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/aotCodeCache.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compilerDefinitions.hpp"
37 #include "compiler/compilerOracle.hpp"
38 #include "compiler/compiler_globals.hpp"
39 #include "compiler/disassembler.hpp"
40 #include "compiler/oopMap.hpp"
41 #include "gc/shared/barrierSet.hpp"
42 #include "gc/shared/c2/barrierSetC2.hpp"
43 #include "jfr/jfrEvents.hpp"
44 #include "jvm_io.h"
45 #include "memory/allocation.hpp"
46 #include "memory/arena.hpp"
47 #include "memory/resourceArea.hpp"
48 #include "opto/addnode.hpp"
49 #include "opto/block.hpp"
50 #include "opto/c2compiler.hpp"
51 #include "opto/callGenerator.hpp"
52 #include "opto/callnode.hpp"
53 #include "opto/castnode.hpp"
54 #include "opto/cfgnode.hpp"
55 #include "opto/chaitin.hpp"
56 #include "opto/compile.hpp"
57 #include "opto/connode.hpp"
58 #include "opto/convertnode.hpp"
59 #include "opto/divnode.hpp"
60 #include "opto/escape.hpp"
61 #include "opto/idealGraphPrinter.hpp"
62 #include "opto/locknode.hpp"
63 #include "opto/loopnode.hpp"
64 #include "opto/machnode.hpp"
65 #include "opto/macro.hpp"
66 #include "opto/matcher.hpp"
67 #include "opto/mathexactnode.hpp"
68 #include "opto/memnode.hpp"
69 #include "opto/mulnode.hpp"
70 #include "opto/narrowptrnode.hpp"
71 #include "opto/node.hpp"
72 #include "opto/opaquenode.hpp"
73 #include "opto/opcodes.hpp"
74 #include "opto/output.hpp"
75 #include "opto/parse.hpp"
76 #include "opto/phaseX.hpp"
77 #include "opto/rootnode.hpp"
78 #include "opto/runtime.hpp"
79 #include "opto/stringopts.hpp"
80 #include "opto/type.hpp"
81 #include "opto/vector.hpp"
82 #include "opto/vectornode.hpp"
83 #include "runtime/globals_extension.hpp"
84 #include "runtime/sharedRuntime.hpp"
85 #include "runtime/signature.hpp"
86 #include "runtime/stubRoutines.hpp"
87 #include "runtime/timer.hpp"
88 #include "utilities/align.hpp"
89 #include "utilities/copy.hpp"
90 #include "utilities/macros.hpp"
91 #include "utilities/resourceHash.hpp"
92
93 // -------------------- Compile::mach_constant_base_node -----------------------
94 // Constant table base node singleton.
95 MachConstantBaseNode* Compile::mach_constant_base_node() {
96 if (_mach_constant_base_node == nullptr) {
97 _mach_constant_base_node = new MachConstantBaseNode();
98 _mach_constant_base_node->add_req(C->root());
99 }
100 return _mach_constant_base_node;
101 }
102
103
104 /// Support for intrinsics.
105
106 // Return the index at which m must be inserted (or already exists).
107 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
108 class IntrinsicDescPair {
109 private:
110 ciMethod* _m;
111 bool _is_virtual;
112 public:
113 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
114 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
115 ciMethod* m= elt->method();
116 ciMethod* key_m = key->_m;
117 if (key_m < m) return -1;
118 else if (key_m > m) return 1;
119 else {
120 bool is_virtual = elt->is_virtual();
121 bool key_virtual = key->_is_virtual;
122 if (key_virtual < is_virtual) return -1;
123 else if (key_virtual > is_virtual) return 1;
124 else return 0;
125 }
126 }
127 };
128 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
129 #ifdef ASSERT
130 for (int i = 1; i < _intrinsics.length(); i++) {
131 CallGenerator* cg1 = _intrinsics.at(i-1);
132 CallGenerator* cg2 = _intrinsics.at(i);
133 assert(cg1->method() != cg2->method()
134 ? cg1->method() < cg2->method()
135 : cg1->is_virtual() < cg2->is_virtual(),
136 "compiler intrinsics list must stay sorted");
137 }
138 #endif
139 IntrinsicDescPair pair(m, is_virtual);
140 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
141 }
142
143 void Compile::register_intrinsic(CallGenerator* cg) {
144 bool found = false;
145 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
146 assert(!found, "registering twice");
147 _intrinsics.insert_before(index, cg);
148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
149 }
150
151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
152 assert(m->is_loaded(), "don't try this on unloaded methods");
153 if (_intrinsics.length() > 0) {
154 bool found = false;
155 int index = intrinsic_insertion_index(m, is_virtual, found);
156 if (found) {
157 return _intrinsics.at(index);
158 }
159 }
160 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
161 if (m->intrinsic_id() != vmIntrinsics::_none &&
162 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
163 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
164 if (cg != nullptr) {
165 // Save it for next time:
166 register_intrinsic(cg);
167 return cg;
168 } else {
169 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
170 }
171 }
172 return nullptr;
173 }
174
175 // Compile::make_vm_intrinsic is defined in library_call.cpp.
176
177 #ifndef PRODUCT
178 // statistics gathering...
179
180 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
182
183 inline int as_int(vmIntrinsics::ID id) {
184 return vmIntrinsics::as_int(id);
185 }
186
187 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
188 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
189 int oflags = _intrinsic_hist_flags[as_int(id)];
190 assert(flags != 0, "what happened?");
191 if (is_virtual) {
192 flags |= _intrinsic_virtual;
193 }
194 bool changed = (flags != oflags);
195 if ((flags & _intrinsic_worked) != 0) {
196 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
197 if (count == 1) {
198 changed = true; // first time
199 }
200 // increment the overall count also:
201 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
202 }
203 if (changed) {
204 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
205 // Something changed about the intrinsic's virtuality.
206 if ((flags & _intrinsic_virtual) != 0) {
207 // This is the first use of this intrinsic as a virtual call.
208 if (oflags != 0) {
209 // We already saw it as a non-virtual, so note both cases.
210 flags |= _intrinsic_both;
211 }
212 } else if ((oflags & _intrinsic_both) == 0) {
213 // This is the first use of this intrinsic as a non-virtual
214 flags |= _intrinsic_both;
215 }
216 }
217 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
218 }
219 // update the overall flags also:
220 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
221 return changed;
222 }
223
224 static char* format_flags(int flags, char* buf) {
225 buf[0] = 0;
226 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
227 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
228 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
229 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
230 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
231 if (buf[0] == 0) strcat(buf, ",");
232 assert(buf[0] == ',', "must be");
233 return &buf[1];
234 }
235
236 void Compile::print_intrinsic_statistics() {
237 char flagsbuf[100];
238 ttyLocker ttyl;
239 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
240 tty->print_cr("Compiler intrinsic usage:");
241 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
242 if (total == 0) total = 1; // avoid div0 in case of no successes
243 #define PRINT_STAT_LINE(name, c, f) \
244 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
245 for (auto id : EnumRange<vmIntrinsicID>{}) {
246 int flags = _intrinsic_hist_flags[as_int(id)];
247 juint count = _intrinsic_hist_count[as_int(id)];
248 if ((flags | count) != 0) {
249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
250 }
251 }
252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
253 if (xtty != nullptr) xtty->tail("statistics");
254 }
255
256 void Compile::print_statistics() {
257 { ttyLocker ttyl;
258 if (xtty != nullptr) xtty->head("statistics type='opto'");
259 Parse::print_statistics();
260 PhaseStringOpts::print_statistics();
261 PhaseCCP::print_statistics();
262 PhaseRegAlloc::print_statistics();
263 PhaseOutput::print_statistics();
264 PhasePeephole::print_statistics();
265 PhaseIdealLoop::print_statistics();
266 ConnectionGraph::print_statistics();
267 PhaseMacroExpand::print_statistics();
268 if (xtty != nullptr) xtty->tail("statistics");
269 }
270 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
271 // put this under its own <statistics> element.
272 print_intrinsic_statistics();
273 }
274 }
275 #endif //PRODUCT
276
277 void Compile::gvn_replace_by(Node* n, Node* nn) {
278 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
279 Node* use = n->last_out(i);
280 bool is_in_table = initial_gvn()->hash_delete(use);
281 uint uses_found = 0;
282 for (uint j = 0; j < use->len(); j++) {
283 if (use->in(j) == n) {
284 if (j < use->req())
285 use->set_req(j, nn);
286 else
287 use->set_prec(j, nn);
288 uses_found++;
289 }
290 }
291 if (is_in_table) {
292 // reinsert into table
293 initial_gvn()->hash_find_insert(use);
294 }
295 record_for_igvn(use);
296 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
297 i -= uses_found; // we deleted 1 or more copies of this edge
298 }
299 }
300
301
302 // Identify all nodes that are reachable from below, useful.
303 // Use breadth-first pass that records state in a Unique_Node_List,
304 // recursive traversal is slower.
305 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
306 int estimated_worklist_size = live_nodes();
307 useful.map( estimated_worklist_size, nullptr ); // preallocate space
308
309 // Initialize worklist
310 if (root() != nullptr) { useful.push(root()); }
311 // If 'top' is cached, declare it useful to preserve cached node
312 if (cached_top_node()) { useful.push(cached_top_node()); }
313
314 // Push all useful nodes onto the list, breadthfirst
315 for( uint next = 0; next < useful.size(); ++next ) {
316 assert( next < unique(), "Unique useful nodes < total nodes");
317 Node *n = useful.at(next);
318 uint max = n->len();
319 for( uint i = 0; i < max; ++i ) {
320 Node *m = n->in(i);
321 if (not_a_node(m)) continue;
322 useful.push(m);
323 }
324 }
325 }
326
327 // Update dead_node_list with any missing dead nodes using useful
328 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
329 void Compile::update_dead_node_list(Unique_Node_List &useful) {
330 uint max_idx = unique();
331 VectorSet& useful_node_set = useful.member_set();
332
333 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
334 // If node with index node_idx is not in useful set,
335 // mark it as dead in dead node list.
336 if (!useful_node_set.test(node_idx)) {
337 record_dead_node(node_idx);
338 }
339 }
340 }
341
342 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
343 int shift = 0;
344 for (int i = 0; i < inlines->length(); i++) {
345 CallGenerator* cg = inlines->at(i);
346 if (useful.member(cg->call_node())) {
347 if (shift > 0) {
348 inlines->at_put(i - shift, cg);
349 }
350 } else {
351 shift++; // skip over the dead element
352 }
353 }
354 if (shift > 0) {
355 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
356 }
357 }
358
359 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
360 assert(dead != nullptr && dead->is_Call(), "sanity");
361 int found = 0;
362 for (int i = 0; i < inlines->length(); i++) {
363 if (inlines->at(i)->call_node() == dead) {
364 inlines->remove_at(i);
365 found++;
366 NOT_DEBUG( break; ) // elements are unique, so exit early
367 }
368 }
369 assert(found <= 1, "not unique");
370 }
371
372 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
373 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
374 for (int i = node_list.length() - 1; i >= 0; i--) {
375 N* node = node_list.at(i);
376 if (!useful.member(node)) {
377 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
378 }
379 }
380 }
381
382 void Compile::remove_useless_node(Node* dead) {
383 remove_modified_node(dead);
384
385 // Constant node that has no out-edges and has only one in-edge from
386 // root is usually dead. However, sometimes reshaping walk makes
387 // it reachable by adding use edges. So, we will NOT count Con nodes
388 // as dead to be conservative about the dead node count at any
389 // given time.
390 if (!dead->is_Con()) {
391 record_dead_node(dead->_idx);
392 }
393 if (dead->is_macro()) {
394 remove_macro_node(dead);
395 }
396 if (dead->is_expensive()) {
397 remove_expensive_node(dead);
398 }
399 if (dead->is_OpaqueTemplateAssertionPredicate()) {
400 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
401 }
402 if (dead->is_ParsePredicate()) {
403 remove_parse_predicate(dead->as_ParsePredicate());
404 }
405 if (dead->for_post_loop_opts_igvn()) {
406 remove_from_post_loop_opts_igvn(dead);
407 }
408 if (dead->for_merge_stores_igvn()) {
409 remove_from_merge_stores_igvn(dead);
410 }
411 if (dead->is_Call()) {
412 remove_useless_late_inlines( &_late_inlines, dead);
413 remove_useless_late_inlines( &_string_late_inlines, dead);
414 remove_useless_late_inlines( &_boxing_late_inlines, dead);
415 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
416
417 if (dead->is_CallStaticJava()) {
418 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
419 }
420 }
421 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
422 bs->unregister_potential_barrier_node(dead);
423 }
424
425 // Disconnect all useless nodes by disconnecting those at the boundary.
426 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
427 uint next = 0;
428 while (next < useful.size()) {
429 Node *n = useful.at(next++);
430 if (n->is_SafePoint()) {
431 // We're done with a parsing phase. Replaced nodes are not valid
432 // beyond that point.
433 n->as_SafePoint()->delete_replaced_nodes();
434 }
435 // Use raw traversal of out edges since this code removes out edges
436 int max = n->outcnt();
437 for (int j = 0; j < max; ++j) {
438 Node* child = n->raw_out(j);
439 if (!useful.member(child)) {
440 assert(!child->is_top() || child != top(),
441 "If top is cached in Compile object it is in useful list");
442 // Only need to remove this out-edge to the useless node
443 n->raw_del_out(j);
444 --j;
445 --max;
446 if (child->is_data_proj_of_pure_function(n)) {
447 worklist.push(n);
448 }
449 }
450 }
451 if (n->outcnt() == 1 && n->has_special_unique_user()) {
452 assert(useful.member(n->unique_out()), "do not push a useless node");
453 worklist.push(n->unique_out());
454 }
455 }
456
457 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
458 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
459 // Remove useless Template Assertion Predicate opaque nodes
460 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
461 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
462 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
463 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
464 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
465 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
466 #ifdef ASSERT
467 if (_modified_nodes != nullptr) {
468 _modified_nodes->remove_useless_nodes(useful.member_set());
469 }
470 #endif
471
472 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
473 bs->eliminate_useless_gc_barriers(useful, this);
474 // clean up the late inline lists
475 remove_useless_late_inlines( &_late_inlines, useful);
476 remove_useless_late_inlines( &_string_late_inlines, useful);
477 remove_useless_late_inlines( &_boxing_late_inlines, useful);
478 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
479 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
480 }
481
482 // ============================================================================
483 //------------------------------CompileWrapper---------------------------------
484 class CompileWrapper : public StackObj {
485 Compile *const _compile;
486 public:
487 CompileWrapper(Compile* compile);
488
489 ~CompileWrapper();
490 };
491
492 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
493 // the Compile* pointer is stored in the current ciEnv:
494 ciEnv* env = compile->env();
495 assert(env == ciEnv::current(), "must already be a ciEnv active");
496 assert(env->compiler_data() == nullptr, "compile already active?");
497 env->set_compiler_data(compile);
498 assert(compile == Compile::current(), "sanity");
499
500 compile->set_type_dict(nullptr);
501 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
502 compile->clone_map().set_clone_idx(0);
503 compile->set_type_last_size(0);
504 compile->set_last_tf(nullptr, nullptr);
505 compile->set_indexSet_arena(nullptr);
506 compile->set_indexSet_free_block_list(nullptr);
507 compile->init_type_arena();
508 Type::Initialize(compile);
509 _compile->begin_method();
510 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
511 }
512 CompileWrapper::~CompileWrapper() {
513 // simulate crash during compilation
514 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
515
516 _compile->end_method();
517 _compile->env()->set_compiler_data(nullptr);
518 }
519
520
521 //----------------------------print_compile_messages---------------------------
522 void Compile::print_compile_messages() {
523 #ifndef PRODUCT
524 // Check if recompiling
525 if (!subsume_loads() && PrintOpto) {
526 // Recompiling without allowing machine instructions to subsume loads
527 tty->print_cr("*********************************************************");
528 tty->print_cr("** Bailout: Recompile without subsuming loads **");
529 tty->print_cr("*********************************************************");
530 }
531 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
532 // Recompiling without escape analysis
533 tty->print_cr("*********************************************************");
534 tty->print_cr("** Bailout: Recompile without escape analysis **");
535 tty->print_cr("*********************************************************");
536 }
537 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
538 // Recompiling without iterative escape analysis
539 tty->print_cr("*********************************************************");
540 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
541 tty->print_cr("*********************************************************");
542 }
543 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
544 // Recompiling without reducing allocation merges
545 tty->print_cr("*********************************************************");
546 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
547 tty->print_cr("*********************************************************");
548 }
549 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
550 // Recompiling without boxing elimination
551 tty->print_cr("*********************************************************");
552 tty->print_cr("** Bailout: Recompile without boxing elimination **");
553 tty->print_cr("*********************************************************");
554 }
555 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
556 // Recompiling without locks coarsening
557 tty->print_cr("*********************************************************");
558 tty->print_cr("** Bailout: Recompile without locks coarsening **");
559 tty->print_cr("*********************************************************");
560 }
561 if (env()->break_at_compile()) {
562 // Open the debugger when compiling this method.
563 tty->print("### Breaking when compiling: ");
564 method()->print_short_name();
565 tty->cr();
566 BREAKPOINT;
567 }
568
569 if( PrintOpto ) {
570 if (is_osr_compilation()) {
571 tty->print("[OSR]%3d", _compile_id);
572 } else {
573 tty->print("%3d", _compile_id);
574 }
575 }
576 #endif
577 }
578
579 #ifndef PRODUCT
580 void Compile::print_phase(const char* phase_name) {
581 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name);
582 }
583
584 void Compile::print_ideal_ir(const char* phase_name) {
585 // keep the following output all in one block
586 // This output goes directly to the tty, not the compiler log.
587 // To enable tools to match it up with the compilation activity,
588 // be sure to tag this tty output with the compile ID.
589
590 // Node dumping can cause a safepoint, which can break the tty lock.
591 // Buffer all node dumps, so that all safepoints happen before we lock.
592 ResourceMark rm;
593 stringStream ss;
594
595 if (_output == nullptr) {
596 ss.print_cr("AFTER: %s", phase_name);
597 // Print out all nodes in ascending order of index.
598 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
599 } else {
600 // Dump the node blockwise if we have a scheduling
601 _output->print_scheduling(&ss);
602 }
603
604 // Check that the lock is not broken by a safepoint.
605 NoSafepointVerifier nsv;
606 ttyLocker ttyl;
607 if (xtty != nullptr) {
608 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
609 compile_id(),
610 is_osr_compilation() ? " compile_kind='osr'" : "",
611 phase_name);
612 }
613
614 tty->print("%s", ss.as_string());
615
616 if (xtty != nullptr) {
617 xtty->tail("ideal");
618 }
619 }
620 #endif
621
622 // ============================================================================
623 //------------------------------Compile standard-------------------------------
624
625 // Compile a method. entry_bci is -1 for normal compilations and indicates
626 // the continuation bci for on stack replacement.
627
628
629 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
630 Options options, DirectiveSet* directive)
631 : Phase(Compiler),
632 _compile_id(ci_env->compile_id()),
633 _options(options),
634 _method(target),
635 _entry_bci(osr_bci),
636 _ilt(nullptr),
637 _stub_function(nullptr),
638 _stub_name(nullptr),
639 _stub_id(-1),
640 _stub_entry_point(nullptr),
641 _max_node_limit(MaxNodeLimit),
642 _post_loop_opts_phase(false),
643 _merge_stores_phase(false),
644 _allow_macro_nodes(true),
645 _inlining_progress(false),
646 _inlining_incrementally(false),
647 _do_cleanup(false),
648 _has_reserved_stack_access(target->has_reserved_stack_access()),
649 #ifndef PRODUCT
650 _igv_idx(0),
651 _trace_opto_output(directive->TraceOptoOutputOption),
652 #endif
653 _has_method_handle_invokes(false),
654 _clinit_barrier_on_entry(false),
655 _has_clinit_barriers(false),
656 _stress_seed(0),
657 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
658 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
659 _env(ci_env),
660 _directive(directive),
661 _log(ci_env->log()),
662 _first_failure_details(nullptr),
663 _intrinsics(comp_arena(), 0, 0, nullptr),
664 _macro_nodes(comp_arena(), 8, 0, nullptr),
665 _parse_predicates(comp_arena(), 8, 0, nullptr),
666 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
667 _expensive_nodes(comp_arena(), 8, 0, nullptr),
668 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
669 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
670 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
671 _coarsened_locks(comp_arena(), 8, 0, nullptr),
672 _congraph(nullptr),
673 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
674 _unique(0),
675 _dead_node_count(0),
676 _dead_node_list(comp_arena()),
677 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
678 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
679 _node_arena(&_node_arena_one),
680 _mach_constant_base_node(nullptr),
681 _Compile_types(mtCompiler, Arena::Tag::tag_type),
682 _initial_gvn(nullptr),
683 _igvn_worklist(nullptr),
684 _types(nullptr),
685 _node_hash(nullptr),
686 _late_inlines(comp_arena(), 2, 0, nullptr),
687 _string_late_inlines(comp_arena(), 2, 0, nullptr),
688 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
689 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
690 _late_inlines_pos(0),
691 _number_of_mh_late_inlines(0),
692 _oom(false),
693 _replay_inline_data(nullptr),
694 _inline_printer(this),
695 _java_calls(0),
696 _inner_loops(0),
697 _interpreter_frame_size(0),
698 _output(nullptr)
699 #ifndef PRODUCT
700 ,
701 _in_dump_cnt(0)
702 #endif
703 {
704 C = this;
705 CompileWrapper cw(this);
706
707 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
708 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
709
710 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
711 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
712 // We can always print a disassembly, either abstract (hex dump) or
713 // with the help of a suitable hsdis library. Thus, we should not
714 // couple print_assembly and print_opto_assembly controls.
