4028 }
4029
4030 // Recheck with a better notion of 'required_outcnt'
4031 if (n->outcnt() != required_outcnt) {
4032 record_method_not_compilable("malformed control flow");
4033 return true; // Not all targets reachable!
4034 }
4035 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4036 CallNode* call = n->in(0)->in(0)->as_Call();
4037 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4038 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4039 assert(call->is_CallStaticJava(), "static call expected");
4040 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4041 uint valid_length_test_input = call->req() - 1;
4042 dead_nodes.push(call->in(valid_length_test_input));
4043 call->del_req(valid_length_test_input); // valid length test useless now
4044 }
4045 }
4046 // Check that I actually visited all kids. Unreached kids
4047 // must be infinite loops.
4048 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4049 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4050 record_method_not_compilable("infinite loop");
4051 return true; // Found unvisited kid; must be unreach
4052 }
4053
4054 // Here so verification code in final_graph_reshaping_walk()
4055 // always see an OuterStripMinedLoopEnd
4056 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4057 IfNode* init_iff = n->as_If();
4058 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4059 n->subsume_by(iff, this);
4060 }
4061 }
4062
4063 while (dead_nodes.size() > 0) {
4064 Node* m = dead_nodes.pop();
4065 if (m->outcnt() == 0 && m != top()) {
4066 for (uint j = 0; j < m->req(); j++) {
4067 Node* in = m->in(j);
4068 if (in != nullptr) {
4069 dead_nodes.push(in);
4070 }
4071 }
4072 m->disconnect_inputs(this);
4073 }
4074 }
4075
4076 #ifdef IA32
4077 // If original bytecodes contained a mixture of floats and doubles
4078 // check if the optimizer has made it homogeneous, item (3).
4079 if (UseSSE == 0 &&
4080 frc.get_float_count() > 32 &&
4081 frc.get_double_count() == 0 &&
4082 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4083 set_24_bit_selection_and_mode(false, true);
4084 }
4085 #endif // IA32
4086
4087 set_java_calls(frc.get_java_call_count());
4088 set_inner_loops(frc.get_inner_loop_count());
4089
4090 // No infinite loops, no reason to bail out.
4091 return false;
4092 }
4093
4094 //-----------------------------too_many_traps----------------------------------
4095 // Report if there are too many traps at the current method and bci.
4096 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4097 bool Compile::too_many_traps(ciMethod* method,
4098 int bci,
4099 Deoptimization::DeoptReason reason) {
4100 if (method->has_trap_at(bci)) {
4101 return true;
4102 }
4103 ciMethodData* md = method->method_data();
4104 if (md->is_empty()) {
4105 // Assume the trap has not occurred, or that it occurred only
4106 // because of a transient condition during start-up in the interpreter.
4107 return false;
4108 }
4109 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4110 if (md->has_trap_at(bci, m, reason) != 0) {
4111 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4112 // Also, if there are multiple reasons, or if there is no per-BCI record,
4113 // assume the worst.
4114 if (log())
4115 log()->elem("observe trap='%s' count='%d'",
4116 Deoptimization::trap_reason_name(reason),
4117 md->trap_count(reason));
4118 return true;
4119 } else {
4120 // Ignore method/bci and see if there have been too many globally.
4121 return too_many_traps(reason, md);
4122 }
4123 }
4124
4125 // Less-accurate variant which does not require a method and bci.
4126 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4127 ciMethodData* logmd) {
4128 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4129 // Too many traps globally.
4130 // Note that we use cumulative trap_count, not just md->trap_count.
4131 if (log()) {
4132 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4133 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4134 Deoptimization::trap_reason_name(reason),
4135 mcount, trap_count(reason));
4136 }
4137 return true;
4138 } else {
4139 // The coast is clear.
4140 return false;
4141 }
4142 }
4143
4144 //--------------------------too_many_recompiles--------------------------------
4145 // Report if there are too many recompiles at the current method and bci.
4146 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4147 // Is not eager to return true, since this will cause the compiler to use
4148 // Action_none for a trap point, to avoid too many recompilations.
4149 bool Compile::too_many_recompiles(ciMethod* method,
4150 int bci,
4151 Deoptimization::DeoptReason reason) {
4152 ciMethodData* md = method->method_data();
4153 if (md->is_empty()) {
4154 // Assume the trap has not occurred, or that it occurred only
4155 // because of a transient condition during start-up in the interpreter.
4156 return false;
4157 }
4158 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4159 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4160 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4161 Deoptimization::DeoptReason per_bc_reason
4162 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4163 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4164 if ((per_bc_reason == Deoptimization::Reason_none
4165 || md->has_trap_at(bci, m, reason) != 0)
4166 // The trap frequency measure we care about is the recompile count:
4167 && md->trap_recompiled_at(bci, m)
4168 && md->overflow_recompile_count() >= bc_cutoff) {
4169 // Do not emit a trap here if it has already caused recompilations.
4170 // Also, if there are multiple reasons, or if there is no per-BCI record,
4171 // assume the worst.
4172 if (log())
4173 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4174 Deoptimization::trap_reason_name(reason),
4175 md->trap_count(reason),
4176 md->overflow_recompile_count());
4177 return true;
4178 } else if (trap_count(reason) != 0
4179 && decompile_count() >= m_cutoff) {
4180 // Too many recompiles globally, and we have seen this sort of trap.
4181 // Use cumulative decompile_count, not just md->decompile_count.
