1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compilerOracle.hpp"
37 #include "compiler/compiler_globals.hpp"
38 #include "compiler/disassembler.hpp"
39 #include "compiler/oopMap.hpp"
40 #include "gc/shared/barrierSet.hpp"
41 #include "gc/shared/c2/barrierSetC2.hpp"
42 #include "jfr/jfrEvents.hpp"
43 #include "jvm_io.h"
44 #include "memory/allocation.hpp"
45 #include "memory/resourceArea.hpp"
46 #include "opto/addnode.hpp"
47 #include "opto/block.hpp"
48 #include "opto/c2compiler.hpp"
49 #include "opto/callGenerator.hpp"
50 #include "opto/callnode.hpp"
51 #include "opto/castnode.hpp"
52 #include "opto/cfgnode.hpp"
53 #include "opto/chaitin.hpp"
54 #include "opto/compile.hpp"
55 #include "opto/connode.hpp"
56 #include "opto/convertnode.hpp"
57 #include "opto/divnode.hpp"
58 #include "opto/escape.hpp"
59 #include "opto/idealGraphPrinter.hpp"
60 #include "opto/locknode.hpp"
61 #include "opto/loopnode.hpp"
62 #include "opto/machnode.hpp"
63 #include "opto/macro.hpp"
64 #include "opto/matcher.hpp"
65 #include "opto/mathexactnode.hpp"
66 #include "opto/memnode.hpp"
67 #include "opto/mulnode.hpp"
68 #include "opto/narrowptrnode.hpp"
69 #include "opto/node.hpp"
70 #include "opto/opcodes.hpp"
71 #include "opto/output.hpp"
72 #include "opto/parse.hpp"
73 #include "opto/phaseX.hpp"
74 #include "opto/rootnode.hpp"
75 #include "opto/runtime.hpp"
76 #include "opto/stringopts.hpp"
77 #include "opto/type.hpp"
78 #include "opto/vector.hpp"
79 #include "opto/vectornode.hpp"
80 #include "runtime/globals_extension.hpp"
81 #include "runtime/sharedRuntime.hpp"
82 #include "runtime/signature.hpp"
83 #include "runtime/stubRoutines.hpp"
84 #include "runtime/timer.hpp"
85 #include "utilities/align.hpp"
86 #include "utilities/copy.hpp"
87 #include "utilities/macros.hpp"
88 #include "utilities/resourceHash.hpp"
89
90 // -------------------- Compile::mach_constant_base_node -----------------------
91 // Constant table base node singleton.
92 MachConstantBaseNode* Compile::mach_constant_base_node() {
93 if (_mach_constant_base_node == nullptr) {
94 _mach_constant_base_node = new MachConstantBaseNode();
95 _mach_constant_base_node->add_req(C->root());
96 }
97 return _mach_constant_base_node;
98 }
99
100
101 /// Support for intrinsics.
102
103 // Return the index at which m must be inserted (or already exists).
104 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
105 class IntrinsicDescPair {
106 private:
107 ciMethod* _m;
108 bool _is_virtual;
109 public:
110 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
111 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
112 ciMethod* m= elt->method();
113 ciMethod* key_m = key->_m;
114 if (key_m < m) return -1;
115 else if (key_m > m) return 1;
116 else {
117 bool is_virtual = elt->is_virtual();
118 bool key_virtual = key->_is_virtual;
119 if (key_virtual < is_virtual) return -1;
120 else if (key_virtual > is_virtual) return 1;
121 else return 0;
122 }
123 }
124 };
125 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
126 #ifdef ASSERT
127 for (int i = 1; i < _intrinsics.length(); i++) {
128 CallGenerator* cg1 = _intrinsics.at(i-1);
129 CallGenerator* cg2 = _intrinsics.at(i);
130 assert(cg1->method() != cg2->method()
131 ? cg1->method() < cg2->method()
132 : cg1->is_virtual() < cg2->is_virtual(),
133 "compiler intrinsics list must stay sorted");
134 }
135 #endif
136 IntrinsicDescPair pair(m, is_virtual);
137 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
138 }
139
140 void Compile::register_intrinsic(CallGenerator* cg) {
141 bool found = false;
142 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
143 assert(!found, "registering twice");
144 _intrinsics.insert_before(index, cg);
145 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
146 }
147
148 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
149 assert(m->is_loaded(), "don't try this on unloaded methods");
150 if (_intrinsics.length() > 0) {
151 bool found = false;
152 int index = intrinsic_insertion_index(m, is_virtual, found);
153 if (found) {
154 return _intrinsics.at(index);
155 }
156 }
157 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
158 if (m->intrinsic_id() != vmIntrinsics::_none &&
159 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
160 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
161 if (cg != nullptr) {
162 // Save it for next time:
163 register_intrinsic(cg);
164 return cg;
165 } else {
166 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
167 }
168 }
169 return nullptr;
170 }
171
172 // Compile::make_vm_intrinsic is defined in library_call.cpp.
173
174 #ifndef PRODUCT
175 // statistics gathering...
176
177 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
179
180 inline int as_int(vmIntrinsics::ID id) {
181 return vmIntrinsics::as_int(id);
182 }
183
184 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
185 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
186 int oflags = _intrinsic_hist_flags[as_int(id)];
187 assert(flags != 0, "what happened?");
188 if (is_virtual) {
189 flags |= _intrinsic_virtual;
190 }
191 bool changed = (flags != oflags);
192 if ((flags & _intrinsic_worked) != 0) {
193 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
194 if (count == 1) {
195 changed = true; // first time
196 }
197 // increment the overall count also:
198 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
199 }
200 if (changed) {
201 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
202 // Something changed about the intrinsic's virtuality.
203 if ((flags & _intrinsic_virtual) != 0) {
204 // This is the first use of this intrinsic as a virtual call.
205 if (oflags != 0) {
206 // We already saw it as a non-virtual, so note both cases.
207 flags |= _intrinsic_both;
208 }
209 } else if ((oflags & _intrinsic_both) == 0) {
210 // This is the first use of this intrinsic as a non-virtual
211 flags |= _intrinsic_both;
212 }
213 }
214 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
215 }
216 // update the overall flags also:
217 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
218 return changed;
219 }
220
221 static char* format_flags(int flags, char* buf) {
222 buf[0] = 0;
223 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
224 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
225 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
226 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
227 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
228 if (buf[0] == 0) strcat(buf, ",");
229 assert(buf[0] == ',', "must be");
230 return &buf[1];
231 }
232
233 void Compile::print_intrinsic_statistics() {
234 char flagsbuf[100];
235 ttyLocker ttyl;
236 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
237 tty->print_cr("Compiler intrinsic usage:");
238 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
239 if (total == 0) total = 1; // avoid div0 in case of no successes
240 #define PRINT_STAT_LINE(name, c, f) \
241 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
242 for (auto id : EnumRange<vmIntrinsicID>{}) {
243 int flags = _intrinsic_hist_flags[as_int(id)];
244 juint count = _intrinsic_hist_count[as_int(id)];
245 if ((flags | count) != 0) {
246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
247 }
248 }
249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
250 if (xtty != nullptr) xtty->tail("statistics");
251 }
252
253 void Compile::print_statistics() {
254 { ttyLocker ttyl;
255 if (xtty != nullptr) xtty->head("statistics type='opto'");
256 Parse::print_statistics();
257 PhaseStringOpts::print_statistics();
258 PhaseCCP::print_statistics();
259 PhaseRegAlloc::print_statistics();
260 PhaseOutput::print_statistics();
261 PhasePeephole::print_statistics();
262 PhaseIdealLoop::print_statistics();
263 ConnectionGraph::print_statistics();
264 PhaseMacroExpand::print_statistics();
265 if (xtty != nullptr) xtty->tail("statistics");
266 }
267 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
268 // put this under its own <statistics> element.
269 print_intrinsic_statistics();
270 }
271 }
272 #endif //PRODUCT
273
274 void Compile::gvn_replace_by(Node* n, Node* nn) {
275 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
276 Node* use = n->last_out(i);
277 bool is_in_table = initial_gvn()->hash_delete(use);
278 uint uses_found = 0;
279 for (uint j = 0; j < use->len(); j++) {
280 if (use->in(j) == n) {
281 if (j < use->req())
282 use->set_req(j, nn);
283 else
284 use->set_prec(j, nn);
285 uses_found++;
286 }
287 }
288 if (is_in_table) {
289 // reinsert into table
290 initial_gvn()->hash_find_insert(use);
291 }
292 record_for_igvn(use);
293 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
294 i -= uses_found; // we deleted 1 or more copies of this edge
295 }
296 }
297
298
299 // Identify all nodes that are reachable from below, useful.
300 // Use breadth-first pass that records state in a Unique_Node_List,
301 // recursive traversal is slower.
302 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
303 int estimated_worklist_size = live_nodes();
304 useful.map( estimated_worklist_size, nullptr ); // preallocate space
305
306 // Initialize worklist
307 if (root() != nullptr) { useful.push(root()); }
308 // If 'top' is cached, declare it useful to preserve cached node
309 if (cached_top_node()) { useful.push(cached_top_node()); }
310
311 // Push all useful nodes onto the list, breadthfirst
312 for( uint next = 0; next < useful.size(); ++next ) {
313 assert( next < unique(), "Unique useful nodes < total nodes");
314 Node *n = useful.at(next);
315 uint max = n->len();
316 for( uint i = 0; i < max; ++i ) {
317 Node *m = n->in(i);
318 if (not_a_node(m)) continue;
319 useful.push(m);
320 }
321 }
322 }
323
324 // Update dead_node_list with any missing dead nodes using useful
325 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
326 void Compile::update_dead_node_list(Unique_Node_List &useful) {
327 uint max_idx = unique();
328 VectorSet& useful_node_set = useful.member_set();
329
330 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
331 // If node with index node_idx is not in useful set,
332 // mark it as dead in dead node list.
333 if (!useful_node_set.test(node_idx)) {
334 record_dead_node(node_idx);
335 }
336 }
337 }
338
339 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
340 int shift = 0;
341 for (int i = 0; i < inlines->length(); i++) {
342 CallGenerator* cg = inlines->at(i);
343 if (useful.member(cg->call_node())) {
344 if (shift > 0) {
345 inlines->at_put(i - shift, cg);
346 }
347 } else {
348 shift++; // skip over the dead element
349 }
350 }
351 if (shift > 0) {
352 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
353 }
354 }
355
356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
357 assert(dead != nullptr && dead->is_Call(), "sanity");
358 int found = 0;
359 for (int i = 0; i < inlines->length(); i++) {
360 if (inlines->at(i)->call_node() == dead) {
361 inlines->remove_at(i);
362 found++;
363 NOT_DEBUG( break; ) // elements are unique, so exit early
364 }
365 }
366 assert(found <= 1, "not unique");
367 }
368
369 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
370 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
371 for (int i = node_list.length() - 1; i >= 0; i--) {
372 N* node = node_list.at(i);
373 if (!useful.member(node)) {
374 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
375 }
376 }
377 }
378
379 void Compile::remove_useless_node(Node* dead) {
380 remove_modified_node(dead);
381
382 // Constant node that has no out-edges and has only one in-edge from
383 // root is usually dead. However, sometimes reshaping walk makes
384 // it reachable by adding use edges. So, we will NOT count Con nodes
385 // as dead to be conservative about the dead node count at any
386 // given time.
387 if (!dead->is_Con()) {
388 record_dead_node(dead->_idx);
389 }
390 if (dead->is_macro()) {
391 remove_macro_node(dead);
392 }
393 if (dead->is_expensive()) {
394 remove_expensive_node(dead);
395 }
396 if (dead->Opcode() == Op_Opaque4) {
397 remove_template_assertion_predicate_opaq(dead);
398 }
399 if (dead->is_ParsePredicate()) {
400 remove_parse_predicate(dead->as_ParsePredicate());
401 }
402 if (dead->for_post_loop_opts_igvn()) {
403 remove_from_post_loop_opts_igvn(dead);
404 }
405 if (dead->is_Call()) {
406 remove_useless_late_inlines( &_late_inlines, dead);
407 remove_useless_late_inlines( &_string_late_inlines, dead);
408 remove_useless_late_inlines( &_boxing_late_inlines, dead);
409 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
410
411 if (dead->is_CallStaticJava()) {
412 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
413 }
414 }
415 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
416 bs->unregister_potential_barrier_node(dead);
417 }
418
419 // Disconnect all useless nodes by disconnecting those at the boundary.
420 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
421 uint next = 0;
422 while (next < useful.size()) {
423 Node *n = useful.at(next++);
424 if (n->is_SafePoint()) {
425 // We're done with a parsing phase. Replaced nodes are not valid
426 // beyond that point.
427 n->as_SafePoint()->delete_replaced_nodes();
428 }
429 // Use raw traversal of out edges since this code removes out edges
430 int max = n->outcnt();
431 for (int j = 0; j < max; ++j) {
432 Node* child = n->raw_out(j);
433 if (!useful.member(child)) {
434 assert(!child->is_top() || child != top(),
435 "If top is cached in Compile object it is in useful list");
436 // Only need to remove this out-edge to the useless node
437 n->raw_del_out(j);
438 --j;
439 --max;
440 }
441 }
442 if (n->outcnt() == 1 && n->has_special_unique_user()) {
443 assert(useful.member(n->unique_out()), "do not push a useless node");
444 worklist.push(n->unique_out());
445 }
446 }
447
448 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
449 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
450 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
451 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
452 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
453 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
454 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
455 #ifdef ASSERT
456 if (_modified_nodes != nullptr) {
457 _modified_nodes->remove_useless_nodes(useful.member_set());
458 }
459 #endif
460
461 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
462 bs->eliminate_useless_gc_barriers(useful, this);
463 // clean up the late inline lists
464 remove_useless_late_inlines( &_late_inlines, useful);
465 remove_useless_late_inlines( &_string_late_inlines, useful);
466 remove_useless_late_inlines( &_boxing_late_inlines, useful);
467 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
468 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
469 }
470
471 // ============================================================================
472 //------------------------------CompileWrapper---------------------------------
473 class CompileWrapper : public StackObj {
474 Compile *const _compile;
475 public:
476 CompileWrapper(Compile* compile);
477
478 ~CompileWrapper();
479 };
480
481 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
482 // the Compile* pointer is stored in the current ciEnv:
483 ciEnv* env = compile->env();
484 assert(env == ciEnv::current(), "must already be a ciEnv active");
485 assert(env->compiler_data() == nullptr, "compile already active?");
486 env->set_compiler_data(compile);
487 assert(compile == Compile::current(), "sanity");
488
489 compile->set_type_dict(nullptr);
490 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
491 compile->clone_map().set_clone_idx(0);
492 compile->set_type_last_size(0);
493 compile->set_last_tf(nullptr, nullptr);
494 compile->set_indexSet_arena(nullptr);
495 compile->set_indexSet_free_block_list(nullptr);
496 compile->init_type_arena();
497 Type::Initialize(compile);
498 _compile->begin_method();
499 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
500 }
501 CompileWrapper::~CompileWrapper() {
502 // simulate crash during compilation
503 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
504
505 _compile->end_method();
506 _compile->env()->set_compiler_data(nullptr);
507 }
508
509
510 //----------------------------print_compile_messages---------------------------
511 void Compile::print_compile_messages() {
512 #ifndef PRODUCT
513 // Check if recompiling
514 if (!subsume_loads() && PrintOpto) {
515 // Recompiling without allowing machine instructions to subsume loads
516 tty->print_cr("*********************************************************");
517 tty->print_cr("** Bailout: Recompile without subsuming loads **");
518 tty->print_cr("*********************************************************");
519 }
520 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
521 // Recompiling without escape analysis
522 tty->print_cr("*********************************************************");
523 tty->print_cr("** Bailout: Recompile without escape analysis **");
524 tty->print_cr("*********************************************************");
525 }
526 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
527 // Recompiling without iterative escape analysis
528 tty->print_cr("*********************************************************");
529 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
530 tty->print_cr("*********************************************************");
531 }
532 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
533 // Recompiling without reducing allocation merges
534 tty->print_cr("*********************************************************");
535 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
536 tty->print_cr("*********************************************************");
537 }
538 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
539 // Recompiling without boxing elimination
540 tty->print_cr("*********************************************************");
541 tty->print_cr("** Bailout: Recompile without boxing elimination **");
542 tty->print_cr("*********************************************************");
543 }
544 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
545 // Recompiling without locks coarsening
546 tty->print_cr("*********************************************************");
547 tty->print_cr("** Bailout: Recompile without locks coarsening **");
548 tty->print_cr("*********************************************************");
549 }
550 if (env()->break_at_compile()) {
551 // Open the debugger when compiling this method.
552 tty->print("### Breaking when compiling: ");
553 method()->print_short_name();
554 tty->cr();
555 BREAKPOINT;
556 }
557
558 if( PrintOpto ) {
559 if (is_osr_compilation()) {
560 tty->print("[OSR]%3d", _compile_id);
561 } else {
562 tty->print("%3d", _compile_id);
563 }
564 }
565 #endif
566 }
567
568 #ifndef PRODUCT
569 void Compile::print_ideal_ir(const char* phase_name) {
570 // keep the following output all in one block
571 // This output goes directly to the tty, not the compiler log.
572 // To enable tools to match it up with the compilation activity,
573 // be sure to tag this tty output with the compile ID.
574
575 // Node dumping can cause a safepoint, which can break the tty lock.
576 // Buffer all node dumps, so that all safepoints happen before we lock.
577 ResourceMark rm;
578 stringStream ss;
579
580 if (_output == nullptr) {
581 ss.print_cr("AFTER: %s", phase_name);
582 // Print out all nodes in ascending order of index.
583 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
584 } else {
585 // Dump the node blockwise if we have a scheduling
586 _output->print_scheduling(&ss);
587 }
588
589 // Check that the lock is not broken by a safepoint.
590 NoSafepointVerifier nsv;
591 ttyLocker ttyl;
592 if (xtty != nullptr) {
593 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
594 compile_id(),
595 is_osr_compilation() ? " compile_kind='osr'" : "",
596 phase_name);
597 }
598
599 tty->print("%s", ss.as_string());
600
601 if (xtty != nullptr) {
602 xtty->tail("ideal");
603 }
604 }
605 #endif
606
607 // ============================================================================
608 //------------------------------Compile standard-------------------------------
609
610 // Compile a method. entry_bci is -1 for normal compilations and indicates
611 // the continuation bci for on stack replacement.
612
613
614 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
615 Options options, DirectiveSet* directive)
616 : Phase(Compiler),
617 _compile_id(ci_env->compile_id()),
618 _options(options),
619 _method(target),
620 _entry_bci(osr_bci),
621 _ilt(nullptr),
622 _stub_function(nullptr),
623 _stub_name(nullptr),
624 _stub_entry_point(nullptr),
625 _max_node_limit(MaxNodeLimit),
626 _post_loop_opts_phase(false),
627 _allow_macro_nodes(true),
628 _inlining_progress(false),
629 _inlining_incrementally(false),
630 _do_cleanup(false),
631 _has_reserved_stack_access(target->has_reserved_stack_access()),
632 #ifndef PRODUCT
633 _igv_idx(0),
634 _trace_opto_output(directive->TraceOptoOutputOption),
635 #endif
636 _has_method_handle_invokes(false),
637 _clinit_barrier_on_entry(false),
638 _stress_seed(0),
639 _comp_arena(mtCompiler),
640 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
641 _env(ci_env),
642 _directive(directive),
643 _log(ci_env->log()),
644 _first_failure_details(nullptr),
645 _intrinsics (comp_arena(), 0, 0, nullptr),
646 _macro_nodes (comp_arena(), 8, 0, nullptr),
647 _parse_predicates (comp_arena(), 8, 0, nullptr),
648 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
649 _expensive_nodes (comp_arena(), 8, 0, nullptr),
650 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
651 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
652 _coarsened_locks (comp_arena(), 8, 0, nullptr),
653 _congraph(nullptr),
654 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
655 _unique(0),
656 _dead_node_count(0),
657 _dead_node_list(comp_arena()),
658 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
659 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
660 _node_arena(&_node_arena_one),
661 _mach_constant_base_node(nullptr),
662 _Compile_types(mtCompiler),
663 _initial_gvn(nullptr),
664 _igvn_worklist(nullptr),
665 _types(nullptr),
666 _node_hash(nullptr),
667 _late_inlines(comp_arena(), 2, 0, nullptr),
668 _string_late_inlines(comp_arena(), 2, 0, nullptr),
669 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
670 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
671 _late_inlines_pos(0),
672 _number_of_mh_late_inlines(0),
673 _oom(false),
674 _print_inlining_stream(new (mtCompiler) stringStream()),
675 _print_inlining_list(nullptr),
676 _print_inlining_idx(0),
677 _print_inlining_output(nullptr),
678 _replay_inline_data(nullptr),
679 _java_calls(0),
680 _inner_loops(0),
681 _interpreter_frame_size(0),
682 _output(nullptr)
683 #ifndef PRODUCT
684 , _in_dump_cnt(0)
685 #endif
686 {
687 C = this;
688 CompileWrapper cw(this);
689
690 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
691 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
692
693 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
694 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
695 // We can always print a disassembly, either abstract (hex dump) or
696 // with the help of a suitable hsdis library. Thus, we should not
697 // couple print_assembly and print_opto_assembly controls.
