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 _stress_seed(0),
639 _comp_arena(mtCompiler),
640 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
641 _env(ci_env),
642 _directive(directive),
643 _log(ci_env->log()),
644 _first_failure_details(nullptr),
645 _intrinsics (comp_arena(), 0, 0, nullptr),
646 _macro_nodes (comp_arena(), 8, 0, nullptr),
647 _parse_predicates (comp_arena(), 8, 0, nullptr),
648 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
649 _expensive_nodes (comp_arena(), 8, 0, nullptr),
650 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
651 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
652 _coarsened_locks (comp_arena(), 8, 0, nullptr),
653 _congraph(nullptr),
654 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
655 _unique(0),
656 _dead_node_count(0),
657 _dead_node_list(comp_arena()),
658 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
659 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
660 _node_arena(&_node_arena_one),
661 _mach_constant_base_node(nullptr),
662 _Compile_types(mtCompiler),
663 _initial_gvn(nullptr),
664 _igvn_worklist(nullptr),
665 _types(nullptr),
666 _node_hash(nullptr),
667 _late_inlines(comp_arena(), 2, 0, nullptr),
668 _string_late_inlines(comp_arena(), 2, 0, nullptr),
669 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
670 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
671 _late_inlines_pos(0),
672 _number_of_mh_late_inlines(0),
673 _oom(false),
674 _print_inlining_stream(new (mtCompiler) stringStream()),
675 _print_inlining_list(nullptr),
676 _print_inlining_idx(0),
677 _print_inlining_output(nullptr),
678 _replay_inline_data(nullptr),
679 _java_calls(0),
680 _inner_loops(0),
681 _interpreter_frame_size(0),
682 _output(nullptr)
683 #ifndef PRODUCT
684 , _in_dump_cnt(0)
685 #endif
686 {
687 C = this;
688 CompileWrapper cw(this);
689
690 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
691 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
692
693 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
694 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
695 // We can always print a disassembly, either abstract (hex dump) or
696 // with the help of a suitable hsdis library. Thus, we should not
697 // couple print_assembly and print_opto_assembly controls.
698 // But: always print opto and regular assembly on compile command 'print'.
699 bool print_assembly = directive->PrintAssemblyOption;
700 set_print_assembly(print_opto_assembly || print_assembly);
701 #else
702 set_print_assembly(false); // must initialize.
703 #endif
704
705 #ifndef PRODUCT
706 set_parsed_irreducible_loop(false);
707 #endif
708
709 if (directive->ReplayInlineOption) {
710 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
711 }
712 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
713 set_print_intrinsics(directive->PrintIntrinsicsOption);
714 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
715
716 if (ProfileTraps) {
717 // Make sure the method being compiled gets its own MDO,
718 // so we can at least track the decompile_count().
719 method()->ensure_method_data();
720 }
721
722 Init(/*do_aliasing=*/ true);
723
724 print_compile_messages();
725
726 _ilt = InlineTree::build_inline_tree_root();
727
728 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
729 assert(num_alias_types() >= AliasIdxRaw, "");
730
731 #define MINIMUM_NODE_HASH 1023
732
733 // GVN that will be run immediately on new nodes
734 uint estimated_size = method()->code_size()*4+64;
735 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
736 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
737 _types = new (comp_arena()) Type_Array(comp_arena());
738 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
739 PhaseGVN gvn;
740 set_initial_gvn(&gvn);
741
742 print_inlining_init();
743 { // Scope for timing the parser
744 TracePhase tp("parse", &timers[_t_parser]);
745
746 // Put top into the hash table ASAP.
747 initial_gvn()->transform(top());
748
749 // Set up tf(), start(), and find a CallGenerator.
750 CallGenerator* cg = nullptr;
751 if (is_osr_compilation()) {
752 const TypeTuple *domain = StartOSRNode::osr_domain();
753 const TypeTuple *range = TypeTuple::make_range(method()->signature());
754 init_tf(TypeFunc::make(domain, range));
755 StartNode* s = new StartOSRNode(root(), domain);
756 initial_gvn()->set_type_bottom(s);
757 init_start(s);
758 cg = CallGenerator::for_osr(method(), entry_bci());
759 } else {
760 // Normal case.
761 init_tf(TypeFunc::make(method()));
762 StartNode* s = new StartNode(root(), tf()->domain());
763 initial_gvn()->set_type_bottom(s);
764 init_start(s);
765 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
766 // With java.lang.ref.reference.get() we must go through the
767 // intrinsic - even when get() is the root
768 // method of the compile - so that, if necessary, the value in
769 // the referent field of the reference object gets recorded by
770 // the pre-barrier code.
771 cg = find_intrinsic(method(), false);
772 }
773 if (cg == nullptr) {
774 float past_uses = method()->interpreter_invocation_count();
775 float expected_uses = past_uses;
776 cg = CallGenerator::for_inline(method(), expected_uses);
777 }
778 }
779 if (failing()) return;
780 if (cg == nullptr) {
781 const char* reason = InlineTree::check_can_parse(method());
782 assert(reason != nullptr, "expect reason for parse failure");
783 stringStream ss;
784 ss.print("cannot parse method: %s", reason);
785 record_method_not_compilable(ss.as_string());
786 return;
787 }
788
789 gvn.set_type(root(), root()->bottom_type());
790
791 JVMState* jvms = build_start_state(start(), tf());
792 if ((jvms = cg->generate(jvms)) == nullptr) {
793 assert(failure_reason() != nullptr, "expect reason for parse failure");
794 stringStream ss;
795 ss.print("method parse failed: %s", failure_reason());
796 record_method_not_compilable(ss.as_string());
797 return;
798 }
799 GraphKit kit(jvms);
800
801 if (!kit.stopped()) {
802 // Accept return values, and transfer control we know not where.
803 // This is done by a special, unique ReturnNode bound to root.
804 return_values(kit.jvms());
805 }
806
807 if (kit.has_exceptions()) {
808 // Any exceptions that escape from this call must be rethrown
809 // to whatever caller is dynamically above us on the stack.
810 // This is done by a special, unique RethrowNode bound to root.
811 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
812 }
813
814 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
815
816 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
817 inline_string_calls(true);
818 }
819
820 if (failing()) return;
821
822 // Remove clutter produced by parsing.
823 if (!failing()) {
824 ResourceMark rm;
825 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
826 }
827 }
828
829 // Note: Large methods are capped off in do_one_bytecode().
830 if (failing()) return;
831
832 // After parsing, node notes are no longer automagic.
833 // They must be propagated by register_new_node_with_optimizer(),
834 // clone(), or the like.
835 set_default_node_notes(nullptr);
836
837 #ifndef PRODUCT
838 if (should_print_igv(1)) {
839 _igv_printer->print_inlining();
840 }
841 #endif
842
843 if (failing()) return;
844 NOT_PRODUCT( verify_graph_edges(); )
845
846 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
847 StressIncrementalInlining || StressMacroExpansion) {
848 initialize_stress_seed(directive);
849 }
850
851 // Now optimize
852 Optimize();
853 if (failing()) return;
854 NOT_PRODUCT( verify_graph_edges(); )
855
856 #ifndef PRODUCT
857 if (should_print_ideal()) {
858 print_ideal_ir("print_ideal");
859 }
860 #endif
861
862 #ifdef ASSERT
863 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
864 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
865 #endif
866
867 // Dump compilation data to replay it.
868 if (directive->DumpReplayOption) {
869 env()->dump_replay_data(_compile_id);
870 }
871 if (directive->DumpInlineOption && (ilt() != nullptr)) {
872 env()->dump_inline_data(_compile_id);
873 }
874
875 // Now that we know the size of all the monitors we can add a fixed slot
876 // for the original deopt pc.
877 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
878 set_fixed_slots(next_slot);
879
880 // Compute when to use implicit null checks. Used by matching trap based
881 // nodes and NullCheck optimization.
882 set_allowed_deopt_reasons();
883
884 // Now generate code
885 Code_Gen();
886 }
887
888 //------------------------------Compile----------------------------------------
889 // Compile a runtime stub
890 Compile::Compile( ciEnv* ci_env,
891 TypeFunc_generator generator,
892 address stub_function,
893 const char *stub_name,
894 int is_fancy_jump,
895 bool pass_tls,
896 bool return_pc,
897 DirectiveSet* directive)
898 : Phase(Compiler),
899 _compile_id(0),
900 _options(Options::for_runtime_stub()),
901 _method(nullptr),
902 _entry_bci(InvocationEntryBci),
903 _stub_function(stub_function),
904 _stub_name(stub_name),
905 _stub_entry_point(nullptr),
906 _max_node_limit(MaxNodeLimit),
907 _post_loop_opts_phase(false),
908 _allow_macro_nodes(true),
909 _inlining_progress(false),
910 _inlining_incrementally(false),
911 _has_reserved_stack_access(false),
912 #ifndef PRODUCT
913 _igv_idx(0),
914 _trace_opto_output(directive->TraceOptoOutputOption),
915 #endif
916 _has_method_handle_invokes(false),
917 _clinit_barrier_on_entry(false),
918 _stress_seed(0),
919 _comp_arena(mtCompiler),
920 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
921 _env(ci_env),
922 _directive(directive),
923 _log(ci_env->log()),
924 _first_failure_details(nullptr),
925 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
926 _congraph(nullptr),
927 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
928 _unique(0),
929 _dead_node_count(0),
930 _dead_node_list(comp_arena()),
931 _node_arena_one(mtCompiler),
932 _node_arena_two(mtCompiler),
933 _node_arena(&_node_arena_one),
934 _mach_constant_base_node(nullptr),
935 _Compile_types(mtCompiler),
936 _initial_gvn(nullptr),
937 _igvn_worklist(nullptr),
938 _types(nullptr),
939 _node_hash(nullptr),
940 _number_of_mh_late_inlines(0),
941 _oom(false),
942 _print_inlining_stream(new (mtCompiler) stringStream()),
943 _print_inlining_list(nullptr),
944 _print_inlining_idx(0),
945 _print_inlining_output(nullptr),
946 _replay_inline_data(nullptr),
947 _java_calls(0),
948 _inner_loops(0),
949 _interpreter_frame_size(0),
950 _output(nullptr),
951 #ifndef PRODUCT
952 _in_dump_cnt(0),
953 #endif
954 _allowed_reasons(0) {
955 C = this;
956
957 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
958 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
959
960 #ifndef PRODUCT
961 set_print_assembly(PrintFrameConverterAssembly);
962 set_parsed_irreducible_loop(false);
963 #else
964 set_print_assembly(false); // Must initialize.
965 #endif
966 set_has_irreducible_loop(false); // no loops
967
968 CompileWrapper cw(this);
969 Init(/*do_aliasing=*/ false);
970 init_tf((*generator)());
971
972 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
973 _types = new (comp_arena()) Type_Array(comp_arena());
974 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
975
976 if (StressLCM || StressGCM) {
977 initialize_stress_seed(directive);
978 }
979
980 {
981 PhaseGVN gvn;
982 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
983 gvn.transform(top());
984
985 GraphKit kit;
986 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
987 }
988
989 NOT_PRODUCT( verify_graph_edges(); )
990
991 Code_Gen();
992 }
993
994 Compile::~Compile() {
995 delete _print_inlining_stream;
996 delete _first_failure_details;
997 };
998
999 //------------------------------Init-------------------------------------------
1000 // Prepare for a single compilation
1001 void Compile::Init(bool aliasing) {
1002 _do_aliasing = aliasing;
1003 _unique = 0;
1004 _regalloc = nullptr;
1005
1006 _tf = nullptr; // filled in later
1007 _top = nullptr; // cached later
1008 _matcher = nullptr; // filled in later
1009 _cfg = nullptr; // filled in later
1010
1011 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1012
1013 _node_note_array = nullptr;
1014 _default_node_notes = nullptr;
1015 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1016
1017 _immutable_memory = nullptr; // filled in at first inquiry
1018
1019 #ifdef ASSERT
1020 _phase_optimize_finished = false;
1021 _exception_backedge = false;
1022 _type_verify = nullptr;
1023 #endif
1024
1025 // Globally visible Nodes
1026 // First set TOP to null to give safe behavior during creation of RootNode
1027 set_cached_top_node(nullptr);
1028 set_root(new RootNode());
1029 // Now that you have a Root to point to, create the real TOP
1030 set_cached_top_node( new ConNode(Type::TOP) );
1031 set_recent_alloc(nullptr, nullptr);
1032
1033 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1034 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1035 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1036 env()->set_dependencies(new Dependencies(env()));
1037
1038 _fixed_slots = 0;
1039 set_has_split_ifs(false);
1040 set_has_loops(false); // first approximation
1041 set_has_stringbuilder(false);
1042 set_has_boxed_value(false);
1043 _trap_can_recompile = false; // no traps emitted yet
1044 _major_progress = true; // start out assuming good things will happen
1045 set_has_unsafe_access(false);
1046 set_max_vector_size(0);
1047 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1048 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1049 set_decompile_count(0);
1050
1051 #ifndef PRODUCT
1052 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1053 #endif
1054
1055 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1056 _loop_opts_cnt = LoopOptsCount;
1057 set_do_inlining(Inline);
1058 set_max_inline_size(MaxInlineSize);
1059 set_freq_inline_size(FreqInlineSize);
1060 set_do_scheduling(OptoScheduling);
1061
1062 set_do_vector_loop(false);
1063 set_has_monitors(false);
1064
1065 if (AllowVectorizeOnDemand) {
1066 if (has_method() && _directive->VectorizeOption) {
1067 set_do_vector_loop(true);
1068 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());})
1069 } else if (has_method() && method()->name() != 0 &&
1070 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1071 set_do_vector_loop(true);
1072 }
1073 }
1074 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1075 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());})
1076
1077 _max_node_limit = _directive->MaxNodeLimitOption;
1078
1079 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1080 set_clinit_barrier_on_entry(true);
1081 }
1082 if (debug_info()->recording_non_safepoints()) {
1083 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1084 (comp_arena(), 8, 0, nullptr));
1085 set_default_node_notes(Node_Notes::make(this));
1086 }
1087
1088 const int grow_ats = 16;
1089 _max_alias_types = grow_ats;
1090 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1091 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1092 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1093 {
1094 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1095 }
1096 // Initialize the first few types.
1097 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1098 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1099 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1100 _num_alias_types = AliasIdxRaw+1;
1101 // Zero out the alias type cache.
1102 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1103 // A null adr_type hits in the cache right away. Preload the right answer.
1104 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1105 }
1106
1107 //---------------------------init_start----------------------------------------
1108 // Install the StartNode on this compile object.
1109 void Compile::init_start(StartNode* s) {
1110 if (failing())
1111 return; // already failing
1112 assert(s == start(), "");
1113 }
1114
1115 /**
1116 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1117 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1118 * the ideal graph.
1119 */
1120 StartNode* Compile::start() const {
1121 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1122 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1123 Node* start = root()->fast_out(i);
1124 if (start->is_Start()) {
1125 return start->as_Start();
1126 }
1127 }
1128 fatal("Did not find Start node!");
1129 return nullptr;
1130 }
1131
1132 //-------------------------------immutable_memory-------------------------------------
1133 // Access immutable memory
1134 Node* Compile::immutable_memory() {
1135 if (_immutable_memory != nullptr) {
1136 return _immutable_memory;
1137 }
1138 StartNode* s = start();
1139 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1140 Node *p = s->fast_out(i);
1141 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1142 _immutable_memory = p;
1143 return _immutable_memory;
1144 }
1145 }
1146 ShouldNotReachHere();
1147 return nullptr;
1148 }
1149
1150 //----------------------set_cached_top_node------------------------------------
1151 // Install the cached top node, and make sure Node::is_top works correctly.
1152 void Compile::set_cached_top_node(Node* tn) {
1153 if (tn != nullptr) verify_top(tn);
1154 Node* old_top = _top;
1155 _top = tn;
1156 // Calling Node::setup_is_top allows the nodes the chance to adjust
1157 // their _out arrays.
1158 if (_top != nullptr) _top->setup_is_top();
1159 if (old_top != nullptr) old_top->setup_is_top();
1160 assert(_top == nullptr || top()->is_top(), "");
1161 }
1162
1163 #ifdef ASSERT
1164 uint Compile::count_live_nodes_by_graph_walk() {
1165 Unique_Node_List useful(comp_arena());
1166 // Get useful node list by walking the graph.
1167 identify_useful_nodes(useful);
1168 return useful.size();
1169 }
1170
1171 void Compile::print_missing_nodes() {
1172
1173 // Return if CompileLog is null and PrintIdealNodeCount is false.
1174 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1175 return;
1176 }
1177
1178 // This is an expensive function. It is executed only when the user
1179 // specifies VerifyIdealNodeCount option or otherwise knows the
1180 // additional work that needs to be done to identify reachable nodes
1181 // by walking the flow graph and find the missing ones using
1182 // _dead_node_list.
1183
1184 Unique_Node_List useful(comp_arena());
1185 // Get useful node list by walking the graph.
1186 identify_useful_nodes(useful);
1187
1188 uint l_nodes = C->live_nodes();
1189 uint l_nodes_by_walk = useful.size();
1190
1191 if (l_nodes != l_nodes_by_walk) {
1192 if (_log != nullptr) {
1193 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1194 _log->stamp();
1195 _log->end_head();
1196 }
1197 VectorSet& useful_member_set = useful.member_set();
1198 int last_idx = l_nodes_by_walk;
1199 for (int i = 0; i < last_idx; i++) {
1200 if (useful_member_set.test(i)) {
1201 if (_dead_node_list.test(i)) {
1202 if (_log != nullptr) {
1203 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1204 }
1205 if (PrintIdealNodeCount) {
1206 // Print the log message to tty
1207 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1208 useful.at(i)->dump();
1209 }
1210 }
1211 }
1212 else if (! _dead_node_list.test(i)) {
1213 if (_log != nullptr) {
1214 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1215 }
1216 if (PrintIdealNodeCount) {
1217 // Print the log message to tty
1218 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1219 }
1220 }
1221 }
1222 if (_log != nullptr) {
1223 _log->tail("mismatched_nodes");
1224 }
1225 }
1226 }
1227 void Compile::record_modified_node(Node* n) {
1228 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1229 _modified_nodes->push(n);
1230 }
1231 }
1232
1233 void Compile::remove_modified_node(Node* n) {
1234 if (_modified_nodes != nullptr) {
1235 _modified_nodes->remove(n);
1236 }
1237 }
1238 #endif
1239
1240 #ifndef PRODUCT
1241 void Compile::verify_top(Node* tn) const {
1242 if (tn != nullptr) {
1243 assert(tn->is_Con(), "top node must be a constant");
1244 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1245 assert(tn->in(0) != nullptr, "must have live top node");
1246 }
1247 }
1248 #endif
1249
1250
1251 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1252
1253 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1254 guarantee(arr != nullptr, "");
1255 int num_blocks = arr->length();
1256 if (grow_by < num_blocks) grow_by = num_blocks;
1257 int num_notes = grow_by * _node_notes_block_size;
1258 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1259 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1260 while (num_notes > 0) {
1261 arr->append(notes);
1262 notes += _node_notes_block_size;
1263 num_notes -= _node_notes_block_size;
1264 }
1265 assert(num_notes == 0, "exact multiple, please");
1266 }
1267
1268 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1269 if (source == nullptr || dest == nullptr) return false;
1270
1271 if (dest->is_Con())
1272 return false; // Do not push debug info onto constants.
