1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compilerOracle.hpp"
37 #include "compiler/compiler_globals.hpp"
38 #include "compiler/disassembler.hpp"
39 #include "compiler/oopMap.hpp"
40 #include "gc/shared/barrierSet.hpp"
41 #include "gc/shared/c2/barrierSetC2.hpp"
42 #include "jfr/jfrEvents.hpp"
43 #include "jvm_io.h"
44 #include "memory/allocation.hpp"
45 #include "memory/resourceArea.hpp"
46 #include "opto/addnode.hpp"
47 #include "opto/block.hpp"
48 #include "opto/c2compiler.hpp"
49 #include "opto/callGenerator.hpp"
50 #include "opto/callnode.hpp"
51 #include "opto/castnode.hpp"
52 #include "opto/cfgnode.hpp"
53 #include "opto/chaitin.hpp"
54 #include "opto/compile.hpp"
55 #include "opto/connode.hpp"
56 #include "opto/convertnode.hpp"
57 #include "opto/divnode.hpp"
58 #include "opto/escape.hpp"
59 #include "opto/idealGraphPrinter.hpp"
60 #include "opto/locknode.hpp"
61 #include "opto/loopnode.hpp"
62 #include "opto/machnode.hpp"
63 #include "opto/macro.hpp"
64 #include "opto/matcher.hpp"
65 #include "opto/mathexactnode.hpp"
66 #include "opto/memnode.hpp"
67 #include "opto/mulnode.hpp"
68 #include "opto/narrowptrnode.hpp"
69 #include "opto/node.hpp"
70 #include "opto/opcodes.hpp"
71 #include "opto/output.hpp"
72 #include "opto/parse.hpp"
73 #include "opto/phaseX.hpp"
74 #include "opto/rootnode.hpp"
75 #include "opto/runtime.hpp"
76 #include "opto/stringopts.hpp"
77 #include "opto/type.hpp"
78 #include "opto/vector.hpp"
79 #include "opto/vectornode.hpp"
80 #include "runtime/globals_extension.hpp"
81 #include "runtime/sharedRuntime.hpp"
82 #include "runtime/signature.hpp"
83 #include "runtime/stubRoutines.hpp"
84 #include "runtime/timer.hpp"
85 #include "utilities/align.hpp"
86 #include "utilities/copy.hpp"
87 #include "utilities/macros.hpp"
88 #include "utilities/resourceHash.hpp"
89
90 // -------------------- Compile::mach_constant_base_node -----------------------
91 // Constant table base node singleton.
92 MachConstantBaseNode* Compile::mach_constant_base_node() {
93 if (_mach_constant_base_node == nullptr) {
94 _mach_constant_base_node = new MachConstantBaseNode();
95 _mach_constant_base_node->add_req(C->root());
96 }
97 return _mach_constant_base_node;
98 }
99
100
101 /// Support for intrinsics.
102
103 // Return the index at which m must be inserted (or already exists).
104 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
105 class IntrinsicDescPair {
106 private:
107 ciMethod* _m;
108 bool _is_virtual;
109 public:
110 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
111 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
112 ciMethod* m= elt->method();
113 ciMethod* key_m = key->_m;
114 if (key_m < m) return -1;
115 else if (key_m > m) return 1;
116 else {
117 bool is_virtual = elt->is_virtual();
118 bool key_virtual = key->_is_virtual;
119 if (key_virtual < is_virtual) return -1;
120 else if (key_virtual > is_virtual) return 1;
121 else return 0;
122 }
123 }
124 };
125 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
126 #ifdef ASSERT
127 for (int i = 1; i < _intrinsics.length(); i++) {
128 CallGenerator* cg1 = _intrinsics.at(i-1);
129 CallGenerator* cg2 = _intrinsics.at(i);
130 assert(cg1->method() != cg2->method()
131 ? cg1->method() < cg2->method()
132 : cg1->is_virtual() < cg2->is_virtual(),
133 "compiler intrinsics list must stay sorted");
134 }
135 #endif
136 IntrinsicDescPair pair(m, is_virtual);
137 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
138 }
139
140 void Compile::register_intrinsic(CallGenerator* cg) {
141 bool found = false;
142 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
143 assert(!found, "registering twice");
144 _intrinsics.insert_before(index, cg);
145 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
146 }
147
148 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
149 assert(m->is_loaded(), "don't try this on unloaded methods");
150 if (_intrinsics.length() > 0) {
151 bool found = false;
152 int index = intrinsic_insertion_index(m, is_virtual, found);
153 if (found) {
154 return _intrinsics.at(index);
155 }
156 }
157 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
158 if (m->intrinsic_id() != vmIntrinsics::_none &&
159 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
160 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
161 if (cg != nullptr) {
162 // Save it for next time:
163 register_intrinsic(cg);
164 return cg;
165 } else {
166 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
167 }
168 }
169 return nullptr;
170 }
171
172 // Compile::make_vm_intrinsic is defined in library_call.cpp.
173
174 #ifndef PRODUCT
175 // statistics gathering...
176
177 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
179
180 inline int as_int(vmIntrinsics::ID id) {
181 return vmIntrinsics::as_int(id);
182 }
183
184 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
185 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
186 int oflags = _intrinsic_hist_flags[as_int(id)];
187 assert(flags != 0, "what happened?");
188 if (is_virtual) {
189 flags |= _intrinsic_virtual;
190 }
191 bool changed = (flags != oflags);
192 if ((flags & _intrinsic_worked) != 0) {
193 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
194 if (count == 1) {
195 changed = true; // first time
196 }
197 // increment the overall count also:
198 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
199 }
200 if (changed) {
201 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
202 // Something changed about the intrinsic's virtuality.
203 if ((flags & _intrinsic_virtual) != 0) {
204 // This is the first use of this intrinsic as a virtual call.
205 if (oflags != 0) {
206 // We already saw it as a non-virtual, so note both cases.
207 flags |= _intrinsic_both;
208 }
209 } else if ((oflags & _intrinsic_both) == 0) {
210 // This is the first use of this intrinsic as a non-virtual
211 flags |= _intrinsic_both;
212 }
213 }
214 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
215 }
216 // update the overall flags also:
217 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
218 return changed;
219 }
220
221 static char* format_flags(int flags, char* buf) {
222 buf[0] = 0;
223 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
224 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
225 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
226 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
227 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
228 if (buf[0] == 0) strcat(buf, ",");
229 assert(buf[0] == ',', "must be");
230 return &buf[1];
231 }
232
233 void Compile::print_intrinsic_statistics() {
234 char flagsbuf[100];
235 ttyLocker ttyl;
236 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
237 tty->print_cr("Compiler intrinsic usage:");
238 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
239 if (total == 0) total = 1; // avoid div0 in case of no successes
240 #define PRINT_STAT_LINE(name, c, f) \
241 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
242 for (auto id : EnumRange<vmIntrinsicID>{}) {
243 int flags = _intrinsic_hist_flags[as_int(id)];
244 juint count = _intrinsic_hist_count[as_int(id)];
245 if ((flags | count) != 0) {
246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
247 }
248 }
249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
250 if (xtty != nullptr) xtty->tail("statistics");
251 }
252
253 void Compile::print_statistics() {
254 { ttyLocker ttyl;
255 if (xtty != nullptr) xtty->head("statistics type='opto'");
256 Parse::print_statistics();
257 PhaseStringOpts::print_statistics();
258 PhaseCCP::print_statistics();
259 PhaseRegAlloc::print_statistics();
260 PhaseOutput::print_statistics();
261 PhasePeephole::print_statistics();
262 PhaseIdealLoop::print_statistics();
263 ConnectionGraph::print_statistics();
264 PhaseMacroExpand::print_statistics();
265 if (xtty != nullptr) xtty->tail("statistics");
266 }
267 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
268 // put this under its own <statistics> element.
269 print_intrinsic_statistics();
270 }
271 }
272 #endif //PRODUCT
273
274 void Compile::gvn_replace_by(Node* n, Node* nn) {
275 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
276 Node* use = n->last_out(i);
277 bool is_in_table = initial_gvn()->hash_delete(use);
278 uint uses_found = 0;
279 for (uint j = 0; j < use->len(); j++) {
280 if (use->in(j) == n) {
281 if (j < use->req())
282 use->set_req(j, nn);
283 else
284 use->set_prec(j, nn);
285 uses_found++;
286 }
287 }
288 if (is_in_table) {
289 // reinsert into table
290 initial_gvn()->hash_find_insert(use);
291 }
292 record_for_igvn(use);
293 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
294 i -= uses_found; // we deleted 1 or more copies of this edge
295 }
296 }
297
298
299 // Identify all nodes that are reachable from below, useful.
300 // Use breadth-first pass that records state in a Unique_Node_List,
301 // recursive traversal is slower.
302 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
303 int estimated_worklist_size = live_nodes();
304 useful.map( estimated_worklist_size, nullptr ); // preallocate space
305
306 // Initialize worklist
307 if (root() != nullptr) { useful.push(root()); }
308 // If 'top' is cached, declare it useful to preserve cached node
309 if (cached_top_node()) { useful.push(cached_top_node()); }
310
311 // Push all useful nodes onto the list, breadthfirst
312 for( uint next = 0; next < useful.size(); ++next ) {
313 assert( next < unique(), "Unique useful nodes < total nodes");
314 Node *n = useful.at(next);
315 uint max = n->len();
316 for( uint i = 0; i < max; ++i ) {
317 Node *m = n->in(i);
318 if (not_a_node(m)) continue;
319 useful.push(m);
320 }
321 }
322 }
323
324 // Update dead_node_list with any missing dead nodes using useful
325 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
326 void Compile::update_dead_node_list(Unique_Node_List &useful) {
327 uint max_idx = unique();
328 VectorSet& useful_node_set = useful.member_set();
329
330 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
331 // If node with index node_idx is not in useful set,
332 // mark it as dead in dead node list.
333 if (!useful_node_set.test(node_idx)) {
334 record_dead_node(node_idx);
335 }
336 }
337 }
338
339 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
340 int shift = 0;
341 for (int i = 0; i < inlines->length(); i++) {
342 CallGenerator* cg = inlines->at(i);
343 if (useful.member(cg->call_node())) {
344 if (shift > 0) {
345 inlines->at_put(i - shift, cg);
346 }
347 } else {
348 shift++; // skip over the dead element
349 }
350 }
351 if (shift > 0) {
352 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
353 }
354 }
355
356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
357 assert(dead != nullptr && dead->is_Call(), "sanity");
358 int found = 0;
359 for (int i = 0; i < inlines->length(); i++) {
360 if (inlines->at(i)->call_node() == dead) {
361 inlines->remove_at(i);
362 found++;
363 NOT_DEBUG( break; ) // elements are unique, so exit early
364 }
365 }
366 assert(found <= 1, "not unique");
367 }
368
369 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
370 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
371 for (int i = node_list.length() - 1; i >= 0; i--) {
372 N* node = node_list.at(i);
373 if (!useful.member(node)) {
374 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
375 }
376 }
377 }
378
379 void Compile::remove_useless_node(Node* dead) {
380 remove_modified_node(dead);
381
382 // Constant node that has no out-edges and has only one in-edge from
383 // root is usually dead. However, sometimes reshaping walk makes
384 // it reachable by adding use edges. So, we will NOT count Con nodes
385 // as dead to be conservative about the dead node count at any
386 // given time.
387 if (!dead->is_Con()) {
388 record_dead_node(dead->_idx);
389 }
390 if (dead->is_macro()) {
391 remove_macro_node(dead);
392 }
393 if (dead->is_expensive()) {
394 remove_expensive_node(dead);
395 }
396 if (dead->Opcode() == Op_Opaque4) {
397 remove_template_assertion_predicate_opaq(dead);
398 }
399 if (dead->is_ParsePredicate()) {
400 remove_parse_predicate(dead->as_ParsePredicate());
401 }
402 if (dead->for_post_loop_opts_igvn()) {
403 remove_from_post_loop_opts_igvn(dead);
404 }
405 if (dead->is_Call()) {
406 remove_useless_late_inlines( &_late_inlines, dead);
407 remove_useless_late_inlines( &_string_late_inlines, dead);
408 remove_useless_late_inlines( &_boxing_late_inlines, dead);
409 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
410
411 if (dead->is_CallStaticJava()) {
412 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
413 }
414 }
415 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
416 bs->unregister_potential_barrier_node(dead);
417 }
418
419 // Disconnect all useless nodes by disconnecting those at the boundary.
420 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
421 uint next = 0;
422 while (next < useful.size()) {
423 Node *n = useful.at(next++);
424 if (n->is_SafePoint()) {
425 // We're done with a parsing phase. Replaced nodes are not valid
426 // beyond that point.
427 n->as_SafePoint()->delete_replaced_nodes();
428 }
429 // Use raw traversal of out edges since this code removes out edges
430 int max = n->outcnt();
431 for (int j = 0; j < max; ++j) {
432 Node* child = n->raw_out(j);
433 if (!useful.member(child)) {
434 assert(!child->is_top() || child != top(),
435 "If top is cached in Compile object it is in useful list");
436 // Only need to remove this out-edge to the useless node
437 n->raw_del_out(j);
438 --j;
439 --max;
440 }
441 }
442 if (n->outcnt() == 1 && n->has_special_unique_user()) {
443 assert(useful.member(n->unique_out()), "do not push a useless node");
444 worklist.push(n->unique_out());
445 }
446 }
447
448 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
449 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
450 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
451 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
452 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
453 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
454 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
455 #ifdef ASSERT
456 if (_modified_nodes != nullptr) {
457 _modified_nodes->remove_useless_nodes(useful.member_set());
458 }
459 #endif
460
461 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
462 bs->eliminate_useless_gc_barriers(useful, this);
463 // clean up the late inline lists
464 remove_useless_late_inlines( &_late_inlines, useful);
465 remove_useless_late_inlines( &_string_late_inlines, useful);
466 remove_useless_late_inlines( &_boxing_late_inlines, useful);
467 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
468 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
469 }
470
471 // ============================================================================
472 //------------------------------CompileWrapper---------------------------------
473 class CompileWrapper : public StackObj {
474 Compile *const _compile;
475 public:
476 CompileWrapper(Compile* compile);
477
478 ~CompileWrapper();
479 };
480
481 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
482 // the Compile* pointer is stored in the current ciEnv:
483 ciEnv* env = compile->env();
484 assert(env == ciEnv::current(), "must already be a ciEnv active");
485 assert(env->compiler_data() == nullptr, "compile already active?");
486 env->set_compiler_data(compile);
487 assert(compile == Compile::current(), "sanity");
488
489 compile->set_type_dict(nullptr);
490 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
491 compile->clone_map().set_clone_idx(0);
492 compile->set_type_last_size(0);
493 compile->set_last_tf(nullptr, nullptr);
494 compile->set_indexSet_arena(nullptr);
495 compile->set_indexSet_free_block_list(nullptr);
496 compile->init_type_arena();
497 Type::Initialize(compile);
498 _compile->begin_method();
499 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
500 }
501 CompileWrapper::~CompileWrapper() {
502 // simulate crash during compilation
503 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
504
505 _compile->end_method();
506 _compile->env()->set_compiler_data(nullptr);
507 }
508
509
510 //----------------------------print_compile_messages---------------------------
511 void Compile::print_compile_messages() {
512 #ifndef PRODUCT
513 // Check if recompiling
514 if (!subsume_loads() && PrintOpto) {
515 // Recompiling without allowing machine instructions to subsume loads
516 tty->print_cr("*********************************************************");
517 tty->print_cr("** Bailout: Recompile without subsuming loads **");
518 tty->print_cr("*********************************************************");
519 }
520 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
521 // Recompiling without escape analysis
522 tty->print_cr("*********************************************************");
523 tty->print_cr("** Bailout: Recompile without escape analysis **");
524 tty->print_cr("*********************************************************");
525 }
526 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
527 // Recompiling without iterative escape analysis
528 tty->print_cr("*********************************************************");
529 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
530 tty->print_cr("*********************************************************");
531 }
532 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
533 // Recompiling without reducing allocation merges
534 tty->print_cr("*********************************************************");
535 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
536 tty->print_cr("*********************************************************");
537 }
538 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
539 // Recompiling without boxing elimination
540 tty->print_cr("*********************************************************");
541 tty->print_cr("** Bailout: Recompile without boxing elimination **");
542 tty->print_cr("*********************************************************");
543 }
544 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
545 // Recompiling without locks coarsening
546 tty->print_cr("*********************************************************");
547 tty->print_cr("** Bailout: Recompile without locks coarsening **");
548 tty->print_cr("*********************************************************");
549 }
550 if (env()->break_at_compile()) {
551 // Open the debugger when compiling this method.
552 tty->print("### Breaking when compiling: ");
553 method()->print_short_name();
554 tty->cr();
555 BREAKPOINT;
556 }
557
558 if( PrintOpto ) {
559 if (is_osr_compilation()) {
560 tty->print("[OSR]%3d", _compile_id);
561 } else {
562 tty->print("%3d", _compile_id);
563 }
564 }
565 #endif
566 }
567
568 #ifndef PRODUCT
569 void Compile::print_ideal_ir(const char* phase_name) {
570 // keep the following output all in one block
571 // This output goes directly to the tty, not the compiler log.
572 // To enable tools to match it up with the compilation activity,
573 // be sure to tag this tty output with the compile ID.
574
575 // Node dumping can cause a safepoint, which can break the tty lock.
576 // Buffer all node dumps, so that all safepoints happen before we lock.
577 ResourceMark rm;
578 stringStream ss;
579
580 if (_output == nullptr) {
581 ss.print_cr("AFTER: %s", phase_name);
582 // Print out all nodes in ascending order of index.
583 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
584 } else {
585 // Dump the node blockwise if we have a scheduling
586 _output->print_scheduling(&ss);
587 }
588
589 // Check that the lock is not broken by a safepoint.
590 NoSafepointVerifier nsv;
591 ttyLocker ttyl;
592 if (xtty != nullptr) {
593 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
594 compile_id(),
595 is_osr_compilation() ? " compile_kind='osr'" : "",
596 phase_name);
597 }
598
599 tty->print("%s", ss.as_string());
600
601 if (xtty != nullptr) {
602 xtty->tail("ideal");
603 }
604 }
605 #endif
606
607 // ============================================================================
608 //------------------------------Compile standard-------------------------------
609
610 // Compile a method. entry_bci is -1 for normal compilations and indicates
611 // the continuation bci for on stack replacement.
612
613
614 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
615 Options options, DirectiveSet* directive)
616 : Phase(Compiler),
617 _compile_id(ci_env->compile_id()),
618 _options(options),
619 _method(target),
620 _entry_bci(osr_bci),
621 _ilt(nullptr),
622 _stub_function(nullptr),
623 _stub_name(nullptr),
624 _stub_entry_point(nullptr),
625 _max_node_limit(MaxNodeLimit),
626 _post_loop_opts_phase(false),
627 _allow_macro_nodes(true),
628 _inlining_progress(false),
629 _inlining_incrementally(false),
630 _do_cleanup(false),
631 _has_reserved_stack_access(target->has_reserved_stack_access()),
632 #ifndef PRODUCT
633 _igv_idx(0),
634 _trace_opto_output(directive->TraceOptoOutputOption),
635 #endif
636 _has_method_handle_invokes(false),
637 _clinit_barrier_on_entry(false),
638 _stress_seed(0),
639 _comp_arena(mtCompiler),
640 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
641 _env(ci_env),
642 _directive(directive),
643 _log(ci_env->log()),
644 _first_failure_details(nullptr),
645 _intrinsics (comp_arena(), 0, 0, nullptr),
646 _macro_nodes (comp_arena(), 8, 0, nullptr),
647 _parse_predicates (comp_arena(), 8, 0, nullptr),
648 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
649 _expensive_nodes (comp_arena(), 8, 0, nullptr),
650 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
651 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
652 _coarsened_locks (comp_arena(), 8, 0, nullptr),
653 _congraph(nullptr),
654 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
655 _unique(0),
656 _dead_node_count(0),
657 _dead_node_list(comp_arena()),
658 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
659 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
660 _node_arena(&_node_arena_one),
661 _mach_constant_base_node(nullptr),
662 _Compile_types(mtCompiler),
663 _initial_gvn(nullptr),
664 _igvn_worklist(nullptr),
665 _types(nullptr),
666 _node_hash(nullptr),
667 _late_inlines(comp_arena(), 2, 0, nullptr),
668 _string_late_inlines(comp_arena(), 2, 0, nullptr),
669 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
670 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
671 _late_inlines_pos(0),
672 _number_of_mh_late_inlines(0),
673 _oom(false),
674 _print_inlining_stream(new (mtCompiler) stringStream()),
675 _print_inlining_list(nullptr),
676 _print_inlining_idx(0),
677 _print_inlining_output(nullptr),
678 _replay_inline_data(nullptr),
679 _java_calls(0),
680 _inner_loops(0),
681 _interpreter_frame_size(0),
682 _output(nullptr)
683 #ifndef PRODUCT
684 , _in_dump_cnt(0)
685 #endif
686 {
687 C = this;
688 CompileWrapper cw(this);
689
690 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
691 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
692
693 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
694 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
695 // We can always print a disassembly, either abstract (hex dump) or
696 // with the help of a suitable hsdis library. Thus, we should not
697 // couple print_assembly and print_opto_assembly controls.