4182 if (log())
4183 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4184 Deoptimization::trap_reason_name(reason),
4185 md->trap_count(reason), trap_count(reason),
4186 md->decompile_count(), decompile_count());
4187 return true;
4188 } else {
4189 // The coast is clear.
4190 return false;
4191 }
4192 }
4193
4194 // Compute when not to trap. Used by matching trap based nodes and
4195 // NullCheck optimization.
4196 void Compile::set_allowed_deopt_reasons() {
4197 _allowed_reasons = 0;
4198 if (is_method_compilation()) {
4199 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4200 assert(rs < BitsPerInt, "recode bit map");
4201 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4202 _allowed_reasons |= nth_bit(rs);
4203 }
4204 }
4205 }
4206 }
4207
4208 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4209 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4210 }
4211
4212 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4213 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4214 }
4215
4216 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4217 if (holder->is_initialized() && !do_clinit_barriers()) {
4218 return false;
4219 }
4220 if (holder->is_being_initialized() || do_clinit_barriers()) {
4221 if (accessing_method->holder() == holder) {
4222 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4223 // <init>, or a static method. In all those cases, there was an initialization
4224 // barrier on the holder klass passed.
4225 if (accessing_method->is_static_initializer() ||
4226 accessing_method->is_object_initializer() ||
4227 accessing_method->is_static()) {
4228 return false;
4229 }
4230 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4231 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4232 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4233 // child class can become fully initialized while its parent class is still being initialized.
4234 if (accessing_method->is_static_initializer()) {
4235 return false;
4236 }
4237 }
4238 ciMethod* root = method(); // the root method of compilation
4239 if (root != accessing_method) {
4240 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4241 }
4242 }
4243 return true;
4244 }
4245
4246 #ifndef PRODUCT
4247 //------------------------------verify_bidirectional_edges---------------------
4248 // For each input edge to a node (ie - for each Use-Def edge), verify that
4249 // there is a corresponding Def-Use edge.
4250 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4251 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4252 uint stack_size = live_nodes() >> 4;
4253 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4254 nstack.push(_root);
4255
4256 while (nstack.size() > 0) {
4257 Node* n = nstack.pop();
4258 if (visited.member(n)) {
4259 continue;
4260 }
4261 visited.push(n);
4262
4263 // Walk over all input edges, checking for correspondence
4264 uint length = n->len();
4265 for (uint i = 0; i < length; i++) {
4266 Node* in = n->in(i);
4267 if (in != nullptr && !visited.member(in)) {
4268 nstack.push(in); // Put it on stack
4269 }
4270 if (in != nullptr && !in->is_top()) {
4271 // Count instances of `next`
4272 int cnt = 0;
4273 for (uint idx = 0; idx < in->_outcnt; idx++) {
4274 if (in->_out[idx] == n) {
4275 cnt++;
4276 }
4277 }
4278 assert(cnt > 0, "Failed to find Def-Use edge.");
4279 // Check for duplicate edges
4280 // walk the input array downcounting the input edges to n
4281 for (uint j = 0; j < length; j++) {
4282 if (n->in(j) == in) {
4283 cnt--;
4284 }
4285 }
4286 assert(cnt == 0, "Mismatched edge count.");
4287 } else if (in == nullptr) {
4288 assert(i == 0 || i >= n->req() ||
4289 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4290 (n->is_Unlock() && i == (n->req() - 1)) ||
4291 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4292 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4293 } else {
4294 assert(in->is_top(), "sanity");
4295 // Nothing to check.
4296 }
4297 }
4298 }
4299 }
4300
4301 //------------------------------verify_graph_edges---------------------------
4302 // Walk the Graph and verify that there is a one-to-one correspondence
4303 // between Use-Def edges and Def-Use edges in the graph.
4304 void Compile::verify_graph_edges(bool no_dead_code) {
4305 if (VerifyGraphEdges) {
4306 Unique_Node_List visited;
4307
4308 // Call graph walk to check edges
4309 verify_bidirectional_edges(visited);
4310 if (no_dead_code) {
4311 // Now make sure that no visited node is used by an unvisited node.
4312 bool dead_nodes = false;
4313 Unique_Node_List checked;
4314 while (visited.size() > 0) {
4315 Node* n = visited.pop();
4316 checked.push(n);
4317 for (uint i = 0; i < n->outcnt(); i++) {
4318 Node* use = n->raw_out(i);
4319 if (checked.member(use)) continue; // already checked
4320 if (visited.member(use)) continue; // already in the graph
4321 if (use->is_Con()) continue; // a dead ConNode is OK
4322 // At this point, we have found a dead node which is DU-reachable.
4323 if (!dead_nodes) {
4324 tty->print_cr("*** Dead nodes reachable via DU edges:");
4325 dead_nodes = true;
4326 }
4327 use->dump(2);
4328 tty->print_cr("---");
4329 checked.push(use); // No repeats; pretend it is now checked.
4330 }
4331 }
4332 assert(!dead_nodes, "using nodes must be reachable from root");
4333 }
4334 }
4335 }
4336 #endif
4337
4338 // The Compile object keeps track of failure reasons separately from the ciEnv.
4339 // This is required because there is not quite a 1-1 relation between the
4340 // ciEnv and its compilation task and the Compile object. Note that one
4341 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4342 // to backtrack and retry without subsuming loads. Other than this backtracking
4343 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4344 // by the logic in C2Compiler.
