1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/macroAssembler.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "ci/ciReplay.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "code/aotCodeCache.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compiler_globals.hpp"
37 #include "compiler/compilerDefinitions.hpp"
38 #include "compiler/compilerOracle.hpp"
39 #include "compiler/disassembler.hpp"
40 #include "compiler/oopMap.hpp"
41 #include "gc/shared/barrierSet.hpp"
42 #include "gc/shared/c2/barrierSetC2.hpp"
43 #include "jfr/jfrEvents.hpp"
44 #include "jvm_io.h"
45 #include "memory/allocation.hpp"
46 #include "memory/arena.hpp"
47 #include "memory/resourceArea.hpp"
48 #include "opto/addnode.hpp"
49 #include "opto/block.hpp"
50 #include "opto/c2compiler.hpp"
51 #include "opto/callGenerator.hpp"
52 #include "opto/callnode.hpp"
53 #include "opto/castnode.hpp"
54 #include "opto/cfgnode.hpp"
55 #include "opto/chaitin.hpp"
56 #include "opto/compile.hpp"
57 #include "opto/connode.hpp"
58 #include "opto/convertnode.hpp"
59 #include "opto/divnode.hpp"
60 #include "opto/escape.hpp"
61 #include "opto/idealGraphPrinter.hpp"
62 #include "opto/locknode.hpp"
63 #include "opto/loopnode.hpp"
64 #include "opto/machnode.hpp"
65 #include "opto/macro.hpp"
66 #include "opto/matcher.hpp"
67 #include "opto/mathexactnode.hpp"
68 #include "opto/memnode.hpp"
69 #include "opto/mulnode.hpp"
70 #include "opto/narrowptrnode.hpp"
71 #include "opto/node.hpp"
72 #include "opto/opaquenode.hpp"
73 #include "opto/opcodes.hpp"
74 #include "opto/output.hpp"
75 #include "opto/parse.hpp"
76 #include "opto/phaseX.hpp"
77 #include "opto/rootnode.hpp"
78 #include "opto/runtime.hpp"
79 #include "opto/stringopts.hpp"
80 #include "opto/type.hpp"
81 #include "opto/vector.hpp"
82 #include "opto/vectornode.hpp"
83 #include "runtime/globals_extension.hpp"
84 #include "runtime/sharedRuntime.hpp"
85 #include "runtime/signature.hpp"
86 #include "runtime/stubRoutines.hpp"
87 #include "runtime/timer.hpp"
88 #include "utilities/align.hpp"
89 #include "utilities/copy.hpp"
90 #include "utilities/hashTable.hpp"
91 #include "utilities/macros.hpp"
92
93 // -------------------- Compile::mach_constant_base_node -----------------------
94 // Constant table base node singleton.
95 MachConstantBaseNode* Compile::mach_constant_base_node() {
96 if (_mach_constant_base_node == nullptr) {
97 _mach_constant_base_node = new MachConstantBaseNode();
98 _mach_constant_base_node->add_req(C->root());
99 }
100 return _mach_constant_base_node;
101 }
102
103
104 /// Support for intrinsics.
105
106 // Return the index at which m must be inserted (or already exists).
107 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
108 class IntrinsicDescPair {
109 private:
110 ciMethod* _m;
111 bool _is_virtual;
112 public:
113 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
114 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
115 ciMethod* m= elt->method();
116 ciMethod* key_m = key->_m;
117 if (key_m < m) return -1;
118 else if (key_m > m) return 1;
119 else {
120 bool is_virtual = elt->is_virtual();
121 bool key_virtual = key->_is_virtual;
122 if (key_virtual < is_virtual) return -1;
123 else if (key_virtual > is_virtual) return 1;
124 else return 0;
125 }
126 }
127 };
128 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
129 #ifdef ASSERT
130 for (int i = 1; i < _intrinsics.length(); i++) {
131 CallGenerator* cg1 = _intrinsics.at(i-1);
132 CallGenerator* cg2 = _intrinsics.at(i);
133 assert(cg1->method() != cg2->method()
134 ? cg1->method() < cg2->method()
135 : cg1->is_virtual() < cg2->is_virtual(),
136 "compiler intrinsics list must stay sorted");
137 }
138 #endif
139 IntrinsicDescPair pair(m, is_virtual);
140 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
141 }
142
143 void Compile::register_intrinsic(CallGenerator* cg) {
144 bool found = false;
145 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
146 assert(!found, "registering twice");
147 _intrinsics.insert_before(index, cg);
148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
149 }
150
151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
152 assert(m->is_loaded(), "don't try this on unloaded methods");
153 if (_intrinsics.length() > 0) {
154 bool found = false;
155 int index = intrinsic_insertion_index(m, is_virtual, found);
156 if (found) {
157 return _intrinsics.at(index);
158 }
159 }
160 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
161 if (m->intrinsic_id() != vmIntrinsics::_none &&
162 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
163 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
164 if (cg != nullptr) {
165 // Save it for next time:
166 register_intrinsic(cg);
167 return cg;
168 } else {
169 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
170 }
171 }
172 return nullptr;
173 }
174
175 // Compile::make_vm_intrinsic is defined in library_call.cpp.
176
177 #ifndef PRODUCT
178 // statistics gathering...
179
180 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
182
183 inline int as_int(vmIntrinsics::ID id) {
184 return vmIntrinsics::as_int(id);
185 }
186
187 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
188 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
189 int oflags = _intrinsic_hist_flags[as_int(id)];
190 assert(flags != 0, "what happened?");
191 if (is_virtual) {
192 flags |= _intrinsic_virtual;
193 }
194 bool changed = (flags != oflags);
195 if ((flags & _intrinsic_worked) != 0) {
196 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
197 if (count == 1) {
198 changed = true; // first time
199 }
200 // increment the overall count also:
201 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
202 }
203 if (changed) {
204 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
205 // Something changed about the intrinsic's virtuality.
206 if ((flags & _intrinsic_virtual) != 0) {
207 // This is the first use of this intrinsic as a virtual call.
208 if (oflags != 0) {
209 // We already saw it as a non-virtual, so note both cases.
210 flags |= _intrinsic_both;
211 }
212 } else if ((oflags & _intrinsic_both) == 0) {
213 // This is the first use of this intrinsic as a non-virtual
214 flags |= _intrinsic_both;
215 }
216 }
217 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
218 }
219 // update the overall flags also:
220 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
221 return changed;
222 }
223
224 static char* format_flags(int flags, char* buf) {
225 buf[0] = 0;
226 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
227 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
228 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
229 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
230 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
231 if (buf[0] == 0) strcat(buf, ",");
232 assert(buf[0] == ',', "must be");
233 return &buf[1];
234 }
235
236 void Compile::print_intrinsic_statistics() {
237 char flagsbuf[100];
238 ttyLocker ttyl;
239 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
240 tty->print_cr("Compiler intrinsic usage:");
241 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
242 if (total == 0) total = 1; // avoid div0 in case of no successes
243 #define PRINT_STAT_LINE(name, c, f) \
244 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
245 for (auto id : EnumRange<vmIntrinsicID>{}) {
246 int flags = _intrinsic_hist_flags[as_int(id)];
247 juint count = _intrinsic_hist_count[as_int(id)];
248 if ((flags | count) != 0) {
249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
250 }
251 }
252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
253 if (xtty != nullptr) xtty->tail("statistics");
254 }
255
256 void Compile::print_statistics() {
257 { ttyLocker ttyl;
258 if (xtty != nullptr) xtty->head("statistics type='opto'");
259 Parse::print_statistics();
260 PhaseStringOpts::print_statistics();
261 PhaseCCP::print_statistics();
262 PhaseRegAlloc::print_statistics();
263 PhaseOutput::print_statistics();
264 PhasePeephole::print_statistics();
265 PhaseIdealLoop::print_statistics();
266 ConnectionGraph::print_statistics();
267 PhaseMacroExpand::print_statistics();
268 if (xtty != nullptr) xtty->tail("statistics");
269 }
270 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
271 // put this under its own <statistics> element.
272 print_intrinsic_statistics();
273 }
274 }
275 #endif //PRODUCT
276
277 void Compile::gvn_replace_by(Node* n, Node* nn) {
278 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
279 Node* use = n->last_out(i);
280 bool is_in_table = initial_gvn()->hash_delete(use);
281 uint uses_found = 0;
282 for (uint j = 0; j < use->len(); j++) {
283 if (use->in(j) == n) {
284 if (j < use->req())
285 use->set_req(j, nn);
286 else
287 use->set_prec(j, nn);
288 uses_found++;
289 }
290 }
291 if (is_in_table) {
292 // reinsert into table
293 initial_gvn()->hash_find_insert(use);
294 }
295 record_for_igvn(use);
296 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
297 i -= uses_found; // we deleted 1 or more copies of this edge
298 }
299 }
300
301
302 // Identify all nodes that are reachable from below, useful.
303 // Use breadth-first pass that records state in a Unique_Node_List,
304 // recursive traversal is slower.
305 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
306 int estimated_worklist_size = live_nodes();
307 useful.map( estimated_worklist_size, nullptr ); // preallocate space
308
309 // Initialize worklist
310 if (root() != nullptr) { useful.push(root()); }
311 // If 'top' is cached, declare it useful to preserve cached node
312 if (cached_top_node()) { useful.push(cached_top_node()); }
313
314 // Push all useful nodes onto the list, breadthfirst
315 for( uint next = 0; next < useful.size(); ++next ) {
316 assert( next < unique(), "Unique useful nodes < total nodes");
317 Node *n = useful.at(next);
318 uint max = n->len();
319 for( uint i = 0; i < max; ++i ) {
320 Node *m = n->in(i);
321 if (not_a_node(m)) continue;
322 useful.push(m);
323 }
324 }
325 }
326
327 // Update dead_node_list with any missing dead nodes using useful
328 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
329 void Compile::update_dead_node_list(Unique_Node_List &useful) {
330 uint max_idx = unique();
331 VectorSet& useful_node_set = useful.member_set();
332
333 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
334 // If node with index node_idx is not in useful set,
335 // mark it as dead in dead node list.
336 if (!useful_node_set.test(node_idx)) {
337 record_dead_node(node_idx);
338 }
339 }
340 }
341
342 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
343 int shift = 0;
344 for (int i = 0; i < inlines->length(); i++) {
345 CallGenerator* cg = inlines->at(i);
346 if (useful.member(cg->call_node())) {
347 if (shift > 0) {
348 inlines->at_put(i - shift, cg);
349 }
350 } else {
351 shift++; // skip over the dead element
352 }
353 }
354 if (shift > 0) {
355 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
356 }
357 }
358
359 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
360 assert(dead != nullptr && dead->is_Call(), "sanity");
361 int found = 0;
362 for (int i = 0; i < inlines->length(); i++) {
363 if (inlines->at(i)->call_node() == dead) {
364 inlines->remove_at(i);
365 found++;
366 NOT_DEBUG( break; ) // elements are unique, so exit early
367 }
368 }
369 assert(found <= 1, "not unique");
370 }
371
372 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
373 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
374 for (int i = node_list.length() - 1; i >= 0; i--) {
375 N* node = node_list.at(i);
376 if (!useful.member(node)) {
377 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
378 }
379 }
380 }
381
382 void Compile::remove_useless_node(Node* dead) {
383 remove_modified_node(dead);
384
385 // Constant node that has no out-edges and has only one in-edge from
386 // root is usually dead. However, sometimes reshaping walk makes
387 // it reachable by adding use edges. So, we will NOT count Con nodes
388 // as dead to be conservative about the dead node count at any
389 // given time.
390 if (!dead->is_Con()) {
391 record_dead_node(dead->_idx);
392 }
393 if (dead->is_macro()) {
394 remove_macro_node(dead);
395 }
396 if (dead->is_expensive()) {
397 remove_expensive_node(dead);
398 }
399 if (dead->is_OpaqueTemplateAssertionPredicate()) {
400 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
401 }
402 if (dead->is_ParsePredicate()) {
403 remove_parse_predicate(dead->as_ParsePredicate());
404 }
405 if (dead->for_post_loop_opts_igvn()) {
406 remove_from_post_loop_opts_igvn(dead);
407 }
408 if (dead->for_merge_stores_igvn()) {
409 remove_from_merge_stores_igvn(dead);
410 }
411 if (dead->is_Call()) {
412 remove_useless_late_inlines( &_late_inlines, dead);
413 remove_useless_late_inlines( &_string_late_inlines, dead);
414 remove_useless_late_inlines( &_boxing_late_inlines, dead);
415 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
416
417 if (dead->is_CallStaticJava()) {
418 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
419 }
420 }
421 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
422 bs->unregister_potential_barrier_node(dead);
423 }
424
425 // Disconnect all useless nodes by disconnecting those at the boundary.
426 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
427 uint next = 0;
428 while (next < useful.size()) {
429 Node *n = useful.at(next++);
430 if (n->is_SafePoint()) {
431 // We're done with a parsing phase. Replaced nodes are not valid
432 // beyond that point.