698 // But: always print opto and regular assembly on compile command 'print'.
699 bool print_assembly = directive->PrintAssemblyOption;
700 set_print_assembly(print_opto_assembly || print_assembly);
701 #else
702 set_print_assembly(false); // must initialize.
703 #endif
704
705 #ifndef PRODUCT
706 set_parsed_irreducible_loop(false);
707 #endif
708
709 if (directive->ReplayInlineOption) {
710 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
711 }
712 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
713 set_print_intrinsics(directive->PrintIntrinsicsOption);
714 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
715
716 if (ProfileTraps) {
717 // Make sure the method being compiled gets its own MDO,
718 // so we can at least track the decompile_count().
719 method()->ensure_method_data();
720 }
721
722 Init(/*do_aliasing=*/ true);
723
724 print_compile_messages();
725
726 _ilt = InlineTree::build_inline_tree_root();
727
728 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
729 assert(num_alias_types() >= AliasIdxRaw, "");
730
731 #define MINIMUM_NODE_HASH 1023
732
733 // GVN that will be run immediately on new nodes
734 uint estimated_size = method()->code_size()*4+64;
735 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
736 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
737 _types = new (comp_arena()) Type_Array(comp_arena());
738 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
739 PhaseGVN gvn;
740 set_initial_gvn(&gvn);
741
742 print_inlining_init();
743 { // Scope for timing the parser
744 TracePhase tp("parse", &timers[_t_parser]);
745
746 // Put top into the hash table ASAP.
747 initial_gvn()->transform(top());
748
749 // Set up tf(), start(), and find a CallGenerator.
750 CallGenerator* cg = nullptr;
751 if (is_osr_compilation()) {
752 const TypeTuple *domain = StartOSRNode::osr_domain();
753 const TypeTuple *range = TypeTuple::make_range(method()->signature());
754 init_tf(TypeFunc::make(domain, range));
755 StartNode* s = new StartOSRNode(root(), domain);
756 initial_gvn()->set_type_bottom(s);
757 init_start(s);
758 cg = CallGenerator::for_osr(method(), entry_bci());
759 } else {
760 // Normal case.
761 init_tf(TypeFunc::make(method()));
762 StartNode* s = new StartNode(root(), tf()->domain());
763 initial_gvn()->set_type_bottom(s);
764 init_start(s);
765 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
766 // With java.lang.ref.reference.get() we must go through the
767 // intrinsic - even when get() is the root
768 // method of the compile - so that, if necessary, the value in
769 // the referent field of the reference object gets recorded by
770 // the pre-barrier code.
771 cg = find_intrinsic(method(), false);
772 }
773 if (cg == nullptr) {
774 float past_uses = method()->interpreter_invocation_count();
775 float expected_uses = past_uses;
776 cg = CallGenerator::for_inline(method(), expected_uses);
777 }
778 }
779 if (failing()) return;
780 if (cg == nullptr) {
781 const char* reason = InlineTree::check_can_parse(method());
782 assert(reason != nullptr, "expect reason for parse failure");
783 stringStream ss;
784 ss.print("cannot parse method: %s", reason);
785 record_method_not_compilable(ss.as_string());
786 return;
787 }
788
789 gvn.set_type(root(), root()->bottom_type());
790
791 JVMState* jvms = build_start_state(start(), tf());
792 if ((jvms = cg->generate(jvms)) == nullptr) {
793 assert(failure_reason() != nullptr, "expect reason for parse failure");
794 stringStream ss;
795 ss.print("method parse failed: %s", failure_reason());
796 record_method_not_compilable(ss.as_string());
797 return;
798 }
799 GraphKit kit(jvms);
800
801 if (!kit.stopped()) {
802 // Accept return values, and transfer control we know not where.
803 // This is done by a special, unique ReturnNode bound to root.
804 return_values(kit.jvms());
805 }
806
807 if (kit.has_exceptions()) {
808 // Any exceptions that escape from this call must be rethrown
809 // to whatever caller is dynamically above us on the stack.
810 // This is done by a special, unique RethrowNode bound to root.
811 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
812 }
813
814 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
815
816 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
817 inline_string_calls(true);
818 }
819
820 if (failing()) return;
821
822 // Remove clutter produced by parsing.
823 if (!failing()) {
824 ResourceMark rm;
825 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
826 }
827 }
828
829 // Note: Large methods are capped off in do_one_bytecode().
830 if (failing()) return;
831
832 // After parsing, node notes are no longer automagic.
833 // They must be propagated by register_new_node_with_optimizer(),
834 // clone(), or the like.
835 set_default_node_notes(nullptr);
836
837 #ifndef PRODUCT
838 if (should_print_igv(1)) {
839 _igv_printer->print_inlining();
840 }
841 #endif
842
843 if (failing()) return;
844 NOT_PRODUCT( verify_graph_edges(); )
845
846 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
847 StressIncrementalInlining || StressMacroExpansion) {
848 initialize_stress_seed(directive);
849 }
850
851 // Now optimize
852 Optimize();
853 if (failing()) return;
854 NOT_PRODUCT( verify_graph_edges(); )
855
856 #ifndef PRODUCT
857 if (should_print_ideal()) {
858 print_ideal_ir("print_ideal");
859 }
860 #endif
861
862 #ifdef ASSERT
863 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
864 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
865 #endif
866
867 // Dump compilation data to replay it.
868 if (directive->DumpReplayOption) {
869 env()->dump_replay_data(_compile_id);
870 }
871 if (directive->DumpInlineOption && (ilt() != nullptr)) {
872 env()->dump_inline_data(_compile_id);
873 }
874
875 // Now that we know the size of all the monitors we can add a fixed slot
876 // for the original deopt pc.
877 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
878 set_fixed_slots(next_slot);
879
880 // Compute when to use implicit null checks. Used by matching trap based
881 // nodes and NullCheck optimization.
882 set_allowed_deopt_reasons();
883
884 // Now generate code
885 Code_Gen();
886 }
887
888 //------------------------------Compile----------------------------------------
889 // Compile a runtime stub
890 Compile::Compile( ciEnv* ci_env,
891 TypeFunc_generator generator,
892 address stub_function,
893 const char *stub_name,
894 int is_fancy_jump,
895 bool pass_tls,
896 bool return_pc,
897 DirectiveSet* directive)
898 : Phase(Compiler),
899 _compile_id(0),
900 _options(Options::for_runtime_stub()),
901 _method(nullptr),
902 _entry_bci(InvocationEntryBci),
903 _stub_function(stub_function),
904 _stub_name(stub_name),
905 _stub_entry_point(nullptr),
906 _max_node_limit(MaxNodeLimit),
907 _post_loop_opts_phase(false),
908 _allow_macro_nodes(true),
909 _inlining_progress(false),
910 _inlining_incrementally(false),
911 _has_reserved_stack_access(false),
912 #ifndef PRODUCT
913 _igv_idx(0),
914 _trace_opto_output(directive->TraceOptoOutputOption),
915 #endif
916 _has_method_handle_invokes(false),
917 _clinit_barrier_on_entry(false),
918 _stress_seed(0),
919 _comp_arena(mtCompiler),
920 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
921 _env(ci_env),
922 _directive(directive),
923 _log(ci_env->log()),
924 _first_failure_details(nullptr),
925 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
926 _congraph(nullptr),
927 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
928 _unique(0),
929 _dead_node_count(0),
930 _dead_node_list(comp_arena()),
931 _node_arena_one(mtCompiler),
932 _node_arena_two(mtCompiler),
933 _node_arena(&_node_arena_one),
934 _mach_constant_base_node(nullptr),
935 _Compile_types(mtCompiler),
936 _initial_gvn(nullptr),
937 _igvn_worklist(nullptr),
938 _types(nullptr),
939 _node_hash(nullptr),
940 _number_of_mh_late_inlines(0),
941 _oom(false),
942 _print_inlining_stream(new (mtCompiler) stringStream()),
943 _print_inlining_list(nullptr),
944 _print_inlining_idx(0),
945 _print_inlining_output(nullptr),
946 _replay_inline_data(nullptr),
947 _java_calls(0),
948 _inner_loops(0),
949 _interpreter_frame_size(0),
950 _output(nullptr),
951 #ifndef PRODUCT
952 _in_dump_cnt(0),
953 #endif
954 _allowed_reasons(0) {
955 C = this;
956
957 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
958 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
959
960 #ifndef PRODUCT
961 set_print_assembly(PrintFrameConverterAssembly);
962 set_parsed_irreducible_loop(false);
963 #else
964 set_print_assembly(false); // Must initialize.
965 #endif
966 set_has_irreducible_loop(false); // no loops
967
968 CompileWrapper cw(this);
969 Init(/*do_aliasing=*/ false);
970 init_tf((*generator)());
971
972 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
973 _types = new (comp_arena()) Type_Array(comp_arena());
974 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
975
976 if (StressLCM || StressGCM) {
977 initialize_stress_seed(directive);
978 }
979
980 {
981 PhaseGVN gvn;
982 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
983 gvn.transform(top());
984
985 GraphKit kit;
986 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
987 }
988
989 NOT_PRODUCT( verify_graph_edges(); )
990
991 Code_Gen();
992 }
993
994 Compile::~Compile() {
995 delete _print_inlining_stream;
996 delete _first_failure_details;
997 };
998
999 //------------------------------Init-------------------------------------------
1000 // Prepare for a single compilation
1001 void Compile::Init(bool aliasing) {
1002 _do_aliasing = aliasing;
1003 _unique = 0;
1004 _regalloc = nullptr;
1005
1006 _tf = nullptr; // filled in later
1007 _top = nullptr; // cached later
1008 _matcher = nullptr; // filled in later
1009 _cfg = nullptr; // filled in later
1010
1011 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1012
1013 _node_note_array = nullptr;
1014 _default_node_notes = nullptr;
1015 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1016
1017 _immutable_memory = nullptr; // filled in at first inquiry
1018
1019 #ifdef ASSERT
1020 _phase_optimize_finished = false;
1021 _exception_backedge = false;
1022 _type_verify = nullptr;
1023 #endif
1024
1025 // Globally visible Nodes
1026 // First set TOP to null to give safe behavior during creation of RootNode
1027 set_cached_top_node(nullptr);
1028 set_root(new RootNode());
1029 // Now that you have a Root to point to, create the real TOP
1030 set_cached_top_node( new ConNode(Type::TOP) );
1031 set_recent_alloc(nullptr, nullptr);
1032
1033 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1034 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1035 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1036 env()->set_dependencies(new Dependencies(env()));
1037
1038 _fixed_slots = 0;
1039 set_has_split_ifs(false);
1040 set_has_loops(false); // first approximation
1041 set_has_stringbuilder(false);
1042 set_has_boxed_value(false);
1043 _trap_can_recompile = false; // no traps emitted yet
1044 _major_progress = true; // start out assuming good things will happen
1045 set_has_unsafe_access(false);
1046 set_max_vector_size(0);
1047 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1048 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1049 set_decompile_count(0);
1050
1051 #ifndef PRODUCT
1052 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1053 #endif
1054
1055 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1056 _loop_opts_cnt = LoopOptsCount;
1057 set_do_inlining(Inline);
1058 set_max_inline_size(MaxInlineSize);
1059 set_freq_inline_size(FreqInlineSize);
1060 set_do_scheduling(OptoScheduling);
1061
1062 set_do_vector_loop(false);
1063 set_has_monitors(false);
1064
1065 if (AllowVectorizeOnDemand) {
1066 if (has_method() && _directive->VectorizeOption) {
1067 set_do_vector_loop(true);
1068 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());})
1069 } else if (has_method() && method()->name() != 0 &&
1070 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1071 set_do_vector_loop(true);
1072 }
1073 }
1074 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1075 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());})
1076
1077 _max_node_limit = _directive->MaxNodeLimitOption;
1078
1079 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1080 set_clinit_barrier_on_entry(true);
1081 }
1082 if (debug_info()->recording_non_safepoints()) {
1083 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1084 (comp_arena(), 8, 0, nullptr));
1085 set_default_node_notes(Node_Notes::make(this));
1086 }
1087
1088 const int grow_ats = 16;
1089 _max_alias_types = grow_ats;
1090 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1091 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1092 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1093 {
1094 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1095 }
1096 // Initialize the first few types.
1097 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1098 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1099 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1100 _num_alias_types = AliasIdxRaw+1;
1101 // Zero out the alias type cache.
1102 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1103 // A null adr_type hits in the cache right away. Preload the right answer.
1104 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1105 }
1106
1107 //---------------------------init_start----------------------------------------
1108 // Install the StartNode on this compile object.
1109 void Compile::init_start(StartNode* s) {
1110 if (failing())
1111 return; // already failing
1112 assert(s == start(), "");
1113 }
1114
1115 /**
1116 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1117 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1118 * the ideal graph.
1119 */
1120 StartNode* Compile::start() const {
1121 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1122 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1123 Node* start = root()->fast_out(i);
1124 if (start->is_Start()) {
1125 return start->as_Start();
1126 }
1127 }
1128 fatal("Did not find Start node!");
1129 return nullptr;
1130 }
1131
1132 //-------------------------------immutable_memory-------------------------------------
1133 // Access immutable memory
1134 Node* Compile::immutable_memory() {
1135 if (_immutable_memory != nullptr) {
1136 return _immutable_memory;
1137 }
1138 StartNode* s = start();
1139 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1140 Node *p = s->fast_out(i);
1141 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1142 _immutable_memory = p;
1143 return _immutable_memory;
1144 }
1145 }
1146 ShouldNotReachHere();
1147 return nullptr;
1148 }
1149
1150 //----------------------set_cached_top_node------------------------------------
1151 // Install the cached top node, and make sure Node::is_top works correctly.
1152 void Compile::set_cached_top_node(Node* tn) {
1153 if (tn != nullptr) verify_top(tn);
1154 Node* old_top = _top;
1155 _top = tn;
1156 // Calling Node::setup_is_top allows the nodes the chance to adjust
1157 // their _out arrays.
1158 if (_top != nullptr) _top->setup_is_top();
1159 if (old_top != nullptr) old_top->setup_is_top();
1160 assert(_top == nullptr || top()->is_top(), "");
1161 }
1162
1163 #ifdef ASSERT
1164 uint Compile::count_live_nodes_by_graph_walk() {
1165 Unique_Node_List useful(comp_arena());
1166 // Get useful node list by walking the graph.
1167 identify_useful_nodes(useful);
1168 return useful.size();
1169 }
1170
1171 void Compile::print_missing_nodes() {
1172
1173 // Return if CompileLog is null and PrintIdealNodeCount is false.
1174 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1175 return;
1176 }
1177
1178 // This is an expensive function. It is executed only when the user
1179 // specifies VerifyIdealNodeCount option or otherwise knows the
1180 // additional work that needs to be done to identify reachable nodes
1181 // by walking the flow graph and find the missing ones using
1182 // _dead_node_list.
1183
1184 Unique_Node_List useful(comp_arena());
1185 // Get useful node list by walking the graph.
1186 identify_useful_nodes(useful);
1187
1188 uint l_nodes = C->live_nodes();
1189 uint l_nodes_by_walk = useful.size();
1190
1191 if (l_nodes != l_nodes_by_walk) {
1192 if (_log != nullptr) {
1193 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1194 _log->stamp();
1195 _log->end_head();
1196 }
1197 VectorSet& useful_member_set = useful.member_set();
1198 int last_idx = l_nodes_by_walk;
1199 for (int i = 0; i < last_idx; i++) {
1200 if (useful_member_set.test(i)) {
1201 if (_dead_node_list.test(i)) {
1202 if (_log != nullptr) {
1203 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1204 }
1205 if (PrintIdealNodeCount) {
1206 // Print the log message to tty
1207 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1208 useful.at(i)->dump();
1209 }
1210 }
1211 }
1212 else if (! _dead_node_list.test(i)) {
1213 if (_log != nullptr) {
1214 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1215 }
1216 if (PrintIdealNodeCount) {
1217 // Print the log message to tty
1218 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1219 }
1220 }
1221 }
1222 if (_log != nullptr) {
1223 _log->tail("mismatched_nodes");
1224 }
1225 }
1226 }
1227 void Compile::record_modified_node(Node* n) {
1228 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1229 _modified_nodes->push(n);
1230 }
1231 }
1232
1233 void Compile::remove_modified_node(Node* n) {
1234 if (_modified_nodes != nullptr) {
1235 _modified_nodes->remove(n);
1236 }
1237 }
1238 #endif
1239
1240 #ifndef PRODUCT
1241 void Compile::verify_top(Node* tn) const {
1242 if (tn != nullptr) {
1243 assert(tn->is_Con(), "top node must be a constant");
1244 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1245 assert(tn->in(0) != nullptr, "must have live top node");
1246 }
1247 }
1248 #endif
1249
1250
1251 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1252
1253 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1254 guarantee(arr != nullptr, "");
1255 int num_blocks = arr->length();
1256 if (grow_by < num_blocks) grow_by = num_blocks;
1257 int num_notes = grow_by * _node_notes_block_size;
1258 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1259 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1260 while (num_notes > 0) {
1261 arr->append(notes);
1262 notes += _node_notes_block_size;
1263 num_notes -= _node_notes_block_size;
1264 }
1265 assert(num_notes == 0, "exact multiple, please");
1266 }
1267
1268 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1269 if (source == nullptr || dest == nullptr) return false;
1270
1271 if (dest->is_Con())
1272 return false; // Do not push debug info onto constants.
1273
1274 #ifdef ASSERT
1275 // Leave a bread crumb trail pointing to the original node:
1276 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1277 dest->set_debug_orig(source);
1278 }
1279 #endif
1280
1281 if (node_note_array() == nullptr)
1282 return false; // Not collecting any notes now.
1283
1284 // This is a copy onto a pre-existing node, which may already have notes.
1285 // If both nodes have notes, do not overwrite any pre-existing notes.
1286 Node_Notes* source_notes = node_notes_at(source->_idx);
1287 if (source_notes == nullptr || source_notes->is_clear()) return false;
1288 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1289 if (dest_notes == nullptr || dest_notes->is_clear()) {
1290 return set_node_notes_at(dest->_idx, source_notes);
1291 }
1292
1293 Node_Notes merged_notes = (*source_notes);
1294 // The order of operations here ensures that dest notes will win...
1295 merged_notes.update_from(dest_notes);
1296 return set_node_notes_at(dest->_idx, &merged_notes);
1297 }
1298
1299
1300 //--------------------------allow_range_check_smearing-------------------------
1301 // Gating condition for coalescing similar range checks.
1302 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1303 // single covering check that is at least as strong as any of them.
1304 // If the optimization succeeds, the simplified (strengthened) range check
1305 // will always succeed. If it fails, we will deopt, and then give up
1306 // on the optimization.
1307 bool Compile::allow_range_check_smearing() const {
1308 // If this method has already thrown a range-check,
1309 // assume it was because we already tried range smearing
1310 // and it failed.
1311 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1312 return !already_trapped;
1313 }
1314
1315
1316 //------------------------------flatten_alias_type-----------------------------
1317 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1318 assert(do_aliasing(), "Aliasing should be enabled");
1319 int offset = tj->offset();
1320 TypePtr::PTR ptr = tj->ptr();
1321
1322 // Known instance (scalarizable allocation) alias only with itself.
1323 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1324 tj->is_oopptr()->is_known_instance();
1325
1326 // Process weird unsafe references.
1327 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1328 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1329 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1330 tj = TypeOopPtr::BOTTOM;
1331 ptr = tj->ptr();
1332 offset = tj->offset();
1333 }
1334
1335 // Array pointers need some flattening
1336 const TypeAryPtr* ta = tj->isa_aryptr();
1337 if (ta && ta->is_stable()) {
1338 // Erase stability property for alias analysis.
1339 tj = ta = ta->cast_to_stable(false);
1340 }
1341 if( ta && is_known_inst ) {
1342 if ( offset != Type::OffsetBot &&
1343 offset > arrayOopDesc::length_offset_in_bytes() ) {
1344 offset = Type::OffsetBot; // Flatten constant access into array body only
1345 tj = ta = ta->
1346 remove_speculative()->
1347 cast_to_ptr_type(ptr)->
1348 with_offset(offset);
1349 }
1350 } else if (ta) {
1351 // For arrays indexed by constant indices, we flatten the alias
1352 // space to include all of the array body. Only the header, klass
1353 // and array length can be accessed un-aliased.
1354 if( offset != Type::OffsetBot ) {
1355 if( ta->const_oop() ) { // MethodData* or Method*
1356 offset = Type::OffsetBot; // Flatten constant access into array body
1357 tj = ta = ta->
1358 remove_speculative()->
1359 cast_to_ptr_type(ptr)->
1360 cast_to_exactness(false)->
1361 with_offset(offset);
1362 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1363 // range is OK as-is.
