1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/macroAssembler.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "ci/ciReplay.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "code/exceptionHandlerTable.hpp"
30 #include "code/nmethod.hpp"
31 #include "compiler/compilationFailureInfo.hpp"
32 #include "compiler/compilationMemoryStatistic.hpp"
33 #include "compiler/compileBroker.hpp"
34 #include "compiler/compileLog.hpp"
35 #include "compiler/compilerOracle.hpp"
36 #include "compiler/compiler_globals.hpp"
37 #include "compiler/disassembler.hpp"
38 #include "compiler/oopMap.hpp"
39 #include "gc/shared/barrierSet.hpp"
40 #include "gc/shared/c2/barrierSetC2.hpp"
41 #include "jfr/jfrEvents.hpp"
42 #include "jvm_io.h"
43 #include "memory/allocation.hpp"
44 #include "memory/arena.hpp"
45 #include "memory/resourceArea.hpp"
46 #include "opto/addnode.hpp"
47 #include "opto/block.hpp"
48 #include "opto/c2compiler.hpp"
49 #include "opto/callGenerator.hpp"
50 #include "opto/callnode.hpp"
51 #include "opto/castnode.hpp"
52 #include "opto/cfgnode.hpp"
53 #include "opto/chaitin.hpp"
54 #include "opto/compile.hpp"
55 #include "opto/connode.hpp"
56 #include "opto/convertnode.hpp"
57 #include "opto/divnode.hpp"
58 #include "opto/escape.hpp"
59 #include "opto/idealGraphPrinter.hpp"
60 #include "opto/locknode.hpp"
61 #include "opto/loopnode.hpp"
62 #include "opto/machnode.hpp"
63 #include "opto/macro.hpp"
64 #include "opto/matcher.hpp"
65 #include "opto/mathexactnode.hpp"
66 #include "opto/memnode.hpp"
67 #include "opto/mulnode.hpp"
68 #include "opto/narrowptrnode.hpp"
69 #include "opto/node.hpp"
70 #include "opto/opcodes.hpp"
71 #include "opto/output.hpp"
72 #include "opto/parse.hpp"
73 #include "opto/phaseX.hpp"
74 #include "opto/rootnode.hpp"
75 #include "opto/runtime.hpp"
76 #include "opto/stringopts.hpp"
77 #include "opto/type.hpp"
78 #include "opto/vector.hpp"
79 #include "opto/vectornode.hpp"
80 #include "runtime/globals_extension.hpp"
81 #include "runtime/sharedRuntime.hpp"
82 #include "runtime/signature.hpp"
83 #include "runtime/stubRoutines.hpp"
84 #include "runtime/timer.hpp"
85 #include "utilities/align.hpp"
86 #include "utilities/copy.hpp"
87 #include "utilities/macros.hpp"
88 #include "utilities/resourceHash.hpp"
89
90 // -------------------- Compile::mach_constant_base_node -----------------------
91 // Constant table base node singleton.
92 MachConstantBaseNode* Compile::mach_constant_base_node() {
93 if (_mach_constant_base_node == nullptr) {
94 _mach_constant_base_node = new MachConstantBaseNode();
95 _mach_constant_base_node->add_req(C->root());
96 }
97 return _mach_constant_base_node;
98 }
99
100
101 /// Support for intrinsics.
102
103 // Return the index at which m must be inserted (or already exists).
104 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
105 class IntrinsicDescPair {
106 private:
107 ciMethod* _m;
108 bool _is_virtual;
109 public:
110 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
111 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
112 ciMethod* m= elt->method();
113 ciMethod* key_m = key->_m;
114 if (key_m < m) return -1;
115 else if (key_m > m) return 1;
116 else {
117 bool is_virtual = elt->is_virtual();
118 bool key_virtual = key->_is_virtual;
119 if (key_virtual < is_virtual) return -1;
120 else if (key_virtual > is_virtual) return 1;
121 else return 0;
122 }
123 }
124 };
125 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
126 #ifdef ASSERT
127 for (int i = 1; i < _intrinsics.length(); i++) {
128 CallGenerator* cg1 = _intrinsics.at(i-1);
129 CallGenerator* cg2 = _intrinsics.at(i);
130 assert(cg1->method() != cg2->method()
131 ? cg1->method() < cg2->method()
132 : cg1->is_virtual() < cg2->is_virtual(),
133 "compiler intrinsics list must stay sorted");
134 }
135 #endif
136 IntrinsicDescPair pair(m, is_virtual);
137 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
138 }
139
140 void Compile::register_intrinsic(CallGenerator* cg) {
141 bool found = false;
142 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
143 assert(!found, "registering twice");
144 _intrinsics.insert_before(index, cg);
145 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
146 }
147
148 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
149 assert(m->is_loaded(), "don't try this on unloaded methods");
150 if (_intrinsics.length() > 0) {
151 bool found = false;
152 int index = intrinsic_insertion_index(m, is_virtual, found);
153 if (found) {
154 return _intrinsics.at(index);
155 }
156 }
157 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
158 if (m->intrinsic_id() != vmIntrinsics::_none &&
159 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
160 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
161 if (cg != nullptr) {
162 // Save it for next time:
163 register_intrinsic(cg);
164 return cg;
165 } else {
166 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
167 }
168 }
169 return nullptr;
170 }
171
172 // Compile::make_vm_intrinsic is defined in library_call.cpp.
173
174 #ifndef PRODUCT
175 // statistics gathering...
176
177 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
179
180 inline int as_int(vmIntrinsics::ID id) {
181 return vmIntrinsics::as_int(id);
182 }
183
184 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
185 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
186 int oflags = _intrinsic_hist_flags[as_int(id)];
187 assert(flags != 0, "what happened?");
188 if (is_virtual) {
189 flags |= _intrinsic_virtual;
190 }
191 bool changed = (flags != oflags);
192 if ((flags & _intrinsic_worked) != 0) {
193 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
194 if (count == 1) {
195 changed = true; // first time
196 }
197 // increment the overall count also:
198 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
199 }
200 if (changed) {
201 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
202 // Something changed about the intrinsic's virtuality.
203 if ((flags & _intrinsic_virtual) != 0) {
204 // This is the first use of this intrinsic as a virtual call.
205 if (oflags != 0) {
206 // We already saw it as a non-virtual, so note both cases.
207 flags |= _intrinsic_both;
208 }
209 } else if ((oflags & _intrinsic_both) == 0) {
210 // This is the first use of this intrinsic as a non-virtual
211 flags |= _intrinsic_both;
212 }
213 }
214 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
215 }
216 // update the overall flags also:
217 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
218 return changed;
219 }
220
221 static char* format_flags(int flags, char* buf) {
222 buf[0] = 0;
223 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
224 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
225 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
226 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
227 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
228 if (buf[0] == 0) strcat(buf, ",");
229 assert(buf[0] == ',', "must be");
230 return &buf[1];
231 }
232
233 void Compile::print_intrinsic_statistics() {
234 char flagsbuf[100];
235 ttyLocker ttyl;
236 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
237 tty->print_cr("Compiler intrinsic usage:");
238 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
239 if (total == 0) total = 1; // avoid div0 in case of no successes
240 #define PRINT_STAT_LINE(name, c, f) \
241 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
242 for (auto id : EnumRange<vmIntrinsicID>{}) {
243 int flags = _intrinsic_hist_flags[as_int(id)];
244 juint count = _intrinsic_hist_count[as_int(id)];
245 if ((flags | count) != 0) {
246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
247 }
248 }
249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
250 if (xtty != nullptr) xtty->tail("statistics");
251 }
252
253 void Compile::print_statistics() {
254 { ttyLocker ttyl;
255 if (xtty != nullptr) xtty->head("statistics type='opto'");
256 Parse::print_statistics();
257 PhaseStringOpts::print_statistics();
258 PhaseCCP::print_statistics();
259 PhaseRegAlloc::print_statistics();
260 PhaseOutput::print_statistics();
261 PhasePeephole::print_statistics();
262 PhaseIdealLoop::print_statistics();
263 ConnectionGraph::print_statistics();
264 PhaseMacroExpand::print_statistics();
265 if (xtty != nullptr) xtty->tail("statistics");
266 }
267 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
268 // put this under its own <statistics> element.
269 print_intrinsic_statistics();
270 }
271 }
272 #endif //PRODUCT
273
274 void Compile::gvn_replace_by(Node* n, Node* nn) {
275 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
276 Node* use = n->last_out(i);
277 bool is_in_table = initial_gvn()->hash_delete(use);
278 uint uses_found = 0;
279 for (uint j = 0; j < use->len(); j++) {
280 if (use->in(j) == n) {
281 if (j < use->req())
282 use->set_req(j, nn);
283 else
284 use->set_prec(j, nn);
285 uses_found++;
286 }
287 }
288 if (is_in_table) {
289 // reinsert into table
290 initial_gvn()->hash_find_insert(use);
291 }
292 record_for_igvn(use);
293 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
294 i -= uses_found; // we deleted 1 or more copies of this edge
295 }
296 }
297
298
299 // Identify all nodes that are reachable from below, useful.
300 // Use breadth-first pass that records state in a Unique_Node_List,
301 // recursive traversal is slower.
302 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
303 int estimated_worklist_size = live_nodes();
304 useful.map( estimated_worklist_size, nullptr ); // preallocate space
305
306 // Initialize worklist
307 if (root() != nullptr) { useful.push(root()); }
308 // If 'top' is cached, declare it useful to preserve cached node
309 if (cached_top_node()) { useful.push(cached_top_node()); }
310
311 // Push all useful nodes onto the list, breadthfirst
312 for( uint next = 0; next < useful.size(); ++next ) {
313 assert( next < unique(), "Unique useful nodes < total nodes");
314 Node *n = useful.at(next);
315 uint max = n->len();
316 for( uint i = 0; i < max; ++i ) {
317 Node *m = n->in(i);
318 if (not_a_node(m)) continue;
319 useful.push(m);
320 }
321 }
322 }
323
324 // Update dead_node_list with any missing dead nodes using useful
325 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
326 void Compile::update_dead_node_list(Unique_Node_List &useful) {
327 uint max_idx = unique();
328 VectorSet& useful_node_set = useful.member_set();
329
330 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
331 // If node with index node_idx is not in useful set,
332 // mark it as dead in dead node list.
333 if (!useful_node_set.test(node_idx)) {
334 record_dead_node(node_idx);
335 }
336 }
337 }
338
339 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
340 int shift = 0;
341 for (int i = 0; i < inlines->length(); i++) {
342 CallGenerator* cg = inlines->at(i);
343 if (useful.member(cg->call_node())) {
344 if (shift > 0) {
345 inlines->at_put(i - shift, cg);
346 }
347 } else {
348 shift++; // skip over the dead element
349 }
350 }
351 if (shift > 0) {
352 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
353 }
354 }
355
356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
357 assert(dead != nullptr && dead->is_Call(), "sanity");
358 int found = 0;
359 for (int i = 0; i < inlines->length(); i++) {
360 if (inlines->at(i)->call_node() == dead) {
361 inlines->remove_at(i);
362 found++;
363 NOT_DEBUG( break; ) // elements are unique, so exit early
364 }
365 }
366 assert(found <= 1, "not unique");
367 }
368
369 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
370 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
371 for (int i = node_list.length() - 1; i >= 0; i--) {
372 N* node = node_list.at(i);
373 if (!useful.member(node)) {
374 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
375 }
376 }
377 }
378
379 void Compile::remove_useless_node(Node* dead) {
380 remove_modified_node(dead);
381
382 // Constant node that has no out-edges and has only one in-edge from
383 // root is usually dead. However, sometimes reshaping walk makes
384 // it reachable by adding use edges. So, we will NOT count Con nodes
385 // as dead to be conservative about the dead node count at any
386 // given time.
387 if (!dead->is_Con()) {
388 record_dead_node(dead->_idx);
389 }
390 if (dead->is_macro()) {
391 remove_macro_node(dead);
392 }
393 if (dead->is_expensive()) {
394 remove_expensive_node(dead);
395 }
396 if (dead->is_OpaqueTemplateAssertionPredicate()) {
397 remove_template_assertion_predicate_opaq(dead);
398 }
399 if (dead->is_ParsePredicate()) {
400 remove_parse_predicate(dead->as_ParsePredicate());
401 }
402 if (dead->for_post_loop_opts_igvn()) {
403 remove_from_post_loop_opts_igvn(dead);
404 }
405 if (dead->for_merge_stores_igvn()) {
406 remove_from_merge_stores_igvn(dead);
407 }
408 if (dead->is_Call()) {
409 remove_useless_late_inlines( &_late_inlines, dead);
410 remove_useless_late_inlines( &_string_late_inlines, dead);
411 remove_useless_late_inlines( &_boxing_late_inlines, dead);
412 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
413
414 if (dead->is_CallStaticJava()) {
415 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
416 }
417 }
418 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
419 bs->unregister_potential_barrier_node(dead);
420 }
421
422 // Disconnect all useless nodes by disconnecting those at the boundary.
423 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
424 uint next = 0;
425 while (next < useful.size()) {
426 Node *n = useful.at(next++);
427 if (n->is_SafePoint()) {
428 // We're done with a parsing phase. Replaced nodes are not valid
429 // beyond that point.
430 n->as_SafePoint()->delete_replaced_nodes();
431 }
432 // Use raw traversal of out edges since this code removes out edges
433 int max = n->outcnt();
434 for (int j = 0; j < max; ++j) {
435 Node* child = n->raw_out(j);
436 if (!useful.member(child)) {
437 assert(!child->is_top() || child != top(),
438 "If top is cached in Compile object it is in useful list");
439 // Only need to remove this out-edge to the useless node
440 n->raw_del_out(j);
441 --j;
442 --max;
443 if (child->is_data_proj_of_pure_function(n)) {
444 worklist.push(n);
445 }
446 }
447 }
448 if (n->outcnt() == 1 && n->has_special_unique_user()) {
449 assert(useful.member(n->unique_out()), "do not push a useless node");
450 worklist.push(n->unique_out());
451 }
452 }
453
454 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
455 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
456 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
457 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
458 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
459 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
460 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
461 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
462 #ifdef ASSERT
463 if (_modified_nodes != nullptr) {
464 _modified_nodes->remove_useless_nodes(useful.member_set());
465 }
466 #endif
467
468 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
469 bs->eliminate_useless_gc_barriers(useful, this);
470 // clean up the late inline lists
471 remove_useless_late_inlines( &_late_inlines, useful);
472 remove_useless_late_inlines( &_string_late_inlines, useful);
473 remove_useless_late_inlines( &_boxing_late_inlines, useful);
474 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
475 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
476 }
477
478 // ============================================================================
479 //------------------------------CompileWrapper---------------------------------
480 class CompileWrapper : public StackObj {
481 Compile *const _compile;
482 public:
483 CompileWrapper(Compile* compile);
484
485 ~CompileWrapper();
486 };
487
488 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
489 // the Compile* pointer is stored in the current ciEnv:
490 ciEnv* env = compile->env();
491 assert(env == ciEnv::current(), "must already be a ciEnv active");
492 assert(env->compiler_data() == nullptr, "compile already active?");
493 env->set_compiler_data(compile);
494 assert(compile == Compile::current(), "sanity");
495
496 compile->set_type_dict(nullptr);
497 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
498 compile->clone_map().set_clone_idx(0);
499 compile->set_type_last_size(0);
500 compile->set_last_tf(nullptr, nullptr);
501 compile->set_indexSet_arena(nullptr);
502 compile->set_indexSet_free_block_list(nullptr);
503 compile->init_type_arena();
504 Type::Initialize(compile);
505 _compile->begin_method();
506 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
507 }
508 CompileWrapper::~CompileWrapper() {
509 // simulate crash during compilation
510 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
511
512 _compile->end_method();
513 _compile->env()->set_compiler_data(nullptr);
514 }
515
516
517 //----------------------------print_compile_messages---------------------------
518 void Compile::print_compile_messages() {
519 #ifndef PRODUCT
520 // Check if recompiling
521 if (!subsume_loads() && PrintOpto) {
522 // Recompiling without allowing machine instructions to subsume loads
523 tty->print_cr("*********************************************************");
524 tty->print_cr("** Bailout: Recompile without subsuming loads **");
525 tty->print_cr("*********************************************************");
526 }
527 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
528 // Recompiling without escape analysis
529 tty->print_cr("*********************************************************");
530 tty->print_cr("** Bailout: Recompile without escape analysis **");
531 tty->print_cr("*********************************************************");
532 }
533 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
534 // Recompiling without iterative escape analysis
535 tty->print_cr("*********************************************************");
536 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
537 tty->print_cr("*********************************************************");
538 }
539 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
540 // Recompiling without reducing allocation merges
541 tty->print_cr("*********************************************************");
542 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
543 tty->print_cr("*********************************************************");
544 }
545 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
546 // Recompiling without boxing elimination
547 tty->print_cr("*********************************************************");
548 tty->print_cr("** Bailout: Recompile without boxing elimination **");
549 tty->print_cr("*********************************************************");
550 }
551 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
552 // Recompiling without locks coarsening
553 tty->print_cr("*********************************************************");
554 tty->print_cr("** Bailout: Recompile without locks coarsening **");
555 tty->print_cr("*********************************************************");
556 }
557 if (env()->break_at_compile()) {
558 // Open the debugger when compiling this method.
559 tty->print("### Breaking when compiling: ");
560 method()->print_short_name();
561 tty->cr();
562 BREAKPOINT;
563 }
564
565 if( PrintOpto ) {
566 if (is_osr_compilation()) {
567 tty->print("[OSR]%3d", _compile_id);
568 } else {
569 tty->print("%3d", _compile_id);
570 }
571 }
572 #endif
573 }
574
575 #ifndef PRODUCT
576 void Compile::print_ideal_ir(const char* phase_name) {
577 // keep the following output all in one block
578 // This output goes directly to the tty, not the compiler log.
579 // To enable tools to match it up with the compilation activity,
580 // be sure to tag this tty output with the compile ID.
581
582 // Node dumping can cause a safepoint, which can break the tty lock.
583 // Buffer all node dumps, so that all safepoints happen before we lock.
584 ResourceMark rm;
585 stringStream ss;
586
587 if (_output == nullptr) {
588 ss.print_cr("AFTER: %s", phase_name);
589 // Print out all nodes in ascending order of index.
590 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
591 } else {
592 // Dump the node blockwise if we have a scheduling
593 _output->print_scheduling(&ss);
594 }
595
596 // Check that the lock is not broken by a safepoint.
597 NoSafepointVerifier nsv;
598 ttyLocker ttyl;
599 if (xtty != nullptr) {
600 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
601 compile_id(),
602 is_osr_compilation() ? " compile_kind='osr'" : "",
603 phase_name);
604 }
605
606 tty->print("%s", ss.as_string());
607
608 if (xtty != nullptr) {
609 xtty->tail("ideal");
610 }
611 }
612 #endif
613
614 // ============================================================================
615 //------------------------------Compile standard-------------------------------
616
617 // Compile a method. entry_bci is -1 for normal compilations and indicates
618 // the continuation bci for on stack replacement.
619
620
621 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
622 Options options, DirectiveSet* directive)
623 : Phase(Compiler),
624 _compile_id(ci_env->compile_id()),
625 _options(options),
626 _method(target),
627 _entry_bci(osr_bci),
628 _ilt(nullptr),
629 _stub_function(nullptr),
630 _stub_name(nullptr),
631 _stub_entry_point(nullptr),
632 _max_node_limit(MaxNodeLimit),
633 _post_loop_opts_phase(false),
634 _merge_stores_phase(false),
635 _allow_macro_nodes(true),
636 _inlining_progress(false),
637 _inlining_incrementally(false),
638 _do_cleanup(false),
639 _has_reserved_stack_access(target->has_reserved_stack_access()),
640 #ifndef PRODUCT
641 _igv_idx(0),
642 _trace_opto_output(directive->TraceOptoOutputOption),
643 #endif
644 _has_method_handle_invokes(false),
645 _clinit_barrier_on_entry(false),
646 _has_clinit_barriers(false),
647 _stress_seed(0),
648 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
649 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
650 _env(ci_env),
651 _directive(directive),
652 _log(ci_env->log()),
653 _first_failure_details(nullptr),
654 _intrinsics(comp_arena(), 0, 0, nullptr),
655 _macro_nodes(comp_arena(), 8, 0, nullptr),
656 _parse_predicates(comp_arena(), 8, 0, nullptr),
657 _template_assertion_predicate_opaqs(comp_arena(), 8, 0, nullptr),
658 _expensive_nodes(comp_arena(), 8, 0, nullptr),
659 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
660 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
661 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
662 _coarsened_locks(comp_arena(), 8, 0, nullptr),
663 _congraph(nullptr),
664 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
665 _unique(0),
666 _dead_node_count(0),
667 _dead_node_list(comp_arena()),
668 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
669 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
670 _node_arena(&_node_arena_one),
671 _mach_constant_base_node(nullptr),
672 _Compile_types(mtCompiler, Arena::Tag::tag_type),
673 _initial_gvn(nullptr),
674 _igvn_worklist(nullptr),
675 _types(nullptr),
676 _node_hash(nullptr),
677 _late_inlines(comp_arena(), 2, 0, nullptr),
678 _string_late_inlines(comp_arena(), 2, 0, nullptr),
679 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
680 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
681 _late_inlines_pos(0),
682 _number_of_mh_late_inlines(0),
683 _oom(false),
684 _replay_inline_data(nullptr),
685 _inline_printer(this),
686 _java_calls(0),
687 _inner_loops(0),
688 _interpreter_frame_size(0),
689 _output(nullptr)
690 #ifndef PRODUCT
691 ,
692 _in_dump_cnt(0)
693 #endif
694 {
695 C = this;
696 CompileWrapper cw(this);
697
698 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
699 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
700
701 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
702 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
703 // We can always print a disassembly, either abstract (hex dump) or
704 // with the help of a suitable hsdis library. Thus, we should not
705 // couple print_assembly and print_opto_assembly controls.