1273
1274 #ifdef ASSERT
1275 // Leave a bread crumb trail pointing to the original node:
1276 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1277 dest->set_debug_orig(source);
1278 }
1279 #endif
1280
1281 if (node_note_array() == nullptr)
1282 return false; // Not collecting any notes now.
1283
1284 // This is a copy onto a pre-existing node, which may already have notes.
1285 // If both nodes have notes, do not overwrite any pre-existing notes.
1286 Node_Notes* source_notes = node_notes_at(source->_idx);
1287 if (source_notes == nullptr || source_notes->is_clear()) return false;
1288 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1289 if (dest_notes == nullptr || dest_notes->is_clear()) {
1290 return set_node_notes_at(dest->_idx, source_notes);
1291 }
1292
1293 Node_Notes merged_notes = (*source_notes);
1294 // The order of operations here ensures that dest notes will win...
1295 merged_notes.update_from(dest_notes);
1296 return set_node_notes_at(dest->_idx, &merged_notes);
1297 }
1298
1299
1300 //--------------------------allow_range_check_smearing-------------------------
1301 // Gating condition for coalescing similar range checks.
1302 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1303 // single covering check that is at least as strong as any of them.
1304 // If the optimization succeeds, the simplified (strengthened) range check
1305 // will always succeed. If it fails, we will deopt, and then give up
1306 // on the optimization.
1307 bool Compile::allow_range_check_smearing() const {
1308 // If this method has already thrown a range-check,
1309 // assume it was because we already tried range smearing
1310 // and it failed.
1311 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1312 return !already_trapped;
1313 }
1314
1315
1316 //------------------------------flatten_alias_type-----------------------------
1317 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1318 assert(do_aliasing(), "Aliasing should be enabled");
1319 int offset = tj->offset();
1320 TypePtr::PTR ptr = tj->ptr();
1321
1322 // Known instance (scalarizable allocation) alias only with itself.
1323 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1324 tj->is_oopptr()->is_known_instance();
1325
1326 // Process weird unsafe references.
1327 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1328 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1329 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1330 tj = TypeOopPtr::BOTTOM;
1331 ptr = tj->ptr();
1332 offset = tj->offset();
1333 }
1334
1335 // Array pointers need some flattening
1336 const TypeAryPtr* ta = tj->isa_aryptr();
1337 if (ta && ta->is_stable()) {
1338 // Erase stability property for alias analysis.
1339 tj = ta = ta->cast_to_stable(false);
1340 }
1341 if( ta && is_known_inst ) {
1342 if ( offset != Type::OffsetBot &&
1343 offset > arrayOopDesc::length_offset_in_bytes() ) {
1344 offset = Type::OffsetBot; // Flatten constant access into array body only
1345 tj = ta = ta->
1346 remove_speculative()->
1347 cast_to_ptr_type(ptr)->
1348 with_offset(offset);
1349 }
1350 } else if (ta) {
1351 // For arrays indexed by constant indices, we flatten the alias
1352 // space to include all of the array body. Only the header, klass
1353 // and array length can be accessed un-aliased.
1354 if( offset != Type::OffsetBot ) {
1355 if( ta->const_oop() ) { // MethodData* or Method*
1356 offset = Type::OffsetBot; // Flatten constant access into array body
1357 tj = ta = ta->
1358 remove_speculative()->
1359 cast_to_ptr_type(ptr)->
1360 cast_to_exactness(false)->
1361 with_offset(offset);
1362 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1363 // range is OK as-is.
1364 tj = ta = TypeAryPtr::RANGE;
1365 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1366 tj = TypeInstPtr::KLASS; // all klass loads look alike
1367 ta = TypeAryPtr::RANGE; // generic ignored junk
1368 ptr = TypePtr::BotPTR;
1369 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1370 tj = TypeInstPtr::MARK;
1371 ta = TypeAryPtr::RANGE; // generic ignored junk
1372 ptr = TypePtr::BotPTR;
1373 } else { // Random constant offset into array body
1374 offset = Type::OffsetBot; // Flatten constant access into array body
1375 tj = ta = ta->
1376 remove_speculative()->
1377 cast_to_ptr_type(ptr)->
1378 cast_to_exactness(false)->
1379 with_offset(offset);
1380 }
1381 }
1382 // Arrays of fixed size alias with arrays of unknown size.
1383 if (ta->size() != TypeInt::POS) {
1384 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1385 tj = ta = ta->
1386 remove_speculative()->
1387 cast_to_ptr_type(ptr)->
1388 with_ary(tary)->
1389 cast_to_exactness(false);
1390 }
1391 // Arrays of known objects become arrays of unknown objects.
1392 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1393 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1394 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1395 }
1396 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1397 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1398 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1399 }
1400 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1401 // cannot be distinguished by bytecode alone.
1402 if (ta->elem() == TypeInt::BOOL) {
1403 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1404 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1405 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1406 }
1407 // During the 2nd round of IterGVN, NotNull castings are removed.
1408 // Make sure the Bottom and NotNull variants alias the same.
1409 // Also, make sure exact and non-exact variants alias the same.
1410 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1411 tj = ta = ta->
1412 remove_speculative()->
1413 cast_to_ptr_type(TypePtr::BotPTR)->
1414 cast_to_exactness(false)->
1415 with_offset(offset);
1416 }
1417 }
1418
1419 // Oop pointers need some flattening
1420 const TypeInstPtr *to = tj->isa_instptr();
1421 if (to && to != TypeOopPtr::BOTTOM) {
1422 ciInstanceKlass* ik = to->instance_klass();
1423 if( ptr == TypePtr::Constant ) {
1424 if (ik != ciEnv::current()->Class_klass() ||
1425 offset < ik->layout_helper_size_in_bytes()) {
1426 // No constant oop pointers (such as Strings); they alias with
1427 // unknown strings.
1428 assert(!is_known_inst, "not scalarizable allocation");
1429 tj = to = to->
1430 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1431 remove_speculative()->
1432 cast_to_ptr_type(TypePtr::BotPTR)->
1433 cast_to_exactness(false);
1434 }
1435 } else if( is_known_inst ) {
1436 tj = to; // Keep NotNull and klass_is_exact for instance type
1437 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1438 // During the 2nd round of IterGVN, NotNull castings are removed.
1439 // Make sure the Bottom and NotNull variants alias the same.
1440 // Also, make sure exact and non-exact variants alias the same.
1441 tj = to = to->
1442 remove_speculative()->
1443 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1444 cast_to_ptr_type(TypePtr::BotPTR)->
1445 cast_to_exactness(false);
1446 }
1447 if (to->speculative() != nullptr) {
1448 tj = to = to->remove_speculative();
1449 }
1450 // Canonicalize the holder of this field
1451 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1452 // First handle header references such as a LoadKlassNode, even if the
1453 // object's klass is unloaded at compile time (4965979).
1454 if (!is_known_inst) { // Do it only for non-instance types
1455 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1456 }
1457 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1458 // Static fields are in the space above the normal instance
1459 // fields in the java.lang.Class instance.
1460 if (ik != ciEnv::current()->Class_klass()) {
1461 to = nullptr;
1462 tj = TypeOopPtr::BOTTOM;
1463 offset = tj->offset();
1464 }
1465 } else {
1466 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1467 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1468 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1469 if( is_known_inst ) {
1470 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id());
1471 } else {
1472 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset);
1473 }
1474 }
1475 }
1476 }
1477
1478 // Klass pointers to object array klasses need some flattening
1479 const TypeKlassPtr *tk = tj->isa_klassptr();
1480 if( tk ) {
1481 // If we are referencing a field within a Klass, we need
1482 // to assume the worst case of an Object. Both exact and
1483 // inexact types must flatten to the same alias class so
1484 // use NotNull as the PTR.
1485 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1486 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1487 env()->Object_klass(),
1488 offset);
1489 }
1490
1491 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1492 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1493 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1494 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1495 } else {
1496 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1497 }
1498 }
1499
1500 // Check for precise loads from the primary supertype array and force them
1501 // to the supertype cache alias index. Check for generic array loads from
1502 // the primary supertype array and also force them to the supertype cache
1503 // alias index. Since the same load can reach both, we need to merge
1504 // these 2 disparate memories into the same alias class. Since the
1505 // primary supertype array is read-only, there's no chance of confusion
1506 // where we bypass an array load and an array store.
1507 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1508 if (offset == Type::OffsetBot ||
1509 (offset >= primary_supers_offset &&
1510 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1511 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1512 offset = in_bytes(Klass::secondary_super_cache_offset());
1513 tj = tk = tk->with_offset(offset);
1514 }
1515 }
1516
1517 // Flatten all Raw pointers together.
1518 if (tj->base() == Type::RawPtr)
1519 tj = TypeRawPtr::BOTTOM;
1520
1521 if (tj->base() == Type::AnyPtr)
1522 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1523
1524 offset = tj->offset();
1525 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1526
1527 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1528 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1529 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1530 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1531 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1532 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1533 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1534 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1535 assert( tj->ptr() != TypePtr::TopPTR &&
1536 tj->ptr() != TypePtr::AnyNull &&
1537 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1538 // assert( tj->ptr() != TypePtr::Constant ||
1539 // tj->base() == Type::RawPtr ||
1540 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1541
1542 return tj;
1543 }
1544
1545 void Compile::AliasType::Init(int i, const TypePtr* at) {
1546 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1547 _index = i;
1548 _adr_type = at;
1549 _field = nullptr;
1550 _element = nullptr;
1551 _is_rewritable = true; // default
1552 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1553 if (atoop != nullptr && atoop->is_known_instance()) {
1554 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1555 _general_index = Compile::current()->get_alias_index(gt);
1556 } else {
1557 _general_index = 0;
1558 }
1559 }
1560
1561 BasicType Compile::AliasType::basic_type() const {
1562 if (element() != nullptr) {
1563 const Type* element = adr_type()->is_aryptr()->elem();
1564 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1565 } if (field() != nullptr) {
1566 return field()->layout_type();
1567 } else {
1568 return T_ILLEGAL; // unknown
1569 }
1570 }
1571
1572 //---------------------------------print_on------------------------------------
1573 #ifndef PRODUCT
1574 void Compile::AliasType::print_on(outputStream* st) {
1575 if (index() < 10)
1576 st->print("@ <%d> ", index());
1577 else st->print("@ <%d>", index());
1578 st->print(is_rewritable() ? " " : " RO");
1579 int offset = adr_type()->offset();
1580 if (offset == Type::OffsetBot)
1581 st->print(" +any");
1582 else st->print(" +%-3d", offset);
1583 st->print(" in ");
1584 adr_type()->dump_on(st);
1585 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1586 if (field() != nullptr && tjp) {
1587 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1588 tjp->offset() != field()->offset_in_bytes()) {
1589 st->print(" != ");
1590 field()->print();
1591 st->print(" ***");
1592 }
1593 }
1594 }
1595
1596 void print_alias_types() {
1597 Compile* C = Compile::current();
1598 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1599 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1600 C->alias_type(idx)->print_on(tty);
1601 tty->cr();
1602 }
1603 }
1604 #endif
1605
1606
1607 //----------------------------probe_alias_cache--------------------------------
1608 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1609 intptr_t key = (intptr_t) adr_type;
1610 key ^= key >> logAliasCacheSize;
1611 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1612 }
1613
1614
1615 //-----------------------------grow_alias_types--------------------------------
1616 void Compile::grow_alias_types() {
1617 const int old_ats = _max_alias_types; // how many before?
1618 const int new_ats = old_ats; // how many more?
1619 const int grow_ats = old_ats+new_ats; // how many now?
1620 _max_alias_types = grow_ats;
1621 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1622 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1623 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1624 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1625 }
1626
1627
1628 //--------------------------------find_alias_type------------------------------
1629 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1630 if (!do_aliasing()) {
1631 return alias_type(AliasIdxBot);
1632 }
1633
1634 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1635 if (ace->_adr_type == adr_type) {
1636 return alias_type(ace->_index);
1637 }
1638
1639 // Handle special cases.
1640 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1641 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1642
1643 // Do it the slow way.
1644 const TypePtr* flat = flatten_alias_type(adr_type);
1645
1646 #ifdef ASSERT
1647 {
1648 ResourceMark rm;
1649 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1650 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1651 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1652 Type::str(adr_type));
1653 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1654 const TypeOopPtr* foop = flat->is_oopptr();
1655 // Scalarizable allocations have exact klass always.
1656 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1657 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1658 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1659 Type::str(foop), Type::str(xoop));
1660 }
1661 }
1662 #endif
1663
1664 int idx = AliasIdxTop;
1665 for (int i = 0; i < num_alias_types(); i++) {
1666 if (alias_type(i)->adr_type() == flat) {
1667 idx = i;
1668 break;
1669 }
1670 }
1671
1672 if (idx == AliasIdxTop) {
1673 if (no_create) return nullptr;
1674 // Grow the array if necessary.
1675 if (_num_alias_types == _max_alias_types) grow_alias_types();
1676 // Add a new alias type.
1677 idx = _num_alias_types++;
1678 _alias_types[idx]->Init(idx, flat);
1679 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1680 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1681 if (flat->isa_instptr()) {
1682 if (flat->offset() == java_lang_Class::klass_offset()
1683 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1684 alias_type(idx)->set_rewritable(false);
1685 }
1686 if (flat->isa_aryptr()) {
1687 #ifdef ASSERT
1688 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1689 // (T_BYTE has the weakest alignment and size restrictions...)
1690 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1691 #endif
1692 if (flat->offset() == TypePtr::OffsetBot) {
1693 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1694 }
1695 }
1696 if (flat->isa_klassptr()) {
1697 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1698 alias_type(idx)->set_rewritable(false);
1699 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1700 alias_type(idx)->set_rewritable(false);
1701 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1702 alias_type(idx)->set_rewritable(false);
1703 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1704 alias_type(idx)->set_rewritable(false);
1705 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1706 alias_type(idx)->set_rewritable(false);
1707 }
1708 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1709 // but the base pointer type is not distinctive enough to identify
1710 // references into JavaThread.)
1711
1712 // Check for final fields.
1713 const TypeInstPtr* tinst = flat->isa_instptr();
1714 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1715 ciField* field;
1716 if (tinst->const_oop() != nullptr &&
1717 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1718 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1719 // static field
1720 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1721 field = k->get_field_by_offset(tinst->offset(), true);
1722 } else {
1723 ciInstanceKlass *k = tinst->instance_klass();
1724 field = k->get_field_by_offset(tinst->offset(), false);
1725 }
1726 assert(field == nullptr ||
1727 original_field == nullptr ||
1728 (field->holder() == original_field->holder() &&
1729 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1730 field->is_static() == original_field->is_static()), "wrong field?");
1731 // Set field() and is_rewritable() attributes.
1732 if (field != nullptr) alias_type(idx)->set_field(field);
1733 }
1734 }
1735
1736 // Fill the cache for next time.
1737 ace->_adr_type = adr_type;
1738 ace->_index = idx;
1739 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1740
1741 // Might as well try to fill the cache for the flattened version, too.
1742 AliasCacheEntry* face = probe_alias_cache(flat);
1743 if (face->_adr_type == nullptr) {
1744 face->_adr_type = flat;
1745 face->_index = idx;
1746 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1747 }
1748
1749 return alias_type(idx);
1750 }
1751
1752
1753 Compile::AliasType* Compile::alias_type(ciField* field) {
1754 const TypeOopPtr* t;
1755 if (field->is_static())
1756 t = TypeInstPtr::make(field->holder()->java_mirror());
1757 else
1758 t = TypeOopPtr::make_from_klass_raw(field->holder());
1759 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1760 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1761 return atp;
1762 }
1763
1764
1765 //------------------------------have_alias_type--------------------------------
1766 bool Compile::have_alias_type(const TypePtr* adr_type) {
1767 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1768 if (ace->_adr_type == adr_type) {
1769 return true;
1770 }
1771
1772 // Handle special cases.
1773 if (adr_type == nullptr) return true;
1774 if (adr_type == TypePtr::BOTTOM) return true;
1775
1776 return find_alias_type(adr_type, true, nullptr) != nullptr;
1777 }
1778
1779 //-----------------------------must_alias--------------------------------------
1780 // True if all values of the given address type are in the given alias category.
1781 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1782 if (alias_idx == AliasIdxBot) return true; // the universal category
1783 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1784 if (alias_idx == AliasIdxTop) return false; // the empty category
1785 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1786
1787 // the only remaining possible overlap is identity
1788 int adr_idx = get_alias_index(adr_type);
1789 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1790 assert(adr_idx == alias_idx ||
1791 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1792 && adr_type != TypeOopPtr::BOTTOM),
1793 "should not be testing for overlap with an unsafe pointer");
1794 return adr_idx == alias_idx;
1795 }
1796
1797 //------------------------------can_alias--------------------------------------
1798 // True if any values of the given address type are in the given alias category.
1799 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1800 if (alias_idx == AliasIdxTop) return false; // the empty category
1801 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1802 // Known instance doesn't alias with bottom memory
1803 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1804 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1805
1806 // the only remaining possible overlap is identity
1807 int adr_idx = get_alias_index(adr_type);
1808 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1809 return adr_idx == alias_idx;
1810 }
1811
1812 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1813 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1814 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1815 if (parse_predicate_count() == 0) {
1816 return;
1817 }
1818 for (int i = 0; i < parse_predicate_count(); i++) {
1819 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1820 parse_predicate->mark_useless();
1821 igvn._worklist.push(parse_predicate);
1822 }
1823 _parse_predicates.clear();
1824 }
1825
1826 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1827 if (!n->for_post_loop_opts_igvn()) {
1828 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1829 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1830 _for_post_loop_igvn.append(n);
1831 }
1832 }
1833
1834 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1835 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1836 _for_post_loop_igvn.remove(n);
1837 }
1838
1839 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1840 // Verify that all previous optimizations produced a valid graph
1841 // at least to this point, even if no loop optimizations were done.