715 // But: always print opto and regular assembly on compile command 'print'.
716 bool print_assembly = directive->PrintAssemblyOption;
717 set_print_assembly(print_opto_assembly || print_assembly);
718 #else
719 set_print_assembly(false); // must initialize.
720 #endif
721
722 #ifndef PRODUCT
723 set_parsed_irreducible_loop(false);
724 #endif
725
726 if (directive->ReplayInlineOption) {
727 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
728 }
729 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
730 set_print_intrinsics(directive->PrintIntrinsicsOption);
731 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
732
733 if (ProfileTraps) {
734 // Make sure the method being compiled gets its own MDO,
735 // so we can at least track the decompile_count().
736 method()->ensure_method_data();
737 }
738
739 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
740 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout ||
741 StressLoopPeeling) {
742 initialize_stress_seed(directive);
743 }
744
745 Init(/*do_aliasing=*/ true);
746
747 print_compile_messages();
748
749 _ilt = InlineTree::build_inline_tree_root();
750
751 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
752 assert(num_alias_types() >= AliasIdxRaw, "");
753
754 #define MINIMUM_NODE_HASH 1023
755
756 // GVN that will be run immediately on new nodes
757 uint estimated_size = method()->code_size()*4+64;
758 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
759 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
760 _types = new (comp_arena()) Type_Array(comp_arena());
761 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
762 PhaseGVN gvn;
763 set_initial_gvn(&gvn);
764
765 { // Scope for timing the parser
766 TracePhase tp(_t_parser);
767
768 // Put top into the hash table ASAP.
769 initial_gvn()->transform(top());
770
771 // Set up tf(), start(), and find a CallGenerator.
772 CallGenerator* cg = nullptr;
773 if (is_osr_compilation()) {
774 const TypeTuple *domain = StartOSRNode::osr_domain();
775 const TypeTuple *range = TypeTuple::make_range(method()->signature());
776 init_tf(TypeFunc::make(domain, range));
777 StartNode* s = new StartOSRNode(root(), domain);
778 initial_gvn()->set_type_bottom(s);
779 verify_start(s);
780 cg = CallGenerator::for_osr(method(), entry_bci());
781 } else {
782 // Normal case.
783 init_tf(TypeFunc::make(method()));
784 StartNode* s = new StartNode(root(), tf()->domain());
785 initial_gvn()->set_type_bottom(s);
786 verify_start(s);
787 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
788 // With java.lang.ref.reference.get() we must go through the
789 // intrinsic - even when get() is the root
790 // method of the compile - so that, if necessary, the value in
791 // the referent field of the reference object gets recorded by
792 // the pre-barrier code.
793 cg = find_intrinsic(method(), false);
794 }
795 if (cg == nullptr) {
796 float past_uses = method()->interpreter_invocation_count();
797 float expected_uses = past_uses;
798 cg = CallGenerator::for_inline(method(), expected_uses);
799 }
800 }
801 if (failing()) return;
802 if (cg == nullptr) {
803 const char* reason = InlineTree::check_can_parse(method());
804 assert(reason != nullptr, "expect reason for parse failure");
805 stringStream ss;
806 ss.print("cannot parse method: %s", reason);
807 record_method_not_compilable(ss.as_string());
808 return;
809 }
810
811 gvn.set_type(root(), root()->bottom_type());
812
813 JVMState* jvms = build_start_state(start(), tf());
814 if ((jvms = cg->generate(jvms)) == nullptr) {
815 assert(failure_reason() != nullptr, "expect reason for parse failure");
816 stringStream ss;
817 ss.print("method parse failed: %s", failure_reason());
818 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
819 return;
820 }
821 GraphKit kit(jvms);
822
823 if (!kit.stopped()) {
824 // Accept return values, and transfer control we know not where.
825 // This is done by a special, unique ReturnNode bound to root.
826 return_values(kit.jvms());
827 }
828
829 if (kit.has_exceptions()) {
830 // Any exceptions that escape from this call must be rethrown
831 // to whatever caller is dynamically above us on the stack.
832 // This is done by a special, unique RethrowNode bound to root.
833 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
834 }
835
836 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
837
838 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
839 inline_string_calls(true);
840 }
841
842 if (failing()) return;
843
844 // Remove clutter produced by parsing.
845 if (!failing()) {
846 ResourceMark rm;
847 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
848 }
849 }
850
851 // Note: Large methods are capped off in do_one_bytecode().
852 if (failing()) return;
853
854 // After parsing, node notes are no longer automagic.
855 // They must be propagated by register_new_node_with_optimizer(),
856 // clone(), or the like.
857 set_default_node_notes(nullptr);
858
859 #ifndef PRODUCT
860 if (should_print_igv(1)) {
861 _igv_printer->print_inlining();
862 }
863 #endif
864
865 if (failing()) return;
866 NOT_PRODUCT( verify_graph_edges(); )
867
868 // Now optimize
869 Optimize();
870 if (failing()) return;
871 NOT_PRODUCT( verify_graph_edges(); )
872
873 #ifndef PRODUCT
874 if (should_print_ideal()) {
875 print_ideal_ir("print_ideal");
876 }
877 #endif
878
879 #ifdef ASSERT
880 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
881 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
882 #endif
883
884 // Dump compilation data to replay it.
885 if (directive->DumpReplayOption) {
886 env()->dump_replay_data(_compile_id);
887 }
888 if (directive->DumpInlineOption && (ilt() != nullptr)) {
889 env()->dump_inline_data(_compile_id);
890 }
891
892 // Now that we know the size of all the monitors we can add a fixed slot
893 // for the original deopt pc.
894 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
895 set_fixed_slots(next_slot);
896
897 // Compute when to use implicit null checks. Used by matching trap based
898 // nodes and NullCheck optimization.
899 set_allowed_deopt_reasons();
900
901 // Now generate code
902 Code_Gen();
903 }
904
905 //------------------------------Compile----------------------------------------
906 // Compile a runtime stub
907 Compile::Compile(ciEnv* ci_env,
908 TypeFunc_generator generator,
909 address stub_function,
910 const char* stub_name,
911 int stub_id,
912 int is_fancy_jump,
913 bool pass_tls,
914 bool return_pc,
915 DirectiveSet* directive)
916 : Phase(Compiler),
917 _compile_id(0),
918 _options(Options::for_runtime_stub()),
919 _method(nullptr),
920 _entry_bci(InvocationEntryBci),
921 _stub_function(stub_function),
922 _stub_name(stub_name),
923 _stub_id(stub_id),
924 _stub_entry_point(nullptr),
925 _max_node_limit(MaxNodeLimit),
926 _post_loop_opts_phase(false),
927 _merge_stores_phase(false),
928 _allow_macro_nodes(true),
929 _inlining_progress(false),
930 _inlining_incrementally(false),
931 _has_reserved_stack_access(false),
932 #ifndef PRODUCT
933 _igv_idx(0),
934 _trace_opto_output(directive->TraceOptoOutputOption),
935 #endif
936 _has_method_handle_invokes(false),
937 _clinit_barrier_on_entry(false),
938 _has_clinit_barriers(false),
939 _stress_seed(0),
940 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
941 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
942 _env(ci_env),
943 _directive(directive),
944 _log(ci_env->log()),
945 _first_failure_details(nullptr),
946 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
947 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
948 _congraph(nullptr),
949 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
950 _unique(0),
951 _dead_node_count(0),
952 _dead_node_list(comp_arena()),
953 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
954 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
955 _node_arena(&_node_arena_one),
956 _mach_constant_base_node(nullptr),
957 _Compile_types(mtCompiler, Arena::Tag::tag_type),
958 _initial_gvn(nullptr),
959 _igvn_worklist(nullptr),
960 _types(nullptr),
961 _node_hash(nullptr),
962 _number_of_mh_late_inlines(0),
963 _oom(false),
964 _replay_inline_data(nullptr),
965 _inline_printer(this),
966 _java_calls(0),
967 _inner_loops(0),
968 _interpreter_frame_size(0),
969 _output(nullptr),
970 #ifndef PRODUCT
971 _in_dump_cnt(0),
972 #endif
973 _allowed_reasons(0) {
974 C = this;
975
976 // try to reuse an existing stub
977 {
978 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, _stub_id, stub_name);
979 if (blob != nullptr) {
980 RuntimeStub* rs = blob->as_runtime_stub();
981 _stub_entry_point = rs->entry_point();
982 return;
983 }
984 }
985
986 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
987 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
988
989 #ifndef PRODUCT
990 set_print_assembly(PrintFrameConverterAssembly);
991 set_parsed_irreducible_loop(false);
992 #else
993 set_print_assembly(false); // Must initialize.
994 #endif
995 set_has_irreducible_loop(false); // no loops
996
997 CompileWrapper cw(this);
998 Init(/*do_aliasing=*/ false);
999 init_tf((*generator)());
1000
1001 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1002 _types = new (comp_arena()) Type_Array(comp_arena());
1003 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1004
1005 if (StressLCM || StressGCM || StressBailout) {
1006 initialize_stress_seed(directive);
1007 }
1008
1009 {
1010 PhaseGVN gvn;
1011 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1012 gvn.transform(top());
1013
1014 GraphKit kit;
1015 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1016 }
1017
1018 NOT_PRODUCT( verify_graph_edges(); )
1019
1020 Code_Gen();
1021 }
1022
1023 Compile::~Compile() {
1024 delete _first_failure_details;
1025 };
1026
1027 //------------------------------Init-------------------------------------------
1028 // Prepare for a single compilation
1029 void Compile::Init(bool aliasing) {
1030 _do_aliasing = aliasing;
1031 _unique = 0;
1032 _regalloc = nullptr;
1033
1034 _tf = nullptr; // filled in later
1035 _top = nullptr; // cached later
1036 _matcher = nullptr; // filled in later
1037 _cfg = nullptr; // filled in later
1038
1039 _node_note_array = nullptr;
1040 _default_node_notes = nullptr;
1041 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1042
1043 _immutable_memory = nullptr; // filled in at first inquiry
1044
1045 #ifdef ASSERT
1046 _phase_optimize_finished = false;
1047 _phase_verify_ideal_loop = false;
1048 _exception_backedge = false;
1049 _type_verify = nullptr;
1050 #endif
1051
1052 // Globally visible Nodes
1053 // First set TOP to null to give safe behavior during creation of RootNode
1054 set_cached_top_node(nullptr);
1055 set_root(new RootNode());
1056 // Now that you have a Root to point to, create the real TOP
1057 set_cached_top_node( new ConNode(Type::TOP) );
1058 set_recent_alloc(nullptr, nullptr);
1059
1060 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1061 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1062 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1063 env()->set_dependencies(new Dependencies(env()));
1064
1065 _fixed_slots = 0;
1066 set_has_split_ifs(false);
1067 set_has_loops(false); // first approximation
1068 set_has_stringbuilder(false);
1069 set_has_boxed_value(false);
1070 _trap_can_recompile = false; // no traps emitted yet
1071 _major_progress = true; // start out assuming good things will happen
1072 set_has_unsafe_access(false);
1073 set_max_vector_size(0);
1074 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1075 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1076 set_decompile_count(0);
1077
1078 #ifndef PRODUCT
1079 _phase_counter = 0;
1080 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1081 #endif
1082
1083 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1084 _loop_opts_cnt = LoopOptsCount;
1085 set_do_inlining(Inline);
1086 set_max_inline_size(MaxInlineSize);
1087 set_freq_inline_size(FreqInlineSize);
1088 set_do_scheduling(OptoScheduling);
1089
1090 set_do_vector_loop(false);
1091 set_has_monitors(false);
1092 set_has_scoped_access(false);
1093
1094 if (AllowVectorizeOnDemand) {
1095 if (has_method() && _directive->VectorizeOption) {
1096 set_do_vector_loop(true);
1097 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());})
1098 } else if (has_method() && method()->name() != nullptr &&
1099 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1100 set_do_vector_loop(true);
1101 }
1102 }
1103 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1104 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());})
1105
1106 _max_node_limit = _directive->MaxNodeLimitOption;
1107
1108 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() &&
1109 (method()->needs_clinit_barrier() || (do_clinit_barriers() && method()->is_static()))) {
1110 set_clinit_barrier_on_entry(true);
1111 if (do_clinit_barriers()) {
1112 set_has_clinit_barriers(true); // Entry clinit barrier is in prolog code.
1113 }
1114 }
1115 if (debug_info()->recording_non_safepoints()) {
1116 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1117 (comp_arena(), 8, 0, nullptr));
1118 set_default_node_notes(Node_Notes::make(this));
1119 }
1120
1121 const int grow_ats = 16;
1122 _max_alias_types = grow_ats;
1123 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1124 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1125 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1126 {
1127 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1128 }
1129 // Initialize the first few types.
1130 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1131 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1132 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1133 _num_alias_types = AliasIdxRaw+1;
1134 // Zero out the alias type cache.
1135 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1136 // A null adr_type hits in the cache right away. Preload the right answer.
1137 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1138 }
1139
1140 #ifdef ASSERT
1141 // Verify that the current StartNode is valid.
1142 void Compile::verify_start(StartNode* s) const {
1143 assert(failing_internal() || s == start(), "should be StartNode");
1144 }
1145 #endif
1146
1147 /**
1148 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1149 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1150 * the ideal graph.
1151 */
1152 StartNode* Compile::start() const {
1153 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1154 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1155 Node* start = root()->fast_out(i);
1156 if (start->is_Start()) {
1157 return start->as_Start();
1158 }
1159 }
1160 fatal("Did not find Start node!");
1161 return nullptr;
1162 }
1163
1164 //-------------------------------immutable_memory-------------------------------------
1165 // Access immutable memory
1166 Node* Compile::immutable_memory() {
1167 if (_immutable_memory != nullptr) {
1168 return _immutable_memory;
1169 }
1170 StartNode* s = start();
1171 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1172 Node *p = s->fast_out(i);
1173 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1174 _immutable_memory = p;
1175 return _immutable_memory;
1176 }
1177 }
1178 ShouldNotReachHere();
1179 return nullptr;
1180 }
1181
1182 //----------------------set_cached_top_node------------------------------------
1183 // Install the cached top node, and make sure Node::is_top works correctly.
1184 void Compile::set_cached_top_node(Node* tn) {
1185 if (tn != nullptr) verify_top(tn);
1186 Node* old_top = _top;
1187 _top = tn;
1188 // Calling Node::setup_is_top allows the nodes the chance to adjust
1189 // their _out arrays.
1190 if (_top != nullptr) _top->setup_is_top();
1191 if (old_top != nullptr) old_top->setup_is_top();
1192 assert(_top == nullptr || top()->is_top(), "");
1193 }
1194
1195 #ifdef ASSERT
1196 uint Compile::count_live_nodes_by_graph_walk() {
1197 Unique_Node_List useful(comp_arena());
1198 // Get useful node list by walking the graph.
1199 identify_useful_nodes(useful);
1200 return useful.size();
1201 }
1202
1203 void Compile::print_missing_nodes() {
1204
1205 // Return if CompileLog is null and PrintIdealNodeCount is false.
1206 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1207 return;
1208 }
1209
1210 // This is an expensive function. It is executed only when the user
1211 // specifies VerifyIdealNodeCount option or otherwise knows the
1212 // additional work that needs to be done to identify reachable nodes
1213 // by walking the flow graph and find the missing ones using
1214 // _dead_node_list.
1215
1216 Unique_Node_List useful(comp_arena());
1217 // Get useful node list by walking the graph.
1218 identify_useful_nodes(useful);
1219
1220 uint l_nodes = C->live_nodes();
1221 uint l_nodes_by_walk = useful.size();
1222
1223 if (l_nodes != l_nodes_by_walk) {
1224 if (_log != nullptr) {
1225 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1226 _log->stamp();
1227 _log->end_head();
1228 }
1229 VectorSet& useful_member_set = useful.member_set();
1230 int last_idx = l_nodes_by_walk;
1231 for (int i = 0; i < last_idx; i++) {
1232 if (useful_member_set.test(i)) {
1233 if (_dead_node_list.test(i)) {
1234 if (_log != nullptr) {
1235 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1236 }
1237 if (PrintIdealNodeCount) {
1238 // Print the log message to tty
1239 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1240 useful.at(i)->dump();
1241 }
1242 }
1243 }
1244 else if (! _dead_node_list.test(i)) {
1245 if (_log != nullptr) {
1246 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1247 }
1248 if (PrintIdealNodeCount) {
1249 // Print the log message to tty
1250 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1251 }
1252 }
1253 }
1254 if (_log != nullptr) {
1255 _log->tail("mismatched_nodes");
1256 }
1257 }
1258 }
1259 void Compile::record_modified_node(Node* n) {
1260 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1261 _modified_nodes->push(n);
1262 }
1263 }
1264
1265 void Compile::remove_modified_node(Node* n) {
1266 if (_modified_nodes != nullptr) {
1267 _modified_nodes->remove(n);
1268 }
1269 }
1270 #endif
1271
1272 #ifndef PRODUCT
1273 void Compile::verify_top(Node* tn) const {
1274 if (tn != nullptr) {
1275 assert(tn->is_Con(), "top node must be a constant");
1276 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1277 assert(tn->in(0) != nullptr, "must have live top node");
1278 }
1279 }
1280 #endif
1281
1282
1283 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1284
1285 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1286 guarantee(arr != nullptr, "");
1287 int num_blocks = arr->length();
1288 if (grow_by < num_blocks) grow_by = num_blocks;
1289 int num_notes = grow_by * _node_notes_block_size;
1290 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1291 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1292 while (num_notes > 0) {
1293 arr->append(notes);
1294 notes += _node_notes_block_size;
1295 num_notes -= _node_notes_block_size;
1296 }
1297 assert(num_notes == 0, "exact multiple, please");
1298 }
1299
1300 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1301 if (source == nullptr || dest == nullptr) return false;
1302
1303 if (dest->is_Con())
1304 return false; // Do not push debug info onto constants.
1305
1306 #ifdef ASSERT
1307 // Leave a bread crumb trail pointing to the original node:
1308 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1309 dest->set_debug_orig(source);
1310 }
1311 #endif
1312
1313 if (node_note_array() == nullptr)
1314 return false; // Not collecting any notes now.
1315
1316 // This is a copy onto a pre-existing node, which may already have notes.
1317 // If both nodes have notes, do not overwrite any pre-existing notes.
1318 Node_Notes* source_notes = node_notes_at(source->_idx);
1319 if (source_notes == nullptr || source_notes->is_clear()) return false;
1320 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1321 if (dest_notes == nullptr || dest_notes->is_clear()) {
1322 return set_node_notes_at(dest->_idx, source_notes);
1323 }
1324
1325 Node_Notes merged_notes = (*source_notes);
1326 // The order of operations here ensures that dest notes will win...
1327 merged_notes.update_from(dest_notes);
1328 return set_node_notes_at(dest->_idx, &merged_notes);
1329 }
1330
1331
1332 //--------------------------allow_range_check_smearing-------------------------
1333 // Gating condition for coalescing similar range checks.
1334 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1335 // single covering check that is at least as strong as any of them.
1336 // If the optimization succeeds, the simplified (strengthened) range check
1337 // will always succeed. If it fails, we will deopt, and then give up
1338 // on the optimization.
1339 bool Compile::allow_range_check_smearing() const {
1340 // If this method has already thrown a range-check,
1341 // assume it was because we already tried range smearing
1342 // and it failed.
1343 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1344 return !already_trapped;
1345 }
1346
1347
1348 //------------------------------flatten_alias_type-----------------------------
1349 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1350 assert(do_aliasing(), "Aliasing should be enabled");
1351 int offset = tj->offset();
1352 TypePtr::PTR ptr = tj->ptr();
1353
1354 // Known instance (scalarizable allocation) alias only with itself.
1355 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1356 tj->is_oopptr()->is_known_instance();
1357
1358 // Process weird unsafe references.