706 // But: always print opto and regular assembly on compile command 'print'.
707 bool print_assembly = directive->PrintAssemblyOption;
708 set_print_assembly(print_opto_assembly || print_assembly);
709 #else
710 set_print_assembly(false); // must initialize.
711 #endif
712
713 #ifndef PRODUCT
714 set_parsed_irreducible_loop(false);
715 #endif
716
717 if (directive->ReplayInlineOption) {
718 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
719 }
720 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
721 set_print_intrinsics(directive->PrintIntrinsicsOption);
722 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
723
724 if (ProfileTraps) {
725 // Make sure the method being compiled gets its own MDO,
726 // so we can at least track the decompile_count().
727 method()->ensure_method_data();
728 }
729
730 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
731 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout) {
732 initialize_stress_seed(directive);
733 }
734
735 Init(/*do_aliasing=*/ true);
736
737 print_compile_messages();
738
739 _ilt = InlineTree::build_inline_tree_root();
740
741 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
742 assert(num_alias_types() >= AliasIdxRaw, "");
743
744 #define MINIMUM_NODE_HASH 1023
745
746 // GVN that will be run immediately on new nodes
747 uint estimated_size = method()->code_size()*4+64;
748 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
749 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
750 _types = new (comp_arena()) Type_Array(comp_arena());
751 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
752 PhaseGVN gvn;
753 set_initial_gvn(&gvn);
754
755 { // Scope for timing the parser
756 TracePhase tp(_t_parser);
757
758 // Put top into the hash table ASAP.
759 initial_gvn()->transform(top());
760
761 // Set up tf(), start(), and find a CallGenerator.
762 CallGenerator* cg = nullptr;
763 if (is_osr_compilation()) {
764 const TypeTuple *domain = StartOSRNode::osr_domain();
765 const TypeTuple *range = TypeTuple::make_range(method()->signature());
766 init_tf(TypeFunc::make(domain, range));
767 StartNode* s = new StartOSRNode(root(), domain);
768 initial_gvn()->set_type_bottom(s);
769 verify_start(s);
770 cg = CallGenerator::for_osr(method(), entry_bci());
771 } else {
772 // Normal case.
773 init_tf(TypeFunc::make(method()));
774 StartNode* s = new StartNode(root(), tf()->domain());
775 initial_gvn()->set_type_bottom(s);
776 verify_start(s);
777 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
778 // With java.lang.ref.reference.get() we must go through the
779 // intrinsic - even when get() is the root
780 // method of the compile - so that, if necessary, the value in
781 // the referent field of the reference object gets recorded by
782 // the pre-barrier code.
783 cg = find_intrinsic(method(), false);
784 }
785 if (cg == nullptr) {
786 float past_uses = method()->interpreter_invocation_count();
787 float expected_uses = past_uses;
788 cg = CallGenerator::for_inline(method(), expected_uses);
789 }
790 }
791 if (failing()) return;
792 if (cg == nullptr) {
793 const char* reason = InlineTree::check_can_parse(method());
794 assert(reason != nullptr, "expect reason for parse failure");
795 stringStream ss;
796 ss.print("cannot parse method: %s", reason);
797 record_method_not_compilable(ss.as_string());
798 return;
799 }
800
801 gvn.set_type(root(), root()->bottom_type());
802
803 JVMState* jvms = build_start_state(start(), tf());
804 if ((jvms = cg->generate(jvms)) == nullptr) {
805 assert(failure_reason() != nullptr, "expect reason for parse failure");
806 stringStream ss;
807 ss.print("method parse failed: %s", failure_reason());
808 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
809 return;
810 }
811 GraphKit kit(jvms);
812
813 if (!kit.stopped()) {
814 // Accept return values, and transfer control we know not where.
815 // This is done by a special, unique ReturnNode bound to root.
816 return_values(kit.jvms());
817 }
818
819 if (kit.has_exceptions()) {
820 // Any exceptions that escape from this call must be rethrown
821 // to whatever caller is dynamically above us on the stack.
822 // This is done by a special, unique RethrowNode bound to root.
823 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
824 }
825
826 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
827
828 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
829 inline_string_calls(true);
830 }
831
832 if (failing()) return;
833
834 // Remove clutter produced by parsing.
835 if (!failing()) {
836 ResourceMark rm;
837 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
838 }
839 }
840
841 // Note: Large methods are capped off in do_one_bytecode().
842 if (failing()) return;
843
844 // After parsing, node notes are no longer automagic.
845 // They must be propagated by register_new_node_with_optimizer(),
846 // clone(), or the like.
847 set_default_node_notes(nullptr);
848
849 #ifndef PRODUCT
850 if (should_print_igv(1)) {
851 _igv_printer->print_inlining();
852 }
853 #endif
854
855 if (failing()) return;
856 NOT_PRODUCT( verify_graph_edges(); )
857
858 // Now optimize
859 Optimize();
860 if (failing()) return;
861 NOT_PRODUCT( verify_graph_edges(); )
862
863 #ifndef PRODUCT
864 if (should_print_ideal()) {
865 print_ideal_ir("print_ideal");
866 }
867 #endif
868
869 #ifdef ASSERT
870 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
871 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
872 #endif
873
874 // Dump compilation data to replay it.
875 if (directive->DumpReplayOption) {
876 env()->dump_replay_data(_compile_id);
877 }
878 if (directive->DumpInlineOption && (ilt() != nullptr)) {
879 env()->dump_inline_data(_compile_id);
880 }
881
882 // Now that we know the size of all the monitors we can add a fixed slot
883 // for the original deopt pc.
884 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
885 set_fixed_slots(next_slot);
886
887 // Compute when to use implicit null checks. Used by matching trap based
888 // nodes and NullCheck optimization.
889 set_allowed_deopt_reasons();
890
891 // Now generate code
892 Code_Gen();
893 }
894
895 //------------------------------Compile----------------------------------------
896 // Compile a runtime stub
897 Compile::Compile(ciEnv* ci_env,
898 TypeFunc_generator generator,
899 address stub_function,
900 const char* stub_name,
901 int is_fancy_jump,
902 bool pass_tls,
903 bool return_pc,
904 DirectiveSet* directive)
905 : Phase(Compiler),
906 _compile_id(0),
907 _options(Options::for_runtime_stub()),
908 _method(nullptr),
909 _entry_bci(InvocationEntryBci),
910 _stub_function(stub_function),
911 _stub_name(stub_name),
912 _stub_entry_point(nullptr),
913 _max_node_limit(MaxNodeLimit),
914 _post_loop_opts_phase(false),
915 _merge_stores_phase(false),
916 _allow_macro_nodes(true),
917 _inlining_progress(false),
918 _inlining_incrementally(false),
919 _has_reserved_stack_access(false),
920 #ifndef PRODUCT
921 _igv_idx(0),
922 _trace_opto_output(directive->TraceOptoOutputOption),
923 #endif
924 _has_method_handle_invokes(false),
925 _clinit_barrier_on_entry(false),
926 _has_clinit_barriers(false),
927 _stress_seed(0),
928 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
929 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
930 _env(ci_env),
931 _directive(directive),
932 _log(ci_env->log()),
933 _first_failure_details(nullptr),
934 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
935 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
936 _congraph(nullptr),
937 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
938 _unique(0),
939 _dead_node_count(0),
940 _dead_node_list(comp_arena()),
941 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
942 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
943 _node_arena(&_node_arena_one),
944 _mach_constant_base_node(nullptr),
945 _Compile_types(mtCompiler, Arena::Tag::tag_type),
946 _initial_gvn(nullptr),
947 _igvn_worklist(nullptr),
948 _types(nullptr),
949 _node_hash(nullptr),
950 _number_of_mh_late_inlines(0),
951 _oom(false),
952 _replay_inline_data(nullptr),
953 _inline_printer(this),
954 _java_calls(0),
955 _inner_loops(0),
956 _interpreter_frame_size(0),
957 _output(nullptr),
958 #ifndef PRODUCT
959 _in_dump_cnt(0),
960 #endif
961 _allowed_reasons(0) {
962 C = this;
963
964 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
965 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
966
967 #ifndef PRODUCT
968 set_print_assembly(PrintFrameConverterAssembly);
969 set_parsed_irreducible_loop(false);
970 #else
971 set_print_assembly(false); // Must initialize.
972 #endif
973 set_has_irreducible_loop(false); // no loops
974
975 CompileWrapper cw(this);
976 Init(/*do_aliasing=*/ false);
977 init_tf((*generator)());
978
979 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
980 _types = new (comp_arena()) Type_Array(comp_arena());
981 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
982
983 if (StressLCM || StressGCM || StressBailout) {
984 initialize_stress_seed(directive);
985 }
986
987 {
988 PhaseGVN gvn;
989 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
990 gvn.transform(top());
991
992 GraphKit kit;
993 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
994 }
995
996 NOT_PRODUCT( verify_graph_edges(); )
997
998 Code_Gen();
999 }
1000
1001 Compile::~Compile() {
1002 delete _first_failure_details;
1003 };
1004
1005 //------------------------------Init-------------------------------------------
1006 // Prepare for a single compilation
1007 void Compile::Init(bool aliasing) {
1008 _do_aliasing = aliasing;
1009 _unique = 0;
1010 _regalloc = nullptr;
1011
1012 _tf = nullptr; // filled in later
1013 _top = nullptr; // cached later
1014 _matcher = nullptr; // filled in later
1015 _cfg = nullptr; // filled in later
1016
1017 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1018
1019 _node_note_array = nullptr;
1020 _default_node_notes = nullptr;
1021 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1022
1023 _immutable_memory = nullptr; // filled in at first inquiry
1024
1025 #ifdef ASSERT
1026 _phase_optimize_finished = false;
1027 _phase_verify_ideal_loop = false;
1028 _exception_backedge = false;
1029 _type_verify = nullptr;
1030 #endif
1031
1032 // Globally visible Nodes
1033 // First set TOP to null to give safe behavior during creation of RootNode
1034 set_cached_top_node(nullptr);
1035 set_root(new RootNode());
1036 // Now that you have a Root to point to, create the real TOP
1037 set_cached_top_node( new ConNode(Type::TOP) );
1038 set_recent_alloc(nullptr, nullptr);
1039
1040 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1041 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1042 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1043 env()->set_dependencies(new Dependencies(env()));
1044
1045 _fixed_slots = 0;
1046 set_has_split_ifs(false);
1047 set_has_loops(false); // first approximation
1048 set_has_stringbuilder(false);
1049 set_has_boxed_value(false);
1050 _trap_can_recompile = false; // no traps emitted yet
1051 _major_progress = true; // start out assuming good things will happen
1052 set_has_unsafe_access(false);
1053 set_max_vector_size(0);
1054 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1055 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1056 set_decompile_count(0);
1057
1058 #ifndef PRODUCT
1059 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1060 #endif
1061
1062 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1063 _loop_opts_cnt = LoopOptsCount;
1064 set_do_inlining(Inline);
1065 set_max_inline_size(MaxInlineSize);
1066 set_freq_inline_size(FreqInlineSize);
1067 set_do_scheduling(OptoScheduling);
1068
1069 set_do_vector_loop(false);
1070 set_has_monitors(false);
1071 set_has_scoped_access(false);
1072
1073 if (AllowVectorizeOnDemand) {
1074 if (has_method() && _directive->VectorizeOption) {
1075 set_do_vector_loop(true);
1076 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());})
1077 } else if (has_method() && method()->name() != nullptr &&
1078 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1079 set_do_vector_loop(true);
1080 }
1081 }
1082 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1083 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());})
1084
1085 _max_node_limit = _directive->MaxNodeLimitOption;
1086
1087 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() &&
1088 (method()->needs_clinit_barrier() || (do_clinit_barriers() && method()->is_static()))) {
1089 set_clinit_barrier_on_entry(true);
1090 if (do_clinit_barriers()) {
1091 set_has_clinit_barriers(true); // Entry clinit barrier is in prolog code.
1092 }
1093 }
1094 if (debug_info()->recording_non_safepoints()) {
1095 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1096 (comp_arena(), 8, 0, nullptr));
1097 set_default_node_notes(Node_Notes::make(this));
1098 }
1099
1100 const int grow_ats = 16;
1101 _max_alias_types = grow_ats;
1102 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1103 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1104 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1105 {
1106 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1107 }
1108 // Initialize the first few types.
1109 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1110 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1111 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1112 _num_alias_types = AliasIdxRaw+1;
1113 // Zero out the alias type cache.
1114 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1115 // A null adr_type hits in the cache right away. Preload the right answer.
1116 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1117 }
1118
1119 #ifdef ASSERT
1120 // Verify that the current StartNode is valid.
1121 void Compile::verify_start(StartNode* s) const {
1122 assert(failing_internal() || s == start(), "should be StartNode");
1123 }
1124 #endif
1125
1126 /**
1127 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1128 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1129 * the ideal graph.
1130 */
1131 StartNode* Compile::start() const {
1132 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1133 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1134 Node* start = root()->fast_out(i);
1135 if (start->is_Start()) {
1136 return start->as_Start();
1137 }
1138 }
1139 fatal("Did not find Start node!");
1140 return nullptr;
1141 }
1142
1143 //-------------------------------immutable_memory-------------------------------------
1144 // Access immutable memory
1145 Node* Compile::immutable_memory() {
1146 if (_immutable_memory != nullptr) {
1147 return _immutable_memory;
1148 }
1149 StartNode* s = start();
1150 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1151 Node *p = s->fast_out(i);
1152 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1153 _immutable_memory = p;
1154 return _immutable_memory;
1155 }
1156 }
1157 ShouldNotReachHere();
1158 return nullptr;
1159 }
1160
1161 //----------------------set_cached_top_node------------------------------------
1162 // Install the cached top node, and make sure Node::is_top works correctly.
1163 void Compile::set_cached_top_node(Node* tn) {
1164 if (tn != nullptr) verify_top(tn);
1165 Node* old_top = _top;
1166 _top = tn;
1167 // Calling Node::setup_is_top allows the nodes the chance to adjust
1168 // their _out arrays.
1169 if (_top != nullptr) _top->setup_is_top();
1170 if (old_top != nullptr) old_top->setup_is_top();
1171 assert(_top == nullptr || top()->is_top(), "");
1172 }
1173
1174 #ifdef ASSERT
1175 uint Compile::count_live_nodes_by_graph_walk() {
1176 Unique_Node_List useful(comp_arena());
1177 // Get useful node list by walking the graph.
1178 identify_useful_nodes(useful);
1179 return useful.size();
1180 }
1181
1182 void Compile::print_missing_nodes() {
1183
1184 // Return if CompileLog is null and PrintIdealNodeCount is false.
1185 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1186 return;
1187 }
1188
1189 // This is an expensive function. It is executed only when the user
1190 // specifies VerifyIdealNodeCount option or otherwise knows the
1191 // additional work that needs to be done to identify reachable nodes
1192 // by walking the flow graph and find the missing ones using
1193 // _dead_node_list.
1194
1195 Unique_Node_List useful(comp_arena());
1196 // Get useful node list by walking the graph.
1197 identify_useful_nodes(useful);
1198
1199 uint l_nodes = C->live_nodes();
1200 uint l_nodes_by_walk = useful.size();
1201
1202 if (l_nodes != l_nodes_by_walk) {
1203 if (_log != nullptr) {
1204 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1205 _log->stamp();
1206 _log->end_head();
1207 }
1208 VectorSet& useful_member_set = useful.member_set();
1209 int last_idx = l_nodes_by_walk;
1210 for (int i = 0; i < last_idx; i++) {
1211 if (useful_member_set.test(i)) {
1212 if (_dead_node_list.test(i)) {
1213 if (_log != nullptr) {
1214 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1215 }
1216 if (PrintIdealNodeCount) {
1217 // Print the log message to tty
1218 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1219 useful.at(i)->dump();
1220 }
1221 }
1222 }
1223 else if (! _dead_node_list.test(i)) {
1224 if (_log != nullptr) {
1225 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1226 }
1227 if (PrintIdealNodeCount) {
1228 // Print the log message to tty
1229 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1230 }
1231 }
1232 }
1233 if (_log != nullptr) {
1234 _log->tail("mismatched_nodes");
1235 }
1236 }
1237 }
1238 void Compile::record_modified_node(Node* n) {
1239 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1240 _modified_nodes->push(n);
1241 }
1242 }
1243
1244 void Compile::remove_modified_node(Node* n) {
1245 if (_modified_nodes != nullptr) {
1246 _modified_nodes->remove(n);
1247 }
1248 }
1249 #endif
1250
1251 #ifndef PRODUCT
1252 void Compile::verify_top(Node* tn) const {
1253 if (tn != nullptr) {
1254 assert(tn->is_Con(), "top node must be a constant");
1255 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1256 assert(tn->in(0) != nullptr, "must have live top node");
1257 }
1258 }
1259 #endif
1260
1261
1262 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1263
1264 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1265 guarantee(arr != nullptr, "");
1266 int num_blocks = arr->length();
1267 if (grow_by < num_blocks) grow_by = num_blocks;
1268 int num_notes = grow_by * _node_notes_block_size;
1269 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1270 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1271 while (num_notes > 0) {
1272 arr->append(notes);
1273 notes += _node_notes_block_size;
1274 num_notes -= _node_notes_block_size;
1275 }
1276 assert(num_notes == 0, "exact multiple, please");
1277 }
1278
1279 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1280 if (source == nullptr || dest == nullptr) return false;
1281
1282 if (dest->is_Con())
1283 return false; // Do not push debug info onto constants.
1284
1285 #ifdef ASSERT
1286 // Leave a bread crumb trail pointing to the original node:
1287 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1288 dest->set_debug_orig(source);
1289 }
1290 #endif
1291
1292 if (node_note_array() == nullptr)
1293 return false; // Not collecting any notes now.
1294
1295 // This is a copy onto a pre-existing node, which may already have notes.
1296 // If both nodes have notes, do not overwrite any pre-existing notes.
1297 Node_Notes* source_notes = node_notes_at(source->_idx);
1298 if (source_notes == nullptr || source_notes->is_clear()) return false;
1299 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1300 if (dest_notes == nullptr || dest_notes->is_clear()) {
1301 return set_node_notes_at(dest->_idx, source_notes);
1302 }
1303
1304 Node_Notes merged_notes = (*source_notes);
1305 // The order of operations here ensures that dest notes will win...
1306 merged_notes.update_from(dest_notes);
1307 return set_node_notes_at(dest->_idx, &merged_notes);
1308 }
1309
1310
1311 //--------------------------allow_range_check_smearing-------------------------
1312 // Gating condition for coalescing similar range checks.
1313 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1314 // single covering check that is at least as strong as any of them.
1315 // If the optimization succeeds, the simplified (strengthened) range check
1316 // will always succeed. If it fails, we will deopt, and then give up
1317 // on the optimization.
1318 bool Compile::allow_range_check_smearing() const {
1319 // If this method has already thrown a range-check,
1320 // assume it was because we already tried range smearing
1321 // and it failed.
1322 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1323 return !already_trapped;
1324 }
1325
1326
1327 //------------------------------flatten_alias_type-----------------------------
1328 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1329 assert(do_aliasing(), "Aliasing should be enabled");
1330 int offset = tj->offset();
1331 TypePtr::PTR ptr = tj->ptr();
1332
1333 // Known instance (scalarizable allocation) alias only with itself.
1334 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1335 tj->is_oopptr()->is_known_instance();
1336
1337 // Process weird unsafe references.
