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