433 n->as_SafePoint()->delete_replaced_nodes();
434 }
435 // Use raw traversal of out edges since this code removes out edges
436 int max = n->outcnt();
437 for (int j = 0; j < max; ++j) {
438 Node* child = n->raw_out(j);
439 if (!useful.member(child)) {
440 assert(!child->is_top() || child != top(),
441 "If top is cached in Compile object it is in useful list");
442 // Only need to remove this out-edge to the useless node
443 n->raw_del_out(j);
444 --j;
445 --max;
446 if (child->is_data_proj_of_pure_function(n)) {
447 worklist.push(n);
448 }
449 }
450 }
451 if (n->outcnt() == 1 && n->has_special_unique_user()) {
452 assert(useful.member(n->unique_out()), "do not push a useless node");
453 worklist.push(n->unique_out());
454 }
455 }
456
457 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
458 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
459 // Remove useless Template Assertion Predicate opaque nodes
460 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
461 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
462 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
463 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
464 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
465 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
466 #ifdef ASSERT
467 if (_modified_nodes != nullptr) {
468 _modified_nodes->remove_useless_nodes(useful.member_set());
469 }
470 #endif
471
472 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
473 bs->eliminate_useless_gc_barriers(useful, this);
474 // clean up the late inline lists
475 remove_useless_late_inlines( &_late_inlines, useful);
476 remove_useless_late_inlines( &_string_late_inlines, useful);
477 remove_useless_late_inlines( &_boxing_late_inlines, useful);
478 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
479 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
480 }
481
482 // ============================================================================
483 //------------------------------CompileWrapper---------------------------------
484 class CompileWrapper : public StackObj {
485 Compile *const _compile;
486 public:
487 CompileWrapper(Compile* compile);
488
489 ~CompileWrapper();
490 };
491
492 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
493 // the Compile* pointer is stored in the current ciEnv:
494 ciEnv* env = compile->env();
495 assert(env == ciEnv::current(), "must already be a ciEnv active");
496 assert(env->compiler_data() == nullptr, "compile already active?");
497 env->set_compiler_data(compile);
498 assert(compile == Compile::current(), "sanity");
499
500 compile->set_type_dict(nullptr);
501 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
502 compile->clone_map().set_clone_idx(0);
503 compile->set_type_last_size(0);
504 compile->set_last_tf(nullptr, nullptr);
505 compile->set_indexSet_arena(nullptr);
506 compile->set_indexSet_free_block_list(nullptr);
507 compile->init_type_arena();
508 Type::Initialize(compile);
509 _compile->begin_method();
510 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
511 }
512 CompileWrapper::~CompileWrapper() {
513 // simulate crash during compilation
514 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
515
516 _compile->end_method();
517 _compile->env()->set_compiler_data(nullptr);
518 }
519
520
521 //----------------------------print_compile_messages---------------------------
522 void Compile::print_compile_messages() {
523 #ifndef PRODUCT
524 // Check if recompiling
525 if (!subsume_loads() && PrintOpto) {
526 // Recompiling without allowing machine instructions to subsume loads
527 tty->print_cr("*********************************************************");
528 tty->print_cr("** Bailout: Recompile without subsuming loads **");
529 tty->print_cr("*********************************************************");
530 }
531 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
532 // Recompiling without escape analysis
533 tty->print_cr("*********************************************************");
534 tty->print_cr("** Bailout: Recompile without escape analysis **");
535 tty->print_cr("*********************************************************");
536 }
537 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
538 // Recompiling without iterative escape analysis
539 tty->print_cr("*********************************************************");
540 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
541 tty->print_cr("*********************************************************");
542 }
543 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
544 // Recompiling without reducing allocation merges
545 tty->print_cr("*********************************************************");
546 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
547 tty->print_cr("*********************************************************");
548 }
549 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
550 // Recompiling without boxing elimination
551 tty->print_cr("*********************************************************");
552 tty->print_cr("** Bailout: Recompile without boxing elimination **");
553 tty->print_cr("*********************************************************");
554 }
555 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
556 // Recompiling without locks coarsening
557 tty->print_cr("*********************************************************");
558 tty->print_cr("** Bailout: Recompile without locks coarsening **");
559 tty->print_cr("*********************************************************");
560 }
561 if (env()->break_at_compile()) {
562 // Open the debugger when compiling this method.
563 tty->print("### Breaking when compiling: ");
564 method()->print_short_name();
565 tty->cr();
566 BREAKPOINT;
567 }
568
569 if( PrintOpto ) {
570 if (is_osr_compilation()) {
571 tty->print("[OSR]%3d", _compile_id);
572 } else {
573 tty->print("%3d", _compile_id);
574 }
575 }
576 #endif
577 }
578
579 #ifndef PRODUCT
580 void Compile::print_phase(const char* phase_name) {
581 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name);
582 }
583
584 void Compile::print_ideal_ir(const char* phase_name) {
585 // keep the following output all in one block
586 // This output goes directly to the tty, not the compiler log.
587 // To enable tools to match it up with the compilation activity,
588 // be sure to tag this tty output with the compile ID.
589
590 // Node dumping can cause a safepoint, which can break the tty lock.
591 // Buffer all node dumps, so that all safepoints happen before we lock.
592 ResourceMark rm;
593 stringStream ss;
594
595 if (_output == nullptr) {
596 ss.print_cr("AFTER: %s", phase_name);
597 // Print out all nodes in ascending order of index.
598 // It is important that we traverse both inputs and outputs of nodes,
599 // so that we reach all nodes that are connected to Root.
600 root()->dump_bfs(MaxNodeLimit, nullptr, "-+S$", &ss);
601 } else {
602 // Dump the node blockwise if we have a scheduling
603 _output->print_scheduling(&ss);
604 }
605
606 // Check that the lock is not broken by a safepoint.
607 NoSafepointVerifier nsv;
608 ttyLocker ttyl;
609 if (xtty != nullptr) {
610 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
611 compile_id(),
612 is_osr_compilation() ? " compile_kind='osr'" : "",
613 phase_name);
614 }
615
616 tty->print("%s", ss.as_string());
617
618 if (xtty != nullptr) {
619 xtty->tail("ideal");
620 }
621 }
622 #endif
623
624 // ============================================================================
625 //------------------------------Compile standard-------------------------------
626
627 // Compile a method. entry_bci is -1 for normal compilations and indicates
628 // the continuation bci for on stack replacement.
629
630
631 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
632 Options options, DirectiveSet* directive)
633 : Phase(Compiler),
634 _compile_id(ci_env->compile_id()),
635 _options(options),
636 _method(target),
637 _entry_bci(osr_bci),
638 _ilt(nullptr),
639 _stub_function(nullptr),
640 _stub_name(nullptr),
641 _stub_id(StubId::NO_STUBID),
642 _stub_entry_point(nullptr),
643 _max_node_limit(MaxNodeLimit),
644 _post_loop_opts_phase(false),
645 _merge_stores_phase(false),
646 _allow_macro_nodes(true),
647 _inlining_progress(false),
648 _inlining_incrementally(false),
649 _do_cleanup(false),
650 _has_reserved_stack_access(target->has_reserved_stack_access()),
651 #ifndef PRODUCT
652 _igv_idx(0),
653 _trace_opto_output(directive->TraceOptoOutputOption),
654 #endif
655 _clinit_barrier_on_entry(false),
656 _stress_seed(0),
657 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
658 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
659 _env(ci_env),
660 _directive(directive),
661 _log(ci_env->log()),
662 _first_failure_details(nullptr),
663 _intrinsics(comp_arena(), 0, 0, nullptr),
664 _macro_nodes(comp_arena(), 8, 0, nullptr),
665 _parse_predicates(comp_arena(), 8, 0, nullptr),
666 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
667 _expensive_nodes(comp_arena(), 8, 0, nullptr),
668 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
669 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
670 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
671 _coarsened_locks(comp_arena(), 8, 0, nullptr),
672 _congraph(nullptr),
673 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
674 _unique(0),
675 _dead_node_count(0),
676 _dead_node_list(comp_arena()),
677 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
678 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
679 _node_arena(&_node_arena_one),
680 _mach_constant_base_node(nullptr),
681 _Compile_types(mtCompiler, Arena::Tag::tag_type),
682 _initial_gvn(nullptr),
683 _igvn_worklist(nullptr),
684 _types(nullptr),
685 _node_hash(nullptr),
686 _late_inlines(comp_arena(), 2, 0, nullptr),
687 _string_late_inlines(comp_arena(), 2, 0, nullptr),
688 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
689 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
690 _late_inlines_pos(0),
691 _has_mh_late_inlines(false),
692 _oom(false),
693 _replay_inline_data(nullptr),
694 _inline_printer(this),
695 _java_calls(0),
696 _inner_loops(0),
697 _FIRST_STACK_mask(comp_arena()),
698 _interpreter_frame_size(0),
699 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
700 _output(nullptr)
701 #ifndef PRODUCT
702 ,
703 _in_dump_cnt(0)
704 #endif
705 {
706 C = this;
707 CompileWrapper cw(this);
708
709 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
710 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
711
712 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
713 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
714 // We can always print a disassembly, either abstract (hex dump) or
715 // with the help of a suitable hsdis library. Thus, we should not
716 // couple print_assembly and print_opto_assembly controls.
717 // But: always print opto and regular assembly on compile command 'print'.
718 bool print_assembly = directive->PrintAssemblyOption;
719 set_print_assembly(print_opto_assembly || print_assembly);
720 #else
721 set_print_assembly(false); // must initialize.
722 #endif
723
724 #ifndef PRODUCT
725 set_parsed_irreducible_loop(false);
726 #endif
727
728 if (directive->ReplayInlineOption) {
729 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
730 }
731 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
732 set_print_intrinsics(directive->PrintIntrinsicsOption);
733 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
734
735 if (ProfileTraps) {
736 // Make sure the method being compiled gets its own MDO,
737 // so we can at least track the decompile_count().
738 method()->ensure_method_data();
739 }
740
741 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
742 StressIncrementalInlining || StressMacroExpansion ||
743 StressMacroElimination || StressUnstableIfTraps ||
744 StressBailout || StressLoopPeeling) {
745 initialize_stress_seed(directive);
746 }
747
748 Init(/*do_aliasing=*/ true);
749
750 print_compile_messages();
751
752 _ilt = InlineTree::build_inline_tree_root();
753
754 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
755 assert(num_alias_types() >= AliasIdxRaw, "");
756
757 #define MINIMUM_NODE_HASH 1023
758
759 // GVN that will be run immediately on new nodes
760 uint estimated_size = method()->code_size()*4+64;
761 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
762 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
763 _types = new (comp_arena()) Type_Array(comp_arena());
764 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
765 PhaseGVN gvn;
766 set_initial_gvn(&gvn);
767
768 { // Scope for timing the parser
769 TracePhase tp(_t_parser);
770
771 // Put top into the hash table ASAP.
772 initial_gvn()->transform(top());
773
774 // Set up tf(), start(), and find a CallGenerator.
775 CallGenerator* cg = nullptr;
776 if (is_osr_compilation()) {
777 const TypeTuple *domain = StartOSRNode::osr_domain();
778 const TypeTuple *range = TypeTuple::make_range(method()->signature());
779 init_tf(TypeFunc::make(domain, range));
780 StartNode* s = new StartOSRNode(root(), domain);
781 initial_gvn()->set_type_bottom(s);
782 verify_start(s);
783 cg = CallGenerator::for_osr(method(), entry_bci());
784 } else {
785 // Normal case.
786 init_tf(TypeFunc::make(method()));
787 StartNode* s = new StartNode(root(), tf()->domain());
788 initial_gvn()->set_type_bottom(s);
789 verify_start(s);
790 float past_uses = method()->interpreter_invocation_count();
791 float expected_uses = past_uses;
792 cg = CallGenerator::for_inline(method(), expected_uses);
793 }
794 if (failing()) return;
795 if (cg == nullptr) {
796 const char* reason = InlineTree::check_can_parse(method());
797 assert(reason != nullptr, "expect reason for parse failure");
798 stringStream ss;
799 ss.print("cannot parse method: %s", reason);
800 record_method_not_compilable(ss.as_string());
801 return;
802 }
803
804 gvn.set_type(root(), root()->bottom_type());
805
806 JVMState* jvms = build_start_state(start(), tf());
807 if ((jvms = cg->generate(jvms)) == nullptr) {
808 assert(failure_reason() != nullptr, "expect reason for parse failure");
809 stringStream ss;
810 ss.print("method parse failed: %s", failure_reason());
811 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
812 return;
813 }
814 GraphKit kit(jvms);
815
816 if (!kit.stopped()) {
817 // Accept return values, and transfer control we know not where.
818 // This is done by a special, unique ReturnNode bound to root.
819 return_values(kit.jvms());
820 }
821
822 if (kit.has_exceptions()) {
823 // Any exceptions that escape from this call must be rethrown
824 // to whatever caller is dynamically above us on the stack.
825 // This is done by a special, unique RethrowNode bound to root.
826 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
827 }
828
829 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
830
831 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
832 inline_string_calls(true);
833 }
834
835 if (failing()) return;
836
837 // Remove clutter produced by parsing.
838 if (!failing()) {
839 ResourceMark rm;
840 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
841 }
842 }
843
844 // Note: Large methods are capped off in do_one_bytecode().
845 if (failing()) return;
846
847 // After parsing, node notes are no longer automagic.
848 // They must be propagated by register_new_node_with_optimizer(),
849 // clone(), or the like.
850 set_default_node_notes(nullptr);
851
852 #ifndef PRODUCT
853 if (should_print_igv(1)) {
854 _igv_printer->print_inlining();
855 }
856 #endif
857
858 if (failing()) return;
859 NOT_PRODUCT( verify_graph_edges(); )
860
861 // Now optimize
862 Optimize();
863 if (failing()) return;
864 NOT_PRODUCT( verify_graph_edges(); )
865
866 #ifndef PRODUCT
867 if (should_print_ideal()) {
868 print_ideal_ir("print_ideal");
869 }
870 #endif
871
872 #ifdef ASSERT
873 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
874 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
875 #endif
876
877 // Dump compilation data to replay it.
878 if (directive->DumpReplayOption) {
879 env()->dump_replay_data(_compile_id);
880 }
881 if (directive->DumpInlineOption && (ilt() != nullptr)) {
882 env()->dump_inline_data(_compile_id);
883 }
884
885 // Now that we know the size of all the monitors we can add a fixed slot
886 // for the original deopt pc.
887 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
888 set_fixed_slots(next_slot);
889
890 // Compute when to use implicit null checks. Used by matching trap based
891 // nodes and NullCheck optimization.
892 set_allowed_deopt_reasons();
893
894 // Now generate code
895 Code_Gen();
896 }
897
898 //------------------------------Compile----------------------------------------
899 // Compile a runtime stub
900 Compile::Compile(ciEnv* ci_env,
901 TypeFunc_generator generator,
902 address stub_function,
903 const char* stub_name,
904 StubId stub_id,
905 int is_fancy_jump,
906 bool pass_tls,
907 bool return_pc,
908 DirectiveSet* directive)
909 : Phase(Compiler),
910 _compile_id(0),
911 _options(Options::for_runtime_stub()),
912 _method(nullptr),
913 _entry_bci(InvocationEntryBci),
914 _stub_function(stub_function),
915 _stub_name(stub_name),
916 _stub_id(stub_id),
917 _stub_entry_point(nullptr),
918 _max_node_limit(MaxNodeLimit),
919 _post_loop_opts_phase(false),
920 _merge_stores_phase(false),
921 _allow_macro_nodes(true),
922 _inlining_progress(false),
923 _inlining_incrementally(false),
924 _has_reserved_stack_access(false),
925 #ifndef PRODUCT
926 _igv_idx(0),
927 _trace_opto_output(directive->TraceOptoOutputOption),
928 #endif
929 _clinit_barrier_on_entry(false),
930 _stress_seed(0),
931 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
932 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
933 _env(ci_env),
934 _directive(directive),
935 _log(ci_env->log()),
936 _first_failure_details(nullptr),
937 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
938 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
939 _congraph(nullptr),
940 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
941 _unique(0),
942 _dead_node_count(0),
943 _dead_node_list(comp_arena()),
944 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
945 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
946 _node_arena(&_node_arena_one),
947 _mach_constant_base_node(nullptr),
948 _Compile_types(mtCompiler, Arena::Tag::tag_type),
949 _initial_gvn(nullptr),
950 _igvn_worklist(nullptr),
951 _types(nullptr),
952 _node_hash(nullptr),
953 _has_mh_late_inlines(false),
954 _oom(false),
955 _replay_inline_data(nullptr),
956 _inline_printer(this),
957 _java_calls(0),
958 _inner_loops(0),
959 _FIRST_STACK_mask(comp_arena()),
960 _interpreter_frame_size(0),
961 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
962 _output(nullptr),
963 #ifndef PRODUCT
964 _in_dump_cnt(0),
965 #endif
966 _allowed_reasons(0) {
967 C = this;
968
969 // try to reuse an existing stub
970 {
971 BlobId blob_id = StubInfo::blob(_stub_id);
972 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
973 if (blob != nullptr) {
974 RuntimeStub* rs = blob->as_runtime_stub();
975 _stub_entry_point = rs->entry_point();
976 return;
977 }
978 }
979
980 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
981 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
982
983 #ifndef PRODUCT
984 set_print_assembly(PrintFrameConverterAssembly);
985 set_parsed_irreducible_loop(false);
986 #else
987 set_print_assembly(false); // Must initialize.