698 // But: always print opto and regular assembly on compile command 'print'.
699 bool print_assembly = directive->PrintAssemblyOption;
700 set_print_assembly(print_opto_assembly || print_assembly);
701 #else
702 set_print_assembly(false); // must initialize.
703 #endif
704
705 #ifndef PRODUCT
706 set_parsed_irreducible_loop(false);
707 #endif
708
709 if (directive->ReplayInlineOption) {
710 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
711 }
712 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
713 set_print_intrinsics(directive->PrintIntrinsicsOption);
714 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
715
716 if (ProfileTraps) {
717 // Make sure the method being compiled gets its own MDO,
718 // so we can at least track the decompile_count().
719 method()->ensure_method_data();
720 }
721
722 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
723 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout) {
724 initialize_stress_seed(directive);
725 }
726
727 Init(/*do_aliasing=*/ true);
728
729 print_compile_messages();
730
731 _ilt = InlineTree::build_inline_tree_root();
732
733 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
734 assert(num_alias_types() >= AliasIdxRaw, "");
735
736 #define MINIMUM_NODE_HASH 1023
737
738 // GVN that will be run immediately on new nodes
739 uint estimated_size = method()->code_size()*4+64;
740 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
741 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
742 _types = new (comp_arena()) Type_Array(comp_arena());
743 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
744 PhaseGVN gvn;
745 set_initial_gvn(&gvn);
746
747 print_inlining_init();
748 { // Scope for timing the parser
749 TracePhase tp("parse", &timers[_t_parser]);
750
751 // Put top into the hash table ASAP.
752 initial_gvn()->transform(top());
753
754 // Set up tf(), start(), and find a CallGenerator.
755 CallGenerator* cg = nullptr;
756 if (is_osr_compilation()) {
757 const TypeTuple *domain = StartOSRNode::osr_domain();
758 const TypeTuple *range = TypeTuple::make_range(method()->signature());
759 init_tf(TypeFunc::make(domain, range));
760 StartNode* s = new StartOSRNode(root(), domain);
761 initial_gvn()->set_type_bottom(s);
762 verify_start(s);
763 cg = CallGenerator::for_osr(method(), entry_bci());
764 } else {
765 // Normal case.
766 init_tf(TypeFunc::make(method()));
767 StartNode* s = new StartNode(root(), tf()->domain());
768 initial_gvn()->set_type_bottom(s);
769 verify_start(s);
770 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
771 // With java.lang.ref.reference.get() we must go through the
772 // intrinsic - even when get() is the root
773 // method of the compile - so that, if necessary, the value in
774 // the referent field of the reference object gets recorded by
775 // the pre-barrier code.
776 cg = find_intrinsic(method(), false);
777 }
778 if (cg == nullptr) {
779 float past_uses = method()->interpreter_invocation_count();
780 float expected_uses = past_uses;
781 cg = CallGenerator::for_inline(method(), expected_uses);
782 }
783 }
784 if (failing()) return;
785 if (cg == nullptr) {
786 const char* reason = InlineTree::check_can_parse(method());
787 assert(reason != nullptr, "expect reason for parse failure");
788 stringStream ss;
789 ss.print("cannot parse method: %s", reason);
790 record_method_not_compilable(ss.as_string());
791 return;
792 }
793
794 gvn.set_type(root(), root()->bottom_type());
795
796 JVMState* jvms = build_start_state(start(), tf());
797 if ((jvms = cg->generate(jvms)) == nullptr) {
798 assert(failure_reason() != nullptr, "expect reason for parse failure");
799 stringStream ss;
800 ss.print("method parse failed: %s", failure_reason());
801 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
802 return;
803 }
804 GraphKit kit(jvms);
805
806 if (!kit.stopped()) {
807 // Accept return values, and transfer control we know not where.
808 // This is done by a special, unique ReturnNode bound to root.
809 return_values(kit.jvms());
810 }
811
812 if (kit.has_exceptions()) {
813 // Any exceptions that escape from this call must be rethrown
814 // to whatever caller is dynamically above us on the stack.
815 // This is done by a special, unique RethrowNode bound to root.
816 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
817 }
818
819 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
820
821 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
822 inline_string_calls(true);
823 }
824
825 if (failing()) return;
826
827 // Remove clutter produced by parsing.
828 if (!failing()) {
829 ResourceMark rm;
830 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
831 }
832 }
833
834 // Note: Large methods are capped off in do_one_bytecode().
835 if (failing()) return;
836
837 // After parsing, node notes are no longer automagic.
838 // They must be propagated by register_new_node_with_optimizer(),
839 // clone(), or the like.
840 set_default_node_notes(nullptr);
841
842 #ifndef PRODUCT
843 if (should_print_igv(1)) {
844 _igv_printer->print_inlining();
845 }
846 #endif
847
848 if (failing()) return;
849 NOT_PRODUCT( verify_graph_edges(); )
850
851 // Now optimize
852 Optimize();
853 if (failing()) return;
854 NOT_PRODUCT( verify_graph_edges(); )
855
856 #ifndef PRODUCT
857 if (should_print_ideal()) {
858 print_ideal_ir("print_ideal");
859 }
860 #endif
861
862 #ifdef ASSERT
863 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
864 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
865 #endif
866
867 // Dump compilation data to replay it.
868 if (directive->DumpReplayOption) {
869 env()->dump_replay_data(_compile_id);
870 }
871 if (directive->DumpInlineOption && (ilt() != nullptr)) {
872 env()->dump_inline_data(_compile_id);
873 }
874
875 // Now that we know the size of all the monitors we can add a fixed slot
876 // for the original deopt pc.
877 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
878 set_fixed_slots(next_slot);
879
880 // Compute when to use implicit null checks. Used by matching trap based
881 // nodes and NullCheck optimization.
882 set_allowed_deopt_reasons();
883
884 // Now generate code
885 Code_Gen();
886 }
887
888 //------------------------------Compile----------------------------------------
889 // Compile a runtime stub
890 Compile::Compile( ciEnv* ci_env,
891 TypeFunc_generator generator,
892 address stub_function,
893 const char *stub_name,
894 int is_fancy_jump,
895 bool pass_tls,
896 bool return_pc,
897 DirectiveSet* directive)
898 : Phase(Compiler),
899 _compile_id(0),
900 _options(Options::for_runtime_stub()),
901 _method(nullptr),
902 _entry_bci(InvocationEntryBci),
903 _stub_function(stub_function),
904 _stub_name(stub_name),
905 _stub_entry_point(nullptr),
906 _max_node_limit(MaxNodeLimit),
907 _post_loop_opts_phase(false),
908 _allow_macro_nodes(true),
909 _inlining_progress(false),
910 _inlining_incrementally(false),
911 _has_reserved_stack_access(false),
912 #ifndef PRODUCT
913 _igv_idx(0),
914 _trace_opto_output(directive->TraceOptoOutputOption),
915 #endif
916 _has_method_handle_invokes(false),
917 _clinit_barrier_on_entry(false),
918 _stress_seed(0),
919 _comp_arena(mtCompiler),
920 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
921 _env(ci_env),
922 _directive(directive),
923 _log(ci_env->log()),
924 _first_failure_details(nullptr),
925 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
926 _congraph(nullptr),
927 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
928 _unique(0),
929 _dead_node_count(0),
930 _dead_node_list(comp_arena()),
931 _node_arena_one(mtCompiler),
932 _node_arena_two(mtCompiler),
933 _node_arena(&_node_arena_one),
934 _mach_constant_base_node(nullptr),
935 _Compile_types(mtCompiler),
936 _initial_gvn(nullptr),
937 _igvn_worklist(nullptr),
938 _types(nullptr),
939 _node_hash(nullptr),
940 _number_of_mh_late_inlines(0),
941 _oom(false),
942 _print_inlining_stream(new (mtCompiler) stringStream()),
943 _print_inlining_list(nullptr),
944 _print_inlining_idx(0),
945 _print_inlining_output(nullptr),
946 _replay_inline_data(nullptr),
947 _java_calls(0),
948 _inner_loops(0),
949 _interpreter_frame_size(0),
950 _output(nullptr),
951 #ifndef PRODUCT
952 _in_dump_cnt(0),
953 #endif
954 _allowed_reasons(0) {
955 C = this;
956
957 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
958 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
959
960 #ifndef PRODUCT
961 set_print_assembly(PrintFrameConverterAssembly);
962 set_parsed_irreducible_loop(false);
963 #else
964 set_print_assembly(false); // Must initialize.
965 #endif
966 set_has_irreducible_loop(false); // no loops
967
968 CompileWrapper cw(this);
969 Init(/*do_aliasing=*/ false);
970 init_tf((*generator)());
971
972 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
973 _types = new (comp_arena()) Type_Array(comp_arena());
974 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
975
976 if (StressLCM || StressGCM || StressBailout) {
977 initialize_stress_seed(directive);
978 }
979
980 {
981 PhaseGVN gvn;
982 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
983 gvn.transform(top());
984
985 GraphKit kit;
986 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
987 }
988
989 NOT_PRODUCT( verify_graph_edges(); )
990
991 Code_Gen();
992 }
993
994 Compile::~Compile() {
995 delete _print_inlining_stream;
996 delete _first_failure_details;
997 };
998
999 //------------------------------Init-------------------------------------------
1000 // Prepare for a single compilation
1001 void Compile::Init(bool aliasing) {
1002 _do_aliasing = aliasing;
1003 _unique = 0;
1004 _regalloc = nullptr;
1005
1006 _tf = nullptr; // filled in later
1007 _top = nullptr; // cached later
1008 _matcher = nullptr; // filled in later
1009 _cfg = nullptr; // filled in later
1010
1011 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1012
1013 _node_note_array = nullptr;
1014 _default_node_notes = nullptr;
1015 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1016
1017 _immutable_memory = nullptr; // filled in at first inquiry
1018
1019 #ifdef ASSERT
1020 _phase_optimize_finished = false;
1021 _phase_verify_ideal_loop = false;
1022 _exception_backedge = false;
1023 _type_verify = nullptr;
1024 #endif
1025
1026 // Globally visible Nodes
1027 // First set TOP to null to give safe behavior during creation of RootNode
1028 set_cached_top_node(nullptr);
1029 set_root(new RootNode());
1030 // Now that you have a Root to point to, create the real TOP
1031 set_cached_top_node( new ConNode(Type::TOP) );
1032 set_recent_alloc(nullptr, nullptr);
1033
1034 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1035 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1036 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1037 env()->set_dependencies(new Dependencies(env()));
1038
1039 _fixed_slots = 0;
1040 set_has_split_ifs(false);
1041 set_has_loops(false); // first approximation
1042 set_has_stringbuilder(false);
1043 set_has_boxed_value(false);
1044 _trap_can_recompile = false; // no traps emitted yet
1045 _major_progress = true; // start out assuming good things will happen
1046 set_has_unsafe_access(false);
1047 set_max_vector_size(0);
1048 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1049 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1050 set_decompile_count(0);
1051
1052 #ifndef PRODUCT
1053 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1054 #endif
1055
1056 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1057 _loop_opts_cnt = LoopOptsCount;
1058 set_do_inlining(Inline);
1059 set_max_inline_size(MaxInlineSize);
1060 set_freq_inline_size(FreqInlineSize);
1061 set_do_scheduling(OptoScheduling);
1062
1063 set_do_vector_loop(false);
1064 set_has_monitors(false);
1065 set_has_scoped_access(false);
1066
1067 if (AllowVectorizeOnDemand) {
1068 if (has_method() && _directive->VectorizeOption) {
1069 set_do_vector_loop(true);
1070 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());})
1071 } else if (has_method() && method()->name() != nullptr &&
1072 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1073 set_do_vector_loop(true);
1074 }
1075 }
1076 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1077 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());})
1078
1079 _max_node_limit = _directive->MaxNodeLimitOption;
1080
1081 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1082 set_clinit_barrier_on_entry(true);
1083 }
1084 if (debug_info()->recording_non_safepoints()) {
1085 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1086 (comp_arena(), 8, 0, nullptr));
1087 set_default_node_notes(Node_Notes::make(this));
1088 }
1089
1090 const int grow_ats = 16;
1091 _max_alias_types = grow_ats;
1092 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1093 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1094 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1095 {
1096 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1097 }
1098 // Initialize the first few types.
1099 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1100 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1101 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1102 _num_alias_types = AliasIdxRaw+1;
1103 // Zero out the alias type cache.
1104 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1105 // A null adr_type hits in the cache right away. Preload the right answer.
1106 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1107 }
1108
1109 #ifdef ASSERT
1110 // Verify that the current StartNode is valid.
1111 void Compile::verify_start(StartNode* s) const {
1112 assert(failing_internal() || s == start(), "should be StartNode");
1113 }
1114 #endif
1115
1116 /**
1117 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1118 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1119 * the ideal graph.
1120 */
1121 StartNode* Compile::start() const {
1122 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1123 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1124 Node* start = root()->fast_out(i);
1125 if (start->is_Start()) {
1126 return start->as_Start();
1127 }
1128 }
1129 fatal("Did not find Start node!");
1130 return nullptr;
1131 }
1132
1133 //-------------------------------immutable_memory-------------------------------------
1134 // Access immutable memory
1135 Node* Compile::immutable_memory() {
1136 if (_immutable_memory != nullptr) {
1137 return _immutable_memory;
1138 }
1139 StartNode* s = start();
1140 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1141 Node *p = s->fast_out(i);
1142 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1143 _immutable_memory = p;
1144 return _immutable_memory;
1145 }
1146 }
1147 ShouldNotReachHere();
1148 return nullptr;
1149 }
1150
1151 //----------------------set_cached_top_node------------------------------------
1152 // Install the cached top node, and make sure Node::is_top works correctly.
1153 void Compile::set_cached_top_node(Node* tn) {
1154 if (tn != nullptr) verify_top(tn);
1155 Node* old_top = _top;
1156 _top = tn;
1157 // Calling Node::setup_is_top allows the nodes the chance to adjust
1158 // their _out arrays.
1159 if (_top != nullptr) _top->setup_is_top();
1160 if (old_top != nullptr) old_top->setup_is_top();
1161 assert(_top == nullptr || top()->is_top(), "");
1162 }
1163
1164 #ifdef ASSERT
1165 uint Compile::count_live_nodes_by_graph_walk() {
1166 Unique_Node_List useful(comp_arena());
1167 // Get useful node list by walking the graph.
1168 identify_useful_nodes(useful);
1169 return useful.size();
1170 }
1171
1172 void Compile::print_missing_nodes() {
1173
1174 // Return if CompileLog is null and PrintIdealNodeCount is false.
1175 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1176 return;
1177 }
1178
1179 // This is an expensive function. It is executed only when the user
1180 // specifies VerifyIdealNodeCount option or otherwise knows the
1181 // additional work that needs to be done to identify reachable nodes
1182 // by walking the flow graph and find the missing ones using
1183 // _dead_node_list.
1184
1185 Unique_Node_List useful(comp_arena());
1186 // Get useful node list by walking the graph.
1187 identify_useful_nodes(useful);
1188
1189 uint l_nodes = C->live_nodes();
1190 uint l_nodes_by_walk = useful.size();
1191
1192 if (l_nodes != l_nodes_by_walk) {
1193 if (_log != nullptr) {
1194 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1195 _log->stamp();
1196 _log->end_head();
1197 }
1198 VectorSet& useful_member_set = useful.member_set();
1199 int last_idx = l_nodes_by_walk;
1200 for (int i = 0; i < last_idx; i++) {
1201 if (useful_member_set.test(i)) {
1202 if (_dead_node_list.test(i)) {
1203 if (_log != nullptr) {
1204 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1205 }
1206 if (PrintIdealNodeCount) {
1207 // Print the log message to tty
1208 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1209 useful.at(i)->dump();
1210 }
1211 }
1212 }
1213 else if (! _dead_node_list.test(i)) {
1214 if (_log != nullptr) {
1215 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1216 }
1217 if (PrintIdealNodeCount) {
1218 // Print the log message to tty
1219 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1220 }
1221 }
1222 }
1223 if (_log != nullptr) {
1224 _log->tail("mismatched_nodes");
1225 }
1226 }
1227 }
1228 void Compile::record_modified_node(Node* n) {
1229 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1230 _modified_nodes->push(n);
1231 }
1232 }
1233
1234 void Compile::remove_modified_node(Node* n) {
1235 if (_modified_nodes != nullptr) {
1236 _modified_nodes->remove(n);
1237 }
1238 }
1239 #endif
1240
1241 #ifndef PRODUCT
1242 void Compile::verify_top(Node* tn) const {
1243 if (tn != nullptr) {
1244 assert(tn->is_Con(), "top node must be a constant");
1245 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1246 assert(tn->in(0) != nullptr, "must have live top node");
1247 }
1248 }
1249 #endif
1250
1251
1252 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1253
1254 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1255 guarantee(arr != nullptr, "");
1256 int num_blocks = arr->length();
1257 if (grow_by < num_blocks) grow_by = num_blocks;
1258 int num_notes = grow_by * _node_notes_block_size;
1259 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1260 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1261 while (num_notes > 0) {
1262 arr->append(notes);
1263 notes += _node_notes_block_size;
1264 num_notes -= _node_notes_block_size;
1265 }
1266 assert(num_notes == 0, "exact multiple, please");
1267 }
1268
1269 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1270 if (source == nullptr || dest == nullptr) return false;
1271
1272 if (dest->is_Con())
1273 return false; // Do not push debug info onto constants.
1274
1275 #ifdef ASSERT
1276 // Leave a bread crumb trail pointing to the original node:
1277 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1278 dest->set_debug_orig(source);
1279 }
1280 #endif
1281
1282 if (node_note_array() == nullptr)
1283 return false; // Not collecting any notes now.
1284
1285 // This is a copy onto a pre-existing node, which may already have notes.
1286 // If both nodes have notes, do not overwrite any pre-existing notes.
1287 Node_Notes* source_notes = node_notes_at(source->_idx);
1288 if (source_notes == nullptr || source_notes->is_clear()) return false;
1289 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1290 if (dest_notes == nullptr || dest_notes->is_clear()) {
1291 return set_node_notes_at(dest->_idx, source_notes);
1292 }
1293
1294 Node_Notes merged_notes = (*source_notes);
1295 // The order of operations here ensures that dest notes will win...
1296 merged_notes.update_from(dest_notes);
1297 return set_node_notes_at(dest->_idx, &merged_notes);
1298 }
1299
1300
1301 //--------------------------allow_range_check_smearing-------------------------
1302 // Gating condition for coalescing similar range checks.
1303 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1304 // single covering check that is at least as strong as any of them.
1305 // If the optimization succeeds, the simplified (strengthened) range check
1306 // will always succeed. If it fails, we will deopt, and then give up
1307 // on the optimization.
1308 bool Compile::allow_range_check_smearing() const {
1309 // If this method has already thrown a range-check,
1310 // assume it was because we already tried range smearing
1311 // and it failed.
1312 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1313 return !already_trapped;
1314 }
1315
1316
1317 //------------------------------flatten_alias_type-----------------------------
1318 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1319 assert(do_aliasing(), "Aliasing should be enabled");
1320 int offset = tj->offset();
1321 TypePtr::PTR ptr = tj->ptr();
1322
1323 // Known instance (scalarizable allocation) alias only with itself.
1324 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1325 tj->is_oopptr()->is_known_instance();
1326
1327 // Process weird unsafe references.
1328 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1329 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1330 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1331 tj = TypeOopPtr::BOTTOM;
1332 ptr = tj->ptr();
1333 offset = tj->offset();
1334 }
1335
1336 // Array pointers need some flattening
1337 const TypeAryPtr* ta = tj->isa_aryptr();
1338 if (ta && ta->is_stable()) {
1339 // Erase stability property for alias analysis.
1340 tj = ta = ta->cast_to_stable(false);
1341 }
1342 if( ta && is_known_inst ) {
1343 if ( offset != Type::OffsetBot &&
1344 offset > arrayOopDesc::length_offset_in_bytes() ) {
1345 offset = Type::OffsetBot; // Flatten constant access into array body only
1346 tj = ta = ta->
1347 remove_speculative()->
1348 cast_to_ptr_type(ptr)->
1349 with_offset(offset);
1350 }
1351 } else if (ta) {
1352 // For arrays indexed by constant indices, we flatten the alias
1353 // space to include all of the array body. Only the header, klass
1354 // and array length can be accessed un-aliased.
1355 if( offset != Type::OffsetBot ) {
1356 if( ta->const_oop() ) { // MethodData* or Method*
1357 offset = Type::OffsetBot; // Flatten constant access into array body
1358 tj = ta = ta->
1359 remove_speculative()->
1360 cast_to_ptr_type(ptr)->
1361 cast_to_exactness(false)->
1362 with_offset(offset);
1363 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1364 // range is OK as-is.