1338 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1339 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1340 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1341 tj = TypeOopPtr::BOTTOM;
1342 ptr = tj->ptr();
1343 offset = tj->offset();
1344 }
1345
1346 // Array pointers need some flattening
1347 const TypeAryPtr* ta = tj->isa_aryptr();
1348 if (ta && ta->is_stable()) {
1349 // Erase stability property for alias analysis.
1350 tj = ta = ta->cast_to_stable(false);
1351 }
1352 if( ta && is_known_inst ) {
1353 if ( offset != Type::OffsetBot &&
1354 offset > arrayOopDesc::length_offset_in_bytes() ) {
1355 offset = Type::OffsetBot; // Flatten constant access into array body only
1356 tj = ta = ta->
1357 remove_speculative()->
1358 cast_to_ptr_type(ptr)->
1359 with_offset(offset);
1360 }
1361 } else if (ta) {
1362 // For arrays indexed by constant indices, we flatten the alias
1363 // space to include all of the array body. Only the header, klass
1364 // and array length can be accessed un-aliased.
1365 if( offset != Type::OffsetBot ) {
1366 if( ta->const_oop() ) { // MethodData* or Method*
1367 offset = Type::OffsetBot; // Flatten constant access into array body
1368 tj = ta = ta->
1369 remove_speculative()->
1370 cast_to_ptr_type(ptr)->
1371 cast_to_exactness(false)->
1372 with_offset(offset);
1373 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1374 // range is OK as-is.
1375 tj = ta = TypeAryPtr::RANGE;
1376 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1377 tj = TypeInstPtr::KLASS; // all klass loads look alike
1378 ta = TypeAryPtr::RANGE; // generic ignored junk
1379 ptr = TypePtr::BotPTR;
1380 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1381 tj = TypeInstPtr::MARK;
1382 ta = TypeAryPtr::RANGE; // generic ignored junk
1383 ptr = TypePtr::BotPTR;
1384 } else { // Random constant offset into array body
1385 offset = Type::OffsetBot; // Flatten constant access into array body
1386 tj = ta = ta->
1387 remove_speculative()->
1388 cast_to_ptr_type(ptr)->
1389 cast_to_exactness(false)->
1390 with_offset(offset);
1391 }
1392 }
1393 // Arrays of fixed size alias with arrays of unknown size.
1394 if (ta->size() != TypeInt::POS) {
1395 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1396 tj = ta = ta->
1397 remove_speculative()->
1398 cast_to_ptr_type(ptr)->
1399 with_ary(tary)->
1400 cast_to_exactness(false);
1401 }
1402 // Arrays of known objects become arrays of unknown objects.
1403 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1404 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1405 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1406 }
1407 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1408 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1409 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1410 }
1411 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1412 // cannot be distinguished by bytecode alone.
1413 if (ta->elem() == TypeInt::BOOL) {
1414 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1415 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1416 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1417 }
1418 // During the 2nd round of IterGVN, NotNull castings are removed.
1419 // Make sure the Bottom and NotNull variants alias the same.
1420 // Also, make sure exact and non-exact variants alias the same.
1421 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1422 tj = ta = ta->
1423 remove_speculative()->
1424 cast_to_ptr_type(TypePtr::BotPTR)->
1425 cast_to_exactness(false)->
1426 with_offset(offset);
1427 }
1428 }
1429
1430 // Oop pointers need some flattening
1431 const TypeInstPtr *to = tj->isa_instptr();
1432 if (to && to != TypeOopPtr::BOTTOM) {
1433 ciInstanceKlass* ik = to->instance_klass();
1434 if( ptr == TypePtr::Constant ) {
1435 if (ik != ciEnv::current()->Class_klass() ||
1436 offset < ik->layout_helper_size_in_bytes()) {
1437 // No constant oop pointers (such as Strings); they alias with
1438 // unknown strings.
1439 assert(!is_known_inst, "not scalarizable allocation");
1440 tj = to = to->
1441 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1442 remove_speculative()->
1443 cast_to_ptr_type(TypePtr::BotPTR)->
1444 cast_to_exactness(false);
1445 }
1446 } else if( is_known_inst ) {
1447 tj = to; // Keep NotNull and klass_is_exact for instance type
1448 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1449 // During the 2nd round of IterGVN, NotNull castings are removed.
1450 // Make sure the Bottom and NotNull variants alias the same.
1451 // Also, make sure exact and non-exact variants alias the same.
1452 tj = to = to->
1453 remove_speculative()->
1454 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1455 cast_to_ptr_type(TypePtr::BotPTR)->
1456 cast_to_exactness(false);
1457 }
1458 if (to->speculative() != nullptr) {
1459 tj = to = to->remove_speculative();
1460 }
1461 // Canonicalize the holder of this field
1462 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1463 // First handle header references such as a LoadKlassNode, even if the
1464 // object's klass is unloaded at compile time (4965979).
1465 if (!is_known_inst) { // Do it only for non-instance types
1466 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1467 }
1468 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1469 // Static fields are in the space above the normal instance
1470 // fields in the java.lang.Class instance.
1471 if (ik != ciEnv::current()->Class_klass()) {
1472 to = nullptr;
1473 tj = TypeOopPtr::BOTTOM;
1474 offset = tj->offset();
1475 }
1476 } else {
1477 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1478 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1479 assert(tj->offset() == offset, "no change to offset expected");
1480 bool xk = to->klass_is_exact();
1481 int instance_id = to->instance_id();
1482
1483 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1484 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1485 // its interfaces are included.
1486 if (xk && ik->equals(canonical_holder)) {
1487 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1488 } else {
1489 assert(xk || !is_known_inst, "Known instance should be exact type");
1490 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1491 }
1492 }
1493 }
1494
1495 // Klass pointers to object array klasses need some flattening
1496 const TypeKlassPtr *tk = tj->isa_klassptr();
1497 if( tk ) {
1498 // If we are referencing a field within a Klass, we need
1499 // to assume the worst case of an Object. Both exact and
1500 // inexact types must flatten to the same alias class so
1501 // use NotNull as the PTR.
1502 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1503 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1504 env()->Object_klass(),
1505 offset);
1506 }
1507
1508 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1509 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1510 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1511 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1512 } else {
1513 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1514 }
1515 }
1516
1517 // Check for precise loads from the primary supertype array and force them
1518 // to the supertype cache alias index. Check for generic array loads from
1519 // the primary supertype array and also force them to the supertype cache
1520 // alias index. Since the same load can reach both, we need to merge
1521 // these 2 disparate memories into the same alias class. Since the
1522 // primary supertype array is read-only, there's no chance of confusion
1523 // where we bypass an array load and an array store.
1524 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1525 if (offset == Type::OffsetBot ||
1526 (offset >= primary_supers_offset &&
1527 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1528 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1529 offset = in_bytes(Klass::secondary_super_cache_offset());
1530 tj = tk = tk->with_offset(offset);
1531 }
1532 }
1533
1534 // Flatten all Raw pointers together.
1535 if (tj->base() == Type::RawPtr)
1536 tj = TypeRawPtr::BOTTOM;
1537
1538 if (tj->base() == Type::AnyPtr)
1539 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1540
1541 offset = tj->offset();
1542 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1543
1544 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1545 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1546 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1547 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1548 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1549 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1550 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1551 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1552 assert( tj->ptr() != TypePtr::TopPTR &&
1553 tj->ptr() != TypePtr::AnyNull &&
1554 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1555 // assert( tj->ptr() != TypePtr::Constant ||
1556 // tj->base() == Type::RawPtr ||
1557 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1558
1559 return tj;
1560 }
1561
1562 void Compile::AliasType::Init(int i, const TypePtr* at) {
1563 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1564 _index = i;
1565 _adr_type = at;
1566 _field = nullptr;
1567 _element = nullptr;
1568 _is_rewritable = true; // default
1569 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1570 if (atoop != nullptr && atoop->is_known_instance()) {
1571 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1572 _general_index = Compile::current()->get_alias_index(gt);
1573 } else {
1574 _general_index = 0;
1575 }
1576 }
1577
1578 BasicType Compile::AliasType::basic_type() const {
1579 if (element() != nullptr) {
1580 const Type* element = adr_type()->is_aryptr()->elem();
1581 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1582 } if (field() != nullptr) {
1583 return field()->layout_type();
1584 } else {
1585 return T_ILLEGAL; // unknown
1586 }
1587 }
1588
1589 //---------------------------------print_on------------------------------------
1590 #ifndef PRODUCT
1591 void Compile::AliasType::print_on(outputStream* st) {
1592 if (index() < 10)
1593 st->print("@ <%d> ", index());
1594 else st->print("@ <%d>", index());
1595 st->print(is_rewritable() ? " " : " RO");
1596 int offset = adr_type()->offset();
1597 if (offset == Type::OffsetBot)
1598 st->print(" +any");
1599 else st->print(" +%-3d", offset);
1600 st->print(" in ");
1601 adr_type()->dump_on(st);
1602 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1603 if (field() != nullptr && tjp) {
1604 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1605 tjp->offset() != field()->offset_in_bytes()) {
1606 st->print(" != ");
1607 field()->print();
1608 st->print(" ***");
1609 }
1610 }
1611 }
1612
1613 void print_alias_types() {
1614 Compile* C = Compile::current();
1615 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1616 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1617 C->alias_type(idx)->print_on(tty);
1618 tty->cr();
1619 }
1620 }
1621 #endif
1622
1623
1624 //----------------------------probe_alias_cache--------------------------------
1625 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1626 intptr_t key = (intptr_t) adr_type;
1627 key ^= key >> logAliasCacheSize;
1628 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1629 }
1630
1631
1632 //-----------------------------grow_alias_types--------------------------------
1633 void Compile::grow_alias_types() {
1634 const int old_ats = _max_alias_types; // how many before?
1635 const int new_ats = old_ats; // how many more?
1636 const int grow_ats = old_ats+new_ats; // how many now?
1637 _max_alias_types = grow_ats;
1638 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1639 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1640 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1641 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1642 }
1643
1644
1645 //--------------------------------find_alias_type------------------------------
1646 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1647 if (!do_aliasing()) {
1648 return alias_type(AliasIdxBot);
1649 }
1650
1651 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1652 if (ace->_adr_type == adr_type) {
1653 return alias_type(ace->_index);
1654 }
1655
1656 // Handle special cases.
1657 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1658 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1659
1660 // Do it the slow way.
1661 const TypePtr* flat = flatten_alias_type(adr_type);
1662
1663 #ifdef ASSERT
1664 {
1665 ResourceMark rm;
1666 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1667 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1668 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1669 Type::str(adr_type));
1670 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1671 const TypeOopPtr* foop = flat->is_oopptr();
1672 // Scalarizable allocations have exact klass always.
1673 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1674 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1675 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1676 Type::str(foop), Type::str(xoop));
1677 }
1678 }
1679 #endif
1680
1681 int idx = AliasIdxTop;
1682 for (int i = 0; i < num_alias_types(); i++) {
1683 if (alias_type(i)->adr_type() == flat) {
1684 idx = i;
1685 break;
1686 }
1687 }
1688
1689 if (idx == AliasIdxTop) {
1690 if (no_create) return nullptr;
1691 // Grow the array if necessary.
1692 if (_num_alias_types == _max_alias_types) grow_alias_types();
1693 // Add a new alias type.
1694 idx = _num_alias_types++;
1695 _alias_types[idx]->Init(idx, flat);
1696 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1697 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1698 if (flat->isa_instptr()) {
1699 if (flat->offset() == java_lang_Class::klass_offset()
1700 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1701 alias_type(idx)->set_rewritable(false);
1702 }
1703 if (flat->isa_aryptr()) {
1704 #ifdef ASSERT
1705 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1706 // (T_BYTE has the weakest alignment and size restrictions...)
1707 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1708 #endif
1709 if (flat->offset() == TypePtr::OffsetBot) {
1710 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1711 }
1712 }
1713 if (flat->isa_klassptr()) {
1714 if (UseCompactObjectHeaders) {
1715 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1716 alias_type(idx)->set_rewritable(false);
1717 }
1718 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1719 alias_type(idx)->set_rewritable(false);
1720 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1721 alias_type(idx)->set_rewritable(false);
1722 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1723 alias_type(idx)->set_rewritable(false);
1724 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1725 alias_type(idx)->set_rewritable(false);
1726 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1727 alias_type(idx)->set_rewritable(false);
1728 }
1729 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1730 // but the base pointer type is not distinctive enough to identify
1731 // references into JavaThread.)
1732
1733 // Check for final fields.
1734 const TypeInstPtr* tinst = flat->isa_instptr();
1735 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1736 ciField* field;
1737 if (tinst->const_oop() != nullptr &&
1738 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1739 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1740 // static field
1741 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1742 field = k->get_field_by_offset(tinst->offset(), true);
1743 } else {
1744 ciInstanceKlass *k = tinst->instance_klass();
1745 field = k->get_field_by_offset(tinst->offset(), false);
1746 }
1747 assert(field == nullptr ||
1748 original_field == nullptr ||
1749 (field->holder() == original_field->holder() &&
1750 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1751 field->is_static() == original_field->is_static()), "wrong field?");
1752 // Set field() and is_rewritable() attributes.
1753 if (field != nullptr) alias_type(idx)->set_field(field);
1754 }
1755 }
1756
1757 // Fill the cache for next time.
1758 ace->_adr_type = adr_type;
1759 ace->_index = idx;
1760 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1761
1762 // Might as well try to fill the cache for the flattened version, too.
1763 AliasCacheEntry* face = probe_alias_cache(flat);
1764 if (face->_adr_type == nullptr) {
1765 face->_adr_type = flat;
1766 face->_index = idx;
1767 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1768 }
1769
1770 return alias_type(idx);
1771 }
1772
1773
1774 Compile::AliasType* Compile::alias_type(ciField* field) {
1775 const TypeOopPtr* t;
1776 if (field->is_static())
1777 t = TypeInstPtr::make(field->holder()->java_mirror());
1778 else
1779 t = TypeOopPtr::make_from_klass_raw(field->holder());
1780 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1781 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1782 return atp;
1783 }
1784
1785
1786 //------------------------------have_alias_type--------------------------------
1787 bool Compile::have_alias_type(const TypePtr* adr_type) {
1788 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1789 if (ace->_adr_type == adr_type) {
1790 return true;
1791 }
1792
1793 // Handle special cases.
1794 if (adr_type == nullptr) return true;
1795 if (adr_type == TypePtr::BOTTOM) return true;
1796
1797 return find_alias_type(adr_type, true, nullptr) != nullptr;
1798 }
1799
1800 //-----------------------------must_alias--------------------------------------
1801 // True if all values of the given address type are in the given alias category.
1802 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1803 if (alias_idx == AliasIdxBot) return true; // the universal category
1804 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1805 if (alias_idx == AliasIdxTop) return false; // the empty category
1806 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1807
1808 // the only remaining possible overlap is identity
1809 int adr_idx = get_alias_index(adr_type);
1810 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1811 assert(adr_idx == alias_idx ||
1812 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1813 && adr_type != TypeOopPtr::BOTTOM),
1814 "should not be testing for overlap with an unsafe pointer");
1815 return adr_idx == alias_idx;
1816 }
1817
1818 //------------------------------can_alias--------------------------------------
1819 // True if any values of the given address type are in the given alias category.
1820 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1821 if (alias_idx == AliasIdxTop) return false; // the empty category
1822 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1823 // Known instance doesn't alias with bottom memory
1824 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1825 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1826
1827 // the only remaining possible overlap is identity
1828 int adr_idx = get_alias_index(adr_type);
1829 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1830 return adr_idx == alias_idx;
1831 }
1832
1833 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1834 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1835 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1836 if (parse_predicate_count() == 0) {
1837 return;
1838 }
1839 for (int i = 0; i < parse_predicate_count(); i++) {
1840 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1841 parse_predicate->mark_useless();
1842 igvn._worklist.push(parse_predicate);
1843 }
1844 _parse_predicates.clear();
1845 }
1846
1847 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1848 if (!n->for_post_loop_opts_igvn()) {
1849 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1850 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1851 _for_post_loop_igvn.append(n);
1852 }
1853 }
1854
1855 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1856 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1857 _for_post_loop_igvn.remove(n);
1858 }
1859
1860 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1861 // Verify that all previous optimizations produced a valid graph
1862 // at least to this point, even if no loop optimizations were done.
1863 PhaseIdealLoop::verify(igvn);
1864
1865 C->set_post_loop_opts_phase(); // no more loop opts allowed
1866
1867 assert(!C->major_progress(), "not cleared");
1868
1869 if (_for_post_loop_igvn.length() > 0) {
1870 while (_for_post_loop_igvn.length() > 0) {
1871 Node* n = _for_post_loop_igvn.pop();
1872 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1873 igvn._worklist.push(n);
1874 }
1875 igvn.optimize();
1876 if (failing()) return;
1877 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1878 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1879
1880 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1881 if (C->major_progress()) {
1882 C->clear_major_progress(); // ensure that major progress is now clear
1883 }
1884 }
1885 }
1886
1887 void Compile::record_for_merge_stores_igvn(Node* n) {
1888 if (!n->for_merge_stores_igvn()) {
1889 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1890 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1891 _for_merge_stores_igvn.append(n);
1892 }
1893 }
1894
1895 void Compile::remove_from_merge_stores_igvn(Node* n) {
1896 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1897 _for_merge_stores_igvn.remove(n);
1898 }
1899
1900 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1901 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1902 // the stores, and we merge the wrong sequence of stores.
1903 // Example:
1904 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1905 // Apply MergeStores:
1906 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1907 // Remove more RangeChecks:
1908 // StoreI [ StoreL ] StoreI
1909 // But now it would have been better to do this instead:
1910 // [ StoreL ] [ StoreL ]
1911 //
1912 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1913 // since we never unset _merge_stores_phase.
1914 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1915 C->set_merge_stores_phase();
1916
1917 if (_for_merge_stores_igvn.length() > 0) {
1918 while (_for_merge_stores_igvn.length() > 0) {
1919 Node* n = _for_merge_stores_igvn.pop();
1920 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1921 igvn._worklist.push(n);
1922 }
1923 igvn.optimize();
1924 if (failing()) return;
1925 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1926 print_method(PHASE_AFTER_MERGE_STORES, 3);
1927 }
1928 }
1929
1930 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1931 if (OptimizeUnstableIf) {
1932 _unstable_if_traps.append(trap);
1933 }
1934 }
1935
1936 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1937 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1938 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1939 Node* n = trap->uncommon_trap();
1940 if (!useful.member(n)) {
1941 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1942 }
1943 }
1944 }
1945
1946 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1947 // or fold-compares case. Return true if succeed or not found.
1948 //
1949 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1950 // false and ask fold-compares to yield.
1951 //
1952 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1953 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1954 // when deoptimization does happen.
1955 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1956 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1957 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1958 if (trap->uncommon_trap() == unc) {
1959 if (yield && trap->modified()) {
1960 return false;
1961 }
1962 _unstable_if_traps.delete_at(i);
1963 break;
1964 }
1965 }
1966 return true;
1967 }
1968
1969 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1970 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1971 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1972 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1973 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1974 CallStaticJavaNode* unc = trap->uncommon_trap();
1975 int next_bci = trap->next_bci();
1976 bool modified = trap->modified();
1977
1978 if (next_bci != -1 && !modified) {
1979 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1980 JVMState* jvms = unc->jvms();
1981 ciMethod* method = jvms->method();
1982 ciBytecodeStream iter(method);
1983
1984 iter.force_bci(jvms->bci());
1985 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1986 Bytecodes::Code c = iter.cur_bc();
1987 Node* lhs = nullptr;
1988 Node* rhs = nullptr;
1989 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1990 lhs = unc->peek_operand(0);
1991 rhs = unc->peek_operand(1);
1992 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1993 lhs = unc->peek_operand(0);
1994 }
1995
1996 ResourceMark rm;
1997 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1998 assert(live_locals.is_valid(), "broken liveness info");
1999 int len = (int)live_locals.size();
2000
2001 for (int i = 0; i < len; i++) {
2002 Node* local = unc->local(jvms, i);
2003 // kill local using the liveness of next_bci.
2004 // give up when the local looks like an operand to secure reexecution.