988 #endif
989 set_has_irreducible_loop(false); // no loops
990
991 CompileWrapper cw(this);
992 Init(/*do_aliasing=*/ false);
993 init_tf((*generator)());
994
995 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
996 _types = new (comp_arena()) Type_Array(comp_arena());
997 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
998
999 if (StressLCM || StressGCM || StressBailout) {
1000 initialize_stress_seed(directive);
1001 }
1002
1003 {
1004 PhaseGVN gvn;
1005 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1006 gvn.transform(top());
1007
1008 GraphKit kit;
1009 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1010 }
1011
1012 NOT_PRODUCT( verify_graph_edges(); )
1013
1014 Code_Gen();
1015 }
1016
1017 Compile::~Compile() {
1018 delete _first_failure_details;
1019 };
1020
1021 //------------------------------Init-------------------------------------------
1022 // Prepare for a single compilation
1023 void Compile::Init(bool aliasing) {
1024 _do_aliasing = aliasing;
1025 _unique = 0;
1026 _regalloc = nullptr;
1027
1028 _tf = nullptr; // filled in later
1029 _top = nullptr; // cached later
1030 _matcher = nullptr; // filled in later
1031 _cfg = nullptr; // filled in later
1032
1033 _node_note_array = nullptr;
1034 _default_node_notes = nullptr;
1035 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1036
1037 _immutable_memory = nullptr; // filled in at first inquiry
1038
1039 #ifdef ASSERT
1040 _phase_optimize_finished = false;
1041 _phase_verify_ideal_loop = false;
1042 _exception_backedge = false;
1043 _type_verify = nullptr;
1044 #endif
1045
1046 // Globally visible Nodes
1047 // First set TOP to null to give safe behavior during creation of RootNode
1048 set_cached_top_node(nullptr);
1049 set_root(new RootNode());
1050 // Now that you have a Root to point to, create the real TOP
1051 set_cached_top_node( new ConNode(Type::TOP) );
1052 set_recent_alloc(nullptr, nullptr);
1053
1054 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1055 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1056 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1057 env()->set_dependencies(new Dependencies(env()));
1058
1059 _fixed_slots = 0;
1060 set_has_split_ifs(false);
1061 set_has_loops(false); // first approximation
1062 set_has_stringbuilder(false);
1063 set_has_boxed_value(false);
1064 _trap_can_recompile = false; // no traps emitted yet
1065 _major_progress = true; // start out assuming good things will happen
1066 set_has_unsafe_access(false);
1067 set_max_vector_size(0);
1068 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1069 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1070 set_decompile_count(0);
1071
1072 #ifndef PRODUCT
1073 _phase_counter = 0;
1074 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1075 #endif
1076
1077 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1078 _loop_opts_cnt = LoopOptsCount;
1079 set_do_inlining(Inline);
1080 set_max_inline_size(MaxInlineSize);
1081 set_freq_inline_size(FreqInlineSize);
1082 set_do_scheduling(OptoScheduling);
1083
1084 set_do_vector_loop(false);
1085 set_has_monitors(false);
1086 set_has_scoped_access(false);
1087
1088 if (AllowVectorizeOnDemand) {
1089 if (has_method() && _directive->VectorizeOption) {
1090 set_do_vector_loop(true);
1091 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());})
1092 } else if (has_method() && method()->name() != nullptr &&
1093 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1094 set_do_vector_loop(true);
1095 }
1096 }
1097 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1098 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());})
1099
1100 _max_node_limit = _directive->MaxNodeLimitOption;
1101
1102 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1103 set_clinit_barrier_on_entry(true);
1104 }
1105 if (debug_info()->recording_non_safepoints()) {
1106 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1107 (comp_arena(), 8, 0, nullptr));
1108 set_default_node_notes(Node_Notes::make(this));
1109 }
1110
1111 const int grow_ats = 16;
1112 _max_alias_types = grow_ats;
1113 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1114 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1115 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1116 {
1117 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1118 }
1119 // Initialize the first few types.
1120 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1121 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1122 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1123 _num_alias_types = AliasIdxRaw+1;
1124 // Zero out the alias type cache.
1125 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1126 // A null adr_type hits in the cache right away. Preload the right answer.
1127 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1128 }
1129
1130 #ifdef ASSERT
1131 // Verify that the current StartNode is valid.
1132 void Compile::verify_start(StartNode* s) const {
1133 assert(failing_internal() || s == start(), "should be StartNode");
1134 }
1135 #endif
1136
1137 /**
1138 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1139 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1140 * the ideal graph.
1141 */
1142 StartNode* Compile::start() const {
1143 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1144 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1145 Node* start = root()->fast_out(i);
1146 if (start->is_Start()) {
1147 return start->as_Start();
1148 }
1149 }
1150 fatal("Did not find Start node!");
1151 return nullptr;
1152 }
1153
1154 //-------------------------------immutable_memory-------------------------------------
1155 // Access immutable memory
1156 Node* Compile::immutable_memory() {
1157 if (_immutable_memory != nullptr) {
1158 return _immutable_memory;
1159 }
1160 StartNode* s = start();
1161 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1162 Node *p = s->fast_out(i);
1163 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1164 _immutable_memory = p;
1165 return _immutable_memory;
1166 }
1167 }
1168 ShouldNotReachHere();
1169 return nullptr;
1170 }
1171
1172 //----------------------set_cached_top_node------------------------------------
1173 // Install the cached top node, and make sure Node::is_top works correctly.
1174 void Compile::set_cached_top_node(Node* tn) {
1175 if (tn != nullptr) verify_top(tn);
1176 Node* old_top = _top;
1177 _top = tn;
1178 // Calling Node::setup_is_top allows the nodes the chance to adjust
1179 // their _out arrays.
1180 if (_top != nullptr) _top->setup_is_top();
1181 if (old_top != nullptr) old_top->setup_is_top();
1182 assert(_top == nullptr || top()->is_top(), "");
1183 }
1184
1185 #ifdef ASSERT
1186 uint Compile::count_live_nodes_by_graph_walk() {
1187 Unique_Node_List useful(comp_arena());
1188 // Get useful node list by walking the graph.
1189 identify_useful_nodes(useful);
1190 return useful.size();
1191 }
1192
1193 void Compile::print_missing_nodes() {
1194
1195 // Return if CompileLog is null and PrintIdealNodeCount is false.
1196 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1197 return;
1198 }
1199
1200 // This is an expensive function. It is executed only when the user
1201 // specifies VerifyIdealNodeCount option or otherwise knows the
1202 // additional work that needs to be done to identify reachable nodes
1203 // by walking the flow graph and find the missing ones using
1204 // _dead_node_list.
1205
1206 Unique_Node_List useful(comp_arena());
1207 // Get useful node list by walking the graph.
1208 identify_useful_nodes(useful);
1209
1210 uint l_nodes = C->live_nodes();
1211 uint l_nodes_by_walk = useful.size();
1212
1213 if (l_nodes != l_nodes_by_walk) {
1214 if (_log != nullptr) {
1215 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1216 _log->stamp();
1217 _log->end_head();
1218 }
1219 VectorSet& useful_member_set = useful.member_set();
1220 int last_idx = l_nodes_by_walk;
1221 for (int i = 0; i < last_idx; i++) {
1222 if (useful_member_set.test(i)) {
1223 if (_dead_node_list.test(i)) {
1224 if (_log != nullptr) {
1225 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1226 }
1227 if (PrintIdealNodeCount) {
1228 // Print the log message to tty
1229 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1230 useful.at(i)->dump();
1231 }
1232 }
1233 }
1234 else if (! _dead_node_list.test(i)) {
1235 if (_log != nullptr) {
1236 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1237 }
1238 if (PrintIdealNodeCount) {
1239 // Print the log message to tty
1240 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1241 }
1242 }
1243 }
1244 if (_log != nullptr) {
1245 _log->tail("mismatched_nodes");
1246 }
1247 }
1248 }
1249 void Compile::record_modified_node(Node* n) {
1250 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1251 _modified_nodes->push(n);
1252 }
1253 }
1254
1255 void Compile::remove_modified_node(Node* n) {
1256 if (_modified_nodes != nullptr) {
1257 _modified_nodes->remove(n);
1258 }
1259 }
1260 #endif
1261
1262 #ifndef PRODUCT
1263 void Compile::verify_top(Node* tn) const {
1264 if (tn != nullptr) {
1265 assert(tn->is_Con(), "top node must be a constant");
1266 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1267 assert(tn->in(0) != nullptr, "must have live top node");
1268 }
1269 }
1270 #endif
1271
1272
1273 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1274
1275 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1276 guarantee(arr != nullptr, "");
1277 int num_blocks = arr->length();
1278 if (grow_by < num_blocks) grow_by = num_blocks;
1279 int num_notes = grow_by * _node_notes_block_size;
1280 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1281 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1282 while (num_notes > 0) {
1283 arr->append(notes);
1284 notes += _node_notes_block_size;
1285 num_notes -= _node_notes_block_size;
1286 }
1287 assert(num_notes == 0, "exact multiple, please");
1288 }
1289
1290 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1291 if (source == nullptr || dest == nullptr) return false;
1292
1293 if (dest->is_Con())
1294 return false; // Do not push debug info onto constants.
1295
1296 #ifdef ASSERT
1297 // Leave a bread crumb trail pointing to the original node:
1298 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1299 dest->set_debug_orig(source);
1300 }
1301 #endif
1302
1303 if (node_note_array() == nullptr)
1304 return false; // Not collecting any notes now.
1305
1306 // This is a copy onto a pre-existing node, which may already have notes.
1307 // If both nodes have notes, do not overwrite any pre-existing notes.
1308 Node_Notes* source_notes = node_notes_at(source->_idx);
1309 if (source_notes == nullptr || source_notes->is_clear()) return false;
1310 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1311 if (dest_notes == nullptr || dest_notes->is_clear()) {
1312 return set_node_notes_at(dest->_idx, source_notes);
1313 }
1314
1315 Node_Notes merged_notes = (*source_notes);
1316 // The order of operations here ensures that dest notes will win...
1317 merged_notes.update_from(dest_notes);
1318 return set_node_notes_at(dest->_idx, &merged_notes);
1319 }
1320
1321
1322 //--------------------------allow_range_check_smearing-------------------------
1323 // Gating condition for coalescing similar range checks.
1324 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1325 // single covering check that is at least as strong as any of them.
1326 // If the optimization succeeds, the simplified (strengthened) range check
1327 // will always succeed. If it fails, we will deopt, and then give up
1328 // on the optimization.
1329 bool Compile::allow_range_check_smearing() const {
1330 // If this method has already thrown a range-check,
1331 // assume it was because we already tried range smearing
1332 // and it failed.
1333 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1334 return !already_trapped;
1335 }
1336
1337
1338 //------------------------------flatten_alias_type-----------------------------
1339 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1340 assert(do_aliasing(), "Aliasing should be enabled");
1341 int offset = tj->offset();
1342 TypePtr::PTR ptr = tj->ptr();
1343
1344 // Known instance (scalarizable allocation) alias only with itself.
1345 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1346 tj->is_oopptr()->is_known_instance();
1347
1348 // Process weird unsafe references.
1349 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1350 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1351 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1352 tj = TypeOopPtr::BOTTOM;
1353 ptr = tj->ptr();
1354 offset = tj->offset();
1355 }
1356
1357 // Array pointers need some flattening
1358 const TypeAryPtr* ta = tj->isa_aryptr();
1359 if (ta && ta->is_stable()) {
1360 // Erase stability property for alias analysis.
1361 tj = ta = ta->cast_to_stable(false);
1362 }
1363 if( ta && is_known_inst ) {
1364 if ( offset != Type::OffsetBot &&
1365 offset > arrayOopDesc::length_offset_in_bytes() ) {
1366 offset = Type::OffsetBot; // Flatten constant access into array body only
1367 tj = ta = ta->
1368 remove_speculative()->
1369 cast_to_ptr_type(ptr)->
1370 with_offset(offset);
1371 }
1372 } else if (ta) {
1373 // For arrays indexed by constant indices, we flatten the alias
1374 // space to include all of the array body. Only the header, klass
1375 // and array length can be accessed un-aliased.
1376 if( offset != Type::OffsetBot ) {
1377 if( ta->const_oop() ) { // MethodData* or Method*
1378 offset = Type::OffsetBot; // Flatten constant access into array body
1379 tj = ta = ta->
1380 remove_speculative()->
1381 cast_to_ptr_type(ptr)->
1382 cast_to_exactness(false)->
1383 with_offset(offset);
1384 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1385 // range is OK as-is.
1386 tj = ta = TypeAryPtr::RANGE;
1387 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1388 tj = TypeInstPtr::KLASS; // all klass loads look alike
1389 ta = TypeAryPtr::RANGE; // generic ignored junk
1390 ptr = TypePtr::BotPTR;
1391 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1392 tj = TypeInstPtr::MARK;
1393 ta = TypeAryPtr::RANGE; // generic ignored junk
1394 ptr = TypePtr::BotPTR;
1395 } else { // Random constant offset into array body
1396 offset = Type::OffsetBot; // Flatten constant access into array body
1397 tj = ta = ta->
1398 remove_speculative()->
1399 cast_to_ptr_type(ptr)->
1400 cast_to_exactness(false)->
1401 with_offset(offset);
1402 }
1403 }
1404 // Arrays of fixed size alias with arrays of unknown size.
1405 if (ta->size() != TypeInt::POS) {
1406 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1407 tj = ta = ta->
1408 remove_speculative()->
1409 cast_to_ptr_type(ptr)->
1410 with_ary(tary)->
1411 cast_to_exactness(false);
1412 }
1413 // Arrays of known objects become arrays of unknown objects.
1414 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1415 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1416 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1417 }
1418 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1419 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1420 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1421 }
1422 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1423 // cannot be distinguished by bytecode alone.
1424 if (ta->elem() == TypeInt::BOOL) {
1425 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1426 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1427 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1428 }
1429 // During the 2nd round of IterGVN, NotNull castings are removed.
1430 // Make sure the Bottom and NotNull variants alias the same.
1431 // Also, make sure exact and non-exact variants alias the same.