1365 tj = ta = TypeAryPtr::RANGE;
1366 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1367 tj = TypeInstPtr::KLASS; // all klass loads look alike
1368 ta = TypeAryPtr::RANGE; // generic ignored junk
1369 ptr = TypePtr::BotPTR;
1370 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1371 tj = TypeInstPtr::MARK;
1372 ta = TypeAryPtr::RANGE; // generic ignored junk
1373 ptr = TypePtr::BotPTR;
1374 } else { // Random constant offset into array body
1375 offset = Type::OffsetBot; // Flatten constant access into array body
1376 tj = ta = ta->
1377 remove_speculative()->
1378 cast_to_ptr_type(ptr)->
1379 cast_to_exactness(false)->
1380 with_offset(offset);
1381 }
1382 }
1383 // Arrays of fixed size alias with arrays of unknown size.
1384 if (ta->size() != TypeInt::POS) {
1385 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1386 tj = ta = ta->
1387 remove_speculative()->
1388 cast_to_ptr_type(ptr)->
1389 with_ary(tary)->
1390 cast_to_exactness(false);
1391 }
1392 // Arrays of known objects become arrays of unknown objects.
1393 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1394 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1395 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1396 }
1397 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1398 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1399 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1400 }
1401 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1402 // cannot be distinguished by bytecode alone.
1403 if (ta->elem() == TypeInt::BOOL) {
1404 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1405 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1406 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1407 }
1408 // During the 2nd round of IterGVN, NotNull castings are removed.
1409 // Make sure the Bottom and NotNull variants alias the same.
1410 // Also, make sure exact and non-exact variants alias the same.
1411 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1412 tj = ta = ta->
1413 remove_speculative()->
1414 cast_to_ptr_type(TypePtr::BotPTR)->
1415 cast_to_exactness(false)->
1416 with_offset(offset);
1417 }
1418 }
1419
1420 // Oop pointers need some flattening
1421 const TypeInstPtr *to = tj->isa_instptr();
1422 if (to && to != TypeOopPtr::BOTTOM) {
1423 ciInstanceKlass* ik = to->instance_klass();
1424 if( ptr == TypePtr::Constant ) {
1425 if (ik != ciEnv::current()->Class_klass() ||
1426 offset < ik->layout_helper_size_in_bytes()) {
1427 // No constant oop pointers (such as Strings); they alias with
1428 // unknown strings.
1429 assert(!is_known_inst, "not scalarizable allocation");
1430 tj = to = to->
1431 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1432 remove_speculative()->
1433 cast_to_ptr_type(TypePtr::BotPTR)->
1434 cast_to_exactness(false);
1435 }
1436 } else if( is_known_inst ) {
1437 tj = to; // Keep NotNull and klass_is_exact for instance type
1438 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1439 // During the 2nd round of IterGVN, NotNull castings are removed.
1440 // Make sure the Bottom and NotNull variants alias the same.
1441 // Also, make sure exact and non-exact variants alias the same.
1442 tj = to = to->
1443 remove_speculative()->
1444 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1445 cast_to_ptr_type(TypePtr::BotPTR)->
1446 cast_to_exactness(false);
1447 }
1448 if (to->speculative() != nullptr) {
1449 tj = to = to->remove_speculative();
1450 }
1451 // Canonicalize the holder of this field
1452 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1453 // First handle header references such as a LoadKlassNode, even if the
1454 // object's klass is unloaded at compile time (4965979).
1455 if (!is_known_inst) { // Do it only for non-instance types
1456 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1457 }
1458 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1459 // Static fields are in the space above the normal instance
1460 // fields in the java.lang.Class instance.
1461 if (ik != ciEnv::current()->Class_klass()) {
1462 to = nullptr;
1463 tj = TypeOopPtr::BOTTOM;
1464 offset = tj->offset();
1465 }
1466 } else {
1467 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1468 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1469 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1470 if( is_known_inst ) {
1471 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id());
1472 } else {
1473 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset);
1474 }
1475 }
1476 }
1477 }
1478
1479 // Klass pointers to object array klasses need some flattening
1480 const TypeKlassPtr *tk = tj->isa_klassptr();
1481 if( tk ) {
1482 // If we are referencing a field within a Klass, we need
1483 // to assume the worst case of an Object. Both exact and
1484 // inexact types must flatten to the same alias class so
1485 // use NotNull as the PTR.
1486 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1487 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1488 env()->Object_klass(),
1489 offset);
1490 }
1491
1492 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1493 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1494 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1495 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1496 } else {
1497 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1498 }
1499 }
1500
1501 // Check for precise loads from the primary supertype array and force them
1502 // to the supertype cache alias index. Check for generic array loads from
1503 // the primary supertype array and also force them to the supertype cache
1504 // alias index. Since the same load can reach both, we need to merge
1505 // these 2 disparate memories into the same alias class. Since the
1506 // primary supertype array is read-only, there's no chance of confusion
1507 // where we bypass an array load and an array store.
1508 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1509 if (offset == Type::OffsetBot ||
1510 (offset >= primary_supers_offset &&
1511 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1512 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1513 offset = in_bytes(Klass::secondary_super_cache_offset());
1514 tj = tk = tk->with_offset(offset);
1515 }
1516 }
1517
1518 // Flatten all Raw pointers together.
1519 if (tj->base() == Type::RawPtr)
1520 tj = TypeRawPtr::BOTTOM;
1521
1522 if (tj->base() == Type::AnyPtr)
1523 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1524
1525 offset = tj->offset();
1526 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1527
1528 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1529 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1530 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1531 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1532 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1533 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1534 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1535 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1536 assert( tj->ptr() != TypePtr::TopPTR &&
1537 tj->ptr() != TypePtr::AnyNull &&
1538 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1539 // assert( tj->ptr() != TypePtr::Constant ||
1540 // tj->base() == Type::RawPtr ||
1541 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1542
1543 return tj;
1544 }
1545
1546 void Compile::AliasType::Init(int i, const TypePtr* at) {
1547 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1548 _index = i;
1549 _adr_type = at;
1550 _field = nullptr;
1551 _element = nullptr;
1552 _is_rewritable = true; // default
1553 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1554 if (atoop != nullptr && atoop->is_known_instance()) {
1555 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1556 _general_index = Compile::current()->get_alias_index(gt);
1557 } else {
1558 _general_index = 0;
1559 }
1560 }
1561
1562 BasicType Compile::AliasType::basic_type() const {
1563 if (element() != nullptr) {
1564 const Type* element = adr_type()->is_aryptr()->elem();
1565 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1566 } if (field() != nullptr) {
1567 return field()->layout_type();
1568 } else {
1569 return T_ILLEGAL; // unknown
1570 }
1571 }
1572
1573 //---------------------------------print_on------------------------------------
1574 #ifndef PRODUCT
1575 void Compile::AliasType::print_on(outputStream* st) {
1576 if (index() < 10)
1577 st->print("@ <%d> ", index());
1578 else st->print("@ <%d>", index());
1579 st->print(is_rewritable() ? " " : " RO");
1580 int offset = adr_type()->offset();
1581 if (offset == Type::OffsetBot)
1582 st->print(" +any");
1583 else st->print(" +%-3d", offset);
1584 st->print(" in ");
1585 adr_type()->dump_on(st);
1586 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1587 if (field() != nullptr && tjp) {
1588 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1589 tjp->offset() != field()->offset_in_bytes()) {
1590 st->print(" != ");
1591 field()->print();
1592 st->print(" ***");
1593 }
1594 }
1595 }
1596
1597 void print_alias_types() {
1598 Compile* C = Compile::current();
1599 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1600 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1601 C->alias_type(idx)->print_on(tty);
1602 tty->cr();
1603 }
1604 }
1605 #endif
1606
1607
1608 //----------------------------probe_alias_cache--------------------------------
1609 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1610 intptr_t key = (intptr_t) adr_type;
1611 key ^= key >> logAliasCacheSize;
1612 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1613 }
1614
1615
1616 //-----------------------------grow_alias_types--------------------------------
1617 void Compile::grow_alias_types() {
1618 const int old_ats = _max_alias_types; // how many before?
1619 const int new_ats = old_ats; // how many more?
1620 const int grow_ats = old_ats+new_ats; // how many now?
1621 _max_alias_types = grow_ats;
1622 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1623 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1624 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1625 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1626 }
1627
1628
1629 //--------------------------------find_alias_type------------------------------
1630 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1631 if (!do_aliasing()) {
1632 return alias_type(AliasIdxBot);
1633 }
1634
1635 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1636 if (ace->_adr_type == adr_type) {
1637 return alias_type(ace->_index);
1638 }
1639
1640 // Handle special cases.
1641 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1642 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1643
1644 // Do it the slow way.
1645 const TypePtr* flat = flatten_alias_type(adr_type);
1646
1647 #ifdef ASSERT
1648 {
1649 ResourceMark rm;
1650 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1651 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1652 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1653 Type::str(adr_type));
1654 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1655 const TypeOopPtr* foop = flat->is_oopptr();
1656 // Scalarizable allocations have exact klass always.
1657 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1658 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1659 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1660 Type::str(foop), Type::str(xoop));
1661 }
1662 }
1663 #endif
1664
1665 int idx = AliasIdxTop;
1666 for (int i = 0; i < num_alias_types(); i++) {
1667 if (alias_type(i)->adr_type() == flat) {
1668 idx = i;
1669 break;
1670 }
1671 }
1672
1673 if (idx == AliasIdxTop) {
1674 if (no_create) return nullptr;
1675 // Grow the array if necessary.
1676 if (_num_alias_types == _max_alias_types) grow_alias_types();
1677 // Add a new alias type.
1678 idx = _num_alias_types++;
1679 _alias_types[idx]->Init(idx, flat);
1680 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1681 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1682 if (flat->isa_instptr()) {
1683 if (flat->offset() == java_lang_Class::klass_offset()
1684 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1685 alias_type(idx)->set_rewritable(false);
1686 }
1687 if (flat->isa_aryptr()) {
1688 #ifdef ASSERT
1689 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1690 // (T_BYTE has the weakest alignment and size restrictions...)
1691 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1692 #endif
1693 if (flat->offset() == TypePtr::OffsetBot) {
1694 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1695 }
1696 }
1697 if (flat->isa_klassptr()) {
1698 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1699 alias_type(idx)->set_rewritable(false);
1700 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1701 alias_type(idx)->set_rewritable(false);
1702 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1703 alias_type(idx)->set_rewritable(false);
1704 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1705 alias_type(idx)->set_rewritable(false);
1706 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1707 alias_type(idx)->set_rewritable(false);
1708 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1709 alias_type(idx)->set_rewritable(false);
1710 }
1711 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1712 // but the base pointer type is not distinctive enough to identify
1713 // references into JavaThread.)
1714
1715 // Check for final fields.
1716 const TypeInstPtr* tinst = flat->isa_instptr();
1717 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1718 ciField* field;
1719 if (tinst->const_oop() != nullptr &&
1720 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1721 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1722 // static field
1723 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1724 field = k->get_field_by_offset(tinst->offset(), true);
1725 } else {
1726 ciInstanceKlass *k = tinst->instance_klass();
1727 field = k->get_field_by_offset(tinst->offset(), false);
1728 }
1729 assert(field == nullptr ||
1730 original_field == nullptr ||
1731 (field->holder() == original_field->holder() &&
1732 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1733 field->is_static() == original_field->is_static()), "wrong field?");
1734 // Set field() and is_rewritable() attributes.
1735 if (field != nullptr) alias_type(idx)->set_field(field);
1736 }
1737 }
1738
1739 // Fill the cache for next time.
1740 ace->_adr_type = adr_type;
1741 ace->_index = idx;
1742 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1743
1744 // Might as well try to fill the cache for the flattened version, too.
1745 AliasCacheEntry* face = probe_alias_cache(flat);
1746 if (face->_adr_type == nullptr) {
1747 face->_adr_type = flat;
1748 face->_index = idx;
1749 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1750 }
1751
1752 return alias_type(idx);
1753 }
1754
1755
1756 Compile::AliasType* Compile::alias_type(ciField* field) {
1757 const TypeOopPtr* t;
1758 if (field->is_static())
1759 t = TypeInstPtr::make(field->holder()->java_mirror());
1760 else
1761 t = TypeOopPtr::make_from_klass_raw(field->holder());
1762 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1763 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1764 return atp;
1765 }
1766
1767
1768 //------------------------------have_alias_type--------------------------------
1769 bool Compile::have_alias_type(const TypePtr* adr_type) {
1770 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1771 if (ace->_adr_type == adr_type) {
1772 return true;
1773 }
1774
1775 // Handle special cases.
1776 if (adr_type == nullptr) return true;
1777 if (adr_type == TypePtr::BOTTOM) return true;
1778
1779 return find_alias_type(adr_type, true, nullptr) != nullptr;
1780 }
1781
1782 //-----------------------------must_alias--------------------------------------
1783 // True if all values of the given address type are in the given alias category.
1784 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1785 if (alias_idx == AliasIdxBot) return true; // the universal category
1786 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1787 if (alias_idx == AliasIdxTop) return false; // the empty category
1788 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1789
1790 // the only remaining possible overlap is identity
1791 int adr_idx = get_alias_index(adr_type);
1792 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1793 assert(adr_idx == alias_idx ||
1794 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1795 && adr_type != TypeOopPtr::BOTTOM),
1796 "should not be testing for overlap with an unsafe pointer");
1797 return adr_idx == alias_idx;
1798 }
1799
1800 //------------------------------can_alias--------------------------------------
1801 // True if any values of the given address type are in the given alias category.
1802 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1803 if (alias_idx == AliasIdxTop) return false; // the empty category
1804 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1805 // Known instance doesn't alias with bottom memory
1806 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1807 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1808
1809 // the only remaining possible overlap is identity
1810 int adr_idx = get_alias_index(adr_type);
1811 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1812 return adr_idx == alias_idx;
1813 }
1814
1815 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1816 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1817 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1818 if (parse_predicate_count() == 0) {
1819 return;
1820 }
1821 for (int i = 0; i < parse_predicate_count(); i++) {
1822 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1823 parse_predicate->mark_useless();
1824 igvn._worklist.push(parse_predicate);
1825 }
1826 _parse_predicates.clear();
1827 }
1828
1829 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1830 if (!n->for_post_loop_opts_igvn()) {
1831 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1832 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1833 _for_post_loop_igvn.append(n);
1834 }
1835 }
1836
1837 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1838 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1839 _for_post_loop_igvn.remove(n);
1840 }
1841
1842 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1843 // Verify that all previous optimizations produced a valid graph
1844 // at least to this point, even if no loop optimizations were done.
1845 PhaseIdealLoop::verify(igvn);
1846
1847 C->set_post_loop_opts_phase(); // no more loop opts allowed
1848
1849 assert(!C->major_progress(), "not cleared");
1850
1851 if (_for_post_loop_igvn.length() > 0) {
1852 while (_for_post_loop_igvn.length() > 0) {
1853 Node* n = _for_post_loop_igvn.pop();
1854 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1855 igvn._worklist.push(n);
1856 }
1857 igvn.optimize();
1858 if (failing()) return;
1859 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1860 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1861
1862 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1863 if (C->major_progress()) {
1864 C->clear_major_progress(); // ensure that major progress is now clear
1865 }
1866 }
1867 }
1868
1869 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1870 if (OptimizeUnstableIf) {
1871 _unstable_if_traps.append(trap);
1872 }
1873 }
1874
1875 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1876 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1877 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1878 Node* n = trap->uncommon_trap();
1879 if (!useful.member(n)) {
1880 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1881 }
1882 }
1883 }
1884
1885 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1886 // or fold-compares case. Return true if succeed or not found.
1887 //
1888 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1889 // false and ask fold-compares to yield.
1890 //
1891 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1892 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1893 // when deoptimization does happen.
1894 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1895 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1896 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1897 if (trap->uncommon_trap() == unc) {
1898 if (yield && trap->modified()) {
1899 return false;
1900 }
1901 _unstable_if_traps.delete_at(i);
1902 break;
1903 }
1904 }
1905 return true;
1906 }
1907
1908 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1909 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1910 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1911 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1912 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1913 CallStaticJavaNode* unc = trap->uncommon_trap();
1914 int next_bci = trap->next_bci();
1915 bool modified = trap->modified();
1916
1917 if (next_bci != -1 && !modified) {
1918 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1919 JVMState* jvms = unc->jvms();
1920 ciMethod* method = jvms->method();
1921 ciBytecodeStream iter(method);
1922
1923 iter.force_bci(jvms->bci());
1924 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1925 Bytecodes::Code c = iter.cur_bc();
1926 Node* lhs = nullptr;
1927 Node* rhs = nullptr;
1928 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1929 lhs = unc->peek_operand(0);
1930 rhs = unc->peek_operand(1);
1931 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1932 lhs = unc->peek_operand(0);
1933 }
1934
1935 ResourceMark rm;
1936 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1937 assert(live_locals.is_valid(), "broken liveness info");
1938 int len = (int)live_locals.size();
1939
1940 for (int i = 0; i < len; i++) {
1941 Node* local = unc->local(jvms, i);
1942 // kill local using the liveness of next_bci.
1943 // give up when the local looks like an operand to secure reexecution.
1944 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1945 uint idx = jvms->locoff() + i;
1946 #ifdef ASSERT
1947 if (PrintOpto && Verbose) {
1948 tty->print("[unstable_if] kill local#%d: ", idx);
1949 local->dump();
1950 tty->cr();
1951 }
1952 #endif
1953 igvn.replace_input_of(unc, idx, top());
1954 modified = true;
1955 }
1956 }
1957 }
1958
1959 // keep the mondified trap for late query
1960 if (modified) {
1961 trap->set_modified();
1962 } else {
1963 _unstable_if_traps.delete_at(i);
1964 }
1965 }
1966 igvn.optimize();
1967 }
1968
1969 // StringOpts and late inlining of string methods
1970 void Compile::inline_string_calls(bool parse_time) {
1971 {
1972 // remove useless nodes to make the usage analysis simpler
1973 ResourceMark rm;
1974 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1975 }
1976
1977 {
1978 ResourceMark rm;
1979 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1980 PhaseStringOpts pso(initial_gvn());
1981 print_method(PHASE_AFTER_STRINGOPTS, 3);
1982 }
1983
1984 // now inline anything that we skipped the first time around
1985 if (!parse_time) {
1986 _late_inlines_pos = _late_inlines.length();
1987 }
1988
1989 while (_string_late_inlines.length() > 0) {
1990 CallGenerator* cg = _string_late_inlines.pop();
1991 cg->do_late_inline();
1992 if (failing()) return;
1993 }
1994 _string_late_inlines.trunc_to(0);
1995 }
1996
1997 // Late inlining of boxing methods
1998 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1999 if (_boxing_late_inlines.length() > 0) {
2000 assert(has_boxed_value(), "inconsistent");
2001
2002 set_inlining_incrementally(true);
2003
2004 igvn_worklist()->ensure_empty(); // should be done with igvn
2005
2006 _late_inlines_pos = _late_inlines.length();
2007
2008 while (_boxing_late_inlines.length() > 0) {
2009 CallGenerator* cg = _boxing_late_inlines.pop();
2010 cg->do_late_inline();
2011 if (failing()) return;
2012 }
2013 _boxing_late_inlines.trunc_to(0);
2014
2015 inline_incrementally_cleanup(igvn);
2016
2017 set_inlining_incrementally(false);
2018 }
2019 }
2020
2021 bool Compile::inline_incrementally_one() {
2022 assert(IncrementalInline, "incremental inlining should be on");
2023
2024 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2025
2026 set_inlining_progress(false);
2027 set_do_cleanup(false);
2028
2029 for (int i = 0; i < _late_inlines.length(); i++) {
2030 _late_inlines_pos = i+1;
2031 CallGenerator* cg = _late_inlines.at(i);
2032 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2033 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2034 cg->do_late_inline();
2035 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2036 if (failing()) {
2037 return false;
2038 } else if (inlining_progress()) {
2039 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2040 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2041 break; // process one call site at a time
2042 }
2043 } else {
2044 // Ignore late inline direct calls when inlining is not allowed.
2045 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2046 }
2047 }
2048 // Remove processed elements.