2005 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2006 uint idx = jvms->locoff() + i;
2007 #ifdef ASSERT
2008 if (PrintOpto && Verbose) {
2009 tty->print("[unstable_if] kill local#%d: ", idx);
2010 local->dump();
2011 tty->cr();
2012 }
2013 #endif
2014 igvn.replace_input_of(unc, idx, top());
2015 modified = true;
2016 }
2017 }
2018 }
2019
2020 // keep the mondified trap for late query
2021 if (modified) {
2022 trap->set_modified();
2023 } else {
2024 _unstable_if_traps.delete_at(i);
2025 }
2026 }
2027 igvn.optimize();
2028 }
2029
2030 // StringOpts and late inlining of string methods
2031 void Compile::inline_string_calls(bool parse_time) {
2032 {
2033 // remove useless nodes to make the usage analysis simpler
2034 ResourceMark rm;
2035 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2036 }
2037
2038 {
2039 ResourceMark rm;
2040 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2041 PhaseStringOpts pso(initial_gvn());
2042 print_method(PHASE_AFTER_STRINGOPTS, 3);
2043 }
2044
2045 // now inline anything that we skipped the first time around
2046 if (!parse_time) {
2047 _late_inlines_pos = _late_inlines.length();
2048 }
2049
2050 while (_string_late_inlines.length() > 0) {
2051 CallGenerator* cg = _string_late_inlines.pop();
2052 cg->do_late_inline();
2053 if (failing()) return;
2054 }
2055 _string_late_inlines.trunc_to(0);
2056 }
2057
2058 // Late inlining of boxing methods
2059 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2060 if (_boxing_late_inlines.length() > 0) {
2061 assert(has_boxed_value(), "inconsistent");
2062
2063 set_inlining_incrementally(true);
2064
2065 igvn_worklist()->ensure_empty(); // should be done with igvn
2066
2067 _late_inlines_pos = _late_inlines.length();
2068
2069 while (_boxing_late_inlines.length() > 0) {
2070 CallGenerator* cg = _boxing_late_inlines.pop();
2071 cg->do_late_inline();
2072 if (failing()) return;
2073 }
2074 _boxing_late_inlines.trunc_to(0);
2075
2076 inline_incrementally_cleanup(igvn);
2077
2078 set_inlining_incrementally(false);
2079 }
2080 }
2081
2082 bool Compile::inline_incrementally_one() {
2083 assert(IncrementalInline, "incremental inlining should be on");
2084
2085 TracePhase tp(_t_incrInline_inline);
2086
2087 set_inlining_progress(false);
2088 set_do_cleanup(false);
2089
2090 for (int i = 0; i < _late_inlines.length(); i++) {
2091 _late_inlines_pos = i+1;
2092 CallGenerator* cg = _late_inlines.at(i);
2093 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2094 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2095 cg->do_late_inline();
2096 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2097 if (failing()) {
2098 return false;
2099 } else if (inlining_progress()) {
2100 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2101 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2102 break; // process one call site at a time
2103 }
2104 } else {
2105 // Ignore late inline direct calls when inlining is not allowed.
2106 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2107 }
2108 }
2109 // Remove processed elements.
2110 _late_inlines.remove_till(_late_inlines_pos);
2111 _late_inlines_pos = 0;
2112
2113 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2114
2115 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2116
2117 set_inlining_progress(false);
2118 set_do_cleanup(false);
2119
2120 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2121 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2122 }
2123
2124 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2125 {
2126 TracePhase tp(_t_incrInline_pru);
2127 ResourceMark rm;
2128 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2129 }
2130 {
2131 TracePhase tp(_t_incrInline_igvn);
2132 igvn.reset_from_gvn(initial_gvn());
2133 igvn.optimize();
2134 if (failing()) return;
2135 }
2136 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2137 }
2138
2139 // Perform incremental inlining until bound on number of live nodes is reached
2140 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2141 TracePhase tp(_t_incrInline);
2142
2143 set_inlining_incrementally(true);
2144 uint low_live_nodes = 0;
2145
2146 while (_late_inlines.length() > 0) {
2147 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2148 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2149 TracePhase tp(_t_incrInline_ideal);
2150 // PhaseIdealLoop is expensive so we only try it once we are
2151 // out of live nodes and we only try it again if the previous
2152 // helped got the number of nodes down significantly
2153 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2154 if (failing()) return;
2155 low_live_nodes = live_nodes();
2156 _major_progress = true;
2157 }
2158
2159 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2160 bool do_print_inlining = print_inlining() || print_intrinsics();
2161 if (do_print_inlining || log() != nullptr) {
2162 // Print inlining message for candidates that we couldn't inline for lack of space.
2163 for (int i = 0; i < _late_inlines.length(); i++) {
2164 CallGenerator* cg = _late_inlines.at(i);
2165 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2166 if (do_print_inlining) {
2167 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2168 }
2169 log_late_inline_failure(cg, msg);
2170 }
2171 }
2172 break; // finish
2173 }
2174 }
2175
2176 igvn_worklist()->ensure_empty(); // should be done with igvn
2177
2178 while (inline_incrementally_one()) {
2179 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2180 }
2181 if (failing()) return;
2182
2183 inline_incrementally_cleanup(igvn);
2184
2185 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2186
2187 if (failing()) return;
2188
2189 if (_late_inlines.length() == 0) {
2190 break; // no more progress
2191 }
2192 }
2193
2194 igvn_worklist()->ensure_empty(); // should be done with igvn
2195
2196 if (_string_late_inlines.length() > 0) {
2197 assert(has_stringbuilder(), "inconsistent");
2198
2199 inline_string_calls(false);
2200
2201 if (failing()) return;
2202
2203 inline_incrementally_cleanup(igvn);
2204 }
2205
2206 set_inlining_incrementally(false);
2207 }
2208
2209 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2210 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2211 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2212 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2213 // as if "inlining_incrementally() == true" were set.
2214 assert(inlining_incrementally() == false, "not allowed");
2215 assert(_modified_nodes == nullptr, "not allowed");
2216 assert(_late_inlines.length() > 0, "sanity");
2217
2218 while (_late_inlines.length() > 0) {
2219 igvn_worklist()->ensure_empty(); // should be done with igvn
2220
2221 while (inline_incrementally_one()) {
2222 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2223 }
2224 if (failing()) return;
2225
2226 inline_incrementally_cleanup(igvn);
2227 }
2228 }
2229
2230 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2231 if (_loop_opts_cnt > 0) {
2232 while (major_progress() && (_loop_opts_cnt > 0)) {
2233 TracePhase tp(_t_idealLoop);
2234 PhaseIdealLoop::optimize(igvn, mode);
2235 _loop_opts_cnt--;
2236 if (failing()) return false;
2237 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2238 }
2239 }
2240 return true;
2241 }
2242
2243 // Remove edges from "root" to each SafePoint at a backward branch.
2244 // They were inserted during parsing (see add_safepoint()) to make
2245 // infinite loops without calls or exceptions visible to root, i.e.,
2246 // useful.
2247 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2248 Node *r = root();
2249 if (r != nullptr) {
2250 for (uint i = r->req(); i < r->len(); ++i) {
2251 Node *n = r->in(i);
2252 if (n != nullptr && n->is_SafePoint()) {
2253 r->rm_prec(i);
2254 if (n->outcnt() == 0) {
2255 igvn.remove_dead_node(n);
2256 }
2257 --i;
2258 }
2259 }
2260 // Parsing may have added top inputs to the root node (Path
2261 // leading to the Halt node proven dead). Make sure we get a
2262 // chance to clean them up.
2263 igvn._worklist.push(r);
2264 igvn.optimize();
2265 }
2266 }
2267
2268 //------------------------------Optimize---------------------------------------
2269 // Given a graph, optimize it.
2270 void Compile::Optimize() {
2271 TracePhase tp(_t_optimizer);
2272
2273 #ifndef PRODUCT
2274 if (env()->break_at_compile()) {
2275 BREAKPOINT;
2276 }
2277
2278 #endif
2279
2280 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2281 #ifdef ASSERT
2282 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2283 #endif
2284
2285 ResourceMark rm;
2286
2287 NOT_PRODUCT( verify_graph_edges(); )
2288
2289 print_method(PHASE_AFTER_PARSING, 1);
2290
2291 {
2292 // Iterative Global Value Numbering, including ideal transforms
2293 // Initialize IterGVN with types and values from parse-time GVN
2294 PhaseIterGVN igvn(initial_gvn());
2295 #ifdef ASSERT
2296 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2297 #endif
2298 {
2299 TracePhase tp(_t_iterGVN);
2300 igvn.optimize();
2301 }
2302
2303 if (failing()) return;
2304
2305 print_method(PHASE_ITER_GVN1, 2);
2306
2307 process_for_unstable_if_traps(igvn);
2308
2309 if (failing()) return;
2310
2311 inline_incrementally(igvn);
2312
2313 print_method(PHASE_INCREMENTAL_INLINE, 2);
2314
2315 if (failing()) return;
2316
2317 if (eliminate_boxing()) {
2318 // Inline valueOf() methods now.
2319 inline_boxing_calls(igvn);
2320
2321 if (failing()) return;
2322
2323 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2324 inline_incrementally(igvn);
2325 }
2326
2327 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2328
2329 if (failing()) return;
2330 }
2331
2332 // Remove the speculative part of types and clean up the graph from
2333 // the extra CastPP nodes whose only purpose is to carry them. Do
2334 // that early so that optimizations are not disrupted by the extra
2335 // CastPP nodes.
2336 remove_speculative_types(igvn);
2337
2338 if (failing()) return;
2339
2340 // No more new expensive nodes will be added to the list from here
2341 // so keep only the actual candidates for optimizations.
2342 cleanup_expensive_nodes(igvn);
2343
2344 if (failing()) return;
2345
2346 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2347 if (EnableVectorSupport && has_vbox_nodes()) {
2348 TracePhase tp(_t_vector);
2349 PhaseVector pv(igvn);
2350 pv.optimize_vector_boxes();
2351 if (failing()) return;
2352 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2353 }
2354 assert(!has_vbox_nodes(), "sanity");
2355
2356 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2357 Compile::TracePhase tp(_t_renumberLive);
2358 igvn_worklist()->ensure_empty(); // should be done with igvn
2359 {
2360 ResourceMark rm;
2361 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2362 }
2363 igvn.reset_from_gvn(initial_gvn());
2364 igvn.optimize();
2365 if (failing()) return;
2366 }
2367
2368 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2369 // safepoints
2370 remove_root_to_sfpts_edges(igvn);
2371
2372 if (failing()) return;
2373
2374 // Perform escape analysis
2375 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2376 if (has_loops()) {
2377 // Cleanup graph (remove dead nodes).
2378 TracePhase tp(_t_idealLoop);
2379 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2380 if (failing()) return;
2381 }
2382 bool progress;
2383 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2384 do {
2385 ConnectionGraph::do_analysis(this, &igvn);
2386
2387 if (failing()) return;
2388
2389 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2390
2391 // Optimize out fields loads from scalar replaceable allocations.
2392 igvn.optimize();
2393 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2394
2395 if (failing()) return;
2396
2397 if (congraph() != nullptr && macro_count() > 0) {
2398 TracePhase tp(_t_macroEliminate);
2399 PhaseMacroExpand mexp(igvn);
2400 mexp.eliminate_macro_nodes();
2401 if (failing()) return;
2402
2403 igvn.set_delay_transform(false);
2404 igvn.optimize();
2405 if (failing()) return;
2406
2407 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2408 }
2409
2410 ConnectionGraph::verify_ram_nodes(this, root());
2411 if (failing()) return;
2412
2413 progress = do_iterative_escape_analysis() &&
2414 (macro_count() < mcount) &&
2415 ConnectionGraph::has_candidates(this);
2416 // Try again if candidates exist and made progress
2417 // by removing some allocations and/or locks.
2418 } while (progress);
2419 }
2420
2421 // Loop transforms on the ideal graph. Range Check Elimination,
2422 // peeling, unrolling, etc.
2423
2424 // Set loop opts counter
2425 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2426 {
2427 TracePhase tp(_t_idealLoop);
2428 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2429 _loop_opts_cnt--;
2430 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2431 if (failing()) return;
2432 }
2433 // Loop opts pass if partial peeling occurred in previous pass
2434 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2435 TracePhase tp(_t_idealLoop);
2436 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2437 _loop_opts_cnt--;
2438 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2439 if (failing()) return;
2440 }
2441 // Loop opts pass for loop-unrolling before CCP
2442 if(major_progress() && (_loop_opts_cnt > 0)) {
2443 TracePhase tp(_t_idealLoop);
2444 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2445 _loop_opts_cnt--;
2446 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2447 }
2448 if (!failing()) {
2449 // Verify that last round of loop opts produced a valid graph
2450 PhaseIdealLoop::verify(igvn);
2451 }
2452 }
2453 if (failing()) return;
2454
2455 // Conditional Constant Propagation;
2456 print_method(PHASE_BEFORE_CCP1, 2);
2457 PhaseCCP ccp( &igvn );
2458 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2459 {
2460 TracePhase tp(_t_ccp);
2461 ccp.do_transform();
2462 }
2463 print_method(PHASE_CCP1, 2);
2464
2465 assert( true, "Break here to ccp.dump_old2new_map()");
2466
2467 // Iterative Global Value Numbering, including ideal transforms
2468 {
2469 TracePhase tp(_t_iterGVN2);
2470 igvn.reset_from_igvn(&ccp);
2471 igvn.optimize();
2472 }
2473 print_method(PHASE_ITER_GVN2, 2);
2474
2475 if (failing()) return;
2476
2477 // Loop transforms on the ideal graph. Range Check Elimination,
2478 // peeling, unrolling, etc.
2479 if (!optimize_loops(igvn, LoopOptsDefault)) {
2480 return;
2481 }
2482
2483 if (failing()) return;
2484
2485 C->clear_major_progress(); // ensure that major progress is now clear
2486
2487 process_for_post_loop_opts_igvn(igvn);
2488
2489 process_for_merge_stores_igvn(igvn);
2490
2491 if (failing()) return;
2492
2493 #ifdef ASSERT
2494 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2495 #endif
2496
2497 {
2498 TracePhase tp(_t_macroExpand);
2499 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2500 PhaseMacroExpand mex(igvn);
2501 if (mex.expand_macro_nodes()) {
2502 assert(failing(), "must bail out w/ explicit message");
2503 return;
2504 }
2505 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2506 }
2507
2508 {
2509 TracePhase tp(_t_barrierExpand);
2510 if (bs->expand_barriers(this, igvn)) {
2511 assert(failing(), "must bail out w/ explicit message");
2512 return;
2513 }
2514 print_method(PHASE_BARRIER_EXPANSION, 2);
2515 }
2516
2517 if (C->max_vector_size() > 0) {
2518 C->optimize_logic_cones(igvn);
2519 igvn.optimize();
2520 if (failing()) return;
2521 }
2522
2523 DEBUG_ONLY( _modified_nodes = nullptr; )
2524
2525 assert(igvn._worklist.size() == 0, "not empty");
2526
2527 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2528
2529 if (_late_inlines.length() > 0) {
2530 // More opportunities to optimize virtual and MH calls.
2531 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2532 process_late_inline_calls_no_inline(igvn);
2533 if (failing()) return;
2534 }
2535 } // (End scope of igvn; run destructor if necessary for asserts.)
2536
2537 check_no_dead_use();
2538
2539 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2540 // to remove hashes to unlock nodes for modifications.
2541 C->node_hash()->clear();
2542
2543 // A method with only infinite loops has no edges entering loops from root
2544 {
2545 TracePhase tp(_t_graphReshaping);
2546 if (final_graph_reshaping()) {
2547 assert(failing(), "must bail out w/ explicit message");
2548 return;
2549 }
2550 }
2551
2552 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2553 DEBUG_ONLY(set_phase_optimize_finished();)
2554 }
2555
2556 #ifdef ASSERT
2557 void Compile::check_no_dead_use() const {
2558 ResourceMark rm;
2559 Unique_Node_List wq;
2560 wq.push(root());
2561 for (uint i = 0; i < wq.size(); ++i) {
2562 Node* n = wq.at(i);
2563 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2564 Node* u = n->fast_out(j);
2565 if (u->outcnt() == 0 && !u->is_Con()) {
2566 u->dump();
2567 fatal("no reachable node should have no use");
2568 }
2569 wq.push(u);
2570 }
2571 }
2572 }
2573 #endif
2574
2575 void Compile::inline_vector_reboxing_calls() {
2576 if (C->_vector_reboxing_late_inlines.length() > 0) {
2577 _late_inlines_pos = C->_late_inlines.length();
2578 while (_vector_reboxing_late_inlines.length() > 0) {
2579 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2580 cg->do_late_inline();
2581 if (failing()) return;
2582 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2583 }
2584 _vector_reboxing_late_inlines.trunc_to(0);
2585 }
2586 }
2587
2588 bool Compile::has_vbox_nodes() {
2589 if (C->_vector_reboxing_late_inlines.length() > 0) {
2590 return true;
2591 }
2592 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2593 Node * n = C->macro_node(macro_idx);
2594 assert(n->is_macro(), "only macro nodes expected here");
2595 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2596 return true;
2597 }
2598 }
2599 return false;
2600 }
2601
2602 //---------------------------- Bitwise operation packing optimization ---------------------------
2603
2604 static bool is_vector_unary_bitwise_op(Node* n) {
2605 return n->Opcode() == Op_XorV &&
2606 VectorNode::is_vector_bitwise_not_pattern(n);
2607 }
2608
2609 static bool is_vector_binary_bitwise_op(Node* n) {
2610 switch (n->Opcode()) {
2611 case Op_AndV:
2612 case Op_OrV:
2613 return true;
2614
2615 case Op_XorV:
2616 return !is_vector_unary_bitwise_op(n);
2617
2618 default:
2619 return false;
2620 }
2621 }
2622
2623 static bool is_vector_ternary_bitwise_op(Node* n) {
2624 return n->Opcode() == Op_MacroLogicV;
2625 }
2626
2627 static bool is_vector_bitwise_op(Node* n) {
2628 return is_vector_unary_bitwise_op(n) ||
2629 is_vector_binary_bitwise_op(n) ||
2630 is_vector_ternary_bitwise_op(n);
2631 }
2632
2633 static bool is_vector_bitwise_cone_root(Node* n) {
2634 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2635 return false;
2636 }
2637 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2638 if (is_vector_bitwise_op(n->fast_out(i))) {
2639 return false;
2640 }
2641 }
2642 return true;
2643 }
2644
2645 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2646 uint cnt = 0;
2647 if (is_vector_bitwise_op(n)) {
2648 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2649 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2650 for (uint i = 1; i < inp_cnt; i++) {
2651 Node* in = n->in(i);
2652 bool skip = VectorNode::is_all_ones_vector(in);
2653 if (!skip && !inputs.member(in)) {
2654 inputs.push(in);
2655 cnt++;
2656 }
2657 }
2658 assert(cnt <= 1, "not unary");
2659 } else {
2660 uint last_req = inp_cnt;
2661 if (is_vector_ternary_bitwise_op(n)) {
2662 last_req = inp_cnt - 1; // skip last input
2663 }
2664 for (uint i = 1; i < last_req; i++) {
2665 Node* def = n->in(i);
2666 if (!inputs.member(def)) {
2667 inputs.push(def);
2668 cnt++;
2669 }
2670 }
2671 }
2672 } else { // not a bitwise operations
2673 if (!inputs.member(n)) {
2674 inputs.push(n);
2675 cnt++;
2676 }
2677 }
2678 return cnt;
2679 }
2680
2681 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2682 Unique_Node_List useful_nodes;
2683 C->identify_useful_nodes(useful_nodes);
2684
2685 for (uint i = 0; i < useful_nodes.size(); i++) {
2686 Node* n = useful_nodes.at(i);
2687 if (is_vector_bitwise_cone_root(n)) {
2688 list.push(n);
2689 }
2690 }
2691 }
2692
2693 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2694 const TypeVect* vt,
2695 Unique_Node_List& partition,
2696 Unique_Node_List& inputs) {
2697 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2698 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2699 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2700
2701 Node* in1 = inputs.at(0);
2702 Node* in2 = inputs.at(1);
2703 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2704
2705 uint func = compute_truth_table(partition, inputs);
2706
2707 Node* pn = partition.at(partition.size() - 1);
2708 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2709 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2710 }
2711
2712 static uint extract_bit(uint func, uint pos) {
2713 return (func & (1 << pos)) >> pos;
2714 }
2715
2716 //
2717 // A macro logic node represents a truth table. It has 4 inputs,
2718 // First three inputs corresponds to 3 columns of a truth table
2719 // and fourth input captures the logic function.