1364 tj = ta = TypeAryPtr::RANGE;
1365 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1366 tj = TypeInstPtr::KLASS; // all klass loads look alike
1367 ta = TypeAryPtr::RANGE; // generic ignored junk
1368 ptr = TypePtr::BotPTR;
1369 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1370 tj = TypeInstPtr::MARK;
1371 ta = TypeAryPtr::RANGE; // generic ignored junk
1372 ptr = TypePtr::BotPTR;
1373 } else { // Random constant offset into array body
1374 offset = Type::OffsetBot; // Flatten constant access into array body
1375 tj = ta = ta->
1376 remove_speculative()->
1377 cast_to_ptr_type(ptr)->
1378 cast_to_exactness(false)->
1379 with_offset(offset);
1380 }
1381 }
1382 // Arrays of fixed size alias with arrays of unknown size.
1383 if (ta->size() != TypeInt::POS) {
1384 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1385 tj = ta = ta->
1386 remove_speculative()->
1387 cast_to_ptr_type(ptr)->
1388 with_ary(tary)->
1389 cast_to_exactness(false);
1390 }
1391 // Arrays of known objects become arrays of unknown objects.
1392 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1393 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1394 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1395 }
1396 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1397 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1398 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1399 }
1400 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1401 // cannot be distinguished by bytecode alone.
1402 if (ta->elem() == TypeInt::BOOL) {
1403 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1404 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1405 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1406 }
1407 // During the 2nd round of IterGVN, NotNull castings are removed.
1408 // Make sure the Bottom and NotNull variants alias the same.
1409 // Also, make sure exact and non-exact variants alias the same.
1410 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1411 tj = ta = ta->
1412 remove_speculative()->
1413 cast_to_ptr_type(TypePtr::BotPTR)->
1414 cast_to_exactness(false)->
1415 with_offset(offset);
1416 }
1417 }
1418
1419 // Oop pointers need some flattening
1420 const TypeInstPtr *to = tj->isa_instptr();
1421 if (to && to != TypeOopPtr::BOTTOM) {
1422 ciInstanceKlass* ik = to->instance_klass();
1423 if( ptr == TypePtr::Constant ) {
1424 if (ik != ciEnv::current()->Class_klass() ||
1425 offset < ik->layout_helper_size_in_bytes()) {
1426 // No constant oop pointers (such as Strings); they alias with
1427 // unknown strings.
1428 assert(!is_known_inst, "not scalarizable allocation");
1429 tj = to = to->
1430 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1431 remove_speculative()->
1432 cast_to_ptr_type(TypePtr::BotPTR)->
1433 cast_to_exactness(false);
1434 }
1435 } else if( is_known_inst ) {
1436 tj = to; // Keep NotNull and klass_is_exact for instance type
1437 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1438 // During the 2nd round of IterGVN, NotNull castings are removed.
1439 // Make sure the Bottom and NotNull variants alias the same.
1440 // Also, make sure exact and non-exact variants alias the same.
1441 tj = to = to->
1442 remove_speculative()->
1443 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1444 cast_to_ptr_type(TypePtr::BotPTR)->
1445 cast_to_exactness(false);
1446 }
1447 if (to->speculative() != nullptr) {
1448 tj = to = to->remove_speculative();
1449 }
1450 // Canonicalize the holder of this field
1451 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1452 // First handle header references such as a LoadKlassNode, even if the
1453 // object's klass is unloaded at compile time (4965979).
1454 if (!is_known_inst) { // Do it only for non-instance types
1455 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1456 }
1457 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1458 // Static fields are in the space above the normal instance
1459 // fields in the java.lang.Class instance.
1460 if (ik != ciEnv::current()->Class_klass()) {
1461 to = nullptr;
1462 tj = TypeOopPtr::BOTTOM;
1463 offset = tj->offset();
1464 }
1465 } else {
1466 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1467 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1468 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1469 if( is_known_inst ) {
1470 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id());
1471 } else {
1472 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset);
1473 }
1474 }
1475 }
1476 }
1477
1478 // Klass pointers to object array klasses need some flattening
1479 const TypeKlassPtr *tk = tj->isa_klassptr();
1480 if( tk ) {
1481 // If we are referencing a field within a Klass, we need
1482 // to assume the worst case of an Object. Both exact and
1483 // inexact types must flatten to the same alias class so
1484 // use NotNull as the PTR.
1485 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1486 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1487 env()->Object_klass(),
1488 offset);
1489 }
1490
1491 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1492 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1493 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1494 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1495 } else {
1496 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1497 }
1498 }
1499
1500 // Check for precise loads from the primary supertype array and force them
1501 // to the supertype cache alias index. Check for generic array loads from
1502 // the primary supertype array and also force them to the supertype cache
1503 // alias index. Since the same load can reach both, we need to merge
1504 // these 2 disparate memories into the same alias class. Since the
1505 // primary supertype array is read-only, there's no chance of confusion
1506 // where we bypass an array load and an array store.
1507 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1508 if (offset == Type::OffsetBot ||
1509 (offset >= primary_supers_offset &&
1510 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1511 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1512 offset = in_bytes(Klass::secondary_super_cache_offset());
1513 tj = tk = tk->with_offset(offset);
1514 }
1515 }
1516
1517 // Flatten all Raw pointers together.
1518 if (tj->base() == Type::RawPtr)
1519 tj = TypeRawPtr::BOTTOM;
1520
1521 if (tj->base() == Type::AnyPtr)
1522 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1523
1524 offset = tj->offset();
1525 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1526
1527 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1528 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1529 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1530 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1531 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1532 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1533 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1534 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1535 assert( tj->ptr() != TypePtr::TopPTR &&
1536 tj->ptr() != TypePtr::AnyNull &&
1537 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1538 // assert( tj->ptr() != TypePtr::Constant ||
1539 // tj->base() == Type::RawPtr ||
1540 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1541
1542 return tj;
1543 }
1544
1545 void Compile::AliasType::Init(int i, const TypePtr* at) {
1546 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1547 _index = i;
1548 _adr_type = at;
1549 _field = nullptr;
1550 _element = nullptr;
1551 _is_rewritable = true; // default
1552 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1553 if (atoop != nullptr && atoop->is_known_instance()) {
1554 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1555 _general_index = Compile::current()->get_alias_index(gt);
1556 } else {
1557 _general_index = 0;
1558 }
1559 }
1560
1561 BasicType Compile::AliasType::basic_type() const {
1562 if (element() != nullptr) {
1563 const Type* element = adr_type()->is_aryptr()->elem();
1564 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1565 } if (field() != nullptr) {
1566 return field()->layout_type();
1567 } else {
1568 return T_ILLEGAL; // unknown
1569 }
1570 }
1571
1572 //---------------------------------print_on------------------------------------
1573 #ifndef PRODUCT
1574 void Compile::AliasType::print_on(outputStream* st) {
1575 if (index() < 10)
1576 st->print("@ <%d> ", index());
1577 else st->print("@ <%d>", index());
1578 st->print(is_rewritable() ? " " : " RO");
1579 int offset = adr_type()->offset();
1580 if (offset == Type::OffsetBot)
1581 st->print(" +any");
1582 else st->print(" +%-3d", offset);
1583 st->print(" in ");
1584 adr_type()->dump_on(st);
1585 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1586 if (field() != nullptr && tjp) {
1587 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1588 tjp->offset() != field()->offset_in_bytes()) {
1589 st->print(" != ");
1590 field()->print();
1591 st->print(" ***");
1592 }
1593 }
1594 }
1595
1596 void print_alias_types() {
1597 Compile* C = Compile::current();
1598 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1599 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1600 C->alias_type(idx)->print_on(tty);
1601 tty->cr();
1602 }
1603 }
1604 #endif
1605
1606
1607 //----------------------------probe_alias_cache--------------------------------
1608 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1609 intptr_t key = (intptr_t) adr_type;
1610 key ^= key >> logAliasCacheSize;
1611 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1612 }
1613
1614
1615 //-----------------------------grow_alias_types--------------------------------
1616 void Compile::grow_alias_types() {
1617 const int old_ats = _max_alias_types; // how many before?
1618 const int new_ats = old_ats; // how many more?
1619 const int grow_ats = old_ats+new_ats; // how many now?
1620 _max_alias_types = grow_ats;
1621 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1622 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1623 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1624 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1625 }
1626
1627
1628 //--------------------------------find_alias_type------------------------------
1629 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1630 if (!do_aliasing()) {
1631 return alias_type(AliasIdxBot);
1632 }
1633
1634 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1635 if (ace->_adr_type == adr_type) {
1636 return alias_type(ace->_index);
1637 }
1638
1639 // Handle special cases.
1640 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1641 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1642
1643 // Do it the slow way.
1644 const TypePtr* flat = flatten_alias_type(adr_type);
1645
1646 #ifdef ASSERT
1647 {
1648 ResourceMark rm;
1649 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1650 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1651 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1652 Type::str(adr_type));
1653 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1654 const TypeOopPtr* foop = flat->is_oopptr();
1655 // Scalarizable allocations have exact klass always.
1656 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1657 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1658 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1659 Type::str(foop), Type::str(xoop));
1660 }
1661 }
1662 #endif
1663
1664 int idx = AliasIdxTop;
1665 for (int i = 0; i < num_alias_types(); i++) {
1666 if (alias_type(i)->adr_type() == flat) {
1667 idx = i;
1668 break;
1669 }
1670 }
1671
1672 if (idx == AliasIdxTop) {
1673 if (no_create) return nullptr;
1674 // Grow the array if necessary.
1675 if (_num_alias_types == _max_alias_types) grow_alias_types();
1676 // Add a new alias type.
1677 idx = _num_alias_types++;
1678 _alias_types[idx]->Init(idx, flat);
1679 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1680 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1681 if (flat->isa_instptr()) {
1682 if (flat->offset() == java_lang_Class::klass_offset()
1683 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1684 alias_type(idx)->set_rewritable(false);
1685 }
1686 if (flat->isa_aryptr()) {
1687 #ifdef ASSERT
1688 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1689 // (T_BYTE has the weakest alignment and size restrictions...)
1690 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1691 #endif
1692 if (flat->offset() == TypePtr::OffsetBot) {
1693 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1694 }
1695 }
1696 if (flat->isa_klassptr()) {
1697 if (UseCompactObjectHeaders) {
1698 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1699 alias_type(idx)->set_rewritable(false);
1700 }
1701 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1702 alias_type(idx)->set_rewritable(false);
1703 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1704 alias_type(idx)->set_rewritable(false);
1705 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1706 alias_type(idx)->set_rewritable(false);
1707 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1708 alias_type(idx)->set_rewritable(false);
1709 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1710 alias_type(idx)->set_rewritable(false);
1711 }
1712 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1713 // but the base pointer type is not distinctive enough to identify
1714 // references into JavaThread.)
1715
1716 // Check for final fields.
1717 const TypeInstPtr* tinst = flat->isa_instptr();
1718 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1719 ciField* field;
1720 if (tinst->const_oop() != nullptr &&
1721 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1722 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1723 // static field
1724 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1725 field = k->get_field_by_offset(tinst->offset(), true);
1726 } else {
1727 ciInstanceKlass *k = tinst->instance_klass();
1728 field = k->get_field_by_offset(tinst->offset(), false);
1729 }
1730 assert(field == nullptr ||
1731 original_field == nullptr ||
1732 (field->holder() == original_field->holder() &&
1733 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1734 field->is_static() == original_field->is_static()), "wrong field?");
1735 // Set field() and is_rewritable() attributes.
1736 if (field != nullptr) alias_type(idx)->set_field(field);
1737 }
1738 }
1739
1740 // Fill the cache for next time.
1741 ace->_adr_type = adr_type;
1742 ace->_index = idx;
1743 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1744
1745 // Might as well try to fill the cache for the flattened version, too.
1746 AliasCacheEntry* face = probe_alias_cache(flat);
1747 if (face->_adr_type == nullptr) {
1748 face->_adr_type = flat;
1749 face->_index = idx;
1750 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1751 }
1752
1753 return alias_type(idx);
1754 }
1755
1756
1757 Compile::AliasType* Compile::alias_type(ciField* field) {
1758 const TypeOopPtr* t;
1759 if (field->is_static())
1760 t = TypeInstPtr::make(field->holder()->java_mirror());
1761 else
1762 t = TypeOopPtr::make_from_klass_raw(field->holder());
1763 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1764 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1765 return atp;
1766 }
1767
1768
1769 //------------------------------have_alias_type--------------------------------
1770 bool Compile::have_alias_type(const TypePtr* adr_type) {
1771 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1772 if (ace->_adr_type == adr_type) {
1773 return true;
1774 }
1775
1776 // Handle special cases.
1777 if (adr_type == nullptr) return true;
1778 if (adr_type == TypePtr::BOTTOM) return true;
1779
1780 return find_alias_type(adr_type, true, nullptr) != nullptr;
1781 }
1782
1783 //-----------------------------must_alias--------------------------------------
1784 // True if all values of the given address type are in the given alias category.
1785 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1786 if (alias_idx == AliasIdxBot) return true; // the universal category
1787 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1788 if (alias_idx == AliasIdxTop) return false; // the empty category
1789 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1790
1791 // the only remaining possible overlap is identity
1792 int adr_idx = get_alias_index(adr_type);
1793 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1794 assert(adr_idx == alias_idx ||
1795 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1796 && adr_type != TypeOopPtr::BOTTOM),
1797 "should not be testing for overlap with an unsafe pointer");
1798 return adr_idx == alias_idx;
1799 }
1800
1801 //------------------------------can_alias--------------------------------------
1802 // True if any values of the given address type are in the given alias category.
1803 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1804 if (alias_idx == AliasIdxTop) return false; // the empty category
1805 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1806 // Known instance doesn't alias with bottom memory
1807 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1808 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1809
1810 // the only remaining possible overlap is identity
1811 int adr_idx = get_alias_index(adr_type);
1812 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1813 return adr_idx == alias_idx;
1814 }
1815
1816 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1817 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1818 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1819 if (parse_predicate_count() == 0) {
1820 return;
1821 }
1822 for (int i = 0; i < parse_predicate_count(); i++) {
1823 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1824 parse_predicate->mark_useless();
1825 igvn._worklist.push(parse_predicate);
1826 }
1827 _parse_predicates.clear();
1828 }
1829
1830 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1831 if (!n->for_post_loop_opts_igvn()) {
1832 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1833 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1834 _for_post_loop_igvn.append(n);
1835 }
1836 }
1837
1838 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1839 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1840 _for_post_loop_igvn.remove(n);
1841 }
1842
1843 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1844 // Verify that all previous optimizations produced a valid graph
1845 // at least to this point, even if no loop optimizations were done.
1846 PhaseIdealLoop::verify(igvn);
1847
1848 C->set_post_loop_opts_phase(); // no more loop opts allowed
1849
1850 assert(!C->major_progress(), "not cleared");
1851
1852 if (_for_post_loop_igvn.length() > 0) {
1853 while (_for_post_loop_igvn.length() > 0) {
1854 Node* n = _for_post_loop_igvn.pop();
1855 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1856 igvn._worklist.push(n);
1857 }
1858 igvn.optimize();
1859 if (failing()) return;
1860 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1861 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1862
1863 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1864 if (C->major_progress()) {
1865 C->clear_major_progress(); // ensure that major progress is now clear
1866 }
1867 }
1868 }
1869
1870 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1871 if (OptimizeUnstableIf) {
1872 _unstable_if_traps.append(trap);
1873 }
1874 }
1875
1876 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1877 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1878 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1879 Node* n = trap->uncommon_trap();
1880 if (!useful.member(n)) {
1881 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1882 }
1883 }
1884 }
1885
1886 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1887 // or fold-compares case. Return true if succeed or not found.
1888 //
1889 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1890 // false and ask fold-compares to yield.
1891 //
1892 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1893 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1894 // when deoptimization does happen.
1895 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1896 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1897 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1898 if (trap->uncommon_trap() == unc) {
1899 if (yield && trap->modified()) {
1900 return false;
1901 }
1902 _unstable_if_traps.delete_at(i);
1903 break;
1904 }
1905 }
1906 return true;
1907 }
1908
1909 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1910 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1911 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1912 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1913 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1914 CallStaticJavaNode* unc = trap->uncommon_trap();
1915 int next_bci = trap->next_bci();
1916 bool modified = trap->modified();
1917
1918 if (next_bci != -1 && !modified) {
1919 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1920 JVMState* jvms = unc->jvms();
1921 ciMethod* method = jvms->method();
1922 ciBytecodeStream iter(method);
1923
1924 iter.force_bci(jvms->bci());
1925 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1926 Bytecodes::Code c = iter.cur_bc();
1927 Node* lhs = nullptr;
1928 Node* rhs = nullptr;
1929 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1930 lhs = unc->peek_operand(0);
1931 rhs = unc->peek_operand(1);
1932 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1933 lhs = unc->peek_operand(0);
1934 }
1935
1936 ResourceMark rm;
1937 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1938 assert(live_locals.is_valid(), "broken liveness info");
1939 int len = (int)live_locals.size();
1940
1941 for (int i = 0; i < len; i++) {
1942 Node* local = unc->local(jvms, i);
1943 // kill local using the liveness of next_bci.
1944 // give up when the local looks like an operand to secure reexecution.
1945 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1946 uint idx = jvms->locoff() + i;
1947 #ifdef ASSERT
1948 if (PrintOpto && Verbose) {
1949 tty->print("[unstable_if] kill local#%d: ", idx);
1950 local->dump();
1951 tty->cr();
1952 }
1953 #endif
1954 igvn.replace_input_of(unc, idx, top());
1955 modified = true;
1956 }
1957 }
1958 }
1959
1960 // keep the mondified trap for late query
1961 if (modified) {
1962 trap->set_modified();
1963 } else {
1964 _unstable_if_traps.delete_at(i);
1965 }
1966 }
1967 igvn.optimize();
1968 }
1969
1970 // StringOpts and late inlining of string methods
1971 void Compile::inline_string_calls(bool parse_time) {
1972 {
1973 // remove useless nodes to make the usage analysis simpler
1974 ResourceMark rm;
1975 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1976 }
1977
1978 {
1979 ResourceMark rm;
1980 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1981 PhaseStringOpts pso(initial_gvn());
1982 print_method(PHASE_AFTER_STRINGOPTS, 3);
1983 }
1984
1985 // now inline anything that we skipped the first time around
1986 if (!parse_time) {
1987 _late_inlines_pos = _late_inlines.length();
1988 }
1989
1990 while (_string_late_inlines.length() > 0) {
1991 CallGenerator* cg = _string_late_inlines.pop();
1992 cg->do_late_inline();
1993 if (failing()) return;
1994 }
1995 _string_late_inlines.trunc_to(0);
1996 }
1997
1998 // Late inlining of boxing methods
1999 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2000 if (_boxing_late_inlines.length() > 0) {
2001 assert(has_boxed_value(), "inconsistent");
2002
2003 set_inlining_incrementally(true);
2004
2005 igvn_worklist()->ensure_empty(); // should be done with igvn
2006
2007 _late_inlines_pos = _late_inlines.length();
2008
2009 while (_boxing_late_inlines.length() > 0) {
2010 CallGenerator* cg = _boxing_late_inlines.pop();
2011 cg->do_late_inline();
2012 if (failing()) return;
2013 }
2014 _boxing_late_inlines.trunc_to(0);
2015
2016 inline_incrementally_cleanup(igvn);
2017
2018 set_inlining_incrementally(false);
2019 }
2020 }
2021
2022 bool Compile::inline_incrementally_one() {
2023 assert(IncrementalInline, "incremental inlining should be on");
2024
2025 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2026
2027 set_inlining_progress(false);
2028 set_do_cleanup(false);
2029
2030 for (int i = 0; i < _late_inlines.length(); i++) {
2031 _late_inlines_pos = i+1;
2032 CallGenerator* cg = _late_inlines.at(i);
2033 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2034 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2035 cg->do_late_inline();
2036 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2037 if (failing()) {
2038 return false;
2039 } else if (inlining_progress()) {
2040 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2041 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2042 break; // process one call site at a time
2043 }
2044 } else {
2045 // Ignore late inline direct calls when inlining is not allowed.
2046 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2047 }
2048 }
2049 // Remove processed elements.
2050 _late_inlines.remove_till(_late_inlines_pos);
2051 _late_inlines_pos = 0;
2052
2053 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2054
2055 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2056
2057 set_inlining_progress(false);
2058 set_do_cleanup(false);
2059
2060 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2061 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2062 }
2063
2064 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2065 {
2066 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2067 ResourceMark rm;
2068 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2069 }
2070 {
2071 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2072 igvn.reset_from_gvn(initial_gvn());
2073 igvn.optimize();
2074 if (failing()) return;
2075 }
2076 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2077 }
2078
2079 // Perform incremental inlining until bound on number of live nodes is reached
2080 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2081 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2082
2083 set_inlining_incrementally(true);
2084 uint low_live_nodes = 0;
2085
2086 while (_late_inlines.length() > 0) {
2087 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2088 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2089 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2090 // PhaseIdealLoop is expensive so we only try it once we are
2091 // out of live nodes and we only try it again if the previous
2092 // helped got the number of nodes down significantly
2093 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2094 if (failing()) return;
2095 low_live_nodes = live_nodes();
2096 _major_progress = true;
2097 }
2098
2099 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2100 bool do_print_inlining = print_inlining() || print_intrinsics();
2101 if (do_print_inlining || log() != nullptr) {
2102 // Print inlining message for candidates that we couldn't inline for lack of space.