2049 _late_inlines.remove_till(_late_inlines_pos);
2050 _late_inlines_pos = 0;
2051
2052 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2053
2054 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2055
2056 set_inlining_progress(false);
2057 set_do_cleanup(false);
2058
2059 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2060 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2061 }
2062
2063 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2064 {
2065 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2066 ResourceMark rm;
2067 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2068 }
2069 {
2070 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2071 igvn.reset_from_gvn(initial_gvn());
2072 igvn.optimize();
2073 if (failing()) return;
2074 }
2075 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2076 }
2077
2078 // Perform incremental inlining until bound on number of live nodes is reached
2079 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2080 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2081
2082 set_inlining_incrementally(true);
2083 uint low_live_nodes = 0;
2084
2085 while (_late_inlines.length() > 0) {
2086 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2087 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2088 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2089 // PhaseIdealLoop is expensive so we only try it once we are
2090 // out of live nodes and we only try it again if the previous
2091 // helped got the number of nodes down significantly
2092 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2093 if (failing()) return;
2094 low_live_nodes = live_nodes();
2095 _major_progress = true;
2096 }
2097
2098 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2099 bool do_print_inlining = print_inlining() || print_intrinsics();
2100 if (do_print_inlining || log() != nullptr) {
2101 // Print inlining message for candidates that we couldn't inline for lack of space.
2102 for (int i = 0; i < _late_inlines.length(); i++) {
2103 CallGenerator* cg = _late_inlines.at(i);
2104 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2105 if (do_print_inlining) {
2106 cg->print_inlining_late(InliningResult::FAILURE, msg);
2107 }
2108 log_late_inline_failure(cg, msg);
2109 }
2110 }
2111 break; // finish
2112 }
2113 }
2114
2115 igvn_worklist()->ensure_empty(); // should be done with igvn
2116
2117 while (inline_incrementally_one()) {
2118 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2119 }
2120 if (failing()) return;
2121
2122 inline_incrementally_cleanup(igvn);
2123
2124 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2125
2126 if (failing()) return;
2127
2128 if (_late_inlines.length() == 0) {
2129 break; // no more progress
2130 }
2131 }
2132
2133 igvn_worklist()->ensure_empty(); // should be done with igvn
2134
2135 if (_string_late_inlines.length() > 0) {
2136 assert(has_stringbuilder(), "inconsistent");
2137
2138 inline_string_calls(false);
2139
2140 if (failing()) return;
2141
2142 inline_incrementally_cleanup(igvn);
2143 }
2144
2145 set_inlining_incrementally(false);
2146 }
2147
2148 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2149 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2150 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2151 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2152 // as if "inlining_incrementally() == true" were set.
2153 assert(inlining_incrementally() == false, "not allowed");
2154 assert(_modified_nodes == nullptr, "not allowed");
2155 assert(_late_inlines.length() > 0, "sanity");
2156
2157 while (_late_inlines.length() > 0) {
2158 igvn_worklist()->ensure_empty(); // should be done with igvn
2159
2160 while (inline_incrementally_one()) {
2161 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2162 }
2163 if (failing()) return;
2164
2165 inline_incrementally_cleanup(igvn);
2166 }
2167 }
2168
2169 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2170 if (_loop_opts_cnt > 0) {
2171 while (major_progress() && (_loop_opts_cnt > 0)) {
2172 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2173 PhaseIdealLoop::optimize(igvn, mode);
2174 _loop_opts_cnt--;
2175 if (failing()) return false;
2176 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2177 }
2178 }
2179 return true;
2180 }
2181
2182 // Remove edges from "root" to each SafePoint at a backward branch.
2183 // They were inserted during parsing (see add_safepoint()) to make
2184 // infinite loops without calls or exceptions visible to root, i.e.,
2185 // useful.
2186 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2187 Node *r = root();
2188 if (r != nullptr) {
2189 for (uint i = r->req(); i < r->len(); ++i) {
2190 Node *n = r->in(i);
2191 if (n != nullptr && n->is_SafePoint()) {
2192 r->rm_prec(i);
2193 if (n->outcnt() == 0) {
2194 igvn.remove_dead_node(n);
2195 }
2196 --i;
2197 }
2198 }
2199 // Parsing may have added top inputs to the root node (Path
2200 // leading to the Halt node proven dead). Make sure we get a
2201 // chance to clean them up.
2202 igvn._worklist.push(r);
2203 igvn.optimize();
2204 }
2205 }
2206
2207 //------------------------------Optimize---------------------------------------
2208 // Given a graph, optimize it.
2209 void Compile::Optimize() {
2210 TracePhase tp("optimizer", &timers[_t_optimizer]);
2211
2212 #ifndef PRODUCT
2213 if (env()->break_at_compile()) {
2214 BREAKPOINT;
2215 }
2216
2217 #endif
2218
2219 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2220 #ifdef ASSERT
2221 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2222 #endif
2223
2224 ResourceMark rm;
2225
2226 print_inlining_reinit();
2227
2228 NOT_PRODUCT( verify_graph_edges(); )
2229
2230 print_method(PHASE_AFTER_PARSING, 1);
2231
2232 {
2233 // Iterative Global Value Numbering, including ideal transforms
2234 // Initialize IterGVN with types and values from parse-time GVN
2235 PhaseIterGVN igvn(initial_gvn());
2236 #ifdef ASSERT
2237 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2238 #endif
2239 {
2240 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2241 igvn.optimize();
2242 }
2243
2244 if (failing()) return;
2245
2246 print_method(PHASE_ITER_GVN1, 2);
2247
2248 process_for_unstable_if_traps(igvn);
2249
2250 if (failing()) return;
2251
2252 inline_incrementally(igvn);
2253
2254 print_method(PHASE_INCREMENTAL_INLINE, 2);
2255
2256 if (failing()) return;
2257
2258 if (eliminate_boxing()) {
2259 // Inline valueOf() methods now.
2260 inline_boxing_calls(igvn);
2261
2262 if (failing()) return;
2263
2264 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2265 inline_incrementally(igvn);
2266 }
2267
2268 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2269
2270 if (failing()) return;
2271 }
2272
2273 // Remove the speculative part of types and clean up the graph from
2274 // the extra CastPP nodes whose only purpose is to carry them. Do
2275 // that early so that optimizations are not disrupted by the extra
2276 // CastPP nodes.
2277 remove_speculative_types(igvn);
2278
2279 if (failing()) return;
2280
2281 // No more new expensive nodes will be added to the list from here
2282 // so keep only the actual candidates for optimizations.
2283 cleanup_expensive_nodes(igvn);
2284
2285 if (failing()) return;
2286
2287 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2288 if (EnableVectorSupport && has_vbox_nodes()) {
2289 TracePhase tp("", &timers[_t_vector]);
2290 PhaseVector pv(igvn);
2291 pv.optimize_vector_boxes();
2292 if (failing()) return;
2293 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2294 }
2295 assert(!has_vbox_nodes(), "sanity");
2296
2297 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2298 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2299 igvn_worklist()->ensure_empty(); // should be done with igvn
2300 {
2301 ResourceMark rm;
2302 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2303 }
2304 igvn.reset_from_gvn(initial_gvn());
2305 igvn.optimize();
2306 if (failing()) return;
2307 }
2308
2309 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2310 // safepoints
2311 remove_root_to_sfpts_edges(igvn);
2312
2313 if (failing()) return;
2314
2315 // Perform escape analysis
2316 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2317 if (has_loops()) {
2318 // Cleanup graph (remove dead nodes).
2319 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2320 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2321 if (failing()) return;
2322 }
2323 bool progress;
2324 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2325 do {
2326 ConnectionGraph::do_analysis(this, &igvn);
2327
2328 if (failing()) return;
2329
2330 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2331
2332 // Optimize out fields loads from scalar replaceable allocations.
2333 igvn.optimize();
2334 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2335
2336 if (failing()) return;
2337
2338 if (congraph() != nullptr && macro_count() > 0) {
2339 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2340 PhaseMacroExpand mexp(igvn);
2341 mexp.eliminate_macro_nodes();
2342 if (failing()) return;
2343
2344 igvn.set_delay_transform(false);
2345 igvn.optimize();
2346 if (failing()) return;
2347
2348 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2349 }
2350
2351 ConnectionGraph::verify_ram_nodes(this, root());
2352 if (failing()) return;
2353
2354 progress = do_iterative_escape_analysis() &&
2355 (macro_count() < mcount) &&
2356 ConnectionGraph::has_candidates(this);
2357 // Try again if candidates exist and made progress
2358 // by removing some allocations and/or locks.
2359 } while (progress);
2360 }
2361
2362 // Loop transforms on the ideal graph. Range Check Elimination,
2363 // peeling, unrolling, etc.
2364
2365 // Set loop opts counter
2366 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2367 {
2368 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2369 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2370 _loop_opts_cnt--;
2371 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2372 if (failing()) return;
2373 }
2374 // Loop opts pass if partial peeling occurred in previous pass
2375 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2376 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2377 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2378 _loop_opts_cnt--;
2379 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2380 if (failing()) return;
2381 }
2382 // Loop opts pass for loop-unrolling before CCP
2383 if(major_progress() && (_loop_opts_cnt > 0)) {
2384 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2385 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2386 _loop_opts_cnt--;
2387 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2388 }
2389 if (!failing()) {
2390 // Verify that last round of loop opts produced a valid graph
2391 PhaseIdealLoop::verify(igvn);
2392 }
2393 }
2394 if (failing()) return;
2395
2396 // Conditional Constant Propagation;
2397 print_method(PHASE_BEFORE_CCP1, 2);
2398 PhaseCCP ccp( &igvn );
2399 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2400 {
2401 TracePhase tp("ccp", &timers[_t_ccp]);
2402 ccp.do_transform();
2403 }
2404 print_method(PHASE_CCP1, 2);
2405
2406 assert( true, "Break here to ccp.dump_old2new_map()");
2407
2408 // Iterative Global Value Numbering, including ideal transforms
2409 {
2410 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2411 igvn.reset_from_igvn(&ccp);
2412 igvn.optimize();
2413 }
2414 print_method(PHASE_ITER_GVN2, 2);
2415
2416 if (failing()) return;
2417
2418 // Loop transforms on the ideal graph. Range Check Elimination,
2419 // peeling, unrolling, etc.
2420 if (!optimize_loops(igvn, LoopOptsDefault)) {
2421 return;
2422 }
2423
2424 if (failing()) return;
2425
2426 C->clear_major_progress(); // ensure that major progress is now clear
2427
2428 process_for_post_loop_opts_igvn(igvn);
2429
2430 if (failing()) return;
2431
2432 #ifdef ASSERT
2433 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2434 #endif
2435
2436 {
2437 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2438 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2439 PhaseMacroExpand mex(igvn);
2440 if (mex.expand_macro_nodes()) {
2441 assert(failing(), "must bail out w/ explicit message");
2442 return;
2443 }
2444 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2445 }
2446
2447 {
2448 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2449 if (bs->expand_barriers(this, igvn)) {
2450 assert(failing(), "must bail out w/ explicit message");
2451 return;
2452 }
2453 print_method(PHASE_BARRIER_EXPANSION, 2);
2454 }
2455
2456 if (C->max_vector_size() > 0) {
2457 C->optimize_logic_cones(igvn);
2458 igvn.optimize();
2459 if (failing()) return;
2460 }
2461
2462 DEBUG_ONLY( _modified_nodes = nullptr; )
2463
2464 assert(igvn._worklist.size() == 0, "not empty");
2465
2466 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2467
2468 if (_late_inlines.length() > 0) {
2469 // More opportunities to optimize virtual and MH calls.
2470 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2471 process_late_inline_calls_no_inline(igvn);
2472 if (failing()) return;
2473 }
2474 } // (End scope of igvn; run destructor if necessary for asserts.)
2475
2476 check_no_dead_use();
2477
2478 process_print_inlining();
2479
2480 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2481 // to remove hashes to unlock nodes for modifications.
2482 C->node_hash()->clear();
2483
2484 // A method with only infinite loops has no edges entering loops from root
2485 {
2486 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2487 if (final_graph_reshaping()) {
2488 assert(failing(), "must bail out w/ explicit message");
2489 return;
2490 }
2491 }
2492
2493 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2494 DEBUG_ONLY(set_phase_optimize_finished();)
2495 }
2496
2497 #ifdef ASSERT
2498 void Compile::check_no_dead_use() const {
2499 ResourceMark rm;
2500 Unique_Node_List wq;
2501 wq.push(root());
2502 for (uint i = 0; i < wq.size(); ++i) {
2503 Node* n = wq.at(i);
2504 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2505 Node* u = n->fast_out(j);
2506 if (u->outcnt() == 0 && !u->is_Con()) {
2507 u->dump();
2508 fatal("no reachable node should have no use");
2509 }
2510 wq.push(u);
2511 }
2512 }
2513 }
2514 #endif
2515
2516 void Compile::inline_vector_reboxing_calls() {
2517 if (C->_vector_reboxing_late_inlines.length() > 0) {
2518 _late_inlines_pos = C->_late_inlines.length();
2519 while (_vector_reboxing_late_inlines.length() > 0) {
2520 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2521 cg->do_late_inline();
2522 if (failing()) return;
2523 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2524 }
2525 _vector_reboxing_late_inlines.trunc_to(0);
2526 }
2527 }
2528
2529 bool Compile::has_vbox_nodes() {
2530 if (C->_vector_reboxing_late_inlines.length() > 0) {
2531 return true;
2532 }
2533 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2534 Node * n = C->macro_node(macro_idx);
2535 assert(n->is_macro(), "only macro nodes expected here");
2536 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2537 return true;
2538 }
2539 }
2540 return false;
2541 }
2542
2543 //---------------------------- Bitwise operation packing optimization ---------------------------
2544
2545 static bool is_vector_unary_bitwise_op(Node* n) {
2546 return n->Opcode() == Op_XorV &&
2547 VectorNode::is_vector_bitwise_not_pattern(n);
2548 }
2549
2550 static bool is_vector_binary_bitwise_op(Node* n) {
2551 switch (n->Opcode()) {
2552 case Op_AndV:
2553 case Op_OrV:
2554 return true;
2555
2556 case Op_XorV:
2557 return !is_vector_unary_bitwise_op(n);
2558
2559 default:
2560 return false;
2561 }
2562 }
2563
2564 static bool is_vector_ternary_bitwise_op(Node* n) {
2565 return n->Opcode() == Op_MacroLogicV;
2566 }
2567
2568 static bool is_vector_bitwise_op(Node* n) {
2569 return is_vector_unary_bitwise_op(n) ||
2570 is_vector_binary_bitwise_op(n) ||
2571 is_vector_ternary_bitwise_op(n);
2572 }
2573
2574 static bool is_vector_bitwise_cone_root(Node* n) {
2575 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2576 return false;
2577 }
2578 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2579 if (is_vector_bitwise_op(n->fast_out(i))) {
2580 return false;
2581 }
2582 }
2583 return true;
2584 }
2585
2586 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2587 uint cnt = 0;
2588 if (is_vector_bitwise_op(n)) {
2589 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2590 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2591 for (uint i = 1; i < inp_cnt; i++) {
2592 Node* in = n->in(i);
2593 bool skip = VectorNode::is_all_ones_vector(in);
2594 if (!skip && !inputs.member(in)) {
2595 inputs.push(in);
2596 cnt++;
2597 }
2598 }
2599 assert(cnt <= 1, "not unary");
2600 } else {
2601 uint last_req = inp_cnt;
2602 if (is_vector_ternary_bitwise_op(n)) {
2603 last_req = inp_cnt - 1; // skip last input
2604 }
2605 for (uint i = 1; i < last_req; i++) {
2606 Node* def = n->in(i);
2607 if (!inputs.member(def)) {
2608 inputs.push(def);
2609 cnt++;
2610 }
2611 }
2612 }
2613 } else { // not a bitwise operations
2614 if (!inputs.member(n)) {
2615 inputs.push(n);
2616 cnt++;
2617 }
2618 }
2619 return cnt;
2620 }
2621
2622 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2623 Unique_Node_List useful_nodes;
2624 C->identify_useful_nodes(useful_nodes);
2625
2626 for (uint i = 0; i < useful_nodes.size(); i++) {
2627 Node* n = useful_nodes.at(i);
2628 if (is_vector_bitwise_cone_root(n)) {
2629 list.push(n);
2630 }
2631 }
2632 }
2633
2634 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2635 const TypeVect* vt,
2636 Unique_Node_List& partition,
2637 Unique_Node_List& inputs) {
2638 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2639 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2640 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2641
2642 Node* in1 = inputs.at(0);
2643 Node* in2 = inputs.at(1);
2644 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2645
2646 uint func = compute_truth_table(partition, inputs);
2647
2648 Node* pn = partition.at(partition.size() - 1);
2649 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2650 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2651 }
2652
2653 static uint extract_bit(uint func, uint pos) {
2654 return (func & (1 << pos)) >> pos;
2655 }
2656
2657 //
2658 // A macro logic node represents a truth table. It has 4 inputs,
2659 // First three inputs corresponds to 3 columns of a truth table
2660 // and fourth input captures the logic function.
2661 //
2662 // eg. fn = (in1 AND in2) OR in3;
2663 //
2664 // MacroNode(in1,in2,in3,fn)
2665 //
2666 // -----------------
2667 // in1 in2 in3 fn
2668 // -----------------
2669 // 0 0 0 0
2670 // 0 0 1 1
2671 // 0 1 0 0
2672 // 0 1 1 1
2673 // 1 0 0 0
2674 // 1 0 1 1
2675 // 1 1 0 1
2676 // 1 1 1 1
2677 //
2678
2679 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2680 int res = 0;
2681 for (int i = 0; i < 8; i++) {
2682 int bit1 = extract_bit(in1, i);
2683 int bit2 = extract_bit(in2, i);
2684 int bit3 = extract_bit(in3, i);
2685
2686 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2687 int func_bit = extract_bit(func, func_bit_pos);
2688
2689 res |= func_bit << i;
2690 }
2691 return res;
2692 }
2693
2694 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2695 assert(n != nullptr, "");
2696 assert(eval_map.contains(n), "absent");
2697 return *(eval_map.get(n));
2698 }
2699
2700 static void eval_operands(Node* n,
2701 uint& func1, uint& func2, uint& func3,
2702 ResourceHashtable<Node*,uint>& eval_map) {
2703 assert(is_vector_bitwise_op(n), "");
2704
2705 if (is_vector_unary_bitwise_op(n)) {
2706 Node* opnd = n->in(1);
2707 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2708 opnd = n->in(2);
2709 }
2710 func1 = eval_operand(opnd, eval_map);
2711 } else if (is_vector_binary_bitwise_op(n)) {
2712 func1 = eval_operand(n->in(1), eval_map);
2713 func2 = eval_operand(n->in(2), eval_map);
2714 } else {
2715 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2716 func1 = eval_operand(n->in(1), eval_map);
2717 func2 = eval_operand(n->in(2), eval_map);
2718 func3 = eval_operand(n->in(3), eval_map);
2719 }
2720 }
2721
2722 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2723 assert(inputs.size() <= 3, "sanity");
2724 ResourceMark rm;
2725 uint res = 0;
2726 ResourceHashtable<Node*,uint> eval_map;
2727
2728 // Populate precomputed functions for inputs.
2729 // Each input corresponds to one column of 3 input truth-table.
2730 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2731 0xCC, // (_, b, _) -> b
2732 0xF0 }; // (a, _, _) -> a
2733 for (uint i = 0; i < inputs.size(); i++) {
2734 eval_map.put(inputs.at(i), input_funcs[2-i]);
2735 }
2736
2737 for (uint i = 0; i < partition.size(); i++) {
2738 Node* n = partition.at(i);
2739
2740 uint func1 = 0, func2 = 0, func3 = 0;
2741 eval_operands(n, func1, func2, func3, eval_map);
2742
2743 switch (n->Opcode()) {
2744 case Op_OrV:
2745 assert(func3 == 0, "not binary");
2746 res = func1 | func2;
2747 break;
2748 case Op_AndV:
2749 assert(func3 == 0, "not binary");
2750 res = func1 & func2;
2751 break;
2752 case Op_XorV:
2753 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2754 assert(func2 == 0 && func3 == 0, "not unary");
2755 res = (~func1) & 0xFF;
2756 } else {
2757 assert(func3 == 0, "not binary");
2758 res = func1 ^ func2;
2759 }
2760 break;
2761 case Op_MacroLogicV:
2762 // Ordering of inputs may change during evaluation of sub-tree
2763 // containing MacroLogic node as a child node, thus a re-evaluation
2764 // makes sure that function is evaluated in context of current
2765 // inputs.
2766 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2767 break;
2768
2769 default: assert(false, "not supported: %s", n->Name());
2770 }
2771 assert(res <= 0xFF, "invalid");
2772 eval_map.put(n, res);
2773 }
2774 return res;
2775 }
2776
2777 // Criteria under which nodes gets packed into a macro logic node:-
2778 // 1) Parent and both child nodes are all unmasked or masked with
2779 // same predicates.
2780 // 2) Masked parent can be packed with left child if it is predicated
2781 // and both have same predicates.
2782 // 3) Masked parent can be packed with right child if its un-predicated
2783 // or has matching predication condition.
2784 // 4) An unmasked parent can be packed with an unmasked child.