2720 //
2721 // eg. fn = (in1 AND in2) OR in3;
2722 //
2723 // MacroNode(in1,in2,in3,fn)
2724 //
2725 // -----------------
2726 // in1 in2 in3 fn
2727 // -----------------
2728 // 0 0 0 0
2729 // 0 0 1 1
2730 // 0 1 0 0
2731 // 0 1 1 1
2732 // 1 0 0 0
2733 // 1 0 1 1
2734 // 1 1 0 1
2735 // 1 1 1 1
2736 //
2737
2738 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2739 int res = 0;
2740 for (int i = 0; i < 8; i++) {
2741 int bit1 = extract_bit(in1, i);
2742 int bit2 = extract_bit(in2, i);
2743 int bit3 = extract_bit(in3, i);
2744
2745 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2746 int func_bit = extract_bit(func, func_bit_pos);
2747
2748 res |= func_bit << i;
2749 }
2750 return res;
2751 }
2752
2753 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2754 assert(n != nullptr, "");
2755 assert(eval_map.contains(n), "absent");
2756 return *(eval_map.get(n));
2757 }
2758
2759 static void eval_operands(Node* n,
2760 uint& func1, uint& func2, uint& func3,
2761 ResourceHashtable<Node*,uint>& eval_map) {
2762 assert(is_vector_bitwise_op(n), "");
2763
2764 if (is_vector_unary_bitwise_op(n)) {
2765 Node* opnd = n->in(1);
2766 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2767 opnd = n->in(2);
2768 }
2769 func1 = eval_operand(opnd, eval_map);
2770 } else if (is_vector_binary_bitwise_op(n)) {
2771 func1 = eval_operand(n->in(1), eval_map);
2772 func2 = eval_operand(n->in(2), eval_map);
2773 } else {
2774 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2775 func1 = eval_operand(n->in(1), eval_map);
2776 func2 = eval_operand(n->in(2), eval_map);
2777 func3 = eval_operand(n->in(3), eval_map);
2778 }
2779 }
2780
2781 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2782 assert(inputs.size() <= 3, "sanity");
2783 ResourceMark rm;
2784 uint res = 0;
2785 ResourceHashtable<Node*,uint> eval_map;
2786
2787 // Populate precomputed functions for inputs.
2788 // Each input corresponds to one column of 3 input truth-table.
2789 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2790 0xCC, // (_, b, _) -> b
2791 0xF0 }; // (a, _, _) -> a
2792 for (uint i = 0; i < inputs.size(); i++) {
2793 eval_map.put(inputs.at(i), input_funcs[2-i]);
2794 }
2795
2796 for (uint i = 0; i < partition.size(); i++) {
2797 Node* n = partition.at(i);
2798
2799 uint func1 = 0, func2 = 0, func3 = 0;
2800 eval_operands(n, func1, func2, func3, eval_map);
2801
2802 switch (n->Opcode()) {
2803 case Op_OrV:
2804 assert(func3 == 0, "not binary");
2805 res = func1 | func2;
2806 break;
2807 case Op_AndV:
2808 assert(func3 == 0, "not binary");
2809 res = func1 & func2;
2810 break;
2811 case Op_XorV:
2812 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2813 assert(func2 == 0 && func3 == 0, "not unary");
2814 res = (~func1) & 0xFF;
2815 } else {
2816 assert(func3 == 0, "not binary");
2817 res = func1 ^ func2;
2818 }
2819 break;
2820 case Op_MacroLogicV:
2821 // Ordering of inputs may change during evaluation of sub-tree
2822 // containing MacroLogic node as a child node, thus a re-evaluation
2823 // makes sure that function is evaluated in context of current
2824 // inputs.
2825 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2826 break;
2827
2828 default: assert(false, "not supported: %s", n->Name());
2829 }
2830 assert(res <= 0xFF, "invalid");
2831 eval_map.put(n, res);
2832 }
2833 return res;
2834 }
2835
2836 // Criteria under which nodes gets packed into a macro logic node:-
2837 // 1) Parent and both child nodes are all unmasked or masked with
2838 // same predicates.
2839 // 2) Masked parent can be packed with left child if it is predicated
2840 // and both have same predicates.
2841 // 3) Masked parent can be packed with right child if its un-predicated
2842 // or has matching predication condition.
2843 // 4) An unmasked parent can be packed with an unmasked child.
2844 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2845 assert(partition.size() == 0, "not empty");
2846 assert(inputs.size() == 0, "not empty");
2847 if (is_vector_ternary_bitwise_op(n)) {
2848 return false;
2849 }
2850
2851 bool is_unary_op = is_vector_unary_bitwise_op(n);
2852 if (is_unary_op) {
2853 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2854 return false; // too few inputs
2855 }
2856
2857 bool pack_left_child = true;
2858 bool pack_right_child = true;
2859
2860 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2861 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2862
2863 int left_child_input_cnt = 0;
2864 int right_child_input_cnt = 0;
2865
2866 bool parent_is_predicated = n->is_predicated_vector();
2867 bool left_child_predicated = n->in(1)->is_predicated_vector();
2868 bool right_child_predicated = n->in(2)->is_predicated_vector();
2869
2870 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2871 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2872 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2873
2874 do {
2875 if (pack_left_child && left_child_LOP &&
2876 ((!parent_is_predicated && !left_child_predicated) ||
2877 ((parent_is_predicated && left_child_predicated &&
2878 parent_pred == left_child_pred)))) {
2879 partition.push(n->in(1));
2880 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2881 } else {
2882 inputs.push(n->in(1));
2883 left_child_input_cnt = 1;
2884 }
2885
2886 if (pack_right_child && right_child_LOP &&
2887 (!right_child_predicated ||
2888 (right_child_predicated && parent_is_predicated &&
2889 parent_pred == right_child_pred))) {
2890 partition.push(n->in(2));
2891 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2892 } else {
2893 inputs.push(n->in(2));
2894 right_child_input_cnt = 1;
2895 }
2896
2897 if (inputs.size() > 3) {
2898 assert(partition.size() > 0, "");
2899 inputs.clear();
2900 partition.clear();
2901 if (left_child_input_cnt > right_child_input_cnt) {
2902 pack_left_child = false;
2903 } else {
2904 pack_right_child = false;
2905 }
2906 } else {
2907 break;
2908 }
2909 } while(true);
2910
2911 if(partition.size()) {
2912 partition.push(n);
2913 }
2914
2915 return (partition.size() == 2 || partition.size() == 3) &&
2916 (inputs.size() == 2 || inputs.size() == 3);
2917 }
2918
2919 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2920 assert(is_vector_bitwise_op(n), "not a root");
2921
2922 visited.set(n->_idx);
2923
2924 // 1) Do a DFS walk over the logic cone.
2925 for (uint i = 1; i < n->req(); i++) {
2926 Node* in = n->in(i);
2927 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2928 process_logic_cone_root(igvn, in, visited);
2929 }
2930 }
2931
2932 // 2) Bottom up traversal: Merge node[s] with
2933 // the parent to form macro logic node.
2934 Unique_Node_List partition;
2935 Unique_Node_List inputs;
2936 if (compute_logic_cone(n, partition, inputs)) {
2937 const TypeVect* vt = n->bottom_type()->is_vect();
2938 Node* pn = partition.at(partition.size() - 1);
2939 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2940 if (mask == nullptr ||
2941 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2942 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2943 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2944 igvn.replace_node(n, macro_logic);
2945 }
2946 }
2947 }
2948
2949 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2950 ResourceMark rm;
2951 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2952 Unique_Node_List list;
2953 collect_logic_cone_roots(list);
2954
2955 while (list.size() > 0) {
2956 Node* n = list.pop();
2957 const TypeVect* vt = n->bottom_type()->is_vect();
2958 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2959 if (supported) {
2960 VectorSet visited(comp_arena());
2961 process_logic_cone_root(igvn, n, visited);
2962 }
2963 }
2964 }
2965 }
2966
2967 //------------------------------Code_Gen---------------------------------------
2968 // Given a graph, generate code for it
2969 void Compile::Code_Gen() {
2970 if (failing()) {
2971 return;
2972 }
2973
2974 // Perform instruction selection. You might think we could reclaim Matcher
2975 // memory PDQ, but actually the Matcher is used in generating spill code.
2976 // Internals of the Matcher (including some VectorSets) must remain live
2977 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2978 // set a bit in reclaimed memory.
2979
2980 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2981 // nodes. Mapping is only valid at the root of each matched subtree.
2982 NOT_PRODUCT( verify_graph_edges(); )
2983
2984 Matcher matcher;
2985 _matcher = &matcher;
2986 {
2987 TracePhase tp(_t_matcher);
2988 matcher.match();
2989 if (failing()) {
2990 return;
2991 }
2992 }
2993 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2994 // nodes. Mapping is only valid at the root of each matched subtree.
2995 NOT_PRODUCT( verify_graph_edges(); )
2996
2997 // If you have too many nodes, or if matching has failed, bail out
2998 check_node_count(0, "out of nodes matching instructions");
2999 if (failing()) {
3000 return;
3001 }
3002
3003 print_method(PHASE_MATCHING, 2);
3004
3005 // Build a proper-looking CFG
3006 PhaseCFG cfg(node_arena(), root(), matcher);
3007 if (failing()) {
3008 return;
3009 }
3010 _cfg = &cfg;
3011 {
3012 TracePhase tp(_t_scheduler);
3013 bool success = cfg.do_global_code_motion();
3014 if (!success) {
3015 return;
3016 }
3017
3018 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3019 NOT_PRODUCT( verify_graph_edges(); )
3020 cfg.verify();
3021 if (failing()) {
3022 return;
3023 }
3024 }
3025
3026 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3027 _regalloc = ®alloc;
3028 {
3029 TracePhase tp(_t_registerAllocation);
3030 // Perform register allocation. After Chaitin, use-def chains are
3031 // no longer accurate (at spill code) and so must be ignored.
3032 // Node->LRG->reg mappings are still accurate.
3033 _regalloc->Register_Allocate();
3034
3035 // Bail out if the allocator builds too many nodes
3036 if (failing()) {
3037 return;
3038 }
3039
3040 print_method(PHASE_REGISTER_ALLOCATION, 2);
3041 }
3042
3043 // Prior to register allocation we kept empty basic blocks in case the
3044 // the allocator needed a place to spill. After register allocation we
3045 // are not adding any new instructions. If any basic block is empty, we
3046 // can now safely remove it.
3047 {
3048 TracePhase tp(_t_blockOrdering);
3049 cfg.remove_empty_blocks();
3050 if (do_freq_based_layout()) {
3051 PhaseBlockLayout layout(cfg);
3052 } else {
3053 cfg.set_loop_alignment();
3054 }
3055 cfg.fixup_flow();
3056 cfg.remove_unreachable_blocks();
3057 cfg.verify_dominator_tree();
3058 print_method(PHASE_BLOCK_ORDERING, 3);
3059 }
3060
3061 // Apply peephole optimizations
3062 if( OptoPeephole ) {
3063 TracePhase tp(_t_peephole);
3064 PhasePeephole peep( _regalloc, cfg);
3065 peep.do_transform();
3066 print_method(PHASE_PEEPHOLE, 3);
3067 }
3068
3069 // Do late expand if CPU requires this.
3070 if (Matcher::require_postalloc_expand) {
3071 TracePhase tp(_t_postalloc_expand);
3072 cfg.postalloc_expand(_regalloc);
3073 print_method(PHASE_POSTALLOC_EXPAND, 3);
3074 }
3075
3076 #ifdef ASSERT
3077 {
3078 CompilationMemoryStatistic::do_test_allocations();
3079 if (failing()) return;
3080 }
3081 #endif
3082
3083 // Convert Nodes to instruction bits in a buffer
3084 {
3085 TracePhase tp(_t_output);
3086 PhaseOutput output;
3087 output.Output();
3088 if (failing()) return;
3089 output.install();
3090 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3091 }
3092
3093 // He's dead, Jim.
3094 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3095 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3096 }
3097
3098 //------------------------------Final_Reshape_Counts---------------------------
3099 // This class defines counters to help identify when a method
3100 // may/must be executed using hardware with only 24-bit precision.
3101 struct Final_Reshape_Counts : public StackObj {
3102 int _call_count; // count non-inlined 'common' calls
3103 int _float_count; // count float ops requiring 24-bit precision
3104 int _double_count; // count double ops requiring more precision
3105 int _java_call_count; // count non-inlined 'java' calls
3106 int _inner_loop_count; // count loops which need alignment
3107 VectorSet _visited; // Visitation flags
3108 Node_List _tests; // Set of IfNodes & PCTableNodes
3109
3110 Final_Reshape_Counts() :
3111 _call_count(0), _float_count(0), _double_count(0),
3112 _java_call_count(0), _inner_loop_count(0) { }
3113
3114 void inc_call_count () { _call_count ++; }
3115 void inc_float_count () { _float_count ++; }
3116 void inc_double_count() { _double_count++; }
3117 void inc_java_call_count() { _java_call_count++; }
3118 void inc_inner_loop_count() { _inner_loop_count++; }
3119
3120 int get_call_count () const { return _call_count ; }
3121 int get_float_count () const { return _float_count ; }
3122 int get_double_count() const { return _double_count; }
3123 int get_java_call_count() const { return _java_call_count; }
3124 int get_inner_loop_count() const { return _inner_loop_count; }
3125 };
3126
3127 //------------------------------final_graph_reshaping_impl----------------------
3128 // Implement items 1-5 from final_graph_reshaping below.
3129 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3130
3131 if ( n->outcnt() == 0 ) return; // dead node
3132 uint nop = n->Opcode();
3133
3134 // Check for 2-input instruction with "last use" on right input.
3135 // Swap to left input. Implements item (2).
3136 if( n->req() == 3 && // two-input instruction
3137 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3138 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3139 n->in(2)->outcnt() == 1 &&// right use IS a last use
3140 !n->in(2)->is_Con() ) { // right use is not a constant
3141 // Check for commutative opcode
3142 switch( nop ) {
3143 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3144 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3145 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3146 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3147 case Op_AndL: case Op_XorL: case Op_OrL:
3148 case Op_AndI: case Op_XorI: case Op_OrI: {
3149 // Move "last use" input to left by swapping inputs
3150 n->swap_edges(1, 2);
3151 break;
3152 }
3153 default:
3154 break;
3155 }
3156 }
3157
3158 #ifdef ASSERT
3159 if( n->is_Mem() ) {
3160 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3161 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3162 // oop will be recorded in oop map if load crosses safepoint
3163 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3164 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3165 "raw memory operations should have control edge");
3166 }
3167 if (n->is_MemBar()) {
3168 MemBarNode* mb = n->as_MemBar();
3169 if (mb->trailing_store() || mb->trailing_load_store()) {
3170 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3171 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3172 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3173 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3174 } else if (mb->leading()) {
3175 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3176 }
3177 }
3178 #endif
3179 // Count FPU ops and common calls, implements item (3)
3180 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3181 if (!gc_handled) {
3182 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3183 }
3184
3185 // Collect CFG split points
3186 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3187 frc._tests.push(n);
3188 }
3189 }
3190
3191 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3192 if (!UseDivMod) {
3193 return;
3194 }
3195
3196 // Check if "a % b" and "a / b" both exist
3197 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3198 if (d == nullptr) {
3199 return;
3200 }
3201
3202 // Replace them with a fused divmod if supported
3203 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3204 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3205 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3206 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3207 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3208 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3209 // DivMod node so the dependency is not lost.
3210 divmod->add_prec_from(n);
3211 divmod->add_prec_from(d);
3212 d->subsume_by(divmod->div_proj(), this);
3213 n->subsume_by(divmod->mod_proj(), this);
3214 } else {
3215 // Replace "a % b" with "a - ((a / b) * b)"
3216 Node* mult = MulNode::make(d, d->in(2), bt);
3217 Node* sub = SubNode::make(d->in(1), mult, bt);
3218 n->subsume_by(sub, this);
3219 }
3220 }
3221
3222 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3223 switch( nop ) {
3224 // Count all float operations that may use FPU
3225 case Op_AddF:
3226 case Op_SubF:
3227 case Op_MulF:
3228 case Op_DivF:
3229 case Op_NegF:
3230 case Op_ModF:
3231 case Op_ConvI2F:
3232 case Op_ConF:
3233 case Op_CmpF:
3234 case Op_CmpF3:
3235 case Op_StoreF:
3236 case Op_LoadF:
3237 // case Op_ConvL2F: // longs are split into 32-bit halves
3238 frc.inc_float_count();
3239 break;
3240
3241 case Op_ConvF2D:
3242 case Op_ConvD2F:
3243 frc.inc_float_count();
3244 frc.inc_double_count();
3245 break;
3246
3247 // Count all double operations that may use FPU
3248 case Op_AddD:
3249 case Op_SubD:
3250 case Op_MulD:
3251 case Op_DivD:
3252 case Op_NegD:
3253 case Op_ModD:
3254 case Op_ConvI2D:
3255 case Op_ConvD2I:
3256 // case Op_ConvL2D: // handled by leaf call
3257 // case Op_ConvD2L: // handled by leaf call
3258 case Op_ConD:
3259 case Op_CmpD:
3260 case Op_CmpD3:
3261 case Op_StoreD:
3262 case Op_LoadD:
3263 case Op_LoadD_unaligned:
3264 frc.inc_double_count();
3265 break;
3266 case Op_Opaque1: // Remove Opaque Nodes before matching
3267 n->subsume_by(n->in(1), this);
3268 break;
3269 case Op_CallStaticJava:
3270 case Op_CallJava:
3271 case Op_CallDynamicJava:
3272 frc.inc_java_call_count(); // Count java call site;
3273 case Op_CallRuntime:
3274 case Op_CallLeaf:
3275 case Op_CallLeafVector:
3276 case Op_CallLeafNoFP: {
3277 assert (n->is_Call(), "");
3278 CallNode *call = n->as_Call();
3279 // Count call sites where the FP mode bit would have to be flipped.
3280 // Do not count uncommon runtime calls:
3281 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3282 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3283 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3284 frc.inc_call_count(); // Count the call site
3285 } else { // See if uncommon argument is shared
3286 Node *n = call->in(TypeFunc::Parms);
3287 int nop = n->Opcode();
3288 // Clone shared simple arguments to uncommon calls, item (1).
3289 if (n->outcnt() > 1 &&
3290 !n->is_Proj() &&
3291 nop != Op_CreateEx &&
3292 nop != Op_CheckCastPP &&
3293 nop != Op_DecodeN &&
3294 nop != Op_DecodeNKlass &&
3295 !n->is_Mem() &&
3296 !n->is_Phi()) {
3297 Node *x = n->clone();
3298 call->set_req(TypeFunc::Parms, x);
3299 }
3300 }
3301 break;
3302 }
3303 case Op_StoreB:
3304 case Op_StoreC:
3305 case Op_StoreI:
3306 case Op_StoreL:
3307 case Op_CompareAndSwapB:
3308 case Op_CompareAndSwapS:
3309 case Op_CompareAndSwapI:
3310 case Op_CompareAndSwapL:
3311 case Op_CompareAndSwapP:
3312 case Op_CompareAndSwapN:
3313 case Op_WeakCompareAndSwapB:
3314 case Op_WeakCompareAndSwapS:
3315 case Op_WeakCompareAndSwapI:
3316 case Op_WeakCompareAndSwapL:
3317 case Op_WeakCompareAndSwapP:
3318 case Op_WeakCompareAndSwapN:
3319 case Op_CompareAndExchangeB:
3320 case Op_CompareAndExchangeS:
3321 case Op_CompareAndExchangeI:
3322 case Op_CompareAndExchangeL:
3323 case Op_CompareAndExchangeP:
3324 case Op_CompareAndExchangeN:
3325 case Op_GetAndAddS:
3326 case Op_GetAndAddB:
3327 case Op_GetAndAddI:
3328 case Op_GetAndAddL:
3329 case Op_GetAndSetS:
3330 case Op_GetAndSetB:
3331 case Op_GetAndSetI:
3332 case Op_GetAndSetL:
3333 case Op_GetAndSetP:
3334 case Op_GetAndSetN:
3335 case Op_StoreP:
3336 case Op_StoreN:
3337 case Op_StoreNKlass:
3338 case Op_LoadB:
3339 case Op_LoadUB:
3340 case Op_LoadUS:
3341 case Op_LoadI:
3342 case Op_LoadKlass:
3343 case Op_LoadNKlass:
3344 case Op_LoadL:
3345 case Op_LoadL_unaligned:
3346 case Op_LoadP:
3347 case Op_LoadN:
3348 case Op_LoadRange:
3349 case Op_LoadS:
3350 break;
3351
3352 case Op_AddP: { // Assert sane base pointers
3353 Node *addp = n->in(AddPNode::Address);
3354 assert( !addp->is_AddP() ||
3355 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3356 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3357 "Base pointers must match (addp %u)", addp->_idx );
3358 #ifdef _LP64
3359 if ((UseCompressedOops || UseCompressedClassPointers) &&
3360 addp->Opcode() == Op_ConP &&
3361 addp == n->in(AddPNode::Base) &&
3362 n->in(AddPNode::Offset)->is_Con()) {
3363 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3364 // on the platform and on the compressed oops mode.
3365 // Use addressing with narrow klass to load with offset on x86.
3366 // Some platforms can use the constant pool to load ConP.
3367 // Do this transformation here since IGVN will convert ConN back to ConP.
3368 const Type* t = addp->bottom_type();
3369 bool is_oop = t->isa_oopptr() != nullptr;
3370 bool is_klass = t->isa_klassptr() != nullptr;
3371
3372 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3373 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3374 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3375 Node* nn = nullptr;
3376
3377 int op = is_oop ? Op_ConN : Op_ConNKlass;
3378
3379 // Look for existing ConN node of the same exact type.