1359 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1360 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1361 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1362 tj = TypeOopPtr::BOTTOM;
1363 ptr = tj->ptr();
1364 offset = tj->offset();
1365 }
1366
1367 // Array pointers need some flattening
1368 const TypeAryPtr* ta = tj->isa_aryptr();
1369 if (ta && ta->is_stable()) {
1370 // Erase stability property for alias analysis.
1371 tj = ta = ta->cast_to_stable(false);
1372 }
1373 if( ta && is_known_inst ) {
1374 if ( offset != Type::OffsetBot &&
1375 offset > arrayOopDesc::length_offset_in_bytes() ) {
1376 offset = Type::OffsetBot; // Flatten constant access into array body only
1377 tj = ta = ta->
1378 remove_speculative()->
1379 cast_to_ptr_type(ptr)->
1380 with_offset(offset);
1381 }
1382 } else if (ta) {
1383 // For arrays indexed by constant indices, we flatten the alias
1384 // space to include all of the array body. Only the header, klass
1385 // and array length can be accessed un-aliased.
1386 if( offset != Type::OffsetBot ) {
1387 if( ta->const_oop() ) { // MethodData* or Method*
1388 offset = Type::OffsetBot; // Flatten constant access into array body
1389 tj = ta = ta->
1390 remove_speculative()->
1391 cast_to_ptr_type(ptr)->
1392 cast_to_exactness(false)->
1393 with_offset(offset);
1394 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1395 // range is OK as-is.
1396 tj = ta = TypeAryPtr::RANGE;
1397 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1398 tj = TypeInstPtr::KLASS; // all klass loads look alike
1399 ta = TypeAryPtr::RANGE; // generic ignored junk
1400 ptr = TypePtr::BotPTR;
1401 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1402 tj = TypeInstPtr::MARK;
1403 ta = TypeAryPtr::RANGE; // generic ignored junk
1404 ptr = TypePtr::BotPTR;
1405 } else { // Random constant offset into array body
1406 offset = Type::OffsetBot; // Flatten constant access into array body
1407 tj = ta = ta->
1408 remove_speculative()->
1409 cast_to_ptr_type(ptr)->
1410 cast_to_exactness(false)->
1411 with_offset(offset);
1412 }
1413 }
1414 // Arrays of fixed size alias with arrays of unknown size.
1415 if (ta->size() != TypeInt::POS) {
1416 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1417 tj = ta = ta->
1418 remove_speculative()->
1419 cast_to_ptr_type(ptr)->
1420 with_ary(tary)->
1421 cast_to_exactness(false);
1422 }
1423 // Arrays of known objects become arrays of unknown objects.
1424 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1425 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1426 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1427 }
1428 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1429 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1430 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1431 }
1432 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1433 // cannot be distinguished by bytecode alone.
1434 if (ta->elem() == TypeInt::BOOL) {
1435 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1436 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1437 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1438 }
1439 // During the 2nd round of IterGVN, NotNull castings are removed.
1440 // Make sure the Bottom and NotNull variants alias the same.
1441 // Also, make sure exact and non-exact variants alias the same.
1442 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1443 tj = ta = ta->
1444 remove_speculative()->
1445 cast_to_ptr_type(TypePtr::BotPTR)->
1446 cast_to_exactness(false)->
1447 with_offset(offset);
1448 }
1449 }
1450
1451 // Oop pointers need some flattening
1452 const TypeInstPtr *to = tj->isa_instptr();
1453 if (to && to != TypeOopPtr::BOTTOM) {
1454 ciInstanceKlass* ik = to->instance_klass();
1455 if( ptr == TypePtr::Constant ) {
1456 if (ik != ciEnv::current()->Class_klass() ||
1457 offset < ik->layout_helper_size_in_bytes()) {
1458 // No constant oop pointers (such as Strings); they alias with
1459 // unknown strings.
1460 assert(!is_known_inst, "not scalarizable allocation");
1461 tj = to = to->
1462 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1463 remove_speculative()->
1464 cast_to_ptr_type(TypePtr::BotPTR)->
1465 cast_to_exactness(false);
1466 }
1467 } else if( is_known_inst ) {
1468 tj = to; // Keep NotNull and klass_is_exact for instance type
1469 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1470 // During the 2nd round of IterGVN, NotNull castings are removed.
1471 // Make sure the Bottom and NotNull variants alias the same.
1472 // Also, make sure exact and non-exact variants alias the same.
1473 tj = to = to->
1474 remove_speculative()->
1475 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1476 cast_to_ptr_type(TypePtr::BotPTR)->
1477 cast_to_exactness(false);
1478 }
1479 if (to->speculative() != nullptr) {
1480 tj = to = to->remove_speculative();
1481 }
1482 // Canonicalize the holder of this field
1483 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1484 // First handle header references such as a LoadKlassNode, even if the
1485 // object's klass is unloaded at compile time (4965979).
1486 if (!is_known_inst) { // Do it only for non-instance types
1487 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1488 }
1489 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1490 // Static fields are in the space above the normal instance
1491 // fields in the java.lang.Class instance.
1492 if (ik != ciEnv::current()->Class_klass()) {
1493 to = nullptr;
1494 tj = TypeOopPtr::BOTTOM;
1495 offset = tj->offset();
1496 }
1497 } else {
1498 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1499 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1500 assert(tj->offset() == offset, "no change to offset expected");
1501 bool xk = to->klass_is_exact();
1502 int instance_id = to->instance_id();
1503
1504 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1505 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1506 // its interfaces are included.
1507 if (xk && ik->equals(canonical_holder)) {
1508 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1509 } else {
1510 assert(xk || !is_known_inst, "Known instance should be exact type");
1511 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1512 }
1513 }
1514 }
1515
1516 // Klass pointers to object array klasses need some flattening
1517 const TypeKlassPtr *tk = tj->isa_klassptr();
1518 if( tk ) {
1519 // If we are referencing a field within a Klass, we need
1520 // to assume the worst case of an Object. Both exact and
1521 // inexact types must flatten to the same alias class so
1522 // use NotNull as the PTR.
1523 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1524 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1525 env()->Object_klass(),
1526 offset);
1527 }
1528
1529 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1530 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1531 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1532 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1533 } else {
1534 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1535 }
1536 }
1537
1538 // Check for precise loads from the primary supertype array and force them
1539 // to the supertype cache alias index. Check for generic array loads from
1540 // the primary supertype array and also force them to the supertype cache
1541 // alias index. Since the same load can reach both, we need to merge
1542 // these 2 disparate memories into the same alias class. Since the
1543 // primary supertype array is read-only, there's no chance of confusion
1544 // where we bypass an array load and an array store.
1545 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1546 if (offset == Type::OffsetBot ||
1547 (offset >= primary_supers_offset &&
1548 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1549 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1550 offset = in_bytes(Klass::secondary_super_cache_offset());
1551 tj = tk = tk->with_offset(offset);
1552 }
1553 }
1554
1555 // Flatten all Raw pointers together.
1556 if (tj->base() == Type::RawPtr)
1557 tj = TypeRawPtr::BOTTOM;
1558
1559 if (tj->base() == Type::AnyPtr)
1560 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1561
1562 offset = tj->offset();
1563 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1564
1565 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1566 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1567 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1568 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1569 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1570 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1571 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1572 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1573 assert( tj->ptr() != TypePtr::TopPTR &&
1574 tj->ptr() != TypePtr::AnyNull &&
1575 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1576 // assert( tj->ptr() != TypePtr::Constant ||
1577 // tj->base() == Type::RawPtr ||
1578 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1579
1580 return tj;
1581 }
1582
1583 void Compile::AliasType::Init(int i, const TypePtr* at) {
1584 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1585 _index = i;
1586 _adr_type = at;
1587 _field = nullptr;
1588 _element = nullptr;
1589 _is_rewritable = true; // default
1590 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1591 if (atoop != nullptr && atoop->is_known_instance()) {
1592 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1593 _general_index = Compile::current()->get_alias_index(gt);
1594 } else {
1595 _general_index = 0;
1596 }
1597 }
1598
1599 BasicType Compile::AliasType::basic_type() const {
1600 if (element() != nullptr) {
1601 const Type* element = adr_type()->is_aryptr()->elem();
1602 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1603 } if (field() != nullptr) {
1604 return field()->layout_type();
1605 } else {
1606 return T_ILLEGAL; // unknown
1607 }
1608 }
1609
1610 //---------------------------------print_on------------------------------------
1611 #ifndef PRODUCT
1612 void Compile::AliasType::print_on(outputStream* st) {
1613 if (index() < 10)
1614 st->print("@ <%d> ", index());
1615 else st->print("@ <%d>", index());
1616 st->print(is_rewritable() ? " " : " RO");
1617 int offset = adr_type()->offset();
1618 if (offset == Type::OffsetBot)
1619 st->print(" +any");
1620 else st->print(" +%-3d", offset);
1621 st->print(" in ");
1622 adr_type()->dump_on(st);
1623 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1624 if (field() != nullptr && tjp) {
1625 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1626 tjp->offset() != field()->offset_in_bytes()) {
1627 st->print(" != ");
1628 field()->print();
1629 st->print(" ***");
1630 }
1631 }
1632 }
1633
1634 void print_alias_types() {
1635 Compile* C = Compile::current();
1636 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1637 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1638 C->alias_type(idx)->print_on(tty);
1639 tty->cr();
1640 }
1641 }
1642 #endif
1643
1644
1645 //----------------------------probe_alias_cache--------------------------------
1646 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1647 intptr_t key = (intptr_t) adr_type;
1648 key ^= key >> logAliasCacheSize;
1649 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1650 }
1651
1652
1653 //-----------------------------grow_alias_types--------------------------------
1654 void Compile::grow_alias_types() {
1655 const int old_ats = _max_alias_types; // how many before?
1656 const int new_ats = old_ats; // how many more?
1657 const int grow_ats = old_ats+new_ats; // how many now?
1658 _max_alias_types = grow_ats;
1659 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1660 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1661 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1662 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1663 }
1664
1665
1666 //--------------------------------find_alias_type------------------------------
1667 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1668 if (!do_aliasing()) {
1669 return alias_type(AliasIdxBot);
1670 }
1671
1672 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1673 if (ace->_adr_type == adr_type) {
1674 return alias_type(ace->_index);
1675 }
1676
1677 // Handle special cases.
1678 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1679 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1680
1681 // Do it the slow way.
1682 const TypePtr* flat = flatten_alias_type(adr_type);
1683
1684 #ifdef ASSERT
1685 {
1686 ResourceMark rm;
1687 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1688 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1689 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1690 Type::str(adr_type));
1691 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1692 const TypeOopPtr* foop = flat->is_oopptr();
1693 // Scalarizable allocations have exact klass always.
1694 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1695 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1696 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1697 Type::str(foop), Type::str(xoop));
1698 }
1699 }
1700 #endif
1701
1702 int idx = AliasIdxTop;
1703 for (int i = 0; i < num_alias_types(); i++) {
1704 if (alias_type(i)->adr_type() == flat) {
1705 idx = i;
1706 break;
1707 }
1708 }
1709
1710 if (idx == AliasIdxTop) {
1711 if (no_create) return nullptr;
1712 // Grow the array if necessary.
1713 if (_num_alias_types == _max_alias_types) grow_alias_types();
1714 // Add a new alias type.
1715 idx = _num_alias_types++;
1716 _alias_types[idx]->Init(idx, flat);
1717 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1718 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1719 if (flat->isa_instptr()) {
1720 if (flat->offset() == java_lang_Class::klass_offset()
1721 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1722 alias_type(idx)->set_rewritable(false);
1723 }
1724 if (flat->isa_aryptr()) {
1725 #ifdef ASSERT
1726 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1727 // (T_BYTE has the weakest alignment and size restrictions...)
1728 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1729 #endif
1730 if (flat->offset() == TypePtr::OffsetBot) {
1731 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1732 }
1733 }
1734 if (flat->isa_klassptr()) {
1735 if (UseCompactObjectHeaders) {
1736 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1737 alias_type(idx)->set_rewritable(false);
1738 }
1739 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1740 alias_type(idx)->set_rewritable(false);
1741 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1742 alias_type(idx)->set_rewritable(false);
1743 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1744 alias_type(idx)->set_rewritable(false);
1745 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1746 alias_type(idx)->set_rewritable(false);
1747 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1748 alias_type(idx)->set_rewritable(false);
1749 }
1750 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1751 // but the base pointer type is not distinctive enough to identify
1752 // references into JavaThread.)
1753
1754 // Check for final fields.
1755 const TypeInstPtr* tinst = flat->isa_instptr();
1756 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1757 ciField* field;
1758 if (tinst->const_oop() != nullptr &&
1759 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1760 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1761 // static field
1762 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1763 field = k->get_field_by_offset(tinst->offset(), true);
1764 } else {
1765 ciInstanceKlass *k = tinst->instance_klass();
1766 field = k->get_field_by_offset(tinst->offset(), false);
1767 }
1768 assert(field == nullptr ||
1769 original_field == nullptr ||
1770 (field->holder() == original_field->holder() &&
1771 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1772 field->is_static() == original_field->is_static()), "wrong field?");
1773 // Set field() and is_rewritable() attributes.
1774 if (field != nullptr) alias_type(idx)->set_field(field);
1775 }
1776 }
1777
1778 // Fill the cache for next time.
1779 ace->_adr_type = adr_type;
1780 ace->_index = idx;
1781 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1782
1783 // Might as well try to fill the cache for the flattened version, too.
1784 AliasCacheEntry* face = probe_alias_cache(flat);
1785 if (face->_adr_type == nullptr) {
1786 face->_adr_type = flat;
1787 face->_index = idx;
1788 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1789 }
1790
1791 return alias_type(idx);
1792 }
1793
1794
1795 Compile::AliasType* Compile::alias_type(ciField* field) {
1796 const TypeOopPtr* t;
1797 if (field->is_static())
1798 t = TypeInstPtr::make(field->holder()->java_mirror());
1799 else
1800 t = TypeOopPtr::make_from_klass_raw(field->holder());
1801 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1802 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1803 return atp;
1804 }
1805
1806
1807 //------------------------------have_alias_type--------------------------------
1808 bool Compile::have_alias_type(const TypePtr* adr_type) {
1809 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1810 if (ace->_adr_type == adr_type) {
1811 return true;
1812 }
1813
1814 // Handle special cases.
1815 if (adr_type == nullptr) return true;
1816 if (adr_type == TypePtr::BOTTOM) return true;
1817
1818 return find_alias_type(adr_type, true, nullptr) != nullptr;
1819 }
1820
1821 //-----------------------------must_alias--------------------------------------
1822 // True if all values of the given address type are in the given alias category.
1823 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1824 if (alias_idx == AliasIdxBot) return true; // the universal category
1825 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1826 if (alias_idx == AliasIdxTop) return false; // the empty category
1827 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1828
1829 // the only remaining possible overlap is identity
1830 int adr_idx = get_alias_index(adr_type);
1831 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1832 assert(adr_idx == alias_idx ||
1833 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1834 && adr_type != TypeOopPtr::BOTTOM),
1835 "should not be testing for overlap with an unsafe pointer");
1836 return adr_idx == alias_idx;
1837 }
1838
1839 //------------------------------can_alias--------------------------------------
1840 // True if any values of the given address type are in the given alias category.
1841 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1842 if (alias_idx == AliasIdxTop) return false; // the empty category
1843 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1844 // Known instance doesn't alias with bottom memory
1845 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1846 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1847
1848 // the only remaining possible overlap is identity
1849 int adr_idx = get_alias_index(adr_type);
1850 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1851 return adr_idx == alias_idx;
1852 }
1853
1854 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1855 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1856 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1857 if (parse_predicate_count() == 0) {
1858 return;
1859 }
1860 for (int i = 0; i < parse_predicate_count(); i++) {
1861 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1862 parse_predicate->mark_useless(igvn);
1863 }
1864 _parse_predicates.clear();
1865 }
1866
1867 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1868 if (!n->for_post_loop_opts_igvn()) {
1869 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1870 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1871 _for_post_loop_igvn.append(n);
1872 }
1873 }
1874
1875 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1876 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1877 _for_post_loop_igvn.remove(n);
1878 }
1879
1880 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1881 // Verify that all previous optimizations produced a valid graph
1882 // at least to this point, even if no loop optimizations were done.
1883 PhaseIdealLoop::verify(igvn);
1884
1885 if (has_loops() || _loop_opts_cnt > 0) {
1886 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1887 }
1888 C->set_post_loop_opts_phase(); // no more loop opts allowed
1889
1890 assert(!C->major_progress(), "not cleared");
1891
1892 if (_for_post_loop_igvn.length() > 0) {
1893 while (_for_post_loop_igvn.length() > 0) {
1894 Node* n = _for_post_loop_igvn.pop();
1895 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1896 igvn._worklist.push(n);
1897 }
1898 igvn.optimize();
1899 if (failing()) return;
1900 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1901 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1902
1903 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1904 if (C->major_progress()) {
1905 C->clear_major_progress(); // ensure that major progress is now clear
1906 }
1907 }
1908 }
1909
1910 void Compile::record_for_merge_stores_igvn(Node* n) {
1911 if (!n->for_merge_stores_igvn()) {
1912 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1913 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1914 _for_merge_stores_igvn.append(n);
1915 }
1916 }
1917
1918 void Compile::remove_from_merge_stores_igvn(Node* n) {
1919 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1920 _for_merge_stores_igvn.remove(n);
1921 }
1922
1923 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1924 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1925 // the stores, and we merge the wrong sequence of stores.
1926 // Example:
1927 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1928 // Apply MergeStores:
1929 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1930 // Remove more RangeChecks:
1931 // StoreI [ StoreL ] StoreI
1932 // But now it would have been better to do this instead:
1933 // [ StoreL ] [ StoreL ]
1934 //
1935 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1936 // since we never unset _merge_stores_phase.
1937 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1938 C->set_merge_stores_phase();
1939
1940 if (_for_merge_stores_igvn.length() > 0) {
1941 while (_for_merge_stores_igvn.length() > 0) {
1942 Node* n = _for_merge_stores_igvn.pop();
1943 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1944 igvn._worklist.push(n);
1945 }
1946 igvn.optimize();
1947 if (failing()) return;
1948 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1949 print_method(PHASE_AFTER_MERGE_STORES, 3);
1950 }
1951 }
1952
1953 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1954 if (OptimizeUnstableIf) {
1955 _unstable_if_traps.append(trap);
1956 }
1957 }
1958
1959 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1960 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1961 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1962 Node* n = trap->uncommon_trap();
1963 if (!useful.member(n)) {
1964 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1965 }
1966 }
1967 }
1968
1969 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1970 // or fold-compares case. Return true if succeed or not found.
1971 //
1972 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1973 // false and ask fold-compares to yield.
1974 //
1975 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1976 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1977 // when deoptimization does happen.
1978 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1979 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1980 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1981 if (trap->uncommon_trap() == unc) {
1982 if (yield && trap->modified()) {
1983 return false;
1984 }
1985 _unstable_if_traps.delete_at(i);
1986 break;
1987 }
1988 }
1989 return true;
1990 }
1991
1992 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1993 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1994 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1995 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1996 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1997 CallStaticJavaNode* unc = trap->uncommon_trap();
1998 int next_bci = trap->next_bci();
1999 bool modified = trap->modified();
2000
2001 if (next_bci != -1 && !modified) {
2002 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2003 JVMState* jvms = unc->jvms();
2004 ciMethod* method = jvms->method();
2005 ciBytecodeStream iter(method);
2006
2007 iter.force_bci(jvms->bci());
2008 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2009 Bytecodes::Code c = iter.cur_bc();
2010 Node* lhs = nullptr;
2011 Node* rhs = nullptr;
2012 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2013 lhs = unc->peek_operand(0);
2014 rhs = unc->peek_operand(1);
2015 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2016 lhs = unc->peek_operand(0);
2017 }
2018
2019 ResourceMark rm;
2020 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2021 assert(live_locals.is_valid(), "broken liveness info");
2022 int len = (int)live_locals.size();
2023
2024 for (int i = 0; i < len; i++) {
2025 Node* local = unc->local(jvms, i);
2026 // kill local using the liveness of next_bci.
2027 // give up when the local looks like an operand to secure reexecution.