2785 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2786 assert(partition.size() == 0, "not empty");
2787 assert(inputs.size() == 0, "not empty");
2788 if (is_vector_ternary_bitwise_op(n)) {
2789 return false;
2790 }
2791
2792 bool is_unary_op = is_vector_unary_bitwise_op(n);
2793 if (is_unary_op) {
2794 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2795 return false; // too few inputs
2796 }
2797
2798 bool pack_left_child = true;
2799 bool pack_right_child = true;
2800
2801 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2802 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2803
2804 int left_child_input_cnt = 0;
2805 int right_child_input_cnt = 0;
2806
2807 bool parent_is_predicated = n->is_predicated_vector();
2808 bool left_child_predicated = n->in(1)->is_predicated_vector();
2809 bool right_child_predicated = n->in(2)->is_predicated_vector();
2810
2811 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2812 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2813 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2814
2815 do {
2816 if (pack_left_child && left_child_LOP &&
2817 ((!parent_is_predicated && !left_child_predicated) ||
2818 ((parent_is_predicated && left_child_predicated &&
2819 parent_pred == left_child_pred)))) {
2820 partition.push(n->in(1));
2821 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2822 } else {
2823 inputs.push(n->in(1));
2824 left_child_input_cnt = 1;
2825 }
2826
2827 if (pack_right_child && right_child_LOP &&
2828 (!right_child_predicated ||
2829 (right_child_predicated && parent_is_predicated &&
2830 parent_pred == right_child_pred))) {
2831 partition.push(n->in(2));
2832 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2833 } else {
2834 inputs.push(n->in(2));
2835 right_child_input_cnt = 1;
2836 }
2837
2838 if (inputs.size() > 3) {
2839 assert(partition.size() > 0, "");
2840 inputs.clear();
2841 partition.clear();
2842 if (left_child_input_cnt > right_child_input_cnt) {
2843 pack_left_child = false;
2844 } else {
2845 pack_right_child = false;
2846 }
2847 } else {
2848 break;
2849 }
2850 } while(true);
2851
2852 if(partition.size()) {
2853 partition.push(n);
2854 }
2855
2856 return (partition.size() == 2 || partition.size() == 3) &&
2857 (inputs.size() == 2 || inputs.size() == 3);
2858 }
2859
2860 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2861 assert(is_vector_bitwise_op(n), "not a root");
2862
2863 visited.set(n->_idx);
2864
2865 // 1) Do a DFS walk over the logic cone.
2866 for (uint i = 1; i < n->req(); i++) {
2867 Node* in = n->in(i);
2868 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2869 process_logic_cone_root(igvn, in, visited);
2870 }
2871 }
2872
2873 // 2) Bottom up traversal: Merge node[s] with
2874 // the parent to form macro logic node.
2875 Unique_Node_List partition;
2876 Unique_Node_List inputs;
2877 if (compute_logic_cone(n, partition, inputs)) {
2878 const TypeVect* vt = n->bottom_type()->is_vect();
2879 Node* pn = partition.at(partition.size() - 1);
2880 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2881 if (mask == nullptr ||
2882 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2883 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2884 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2885 igvn.replace_node(n, macro_logic);
2886 }
2887 }
2888 }
2889
2890 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2891 ResourceMark rm;
2892 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2893 Unique_Node_List list;
2894 collect_logic_cone_roots(list);
2895
2896 while (list.size() > 0) {
2897 Node* n = list.pop();
2898 const TypeVect* vt = n->bottom_type()->is_vect();
2899 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2900 if (supported) {
2901 VectorSet visited(comp_arena());
2902 process_logic_cone_root(igvn, n, visited);
2903 }
2904 }
2905 }
2906 }
2907
2908 //------------------------------Code_Gen---------------------------------------
2909 // Given a graph, generate code for it
2910 void Compile::Code_Gen() {
2911 if (failing()) {
2912 return;
2913 }
2914
2915 // Perform instruction selection. You might think we could reclaim Matcher
2916 // memory PDQ, but actually the Matcher is used in generating spill code.
2917 // Internals of the Matcher (including some VectorSets) must remain live
2918 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2919 // set a bit in reclaimed memory.
2920
2921 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2922 // nodes. Mapping is only valid at the root of each matched subtree.
2923 NOT_PRODUCT( verify_graph_edges(); )
2924
2925 Matcher matcher;
2926 _matcher = &matcher;
2927 {
2928 TracePhase tp("matcher", &timers[_t_matcher]);
2929 matcher.match();
2930 if (failing()) {
2931 return;
2932 }
2933 }
2934 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2935 // nodes. Mapping is only valid at the root of each matched subtree.
2936 NOT_PRODUCT( verify_graph_edges(); )
2937
2938 // If you have too many nodes, or if matching has failed, bail out
2939 check_node_count(0, "out of nodes matching instructions");
2940 if (failing()) {
2941 return;
2942 }
2943
2944 print_method(PHASE_MATCHING, 2);
2945
2946 // Build a proper-looking CFG
2947 PhaseCFG cfg(node_arena(), root(), matcher);
2948 if (failing()) {
2949 return;
2950 }
2951 _cfg = &cfg;
2952 {
2953 TracePhase tp("scheduler", &timers[_t_scheduler]);
2954 bool success = cfg.do_global_code_motion();
2955 if (!success) {
2956 return;
2957 }
2958
2959 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2960 NOT_PRODUCT( verify_graph_edges(); )
2961 cfg.verify();
2962 }
2963
2964 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2965 _regalloc = ®alloc;
2966 {
2967 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2968 // Perform register allocation. After Chaitin, use-def chains are
2969 // no longer accurate (at spill code) and so must be ignored.
2970 // Node->LRG->reg mappings are still accurate.
2971 _regalloc->Register_Allocate();
2972
2973 // Bail out if the allocator builds too many nodes
2974 if (failing()) {
2975 return;
2976 }
2977
2978 print_method(PHASE_REGISTER_ALLOCATION, 2);
2979 }
2980
2981 // Prior to register allocation we kept empty basic blocks in case the
2982 // the allocator needed a place to spill. After register allocation we
2983 // are not adding any new instructions. If any basic block is empty, we
2984 // can now safely remove it.
2985 {
2986 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2987 cfg.remove_empty_blocks();
2988 if (do_freq_based_layout()) {
2989 PhaseBlockLayout layout(cfg);
2990 } else {
2991 cfg.set_loop_alignment();
2992 }
2993 cfg.fixup_flow();
2994 cfg.remove_unreachable_blocks();
2995 cfg.verify_dominator_tree();
2996 print_method(PHASE_BLOCK_ORDERING, 3);
2997 }
2998
2999 // Apply peephole optimizations
3000 if( OptoPeephole ) {
3001 TracePhase tp("peephole", &timers[_t_peephole]);
3002 PhasePeephole peep( _regalloc, cfg);
3003 peep.do_transform();
3004 print_method(PHASE_PEEPHOLE, 3);
3005 }
3006
3007 // Do late expand if CPU requires this.
3008 if (Matcher::require_postalloc_expand) {
3009 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3010 cfg.postalloc_expand(_regalloc);
3011 print_method(PHASE_POSTALLOC_EXPAND, 3);
3012 }
3013
3014 // Convert Nodes to instruction bits in a buffer
3015 {
3016 TracePhase tp("output", &timers[_t_output]);
3017 PhaseOutput output;
3018 output.Output();
3019 if (failing()) return;
3020 output.install();
3021 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3022 }
3023
3024 // He's dead, Jim.
3025 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3026 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3027 }
3028
3029 //------------------------------Final_Reshape_Counts---------------------------
3030 // This class defines counters to help identify when a method
3031 // may/must be executed using hardware with only 24-bit precision.
3032 struct Final_Reshape_Counts : public StackObj {
3033 int _call_count; // count non-inlined 'common' calls
3034 int _float_count; // count float ops requiring 24-bit precision
3035 int _double_count; // count double ops requiring more precision
3036 int _java_call_count; // count non-inlined 'java' calls
3037 int _inner_loop_count; // count loops which need alignment
3038 VectorSet _visited; // Visitation flags
3039 Node_List _tests; // Set of IfNodes & PCTableNodes
3040
3041 Final_Reshape_Counts() :
3042 _call_count(0), _float_count(0), _double_count(0),
3043 _java_call_count(0), _inner_loop_count(0) { }
3044
3045 void inc_call_count () { _call_count ++; }
3046 void inc_float_count () { _float_count ++; }
3047 void inc_double_count() { _double_count++; }
3048 void inc_java_call_count() { _java_call_count++; }
3049 void inc_inner_loop_count() { _inner_loop_count++; }
3050
3051 int get_call_count () const { return _call_count ; }
3052 int get_float_count () const { return _float_count ; }
3053 int get_double_count() const { return _double_count; }
3054 int get_java_call_count() const { return _java_call_count; }
3055 int get_inner_loop_count() const { return _inner_loop_count; }
3056 };
3057
3058 // Eliminate trivially redundant StoreCMs and accumulate their
3059 // precedence edges.
3060 void Compile::eliminate_redundant_card_marks(Node* n) {
3061 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3062 if (n->in(MemNode::Address)->outcnt() > 1) {
3063 // There are multiple users of the same address so it might be
3064 // possible to eliminate some of the StoreCMs
3065 Node* mem = n->in(MemNode::Memory);
3066 Node* adr = n->in(MemNode::Address);
3067 Node* val = n->in(MemNode::ValueIn);
3068 Node* prev = n;
3069 bool done = false;
3070 // Walk the chain of StoreCMs eliminating ones that match. As
3071 // long as it's a chain of single users then the optimization is
3072 // safe. Eliminating partially redundant StoreCMs would require
3073 // cloning copies down the other paths.
3074 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3075 if (adr == mem->in(MemNode::Address) &&
3076 val == mem->in(MemNode::ValueIn)) {
3077 // redundant StoreCM
3078 if (mem->req() > MemNode::OopStore) {
3079 // Hasn't been processed by this code yet.
3080 n->add_prec(mem->in(MemNode::OopStore));
3081 } else {
3082 // Already converted to precedence edge
3083 for (uint i = mem->req(); i < mem->len(); i++) {
3084 // Accumulate any precedence edges
3085 if (mem->in(i) != nullptr) {
3086 n->add_prec(mem->in(i));
3087 }
3088 }
3089 // Everything above this point has been processed.
3090 done = true;
3091 }
3092 // Eliminate the previous StoreCM
3093 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3094 assert(mem->outcnt() == 0, "should be dead");
3095 mem->disconnect_inputs(this);
3096 } else {
3097 prev = mem;
3098 }
3099 mem = prev->in(MemNode::Memory);
3100 }
3101 }
3102 }
3103
3104 //------------------------------final_graph_reshaping_impl----------------------
3105 // Implement items 1-5 from final_graph_reshaping below.
3106 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3107
3108 if ( n->outcnt() == 0 ) return; // dead node
3109 uint nop = n->Opcode();
3110
3111 // Check for 2-input instruction with "last use" on right input.
3112 // Swap to left input. Implements item (2).
3113 if( n->req() == 3 && // two-input instruction
3114 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3115 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3116 n->in(2)->outcnt() == 1 &&// right use IS a last use
3117 !n->in(2)->is_Con() ) { // right use is not a constant
3118 // Check for commutative opcode
3119 switch( nop ) {
3120 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3121 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3122 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3123 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3124 case Op_AndL: case Op_XorL: case Op_OrL:
3125 case Op_AndI: case Op_XorI: case Op_OrI: {
3126 // Move "last use" input to left by swapping inputs
3127 n->swap_edges(1, 2);
3128 break;
3129 }
3130 default:
3131 break;
3132 }
3133 }
3134
3135 #ifdef ASSERT
3136 if( n->is_Mem() ) {
3137 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3138 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3139 // oop will be recorded in oop map if load crosses safepoint
3140 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3141 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3142 "raw memory operations should have control edge");
3143 }
3144 if (n->is_MemBar()) {
3145 MemBarNode* mb = n->as_MemBar();
3146 if (mb->trailing_store() || mb->trailing_load_store()) {
3147 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3148 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3149 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3150 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3151 } else if (mb->leading()) {
3152 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3153 }
3154 }
3155 #endif
3156 // Count FPU ops and common calls, implements item (3)
3157 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3158 if (!gc_handled) {
3159 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3160 }
3161
3162 // Collect CFG split points
3163 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3164 frc._tests.push(n);
3165 }
3166 }
3167
3168 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3169 if (!UseDivMod) {
3170 return;
3171 }
3172
3173 // Check if "a % b" and "a / b" both exist
3174 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3175 if (d == nullptr) {
3176 return;
3177 }
3178
3179 // Replace them with a fused divmod if supported
3180 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3181 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3182 d->subsume_by(divmod->div_proj(), this);
3183 n->subsume_by(divmod->mod_proj(), this);
3184 } else {
3185 // Replace "a % b" with "a - ((a / b) * b)"
3186 Node* mult = MulNode::make(d, d->in(2), bt);
3187 Node* sub = SubNode::make(d->in(1), mult, bt);
3188 n->subsume_by(sub, this);
3189 }
3190 }
3191
3192 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3193 switch( nop ) {
3194 // Count all float operations that may use FPU
3195 case Op_AddF:
3196 case Op_SubF:
3197 case Op_MulF:
3198 case Op_DivF:
3199 case Op_NegF:
3200 case Op_ModF:
3201 case Op_ConvI2F:
3202 case Op_ConF:
3203 case Op_CmpF:
3204 case Op_CmpF3:
3205 case Op_StoreF:
3206 case Op_LoadF:
3207 // case Op_ConvL2F: // longs are split into 32-bit halves
3208 frc.inc_float_count();
3209 break;
3210
3211 case Op_ConvF2D:
3212 case Op_ConvD2F:
3213 frc.inc_float_count();
3214 frc.inc_double_count();
3215 break;
3216
3217 // Count all double operations that may use FPU
3218 case Op_AddD:
3219 case Op_SubD:
3220 case Op_MulD:
3221 case Op_DivD:
3222 case Op_NegD:
3223 case Op_ModD:
3224 case Op_ConvI2D:
3225 case Op_ConvD2I:
3226 // case Op_ConvL2D: // handled by leaf call
3227 // case Op_ConvD2L: // handled by leaf call
3228 case Op_ConD:
3229 case Op_CmpD:
3230 case Op_CmpD3:
3231 case Op_StoreD:
3232 case Op_LoadD:
3233 case Op_LoadD_unaligned:
3234 frc.inc_double_count();
3235 break;
3236 case Op_Opaque1: // Remove Opaque Nodes before matching
3237 n->subsume_by(n->in(1), this);
3238 break;
3239 case Op_CallStaticJava:
3240 case Op_CallJava:
3241 case Op_CallDynamicJava:
3242 frc.inc_java_call_count(); // Count java call site;
3243 case Op_CallRuntime:
3244 case Op_CallLeaf:
3245 case Op_CallLeafVector:
3246 case Op_CallLeafNoFP: {
3247 assert (n->is_Call(), "");
3248 CallNode *call = n->as_Call();
3249 // Count call sites where the FP mode bit would have to be flipped.
3250 // Do not count uncommon runtime calls:
3251 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3252 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3253 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3254 frc.inc_call_count(); // Count the call site
3255 } else { // See if uncommon argument is shared
3256 Node *n = call->in(TypeFunc::Parms);
3257 int nop = n->Opcode();
3258 // Clone shared simple arguments to uncommon calls, item (1).
3259 if (n->outcnt() > 1 &&
3260 !n->is_Proj() &&
3261 nop != Op_CreateEx &&
3262 nop != Op_CheckCastPP &&
3263 nop != Op_DecodeN &&
3264 nop != Op_DecodeNKlass &&
3265 !n->is_Mem() &&
3266 !n->is_Phi()) {
3267 Node *x = n->clone();
3268 call->set_req(TypeFunc::Parms, x);
3269 }
3270 }
3271 break;
3272 }
3273
3274 case Op_StoreCM:
3275 {
3276 // Convert OopStore dependence into precedence edge
3277 Node* prec = n->in(MemNode::OopStore);
3278 n->del_req(MemNode::OopStore);
3279 n->add_prec(prec);
3280 eliminate_redundant_card_marks(n);
3281 }
3282
3283 // fall through
3284
3285 case Op_StoreB:
3286 case Op_StoreC:
3287 case Op_StoreI:
3288 case Op_StoreL:
3289 case Op_CompareAndSwapB:
3290 case Op_CompareAndSwapS:
3291 case Op_CompareAndSwapI:
3292 case Op_CompareAndSwapL:
3293 case Op_CompareAndSwapP:
3294 case Op_CompareAndSwapN:
3295 case Op_WeakCompareAndSwapB:
3296 case Op_WeakCompareAndSwapS:
3297 case Op_WeakCompareAndSwapI:
3298 case Op_WeakCompareAndSwapL:
3299 case Op_WeakCompareAndSwapP:
3300 case Op_WeakCompareAndSwapN:
3301 case Op_CompareAndExchangeB:
3302 case Op_CompareAndExchangeS:
3303 case Op_CompareAndExchangeI:
3304 case Op_CompareAndExchangeL:
3305 case Op_CompareAndExchangeP:
3306 case Op_CompareAndExchangeN:
3307 case Op_GetAndAddS:
3308 case Op_GetAndAddB:
3309 case Op_GetAndAddI:
3310 case Op_GetAndAddL:
3311 case Op_GetAndSetS:
3312 case Op_GetAndSetB:
3313 case Op_GetAndSetI:
3314 case Op_GetAndSetL:
3315 case Op_GetAndSetP:
3316 case Op_GetAndSetN:
3317 case Op_StoreP:
3318 case Op_StoreN:
3319 case Op_StoreNKlass:
3320 case Op_LoadB:
3321 case Op_LoadUB:
3322 case Op_LoadUS:
3323 case Op_LoadI:
3324 case Op_LoadKlass:
3325 case Op_LoadNKlass:
3326 case Op_LoadL:
3327 case Op_LoadL_unaligned:
3328 case Op_LoadP:
3329 case Op_LoadN:
3330 case Op_LoadRange:
3331 case Op_LoadS:
3332 break;
3333
3334 case Op_AddP: { // Assert sane base pointers
3335 Node *addp = n->in(AddPNode::Address);
3336 assert( !addp->is_AddP() ||
3337 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3338 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3339 "Base pointers must match (addp %u)", addp->_idx );
3340 #ifdef _LP64
3341 if ((UseCompressedOops || UseCompressedClassPointers) &&
3342 addp->Opcode() == Op_ConP &&
3343 addp == n->in(AddPNode::Base) &&
3344 n->in(AddPNode::Offset)->is_Con()) {
3345 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3346 // on the platform and on the compressed oops mode.
3347 // Use addressing with narrow klass to load with offset on x86.
3348 // Some platforms can use the constant pool to load ConP.
3349 // Do this transformation here since IGVN will convert ConN back to ConP.
3350 const Type* t = addp->bottom_type();
3351 bool is_oop = t->isa_oopptr() != nullptr;
3352 bool is_klass = t->isa_klassptr() != nullptr;
3353
3354 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3355 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3356 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3357 Node* nn = nullptr;
3358
3359 int op = is_oop ? Op_ConN : Op_ConNKlass;
3360
3361 // Look for existing ConN node of the same exact type.
3362 Node* r = root();
3363 uint cnt = r->outcnt();
3364 for (uint i = 0; i < cnt; i++) {
3365 Node* m = r->raw_out(i);
3366 if (m!= nullptr && m->Opcode() == op &&
3367 m->bottom_type()->make_ptr() == t) {
3368 nn = m;
3369 break;
3370 }
3371 }
3372 if (nn != nullptr) {
3373 // Decode a narrow oop to match address
3374 // [R12 + narrow_oop_reg<<3 + offset]
3375 if (is_oop) {
3376 nn = new DecodeNNode(nn, t);
3377 } else {
3378 nn = new DecodeNKlassNode(nn, t);
3379 }
3380 // Check for succeeding AddP which uses the same Base.
3381 // Otherwise we will run into the assertion above when visiting that guy.
3382 for (uint i = 0; i < n->outcnt(); ++i) {
3383 Node *out_i = n->raw_out(i);
3384 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3385 out_i->set_req(AddPNode::Base, nn);
3386 #ifdef ASSERT
3387 for (uint j = 0; j < out_i->outcnt(); ++j) {
3388 Node *out_j = out_i->raw_out(j);
3389 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3390 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3391 }
3392 #endif
3393 }
3394 }
3395 n->set_req(AddPNode::Base, nn);
3396 n->set_req(AddPNode::Address, nn);
3397 if (addp->outcnt() == 0) {
3398 addp->disconnect_inputs(this);
3399 }
3400 }
3401 }
3402 }
3403 #endif
3404 break;
3405 }
3406
3407 case Op_CastPP: {
3408 // Remove CastPP nodes to gain more freedom during scheduling but
3409 // keep the dependency they encode as control or precedence edges
3410 // (if control is set already) on memory operations. Some CastPP
3411 // nodes don't have a control (don't carry a dependency): skip
3412 // those.