2103 for (int i = 0; i < _late_inlines.length(); i++) {
2104 CallGenerator* cg = _late_inlines.at(i);
2105 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2106 if (do_print_inlining) {
2107 cg->print_inlining_late(InliningResult::FAILURE, msg);
2108 }
2109 log_late_inline_failure(cg, msg);
2110 }
2111 }
2112 break; // finish
2113 }
2114 }
2115
2116 igvn_worklist()->ensure_empty(); // should be done with igvn
2117
2118 while (inline_incrementally_one()) {
2119 assert(!failing(), "inconsistent");
2120 }
2121 if (failing()) return;
2122
2123 inline_incrementally_cleanup(igvn);
2124
2125 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2126
2127 if (failing()) return;
2128
2129 if (_late_inlines.length() == 0) {
2130 break; // no more progress
2131 }
2132 }
2133
2134 igvn_worklist()->ensure_empty(); // should be done with igvn
2135
2136 if (_string_late_inlines.length() > 0) {
2137 assert(has_stringbuilder(), "inconsistent");
2138
2139 inline_string_calls(false);
2140
2141 if (failing()) return;
2142
2143 inline_incrementally_cleanup(igvn);
2144 }
2145
2146 set_inlining_incrementally(false);
2147 }
2148
2149 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2150 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2151 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2152 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2153 // as if "inlining_incrementally() == true" were set.
2154 assert(inlining_incrementally() == false, "not allowed");
2155 assert(_modified_nodes == nullptr, "not allowed");
2156 assert(_late_inlines.length() > 0, "sanity");
2157
2158 while (_late_inlines.length() > 0) {
2159 igvn_worklist()->ensure_empty(); // should be done with igvn
2160
2161 while (inline_incrementally_one()) {
2162 assert(!failing(), "inconsistent");
2163 }
2164 if (failing()) return;
2165
2166 inline_incrementally_cleanup(igvn);
2167 }
2168 }
2169
2170 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2171 if (_loop_opts_cnt > 0) {
2172 while (major_progress() && (_loop_opts_cnt > 0)) {
2173 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2174 PhaseIdealLoop::optimize(igvn, mode);
2175 _loop_opts_cnt--;
2176 if (failing()) return false;
2177 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2178 }
2179 }
2180 return true;
2181 }
2182
2183 // Remove edges from "root" to each SafePoint at a backward branch.
2184 // They were inserted during parsing (see add_safepoint()) to make
2185 // infinite loops without calls or exceptions visible to root, i.e.,
2186 // useful.
2187 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2188 Node *r = root();
2189 if (r != nullptr) {
2190 for (uint i = r->req(); i < r->len(); ++i) {
2191 Node *n = r->in(i);
2192 if (n != nullptr && n->is_SafePoint()) {
2193 r->rm_prec(i);
2194 if (n->outcnt() == 0) {
2195 igvn.remove_dead_node(n);
2196 }
2197 --i;
2198 }
2199 }
2200 // Parsing may have added top inputs to the root node (Path
2201 // leading to the Halt node proven dead). Make sure we get a
2202 // chance to clean them up.
2203 igvn._worklist.push(r);
2204 igvn.optimize();
2205 }
2206 }
2207
2208 //------------------------------Optimize---------------------------------------
2209 // Given a graph, optimize it.
2210 void Compile::Optimize() {
2211 TracePhase tp("optimizer", &timers[_t_optimizer]);
2212
2213 #ifndef PRODUCT
2214 if (env()->break_at_compile()) {
2215 BREAKPOINT;
2216 }
2217
2218 #endif
2219
2220 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2221 #ifdef ASSERT
2222 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2223 #endif
2224
2225 ResourceMark rm;
2226
2227 print_inlining_reinit();
2228
2229 NOT_PRODUCT( verify_graph_edges(); )
2230
2231 print_method(PHASE_AFTER_PARSING, 1);
2232
2233 {
2234 // Iterative Global Value Numbering, including ideal transforms
2235 // Initialize IterGVN with types and values from parse-time GVN
2236 PhaseIterGVN igvn(initial_gvn());
2237 #ifdef ASSERT
2238 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2239 #endif
2240 {
2241 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2242 igvn.optimize();
2243 }
2244
2245 if (failing()) return;
2246
2247 print_method(PHASE_ITER_GVN1, 2);
2248
2249 process_for_unstable_if_traps(igvn);
2250
2251 if (failing()) return;
2252
2253 inline_incrementally(igvn);
2254
2255 print_method(PHASE_INCREMENTAL_INLINE, 2);
2256
2257 if (failing()) return;
2258
2259 if (eliminate_boxing()) {
2260 // Inline valueOf() methods now.
2261 inline_boxing_calls(igvn);
2262
2263 if (failing()) return;
2264
2265 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2266 inline_incrementally(igvn);
2267 }
2268
2269 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2270
2271 if (failing()) return;
2272 }
2273
2274 // Remove the speculative part of types and clean up the graph from
2275 // the extra CastPP nodes whose only purpose is to carry them. Do
2276 // that early so that optimizations are not disrupted by the extra
2277 // CastPP nodes.
2278 remove_speculative_types(igvn);
2279
2280 if (failing()) return;
2281
2282 // No more new expensive nodes will be added to the list from here
2283 // so keep only the actual candidates for optimizations.
2284 cleanup_expensive_nodes(igvn);
2285
2286 if (failing()) return;
2287
2288 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2289 if (EnableVectorSupport && has_vbox_nodes()) {
2290 TracePhase tp("", &timers[_t_vector]);
2291 PhaseVector pv(igvn);
2292 pv.optimize_vector_boxes();
2293 if (failing()) return;
2294 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2295 }
2296 assert(!has_vbox_nodes(), "sanity");
2297
2298 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2299 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2300 igvn_worklist()->ensure_empty(); // should be done with igvn
2301 {
2302 ResourceMark rm;
2303 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2304 }
2305 igvn.reset_from_gvn(initial_gvn());
2306 igvn.optimize();
2307 if (failing()) return;
2308 }
2309
2310 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2311 // safepoints
2312 remove_root_to_sfpts_edges(igvn);
2313
2314 if (failing()) return;
2315
2316 // Perform escape analysis
2317 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2318 if (has_loops()) {
2319 // Cleanup graph (remove dead nodes).
2320 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2321 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2322 if (failing()) return;
2323 }
2324 bool progress;
2325 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2326 do {
2327 ConnectionGraph::do_analysis(this, &igvn);
2328
2329 if (failing()) return;
2330
2331 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2332
2333 // Optimize out fields loads from scalar replaceable allocations.
2334 igvn.optimize();
2335 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2336
2337 if (failing()) return;
2338
2339 if (congraph() != nullptr && macro_count() > 0) {
2340 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2341 PhaseMacroExpand mexp(igvn);
2342 mexp.eliminate_macro_nodes();
2343 if (failing()) return;
2344
2345 igvn.set_delay_transform(false);
2346 igvn.optimize();
2347 if (failing()) return;
2348
2349 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2350 }
2351
2352 ConnectionGraph::verify_ram_nodes(this, root());
2353 if (failing()) return;
2354
2355 progress = do_iterative_escape_analysis() &&
2356 (macro_count() < mcount) &&
2357 ConnectionGraph::has_candidates(this);
2358 // Try again if candidates exist and made progress
2359 // by removing some allocations and/or locks.
2360 } while (progress);
2361 }
2362
2363 // Loop transforms on the ideal graph. Range Check Elimination,
2364 // peeling, unrolling, etc.
2365
2366 // Set loop opts counter
2367 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2368 {
2369 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2370 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2371 _loop_opts_cnt--;
2372 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2373 if (failing()) return;
2374 }
2375 // Loop opts pass if partial peeling occurred in previous pass
2376 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2377 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2378 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2379 _loop_opts_cnt--;
2380 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2381 if (failing()) return;
2382 }
2383 // Loop opts pass for loop-unrolling before CCP
2384 if(major_progress() && (_loop_opts_cnt > 0)) {
2385 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2386 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2387 _loop_opts_cnt--;
2388 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2389 }
2390 if (!failing()) {
2391 // Verify that last round of loop opts produced a valid graph
2392 PhaseIdealLoop::verify(igvn);
2393 }
2394 }
2395 if (failing()) return;
2396
2397 // Conditional Constant Propagation;
2398 print_method(PHASE_BEFORE_CCP1, 2);
2399 PhaseCCP ccp( &igvn );
2400 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2401 {
2402 TracePhase tp("ccp", &timers[_t_ccp]);
2403 ccp.do_transform();
2404 }
2405 print_method(PHASE_CCP1, 2);
2406
2407 assert( true, "Break here to ccp.dump_old2new_map()");
2408
2409 // Iterative Global Value Numbering, including ideal transforms
2410 {
2411 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2412 igvn.reset_from_igvn(&ccp);
2413 igvn.optimize();
2414 }
2415 print_method(PHASE_ITER_GVN2, 2);
2416
2417 if (failing()) return;
2418
2419 // Loop transforms on the ideal graph. Range Check Elimination,
2420 // peeling, unrolling, etc.
2421 if (!optimize_loops(igvn, LoopOptsDefault)) {
2422 return;
2423 }
2424
2425 if (failing()) return;
2426
2427 C->clear_major_progress(); // ensure that major progress is now clear
2428
2429 process_for_post_loop_opts_igvn(igvn);
2430
2431 if (failing()) return;
2432
2433 #ifdef ASSERT
2434 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2435 #endif
2436
2437 {
2438 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2439 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2440 PhaseMacroExpand mex(igvn);
2441 if (mex.expand_macro_nodes()) {
2442 assert(failing(), "must bail out w/ explicit message");
2443 return;
2444 }
2445 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2446 }
2447
2448 {
2449 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2450 if (bs->expand_barriers(this, igvn)) {
2451 assert(failing(), "must bail out w/ explicit message");
2452 return;
2453 }
2454 print_method(PHASE_BARRIER_EXPANSION, 2);
2455 }
2456
2457 if (C->max_vector_size() > 0) {
2458 C->optimize_logic_cones(igvn);
2459 igvn.optimize();
2460 if (failing()) return;
2461 }
2462
2463 DEBUG_ONLY( _modified_nodes = nullptr; )
2464
2465 assert(igvn._worklist.size() == 0, "not empty");
2466
2467 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2468
2469 if (_late_inlines.length() > 0) {
2470 // More opportunities to optimize virtual and MH calls.
2471 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2472 process_late_inline_calls_no_inline(igvn);
2473 if (failing()) return;
2474 }
2475 } // (End scope of igvn; run destructor if necessary for asserts.)
2476
2477 check_no_dead_use();
2478
2479 process_print_inlining();
2480
2481 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2482 // to remove hashes to unlock nodes for modifications.
2483 C->node_hash()->clear();
2484
2485 // A method with only infinite loops has no edges entering loops from root
2486 {
2487 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2488 if (final_graph_reshaping()) {
2489 assert(failing(), "must bail out w/ explicit message");
2490 return;
2491 }
2492 }
2493
2494 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2495 DEBUG_ONLY(set_phase_optimize_finished();)
2496 }
2497
2498 #ifdef ASSERT
2499 void Compile::check_no_dead_use() const {
2500 ResourceMark rm;
2501 Unique_Node_List wq;
2502 wq.push(root());
2503 for (uint i = 0; i < wq.size(); ++i) {
2504 Node* n = wq.at(i);
2505 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2506 Node* u = n->fast_out(j);
2507 if (u->outcnt() == 0 && !u->is_Con()) {
2508 u->dump();
2509 fatal("no reachable node should have no use");
2510 }
2511 wq.push(u);
2512 }
2513 }
2514 }
2515 #endif
2516
2517 void Compile::inline_vector_reboxing_calls() {
2518 if (C->_vector_reboxing_late_inlines.length() > 0) {
2519 _late_inlines_pos = C->_late_inlines.length();
2520 while (_vector_reboxing_late_inlines.length() > 0) {
2521 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2522 cg->do_late_inline();
2523 if (failing()) return;
2524 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2525 }
2526 _vector_reboxing_late_inlines.trunc_to(0);
2527 }
2528 }
2529
2530 bool Compile::has_vbox_nodes() {
2531 if (C->_vector_reboxing_late_inlines.length() > 0) {
2532 return true;
2533 }
2534 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2535 Node * n = C->macro_node(macro_idx);
2536 assert(n->is_macro(), "only macro nodes expected here");
2537 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2538 return true;
2539 }
2540 }
2541 return false;
2542 }
2543
2544 //---------------------------- Bitwise operation packing optimization ---------------------------
2545
2546 static bool is_vector_unary_bitwise_op(Node* n) {
2547 return n->Opcode() == Op_XorV &&
2548 VectorNode::is_vector_bitwise_not_pattern(n);
2549 }
2550
2551 static bool is_vector_binary_bitwise_op(Node* n) {
2552 switch (n->Opcode()) {
2553 case Op_AndV:
2554 case Op_OrV:
2555 return true;
2556
2557 case Op_XorV:
2558 return !is_vector_unary_bitwise_op(n);
2559
2560 default:
2561 return false;
2562 }
2563 }
2564
2565 static bool is_vector_ternary_bitwise_op(Node* n) {
2566 return n->Opcode() == Op_MacroLogicV;
2567 }
2568
2569 static bool is_vector_bitwise_op(Node* n) {
2570 return is_vector_unary_bitwise_op(n) ||
2571 is_vector_binary_bitwise_op(n) ||
2572 is_vector_ternary_bitwise_op(n);
2573 }
2574
2575 static bool is_vector_bitwise_cone_root(Node* n) {
2576 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2577 return false;
2578 }
2579 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2580 if (is_vector_bitwise_op(n->fast_out(i))) {
2581 return false;
2582 }
2583 }
2584 return true;
2585 }
2586
2587 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2588 uint cnt = 0;
2589 if (is_vector_bitwise_op(n)) {
2590 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2591 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2592 for (uint i = 1; i < inp_cnt; i++) {
2593 Node* in = n->in(i);
2594 bool skip = VectorNode::is_all_ones_vector(in);
2595 if (!skip && !inputs.member(in)) {
2596 inputs.push(in);
2597 cnt++;
2598 }
2599 }
2600 assert(cnt <= 1, "not unary");
2601 } else {
2602 uint last_req = inp_cnt;
2603 if (is_vector_ternary_bitwise_op(n)) {
2604 last_req = inp_cnt - 1; // skip last input
2605 }
2606 for (uint i = 1; i < last_req; i++) {
2607 Node* def = n->in(i);
2608 if (!inputs.member(def)) {
2609 inputs.push(def);
2610 cnt++;
2611 }
2612 }
2613 }
2614 } else { // not a bitwise operations
2615 if (!inputs.member(n)) {
2616 inputs.push(n);
2617 cnt++;
2618 }
2619 }
2620 return cnt;
2621 }
2622
2623 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2624 Unique_Node_List useful_nodes;
2625 C->identify_useful_nodes(useful_nodes);
2626
2627 for (uint i = 0; i < useful_nodes.size(); i++) {
2628 Node* n = useful_nodes.at(i);
2629 if (is_vector_bitwise_cone_root(n)) {
2630 list.push(n);
2631 }
2632 }
2633 }
2634
2635 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2636 const TypeVect* vt,
2637 Unique_Node_List& partition,
2638 Unique_Node_List& inputs) {
2639 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2640 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2641 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2642
2643 Node* in1 = inputs.at(0);
2644 Node* in2 = inputs.at(1);
2645 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2646
2647 uint func = compute_truth_table(partition, inputs);
2648
2649 Node* pn = partition.at(partition.size() - 1);
2650 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2651 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2652 }
2653
2654 static uint extract_bit(uint func, uint pos) {
2655 return (func & (1 << pos)) >> pos;
2656 }
2657
2658 //
2659 // A macro logic node represents a truth table. It has 4 inputs,
2660 // First three inputs corresponds to 3 columns of a truth table
2661 // and fourth input captures the logic function.
2662 //
2663 // eg. fn = (in1 AND in2) OR in3;
2664 //
2665 // MacroNode(in1,in2,in3,fn)
2666 //
2667 // -----------------
2668 // in1 in2 in3 fn
2669 // -----------------
2670 // 0 0 0 0
2671 // 0 0 1 1
2672 // 0 1 0 0
2673 // 0 1 1 1
2674 // 1 0 0 0
2675 // 1 0 1 1
2676 // 1 1 0 1
2677 // 1 1 1 1
2678 //
2679
2680 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2681 int res = 0;
2682 for (int i = 0; i < 8; i++) {
2683 int bit1 = extract_bit(in1, i);
2684 int bit2 = extract_bit(in2, i);
2685 int bit3 = extract_bit(in3, i);
2686
2687 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2688 int func_bit = extract_bit(func, func_bit_pos);
2689
2690 res |= func_bit << i;
2691 }
2692 return res;
2693 }
2694
2695 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2696 assert(n != nullptr, "");
2697 assert(eval_map.contains(n), "absent");
2698 return *(eval_map.get(n));
2699 }
2700
2701 static void eval_operands(Node* n,
2702 uint& func1, uint& func2, uint& func3,
2703 ResourceHashtable<Node*,uint>& eval_map) {
2704 assert(is_vector_bitwise_op(n), "");
2705
2706 if (is_vector_unary_bitwise_op(n)) {
2707 Node* opnd = n->in(1);
2708 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2709 opnd = n->in(2);
2710 }
2711 func1 = eval_operand(opnd, eval_map);
2712 } else if (is_vector_binary_bitwise_op(n)) {
2713 func1 = eval_operand(n->in(1), eval_map);
2714 func2 = eval_operand(n->in(2), eval_map);
2715 } else {
2716 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2717 func1 = eval_operand(n->in(1), eval_map);
2718 func2 = eval_operand(n->in(2), eval_map);
2719 func3 = eval_operand(n->in(3), eval_map);
2720 }
2721 }
2722
2723 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2724 assert(inputs.size() <= 3, "sanity");
2725 ResourceMark rm;
2726 uint res = 0;
2727 ResourceHashtable<Node*,uint> eval_map;
2728
2729 // Populate precomputed functions for inputs.
2730 // Each input corresponds to one column of 3 input truth-table.
2731 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2732 0xCC, // (_, b, _) -> b
2733 0xF0 }; // (a, _, _) -> a
2734 for (uint i = 0; i < inputs.size(); i++) {
2735 eval_map.put(inputs.at(i), input_funcs[2-i]);
2736 }
2737
2738 for (uint i = 0; i < partition.size(); i++) {
2739 Node* n = partition.at(i);
2740
2741 uint func1 = 0, func2 = 0, func3 = 0;
2742 eval_operands(n, func1, func2, func3, eval_map);
2743
2744 switch (n->Opcode()) {
2745 case Op_OrV:
2746 assert(func3 == 0, "not binary");
2747 res = func1 | func2;
2748 break;
2749 case Op_AndV:
2750 assert(func3 == 0, "not binary");
2751 res = func1 & func2;
2752 break;
2753 case Op_XorV:
2754 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2755 assert(func2 == 0 && func3 == 0, "not unary");
2756 res = (~func1) & 0xFF;
2757 } else {
2758 assert(func3 == 0, "not binary");
2759 res = func1 ^ func2;
2760 }
2761 break;
2762 case Op_MacroLogicV:
2763 // Ordering of inputs may change during evaluation of sub-tree
2764 // containing MacroLogic node as a child node, thus a re-evaluation
2765 // makes sure that function is evaluated in context of current
2766 // inputs.
2767 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2768 break;
2769
2770 default: assert(false, "not supported: %s", n->Name());
2771 }
2772 assert(res <= 0xFF, "invalid");
2773 eval_map.put(n, res);
2774 }
2775 return res;
2776 }
2777
2778 // Criteria under which nodes gets packed into a macro logic node:-
2779 // 1) Parent and both child nodes are all unmasked or masked with
2780 // same predicates.
2781 // 2) Masked parent can be packed with left child if it is predicated
2782 // and both have same predicates.
2783 // 3) Masked parent can be packed with right child if its un-predicated
2784 // or has matching predication condition.
2785 // 4) An unmasked parent can be packed with an unmasked child.