1842 PhaseIdealLoop::verify(igvn);
1843
1844 C->set_post_loop_opts_phase(); // no more loop opts allowed
1845
1846 assert(!C->major_progress(), "not cleared");
1847
1848 if (_for_post_loop_igvn.length() > 0) {
1849 while (_for_post_loop_igvn.length() > 0) {
1850 Node* n = _for_post_loop_igvn.pop();
1851 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1852 igvn._worklist.push(n);
1853 }
1854 igvn.optimize();
1855 if (failing()) return;
1856 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1857 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1858
1859 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1860 if (C->major_progress()) {
1861 C->clear_major_progress(); // ensure that major progress is now clear
1862 }
1863 }
1864 }
1865
1866 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1867 if (OptimizeUnstableIf) {
1868 _unstable_if_traps.append(trap);
1869 }
1870 }
1871
1872 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1873 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1874 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1875 Node* n = trap->uncommon_trap();
1876 if (!useful.member(n)) {
1877 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1878 }
1879 }
1880 }
1881
1882 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1883 // or fold-compares case. Return true if succeed or not found.
1884 //
1885 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1886 // false and ask fold-compares to yield.
1887 //
1888 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1889 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1890 // when deoptimization does happen.
1891 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1892 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1893 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1894 if (trap->uncommon_trap() == unc) {
1895 if (yield && trap->modified()) {
1896 return false;
1897 }
1898 _unstable_if_traps.delete_at(i);
1899 break;
1900 }
1901 }
1902 return true;
1903 }
1904
1905 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1906 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1907 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1908 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1909 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1910 CallStaticJavaNode* unc = trap->uncommon_trap();
1911 int next_bci = trap->next_bci();
1912 bool modified = trap->modified();
1913
1914 if (next_bci != -1 && !modified) {
1915 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1916 JVMState* jvms = unc->jvms();
1917 ciMethod* method = jvms->method();
1918 ciBytecodeStream iter(method);
1919
1920 iter.force_bci(jvms->bci());
1921 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1922 Bytecodes::Code c = iter.cur_bc();
1923 Node* lhs = nullptr;
1924 Node* rhs = nullptr;
1925 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1926 lhs = unc->peek_operand(0);
1927 rhs = unc->peek_operand(1);
1928 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1929 lhs = unc->peek_operand(0);
1930 }
1931
1932 ResourceMark rm;
1933 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1934 assert(live_locals.is_valid(), "broken liveness info");
1935 int len = (int)live_locals.size();
1936
1937 for (int i = 0; i < len; i++) {
1938 Node* local = unc->local(jvms, i);
1939 // kill local using the liveness of next_bci.
1940 // give up when the local looks like an operand to secure reexecution.
1941 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1942 uint idx = jvms->locoff() + i;
1943 #ifdef ASSERT
1944 if (PrintOpto && Verbose) {
1945 tty->print("[unstable_if] kill local#%d: ", idx);
1946 local->dump();
1947 tty->cr();
1948 }
1949 #endif
1950 igvn.replace_input_of(unc, idx, top());
1951 modified = true;
1952 }
1953 }
1954 }
1955
1956 // keep the mondified trap for late query
1957 if (modified) {
1958 trap->set_modified();
1959 } else {
1960 _unstable_if_traps.delete_at(i);
1961 }
1962 }
1963 igvn.optimize();
1964 }
1965
1966 // StringOpts and late inlining of string methods
1967 void Compile::inline_string_calls(bool parse_time) {
1968 {
1969 // remove useless nodes to make the usage analysis simpler
1970 ResourceMark rm;
1971 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1972 }
1973
1974 {
1975 ResourceMark rm;
1976 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1977 PhaseStringOpts pso(initial_gvn());
1978 print_method(PHASE_AFTER_STRINGOPTS, 3);
1979 }
1980
1981 // now inline anything that we skipped the first time around
1982 if (!parse_time) {
1983 _late_inlines_pos = _late_inlines.length();
1984 }
1985
1986 while (_string_late_inlines.length() > 0) {
1987 CallGenerator* cg = _string_late_inlines.pop();
1988 cg->do_late_inline();
1989 if (failing()) return;
1990 }
1991 _string_late_inlines.trunc_to(0);
1992 }
1993
1994 // Late inlining of boxing methods
1995 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1996 if (_boxing_late_inlines.length() > 0) {
1997 assert(has_boxed_value(), "inconsistent");
1998
1999 set_inlining_incrementally(true);
2000
2001 igvn_worklist()->ensure_empty(); // should be done with igvn
2002
2003 _late_inlines_pos = _late_inlines.length();
2004
2005 while (_boxing_late_inlines.length() > 0) {
2006 CallGenerator* cg = _boxing_late_inlines.pop();
2007 cg->do_late_inline();
2008 if (failing()) return;
2009 }
2010 _boxing_late_inlines.trunc_to(0);
2011
2012 inline_incrementally_cleanup(igvn);
2013
2014 set_inlining_incrementally(false);
2015 }
2016 }
2017
2018 bool Compile::inline_incrementally_one() {
2019 assert(IncrementalInline, "incremental inlining should be on");
2020
2021 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2022
2023 set_inlining_progress(false);
2024 set_do_cleanup(false);
2025
2026 for (int i = 0; i < _late_inlines.length(); i++) {
2027 _late_inlines_pos = i+1;
2028 CallGenerator* cg = _late_inlines.at(i);
2029 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2030 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2031 cg->do_late_inline();
2032 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2033 if (failing()) {
2034 return false;
2035 } else if (inlining_progress()) {
2036 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2037 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2038 break; // process one call site at a time
2039 }
2040 } else {
2041 // Ignore late inline direct calls when inlining is not allowed.
2042 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2043 }
2044 }
2045 // Remove processed elements.
2046 _late_inlines.remove_till(_late_inlines_pos);
2047 _late_inlines_pos = 0;
2048
2049 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2050
2051 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2052
2053 set_inlining_progress(false);
2054 set_do_cleanup(false);
2055
2056 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2057 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2058 }
2059
2060 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2061 {
2062 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2063 ResourceMark rm;
2064 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2065 }
2066 {
2067 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2068 igvn.reset_from_gvn(initial_gvn());
2069 igvn.optimize();
2070 if (failing()) return;
2071 }
2072 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2073 }
2074
2075 // Perform incremental inlining until bound on number of live nodes is reached
2076 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2077 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2078
2079 set_inlining_incrementally(true);
2080 uint low_live_nodes = 0;
2081
2082 while (_late_inlines.length() > 0) {
2083 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2084 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2085 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2086 // PhaseIdealLoop is expensive so we only try it once we are
2087 // out of live nodes and we only try it again if the previous
2088 // helped got the number of nodes down significantly
2089 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2090 if (failing()) return;
2091 low_live_nodes = live_nodes();
2092 _major_progress = true;
2093 }
2094
2095 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2096 bool do_print_inlining = print_inlining() || print_intrinsics();
2097 if (do_print_inlining || log() != nullptr) {
2098 // Print inlining message for candidates that we couldn't inline for lack of space.
2099 for (int i = 0; i < _late_inlines.length(); i++) {
2100 CallGenerator* cg = _late_inlines.at(i);
2101 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2102 if (do_print_inlining) {
2103 cg->print_inlining_late(InliningResult::FAILURE, msg);
2104 }
2105 log_late_inline_failure(cg, msg);
2106 }
2107 }
2108 break; // finish
2109 }
2110 }
2111
2112 igvn_worklist()->ensure_empty(); // should be done with igvn
2113
2114 while (inline_incrementally_one()) {
2115 assert(!failing(), "inconsistent");
2116 }
2117 if (failing()) return;
2118
2119 inline_incrementally_cleanup(igvn);
2120
2121 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2122
2123 if (failing()) return;
2124
2125 if (_late_inlines.length() == 0) {
2126 break; // no more progress
2127 }
2128 }
2129
2130 igvn_worklist()->ensure_empty(); // should be done with igvn
2131
2132 if (_string_late_inlines.length() > 0) {
2133 assert(has_stringbuilder(), "inconsistent");
2134
2135 inline_string_calls(false);
2136
2137 if (failing()) return;
2138
2139 inline_incrementally_cleanup(igvn);
2140 }
2141
2142 set_inlining_incrementally(false);
2143 }
2144
2145 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2146 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2147 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2148 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2149 // as if "inlining_incrementally() == true" were set.
2150 assert(inlining_incrementally() == false, "not allowed");
2151 assert(_modified_nodes == nullptr, "not allowed");
2152 assert(_late_inlines.length() > 0, "sanity");
2153
2154 while (_late_inlines.length() > 0) {
2155 igvn_worklist()->ensure_empty(); // should be done with igvn
2156
2157 while (inline_incrementally_one()) {
2158 assert(!failing(), "inconsistent");
2159 }
2160 if (failing()) return;
2161
2162 inline_incrementally_cleanup(igvn);
2163 }
2164 }
2165
2166 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2167 if (_loop_opts_cnt > 0) {
2168 while (major_progress() && (_loop_opts_cnt > 0)) {
2169 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2170 PhaseIdealLoop::optimize(igvn, mode);
2171 _loop_opts_cnt--;
2172 if (failing()) return false;
2173 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2174 }
2175 }
2176 return true;
2177 }
2178
2179 // Remove edges from "root" to each SafePoint at a backward branch.
2180 // They were inserted during parsing (see add_safepoint()) to make
2181 // infinite loops without calls or exceptions visible to root, i.e.,
2182 // useful.
2183 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2184 Node *r = root();
2185 if (r != nullptr) {
2186 for (uint i = r->req(); i < r->len(); ++i) {
2187 Node *n = r->in(i);
2188 if (n != nullptr && n->is_SafePoint()) {
2189 r->rm_prec(i);
2190 if (n->outcnt() == 0) {
2191 igvn.remove_dead_node(n);
2192 }
2193 --i;
2194 }
2195 }
2196 // Parsing may have added top inputs to the root node (Path
2197 // leading to the Halt node proven dead). Make sure we get a
2198 // chance to clean them up.
2199 igvn._worklist.push(r);
2200 igvn.optimize();
2201 }
2202 }
2203
2204 //------------------------------Optimize---------------------------------------
2205 // Given a graph, optimize it.
2206 void Compile::Optimize() {
2207 TracePhase tp("optimizer", &timers[_t_optimizer]);
2208
2209 #ifndef PRODUCT
2210 if (env()->break_at_compile()) {
2211 BREAKPOINT;
2212 }
2213
2214 #endif
2215
2216 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2217 #ifdef ASSERT
2218 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2219 #endif
2220
2221 ResourceMark rm;
2222
2223 print_inlining_reinit();
2224
2225 NOT_PRODUCT( verify_graph_edges(); )
2226
2227 print_method(PHASE_AFTER_PARSING, 1);
2228
2229 {
2230 // Iterative Global Value Numbering, including ideal transforms
2231 // Initialize IterGVN with types and values from parse-time GVN
2232 PhaseIterGVN igvn(initial_gvn());
2233 #ifdef ASSERT
2234 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2235 #endif
2236 {
2237 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2238 igvn.optimize();
2239 }
2240
2241 if (failing()) return;
2242
2243 print_method(PHASE_ITER_GVN1, 2);
2244
2245 process_for_unstable_if_traps(igvn);
2246
2247 if (failing()) return;
2248
2249 inline_incrementally(igvn);
2250
2251 print_method(PHASE_INCREMENTAL_INLINE, 2);
2252
2253 if (failing()) return;
2254
2255 if (eliminate_boxing()) {
2256 // Inline valueOf() methods now.
2257 inline_boxing_calls(igvn);
2258
2259 if (failing()) return;
2260
2261 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2262 inline_incrementally(igvn);
2263 }
2264
2265 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2266
2267 if (failing()) return;
2268 }
2269
2270 // Remove the speculative part of types and clean up the graph from
2271 // the extra CastPP nodes whose only purpose is to carry them. Do
2272 // that early so that optimizations are not disrupted by the extra
2273 // CastPP nodes.
2274 remove_speculative_types(igvn);
2275
2276 if (failing()) return;
2277
2278 // No more new expensive nodes will be added to the list from here
2279 // so keep only the actual candidates for optimizations.
2280 cleanup_expensive_nodes(igvn);
2281
2282 if (failing()) return;
2283
2284 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2285 if (EnableVectorSupport && has_vbox_nodes()) {
2286 TracePhase tp("", &timers[_t_vector]);
2287 PhaseVector pv(igvn);
2288 pv.optimize_vector_boxes();
2289 if (failing()) return;
2290 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2291 }
2292 assert(!has_vbox_nodes(), "sanity");
2293
2294 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2295 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2296 igvn_worklist()->ensure_empty(); // should be done with igvn
2297 {
2298 ResourceMark rm;
2299 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2300 }
2301 igvn.reset_from_gvn(initial_gvn());
2302 igvn.optimize();
2303 if (failing()) return;
2304 }
2305
2306 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2307 // safepoints
2308 remove_root_to_sfpts_edges(igvn);
2309
2310 if (failing()) return;
2311
2312 // Perform escape analysis
2313 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2314 if (has_loops()) {
2315 // Cleanup graph (remove dead nodes).
2316 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2317 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2318 if (failing()) return;
2319 }
2320 bool progress;
2321 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2322 do {
2323 ConnectionGraph::do_analysis(this, &igvn);
2324
2325 if (failing()) return;
2326
2327 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2328
2329 // Optimize out fields loads from scalar replaceable allocations.
2330 igvn.optimize();
2331 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2332
2333 if (failing()) return;
2334
2335 if (congraph() != nullptr && macro_count() > 0) {
2336 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2337 PhaseMacroExpand mexp(igvn);
2338 mexp.eliminate_macro_nodes();
2339 if (failing()) return;
2340
2341 igvn.set_delay_transform(false);
2342 igvn.optimize();
2343 if (failing()) return;
2344
2345 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2346 }
2347
2348 ConnectionGraph::verify_ram_nodes(this, root());
2349 if (failing()) return;
2350
2351 progress = do_iterative_escape_analysis() &&
2352 (macro_count() < mcount) &&
2353 ConnectionGraph::has_candidates(this);
2354 // Try again if candidates exist and made progress
2355 // by removing some allocations and/or locks.
2356 } while (progress);
2357 }
2358
2359 // Loop transforms on the ideal graph. Range Check Elimination,
2360 // peeling, unrolling, etc.
2361
2362 // Set loop opts counter
2363 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2364 {
2365 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2366 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2367 _loop_opts_cnt--;
2368 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2369 if (failing()) return;
2370 }
2371 // Loop opts pass if partial peeling occurred in previous pass
2372 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2373 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2374 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2375 _loop_opts_cnt--;
2376 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2377 if (failing()) return;
2378 }
2379 // Loop opts pass for loop-unrolling before CCP
2380 if(major_progress() && (_loop_opts_cnt > 0)) {
2381 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2382 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2383 _loop_opts_cnt--;
2384 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2385 }
2386 if (!failing()) {
2387 // Verify that last round of loop opts produced a valid graph
2388 PhaseIdealLoop::verify(igvn);
2389 }
2390 }
2391 if (failing()) return;
2392
2393 // Conditional Constant Propagation;
2394 print_method(PHASE_BEFORE_CCP1, 2);
2395 PhaseCCP ccp( &igvn );
2396 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2397 {
2398 TracePhase tp("ccp", &timers[_t_ccp]);
2399 ccp.do_transform();
2400 }
2401 print_method(PHASE_CCP1, 2);
2402
2403 assert( true, "Break here to ccp.dump_old2new_map()");
2404
2405 // Iterative Global Value Numbering, including ideal transforms
2406 {
2407 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2408 igvn.reset_from_igvn(&ccp);
2409 igvn.optimize();
2410 }
2411 print_method(PHASE_ITER_GVN2, 2);
2412
2413 if (failing()) return;
2414
2415 // Loop transforms on the ideal graph. Range Check Elimination,
2416 // peeling, unrolling, etc.
2417 if (!optimize_loops(igvn, LoopOptsDefault)) {
2418 return;
2419 }
2420
2421 if (failing()) return;
2422
2423 C->clear_major_progress(); // ensure that major progress is now clear
2424
2425 process_for_post_loop_opts_igvn(igvn);
2426
2427 if (failing()) return;
2428
2429 #ifdef ASSERT
2430 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2431 #endif
2432
2433 {
2434 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2435 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2436 PhaseMacroExpand mex(igvn);
2437 if (mex.expand_macro_nodes()) {
2438 assert(failing(), "must bail out w/ explicit message");
2439 return;
2440 }
2441 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2442 }
2443
2444 {
2445 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2446 if (bs->expand_barriers(this, igvn)) {
2447 assert(failing(), "must bail out w/ explicit message");
2448 return;
2449 }
2450 print_method(PHASE_BARRIER_EXPANSION, 2);
2451 }
2452
2453 if (C->max_vector_size() > 0) {
2454 C->optimize_logic_cones(igvn);
2455 igvn.optimize();
2456 if (failing()) return;
2457 }
2458
2459 DEBUG_ONLY( _modified_nodes = nullptr; )
2460
2461 assert(igvn._worklist.size() == 0, "not empty");
2462
2463 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2464
2465 if (_late_inlines.length() > 0) {
2466 // More opportunities to optimize virtual and MH calls.
2467 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2468 process_late_inline_calls_no_inline(igvn);
2469 if (failing()) return;
2470 }
2471 } // (End scope of igvn; run destructor if necessary for asserts.)
2472
2473 check_no_dead_use();
2474
2475 process_print_inlining();
2476
2477 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2478 // to remove hashes to unlock nodes for modifications.