3380 Node* r = root();
3381 uint cnt = r->outcnt();
3382 for (uint i = 0; i < cnt; i++) {
3383 Node* m = r->raw_out(i);
3384 if (m!= nullptr && m->Opcode() == op &&
3385 m->bottom_type()->make_ptr() == t) {
3386 nn = m;
3387 break;
3388 }
3389 }
3390 if (nn != nullptr) {
3391 // Decode a narrow oop to match address
3392 // [R12 + narrow_oop_reg<<3 + offset]
3393 if (is_oop) {
3394 nn = new DecodeNNode(nn, t);
3395 } else {
3396 nn = new DecodeNKlassNode(nn, t);
3397 }
3398 // Check for succeeding AddP which uses the same Base.
3399 // Otherwise we will run into the assertion above when visiting that guy.
3400 for (uint i = 0; i < n->outcnt(); ++i) {
3401 Node *out_i = n->raw_out(i);
3402 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3403 out_i->set_req(AddPNode::Base, nn);
3404 #ifdef ASSERT
3405 for (uint j = 0; j < out_i->outcnt(); ++j) {
3406 Node *out_j = out_i->raw_out(j);
3407 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3408 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3409 }
3410 #endif
3411 }
3412 }
3413 n->set_req(AddPNode::Base, nn);
3414 n->set_req(AddPNode::Address, nn);
3415 if (addp->outcnt() == 0) {
3416 addp->disconnect_inputs(this);
3417 }
3418 }
3419 }
3420 }
3421 #endif
3422 break;
3423 }
3424
3425 case Op_CastPP: {
3426 // Remove CastPP nodes to gain more freedom during scheduling but
3427 // keep the dependency they encode as control or precedence edges
3428 // (if control is set already) on memory operations. Some CastPP
3429 // nodes don't have a control (don't carry a dependency): skip
3430 // those.
3431 if (n->in(0) != nullptr) {
3432 ResourceMark rm;
3433 Unique_Node_List wq;
3434 wq.push(n);
3435 for (uint next = 0; next < wq.size(); ++next) {
3436 Node *m = wq.at(next);
3437 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3438 Node* use = m->fast_out(i);
3439 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3440 use->ensure_control_or_add_prec(n->in(0));
3441 } else {
3442 switch(use->Opcode()) {
3443 case Op_AddP:
3444 case Op_DecodeN:
3445 case Op_DecodeNKlass:
3446 case Op_CheckCastPP:
3447 case Op_CastPP:
3448 wq.push(use);
3449 break;
3450 }
3451 }
3452 }
3453 }
3454 }
3455 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3456 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3457 Node* in1 = n->in(1);
3458 const Type* t = n->bottom_type();
3459 Node* new_in1 = in1->clone();
3460 new_in1->as_DecodeN()->set_type(t);
3461
3462 if (!Matcher::narrow_oop_use_complex_address()) {
3463 //
3464 // x86, ARM and friends can handle 2 adds in addressing mode
3465 // and Matcher can fold a DecodeN node into address by using
3466 // a narrow oop directly and do implicit null check in address:
3467 //
3468 // [R12 + narrow_oop_reg<<3 + offset]
3469 // NullCheck narrow_oop_reg
3470 //
3471 // On other platforms (Sparc) we have to keep new DecodeN node and
3472 // use it to do implicit null check in address:
3473 //
3474 // decode_not_null narrow_oop_reg, base_reg
3475 // [base_reg + offset]
3476 // NullCheck base_reg
3477 //
3478 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3479 // to keep the information to which null check the new DecodeN node
3480 // corresponds to use it as value in implicit_null_check().
3481 //
3482 new_in1->set_req(0, n->in(0));
3483 }
3484
3485 n->subsume_by(new_in1, this);
3486 if (in1->outcnt() == 0) {
3487 in1->disconnect_inputs(this);
3488 }
3489 } else {
3490 n->subsume_by(n->in(1), this);
3491 if (n->outcnt() == 0) {
3492 n->disconnect_inputs(this);
3493 }
3494 }
3495 break;
3496 }
3497 case Op_CastII: {
3498 n->as_CastII()->remove_range_check_cast(this);
3499 break;
3500 }
3501 #ifdef _LP64
3502 case Op_CmpP:
3503 // Do this transformation here to preserve CmpPNode::sub() and
3504 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3505 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3506 Node* in1 = n->in(1);
3507 Node* in2 = n->in(2);
3508 if (!in1->is_DecodeNarrowPtr()) {
3509 in2 = in1;
3510 in1 = n->in(2);
3511 }
3512 assert(in1->is_DecodeNarrowPtr(), "sanity");
3513
3514 Node* new_in2 = nullptr;
3515 if (in2->is_DecodeNarrowPtr()) {
3516 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3517 new_in2 = in2->in(1);
3518 } else if (in2->Opcode() == Op_ConP) {
3519 const Type* t = in2->bottom_type();
3520 if (t == TypePtr::NULL_PTR) {
3521 assert(in1->is_DecodeN(), "compare klass to null?");
3522 // Don't convert CmpP null check into CmpN if compressed
3523 // oops implicit null check is not generated.
3524 // This will allow to generate normal oop implicit null check.
3525 if (Matcher::gen_narrow_oop_implicit_null_checks())
3526 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3527 //
3528 // This transformation together with CastPP transformation above
3529 // will generated code for implicit null checks for compressed oops.
3530 //
3531 // The original code after Optimize()
3532 //
3533 // LoadN memory, narrow_oop_reg
3534 // decode narrow_oop_reg, base_reg
3535 // CmpP base_reg, nullptr
3536 // CastPP base_reg // NotNull
3537 // Load [base_reg + offset], val_reg
3538 //
3539 // after these transformations will be
3540 //
3541 // LoadN memory, narrow_oop_reg
3542 // CmpN narrow_oop_reg, nullptr
3543 // decode_not_null narrow_oop_reg, base_reg
3544 // Load [base_reg + offset], val_reg
3545 //
3546 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3547 // since narrow oops can be used in debug info now (see the code in
3548 // final_graph_reshaping_walk()).
3549 //
3550 // At the end the code will be matched to
3551 // on x86:
3552 //
3553 // Load_narrow_oop memory, narrow_oop_reg
3554 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3555 // NullCheck narrow_oop_reg
3556 //
3557 // and on sparc:
3558 //
3559 // Load_narrow_oop memory, narrow_oop_reg
3560 // decode_not_null narrow_oop_reg, base_reg
3561 // Load [base_reg + offset], val_reg
3562 // NullCheck base_reg
3563 //
3564 } else if (t->isa_oopptr()) {
3565 new_in2 = ConNode::make(t->make_narrowoop());
3566 } else if (t->isa_klassptr()) {
3567 new_in2 = ConNode::make(t->make_narrowklass());
3568 }
3569 }
3570 if (new_in2 != nullptr) {
3571 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3572 n->subsume_by(cmpN, this);
3573 if (in1->outcnt() == 0) {
3574 in1->disconnect_inputs(this);
3575 }
3576 if (in2->outcnt() == 0) {
3577 in2->disconnect_inputs(this);
3578 }
3579 }
3580 }
3581 break;
3582
3583 case Op_DecodeN:
3584 case Op_DecodeNKlass:
3585 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3586 // DecodeN could be pinned when it can't be fold into
3587 // an address expression, see the code for Op_CastPP above.
3588 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3589 break;
3590
3591 case Op_EncodeP:
3592 case Op_EncodePKlass: {
3593 Node* in1 = n->in(1);
3594 if (in1->is_DecodeNarrowPtr()) {
3595 n->subsume_by(in1->in(1), this);
3596 } else if (in1->Opcode() == Op_ConP) {
3597 const Type* t = in1->bottom_type();
3598 if (t == TypePtr::NULL_PTR) {
3599 assert(t->isa_oopptr(), "null klass?");
3600 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3601 } else if (t->isa_oopptr()) {
3602 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3603 } else if (t->isa_klassptr()) {
3604 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3605 }
3606 }
3607 if (in1->outcnt() == 0) {
3608 in1->disconnect_inputs(this);
3609 }
3610 break;
3611 }
3612
3613 case Op_Proj: {
3614 if (OptimizeStringConcat || IncrementalInline) {
3615 ProjNode* proj = n->as_Proj();
3616 if (proj->_is_io_use) {
3617 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3618 // Separate projections were used for the exception path which
3619 // are normally removed by a late inline. If it wasn't inlined
3620 // then they will hang around and should just be replaced with
3621 // the original one. Merge them.
3622 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3623 if (non_io_proj != nullptr) {
3624 proj->subsume_by(non_io_proj , this);
3625 }
3626 }
3627 }
3628 break;
3629 }
3630
3631 case Op_Phi:
3632 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3633 // The EncodeP optimization may create Phi with the same edges
3634 // for all paths. It is not handled well by Register Allocator.
3635 Node* unique_in = n->in(1);
3636 assert(unique_in != nullptr, "");
3637 uint cnt = n->req();
3638 for (uint i = 2; i < cnt; i++) {
3639 Node* m = n->in(i);
3640 assert(m != nullptr, "");
3641 if (unique_in != m)
3642 unique_in = nullptr;
3643 }
3644 if (unique_in != nullptr) {
3645 n->subsume_by(unique_in, this);
3646 }
3647 }
3648 break;
3649
3650 #endif
3651
3652 case Op_ModI:
3653 handle_div_mod_op(n, T_INT, false);
3654 break;
3655
3656 case Op_ModL:
3657 handle_div_mod_op(n, T_LONG, false);
3658 break;
3659
3660 case Op_UModI:
3661 handle_div_mod_op(n, T_INT, true);
3662 break;
3663
3664 case Op_UModL:
3665 handle_div_mod_op(n, T_LONG, true);
3666 break;
3667
3668 case Op_LoadVector:
3669 case Op_StoreVector:
3670 #ifdef ASSERT
3671 // Add VerifyVectorAlignment node between adr and load / store.
3672 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3673 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3674 n->as_StoreVector()->must_verify_alignment();
3675 if (must_verify_alignment) {
3676 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3677 n->as_StoreVector()->memory_size();
3678 // The memory access should be aligned to the vector width in bytes.
3679 // However, the underlying array is possibly less well aligned, but at least
3680 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3681 // a loop we can expect at least the following alignment:
3682 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3683 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3684 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3685 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3686 jlong mask = guaranteed_alignment - 1;
3687 Node* mask_con = ConLNode::make(mask);
3688 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3689 n->set_req(MemNode::Address, va);
3690 }
3691 }
3692 #endif
3693 break;
3694
3695 case Op_LoadVectorGather:
3696 case Op_StoreVectorScatter:
3697 case Op_LoadVectorGatherMasked:
3698 case Op_StoreVectorScatterMasked:
3699 case Op_VectorCmpMasked:
3700 case Op_VectorMaskGen:
3701 case Op_LoadVectorMasked:
3702 case Op_StoreVectorMasked:
3703 break;
3704
3705 case Op_AddReductionVI:
3706 case Op_AddReductionVL:
3707 case Op_AddReductionVF:
3708 case Op_AddReductionVD:
3709 case Op_MulReductionVI:
3710 case Op_MulReductionVL:
3711 case Op_MulReductionVF:
3712 case Op_MulReductionVD:
3713 case Op_MinReductionV:
3714 case Op_MaxReductionV:
3715 case Op_AndReductionV:
3716 case Op_OrReductionV:
3717 case Op_XorReductionV:
3718 break;
3719
3720 case Op_PackB:
3721 case Op_PackS:
3722 case Op_PackI:
3723 case Op_PackF:
3724 case Op_PackL:
3725 case Op_PackD:
3726 if (n->req()-1 > 2) {
3727 // Replace many operand PackNodes with a binary tree for matching
3728 PackNode* p = (PackNode*) n;
3729 Node* btp = p->binary_tree_pack(1, n->req());
3730 n->subsume_by(btp, this);
3731 }
3732 break;
3733 case Op_Loop:
3734 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3735 case Op_CountedLoop:
3736 case Op_LongCountedLoop:
3737 case Op_OuterStripMinedLoop:
3738 if (n->as_Loop()->is_inner_loop()) {
3739 frc.inc_inner_loop_count();
3740 }
3741 n->as_Loop()->verify_strip_mined(0);
3742 break;
3743 case Op_LShiftI:
3744 case Op_RShiftI:
3745 case Op_URShiftI:
3746 case Op_LShiftL:
3747 case Op_RShiftL:
3748 case Op_URShiftL:
3749 if (Matcher::need_masked_shift_count) {
3750 // The cpu's shift instructions don't restrict the count to the
3751 // lower 5/6 bits. We need to do the masking ourselves.
3752 Node* in2 = n->in(2);
3753 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3754 const TypeInt* t = in2->find_int_type();
3755 if (t != nullptr && t->is_con()) {
3756 juint shift = t->get_con();
3757 if (shift > mask) { // Unsigned cmp
3758 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3759 }
3760 } else {
3761 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3762 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3763 n->set_req(2, shift);
3764 }
3765 }
3766 if (in2->outcnt() == 0) { // Remove dead node
3767 in2->disconnect_inputs(this);
3768 }
3769 }
3770 break;
3771 case Op_MemBarStoreStore:
3772 case Op_MemBarRelease:
3773 // Break the link with AllocateNode: it is no longer useful and
3774 // confuses register allocation.
3775 if (n->req() > MemBarNode::Precedent) {
3776 n->set_req(MemBarNode::Precedent, top());
3777 }
3778 break;
3779 case Op_MemBarAcquire: {
3780 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3781 // At parse time, the trailing MemBarAcquire for a volatile load
3782 // is created with an edge to the load. After optimizations,
3783 // that input may be a chain of Phis. If those phis have no
3784 // other use, then the MemBarAcquire keeps them alive and
3785 // register allocation can be confused.
3786 dead_nodes.push(n->in(MemBarNode::Precedent));
3787 n->set_req(MemBarNode::Precedent, top());
3788 }
3789 break;
3790 }
3791 case Op_Blackhole:
3792 break;
3793 case Op_RangeCheck: {
3794 RangeCheckNode* rc = n->as_RangeCheck();
3795 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3796 n->subsume_by(iff, this);
3797 frc._tests.push(iff);
3798 break;
3799 }
3800 case Op_ConvI2L: {
3801 if (!Matcher::convi2l_type_required) {
3802 // Code generation on some platforms doesn't need accurate
3803 // ConvI2L types. Widening the type can help remove redundant
3804 // address computations.
3805 n->as_Type()->set_type(TypeLong::INT);
3806 ResourceMark rm;
3807 Unique_Node_List wq;
3808 wq.push(n);
3809 for (uint next = 0; next < wq.size(); next++) {
3810 Node *m = wq.at(next);
3811
3812 for(;;) {
3813 // Loop over all nodes with identical inputs edges as m
3814 Node* k = m->find_similar(m->Opcode());
3815 if (k == nullptr) {
3816 break;
3817 }
3818 // Push their uses so we get a chance to remove node made
3819 // redundant
3820 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3821 Node* u = k->fast_out(i);
3822 if (u->Opcode() == Op_LShiftL ||
3823 u->Opcode() == Op_AddL ||
3824 u->Opcode() == Op_SubL ||
3825 u->Opcode() == Op_AddP) {
3826 wq.push(u);
3827 }
3828 }
3829 // Replace all nodes with identical edges as m with m
3830 k->subsume_by(m, this);
3831 }
3832 }
3833 }
3834 break;
3835 }
3836 case Op_CmpUL: {
3837 if (!Matcher::has_match_rule(Op_CmpUL)) {
3838 // No support for unsigned long comparisons
3839 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3840 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3841 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3842 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3843 Node* andl = new AndLNode(orl, remove_sign_mask);
3844 Node* cmp = new CmpLNode(andl, n->in(2));
3845 n->subsume_by(cmp, this);
3846 }
3847 break;
3848 }
3849 #ifdef ASSERT
3850 case Op_ConNKlass: {
3851 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3852 ciKlass* klass = tp->is_klassptr()->exact_klass();
3853 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3854 break;
3855 }
3856 #endif
3857 default:
3858 assert(!n->is_Call(), "");
3859 assert(!n->is_Mem(), "");
3860 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3861 break;
3862 }
3863 }
3864
3865 //------------------------------final_graph_reshaping_walk---------------------
3866 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3867 // requires that the walk visits a node's inputs before visiting the node.
3868 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3869 Unique_Node_List sfpt;
3870
3871 frc._visited.set(root->_idx); // first, mark node as visited
3872 uint cnt = root->req();
3873 Node *n = root;
3874 uint i = 0;
3875 while (true) {
3876 if (i < cnt) {
3877 // Place all non-visited non-null inputs onto stack
3878 Node* m = n->in(i);
3879 ++i;
3880 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3881 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3882 // compute worst case interpreter size in case of a deoptimization
3883 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3884
3885 sfpt.push(m);
3886 }
3887 cnt = m->req();
3888 nstack.push(n, i); // put on stack parent and next input's index
3889 n = m;
3890 i = 0;
3891 }
3892 } else {
3893 // Now do post-visit work
3894 final_graph_reshaping_impl(n, frc, dead_nodes);
3895 if (nstack.is_empty())
3896 break; // finished
3897 n = nstack.node(); // Get node from stack
3898 cnt = n->req();
3899 i = nstack.index();
3900 nstack.pop(); // Shift to the next node on stack
3901 }
3902 }
3903
3904 // Skip next transformation if compressed oops are not used.
3905 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3906 (!UseCompressedOops && !UseCompressedClassPointers))
3907 return;
3908
3909 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3910 // It could be done for an uncommon traps or any safepoints/calls
3911 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3912 while (sfpt.size() > 0) {
3913 n = sfpt.pop();
3914 JVMState *jvms = n->as_SafePoint()->jvms();
3915 assert(jvms != nullptr, "sanity");
3916 int start = jvms->debug_start();
3917 int end = n->req();
3918 bool is_uncommon = (n->is_CallStaticJava() &&
3919 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3920 for (int j = start; j < end; j++) {
3921 Node* in = n->in(j);
3922 if (in->is_DecodeNarrowPtr()) {
3923 bool safe_to_skip = true;
3924 if (!is_uncommon ) {
3925 // Is it safe to skip?
3926 for (uint i = 0; i < in->outcnt(); i++) {
3927 Node* u = in->raw_out(i);
3928 if (!u->is_SafePoint() ||
3929 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3930 safe_to_skip = false;
3931 }
3932 }
3933 }
3934 if (safe_to_skip) {
3935 n->set_req(j, in->in(1));
3936 }
3937 if (in->outcnt() == 0) {
3938 in->disconnect_inputs(this);
3939 }
3940 }
3941 }
3942 }
3943 }
3944
3945 //------------------------------final_graph_reshaping--------------------------
3946 // Final Graph Reshaping.
3947 //
3948 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3949 // and not commoned up and forced early. Must come after regular
3950 // optimizations to avoid GVN undoing the cloning. Clone constant
3951 // inputs to Loop Phis; these will be split by the allocator anyways.
3952 // Remove Opaque nodes.
3953 // (2) Move last-uses by commutative operations to the left input to encourage
3954 // Intel update-in-place two-address operations and better register usage
3955 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3956 // calls canonicalizing them back.
3957 // (3) Count the number of double-precision FP ops, single-precision FP ops
3958 // and call sites. On Intel, we can get correct rounding either by
3959 // forcing singles to memory (requires extra stores and loads after each
3960 // FP bytecode) or we can set a rounding mode bit (requires setting and
3961 // clearing the mode bit around call sites). The mode bit is only used
3962 // if the relative frequency of single FP ops to calls is low enough.
3963 // This is a key transform for SPEC mpeg_audio.
3964 // (4) Detect infinite loops; blobs of code reachable from above but not
3965 // below. Several of the Code_Gen algorithms fail on such code shapes,
3966 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3967 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3968 // Detection is by looking for IfNodes where only 1 projection is
3969 // reachable from below or CatchNodes missing some targets.
3970 // (5) Assert for insane oop offsets in debug mode.
3971
3972 bool Compile::final_graph_reshaping() {
3973 // an infinite loop may have been eliminated by the optimizer,
3974 // in which case the graph will be empty.
3975 if (root()->req() == 1) {
3976 // Do not compile method that is only a trivial infinite loop,
3977 // since the content of the loop may have been eliminated.
3978 record_method_not_compilable("trivial infinite loop");
3979 return true;
3980 }
3981
3982 // Expensive nodes have their control input set to prevent the GVN
3983 // from freely commoning them. There's no GVN beyond this point so
3984 // no need to keep the control input. We want the expensive nodes to
3985 // be freely moved to the least frequent code path by gcm.