4345 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4346 if (log() != nullptr) {
4347 log()->elem("failure reason='%s' phase='compile'", reason);
4348 }
4349 if (_failure_reason.get() == nullptr) {
4350 // Record the first failure reason.
4351 _failure_reason.set(reason);
4352 if (CaptureBailoutInformation) {
4353 _first_failure_details = new CompilationFailureInfo(reason);
4354 }
4355 } else {
4356 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4357 }
4358
4359 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4360 C->print_method(PHASE_FAILURE, 1);
4361 }
4362 _root = nullptr; // flush the graph, too
4363 }
4364
4365 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4366 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4367 _compile(Compile::current()),
4368 _log(nullptr),
4369 _phase_name(name),
4370 _dolog(CITimeVerbose)
4371 {
4372 assert(_compile != nullptr, "sanity check");
4373 if (_dolog) {
4374 _log = _compile->log();
4375 }
4376 if (_log != nullptr) {
4377 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4378 _log->stamp();
4379 _log->end_head();
4380 }
4381 }
4382
4383 Compile::TracePhase::~TracePhase() {
4384 if (_compile->failing_internal()) {
4385 return; // timing code, not stressing bailouts.
4386 }
4387 #ifdef ASSERT
4388 if (PrintIdealNodeCount) {
4389 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4390 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4391 }
4392
4393 if (VerifyIdealNodeCount) {
4394 _compile->print_missing_nodes();
4395 }
4396 #endif
4397
4398 if (_log != nullptr) {
4399 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4400 }
4401 }
4402
4403 //----------------------------static_subtype_check-----------------------------
4404 // Shortcut important common cases when superklass is exact:
4405 // (0) superklass is java.lang.Object (can occur in reflective code)
4406 // (1) subklass is already limited to a subtype of superklass => always ok
4407 // (2) subklass does not overlap with superklass => always fail
4408 // (3) superklass has NO subtypes and we can check with a simple compare.
4409 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4410 if (skip) {
4411 return SSC_full_test; // Let caller generate the general case.
4412 }
4413
4414 if (subk->is_java_subtype_of(superk)) {
4415 return SSC_always_true; // (0) and (1) this test cannot fail
4416 }
4417
4418 if (!subk->maybe_java_subtype_of(superk)) {
4419 return SSC_always_false; // (2) true path dead; no dynamic test needed
4420 }
4421
4422 const Type* superelem = superk;
4423 if (superk->isa_aryklassptr()) {
4424 int ignored;
4425 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4426 }
4427
4428 if (superelem->isa_instklassptr()) {
4429 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4430 if (!ik->has_subklass()) {
4431 if (!ik->is_final()) {
4432 // Add a dependency if there is a chance of a later subclass.
4433 dependencies()->assert_leaf_type(ik);
4434 }
4435 if (!superk->maybe_java_subtype_of(subk)) {
4436 return SSC_always_false;
4437 }
4438 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4439 }
4440 } else {
4441 // A primitive array type has no subtypes.
4442 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4443 }
4444
4445 return SSC_full_test;
4446 }
4447
4448 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4449 #ifdef _LP64
4450 // The scaled index operand to AddP must be a clean 64-bit value.
4451 // Java allows a 32-bit int to be incremented to a negative
4452 // value, which appears in a 64-bit register as a large
4453 // positive number. Using that large positive number as an
4454 // operand in pointer arithmetic has bad consequences.
4455 // On the other hand, 32-bit overflow is rare, and the possibility
4456 // can often be excluded, if we annotate the ConvI2L node with
4457 // a type assertion that its value is known to be a small positive
4458 // number. (The prior range check has ensured this.)
4459 // This assertion is used by ConvI2LNode::Ideal.
4460 int index_max = max_jint - 1; // array size is max_jint, index is one less
4461 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4462 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4463 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4464 #endif
4465 return idx;
4466 }
4467
4468 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4469 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4470 if (ctrl != nullptr) {
4471 // Express control dependency by a CastII node with a narrow type.
4472 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4473 // node from floating above the range check during loop optimizations. Otherwise, the
4474 // ConvI2L node may be eliminated independently of the range check, causing the data path
4475 // to become TOP while the control path is still there (although it's unreachable).
4476 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4477 value = phase->transform(value);
4478 }
4479 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4480 return phase->transform(new ConvI2LNode(value, ltype));
4481 }
4482
4483 // The message about the current inlining is accumulated in
4484 // _print_inlining_stream and transferred into the _print_inlining_list
4485 // once we know whether inlining succeeds or not. For regular
4486 // inlining, messages are appended to the buffer pointed by
4487 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4488 // a new buffer is added after _print_inlining_idx in the list. This
4489 // way we can update the inlining message for late inlining call site
4490 // when the inlining is attempted again.
4491 void Compile::print_inlining_init() {
4492 if (print_inlining() || print_intrinsics()) {
4493 // print_inlining_init is actually called several times.
4494 print_inlining_reset();
4495 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4496 }
4497 }
4498
4499 void Compile::print_inlining_reinit() {
4500 if (print_inlining() || print_intrinsics()) {
4501 print_inlining_reset();
4502 }
4503 }
4504
4505 void Compile::print_inlining_reset() {
4506 _print_inlining_stream->reset();
4507 }
4508
4509 void Compile::print_inlining_commit() {
4510 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4511 // Transfer the message from _print_inlining_stream to the current
4512 // _print_inlining_list buffer and clear _print_inlining_stream.