2479 C->node_hash()->clear();
2480
2481 // A method with only infinite loops has no edges entering loops from root
2482 {
2483 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2484 if (final_graph_reshaping()) {
2485 assert(failing(), "must bail out w/ explicit message");
2486 return;
2487 }
2488 }
2489
2490 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2491 DEBUG_ONLY(set_phase_optimize_finished();)
2492 }
2493
2494 #ifdef ASSERT
2495 void Compile::check_no_dead_use() const {
2496 ResourceMark rm;
2497 Unique_Node_List wq;
2498 wq.push(root());
2499 for (uint i = 0; i < wq.size(); ++i) {
2500 Node* n = wq.at(i);
2501 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2502 Node* u = n->fast_out(j);
2503 if (u->outcnt() == 0 && !u->is_Con()) {
2504 u->dump();
2505 fatal("no reachable node should have no use");
2506 }
2507 wq.push(u);
2508 }
2509 }
2510 }
2511 #endif
2512
2513 void Compile::inline_vector_reboxing_calls() {
2514 if (C->_vector_reboxing_late_inlines.length() > 0) {
2515 _late_inlines_pos = C->_late_inlines.length();
2516 while (_vector_reboxing_late_inlines.length() > 0) {
2517 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2518 cg->do_late_inline();
2519 if (failing()) return;
2520 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2521 }
2522 _vector_reboxing_late_inlines.trunc_to(0);
2523 }
2524 }
2525
2526 bool Compile::has_vbox_nodes() {
2527 if (C->_vector_reboxing_late_inlines.length() > 0) {
2528 return true;
2529 }
2530 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2531 Node * n = C->macro_node(macro_idx);
2532 assert(n->is_macro(), "only macro nodes expected here");
2533 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2534 return true;
2535 }
2536 }
2537 return false;
2538 }
2539
2540 //---------------------------- Bitwise operation packing optimization ---------------------------
2541
2542 static bool is_vector_unary_bitwise_op(Node* n) {
2543 return n->Opcode() == Op_XorV &&
2544 VectorNode::is_vector_bitwise_not_pattern(n);
2545 }
2546
2547 static bool is_vector_binary_bitwise_op(Node* n) {
2548 switch (n->Opcode()) {
2549 case Op_AndV:
2550 case Op_OrV:
2551 return true;
2552
2553 case Op_XorV:
2554 return !is_vector_unary_bitwise_op(n);
2555
2556 default:
2557 return false;
2558 }
2559 }
2560
2561 static bool is_vector_ternary_bitwise_op(Node* n) {
2562 return n->Opcode() == Op_MacroLogicV;
2563 }
2564
2565 static bool is_vector_bitwise_op(Node* n) {
2566 return is_vector_unary_bitwise_op(n) ||
2567 is_vector_binary_bitwise_op(n) ||
2568 is_vector_ternary_bitwise_op(n);
2569 }
2570
2571 static bool is_vector_bitwise_cone_root(Node* n) {
2572 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2573 return false;
2574 }
2575 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2576 if (is_vector_bitwise_op(n->fast_out(i))) {
2577 return false;
2578 }
2579 }
2580 return true;
2581 }
2582
2583 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2584 uint cnt = 0;
2585 if (is_vector_bitwise_op(n)) {
2586 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2587 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2588 for (uint i = 1; i < inp_cnt; i++) {
2589 Node* in = n->in(i);
2590 bool skip = VectorNode::is_all_ones_vector(in);
2591 if (!skip && !inputs.member(in)) {
2592 inputs.push(in);
2593 cnt++;
2594 }
2595 }
2596 assert(cnt <= 1, "not unary");
2597 } else {
2598 uint last_req = inp_cnt;
2599 if (is_vector_ternary_bitwise_op(n)) {
2600 last_req = inp_cnt - 1; // skip last input
2601 }
2602 for (uint i = 1; i < last_req; i++) {
2603 Node* def = n->in(i);
2604 if (!inputs.member(def)) {
2605 inputs.push(def);
2606 cnt++;
2607 }
2608 }
2609 }
2610 } else { // not a bitwise operations
2611 if (!inputs.member(n)) {
2612 inputs.push(n);
2613 cnt++;
2614 }
2615 }
2616 return cnt;
2617 }
2618
2619 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2620 Unique_Node_List useful_nodes;
2621 C->identify_useful_nodes(useful_nodes);
2622
2623 for (uint i = 0; i < useful_nodes.size(); i++) {
2624 Node* n = useful_nodes.at(i);
2625 if (is_vector_bitwise_cone_root(n)) {
2626 list.push(n);
2627 }
2628 }
2629 }
2630
2631 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2632 const TypeVect* vt,
2633 Unique_Node_List& partition,
2634 Unique_Node_List& inputs) {
2635 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2636 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2637 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2638
2639 Node* in1 = inputs.at(0);
2640 Node* in2 = inputs.at(1);
2641 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2642
2643 uint func = compute_truth_table(partition, inputs);
2644
2645 Node* pn = partition.at(partition.size() - 1);
2646 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2647 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2648 }
2649
2650 static uint extract_bit(uint func, uint pos) {
2651 return (func & (1 << pos)) >> pos;
2652 }
2653
2654 //
2655 // A macro logic node represents a truth table. It has 4 inputs,
2656 // First three inputs corresponds to 3 columns of a truth table
2657 // and fourth input captures the logic function.
2658 //
2659 // eg. fn = (in1 AND in2) OR in3;
2660 //
2661 // MacroNode(in1,in2,in3,fn)
2662 //
2663 // -----------------
2664 // in1 in2 in3 fn
2665 // -----------------
2666 // 0 0 0 0
2667 // 0 0 1 1
2668 // 0 1 0 0
2669 // 0 1 1 1
2670 // 1 0 0 0
2671 // 1 0 1 1
2672 // 1 1 0 1
2673 // 1 1 1 1
2674 //
2675
2676 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2677 int res = 0;
2678 for (int i = 0; i < 8; i++) {
2679 int bit1 = extract_bit(in1, i);
2680 int bit2 = extract_bit(in2, i);
2681 int bit3 = extract_bit(in3, i);
2682
2683 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2684 int func_bit = extract_bit(func, func_bit_pos);
2685
2686 res |= func_bit << i;
2687 }
2688 return res;
2689 }
2690
2691 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2692 assert(n != nullptr, "");
2693 assert(eval_map.contains(n), "absent");
2694 return *(eval_map.get(n));
2695 }
2696
2697 static void eval_operands(Node* n,
2698 uint& func1, uint& func2, uint& func3,
2699 ResourceHashtable<Node*,uint>& eval_map) {
2700 assert(is_vector_bitwise_op(n), "");
2701
2702 if (is_vector_unary_bitwise_op(n)) {
2703 Node* opnd = n->in(1);
2704 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2705 opnd = n->in(2);
2706 }
2707 func1 = eval_operand(opnd, eval_map);
2708 } else if (is_vector_binary_bitwise_op(n)) {
2709 func1 = eval_operand(n->in(1), eval_map);
2710 func2 = eval_operand(n->in(2), eval_map);
2711 } else {
2712 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2713 func1 = eval_operand(n->in(1), eval_map);
2714 func2 = eval_operand(n->in(2), eval_map);
2715 func3 = eval_operand(n->in(3), eval_map);
2716 }
2717 }
2718
2719 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2720 assert(inputs.size() <= 3, "sanity");
2721 ResourceMark rm;
2722 uint res = 0;
2723 ResourceHashtable<Node*,uint> eval_map;
2724
2725 // Populate precomputed functions for inputs.
2726 // Each input corresponds to one column of 3 input truth-table.
2727 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2728 0xCC, // (_, b, _) -> b
2729 0xF0 }; // (a, _, _) -> a
2730 for (uint i = 0; i < inputs.size(); i++) {
2731 eval_map.put(inputs.at(i), input_funcs[2-i]);
2732 }
2733
2734 for (uint i = 0; i < partition.size(); i++) {
2735 Node* n = partition.at(i);
2736
2737 uint func1 = 0, func2 = 0, func3 = 0;
2738 eval_operands(n, func1, func2, func3, eval_map);
2739
2740 switch (n->Opcode()) {
2741 case Op_OrV:
2742 assert(func3 == 0, "not binary");
2743 res = func1 | func2;
2744 break;
2745 case Op_AndV:
2746 assert(func3 == 0, "not binary");
2747 res = func1 & func2;
2748 break;
2749 case Op_XorV:
2750 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2751 assert(func2 == 0 && func3 == 0, "not unary");
2752 res = (~func1) & 0xFF;
2753 } else {
2754 assert(func3 == 0, "not binary");
2755 res = func1 ^ func2;
2756 }
2757 break;
2758 case Op_MacroLogicV:
2759 // Ordering of inputs may change during evaluation of sub-tree
2760 // containing MacroLogic node as a child node, thus a re-evaluation
2761 // makes sure that function is evaluated in context of current
2762 // inputs.
2763 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2764 break;
2765
2766 default: assert(false, "not supported: %s", n->Name());
2767 }
2768 assert(res <= 0xFF, "invalid");
2769 eval_map.put(n, res);
2770 }
2771 return res;
2772 }
2773
2774 // Criteria under which nodes gets packed into a macro logic node:-
2775 // 1) Parent and both child nodes are all unmasked or masked with
2776 // same predicates.
2777 // 2) Masked parent can be packed with left child if it is predicated
2778 // and both have same predicates.
2779 // 3) Masked parent can be packed with right child if its un-predicated
2780 // or has matching predication condition.
2781 // 4) An unmasked parent can be packed with an unmasked child.
2782 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2783 assert(partition.size() == 0, "not empty");
2784 assert(inputs.size() == 0, "not empty");
2785 if (is_vector_ternary_bitwise_op(n)) {
2786 return false;
2787 }
2788
2789 bool is_unary_op = is_vector_unary_bitwise_op(n);
2790 if (is_unary_op) {
2791 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2792 return false; // too few inputs
2793 }
2794
2795 bool pack_left_child = true;
2796 bool pack_right_child = true;
2797
2798 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2799 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2800
2801 int left_child_input_cnt = 0;
2802 int right_child_input_cnt = 0;
2803
2804 bool parent_is_predicated = n->is_predicated_vector();
2805 bool left_child_predicated = n->in(1)->is_predicated_vector();
2806 bool right_child_predicated = n->in(2)->is_predicated_vector();
2807
2808 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2809 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2810 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2811
2812 do {
2813 if (pack_left_child && left_child_LOP &&
2814 ((!parent_is_predicated && !left_child_predicated) ||
2815 ((parent_is_predicated && left_child_predicated &&
2816 parent_pred == left_child_pred)))) {
2817 partition.push(n->in(1));
2818 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2819 } else {
2820 inputs.push(n->in(1));
2821 left_child_input_cnt = 1;
2822 }
2823
2824 if (pack_right_child && right_child_LOP &&
2825 (!right_child_predicated ||
2826 (right_child_predicated && parent_is_predicated &&
2827 parent_pred == right_child_pred))) {
2828 partition.push(n->in(2));
2829 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2830 } else {
2831 inputs.push(n->in(2));
2832 right_child_input_cnt = 1;
2833 }
2834
2835 if (inputs.size() > 3) {
2836 assert(partition.size() > 0, "");
2837 inputs.clear();
2838 partition.clear();
2839 if (left_child_input_cnt > right_child_input_cnt) {
2840 pack_left_child = false;
2841 } else {
2842 pack_right_child = false;
2843 }
2844 } else {
2845 break;
2846 }
2847 } while(true);
2848
2849 if(partition.size()) {
2850 partition.push(n);
2851 }
2852
2853 return (partition.size() == 2 || partition.size() == 3) &&
2854 (inputs.size() == 2 || inputs.size() == 3);
2855 }
2856
2857 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2858 assert(is_vector_bitwise_op(n), "not a root");
2859
2860 visited.set(n->_idx);
2861
2862 // 1) Do a DFS walk over the logic cone.
2863 for (uint i = 1; i < n->req(); i++) {
2864 Node* in = n->in(i);
2865 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2866 process_logic_cone_root(igvn, in, visited);
2867 }
2868 }
2869
2870 // 2) Bottom up traversal: Merge node[s] with
2871 // the parent to form macro logic node.
2872 Unique_Node_List partition;
2873 Unique_Node_List inputs;
2874 if (compute_logic_cone(n, partition, inputs)) {
2875 const TypeVect* vt = n->bottom_type()->is_vect();
2876 Node* pn = partition.at(partition.size() - 1);
2877 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2878 if (mask == nullptr ||
2879 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2880 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2881 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2882 igvn.replace_node(n, macro_logic);
2883 }
2884 }
2885 }
2886
2887 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2888 ResourceMark rm;
2889 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2890 Unique_Node_List list;
2891 collect_logic_cone_roots(list);
2892
2893 while (list.size() > 0) {
2894 Node* n = list.pop();
2895 const TypeVect* vt = n->bottom_type()->is_vect();
2896 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2897 if (supported) {
2898 VectorSet visited(comp_arena());
2899 process_logic_cone_root(igvn, n, visited);
2900 }
2901 }
2902 }
2903 }
2904
2905 //------------------------------Code_Gen---------------------------------------
2906 // Given a graph, generate code for it
2907 void Compile::Code_Gen() {
2908 if (failing()) {
2909 return;
2910 }
2911
2912 // Perform instruction selection. You might think we could reclaim Matcher
2913 // memory PDQ, but actually the Matcher is used in generating spill code.
2914 // Internals of the Matcher (including some VectorSets) must remain live
2915 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2916 // set a bit in reclaimed memory.
2917
2918 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2919 // nodes. Mapping is only valid at the root of each matched subtree.
2920 NOT_PRODUCT( verify_graph_edges(); )
2921
2922 Matcher matcher;
2923 _matcher = &matcher;
2924 {
2925 TracePhase tp("matcher", &timers[_t_matcher]);
2926 matcher.match();
2927 if (failing()) {
2928 return;
2929 }
2930 }
2931 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2932 // nodes. Mapping is only valid at the root of each matched subtree.
2933 NOT_PRODUCT( verify_graph_edges(); )
2934
2935 // If you have too many nodes, or if matching has failed, bail out
2936 check_node_count(0, "out of nodes matching instructions");
2937 if (failing()) {
2938 return;
2939 }
2940
2941 print_method(PHASE_MATCHING, 2);
2942
2943 // Build a proper-looking CFG
2944 PhaseCFG cfg(node_arena(), root(), matcher);
2945 _cfg = &cfg;
2946 {
2947 TracePhase tp("scheduler", &timers[_t_scheduler]);
2948 bool success = cfg.do_global_code_motion();
2949 if (!success) {
2950 return;
2951 }
2952
2953 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2954 NOT_PRODUCT( verify_graph_edges(); )
2955 cfg.verify();
2956 }
2957
2958 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2959 _regalloc = ®alloc;
2960 {
2961 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2962 // Perform register allocation. After Chaitin, use-def chains are
2963 // no longer accurate (at spill code) and so must be ignored.
2964 // Node->LRG->reg mappings are still accurate.
2965 _regalloc->Register_Allocate();
2966
2967 // Bail out if the allocator builds too many nodes
2968 if (failing()) {
2969 return;
2970 }
2971
2972 print_method(PHASE_REGISTER_ALLOCATION, 2);
2973 }
2974
2975 // Prior to register allocation we kept empty basic blocks in case the
2976 // the allocator needed a place to spill. After register allocation we
2977 // are not adding any new instructions. If any basic block is empty, we
2978 // can now safely remove it.
2979 {
2980 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2981 cfg.remove_empty_blocks();
2982 if (do_freq_based_layout()) {
2983 PhaseBlockLayout layout(cfg);
2984 } else {
2985 cfg.set_loop_alignment();
2986 }
2987 cfg.fixup_flow();
2988 cfg.remove_unreachable_blocks();
2989 cfg.verify_dominator_tree();
2990 print_method(PHASE_BLOCK_ORDERING, 3);
2991 }
2992
2993 // Apply peephole optimizations
2994 if( OptoPeephole ) {
2995 TracePhase tp("peephole", &timers[_t_peephole]);
2996 PhasePeephole peep( _regalloc, cfg);
2997 peep.do_transform();
2998 print_method(PHASE_PEEPHOLE, 3);
2999 }
3000
3001 // Do late expand if CPU requires this.
3002 if (Matcher::require_postalloc_expand) {
3003 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3004 cfg.postalloc_expand(_regalloc);
3005 print_method(PHASE_POSTALLOC_EXPAND, 3);
3006 }
3007
3008 // Convert Nodes to instruction bits in a buffer
3009 {
3010 TracePhase tp("output", &timers[_t_output]);
3011 PhaseOutput output;
3012 output.Output();
3013 if (failing()) return;
3014 output.install();
3015 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3016 }
3017
3018 // He's dead, Jim.
3019 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3020 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3021 }
3022
3023 //------------------------------Final_Reshape_Counts---------------------------
3024 // This class defines counters to help identify when a method
3025 // may/must be executed using hardware with only 24-bit precision.
3026 struct Final_Reshape_Counts : public StackObj {
3027 int _call_count; // count non-inlined 'common' calls
3028 int _float_count; // count float ops requiring 24-bit precision
3029 int _double_count; // count double ops requiring more precision
3030 int _java_call_count; // count non-inlined 'java' calls
3031 int _inner_loop_count; // count loops which need alignment
3032 VectorSet _visited; // Visitation flags
3033 Node_List _tests; // Set of IfNodes & PCTableNodes
3034
3035 Final_Reshape_Counts() :
3036 _call_count(0), _float_count(0), _double_count(0),
3037 _java_call_count(0), _inner_loop_count(0) { }
3038
3039 void inc_call_count () { _call_count ++; }
3040 void inc_float_count () { _float_count ++; }
3041 void inc_double_count() { _double_count++; }
3042 void inc_java_call_count() { _java_call_count++; }
3043 void inc_inner_loop_count() { _inner_loop_count++; }
3044
3045 int get_call_count () const { return _call_count ; }
3046 int get_float_count () const { return _float_count ; }
3047 int get_double_count() const { return _double_count; }
3048 int get_java_call_count() const { return _java_call_count; }
3049 int get_inner_loop_count() const { return _inner_loop_count; }
3050 };
3051
3052 // Eliminate trivially redundant StoreCMs and accumulate their
3053 // precedence edges.
3054 void Compile::eliminate_redundant_card_marks(Node* n) {
3055 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3056 if (n->in(MemNode::Address)->outcnt() > 1) {
3057 // There are multiple users of the same address so it might be
3058 // possible to eliminate some of the StoreCMs
3059 Node* mem = n->in(MemNode::Memory);
3060 Node* adr = n->in(MemNode::Address);
3061 Node* val = n->in(MemNode::ValueIn);
3062 Node* prev = n;
3063 bool done = false;
3064 // Walk the chain of StoreCMs eliminating ones that match. As
3065 // long as it's a chain of single users then the optimization is
3066 // safe. Eliminating partially redundant StoreCMs would require
3067 // cloning copies down the other paths.
3068 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3069 if (adr == mem->in(MemNode::Address) &&
3070 val == mem->in(MemNode::ValueIn)) {
3071 // redundant StoreCM
3072 if (mem->req() > MemNode::OopStore) {
3073 // Hasn't been processed by this code yet.
3074 n->add_prec(mem->in(MemNode::OopStore));
3075 } else {
3076 // Already converted to precedence edge
3077 for (uint i = mem->req(); i < mem->len(); i++) {
3078 // Accumulate any precedence edges
3079 if (mem->in(i) != nullptr) {
3080 n->add_prec(mem->in(i));
3081 }
3082 }
3083 // Everything above this point has been processed.
3084 done = true;
3085 }
3086 // Eliminate the previous StoreCM
3087 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3088 assert(mem->outcnt() == 0, "should be dead");
3089 mem->disconnect_inputs(this);
3090 } else {
3091 prev = mem;
3092 }
3093 mem = prev->in(MemNode::Memory);
3094 }
3095 }
3096 }
3097
3098 //------------------------------final_graph_reshaping_impl----------------------
3099 // Implement items 1-5 from final_graph_reshaping below.