3986 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3987 for (int i = 0; i < expensive_count(); i++) {
3988 _expensive_nodes.at(i)->set_req(0, nullptr);
3989 }
3990
3991 Final_Reshape_Counts frc;
3992
3993 // Visit everybody reachable!
3994 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3995 Node_Stack nstack(live_nodes() >> 1);
3996 Unique_Node_List dead_nodes;
3997 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
3998
3999 // Check for unreachable (from below) code (i.e., infinite loops).
4000 for( uint i = 0; i < frc._tests.size(); i++ ) {
4001 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4002 // Get number of CFG targets.
4003 // Note that PCTables include exception targets after calls.
4004 uint required_outcnt = n->required_outcnt();
4005 if (n->outcnt() != required_outcnt) {
4006 // Check for a few special cases. Rethrow Nodes never take the
4007 // 'fall-thru' path, so expected kids is 1 less.
4008 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4009 if (n->in(0)->in(0)->is_Call()) {
4010 CallNode* call = n->in(0)->in(0)->as_Call();
4011 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4012 required_outcnt--; // Rethrow always has 1 less kid
4013 } else if (call->req() > TypeFunc::Parms &&
4014 call->is_CallDynamicJava()) {
4015 // Check for null receiver. In such case, the optimizer has
4016 // detected that the virtual call will always result in a null
4017 // pointer exception. The fall-through projection of this CatchNode
4018 // will not be populated.
4019 Node* arg0 = call->in(TypeFunc::Parms);
4020 if (arg0->is_Type() &&
4021 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4022 required_outcnt--;
4023 }
4024 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4025 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4026 // Check for illegal array length. In such case, the optimizer has
4027 // detected that the allocation attempt will always result in an
4028 // exception. There is no fall-through projection of this CatchNode .
4029 assert(call->is_CallStaticJava(), "static call expected");
4030 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4031 uint valid_length_test_input = call->req() - 1;
4032 Node* valid_length_test = call->in(valid_length_test_input);
4033 call->del_req(valid_length_test_input);
4034 if (valid_length_test->find_int_con(1) == 0) {
4035 required_outcnt--;
4036 }
4037 dead_nodes.push(valid_length_test);
4038 assert(n->outcnt() == required_outcnt, "malformed control flow");
4039 continue;
4040 }
4041 }
4042 }
4043
4044 // Recheck with a better notion of 'required_outcnt'
4045 if (n->outcnt() != required_outcnt) {
4046 record_method_not_compilable("malformed control flow");
4047 return true; // Not all targets reachable!
4048 }
4049 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4050 CallNode* call = n->in(0)->in(0)->as_Call();
4051 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4052 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4053 assert(call->is_CallStaticJava(), "static call expected");
4054 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4055 uint valid_length_test_input = call->req() - 1;
4056 dead_nodes.push(call->in(valid_length_test_input));
4057 call->del_req(valid_length_test_input); // valid length test useless now
4058 }
4059 }
4060 // Check that I actually visited all kids. Unreached kids
4061 // must be infinite loops.
4062 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4063 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4064 record_method_not_compilable("infinite loop");
4065 return true; // Found unvisited kid; must be unreach
4066 }
4067
4068 // Here so verification code in final_graph_reshaping_walk()
4069 // always see an OuterStripMinedLoopEnd
4070 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4071 IfNode* init_iff = n->as_If();
4072 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4073 n->subsume_by(iff, this);
4074 }
4075 }
4076
4077 while (dead_nodes.size() > 0) {
4078 Node* m = dead_nodes.pop();
4079 if (m->outcnt() == 0 && m != top()) {
4080 for (uint j = 0; j < m->req(); j++) {
4081 Node* in = m->in(j);
4082 if (in != nullptr) {
4083 dead_nodes.push(in);
4084 }
4085 }
4086 m->disconnect_inputs(this);
4087 }
4088 }
4089
4090 #ifdef IA32
4091 // If original bytecodes contained a mixture of floats and doubles
4092 // check if the optimizer has made it homogeneous, item (3).
4093 if (UseSSE == 0 &&
4094 frc.get_float_count() > 32 &&
4095 frc.get_double_count() == 0 &&
4096 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4097 set_24_bit_selection_and_mode(false, true);
4098 }
4099 #endif // IA32
4100
4101 set_java_calls(frc.get_java_call_count());
4102 set_inner_loops(frc.get_inner_loop_count());
4103
4104 // No infinite loops, no reason to bail out.
4105 return false;
4106 }
4107
4108 //-----------------------------too_many_traps----------------------------------
4109 // Report if there are too many traps at the current method and bci.
4110 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4111 bool Compile::too_many_traps(ciMethod* method,
4112 int bci,
4113 Deoptimization::DeoptReason reason) {
4114 if (method->has_trap_at(bci)) {
4115 return true;
4116 }
4117 ciMethodData* md = method->method_data();
4118 if (md->is_empty()) {
4119 // Assume the trap has not occurred, or that it occurred only
4120 // because of a transient condition during start-up in the interpreter.
4121 return false;
4122 }
4123 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4124 if (md->has_trap_at(bci, m, reason) != 0) {
4125 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4126 // Also, if there are multiple reasons, or if there is no per-BCI record,
4127 // assume the worst.
4128 if (log())
4129 log()->elem("observe trap='%s' count='%d'",
4130 Deoptimization::trap_reason_name(reason),
4131 md->trap_count(reason));
4132 return true;
4133 } else {
4134 // Ignore method/bci and see if there have been too many globally.
4135 return too_many_traps(reason, md);
4136 }
4137 }
4138
4139 // Less-accurate variant which does not require a method and bci.
4140 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4141 ciMethodData* logmd) {
4142 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4143 // Too many traps globally.
4144 // Note that we use cumulative trap_count, not just md->trap_count.
4145 if (log()) {
4146 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4147 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4148 Deoptimization::trap_reason_name(reason),
4149 mcount, trap_count(reason));
4150 }
4151 return true;
4152 } else {
4153 // The coast is clear.
4154 return false;
4155 }
4156 }
4157
4158 //--------------------------too_many_recompiles--------------------------------
4159 // Report if there are too many recompiles at the current method and bci.
4160 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4161 // Is not eager to return true, since this will cause the compiler to use
4162 // Action_none for a trap point, to avoid too many recompilations.
4163 bool Compile::too_many_recompiles(ciMethod* method,
4164 int bci,
4165 Deoptimization::DeoptReason reason) {
4166 ciMethodData* md = method->method_data();
4167 if (md->is_empty()) {
4168 // Assume the trap has not occurred, or that it occurred only
4169 // because of a transient condition during start-up in the interpreter.
4170 return false;
4171 }
4172 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4173 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4174 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4175 Deoptimization::DeoptReason per_bc_reason
4176 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4177 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4178 if ((per_bc_reason == Deoptimization::Reason_none
4179 || md->has_trap_at(bci, m, reason) != 0)
4180 // The trap frequency measure we care about is the recompile count:
4181 && md->trap_recompiled_at(bci, m)
4182 && md->overflow_recompile_count() >= bc_cutoff) {
4183 // Do not emit a trap here if it has already caused recompilations.
4184 // Also, if there are multiple reasons, or if there is no per-BCI record,
4185 // assume the worst.
4186 if (log())
4187 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4188 Deoptimization::trap_reason_name(reason),
4189 md->trap_count(reason),
4190 md->overflow_recompile_count());
4191 return true;
4192 } else if (trap_count(reason) != 0
4193 && decompile_count() >= m_cutoff) {
4194 // Too many recompiles globally, and we have seen this sort of trap.
4195 // Use cumulative decompile_count, not just md->decompile_count.
4196 if (log())
4197 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4198 Deoptimization::trap_reason_name(reason),
4199 md->trap_count(reason), trap_count(reason),
4200 md->decompile_count(), decompile_count());
4201 return true;
4202 } else {
4203 // The coast is clear.
4204 return false;
4205 }
4206 }
4207
4208 // Compute when not to trap. Used by matching trap based nodes and
4209 // NullCheck optimization.
4210 void Compile::set_allowed_deopt_reasons() {
4211 _allowed_reasons = 0;
4212 if (is_method_compilation()) {
4213 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4214 assert(rs < BitsPerInt, "recode bit map");
4215 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4216 _allowed_reasons |= nth_bit(rs);
4217 }
4218 }
4219 }
4220 }
4221
4222 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4223 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4224 }
4225
4226 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4227 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4228 }
4229
4230 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4231 if (holder->is_initialized() && !do_clinit_barriers()) {
4232 return false;
4233 }
4234 if (holder->is_being_initialized() || do_clinit_barriers()) {
4235 if (accessing_method->holder() == holder) {
4236 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4237 // <init>, or a static method. In all those cases, there was an initialization
4238 // barrier on the holder klass passed.
4239 if (accessing_method->is_static_initializer() ||
4240 accessing_method->is_object_initializer() ||
4241 accessing_method->is_static()) {
4242 return false;
4243 }
4244 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4245 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4246 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4247 // child class can become fully initialized while its parent class is still being initialized.
4248 if (accessing_method->is_static_initializer()) {
4249 return false;
4250 }
4251 }
4252 ciMethod* root = method(); // the root method of compilation
4253 if (root != accessing_method) {
4254 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4255 }
4256 }
4257 return true;
4258 }
4259
4260 #ifndef PRODUCT
4261 //------------------------------verify_bidirectional_edges---------------------
4262 // For each input edge to a node (ie - for each Use-Def edge), verify that
4263 // there is a corresponding Def-Use edge.
4264 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4265 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4266 uint stack_size = live_nodes() >> 4;
4267 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4268 nstack.push(_root);
4269
4270 while (nstack.size() > 0) {
4271 Node* n = nstack.pop();
4272 if (visited.member(n)) {
4273 continue;
4274 }
4275 visited.push(n);
4276
4277 // Walk over all input edges, checking for correspondence
4278 uint length = n->len();
4279 for (uint i = 0; i < length; i++) {
4280 Node* in = n->in(i);
4281 if (in != nullptr && !visited.member(in)) {
4282 nstack.push(in); // Put it on stack
4283 }
4284 if (in != nullptr && !in->is_top()) {
4285 // Count instances of `next`
4286 int cnt = 0;
4287 for (uint idx = 0; idx < in->_outcnt; idx++) {
4288 if (in->_out[idx] == n) {
4289 cnt++;
4290 }
4291 }
4292 assert(cnt > 0, "Failed to find Def-Use edge.");
4293 // Check for duplicate edges
4294 // walk the input array downcounting the input edges to n
4295 for (uint j = 0; j < length; j++) {
4296 if (n->in(j) == in) {
4297 cnt--;
4298 }
4299 }
4300 assert(cnt == 0, "Mismatched edge count.");
4301 } else if (in == nullptr) {
4302 assert(i == 0 || i >= n->req() ||
4303 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4304 (n->is_Unlock() && i == (n->req() - 1)) ||
4305 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4306 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4307 } else {
4308 assert(in->is_top(), "sanity");
4309 // Nothing to check.
4310 }
4311 }
4312 }
4313 }
4314
4315 //------------------------------verify_graph_edges---------------------------
4316 // Walk the Graph and verify that there is a one-to-one correspondence
4317 // between Use-Def edges and Def-Use edges in the graph.
4318 void Compile::verify_graph_edges(bool no_dead_code) {
4319 if (VerifyGraphEdges) {
4320 Unique_Node_List visited;
4321
4322 // Call graph walk to check edges
4323 verify_bidirectional_edges(visited);
4324 if (no_dead_code) {
4325 // Now make sure that no visited node is used by an unvisited node.
4326 bool dead_nodes = false;
4327 Unique_Node_List checked;
4328 while (visited.size() > 0) {
4329 Node* n = visited.pop();
4330 checked.push(n);
4331 for (uint i = 0; i < n->outcnt(); i++) {
4332 Node* use = n->raw_out(i);
4333 if (checked.member(use)) continue; // already checked
4334 if (visited.member(use)) continue; // already in the graph
4335 if (use->is_Con()) continue; // a dead ConNode is OK
4336 // At this point, we have found a dead node which is DU-reachable.
4337 if (!dead_nodes) {
4338 tty->print_cr("*** Dead nodes reachable via DU edges:");
4339 dead_nodes = true;
4340 }
4341 use->dump(2);
4342 tty->print_cr("---");
4343 checked.push(use); // No repeats; pretend it is now checked.
4344 }
4345 }
4346 assert(!dead_nodes, "using nodes must be reachable from root");
4347 }
4348 }
4349 }
4350 #endif
4351
4352 // The Compile object keeps track of failure reasons separately from the ciEnv.
4353 // This is required because there is not quite a 1-1 relation between the
4354 // ciEnv and its compilation task and the Compile object. Note that one
4355 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4356 // to backtrack and retry without subsuming loads. Other than this backtracking
4357 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4358 // by the logic in C2Compiler.
4359 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4360 if (log() != nullptr) {
4361 log()->elem("failure reason='%s' phase='compile'", reason);
4362 }
4363 if (_failure_reason.get() == nullptr) {
4364 // Record the first failure reason.
4365 _failure_reason.set(reason);
4366 if (CaptureBailoutInformation) {
4367 _first_failure_details = new CompilationFailureInfo(reason);
4368 }
4369 } else {
4370 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4371 }
4372
4373 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4374 C->print_method(PHASE_FAILURE, 1);
4375 }
4376 _root = nullptr; // flush the graph, too
4377 }
4378
4379 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4380 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4381 _compile(Compile::current()),
4382 _log(nullptr),
4383 _dolog(CITimeVerbose)
4384 {
4385 assert(_compile != nullptr, "sanity check");
4386 assert(id != PhaseTraceId::_t_none, "Don't use none");
4387 if (_dolog) {
4388 _log = _compile->log();
4389 }
4390 if (_log != nullptr) {
4391 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4392 _log->stamp();
4393 _log->end_head();
4394 }
4395
4396 // Inform memory statistic, if enabled
4397 if (CompilationMemoryStatistic::enabled()) {
4398 CompilationMemoryStatistic::on_phase_start((int)id, name);
4399 }
4400 }
4401
4402 Compile::TracePhase::TracePhase(PhaseTraceId id)
4403 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4404
4405 Compile::TracePhase::~TracePhase() {
4406
4407 // Inform memory statistic, if enabled
4408 if (CompilationMemoryStatistic::enabled()) {
4409 CompilationMemoryStatistic::on_phase_end();
4410 }
4411
4412 if (_compile->failing_internal()) {
4413 if (_log != nullptr) {
4414 _log->done("phase");
4415 }
4416 return; // timing code, not stressing bailouts.
4417 }
4418 #ifdef ASSERT
4419 if (PrintIdealNodeCount) {
4420 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4421 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4422 }
4423
4424 if (VerifyIdealNodeCount) {
4425 _compile->print_missing_nodes();
4426 }
4427 #endif
4428
4429 if (_log != nullptr) {
4430 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4431 }
4432 }
4433
4434 //----------------------------static_subtype_check-----------------------------
4435 // Shortcut important common cases when superklass is exact:
4436 // (0) superklass is java.lang.Object (can occur in reflective code)
4437 // (1) subklass is already limited to a subtype of superklass => always ok
4438 // (2) subklass does not overlap with superklass => always fail
4439 // (3) superklass has NO subtypes and we can check with a simple compare.
4440 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4441 if (skip) {
4442 return SSC_full_test; // Let caller generate the general case.
4443 }
4444
4445 if (subk->is_java_subtype_of(superk)) {
4446 return SSC_always_true; // (0) and (1) this test cannot fail
4447 }
4448
4449 if (!subk->maybe_java_subtype_of(superk)) {
4450 return SSC_always_false; // (2) true path dead; no dynamic test needed
4451 }
4452
4453 const Type* superelem = superk;
4454 if (superk->isa_aryklassptr()) {
4455 int ignored;
4456 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4457 }
4458
4459 if (superelem->isa_instklassptr()) {
4460 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4461 if (!ik->has_subklass()) {
4462 if (!ik->is_final()) {
4463 // Add a dependency if there is a chance of a later subclass.
4464 dependencies()->assert_leaf_type(ik);
4465 }
4466 if (!superk->maybe_java_subtype_of(subk)) {
4467 return SSC_always_false;
4468 }
4469 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4470 }
4471 } else {
4472 // A primitive array type has no subtypes.
4473 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4474 }
4475
4476 return SSC_full_test;
4477 }
4478
4479 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4480 #ifdef _LP64
4481 // The scaled index operand to AddP must be a clean 64-bit value.
4482 // Java allows a 32-bit int to be incremented to a negative
4483 // value, which appears in a 64-bit register as a large
4484 // positive number. Using that large positive number as an
4485 // operand in pointer arithmetic has bad consequences.
4486 // On the other hand, 32-bit overflow is rare, and the possibility
4487 // can often be excluded, if we annotate the ConvI2L node with
4488 // a type assertion that its value is known to be a small positive
4489 // number. (The prior range check has ensured this.)
4490 // This assertion is used by ConvI2LNode::Ideal.
4491 int index_max = max_jint - 1; // array size is max_jint, index is one less
4492 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4493 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4494 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4495 #endif
4496 return idx;
4497 }
4498
4499 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4500 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4501 if (ctrl != nullptr) {
4502 // Express control dependency by a CastII node with a narrow type.
4503 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4504 // node from floating above the range check during loop optimizations. Otherwise, the
4505 // ConvI2L node may be eliminated independently of the range check, causing the data path
4506 // to become TOP while the control path is still there (although it's unreachable).
4507 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4508 value = phase->transform(value);
4509 }
4510 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4511 return phase->transform(new ConvI2LNode(value, ltype));
4512 }
4513
4514 void Compile::dump_print_inlining() {
4515 inline_printer()->print_on(tty);
4516 }
4517
4518 void Compile::log_late_inline(CallGenerator* cg) {
4519 if (log() != nullptr) {
4520 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4521 cg->unique_id());
4522 JVMState* p = cg->call_node()->jvms();
4523 while (p != nullptr) {
4524 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4525 p = p->caller();
4526 }
4527 log()->tail("late_inline");
4528 }
4529 }
4530
4531 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4532 log_late_inline(cg);
4533 if (log() != nullptr) {
4534 log()->inline_fail(msg);
4535 }
4536 }
4537
4538 void Compile::log_inline_id(CallGenerator* cg) {
4539 if (log() != nullptr) {
4540 // The LogCompilation tool needs a unique way to identify late
4541 // inline call sites. This id must be unique for this call site in
4542 // this compilation. Try to have it unique across compilations as
4543 // well because it can be convenient when grepping through the log
4544 // file.
4545 // Distinguish OSR compilations from others in case CICountOSR is
4546 // on.
4547 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4548 cg->set_unique_id(id);
4549 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4550 }
4551 }
4552
4553 void Compile::log_inline_failure(const char* msg) {
4554 if (C->log() != nullptr) {
4555 C->log()->inline_fail(msg);
4556 }
4557 }
4558
4559
4560 // Dump inlining replay data to the stream.
4561 // Don't change thread state and acquire any locks.
4562 void Compile::dump_inline_data(outputStream* out) {
4563 InlineTree* inl_tree = ilt();
4564 if (inl_tree != nullptr) {
4565 out->print(" inline %d", inl_tree->count());
4566 inl_tree->dump_replay_data(out);
4567 }
4568 }
4569
4570 void Compile::dump_inline_data_reduced(outputStream* out) {
4571 assert(ReplayReduce, "");
4572
4573 InlineTree* inl_tree = ilt();
4574 if (inl_tree == nullptr) {
4575 return;
4576 }
4577 // Enable iterative replay file reduction
4578 // Output "compile" lines for depth 1 subtrees,
4579 // simulating that those trees were compiled
4580 // instead of inlined.
4581 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4582 InlineTree* sub = inl_tree->subtrees().at(i);
4583 if (sub->inline_level() != 1) {
4584 continue;
4585 }
4586
4587 ciMethod* method = sub->method();
4588 int entry_bci = -1;
4589 int comp_level = env()->task()->comp_level();
4590 out->print("compile ");
4591 method->dump_name_as_ascii(out);
4592 out->print(" %d %d", entry_bci, comp_level);
4593 out->print(" inline %d", sub->count());
4594 sub->dump_replay_data(out, -1);
4595 out->cr();
4596 }
4597 }
4598
4599 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4600 if (n1->Opcode() < n2->Opcode()) return -1;
4601 else if (n1->Opcode() > n2->Opcode()) return 1;
4602
4603 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4604 for (uint i = 1; i < n1->req(); i++) {
4605 if (n1->in(i) < n2->in(i)) return -1;
4606 else if (n1->in(i) > n2->in(i)) return 1;
4607 }
4608
4609 return 0;
4610 }
4611
4612 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4613 Node* n1 = *n1p;
4614 Node* n2 = *n2p;
4615
4616 return cmp_expensive_nodes(n1, n2);
4617 }
4618
4619 void Compile::sort_expensive_nodes() {
4620 if (!expensive_nodes_sorted()) {
4621 _expensive_nodes.sort(cmp_expensive_nodes);
4622 }
4623 }
4624
4625 bool Compile::expensive_nodes_sorted() const {
4626 for (int i = 1; i < _expensive_nodes.length(); i++) {
4627 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4628 return false;
4629 }
4630 }
4631 return true;
4632 }
4633
4634 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4635 if (_expensive_nodes.length() == 0) {
4636 return false;
4637 }
4638
4639 assert(OptimizeExpensiveOps, "optimization off?");
4640
4641 // Take this opportunity to remove dead nodes from the list
4642 int j = 0;
4643 for (int i = 0; i < _expensive_nodes.length(); i++) {
4644 Node* n = _expensive_nodes.at(i);
4645 if (!n->is_unreachable(igvn)) {
4646 assert(n->is_expensive(), "should be expensive");
4647 _expensive_nodes.at_put(j, n);
4648 j++;
4649 }
4650 }
4651 _expensive_nodes.trunc_to(j);
4652
4653 // Then sort the list so that similar nodes are next to each other
4654 // and check for at least two nodes of identical kind with same data
4655 // inputs.