4513 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4514 print_inlining_reset();
4515 }
4516
4517 void Compile::print_inlining_push() {
4518 // Add new buffer to the _print_inlining_list at current position
4519 _print_inlining_idx++;
4520 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4521 }
4522
4523 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4524 return _print_inlining_list->at(_print_inlining_idx);
4525 }
4526
4527 void Compile::print_inlining_update(CallGenerator* cg) {
4528 if (print_inlining() || print_intrinsics()) {
4529 if (cg->is_late_inline()) {
4530 if (print_inlining_current()->cg() != cg &&
4531 (print_inlining_current()->cg() != nullptr ||
4532 print_inlining_current()->ss()->size() != 0)) {
4533 print_inlining_push();
4534 }
4535 print_inlining_commit();
4536 print_inlining_current()->set_cg(cg);
4537 } else {
4538 if (print_inlining_current()->cg() != nullptr) {
4539 print_inlining_push();
4540 }
4541 print_inlining_commit();
4542 }
4543 }
4544 }
4545
4546 void Compile::print_inlining_move_to(CallGenerator* cg) {
4547 // We resume inlining at a late inlining call site. Locate the
4548 // corresponding inlining buffer so that we can update it.
4549 if (print_inlining() || print_intrinsics()) {
4550 for (int i = 0; i < _print_inlining_list->length(); i++) {
4551 if (_print_inlining_list->at(i)->cg() == cg) {
4552 _print_inlining_idx = i;
4553 return;
4554 }
4555 }
4556 ShouldNotReachHere();
4557 }
4558 }
4559
4560 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4561 if (print_inlining() || print_intrinsics()) {
4562 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4563 assert(print_inlining_current()->cg() == cg, "wrong entry");
4564 // replace message with new message
4565 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4566 print_inlining_commit();
4567 print_inlining_current()->set_cg(cg);
4568 }
4569 }
4570
4571 void Compile::print_inlining_assert_ready() {
4572 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4573 }
4574
4575 void Compile::process_print_inlining() {
4576 assert(_late_inlines.length() == 0, "not drained yet");
4577 if (print_inlining() || print_intrinsics()) {
4578 ResourceMark rm;
4579 stringStream ss;
4580 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4581 for (int i = 0; i < _print_inlining_list->length(); i++) {
4582 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4583 ss.print("%s", pib->ss()->freeze());
4584 delete pib;
4585 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4586 }
4587 // Reset _print_inlining_list, it only contains destructed objects.
4588 // It is on the arena, so it will be freed when the arena is reset.
4589 _print_inlining_list = nullptr;
4590 // _print_inlining_stream won't be used anymore, either.
4591 print_inlining_reset();
4592 size_t end = ss.size();
4593 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4594 strncpy(_print_inlining_output, ss.freeze(), end+1);
4595 _print_inlining_output[end] = 0;
4596 }
4597 }
4598
4599 void Compile::dump_print_inlining() {
4600 if (_print_inlining_output != nullptr) {
4601 tty->print_raw(_print_inlining_output);
4602 }
4603 }
4604
4605 void Compile::log_late_inline(CallGenerator* cg) {
4606 if (log() != nullptr) {
4607 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4608 cg->unique_id());
4609 JVMState* p = cg->call_node()->jvms();
4610 while (p != nullptr) {
4611 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4612 p = p->caller();
4613 }
4614 log()->tail("late_inline");
4615 }
4616 }
4617
4618 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4619 log_late_inline(cg);
4620 if (log() != nullptr) {
4621 log()->inline_fail(msg);
4622 }
4623 }
4624
4625 void Compile::log_inline_id(CallGenerator* cg) {
4626 if (log() != nullptr) {
4627 // The LogCompilation tool needs a unique way to identify late
4628 // inline call sites. This id must be unique for this call site in
4629 // this compilation. Try to have it unique across compilations as
4630 // well because it can be convenient when grepping through the log
4631 // file.
4632 // Distinguish OSR compilations from others in case CICountOSR is
4633 // on.
4634 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4635 cg->set_unique_id(id);
4636 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4637 }
4638 }
4639
4640 void Compile::log_inline_failure(const char* msg) {
4641 if (C->log() != nullptr) {
4642 C->log()->inline_fail(msg);
4643 }
4644 }
4645
4646
4647 // Dump inlining replay data to the stream.
4648 // Don't change thread state and acquire any locks.
4649 void Compile::dump_inline_data(outputStream* out) {
4650 InlineTree* inl_tree = ilt();
4651 if (inl_tree != nullptr) {
4652 out->print(" inline %d", inl_tree->count());
4653 inl_tree->dump_replay_data(out);
4654 }
4655 }
4656
4657 void Compile::dump_inline_data_reduced(outputStream* out) {
4658 assert(ReplayReduce, "");
4659
4660 InlineTree* inl_tree = ilt();
4661 if (inl_tree == nullptr) {
4662 return;
4663 }
4664 // Enable iterative replay file reduction
4665 // Output "compile" lines for depth 1 subtrees,
4666 // simulating that those trees were compiled
4667 // instead of inlined.