2786 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2787 assert(partition.size() == 0, "not empty");
2788 assert(inputs.size() == 0, "not empty");
2789 if (is_vector_ternary_bitwise_op(n)) {
2790 return false;
2791 }
2792
2793 bool is_unary_op = is_vector_unary_bitwise_op(n);
2794 if (is_unary_op) {
2795 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2796 return false; // too few inputs
2797 }
2798
2799 bool pack_left_child = true;
2800 bool pack_right_child = true;
2801
2802 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2803 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2804
2805 int left_child_input_cnt = 0;
2806 int right_child_input_cnt = 0;
2807
2808 bool parent_is_predicated = n->is_predicated_vector();
2809 bool left_child_predicated = n->in(1)->is_predicated_vector();
2810 bool right_child_predicated = n->in(2)->is_predicated_vector();
2811
2812 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2813 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2814 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2815
2816 do {
2817 if (pack_left_child && left_child_LOP &&
2818 ((!parent_is_predicated && !left_child_predicated) ||
2819 ((parent_is_predicated && left_child_predicated &&
2820 parent_pred == left_child_pred)))) {
2821 partition.push(n->in(1));
2822 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2823 } else {
2824 inputs.push(n->in(1));
2825 left_child_input_cnt = 1;
2826 }
2827
2828 if (pack_right_child && right_child_LOP &&
2829 (!right_child_predicated ||
2830 (right_child_predicated && parent_is_predicated &&
2831 parent_pred == right_child_pred))) {
2832 partition.push(n->in(2));
2833 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2834 } else {
2835 inputs.push(n->in(2));
2836 right_child_input_cnt = 1;
2837 }
2838
2839 if (inputs.size() > 3) {
2840 assert(partition.size() > 0, "");
2841 inputs.clear();
2842 partition.clear();
2843 if (left_child_input_cnt > right_child_input_cnt) {
2844 pack_left_child = false;
2845 } else {
2846 pack_right_child = false;
2847 }
2848 } else {
2849 break;
2850 }
2851 } while(true);
2852
2853 if(partition.size()) {
2854 partition.push(n);
2855 }
2856
2857 return (partition.size() == 2 || partition.size() == 3) &&
2858 (inputs.size() == 2 || inputs.size() == 3);
2859 }
2860
2861 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2862 assert(is_vector_bitwise_op(n), "not a root");
2863
2864 visited.set(n->_idx);
2865
2866 // 1) Do a DFS walk over the logic cone.
2867 for (uint i = 1; i < n->req(); i++) {
2868 Node* in = n->in(i);
2869 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2870 process_logic_cone_root(igvn, in, visited);
2871 }
2872 }
2873
2874 // 2) Bottom up traversal: Merge node[s] with
2875 // the parent to form macro logic node.
2876 Unique_Node_List partition;
2877 Unique_Node_List inputs;
2878 if (compute_logic_cone(n, partition, inputs)) {
2879 const TypeVect* vt = n->bottom_type()->is_vect();
2880 Node* pn = partition.at(partition.size() - 1);
2881 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2882 if (mask == nullptr ||
2883 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2884 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2885 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2886 igvn.replace_node(n, macro_logic);
2887 }
2888 }
2889 }
2890
2891 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2892 ResourceMark rm;
2893 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2894 Unique_Node_List list;
2895 collect_logic_cone_roots(list);
2896
2897 while (list.size() > 0) {
2898 Node* n = list.pop();
2899 const TypeVect* vt = n->bottom_type()->is_vect();
2900 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2901 if (supported) {
2902 VectorSet visited(comp_arena());
2903 process_logic_cone_root(igvn, n, visited);
2904 }
2905 }
2906 }
2907 }
2908
2909 //------------------------------Code_Gen---------------------------------------
2910 // Given a graph, generate code for it
2911 void Compile::Code_Gen() {
2912 if (failing()) {
2913 return;
2914 }
2915
2916 // Perform instruction selection. You might think we could reclaim Matcher
2917 // memory PDQ, but actually the Matcher is used in generating spill code.
2918 // Internals of the Matcher (including some VectorSets) must remain live
2919 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2920 // set a bit in reclaimed memory.
2921
2922 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2923 // nodes. Mapping is only valid at the root of each matched subtree.
2924 NOT_PRODUCT( verify_graph_edges(); )
2925
2926 Matcher matcher;
2927 _matcher = &matcher;
2928 {
2929 TracePhase tp("matcher", &timers[_t_matcher]);
2930 matcher.match();
2931 if (failing()) {
2932 return;
2933 }
2934 }
2935 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2936 // nodes. Mapping is only valid at the root of each matched subtree.
2937 NOT_PRODUCT( verify_graph_edges(); )
2938
2939 // If you have too many nodes, or if matching has failed, bail out
2940 check_node_count(0, "out of nodes matching instructions");
2941 if (failing()) {
2942 return;
2943 }
2944
2945 print_method(PHASE_MATCHING, 2);
2946
2947 // Build a proper-looking CFG
2948 PhaseCFG cfg(node_arena(), root(), matcher);
2949 _cfg = &cfg;
2950 {
2951 TracePhase tp("scheduler", &timers[_t_scheduler]);
2952 bool success = cfg.do_global_code_motion();
2953 if (!success) {
2954 return;
2955 }
2956
2957 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2958 NOT_PRODUCT( verify_graph_edges(); )
2959 cfg.verify();
2960 }
2961
2962 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2963 _regalloc = ®alloc;
2964 {
2965 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2966 // Perform register allocation. After Chaitin, use-def chains are
2967 // no longer accurate (at spill code) and so must be ignored.
2968 // Node->LRG->reg mappings are still accurate.
2969 _regalloc->Register_Allocate();
2970
2971 // Bail out if the allocator builds too many nodes
2972 if (failing()) {
2973 return;
2974 }
2975
2976 print_method(PHASE_REGISTER_ALLOCATION, 2);
2977 }
2978
2979 // Prior to register allocation we kept empty basic blocks in case the
2980 // the allocator needed a place to spill. After register allocation we
2981 // are not adding any new instructions. If any basic block is empty, we
2982 // can now safely remove it.
2983 {
2984 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2985 cfg.remove_empty_blocks();
2986 if (do_freq_based_layout()) {
2987 PhaseBlockLayout layout(cfg);
2988 } else {
2989 cfg.set_loop_alignment();
2990 }
2991 cfg.fixup_flow();
2992 cfg.remove_unreachable_blocks();
2993 cfg.verify_dominator_tree();
2994 print_method(PHASE_BLOCK_ORDERING, 3);
2995 }
2996
2997 // Apply peephole optimizations
2998 if( OptoPeephole ) {
2999 TracePhase tp("peephole", &timers[_t_peephole]);
3000 PhasePeephole peep( _regalloc, cfg);
3001 peep.do_transform();
3002 print_method(PHASE_PEEPHOLE, 3);
3003 }
3004
3005 // Do late expand if CPU requires this.
3006 if (Matcher::require_postalloc_expand) {
3007 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3008 cfg.postalloc_expand(_regalloc);
3009 print_method(PHASE_POSTALLOC_EXPAND, 3);
3010 }
3011
3012 // Convert Nodes to instruction bits in a buffer
3013 {
3014 TracePhase tp("output", &timers[_t_output]);
3015 PhaseOutput output;
3016 output.Output();
3017 if (failing()) return;
3018 output.install();
3019 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3020 }
3021
3022 // He's dead, Jim.
3023 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3024 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3025 }
3026
3027 //------------------------------Final_Reshape_Counts---------------------------
3028 // This class defines counters to help identify when a method
3029 // may/must be executed using hardware with only 24-bit precision.
3030 struct Final_Reshape_Counts : public StackObj {
3031 int _call_count; // count non-inlined 'common' calls
3032 int _float_count; // count float ops requiring 24-bit precision
3033 int _double_count; // count double ops requiring more precision
3034 int _java_call_count; // count non-inlined 'java' calls
3035 int _inner_loop_count; // count loops which need alignment
3036 VectorSet _visited; // Visitation flags
3037 Node_List _tests; // Set of IfNodes & PCTableNodes
3038
3039 Final_Reshape_Counts() :
3040 _call_count(0), _float_count(0), _double_count(0),
3041 _java_call_count(0), _inner_loop_count(0) { }
3042
3043 void inc_call_count () { _call_count ++; }
3044 void inc_float_count () { _float_count ++; }
3045 void inc_double_count() { _double_count++; }
3046 void inc_java_call_count() { _java_call_count++; }
3047 void inc_inner_loop_count() { _inner_loop_count++; }
3048
3049 int get_call_count () const { return _call_count ; }
3050 int get_float_count () const { return _float_count ; }
3051 int get_double_count() const { return _double_count; }
3052 int get_java_call_count() const { return _java_call_count; }
3053 int get_inner_loop_count() const { return _inner_loop_count; }
3054 };
3055
3056 // Eliminate trivially redundant StoreCMs and accumulate their
3057 // precedence edges.
3058 void Compile::eliminate_redundant_card_marks(Node* n) {
3059 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3060 if (n->in(MemNode::Address)->outcnt() > 1) {
3061 // There are multiple users of the same address so it might be
3062 // possible to eliminate some of the StoreCMs
3063 Node* mem = n->in(MemNode::Memory);
3064 Node* adr = n->in(MemNode::Address);
3065 Node* val = n->in(MemNode::ValueIn);
3066 Node* prev = n;
3067 bool done = false;
3068 // Walk the chain of StoreCMs eliminating ones that match. As
3069 // long as it's a chain of single users then the optimization is
3070 // safe. Eliminating partially redundant StoreCMs would require
3071 // cloning copies down the other paths.
3072 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3073 if (adr == mem->in(MemNode::Address) &&
3074 val == mem->in(MemNode::ValueIn)) {
3075 // redundant StoreCM
3076 if (mem->req() > MemNode::OopStore) {
3077 // Hasn't been processed by this code yet.
3078 n->add_prec(mem->in(MemNode::OopStore));
3079 } else {
3080 // Already converted to precedence edge
3081 for (uint i = mem->req(); i < mem->len(); i++) {
3082 // Accumulate any precedence edges
3083 if (mem->in(i) != nullptr) {
3084 n->add_prec(mem->in(i));
3085 }
3086 }
3087 // Everything above this point has been processed.
3088 done = true;
3089 }
3090 // Eliminate the previous StoreCM
3091 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3092 assert(mem->outcnt() == 0, "should be dead");
3093 mem->disconnect_inputs(this);
3094 } else {
3095 prev = mem;
3096 }
3097 mem = prev->in(MemNode::Memory);
3098 }
3099 }
3100 }
3101
3102 //------------------------------final_graph_reshaping_impl----------------------
3103 // Implement items 1-5 from final_graph_reshaping below.
3104 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3105
3106 if ( n->outcnt() == 0 ) return; // dead node
3107 uint nop = n->Opcode();
3108
3109 // Check for 2-input instruction with "last use" on right input.
3110 // Swap to left input. Implements item (2).
3111 if( n->req() == 3 && // two-input instruction
3112 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3113 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3114 n->in(2)->outcnt() == 1 &&// right use IS a last use
3115 !n->in(2)->is_Con() ) { // right use is not a constant
3116 // Check for commutative opcode
3117 switch( nop ) {
3118 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3119 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3120 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3121 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3122 case Op_AndL: case Op_XorL: case Op_OrL:
3123 case Op_AndI: case Op_XorI: case Op_OrI: {
3124 // Move "last use" input to left by swapping inputs
3125 n->swap_edges(1, 2);
3126 break;
3127 }
3128 default:
3129 break;
3130 }
3131 }
3132
3133 #ifdef ASSERT
3134 if( n->is_Mem() ) {
3135 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3136 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3137 // oop will be recorded in oop map if load crosses safepoint
3138 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3139 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3140 "raw memory operations should have control edge");
3141 }
3142 if (n->is_MemBar()) {
3143 MemBarNode* mb = n->as_MemBar();
3144 if (mb->trailing_store() || mb->trailing_load_store()) {
3145 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3146 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3147 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3148 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3149 } else if (mb->leading()) {
3150 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3151 }
3152 }
3153 #endif
3154 // Count FPU ops and common calls, implements item (3)
3155 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3156 if (!gc_handled) {
3157 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3158 }
3159
3160 // Collect CFG split points
3161 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3162 frc._tests.push(n);
3163 }
3164 }
3165
3166 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3167 switch( nop ) {
3168 // Count all float operations that may use FPU
3169 case Op_AddF:
3170 case Op_SubF:
3171 case Op_MulF:
3172 case Op_DivF:
3173 case Op_NegF:
3174 case Op_ModF:
3175 case Op_ConvI2F:
3176 case Op_ConF:
3177 case Op_CmpF:
3178 case Op_CmpF3:
3179 case Op_StoreF:
3180 case Op_LoadF:
3181 // case Op_ConvL2F: // longs are split into 32-bit halves
3182 frc.inc_float_count();
3183 break;
3184
3185 case Op_ConvF2D:
3186 case Op_ConvD2F:
3187 frc.inc_float_count();
3188 frc.inc_double_count();
3189 break;
3190
3191 // Count all double operations that may use FPU
3192 case Op_AddD:
3193 case Op_SubD:
3194 case Op_MulD:
3195 case Op_DivD:
3196 case Op_NegD:
3197 case Op_ModD:
3198 case Op_ConvI2D:
3199 case Op_ConvD2I:
3200 // case Op_ConvL2D: // handled by leaf call
3201 // case Op_ConvD2L: // handled by leaf call
3202 case Op_ConD:
3203 case Op_CmpD:
3204 case Op_CmpD3:
3205 case Op_StoreD:
3206 case Op_LoadD:
3207 case Op_LoadD_unaligned:
3208 frc.inc_double_count();
3209 break;
3210 case Op_Opaque1: // Remove Opaque Nodes before matching
3211 n->subsume_by(n->in(1), this);
3212 break;
3213 case Op_CallStaticJava:
3214 case Op_CallJava:
3215 case Op_CallDynamicJava:
3216 frc.inc_java_call_count(); // Count java call site;
3217 case Op_CallRuntime:
3218 case Op_CallLeaf:
3219 case Op_CallLeafVector:
3220 case Op_CallLeafNoFP: {
3221 assert (n->is_Call(), "");
3222 CallNode *call = n->as_Call();
3223 // Count call sites where the FP mode bit would have to be flipped.
3224 // Do not count uncommon runtime calls:
3225 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3226 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3227 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3228 frc.inc_call_count(); // Count the call site
3229 } else { // See if uncommon argument is shared
3230 Node *n = call->in(TypeFunc::Parms);
3231 int nop = n->Opcode();
3232 // Clone shared simple arguments to uncommon calls, item (1).
3233 if (n->outcnt() > 1 &&
3234 !n->is_Proj() &&
3235 nop != Op_CreateEx &&
3236 nop != Op_CheckCastPP &&
3237 nop != Op_DecodeN &&
3238 nop != Op_DecodeNKlass &&
3239 !n->is_Mem() &&
3240 !n->is_Phi()) {
3241 Node *x = n->clone();
3242 call->set_req(TypeFunc::Parms, x);
3243 }
3244 }
3245 break;
3246 }
3247
3248 case Op_StoreCM:
3249 {
3250 // Convert OopStore dependence into precedence edge
3251 Node* prec = n->in(MemNode::OopStore);
3252 n->del_req(MemNode::OopStore);
3253 n->add_prec(prec);
3254 eliminate_redundant_card_marks(n);
3255 }
3256
3257 // fall through
3258
3259 case Op_StoreB:
3260 case Op_StoreC:
3261 case Op_StoreI:
3262 case Op_StoreL:
3263 case Op_CompareAndSwapB:
3264 case Op_CompareAndSwapS:
3265 case Op_CompareAndSwapI:
3266 case Op_CompareAndSwapL:
3267 case Op_CompareAndSwapP:
3268 case Op_CompareAndSwapN:
3269 case Op_WeakCompareAndSwapB:
3270 case Op_WeakCompareAndSwapS:
3271 case Op_WeakCompareAndSwapI:
3272 case Op_WeakCompareAndSwapL:
3273 case Op_WeakCompareAndSwapP:
3274 case Op_WeakCompareAndSwapN:
3275 case Op_CompareAndExchangeB:
3276 case Op_CompareAndExchangeS:
3277 case Op_CompareAndExchangeI:
3278 case Op_CompareAndExchangeL:
3279 case Op_CompareAndExchangeP:
3280 case Op_CompareAndExchangeN:
3281 case Op_GetAndAddS:
3282 case Op_GetAndAddB:
3283 case Op_GetAndAddI:
3284 case Op_GetAndAddL:
3285 case Op_GetAndSetS:
3286 case Op_GetAndSetB:
3287 case Op_GetAndSetI:
3288 case Op_GetAndSetL:
3289 case Op_GetAndSetP:
3290 case Op_GetAndSetN:
3291 case Op_StoreP:
3292 case Op_StoreN:
3293 case Op_StoreNKlass:
3294 case Op_LoadB:
3295 case Op_LoadUB:
3296 case Op_LoadUS:
3297 case Op_LoadI:
3298 case Op_LoadKlass:
3299 case Op_LoadNKlass:
3300 case Op_LoadL:
3301 case Op_LoadL_unaligned:
3302 case Op_LoadP:
3303 case Op_LoadN:
3304 case Op_LoadRange:
3305 case Op_LoadS:
3306 break;
3307
3308 case Op_AddP: { // Assert sane base pointers
3309 Node *addp = n->in(AddPNode::Address);
3310 assert( !addp->is_AddP() ||
3311 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3312 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3313 "Base pointers must match (addp %u)", addp->_idx );
3314 #ifdef _LP64
3315 if ((UseCompressedOops || UseCompressedClassPointers) &&
3316 addp->Opcode() == Op_ConP &&
3317 addp == n->in(AddPNode::Base) &&
3318 n->in(AddPNode::Offset)->is_Con()) {
3319 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3320 // on the platform and on the compressed oops mode.
3321 // Use addressing with narrow klass to load with offset on x86.
3322 // Some platforms can use the constant pool to load ConP.
3323 // Do this transformation here since IGVN will convert ConN back to ConP.
3324 const Type* t = addp->bottom_type();
3325 bool is_oop = t->isa_oopptr() != nullptr;
3326 bool is_klass = t->isa_klassptr() != nullptr;
3327
3328 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3329 (is_klass && Matcher::const_klass_prefer_decode())) {
3330 Node* nn = nullptr;
3331
3332 int op = is_oop ? Op_ConN : Op_ConNKlass;
3333
3334 // Look for existing ConN node of the same exact type.
3335 Node* r = root();
3336 uint cnt = r->outcnt();
3337 for (uint i = 0; i < cnt; i++) {
3338 Node* m = r->raw_out(i);
3339 if (m!= nullptr && m->Opcode() == op &&
3340 m->bottom_type()->make_ptr() == t) {
3341 nn = m;
3342 break;
3343 }
3344 }
3345 if (nn != nullptr) {
3346 // Decode a narrow oop to match address
3347 // [R12 + narrow_oop_reg<<3 + offset]
3348 if (is_oop) {
3349 nn = new DecodeNNode(nn, t);
3350 } else {
3351 nn = new DecodeNKlassNode(nn, t);
3352 }
3353 // Check for succeeding AddP which uses the same Base.
3354 // Otherwise we will run into the assertion above when visiting that guy.
3355 for (uint i = 0; i < n->outcnt(); ++i) {
3356 Node *out_i = n->raw_out(i);
3357 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3358 out_i->set_req(AddPNode::Base, nn);
3359 #ifdef ASSERT
3360 for (uint j = 0; j < out_i->outcnt(); ++j) {
3361 Node *out_j = out_i->raw_out(j);
3362 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3363 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3364 }
3365 #endif
3366 }
3367 }
3368 n->set_req(AddPNode::Base, nn);
3369 n->set_req(AddPNode::Address, nn);
3370 if (addp->outcnt() == 0) {
3371 addp->disconnect_inputs(this);
3372 }
3373 }
3374 }
3375 }
3376 #endif
3377 break;
3378 }
3379
3380 case Op_CastPP: {
3381 // Remove CastPP nodes to gain more freedom during scheduling but
3382 // keep the dependency they encode as control or precedence edges
3383 // (if control is set already) on memory operations. Some CastPP
3384 // nodes don't have a control (don't carry a dependency): skip
3385 // those.
3386 if (n->in(0) != nullptr) {
3387 ResourceMark rm;
3388 Unique_Node_List wq;
3389 wq.push(n);
3390 for (uint next = 0; next < wq.size(); ++next) {
3391 Node *m = wq.at(next);
3392 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3393 Node* use = m->fast_out(i);
3394 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3395 use->ensure_control_or_add_prec(n->in(0));
3396 } else {
3397 switch(use->Opcode()) {
3398 case Op_AddP:
3399 case Op_DecodeN:
3400 case Op_DecodeNKlass:
3401 case Op_CheckCastPP:
3402 case Op_CastPP:
3403 wq.push(use);
3404 break;
3405 }
3406 }
3407 }
3408 }
3409 }
3410 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3411 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3412 Node* in1 = n->in(1);
3413 const Type* t = n->bottom_type();
3414 Node* new_in1 = in1->clone();
3415 new_in1->as_DecodeN()->set_type(t);
3416
3417 if (!Matcher::narrow_oop_use_complex_address()) {
3418 //
3419 // x86, ARM and friends can handle 2 adds in addressing mode
3420 // and Matcher can fold a DecodeN node into address by using
3421 // a narrow oop directly and do implicit null check in address:
3422 //
3423 // [R12 + narrow_oop_reg<<3 + offset]
3424 // NullCheck narrow_oop_reg
3425 //
3426 // On other platforms (Sparc) we have to keep new DecodeN node and
3427 // use it to do implicit null check in address:
3428 //
3429 // decode_not_null narrow_oop_reg, base_reg
3430 // [base_reg + offset]
3431 // NullCheck base_reg
3432 //
3433 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3434 // to keep the information to which null check the new DecodeN node
3435 // corresponds to use it as value in implicit_null_check().