1432 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1433 tj = ta = ta->
1434 remove_speculative()->
1435 cast_to_ptr_type(TypePtr::BotPTR)->
1436 cast_to_exactness(false)->
1437 with_offset(offset);
1438 }
1439 }
1440
1441 // Oop pointers need some flattening
1442 const TypeInstPtr *to = tj->isa_instptr();
1443 if (to && to != TypeOopPtr::BOTTOM) {
1444 ciInstanceKlass* ik = to->instance_klass();
1445 if( ptr == TypePtr::Constant ) {
1446 if (ik != ciEnv::current()->Class_klass() ||
1447 offset < ik->layout_helper_size_in_bytes()) {
1448 // No constant oop pointers (such as Strings); they alias with
1449 // unknown strings.
1450 assert(!is_known_inst, "not scalarizable allocation");
1451 tj = to = to->
1452 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1453 remove_speculative()->
1454 cast_to_ptr_type(TypePtr::BotPTR)->
1455 cast_to_exactness(false);
1456 }
1457 } else if( is_known_inst ) {
1458 tj = to; // Keep NotNull and klass_is_exact for instance type
1459 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1460 // During the 2nd round of IterGVN, NotNull castings are removed.
1461 // Make sure the Bottom and NotNull variants alias the same.
1462 // Also, make sure exact and non-exact variants alias the same.
1463 tj = to = to->
1464 remove_speculative()->
1465 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1466 cast_to_ptr_type(TypePtr::BotPTR)->
1467 cast_to_exactness(false);
1468 }
1469 if (to->speculative() != nullptr) {
1470 tj = to = to->remove_speculative();
1471 }
1472 // Canonicalize the holder of this field
1473 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1474 // First handle header references such as a LoadKlassNode, even if the
1475 // object's klass is unloaded at compile time (4965979).
1476 if (!is_known_inst) { // Do it only for non-instance types
1477 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1478 }
1479 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1480 // Static fields are in the space above the normal instance
1481 // fields in the java.lang.Class instance.
1482 if (ik != ciEnv::current()->Class_klass()) {
1483 to = nullptr;
1484 tj = TypeOopPtr::BOTTOM;
1485 offset = tj->offset();
1486 }
1487 } else {
1488 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1489 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1490 assert(tj->offset() == offset, "no change to offset expected");
1491 bool xk = to->klass_is_exact();
1492 int instance_id = to->instance_id();
1493
1494 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1495 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1496 // its interfaces are included.
1497 if (xk && ik->equals(canonical_holder)) {
1498 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1499 } else {
1500 assert(xk || !is_known_inst, "Known instance should be exact type");
1501 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1502 }
1503 }
1504 }
1505
1506 // Klass pointers to object array klasses need some flattening
1507 const TypeKlassPtr *tk = tj->isa_klassptr();
1508 if( tk ) {
1509 // If we are referencing a field within a Klass, we need
1510 // to assume the worst case of an Object. Both exact and
1511 // inexact types must flatten to the same alias class so
1512 // use NotNull as the PTR.
1513 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1514 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1515 env()->Object_klass(),
1516 offset);
1517 }
1518
1519 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1520 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1521 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1522 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1523 } else {
1524 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1525 }
1526 }
1527
1528 // Check for precise loads from the primary supertype array and force them
1529 // to the supertype cache alias index. Check for generic array loads from
1530 // the primary supertype array and also force them to the supertype cache
1531 // alias index. Since the same load can reach both, we need to merge
1532 // these 2 disparate memories into the same alias class. Since the
1533 // primary supertype array is read-only, there's no chance of confusion
1534 // where we bypass an array load and an array store.
1535 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1536 if (offset == Type::OffsetBot ||
1537 (offset >= primary_supers_offset &&
1538 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1539 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1540 offset = in_bytes(Klass::secondary_super_cache_offset());
1541 tj = tk = tk->with_offset(offset);
1542 }
1543 }
1544
1545 // Flatten all Raw pointers together.
1546 if (tj->base() == Type::RawPtr)
1547 tj = TypeRawPtr::BOTTOM;
1548
1549 if (tj->base() == Type::AnyPtr)
1550 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1551
1552 offset = tj->offset();
1553 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1554
1555 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1556 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1557 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1558 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1559 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1560 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1561 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1562 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1563 assert( tj->ptr() != TypePtr::TopPTR &&
1564 tj->ptr() != TypePtr::AnyNull &&
1565 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1566 // assert( tj->ptr() != TypePtr::Constant ||
1567 // tj->base() == Type::RawPtr ||
1568 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1569
1570 return tj;
1571 }
1572
1573 void Compile::AliasType::Init(int i, const TypePtr* at) {
1574 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1575 _index = i;
1576 _adr_type = at;
1577 _field = nullptr;
1578 _element = nullptr;
1579 _is_rewritable = true; // default
1580 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1581 if (atoop != nullptr && atoop->is_known_instance()) {
1582 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1583 _general_index = Compile::current()->get_alias_index(gt);
1584 } else {
1585 _general_index = 0;
1586 }
1587 }
1588
1589 BasicType Compile::AliasType::basic_type() const {
1590 if (element() != nullptr) {
1591 const Type* element = adr_type()->is_aryptr()->elem();
1592 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1593 } if (field() != nullptr) {
1594 return field()->layout_type();
1595 } else {
1596 return T_ILLEGAL; // unknown
1597 }
1598 }
1599
1600 //---------------------------------print_on------------------------------------
1601 #ifndef PRODUCT
1602 void Compile::AliasType::print_on(outputStream* st) {
1603 if (index() < 10)
1604 st->print("@ <%d> ", index());
1605 else st->print("@ <%d>", index());
1606 st->print(is_rewritable() ? " " : " RO");
1607 int offset = adr_type()->offset();
1608 if (offset == Type::OffsetBot)
1609 st->print(" +any");
1610 else st->print(" +%-3d", offset);
1611 st->print(" in ");
1612 adr_type()->dump_on(st);
1613 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1614 if (field() != nullptr && tjp) {
1615 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1616 tjp->offset() != field()->offset_in_bytes()) {
1617 st->print(" != ");
1618 field()->print();
1619 st->print(" ***");
1620 }
1621 }
1622 }
1623
1624 void print_alias_types() {
1625 Compile* C = Compile::current();
1626 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1627 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1628 C->alias_type(idx)->print_on(tty);
1629 tty->cr();
1630 }
1631 }
1632 #endif
1633
1634
1635 //----------------------------probe_alias_cache--------------------------------
1636 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1637 intptr_t key = (intptr_t) adr_type;
1638 key ^= key >> logAliasCacheSize;
1639 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1640 }
1641
1642
1643 //-----------------------------grow_alias_types--------------------------------
1644 void Compile::grow_alias_types() {
1645 const int old_ats = _max_alias_types; // how many before?
1646 const int new_ats = old_ats; // how many more?
1647 const int grow_ats = old_ats+new_ats; // how many now?
1648 _max_alias_types = grow_ats;
1649 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1650 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1651 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1652 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1653 }
1654
1655
1656 //--------------------------------find_alias_type------------------------------
1657 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1658 if (!do_aliasing()) {
1659 return alias_type(AliasIdxBot);
1660 }
1661
1662 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1663 if (ace->_adr_type == adr_type) {
1664 return alias_type(ace->_index);
1665 }
1666
1667 // Handle special cases.
1668 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1669 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1670
1671 // Do it the slow way.
1672 const TypePtr* flat = flatten_alias_type(adr_type);
1673
1674 #ifdef ASSERT
1675 {
1676 ResourceMark rm;
1677 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1678 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1679 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1680 Type::str(adr_type));
1681 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1682 const TypeOopPtr* foop = flat->is_oopptr();
1683 // Scalarizable allocations have exact klass always.
1684 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1685 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1686 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1687 Type::str(foop), Type::str(xoop));
1688 }
1689 }
1690 #endif
1691
1692 int idx = AliasIdxTop;
1693 for (int i = 0; i < num_alias_types(); i++) {
1694 if (alias_type(i)->adr_type() == flat) {
1695 idx = i;
1696 break;
1697 }
1698 }
1699
1700 if (idx == AliasIdxTop) {
1701 if (no_create) return nullptr;
1702 // Grow the array if necessary.
1703 if (_num_alias_types == _max_alias_types) grow_alias_types();
1704 // Add a new alias type.
1705 idx = _num_alias_types++;
1706 _alias_types[idx]->Init(idx, flat);
1707 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1708 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1709 if (flat->isa_instptr()) {
1710 if (flat->offset() == java_lang_Class::klass_offset()
1711 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1712 alias_type(idx)->set_rewritable(false);
1713 }
1714 if (flat->isa_aryptr()) {
1715 #ifdef ASSERT
1716 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1717 // (T_BYTE has the weakest alignment and size restrictions...)
1718 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1719 #endif
1720 if (flat->offset() == TypePtr::OffsetBot) {
1721 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1722 }
1723 }
1724 if (flat->isa_klassptr()) {
1725 if (UseCompactObjectHeaders) {
1726 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1727 alias_type(idx)->set_rewritable(false);
1728 }
1729 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1730 alias_type(idx)->set_rewritable(false);
1731 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1732 alias_type(idx)->set_rewritable(false);
1733 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1734 alias_type(idx)->set_rewritable(false);
1735 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1736 alias_type(idx)->set_rewritable(false);
1737 }
1738
1739 if (flat->isa_instklassptr()) {
1740 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1741 alias_type(idx)->set_rewritable(false);
1742 }
1743 }
1744 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1745 // but the base pointer type is not distinctive enough to identify
1746 // references into JavaThread.)
1747
1748 // Check for final fields.
1749 const TypeInstPtr* tinst = flat->isa_instptr();
1750 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1751 ciField* field;
1752 if (tinst->const_oop() != nullptr &&
1753 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1754 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1755 // static field
1756 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1757 field = k->get_field_by_offset(tinst->offset(), true);
1758 } else {
1759 ciInstanceKlass *k = tinst->instance_klass();
1760 field = k->get_field_by_offset(tinst->offset(), false);
1761 }
1762 assert(field == nullptr ||
1763 original_field == nullptr ||
1764 (field->holder() == original_field->holder() &&
1765 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1766 field->is_static() == original_field->is_static()), "wrong field?");
1767 // Set field() and is_rewritable() attributes.
1768 if (field != nullptr) alias_type(idx)->set_field(field);
1769 }
1770 }
1771
1772 // Fill the cache for next time.
1773 ace->_adr_type = adr_type;
1774 ace->_index = idx;
1775 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1776
1777 // Might as well try to fill the cache for the flattened version, too.
1778 AliasCacheEntry* face = probe_alias_cache(flat);
1779 if (face->_adr_type == nullptr) {
1780 face->_adr_type = flat;
1781 face->_index = idx;
1782 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1783 }
1784
1785 return alias_type(idx);
1786 }
1787
1788
1789 Compile::AliasType* Compile::alias_type(ciField* field) {
1790 const TypeOopPtr* t;
1791 if (field->is_static())
1792 t = TypeInstPtr::make(field->holder()->java_mirror());
1793 else
1794 t = TypeOopPtr::make_from_klass_raw(field->holder());
1795 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1796 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1797 return atp;
1798 }
1799
1800
1801 //------------------------------have_alias_type--------------------------------
1802 bool Compile::have_alias_type(const TypePtr* adr_type) {
1803 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1804 if (ace->_adr_type == adr_type) {
1805 return true;
1806 }
1807
1808 // Handle special cases.
1809 if (adr_type == nullptr) return true;
1810 if (adr_type == TypePtr::BOTTOM) return true;
1811
1812 return find_alias_type(adr_type, true, nullptr) != nullptr;
1813 }
1814
1815 //-----------------------------must_alias--------------------------------------
1816 // True if all values of the given address type are in the given alias category.
1817 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1818 if (alias_idx == AliasIdxBot) return true; // the universal category
1819 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1820 if (alias_idx == AliasIdxTop) return false; // the empty category
1821 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1822
1823 // the only remaining possible overlap is identity
1824 int adr_idx = get_alias_index(adr_type);
1825 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1826 assert(adr_idx == alias_idx ||
1827 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1828 && adr_type != TypeOopPtr::BOTTOM),
1829 "should not be testing for overlap with an unsafe pointer");
1830 return adr_idx == alias_idx;
1831 }
1832
1833 //------------------------------can_alias--------------------------------------
1834 // True if any values of the given address type are in the given alias category.
1835 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1836 if (alias_idx == AliasIdxTop) return false; // the empty category
1837 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1838 // Known instance doesn't alias with bottom memory
1839 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1840 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1841
1842 // the only remaining possible overlap is identity
1843 int adr_idx = get_alias_index(adr_type);
1844 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1845 return adr_idx == alias_idx;
1846 }
1847
1848 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1849 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1850 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1851 if (parse_predicate_count() == 0) {
1852 return;
1853 }
1854 for (int i = 0; i < parse_predicate_count(); i++) {
1855 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1856 parse_predicate->mark_useless(igvn);
1857 }
1858 _parse_predicates.clear();
1859 }
1860
1861 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1862 if (!n->for_post_loop_opts_igvn()) {
1863 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1864 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1865 _for_post_loop_igvn.append(n);
1866 }
1867 }
1868
1869 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1870 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1871 _for_post_loop_igvn.remove(n);
1872 }
1873
1874 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1875 // Verify that all previous optimizations produced a valid graph
1876 // at least to this point, even if no loop optimizations were done.
1877 PhaseIdealLoop::verify(igvn);
1878
1879 if (has_loops() || _loop_opts_cnt > 0) {
1880 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1881 }
1882 C->set_post_loop_opts_phase(); // no more loop opts allowed
1883
1884 assert(!C->major_progress(), "not cleared");
1885
1886 if (_for_post_loop_igvn.length() > 0) {
1887 while (_for_post_loop_igvn.length() > 0) {
1888 Node* n = _for_post_loop_igvn.pop();
1889 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1890 igvn._worklist.push(n);
1891 }
1892 igvn.optimize();
1893 if (failing()) return;
1894 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1895 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1896
1897 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1898 if (C->major_progress()) {
1899 C->clear_major_progress(); // ensure that major progress is now clear
1900 }
1901 }
1902 }
1903
1904 void Compile::record_for_merge_stores_igvn(Node* n) {
1905 if (!n->for_merge_stores_igvn()) {
1906 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1907 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1908 _for_merge_stores_igvn.append(n);
1909 }
1910 }
1911
1912 void Compile::remove_from_merge_stores_igvn(Node* n) {
1913 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1914 _for_merge_stores_igvn.remove(n);
1915 }
1916
1917 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1918 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1919 // the stores, and we merge the wrong sequence of stores.
1920 // Example:
1921 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1922 // Apply MergeStores:
1923 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1924 // Remove more RangeChecks:
1925 // StoreI [ StoreL ] StoreI
1926 // But now it would have been better to do this instead:
1927 // [ StoreL ] [ StoreL ]
1928 //
1929 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1930 // since we never unset _merge_stores_phase.