2028 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2029 uint idx = jvms->locoff() + i;
2030 #ifdef ASSERT
2031 if (PrintOpto && Verbose) {
2032 tty->print("[unstable_if] kill local#%d: ", idx);
2033 local->dump();
2034 tty->cr();
2035 }
2036 #endif
2037 igvn.replace_input_of(unc, idx, top());
2038 modified = true;
2039 }
2040 }
2041 }
2042
2043 // keep the mondified trap for late query
2044 if (modified) {
2045 trap->set_modified();
2046 } else {
2047 _unstable_if_traps.delete_at(i);
2048 }
2049 }
2050 igvn.optimize();
2051 }
2052
2053 // StringOpts and late inlining of string methods
2054 void Compile::inline_string_calls(bool parse_time) {
2055 {
2056 // remove useless nodes to make the usage analysis simpler
2057 ResourceMark rm;
2058 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2059 }
2060
2061 {
2062 ResourceMark rm;
2063 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2064 PhaseStringOpts pso(initial_gvn());
2065 print_method(PHASE_AFTER_STRINGOPTS, 3);
2066 }
2067
2068 // now inline anything that we skipped the first time around
2069 if (!parse_time) {
2070 _late_inlines_pos = _late_inlines.length();
2071 }
2072
2073 while (_string_late_inlines.length() > 0) {
2074 CallGenerator* cg = _string_late_inlines.pop();
2075 cg->do_late_inline();
2076 if (failing()) return;
2077 }
2078 _string_late_inlines.trunc_to(0);
2079 }
2080
2081 // Late inlining of boxing methods
2082 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2083 if (_boxing_late_inlines.length() > 0) {
2084 assert(has_boxed_value(), "inconsistent");
2085
2086 set_inlining_incrementally(true);
2087
2088 igvn_worklist()->ensure_empty(); // should be done with igvn
2089
2090 _late_inlines_pos = _late_inlines.length();
2091
2092 while (_boxing_late_inlines.length() > 0) {
2093 CallGenerator* cg = _boxing_late_inlines.pop();
2094 cg->do_late_inline();
2095 if (failing()) return;
2096 }
2097 _boxing_late_inlines.trunc_to(0);
2098
2099 inline_incrementally_cleanup(igvn);
2100
2101 set_inlining_incrementally(false);
2102 }
2103 }
2104
2105 bool Compile::inline_incrementally_one() {
2106 assert(IncrementalInline, "incremental inlining should be on");
2107
2108 TracePhase tp(_t_incrInline_inline);
2109
2110 set_inlining_progress(false);
2111 set_do_cleanup(false);
2112
2113 for (int i = 0; i < _late_inlines.length(); i++) {
2114 _late_inlines_pos = i+1;
2115 CallGenerator* cg = _late_inlines.at(i);
2116 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2117 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2118 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2119 cg->do_late_inline();
2120 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2121 if (failing()) {
2122 return false;
2123 } else if (inlining_progress()) {
2124 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2125 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2126 break; // process one call site at a time
2127 } else {
2128 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2129 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2130 // Avoid potential infinite loop if node already in the IGVN list
2131 assert(false, "scheduled for IGVN during inlining attempt");
2132 } else {
2133 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2134 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2135 cg->call_node()->set_generator(cg);
2136 }
2137 }
2138 } else {
2139 // Ignore late inline direct calls when inlining is not allowed.
2140 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2141 }
2142 }
2143 // Remove processed elements.
2144 _late_inlines.remove_till(_late_inlines_pos);
2145 _late_inlines_pos = 0;
2146
2147 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2148
2149 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2150
2151 set_inlining_progress(false);
2152 set_do_cleanup(false);
2153
2154 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2155 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2156 }
2157
2158 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2159 {
2160 TracePhase tp(_t_incrInline_pru);
2161 ResourceMark rm;
2162 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2163 }
2164 {
2165 TracePhase tp(_t_incrInline_igvn);
2166 igvn.reset_from_gvn(initial_gvn());
2167 igvn.optimize();
2168 if (failing()) return;
2169 }
2170 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2171 }
2172
2173 // Perform incremental inlining until bound on number of live nodes is reached
2174 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2175 TracePhase tp(_t_incrInline);
2176
2177 set_inlining_incrementally(true);
2178 uint low_live_nodes = 0;
2179
2180 while (_late_inlines.length() > 0) {
2181 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2182 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2183 TracePhase tp(_t_incrInline_ideal);
2184 // PhaseIdealLoop is expensive so we only try it once we are
2185 // out of live nodes and we only try it again if the previous
2186 // helped got the number of nodes down significantly
2187 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2188 if (failing()) return;
2189 low_live_nodes = live_nodes();
2190 _major_progress = true;
2191 }
2192
2193 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2194 bool do_print_inlining = print_inlining() || print_intrinsics();
2195 if (do_print_inlining || log() != nullptr) {
2196 // Print inlining message for candidates that we couldn't inline for lack of space.
2197 for (int i = 0; i < _late_inlines.length(); i++) {
2198 CallGenerator* cg = _late_inlines.at(i);
2199 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2200 if (do_print_inlining) {
2201 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2202 }
2203 log_late_inline_failure(cg, msg);
2204 }
2205 }
2206 break; // finish
2207 }
2208 }
2209
2210 igvn_worklist()->ensure_empty(); // should be done with igvn
2211
2212 while (inline_incrementally_one()) {
2213 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2214 }
2215 if (failing()) return;
2216
2217 inline_incrementally_cleanup(igvn);
2218
2219 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2220
2221 if (failing()) return;
2222
2223 if (_late_inlines.length() == 0) {
2224 break; // no more progress
2225 }
2226 }
2227
2228 igvn_worklist()->ensure_empty(); // should be done with igvn
2229
2230 if (_string_late_inlines.length() > 0) {
2231 assert(has_stringbuilder(), "inconsistent");
2232
2233 inline_string_calls(false);
2234
2235 if (failing()) return;
2236
2237 inline_incrementally_cleanup(igvn);
2238 }
2239
2240 set_inlining_incrementally(false);
2241 }
2242
2243 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2244 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2245 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2246 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2247 // as if "inlining_incrementally() == true" were set.
2248 assert(inlining_incrementally() == false, "not allowed");
2249 assert(_modified_nodes == nullptr, "not allowed");
2250 assert(_late_inlines.length() > 0, "sanity");
2251
2252 while (_late_inlines.length() > 0) {
2253 igvn_worklist()->ensure_empty(); // should be done with igvn
2254
2255 while (inline_incrementally_one()) {
2256 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2257 }
2258 if (failing()) return;
2259
2260 inline_incrementally_cleanup(igvn);
2261 }
2262 }
2263
2264 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2265 if (_loop_opts_cnt > 0) {
2266 while (major_progress() && (_loop_opts_cnt > 0)) {
2267 TracePhase tp(_t_idealLoop);
2268 PhaseIdealLoop::optimize(igvn, mode);
2269 _loop_opts_cnt--;
2270 if (failing()) return false;
2271 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2272 }
2273 }
2274 return true;
2275 }
2276
2277 // Remove edges from "root" to each SafePoint at a backward branch.
2278 // They were inserted during parsing (see add_safepoint()) to make
2279 // infinite loops without calls or exceptions visible to root, i.e.,
2280 // useful.
2281 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2282 Node *r = root();
2283 if (r != nullptr) {
2284 for (uint i = r->req(); i < r->len(); ++i) {
2285 Node *n = r->in(i);
2286 if (n != nullptr && n->is_SafePoint()) {
2287 r->rm_prec(i);
2288 if (n->outcnt() == 0) {
2289 igvn.remove_dead_node(n);
2290 }
2291 --i;
2292 }
2293 }
2294 // Parsing may have added top inputs to the root node (Path
2295 // leading to the Halt node proven dead). Make sure we get a
2296 // chance to clean them up.
2297 igvn._worklist.push(r);
2298 igvn.optimize();
2299 }
2300 }
2301
2302 //------------------------------Optimize---------------------------------------
2303 // Given a graph, optimize it.
2304 void Compile::Optimize() {
2305 TracePhase tp(_t_optimizer);
2306
2307 #ifndef PRODUCT
2308 if (env()->break_at_compile()) {
2309 BREAKPOINT;
2310 }
2311
2312 #endif
2313
2314 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2315 #ifdef ASSERT
2316 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2317 #endif
2318
2319 ResourceMark rm;
2320
2321 NOT_PRODUCT( verify_graph_edges(); )
2322
2323 print_method(PHASE_AFTER_PARSING, 1);
2324
2325 {
2326 // Iterative Global Value Numbering, including ideal transforms
2327 // Initialize IterGVN with types and values from parse-time GVN
2328 PhaseIterGVN igvn(initial_gvn());
2329 #ifdef ASSERT
2330 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2331 #endif
2332 {
2333 TracePhase tp(_t_iterGVN);
2334 igvn.optimize();
2335 }
2336
2337 if (failing()) return;
2338
2339 print_method(PHASE_ITER_GVN1, 2);
2340
2341 process_for_unstable_if_traps(igvn);
2342
2343 if (failing()) return;
2344
2345 inline_incrementally(igvn);
2346
2347 print_method(PHASE_INCREMENTAL_INLINE, 2);
2348
2349 if (failing()) return;
2350
2351 if (eliminate_boxing()) {
2352 // Inline valueOf() methods now.
2353 inline_boxing_calls(igvn);
2354
2355 if (failing()) return;
2356
2357 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2358 inline_incrementally(igvn);
2359 }
2360
2361 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2362
2363 if (failing()) return;
2364 }
2365
2366 // Remove the speculative part of types and clean up the graph from
2367 // the extra CastPP nodes whose only purpose is to carry them. Do
2368 // that early so that optimizations are not disrupted by the extra
2369 // CastPP nodes.
2370 remove_speculative_types(igvn);
2371
2372 if (failing()) return;
2373
2374 // No more new expensive nodes will be added to the list from here
2375 // so keep only the actual candidates for optimizations.
2376 cleanup_expensive_nodes(igvn);
2377
2378 if (failing()) return;
2379
2380 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2381 if (EnableVectorSupport && has_vbox_nodes()) {
2382 TracePhase tp(_t_vector);
2383 PhaseVector pv(igvn);
2384 pv.optimize_vector_boxes();
2385 if (failing()) return;
2386 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2387 }
2388 assert(!has_vbox_nodes(), "sanity");
2389
2390 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2391 Compile::TracePhase tp(_t_renumberLive);
2392 igvn_worklist()->ensure_empty(); // should be done with igvn
2393 {
2394 ResourceMark rm;
2395 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2396 }
2397 igvn.reset_from_gvn(initial_gvn());
2398 igvn.optimize();
2399 if (failing()) return;
2400 }
2401
2402 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2403 // safepoints
2404 remove_root_to_sfpts_edges(igvn);
2405
2406 if (failing()) return;
2407
2408 if (has_loops()) {
2409 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2410 }
2411
2412 // Perform escape analysis
2413 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2414 if (has_loops()) {
2415 // Cleanup graph (remove dead nodes).
2416 TracePhase tp(_t_idealLoop);
2417 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2418 if (failing()) return;
2419 }
2420 bool progress;
2421 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2422 do {
2423 ConnectionGraph::do_analysis(this, &igvn);
2424
2425 if (failing()) return;
2426
2427 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2428
2429 // Optimize out fields loads from scalar replaceable allocations.
2430 igvn.optimize();
2431 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2432
2433 if (failing()) return;
2434
2435 if (congraph() != nullptr && macro_count() > 0) {
2436 TracePhase tp(_t_macroEliminate);
2437 PhaseMacroExpand mexp(igvn);
2438 mexp.eliminate_macro_nodes();
2439 if (failing()) return;
2440
2441 igvn.set_delay_transform(false);
2442 igvn.optimize();
2443 if (failing()) return;
2444
2445 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2446 }
2447
2448 ConnectionGraph::verify_ram_nodes(this, root());
2449 if (failing()) return;
2450
2451 progress = do_iterative_escape_analysis() &&
2452 (macro_count() < mcount) &&
2453 ConnectionGraph::has_candidates(this);
2454 // Try again if candidates exist and made progress
2455 // by removing some allocations and/or locks.
2456 } while (progress);
2457 }
2458
2459 // Loop transforms on the ideal graph. Range Check Elimination,
2460 // peeling, unrolling, etc.
2461
2462 // Set loop opts counter
2463 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2464 {
2465 TracePhase tp(_t_idealLoop);
2466 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2467 _loop_opts_cnt--;
2468 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2469 if (failing()) return;
2470 }
2471 // Loop opts pass if partial peeling occurred in previous pass
2472 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2473 TracePhase tp(_t_idealLoop);
2474 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2475 _loop_opts_cnt--;
2476 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2477 if (failing()) return;
2478 }
2479 // Loop opts pass for loop-unrolling before CCP
2480 if(major_progress() && (_loop_opts_cnt > 0)) {
2481 TracePhase tp(_t_idealLoop);
2482 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2483 _loop_opts_cnt--;
2484 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2485 }
2486 if (!failing()) {
2487 // Verify that last round of loop opts produced a valid graph
2488 PhaseIdealLoop::verify(igvn);
2489 }
2490 }
2491 if (failing()) return;
2492
2493 // Conditional Constant Propagation;
2494 print_method(PHASE_BEFORE_CCP1, 2);
2495 PhaseCCP ccp( &igvn );
2496 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2497 {
2498 TracePhase tp(_t_ccp);
2499 ccp.do_transform();
2500 }
2501 print_method(PHASE_CCP1, 2);
2502
2503 assert( true, "Break here to ccp.dump_old2new_map()");
2504
2505 // Iterative Global Value Numbering, including ideal transforms
2506 {
2507 TracePhase tp(_t_iterGVN2);
2508 igvn.reset_from_igvn(&ccp);
2509 igvn.optimize();
2510 }
2511 print_method(PHASE_ITER_GVN2, 2);
2512
2513 if (failing()) return;
2514
2515 // Loop transforms on the ideal graph. Range Check Elimination,
2516 // peeling, unrolling, etc.
2517 if (!optimize_loops(igvn, LoopOptsDefault)) {
2518 return;
2519 }
2520
2521 if (failing()) return;
2522
2523 C->clear_major_progress(); // ensure that major progress is now clear
2524
2525 process_for_post_loop_opts_igvn(igvn);
2526
2527 process_for_merge_stores_igvn(igvn);
2528
2529 if (failing()) return;
2530
2531 #ifdef ASSERT
2532 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2533 #endif
2534
2535 {
2536 TracePhase tp(_t_macroExpand);
2537 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2538 PhaseMacroExpand mex(igvn);
2539 if (mex.expand_macro_nodes()) {
2540 assert(failing(), "must bail out w/ explicit message");
2541 return;
2542 }
2543 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2544 }
2545
2546 {
2547 TracePhase tp(_t_barrierExpand);
2548 if (bs->expand_barriers(this, igvn)) {
2549 assert(failing(), "must bail out w/ explicit message");
2550 return;
2551 }
2552 print_method(PHASE_BARRIER_EXPANSION, 2);
2553 }
2554
2555 if (C->max_vector_size() > 0) {
2556 C->optimize_logic_cones(igvn);
2557 igvn.optimize();
2558 if (failing()) return;
2559 }
2560
2561 DEBUG_ONLY( _modified_nodes = nullptr; )
2562
2563 assert(igvn._worklist.size() == 0, "not empty");
2564
2565 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2566
2567 if (_late_inlines.length() > 0) {
2568 // More opportunities to optimize virtual and MH calls.
2569 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2570 process_late_inline_calls_no_inline(igvn);
2571 if (failing()) return;
2572 }
2573 } // (End scope of igvn; run destructor if necessary for asserts.)
2574
2575 check_no_dead_use();
2576
2577 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2578 // to remove hashes to unlock nodes for modifications.
2579 C->node_hash()->clear();
2580
2581 // A method with only infinite loops has no edges entering loops from root
2582 {
2583 TracePhase tp(_t_graphReshaping);
2584 if (final_graph_reshaping()) {
2585 assert(failing(), "must bail out w/ explicit message");
2586 return;
2587 }
2588 }
2589
2590 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2591 DEBUG_ONLY(set_phase_optimize_finished();)
2592 }
2593
2594 #ifdef ASSERT
2595 void Compile::check_no_dead_use() const {
2596 ResourceMark rm;
2597 Unique_Node_List wq;
2598 wq.push(root());
2599 for (uint i = 0; i < wq.size(); ++i) {
2600 Node* n = wq.at(i);
2601 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2602 Node* u = n->fast_out(j);
2603 if (u->outcnt() == 0 && !u->is_Con()) {
2604 u->dump();
2605 fatal("no reachable node should have no use");
2606 }
2607 wq.push(u);
2608 }
2609 }
2610 }
2611 #endif
2612
2613 void Compile::inline_vector_reboxing_calls() {
2614 if (C->_vector_reboxing_late_inlines.length() > 0) {
2615 _late_inlines_pos = C->_late_inlines.length();
2616 while (_vector_reboxing_late_inlines.length() > 0) {
2617 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2618 cg->do_late_inline();
2619 if (failing()) return;
2620 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2621 }
2622 _vector_reboxing_late_inlines.trunc_to(0);
2623 }
2624 }
2625
2626 bool Compile::has_vbox_nodes() {
2627 if (C->_vector_reboxing_late_inlines.length() > 0) {
2628 return true;
2629 }
2630 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2631 Node * n = C->macro_node(macro_idx);
2632 assert(n->is_macro(), "only macro nodes expected here");
2633 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2634 return true;
2635 }
2636 }
2637 return false;
2638 }
2639
2640 //---------------------------- Bitwise operation packing optimization ---------------------------
2641
2642 static bool is_vector_unary_bitwise_op(Node* n) {
2643 return n->Opcode() == Op_XorV &&
2644 VectorNode::is_vector_bitwise_not_pattern(n);
2645 }
2646
2647 static bool is_vector_binary_bitwise_op(Node* n) {
2648 switch (n->Opcode()) {
2649 case Op_AndV:
2650 case Op_OrV:
2651 return true;
2652
2653 case Op_XorV:
2654 return !is_vector_unary_bitwise_op(n);
2655
2656 default:
2657 return false;
2658 }
2659 }
2660
2661 static bool is_vector_ternary_bitwise_op(Node* n) {
2662 return n->Opcode() == Op_MacroLogicV;
2663 }
2664
2665 static bool is_vector_bitwise_op(Node* n) {
2666 return is_vector_unary_bitwise_op(n) ||
2667 is_vector_binary_bitwise_op(n) ||
2668 is_vector_ternary_bitwise_op(n);
2669 }
2670
2671 static bool is_vector_bitwise_cone_root(Node* n) {
2672 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2673 return false;
2674 }
2675 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2676 if (is_vector_bitwise_op(n->fast_out(i))) {
2677 return false;
2678 }
2679 }
2680 return true;
2681 }
2682
2683 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2684 uint cnt = 0;
2685 if (is_vector_bitwise_op(n)) {
2686 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2687 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2688 for (uint i = 1; i < inp_cnt; i++) {
2689 Node* in = n->in(i);
2690 bool skip = VectorNode::is_all_ones_vector(in);
2691 if (!skip && !inputs.member(in)) {
2692 inputs.push(in);
2693 cnt++;
2694 }
2695 }
2696 assert(cnt <= 1, "not unary");
2697 } else {
2698 uint last_req = inp_cnt;
2699 if (is_vector_ternary_bitwise_op(n)) {
2700 last_req = inp_cnt - 1; // skip last input
2701 }
2702 for (uint i = 1; i < last_req; i++) {
2703 Node* def = n->in(i);
2704 if (!inputs.member(def)) {
2705 inputs.push(def);
2706 cnt++;
2707 }
2708 }
2709 }
2710 } else { // not a bitwise operations
2711 if (!inputs.member(n)) {
2712 inputs.push(n);
2713 cnt++;
2714 }
2715 }
2716 return cnt;
2717 }
2718
2719 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2720 Unique_Node_List useful_nodes;
2721 C->identify_useful_nodes(useful_nodes);
2722
2723 for (uint i = 0; i < useful_nodes.size(); i++) {
2724 Node* n = useful_nodes.at(i);
2725 if (is_vector_bitwise_cone_root(n)) {
2726 list.push(n);
2727 }
2728 }
2729 }
2730
2731 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2732 const TypeVect* vt,
2733 Unique_Node_List& partition,
2734 Unique_Node_List& inputs) {
2735 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2736 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2737 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2738
2739 Node* in1 = inputs.at(0);
2740 Node* in2 = inputs.at(1);
2741 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2742
2743 uint func = compute_truth_table(partition, inputs);
2744
2745 Node* pn = partition.at(partition.size() - 1);
2746 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2747 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2748 }
2749
2750 static uint extract_bit(uint func, uint pos) {
2751 return (func & (1 << pos)) >> pos;
2752 }
2753
2754 //
2755 // A macro logic node represents a truth table. It has 4 inputs,
2756 // First three inputs corresponds to 3 columns of a truth table
2757 // and fourth input captures the logic function.