3436 //
3437 new_in1->set_req(0, n->in(0));
3438 }
3439
3440 n->subsume_by(new_in1, this);
3441 if (in1->outcnt() == 0) {
3442 in1->disconnect_inputs(this);
3443 }
3444 } else {
3445 n->subsume_by(n->in(1), this);
3446 if (n->outcnt() == 0) {
3447 n->disconnect_inputs(this);
3448 }
3449 }
3450 break;
3451 }
3452 #ifdef _LP64
3453 case Op_CmpP:
3454 // Do this transformation here to preserve CmpPNode::sub() and
3455 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3456 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3457 Node* in1 = n->in(1);
3458 Node* in2 = n->in(2);
3459 if (!in1->is_DecodeNarrowPtr()) {
3460 in2 = in1;
3461 in1 = n->in(2);
3462 }
3463 assert(in1->is_DecodeNarrowPtr(), "sanity");
3464
3465 Node* new_in2 = nullptr;
3466 if (in2->is_DecodeNarrowPtr()) {
3467 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3468 new_in2 = in2->in(1);
3469 } else if (in2->Opcode() == Op_ConP) {
3470 const Type* t = in2->bottom_type();
3471 if (t == TypePtr::NULL_PTR) {
3472 assert(in1->is_DecodeN(), "compare klass to null?");
3473 // Don't convert CmpP null check into CmpN if compressed
3474 // oops implicit null check is not generated.
3475 // This will allow to generate normal oop implicit null check.
3476 if (Matcher::gen_narrow_oop_implicit_null_checks())
3477 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3478 //
3479 // This transformation together with CastPP transformation above
3480 // will generated code for implicit null checks for compressed oops.
3481 //
3482 // The original code after Optimize()
3483 //
3484 // LoadN memory, narrow_oop_reg
3485 // decode narrow_oop_reg, base_reg
3486 // CmpP base_reg, nullptr
3487 // CastPP base_reg // NotNull
3488 // Load [base_reg + offset], val_reg
3489 //
3490 // after these transformations will be
3491 //
3492 // LoadN memory, narrow_oop_reg
3493 // CmpN narrow_oop_reg, nullptr
3494 // decode_not_null narrow_oop_reg, base_reg
3495 // Load [base_reg + offset], val_reg
3496 //
3497 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3498 // since narrow oops can be used in debug info now (see the code in
3499 // final_graph_reshaping_walk()).
3500 //
3501 // At the end the code will be matched to
3502 // on x86:
3503 //
3504 // Load_narrow_oop memory, narrow_oop_reg
3505 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3506 // NullCheck narrow_oop_reg
3507 //
3508 // and on sparc:
3509 //
3510 // Load_narrow_oop memory, narrow_oop_reg
3511 // decode_not_null narrow_oop_reg, base_reg
3512 // Load [base_reg + offset], val_reg
3513 // NullCheck base_reg
3514 //
3515 } else if (t->isa_oopptr()) {
3516 new_in2 = ConNode::make(t->make_narrowoop());
3517 } else if (t->isa_klassptr()) {
3518 new_in2 = ConNode::make(t->make_narrowklass());
3519 }
3520 }
3521 if (new_in2 != nullptr) {
3522 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3523 n->subsume_by(cmpN, this);
3524 if (in1->outcnt() == 0) {
3525 in1->disconnect_inputs(this);
3526 }
3527 if (in2->outcnt() == 0) {
3528 in2->disconnect_inputs(this);
3529 }
3530 }
3531 }
3532 break;
3533
3534 case Op_DecodeN:
3535 case Op_DecodeNKlass:
3536 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3537 // DecodeN could be pinned when it can't be fold into
3538 // an address expression, see the code for Op_CastPP above.
3539 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3540 break;
3541
3542 case Op_EncodeP:
3543 case Op_EncodePKlass: {
3544 Node* in1 = n->in(1);
3545 if (in1->is_DecodeNarrowPtr()) {
3546 n->subsume_by(in1->in(1), this);
3547 } else if (in1->Opcode() == Op_ConP) {
3548 const Type* t = in1->bottom_type();
3549 if (t == TypePtr::NULL_PTR) {
3550 assert(t->isa_oopptr(), "null klass?");
3551 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3552 } else if (t->isa_oopptr()) {
3553 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3554 } else if (t->isa_klassptr()) {
3555 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3556 }
3557 }
3558 if (in1->outcnt() == 0) {
3559 in1->disconnect_inputs(this);
3560 }
3561 break;
3562 }
3563
3564 case Op_Proj: {
3565 if (OptimizeStringConcat || IncrementalInline) {
3566 ProjNode* proj = n->as_Proj();
3567 if (proj->_is_io_use) {
3568 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3569 // Separate projections were used for the exception path which
3570 // are normally removed by a late inline. If it wasn't inlined
3571 // then they will hang around and should just be replaced with
3572 // the original one. Merge them.
3573 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3574 if (non_io_proj != nullptr) {
3575 proj->subsume_by(non_io_proj , this);
3576 }
3577 }
3578 }
3579 break;
3580 }
3581
3582 case Op_Phi:
3583 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3584 // The EncodeP optimization may create Phi with the same edges
3585 // for all paths. It is not handled well by Register Allocator.
3586 Node* unique_in = n->in(1);
3587 assert(unique_in != nullptr, "");
3588 uint cnt = n->req();
3589 for (uint i = 2; i < cnt; i++) {
3590 Node* m = n->in(i);
3591 assert(m != nullptr, "");
3592 if (unique_in != m)
3593 unique_in = nullptr;
3594 }
3595 if (unique_in != nullptr) {
3596 n->subsume_by(unique_in, this);
3597 }
3598 }
3599 break;
3600
3601 #endif
3602
3603 #ifdef ASSERT
3604 case Op_CastII:
3605 // Verify that all range check dependent CastII nodes were removed.
3606 if (n->isa_CastII()->has_range_check()) {
3607 n->dump(3);
3608 assert(false, "Range check dependent CastII node was not removed");
3609 }
3610 break;
3611 #endif
3612
3613 case Op_ModI:
3614 if (UseDivMod) {
3615 // Check if a%b and a/b both exist
3616 Node* d = n->find_similar(Op_DivI);
3617 if (d) {
3618 // Replace them with a fused divmod if supported
3619 if (Matcher::has_match_rule(Op_DivModI)) {
3620 DivModINode* divmod = DivModINode::make(n);
3621 d->subsume_by(divmod->div_proj(), this);
3622 n->subsume_by(divmod->mod_proj(), this);
3623 } else {
3624 // replace a%b with a-((a/b)*b)
3625 Node* mult = new MulINode(d, d->in(2));
3626 Node* sub = new SubINode(d->in(1), mult);
3627 n->subsume_by(sub, this);
3628 }
3629 }
3630 }
3631 break;
3632
3633 case Op_ModL:
3634 if (UseDivMod) {
3635 // Check if a%b and a/b both exist
3636 Node* d = n->find_similar(Op_DivL);
3637 if (d) {
3638 // Replace them with a fused divmod if supported
3639 if (Matcher::has_match_rule(Op_DivModL)) {
3640 DivModLNode* divmod = DivModLNode::make(n);
3641 d->subsume_by(divmod->div_proj(), this);
3642 n->subsume_by(divmod->mod_proj(), this);
3643 } else {
3644 // replace a%b with a-((a/b)*b)
3645 Node* mult = new MulLNode(d, d->in(2));
3646 Node* sub = new SubLNode(d->in(1), mult);
3647 n->subsume_by(sub, this);
3648 }
3649 }
3650 }
3651 break;
3652
3653 case Op_UModI:
3654 if (UseDivMod) {
3655 // Check if a%b and a/b both exist
3656 Node* d = n->find_similar(Op_UDivI);
3657 if (d) {
3658 // Replace them with a fused unsigned divmod if supported
3659 if (Matcher::has_match_rule(Op_UDivModI)) {
3660 UDivModINode* divmod = UDivModINode::make(n);
3661 d->subsume_by(divmod->div_proj(), this);
3662 n->subsume_by(divmod->mod_proj(), this);
3663 } else {
3664 // replace a%b with a-((a/b)*b)
3665 Node* mult = new MulINode(d, d->in(2));
3666 Node* sub = new SubINode(d->in(1), mult);
3667 n->subsume_by(sub, this);
3668 }
3669 }
3670 }
3671 break;
3672
3673 case Op_UModL:
3674 if (UseDivMod) {
3675 // Check if a%b and a/b both exist
3676 Node* d = n->find_similar(Op_UDivL);
3677 if (d) {
3678 // Replace them with a fused unsigned divmod if supported
3679 if (Matcher::has_match_rule(Op_UDivModL)) {
3680 UDivModLNode* divmod = UDivModLNode::make(n);
3681 d->subsume_by(divmod->div_proj(), this);
3682 n->subsume_by(divmod->mod_proj(), this);
3683 } else {
3684 // replace a%b with a-((a/b)*b)
3685 Node* mult = new MulLNode(d, d->in(2));
3686 Node* sub = new SubLNode(d->in(1), mult);
3687 n->subsume_by(sub, this);
3688 }
3689 }
3690 }
3691 break;
3692
3693 case Op_LoadVector:
3694 case Op_StoreVector:
3695 #ifdef ASSERT
3696 // Add VerifyVectorAlignment node between adr and load / store.
3697 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3698 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3699 n->as_StoreVector()->must_verify_alignment();
3700 if (must_verify_alignment) {
3701 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3702 n->as_StoreVector()->memory_size();
3703 // The memory access should be aligned to the vector width in bytes.
3704 // However, the underlying array is possibly less well aligned, but at least
3705 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3706 // a loop we can expect at least the following alignment:
3707 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3708 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3709 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3710 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3711 jlong mask = guaranteed_alignment - 1;
3712 Node* mask_con = ConLNode::make(mask);
3713 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3714 n->set_req(MemNode::Address, va);
3715 }
3716 }
3717 #endif
3718 break;
3719
3720 case Op_LoadVectorGather:
3721 case Op_StoreVectorScatter:
3722 case Op_LoadVectorGatherMasked:
3723 case Op_StoreVectorScatterMasked:
3724 case Op_VectorCmpMasked:
3725 case Op_VectorMaskGen:
3726 case Op_LoadVectorMasked:
3727 case Op_StoreVectorMasked:
3728 break;
3729
3730 case Op_AddReductionVI:
3731 case Op_AddReductionVL:
3732 case Op_AddReductionVF:
3733 case Op_AddReductionVD:
3734 case Op_MulReductionVI:
3735 case Op_MulReductionVL:
3736 case Op_MulReductionVF:
3737 case Op_MulReductionVD:
3738 case Op_MinReductionV:
3739 case Op_MaxReductionV:
3740 case Op_AndReductionV:
3741 case Op_OrReductionV:
3742 case Op_XorReductionV:
3743 break;
3744
3745 case Op_PackB:
3746 case Op_PackS:
3747 case Op_PackI:
3748 case Op_PackF:
3749 case Op_PackL:
3750 case Op_PackD:
3751 if (n->req()-1 > 2) {
3752 // Replace many operand PackNodes with a binary tree for matching
3753 PackNode* p = (PackNode*) n;
3754 Node* btp = p->binary_tree_pack(1, n->req());
3755 n->subsume_by(btp, this);
3756 }
3757 break;
3758 case Op_Loop:
3759 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3760 case Op_CountedLoop:
3761 case Op_LongCountedLoop:
3762 case Op_OuterStripMinedLoop:
3763 if (n->as_Loop()->is_inner_loop()) {
3764 frc.inc_inner_loop_count();
3765 }
3766 n->as_Loop()->verify_strip_mined(0);
3767 break;
3768 case Op_LShiftI:
3769 case Op_RShiftI:
3770 case Op_URShiftI:
3771 case Op_LShiftL:
3772 case Op_RShiftL:
3773 case Op_URShiftL:
3774 if (Matcher::need_masked_shift_count) {
3775 // The cpu's shift instructions don't restrict the count to the
3776 // lower 5/6 bits. We need to do the masking ourselves.
3777 Node* in2 = n->in(2);
3778 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3779 const TypeInt* t = in2->find_int_type();
3780 if (t != nullptr && t->is_con()) {
3781 juint shift = t->get_con();
3782 if (shift > mask) { // Unsigned cmp
3783 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3784 }
3785 } else {
3786 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3787 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3788 n->set_req(2, shift);
3789 }
3790 }
3791 if (in2->outcnt() == 0) { // Remove dead node
3792 in2->disconnect_inputs(this);
3793 }
3794 }
3795 break;
3796 case Op_MemBarStoreStore:
3797 case Op_MemBarRelease:
3798 // Break the link with AllocateNode: it is no longer useful and
3799 // confuses register allocation.
3800 if (n->req() > MemBarNode::Precedent) {
3801 n->set_req(MemBarNode::Precedent, top());
3802 }
3803 break;
3804 case Op_MemBarAcquire: {
3805 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3806 // At parse time, the trailing MemBarAcquire for a volatile load
3807 // is created with an edge to the load. After optimizations,
3808 // that input may be a chain of Phis. If those phis have no
3809 // other use, then the MemBarAcquire keeps them alive and
3810 // register allocation can be confused.
3811 dead_nodes.push(n->in(MemBarNode::Precedent));
3812 n->set_req(MemBarNode::Precedent, top());
3813 }
3814 break;
3815 }
3816 case Op_Blackhole:
3817 break;
3818 case Op_RangeCheck: {
3819 RangeCheckNode* rc = n->as_RangeCheck();
3820 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3821 n->subsume_by(iff, this);
3822 frc._tests.push(iff);
3823 break;
3824 }
3825 case Op_ConvI2L: {
3826 if (!Matcher::convi2l_type_required) {
3827 // Code generation on some platforms doesn't need accurate
3828 // ConvI2L types. Widening the type can help remove redundant
3829 // address computations.
3830 n->as_Type()->set_type(TypeLong::INT);
3831 ResourceMark rm;
3832 Unique_Node_List wq;
3833 wq.push(n);
3834 for (uint next = 0; next < wq.size(); next++) {
3835 Node *m = wq.at(next);
3836
3837 for(;;) {
3838 // Loop over all nodes with identical inputs edges as m
3839 Node* k = m->find_similar(m->Opcode());
3840 if (k == nullptr) {
3841 break;
3842 }
3843 // Push their uses so we get a chance to remove node made
3844 // redundant
3845 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3846 Node* u = k->fast_out(i);
3847 if (u->Opcode() == Op_LShiftL ||
3848 u->Opcode() == Op_AddL ||
3849 u->Opcode() == Op_SubL ||
3850 u->Opcode() == Op_AddP) {
3851 wq.push(u);
3852 }
3853 }
3854 // Replace all nodes with identical edges as m with m
3855 k->subsume_by(m, this);
3856 }
3857 }
3858 }
3859 break;
3860 }
3861 case Op_CmpUL: {
3862 if (!Matcher::has_match_rule(Op_CmpUL)) {
3863 // No support for unsigned long comparisons
3864 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3865 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3866 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3867 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3868 Node* andl = new AndLNode(orl, remove_sign_mask);
3869 Node* cmp = new CmpLNode(andl, n->in(2));
3870 n->subsume_by(cmp, this);
3871 }
3872 break;
3873 }
3874 default:
3875 assert(!n->is_Call(), "");
3876 assert(!n->is_Mem(), "");
3877 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3878 break;
3879 }
3880 }
3881
3882 //------------------------------final_graph_reshaping_walk---------------------
3883 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3884 // requires that the walk visits a node's inputs before visiting the node.
3885 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3886 Unique_Node_List sfpt;
3887
3888 frc._visited.set(root->_idx); // first, mark node as visited
3889 uint cnt = root->req();
3890 Node *n = root;
3891 uint i = 0;
3892 while (true) {
3893 if (i < cnt) {
3894 // Place all non-visited non-null inputs onto stack
3895 Node* m = n->in(i);
3896 ++i;
3897 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3898 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3899 // compute worst case interpreter size in case of a deoptimization
3900 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3901
3902 sfpt.push(m);
3903 }
3904 cnt = m->req();
3905 nstack.push(n, i); // put on stack parent and next input's index
3906 n = m;
3907 i = 0;
3908 }
3909 } else {
3910 // Now do post-visit work
3911 final_graph_reshaping_impl(n, frc, dead_nodes);
3912 if (nstack.is_empty())
3913 break; // finished
3914 n = nstack.node(); // Get node from stack
3915 cnt = n->req();
3916 i = nstack.index();
3917 nstack.pop(); // Shift to the next node on stack
3918 }
3919 }
3920
3921 // Skip next transformation if compressed oops are not used.
3922 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3923 (!UseCompressedOops && !UseCompressedClassPointers))
3924 return;
3925
3926 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3927 // It could be done for an uncommon traps or any safepoints/calls
3928 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3929 while (sfpt.size() > 0) {
3930 n = sfpt.pop();
3931 JVMState *jvms = n->as_SafePoint()->jvms();
3932 assert(jvms != nullptr, "sanity");
3933 int start = jvms->debug_start();
3934 int end = n->req();
3935 bool is_uncommon = (n->is_CallStaticJava() &&
3936 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3937 for (int j = start; j < end; j++) {
3938 Node* in = n->in(j);
3939 if (in->is_DecodeNarrowPtr()) {
3940 bool safe_to_skip = true;
3941 if (!is_uncommon ) {
3942 // Is it safe to skip?
3943 for (uint i = 0; i < in->outcnt(); i++) {
3944 Node* u = in->raw_out(i);
3945 if (!u->is_SafePoint() ||
3946 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3947 safe_to_skip = false;
3948 }
3949 }
3950 }
3951 if (safe_to_skip) {
3952 n->set_req(j, in->in(1));
3953 }
3954 if (in->outcnt() == 0) {
3955 in->disconnect_inputs(this);
3956 }
3957 }
3958 }
3959 }
3960 }
3961
3962 //------------------------------final_graph_reshaping--------------------------
3963 // Final Graph Reshaping.
3964 //
3965 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3966 // and not commoned up and forced early. Must come after regular
3967 // optimizations to avoid GVN undoing the cloning. Clone constant
3968 // inputs to Loop Phis; these will be split by the allocator anyways.
3969 // Remove Opaque nodes.
3970 // (2) Move last-uses by commutative operations to the left input to encourage
3971 // Intel update-in-place two-address operations and better register usage
3972 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3973 // calls canonicalizing them back.
3974 // (3) Count the number of double-precision FP ops, single-precision FP ops
3975 // and call sites. On Intel, we can get correct rounding either by
3976 // forcing singles to memory (requires extra stores and loads after each
3977 // FP bytecode) or we can set a rounding mode bit (requires setting and
3978 // clearing the mode bit around call sites). The mode bit is only used
3979 // if the relative frequency of single FP ops to calls is low enough.
3980 // This is a key transform for SPEC mpeg_audio.
3981 // (4) Detect infinite loops; blobs of code reachable from above but not
3982 // below. Several of the Code_Gen algorithms fail on such code shapes,
3983 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3984 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3985 // Detection is by looking for IfNodes where only 1 projection is
3986 // reachable from below or CatchNodes missing some targets.
3987 // (5) Assert for insane oop offsets in debug mode.
3988
3989 bool Compile::final_graph_reshaping() {
3990 // an infinite loop may have been eliminated by the optimizer,
3991 // in which case the graph will be empty.
3992 if (root()->req() == 1) {
3993 // Do not compile method that is only a trivial infinite loop,
3994 // since the content of the loop may have been eliminated.
3995 record_method_not_compilable("trivial infinite loop");
3996 return true;
3997 }
3998
3999 // Expensive nodes have their control input set to prevent the GVN
4000 // from freely commoning them. There's no GVN beyond this point so
4001 // no need to keep the control input. We want the expensive nodes to
4002 // be freely moved to the least frequent code path by gcm.
4003 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4004 for (int i = 0; i < expensive_count(); i++) {
4005 _expensive_nodes.at(i)->set_req(0, nullptr);
4006 }
4007
4008 Final_Reshape_Counts frc;
4009
4010 // Visit everybody reachable!
4011 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4012 Node_Stack nstack(live_nodes() >> 1);
4013 Unique_Node_List dead_nodes;
4014 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4015
4016 // Check for unreachable (from below) code (i.e., infinite loops).
4017 for( uint i = 0; i < frc._tests.size(); i++ ) {
4018 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4019 // Get number of CFG targets.
4020 // Note that PCTables include exception targets after calls.
4021 uint required_outcnt = n->required_outcnt();
4022 if (n->outcnt() != required_outcnt) {
4023 // Check for a few special cases. Rethrow Nodes never take the
4024 // 'fall-thru' path, so expected kids is 1 less.
4025 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4026 if (n->in(0)->in(0)->is_Call()) {
4027 CallNode* call = n->in(0)->in(0)->as_Call();
4028 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4029 required_outcnt--; // Rethrow always has 1 less kid
4030 } else if (call->req() > TypeFunc::Parms &&
4031 call->is_CallDynamicJava()) {
4032 // Check for null receiver. In such case, the optimizer has
4033 // detected that the virtual call will always result in a null
4034 // pointer exception. The fall-through projection of this CatchNode
4035 // will not be populated.