3413 if (n->in(0) != nullptr) {
3414 ResourceMark rm;
3415 Unique_Node_List wq;
3416 wq.push(n);
3417 for (uint next = 0; next < wq.size(); ++next) {
3418 Node *m = wq.at(next);
3419 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3420 Node* use = m->fast_out(i);
3421 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3422 use->ensure_control_or_add_prec(n->in(0));
3423 } else {
3424 switch(use->Opcode()) {
3425 case Op_AddP:
3426 case Op_DecodeN:
3427 case Op_DecodeNKlass:
3428 case Op_CheckCastPP:
3429 case Op_CastPP:
3430 wq.push(use);
3431 break;
3432 }
3433 }
3434 }
3435 }
3436 }
3437 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3438 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3439 Node* in1 = n->in(1);
3440 const Type* t = n->bottom_type();
3441 Node* new_in1 = in1->clone();
3442 new_in1->as_DecodeN()->set_type(t);
3443
3444 if (!Matcher::narrow_oop_use_complex_address()) {
3445 //
3446 // x86, ARM and friends can handle 2 adds in addressing mode
3447 // and Matcher can fold a DecodeN node into address by using
3448 // a narrow oop directly and do implicit null check in address:
3449 //
3450 // [R12 + narrow_oop_reg<<3 + offset]
3451 // NullCheck narrow_oop_reg
3452 //
3453 // On other platforms (Sparc) we have to keep new DecodeN node and
3454 // use it to do implicit null check in address:
3455 //
3456 // decode_not_null narrow_oop_reg, base_reg
3457 // [base_reg + offset]
3458 // NullCheck base_reg
3459 //
3460 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3461 // to keep the information to which null check the new DecodeN node
3462 // corresponds to use it as value in implicit_null_check().
3463 //
3464 new_in1->set_req(0, n->in(0));
3465 }
3466
3467 n->subsume_by(new_in1, this);
3468 if (in1->outcnt() == 0) {
3469 in1->disconnect_inputs(this);
3470 }
3471 } else {
3472 n->subsume_by(n->in(1), this);
3473 if (n->outcnt() == 0) {
3474 n->disconnect_inputs(this);
3475 }
3476 }
3477 break;
3478 }
3479 #ifdef _LP64
3480 case Op_CmpP:
3481 // Do this transformation here to preserve CmpPNode::sub() and
3482 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3483 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3484 Node* in1 = n->in(1);
3485 Node* in2 = n->in(2);
3486 if (!in1->is_DecodeNarrowPtr()) {
3487 in2 = in1;
3488 in1 = n->in(2);
3489 }
3490 assert(in1->is_DecodeNarrowPtr(), "sanity");
3491
3492 Node* new_in2 = nullptr;
3493 if (in2->is_DecodeNarrowPtr()) {
3494 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3495 new_in2 = in2->in(1);
3496 } else if (in2->Opcode() == Op_ConP) {
3497 const Type* t = in2->bottom_type();
3498 if (t == TypePtr::NULL_PTR) {
3499 assert(in1->is_DecodeN(), "compare klass to null?");
3500 // Don't convert CmpP null check into CmpN if compressed
3501 // oops implicit null check is not generated.
3502 // This will allow to generate normal oop implicit null check.
3503 if (Matcher::gen_narrow_oop_implicit_null_checks())
3504 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3505 //
3506 // This transformation together with CastPP transformation above
3507 // will generated code for implicit null checks for compressed oops.
3508 //
3509 // The original code after Optimize()
3510 //
3511 // LoadN memory, narrow_oop_reg
3512 // decode narrow_oop_reg, base_reg
3513 // CmpP base_reg, nullptr
3514 // CastPP base_reg // NotNull
3515 // Load [base_reg + offset], val_reg
3516 //
3517 // after these transformations will be
3518 //
3519 // LoadN memory, narrow_oop_reg
3520 // CmpN narrow_oop_reg, nullptr
3521 // decode_not_null narrow_oop_reg, base_reg
3522 // Load [base_reg + offset], val_reg
3523 //
3524 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3525 // since narrow oops can be used in debug info now (see the code in
3526 // final_graph_reshaping_walk()).
3527 //
3528 // At the end the code will be matched to
3529 // on x86:
3530 //
3531 // Load_narrow_oop memory, narrow_oop_reg
3532 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3533 // NullCheck narrow_oop_reg
3534 //
3535 // and on sparc:
3536 //
3537 // Load_narrow_oop memory, narrow_oop_reg
3538 // decode_not_null narrow_oop_reg, base_reg
3539 // Load [base_reg + offset], val_reg
3540 // NullCheck base_reg
3541 //
3542 } else if (t->isa_oopptr()) {
3543 new_in2 = ConNode::make(t->make_narrowoop());
3544 } else if (t->isa_klassptr()) {
3545 new_in2 = ConNode::make(t->make_narrowklass());
3546 }
3547 }
3548 if (new_in2 != nullptr) {
3549 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3550 n->subsume_by(cmpN, this);
3551 if (in1->outcnt() == 0) {
3552 in1->disconnect_inputs(this);
3553 }
3554 if (in2->outcnt() == 0) {
3555 in2->disconnect_inputs(this);
3556 }
3557 }
3558 }
3559 break;
3560
3561 case Op_DecodeN:
3562 case Op_DecodeNKlass:
3563 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3564 // DecodeN could be pinned when it can't be fold into
3565 // an address expression, see the code for Op_CastPP above.
3566 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3567 break;
3568
3569 case Op_EncodeP:
3570 case Op_EncodePKlass: {
3571 Node* in1 = n->in(1);
3572 if (in1->is_DecodeNarrowPtr()) {
3573 n->subsume_by(in1->in(1), this);
3574 } else if (in1->Opcode() == Op_ConP) {
3575 const Type* t = in1->bottom_type();
3576 if (t == TypePtr::NULL_PTR) {
3577 assert(t->isa_oopptr(), "null klass?");
3578 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3579 } else if (t->isa_oopptr()) {
3580 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3581 } else if (t->isa_klassptr()) {
3582 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3583 }
3584 }
3585 if (in1->outcnt() == 0) {
3586 in1->disconnect_inputs(this);
3587 }
3588 break;
3589 }
3590
3591 case Op_Proj: {
3592 if (OptimizeStringConcat || IncrementalInline) {
3593 ProjNode* proj = n->as_Proj();
3594 if (proj->_is_io_use) {
3595 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3596 // Separate projections were used for the exception path which
3597 // are normally removed by a late inline. If it wasn't inlined
3598 // then they will hang around and should just be replaced with
3599 // the original one. Merge them.
3600 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3601 if (non_io_proj != nullptr) {
3602 proj->subsume_by(non_io_proj , this);
3603 }
3604 }
3605 }
3606 break;
3607 }
3608
3609 case Op_Phi:
3610 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3611 // The EncodeP optimization may create Phi with the same edges
3612 // for all paths. It is not handled well by Register Allocator.
3613 Node* unique_in = n->in(1);
3614 assert(unique_in != nullptr, "");
3615 uint cnt = n->req();
3616 for (uint i = 2; i < cnt; i++) {
3617 Node* m = n->in(i);
3618 assert(m != nullptr, "");
3619 if (unique_in != m)
3620 unique_in = nullptr;
3621 }
3622 if (unique_in != nullptr) {
3623 n->subsume_by(unique_in, this);
3624 }
3625 }
3626 break;
3627
3628 #endif
3629
3630 #ifdef ASSERT
3631 case Op_CastII:
3632 // Verify that all range check dependent CastII nodes were removed.
3633 if (n->isa_CastII()->has_range_check()) {
3634 n->dump(3);
3635 assert(false, "Range check dependent CastII node was not removed");
3636 }
3637 break;
3638 #endif
3639
3640 case Op_ModI:
3641 handle_div_mod_op(n, T_INT, false);
3642 break;
3643
3644 case Op_ModL:
3645 handle_div_mod_op(n, T_LONG, false);
3646 break;
3647
3648 case Op_UModI:
3649 handle_div_mod_op(n, T_INT, true);
3650 break;
3651
3652 case Op_UModL:
3653 handle_div_mod_op(n, T_LONG, true);
3654 break;
3655
3656 case Op_LoadVector:
3657 case Op_StoreVector:
3658 #ifdef ASSERT
3659 // Add VerifyVectorAlignment node between adr and load / store.
3660 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3661 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3662 n->as_StoreVector()->must_verify_alignment();
3663 if (must_verify_alignment) {
3664 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3665 n->as_StoreVector()->memory_size();
3666 // The memory access should be aligned to the vector width in bytes.
3667 // However, the underlying array is possibly less well aligned, but at least
3668 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3669 // a loop we can expect at least the following alignment:
3670 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3671 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3672 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3673 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3674 jlong mask = guaranteed_alignment - 1;
3675 Node* mask_con = ConLNode::make(mask);
3676 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3677 n->set_req(MemNode::Address, va);
3678 }
3679 }
3680 #endif
3681 break;
3682
3683 case Op_LoadVectorGather:
3684 case Op_StoreVectorScatter:
3685 case Op_LoadVectorGatherMasked:
3686 case Op_StoreVectorScatterMasked:
3687 case Op_VectorCmpMasked:
3688 case Op_VectorMaskGen:
3689 case Op_LoadVectorMasked:
3690 case Op_StoreVectorMasked:
3691 break;
3692
3693 case Op_AddReductionVI:
3694 case Op_AddReductionVL:
3695 case Op_AddReductionVF:
3696 case Op_AddReductionVD:
3697 case Op_MulReductionVI:
3698 case Op_MulReductionVL:
3699 case Op_MulReductionVF:
3700 case Op_MulReductionVD:
3701 case Op_MinReductionV:
3702 case Op_MaxReductionV:
3703 case Op_AndReductionV:
3704 case Op_OrReductionV:
3705 case Op_XorReductionV:
3706 break;
3707
3708 case Op_PackB:
3709 case Op_PackS:
3710 case Op_PackI:
3711 case Op_PackF:
3712 case Op_PackL:
3713 case Op_PackD:
3714 if (n->req()-1 > 2) {
3715 // Replace many operand PackNodes with a binary tree for matching
3716 PackNode* p = (PackNode*) n;
3717 Node* btp = p->binary_tree_pack(1, n->req());
3718 n->subsume_by(btp, this);
3719 }
3720 break;
3721 case Op_Loop:
3722 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3723 case Op_CountedLoop:
3724 case Op_LongCountedLoop:
3725 case Op_OuterStripMinedLoop:
3726 if (n->as_Loop()->is_inner_loop()) {
3727 frc.inc_inner_loop_count();
3728 }
3729 n->as_Loop()->verify_strip_mined(0);
3730 break;
3731 case Op_LShiftI:
3732 case Op_RShiftI:
3733 case Op_URShiftI:
3734 case Op_LShiftL:
3735 case Op_RShiftL:
3736 case Op_URShiftL:
3737 if (Matcher::need_masked_shift_count) {
3738 // The cpu's shift instructions don't restrict the count to the
3739 // lower 5/6 bits. We need to do the masking ourselves.
3740 Node* in2 = n->in(2);
3741 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3742 const TypeInt* t = in2->find_int_type();
3743 if (t != nullptr && t->is_con()) {
3744 juint shift = t->get_con();
3745 if (shift > mask) { // Unsigned cmp
3746 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3747 }
3748 } else {
3749 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3750 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3751 n->set_req(2, shift);
3752 }
3753 }
3754 if (in2->outcnt() == 0) { // Remove dead node
3755 in2->disconnect_inputs(this);
3756 }
3757 }
3758 break;
3759 case Op_MemBarStoreStore:
3760 case Op_MemBarRelease:
3761 // Break the link with AllocateNode: it is no longer useful and
3762 // confuses register allocation.
3763 if (n->req() > MemBarNode::Precedent) {
3764 n->set_req(MemBarNode::Precedent, top());
3765 }
3766 break;
3767 case Op_MemBarAcquire: {
3768 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3769 // At parse time, the trailing MemBarAcquire for a volatile load
3770 // is created with an edge to the load. After optimizations,
3771 // that input may be a chain of Phis. If those phis have no
3772 // other use, then the MemBarAcquire keeps them alive and
3773 // register allocation can be confused.
3774 dead_nodes.push(n->in(MemBarNode::Precedent));
3775 n->set_req(MemBarNode::Precedent, top());
3776 }
3777 break;
3778 }
3779 case Op_Blackhole:
3780 break;
3781 case Op_RangeCheck: {
3782 RangeCheckNode* rc = n->as_RangeCheck();
3783 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3784 n->subsume_by(iff, this);
3785 frc._tests.push(iff);
3786 break;
3787 }
3788 case Op_ConvI2L: {
3789 if (!Matcher::convi2l_type_required) {
3790 // Code generation on some platforms doesn't need accurate
3791 // ConvI2L types. Widening the type can help remove redundant
3792 // address computations.
3793 n->as_Type()->set_type(TypeLong::INT);
3794 ResourceMark rm;
3795 Unique_Node_List wq;
3796 wq.push(n);
3797 for (uint next = 0; next < wq.size(); next++) {
3798 Node *m = wq.at(next);
3799
3800 for(;;) {
3801 // Loop over all nodes with identical inputs edges as m
3802 Node* k = m->find_similar(m->Opcode());
3803 if (k == nullptr) {
3804 break;
3805 }
3806 // Push their uses so we get a chance to remove node made
3807 // redundant
3808 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3809 Node* u = k->fast_out(i);
3810 if (u->Opcode() == Op_LShiftL ||
3811 u->Opcode() == Op_AddL ||
3812 u->Opcode() == Op_SubL ||
3813 u->Opcode() == Op_AddP) {
3814 wq.push(u);
3815 }
3816 }
3817 // Replace all nodes with identical edges as m with m
3818 k->subsume_by(m, this);
3819 }
3820 }
3821 }
3822 break;
3823 }
3824 case Op_CmpUL: {
3825 if (!Matcher::has_match_rule(Op_CmpUL)) {
3826 // No support for unsigned long comparisons
3827 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3828 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3829 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3830 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3831 Node* andl = new AndLNode(orl, remove_sign_mask);
3832 Node* cmp = new CmpLNode(andl, n->in(2));
3833 n->subsume_by(cmp, this);
3834 }
3835 break;
3836 }
3837 default:
3838 assert(!n->is_Call(), "");
3839 assert(!n->is_Mem(), "");
3840 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3841 break;
3842 }
3843 }
3844
3845 //------------------------------final_graph_reshaping_walk---------------------
3846 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3847 // requires that the walk visits a node's inputs before visiting the node.
3848 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3849 Unique_Node_List sfpt;
3850
3851 frc._visited.set(root->_idx); // first, mark node as visited
3852 uint cnt = root->req();
3853 Node *n = root;
3854 uint i = 0;
3855 while (true) {
3856 if (i < cnt) {
3857 // Place all non-visited non-null inputs onto stack
3858 Node* m = n->in(i);
3859 ++i;
3860 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3861 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3862 // compute worst case interpreter size in case of a deoptimization
3863 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3864
3865 sfpt.push(m);
3866 }
3867 cnt = m->req();
3868 nstack.push(n, i); // put on stack parent and next input's index
3869 n = m;
3870 i = 0;
3871 }
3872 } else {
3873 // Now do post-visit work
3874 final_graph_reshaping_impl(n, frc, dead_nodes);
3875 if (nstack.is_empty())
3876 break; // finished
3877 n = nstack.node(); // Get node from stack
3878 cnt = n->req();
3879 i = nstack.index();
3880 nstack.pop(); // Shift to the next node on stack
3881 }
3882 }
3883
3884 // Skip next transformation if compressed oops are not used.
3885 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3886 (!UseCompressedOops && !UseCompressedClassPointers))
3887 return;
3888
3889 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3890 // It could be done for an uncommon traps or any safepoints/calls
3891 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3892 while (sfpt.size() > 0) {
3893 n = sfpt.pop();
3894 JVMState *jvms = n->as_SafePoint()->jvms();
3895 assert(jvms != nullptr, "sanity");
3896 int start = jvms->debug_start();
3897 int end = n->req();
3898 bool is_uncommon = (n->is_CallStaticJava() &&
3899 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3900 for (int j = start; j < end; j++) {
3901 Node* in = n->in(j);
3902 if (in->is_DecodeNarrowPtr()) {
3903 bool safe_to_skip = true;
3904 if (!is_uncommon ) {
3905 // Is it safe to skip?
3906 for (uint i = 0; i < in->outcnt(); i++) {
3907 Node* u = in->raw_out(i);
3908 if (!u->is_SafePoint() ||
3909 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3910 safe_to_skip = false;
3911 }
3912 }
3913 }
3914 if (safe_to_skip) {
3915 n->set_req(j, in->in(1));
3916 }
3917 if (in->outcnt() == 0) {
3918 in->disconnect_inputs(this);
3919 }
3920 }
3921 }
3922 }
3923 }
3924
3925 //------------------------------final_graph_reshaping--------------------------
3926 // Final Graph Reshaping.
3927 //
3928 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3929 // and not commoned up and forced early. Must come after regular
3930 // optimizations to avoid GVN undoing the cloning. Clone constant
3931 // inputs to Loop Phis; these will be split by the allocator anyways.
3932 // Remove Opaque nodes.
3933 // (2) Move last-uses by commutative operations to the left input to encourage
3934 // Intel update-in-place two-address operations and better register usage
3935 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3936 // calls canonicalizing them back.
3937 // (3) Count the number of double-precision FP ops, single-precision FP ops
3938 // and call sites. On Intel, we can get correct rounding either by
3939 // forcing singles to memory (requires extra stores and loads after each
3940 // FP bytecode) or we can set a rounding mode bit (requires setting and
3941 // clearing the mode bit around call sites). The mode bit is only used
3942 // if the relative frequency of single FP ops to calls is low enough.
3943 // This is a key transform for SPEC mpeg_audio.
3944 // (4) Detect infinite loops; blobs of code reachable from above but not
3945 // below. Several of the Code_Gen algorithms fail on such code shapes,
3946 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3947 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3948 // Detection is by looking for IfNodes where only 1 projection is
3949 // reachable from below or CatchNodes missing some targets.
3950 // (5) Assert for insane oop offsets in debug mode.
3951
3952 bool Compile::final_graph_reshaping() {
3953 // an infinite loop may have been eliminated by the optimizer,
3954 // in which case the graph will be empty.
3955 if (root()->req() == 1) {
3956 // Do not compile method that is only a trivial infinite loop,
3957 // since the content of the loop may have been eliminated.
3958 record_method_not_compilable("trivial infinite loop");
3959 return true;
3960 }
3961
3962 // Expensive nodes have their control input set to prevent the GVN
3963 // from freely commoning them. There's no GVN beyond this point so
3964 // no need to keep the control input. We want the expensive nodes to
3965 // be freely moved to the least frequent code path by gcm.
3966 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3967 for (int i = 0; i < expensive_count(); i++) {
3968 _expensive_nodes.at(i)->set_req(0, nullptr);
3969 }
3970
3971 Final_Reshape_Counts frc;
3972
3973 // Visit everybody reachable!
3974 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3975 Node_Stack nstack(live_nodes() >> 1);
3976 Unique_Node_List dead_nodes;
3977 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
3978
3979 // Check for unreachable (from below) code (i.e., infinite loops).
3980 for( uint i = 0; i < frc._tests.size(); i++ ) {
3981 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3982 // Get number of CFG targets.
3983 // Note that PCTables include exception targets after calls.
3984 uint required_outcnt = n->required_outcnt();
3985 if (n->outcnt() != required_outcnt) {
3986 // Check for a few special cases. Rethrow Nodes never take the
3987 // 'fall-thru' path, so expected kids is 1 less.
3988 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3989 if (n->in(0)->in(0)->is_Call()) {
3990 CallNode* call = n->in(0)->in(0)->as_Call();
3991 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3992 required_outcnt--; // Rethrow always has 1 less kid
3993 } else if (call->req() > TypeFunc::Parms &&
3994 call->is_CallDynamicJava()) {
3995 // Check for null receiver. In such case, the optimizer has
3996 // detected that the virtual call will always result in a null
3997 // pointer exception. The fall-through projection of this CatchNode
3998 // will not be populated.
3999 Node* arg0 = call->in(TypeFunc::Parms);
4000 if (arg0->is_Type() &&
4001 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4002 required_outcnt--;
4003 }
4004 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4005 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4006 // Check for illegal array length. In such case, the optimizer has
4007 // detected that the allocation attempt will always result in an
4008 // exception. There is no fall-through projection of this CatchNode .
4009 assert(call->is_CallStaticJava(), "static call expected");
4010 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4011 uint valid_length_test_input = call->req() - 1;
4012 Node* valid_length_test = call->in(valid_length_test_input);
4013 call->del_req(valid_length_test_input);
4014 if (valid_length_test->find_int_con(1) == 0) {
4015 required_outcnt--;
4016 }
4017 dead_nodes.push(valid_length_test);
4018 assert(n->outcnt() == required_outcnt, "malformed control flow");
4019 continue;
4020 }
4021 }
4022 }
4023
4024 // Recheck with a better notion of 'required_outcnt'
4025 if (n->outcnt() != required_outcnt) {
4026 record_method_not_compilable("malformed control flow");
4027 return true; // Not all targets reachable!