1931 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1932 C->set_merge_stores_phase();
1933
1934 if (_for_merge_stores_igvn.length() > 0) {
1935 while (_for_merge_stores_igvn.length() > 0) {
1936 Node* n = _for_merge_stores_igvn.pop();
1937 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1938 igvn._worklist.push(n);
1939 }
1940 igvn.optimize();
1941 if (failing()) return;
1942 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1943 print_method(PHASE_AFTER_MERGE_STORES, 3);
1944 }
1945 }
1946
1947 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1948 if (OptimizeUnstableIf) {
1949 _unstable_if_traps.append(trap);
1950 }
1951 }
1952
1953 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1954 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1955 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1956 Node* n = trap->uncommon_trap();
1957 if (!useful.member(n)) {
1958 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1959 }
1960 }
1961 }
1962
1963 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1964 // or fold-compares case. Return true if succeed or not found.
1965 //
1966 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1967 // false and ask fold-compares to yield.
1968 //
1969 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1970 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1971 // when deoptimization does happen.
1972 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1973 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1974 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1975 if (trap->uncommon_trap() == unc) {
1976 if (yield && trap->modified()) {
1977 return false;
1978 }
1979 _unstable_if_traps.delete_at(i);
1980 break;
1981 }
1982 }
1983 return true;
1984 }
1985
1986 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1987 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1988 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1989 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1990 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1991 CallStaticJavaNode* unc = trap->uncommon_trap();
1992 int next_bci = trap->next_bci();
1993 bool modified = trap->modified();
1994
1995 if (next_bci != -1 && !modified) {
1996 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1997 JVMState* jvms = unc->jvms();
1998 ciMethod* method = jvms->method();
1999 ciBytecodeStream iter(method);
2000
2001 iter.force_bci(jvms->bci());
2002 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2003 Bytecodes::Code c = iter.cur_bc();
2004 Node* lhs = nullptr;
2005 Node* rhs = nullptr;
2006 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2007 lhs = unc->peek_operand(0);
2008 rhs = unc->peek_operand(1);
2009 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2010 lhs = unc->peek_operand(0);
2011 }
2012
2013 ResourceMark rm;
2014 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2015 assert(live_locals.is_valid(), "broken liveness info");
2016 int len = (int)live_locals.size();
2017
2018 for (int i = 0; i < len; i++) {
2019 Node* local = unc->local(jvms, i);
2020 // kill local using the liveness of next_bci.
2021 // give up when the local looks like an operand to secure reexecution.
2022 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2023 uint idx = jvms->locoff() + i;
2024 #ifdef ASSERT
2025 if (PrintOpto && Verbose) {
2026 tty->print("[unstable_if] kill local#%d: ", idx);
2027 local->dump();
2028 tty->cr();
2029 }
2030 #endif
2031 igvn.replace_input_of(unc, idx, top());
2032 modified = true;
2033 }
2034 }
2035 }
2036
2037 // keep the mondified trap for late query
2038 if (modified) {
2039 trap->set_modified();
2040 } else {
2041 _unstable_if_traps.delete_at(i);
2042 }
2043 }
2044 igvn.optimize();
2045 }
2046
2047 // StringOpts and late inlining of string methods
2048 void Compile::inline_string_calls(bool parse_time) {
2049 {
2050 // remove useless nodes to make the usage analysis simpler
2051 ResourceMark rm;
2052 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2053 }
2054
2055 {
2056 ResourceMark rm;
2057 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2058 PhaseStringOpts pso(initial_gvn());
2059 print_method(PHASE_AFTER_STRINGOPTS, 3);
2060 }
2061
2062 // now inline anything that we skipped the first time around
2063 if (!parse_time) {
2064 _late_inlines_pos = _late_inlines.length();
2065 }
2066
2067 while (_string_late_inlines.length() > 0) {
2068 CallGenerator* cg = _string_late_inlines.pop();
2069 cg->do_late_inline();
2070 if (failing()) return;
2071 }
2072 _string_late_inlines.trunc_to(0);
2073 }
2074
2075 // Late inlining of boxing methods
2076 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2077 if (_boxing_late_inlines.length() > 0) {
2078 assert(has_boxed_value(), "inconsistent");
2079
2080 set_inlining_incrementally(true);
2081
2082 igvn_worklist()->ensure_empty(); // should be done with igvn
2083
2084 _late_inlines_pos = _late_inlines.length();
2085
2086 while (_boxing_late_inlines.length() > 0) {
2087 CallGenerator* cg = _boxing_late_inlines.pop();
2088 cg->do_late_inline();
2089 if (failing()) return;
2090 }
2091 _boxing_late_inlines.trunc_to(0);
2092
2093 inline_incrementally_cleanup(igvn);
2094
2095 set_inlining_incrementally(false);
2096 }
2097 }
2098
2099 bool Compile::inline_incrementally_one() {
2100 assert(IncrementalInline, "incremental inlining should be on");
2101
2102 TracePhase tp(_t_incrInline_inline);
2103
2104 set_inlining_progress(false);
2105 set_do_cleanup(false);
2106
2107 for (int i = 0; i < _late_inlines.length(); i++) {
2108 _late_inlines_pos = i+1;
2109 CallGenerator* cg = _late_inlines.at(i);
2110 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2111 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2112 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2113 if (should_stress_inlining()) {
2114 // randomly add repeated inline attempt if stress-inlining
2115 cg->call_node()->set_generator(cg);
2116 C->igvn_worklist()->push(cg->call_node());
2117 continue;
2118 }
2119 cg->do_late_inline();
2120 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2121 if (failing()) {
2122 return false;
2123 } else if (inlining_progress()) {
2124 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2125 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2126 break; // process one call site at a time
2127 } else {
2128 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2129 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2130 // Avoid potential infinite loop if node already in the IGVN list
2131 assert(false, "scheduled for IGVN during inlining attempt");
2132 } else {
2133 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2134 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2135 cg->call_node()->set_generator(cg);
2136 }
2137 }
2138 } else {
2139 // Ignore late inline direct calls when inlining is not allowed.
2140 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2141 }
2142 }
2143 // Remove processed elements.
2144 _late_inlines.remove_till(_late_inlines_pos);
2145 _late_inlines_pos = 0;
2146
2147 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2148
2149 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2150
2151 set_inlining_progress(false);
2152 set_do_cleanup(false);
2153
2154 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2155 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2156 }
2157
2158 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2159 {
2160 TracePhase tp(_t_incrInline_pru);
2161 ResourceMark rm;
2162 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2163 }
2164 {
2165 TracePhase tp(_t_incrInline_igvn);
2166 igvn.reset();
2167 igvn.optimize();
2168 if (failing()) return;
2169 }
2170 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2171 }
2172
2173 // Perform incremental inlining until bound on number of live nodes is reached
2174 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2175 TracePhase tp(_t_incrInline);
2176
2177 set_inlining_incrementally(true);
2178 uint low_live_nodes = 0;
2179
2180 while (_late_inlines.length() > 0) {
2181 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2182 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2183 TracePhase tp(_t_incrInline_ideal);
2184 // PhaseIdealLoop is expensive so we only try it once we are
2185 // out of live nodes and we only try it again if the previous
2186 // helped got the number of nodes down significantly
2187 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2188 if (failing()) return;
2189 low_live_nodes = live_nodes();
2190 _major_progress = true;
2191 }
2192
2193 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2194 bool do_print_inlining = print_inlining() || print_intrinsics();
2195 if (do_print_inlining || log() != nullptr) {
2196 // Print inlining message for candidates that we couldn't inline for lack of space.
2197 for (int i = 0; i < _late_inlines.length(); i++) {
2198 CallGenerator* cg = _late_inlines.at(i);
2199 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2200 if (do_print_inlining) {
2201 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2202 }
2203 log_late_inline_failure(cg, msg);
2204 }
2205 }
2206 break; // finish
2207 }
2208 }
2209
2210 igvn_worklist()->ensure_empty(); // should be done with igvn
2211
2212 while (inline_incrementally_one()) {
2213 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2214 }
2215 if (failing()) return;
2216
2217 inline_incrementally_cleanup(igvn);
2218
2219 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2220
2221 if (failing()) return;
2222
2223 if (_late_inlines.length() == 0) {
2224 break; // no more progress
2225 }
2226 }
2227
2228 igvn_worklist()->ensure_empty(); // should be done with igvn
2229
2230 if (_string_late_inlines.length() > 0) {
2231 assert(has_stringbuilder(), "inconsistent");
2232
2233 inline_string_calls(false);
2234
2235 if (failing()) return;
2236
2237 inline_incrementally_cleanup(igvn);
2238 }
2239
2240 set_inlining_incrementally(false);
2241 }
2242
2243 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2244 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2245 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2246 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2247 // as if "inlining_incrementally() == true" were set.
2248 assert(inlining_incrementally() == false, "not allowed");
2249 assert(_modified_nodes == nullptr, "not allowed");
2250 assert(_late_inlines.length() > 0, "sanity");
2251
2252 while (_late_inlines.length() > 0) {
2253 igvn_worklist()->ensure_empty(); // should be done with igvn
2254
2255 while (inline_incrementally_one()) {
2256 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2257 }
2258 if (failing()) return;
2259
2260 inline_incrementally_cleanup(igvn);
2261 }
2262 }
2263
2264 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2265 if (_loop_opts_cnt > 0) {
2266 while (major_progress() && (_loop_opts_cnt > 0)) {
2267 TracePhase tp(_t_idealLoop);
2268 PhaseIdealLoop::optimize(igvn, mode);
2269 _loop_opts_cnt--;
2270 if (failing()) return false;
2271 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2272 }
2273 }
2274 return true;
2275 }
2276
2277 // Remove edges from "root" to each SafePoint at a backward branch.
2278 // They were inserted during parsing (see add_safepoint()) to make
2279 // infinite loops without calls or exceptions visible to root, i.e.,
2280 // useful.
2281 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2282 Node *r = root();
2283 if (r != nullptr) {
2284 for (uint i = r->req(); i < r->len(); ++i) {
2285 Node *n = r->in(i);
2286 if (n != nullptr && n->is_SafePoint()) {
2287 r->rm_prec(i);
2288 if (n->outcnt() == 0) {
2289 igvn.remove_dead_node(n);
2290 }
2291 --i;
2292 }
2293 }
2294 // Parsing may have added top inputs to the root node (Path
2295 // leading to the Halt node proven dead). Make sure we get a
2296 // chance to clean them up.
2297 igvn._worklist.push(r);
2298 igvn.optimize();
2299 }
2300 }
2301
2302 //------------------------------Optimize---------------------------------------
2303 // Given a graph, optimize it.
2304 void Compile::Optimize() {
2305 TracePhase tp(_t_optimizer);
2306
2307 #ifndef PRODUCT
2308 if (env()->break_at_compile()) {
2309 BREAKPOINT;
2310 }
2311
2312 #endif
2313
2314 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2315 #ifdef ASSERT
2316 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2317 #endif
2318
2319 ResourceMark rm;
2320
2321 NOT_PRODUCT( verify_graph_edges(); )
2322
2323 print_method(PHASE_AFTER_PARSING, 1);
2324
2325 {
2326 // Iterative Global Value Numbering, including ideal transforms
2327 PhaseIterGVN igvn;
2328 #ifdef ASSERT
2329 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2330 #endif
2331 {
2332 TracePhase tp(_t_iterGVN);
2333 igvn.optimize();
2334 }
2335
2336 if (failing()) return;
2337
2338 print_method(PHASE_ITER_GVN1, 2);
2339
2340 process_for_unstable_if_traps(igvn);
2341
2342 if (failing()) return;
2343
2344 inline_incrementally(igvn);
2345
2346 print_method(PHASE_INCREMENTAL_INLINE, 2);
2347
2348 if (failing()) return;
2349
2350 if (eliminate_boxing()) {
2351 // Inline valueOf() methods now.
2352 inline_boxing_calls(igvn);
2353
2354 if (failing()) return;
2355
2356 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2357 inline_incrementally(igvn);
2358 }
2359
2360 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2361
2362 if (failing()) return;
2363 }
2364
2365 // Remove the speculative part of types and clean up the graph from
2366 // the extra CastPP nodes whose only purpose is to carry them. Do
2367 // that early so that optimizations are not disrupted by the extra
2368 // CastPP nodes.
2369 remove_speculative_types(igvn);
2370
2371 if (failing()) return;
2372
2373 // No more new expensive nodes will be added to the list from here
2374 // so keep only the actual candidates for optimizations.
2375 cleanup_expensive_nodes(igvn);
2376
2377 if (failing()) return;
2378
2379 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2380 if (EnableVectorSupport && has_vbox_nodes()) {
2381 TracePhase tp(_t_vector);
2382 PhaseVector pv(igvn);
2383 pv.optimize_vector_boxes();
2384 if (failing()) return;
2385 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2386 }
2387 assert(!has_vbox_nodes(), "sanity");
2388
2389 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2390 Compile::TracePhase tp(_t_renumberLive);
2391 igvn_worklist()->ensure_empty(); // should be done with igvn
2392 {
2393 ResourceMark rm;
2394 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2395 }
2396 igvn.reset();
2397 igvn.optimize();
2398 if (failing()) return;
2399 }
2400
2401 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2402 // safepoints
2403 remove_root_to_sfpts_edges(igvn);
2404
2405 if (failing()) return;
2406
2407 if (has_loops()) {
2408 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2409 }
2410
2411 // Perform escape analysis
2412 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2413 if (has_loops()) {
2414 // Cleanup graph (remove dead nodes).
2415 TracePhase tp(_t_idealLoop);
2416 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2417 if (failing()) return;
2418 }
2419 bool progress;
2420 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2421 do {
2422 ConnectionGraph::do_analysis(this, &igvn);
2423
2424 if (failing()) return;
2425
2426 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2427
2428 // Optimize out fields loads from scalar replaceable allocations.
2429 igvn.optimize();
2430 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2431
2432 if (failing()) return;
2433
2434 if (congraph() != nullptr && macro_count() > 0) {
2435 TracePhase tp(_t_macroEliminate);
2436 PhaseMacroExpand mexp(igvn);
2437 mexp.eliminate_macro_nodes();
2438 if (failing()) return;
2439 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2440
2441 igvn.set_delay_transform(false);
2442 igvn.optimize();
2443 if (failing()) return;
2444
2445 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2446 }
2447
2448 ConnectionGraph::verify_ram_nodes(this, root());
2449 if (failing()) return;
2450
2451 progress = do_iterative_escape_analysis() &&
2452 (macro_count() < mcount) &&
2453 ConnectionGraph::has_candidates(this);
2454 // Try again if candidates exist and made progress
2455 // by removing some allocations and/or locks.
2456 } while (progress);
2457 }
2458
2459 // Loop transforms on the ideal graph. Range Check Elimination,
2460 // peeling, unrolling, etc.