4668 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4669 InlineTree* sub = inl_tree->subtrees().at(i);
4670 if (sub->inline_level() != 1) {
4671 continue;
4672 }
4673
4674 ciMethod* method = sub->method();
4675 int entry_bci = -1;
4676 int comp_level = env()->task()->comp_level();
4677 out->print("compile ");
4678 method->dump_name_as_ascii(out);
4679 out->print(" %d %d", entry_bci, comp_level);
4680 out->print(" inline %d", sub->count());
4681 sub->dump_replay_data(out, -1);
4682 out->cr();
4683 }
4684 }
4685
4686 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4687 if (n1->Opcode() < n2->Opcode()) return -1;
4688 else if (n1->Opcode() > n2->Opcode()) return 1;
4689
4690 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4691 for (uint i = 1; i < n1->req(); i++) {
4692 if (n1->in(i) < n2->in(i)) return -1;
4693 else if (n1->in(i) > n2->in(i)) return 1;
4694 }
4695
4696 return 0;
4697 }
4698
4699 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4700 Node* n1 = *n1p;
4701 Node* n2 = *n2p;
4702
4703 return cmp_expensive_nodes(n1, n2);
4704 }
4705
4706 void Compile::sort_expensive_nodes() {
4707 if (!expensive_nodes_sorted()) {
4708 _expensive_nodes.sort(cmp_expensive_nodes);
4709 }
4710 }
4711
4712 bool Compile::expensive_nodes_sorted() const {
4713 for (int i = 1; i < _expensive_nodes.length(); i++) {
4714 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4715 return false;
4716 }
4717 }
4718 return true;
4719 }
4720
4721 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4722 if (_expensive_nodes.length() == 0) {
4723 return false;
4724 }
4725
4726 assert(OptimizeExpensiveOps, "optimization off?");
4727
4728 // Take this opportunity to remove dead nodes from the list
4729 int j = 0;
4730 for (int i = 0; i < _expensive_nodes.length(); i++) {
4731 Node* n = _expensive_nodes.at(i);
4732 if (!n->is_unreachable(igvn)) {
4733 assert(n->is_expensive(), "should be expensive");
4734 _expensive_nodes.at_put(j, n);
4735 j++;
4736 }
4737 }
4738 _expensive_nodes.trunc_to(j);
4739
4740 // Then sort the list so that similar nodes are next to each other
4741 // and check for at least two nodes of identical kind with same data
4742 // inputs.
4743 sort_expensive_nodes();
4744
4745 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4746 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4747 return true;
4748 }
4749 }
4750
4751 return false;
4752 }
4753
4754 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4755 if (_expensive_nodes.length() == 0) {
4756 return;
4757 }
4758
4759 assert(OptimizeExpensiveOps, "optimization off?");
4760
4761 // Sort to bring similar nodes next to each other and clear the
4762 // control input of nodes for which there's only a single copy.
4763 sort_expensive_nodes();
4764
4765 int j = 0;
4766 int identical = 0;
4767 int i = 0;
4768 bool modified = false;
4769 for (; i < _expensive_nodes.length()-1; i++) {
4770 assert(j <= i, "can't write beyond current index");
4771 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4772 identical++;
4773 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4774 continue;
4775 }
4776 if (identical > 0) {
4777 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4778 identical = 0;
4779 } else {
4780 Node* n = _expensive_nodes.at(i);
4781 igvn.replace_input_of(n, 0, nullptr);
4782 igvn.hash_insert(n);
4783 modified = true;
4784 }
4785 }
4786 if (identical > 0) {
4787 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4788 } else if (_expensive_nodes.length() >= 1) {
4789 Node* n = _expensive_nodes.at(i);
4790 igvn.replace_input_of(n, 0, nullptr);
4791 igvn.hash_insert(n);
4792 modified = true;
4793 }
4794 _expensive_nodes.trunc_to(j);
4795 if (modified) {
4796 igvn.optimize();
4797 }
4798 }
4799
4800 void Compile::add_expensive_node(Node * n) {
4801 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4802 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4803 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4804 if (OptimizeExpensiveOps) {
4805 _expensive_nodes.append(n);
4806 } else {
4807 // Clear control input and let IGVN optimize expensive nodes if
4808 // OptimizeExpensiveOps is off.
4809 n->set_req(0, nullptr);
4810 }
4811 }
4812
4813 /**
4814 * Track coarsened Lock and Unlock nodes.
4815 */
4816
4817 class Lock_List : public Node_List {
4818 uint _origin_cnt;
4819 public:
4820 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4821 uint origin_cnt() const { return _origin_cnt; }
4822 };
4823
4824 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4825 int length = locks.length();
4826 if (length > 0) {
4827 // Have to keep this list until locks elimination during Macro nodes elimination.
4828 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4829 AbstractLockNode* alock = locks.at(0);
4830 BoxLockNode* box = alock->box_node()->as_BoxLock();
4831 for (int i = 0; i < length; i++) {
4832 AbstractLockNode* lock = locks.at(i);
4833 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4834 locks_list->push(lock);
4835 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4836 if (this_box != box) {
4837 // Locking regions (BoxLock) could be Unbalanced here:
4838 // - its coarsened locks were eliminated in earlier
4839 // macro nodes elimination followed by loop unroll
4840 // - it is OSR locking region (no Lock node)
4841 // Preserve Unbalanced status in such cases.