4028 }
4029 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4030 CallNode* call = n->in(0)->in(0)->as_Call();
4031 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4032 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4033 assert(call->is_CallStaticJava(), "static call expected");
4034 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4035 uint valid_length_test_input = call->req() - 1;
4036 dead_nodes.push(call->in(valid_length_test_input));
4037 call->del_req(valid_length_test_input); // valid length test useless now
4038 }
4039 }
4040 // Check that I actually visited all kids. Unreached kids
4041 // must be infinite loops.
4042 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4043 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4044 record_method_not_compilable("infinite loop");
4045 return true; // Found unvisited kid; must be unreach
4046 }
4047
4048 // Here so verification code in final_graph_reshaping_walk()
4049 // always see an OuterStripMinedLoopEnd
4050 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4051 IfNode* init_iff = n->as_If();
4052 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4053 n->subsume_by(iff, this);
4054 }
4055 }
4056
4057 while (dead_nodes.size() > 0) {
4058 Node* m = dead_nodes.pop();
4059 if (m->outcnt() == 0 && m != top()) {
4060 for (uint j = 0; j < m->req(); j++) {
4061 Node* in = m->in(j);
4062 if (in != nullptr) {
4063 dead_nodes.push(in);
4064 }
4065 }
4066 m->disconnect_inputs(this);
4067 }
4068 }
4069
4070 #ifdef IA32
4071 // If original bytecodes contained a mixture of floats and doubles
4072 // check if the optimizer has made it homogeneous, item (3).
4073 if (UseSSE == 0 &&
4074 frc.get_float_count() > 32 &&
4075 frc.get_double_count() == 0 &&
4076 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4077 set_24_bit_selection_and_mode(false, true);
4078 }
4079 #endif // IA32
4080
4081 set_java_calls(frc.get_java_call_count());
4082 set_inner_loops(frc.get_inner_loop_count());
4083
4084 // No infinite loops, no reason to bail out.
4085 return false;
4086 }
4087
4088 //-----------------------------too_many_traps----------------------------------
4089 // Report if there are too many traps at the current method and bci.
4090 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4091 bool Compile::too_many_traps(ciMethod* method,
4092 int bci,
4093 Deoptimization::DeoptReason reason) {
4094 ciMethodData* md = method->method_data();
4095 if (md->is_empty()) {
4096 // Assume the trap has not occurred, or that it occurred only
4097 // because of a transient condition during start-up in the interpreter.
4098 return false;
4099 }
4100 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4101 if (md->has_trap_at(bci, m, reason) != 0) {
4102 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4103 // Also, if there are multiple reasons, or if there is no per-BCI record,
4104 // assume the worst.
4105 if (log())
4106 log()->elem("observe trap='%s' count='%d'",
4107 Deoptimization::trap_reason_name(reason),
4108 md->trap_count(reason));
4109 return true;
4110 } else {
4111 // Ignore method/bci and see if there have been too many globally.
4112 return too_many_traps(reason, md);
4113 }
4114 }
4115
4116 // Less-accurate variant which does not require a method and bci.
4117 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4118 ciMethodData* logmd) {
4119 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4120 // Too many traps globally.
4121 // Note that we use cumulative trap_count, not just md->trap_count.
4122 if (log()) {
4123 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4124 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4125 Deoptimization::trap_reason_name(reason),
4126 mcount, trap_count(reason));
4127 }
4128 return true;
4129 } else {
4130 // The coast is clear.
4131 return false;
4132 }
4133 }
4134
4135 //--------------------------too_many_recompiles--------------------------------
4136 // Report if there are too many recompiles at the current method and bci.
4137 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4138 // Is not eager to return true, since this will cause the compiler to use
4139 // Action_none for a trap point, to avoid too many recompilations.
4140 bool Compile::too_many_recompiles(ciMethod* method,
4141 int bci,
4142 Deoptimization::DeoptReason reason) {
4143 ciMethodData* md = method->method_data();
4144 if (md->is_empty()) {
4145 // Assume the trap has not occurred, or that it occurred only
4146 // because of a transient condition during start-up in the interpreter.
4147 return false;
4148 }
4149 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4150 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4151 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4152 Deoptimization::DeoptReason per_bc_reason
4153 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4154 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4155 if ((per_bc_reason == Deoptimization::Reason_none
4156 || md->has_trap_at(bci, m, reason) != 0)
4157 // The trap frequency measure we care about is the recompile count:
4158 && md->trap_recompiled_at(bci, m)
4159 && md->overflow_recompile_count() >= bc_cutoff) {
4160 // Do not emit a trap here if it has already caused recompilations.
4161 // Also, if there are multiple reasons, or if there is no per-BCI record,
4162 // assume the worst.
4163 if (log())
4164 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4165 Deoptimization::trap_reason_name(reason),
4166 md->trap_count(reason),
4167 md->overflow_recompile_count());
4168 return true;
4169 } else if (trap_count(reason) != 0
4170 && decompile_count() >= m_cutoff) {
4171 // Too many recompiles globally, and we have seen this sort of trap.
4172 // Use cumulative decompile_count, not just md->decompile_count.
4173 if (log())
4174 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4175 Deoptimization::trap_reason_name(reason),
4176 md->trap_count(reason), trap_count(reason),
4177 md->decompile_count(), decompile_count());
4178 return true;
4179 } else {
4180 // The coast is clear.
4181 return false;
4182 }
4183 }
4184
4185 // Compute when not to trap. Used by matching trap based nodes and
4186 // NullCheck optimization.
4187 void Compile::set_allowed_deopt_reasons() {
4188 _allowed_reasons = 0;
4189 if (is_method_compilation()) {
4190 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4191 assert(rs < BitsPerInt, "recode bit map");
4192 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4193 _allowed_reasons |= nth_bit(rs);
4194 }
4195 }
4196 }
4197 }
4198
4199 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4200 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4201 }
4202
4203 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4204 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4205 }
4206
4207 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4208 if (holder->is_initialized()) {
4209 return false;
4210 }
4211 if (holder->is_being_initialized()) {
4212 if (accessing_method->holder() == holder) {
4213 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4214 // <init>, or a static method. In all those cases, there was an initialization
4215 // barrier on the holder klass passed.
4216 if (accessing_method->is_static_initializer() ||
4217 accessing_method->is_object_initializer() ||
4218 accessing_method->is_static()) {
4219 return false;
4220 }
4221 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4222 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4223 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4224 // child class can become fully initialized while its parent class is still being initialized.
4225 if (accessing_method->is_static_initializer()) {
4226 return false;
4227 }
4228 }
4229 ciMethod* root = method(); // the root method of compilation
4230 if (root != accessing_method) {
4231 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4232 }
4233 }
4234 return true;
4235 }
4236
4237 #ifndef PRODUCT
4238 //------------------------------verify_bidirectional_edges---------------------
4239 // For each input edge to a node (ie - for each Use-Def edge), verify that
4240 // there is a corresponding Def-Use edge.
4241 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4242 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4243 uint stack_size = live_nodes() >> 4;
4244 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4245 nstack.push(_root);
4246
4247 while (nstack.size() > 0) {
4248 Node* n = nstack.pop();
4249 if (visited.member(n)) {
4250 continue;
4251 }
4252 visited.push(n);
4253
4254 // Walk over all input edges, checking for correspondence
4255 uint length = n->len();
4256 for (uint i = 0; i < length; i++) {
4257 Node* in = n->in(i);
4258 if (in != nullptr && !visited.member(in)) {
4259 nstack.push(in); // Put it on stack
4260 }
4261 if (in != nullptr && !in->is_top()) {
4262 // Count instances of `next`
4263 int cnt = 0;
4264 for (uint idx = 0; idx < in->_outcnt; idx++) {
4265 if (in->_out[idx] == n) {
4266 cnt++;
4267 }
4268 }
4269 assert(cnt > 0, "Failed to find Def-Use edge.");
4270 // Check for duplicate edges
4271 // walk the input array downcounting the input edges to n
4272 for (uint j = 0; j < length; j++) {
4273 if (n->in(j) == in) {
4274 cnt--;
4275 }
4276 }
4277 assert(cnt == 0, "Mismatched edge count.");
4278 } else if (in == nullptr) {
4279 assert(i == 0 || i >= n->req() ||
4280 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4281 (n->is_Unlock() && i == (n->req() - 1)) ||
4282 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4283 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4284 } else {
4285 assert(in->is_top(), "sanity");
4286 // Nothing to check.
4287 }
4288 }
4289 }
4290 }
4291
4292 //------------------------------verify_graph_edges---------------------------
4293 // Walk the Graph and verify that there is a one-to-one correspondence
4294 // between Use-Def edges and Def-Use edges in the graph.
4295 void Compile::verify_graph_edges(bool no_dead_code) {
4296 if (VerifyGraphEdges) {
4297 Unique_Node_List visited;
4298
4299 // Call graph walk to check edges
4300 verify_bidirectional_edges(visited);
4301 if (no_dead_code) {
4302 // Now make sure that no visited node is used by an unvisited node.
4303 bool dead_nodes = false;
4304 Unique_Node_List checked;
4305 while (visited.size() > 0) {
4306 Node* n = visited.pop();
4307 checked.push(n);
4308 for (uint i = 0; i < n->outcnt(); i++) {
4309 Node* use = n->raw_out(i);
4310 if (checked.member(use)) continue; // already checked
4311 if (visited.member(use)) continue; // already in the graph
4312 if (use->is_Con()) continue; // a dead ConNode is OK
4313 // At this point, we have found a dead node which is DU-reachable.
4314 if (!dead_nodes) {
4315 tty->print_cr("*** Dead nodes reachable via DU edges:");
4316 dead_nodes = true;
4317 }
4318 use->dump(2);
4319 tty->print_cr("---");
4320 checked.push(use); // No repeats; pretend it is now checked.
4321 }
4322 }
4323 assert(!dead_nodes, "using nodes must be reachable from root");
4324 }
4325 }
4326 }
4327 #endif
4328
4329 // The Compile object keeps track of failure reasons separately from the ciEnv.
4330 // This is required because there is not quite a 1-1 relation between the
4331 // ciEnv and its compilation task and the Compile object. Note that one
4332 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4333 // to backtrack and retry without subsuming loads. Other than this backtracking
4334 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4335 // by the logic in C2Compiler.
4336 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4337 if (log() != nullptr) {
4338 log()->elem("failure reason='%s' phase='compile'", reason);
4339 }
4340 if (_failure_reason.get() == nullptr) {
4341 // Record the first failure reason.
4342 _failure_reason.set(reason);
4343 if (CaptureBailoutInformation) {
4344 _first_failure_details = new CompilationFailureInfo(reason);
4345 }
4346 } else {
4347 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4348 }
4349
4350 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4351 C->print_method(PHASE_FAILURE, 1);
4352 }
4353 _root = nullptr; // flush the graph, too
4354 }
4355
4356 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4357 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4358 _compile(Compile::current()),
4359 _log(nullptr),
4360 _phase_name(name),
4361 _dolog(CITimeVerbose)
4362 {
4363 assert(_compile != nullptr, "sanity check");
4364 if (_dolog) {
4365 _log = _compile->log();
4366 }
4367 if (_log != nullptr) {
4368 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4369 _log->stamp();
4370 _log->end_head();
4371 }
4372 }
4373
4374 Compile::TracePhase::~TracePhase() {
4375 if (_compile->failing_internal()) {
4376 return; // timing code, not stressing bailouts.
4377 }
4378 #ifdef ASSERT
4379 if (PrintIdealNodeCount) {
4380 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4381 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4382 }
4383
4384 if (VerifyIdealNodeCount) {
4385 _compile->print_missing_nodes();
4386 }
4387 #endif
4388
4389 if (_log != nullptr) {
4390 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4391 }
4392 }
4393
4394 //----------------------------static_subtype_check-----------------------------
4395 // Shortcut important common cases when superklass is exact:
4396 // (0) superklass is java.lang.Object (can occur in reflective code)
4397 // (1) subklass is already limited to a subtype of superklass => always ok
4398 // (2) subklass does not overlap with superklass => always fail
4399 // (3) superklass has NO subtypes and we can check with a simple compare.
4400 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4401 if (skip) {
4402 return SSC_full_test; // Let caller generate the general case.
4403 }
4404
4405 if (subk->is_java_subtype_of(superk)) {
4406 return SSC_always_true; // (0) and (1) this test cannot fail
4407 }
4408
4409 if (!subk->maybe_java_subtype_of(superk)) {
4410 return SSC_always_false; // (2) true path dead; no dynamic test needed
4411 }
4412
4413 const Type* superelem = superk;
4414 if (superk->isa_aryklassptr()) {
4415 int ignored;
4416 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4417 }
4418
4419 if (superelem->isa_instklassptr()) {
4420 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4421 if (!ik->has_subklass()) {
4422 if (!ik->is_final()) {
4423 // Add a dependency if there is a chance of a later subclass.
4424 dependencies()->assert_leaf_type(ik);
4425 }
4426 if (!superk->maybe_java_subtype_of(subk)) {
4427 return SSC_always_false;
4428 }
4429 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4430 }
4431 } else {
4432 // A primitive array type has no subtypes.
4433 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4434 }
4435
4436 return SSC_full_test;
4437 }
4438
4439 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4440 #ifdef _LP64
4441 // The scaled index operand to AddP must be a clean 64-bit value.
4442 // Java allows a 32-bit int to be incremented to a negative
4443 // value, which appears in a 64-bit register as a large
4444 // positive number. Using that large positive number as an
4445 // operand in pointer arithmetic has bad consequences.
4446 // On the other hand, 32-bit overflow is rare, and the possibility
4447 // can often be excluded, if we annotate the ConvI2L node with
4448 // a type assertion that its value is known to be a small positive
4449 // number. (The prior range check has ensured this.)
4450 // This assertion is used by ConvI2LNode::Ideal.
4451 int index_max = max_jint - 1; // array size is max_jint, index is one less
4452 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4453 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4454 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4455 #endif
4456 return idx;
4457 }
4458
4459 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4460 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4461 if (ctrl != nullptr) {
4462 // Express control dependency by a CastII node with a narrow type.
4463 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4464 // node from floating above the range check during loop optimizations. Otherwise, the
4465 // ConvI2L node may be eliminated independently of the range check, causing the data path
4466 // to become TOP while the control path is still there (although it's unreachable).
4467 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4468 value = phase->transform(value);
4469 }
4470 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4471 return phase->transform(new ConvI2LNode(value, ltype));
4472 }
4473
4474 // The message about the current inlining is accumulated in
4475 // _print_inlining_stream and transferred into the _print_inlining_list
4476 // once we know whether inlining succeeds or not. For regular
4477 // inlining, messages are appended to the buffer pointed by
4478 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4479 // a new buffer is added after _print_inlining_idx in the list. This
4480 // way we can update the inlining message for late inlining call site
4481 // when the inlining is attempted again.
4482 void Compile::print_inlining_init() {
4483 if (print_inlining() || print_intrinsics()) {
4484 // print_inlining_init is actually called several times.
4485 print_inlining_reset();
4486 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4487 }
4488 }
4489
4490 void Compile::print_inlining_reinit() {
4491 if (print_inlining() || print_intrinsics()) {
4492 print_inlining_reset();
4493 }
4494 }
4495
4496 void Compile::print_inlining_reset() {
4497 _print_inlining_stream->reset();
4498 }
4499
4500 void Compile::print_inlining_commit() {
4501 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4502 // Transfer the message from _print_inlining_stream to the current
4503 // _print_inlining_list buffer and clear _print_inlining_stream.
4504 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4505 print_inlining_reset();
4506 }
4507
4508 void Compile::print_inlining_push() {
4509 // Add new buffer to the _print_inlining_list at current position
4510 _print_inlining_idx++;
4511 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4512 }
4513
4514 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4515 return _print_inlining_list->at(_print_inlining_idx);
4516 }
4517
4518 void Compile::print_inlining_update(CallGenerator* cg) {
4519 if (print_inlining() || print_intrinsics()) {
4520 if (cg->is_late_inline()) {
4521 if (print_inlining_current()->cg() != cg &&
4522 (print_inlining_current()->cg() != nullptr ||
4523 print_inlining_current()->ss()->size() != 0)) {
4524 print_inlining_push();
4525 }
4526 print_inlining_commit();
4527 print_inlining_current()->set_cg(cg);
4528 } else {
4529 if (print_inlining_current()->cg() != nullptr) {
4530 print_inlining_push();
4531 }
4532 print_inlining_commit();
4533 }
4534 }
4535 }
4536
4537 void Compile::print_inlining_move_to(CallGenerator* cg) {
4538 // We resume inlining at a late inlining call site. Locate the
4539 // corresponding inlining buffer so that we can update it.
4540 if (print_inlining() || print_intrinsics()) {
4541 for (int i = 0; i < _print_inlining_list->length(); i++) {
4542 if (_print_inlining_list->at(i)->cg() == cg) {
4543 _print_inlining_idx = i;
4544 return;
4545 }
4546 }
4547 ShouldNotReachHere();
4548 }
4549 }
4550
4551 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4552 if (print_inlining() || print_intrinsics()) {
4553 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4554 assert(print_inlining_current()->cg() == cg, "wrong entry");
4555 // replace message with new message
4556 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4557 print_inlining_commit();
4558 print_inlining_current()->set_cg(cg);
4559 }
4560 }
4561
4562 void Compile::print_inlining_assert_ready() {
4563 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4564 }
4565
4566 void Compile::process_print_inlining() {
4567 assert(_late_inlines.length() == 0, "not drained yet");
4568 if (print_inlining() || print_intrinsics()) {
4569 ResourceMark rm;
4570 stringStream ss;
4571 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4572 for (int i = 0; i < _print_inlining_list->length(); i++) {
4573 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4574 ss.print("%s", pib->ss()->freeze());
4575 delete pib;
4576 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4577 }
4578 // Reset _print_inlining_list, it only contains destructed objects.
4579 // It is on the arena, so it will be freed when the arena is reset.
4580 _print_inlining_list = nullptr;
4581 // _print_inlining_stream won't be used anymore, either.
4582 print_inlining_reset();
4583 size_t end = ss.size();
4584 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4585 strncpy(_print_inlining_output, ss.freeze(), end+1);
4586 _print_inlining_output[end] = 0;
4587 }
4588 }
4589
4590 void Compile::dump_print_inlining() {
4591 if (_print_inlining_output != nullptr) {
4592 tty->print_raw(_print_inlining_output);
4593 }
4594 }
4595
4596 void Compile::log_late_inline(CallGenerator* cg) {
4597 if (log() != nullptr) {
4598 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4599 cg->unique_id());
4600 JVMState* p = cg->call_node()->jvms();
4601 while (p != nullptr) {
4602 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4603 p = p->caller();
4604 }
4605 log()->tail("late_inline");
4606 }
4607 }
4608
4609 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4610 log_late_inline(cg);
4611 if (log() != nullptr) {
4612 log()->inline_fail(msg);
4613 }
4614 }
4615
4616 void Compile::log_inline_id(CallGenerator* cg) {
4617 if (log() != nullptr) {
4618 // The LogCompilation tool needs a unique way to identify late
4619 // inline call sites. This id must be unique for this call site in
4620 // this compilation. Try to have it unique across compilations as
4621 // well because it can be convenient when grepping through the log
4622 // file.
4623 // Distinguish OSR compilations from others in case CICountOSR is
4624 // on.
4625 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4626 cg->set_unique_id(id);
4627 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4628 }
4629 }
4630
4631 void Compile::log_inline_failure(const char* msg) {
4632 if (C->log() != nullptr) {
4633 C->log()->inline_fail(msg);
4634 }
4635 }
4636
4637
4638 // Dump inlining replay data to the stream.
4639 // Don't change thread state and acquire any locks.
4640 void Compile::dump_inline_data(outputStream* out) {
4641 InlineTree* inl_tree = ilt();
4642 if (inl_tree != nullptr) {
4643 out->print(" inline %d", inl_tree->count());
4644 inl_tree->dump_replay_data(out);
4645 }
4646 }
4647
4648 void Compile::dump_inline_data_reduced(outputStream* out) {
4649 assert(ReplayReduce, "");
4650
4651 InlineTree* inl_tree = ilt();
4652 if (inl_tree == nullptr) {
4653 return;
4654 }
4655 // Enable iterative replay file reduction
4656 // Output "compile" lines for depth 1 subtrees,
4657 // simulating that those trees were compiled
4658 // instead of inlined.