2758 //
2759 // eg. fn = (in1 AND in2) OR in3;
2760 //
2761 // MacroNode(in1,in2,in3,fn)
2762 //
2763 // -----------------
2764 // in1 in2 in3 fn
2765 // -----------------
2766 // 0 0 0 0
2767 // 0 0 1 1
2768 // 0 1 0 0
2769 // 0 1 1 1
2770 // 1 0 0 0
2771 // 1 0 1 1
2772 // 1 1 0 1
2773 // 1 1 1 1
2774 //
2775
2776 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2777 int res = 0;
2778 for (int i = 0; i < 8; i++) {
2779 int bit1 = extract_bit(in1, i);
2780 int bit2 = extract_bit(in2, i);
2781 int bit3 = extract_bit(in3, i);
2782
2783 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2784 int func_bit = extract_bit(func, func_bit_pos);
2785
2786 res |= func_bit << i;
2787 }
2788 return res;
2789 }
2790
2791 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2792 assert(n != nullptr, "");
2793 assert(eval_map.contains(n), "absent");
2794 return *(eval_map.get(n));
2795 }
2796
2797 static void eval_operands(Node* n,
2798 uint& func1, uint& func2, uint& func3,
2799 ResourceHashtable<Node*,uint>& eval_map) {
2800 assert(is_vector_bitwise_op(n), "");
2801
2802 if (is_vector_unary_bitwise_op(n)) {
2803 Node* opnd = n->in(1);
2804 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2805 opnd = n->in(2);
2806 }
2807 func1 = eval_operand(opnd, eval_map);
2808 } else if (is_vector_binary_bitwise_op(n)) {
2809 func1 = eval_operand(n->in(1), eval_map);
2810 func2 = eval_operand(n->in(2), eval_map);
2811 } else {
2812 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2813 func1 = eval_operand(n->in(1), eval_map);
2814 func2 = eval_operand(n->in(2), eval_map);
2815 func3 = eval_operand(n->in(3), eval_map);
2816 }
2817 }
2818
2819 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2820 assert(inputs.size() <= 3, "sanity");
2821 ResourceMark rm;
2822 uint res = 0;
2823 ResourceHashtable<Node*,uint> eval_map;
2824
2825 // Populate precomputed functions for inputs.
2826 // Each input corresponds to one column of 3 input truth-table.
2827 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2828 0xCC, // (_, b, _) -> b
2829 0xF0 }; // (a, _, _) -> a
2830 for (uint i = 0; i < inputs.size(); i++) {
2831 eval_map.put(inputs.at(i), input_funcs[2-i]);
2832 }
2833
2834 for (uint i = 0; i < partition.size(); i++) {
2835 Node* n = partition.at(i);
2836
2837 uint func1 = 0, func2 = 0, func3 = 0;
2838 eval_operands(n, func1, func2, func3, eval_map);
2839
2840 switch (n->Opcode()) {
2841 case Op_OrV:
2842 assert(func3 == 0, "not binary");
2843 res = func1 | func2;
2844 break;
2845 case Op_AndV:
2846 assert(func3 == 0, "not binary");
2847 res = func1 & func2;
2848 break;
2849 case Op_XorV:
2850 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2851 assert(func2 == 0 && func3 == 0, "not unary");
2852 res = (~func1) & 0xFF;
2853 } else {
2854 assert(func3 == 0, "not binary");
2855 res = func1 ^ func2;
2856 }
2857 break;
2858 case Op_MacroLogicV:
2859 // Ordering of inputs may change during evaluation of sub-tree
2860 // containing MacroLogic node as a child node, thus a re-evaluation
2861 // makes sure that function is evaluated in context of current
2862 // inputs.
2863 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2864 break;
2865
2866 default: assert(false, "not supported: %s", n->Name());
2867 }
2868 assert(res <= 0xFF, "invalid");
2869 eval_map.put(n, res);
2870 }
2871 return res;
2872 }
2873
2874 // Criteria under which nodes gets packed into a macro logic node:-
2875 // 1) Parent and both child nodes are all unmasked or masked with
2876 // same predicates.
2877 // 2) Masked parent can be packed with left child if it is predicated
2878 // and both have same predicates.
2879 // 3) Masked parent can be packed with right child if its un-predicated
2880 // or has matching predication condition.
2881 // 4) An unmasked parent can be packed with an unmasked child.
2882 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2883 assert(partition.size() == 0, "not empty");
2884 assert(inputs.size() == 0, "not empty");
2885 if (is_vector_ternary_bitwise_op(n)) {
2886 return false;
2887 }
2888
2889 bool is_unary_op = is_vector_unary_bitwise_op(n);
2890 if (is_unary_op) {
2891 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2892 return false; // too few inputs
2893 }
2894
2895 bool pack_left_child = true;
2896 bool pack_right_child = true;
2897
2898 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2899 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2900
2901 int left_child_input_cnt = 0;
2902 int right_child_input_cnt = 0;
2903
2904 bool parent_is_predicated = n->is_predicated_vector();
2905 bool left_child_predicated = n->in(1)->is_predicated_vector();
2906 bool right_child_predicated = n->in(2)->is_predicated_vector();
2907
2908 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2909 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2910 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2911
2912 do {
2913 if (pack_left_child && left_child_LOP &&
2914 ((!parent_is_predicated && !left_child_predicated) ||
2915 ((parent_is_predicated && left_child_predicated &&
2916 parent_pred == left_child_pred)))) {
2917 partition.push(n->in(1));
2918 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2919 } else {
2920 inputs.push(n->in(1));
2921 left_child_input_cnt = 1;
2922 }
2923
2924 if (pack_right_child && right_child_LOP &&
2925 (!right_child_predicated ||
2926 (right_child_predicated && parent_is_predicated &&
2927 parent_pred == right_child_pred))) {
2928 partition.push(n->in(2));
2929 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2930 } else {
2931 inputs.push(n->in(2));
2932 right_child_input_cnt = 1;
2933 }
2934
2935 if (inputs.size() > 3) {
2936 assert(partition.size() > 0, "");
2937 inputs.clear();
2938 partition.clear();
2939 if (left_child_input_cnt > right_child_input_cnt) {
2940 pack_left_child = false;
2941 } else {
2942 pack_right_child = false;
2943 }
2944 } else {
2945 break;
2946 }
2947 } while(true);
2948
2949 if(partition.size()) {
2950 partition.push(n);
2951 }
2952
2953 return (partition.size() == 2 || partition.size() == 3) &&
2954 (inputs.size() == 2 || inputs.size() == 3);
2955 }
2956
2957 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2958 assert(is_vector_bitwise_op(n), "not a root");
2959
2960 visited.set(n->_idx);
2961
2962 // 1) Do a DFS walk over the logic cone.
2963 for (uint i = 1; i < n->req(); i++) {
2964 Node* in = n->in(i);
2965 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2966 process_logic_cone_root(igvn, in, visited);
2967 }
2968 }
2969
2970 // 2) Bottom up traversal: Merge node[s] with
2971 // the parent to form macro logic node.
2972 Unique_Node_List partition;
2973 Unique_Node_List inputs;
2974 if (compute_logic_cone(n, partition, inputs)) {
2975 const TypeVect* vt = n->bottom_type()->is_vect();
2976 Node* pn = partition.at(partition.size() - 1);
2977 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2978 if (mask == nullptr ||
2979 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2980 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2981 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2982 igvn.replace_node(n, macro_logic);
2983 }
2984 }
2985 }
2986
2987 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2988 ResourceMark rm;
2989 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2990 Unique_Node_List list;
2991 collect_logic_cone_roots(list);
2992
2993 while (list.size() > 0) {
2994 Node* n = list.pop();
2995 const TypeVect* vt = n->bottom_type()->is_vect();
2996 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2997 if (supported) {
2998 VectorSet visited(comp_arena());
2999 process_logic_cone_root(igvn, n, visited);
3000 }
3001 }
3002 }
3003 }
3004
3005 //------------------------------Code_Gen---------------------------------------
3006 // Given a graph, generate code for it
3007 void Compile::Code_Gen() {
3008 if (failing()) {
3009 return;
3010 }
3011
3012 // Perform instruction selection. You might think we could reclaim Matcher
3013 // memory PDQ, but actually the Matcher is used in generating spill code.
3014 // Internals of the Matcher (including some VectorSets) must remain live
3015 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3016 // set a bit in reclaimed memory.
3017
3018 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3019 // nodes. Mapping is only valid at the root of each matched subtree.
3020 NOT_PRODUCT( verify_graph_edges(); )
3021
3022 Matcher matcher;
3023 _matcher = &matcher;
3024 {
3025 TracePhase tp(_t_matcher);
3026 matcher.match();
3027 if (failing()) {
3028 return;
3029 }
3030 }
3031 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3032 // nodes. Mapping is only valid at the root of each matched subtree.
3033 NOT_PRODUCT( verify_graph_edges(); )
3034
3035 // If you have too many nodes, or if matching has failed, bail out
3036 check_node_count(0, "out of nodes matching instructions");
3037 if (failing()) {
3038 return;
3039 }
3040
3041 print_method(PHASE_MATCHING, 2);
3042
3043 // Build a proper-looking CFG
3044 PhaseCFG cfg(node_arena(), root(), matcher);
3045 if (failing()) {
3046 return;
3047 }
3048 _cfg = &cfg;
3049 {
3050 TracePhase tp(_t_scheduler);
3051 bool success = cfg.do_global_code_motion();
3052 if (!success) {
3053 return;
3054 }
3055
3056 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3057 NOT_PRODUCT( verify_graph_edges(); )
3058 cfg.verify();
3059 if (failing()) {
3060 return;
3061 }
3062 }
3063
3064 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3065 _regalloc = ®alloc;
3066 {
3067 TracePhase tp(_t_registerAllocation);
3068 // Perform register allocation. After Chaitin, use-def chains are
3069 // no longer accurate (at spill code) and so must be ignored.
3070 // Node->LRG->reg mappings are still accurate.
3071 _regalloc->Register_Allocate();
3072
3073 // Bail out if the allocator builds too many nodes
3074 if (failing()) {
3075 return;
3076 }
3077
3078 print_method(PHASE_REGISTER_ALLOCATION, 2);
3079 }
3080
3081 // Prior to register allocation we kept empty basic blocks in case the
3082 // the allocator needed a place to spill. After register allocation we
3083 // are not adding any new instructions. If any basic block is empty, we
3084 // can now safely remove it.
3085 {
3086 TracePhase tp(_t_blockOrdering);
3087 cfg.remove_empty_blocks();
3088 if (do_freq_based_layout()) {
3089 PhaseBlockLayout layout(cfg);
3090 } else {
3091 cfg.set_loop_alignment();
3092 }
3093 cfg.fixup_flow();
3094 cfg.remove_unreachable_blocks();
3095 cfg.verify_dominator_tree();
3096 print_method(PHASE_BLOCK_ORDERING, 3);
3097 }
3098
3099 // Apply peephole optimizations
3100 if( OptoPeephole ) {
3101 TracePhase tp(_t_peephole);
3102 PhasePeephole peep( _regalloc, cfg);
3103 peep.do_transform();
3104 print_method(PHASE_PEEPHOLE, 3);
3105 }
3106
3107 // Do late expand if CPU requires this.
3108 if (Matcher::require_postalloc_expand) {
3109 TracePhase tp(_t_postalloc_expand);
3110 cfg.postalloc_expand(_regalloc);
3111 print_method(PHASE_POSTALLOC_EXPAND, 3);
3112 }
3113
3114 #ifdef ASSERT
3115 {
3116 CompilationMemoryStatistic::do_test_allocations();
3117 if (failing()) return;
3118 }
3119 #endif
3120
3121 // Convert Nodes to instruction bits in a buffer
3122 {
3123 TracePhase tp(_t_output);
3124 PhaseOutput output;
3125 output.Output();
3126 if (failing()) return;
3127 output.install();
3128 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3129 }
3130
3131 // He's dead, Jim.
3132 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3133 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3134 }
3135
3136 //------------------------------Final_Reshape_Counts---------------------------
3137 // This class defines counters to help identify when a method
3138 // may/must be executed using hardware with only 24-bit precision.
3139 struct Final_Reshape_Counts : public StackObj {
3140 int _call_count; // count non-inlined 'common' calls
3141 int _float_count; // count float ops requiring 24-bit precision
3142 int _double_count; // count double ops requiring more precision
3143 int _java_call_count; // count non-inlined 'java' calls
3144 int _inner_loop_count; // count loops which need alignment
3145 VectorSet _visited; // Visitation flags
3146 Node_List _tests; // Set of IfNodes & PCTableNodes
3147
3148 Final_Reshape_Counts() :
3149 _call_count(0), _float_count(0), _double_count(0),
3150 _java_call_count(0), _inner_loop_count(0) { }
3151
3152 void inc_call_count () { _call_count ++; }
3153 void inc_float_count () { _float_count ++; }
3154 void inc_double_count() { _double_count++; }
3155 void inc_java_call_count() { _java_call_count++; }
3156 void inc_inner_loop_count() { _inner_loop_count++; }
3157
3158 int get_call_count () const { return _call_count ; }
3159 int get_float_count () const { return _float_count ; }
3160 int get_double_count() const { return _double_count; }
3161 int get_java_call_count() const { return _java_call_count; }
3162 int get_inner_loop_count() const { return _inner_loop_count; }
3163 };
3164
3165 //------------------------------final_graph_reshaping_impl----------------------
3166 // Implement items 1-5 from final_graph_reshaping below.
3167 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3168
3169 if ( n->outcnt() == 0 ) return; // dead node
3170 uint nop = n->Opcode();
3171
3172 // Check for 2-input instruction with "last use" on right input.
3173 // Swap to left input. Implements item (2).
3174 if( n->req() == 3 && // two-input instruction
3175 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3176 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3177 n->in(2)->outcnt() == 1 &&// right use IS a last use
3178 !n->in(2)->is_Con() ) { // right use is not a constant
3179 // Check for commutative opcode
3180 switch( nop ) {
3181 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3182 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3183 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3184 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3185 case Op_AndL: case Op_XorL: case Op_OrL:
3186 case Op_AndI: case Op_XorI: case Op_OrI: {
3187 // Move "last use" input to left by swapping inputs
3188 n->swap_edges(1, 2);
3189 break;
3190 }
3191 default:
3192 break;
3193 }
3194 }
3195
3196 #ifdef ASSERT
3197 if( n->is_Mem() ) {
3198 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3199 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3200 // oop will be recorded in oop map if load crosses safepoint
3201 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3202 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3203 "raw memory operations should have control edge");
3204 }
3205 if (n->is_MemBar()) {
3206 MemBarNode* mb = n->as_MemBar();
3207 if (mb->trailing_store() || mb->trailing_load_store()) {
3208 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3209 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3210 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3211 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3212 } else if (mb->leading()) {
3213 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3214 }
3215 }
3216 #endif
3217 // Count FPU ops and common calls, implements item (3)
3218 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3219 if (!gc_handled) {
3220 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3221 }
3222
3223 // Collect CFG split points
3224 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3225 frc._tests.push(n);
3226 }
3227 }
3228
3229 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3230 if (!UseDivMod) {
3231 return;
3232 }
3233
3234 // Check if "a % b" and "a / b" both exist
3235 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3236 if (d == nullptr) {
3237 return;
3238 }
3239
3240 // Replace them with a fused divmod if supported
3241 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3242 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3243 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3244 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3245 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3246 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3247 // DivMod node so the dependency is not lost.
3248 divmod->add_prec_from(n);
3249 divmod->add_prec_from(d);
3250 d->subsume_by(divmod->div_proj(), this);
3251 n->subsume_by(divmod->mod_proj(), this);
3252 } else {
3253 // Replace "a % b" with "a - ((a / b) * b)"
3254 Node* mult = MulNode::make(d, d->in(2), bt);
3255 Node* sub = SubNode::make(d->in(1), mult, bt);
3256 n->subsume_by(sub, this);
3257 }
3258 }
3259
3260 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3261 switch( nop ) {
3262 // Count all float operations that may use FPU
3263 case Op_AddF:
3264 case Op_SubF:
3265 case Op_MulF:
3266 case Op_DivF:
3267 case Op_NegF:
3268 case Op_ModF:
3269 case Op_ConvI2F:
3270 case Op_ConF:
3271 case Op_CmpF:
3272 case Op_CmpF3:
3273 case Op_StoreF:
3274 case Op_LoadF:
3275 // case Op_ConvL2F: // longs are split into 32-bit halves
3276 frc.inc_float_count();
3277 break;
3278
3279 case Op_ConvF2D:
3280 case Op_ConvD2F:
3281 frc.inc_float_count();
3282 frc.inc_double_count();
3283 break;
3284
3285 // Count all double operations that may use FPU
3286 case Op_AddD:
3287 case Op_SubD:
3288 case Op_MulD:
3289 case Op_DivD:
3290 case Op_NegD:
3291 case Op_ModD:
3292 case Op_ConvI2D:
3293 case Op_ConvD2I:
3294 // case Op_ConvL2D: // handled by leaf call
3295 // case Op_ConvD2L: // handled by leaf call
3296 case Op_ConD:
3297 case Op_CmpD:
3298 case Op_CmpD3:
3299 case Op_StoreD:
3300 case Op_LoadD:
3301 case Op_LoadD_unaligned:
3302 frc.inc_double_count();
3303 break;
3304 case Op_Opaque1: // Remove Opaque Nodes before matching
3305 n->subsume_by(n->in(1), this);
3306 break;
3307 case Op_CallStaticJava:
3308 case Op_CallJava:
3309 case Op_CallDynamicJava:
3310 frc.inc_java_call_count(); // Count java call site;
3311 case Op_CallRuntime:
3312 case Op_CallLeaf:
3313 case Op_CallLeafVector:
3314 case Op_CallLeafNoFP: {
3315 assert (n->is_Call(), "");
3316 CallNode *call = n->as_Call();
3317 // Count call sites where the FP mode bit would have to be flipped.
3318 // Do not count uncommon runtime calls:
3319 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3320 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3321 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3322 frc.inc_call_count(); // Count the call site
3323 } else { // See if uncommon argument is shared
3324 Node *n = call->in(TypeFunc::Parms);
3325 int nop = n->Opcode();
3326 // Clone shared simple arguments to uncommon calls, item (1).
3327 if (n->outcnt() > 1 &&
3328 !n->is_Proj() &&
3329 nop != Op_CreateEx &&
3330 nop != Op_CheckCastPP &&
3331 nop != Op_DecodeN &&
3332 nop != Op_DecodeNKlass &&
3333 !n->is_Mem() &&
3334 !n->is_Phi()) {
3335 Node *x = n->clone();
3336 call->set_req(TypeFunc::Parms, x);
3337 }
3338 }
3339 break;
3340 }
3341 case Op_StoreB:
3342 case Op_StoreC:
3343 case Op_StoreI:
3344 case Op_StoreL:
3345 case Op_CompareAndSwapB:
3346 case Op_CompareAndSwapS:
3347 case Op_CompareAndSwapI:
3348 case Op_CompareAndSwapL:
3349 case Op_CompareAndSwapP:
3350 case Op_CompareAndSwapN:
3351 case Op_WeakCompareAndSwapB:
3352 case Op_WeakCompareAndSwapS:
3353 case Op_WeakCompareAndSwapI:
3354 case Op_WeakCompareAndSwapL:
3355 case Op_WeakCompareAndSwapP:
3356 case Op_WeakCompareAndSwapN:
3357 case Op_CompareAndExchangeB:
3358 case Op_CompareAndExchangeS:
3359 case Op_CompareAndExchangeI:
3360 case Op_CompareAndExchangeL:
3361 case Op_CompareAndExchangeP:
3362 case Op_CompareAndExchangeN:
3363 case Op_GetAndAddS:
3364 case Op_GetAndAddB:
3365 case Op_GetAndAddI:
3366 case Op_GetAndAddL:
3367 case Op_GetAndSetS:
3368 case Op_GetAndSetB:
3369 case Op_GetAndSetI:
3370 case Op_GetAndSetL:
3371 case Op_GetAndSetP:
3372 case Op_GetAndSetN:
3373 case Op_StoreP:
3374 case Op_StoreN:
3375 case Op_StoreNKlass:
3376 case Op_LoadB:
3377 case Op_LoadUB:
3378 case Op_LoadUS:
3379 case Op_LoadI:
3380 case Op_LoadKlass:
3381 case Op_LoadNKlass:
3382 case Op_LoadL:
3383 case Op_LoadL_unaligned:
3384 case Op_LoadP:
3385 case Op_LoadN:
3386 case Op_LoadRange:
3387 case Op_LoadS:
3388 break;
3389
3390 case Op_AddP: { // Assert sane base pointers
3391 Node *addp = n->in(AddPNode::Address);
3392 assert( !addp->is_AddP() ||
3393 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3394 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3395 "Base pointers must match (addp %u)", addp->_idx );
3396 #ifdef _LP64
3397 if ((UseCompressedOops || UseCompressedClassPointers) &&
3398 addp->Opcode() == Op_ConP &&
3399 addp == n->in(AddPNode::Base) &&
3400 n->in(AddPNode::Offset)->is_Con()) {
3401 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3402 // on the platform and on the compressed oops mode.
3403 // Use addressing with narrow klass to load with offset on x86.
3404 // Some platforms can use the constant pool to load ConP.
3405 // Do this transformation here since IGVN will convert ConN back to ConP.