3100 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3101
3102 if ( n->outcnt() == 0 ) return; // dead node
3103 uint nop = n->Opcode();
3104
3105 // Check for 2-input instruction with "last use" on right input.
3106 // Swap to left input. Implements item (2).
3107 if( n->req() == 3 && // two-input instruction
3108 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3109 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3110 n->in(2)->outcnt() == 1 &&// right use IS a last use
3111 !n->in(2)->is_Con() ) { // right use is not a constant
3112 // Check for commutative opcode
3113 switch( nop ) {
3114 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3115 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3116 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3117 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3118 case Op_AndL: case Op_XorL: case Op_OrL:
3119 case Op_AndI: case Op_XorI: case Op_OrI: {
3120 // Move "last use" input to left by swapping inputs
3121 n->swap_edges(1, 2);
3122 break;
3123 }
3124 default:
3125 break;
3126 }
3127 }
3128
3129 #ifdef ASSERT
3130 if( n->is_Mem() ) {
3131 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3132 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3133 // oop will be recorded in oop map if load crosses safepoint
3134 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3135 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3136 "raw memory operations should have control edge");
3137 }
3138 if (n->is_MemBar()) {
3139 MemBarNode* mb = n->as_MemBar();
3140 if (mb->trailing_store() || mb->trailing_load_store()) {
3141 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3142 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3143 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3144 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3145 } else if (mb->leading()) {
3146 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3147 }
3148 }
3149 #endif
3150 // Count FPU ops and common calls, implements item (3)
3151 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3152 if (!gc_handled) {
3153 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3154 }
3155
3156 // Collect CFG split points
3157 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3158 frc._tests.push(n);
3159 }
3160 }
3161
3162 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3163 switch( nop ) {
3164 // Count all float operations that may use FPU
3165 case Op_AddF:
3166 case Op_SubF:
3167 case Op_MulF:
3168 case Op_DivF:
3169 case Op_NegF:
3170 case Op_ModF:
3171 case Op_ConvI2F:
3172 case Op_ConF:
3173 case Op_CmpF:
3174 case Op_CmpF3:
3175 case Op_StoreF:
3176 case Op_LoadF:
3177 // case Op_ConvL2F: // longs are split into 32-bit halves
3178 frc.inc_float_count();
3179 break;
3180
3181 case Op_ConvF2D:
3182 case Op_ConvD2F:
3183 frc.inc_float_count();
3184 frc.inc_double_count();
3185 break;
3186
3187 // Count all double operations that may use FPU
3188 case Op_AddD:
3189 case Op_SubD:
3190 case Op_MulD:
3191 case Op_DivD:
3192 case Op_NegD:
3193 case Op_ModD:
3194 case Op_ConvI2D:
3195 case Op_ConvD2I:
3196 // case Op_ConvL2D: // handled by leaf call
3197 // case Op_ConvD2L: // handled by leaf call
3198 case Op_ConD:
3199 case Op_CmpD:
3200 case Op_CmpD3:
3201 case Op_StoreD:
3202 case Op_LoadD:
3203 case Op_LoadD_unaligned:
3204 frc.inc_double_count();
3205 break;
3206 case Op_Opaque1: // Remove Opaque Nodes before matching
3207 n->subsume_by(n->in(1), this);
3208 break;
3209 case Op_CallStaticJava:
3210 case Op_CallJava:
3211 case Op_CallDynamicJava:
3212 frc.inc_java_call_count(); // Count java call site;
3213 case Op_CallRuntime:
3214 case Op_CallLeaf:
3215 case Op_CallLeafVector:
3216 case Op_CallLeafNoFP: {
3217 assert (n->is_Call(), "");
3218 CallNode *call = n->as_Call();
3219 // Count call sites where the FP mode bit would have to be flipped.
3220 // Do not count uncommon runtime calls:
3221 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3222 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3223 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3224 frc.inc_call_count(); // Count the call site
3225 } else { // See if uncommon argument is shared
3226 Node *n = call->in(TypeFunc::Parms);
3227 int nop = n->Opcode();
3228 // Clone shared simple arguments to uncommon calls, item (1).
3229 if (n->outcnt() > 1 &&
3230 !n->is_Proj() &&
3231 nop != Op_CreateEx &&
3232 nop != Op_CheckCastPP &&
3233 nop != Op_DecodeN &&
3234 nop != Op_DecodeNKlass &&
3235 !n->is_Mem() &&
3236 !n->is_Phi()) {
3237 Node *x = n->clone();
3238 call->set_req(TypeFunc::Parms, x);
3239 }
3240 }
3241 break;
3242 }
3243
3244 case Op_StoreCM:
3245 {
3246 // Convert OopStore dependence into precedence edge
3247 Node* prec = n->in(MemNode::OopStore);
3248 n->del_req(MemNode::OopStore);
3249 n->add_prec(prec);
3250 eliminate_redundant_card_marks(n);
3251 }
3252
3253 // fall through
3254
3255 case Op_StoreB:
3256 case Op_StoreC:
3257 case Op_StoreI:
3258 case Op_StoreL:
3259 case Op_CompareAndSwapB:
3260 case Op_CompareAndSwapS:
3261 case Op_CompareAndSwapI:
3262 case Op_CompareAndSwapL:
3263 case Op_CompareAndSwapP:
3264 case Op_CompareAndSwapN:
3265 case Op_WeakCompareAndSwapB:
3266 case Op_WeakCompareAndSwapS:
3267 case Op_WeakCompareAndSwapI:
3268 case Op_WeakCompareAndSwapL:
3269 case Op_WeakCompareAndSwapP:
3270 case Op_WeakCompareAndSwapN:
3271 case Op_CompareAndExchangeB:
3272 case Op_CompareAndExchangeS:
3273 case Op_CompareAndExchangeI:
3274 case Op_CompareAndExchangeL:
3275 case Op_CompareAndExchangeP:
3276 case Op_CompareAndExchangeN:
3277 case Op_GetAndAddS:
3278 case Op_GetAndAddB:
3279 case Op_GetAndAddI:
3280 case Op_GetAndAddL:
3281 case Op_GetAndSetS:
3282 case Op_GetAndSetB:
3283 case Op_GetAndSetI:
3284 case Op_GetAndSetL:
3285 case Op_GetAndSetP:
3286 case Op_GetAndSetN:
3287 case Op_StoreP:
3288 case Op_StoreN:
3289 case Op_StoreNKlass:
3290 case Op_LoadB:
3291 case Op_LoadUB:
3292 case Op_LoadUS:
3293 case Op_LoadI:
3294 case Op_LoadKlass:
3295 case Op_LoadNKlass:
3296 case Op_LoadL:
3297 case Op_LoadL_unaligned:
3298 case Op_LoadP:
3299 case Op_LoadN:
3300 case Op_LoadRange:
3301 case Op_LoadS:
3302 break;
3303
3304 case Op_AddP: { // Assert sane base pointers
3305 Node *addp = n->in(AddPNode::Address);
3306 assert( !addp->is_AddP() ||
3307 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3308 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3309 "Base pointers must match (addp %u)", addp->_idx );
3310 #ifdef _LP64
3311 if ((UseCompressedOops || UseCompressedClassPointers) &&
3312 addp->Opcode() == Op_ConP &&
3313 addp == n->in(AddPNode::Base) &&
3314 n->in(AddPNode::Offset)->is_Con()) {
3315 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3316 // on the platform and on the compressed oops mode.
3317 // Use addressing with narrow klass to load with offset on x86.
3318 // Some platforms can use the constant pool to load ConP.
3319 // Do this transformation here since IGVN will convert ConN back to ConP.
3320 const Type* t = addp->bottom_type();
3321 bool is_oop = t->isa_oopptr() != nullptr;
3322 bool is_klass = t->isa_klassptr() != nullptr;
3323
3324 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3325 (is_klass && Matcher::const_klass_prefer_decode())) {
3326 Node* nn = nullptr;
3327
3328 int op = is_oop ? Op_ConN : Op_ConNKlass;
3329
3330 // Look for existing ConN node of the same exact type.
3331 Node* r = root();
3332 uint cnt = r->outcnt();
3333 for (uint i = 0; i < cnt; i++) {
3334 Node* m = r->raw_out(i);
3335 if (m!= nullptr && m->Opcode() == op &&
3336 m->bottom_type()->make_ptr() == t) {
3337 nn = m;
3338 break;
3339 }
3340 }
3341 if (nn != nullptr) {
3342 // Decode a narrow oop to match address
3343 // [R12 + narrow_oop_reg<<3 + offset]
3344 if (is_oop) {
3345 nn = new DecodeNNode(nn, t);
3346 } else {
3347 nn = new DecodeNKlassNode(nn, t);
3348 }
3349 // Check for succeeding AddP which uses the same Base.
3350 // Otherwise we will run into the assertion above when visiting that guy.
3351 for (uint i = 0; i < n->outcnt(); ++i) {
3352 Node *out_i = n->raw_out(i);
3353 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3354 out_i->set_req(AddPNode::Base, nn);
3355 #ifdef ASSERT
3356 for (uint j = 0; j < out_i->outcnt(); ++j) {
3357 Node *out_j = out_i->raw_out(j);
3358 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3359 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3360 }
3361 #endif
3362 }
3363 }
3364 n->set_req(AddPNode::Base, nn);
3365 n->set_req(AddPNode::Address, nn);
3366 if (addp->outcnt() == 0) {
3367 addp->disconnect_inputs(this);
3368 }
3369 }
3370 }
3371 }
3372 #endif
3373 break;
3374 }
3375
3376 case Op_CastPP: {
3377 // Remove CastPP nodes to gain more freedom during scheduling but
3378 // keep the dependency they encode as control or precedence edges
3379 // (if control is set already) on memory operations. Some CastPP
3380 // nodes don't have a control (don't carry a dependency): skip
3381 // those.
3382 if (n->in(0) != nullptr) {
3383 ResourceMark rm;
3384 Unique_Node_List wq;
3385 wq.push(n);
3386 for (uint next = 0; next < wq.size(); ++next) {
3387 Node *m = wq.at(next);
3388 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3389 Node* use = m->fast_out(i);
3390 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3391 use->ensure_control_or_add_prec(n->in(0));
3392 } else {
3393 switch(use->Opcode()) {
3394 case Op_AddP:
3395 case Op_DecodeN:
3396 case Op_DecodeNKlass:
3397 case Op_CheckCastPP:
3398 case Op_CastPP:
3399 wq.push(use);
3400 break;
3401 }
3402 }
3403 }
3404 }
3405 }
3406 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3407 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3408 Node* in1 = n->in(1);
3409 const Type* t = n->bottom_type();
3410 Node* new_in1 = in1->clone();
3411 new_in1->as_DecodeN()->set_type(t);
3412
3413 if (!Matcher::narrow_oop_use_complex_address()) {
3414 //
3415 // x86, ARM and friends can handle 2 adds in addressing mode
3416 // and Matcher can fold a DecodeN node into address by using
3417 // a narrow oop directly and do implicit null check in address:
3418 //
3419 // [R12 + narrow_oop_reg<<3 + offset]
3420 // NullCheck narrow_oop_reg
3421 //
3422 // On other platforms (Sparc) we have to keep new DecodeN node and
3423 // use it to do implicit null check in address:
3424 //
3425 // decode_not_null narrow_oop_reg, base_reg
3426 // [base_reg + offset]
3427 // NullCheck base_reg
3428 //
3429 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3430 // to keep the information to which null check the new DecodeN node
3431 // corresponds to use it as value in implicit_null_check().
3432 //
3433 new_in1->set_req(0, n->in(0));
3434 }
3435
3436 n->subsume_by(new_in1, this);
3437 if (in1->outcnt() == 0) {
3438 in1->disconnect_inputs(this);
3439 }
3440 } else {
3441 n->subsume_by(n->in(1), this);
3442 if (n->outcnt() == 0) {
3443 n->disconnect_inputs(this);
3444 }
3445 }
3446 break;
3447 }
3448 #ifdef _LP64
3449 case Op_CmpP:
3450 // Do this transformation here to preserve CmpPNode::sub() and
3451 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3452 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3453 Node* in1 = n->in(1);
3454 Node* in2 = n->in(2);
3455 if (!in1->is_DecodeNarrowPtr()) {
3456 in2 = in1;
3457 in1 = n->in(2);
3458 }
3459 assert(in1->is_DecodeNarrowPtr(), "sanity");
3460
3461 Node* new_in2 = nullptr;
3462 if (in2->is_DecodeNarrowPtr()) {
3463 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3464 new_in2 = in2->in(1);
3465 } else if (in2->Opcode() == Op_ConP) {
3466 const Type* t = in2->bottom_type();
3467 if (t == TypePtr::NULL_PTR) {
3468 assert(in1->is_DecodeN(), "compare klass to null?");
3469 // Don't convert CmpP null check into CmpN if compressed
3470 // oops implicit null check is not generated.
3471 // This will allow to generate normal oop implicit null check.
3472 if (Matcher::gen_narrow_oop_implicit_null_checks())
3473 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3474 //
3475 // This transformation together with CastPP transformation above
3476 // will generated code for implicit null checks for compressed oops.
3477 //
3478 // The original code after Optimize()
3479 //
3480 // LoadN memory, narrow_oop_reg
3481 // decode narrow_oop_reg, base_reg
3482 // CmpP base_reg, nullptr
3483 // CastPP base_reg // NotNull
3484 // Load [base_reg + offset], val_reg
3485 //
3486 // after these transformations will be
3487 //
3488 // LoadN memory, narrow_oop_reg
3489 // CmpN narrow_oop_reg, nullptr
3490 // decode_not_null narrow_oop_reg, base_reg
3491 // Load [base_reg + offset], val_reg
3492 //
3493 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3494 // since narrow oops can be used in debug info now (see the code in
3495 // final_graph_reshaping_walk()).
3496 //
3497 // At the end the code will be matched to
3498 // on x86:
3499 //
3500 // Load_narrow_oop memory, narrow_oop_reg
3501 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3502 // NullCheck narrow_oop_reg
3503 //
3504 // and on sparc:
3505 //
3506 // Load_narrow_oop memory, narrow_oop_reg
3507 // decode_not_null narrow_oop_reg, base_reg
3508 // Load [base_reg + offset], val_reg
3509 // NullCheck base_reg
3510 //
3511 } else if (t->isa_oopptr()) {
3512 new_in2 = ConNode::make(t->make_narrowoop());
3513 } else if (t->isa_klassptr()) {
3514 new_in2 = ConNode::make(t->make_narrowklass());
3515 }
3516 }
3517 if (new_in2 != nullptr) {
3518 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3519 n->subsume_by(cmpN, this);
3520 if (in1->outcnt() == 0) {
3521 in1->disconnect_inputs(this);
3522 }
3523 if (in2->outcnt() == 0) {
3524 in2->disconnect_inputs(this);
3525 }
3526 }
3527 }
3528 break;
3529
3530 case Op_DecodeN:
3531 case Op_DecodeNKlass:
3532 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3533 // DecodeN could be pinned when it can't be fold into
3534 // an address expression, see the code for Op_CastPP above.
3535 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3536 break;
3537
3538 case Op_EncodeP:
3539 case Op_EncodePKlass: {
3540 Node* in1 = n->in(1);
3541 if (in1->is_DecodeNarrowPtr()) {
3542 n->subsume_by(in1->in(1), this);
3543 } else if (in1->Opcode() == Op_ConP) {
3544 const Type* t = in1->bottom_type();
3545 if (t == TypePtr::NULL_PTR) {
3546 assert(t->isa_oopptr(), "null klass?");
3547 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3548 } else if (t->isa_oopptr()) {
3549 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3550 } else if (t->isa_klassptr()) {
3551 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3552 }
3553 }
3554 if (in1->outcnt() == 0) {
3555 in1->disconnect_inputs(this);
3556 }
3557 break;
3558 }
3559
3560 case Op_Proj: {
3561 if (OptimizeStringConcat || IncrementalInline) {
3562 ProjNode* proj = n->as_Proj();
3563 if (proj->_is_io_use) {
3564 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3565 // Separate projections were used for the exception path which
3566 // are normally removed by a late inline. If it wasn't inlined
3567 // then they will hang around and should just be replaced with
3568 // the original one. Merge them.
3569 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3570 if (non_io_proj != nullptr) {
3571 proj->subsume_by(non_io_proj , this);
3572 }
3573 }
3574 }
3575 break;
3576 }
3577
3578 case Op_Phi:
3579 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3580 // The EncodeP optimization may create Phi with the same edges
3581 // for all paths. It is not handled well by Register Allocator.
3582 Node* unique_in = n->in(1);
3583 assert(unique_in != nullptr, "");
3584 uint cnt = n->req();
3585 for (uint i = 2; i < cnt; i++) {
3586 Node* m = n->in(i);
3587 assert(m != nullptr, "");
3588 if (unique_in != m)
3589 unique_in = nullptr;
3590 }
3591 if (unique_in != nullptr) {
3592 n->subsume_by(unique_in, this);
3593 }
3594 }
3595 break;
3596
3597 #endif
3598
3599 #ifdef ASSERT
3600 case Op_CastII:
3601 // Verify that all range check dependent CastII nodes were removed.