2461
2462 // Set loop opts counter
2463 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2464 {
2465 TracePhase tp(_t_idealLoop);
2466 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2467 _loop_opts_cnt--;
2468 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2469 if (failing()) return;
2470 }
2471 // Loop opts pass if partial peeling occurred in previous pass
2472 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2473 TracePhase tp(_t_idealLoop);
2474 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2475 _loop_opts_cnt--;
2476 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2477 if (failing()) return;
2478 }
2479 // Loop opts pass for loop-unrolling before CCP
2480 if(major_progress() && (_loop_opts_cnt > 0)) {
2481 TracePhase tp(_t_idealLoop);
2482 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2483 _loop_opts_cnt--;
2484 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2485 }
2486 if (!failing()) {
2487 // Verify that last round of loop opts produced a valid graph
2488 PhaseIdealLoop::verify(igvn);
2489 }
2490 }
2491 if (failing()) return;
2492
2493 // Conditional Constant Propagation;
2494 print_method(PHASE_BEFORE_CCP1, 2);
2495 PhaseCCP ccp( &igvn );
2496 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2497 {
2498 TracePhase tp(_t_ccp);
2499 ccp.do_transform();
2500 }
2501 print_method(PHASE_CCP1, 2);
2502
2503 assert( true, "Break here to ccp.dump_old2new_map()");
2504
2505 // Iterative Global Value Numbering, including ideal transforms
2506 {
2507 TracePhase tp(_t_iterGVN2);
2508 igvn.reset_from_igvn(&ccp);
2509 igvn.optimize();
2510 }
2511 print_method(PHASE_ITER_GVN2, 2);
2512
2513 if (failing()) return;
2514
2515 // Loop transforms on the ideal graph. Range Check Elimination,
2516 // peeling, unrolling, etc.
2517 if (!optimize_loops(igvn, LoopOptsDefault)) {
2518 return;
2519 }
2520
2521 if (failing()) return;
2522
2523 C->clear_major_progress(); // ensure that major progress is now clear
2524
2525 process_for_post_loop_opts_igvn(igvn);
2526
2527 process_for_merge_stores_igvn(igvn);
2528
2529 if (failing()) return;
2530
2531 #ifdef ASSERT
2532 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2533 #endif
2534
2535 {
2536 TracePhase tp(_t_macroExpand);
2537 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2538 PhaseMacroExpand mex(igvn);
2539 // Do not allow new macro nodes once we start to eliminate and expand
2540 C->reset_allow_macro_nodes();
2541 // Last attempt to eliminate macro nodes before expand
2542 mex.eliminate_macro_nodes();
2543 if (failing()) {
2544 return;
2545 }
2546 mex.eliminate_opaque_looplimit_macro_nodes();
2547 if (failing()) {
2548 return;
2549 }
2550 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2551 if (mex.expand_macro_nodes()) {
2552 assert(failing(), "must bail out w/ explicit message");
2553 return;
2554 }
2555 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2556 }
2557
2558 {
2559 TracePhase tp(_t_barrierExpand);
2560 if (bs->expand_barriers(this, igvn)) {
2561 assert(failing(), "must bail out w/ explicit message");
2562 return;
2563 }
2564 print_method(PHASE_BARRIER_EXPANSION, 2);
2565 }
2566
2567 if (C->max_vector_size() > 0) {
2568 C->optimize_logic_cones(igvn);
2569 igvn.optimize();
2570 if (failing()) return;
2571 }
2572
2573 DEBUG_ONLY( _modified_nodes = nullptr; )
2574
2575 assert(igvn._worklist.size() == 0, "not empty");
2576
2577 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2578
2579 if (_late_inlines.length() > 0) {
2580 // More opportunities to optimize virtual and MH calls.
2581 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2582 process_late_inline_calls_no_inline(igvn);
2583 if (failing()) return;
2584 }
2585 } // (End scope of igvn; run destructor if necessary for asserts.)
2586
2587 check_no_dead_use();
2588
2589 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2590 // to remove hashes to unlock nodes for modifications.
2591 C->node_hash()->clear();
2592
2593 // A method with only infinite loops has no edges entering loops from root
2594 {
2595 TracePhase tp(_t_graphReshaping);
2596 if (final_graph_reshaping()) {
2597 assert(failing(), "must bail out w/ explicit message");
2598 return;
2599 }
2600 }
2601
2602 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2603 DEBUG_ONLY(set_phase_optimize_finished();)
2604 }
2605
2606 #ifdef ASSERT
2607 void Compile::check_no_dead_use() const {
2608 ResourceMark rm;
2609 Unique_Node_List wq;
2610 wq.push(root());
2611 for (uint i = 0; i < wq.size(); ++i) {
2612 Node* n = wq.at(i);
2613 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2614 Node* u = n->fast_out(j);
2615 if (u->outcnt() == 0 && !u->is_Con()) {
2616 u->dump();
2617 fatal("no reachable node should have no use");
2618 }
2619 wq.push(u);
2620 }
2621 }
2622 }
2623 #endif
2624
2625 void Compile::inline_vector_reboxing_calls() {
2626 if (C->_vector_reboxing_late_inlines.length() > 0) {
2627 _late_inlines_pos = C->_late_inlines.length();
2628 while (_vector_reboxing_late_inlines.length() > 0) {
2629 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2630 cg->do_late_inline();
2631 if (failing()) return;
2632 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2633 }
2634 _vector_reboxing_late_inlines.trunc_to(0);
2635 }
2636 }
2637
2638 bool Compile::has_vbox_nodes() {
2639 if (C->_vector_reboxing_late_inlines.length() > 0) {
2640 return true;
2641 }
2642 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2643 Node * n = C->macro_node(macro_idx);
2644 assert(n->is_macro(), "only macro nodes expected here");
2645 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2646 return true;
2647 }
2648 }
2649 return false;
2650 }
2651
2652 //---------------------------- Bitwise operation packing optimization ---------------------------
2653
2654 static bool is_vector_unary_bitwise_op(Node* n) {
2655 return n->Opcode() == Op_XorV &&
2656 VectorNode::is_vector_bitwise_not_pattern(n);
2657 }
2658
2659 static bool is_vector_binary_bitwise_op(Node* n) {
2660 switch (n->Opcode()) {
2661 case Op_AndV:
2662 case Op_OrV:
2663 return true;
2664
2665 case Op_XorV:
2666 return !is_vector_unary_bitwise_op(n);
2667
2668 default:
2669 return false;
2670 }
2671 }
2672
2673 static bool is_vector_ternary_bitwise_op(Node* n) {
2674 return n->Opcode() == Op_MacroLogicV;
2675 }
2676
2677 static bool is_vector_bitwise_op(Node* n) {
2678 return is_vector_unary_bitwise_op(n) ||
2679 is_vector_binary_bitwise_op(n) ||
2680 is_vector_ternary_bitwise_op(n);
2681 }
2682
2683 static bool is_vector_bitwise_cone_root(Node* n) {
2684 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2685 return false;
2686 }
2687 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2688 if (is_vector_bitwise_op(n->fast_out(i))) {
2689 return false;
2690 }
2691 }
2692 return true;
2693 }
2694
2695 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2696 uint cnt = 0;
2697 if (is_vector_bitwise_op(n)) {
2698 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2699 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2700 for (uint i = 1; i < inp_cnt; i++) {
2701 Node* in = n->in(i);
2702 bool skip = VectorNode::is_all_ones_vector(in);
2703 if (!skip && !inputs.member(in)) {
2704 inputs.push(in);
2705 cnt++;
2706 }
2707 }
2708 assert(cnt <= 1, "not unary");
2709 } else {
2710 uint last_req = inp_cnt;
2711 if (is_vector_ternary_bitwise_op(n)) {
2712 last_req = inp_cnt - 1; // skip last input
2713 }
2714 for (uint i = 1; i < last_req; i++) {
2715 Node* def = n->in(i);
2716 if (!inputs.member(def)) {
2717 inputs.push(def);
2718 cnt++;
2719 }
2720 }
2721 }
2722 } else { // not a bitwise operations
2723 if (!inputs.member(n)) {
2724 inputs.push(n);
2725 cnt++;
2726 }
2727 }
2728 return cnt;
2729 }
2730
2731 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2732 Unique_Node_List useful_nodes;
2733 C->identify_useful_nodes(useful_nodes);
2734
2735 for (uint i = 0; i < useful_nodes.size(); i++) {
2736 Node* n = useful_nodes.at(i);
2737 if (is_vector_bitwise_cone_root(n)) {
2738 list.push(n);
2739 }
2740 }
2741 }
2742
2743 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2744 const TypeVect* vt,
2745 Unique_Node_List& partition,
2746 Unique_Node_List& inputs) {
2747 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2748 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2749 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2750
2751 Node* in1 = inputs.at(0);
2752 Node* in2 = inputs.at(1);
2753 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2754
2755 uint func = compute_truth_table(partition, inputs);
2756
2757 Node* pn = partition.at(partition.size() - 1);
2758 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2759 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2760 }
2761
2762 static uint extract_bit(uint func, uint pos) {
2763 return (func & (1 << pos)) >> pos;
2764 }
2765
2766 //
2767 // A macro logic node represents a truth table. It has 4 inputs,
2768 // First three inputs corresponds to 3 columns of a truth table
2769 // and fourth input captures the logic function.
2770 //
2771 // eg. fn = (in1 AND in2) OR in3;
2772 //
2773 // MacroNode(in1,in2,in3,fn)
2774 //
2775 // -----------------
2776 // in1 in2 in3 fn
2777 // -----------------
2778 // 0 0 0 0
2779 // 0 0 1 1
2780 // 0 1 0 0
2781 // 0 1 1 1
2782 // 1 0 0 0
2783 // 1 0 1 1
2784 // 1 1 0 1
2785 // 1 1 1 1
2786 //
2787
2788 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2789 int res = 0;
2790 for (int i = 0; i < 8; i++) {
2791 int bit1 = extract_bit(in1, i);
2792 int bit2 = extract_bit(in2, i);
2793 int bit3 = extract_bit(in3, i);
2794
2795 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2796 int func_bit = extract_bit(func, func_bit_pos);
2797
2798 res |= func_bit << i;
2799 }
2800 return res;
2801 }
2802
2803 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
2804 assert(n != nullptr, "");
2805 assert(eval_map.contains(n), "absent");
2806 return *(eval_map.get(n));
2807 }
2808
2809 static void eval_operands(Node* n,
2810 uint& func1, uint& func2, uint& func3,
2811 HashTable<Node*,uint>& eval_map) {
2812 assert(is_vector_bitwise_op(n), "");
2813
2814 if (is_vector_unary_bitwise_op(n)) {
2815 Node* opnd = n->in(1);
2816 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2817 opnd = n->in(2);
2818 }
2819 func1 = eval_operand(opnd, eval_map);
2820 } else if (is_vector_binary_bitwise_op(n)) {
2821 func1 = eval_operand(n->in(1), eval_map);
2822 func2 = eval_operand(n->in(2), eval_map);
2823 } else {
2824 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2825 func1 = eval_operand(n->in(1), eval_map);
2826 func2 = eval_operand(n->in(2), eval_map);
2827 func3 = eval_operand(n->in(3), eval_map);
2828 }
2829 }
2830
2831 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2832 assert(inputs.size() <= 3, "sanity");
2833 ResourceMark rm;
2834 uint res = 0;
2835 HashTable<Node*,uint> eval_map;
2836
2837 // Populate precomputed functions for inputs.
2838 // Each input corresponds to one column of 3 input truth-table.
2839 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2840 0xCC, // (_, b, _) -> b
2841 0xF0 }; // (a, _, _) -> a
2842 for (uint i = 0; i < inputs.size(); i++) {
2843 eval_map.put(inputs.at(i), input_funcs[2-i]);
2844 }
2845
2846 for (uint i = 0; i < partition.size(); i++) {
2847 Node* n = partition.at(i);
2848
2849 uint func1 = 0, func2 = 0, func3 = 0;
2850 eval_operands(n, func1, func2, func3, eval_map);
2851
2852 switch (n->Opcode()) {
2853 case Op_OrV:
2854 assert(func3 == 0, "not binary");
2855 res = func1 | func2;
2856 break;
2857 case Op_AndV:
2858 assert(func3 == 0, "not binary");
2859 res = func1 & func2;
2860 break;
2861 case Op_XorV:
2862 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2863 assert(func2 == 0 && func3 == 0, "not unary");
2864 res = (~func1) & 0xFF;
2865 } else {
2866 assert(func3 == 0, "not binary");
2867 res = func1 ^ func2;
2868 }
2869 break;
2870 case Op_MacroLogicV:
2871 // Ordering of inputs may change during evaluation of sub-tree
2872 // containing MacroLogic node as a child node, thus a re-evaluation
2873 // makes sure that function is evaluated in context of current
2874 // inputs.
2875 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2876 break;
2877
2878 default: assert(false, "not supported: %s", n->Name());
2879 }
2880 assert(res <= 0xFF, "invalid");
2881 eval_map.put(n, res);
2882 }
2883 return res;
2884 }
2885
2886 // Criteria under which nodes gets packed into a macro logic node:-
2887 // 1) Parent and both child nodes are all unmasked or masked with
2888 // same predicates.
2889 // 2) Masked parent can be packed with left child if it is predicated
2890 // and both have same predicates.
2891 // 3) Masked parent can be packed with right child if its un-predicated
2892 // or has matching predication condition.
2893 // 4) An unmasked parent can be packed with an unmasked child.
2894 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2895 assert(partition.size() == 0, "not empty");
2896 assert(inputs.size() == 0, "not empty");
2897 if (is_vector_ternary_bitwise_op(n)) {
2898 return false;
2899 }
2900
2901 bool is_unary_op = is_vector_unary_bitwise_op(n);
2902 if (is_unary_op) {
2903 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2904 return false; // too few inputs
2905 }
2906
2907 bool pack_left_child = true;
2908 bool pack_right_child = true;
2909
2910 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2911 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2912
2913 int left_child_input_cnt = 0;
2914 int right_child_input_cnt = 0;
2915
2916 bool parent_is_predicated = n->is_predicated_vector();
2917 bool left_child_predicated = n->in(1)->is_predicated_vector();
2918 bool right_child_predicated = n->in(2)->is_predicated_vector();
2919
2920 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2921 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2922 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2923
2924 do {
2925 if (pack_left_child && left_child_LOP &&
2926 ((!parent_is_predicated && !left_child_predicated) ||
2927 ((parent_is_predicated && left_child_predicated &&
2928 parent_pred == left_child_pred)))) {
2929 partition.push(n->in(1));
2930 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2931 } else {
2932 inputs.push(n->in(1));
2933 left_child_input_cnt = 1;
2934 }
2935
2936 if (pack_right_child && right_child_LOP &&
2937 (!right_child_predicated ||
2938 (right_child_predicated && parent_is_predicated &&
2939 parent_pred == right_child_pred))) {
2940 partition.push(n->in(2));
2941 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2942 } else {
2943 inputs.push(n->in(2));
2944 right_child_input_cnt = 1;
2945 }
2946
2947 if (inputs.size() > 3) {
2948 assert(partition.size() > 0, "");
2949 inputs.clear();
2950 partition.clear();
2951 if (left_child_input_cnt > right_child_input_cnt) {
2952 pack_left_child = false;
2953 } else {
2954 pack_right_child = false;
2955 }
2956 } else {
2957 break;
2958 }
2959 } while(true);
2960
2961 if(partition.size()) {
2962 partition.push(n);
2963 }
2964
2965 return (partition.size() == 2 || partition.size() == 3) &&
2966 (inputs.size() == 2 || inputs.size() == 3);
2967 }
2968
2969 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2970 assert(is_vector_bitwise_op(n), "not a root");
2971
2972 visited.set(n->_idx);
2973
2974 // 1) Do a DFS walk over the logic cone.