4659 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4660 InlineTree* sub = inl_tree->subtrees().at(i);
4661 if (sub->inline_level() != 1) {
4662 continue;
4663 }
4664
4665 ciMethod* method = sub->method();
4666 int entry_bci = -1;
4667 int comp_level = env()->task()->comp_level();
4668 out->print("compile ");
4669 method->dump_name_as_ascii(out);
4670 out->print(" %d %d", entry_bci, comp_level);
4671 out->print(" inline %d", sub->count());
4672 sub->dump_replay_data(out, -1);
4673 out->cr();
4674 }
4675 }
4676
4677 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4678 if (n1->Opcode() < n2->Opcode()) return -1;
4679 else if (n1->Opcode() > n2->Opcode()) return 1;
4680
4681 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4682 for (uint i = 1; i < n1->req(); i++) {
4683 if (n1->in(i) < n2->in(i)) return -1;
4684 else if (n1->in(i) > n2->in(i)) return 1;
4685 }
4686
4687 return 0;
4688 }
4689
4690 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4691 Node* n1 = *n1p;
4692 Node* n2 = *n2p;
4693
4694 return cmp_expensive_nodes(n1, n2);
4695 }
4696
4697 void Compile::sort_expensive_nodes() {
4698 if (!expensive_nodes_sorted()) {
4699 _expensive_nodes.sort(cmp_expensive_nodes);
4700 }
4701 }
4702
4703 bool Compile::expensive_nodes_sorted() const {
4704 for (int i = 1; i < _expensive_nodes.length(); i++) {
4705 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4706 return false;
4707 }
4708 }
4709 return true;
4710 }
4711
4712 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4713 if (_expensive_nodes.length() == 0) {
4714 return false;
4715 }
4716
4717 assert(OptimizeExpensiveOps, "optimization off?");
4718
4719 // Take this opportunity to remove dead nodes from the list
4720 int j = 0;
4721 for (int i = 0; i < _expensive_nodes.length(); i++) {
4722 Node* n = _expensive_nodes.at(i);
4723 if (!n->is_unreachable(igvn)) {
4724 assert(n->is_expensive(), "should be expensive");
4725 _expensive_nodes.at_put(j, n);
4726 j++;
4727 }
4728 }
4729 _expensive_nodes.trunc_to(j);
4730
4731 // Then sort the list so that similar nodes are next to each other
4732 // and check for at least two nodes of identical kind with same data
4733 // inputs.
4734 sort_expensive_nodes();
4735
4736 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4737 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4738 return true;
4739 }
4740 }
4741
4742 return false;
4743 }
4744
4745 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4746 if (_expensive_nodes.length() == 0) {
4747 return;
4748 }
4749
4750 assert(OptimizeExpensiveOps, "optimization off?");
4751
4752 // Sort to bring similar nodes next to each other and clear the
4753 // control input of nodes for which there's only a single copy.
4754 sort_expensive_nodes();
4755
4756 int j = 0;
4757 int identical = 0;
4758 int i = 0;
4759 bool modified = false;
4760 for (; i < _expensive_nodes.length()-1; i++) {
4761 assert(j <= i, "can't write beyond current index");
4762 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4763 identical++;
4764 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4765 continue;
4766 }
4767 if (identical > 0) {
4768 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4769 identical = 0;
4770 } else {
4771 Node* n = _expensive_nodes.at(i);
4772 igvn.replace_input_of(n, 0, nullptr);
4773 igvn.hash_insert(n);
4774 modified = true;
4775 }
4776 }
4777 if (identical > 0) {
4778 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4779 } else if (_expensive_nodes.length() >= 1) {
4780 Node* n = _expensive_nodes.at(i);
4781 igvn.replace_input_of(n, 0, nullptr);
4782 igvn.hash_insert(n);
4783 modified = true;
4784 }
4785 _expensive_nodes.trunc_to(j);
4786 if (modified) {
4787 igvn.optimize();
4788 }
4789 }
4790
4791 void Compile::add_expensive_node(Node * n) {
4792 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4793 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4794 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4795 if (OptimizeExpensiveOps) {
4796 _expensive_nodes.append(n);
4797 } else {
4798 // Clear control input and let IGVN optimize expensive nodes if
4799 // OptimizeExpensiveOps is off.
4800 n->set_req(0, nullptr);
4801 }
4802 }
4803
4804 /**
4805 * Track coarsened Lock and Unlock nodes.
4806 */
4807
4808 class Lock_List : public Node_List {
4809 uint _origin_cnt;
4810 public:
4811 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4812 uint origin_cnt() const { return _origin_cnt; }
4813 };
4814
4815 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4816 int length = locks.length();
4817 if (length > 0) {
4818 // Have to keep this list until locks elimination during Macro nodes elimination.
4819 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4820 AbstractLockNode* alock = locks.at(0);
4821 BoxLockNode* box = alock->box_node()->as_BoxLock();
4822 for (int i = 0; i < length; i++) {
4823 AbstractLockNode* lock = locks.at(i);
4824 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4825 locks_list->push(lock);
4826 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4827 if (this_box != box) {
4828 // Locking regions (BoxLock) could be Unbalanced here:
4829 // - its coarsened locks were eliminated in earlier
4830 // macro nodes elimination followed by loop unroll
4831 // - it is OSR locking region (no Lock node)
4832 // Preserve Unbalanced status in such cases.
4833 if (!this_box->is_unbalanced()) {
4834 this_box->set_coarsened();
4835 }
4836 if (!box->is_unbalanced()) {
4837 box->set_coarsened();
4838 }
4839 }
4840 }
4841 _coarsened_locks.append(locks_list);
4842 }
4843 }
4844
4845 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4846 int count = coarsened_count();
4847 for (int i = 0; i < count; i++) {
4848 Node_List* locks_list = _coarsened_locks.at(i);
4849 for (uint j = 0; j < locks_list->size(); j++) {
4850 Node* lock = locks_list->at(j);
4851 assert(lock->is_AbstractLock(), "sanity");
4852 if (!useful.member(lock)) {
4853 locks_list->yank(lock);
4854 }
4855 }
4856 }
4857 }
4858
4859 void Compile::remove_coarsened_lock(Node* n) {
4860 if (n->is_AbstractLock()) {
4861 int count = coarsened_count();
4862 for (int i = 0; i < count; i++) {
4863 Node_List* locks_list = _coarsened_locks.at(i);
4864 locks_list->yank(n);
4865 }
4866 }
4867 }
4868
4869 bool Compile::coarsened_locks_consistent() {
4870 int count = coarsened_count();
4871 for (int i = 0; i < count; i++) {
4872 bool unbalanced = false;
4873 bool modified = false; // track locks kind modifications
4874 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4875 uint size = locks_list->size();
4876 if (size == 0) {
4877 unbalanced = false; // All locks were eliminated - good
4878 } else if (size != locks_list->origin_cnt()) {
4879 unbalanced = true; // Some locks were removed from list
4880 } else {
4881 for (uint j = 0; j < size; j++) {
4882 Node* lock = locks_list->at(j);
4883 // All nodes in group should have the same state (modified or not)
4884 if (!lock->as_AbstractLock()->is_coarsened()) {
4885 if (j == 0) {
4886 // first on list was modified, the rest should be too for consistency
4887 modified = true;
4888 } else if (!modified) {
4889 // this lock was modified but previous locks on the list were not
4890 unbalanced = true;
4891 break;
4892 }
4893 } else if (modified) {
4894 // previous locks on list were modified but not this lock
4895 unbalanced = true;
4896 break;
4897 }
4898 }
4899 }
4900 if (unbalanced) {
4901 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4902 #ifdef ASSERT
4903 if (PrintEliminateLocks) {
4904 tty->print_cr("=== unbalanced coarsened locks ===");
4905 for (uint l = 0; l < size; l++) {
4906 locks_list->at(l)->dump();
4907 }
4908 }
4909 #endif
4910 record_failure(C2Compiler::retry_no_locks_coarsening());
4911 return false;
4912 }
4913 }
4914 return true;
4915 }
4916
4917 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4918 // locks coarsening optimization removed Lock/Unlock nodes from them.
4919 // Such regions become unbalanced because coarsening only removes part
4920 // of Lock/Unlock nodes in region. As result we can't execute other
4921 // locks elimination optimizations which assume all code paths have
4922 // corresponding pair of Lock/Unlock nodes - they are balanced.
4923 void Compile::mark_unbalanced_boxes() const {
4924 int count = coarsened_count();
4925 for (int i = 0; i < count; i++) {
4926 Node_List* locks_list = _coarsened_locks.at(i);
4927 uint size = locks_list->size();
4928 if (size > 0) {
4929 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4930 BoxLockNode* box = alock->box_node()->as_BoxLock();
4931 if (alock->is_coarsened()) {
4932 // coarsened_locks_consistent(), which is called before this method, verifies
4933 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4934 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4935 for (uint j = 1; j < size; j++) {
4936 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4937 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4938 if (box != this_box) {
4939 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4940 box->set_unbalanced();
4941 this_box->set_unbalanced();
4942 }
4943 }
4944 }
4945 }
4946 }
4947 }
4948
4949 /**
4950 * Remove the speculative part of types and clean up the graph
4951 */
4952 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4953 if (UseTypeSpeculation) {
4954 Unique_Node_List worklist;
4955 worklist.push(root());
4956 int modified = 0;
4957 // Go over all type nodes that carry a speculative type, drop the
4958 // speculative part of the type and enqueue the node for an igvn
4959 // which may optimize it out.
4960 for (uint next = 0; next < worklist.size(); ++next) {
4961 Node *n = worklist.at(next);
4962 if (n->is_Type()) {
4963 TypeNode* tn = n->as_Type();
4964 const Type* t = tn->type();
4965 const Type* t_no_spec = t->remove_speculative();
4966 if (t_no_spec != t) {
4967 bool in_hash = igvn.hash_delete(n);
4968 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4969 tn->set_type(t_no_spec);
4970 igvn.hash_insert(n);
4971 igvn._worklist.push(n); // give it a chance to go away
4972 modified++;
4973 }
4974 }
4975 // Iterate over outs - endless loops is unreachable from below
4976 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4977 Node *m = n->fast_out(i);
4978 if (not_a_node(m)) {
4979 continue;
4980 }
4981 worklist.push(m);
4982 }
4983 }
4984 // Drop the speculative part of all types in the igvn's type table
4985 igvn.remove_speculative_types();
4986 if (modified > 0) {
4987 igvn.optimize();
4988 if (failing()) return;
4989 }
4990 #ifdef ASSERT
4991 // Verify that after the IGVN is over no speculative type has resurfaced
4992 worklist.clear();
4993 worklist.push(root());
4994 for (uint next = 0; next < worklist.size(); ++next) {
4995 Node *n = worklist.at(next);
4996 const Type* t = igvn.type_or_null(n);
4997 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4998 if (n->is_Type()) {
4999 t = n->as_Type()->type();
5000 assert(t == t->remove_speculative(), "no more speculative types");
5001 }
5002 // Iterate over outs - endless loops is unreachable from below
5003 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5004 Node *m = n->fast_out(i);
5005 if (not_a_node(m)) {
5006 continue;
5007 }
5008 worklist.push(m);
5009 }
5010 }
5011 igvn.check_no_speculative_types();
5012 #endif
5013 }
5014 }
5015
5016 // Auxiliary methods to support randomized stressing/fuzzing.
5017
5018 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5019 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5020 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5021 FLAG_SET_ERGO(StressSeed, _stress_seed);
5022 } else {
5023 _stress_seed = StressSeed;
5024 }
5025 if (_log != nullptr) {
5026 _log->elem("stress_test seed='%u'", _stress_seed);
5027 }
5028 }
5029
5030 int Compile::random() {
5031 _stress_seed = os::next_random(_stress_seed);
5032 return static_cast<int>(_stress_seed);
5033 }
5034
5035 // This method can be called the arbitrary number of times, with current count
5036 // as the argument. The logic allows selecting a single candidate from the
5037 // running list of candidates as follows:
5038 // int count = 0;
5039 // Cand* selected = null;
5040 // while(cand = cand->next()) {
5041 // if (randomized_select(++count)) {
5042 // selected = cand;
5043 // }
5044 // }
5045 //
5046 // Including count equalizes the chances any candidate is "selected".
5047 // This is useful when we don't have the complete list of candidates to choose
5048 // from uniformly. In this case, we need to adjust the randomicity of the
5049 // selection, or else we will end up biasing the selection towards the latter
5050 // candidates.
5051 //
5052 // Quick back-envelope calculation shows that for the list of n candidates
5053 // the equal probability for the candidate to persist as "best" can be
5054 // achieved by replacing it with "next" k-th candidate with the probability
5055 // of 1/k. It can be easily shown that by the end of the run, the
5056 // probability for any candidate is converged to 1/n, thus giving the
5057 // uniform distribution among all the candidates.
5058 //
5059 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5060 #define RANDOMIZED_DOMAIN_POW 29
5061 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5062 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5063 bool Compile::randomized_select(int count) {
5064 assert(count > 0, "only positive");
5065 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5066 }
5067
5068 #ifdef ASSERT
5069 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5070 bool Compile::fail_randomly() {
5071 if ((random() % StressBailoutMean) != 0) {
5072 return false;
5073 }
5074 record_failure("StressBailout");
5075 return true;
5076 }
5077
5078 bool Compile::failure_is_artificial() {
5079 assert(failing_internal(), "should be failing");
5080 return C->failure_reason_is("StressBailout");
5081 }
5082 #endif
5083
5084 CloneMap& Compile::clone_map() { return _clone_map; }
5085 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5086
5087 void NodeCloneInfo::dump_on(outputStream* st) const {
5088 st->print(" {%d:%d} ", idx(), gen());
5089 }
5090
5091 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5092 uint64_t val = value(old->_idx);
5093 NodeCloneInfo cio(val);
5094 assert(val != 0, "old node should be in the map");
5095 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5096 insert(nnn->_idx, cin.get());
5097 #ifndef PRODUCT
5098 if (is_debug()) {
5099 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5100 }
5101 #endif
5102 }
5103
5104 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5105 NodeCloneInfo cio(value(old->_idx));
5106 if (cio.get() == 0) {
5107 cio.set(old->_idx, 0);
5108 insert(old->_idx, cio.get());
5109 #ifndef PRODUCT
5110 if (is_debug()) {
5111 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5112 }
5113 #endif
5114 }
5115 clone(old, nnn, gen);
5116 }
5117
5118 int CloneMap::max_gen() const {
5119 int g = 0;
5120 DictI di(_dict);
5121 for(; di.test(); ++di) {
5122 int t = gen(di._key);
5123 if (g < t) {
5124 g = t;
5125 #ifndef PRODUCT
5126 if (is_debug()) {
5127 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5128 }
5129 #endif
5130 }
5131 }
5132 return g;
5133 }
5134
5135 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5136 uint64_t val = value(key);
5137 if (val != 0) {
5138 NodeCloneInfo ni(val);
5139 ni.dump_on(st);
5140 }
5141 }
5142
5143 void Compile::shuffle_macro_nodes() {
5144 if (_macro_nodes.length() < 2) {
5145 return;
5146 }
5147 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5148 uint j = C->random() % (i + 1);
5149 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5150 }
5151 }
5152
5153 // Move Allocate nodes to the start of the list
5154 void Compile::sort_macro_nodes() {
5155 int count = macro_count();
5156 int allocates = 0;
5157 for (int i = 0; i < count; i++) {
5158 Node* n = macro_node(i);
5159 if (n->is_Allocate()) {
5160 if (i != allocates) {
5161 Node* tmp = macro_node(allocates);
5162 _macro_nodes.at_put(allocates, n);
5163 _macro_nodes.at_put(i, tmp);
5164 }
5165 allocates++;
5166 }
5167 }
5168 }
5169
5170 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5171 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5172 EventCompilerPhase event(UNTIMED);
5173 if (event.should_commit()) {
5174 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5175 }
5176 #ifndef PRODUCT
5177 ResourceMark rm;
5178 stringStream ss;
5179 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5180 int iter = ++_igv_phase_iter[cpt];
5181 if (iter > 1) {
5182 ss.print(" %d", iter);
5183 }
5184 if (n != nullptr) {
5185 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5186 if (n->is_Call()) {
5187 CallNode* call = n->as_Call();
5188 if (call->_name != nullptr) {
5189 // E.g. uncommon traps etc.
5190 ss.print(" - %s", call->_name);
5191 } else if (call->is_CallJava()) {
5192 CallJavaNode* call_java = call->as_CallJava();
5193 if (call_java->method() != nullptr) {
5194 ss.print(" -");
5195 call_java->method()->print_short_name(&ss);
5196 }
5197 }
5198 }
5199 }
5200
5201 const char* name = ss.as_string();
5202 if (should_print_igv(level)) {
5203 _igv_printer->print_method(name, level);
5204 }
5205 if (should_print_phase(cpt)) {
5206 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5207 }
5208 #endif
5209 C->_latest_stage_start_counter.stamp();
5210 }
5211
5212 // Only used from CompileWrapper
5213 void Compile::begin_method() {
5214 #ifndef PRODUCT
5215 if (_method != nullptr && should_print_igv(1)) {
5216 _igv_printer->begin_method();
5217 }
5218 #endif
5219 C->_latest_stage_start_counter.stamp();
5220 }
5221
5222 // Only used from CompileWrapper
5223 void Compile::end_method() {
5224 EventCompilerPhase event(UNTIMED);
5225 if (event.should_commit()) {
5226 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5227 }
5228
5229 #ifndef PRODUCT
5230 if (_method != nullptr && should_print_igv(1)) {
5231 _igv_printer->end_method();
5232 }
5233 #endif
5234 }
5235
5236 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5237 #ifndef PRODUCT
5238 if (_directive->should_print_phase(cpt)) {
5239 return true;
5240 }
5241 #endif
5242 return false;
5243 }
5244
5245 bool Compile::should_print_igv(const int level) {
5246 #ifndef PRODUCT
5247 if (PrintIdealGraphLevel < 0) { // disabled by the user
5248 return false;
5249 }
5250
5251 bool need = directive()->IGVPrintLevelOption >= level;
5252 if (need && !_igv_printer) {
5253 _igv_printer = IdealGraphPrinter::printer();
5254 _igv_printer->set_compile(this);
5255 }
5256 return need;
5257 #else
5258 return false;
5259 #endif
5260 }
5261
5262 #ifndef PRODUCT
5263 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5264 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5265
5266 // Called from debugger. Prints method to the default file with the default phase name.
5267 // This works regardless of any Ideal Graph Visualizer flags set or not.
5268 void igv_print() {
5269 Compile::current()->igv_print_method_to_file();
5270 }
5271
5272 // Same as igv_print() above but with a specified phase name.
5273 void igv_print(const char* phase_name) {
5274 Compile::current()->igv_print_method_to_file(phase_name);
5275 }
5276
5277 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5278 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5279 // This works regardless of any Ideal Graph Visualizer flags set or not.
5280 void igv_print(bool network) {
5281 if (network) {
5282 Compile::current()->igv_print_method_to_network();
5283 } else {
5284 Compile::current()->igv_print_method_to_file();
5285 }
5286 }
5287
5288 // Same as igv_print(bool network) above but with a specified phase name.
5289 void igv_print(bool network, const char* phase_name) {
5290 if (network) {
5291 Compile::current()->igv_print_method_to_network(phase_name);
5292 } else {
5293 Compile::current()->igv_print_method_to_file(phase_name);
5294 }
5295 }
5296
5297 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5298 void igv_print_default() {
5299 Compile::current()->print_method(PHASE_DEBUG, 0);
5300 }
5301
5302 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5303 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5304 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5305 void igv_append() {
5306 Compile::current()->igv_print_method_to_file("Debug", true);
5307 }
5308
5309 // Same as igv_append() above but with a specified phase name.
5310 void igv_append(const char* phase_name) {
5311 Compile::current()->igv_print_method_to_file(phase_name, true);
5312 }
5313
5314 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5315 const char* file_name = "custom_debug.xml";
5316 if (_debug_file_printer == nullptr) {
5317 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5318 } else {
5319 _debug_file_printer->update_compiled_method(C->method());
5320 }
5321 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5322 _debug_file_printer->print(phase_name, (Node*)C->root());
5323 }
5324
5325 void Compile::igv_print_method_to_network(const char* phase_name) {
5326 if (_debug_network_printer == nullptr) {
5327 _debug_network_printer = new IdealGraphPrinter(C);
5328 } else {
5329 _debug_network_printer->update_compiled_method(C->method());
5330 }
5331 tty->print_cr("Method printed over network stream to IGV");
5332 _debug_network_printer->print(phase_name, (Node*)C->root());
5333 }
5334 #endif
5335
5336 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5337 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5338 return value;
5339 }
5340 Node* result = nullptr;
5341 if (bt == T_BYTE) {
5342 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5343 result = new RShiftINode(result, phase->intcon(24));
5344 } else if (bt == T_BOOLEAN) {
5345 result = new AndINode(value, phase->intcon(0xFF));
5346 } else if (bt == T_CHAR) {
5347 result = new AndINode(value,phase->intcon(0xFFFF));
5348 } else {
5349 assert(bt == T_SHORT, "unexpected narrow type");
5350 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5351 result = new RShiftINode(result, phase->intcon(16));
5352 }
5353 if (transform_res) {
5354 result = phase->transform(result);
5355 }
5356 return result;
5357 }
5358
5359 void Compile::record_method_not_compilable_oom() {
5360 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5361 }