4036 Node* arg0 = call->in(TypeFunc::Parms);
4037 if (arg0->is_Type() &&
4038 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4039 required_outcnt--;
4040 }
4041 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4042 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4043 // Check for illegal array length. In such case, the optimizer has
4044 // detected that the allocation attempt will always result in an
4045 // exception. There is no fall-through projection of this CatchNode .
4046 assert(call->is_CallStaticJava(), "static call expected");
4047 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4048 uint valid_length_test_input = call->req() - 1;
4049 Node* valid_length_test = call->in(valid_length_test_input);
4050 call->del_req(valid_length_test_input);
4051 if (valid_length_test->find_int_con(1) == 0) {
4052 required_outcnt--;
4053 }
4054 dead_nodes.push(valid_length_test);
4055 assert(n->outcnt() == required_outcnt, "malformed control flow");
4056 continue;
4057 }
4058 }
4059 }
4060
4061 // Recheck with a better notion of 'required_outcnt'
4062 if (n->outcnt() != required_outcnt) {
4063 record_method_not_compilable("malformed control flow");
4064 return true; // Not all targets reachable!
4065 }
4066 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4067 CallNode* call = n->in(0)->in(0)->as_Call();
4068 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4069 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4070 assert(call->is_CallStaticJava(), "static call expected");
4071 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4072 uint valid_length_test_input = call->req() - 1;
4073 dead_nodes.push(call->in(valid_length_test_input));
4074 call->del_req(valid_length_test_input); // valid length test useless now
4075 }
4076 }
4077 // Check that I actually visited all kids. Unreached kids
4078 // must be infinite loops.
4079 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4080 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4081 record_method_not_compilable("infinite loop");
4082 return true; // Found unvisited kid; must be unreach
4083 }
4084
4085 // Here so verification code in final_graph_reshaping_walk()
4086 // always see an OuterStripMinedLoopEnd
4087 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4088 IfNode* init_iff = n->as_If();
4089 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4090 n->subsume_by(iff, this);
4091 }
4092 }
4093
4094 while (dead_nodes.size() > 0) {
4095 Node* m = dead_nodes.pop();
4096 if (m->outcnt() == 0 && m != top()) {
4097 for (uint j = 0; j < m->req(); j++) {
4098 Node* in = m->in(j);
4099 if (in != nullptr) {
4100 dead_nodes.push(in);
4101 }
4102 }
4103 m->disconnect_inputs(this);
4104 }
4105 }
4106
4107 #ifdef IA32
4108 // If original bytecodes contained a mixture of floats and doubles
4109 // check if the optimizer has made it homogeneous, item (3).
4110 if (UseSSE == 0 &&
4111 frc.get_float_count() > 32 &&
4112 frc.get_double_count() == 0 &&
4113 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4114 set_24_bit_selection_and_mode(false, true);
4115 }
4116 #endif // IA32
4117
4118 set_java_calls(frc.get_java_call_count());
4119 set_inner_loops(frc.get_inner_loop_count());
4120
4121 // No infinite loops, no reason to bail out.
4122 return false;
4123 }
4124
4125 //-----------------------------too_many_traps----------------------------------
4126 // Report if there are too many traps at the current method and bci.
4127 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4128 bool Compile::too_many_traps(ciMethod* method,
4129 int bci,
4130 Deoptimization::DeoptReason reason) {
4131 ciMethodData* md = method->method_data();
4132 if (md->is_empty()) {
4133 // Assume the trap has not occurred, or that it occurred only
4134 // because of a transient condition during start-up in the interpreter.
4135 return false;
4136 }
4137 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4138 if (md->has_trap_at(bci, m, reason) != 0) {
4139 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4140 // Also, if there are multiple reasons, or if there is no per-BCI record,
4141 // assume the worst.
4142 if (log())
4143 log()->elem("observe trap='%s' count='%d'",
4144 Deoptimization::trap_reason_name(reason),
4145 md->trap_count(reason));
4146 return true;
4147 } else {
4148 // Ignore method/bci and see if there have been too many globally.
4149 return too_many_traps(reason, md);
4150 }
4151 }
4152
4153 // Less-accurate variant which does not require a method and bci.
4154 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4155 ciMethodData* logmd) {
4156 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4157 // Too many traps globally.
4158 // Note that we use cumulative trap_count, not just md->trap_count.
4159 if (log()) {
4160 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4161 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4162 Deoptimization::trap_reason_name(reason),
4163 mcount, trap_count(reason));
4164 }
4165 return true;
4166 } else {
4167 // The coast is clear.
4168 return false;
4169 }
4170 }
4171
4172 //--------------------------too_many_recompiles--------------------------------
4173 // Report if there are too many recompiles at the current method and bci.
4174 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4175 // Is not eager to return true, since this will cause the compiler to use
4176 // Action_none for a trap point, to avoid too many recompilations.
4177 bool Compile::too_many_recompiles(ciMethod* method,
4178 int bci,
4179 Deoptimization::DeoptReason reason) {
4180 ciMethodData* md = method->method_data();
4181 if (md->is_empty()) {
4182 // Assume the trap has not occurred, or that it occurred only
4183 // because of a transient condition during start-up in the interpreter.
4184 return false;
4185 }
4186 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4187 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4188 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4189 Deoptimization::DeoptReason per_bc_reason
4190 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4191 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4192 if ((per_bc_reason == Deoptimization::Reason_none
4193 || md->has_trap_at(bci, m, reason) != 0)
4194 // The trap frequency measure we care about is the recompile count:
4195 && md->trap_recompiled_at(bci, m)
4196 && md->overflow_recompile_count() >= bc_cutoff) {
4197 // Do not emit a trap here if it has already caused recompilations.
4198 // Also, if there are multiple reasons, or if there is no per-BCI record,
4199 // assume the worst.
4200 if (log())
4201 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4202 Deoptimization::trap_reason_name(reason),
4203 md->trap_count(reason),
4204 md->overflow_recompile_count());
4205 return true;
4206 } else if (trap_count(reason) != 0
4207 && decompile_count() >= m_cutoff) {
4208 // Too many recompiles globally, and we have seen this sort of trap.
4209 // Use cumulative decompile_count, not just md->decompile_count.
4210 if (log())
4211 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4212 Deoptimization::trap_reason_name(reason),
4213 md->trap_count(reason), trap_count(reason),
4214 md->decompile_count(), decompile_count());
4215 return true;
4216 } else {
4217 // The coast is clear.
4218 return false;
4219 }
4220 }
4221
4222 // Compute when not to trap. Used by matching trap based nodes and
4223 // NullCheck optimization.
4224 void Compile::set_allowed_deopt_reasons() {
4225 _allowed_reasons = 0;
4226 if (is_method_compilation()) {
4227 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4228 assert(rs < BitsPerInt, "recode bit map");
4229 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4230 _allowed_reasons |= nth_bit(rs);
4231 }
4232 }
4233 }
4234 }
4235
4236 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4237 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4238 }
4239
4240 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4241 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4242 }
4243
4244 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4245 if (holder->is_initialized()) {
4246 return false;
4247 }
4248 if (holder->is_being_initialized()) {
4249 if (accessing_method->holder() == holder) {
4250 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4251 // <init>, or a static method. In all those cases, there was an initialization
4252 // barrier on the holder klass passed.
4253 if (accessing_method->is_static_initializer() ||
4254 accessing_method->is_object_initializer() ||
4255 accessing_method->is_static()) {
4256 return false;
4257 }
4258 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4259 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4260 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4261 // child class can become fully initialized while its parent class is still being initialized.
4262 if (accessing_method->is_static_initializer()) {
4263 return false;
4264 }
4265 }
4266 ciMethod* root = method(); // the root method of compilation
4267 if (root != accessing_method) {
4268 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4269 }
4270 }
4271 return true;
4272 }
4273
4274 #ifndef PRODUCT
4275 //------------------------------verify_bidirectional_edges---------------------
4276 // For each input edge to a node (ie - for each Use-Def edge), verify that
4277 // there is a corresponding Def-Use edge.
4278 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4279 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4280 uint stack_size = live_nodes() >> 4;
4281 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4282 nstack.push(_root);
4283
4284 while (nstack.size() > 0) {
4285 Node* n = nstack.pop();
4286 if (visited.member(n)) {
4287 continue;
4288 }
4289 visited.push(n);
4290
4291 // Walk over all input edges, checking for correspondence
4292 uint length = n->len();
4293 for (uint i = 0; i < length; i++) {
4294 Node* in = n->in(i);
4295 if (in != nullptr && !visited.member(in)) {
4296 nstack.push(in); // Put it on stack
4297 }
4298 if (in != nullptr && !in->is_top()) {
4299 // Count instances of `next`
4300 int cnt = 0;
4301 for (uint idx = 0; idx < in->_outcnt; idx++) {
4302 if (in->_out[idx] == n) {
4303 cnt++;
4304 }
4305 }
4306 assert(cnt > 0, "Failed to find Def-Use edge.");
4307 // Check for duplicate edges
4308 // walk the input array downcounting the input edges to n
4309 for (uint j = 0; j < length; j++) {
4310 if (n->in(j) == in) {
4311 cnt--;
4312 }
4313 }
4314 assert(cnt == 0, "Mismatched edge count.");
4315 } else if (in == nullptr) {
4316 assert(i == 0 || i >= n->req() ||
4317 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4318 (n->is_Unlock() && i == (n->req() - 1)) ||
4319 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4320 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4321 } else {
4322 assert(in->is_top(), "sanity");
4323 // Nothing to check.
4324 }
4325 }
4326 }
4327 }
4328
4329 //------------------------------verify_graph_edges---------------------------
4330 // Walk the Graph and verify that there is a one-to-one correspondence
4331 // between Use-Def edges and Def-Use edges in the graph.
4332 void Compile::verify_graph_edges(bool no_dead_code) {
4333 if (VerifyGraphEdges) {
4334 Unique_Node_List visited;
4335
4336 // Call graph walk to check edges
4337 verify_bidirectional_edges(visited);
4338 if (no_dead_code) {
4339 // Now make sure that no visited node is used by an unvisited node.
4340 bool dead_nodes = false;
4341 Unique_Node_List checked;
4342 while (visited.size() > 0) {
4343 Node* n = visited.pop();
4344 checked.push(n);
4345 for (uint i = 0; i < n->outcnt(); i++) {
4346 Node* use = n->raw_out(i);
4347 if (checked.member(use)) continue; // already checked
4348 if (visited.member(use)) continue; // already in the graph
4349 if (use->is_Con()) continue; // a dead ConNode is OK
4350 // At this point, we have found a dead node which is DU-reachable.
4351 if (!dead_nodes) {
4352 tty->print_cr("*** Dead nodes reachable via DU edges:");
4353 dead_nodes = true;
4354 }
4355 use->dump(2);
4356 tty->print_cr("---");
4357 checked.push(use); // No repeats; pretend it is now checked.
4358 }
4359 }
4360 assert(!dead_nodes, "using nodes must be reachable from root");
4361 }
4362 }
4363 }
4364 #endif
4365
4366 // The Compile object keeps track of failure reasons separately from the ciEnv.
4367 // This is required because there is not quite a 1-1 relation between the
4368 // ciEnv and its compilation task and the Compile object. Note that one
4369 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4370 // to backtrack and retry without subsuming loads. Other than this backtracking
4371 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4372 // by the logic in C2Compiler.
4373 void Compile::record_failure(const char* reason) {
4374 if (log() != nullptr) {
4375 log()->elem("failure reason='%s' phase='compile'", reason);
4376 }
4377 if (_failure_reason.get() == nullptr) {
4378 // Record the first failure reason.
4379 _failure_reason.set(reason);
4380 if (CaptureBailoutInformation) {
4381 _first_failure_details = new CompilationFailureInfo(reason);
4382 }
4383 }
4384
4385 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4386 C->print_method(PHASE_FAILURE, 1);
4387 }
4388 _root = nullptr; // flush the graph, too
4389 }
4390
4391 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4392 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4393 _compile(Compile::current()),
4394 _log(nullptr),
4395 _phase_name(name),
4396 _dolog(CITimeVerbose)
4397 {
4398 assert(_compile != nullptr, "sanity check");
4399 if (_dolog) {
4400 _log = _compile->log();
4401 }
4402 if (_log != nullptr) {
4403 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4404 _log->stamp();
4405 _log->end_head();
4406 }
4407 }
4408
4409 Compile::TracePhase::~TracePhase() {
4410 if (_compile->failing()) return;
4411 #ifdef ASSERT
4412 if (PrintIdealNodeCount) {
4413 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4414 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4415 }
4416
4417 if (VerifyIdealNodeCount) {
4418 _compile->print_missing_nodes();
4419 }
4420 #endif
4421
4422 if (_log != nullptr) {
4423 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4424 }
4425 }
4426
4427 //----------------------------static_subtype_check-----------------------------
4428 // Shortcut important common cases when superklass is exact:
4429 // (0) superklass is java.lang.Object (can occur in reflective code)
4430 // (1) subklass is already limited to a subtype of superklass => always ok
4431 // (2) subklass does not overlap with superklass => always fail
4432 // (3) superklass has NO subtypes and we can check with a simple compare.
4433 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4434 if (skip) {
4435 return SSC_full_test; // Let caller generate the general case.
4436 }
4437
4438 if (subk->is_java_subtype_of(superk)) {
4439 return SSC_always_true; // (0) and (1) this test cannot fail
4440 }
4441
4442 if (!subk->maybe_java_subtype_of(superk)) {
4443 return SSC_always_false; // (2) true path dead; no dynamic test needed
4444 }
4445
4446 const Type* superelem = superk;
4447 if (superk->isa_aryklassptr()) {
4448 int ignored;
4449 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4450 }
4451
4452 if (superelem->isa_instklassptr()) {
4453 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4454 if (!ik->has_subklass()) {
4455 if (!ik->is_final()) {
4456 // Add a dependency if there is a chance of a later subclass.
4457 dependencies()->assert_leaf_type(ik);
4458 }
4459 if (!superk->maybe_java_subtype_of(subk)) {
4460 return SSC_always_false;
4461 }
4462 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4463 }
4464 } else {
4465 // A primitive array type has no subtypes.
4466 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4467 }
4468
4469 return SSC_full_test;
4470 }
4471
4472 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4473 #ifdef _LP64
4474 // The scaled index operand to AddP must be a clean 64-bit value.
4475 // Java allows a 32-bit int to be incremented to a negative
4476 // value, which appears in a 64-bit register as a large
4477 // positive number. Using that large positive number as an
4478 // operand in pointer arithmetic has bad consequences.
4479 // On the other hand, 32-bit overflow is rare, and the possibility
4480 // can often be excluded, if we annotate the ConvI2L node with
4481 // a type assertion that its value is known to be a small positive
4482 // number. (The prior range check has ensured this.)
4483 // This assertion is used by ConvI2LNode::Ideal.
4484 int index_max = max_jint - 1; // array size is max_jint, index is one less
4485 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4486 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4487 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4488 #endif
4489 return idx;
4490 }
4491
4492 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4493 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4494 if (ctrl != nullptr) {
4495 // Express control dependency by a CastII node with a narrow type.
4496 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4497 // node from floating above the range check during loop optimizations. Otherwise, the
4498 // ConvI2L node may be eliminated independently of the range check, causing the data path
4499 // to become TOP while the control path is still there (although it's unreachable).
4500 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4501 value = phase->transform(value);
4502 }
4503 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4504 return phase->transform(new ConvI2LNode(value, ltype));
4505 }
4506
4507 // The message about the current inlining is accumulated in
4508 // _print_inlining_stream and transferred into the _print_inlining_list
4509 // once we know whether inlining succeeds or not. For regular
4510 // inlining, messages are appended to the buffer pointed by
4511 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4512 // a new buffer is added after _print_inlining_idx in the list. This
4513 // way we can update the inlining message for late inlining call site
4514 // when the inlining is attempted again.
4515 void Compile::print_inlining_init() {
4516 if (print_inlining() || print_intrinsics()) {
4517 // print_inlining_init is actually called several times.
4518 print_inlining_reset();
4519 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4520 }
4521 }
4522
4523 void Compile::print_inlining_reinit() {
4524 if (print_inlining() || print_intrinsics()) {
4525 print_inlining_reset();
4526 }
4527 }
4528
4529 void Compile::print_inlining_reset() {
4530 _print_inlining_stream->reset();
4531 }
4532
4533 void Compile::print_inlining_commit() {
4534 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4535 // Transfer the message from _print_inlining_stream to the current
4536 // _print_inlining_list buffer and clear _print_inlining_stream.
4537 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4538 print_inlining_reset();
4539 }
4540
4541 void Compile::print_inlining_push() {
4542 // Add new buffer to the _print_inlining_list at current position
4543 _print_inlining_idx++;
4544 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4545 }
4546
4547 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4548 return _print_inlining_list->at(_print_inlining_idx);
4549 }
4550
4551 void Compile::print_inlining_update(CallGenerator* cg) {
4552 if (print_inlining() || print_intrinsics()) {
4553 if (cg->is_late_inline()) {
4554 if (print_inlining_current()->cg() != cg &&
4555 (print_inlining_current()->cg() != nullptr ||
4556 print_inlining_current()->ss()->size() != 0)) {
4557 print_inlining_push();
4558 }
4559 print_inlining_commit();
4560 print_inlining_current()->set_cg(cg);
4561 } else {
4562 if (print_inlining_current()->cg() != nullptr) {
4563 print_inlining_push();
4564 }
4565 print_inlining_commit();
4566 }
4567 }
4568 }
4569
4570 void Compile::print_inlining_move_to(CallGenerator* cg) {
4571 // We resume inlining at a late inlining call site. Locate the
4572 // corresponding inlining buffer so that we can update it.
4573 if (print_inlining() || print_intrinsics()) {
4574 for (int i = 0; i < _print_inlining_list->length(); i++) {
4575 if (_print_inlining_list->at(i)->cg() == cg) {
4576 _print_inlining_idx = i;
4577 return;
4578 }
4579 }
4580 ShouldNotReachHere();
4581 }
4582 }
4583
4584 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4585 if (print_inlining() || print_intrinsics()) {
4586 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4587 assert(print_inlining_current()->cg() == cg, "wrong entry");
4588 // replace message with new message
4589 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4590 print_inlining_commit();
4591 print_inlining_current()->set_cg(cg);
4592 }
4593 }
4594
4595 void Compile::print_inlining_assert_ready() {
4596 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4597 }
4598
4599 void Compile::process_print_inlining() {
4600 assert(_late_inlines.length() == 0, "not drained yet");
4601 if (print_inlining() || print_intrinsics()) {
4602 ResourceMark rm;
4603 stringStream ss;
4604 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4605 for (int i = 0; i < _print_inlining_list->length(); i++) {
4606 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4607 ss.print("%s", pib->ss()->freeze());
4608 delete pib;
4609 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4610 }
4611 // Reset _print_inlining_list, it only contains destructed objects.
4612 // It is on the arena, so it will be freed when the arena is reset.
4613 _print_inlining_list = nullptr;
4614 // _print_inlining_stream won't be used anymore, either.
4615 print_inlining_reset();
4616 size_t end = ss.size();
4617 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4618 strncpy(_print_inlining_output, ss.freeze(), end+1);
4619 _print_inlining_output[end] = 0;
4620 }
4621 }
4622
4623 void Compile::dump_print_inlining() {
4624 if (_print_inlining_output != nullptr) {
4625 tty->print_raw(_print_inlining_output);
4626 }
4627 }
4628
4629 void Compile::log_late_inline(CallGenerator* cg) {
4630 if (log() != nullptr) {
4631 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4632 cg->unique_id());
4633 JVMState* p = cg->call_node()->jvms();
4634 while (p != nullptr) {
4635 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4636 p = p->caller();
4637 }
4638 log()->tail("late_inline");
4639 }
4640 }
4641
4642 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4643 log_late_inline(cg);
4644 if (log() != nullptr) {
4645 log()->inline_fail(msg);
4646 }
4647 }
4648
4649 void Compile::log_inline_id(CallGenerator* cg) {
4650 if (log() != nullptr) {
4651 // The LogCompilation tool needs a unique way to identify late
4652 // inline call sites. This id must be unique for this call site in
4653 // this compilation. Try to have it unique across compilations as
4654 // well because it can be convenient when grepping through the log
4655 // file.
4656 // Distinguish OSR compilations from others in case CICountOSR is
4657 // on.
4658 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4659 cg->set_unique_id(id);
4660 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4661 }
4662 }
4663
4664 void Compile::log_inline_failure(const char* msg) {
4665 if (C->log() != nullptr) {
4666 C->log()->inline_fail(msg);
4667 }
4668 }
4669
4670
4671 // Dump inlining replay data to the stream.
4672 // Don't change thread state and acquire any locks.