3602 if (n->isa_CastII()->has_range_check()) {
3603 n->dump(3);
3604 assert(false, "Range check dependent CastII node was not removed");
3605 }
3606 break;
3607 #endif
3608
3609 case Op_ModI:
3610 if (UseDivMod) {
3611 // Check if a%b and a/b both exist
3612 Node* d = n->find_similar(Op_DivI);
3613 if (d) {
3614 // Replace them with a fused divmod if supported
3615 if (Matcher::has_match_rule(Op_DivModI)) {
3616 DivModINode* divmod = DivModINode::make(n);
3617 d->subsume_by(divmod->div_proj(), this);
3618 n->subsume_by(divmod->mod_proj(), this);
3619 } else {
3620 // replace a%b with a-((a/b)*b)
3621 Node* mult = new MulINode(d, d->in(2));
3622 Node* sub = new SubINode(d->in(1), mult);
3623 n->subsume_by(sub, this);
3624 }
3625 }
3626 }
3627 break;
3628
3629 case Op_ModL:
3630 if (UseDivMod) {
3631 // Check if a%b and a/b both exist
3632 Node* d = n->find_similar(Op_DivL);
3633 if (d) {
3634 // Replace them with a fused divmod if supported
3635 if (Matcher::has_match_rule(Op_DivModL)) {
3636 DivModLNode* divmod = DivModLNode::make(n);
3637 d->subsume_by(divmod->div_proj(), this);
3638 n->subsume_by(divmod->mod_proj(), this);
3639 } else {
3640 // replace a%b with a-((a/b)*b)
3641 Node* mult = new MulLNode(d, d->in(2));
3642 Node* sub = new SubLNode(d->in(1), mult);
3643 n->subsume_by(sub, this);
3644 }
3645 }
3646 }
3647 break;
3648
3649 case Op_UModI:
3650 if (UseDivMod) {
3651 // Check if a%b and a/b both exist
3652 Node* d = n->find_similar(Op_UDivI);
3653 if (d) {
3654 // Replace them with a fused unsigned divmod if supported
3655 if (Matcher::has_match_rule(Op_UDivModI)) {
3656 UDivModINode* divmod = UDivModINode::make(n);
3657 d->subsume_by(divmod->div_proj(), this);
3658 n->subsume_by(divmod->mod_proj(), this);
3659 } else {
3660 // replace a%b with a-((a/b)*b)
3661 Node* mult = new MulINode(d, d->in(2));
3662 Node* sub = new SubINode(d->in(1), mult);
3663 n->subsume_by(sub, this);
3664 }
3665 }
3666 }
3667 break;
3668
3669 case Op_UModL:
3670 if (UseDivMod) {
3671 // Check if a%b and a/b both exist
3672 Node* d = n->find_similar(Op_UDivL);
3673 if (d) {
3674 // Replace them with a fused unsigned divmod if supported
3675 if (Matcher::has_match_rule(Op_UDivModL)) {
3676 UDivModLNode* divmod = UDivModLNode::make(n);
3677 d->subsume_by(divmod->div_proj(), this);
3678 n->subsume_by(divmod->mod_proj(), this);
3679 } else {
3680 // replace a%b with a-((a/b)*b)
3681 Node* mult = new MulLNode(d, d->in(2));
3682 Node* sub = new SubLNode(d->in(1), mult);
3683 n->subsume_by(sub, this);
3684 }
3685 }
3686 }
3687 break;
3688
3689 case Op_LoadVector:
3690 case Op_StoreVector:
3691 #ifdef ASSERT
3692 // Add VerifyVectorAlignment node between adr and load / store.
3693 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3694 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3695 n->as_StoreVector()->must_verify_alignment();
3696 if (must_verify_alignment) {
3697 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3698 n->as_StoreVector()->memory_size();
3699 // The memory access should be aligned to the vector width in bytes.
3700 // However, the underlying array is possibly less well aligned, but at least
3701 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3702 // a loop we can expect at least the following alignment:
3703 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3704 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3705 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3706 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3707 jlong mask = guaranteed_alignment - 1;
3708 Node* mask_con = ConLNode::make(mask);
3709 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3710 n->set_req(MemNode::Address, va);
3711 }
3712 }
3713 #endif
3714 break;
3715
3716 case Op_LoadVectorGather:
3717 case Op_StoreVectorScatter:
3718 case Op_LoadVectorGatherMasked:
3719 case Op_StoreVectorScatterMasked:
3720 case Op_VectorCmpMasked:
3721 case Op_VectorMaskGen:
3722 case Op_LoadVectorMasked:
3723 case Op_StoreVectorMasked:
3724 break;
3725
3726 case Op_AddReductionVI:
3727 case Op_AddReductionVL:
3728 case Op_AddReductionVF:
3729 case Op_AddReductionVD:
3730 case Op_MulReductionVI:
3731 case Op_MulReductionVL:
3732 case Op_MulReductionVF:
3733 case Op_MulReductionVD:
3734 case Op_MinReductionV:
3735 case Op_MaxReductionV:
3736 case Op_AndReductionV:
3737 case Op_OrReductionV:
3738 case Op_XorReductionV:
3739 break;
3740
3741 case Op_PackB:
3742 case Op_PackS:
3743 case Op_PackI:
3744 case Op_PackF:
3745 case Op_PackL:
3746 case Op_PackD:
3747 if (n->req()-1 > 2) {
3748 // Replace many operand PackNodes with a binary tree for matching
3749 PackNode* p = (PackNode*) n;
3750 Node* btp = p->binary_tree_pack(1, n->req());
3751 n->subsume_by(btp, this);
3752 }
3753 break;
3754 case Op_Loop:
3755 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3756 case Op_CountedLoop:
3757 case Op_LongCountedLoop:
3758 case Op_OuterStripMinedLoop:
3759 if (n->as_Loop()->is_inner_loop()) {
3760 frc.inc_inner_loop_count();
3761 }
3762 n->as_Loop()->verify_strip_mined(0);
3763 break;
3764 case Op_LShiftI:
3765 case Op_RShiftI:
3766 case Op_URShiftI:
3767 case Op_LShiftL:
3768 case Op_RShiftL:
3769 case Op_URShiftL:
3770 if (Matcher::need_masked_shift_count) {
3771 // The cpu's shift instructions don't restrict the count to the
3772 // lower 5/6 bits. We need to do the masking ourselves.
3773 Node* in2 = n->in(2);
3774 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3775 const TypeInt* t = in2->find_int_type();
3776 if (t != nullptr && t->is_con()) {
3777 juint shift = t->get_con();
3778 if (shift > mask) { // Unsigned cmp
3779 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3780 }
3781 } else {
3782 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3783 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3784 n->set_req(2, shift);
3785 }
3786 }
3787 if (in2->outcnt() == 0) { // Remove dead node
3788 in2->disconnect_inputs(this);
3789 }
3790 }
3791 break;
3792 case Op_MemBarStoreStore:
3793 case Op_MemBarRelease:
3794 // Break the link with AllocateNode: it is no longer useful and
3795 // confuses register allocation.
3796 if (n->req() > MemBarNode::Precedent) {
3797 n->set_req(MemBarNode::Precedent, top());
3798 }
3799 break;
3800 case Op_MemBarAcquire: {
3801 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3802 // At parse time, the trailing MemBarAcquire for a volatile load
3803 // is created with an edge to the load. After optimizations,
3804 // that input may be a chain of Phis. If those phis have no
3805 // other use, then the MemBarAcquire keeps them alive and
3806 // register allocation can be confused.
3807 dead_nodes.push(n->in(MemBarNode::Precedent));
3808 n->set_req(MemBarNode::Precedent, top());
3809 }
3810 break;
3811 }
3812 case Op_Blackhole:
3813 break;
3814 case Op_RangeCheck: {
3815 RangeCheckNode* rc = n->as_RangeCheck();
3816 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3817 n->subsume_by(iff, this);
3818 frc._tests.push(iff);
3819 break;
3820 }
3821 case Op_ConvI2L: {
3822 if (!Matcher::convi2l_type_required) {
3823 // Code generation on some platforms doesn't need accurate
3824 // ConvI2L types. Widening the type can help remove redundant
3825 // address computations.
3826 n->as_Type()->set_type(TypeLong::INT);
3827 ResourceMark rm;
3828 Unique_Node_List wq;
3829 wq.push(n);
3830 for (uint next = 0; next < wq.size(); next++) {
3831 Node *m = wq.at(next);
3832
3833 for(;;) {
3834 // Loop over all nodes with identical inputs edges as m
3835 Node* k = m->find_similar(m->Opcode());
3836 if (k == nullptr) {
3837 break;
3838 }
3839 // Push their uses so we get a chance to remove node made
3840 // redundant
3841 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3842 Node* u = k->fast_out(i);
3843 if (u->Opcode() == Op_LShiftL ||
3844 u->Opcode() == Op_AddL ||
3845 u->Opcode() == Op_SubL ||
3846 u->Opcode() == Op_AddP) {
3847 wq.push(u);
3848 }
3849 }
3850 // Replace all nodes with identical edges as m with m
3851 k->subsume_by(m, this);
3852 }
3853 }
3854 }
3855 break;
3856 }
3857 case Op_CmpUL: {
3858 if (!Matcher::has_match_rule(Op_CmpUL)) {
3859 // No support for unsigned long comparisons
3860 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3861 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3862 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3863 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3864 Node* andl = new AndLNode(orl, remove_sign_mask);
3865 Node* cmp = new CmpLNode(andl, n->in(2));
3866 n->subsume_by(cmp, this);
3867 }
3868 break;
3869 }
3870 default:
3871 assert(!n->is_Call(), "");
3872 assert(!n->is_Mem(), "");
3873 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3874 break;
3875 }
3876 }
3877
3878 //------------------------------final_graph_reshaping_walk---------------------
3879 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3880 // requires that the walk visits a node's inputs before visiting the node.
3881 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3882 Unique_Node_List sfpt;
3883
3884 frc._visited.set(root->_idx); // first, mark node as visited
3885 uint cnt = root->req();
3886 Node *n = root;
3887 uint i = 0;
3888 while (true) {
3889 if (i < cnt) {
3890 // Place all non-visited non-null inputs onto stack
3891 Node* m = n->in(i);
3892 ++i;
3893 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3894 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3895 // compute worst case interpreter size in case of a deoptimization
3896 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3897
3898 sfpt.push(m);
3899 }
3900 cnt = m->req();
3901 nstack.push(n, i); // put on stack parent and next input's index
3902 n = m;
3903 i = 0;
3904 }
3905 } else {
3906 // Now do post-visit work
3907 final_graph_reshaping_impl(n, frc, dead_nodes);
3908 if (nstack.is_empty())
3909 break; // finished
3910 n = nstack.node(); // Get node from stack
3911 cnt = n->req();
3912 i = nstack.index();
3913 nstack.pop(); // Shift to the next node on stack
3914 }
3915 }
3916
3917 // Skip next transformation if compressed oops are not used.
3918 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3919 (!UseCompressedOops && !UseCompressedClassPointers))
3920 return;
3921
3922 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3923 // It could be done for an uncommon traps or any safepoints/calls
3924 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3925 while (sfpt.size() > 0) {
3926 n = sfpt.pop();
3927 JVMState *jvms = n->as_SafePoint()->jvms();
3928 assert(jvms != nullptr, "sanity");
3929 int start = jvms->debug_start();
3930 int end = n->req();
3931 bool is_uncommon = (n->is_CallStaticJava() &&
3932 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3933 for (int j = start; j < end; j++) {
3934 Node* in = n->in(j);
3935 if (in->is_DecodeNarrowPtr()) {
3936 bool safe_to_skip = true;
3937 if (!is_uncommon ) {
3938 // Is it safe to skip?
3939 for (uint i = 0; i < in->outcnt(); i++) {
3940 Node* u = in->raw_out(i);
3941 if (!u->is_SafePoint() ||
3942 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3943 safe_to_skip = false;
3944 }
3945 }
3946 }
3947 if (safe_to_skip) {
3948 n->set_req(j, in->in(1));
3949 }
3950 if (in->outcnt() == 0) {
3951 in->disconnect_inputs(this);
3952 }
3953 }
3954 }
3955 }
3956 }
3957
3958 //------------------------------final_graph_reshaping--------------------------
3959 // Final Graph Reshaping.
3960 //
3961 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3962 // and not commoned up and forced early. Must come after regular
3963 // optimizations to avoid GVN undoing the cloning. Clone constant
3964 // inputs to Loop Phis; these will be split by the allocator anyways.
3965 // Remove Opaque nodes.
3966 // (2) Move last-uses by commutative operations to the left input to encourage
3967 // Intel update-in-place two-address operations and better register usage
3968 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3969 // calls canonicalizing them back.
3970 // (3) Count the number of double-precision FP ops, single-precision FP ops
3971 // and call sites. On Intel, we can get correct rounding either by
3972 // forcing singles to memory (requires extra stores and loads after each
3973 // FP bytecode) or we can set a rounding mode bit (requires setting and
3974 // clearing the mode bit around call sites). The mode bit is only used
3975 // if the relative frequency of single FP ops to calls is low enough.
3976 // This is a key transform for SPEC mpeg_audio.
3977 // (4) Detect infinite loops; blobs of code reachable from above but not
3978 // below. Several of the Code_Gen algorithms fail on such code shapes,
3979 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3980 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3981 // Detection is by looking for IfNodes where only 1 projection is
3982 // reachable from below or CatchNodes missing some targets.
3983 // (5) Assert for insane oop offsets in debug mode.
3984
3985 bool Compile::final_graph_reshaping() {
3986 // an infinite loop may have been eliminated by the optimizer,
3987 // in which case the graph will be empty.
3988 if (root()->req() == 1) {
3989 // Do not compile method that is only a trivial infinite loop,
3990 // since the content of the loop may have been eliminated.
3991 record_method_not_compilable("trivial infinite loop");
3992 return true;
3993 }
3994
3995 // Expensive nodes have their control input set to prevent the GVN
3996 // from freely commoning them. There's no GVN beyond this point so
3997 // no need to keep the control input. We want the expensive nodes to
3998 // be freely moved to the least frequent code path by gcm.
3999 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4000 for (int i = 0; i < expensive_count(); i++) {
4001 _expensive_nodes.at(i)->set_req(0, nullptr);
4002 }
4003
4004 Final_Reshape_Counts frc;
4005
4006 // Visit everybody reachable!
4007 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4008 Node_Stack nstack(live_nodes() >> 1);
4009 Unique_Node_List dead_nodes;
4010 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4011
4012 // Check for unreachable (from below) code (i.e., infinite loops).
4013 for( uint i = 0; i < frc._tests.size(); i++ ) {
4014 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4015 // Get number of CFG targets.
4016 // Note that PCTables include exception targets after calls.
4017 uint required_outcnt = n->required_outcnt();
4018 if (n->outcnt() != required_outcnt) {
4019 // Check for a few special cases. Rethrow Nodes never take the
4020 // 'fall-thru' path, so expected kids is 1 less.
4021 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4022 if (n->in(0)->in(0)->is_Call()) {
4023 CallNode* call = n->in(0)->in(0)->as_Call();
4024 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4025 required_outcnt--; // Rethrow always has 1 less kid
4026 } else if (call->req() > TypeFunc::Parms &&
4027 call->is_CallDynamicJava()) {
4028 // Check for null receiver. In such case, the optimizer has
4029 // detected that the virtual call will always result in a null
4030 // pointer exception. The fall-through projection of this CatchNode
4031 // will not be populated.
4032 Node* arg0 = call->in(TypeFunc::Parms);
4033 if (arg0->is_Type() &&
4034 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4035 required_outcnt--;
4036 }
4037 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4038 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4039 // Check for illegal array length. In such case, the optimizer has
4040 // detected that the allocation attempt will always result in an
4041 // exception. There is no fall-through projection of this CatchNode .
4042 assert(call->is_CallStaticJava(), "static call expected");
4043 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4044 uint valid_length_test_input = call->req() - 1;
4045 Node* valid_length_test = call->in(valid_length_test_input);
4046 call->del_req(valid_length_test_input);
4047 if (valid_length_test->find_int_con(1) == 0) {
4048 required_outcnt--;
4049 }
4050 dead_nodes.push(valid_length_test);
4051 assert(n->outcnt() == required_outcnt, "malformed control flow");
4052 continue;
4053 }
4054 }
4055 }
4056
4057 // Recheck with a better notion of 'required_outcnt'
4058 if (n->outcnt() != required_outcnt) {
4059 record_method_not_compilable("malformed control flow");
4060 return true; // Not all targets reachable!
4061 }
4062 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4063 CallNode* call = n->in(0)->in(0)->as_Call();
4064 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4065 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4066 assert(call->is_CallStaticJava(), "static call expected");
4067 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4068 uint valid_length_test_input = call->req() - 1;
4069 dead_nodes.push(call->in(valid_length_test_input));
4070 call->del_req(valid_length_test_input); // valid length test useless now
4071 }
4072 }
4073 // Check that I actually visited all kids. Unreached kids
4074 // must be infinite loops.
4075 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4076 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4077 record_method_not_compilable("infinite loop");
4078 return true; // Found unvisited kid; must be unreach
4079 }
4080
4081 // Here so verification code in final_graph_reshaping_walk()
4082 // always see an OuterStripMinedLoopEnd
4083 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4084 IfNode* init_iff = n->as_If();
4085 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4086 n->subsume_by(iff, this);
4087 }
4088 }
4089
4090 while (dead_nodes.size() > 0) {
4091 Node* m = dead_nodes.pop();
4092 if (m->outcnt() == 0 && m != top()) {
4093 for (uint j = 0; j < m->req(); j++) {
4094 Node* in = m->in(j);
4095 if (in != nullptr) {
4096 dead_nodes.push(in);
4097 }
4098 }
4099 m->disconnect_inputs(this);
4100 }
4101 }
4102
4103 #ifdef IA32
4104 // If original bytecodes contained a mixture of floats and doubles
4105 // check if the optimizer has made it homogeneous, item (3).
4106 if (UseSSE == 0 &&
4107 frc.get_float_count() > 32 &&
4108 frc.get_double_count() == 0 &&
4109 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4110 set_24_bit_selection_and_mode(false, true);
4111 }
4112 #endif // IA32
4113
4114 set_java_calls(frc.get_java_call_count());
4115 set_inner_loops(frc.get_inner_loop_count());
4116
4117 // No infinite loops, no reason to bail out.
4118 return false;
4119 }
4120
4121 //-----------------------------too_many_traps----------------------------------
4122 // Report if there are too many traps at the current method and bci.
4123 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4124 bool Compile::too_many_traps(ciMethod* method,
4125 int bci,
4126 Deoptimization::DeoptReason reason) {
4127 ciMethodData* md = method->method_data();
4128 if (md->is_empty()) {
4129 // Assume the trap has not occurred, or that it occurred only
4130 // because of a transient condition during start-up in the interpreter.
4131 return false;
4132 }
4133 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4134 if (md->has_trap_at(bci, m, reason) != 0) {
4135 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4136 // Also, if there are multiple reasons, or if there is no per-BCI record,
4137 // assume the worst.
4138 if (log())
4139 log()->elem("observe trap='%s' count='%d'",
4140 Deoptimization::trap_reason_name(reason),
4141 md->trap_count(reason));
4142 return true;
4143 } else {
4144 // Ignore method/bci and see if there have been too many globally.
4145 return too_many_traps(reason, md);
4146 }
4147 }
4148
4149 // Less-accurate variant which does not require a method and bci.
4150 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4151 ciMethodData* logmd) {
4152 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4153 // Too many traps globally.
4154 // Note that we use cumulative trap_count, not just md->trap_count.
4155 if (log()) {
4156 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4157 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4158 Deoptimization::trap_reason_name(reason),
4159 mcount, trap_count(reason));
4160 }
4161 return true;
4162 } else {
4163 // The coast is clear.
4164 return false;
4165 }
4166 }
4167
4168 //--------------------------too_many_recompiles--------------------------------
4169 // Report if there are too many recompiles at the current method and bci.
4170 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4171 // Is not eager to return true, since this will cause the compiler to use
4172 // Action_none for a trap point, to avoid too many recompilations.