2975 for (uint i = 1; i < n->req(); i++) {
2976 Node* in = n->in(i);
2977 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2978 process_logic_cone_root(igvn, in, visited);
2979 }
2980 }
2981
2982 // 2) Bottom up traversal: Merge node[s] with
2983 // the parent to form macro logic node.
2984 Unique_Node_List partition;
2985 Unique_Node_List inputs;
2986 if (compute_logic_cone(n, partition, inputs)) {
2987 const TypeVect* vt = n->bottom_type()->is_vect();
2988 Node* pn = partition.at(partition.size() - 1);
2989 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2990 if (mask == nullptr ||
2991 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2992 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2993 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2994 igvn.replace_node(n, macro_logic);
2995 }
2996 }
2997 }
2998
2999 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3000 ResourceMark rm;
3001 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3002 Unique_Node_List list;
3003 collect_logic_cone_roots(list);
3004
3005 while (list.size() > 0) {
3006 Node* n = list.pop();
3007 const TypeVect* vt = n->bottom_type()->is_vect();
3008 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3009 if (supported) {
3010 VectorSet visited(comp_arena());
3011 process_logic_cone_root(igvn, n, visited);
3012 }
3013 }
3014 }
3015 }
3016
3017 //------------------------------Code_Gen---------------------------------------
3018 // Given a graph, generate code for it
3019 void Compile::Code_Gen() {
3020 if (failing()) {
3021 return;
3022 }
3023
3024 // Perform instruction selection. You might think we could reclaim Matcher
3025 // memory PDQ, but actually the Matcher is used in generating spill code.
3026 // Internals of the Matcher (including some VectorSets) must remain live
3027 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3028 // set a bit in reclaimed memory.
3029
3030 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3031 // nodes. Mapping is only valid at the root of each matched subtree.
3032 NOT_PRODUCT( verify_graph_edges(); )
3033
3034 Matcher matcher;
3035 _matcher = &matcher;
3036 {
3037 TracePhase tp(_t_matcher);
3038 matcher.match();
3039 if (failing()) {
3040 return;
3041 }
3042 }
3043 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3044 // nodes. Mapping is only valid at the root of each matched subtree.
3045 NOT_PRODUCT( verify_graph_edges(); )
3046
3047 // If you have too many nodes, or if matching has failed, bail out
3048 check_node_count(0, "out of nodes matching instructions");
3049 if (failing()) {
3050 return;
3051 }
3052
3053 print_method(PHASE_MATCHING, 2);
3054
3055 // Build a proper-looking CFG
3056 PhaseCFG cfg(node_arena(), root(), matcher);
3057 if (failing()) {
3058 return;
3059 }
3060 _cfg = &cfg;
3061 {
3062 TracePhase tp(_t_scheduler);
3063 bool success = cfg.do_global_code_motion();
3064 if (!success) {
3065 return;
3066 }
3067
3068 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3069 NOT_PRODUCT( verify_graph_edges(); )
3070 cfg.verify();
3071 if (failing()) {
3072 return;
3073 }
3074 }
3075
3076 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3077 _regalloc = ®alloc;
3078 {
3079 TracePhase tp(_t_registerAllocation);
3080 // Perform register allocation. After Chaitin, use-def chains are
3081 // no longer accurate (at spill code) and so must be ignored.
3082 // Node->LRG->reg mappings are still accurate.
3083 _regalloc->Register_Allocate();
3084
3085 // Bail out if the allocator builds too many nodes
3086 if (failing()) {
3087 return;
3088 }
3089
3090 print_method(PHASE_REGISTER_ALLOCATION, 2);
3091 }
3092
3093 // Prior to register allocation we kept empty basic blocks in case the
3094 // the allocator needed a place to spill. After register allocation we
3095 // are not adding any new instructions. If any basic block is empty, we
3096 // can now safely remove it.
3097 {
3098 TracePhase tp(_t_blockOrdering);
3099 cfg.remove_empty_blocks();
3100 if (do_freq_based_layout()) {
3101 PhaseBlockLayout layout(cfg);
3102 } else {
3103 cfg.set_loop_alignment();
3104 }
3105 cfg.fixup_flow();
3106 cfg.remove_unreachable_blocks();
3107 cfg.verify_dominator_tree();
3108 print_method(PHASE_BLOCK_ORDERING, 3);
3109 }
3110
3111 // Apply peephole optimizations
3112 if( OptoPeephole ) {
3113 TracePhase tp(_t_peephole);
3114 PhasePeephole peep( _regalloc, cfg);
3115 peep.do_transform();
3116 print_method(PHASE_PEEPHOLE, 3);
3117 }
3118
3119 // Do late expand if CPU requires this.
3120 if (Matcher::require_postalloc_expand) {
3121 TracePhase tp(_t_postalloc_expand);
3122 cfg.postalloc_expand(_regalloc);
3123 print_method(PHASE_POSTALLOC_EXPAND, 3);
3124 }
3125
3126 #ifdef ASSERT
3127 {
3128 CompilationMemoryStatistic::do_test_allocations();
3129 if (failing()) return;
3130 }
3131 #endif
3132
3133 // Convert Nodes to instruction bits in a buffer
3134 {
3135 TracePhase tp(_t_output);
3136 PhaseOutput output;
3137 output.Output();
3138 if (failing()) return;
3139 output.install();
3140 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3141 }
3142
3143 // He's dead, Jim.
3144 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3145 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3146 }
3147
3148 //------------------------------Final_Reshape_Counts---------------------------
3149 // This class defines counters to help identify when a method
3150 // may/must be executed using hardware with only 24-bit precision.
3151 struct Final_Reshape_Counts : public StackObj {
3152 int _call_count; // count non-inlined 'common' calls
3153 int _float_count; // count float ops requiring 24-bit precision
3154 int _double_count; // count double ops requiring more precision
3155 int _java_call_count; // count non-inlined 'java' calls
3156 int _inner_loop_count; // count loops which need alignment
3157 VectorSet _visited; // Visitation flags
3158 Node_List _tests; // Set of IfNodes & PCTableNodes
3159
3160 Final_Reshape_Counts() :
3161 _call_count(0), _float_count(0), _double_count(0),
3162 _java_call_count(0), _inner_loop_count(0) { }
3163
3164 void inc_call_count () { _call_count ++; }
3165 void inc_float_count () { _float_count ++; }
3166 void inc_double_count() { _double_count++; }
3167 void inc_java_call_count() { _java_call_count++; }
3168 void inc_inner_loop_count() { _inner_loop_count++; }
3169
3170 int get_call_count () const { return _call_count ; }
3171 int get_float_count () const { return _float_count ; }
3172 int get_double_count() const { return _double_count; }
3173 int get_java_call_count() const { return _java_call_count; }
3174 int get_inner_loop_count() const { return _inner_loop_count; }
3175 };
3176
3177 //------------------------------final_graph_reshaping_impl----------------------
3178 // Implement items 1-5 from final_graph_reshaping below.
3179 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3180
3181 if ( n->outcnt() == 0 ) return; // dead node
3182 uint nop = n->Opcode();
3183
3184 // Check for 2-input instruction with "last use" on right input.
3185 // Swap to left input. Implements item (2).
3186 if( n->req() == 3 && // two-input instruction
3187 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3188 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3189 n->in(2)->outcnt() == 1 &&// right use IS a last use
3190 !n->in(2)->is_Con() ) { // right use is not a constant
3191 // Check for commutative opcode
3192 switch( nop ) {
3193 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3194 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3195 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3196 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3197 case Op_AndL: case Op_XorL: case Op_OrL:
3198 case Op_AndI: case Op_XorI: case Op_OrI: {
3199 // Move "last use" input to left by swapping inputs
3200 n->swap_edges(1, 2);
3201 break;
3202 }
3203 default:
3204 break;
3205 }
3206 }
3207
3208 #ifdef ASSERT
3209 if( n->is_Mem() ) {
3210 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3211 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3212 // oop will be recorded in oop map if load crosses safepoint
3213 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3214 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3215 "raw memory operations should have control edge");
3216 }
3217 if (n->is_MemBar()) {
3218 MemBarNode* mb = n->as_MemBar();
3219 if (mb->trailing_store() || mb->trailing_load_store()) {
3220 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3221 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3222 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3223 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3224 } else if (mb->leading()) {
3225 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3226 }
3227 }
3228 #endif
3229 // Count FPU ops and common calls, implements item (3)
3230 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3231 if (!gc_handled) {
3232 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3233 }
3234
3235 // Collect CFG split points
3236 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3237 frc._tests.push(n);
3238 }
3239 }
3240
3241 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3242 if (!UseDivMod) {
3243 return;
3244 }
3245
3246 // Check if "a % b" and "a / b" both exist
3247 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3248 if (d == nullptr) {
3249 return;
3250 }
3251
3252 // Replace them with a fused divmod if supported
3253 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3254 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3255 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3256 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3257 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3258 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3259 // DivMod node so the dependency is not lost.
3260 divmod->add_prec_from(n);
3261 divmod->add_prec_from(d);
3262 d->subsume_by(divmod->div_proj(), this);
3263 n->subsume_by(divmod->mod_proj(), this);
3264 } else {
3265 // Replace "a % b" with "a - ((a / b) * b)"
3266 Node* mult = MulNode::make(d, d->in(2), bt);
3267 Node* sub = SubNode::make(d->in(1), mult, bt);
3268 n->subsume_by(sub, this);
3269 }
3270 }
3271
3272 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3273 switch( nop ) {
3274 // Count all float operations that may use FPU
3275 case Op_AddF:
3276 case Op_SubF:
3277 case Op_MulF:
3278 case Op_DivF:
3279 case Op_NegF:
3280 case Op_ModF:
3281 case Op_ConvI2F:
3282 case Op_ConF:
3283 case Op_CmpF:
3284 case Op_CmpF3:
3285 case Op_StoreF:
3286 case Op_LoadF:
3287 // case Op_ConvL2F: // longs are split into 32-bit halves
3288 frc.inc_float_count();
3289 break;
3290
3291 case Op_ConvF2D:
3292 case Op_ConvD2F:
3293 frc.inc_float_count();
3294 frc.inc_double_count();
3295 break;
3296
3297 // Count all double operations that may use FPU
3298 case Op_AddD:
3299 case Op_SubD:
3300 case Op_MulD:
3301 case Op_DivD:
3302 case Op_NegD:
3303 case Op_ModD:
3304 case Op_ConvI2D:
3305 case Op_ConvD2I:
3306 // case Op_ConvL2D: // handled by leaf call
3307 // case Op_ConvD2L: // handled by leaf call
3308 case Op_ConD:
3309 case Op_CmpD:
3310 case Op_CmpD3:
3311 case Op_StoreD:
3312 case Op_LoadD:
3313 case Op_LoadD_unaligned:
3314 frc.inc_double_count();
3315 break;
3316 case Op_Opaque1: // Remove Opaque Nodes before matching
3317 n->subsume_by(n->in(1), this);
3318 break;
3319 case Op_CallLeafPure: {
3320 // If the pure call is not supported, then lower to a CallLeaf.
3321 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3322 CallNode* call = n->as_Call();
3323 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3324 call->_name, TypeRawPtr::BOTTOM);
3325 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3326 new_call->init_req(TypeFunc::I_O, C->top());
3327 new_call->init_req(TypeFunc::Memory, C->top());
3328 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3329 new_call->init_req(TypeFunc::FramePtr, C->top());
3330 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
3331 new_call->init_req(i, call->in(i));
3332 }
3333 n->subsume_by(new_call, this);
3334 }
3335 frc.inc_call_count();
3336 break;
3337 }
3338 case Op_CallStaticJava:
3339 case Op_CallJava:
3340 case Op_CallDynamicJava:
3341 frc.inc_java_call_count(); // Count java call site;
3342 case Op_CallRuntime:
3343 case Op_CallLeaf:
3344 case Op_CallLeafVector:
3345 case Op_CallLeafNoFP: {
3346 assert (n->is_Call(), "");
3347 CallNode *call = n->as_Call();
3348 // Count call sites where the FP mode bit would have to be flipped.
3349 // Do not count uncommon runtime calls:
3350 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3351 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3352 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3353 frc.inc_call_count(); // Count the call site
3354 } else { // See if uncommon argument is shared
3355 Node *n = call->in(TypeFunc::Parms);
3356 int nop = n->Opcode();
3357 // Clone shared simple arguments to uncommon calls, item (1).
3358 if (n->outcnt() > 1 &&
3359 !n->is_Proj() &&
3360 nop != Op_CreateEx &&
3361 nop != Op_CheckCastPP &&
3362 nop != Op_DecodeN &&
3363 nop != Op_DecodeNKlass &&
3364 !n->is_Mem() &&
3365 !n->is_Phi()) {
3366 Node *x = n->clone();
3367 call->set_req(TypeFunc::Parms, x);
3368 }
3369 }
3370 break;
3371 }
3372 case Op_StoreB:
3373 case Op_StoreC:
3374 case Op_StoreI:
3375 case Op_StoreL:
3376 case Op_CompareAndSwapB:
3377 case Op_CompareAndSwapS:
3378 case Op_CompareAndSwapI:
3379 case Op_CompareAndSwapL:
3380 case Op_CompareAndSwapP:
3381 case Op_CompareAndSwapN:
3382 case Op_WeakCompareAndSwapB:
3383 case Op_WeakCompareAndSwapS:
3384 case Op_WeakCompareAndSwapI:
3385 case Op_WeakCompareAndSwapL:
3386 case Op_WeakCompareAndSwapP:
3387 case Op_WeakCompareAndSwapN:
3388 case Op_CompareAndExchangeB:
3389 case Op_CompareAndExchangeS:
3390 case Op_CompareAndExchangeI:
3391 case Op_CompareAndExchangeL:
3392 case Op_CompareAndExchangeP:
3393 case Op_CompareAndExchangeN:
3394 case Op_GetAndAddS:
3395 case Op_GetAndAddB:
3396 case Op_GetAndAddI:
3397 case Op_GetAndAddL:
3398 case Op_GetAndSetS:
3399 case Op_GetAndSetB:
3400 case Op_GetAndSetI:
3401 case Op_GetAndSetL:
3402 case Op_GetAndSetP:
3403 case Op_GetAndSetN:
3404 case Op_StoreP:
3405 case Op_StoreN:
3406 case Op_StoreNKlass:
3407 case Op_LoadB:
3408 case Op_LoadUB:
3409 case Op_LoadUS:
3410 case Op_LoadI:
3411 case Op_LoadKlass:
3412 case Op_LoadNKlass:
3413 case Op_LoadL:
3414 case Op_LoadL_unaligned:
3415 case Op_LoadP:
3416 case Op_LoadN:
3417 case Op_LoadRange:
3418 case Op_LoadS:
3419 break;
3420
3421 case Op_AddP: { // Assert sane base pointers
3422 Node *addp = n->in(AddPNode::Address);
3423 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
3424 #ifdef _LP64
3425 if ((UseCompressedOops || UseCompressedClassPointers) &&
3426 addp->Opcode() == Op_ConP &&
3427 addp == n->in(AddPNode::Base) &&
3428 n->in(AddPNode::Offset)->is_Con()) {
3429 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3430 // on the platform and on the compressed oops mode.