4842 if (!this_box->is_unbalanced()) {
4843 this_box->set_coarsened();
4844 }
4845 if (!box->is_unbalanced()) {
4846 box->set_coarsened();
4847 }
4848 }
4849 }
4850 _coarsened_locks.append(locks_list);
4851 }
4852 }
4853
4854 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4855 int count = coarsened_count();
4856 for (int i = 0; i < count; i++) {
4857 Node_List* locks_list = _coarsened_locks.at(i);
4858 for (uint j = 0; j < locks_list->size(); j++) {
4859 Node* lock = locks_list->at(j);
4860 assert(lock->is_AbstractLock(), "sanity");
4861 if (!useful.member(lock)) {
4862 locks_list->yank(lock);
4863 }
4864 }
4865 }
4866 }
4867
4868 void Compile::remove_coarsened_lock(Node* n) {
4869 if (n->is_AbstractLock()) {
4870 int count = coarsened_count();
4871 for (int i = 0; i < count; i++) {
4872 Node_List* locks_list = _coarsened_locks.at(i);
4873 locks_list->yank(n);
4874 }
4875 }
4876 }
4877
4878 bool Compile::coarsened_locks_consistent() {
4879 int count = coarsened_count();
4880 for (int i = 0; i < count; i++) {
4881 bool unbalanced = false;
4882 bool modified = false; // track locks kind modifications
4883 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4884 uint size = locks_list->size();
4885 if (size == 0) {
4886 unbalanced = false; // All locks were eliminated - good
4887 } else if (size != locks_list->origin_cnt()) {
4888 unbalanced = true; // Some locks were removed from list
4889 } else {
4890 for (uint j = 0; j < size; j++) {
4891 Node* lock = locks_list->at(j);
4892 // All nodes in group should have the same state (modified or not)
4893 if (!lock->as_AbstractLock()->is_coarsened()) {
4894 if (j == 0) {
4895 // first on list was modified, the rest should be too for consistency
4896 modified = true;
4897 } else if (!modified) {
4898 // this lock was modified but previous locks on the list were not
4899 unbalanced = true;
4900 break;
4901 }
4902 } else if (modified) {
4903 // previous locks on list were modified but not this lock
4904 unbalanced = true;
4905 break;
4906 }
4907 }
4908 }
4909 if (unbalanced) {
4910 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4911 #ifdef ASSERT
4912 if (PrintEliminateLocks) {
4913 tty->print_cr("=== unbalanced coarsened locks ===");
4914 for (uint l = 0; l < size; l++) {
4915 locks_list->at(l)->dump();
4916 }
4917 }
4918 #endif
4919 record_failure(C2Compiler::retry_no_locks_coarsening());
4920 return false;
4921 }
4922 }
4923 return true;
4924 }
4925
4926 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4927 // locks coarsening optimization removed Lock/Unlock nodes from them.
4928 // Such regions become unbalanced because coarsening only removes part
4929 // of Lock/Unlock nodes in region. As result we can't execute other
4930 // locks elimination optimizations which assume all code paths have
4931 // corresponding pair of Lock/Unlock nodes - they are balanced.
4932 void Compile::mark_unbalanced_boxes() const {
4933 int count = coarsened_count();
4934 for (int i = 0; i < count; i++) {
4935 Node_List* locks_list = _coarsened_locks.at(i);
4936 uint size = locks_list->size();
4937 if (size > 0) {
4938 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4939 BoxLockNode* box = alock->box_node()->as_BoxLock();
4940 if (alock->is_coarsened()) {
4941 // coarsened_locks_consistent(), which is called before this method, verifies
4942 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4943 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4944 for (uint j = 1; j < size; j++) {
4945 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4946 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4947 if (box != this_box) {
4948 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4949 box->set_unbalanced();
4950 this_box->set_unbalanced();
4951 }
4952 }
4953 }
4954 }
4955 }
4956 }
4957
4958 /**
4959 * Remove the speculative part of types and clean up the graph
4960 */
4961 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4962 if (UseTypeSpeculation) {
4963 Unique_Node_List worklist;
4964 worklist.push(root());
4965 int modified = 0;
4966 // Go over all type nodes that carry a speculative type, drop the
4967 // speculative part of the type and enqueue the node for an igvn
4968 // which may optimize it out.
4969 for (uint next = 0; next < worklist.size(); ++next) {
4970 Node *n = worklist.at(next);
4971 if (n->is_Type()) {
4972 TypeNode* tn = n->as_Type();
4973 const Type* t = tn->type();
4974 const Type* t_no_spec = t->remove_speculative();
4975 if (t_no_spec != t) {
4976 bool in_hash = igvn.hash_delete(n);
4977 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4978 tn->set_type(t_no_spec);
4979 igvn.hash_insert(n);
4980 igvn._worklist.push(n); // give it a chance to go away
4981 modified++;
4982 }
4983 }
4984 // Iterate over outs - endless loops is unreachable from below
4985 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4986 Node *m = n->fast_out(i);
4987 if (not_a_node(m)) {
4988 continue;
4989 }
4990 worklist.push(m);
4991 }
4992 }
4993 // Drop the speculative part of all types in the igvn's type table
4994 igvn.remove_speculative_types();
4995 if (modified > 0) {
4996 igvn.optimize();
4997 if (failing()) return;
4998 }
4999 #ifdef ASSERT
5000 // Verify that after the IGVN is over no speculative type has resurfaced
5001 worklist.clear();
5002 worklist.push(root());
5003 for (uint next = 0; next < worklist.size(); ++next) {
5004 Node *n = worklist.at(next);
5005 const Type* t = igvn.type_or_null(n);
5006 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5007 if (n->is_Type()) {
5008 t = n->as_Type()->type();
5009 assert(t == t->remove_speculative(), "no more speculative types");
5010 }
5011 // Iterate over outs - endless loops is unreachable from below
5012 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5013 Node *m = n->fast_out(i);
5014 if (not_a_node(m)) {
5015 continue;
5016 }
5017 worklist.push(m);
5018 }
5019 }
5020 igvn.check_no_speculative_types();
5021 #endif
5022 }
5023 }
5024
5025 // Auxiliary methods to support randomized stressing/fuzzing.