4173 bool Compile::too_many_recompiles(ciMethod* method,
4174 int bci,
4175 Deoptimization::DeoptReason reason) {
4176 ciMethodData* md = method->method_data();
4177 if (md->is_empty()) {
4178 // Assume the trap has not occurred, or that it occurred only
4179 // because of a transient condition during start-up in the interpreter.
4180 return false;
4181 }
4182 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4183 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4184 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4185 Deoptimization::DeoptReason per_bc_reason
4186 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4187 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4188 if ((per_bc_reason == Deoptimization::Reason_none
4189 || md->has_trap_at(bci, m, reason) != 0)
4190 // The trap frequency measure we care about is the recompile count:
4191 && md->trap_recompiled_at(bci, m)
4192 && md->overflow_recompile_count() >= bc_cutoff) {
4193 // Do not emit a trap here if it has already caused recompilations.
4194 // Also, if there are multiple reasons, or if there is no per-BCI record,
4195 // assume the worst.
4196 if (log())
4197 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4198 Deoptimization::trap_reason_name(reason),
4199 md->trap_count(reason),
4200 md->overflow_recompile_count());
4201 return true;
4202 } else if (trap_count(reason) != 0
4203 && decompile_count() >= m_cutoff) {
4204 // Too many recompiles globally, and we have seen this sort of trap.
4205 // Use cumulative decompile_count, not just md->decompile_count.
4206 if (log())
4207 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4208 Deoptimization::trap_reason_name(reason),
4209 md->trap_count(reason), trap_count(reason),
4210 md->decompile_count(), decompile_count());
4211 return true;
4212 } else {
4213 // The coast is clear.
4214 return false;
4215 }
4216 }
4217
4218 // Compute when not to trap. Used by matching trap based nodes and
4219 // NullCheck optimization.
4220 void Compile::set_allowed_deopt_reasons() {
4221 _allowed_reasons = 0;
4222 if (is_method_compilation()) {
4223 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4224 assert(rs < BitsPerInt, "recode bit map");
4225 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4226 _allowed_reasons |= nth_bit(rs);
4227 }
4228 }
4229 }
4230 }
4231
4232 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4233 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4234 }
4235
4236 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4237 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4238 }
4239
4240 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4241 if (holder->is_initialized()) {
4242 return false;
4243 }
4244 if (holder->is_being_initialized()) {
4245 if (accessing_method->holder() == holder) {
4246 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4247 // <init>, or a static method. In all those cases, there was an initialization
4248 // barrier on the holder klass passed.
4249 if (accessing_method->is_static_initializer() ||
4250 accessing_method->is_object_initializer() ||
4251 accessing_method->is_static()) {
4252 return false;
4253 }
4254 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4255 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4256 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4257 // child class can become fully initialized while its parent class is still being initialized.
4258 if (accessing_method->is_static_initializer()) {
4259 return false;
4260 }
4261 }
4262 ciMethod* root = method(); // the root method of compilation
4263 if (root != accessing_method) {
4264 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4265 }
4266 }
4267 return true;
4268 }
4269
4270 #ifndef PRODUCT
4271 //------------------------------verify_bidirectional_edges---------------------
4272 // For each input edge to a node (ie - for each Use-Def edge), verify that
4273 // there is a corresponding Def-Use edge.
4274 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4275 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4276 uint stack_size = live_nodes() >> 4;
4277 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4278 nstack.push(_root);
4279
4280 while (nstack.size() > 0) {
4281 Node* n = nstack.pop();
4282 if (visited.member(n)) {
4283 continue;
4284 }
4285 visited.push(n);
4286
4287 // Walk over all input edges, checking for correspondence
4288 uint length = n->len();
4289 for (uint i = 0; i < length; i++) {
4290 Node* in = n->in(i);
4291 if (in != nullptr && !visited.member(in)) {
4292 nstack.push(in); // Put it on stack
4293 }
4294 if (in != nullptr && !in->is_top()) {
4295 // Count instances of `next`
4296 int cnt = 0;
4297 for (uint idx = 0; idx < in->_outcnt; idx++) {
4298 if (in->_out[idx] == n) {
4299 cnt++;
4300 }
4301 }
4302 assert(cnt > 0, "Failed to find Def-Use edge.");
4303 // Check for duplicate edges
4304 // walk the input array downcounting the input edges to n
4305 for (uint j = 0; j < length; j++) {
4306 if (n->in(j) == in) {
4307 cnt--;
4308 }
4309 }
4310 assert(cnt == 0, "Mismatched edge count.");
4311 } else if (in == nullptr) {
4312 assert(i == 0 || i >= n->req() ||
4313 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4314 (n->is_Unlock() && i == (n->req() - 1)) ||
4315 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4316 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4317 } else {
4318 assert(in->is_top(), "sanity");
4319 // Nothing to check.
4320 }
4321 }
4322 }
4323 }
4324
4325 //------------------------------verify_graph_edges---------------------------
4326 // Walk the Graph and verify that there is a one-to-one correspondence
4327 // between Use-Def edges and Def-Use edges in the graph.
4328 void Compile::verify_graph_edges(bool no_dead_code) {
4329 if (VerifyGraphEdges) {
4330 Unique_Node_List visited;
4331
4332 // Call graph walk to check edges
4333 verify_bidirectional_edges(visited);
4334 if (no_dead_code) {
4335 // Now make sure that no visited node is used by an unvisited node.
4336 bool dead_nodes = false;
4337 Unique_Node_List checked;
4338 while (visited.size() > 0) {
4339 Node* n = visited.pop();
4340 checked.push(n);
4341 for (uint i = 0; i < n->outcnt(); i++) {
4342 Node* use = n->raw_out(i);
4343 if (checked.member(use)) continue; // already checked
4344 if (visited.member(use)) continue; // already in the graph
4345 if (use->is_Con()) continue; // a dead ConNode is OK
4346 // At this point, we have found a dead node which is DU-reachable.
4347 if (!dead_nodes) {
4348 tty->print_cr("*** Dead nodes reachable via DU edges:");
4349 dead_nodes = true;
4350 }
4351 use->dump(2);
4352 tty->print_cr("---");
4353 checked.push(use); // No repeats; pretend it is now checked.
4354 }
4355 }
4356 assert(!dead_nodes, "using nodes must be reachable from root");
4357 }
4358 }
4359 }
4360 #endif
4361
4362 // The Compile object keeps track of failure reasons separately from the ciEnv.
4363 // This is required because there is not quite a 1-1 relation between the
4364 // ciEnv and its compilation task and the Compile object. Note that one
4365 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4366 // to backtrack and retry without subsuming loads. Other than this backtracking
4367 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4368 // by the logic in C2Compiler.
4369 void Compile::record_failure(const char* reason) {
4370 if (log() != nullptr) {
4371 log()->elem("failure reason='%s' phase='compile'", reason);
4372 }
4373 if (_failure_reason.get() == nullptr) {
4374 // Record the first failure reason.
4375 _failure_reason.set(reason);
4376 if (CaptureBailoutInformation) {
4377 _first_failure_details = new CompilationFailureInfo(reason);
4378 }
4379 }
4380
4381 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4382 C->print_method(PHASE_FAILURE, 1);
4383 }
4384 _root = nullptr; // flush the graph, too
4385 }
4386
4387 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4388 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4389 _compile(Compile::current()),
4390 _log(nullptr),
4391 _phase_name(name),
4392 _dolog(CITimeVerbose)
4393 {
4394 assert(_compile != nullptr, "sanity check");
4395 if (_dolog) {
4396 _log = _compile->log();
4397 }
4398 if (_log != nullptr) {
4399 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4400 _log->stamp();
4401 _log->end_head();
4402 }
4403 }
4404
4405 Compile::TracePhase::~TracePhase() {
4406 if (_compile->failing()) return;
4407 #ifdef ASSERT
4408 if (PrintIdealNodeCount) {
4409 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4410 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4411 }
4412
4413 if (VerifyIdealNodeCount) {
4414 _compile->print_missing_nodes();
4415 }
4416 #endif
4417
4418 if (_log != nullptr) {
4419 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4420 }
4421 }
4422
4423 //----------------------------static_subtype_check-----------------------------
4424 // Shortcut important common cases when superklass is exact:
4425 // (0) superklass is java.lang.Object (can occur in reflective code)
4426 // (1) subklass is already limited to a subtype of superklass => always ok
4427 // (2) subklass does not overlap with superklass => always fail
4428 // (3) superklass has NO subtypes and we can check with a simple compare.
4429 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4430 if (skip) {
4431 return SSC_full_test; // Let caller generate the general case.
4432 }
4433
4434 if (subk->is_java_subtype_of(superk)) {
4435 return SSC_always_true; // (0) and (1) this test cannot fail
4436 }
4437
4438 if (!subk->maybe_java_subtype_of(superk)) {
4439 return SSC_always_false; // (2) true path dead; no dynamic test needed
4440 }
4441
4442 const Type* superelem = superk;
4443 if (superk->isa_aryklassptr()) {
4444 int ignored;
4445 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4446 }
4447
4448 if (superelem->isa_instklassptr()) {
4449 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4450 if (!ik->has_subklass()) {
4451 if (!ik->is_final()) {
4452 // Add a dependency if there is a chance of a later subclass.
4453 dependencies()->assert_leaf_type(ik);
4454 }
4455 if (!superk->maybe_java_subtype_of(subk)) {
4456 return SSC_always_false;
4457 }
4458 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4459 }
4460 } else {
4461 // A primitive array type has no subtypes.
4462 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4463 }
4464
4465 return SSC_full_test;
4466 }
4467
4468 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4469 #ifdef _LP64
4470 // The scaled index operand to AddP must be a clean 64-bit value.
4471 // Java allows a 32-bit int to be incremented to a negative
4472 // value, which appears in a 64-bit register as a large
4473 // positive number. Using that large positive number as an
4474 // operand in pointer arithmetic has bad consequences.
4475 // On the other hand, 32-bit overflow is rare, and the possibility
4476 // can often be excluded, if we annotate the ConvI2L node with
4477 // a type assertion that its value is known to be a small positive
4478 // number. (The prior range check has ensured this.)
4479 // This assertion is used by ConvI2LNode::Ideal.
4480 int index_max = max_jint - 1; // array size is max_jint, index is one less
4481 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4482 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4483 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4484 #endif
4485 return idx;
4486 }
4487
4488 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4489 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4490 if (ctrl != nullptr) {
4491 // Express control dependency by a CastII node with a narrow type.
4492 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4493 // node from floating above the range check during loop optimizations. Otherwise, the
4494 // ConvI2L node may be eliminated independently of the range check, causing the data path
4495 // to become TOP while the control path is still there (although it's unreachable).
4496 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4497 value = phase->transform(value);
4498 }
4499 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4500 return phase->transform(new ConvI2LNode(value, ltype));
4501 }
4502
4503 // The message about the current inlining is accumulated in
4504 // _print_inlining_stream and transferred into the _print_inlining_list
4505 // once we know whether inlining succeeds or not. For regular
4506 // inlining, messages are appended to the buffer pointed by
4507 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4508 // a new buffer is added after _print_inlining_idx in the list. This
4509 // way we can update the inlining message for late inlining call site
4510 // when the inlining is attempted again.
4511 void Compile::print_inlining_init() {
4512 if (print_inlining() || print_intrinsics()) {
4513 // print_inlining_init is actually called several times.
4514 print_inlining_reset();
4515 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4516 }
4517 }
4518
4519 void Compile::print_inlining_reinit() {
4520 if (print_inlining() || print_intrinsics()) {
4521 print_inlining_reset();
4522 }
4523 }
4524
4525 void Compile::print_inlining_reset() {
4526 _print_inlining_stream->reset();
4527 }
4528
4529 void Compile::print_inlining_commit() {
4530 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4531 // Transfer the message from _print_inlining_stream to the current
4532 // _print_inlining_list buffer and clear _print_inlining_stream.
4533 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4534 print_inlining_reset();
4535 }
4536
4537 void Compile::print_inlining_push() {
4538 // Add new buffer to the _print_inlining_list at current position
4539 _print_inlining_idx++;
4540 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4541 }
4542
4543 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4544 return _print_inlining_list->at(_print_inlining_idx);
4545 }
4546
4547 void Compile::print_inlining_update(CallGenerator* cg) {
4548 if (print_inlining() || print_intrinsics()) {
4549 if (cg->is_late_inline()) {
4550 if (print_inlining_current()->cg() != cg &&
4551 (print_inlining_current()->cg() != nullptr ||
4552 print_inlining_current()->ss()->size() != 0)) {
4553 print_inlining_push();
4554 }
4555 print_inlining_commit();
4556 print_inlining_current()->set_cg(cg);
4557 } else {
4558 if (print_inlining_current()->cg() != nullptr) {
4559 print_inlining_push();
4560 }
4561 print_inlining_commit();
4562 }
4563 }
4564 }
4565
4566 void Compile::print_inlining_move_to(CallGenerator* cg) {
4567 // We resume inlining at a late inlining call site. Locate the
4568 // corresponding inlining buffer so that we can update it.
4569 if (print_inlining() || print_intrinsics()) {
4570 for (int i = 0; i < _print_inlining_list->length(); i++) {
4571 if (_print_inlining_list->at(i)->cg() == cg) {
4572 _print_inlining_idx = i;
4573 return;
4574 }
4575 }
4576 ShouldNotReachHere();
4577 }
4578 }
4579
4580 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4581 if (print_inlining() || print_intrinsics()) {
4582 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4583 assert(print_inlining_current()->cg() == cg, "wrong entry");
4584 // replace message with new message
4585 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4586 print_inlining_commit();
4587 print_inlining_current()->set_cg(cg);
4588 }
4589 }
4590
4591 void Compile::print_inlining_assert_ready() {
4592 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4593 }
4594
4595 void Compile::process_print_inlining() {
4596 assert(_late_inlines.length() == 0, "not drained yet");
4597 if (print_inlining() || print_intrinsics()) {
4598 ResourceMark rm;
4599 stringStream ss;
4600 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4601 for (int i = 0; i < _print_inlining_list->length(); i++) {
4602 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4603 ss.print("%s", pib->ss()->freeze());
4604 delete pib;
4605 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4606 }
4607 // Reset _print_inlining_list, it only contains destructed objects.
4608 // It is on the arena, so it will be freed when the arena is reset.
4609 _print_inlining_list = nullptr;
4610 // _print_inlining_stream won't be used anymore, either.
4611 print_inlining_reset();
4612 size_t end = ss.size();
4613 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4614 strncpy(_print_inlining_output, ss.freeze(), end+1);
4615 _print_inlining_output[end] = 0;
4616 }
4617 }
4618
4619 void Compile::dump_print_inlining() {
4620 if (_print_inlining_output != nullptr) {
4621 tty->print_raw(_print_inlining_output);
4622 }
4623 }
4624
4625 void Compile::log_late_inline(CallGenerator* cg) {
4626 if (log() != nullptr) {
4627 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4628 cg->unique_id());
4629 JVMState* p = cg->call_node()->jvms();
4630 while (p != nullptr) {
4631 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4632 p = p->caller();
4633 }
4634 log()->tail("late_inline");
4635 }
4636 }
4637
4638 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4639 log_late_inline(cg);
4640 if (log() != nullptr) {
4641 log()->inline_fail(msg);
4642 }
4643 }
4644
4645 void Compile::log_inline_id(CallGenerator* cg) {
4646 if (log() != nullptr) {
4647 // The LogCompilation tool needs a unique way to identify late
4648 // inline call sites. This id must be unique for this call site in
4649 // this compilation. Try to have it unique across compilations as
4650 // well because it can be convenient when grepping through the log
4651 // file.
4652 // Distinguish OSR compilations from others in case CICountOSR is
4653 // on.
4654 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4655 cg->set_unique_id(id);
4656 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4657 }
4658 }
4659
4660 void Compile::log_inline_failure(const char* msg) {
4661 if (C->log() != nullptr) {
4662 C->log()->inline_fail(msg);
4663 }
4664 }
4665
4666
4667 // Dump inlining replay data to the stream.
4668 // Don't change thread state and acquire any locks.
4669 void Compile::dump_inline_data(outputStream* out) {
4670 InlineTree* inl_tree = ilt();
4671 if (inl_tree != nullptr) {
4672 out->print(" inline %d", inl_tree->count());
4673 inl_tree->dump_replay_data(out);
4674 }
4675 }
4676
4677 void Compile::dump_inline_data_reduced(outputStream* out) {
4678 assert(ReplayReduce, "");
4679
4680 InlineTree* inl_tree = ilt();
4681 if (inl_tree == nullptr) {
4682 return;
4683 }
4684 // Enable iterative replay file reduction
4685 // Output "compile" lines for depth 1 subtrees,
4686 // simulating that those trees were compiled
4687 // instead of inlined.
4688 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4689 InlineTree* sub = inl_tree->subtrees().at(i);
4690 if (sub->inline_level() != 1) {
4691 continue;
4692 }
4693
4694 ciMethod* method = sub->method();
4695 int entry_bci = -1;
4696 int comp_level = env()->task()->comp_level();
4697 out->print("compile ");
4698 method->dump_name_as_ascii(out);
4699 out->print(" %d %d", entry_bci, comp_level);
4700 out->print(" inline %d", sub->count());
4701 sub->dump_replay_data(out, -1);
4702 out->cr();
4703 }
4704 }
4705
4706 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4707 if (n1->Opcode() < n2->Opcode()) return -1;
4708 else if (n1->Opcode() > n2->Opcode()) return 1;
4709
4710 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4711 for (uint i = 1; i < n1->req(); i++) {
4712 if (n1->in(i) < n2->in(i)) return -1;
4713 else if (n1->in(i) > n2->in(i)) return 1;
4714 }
4715
4716 return 0;
4717 }
4718
4719 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4720 Node* n1 = *n1p;
4721 Node* n2 = *n2p;
4722
4723 return cmp_expensive_nodes(n1, n2);
4724 }
4725
4726 void Compile::sort_expensive_nodes() {
4727 if (!expensive_nodes_sorted()) {
4728 _expensive_nodes.sort(cmp_expensive_nodes);
4729 }
4730 }
4731
4732 bool Compile::expensive_nodes_sorted() const {
4733 for (int i = 1; i < _expensive_nodes.length(); i++) {
4734 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4735 return false;
4736 }
4737 }
4738 return true;
4739 }
4740
4741 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4742 if (_expensive_nodes.length() == 0) {
4743 return false;
4744 }
4745
4746 assert(OptimizeExpensiveOps, "optimization off?");
4747
4748 // Take this opportunity to remove dead nodes from the list
4749 int j = 0;
4750 for (int i = 0; i < _expensive_nodes.length(); i++) {
4751 Node* n = _expensive_nodes.at(i);
4752 if (!n->is_unreachable(igvn)) {
4753 assert(n->is_expensive(), "should be expensive");
4754 _expensive_nodes.at_put(j, n);
4755 j++;
4756 }
4757 }
4758 _expensive_nodes.trunc_to(j);
4759
4760 // Then sort the list so that similar nodes are next to each other
4761 // and check for at least two nodes of identical kind with same data
4762 // inputs.