3406 const Type* t = addp->bottom_type();
3407 bool is_oop = t->isa_oopptr() != nullptr;
3408 bool is_klass = t->isa_klassptr() != nullptr;
3409
3410 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3411 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3412 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3413 Node* nn = nullptr;
3414
3415 int op = is_oop ? Op_ConN : Op_ConNKlass;
3416
3417 // Look for existing ConN node of the same exact type.
3418 Node* r = root();
3419 uint cnt = r->outcnt();
3420 for (uint i = 0; i < cnt; i++) {
3421 Node* m = r->raw_out(i);
3422 if (m!= nullptr && m->Opcode() == op &&
3423 m->bottom_type()->make_ptr() == t) {
3424 nn = m;
3425 break;
3426 }
3427 }
3428 if (nn != nullptr) {
3429 // Decode a narrow oop to match address
3430 // [R12 + narrow_oop_reg<<3 + offset]
3431 if (is_oop) {
3432 nn = new DecodeNNode(nn, t);
3433 } else {
3434 nn = new DecodeNKlassNode(nn, t);
3435 }
3436 // Check for succeeding AddP which uses the same Base.
3437 // Otherwise we will run into the assertion above when visiting that guy.
3438 for (uint i = 0; i < n->outcnt(); ++i) {
3439 Node *out_i = n->raw_out(i);
3440 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3441 out_i->set_req(AddPNode::Base, nn);
3442 #ifdef ASSERT
3443 for (uint j = 0; j < out_i->outcnt(); ++j) {
3444 Node *out_j = out_i->raw_out(j);
3445 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3446 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3447 }
3448 #endif
3449 }
3450 }
3451 n->set_req(AddPNode::Base, nn);
3452 n->set_req(AddPNode::Address, nn);
3453 if (addp->outcnt() == 0) {
3454 addp->disconnect_inputs(this);
3455 }
3456 }
3457 }
3458 }
3459 #endif
3460 break;
3461 }
3462
3463 case Op_CastPP: {
3464 // Remove CastPP nodes to gain more freedom during scheduling but
3465 // keep the dependency they encode as control or precedence edges
3466 // (if control is set already) on memory operations. Some CastPP
3467 // nodes don't have a control (don't carry a dependency): skip
3468 // those.
3469 if (n->in(0) != nullptr) {
3470 ResourceMark rm;
3471 Unique_Node_List wq;
3472 wq.push(n);
3473 for (uint next = 0; next < wq.size(); ++next) {
3474 Node *m = wq.at(next);
3475 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3476 Node* use = m->fast_out(i);
3477 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3478 use->ensure_control_or_add_prec(n->in(0));
3479 } else {
3480 switch(use->Opcode()) {
3481 case Op_AddP:
3482 case Op_DecodeN:
3483 case Op_DecodeNKlass:
3484 case Op_CheckCastPP:
3485 case Op_CastPP:
3486 wq.push(use);
3487 break;
3488 }
3489 }
3490 }
3491 }
3492 }
3493 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3494 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3495 Node* in1 = n->in(1);
3496 const Type* t = n->bottom_type();
3497 Node* new_in1 = in1->clone();
3498 new_in1->as_DecodeN()->set_type(t);
3499
3500 if (!Matcher::narrow_oop_use_complex_address()) {
3501 //
3502 // x86, ARM and friends can handle 2 adds in addressing mode
3503 // and Matcher can fold a DecodeN node into address by using
3504 // a narrow oop directly and do implicit null check in address:
3505 //
3506 // [R12 + narrow_oop_reg<<3 + offset]
3507 // NullCheck narrow_oop_reg
3508 //
3509 // On other platforms (Sparc) we have to keep new DecodeN node and
3510 // use it to do implicit null check in address:
3511 //
3512 // decode_not_null narrow_oop_reg, base_reg
3513 // [base_reg + offset]
3514 // NullCheck base_reg
3515 //
3516 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3517 // to keep the information to which null check the new DecodeN node
3518 // corresponds to use it as value in implicit_null_check().
3519 //
3520 new_in1->set_req(0, n->in(0));
3521 }
3522
3523 n->subsume_by(new_in1, this);
3524 if (in1->outcnt() == 0) {
3525 in1->disconnect_inputs(this);
3526 }
3527 } else {
3528 n->subsume_by(n->in(1), this);
3529 if (n->outcnt() == 0) {
3530 n->disconnect_inputs(this);
3531 }
3532 }
3533 break;
3534 }
3535 case Op_CastII: {
3536 n->as_CastII()->remove_range_check_cast(this);
3537 break;
3538 }
3539 #ifdef _LP64
3540 case Op_CmpP:
3541 // Do this transformation here to preserve CmpPNode::sub() and
3542 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3543 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3544 Node* in1 = n->in(1);
3545 Node* in2 = n->in(2);
3546 if (!in1->is_DecodeNarrowPtr()) {
3547 in2 = in1;
3548 in1 = n->in(2);
3549 }
3550 assert(in1->is_DecodeNarrowPtr(), "sanity");
3551
3552 Node* new_in2 = nullptr;
3553 if (in2->is_DecodeNarrowPtr()) {
3554 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3555 new_in2 = in2->in(1);
3556 } else if (in2->Opcode() == Op_ConP) {
3557 const Type* t = in2->bottom_type();
3558 if (t == TypePtr::NULL_PTR) {
3559 assert(in1->is_DecodeN(), "compare klass to null?");
3560 // Don't convert CmpP null check into CmpN if compressed
3561 // oops implicit null check is not generated.
3562 // This will allow to generate normal oop implicit null check.
3563 if (Matcher::gen_narrow_oop_implicit_null_checks())
3564 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3565 //
3566 // This transformation together with CastPP transformation above
3567 // will generated code for implicit null checks for compressed oops.
3568 //
3569 // The original code after Optimize()
3570 //
3571 // LoadN memory, narrow_oop_reg
3572 // decode narrow_oop_reg, base_reg
3573 // CmpP base_reg, nullptr
3574 // CastPP base_reg // NotNull
3575 // Load [base_reg + offset], val_reg
3576 //
3577 // after these transformations will be
3578 //
3579 // LoadN memory, narrow_oop_reg
3580 // CmpN narrow_oop_reg, nullptr
3581 // decode_not_null narrow_oop_reg, base_reg
3582 // Load [base_reg + offset], val_reg
3583 //
3584 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3585 // since narrow oops can be used in debug info now (see the code in
3586 // final_graph_reshaping_walk()).
3587 //
3588 // At the end the code will be matched to
3589 // on x86:
3590 //
3591 // Load_narrow_oop memory, narrow_oop_reg
3592 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3593 // NullCheck narrow_oop_reg
3594 //
3595 // and on sparc:
3596 //
3597 // Load_narrow_oop memory, narrow_oop_reg
3598 // decode_not_null narrow_oop_reg, base_reg
3599 // Load [base_reg + offset], val_reg
3600 // NullCheck base_reg
3601 //
3602 } else if (t->isa_oopptr()) {
3603 new_in2 = ConNode::make(t->make_narrowoop());
3604 } else if (t->isa_klassptr()) {
3605 new_in2 = ConNode::make(t->make_narrowklass());
3606 }
3607 }
3608 if (new_in2 != nullptr) {
3609 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3610 n->subsume_by(cmpN, this);
3611 if (in1->outcnt() == 0) {
3612 in1->disconnect_inputs(this);
3613 }
3614 if (in2->outcnt() == 0) {
3615 in2->disconnect_inputs(this);
3616 }
3617 }
3618 }
3619 break;
3620
3621 case Op_DecodeN:
3622 case Op_DecodeNKlass:
3623 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3624 // DecodeN could be pinned when it can't be fold into
3625 // an address expression, see the code for Op_CastPP above.
3626 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3627 break;
3628
3629 case Op_EncodeP:
3630 case Op_EncodePKlass: {
3631 Node* in1 = n->in(1);
3632 if (in1->is_DecodeNarrowPtr()) {
3633 n->subsume_by(in1->in(1), this);
3634 } else if (in1->Opcode() == Op_ConP) {
3635 const Type* t = in1->bottom_type();
3636 if (t == TypePtr::NULL_PTR) {
3637 assert(t->isa_oopptr(), "null klass?");
3638 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3639 } else if (t->isa_oopptr()) {
3640 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3641 } else if (t->isa_klassptr()) {
3642 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3643 }
3644 }
3645 if (in1->outcnt() == 0) {
3646 in1->disconnect_inputs(this);
3647 }
3648 break;
3649 }
3650
3651 case Op_Proj: {
3652 if (OptimizeStringConcat || IncrementalInline) {
3653 ProjNode* proj = n->as_Proj();
3654 if (proj->_is_io_use) {
3655 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3656 // Separate projections were used for the exception path which
3657 // are normally removed by a late inline. If it wasn't inlined
3658 // then they will hang around and should just be replaced with
3659 // the original one. Merge them.
3660 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3661 if (non_io_proj != nullptr) {
3662 proj->subsume_by(non_io_proj , this);
3663 }
3664 }
3665 }
3666 break;
3667 }
3668
3669 case Op_Phi:
3670 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3671 // The EncodeP optimization may create Phi with the same edges
3672 // for all paths. It is not handled well by Register Allocator.
3673 Node* unique_in = n->in(1);
3674 assert(unique_in != nullptr, "");
3675 uint cnt = n->req();
3676 for (uint i = 2; i < cnt; i++) {
3677 Node* m = n->in(i);
3678 assert(m != nullptr, "");
3679 if (unique_in != m)
3680 unique_in = nullptr;
3681 }
3682 if (unique_in != nullptr) {
3683 n->subsume_by(unique_in, this);
3684 }
3685 }
3686 break;
3687
3688 #endif
3689
3690 case Op_ModI:
3691 handle_div_mod_op(n, T_INT, false);
3692 break;
3693
3694 case Op_ModL:
3695 handle_div_mod_op(n, T_LONG, false);
3696 break;
3697
3698 case Op_UModI:
3699 handle_div_mod_op(n, T_INT, true);
3700 break;
3701
3702 case Op_UModL:
3703 handle_div_mod_op(n, T_LONG, true);
3704 break;
3705
3706 case Op_LoadVector:
3707 case Op_StoreVector:
3708 #ifdef ASSERT
3709 // Add VerifyVectorAlignment node between adr and load / store.
3710 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3711 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3712 n->as_StoreVector()->must_verify_alignment();
3713 if (must_verify_alignment) {
3714 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3715 n->as_StoreVector()->memory_size();
3716 // The memory access should be aligned to the vector width in bytes.
3717 // However, the underlying array is possibly less well aligned, but at least
3718 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3719 // a loop we can expect at least the following alignment:
3720 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3721 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3722 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3723 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3724 jlong mask = guaranteed_alignment - 1;
3725 Node* mask_con = ConLNode::make(mask);
3726 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3727 n->set_req(MemNode::Address, va);
3728 }
3729 }
3730 #endif
3731 break;
3732
3733 case Op_LoadVectorGather:
3734 case Op_StoreVectorScatter:
3735 case Op_LoadVectorGatherMasked:
3736 case Op_StoreVectorScatterMasked:
3737 case Op_VectorCmpMasked:
3738 case Op_VectorMaskGen:
3739 case Op_LoadVectorMasked:
3740 case Op_StoreVectorMasked:
3741 break;
3742
3743 case Op_AddReductionVI:
3744 case Op_AddReductionVL:
3745 case Op_AddReductionVF:
3746 case Op_AddReductionVD:
3747 case Op_MulReductionVI:
3748 case Op_MulReductionVL:
3749 case Op_MulReductionVF:
3750 case Op_MulReductionVD:
3751 case Op_MinReductionV:
3752 case Op_MaxReductionV:
3753 case Op_AndReductionV:
3754 case Op_OrReductionV:
3755 case Op_XorReductionV:
3756 break;
3757
3758 case Op_PackB:
3759 case Op_PackS:
3760 case Op_PackI:
3761 case Op_PackF:
3762 case Op_PackL:
3763 case Op_PackD:
3764 if (n->req()-1 > 2) {
3765 // Replace many operand PackNodes with a binary tree for matching
3766 PackNode* p = (PackNode*) n;
3767 Node* btp = p->binary_tree_pack(1, n->req());
3768 n->subsume_by(btp, this);
3769 }
3770 break;
3771 case Op_Loop:
3772 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3773 case Op_CountedLoop:
3774 case Op_LongCountedLoop:
3775 case Op_OuterStripMinedLoop:
3776 if (n->as_Loop()->is_inner_loop()) {
3777 frc.inc_inner_loop_count();
3778 }
3779 n->as_Loop()->verify_strip_mined(0);
3780 break;
3781 case Op_LShiftI:
3782 case Op_RShiftI:
3783 case Op_URShiftI:
3784 case Op_LShiftL:
3785 case Op_RShiftL:
3786 case Op_URShiftL:
3787 if (Matcher::need_masked_shift_count) {
3788 // The cpu's shift instructions don't restrict the count to the
3789 // lower 5/6 bits. We need to do the masking ourselves.
3790 Node* in2 = n->in(2);
3791 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3792 const TypeInt* t = in2->find_int_type();
3793 if (t != nullptr && t->is_con()) {
3794 juint shift = t->get_con();
3795 if (shift > mask) { // Unsigned cmp
3796 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3797 }
3798 } else {
3799 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3800 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3801 n->set_req(2, shift);
3802 }
3803 }
3804 if (in2->outcnt() == 0) { // Remove dead node
3805 in2->disconnect_inputs(this);
3806 }
3807 }
3808 break;
3809 case Op_MemBarStoreStore:
3810 case Op_MemBarRelease:
3811 // Break the link with AllocateNode: it is no longer useful and
3812 // confuses register allocation.
3813 if (n->req() > MemBarNode::Precedent) {
3814 n->set_req(MemBarNode::Precedent, top());
3815 }
3816 break;
3817 case Op_MemBarAcquire: {
3818 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3819 // At parse time, the trailing MemBarAcquire for a volatile load
3820 // is created with an edge to the load. After optimizations,
3821 // that input may be a chain of Phis. If those phis have no
3822 // other use, then the MemBarAcquire keeps them alive and
3823 // register allocation can be confused.
3824 dead_nodes.push(n->in(MemBarNode::Precedent));
3825 n->set_req(MemBarNode::Precedent, top());
3826 }
3827 break;
3828 }
3829 case Op_Blackhole:
3830 break;
3831 case Op_RangeCheck: {
3832 RangeCheckNode* rc = n->as_RangeCheck();
3833 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3834 n->subsume_by(iff, this);
3835 frc._tests.push(iff);
3836 break;
3837 }
3838 case Op_ConvI2L: {
3839 if (!Matcher::convi2l_type_required) {
3840 // Code generation on some platforms doesn't need accurate
3841 // ConvI2L types. Widening the type can help remove redundant
3842 // address computations.
3843 n->as_Type()->set_type(TypeLong::INT);
3844 ResourceMark rm;
3845 Unique_Node_List wq;
3846 wq.push(n);
3847 for (uint next = 0; next < wq.size(); next++) {
3848 Node *m = wq.at(next);
3849
3850 for(;;) {
3851 // Loop over all nodes with identical inputs edges as m
3852 Node* k = m->find_similar(m->Opcode());
3853 if (k == nullptr) {
3854 break;
3855 }
3856 // Push their uses so we get a chance to remove node made
3857 // redundant
3858 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3859 Node* u = k->fast_out(i);
3860 if (u->Opcode() == Op_LShiftL ||
3861 u->Opcode() == Op_AddL ||
3862 u->Opcode() == Op_SubL ||
3863 u->Opcode() == Op_AddP) {
3864 wq.push(u);
3865 }
3866 }
3867 // Replace all nodes with identical edges as m with m
3868 k->subsume_by(m, this);
3869 }
3870 }
3871 }
3872 break;
3873 }
3874 case Op_CmpUL: {
3875 if (!Matcher::has_match_rule(Op_CmpUL)) {
3876 // No support for unsigned long comparisons
3877 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3878 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3879 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3880 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3881 Node* andl = new AndLNode(orl, remove_sign_mask);
3882 Node* cmp = new CmpLNode(andl, n->in(2));
3883 n->subsume_by(cmp, this);
3884 }
3885 break;
3886 }
3887 #ifdef ASSERT
3888 case Op_ConNKlass: {
3889 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3890 ciKlass* klass = tp->is_klassptr()->exact_klass();
3891 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3892 break;
3893 }
3894 #endif
3895 default:
3896 assert(!n->is_Call(), "");
3897 assert(!n->is_Mem(), "");
3898 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3899 break;
3900 }
3901 }
3902
3903 //------------------------------final_graph_reshaping_walk---------------------
3904 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3905 // requires that the walk visits a node's inputs before visiting the node.
3906 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3907 Unique_Node_List sfpt;
3908
3909 frc._visited.set(root->_idx); // first, mark node as visited
3910 uint cnt = root->req();
3911 Node *n = root;
3912 uint i = 0;
3913 while (true) {
3914 if (i < cnt) {
3915 // Place all non-visited non-null inputs onto stack
3916 Node* m = n->in(i);
3917 ++i;
3918 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3919 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3920 // compute worst case interpreter size in case of a deoptimization
3921 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3922
3923 sfpt.push(m);
3924 }
3925 cnt = m->req();
3926 nstack.push(n, i); // put on stack parent and next input's index
3927 n = m;
3928 i = 0;
3929 }
3930 } else {
3931 // Now do post-visit work
3932 final_graph_reshaping_impl(n, frc, dead_nodes);
3933 if (nstack.is_empty())
3934 break; // finished
3935 n = nstack.node(); // Get node from stack
3936 cnt = n->req();
3937 i = nstack.index();
3938 nstack.pop(); // Shift to the next node on stack
3939 }
3940 }
3941
3942 // Skip next transformation if compressed oops are not used.
3943 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3944 (!UseCompressedOops && !UseCompressedClassPointers))
3945 return;
3946
3947 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3948 // It could be done for an uncommon traps or any safepoints/calls
3949 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3950 while (sfpt.size() > 0) {
3951 n = sfpt.pop();
3952 JVMState *jvms = n->as_SafePoint()->jvms();
3953 assert(jvms != nullptr, "sanity");
3954 int start = jvms->debug_start();
3955 int end = n->req();
3956 bool is_uncommon = (n->is_CallStaticJava() &&
3957 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3958 for (int j = start; j < end; j++) {
3959 Node* in = n->in(j);
3960 if (in->is_DecodeNarrowPtr()) {
3961 bool safe_to_skip = true;
3962 if (!is_uncommon ) {
3963 // Is it safe to skip?
3964 for (uint i = 0; i < in->outcnt(); i++) {
3965 Node* u = in->raw_out(i);
3966 if (!u->is_SafePoint() ||
3967 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3968 safe_to_skip = false;
3969 }
3970 }
3971 }
3972 if (safe_to_skip) {
3973 n->set_req(j, in->in(1));
3974 }
3975 if (in->outcnt() == 0) {
3976 in->disconnect_inputs(this);
3977 }
3978 }
3979 }
3980 }
3981 }
3982
3983 //------------------------------final_graph_reshaping--------------------------
3984 // Final Graph Reshaping.
3985 //
3986 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3987 // and not commoned up and forced early. Must come after regular
3988 // optimizations to avoid GVN undoing the cloning. Clone constant
3989 // inputs to Loop Phis; these will be split by the allocator anyways.
3990 // Remove Opaque nodes.
3991 // (2) Move last-uses by commutative operations to the left input to encourage
3992 // Intel update-in-place two-address operations and better register usage
3993 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3994 // calls canonicalizing them back.
3995 // (3) Count the number of double-precision FP ops, single-precision FP ops
3996 // and call sites. On Intel, we can get correct rounding either by
3997 // forcing singles to memory (requires extra stores and loads after each
3998 // FP bytecode) or we can set a rounding mode bit (requires setting and
3999 // clearing the mode bit around call sites). The mode bit is only used
4000 // if the relative frequency of single FP ops to calls is low enough.
4001 // This is a key transform for SPEC mpeg_audio.
4002 // (4) Detect infinite loops; blobs of code reachable from above but not
4003 // below. Several of the Code_Gen algorithms fail on such code shapes,
4004 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4005 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4006 // Detection is by looking for IfNodes where only 1 projection is
4007 // reachable from below or CatchNodes missing some targets.
4008 // (5) Assert for insane oop offsets in debug mode.
4009
4010 bool Compile::final_graph_reshaping() {
4011 // an infinite loop may have been eliminated by the optimizer,
4012 // in which case the graph will be empty.
4013 if (root()->req() == 1) {
4014 // Do not compile method that is only a trivial infinite loop,
4015 // since the content of the loop may have been eliminated.
4016 record_method_not_compilable("trivial infinite loop");
4017 return true;
4018 }
4019
4020 // Expensive nodes have their control input set to prevent the GVN
4021 // from freely commoning them. There's no GVN beyond this point so
4022 // no need to keep the control input. We want the expensive nodes to
4023 // be freely moved to the least frequent code path by gcm.