5026
5027 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5028 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5029 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5030 FLAG_SET_ERGO(StressSeed, _stress_seed);
5031 } else {
5032 _stress_seed = StressSeed;
5033 }
5034 if (_log != nullptr) {
5035 _log->elem("stress_test seed='%u'", _stress_seed);
5036 }
5037 }
5038
5039 int Compile::random() {
5040 _stress_seed = os::next_random(_stress_seed);
5041 return static_cast<int>(_stress_seed);
5042 }
5043
5044 // This method can be called the arbitrary number of times, with current count
5045 // as the argument. The logic allows selecting a single candidate from the
5046 // running list of candidates as follows:
5047 // int count = 0;
5048 // Cand* selected = null;
5049 // while(cand = cand->next()) {
5050 // if (randomized_select(++count)) {
5051 // selected = cand;
5052 // }
5053 // }
5054 //
5055 // Including count equalizes the chances any candidate is "selected".
5056 // This is useful when we don't have the complete list of candidates to choose
5057 // from uniformly. In this case, we need to adjust the randomicity of the
5058 // selection, or else we will end up biasing the selection towards the latter
5059 // candidates.
5060 //
5061 // Quick back-envelope calculation shows that for the list of n candidates
5062 // the equal probability for the candidate to persist as "best" can be
5063 // achieved by replacing it with "next" k-th candidate with the probability
5064 // of 1/k. It can be easily shown that by the end of the run, the
5065 // probability for any candidate is converged to 1/n, thus giving the
5066 // uniform distribution among all the candidates.
5067 //
5068 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5069 #define RANDOMIZED_DOMAIN_POW 29
5070 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5071 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5072 bool Compile::randomized_select(int count) {
5073 assert(count > 0, "only positive");
5074 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5075 }
5076
5077 #ifdef ASSERT
5078 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5079 bool Compile::fail_randomly() {
5080 if ((random() % StressBailoutMean) != 0) {
5081 return false;
5082 }
5083 record_failure("StressBailout");
5084 return true;
5085 }
5086
5087 bool Compile::failure_is_artificial() {
5088 assert(failing_internal(), "should be failing");
5089 return C->failure_reason_is("StressBailout");
5090 }
5091 #endif
5092
5093 CloneMap& Compile::clone_map() { return _clone_map; }
5094 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5095
5096 void NodeCloneInfo::dump_on(outputStream* st) const {
5097 st->print(" {%d:%d} ", idx(), gen());
5098 }
5099
5100 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5101 uint64_t val = value(old->_idx);
5102 NodeCloneInfo cio(val);
5103 assert(val != 0, "old node should be in the map");
5104 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5105 insert(nnn->_idx, cin.get());
5106 #ifndef PRODUCT
5107 if (is_debug()) {
5108 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5109 }
5110 #endif
5111 }
5112
5113 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5114 NodeCloneInfo cio(value(old->_idx));
5115 if (cio.get() == 0) {
5116 cio.set(old->_idx, 0);
5117 insert(old->_idx, cio.get());
5118 #ifndef PRODUCT
5119 if (is_debug()) {
5120 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5121 }
5122 #endif
5123 }
5124 clone(old, nnn, gen);
5125 }
5126
5127 int CloneMap::max_gen() const {
5128 int g = 0;
5129 DictI di(_dict);
5130 for(; di.test(); ++di) {
5131 int t = gen(di._key);
5132 if (g < t) {
5133 g = t;
5134 #ifndef PRODUCT
5135 if (is_debug()) {
5136 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5137 }
5138 #endif
5139 }
5140 }
5141 return g;
5142 }
5143
5144 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5145 uint64_t val = value(key);
5146 if (val != 0) {
5147 NodeCloneInfo ni(val);
5148 ni.dump_on(st);
5149 }
5150 }
5151
5152 void Compile::shuffle_macro_nodes() {
5153 if (_macro_nodes.length() < 2) {
5154 return;
5155 }
5156 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5157 uint j = C->random() % (i + 1);
5158 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5159 }
5160 }
5161
5162 // Move Allocate nodes to the start of the list
5163 void Compile::sort_macro_nodes() {
5164 int count = macro_count();
5165 int allocates = 0;
5166 for (int i = 0; i < count; i++) {
5167 Node* n = macro_node(i);
5168 if (n->is_Allocate()) {
5169 if (i != allocates) {
5170 Node* tmp = macro_node(allocates);
5171 _macro_nodes.at_put(allocates, n);
5172 _macro_nodes.at_put(i, tmp);
5173 }
5174 allocates++;
5175 }
5176 }
5177 }
5178
5179 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5180 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5181 EventCompilerPhase event(UNTIMED);
5182 if (event.should_commit()) {
5183 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5184 }
5185 #ifndef PRODUCT
5186 ResourceMark rm;
5187 stringStream ss;
5188 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5189 int iter = ++_igv_phase_iter[cpt];
5190 if (iter > 1) {
5191 ss.print(" %d", iter);
5192 }
5193 if (n != nullptr) {
5194 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5195 if (n->is_Call()) {
5196 CallNode* call = n->as_Call();
5197 if (call->_name != nullptr) {
5198 // E.g. uncommon traps etc.