4763 sort_expensive_nodes();
4764
4765 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4766 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4767 return true;
4768 }
4769 }
4770
4771 return false;
4772 }
4773
4774 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4775 if (_expensive_nodes.length() == 0) {
4776 return;
4777 }
4778
4779 assert(OptimizeExpensiveOps, "optimization off?");
4780
4781 // Sort to bring similar nodes next to each other and clear the
4782 // control input of nodes for which there's only a single copy.
4783 sort_expensive_nodes();
4784
4785 int j = 0;
4786 int identical = 0;
4787 int i = 0;
4788 bool modified = false;
4789 for (; i < _expensive_nodes.length()-1; i++) {
4790 assert(j <= i, "can't write beyond current index");
4791 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4792 identical++;
4793 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4794 continue;
4795 }
4796 if (identical > 0) {
4797 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4798 identical = 0;
4799 } else {
4800 Node* n = _expensive_nodes.at(i);
4801 igvn.replace_input_of(n, 0, nullptr);
4802 igvn.hash_insert(n);
4803 modified = true;
4804 }
4805 }
4806 if (identical > 0) {
4807 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4808 } else if (_expensive_nodes.length() >= 1) {
4809 Node* n = _expensive_nodes.at(i);
4810 igvn.replace_input_of(n, 0, nullptr);
4811 igvn.hash_insert(n);
4812 modified = true;
4813 }
4814 _expensive_nodes.trunc_to(j);
4815 if (modified) {
4816 igvn.optimize();
4817 }
4818 }
4819
4820 void Compile::add_expensive_node(Node * n) {
4821 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4822 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4823 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4824 if (OptimizeExpensiveOps) {
4825 _expensive_nodes.append(n);
4826 } else {
4827 // Clear control input and let IGVN optimize expensive nodes if
4828 // OptimizeExpensiveOps is off.
4829 n->set_req(0, nullptr);
4830 }
4831 }
4832
4833 /**
4834 * Track coarsened Lock and Unlock nodes.
4835 */
4836
4837 class Lock_List : public Node_List {
4838 uint _origin_cnt;
4839 public:
4840 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4841 uint origin_cnt() const { return _origin_cnt; }
4842 };
4843
4844 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4845 int length = locks.length();
4846 if (length > 0) {
4847 // Have to keep this list until locks elimination during Macro nodes elimination.
4848 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4849 AbstractLockNode* alock = locks.at(0);
4850 BoxLockNode* box = alock->box_node()->as_BoxLock();
4851 for (int i = 0; i < length; i++) {
4852 AbstractLockNode* lock = locks.at(i);
4853 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4854 locks_list->push(lock);
4855 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4856 if (this_box != box) {
4857 // Locking regions (BoxLock) could be Unbalanced here:
4858 // - its coarsened locks were eliminated in earlier
4859 // macro nodes elimination followed by loop unroll
4860 // - it is OSR locking region (no Lock node)
4861 // Preserve Unbalanced status in such cases.
4862 if (!this_box->is_unbalanced()) {
4863 this_box->set_coarsened();
4864 }
4865 if (!box->is_unbalanced()) {
4866 box->set_coarsened();
4867 }
4868 }
4869 }
4870 _coarsened_locks.append(locks_list);
4871 }
4872 }
4873
4874 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4875 int count = coarsened_count();
4876 for (int i = 0; i < count; i++) {
4877 Node_List* locks_list = _coarsened_locks.at(i);
4878 for (uint j = 0; j < locks_list->size(); j++) {
4879 Node* lock = locks_list->at(j);
4880 assert(lock->is_AbstractLock(), "sanity");
4881 if (!useful.member(lock)) {
4882 locks_list->yank(lock);
4883 }
4884 }
4885 }
4886 }
4887
4888 void Compile::remove_coarsened_lock(Node* n) {
4889 if (n->is_AbstractLock()) {
4890 int count = coarsened_count();
4891 for (int i = 0; i < count; i++) {
4892 Node_List* locks_list = _coarsened_locks.at(i);
4893 locks_list->yank(n);
4894 }
4895 }
4896 }
4897
4898 bool Compile::coarsened_locks_consistent() {
4899 int count = coarsened_count();
4900 for (int i = 0; i < count; i++) {
4901 bool unbalanced = false;
4902 bool modified = false; // track locks kind modifications
4903 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4904 uint size = locks_list->size();
4905 if (size == 0) {
4906 unbalanced = false; // All locks were eliminated - good
4907 } else if (size != locks_list->origin_cnt()) {
4908 unbalanced = true; // Some locks were removed from list
4909 } else {
4910 for (uint j = 0; j < size; j++) {
4911 Node* lock = locks_list->at(j);
4912 // All nodes in group should have the same state (modified or not)
4913 if (!lock->as_AbstractLock()->is_coarsened()) {
4914 if (j == 0) {
4915 // first on list was modified, the rest should be too for consistency
4916 modified = true;
4917 } else if (!modified) {
4918 // this lock was modified but previous locks on the list were not
4919 unbalanced = true;
4920 break;
4921 }
4922 } else if (modified) {
4923 // previous locks on list were modified but not this lock
4924 unbalanced = true;
4925 break;
4926 }
4927 }
4928 }
4929 if (unbalanced) {
4930 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4931 #ifdef ASSERT
4932 if (PrintEliminateLocks) {
4933 tty->print_cr("=== unbalanced coarsened locks ===");
4934 for (uint l = 0; l < size; l++) {
4935 locks_list->at(l)->dump();
4936 }
4937 }
4938 #endif
4939 record_failure(C2Compiler::retry_no_locks_coarsening());
4940 return false;
4941 }
4942 }
4943 return true;
4944 }
4945
4946 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4947 // locks coarsening optimization removed Lock/Unlock nodes from them.
4948 // Such regions become unbalanced because coarsening only removes part
4949 // of Lock/Unlock nodes in region. As result we can't execute other
4950 // locks elimination optimizations which assume all code paths have
4951 // corresponding pair of Lock/Unlock nodes - they are balanced.
4952 void Compile::mark_unbalanced_boxes() const {
4953 int count = coarsened_count();
4954 for (int i = 0; i < count; i++) {
4955 Node_List* locks_list = _coarsened_locks.at(i);
4956 uint size = locks_list->size();
4957 if (size > 0) {
4958 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4959 BoxLockNode* box = alock->box_node()->as_BoxLock();
4960 if (alock->is_coarsened()) {
4961 // coarsened_locks_consistent(), which is called before this method, verifies
4962 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4963 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4964 for (uint j = 1; j < size; j++) {
4965 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4966 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4967 if (box != this_box) {
4968 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4969 box->set_unbalanced();
4970 this_box->set_unbalanced();
4971 }
4972 }
4973 }
4974 }
4975 }
4976 }
4977
4978 /**
4979 * Remove the speculative part of types and clean up the graph
4980 */
4981 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4982 if (UseTypeSpeculation) {
4983 Unique_Node_List worklist;
4984 worklist.push(root());
4985 int modified = 0;
4986 // Go over all type nodes that carry a speculative type, drop the
4987 // speculative part of the type and enqueue the node for an igvn
4988 // which may optimize it out.
4989 for (uint next = 0; next < worklist.size(); ++next) {
4990 Node *n = worklist.at(next);
4991 if (n->is_Type()) {
4992 TypeNode* tn = n->as_Type();
4993 const Type* t = tn->type();
4994 const Type* t_no_spec = t->remove_speculative();
4995 if (t_no_spec != t) {
4996 bool in_hash = igvn.hash_delete(n);
4997 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4998 tn->set_type(t_no_spec);
4999 igvn.hash_insert(n);
5000 igvn._worklist.push(n); // give it a chance to go away
5001 modified++;
5002 }
5003 }
5004 // Iterate over outs - endless loops is unreachable from below
5005 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5006 Node *m = n->fast_out(i);
5007 if (not_a_node(m)) {
5008 continue;
5009 }
5010 worklist.push(m);
5011 }
5012 }
5013 // Drop the speculative part of all types in the igvn's type table
5014 igvn.remove_speculative_types();
5015 if (modified > 0) {
5016 igvn.optimize();
5017 if (failing()) return;
5018 }
5019 #ifdef ASSERT
5020 // Verify that after the IGVN is over no speculative type has resurfaced
5021 worklist.clear();
5022 worklist.push(root());
5023 for (uint next = 0; next < worklist.size(); ++next) {
5024 Node *n = worklist.at(next);
5025 const Type* t = igvn.type_or_null(n);
5026 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5027 if (n->is_Type()) {
5028 t = n->as_Type()->type();
5029 assert(t == t->remove_speculative(), "no more speculative types");
5030 }
5031 // Iterate over outs - endless loops is unreachable from below
5032 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5033 Node *m = n->fast_out(i);
5034 if (not_a_node(m)) {
5035 continue;
5036 }
5037 worklist.push(m);
5038 }
5039 }
5040 igvn.check_no_speculative_types();
5041 #endif
5042 }
5043 }
5044
5045 // Auxiliary methods to support randomized stressing/fuzzing.
5046
5047 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5048 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5049 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5050 FLAG_SET_ERGO(StressSeed, _stress_seed);
5051 } else {
5052 _stress_seed = StressSeed;
5053 }
5054 if (_log != nullptr) {
5055 _log->elem("stress_test seed='%u'", _stress_seed);
5056 }
5057 }
5058
5059 int Compile::random() {
5060 _stress_seed = os::next_random(_stress_seed);
5061 return static_cast<int>(_stress_seed);
5062 }
5063
5064 // This method can be called the arbitrary number of times, with current count
5065 // as the argument. The logic allows selecting a single candidate from the
5066 // running list of candidates as follows:
5067 // int count = 0;
5068 // Cand* selected = null;
5069 // while(cand = cand->next()) {
5070 // if (randomized_select(++count)) {
5071 // selected = cand;
5072 // }
5073 // }
5074 //
5075 // Including count equalizes the chances any candidate is "selected".
5076 // This is useful when we don't have the complete list of candidates to choose
5077 // from uniformly. In this case, we need to adjust the randomicity of the
5078 // selection, or else we will end up biasing the selection towards the latter
5079 // candidates.
5080 //
5081 // Quick back-envelope calculation shows that for the list of n candidates
5082 // the equal probability for the candidate to persist as "best" can be
5083 // achieved by replacing it with "next" k-th candidate with the probability
5084 // of 1/k. It can be easily shown that by the end of the run, the
5085 // probability for any candidate is converged to 1/n, thus giving the
5086 // uniform distribution among all the candidates.
5087 //
5088 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5089 #define RANDOMIZED_DOMAIN_POW 29
5090 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5091 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5092 bool Compile::randomized_select(int count) {
5093 assert(count > 0, "only positive");
5094 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5095 }
5096
5097 CloneMap& Compile::clone_map() { return _clone_map; }
5098 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5099
5100 void NodeCloneInfo::dump_on(outputStream* st) const {
5101 st->print(" {%d:%d} ", idx(), gen());
5102 }
5103
5104 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5105 uint64_t val = value(old->_idx);
5106 NodeCloneInfo cio(val);
5107 assert(val != 0, "old node should be in the map");
5108 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5109 insert(nnn->_idx, cin.get());
5110 #ifndef PRODUCT
5111 if (is_debug()) {
5112 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5113 }
5114 #endif
5115 }
5116
5117 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5118 NodeCloneInfo cio(value(old->_idx));
5119 if (cio.get() == 0) {
5120 cio.set(old->_idx, 0);
5121 insert(old->_idx, cio.get());
5122 #ifndef PRODUCT
5123 if (is_debug()) {
5124 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5125 }
5126 #endif
5127 }
5128 clone(old, nnn, gen);
5129 }
5130
5131 int CloneMap::max_gen() const {
5132 int g = 0;
5133 DictI di(_dict);
5134 for(; di.test(); ++di) {
5135 int t = gen(di._key);
5136 if (g < t) {
5137 g = t;
5138 #ifndef PRODUCT
5139 if (is_debug()) {
5140 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5141 }
5142 #endif
5143 }
5144 }
5145 return g;
5146 }
5147
5148 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5149 uint64_t val = value(key);
5150 if (val != 0) {
5151 NodeCloneInfo ni(val);
5152 ni.dump_on(st);
5153 }
5154 }
5155
5156 void Compile::shuffle_macro_nodes() {
5157 if (_macro_nodes.length() < 2) {
5158 return;
5159 }
5160 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5161 uint j = C->random() % (i + 1);
5162 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5163 }
5164 }
5165
5166 // Move Allocate nodes to the start of the list
5167 void Compile::sort_macro_nodes() {
5168 int count = macro_count();
5169 int allocates = 0;
5170 for (int i = 0; i < count; i++) {
5171 Node* n = macro_node(i);
5172 if (n->is_Allocate()) {
5173 if (i != allocates) {
5174 Node* tmp = macro_node(allocates);
5175 _macro_nodes.at_put(allocates, n);
5176 _macro_nodes.at_put(i, tmp);
5177 }
5178 allocates++;
5179 }
5180 }
5181 }
5182
5183 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5184 if (failing()) { return; }
5185 EventCompilerPhase event(UNTIMED);
5186 if (event.should_commit()) {
5187 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5188 }
5189 #ifndef PRODUCT
5190 ResourceMark rm;
5191 stringStream ss;
5192 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5193 int iter = ++_igv_phase_iter[cpt];
5194 if (iter > 1) {
5195 ss.print(" %d", iter);
5196 }
5197 if (n != nullptr) {
5198 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5199 }
5200
5201 const char* name = ss.as_string();
5202 if (should_print_igv(level)) {
5203 _igv_printer->print_method(name, level);
5204 }
5205 if (should_print_phase(cpt)) {
5206 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5207 }
5208 #endif
5209 C->_latest_stage_start_counter.stamp();
5210 }
5211
5212 // Only used from CompileWrapper
5213 void Compile::begin_method() {
5214 #ifndef PRODUCT
5215 if (_method != nullptr && should_print_igv(1)) {
5216 _igv_printer->begin_method();
5217 }
5218 #endif
5219 C->_latest_stage_start_counter.stamp();
5220 }
5221
5222 // Only used from CompileWrapper
5223 void Compile::end_method() {
5224 EventCompilerPhase event(UNTIMED);
5225 if (event.should_commit()) {
5226 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5227 }
5228
5229 #ifndef PRODUCT
5230 if (_method != nullptr && should_print_igv(1)) {
5231 _igv_printer->end_method();
5232 }
5233 #endif
5234 }
5235
5236 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5237 #ifndef PRODUCT
5238 if (_directive->should_print_phase(cpt)) {
5239 return true;
5240 }
5241 #endif
5242 return false;
5243 }
5244
5245 bool Compile::should_print_igv(const int level) {
5246 #ifndef PRODUCT
5247 if (PrintIdealGraphLevel < 0) { // disabled by the user
5248 return false;
5249 }
5250
5251 bool need = directive()->IGVPrintLevelOption >= level;
5252 if (need && !_igv_printer) {
5253 _igv_printer = IdealGraphPrinter::printer();
5254 _igv_printer->set_compile(this);
5255 }
5256 return need;
5257 #else
5258 return false;
5259 #endif
5260 }
5261
5262 #ifndef PRODUCT
5263 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5264 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5265
5266 // Called from debugger. Prints method to the default file with the default phase name.
5267 // This works regardless of any Ideal Graph Visualizer flags set or not.
5268 void igv_print() {
5269 Compile::current()->igv_print_method_to_file();
5270 }
5271
5272 // Same as igv_print() above but with a specified phase name.
5273 void igv_print(const char* phase_name) {
5274 Compile::current()->igv_print_method_to_file(phase_name);
5275 }
5276
5277 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5278 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5279 // This works regardless of any Ideal Graph Visualizer flags set or not.
5280 void igv_print(bool network) {
5281 if (network) {
5282 Compile::current()->igv_print_method_to_network();
5283 } else {
5284 Compile::current()->igv_print_method_to_file();
5285 }
5286 }
5287
5288 // Same as igv_print(bool network) above but with a specified phase name.
5289 void igv_print(bool network, const char* phase_name) {
5290 if (network) {
5291 Compile::current()->igv_print_method_to_network(phase_name);
5292 } else {
5293 Compile::current()->igv_print_method_to_file(phase_name);
5294 }
5295 }
5296
5297 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5298 void igv_print_default() {
5299 Compile::current()->print_method(PHASE_DEBUG, 0);
5300 }
5301
5302 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5303 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5304 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5305 void igv_append() {
5306 Compile::current()->igv_print_method_to_file("Debug", true);
5307 }
5308
5309 // Same as igv_append() above but with a specified phase name.
5310 void igv_append(const char* phase_name) {
5311 Compile::current()->igv_print_method_to_file(phase_name, true);
5312 }
5313
5314 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5315 const char* file_name = "custom_debug.xml";
5316 if (_debug_file_printer == nullptr) {
5317 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5318 } else {
5319 _debug_file_printer->update_compiled_method(C->method());
5320 }
5321 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5322 _debug_file_printer->print(phase_name, (Node*)C->root());
5323 }
5324
5325 void Compile::igv_print_method_to_network(const char* phase_name) {
5326 if (_debug_network_printer == nullptr) {
5327 _debug_network_printer = new IdealGraphPrinter(C);
5328 } else {
5329 _debug_network_printer->update_compiled_method(C->method());
5330 }
5331 tty->print_cr("Method printed over network stream to IGV");
5332 _debug_network_printer->print(phase_name, (Node*)C->root());
5333 }
5334 #endif
5335
5336 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5337 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5338 return value;
5339 }
5340 Node* result = nullptr;
5341 if (bt == T_BYTE) {
5342 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5343 result = new RShiftINode(result, phase->intcon(24));
5344 } else if (bt == T_BOOLEAN) {
5345 result = new AndINode(value, phase->intcon(0xFF));
5346 } else if (bt == T_CHAR) {
5347 result = new AndINode(value,phase->intcon(0xFFFF));
5348 } else {
5349 assert(bt == T_SHORT, "unexpected narrow type");
5350 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5351 result = new RShiftINode(result, phase->intcon(16));
5352 }
5353 if (transform_res) {
5354 result = phase->transform(result);
5355 }
5356 return result;
5357 }
5358
5359 void Compile::record_method_not_compilable_oom() {
5360 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5361 }