4024 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4025 for (int i = 0; i < expensive_count(); i++) {
4026 _expensive_nodes.at(i)->set_req(0, nullptr);
4027 }
4028
4029 Final_Reshape_Counts frc;
4030
4031 // Visit everybody reachable!
4032 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4033 Node_Stack nstack(live_nodes() >> 1);
4034 Unique_Node_List dead_nodes;
4035 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4036
4037 // Check for unreachable (from below) code (i.e., infinite loops).
4038 for( uint i = 0; i < frc._tests.size(); i++ ) {
4039 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4040 // Get number of CFG targets.
4041 // Note that PCTables include exception targets after calls.
4042 uint required_outcnt = n->required_outcnt();
4043 if (n->outcnt() != required_outcnt) {
4044 // Check for a few special cases. Rethrow Nodes never take the
4045 // 'fall-thru' path, so expected kids is 1 less.
4046 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4047 if (n->in(0)->in(0)->is_Call()) {
4048 CallNode* call = n->in(0)->in(0)->as_Call();
4049 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4050 required_outcnt--; // Rethrow always has 1 less kid
4051 } else if (call->req() > TypeFunc::Parms &&
4052 call->is_CallDynamicJava()) {
4053 // Check for null receiver. In such case, the optimizer has
4054 // detected that the virtual call will always result in a null
4055 // pointer exception. The fall-through projection of this CatchNode
4056 // will not be populated.
4057 Node* arg0 = call->in(TypeFunc::Parms);
4058 if (arg0->is_Type() &&
4059 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4060 required_outcnt--;
4061 }
4062 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4063 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4064 // Check for illegal array length. In such case, the optimizer has
4065 // detected that the allocation attempt will always result in an
4066 // exception. There is no fall-through projection of this CatchNode .
4067 assert(call->is_CallStaticJava(), "static call expected");
4068 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4069 uint valid_length_test_input = call->req() - 1;
4070 Node* valid_length_test = call->in(valid_length_test_input);
4071 call->del_req(valid_length_test_input);
4072 if (valid_length_test->find_int_con(1) == 0) {
4073 required_outcnt--;
4074 }
4075 dead_nodes.push(valid_length_test);
4076 assert(n->outcnt() == required_outcnt, "malformed control flow");
4077 continue;
4078 }
4079 }
4080 }
4081
4082 // Recheck with a better notion of 'required_outcnt'
4083 if (n->outcnt() != required_outcnt) {
4084 record_method_not_compilable("malformed control flow");
4085 return true; // Not all targets reachable!
4086 }
4087 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4088 CallNode* call = n->in(0)->in(0)->as_Call();
4089 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4090 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4091 assert(call->is_CallStaticJava(), "static call expected");
4092 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4093 uint valid_length_test_input = call->req() - 1;
4094 dead_nodes.push(call->in(valid_length_test_input));
4095 call->del_req(valid_length_test_input); // valid length test useless now
4096 }
4097 }
4098 // Check that I actually visited all kids. Unreached kids
4099 // must be infinite loops.
4100 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4101 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4102 record_method_not_compilable("infinite loop");
4103 return true; // Found unvisited kid; must be unreach
4104 }
4105
4106 // Here so verification code in final_graph_reshaping_walk()
4107 // always see an OuterStripMinedLoopEnd
4108 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4109 IfNode* init_iff = n->as_If();
4110 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4111 n->subsume_by(iff, this);
4112 }
4113 }
4114
4115 while (dead_nodes.size() > 0) {
4116 Node* m = dead_nodes.pop();
4117 if (m->outcnt() == 0 && m != top()) {
4118 for (uint j = 0; j < m->req(); j++) {
4119 Node* in = m->in(j);
4120 if (in != nullptr) {
4121 dead_nodes.push(in);
4122 }
4123 }
4124 m->disconnect_inputs(this);
4125 }
4126 }
4127
4128 set_java_calls(frc.get_java_call_count());
4129 set_inner_loops(frc.get_inner_loop_count());
4130
4131 // No infinite loops, no reason to bail out.
4132 return false;
4133 }
4134
4135 //-----------------------------too_many_traps----------------------------------
4136 // Report if there are too many traps at the current method and bci.
4137 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4138 bool Compile::too_many_traps(ciMethod* method,
4139 int bci,
4140 Deoptimization::DeoptReason reason) {
4141 if (method->has_trap_at(bci)) {
4142 return true;
4143 }
4144 if (PreloadReduceTraps && for_preload()) {
4145 // Preload code should not have traps, if possible.
4146 return true;
4147 }
4148 ciMethodData* md = method->method_data();
4149 if (md->is_empty()) {
4150 // Assume the trap has not occurred, or that it occurred only
4151 // because of a transient condition during start-up in the interpreter.
4152 return false;
4153 }
4154 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4155 if (md->has_trap_at(bci, m, reason) != 0) {
4156 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4157 // Also, if there are multiple reasons, or if there is no per-BCI record,
4158 // assume the worst.
4159 if (log())
4160 log()->elem("observe trap='%s' count='%d'",
4161 Deoptimization::trap_reason_name(reason),
4162 md->trap_count(reason));
4163 return true;
4164 } else {
4165 // Ignore method/bci and see if there have been too many globally.
4166 return too_many_traps(reason, md);
4167 }
4168 }
4169
4170 // Less-accurate variant which does not require a method and bci.
4171 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4172 ciMethodData* logmd) {
4173 if (PreloadReduceTraps && for_preload()) {
4174 // Preload code should not have traps, if possible.
4175 return true;
4176 }
4177 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4178 // Too many traps globally.
4179 // Note that we use cumulative trap_count, not just md->trap_count.
4180 if (log()) {
4181 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4182 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4183 Deoptimization::trap_reason_name(reason),
4184 mcount, trap_count(reason));
4185 }
4186 return true;
4187 } else {
4188 // The coast is clear.
4189 return false;
4190 }
4191 }
4192
4193 //--------------------------too_many_recompiles--------------------------------
4194 // Report if there are too many recompiles at the current method and bci.
4195 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4196 // Is not eager to return true, since this will cause the compiler to use
4197 // Action_none for a trap point, to avoid too many recompilations.
4198 bool Compile::too_many_recompiles(ciMethod* method,
4199 int bci,
4200 Deoptimization::DeoptReason reason) {
4201 ciMethodData* md = method->method_data();
4202 if (md->is_empty()) {
4203 // Assume the trap has not occurred, or that it occurred only
4204 // because of a transient condition during start-up in the interpreter.
4205 return false;
4206 }
4207 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4208 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4209 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4210 Deoptimization::DeoptReason per_bc_reason
4211 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4212 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4213 if ((per_bc_reason == Deoptimization::Reason_none
4214 || md->has_trap_at(bci, m, reason) != 0)
4215 // The trap frequency measure we care about is the recompile count:
4216 && md->trap_recompiled_at(bci, m)
4217 && md->overflow_recompile_count() >= bc_cutoff) {
4218 // Do not emit a trap here if it has already caused recompilations.
4219 // Also, if there are multiple reasons, or if there is no per-BCI record,
4220 // assume the worst.
4221 if (log())
4222 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4223 Deoptimization::trap_reason_name(reason),
4224 md->trap_count(reason),
4225 md->overflow_recompile_count());
4226 return true;
4227 } else if (trap_count(reason) != 0
4228 && decompile_count() >= m_cutoff) {
4229 // Too many recompiles globally, and we have seen this sort of trap.
4230 // Use cumulative decompile_count, not just md->decompile_count.
4231 if (log())
4232 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4233 Deoptimization::trap_reason_name(reason),
4234 md->trap_count(reason), trap_count(reason),
4235 md->decompile_count(), decompile_count());
4236 return true;
4237 } else {
4238 // The coast is clear.
4239 return false;
4240 }
4241 }
4242
4243 // Compute when not to trap. Used by matching trap based nodes and
4244 // NullCheck optimization.
4245 void Compile::set_allowed_deopt_reasons() {
4246 _allowed_reasons = 0;
4247 if (is_method_compilation()) {
4248 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4249 assert(rs < BitsPerInt, "recode bit map");
4250 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4251 _allowed_reasons |= nth_bit(rs);
4252 }
4253 }
4254 }
4255 }
4256
4257 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4258 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4259 }
4260
4261 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4262 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4263 }
4264
4265 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4266 if (holder->is_initialized() && !do_clinit_barriers()) {
4267 return false;
4268 }
4269 if (holder->is_being_initialized() || do_clinit_barriers()) {
4270 if (accessing_method->holder() == holder) {
4271 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4272 // <init>, or a static method. In all those cases, there was an initialization
4273 // barrier on the holder klass passed.
4274 if (accessing_method->is_static_initializer() ||
4275 accessing_method->is_object_initializer() ||
4276 accessing_method->is_static()) {
4277 return false;
4278 }
4279 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4280 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4281 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4282 // child class can become fully initialized while its parent class is still being initialized.
4283 if (accessing_method->is_static_initializer()) {
4284 return false;
4285 }
4286 }
4287 ciMethod* root = method(); // the root method of compilation
4288 if (root != accessing_method) {
4289 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4290 }
4291 }
4292 return true;
4293 }
4294
4295 #ifndef PRODUCT
4296 //------------------------------verify_bidirectional_edges---------------------
4297 // For each input edge to a node (ie - for each Use-Def edge), verify that
4298 // there is a corresponding Def-Use edge.
4299 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4300 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4301 uint stack_size = live_nodes() >> 4;
4302 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4303 if (root_and_safepoints != nullptr) {
4304 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4305 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4306 Node* root_or_safepoint = root_and_safepoints->at(i);
4307 // If the node is a safepoint, let's check if it still has a control input
4308 // Lack of control input signifies that this node was killed by CCP or
4309 // recursively by remove_globally_dead_node and it shouldn't be a starting
4310 // point.
4311 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4312 nstack.push(root_or_safepoint);
4313 }
4314 }
4315 } else {
4316 nstack.push(_root);
4317 }
4318
4319 while (nstack.size() > 0) {
4320 Node* n = nstack.pop();
4321 if (visited.member(n)) {
4322 continue;
4323 }
4324 visited.push(n);
4325
4326 // Walk over all input edges, checking for correspondence
4327 uint length = n->len();
4328 for (uint i = 0; i < length; i++) {
4329 Node* in = n->in(i);
4330 if (in != nullptr && !visited.member(in)) {
4331 nstack.push(in); // Put it on stack
4332 }
4333 if (in != nullptr && !in->is_top()) {
4334 // Count instances of `next`
4335 int cnt = 0;
4336 for (uint idx = 0; idx < in->_outcnt; idx++) {
4337 if (in->_out[idx] == n) {
4338 cnt++;
4339 }
4340 }
4341 assert(cnt > 0, "Failed to find Def-Use edge.");
4342 // Check for duplicate edges
4343 // walk the input array downcounting the input edges to n
4344 for (uint j = 0; j < length; j++) {
4345 if (n->in(j) == in) {
4346 cnt--;
4347 }
4348 }
4349 assert(cnt == 0, "Mismatched edge count.");
4350 } else if (in == nullptr) {
4351 assert(i == 0 || i >= n->req() ||
4352 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4353 (n->is_Unlock() && i == (n->req() - 1)) ||
4354 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4355 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4356 } else {
4357 assert(in->is_top(), "sanity");
4358 // Nothing to check.
4359 }
4360 }
4361 }
4362 }
4363
4364 //------------------------------verify_graph_edges---------------------------
4365 // Walk the Graph and verify that there is a one-to-one correspondence
4366 // between Use-Def edges and Def-Use edges in the graph.
4367 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4368 if (VerifyGraphEdges) {
4369 Unique_Node_List visited;
4370
4371 // Call graph walk to check edges
4372 verify_bidirectional_edges(visited, root_and_safepoints);
4373 if (no_dead_code) {
4374 // Now make sure that no visited node is used by an unvisited node.
4375 bool dead_nodes = false;
4376 Unique_Node_List checked;
4377 while (visited.size() > 0) {
4378 Node* n = visited.pop();
4379 checked.push(n);
4380 for (uint i = 0; i < n->outcnt(); i++) {
4381 Node* use = n->raw_out(i);
4382 if (checked.member(use)) continue; // already checked
4383 if (visited.member(use)) continue; // already in the graph
4384 if (use->is_Con()) continue; // a dead ConNode is OK
4385 // At this point, we have found a dead node which is DU-reachable.
4386 if (!dead_nodes) {
4387 tty->print_cr("*** Dead nodes reachable via DU edges:");
4388 dead_nodes = true;
4389 }
4390 use->dump(2);
4391 tty->print_cr("---");
4392 checked.push(use); // No repeats; pretend it is now checked.
4393 }
4394 }
4395 assert(!dead_nodes, "using nodes must be reachable from root");
4396 }
4397 }
4398 }
4399 #endif
4400
4401 // The Compile object keeps track of failure reasons separately from the ciEnv.
4402 // This is required because there is not quite a 1-1 relation between the
4403 // ciEnv and its compilation task and the Compile object. Note that one
4404 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4405 // to backtrack and retry without subsuming loads. Other than this backtracking
4406 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4407 // by the logic in C2Compiler.
4408 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4409 if (log() != nullptr) {
4410 log()->elem("failure reason='%s' phase='compile'", reason);
4411 }
4412 if (_failure_reason.get() == nullptr) {
4413 // Record the first failure reason.
4414 _failure_reason.set(reason);
4415 if (CaptureBailoutInformation) {
4416 _first_failure_details = new CompilationFailureInfo(reason);
4417 }
4418 } else {
4419 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4420 }
4421
4422 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4423 C->print_method(PHASE_FAILURE, 1);
4424 }
4425 _root = nullptr; // flush the graph, too
4426 }
4427
4428 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4429 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4430 _compile(Compile::current()),
4431 _log(nullptr),
4432 _dolog(CITimeVerbose)
4433 {
4434 assert(_compile != nullptr, "sanity check");
4435 assert(id != PhaseTraceId::_t_none, "Don't use none");
4436 if (_dolog) {
4437 _log = _compile->log();
4438 }
4439 if (_log != nullptr) {
4440 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4441 _log->stamp();
4442 _log->end_head();
4443 }
4444
4445 // Inform memory statistic, if enabled
4446 if (CompilationMemoryStatistic::enabled()) {
4447 CompilationMemoryStatistic::on_phase_start((int)id, name);
4448 }
4449 }
4450
4451 Compile::TracePhase::TracePhase(PhaseTraceId id)
4452 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4453
4454 Compile::TracePhase::~TracePhase() {
4455
4456 // Inform memory statistic, if enabled
4457 if (CompilationMemoryStatistic::enabled()) {
4458 CompilationMemoryStatistic::on_phase_end();
4459 }
4460
4461 if (_compile->failing_internal()) {
4462 if (_log != nullptr) {
4463 _log->done("phase");
4464 }
4465 return; // timing code, not stressing bailouts.
4466 }
4467 #ifdef ASSERT
4468 if (PrintIdealNodeCount) {
4469 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4470 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4471 }
4472
4473 if (VerifyIdealNodeCount) {
4474 _compile->print_missing_nodes();
4475 }
4476 #endif
4477
4478 if (_log != nullptr) {
4479 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4480 }
4481 }
4482
4483 //----------------------------static_subtype_check-----------------------------
4484 // Shortcut important common cases when superklass is exact:
4485 // (0) superklass is java.lang.Object (can occur in reflective code)
4486 // (1) subklass is already limited to a subtype of superklass => always ok
4487 // (2) subklass does not overlap with superklass => always fail
4488 // (3) superklass has NO subtypes and we can check with a simple compare.
4489 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4490 if (skip) {
4491 return SSC_full_test; // Let caller generate the general case.
4492 }
4493
4494 if (subk->is_java_subtype_of(superk)) {
4495 return SSC_always_true; // (0) and (1) this test cannot fail
4496 }
4497
4498 if (!subk->maybe_java_subtype_of(superk)) {
4499 return SSC_always_false; // (2) true path dead; no dynamic test needed
4500 }
4501
4502 const Type* superelem = superk;
4503 if (superk->isa_aryklassptr()) {
4504 int ignored;
4505 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4506 }
4507
4508 if (superelem->isa_instklassptr()) {
4509 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4510 if (!ik->has_subklass()) {
4511 if (!ik->is_final()) {
4512 // Add a dependency if there is a chance of a later subclass.
4513 dependencies()->assert_leaf_type(ik);
4514 }
4515 if (!superk->maybe_java_subtype_of(subk)) {
4516 return SSC_always_false;
4517 }
4518 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4519 }
4520 } else {
4521 // A primitive array type has no subtypes.
4522 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4523 }
4524
4525 return SSC_full_test;
4526 }
4527
4528 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4529 #ifdef _LP64
4530 // The scaled index operand to AddP must be a clean 64-bit value.
4531 // Java allows a 32-bit int to be incremented to a negative
4532 // value, which appears in a 64-bit register as a large
4533 // positive number. Using that large positive number as an
4534 // operand in pointer arithmetic has bad consequences.
4535 // On the other hand, 32-bit overflow is rare, and the possibility
4536 // can often be excluded, if we annotate the ConvI2L node with
4537 // a type assertion that its value is known to be a small positive
4538 // number. (The prior range check has ensured this.)
4539 // This assertion is used by ConvI2LNode::Ideal.
4540 int index_max = max_jint - 1; // array size is max_jint, index is one less
4541 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4542 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4543 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4544 #endif
4545 return idx;
4546 }
4547
4548 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4549 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4550 if (ctrl != nullptr) {
4551 // Express control dependency by a CastII node with a narrow type.
4552 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4553 // node from floating above the range check during loop optimizations. Otherwise, the
4554 // ConvI2L node may be eliminated independently of the range check, causing the data path
4555 // to become TOP while the control path is still there (although it's unreachable).
4556 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4557 value = phase->transform(value);
4558 }
4559 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4560 return phase->transform(new ConvI2LNode(value, ltype));
4561 }
4562
4563 void Compile::dump_print_inlining() {
4564 inline_printer()->print_on(tty);
4565 }
4566
4567 void Compile::log_late_inline(CallGenerator* cg) {
4568 if (log() != nullptr) {
4569 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4570 cg->unique_id());
4571 JVMState* p = cg->call_node()->jvms();
4572 while (p != nullptr) {
4573 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4574 p = p->caller();
4575 }
4576 log()->tail("late_inline");
4577 }
4578 }
4579
4580 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4581 log_late_inline(cg);
4582 if (log() != nullptr) {
4583 log()->inline_fail(msg);
4584 }
4585 }
4586
4587 void Compile::log_inline_id(CallGenerator* cg) {
4588 if (log() != nullptr) {
4589 // The LogCompilation tool needs a unique way to identify late
4590 // inline call sites. This id must be unique for this call site in
4591 // this compilation. Try to have it unique across compilations as
4592 // well because it can be convenient when grepping through the log
4593 // file.
4594 // Distinguish OSR compilations from others in case CICountOSR is
4595 // on.
4596 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4597 cg->set_unique_id(id);
4598 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4599 }
4600 }
4601
4602 void Compile::log_inline_failure(const char* msg) {
4603 if (C->log() != nullptr) {
4604 C->log()->inline_fail(msg);
4605 }
4606 }
4607
4608
4609 // Dump inlining replay data to the stream.
4610 // Don't change thread state and acquire any locks.
4611 void Compile::dump_inline_data(outputStream* out) {
4612 InlineTree* inl_tree = ilt();
4613 if (inl_tree != nullptr) {
4614 out->print(" inline %d", inl_tree->count());
4615 inl_tree->dump_replay_data(out);
4616 }
4617 }
4618
4619 void Compile::dump_inline_data_reduced(outputStream* out) {
4620 assert(ReplayReduce, "");
4621
4622 InlineTree* inl_tree = ilt();
4623 if (inl_tree == nullptr) {
4624 return;
4625 }
4626 // Enable iterative replay file reduction
4627 // Output "compile" lines for depth 1 subtrees,
4628 // simulating that those trees were compiled
4629 // instead of inlined.