4656 sort_expensive_nodes();
4657
4658 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4659 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4660 return true;
4661 }
4662 }
4663
4664 return false;
4665 }
4666
4667 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4668 if (_expensive_nodes.length() == 0) {
4669 return;
4670 }
4671
4672 assert(OptimizeExpensiveOps, "optimization off?");
4673
4674 // Sort to bring similar nodes next to each other and clear the
4675 // control input of nodes for which there's only a single copy.
4676 sort_expensive_nodes();
4677
4678 int j = 0;
4679 int identical = 0;
4680 int i = 0;
4681 bool modified = false;
4682 for (; i < _expensive_nodes.length()-1; i++) {
4683 assert(j <= i, "can't write beyond current index");
4684 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4685 identical++;
4686 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4687 continue;
4688 }
4689 if (identical > 0) {
4690 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4691 identical = 0;
4692 } else {
4693 Node* n = _expensive_nodes.at(i);
4694 igvn.replace_input_of(n, 0, nullptr);
4695 igvn.hash_insert(n);
4696 modified = true;
4697 }
4698 }
4699 if (identical > 0) {
4700 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4701 } else if (_expensive_nodes.length() >= 1) {
4702 Node* n = _expensive_nodes.at(i);
4703 igvn.replace_input_of(n, 0, nullptr);
4704 igvn.hash_insert(n);
4705 modified = true;
4706 }
4707 _expensive_nodes.trunc_to(j);
4708 if (modified) {
4709 igvn.optimize();
4710 }
4711 }
4712
4713 void Compile::add_expensive_node(Node * n) {
4714 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4715 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4716 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4717 if (OptimizeExpensiveOps) {
4718 _expensive_nodes.append(n);
4719 } else {
4720 // Clear control input and let IGVN optimize expensive nodes if
4721 // OptimizeExpensiveOps is off.
4722 n->set_req(0, nullptr);
4723 }
4724 }
4725
4726 /**
4727 * Track coarsened Lock and Unlock nodes.
4728 */
4729
4730 class Lock_List : public Node_List {
4731 uint _origin_cnt;
4732 public:
4733 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4734 uint origin_cnt() const { return _origin_cnt; }
4735 };
4736
4737 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4738 int length = locks.length();
4739 if (length > 0) {
4740 // Have to keep this list until locks elimination during Macro nodes elimination.
4741 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4742 AbstractLockNode* alock = locks.at(0);
4743 BoxLockNode* box = alock->box_node()->as_BoxLock();
4744 for (int i = 0; i < length; i++) {
4745 AbstractLockNode* lock = locks.at(i);
4746 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4747 locks_list->push(lock);
4748 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4749 if (this_box != box) {
4750 // Locking regions (BoxLock) could be Unbalanced here:
4751 // - its coarsened locks were eliminated in earlier
4752 // macro nodes elimination followed by loop unroll
4753 // - it is OSR locking region (no Lock node)
4754 // Preserve Unbalanced status in such cases.
4755 if (!this_box->is_unbalanced()) {
4756 this_box->set_coarsened();
4757 }
4758 if (!box->is_unbalanced()) {
4759 box->set_coarsened();
4760 }
4761 }
4762 }
4763 _coarsened_locks.append(locks_list);
4764 }
4765 }
4766
4767 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4768 int count = coarsened_count();
4769 for (int i = 0; i < count; i++) {
4770 Node_List* locks_list = _coarsened_locks.at(i);
4771 for (uint j = 0; j < locks_list->size(); j++) {
4772 Node* lock = locks_list->at(j);
4773 assert(lock->is_AbstractLock(), "sanity");
4774 if (!useful.member(lock)) {
4775 locks_list->yank(lock);
4776 }
4777 }
4778 }
4779 }
4780
4781 void Compile::remove_coarsened_lock(Node* n) {
4782 if (n->is_AbstractLock()) {
4783 int count = coarsened_count();
4784 for (int i = 0; i < count; i++) {
4785 Node_List* locks_list = _coarsened_locks.at(i);
4786 locks_list->yank(n);
4787 }
4788 }
4789 }
4790
4791 bool Compile::coarsened_locks_consistent() {
4792 int count = coarsened_count();
4793 for (int i = 0; i < count; i++) {
4794 bool unbalanced = false;
4795 bool modified = false; // track locks kind modifications
4796 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4797 uint size = locks_list->size();
4798 if (size == 0) {
4799 unbalanced = false; // All locks were eliminated - good
4800 } else if (size != locks_list->origin_cnt()) {
4801 unbalanced = true; // Some locks were removed from list
4802 } else {
4803 for (uint j = 0; j < size; j++) {
4804 Node* lock = locks_list->at(j);
4805 // All nodes in group should have the same state (modified or not)
4806 if (!lock->as_AbstractLock()->is_coarsened()) {
4807 if (j == 0) {
4808 // first on list was modified, the rest should be too for consistency
4809 modified = true;
4810 } else if (!modified) {
4811 // this lock was modified but previous locks on the list were not
4812 unbalanced = true;
4813 break;
4814 }
4815 } else if (modified) {
4816 // previous locks on list were modified but not this lock
4817 unbalanced = true;
4818 break;
4819 }
4820 }
4821 }
4822 if (unbalanced) {
4823 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4824 #ifdef ASSERT
4825 if (PrintEliminateLocks) {
4826 tty->print_cr("=== unbalanced coarsened locks ===");
4827 for (uint l = 0; l < size; l++) {
4828 locks_list->at(l)->dump();
4829 }
4830 }
4831 #endif
4832 record_failure(C2Compiler::retry_no_locks_coarsening());
4833 return false;
4834 }
4835 }
4836 return true;
4837 }
4838
4839 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4840 // locks coarsening optimization removed Lock/Unlock nodes from them.
4841 // Such regions become unbalanced because coarsening only removes part
4842 // of Lock/Unlock nodes in region. As result we can't execute other
4843 // locks elimination optimizations which assume all code paths have
4844 // corresponding pair of Lock/Unlock nodes - they are balanced.
4845 void Compile::mark_unbalanced_boxes() const {
4846 int count = coarsened_count();
4847 for (int i = 0; i < count; i++) {
4848 Node_List* locks_list = _coarsened_locks.at(i);
4849 uint size = locks_list->size();
4850 if (size > 0) {
4851 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4852 BoxLockNode* box = alock->box_node()->as_BoxLock();
4853 if (alock->is_coarsened()) {
4854 // coarsened_locks_consistent(), which is called before this method, verifies
4855 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4856 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4857 for (uint j = 1; j < size; j++) {
4858 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4859 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4860 if (box != this_box) {
4861 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4862 box->set_unbalanced();
4863 this_box->set_unbalanced();
4864 }
4865 }
4866 }
4867 }
4868 }
4869 }
4870
4871 /**
4872 * Remove the speculative part of types and clean up the graph
4873 */
4874 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4875 if (UseTypeSpeculation) {
4876 Unique_Node_List worklist;
4877 worklist.push(root());
4878 int modified = 0;
4879 // Go over all type nodes that carry a speculative type, drop the
4880 // speculative part of the type and enqueue the node for an igvn
4881 // which may optimize it out.
4882 for (uint next = 0; next < worklist.size(); ++next) {
4883 Node *n = worklist.at(next);
4884 if (n->is_Type()) {
4885 TypeNode* tn = n->as_Type();
4886 const Type* t = tn->type();
4887 const Type* t_no_spec = t->remove_speculative();
4888 if (t_no_spec != t) {
4889 bool in_hash = igvn.hash_delete(n);
4890 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4891 tn->set_type(t_no_spec);
4892 igvn.hash_insert(n);
4893 igvn._worklist.push(n); // give it a chance to go away
4894 modified++;
4895 }
4896 }
4897 // Iterate over outs - endless loops is unreachable from below
4898 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4899 Node *m = n->fast_out(i);
4900 if (not_a_node(m)) {
4901 continue;
4902 }
4903 worklist.push(m);
4904 }
4905 }
4906 // Drop the speculative part of all types in the igvn's type table
4907 igvn.remove_speculative_types();
4908 if (modified > 0) {
4909 igvn.optimize();
4910 if (failing()) return;
4911 }
4912 #ifdef ASSERT
4913 // Verify that after the IGVN is over no speculative type has resurfaced
4914 worklist.clear();
4915 worklist.push(root());
4916 for (uint next = 0; next < worklist.size(); ++next) {
4917 Node *n = worklist.at(next);
4918 const Type* t = igvn.type_or_null(n);
4919 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4920 if (n->is_Type()) {
4921 t = n->as_Type()->type();
4922 assert(t == t->remove_speculative(), "no more speculative types");
4923 }
4924 // Iterate over outs - endless loops is unreachable from below
4925 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4926 Node *m = n->fast_out(i);
4927 if (not_a_node(m)) {
4928 continue;
4929 }
4930 worklist.push(m);
4931 }
4932 }
4933 igvn.check_no_speculative_types();
4934 #endif
4935 }
4936 }
4937
4938 // Auxiliary methods to support randomized stressing/fuzzing.
4939
4940 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
4941 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
4942 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
4943 FLAG_SET_ERGO(StressSeed, _stress_seed);
4944 } else {
4945 _stress_seed = StressSeed;
4946 }
4947 if (_log != nullptr) {
4948 _log->elem("stress_test seed='%u'", _stress_seed);
4949 }
4950 }
4951
4952 int Compile::random() {
4953 _stress_seed = os::next_random(_stress_seed);
4954 return static_cast<int>(_stress_seed);
4955 }
4956
4957 // This method can be called the arbitrary number of times, with current count
4958 // as the argument. The logic allows selecting a single candidate from the
4959 // running list of candidates as follows:
4960 // int count = 0;
4961 // Cand* selected = null;
4962 // while(cand = cand->next()) {
4963 // if (randomized_select(++count)) {
4964 // selected = cand;
4965 // }
4966 // }
4967 //
4968 // Including count equalizes the chances any candidate is "selected".
4969 // This is useful when we don't have the complete list of candidates to choose
4970 // from uniformly. In this case, we need to adjust the randomicity of the
4971 // selection, or else we will end up biasing the selection towards the latter
4972 // candidates.
4973 //
4974 // Quick back-envelope calculation shows that for the list of n candidates
4975 // the equal probability for the candidate to persist as "best" can be
4976 // achieved by replacing it with "next" k-th candidate with the probability
4977 // of 1/k. It can be easily shown that by the end of the run, the
4978 // probability for any candidate is converged to 1/n, thus giving the
4979 // uniform distribution among all the candidates.
4980 //
4981 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4982 #define RANDOMIZED_DOMAIN_POW 29
4983 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4984 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4985 bool Compile::randomized_select(int count) {
4986 assert(count > 0, "only positive");
4987 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4988 }
4989
4990 #ifdef ASSERT
4991 // Failures are geometrically distributed with probability 1/StressBailoutMean.
4992 bool Compile::fail_randomly() {
4993 if ((random() % StressBailoutMean) != 0) {
4994 return false;
4995 }
4996 record_failure("StressBailout");
4997 return true;
4998 }
4999
5000 bool Compile::failure_is_artificial() {
5001 return C->failure_reason_is("StressBailout");
5002 }
5003 #endif
5004
5005 CloneMap& Compile::clone_map() { return _clone_map; }
5006 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5007
5008 void NodeCloneInfo::dump_on(outputStream* st) const {
5009 st->print(" {%d:%d} ", idx(), gen());
5010 }
5011
5012 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5013 uint64_t val = value(old->_idx);
5014 NodeCloneInfo cio(val);
5015 assert(val != 0, "old node should be in the map");
5016 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5017 insert(nnn->_idx, cin.get());
5018 #ifndef PRODUCT
5019 if (is_debug()) {
5020 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5021 }
5022 #endif
5023 }
5024
5025 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5026 NodeCloneInfo cio(value(old->_idx));
5027 if (cio.get() == 0) {
5028 cio.set(old->_idx, 0);
5029 insert(old->_idx, cio.get());
5030 #ifndef PRODUCT
5031 if (is_debug()) {
5032 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5033 }
5034 #endif
5035 }
5036 clone(old, nnn, gen);
5037 }
5038
5039 int CloneMap::max_gen() const {
5040 int g = 0;
5041 DictI di(_dict);
5042 for(; di.test(); ++di) {
5043 int t = gen(di._key);
5044 if (g < t) {
5045 g = t;
5046 #ifndef PRODUCT
5047 if (is_debug()) {
5048 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5049 }
5050 #endif
5051 }
5052 }
5053 return g;
5054 }
5055
5056 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5057 uint64_t val = value(key);
5058 if (val != 0) {
5059 NodeCloneInfo ni(val);
5060 ni.dump_on(st);
5061 }
5062 }
5063
5064 void Compile::shuffle_macro_nodes() {
5065 if (_macro_nodes.length() < 2) {
5066 return;
5067 }
5068 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5069 uint j = C->random() % (i + 1);
5070 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5071 }
5072 }
5073
5074 // Move Allocate nodes to the start of the list
5075 void Compile::sort_macro_nodes() {
5076 int count = macro_count();
5077 int allocates = 0;
5078 for (int i = 0; i < count; i++) {
5079 Node* n = macro_node(i);
5080 if (n->is_Allocate()) {
5081 if (i != allocates) {
5082 Node* tmp = macro_node(allocates);
5083 _macro_nodes.at_put(allocates, n);
5084 _macro_nodes.at_put(i, tmp);
5085 }
5086 allocates++;
5087 }
5088 }
5089 }
5090
5091 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5092 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5093 EventCompilerPhase event(UNTIMED);
5094 if (event.should_commit()) {
5095 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5096 }
5097 #ifndef PRODUCT
5098 ResourceMark rm;
5099 stringStream ss;
5100 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5101 int iter = ++_igv_phase_iter[cpt];
5102 if (iter > 1) {
5103 ss.print(" %d", iter);
5104 }
5105 if (n != nullptr) {
5106 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5107 if (n->is_Call()) {
5108 CallNode* call = n->as_Call();
5109 if (call->_name != nullptr) {
5110 // E.g. uncommon traps etc.
5111 ss.print(" - %s", call->_name);
5112 } else if (call->is_CallJava()) {
5113 CallJavaNode* call_java = call->as_CallJava();
5114 if (call_java->method() != nullptr) {
5115 ss.print(" -");
5116 call_java->method()->print_short_name(&ss);
5117 }
5118 }
5119 }
5120 }
5121
5122 const char* name = ss.as_string();
5123 if (should_print_igv(level)) {
5124 _igv_printer->print_graph(name);
5125 }
5126 if (should_print_phase(cpt)) {
5127 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5128 }
5129 #endif
5130 C->_latest_stage_start_counter.stamp();
5131 }
5132
5133 // Only used from CompileWrapper
5134 void Compile::begin_method() {
5135 #ifndef PRODUCT
5136 if (_method != nullptr && should_print_igv(1)) {
5137 _igv_printer->begin_method();
5138 }
5139 #endif
5140 C->_latest_stage_start_counter.stamp();
5141 }
5142
5143 // Only used from CompileWrapper
5144 void Compile::end_method() {
5145 EventCompilerPhase event(UNTIMED);
5146 if (event.should_commit()) {
5147 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5148 }
5149
5150 #ifndef PRODUCT
5151 if (_method != nullptr && should_print_igv(1)) {
5152 _igv_printer->end_method();
5153 }
5154 #endif
5155 }
5156
5157 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5158 #ifndef PRODUCT
5159 if (_directive->should_print_phase(cpt)) {
5160 return true;
5161 }
5162 #endif
5163 return false;
5164 }
5165
5166 #ifndef PRODUCT
5167 void Compile::init_igv() {
5168 if (_igv_printer == nullptr) {
5169 _igv_printer = IdealGraphPrinter::printer();
5170 _igv_printer->set_compile(this);
5171 }
5172 }
5173 #endif
5174
5175 bool Compile::should_print_igv(const int level) {
5176 #ifndef PRODUCT
5177 if (PrintIdealGraphLevel < 0) { // disabled by the user
5178 return false;
5179 }
5180
5181 bool need = directive()->IGVPrintLevelOption >= level;
5182 if (need) {
5183 Compile::init_igv();
5184 }
5185 return need;
5186 #else
5187 return false;
5188 #endif
5189 }
5190
5191 #ifndef PRODUCT
5192 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5193 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5194
5195 // Called from debugger. Prints method to the default file with the default phase name.
5196 // This works regardless of any Ideal Graph Visualizer flags set or not.
5197 void igv_print() {
5198 Compile::current()->igv_print_method_to_file();
5199 }
5200
5201 // Same as igv_print() above but with a specified phase name.
5202 void igv_print(const char* phase_name) {
5203 Compile::current()->igv_print_method_to_file(phase_name);
5204 }
5205
5206 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5207 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5208 // This works regardless of any Ideal Graph Visualizer flags set or not.
5209 void igv_print(bool network) {
5210 if (network) {
5211 Compile::current()->igv_print_method_to_network();
5212 } else {
5213 Compile::current()->igv_print_method_to_file();
5214 }
5215 }
5216
5217 // Same as igv_print(bool network) above but with a specified phase name.
5218 void igv_print(bool network, const char* phase_name) {
5219 if (network) {
5220 Compile::current()->igv_print_method_to_network(phase_name);
5221 } else {
5222 Compile::current()->igv_print_method_to_file(phase_name);
5223 }
5224 }
5225
5226 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5227 void igv_print_default() {
5228 Compile::current()->print_method(PHASE_DEBUG, 0);
5229 }
5230
5231 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5232 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5233 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5234 void igv_append() {
5235 Compile::current()->igv_print_method_to_file("Debug", true);
5236 }
5237
5238 // Same as igv_append() above but with a specified phase name.
5239 void igv_append(const char* phase_name) {
5240 Compile::current()->igv_print_method_to_file(phase_name, true);
5241 }
5242
5243 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5244 const char* file_name = "custom_debug.xml";
5245 if (_debug_file_printer == nullptr) {
5246 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5247 } else {
5248 _debug_file_printer->update_compiled_method(C->method());
5249 }
5250 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5251 _debug_file_printer->print_graph(phase_name);
5252 }
5253
5254 void Compile::igv_print_method_to_network(const char* phase_name) {
5255 ResourceMark rm;
5256 GrowableArray<const Node*> empty_list;
5257 igv_print_graph_to_network(phase_name, (Node*) C->root(), empty_list);
5258 }
5259
5260 void Compile::igv_print_graph_to_network(const char* name, Node* node, GrowableArray<const Node*>& visible_nodes) {
5261 if (_debug_network_printer == nullptr) {
5262 _debug_network_printer = new IdealGraphPrinter(C);
5263 } else {
5264 _debug_network_printer->update_compiled_method(C->method());
5265 }
5266 tty->print_cr("Method printed over network stream to IGV");
5267 _debug_network_printer->print(name, C->root(), visible_nodes);
5268 }
5269 #endif
5270
5271 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5272 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5273 return value;
5274 }
5275 Node* result = nullptr;
5276 if (bt == T_BYTE) {
5277 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5278 result = new RShiftINode(result, phase->intcon(24));
5279 } else if (bt == T_BOOLEAN) {
5280 result = new AndINode(value, phase->intcon(0xFF));
5281 } else if (bt == T_CHAR) {
5282 result = new AndINode(value,phase->intcon(0xFFFF));
5283 } else {
5284 assert(bt == T_SHORT, "unexpected narrow type");
5285 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5286 result = new RShiftINode(result, phase->intcon(16));
5287 }
5288 if (transform_res) {
5289 result = phase->transform(result);
5290 }
5291 return result;
5292 }
5293
5294 void Compile::record_method_not_compilable_oom() {
5295 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5296 }