4673 void Compile::dump_inline_data(outputStream* out) {
4674 InlineTree* inl_tree = ilt();
4675 if (inl_tree != nullptr) {
4676 out->print(" inline %d", inl_tree->count());
4677 inl_tree->dump_replay_data(out);
4678 }
4679 }
4680
4681 void Compile::dump_inline_data_reduced(outputStream* out) {
4682 assert(ReplayReduce, "");
4683
4684 InlineTree* inl_tree = ilt();
4685 if (inl_tree == nullptr) {
4686 return;
4687 }
4688 // Enable iterative replay file reduction
4689 // Output "compile" lines for depth 1 subtrees,
4690 // simulating that those trees were compiled
4691 // instead of inlined.
4692 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4693 InlineTree* sub = inl_tree->subtrees().at(i);
4694 if (sub->inline_level() != 1) {
4695 continue;
4696 }
4697
4698 ciMethod* method = sub->method();
4699 int entry_bci = -1;
4700 int comp_level = env()->task()->comp_level();
4701 out->print("compile ");
4702 method->dump_name_as_ascii(out);
4703 out->print(" %d %d", entry_bci, comp_level);
4704 out->print(" inline %d", sub->count());
4705 sub->dump_replay_data(out, -1);
4706 out->cr();
4707 }
4708 }
4709
4710 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4711 if (n1->Opcode() < n2->Opcode()) return -1;
4712 else if (n1->Opcode() > n2->Opcode()) return 1;
4713
4714 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4715 for (uint i = 1; i < n1->req(); i++) {
4716 if (n1->in(i) < n2->in(i)) return -1;
4717 else if (n1->in(i) > n2->in(i)) return 1;
4718 }
4719
4720 return 0;
4721 }
4722
4723 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4724 Node* n1 = *n1p;
4725 Node* n2 = *n2p;
4726
4727 return cmp_expensive_nodes(n1, n2);
4728 }
4729
4730 void Compile::sort_expensive_nodes() {
4731 if (!expensive_nodes_sorted()) {
4732 _expensive_nodes.sort(cmp_expensive_nodes);
4733 }
4734 }
4735
4736 bool Compile::expensive_nodes_sorted() const {
4737 for (int i = 1; i < _expensive_nodes.length(); i++) {
4738 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4739 return false;
4740 }
4741 }
4742 return true;
4743 }
4744
4745 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4746 if (_expensive_nodes.length() == 0) {
4747 return false;
4748 }
4749
4750 assert(OptimizeExpensiveOps, "optimization off?");
4751
4752 // Take this opportunity to remove dead nodes from the list
4753 int j = 0;
4754 for (int i = 0; i < _expensive_nodes.length(); i++) {
4755 Node* n = _expensive_nodes.at(i);
4756 if (!n->is_unreachable(igvn)) {
4757 assert(n->is_expensive(), "should be expensive");
4758 _expensive_nodes.at_put(j, n);
4759 j++;
4760 }
4761 }
4762 _expensive_nodes.trunc_to(j);
4763
4764 // Then sort the list so that similar nodes are next to each other
4765 // and check for at least two nodes of identical kind with same data
4766 // inputs.
4767 sort_expensive_nodes();
4768
4769 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4770 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4771 return true;
4772 }
4773 }
4774
4775 return false;
4776 }
4777
4778 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4779 if (_expensive_nodes.length() == 0) {
4780 return;
4781 }
4782
4783 assert(OptimizeExpensiveOps, "optimization off?");
4784
4785 // Sort to bring similar nodes next to each other and clear the
4786 // control input of nodes for which there's only a single copy.
4787 sort_expensive_nodes();
4788
4789 int j = 0;
4790 int identical = 0;
4791 int i = 0;
4792 bool modified = false;
4793 for (; i < _expensive_nodes.length()-1; i++) {
4794 assert(j <= i, "can't write beyond current index");
4795 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4796 identical++;
4797 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4798 continue;
4799 }
4800 if (identical > 0) {
4801 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4802 identical = 0;
4803 } else {
4804 Node* n = _expensive_nodes.at(i);
4805 igvn.replace_input_of(n, 0, nullptr);
4806 igvn.hash_insert(n);
4807 modified = true;
4808 }
4809 }
4810 if (identical > 0) {
4811 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4812 } else if (_expensive_nodes.length() >= 1) {
4813 Node* n = _expensive_nodes.at(i);
4814 igvn.replace_input_of(n, 0, nullptr);
4815 igvn.hash_insert(n);
4816 modified = true;
4817 }
4818 _expensive_nodes.trunc_to(j);
4819 if (modified) {
4820 igvn.optimize();
4821 }
4822 }
4823
4824 void Compile::add_expensive_node(Node * n) {
4825 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4826 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4827 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4828 if (OptimizeExpensiveOps) {
4829 _expensive_nodes.append(n);
4830 } else {
4831 // Clear control input and let IGVN optimize expensive nodes if
4832 // OptimizeExpensiveOps is off.
4833 n->set_req(0, nullptr);
4834 }
4835 }
4836
4837 /**
4838 * Track coarsened Lock and Unlock nodes.
4839 */
4840
4841 class Lock_List : public Node_List {
4842 uint _origin_cnt;
4843 public:
4844 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4845 uint origin_cnt() const { return _origin_cnt; }
4846 };
4847
4848 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4849 int length = locks.length();
4850 if (length > 0) {
4851 // Have to keep this list until locks elimination during Macro nodes elimination.
4852 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4853 AbstractLockNode* alock = locks.at(0);
4854 BoxLockNode* box = alock->box_node()->as_BoxLock();
4855 for (int i = 0; i < length; i++) {
4856 AbstractLockNode* lock = locks.at(i);
4857 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4858 locks_list->push(lock);
4859 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4860 if (this_box != box) {
4861 // Locking regions (BoxLock) could be Unbalanced here:
4862 // - its coarsened locks were eliminated in earlier
4863 // macro nodes elimination followed by loop unroll
4864 // - it is OSR locking region (no Lock node)
4865 // Preserve Unbalanced status in such cases.
4866 if (!this_box->is_unbalanced()) {
4867 this_box->set_coarsened();
4868 }
4869 if (!box->is_unbalanced()) {
4870 box->set_coarsened();
4871 }
4872 }
4873 }
4874 _coarsened_locks.append(locks_list);
4875 }
4876 }
4877
4878 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4879 int count = coarsened_count();
4880 for (int i = 0; i < count; i++) {
4881 Node_List* locks_list = _coarsened_locks.at(i);
4882 for (uint j = 0; j < locks_list->size(); j++) {
4883 Node* lock = locks_list->at(j);
4884 assert(lock->is_AbstractLock(), "sanity");
4885 if (!useful.member(lock)) {
4886 locks_list->yank(lock);
4887 }
4888 }
4889 }
4890 }
4891
4892 void Compile::remove_coarsened_lock(Node* n) {
4893 if (n->is_AbstractLock()) {
4894 int count = coarsened_count();
4895 for (int i = 0; i < count; i++) {
4896 Node_List* locks_list = _coarsened_locks.at(i);
4897 locks_list->yank(n);
4898 }
4899 }
4900 }
4901
4902 bool Compile::coarsened_locks_consistent() {
4903 int count = coarsened_count();
4904 for (int i = 0; i < count; i++) {
4905 bool unbalanced = false;
4906 bool modified = false; // track locks kind modifications
4907 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4908 uint size = locks_list->size();
4909 if (size == 0) {
4910 unbalanced = false; // All locks were eliminated - good
4911 } else if (size != locks_list->origin_cnt()) {
4912 unbalanced = true; // Some locks were removed from list
4913 } else {
4914 for (uint j = 0; j < size; j++) {
4915 Node* lock = locks_list->at(j);
4916 // All nodes in group should have the same state (modified or not)
4917 if (!lock->as_AbstractLock()->is_coarsened()) {
4918 if (j == 0) {
4919 // first on list was modified, the rest should be too for consistency
4920 modified = true;
4921 } else if (!modified) {
4922 // this lock was modified but previous locks on the list were not
4923 unbalanced = true;
4924 break;
4925 }
4926 } else if (modified) {
4927 // previous locks on list were modified but not this lock
4928 unbalanced = true;
4929 break;
4930 }
4931 }
4932 }
4933 if (unbalanced) {
4934 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4935 #ifdef ASSERT
4936 if (PrintEliminateLocks) {
4937 tty->print_cr("=== unbalanced coarsened locks ===");
4938 for (uint l = 0; l < size; l++) {
4939 locks_list->at(l)->dump();
4940 }
4941 }
4942 #endif
4943 record_failure(C2Compiler::retry_no_locks_coarsening());
4944 return false;
4945 }
4946 }
4947 return true;
4948 }
4949
4950 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4951 // locks coarsening optimization removed Lock/Unlock nodes from them.
4952 // Such regions become unbalanced because coarsening only removes part
4953 // of Lock/Unlock nodes in region. As result we can't execute other
4954 // locks elimination optimizations which assume all code paths have
4955 // corresponding pair of Lock/Unlock nodes - they are balanced.
4956 void Compile::mark_unbalanced_boxes() const {
4957 int count = coarsened_count();
4958 for (int i = 0; i < count; i++) {
4959 Node_List* locks_list = _coarsened_locks.at(i);
4960 uint size = locks_list->size();
4961 if (size > 0) {
4962 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4963 BoxLockNode* box = alock->box_node()->as_BoxLock();
4964 if (alock->is_coarsened()) {
4965 // coarsened_locks_consistent(), which is called before this method, verifies
4966 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4967 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4968 for (uint j = 1; j < size; j++) {
4969 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4970 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4971 if (box != this_box) {
4972 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4973 box->set_unbalanced();
4974 this_box->set_unbalanced();
4975 }
4976 }
4977 }
4978 }
4979 }
4980 }
4981
4982 /**
4983 * Remove the speculative part of types and clean up the graph
4984 */
4985 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4986 if (UseTypeSpeculation) {
4987 Unique_Node_List worklist;
4988 worklist.push(root());
4989 int modified = 0;
4990 // Go over all type nodes that carry a speculative type, drop the
4991 // speculative part of the type and enqueue the node for an igvn
4992 // which may optimize it out.
4993 for (uint next = 0; next < worklist.size(); ++next) {
4994 Node *n = worklist.at(next);
4995 if (n->is_Type()) {
4996 TypeNode* tn = n->as_Type();
4997 const Type* t = tn->type();
4998 const Type* t_no_spec = t->remove_speculative();
4999 if (t_no_spec != t) {
5000 bool in_hash = igvn.hash_delete(n);
5001 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5002 tn->set_type(t_no_spec);
5003 igvn.hash_insert(n);
5004 igvn._worklist.push(n); // give it a chance to go away
5005 modified++;
5006 }
5007 }
5008 // Iterate over outs - endless loops is unreachable from below
5009 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5010 Node *m = n->fast_out(i);
5011 if (not_a_node(m)) {
5012 continue;
5013 }
5014 worklist.push(m);
5015 }
5016 }
5017 // Drop the speculative part of all types in the igvn's type table
5018 igvn.remove_speculative_types();
5019 if (modified > 0) {
5020 igvn.optimize();
5021 if (failing()) return;
5022 }
5023 #ifdef ASSERT
5024 // Verify that after the IGVN is over no speculative type has resurfaced
5025 worklist.clear();
5026 worklist.push(root());
5027 for (uint next = 0; next < worklist.size(); ++next) {
5028 Node *n = worklist.at(next);
5029 const Type* t = igvn.type_or_null(n);
5030 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5031 if (n->is_Type()) {
5032 t = n->as_Type()->type();
5033 assert(t == t->remove_speculative(), "no more speculative types");
5034 }
5035 // Iterate over outs - endless loops is unreachable from below
5036 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5037 Node *m = n->fast_out(i);
5038 if (not_a_node(m)) {
5039 continue;
5040 }
5041 worklist.push(m);
5042 }
5043 }
5044 igvn.check_no_speculative_types();
5045 #endif
5046 }
5047 }
5048
5049 // Auxiliary methods to support randomized stressing/fuzzing.
5050
5051 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5052 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5053 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5054 FLAG_SET_ERGO(StressSeed, _stress_seed);
5055 } else {
5056 _stress_seed = StressSeed;
5057 }
5058 if (_log != nullptr) {
5059 _log->elem("stress_test seed='%u'", _stress_seed);
5060 }
5061 }
5062
5063 int Compile::random() {
5064 _stress_seed = os::next_random(_stress_seed);
5065 return static_cast<int>(_stress_seed);
5066 }
5067
5068 // This method can be called the arbitrary number of times, with current count
5069 // as the argument. The logic allows selecting a single candidate from the
5070 // running list of candidates as follows:
5071 // int count = 0;
5072 // Cand* selected = null;
5073 // while(cand = cand->next()) {
5074 // if (randomized_select(++count)) {
5075 // selected = cand;
5076 // }
5077 // }
5078 //
5079 // Including count equalizes the chances any candidate is "selected".
5080 // This is useful when we don't have the complete list of candidates to choose
5081 // from uniformly. In this case, we need to adjust the randomicity of the
5082 // selection, or else we will end up biasing the selection towards the latter
5083 // candidates.
5084 //
5085 // Quick back-envelope calculation shows that for the list of n candidates
5086 // the equal probability for the candidate to persist as "best" can be
5087 // achieved by replacing it with "next" k-th candidate with the probability
5088 // of 1/k. It can be easily shown that by the end of the run, the
5089 // probability for any candidate is converged to 1/n, thus giving the
5090 // uniform distribution among all the candidates.
5091 //
5092 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5093 #define RANDOMIZED_DOMAIN_POW 29
5094 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5095 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5096 bool Compile::randomized_select(int count) {
5097 assert(count > 0, "only positive");
5098 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5099 }
5100
5101 CloneMap& Compile::clone_map() { return _clone_map; }
5102 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5103
5104 void NodeCloneInfo::dump_on(outputStream* st) const {
5105 st->print(" {%d:%d} ", idx(), gen());
5106 }
5107
5108 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5109 uint64_t val = value(old->_idx);
5110 NodeCloneInfo cio(val);
5111 assert(val != 0, "old node should be in the map");
5112 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5113 insert(nnn->_idx, cin.get());
5114 #ifndef PRODUCT
5115 if (is_debug()) {
5116 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5117 }
5118 #endif
5119 }
5120
5121 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5122 NodeCloneInfo cio(value(old->_idx));
5123 if (cio.get() == 0) {
5124 cio.set(old->_idx, 0);
5125 insert(old->_idx, cio.get());
5126 #ifndef PRODUCT
5127 if (is_debug()) {
5128 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5129 }
5130 #endif
5131 }
5132 clone(old, nnn, gen);
5133 }
5134
5135 int CloneMap::max_gen() const {
5136 int g = 0;
5137 DictI di(_dict);
5138 for(; di.test(); ++di) {
5139 int t = gen(di._key);
5140 if (g < t) {
5141 g = t;
5142 #ifndef PRODUCT
5143 if (is_debug()) {
5144 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5145 }
5146 #endif
5147 }
5148 }
5149 return g;
5150 }
5151
5152 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5153 uint64_t val = value(key);
5154 if (val != 0) {
5155 NodeCloneInfo ni(val);
5156 ni.dump_on(st);
5157 }
5158 }
5159
5160 void Compile::shuffle_macro_nodes() {
5161 if (_macro_nodes.length() < 2) {
5162 return;
5163 }
5164 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5165 uint j = C->random() % (i + 1);
5166 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5167 }
5168 }
5169
5170 // Move Allocate nodes to the start of the list
5171 void Compile::sort_macro_nodes() {
5172 int count = macro_count();
5173 int allocates = 0;
5174 for (int i = 0; i < count; i++) {
5175 Node* n = macro_node(i);
5176 if (n->is_Allocate()) {
5177 if (i != allocates) {
5178 Node* tmp = macro_node(allocates);
5179 _macro_nodes.at_put(allocates, n);
5180 _macro_nodes.at_put(i, tmp);
5181 }
5182 allocates++;
5183 }
5184 }
5185 }
5186
5187 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5188 if (failing()) { return; }
5189 EventCompilerPhase event(UNTIMED);
5190 if (event.should_commit()) {
5191 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5192 }
5193 #ifndef PRODUCT
5194 ResourceMark rm;
5195 stringStream ss;
5196 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5197 int iter = ++_igv_phase_iter[cpt];
5198 if (iter > 1) {
5199 ss.print(" %d", iter);
5200 }
5201 if (n != nullptr) {
5202 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5203 }
5204
5205 const char* name = ss.as_string();
5206 if (should_print_igv(level)) {
5207 _igv_printer->print_method(name, level);
5208 }
5209 if (should_print_phase(cpt)) {
5210 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5211 }
5212 #endif
5213 C->_latest_stage_start_counter.stamp();
5214 }
5215
5216 // Only used from CompileWrapper
5217 void Compile::begin_method() {
5218 #ifndef PRODUCT
5219 if (_method != nullptr && should_print_igv(1)) {
5220 _igv_printer->begin_method();
5221 }
5222 #endif
5223 C->_latest_stage_start_counter.stamp();
5224 }
5225
5226 // Only used from CompileWrapper
5227 void Compile::end_method() {
5228 EventCompilerPhase event(UNTIMED);
5229 if (event.should_commit()) {
5230 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5231 }
5232
5233 #ifndef PRODUCT
5234 if (_method != nullptr && should_print_igv(1)) {
5235 _igv_printer->end_method();
5236 }
5237 #endif
5238 }
5239
5240 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5241 #ifndef PRODUCT
5242 if (_directive->should_print_phase(cpt)) {
5243 return true;
5244 }
5245 #endif
5246 return false;
5247 }
5248
5249 bool Compile::should_print_igv(const int level) {
5250 #ifndef PRODUCT
5251 if (PrintIdealGraphLevel < 0) { // disabled by the user
5252 return false;
5253 }
5254
5255 bool need = directive()->IGVPrintLevelOption >= level;
5256 if (need && !_igv_printer) {
5257 _igv_printer = IdealGraphPrinter::printer();
5258 _igv_printer->set_compile(this);
5259 }
5260 return need;
5261 #else
5262 return false;
5263 #endif
5264 }
5265
5266 #ifndef PRODUCT
5267 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5268 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5269
5270 // Called from debugger. Prints method to the default file with the default phase name.
5271 // This works regardless of any Ideal Graph Visualizer flags set or not.
5272 void igv_print() {
5273 Compile::current()->igv_print_method_to_file();
5274 }
5275
5276 // Same as igv_print() above but with a specified phase name.
5277 void igv_print(const char* phase_name) {
5278 Compile::current()->igv_print_method_to_file(phase_name);
5279 }
5280
5281 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5282 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5283 // This works regardless of any Ideal Graph Visualizer flags set or not.
5284 void igv_print(bool network) {
5285 if (network) {
5286 Compile::current()->igv_print_method_to_network();
5287 } else {
5288 Compile::current()->igv_print_method_to_file();
5289 }
5290 }
5291
5292 // Same as igv_print(bool network) above but with a specified phase name.
5293 void igv_print(bool network, const char* phase_name) {
5294 if (network) {
5295 Compile::current()->igv_print_method_to_network(phase_name);
5296 } else {
5297 Compile::current()->igv_print_method_to_file(phase_name);
5298 }
5299 }
5300
5301 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5302 void igv_print_default() {
5303 Compile::current()->print_method(PHASE_DEBUG, 0);
5304 }
5305
5306 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5307 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5308 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5309 void igv_append() {
5310 Compile::current()->igv_print_method_to_file("Debug", true);
5311 }
5312
5313 // Same as igv_append() above but with a specified phase name.
5314 void igv_append(const char* phase_name) {
5315 Compile::current()->igv_print_method_to_file(phase_name, true);
5316 }
5317
5318 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5319 const char* file_name = "custom_debug.xml";
5320 if (_debug_file_printer == nullptr) {
5321 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5322 } else {
5323 _debug_file_printer->update_compiled_method(C->method());
5324 }
5325 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5326 _debug_file_printer->print(phase_name, (Node*)C->root());
5327 }
5328
5329 void Compile::igv_print_method_to_network(const char* phase_name) {
5330 if (_debug_network_printer == nullptr) {
5331 _debug_network_printer = new IdealGraphPrinter(C);
5332 } else {
5333 _debug_network_printer->update_compiled_method(C->method());
5334 }
5335 tty->print_cr("Method printed over network stream to IGV");
5336 _debug_network_printer->print(phase_name, (Node*)C->root());
5337 }
5338 #endif
5339
5340 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5341 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5342 return value;
5343 }
5344 Node* result = nullptr;
5345 if (bt == T_BYTE) {
5346 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5347 result = new RShiftINode(result, phase->intcon(24));
5348 } else if (bt == T_BOOLEAN) {
5349 result = new AndINode(value, phase->intcon(0xFF));
5350 } else if (bt == T_CHAR) {
5351 result = new AndINode(value,phase->intcon(0xFFFF));
5352 } else {
5353 assert(bt == T_SHORT, "unexpected narrow type");
5354 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5355 result = new RShiftINode(result, phase->intcon(16));
5356 }
5357 if (transform_res) {
5358 result = phase->transform(result);
5359 }
5360 return result;
5361 }
5362
5363 void Compile::record_method_not_compilable_oom() {
5364 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5365 }