4630 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4631 InlineTree* sub = inl_tree->subtrees().at(i);
4632 if (sub->inline_level() != 1) {
4633 continue;
4634 }
4635
4636 ciMethod* method = sub->method();
4637 int entry_bci = -1;
4638 int comp_level = env()->task()->comp_level();
4639 out->print("compile ");
4640 method->dump_name_as_ascii(out);
4641 out->print(" %d %d", entry_bci, comp_level);
4642 out->print(" inline %d", sub->count());
4643 sub->dump_replay_data(out, -1);
4644 out->cr();
4645 }
4646 }
4647
4648 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4649 if (n1->Opcode() < n2->Opcode()) return -1;
4650 else if (n1->Opcode() > n2->Opcode()) return 1;
4651
4652 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4653 for (uint i = 1; i < n1->req(); i++) {
4654 if (n1->in(i) < n2->in(i)) return -1;
4655 else if (n1->in(i) > n2->in(i)) return 1;
4656 }
4657
4658 return 0;
4659 }
4660
4661 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4662 Node* n1 = *n1p;
4663 Node* n2 = *n2p;
4664
4665 return cmp_expensive_nodes(n1, n2);
4666 }
4667
4668 void Compile::sort_expensive_nodes() {
4669 if (!expensive_nodes_sorted()) {
4670 _expensive_nodes.sort(cmp_expensive_nodes);
4671 }
4672 }
4673
4674 bool Compile::expensive_nodes_sorted() const {
4675 for (int i = 1; i < _expensive_nodes.length(); i++) {
4676 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4677 return false;
4678 }
4679 }
4680 return true;
4681 }
4682
4683 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4684 if (_expensive_nodes.length() == 0) {
4685 return false;
4686 }
4687
4688 assert(OptimizeExpensiveOps, "optimization off?");
4689
4690 // Take this opportunity to remove dead nodes from the list
4691 int j = 0;
4692 for (int i = 0; i < _expensive_nodes.length(); i++) {
4693 Node* n = _expensive_nodes.at(i);
4694 if (!n->is_unreachable(igvn)) {
4695 assert(n->is_expensive(), "should be expensive");
4696 _expensive_nodes.at_put(j, n);
4697 j++;
4698 }
4699 }
4700 _expensive_nodes.trunc_to(j);
4701
4702 // Then sort the list so that similar nodes are next to each other
4703 // and check for at least two nodes of identical kind with same data
4704 // inputs.
4705 sort_expensive_nodes();
4706
4707 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4708 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4709 return true;
4710 }
4711 }
4712
4713 return false;
4714 }
4715
4716 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4717 if (_expensive_nodes.length() == 0) {
4718 return;
4719 }
4720
4721 assert(OptimizeExpensiveOps, "optimization off?");
4722
4723 // Sort to bring similar nodes next to each other and clear the
4724 // control input of nodes for which there's only a single copy.
4725 sort_expensive_nodes();
4726
4727 int j = 0;
4728 int identical = 0;
4729 int i = 0;
4730 bool modified = false;
4731 for (; i < _expensive_nodes.length()-1; i++) {
4732 assert(j <= i, "can't write beyond current index");
4733 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4734 identical++;
4735 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4736 continue;
4737 }
4738 if (identical > 0) {
4739 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4740 identical = 0;
4741 } else {
4742 Node* n = _expensive_nodes.at(i);
4743 igvn.replace_input_of(n, 0, nullptr);
4744 igvn.hash_insert(n);
4745 modified = true;
4746 }
4747 }
4748 if (identical > 0) {
4749 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4750 } else if (_expensive_nodes.length() >= 1) {
4751 Node* n = _expensive_nodes.at(i);
4752 igvn.replace_input_of(n, 0, nullptr);
4753 igvn.hash_insert(n);
4754 modified = true;
4755 }
4756 _expensive_nodes.trunc_to(j);
4757 if (modified) {
4758 igvn.optimize();
4759 }
4760 }
4761
4762 void Compile::add_expensive_node(Node * n) {
4763 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4764 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4765 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4766 if (OptimizeExpensiveOps) {
4767 _expensive_nodes.append(n);
4768 } else {
4769 // Clear control input and let IGVN optimize expensive nodes if
4770 // OptimizeExpensiveOps is off.
4771 n->set_req(0, nullptr);
4772 }
4773 }
4774
4775 /**
4776 * Track coarsened Lock and Unlock nodes.
4777 */
4778
4779 class Lock_List : public Node_List {
4780 uint _origin_cnt;
4781 public:
4782 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4783 uint origin_cnt() const { return _origin_cnt; }
4784 };
4785
4786 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4787 int length = locks.length();
4788 if (length > 0) {
4789 // Have to keep this list until locks elimination during Macro nodes elimination.
4790 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4791 AbstractLockNode* alock = locks.at(0);
4792 BoxLockNode* box = alock->box_node()->as_BoxLock();
4793 for (int i = 0; i < length; i++) {
4794 AbstractLockNode* lock = locks.at(i);
4795 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4796 locks_list->push(lock);
4797 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4798 if (this_box != box) {
4799 // Locking regions (BoxLock) could be Unbalanced here:
4800 // - its coarsened locks were eliminated in earlier
4801 // macro nodes elimination followed by loop unroll
4802 // - it is OSR locking region (no Lock node)
4803 // Preserve Unbalanced status in such cases.
4804 if (!this_box->is_unbalanced()) {
4805 this_box->set_coarsened();
4806 }
4807 if (!box->is_unbalanced()) {
4808 box->set_coarsened();
4809 }
4810 }
4811 }
4812 _coarsened_locks.append(locks_list);
4813 }
4814 }
4815
4816 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4817 int count = coarsened_count();
4818 for (int i = 0; i < count; i++) {
4819 Node_List* locks_list = _coarsened_locks.at(i);
4820 for (uint j = 0; j < locks_list->size(); j++) {
4821 Node* lock = locks_list->at(j);
4822 assert(lock->is_AbstractLock(), "sanity");
4823 if (!useful.member(lock)) {
4824 locks_list->yank(lock);
4825 }
4826 }
4827 }
4828 }
4829
4830 void Compile::remove_coarsened_lock(Node* n) {
4831 if (n->is_AbstractLock()) {
4832 int count = coarsened_count();
4833 for (int i = 0; i < count; i++) {
4834 Node_List* locks_list = _coarsened_locks.at(i);
4835 locks_list->yank(n);
4836 }
4837 }
4838 }
4839
4840 bool Compile::coarsened_locks_consistent() {
4841 int count = coarsened_count();
4842 for (int i = 0; i < count; i++) {
4843 bool unbalanced = false;
4844 bool modified = false; // track locks kind modifications
4845 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4846 uint size = locks_list->size();
4847 if (size == 0) {
4848 unbalanced = false; // All locks were eliminated - good
4849 } else if (size != locks_list->origin_cnt()) {
4850 unbalanced = true; // Some locks were removed from list
4851 } else {
4852 for (uint j = 0; j < size; j++) {
4853 Node* lock = locks_list->at(j);
4854 // All nodes in group should have the same state (modified or not)
4855 if (!lock->as_AbstractLock()->is_coarsened()) {
4856 if (j == 0) {
4857 // first on list was modified, the rest should be too for consistency
4858 modified = true;
4859 } else if (!modified) {
4860 // this lock was modified but previous locks on the list were not
4861 unbalanced = true;
4862 break;
4863 }
4864 } else if (modified) {
4865 // previous locks on list were modified but not this lock
4866 unbalanced = true;
4867 break;
4868 }
4869 }
4870 }
4871 if (unbalanced) {
4872 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4873 #ifdef ASSERT
4874 if (PrintEliminateLocks) {
4875 tty->print_cr("=== unbalanced coarsened locks ===");
4876 for (uint l = 0; l < size; l++) {
4877 locks_list->at(l)->dump();
4878 }
4879 }
4880 #endif
4881 record_failure(C2Compiler::retry_no_locks_coarsening());
4882 return false;
4883 }
4884 }
4885 return true;
4886 }
4887
4888 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4889 // locks coarsening optimization removed Lock/Unlock nodes from them.
4890 // Such regions become unbalanced because coarsening only removes part
4891 // of Lock/Unlock nodes in region. As result we can't execute other
4892 // locks elimination optimizations which assume all code paths have
4893 // corresponding pair of Lock/Unlock nodes - they are balanced.
4894 void Compile::mark_unbalanced_boxes() const {
4895 int count = coarsened_count();
4896 for (int i = 0; i < count; i++) {
4897 Node_List* locks_list = _coarsened_locks.at(i);
4898 uint size = locks_list->size();
4899 if (size > 0) {
4900 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4901 BoxLockNode* box = alock->box_node()->as_BoxLock();
4902 if (alock->is_coarsened()) {
4903 // coarsened_locks_consistent(), which is called before this method, verifies
4904 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4905 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4906 for (uint j = 1; j < size; j++) {
4907 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4908 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4909 if (box != this_box) {
4910 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4911 box->set_unbalanced();
4912 this_box->set_unbalanced();
4913 }
4914 }
4915 }
4916 }
4917 }
4918 }
4919
4920 /**
4921 * Remove the speculative part of types and clean up the graph
4922 */
4923 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4924 if (UseTypeSpeculation) {
4925 Unique_Node_List worklist;
4926 worklist.push(root());
4927 int modified = 0;
4928 // Go over all type nodes that carry a speculative type, drop the
4929 // speculative part of the type and enqueue the node for an igvn
4930 // which may optimize it out.
4931 for (uint next = 0; next < worklist.size(); ++next) {
4932 Node *n = worklist.at(next);
4933 if (n->is_Type()) {
4934 TypeNode* tn = n->as_Type();
4935 const Type* t = tn->type();
4936 const Type* t_no_spec = t->remove_speculative();
4937 if (t_no_spec != t) {
4938 bool in_hash = igvn.hash_delete(n);
4939 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4940 tn->set_type(t_no_spec);
4941 igvn.hash_insert(n);
4942 igvn._worklist.push(n); // give it a chance to go away
4943 modified++;
4944 }
4945 }
4946 // Iterate over outs - endless loops is unreachable from below
4947 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4948 Node *m = n->fast_out(i);
4949 if (not_a_node(m)) {
4950 continue;
4951 }
4952 worklist.push(m);
4953 }
4954 }
4955 // Drop the speculative part of all types in the igvn's type table
4956 igvn.remove_speculative_types();
4957 if (modified > 0) {
4958 igvn.optimize();
4959 if (failing()) return;
4960 }
4961 #ifdef ASSERT
4962 // Verify that after the IGVN is over no speculative type has resurfaced
4963 worklist.clear();
4964 worklist.push(root());
4965 for (uint next = 0; next < worklist.size(); ++next) {
4966 Node *n = worklist.at(next);
4967 const Type* t = igvn.type_or_null(n);
4968 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4969 if (n->is_Type()) {
4970 t = n->as_Type()->type();
4971 assert(t == t->remove_speculative(), "no more speculative types");
4972 }
4973 // Iterate over outs - endless loops is unreachable from below
4974 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4975 Node *m = n->fast_out(i);
4976 if (not_a_node(m)) {
4977 continue;
4978 }
4979 worklist.push(m);
4980 }
4981 }
4982 igvn.check_no_speculative_types();
4983 #endif
4984 }
4985 }
4986
4987 // Auxiliary methods to support randomized stressing/fuzzing.
4988
4989 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
4990 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
4991 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
4992 FLAG_SET_ERGO(StressSeed, _stress_seed);
4993 } else {
4994 _stress_seed = StressSeed;
4995 }
4996 if (_log != nullptr) {
4997 _log->elem("stress_test seed='%u'", _stress_seed);
4998 }
4999 }
5000
5001 int Compile::random() {
5002 _stress_seed = os::next_random(_stress_seed);
5003 return static_cast<int>(_stress_seed);
5004 }
5005
5006 // This method can be called the arbitrary number of times, with current count
5007 // as the argument. The logic allows selecting a single candidate from the
5008 // running list of candidates as follows:
5009 // int count = 0;
5010 // Cand* selected = null;
5011 // while(cand = cand->next()) {
5012 // if (randomized_select(++count)) {
5013 // selected = cand;
5014 // }
5015 // }
5016 //
5017 // Including count equalizes the chances any candidate is "selected".
5018 // This is useful when we don't have the complete list of candidates to choose
5019 // from uniformly. In this case, we need to adjust the randomicity of the
5020 // selection, or else we will end up biasing the selection towards the latter
5021 // candidates.
5022 //
5023 // Quick back-envelope calculation shows that for the list of n candidates
5024 // the equal probability for the candidate to persist as "best" can be
5025 // achieved by replacing it with "next" k-th candidate with the probability
5026 // of 1/k. It can be easily shown that by the end of the run, the
5027 // probability for any candidate is converged to 1/n, thus giving the
5028 // uniform distribution among all the candidates.
5029 //
5030 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5031 #define RANDOMIZED_DOMAIN_POW 29
5032 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5033 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5034 bool Compile::randomized_select(int count) {
5035 assert(count > 0, "only positive");
5036 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5037 }
5038
5039 #ifdef ASSERT
5040 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5041 bool Compile::fail_randomly() {
5042 if ((random() % StressBailoutMean) != 0) {
5043 return false;
5044 }
5045 record_failure("StressBailout");
5046 return true;
5047 }
5048
5049 bool Compile::failure_is_artificial() {
5050 return C->failure_reason_is("StressBailout");
5051 }
5052 #endif
5053
5054 CloneMap& Compile::clone_map() { return _clone_map; }
5055 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5056
5057 void NodeCloneInfo::dump_on(outputStream* st) const {
5058 st->print(" {%d:%d} ", idx(), gen());
5059 }
5060
5061 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5062 uint64_t val = value(old->_idx);
5063 NodeCloneInfo cio(val);
5064 assert(val != 0, "old node should be in the map");
5065 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5066 insert(nnn->_idx, cin.get());
5067 #ifndef PRODUCT
5068 if (is_debug()) {
5069 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5070 }
5071 #endif
5072 }
5073
5074 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5075 NodeCloneInfo cio(value(old->_idx));
5076 if (cio.get() == 0) {
5077 cio.set(old->_idx, 0);
5078 insert(old->_idx, cio.get());
5079 #ifndef PRODUCT
5080 if (is_debug()) {
5081 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5082 }
5083 #endif
5084 }
5085 clone(old, nnn, gen);
5086 }
5087
5088 int CloneMap::max_gen() const {
5089 int g = 0;
5090 DictI di(_dict);
5091 for(; di.test(); ++di) {
5092 int t = gen(di._key);
5093 if (g < t) {
5094 g = t;
5095 #ifndef PRODUCT
5096 if (is_debug()) {
5097 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5098 }
5099 #endif
5100 }
5101 }
5102 return g;
5103 }
5104
5105 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5106 uint64_t val = value(key);
5107 if (val != 0) {
5108 NodeCloneInfo ni(val);
5109 ni.dump_on(st);
5110 }
5111 }
5112
5113 void Compile::shuffle_macro_nodes() {
5114 if (_macro_nodes.length() < 2) {
5115 return;
5116 }
5117 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5118 uint j = C->random() % (i + 1);
5119 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5120 }
5121 }
5122
5123 // Move Allocate nodes to the start of the list
5124 void Compile::sort_macro_nodes() {
5125 int count = macro_count();
5126 int allocates = 0;
5127 for (int i = 0; i < count; i++) {
5128 Node* n = macro_node(i);
5129 if (n->is_Allocate()) {
5130 if (i != allocates) {
5131 Node* tmp = macro_node(allocates);
5132 _macro_nodes.at_put(allocates, n);
5133 _macro_nodes.at_put(i, tmp);
5134 }
5135 allocates++;
5136 }
5137 }
5138 }
5139
5140 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5141 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5142 EventCompilerPhase event(UNTIMED);
5143 if (event.should_commit()) {
5144 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5145 }
5146 #ifndef PRODUCT
5147 ResourceMark rm;
5148 stringStream ss;
5149 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5150 int iter = ++_igv_phase_iter[cpt];
5151 if (iter > 1) {
5152 ss.print(" %d", iter);
5153 }
5154 if (n != nullptr) {
5155 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5156 if (n->is_Call()) {
5157 CallNode* call = n->as_Call();
5158 if (call->_name != nullptr) {
5159 // E.g. uncommon traps etc.
5160 ss.print(" - %s", call->_name);
5161 } else if (call->is_CallJava()) {
5162 CallJavaNode* call_java = call->as_CallJava();
5163 if (call_java->method() != nullptr) {
5164 ss.print(" -");
5165 call_java->method()->print_short_name(&ss);
5166 }
5167 }
5168 }
5169 }
5170
5171 const char* name = ss.as_string();
5172 if (should_print_igv(level)) {
5173 _igv_printer->print_graph(name);
5174 }
5175 if (should_print_phase(level)) {
5176 print_phase(name);
5177 }
5178 if (should_print_ideal_phase(cpt)) {
5179 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5180 }
5181 #endif
5182 C->_latest_stage_start_counter.stamp();
5183 }
5184
5185 // Only used from CompileWrapper
5186 void Compile::begin_method() {
5187 #ifndef PRODUCT
5188 if (_method != nullptr && should_print_igv(1)) {
5189 _igv_printer->begin_method();
5190 }
5191 #endif
5192 C->_latest_stage_start_counter.stamp();
5193 }
5194
5195 // Only used from CompileWrapper
5196 void Compile::end_method() {
5197 EventCompilerPhase event(UNTIMED);
5198 if (event.should_commit()) {
5199 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5200 }
5201
5202 #ifndef PRODUCT
5203 if (_method != nullptr && should_print_igv(1)) {
5204 _igv_printer->end_method();
5205 }
5206 #endif
5207 }
5208
5209 #ifndef PRODUCT
5210 bool Compile::should_print_phase(const int level) const {
5211 return PrintPhaseLevel > 0 && directive()->PhasePrintLevelOption >= level &&
5212 _method != nullptr; // Do not print phases for stubs.
5213 }
5214
5215 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5216 return _directive->should_print_ideal_phase(cpt);
5217 }
5218
5219 void Compile::init_igv() {
5220 if (_igv_printer == nullptr) {
5221 _igv_printer = IdealGraphPrinter::printer();
5222 _igv_printer->set_compile(this);
5223 }
5224 }
5225
5226 bool Compile::should_print_igv(const int level) {
5227 PRODUCT_RETURN_(return false;);
5228
5229 if (PrintIdealGraphLevel < 0) { // disabled by the user
5230 return false;
5231 }
5232
5233 bool need = directive()->IGVPrintLevelOption >= level;
5234 if (need) {
5235 Compile::init_igv();
5236 }
5237 return need;
5238 }
5239
5240 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5241 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5242
5243 // Called from debugger. Prints method to the default file with the default phase name.
5244 // This works regardless of any Ideal Graph Visualizer flags set or not.
5245 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5246 void igv_print(void* sp, void* fp, void* pc) {
5247 frame fr(sp, fp, pc);
5248 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5249 }
5250
5251 // Same as igv_print() above but with a specified phase name.
5252 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5253 frame fr(sp, fp, pc);
5254 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5255 }
5256
5257 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5258 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5259 // This works regardless of any Ideal Graph Visualizer flags set or not.
5260 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5261 void igv_print(bool network, void* sp, void* fp, void* pc) {
5262 frame fr(sp, fp, pc);
5263 if (network) {
5264 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5265 } else {
5266 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5267 }
5268 }
5269
5270 // Same as igv_print(bool network, ...) above but with a specified phase name.
5271 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5272 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5273 frame fr(sp, fp, pc);
5274 if (network) {
5275 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5276 } else {
5277 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5278 }
5279 }
5280
5281 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5282 void igv_print_default() {
5283 Compile::current()->print_method(PHASE_DEBUG, 0);
5284 }
5285
5286 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5287 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5288 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5289 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5290 void igv_append(void* sp, void* fp, void* pc) {
5291 frame fr(sp, fp, pc);
5292 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5293 }
5294
5295 // Same as igv_append(...) above but with a specified phase name.
5296 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5297 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5298 frame fr(sp, fp, pc);
5299 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5300 }
5301
5302 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5303 const char* file_name = "custom_debug.xml";
5304 if (_debug_file_printer == nullptr) {
5305 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5306 } else {
5307 _debug_file_printer->update_compiled_method(C->method());
5308 }
5309 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5310 _debug_file_printer->print_graph(phase_name, fr);
5311 }
5312
5313 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5314 ResourceMark rm;
5315 GrowableArray<const Node*> empty_list;
5316 igv_print_graph_to_network(phase_name, empty_list, fr);
5317 }
5318
5319 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5320 if (_debug_network_printer == nullptr) {
5321 _debug_network_printer = new IdealGraphPrinter(C);
5322 } else {
5323 _debug_network_printer->update_compiled_method(C->method());
5324 }
5325 tty->print_cr("Method printed over network stream to IGV");
5326 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5327 }
5328 #endif // !PRODUCT
5329
5330 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5331 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5332 return value;
5333 }
5334 Node* result = nullptr;
5335 if (bt == T_BYTE) {
5336 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5337 result = new RShiftINode(result, phase->intcon(24));
5338 } else if (bt == T_BOOLEAN) {
5339 result = new AndINode(value, phase->intcon(0xFF));
5340 } else if (bt == T_CHAR) {
5341 result = new AndINode(value,phase->intcon(0xFFFF));
5342 } else {
5343 assert(bt == T_SHORT, "unexpected narrow type");
5344 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5345 result = new RShiftINode(result, phase->intcon(16));
5346 }
5347 if (transform_res) {
5348 result = phase->transform(result);
5349 }
5350 return result;
5351 }
5352
5353 void Compile::record_method_not_compilable_oom() {
5354 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5355 }