5199 ss.print(" - %s", call->_name);
5200 } else if (call->is_CallJava()) {
5201 CallJavaNode* call_java = call->as_CallJava();
5202 if (call_java->method() != nullptr) {
5203 ss.print(" -");
5204 call_java->method()->print_short_name(&ss);
5205 }
5206 }
5207 }
5208 }
5209
5210 const char* name = ss.as_string();
5211 if (should_print_igv(level)) {
5212 _igv_printer->print_method(name, level);
5213 }
5214 if (should_print_phase(cpt)) {
5215 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5216 }
5217 #endif
5218 C->_latest_stage_start_counter.stamp();
5219 }
5220
5221 // Only used from CompileWrapper
5222 void Compile::begin_method() {
5223 #ifndef PRODUCT
5224 if (_method != nullptr && should_print_igv(1)) {
5225 _igv_printer->begin_method();
5226 }
5227 #endif
5228 C->_latest_stage_start_counter.stamp();
5229 }
5230
5231 // Only used from CompileWrapper
5232 void Compile::end_method() {
5233 EventCompilerPhase event(UNTIMED);
5234 if (event.should_commit()) {
5235 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5236 }
5237
5238 #ifndef PRODUCT
5239 if (_method != nullptr && should_print_igv(1)) {
5240 _igv_printer->end_method();
5241 }
5242 #endif
5243 }
5244
5245 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5246 #ifndef PRODUCT
5247 if (_directive->should_print_phase(cpt)) {
5248 return true;
5249 }
5250 #endif
5251 return false;
5252 }
5253
5254 bool Compile::should_print_igv(const int level) {
5255 #ifndef PRODUCT
5256 if (PrintIdealGraphLevel < 0) { // disabled by the user
5257 return false;
5258 }
5259
5260 bool need = directive()->IGVPrintLevelOption >= level;
5261 if (need && !_igv_printer) {
5262 _igv_printer = IdealGraphPrinter::printer();
5263 _igv_printer->set_compile(this);
5264 }
5265 return need;
5266 #else
5267 return false;
5268 #endif
5269 }
5270
5271 #ifndef PRODUCT
5272 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5273 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5274
5275 // Called from debugger. Prints method to the default file with the default phase name.
5276 // This works regardless of any Ideal Graph Visualizer flags set or not.
5277 void igv_print() {
5278 Compile::current()->igv_print_method_to_file();
5279 }
5280
5281 // Same as igv_print() above but with a specified phase name.
5282 void igv_print(const char* phase_name) {
5283 Compile::current()->igv_print_method_to_file(phase_name);
5284 }
5285
5286 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5287 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5288 // This works regardless of any Ideal Graph Visualizer flags set or not.
5289 void igv_print(bool network) {
5290 if (network) {
5291 Compile::current()->igv_print_method_to_network();
5292 } else {
5293 Compile::current()->igv_print_method_to_file();
5294 }
5295 }
5296
5297 // Same as igv_print(bool network) above but with a specified phase name.
5298 void igv_print(bool network, const char* phase_name) {
5299 if (network) {
5300 Compile::current()->igv_print_method_to_network(phase_name);
5301 } else {
5302 Compile::current()->igv_print_method_to_file(phase_name);
5303 }
5304 }
5305
5306 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5307 void igv_print_default() {
5308 Compile::current()->print_method(PHASE_DEBUG, 0);
5309 }
5310
5311 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5312 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5313 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5314 void igv_append() {
5315 Compile::current()->igv_print_method_to_file("Debug", true);
5316 }
5317
5318 // Same as igv_append() above but with a specified phase name.
5319 void igv_append(const char* phase_name) {
5320 Compile::current()->igv_print_method_to_file(phase_name, true);
5321 }
5322
5323 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5324 const char* file_name = "custom_debug.xml";
5325 if (_debug_file_printer == nullptr) {
5326 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5327 } else {
5328 _debug_file_printer->update_compiled_method(C->method());
5329 }
5330 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5331 _debug_file_printer->print(phase_name, (Node*)C->root());
5332 }
5333
5334 void Compile::igv_print_method_to_network(const char* phase_name) {
5335 if (_debug_network_printer == nullptr) {
5336 _debug_network_printer = new IdealGraphPrinter(C);
5337 } else {
5338 _debug_network_printer->update_compiled_method(C->method());
5339 }
5340 tty->print_cr("Method printed over network stream to IGV");
5341 _debug_network_printer->print(phase_name, (Node*)C->root());
5342 }
5343 #endif
5344
5345 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5346 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5347 return value;
5348 }
5349 Node* result = nullptr;
5350 if (bt == T_BYTE) {
5351 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5352 result = new RShiftINode(result, phase->intcon(24));
5353 } else if (bt == T_BOOLEAN) {
5354 result = new AndINode(value, phase->intcon(0xFF));
5355 } else if (bt == T_CHAR) {
5356 result = new AndINode(value,phase->intcon(0xFFFF));
5357 } else {
5358 assert(bt == T_SHORT, "unexpected narrow type");
5359 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5360 result = new RShiftINode(result, phase->intcon(16));
5361 }
5362 if (transform_res) {
5363 result = phase->transform(result);
5364 }
5365 return result;
5366 }
5367
5368 void Compile::record_method_not_compilable_oom() {
5369 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5370 }