3431 // Use addressing with narrow klass to load with offset on x86.
3432 // Some platforms can use the constant pool to load ConP.
3433 // Do this transformation here since IGVN will convert ConN back to ConP.
3434 const Type* t = addp->bottom_type();
3435 bool is_oop = t->isa_oopptr() != nullptr;
3436 bool is_klass = t->isa_klassptr() != nullptr;
3437
3438 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3439 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3440 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3441 Node* nn = nullptr;
3442
3443 int op = is_oop ? Op_ConN : Op_ConNKlass;
3444
3445 // Look for existing ConN node of the same exact type.
3446 Node* r = root();
3447 uint cnt = r->outcnt();
3448 for (uint i = 0; i < cnt; i++) {
3449 Node* m = r->raw_out(i);
3450 if (m!= nullptr && m->Opcode() == op &&
3451 m->bottom_type()->make_ptr() == t) {
3452 nn = m;
3453 break;
3454 }
3455 }
3456 if (nn != nullptr) {
3457 // Decode a narrow oop to match address
3458 // [R12 + narrow_oop_reg<<3 + offset]
3459 if (is_oop) {
3460 nn = new DecodeNNode(nn, t);
3461 } else {
3462 nn = new DecodeNKlassNode(nn, t);
3463 }
3464 // Check for succeeding AddP which uses the same Base.
3465 // Otherwise we will run into the assertion above when visiting that guy.
3466 for (uint i = 0; i < n->outcnt(); ++i) {
3467 Node *out_i = n->raw_out(i);
3468 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3469 out_i->set_req(AddPNode::Base, nn);
3470 #ifdef ASSERT
3471 for (uint j = 0; j < out_i->outcnt(); ++j) {
3472 Node *out_j = out_i->raw_out(j);
3473 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3474 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3475 }
3476 #endif
3477 }
3478 }
3479 n->set_req(AddPNode::Base, nn);
3480 n->set_req(AddPNode::Address, nn);
3481 if (addp->outcnt() == 0) {
3482 addp->disconnect_inputs(this);
3483 }
3484 }
3485 }
3486 }
3487 #endif
3488 break;
3489 }
3490
3491 case Op_CastPP: {
3492 // Remove CastPP nodes to gain more freedom during scheduling but
3493 // keep the dependency they encode as control or precedence edges
3494 // (if control is set already) on memory operations. Some CastPP
3495 // nodes don't have a control (don't carry a dependency): skip
3496 // those.
3497 if (n->in(0) != nullptr) {
3498 ResourceMark rm;
3499 Unique_Node_List wq;
3500 wq.push(n);
3501 for (uint next = 0; next < wq.size(); ++next) {
3502 Node *m = wq.at(next);
3503 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3504 Node* use = m->fast_out(i);
3505 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3506 use->ensure_control_or_add_prec(n->in(0));
3507 } else {
3508 switch(use->Opcode()) {
3509 case Op_AddP:
3510 case Op_DecodeN:
3511 case Op_DecodeNKlass:
3512 case Op_CheckCastPP:
3513 case Op_CastPP:
3514 wq.push(use);
3515 break;
3516 }
3517 }
3518 }
3519 }
3520 }
3521 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3522 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3523 Node* in1 = n->in(1);
3524 const Type* t = n->bottom_type();
3525 Node* new_in1 = in1->clone();
3526 new_in1->as_DecodeN()->set_type(t);
3527
3528 if (!Matcher::narrow_oop_use_complex_address()) {
3529 //
3530 // x86, ARM and friends can handle 2 adds in addressing mode
3531 // and Matcher can fold a DecodeN node into address by using
3532 // a narrow oop directly and do implicit null check in address:
3533 //
3534 // [R12 + narrow_oop_reg<<3 + offset]
3535 // NullCheck narrow_oop_reg
3536 //
3537 // On other platforms (Sparc) we have to keep new DecodeN node and
3538 // use it to do implicit null check in address:
3539 //
3540 // decode_not_null narrow_oop_reg, base_reg
3541 // [base_reg + offset]
3542 // NullCheck base_reg
3543 //
3544 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3545 // to keep the information to which null check the new DecodeN node
3546 // corresponds to use it as value in implicit_null_check().
3547 //
3548 new_in1->set_req(0, n->in(0));
3549 }
3550
3551 n->subsume_by(new_in1, this);
3552 if (in1->outcnt() == 0) {
3553 in1->disconnect_inputs(this);
3554 }
3555 } else {
3556 n->subsume_by(n->in(1), this);
3557 if (n->outcnt() == 0) {
3558 n->disconnect_inputs(this);
3559 }
3560 }
3561 break;
3562 }
3563 case Op_CastII: {
3564 n->as_CastII()->remove_range_check_cast(this);
3565 break;
3566 }
3567 #ifdef _LP64
3568 case Op_CmpP:
3569 // Do this transformation here to preserve CmpPNode::sub() and
3570 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3571 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3572 Node* in1 = n->in(1);
3573 Node* in2 = n->in(2);
3574 if (!in1->is_DecodeNarrowPtr()) {
3575 in2 = in1;
3576 in1 = n->in(2);
3577 }
3578 assert(in1->is_DecodeNarrowPtr(), "sanity");
3579
3580 Node* new_in2 = nullptr;
3581 if (in2->is_DecodeNarrowPtr()) {
3582 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3583 new_in2 = in2->in(1);
3584 } else if (in2->Opcode() == Op_ConP) {
3585 const Type* t = in2->bottom_type();
3586 if (t == TypePtr::NULL_PTR) {
3587 assert(in1->is_DecodeN(), "compare klass to null?");
3588 // Don't convert CmpP null check into CmpN if compressed
3589 // oops implicit null check is not generated.
3590 // This will allow to generate normal oop implicit null check.
3591 if (Matcher::gen_narrow_oop_implicit_null_checks())
3592 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3593 //
3594 // This transformation together with CastPP transformation above
3595 // will generated code for implicit null checks for compressed oops.
3596 //
3597 // The original code after Optimize()
3598 //
3599 // LoadN memory, narrow_oop_reg
3600 // decode narrow_oop_reg, base_reg
3601 // CmpP base_reg, nullptr
3602 // CastPP base_reg // NotNull
3603 // Load [base_reg + offset], val_reg
3604 //
3605 // after these transformations will be
3606 //
3607 // LoadN memory, narrow_oop_reg
3608 // CmpN narrow_oop_reg, nullptr
3609 // decode_not_null narrow_oop_reg, base_reg
3610 // Load [base_reg + offset], val_reg
3611 //
3612 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3613 // since narrow oops can be used in debug info now (see the code in
3614 // final_graph_reshaping_walk()).
3615 //
3616 // At the end the code will be matched to
3617 // on x86:
3618 //
3619 // Load_narrow_oop memory, narrow_oop_reg
3620 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3621 // NullCheck narrow_oop_reg
3622 //
3623 // and on sparc:
3624 //
3625 // Load_narrow_oop memory, narrow_oop_reg
3626 // decode_not_null narrow_oop_reg, base_reg
3627 // Load [base_reg + offset], val_reg
3628 // NullCheck base_reg
3629 //
3630 } else if (t->isa_oopptr()) {
3631 new_in2 = ConNode::make(t->make_narrowoop());
3632 } else if (t->isa_klassptr()) {
3633 ciKlass* klass = t->is_klassptr()->exact_klass();
3634 if (klass->is_in_encoding_range()) {
3635 new_in2 = ConNode::make(t->make_narrowklass());
3636 }
3637 }
3638 }
3639 if (new_in2 != nullptr) {
3640 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3641 n->subsume_by(cmpN, this);
3642 if (in1->outcnt() == 0) {
3643 in1->disconnect_inputs(this);
3644 }
3645 if (in2->outcnt() == 0) {
3646 in2->disconnect_inputs(this);
3647 }
3648 }
3649 }
3650 break;
3651
3652 case Op_DecodeN:
3653 case Op_DecodeNKlass:
3654 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3655 // DecodeN could be pinned when it can't be fold into
3656 // an address expression, see the code for Op_CastPP above.
3657 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3658 break;
3659
3660 case Op_EncodeP:
3661 case Op_EncodePKlass: {
3662 Node* in1 = n->in(1);
3663 if (in1->is_DecodeNarrowPtr()) {
3664 n->subsume_by(in1->in(1), this);
3665 } else if (in1->Opcode() == Op_ConP) {
3666 const Type* t = in1->bottom_type();
3667 if (t == TypePtr::NULL_PTR) {
3668 assert(t->isa_oopptr(), "null klass?");
3669 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3670 } else if (t->isa_oopptr()) {
3671 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3672 } else if (t->isa_klassptr()) {
3673 ciKlass* klass = t->is_klassptr()->exact_klass();
3674 if (klass->is_in_encoding_range()) {
3675 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3676 } else {
3677 assert(false, "unencodable klass in ConP -> EncodeP");
3678 C->record_failure("unencodable klass in ConP -> EncodeP");
3679 }
3680 }
3681 }
3682 if (in1->outcnt() == 0) {
3683 in1->disconnect_inputs(this);
3684 }
3685 break;
3686 }
3687
3688 case Op_Proj: {
3689 if (OptimizeStringConcat || IncrementalInline) {
3690 ProjNode* proj = n->as_Proj();
3691 if (proj->_is_io_use) {
3692 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3693 // Separate projections were used for the exception path which
3694 // are normally removed by a late inline. If it wasn't inlined
3695 // then they will hang around and should just be replaced with
3696 // the original one. Merge them.
3697 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3698 if (non_io_proj != nullptr) {
3699 proj->subsume_by(non_io_proj , this);
3700 }
3701 }
3702 }
3703 break;
3704 }
3705
3706 case Op_Phi:
3707 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3708 // The EncodeP optimization may create Phi with the same edges
3709 // for all paths. It is not handled well by Register Allocator.
3710 Node* unique_in = n->in(1);
3711 assert(unique_in != nullptr, "");
3712 uint cnt = n->req();
3713 for (uint i = 2; i < cnt; i++) {
3714 Node* m = n->in(i);
3715 assert(m != nullptr, "");
3716 if (unique_in != m)
3717 unique_in = nullptr;
3718 }
3719 if (unique_in != nullptr) {
3720 n->subsume_by(unique_in, this);
3721 }
3722 }
3723 break;
3724
3725 #endif
3726
3727 case Op_ModI:
3728 handle_div_mod_op(n, T_INT, false);
3729 break;
3730
3731 case Op_ModL:
3732 handle_div_mod_op(n, T_LONG, false);
3733 break;
3734
3735 case Op_UModI:
3736 handle_div_mod_op(n, T_INT, true);
3737 break;
3738
3739 case Op_UModL:
3740 handle_div_mod_op(n, T_LONG, true);
3741 break;
3742
3743 case Op_LoadVector:
3744 case Op_StoreVector:
3745 #ifdef ASSERT
3746 // Add VerifyVectorAlignment node between adr and load / store.
3747 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3748 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3749 n->as_StoreVector()->must_verify_alignment();
3750 if (must_verify_alignment) {
3751 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3752 n->as_StoreVector()->memory_size();
3753 // The memory access should be aligned to the vector width in bytes.
3754 // However, the underlying array is possibly less well aligned, but at least
3755 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3756 // a loop we can expect at least the following alignment:
3757 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3758 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3759 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3760 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3761 jlong mask = guaranteed_alignment - 1;
3762 Node* mask_con = ConLNode::make(mask);
3763 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3764 n->set_req(MemNode::Address, va);
3765 }
3766 }
3767 #endif
3768 break;
3769
3770 case Op_LoadVectorGather:
3771 case Op_StoreVectorScatter:
3772 case Op_LoadVectorGatherMasked:
3773 case Op_StoreVectorScatterMasked:
3774 case Op_VectorCmpMasked:
3775 case Op_VectorMaskGen:
3776 case Op_LoadVectorMasked:
3777 case Op_StoreVectorMasked:
3778 break;
3779
3780 case Op_AddReductionVI:
3781 case Op_AddReductionVL:
3782 case Op_AddReductionVF:
3783 case Op_AddReductionVD:
3784 case Op_MulReductionVI:
3785 case Op_MulReductionVL:
3786 case Op_MulReductionVF:
3787 case Op_MulReductionVD:
3788 case Op_MinReductionV:
3789 case Op_MaxReductionV:
3790 case Op_AndReductionV:
3791 case Op_OrReductionV:
3792 case Op_XorReductionV:
3793 break;
3794
3795 case Op_PackB:
3796 case Op_PackS:
3797 case Op_PackI:
3798 case Op_PackF:
3799 case Op_PackL:
3800 case Op_PackD:
3801 if (n->req()-1 > 2) {
3802 // Replace many operand PackNodes with a binary tree for matching
3803 PackNode* p = (PackNode*) n;
3804 Node* btp = p->binary_tree_pack(1, n->req());
3805 n->subsume_by(btp, this);
3806 }
3807 break;
3808 case Op_Loop:
3809 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3810 case Op_CountedLoop:
3811 case Op_LongCountedLoop:
3812 case Op_OuterStripMinedLoop:
3813 if (n->as_Loop()->is_inner_loop()) {
3814 frc.inc_inner_loop_count();
3815 }
3816 n->as_Loop()->verify_strip_mined(0);
3817 break;
3818 case Op_LShiftI:
3819 case Op_RShiftI:
3820 case Op_URShiftI:
3821 case Op_LShiftL:
3822 case Op_RShiftL:
3823 case Op_URShiftL:
3824 if (Matcher::need_masked_shift_count) {
3825 // The cpu's shift instructions don't restrict the count to the
3826 // lower 5/6 bits. We need to do the masking ourselves.
3827 Node* in2 = n->in(2);
3828 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3829 const TypeInt* t = in2->find_int_type();
3830 if (t != nullptr && t->is_con()) {
3831 juint shift = t->get_con();
3832 if (shift > mask) { // Unsigned cmp
3833 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3834 }
3835 } else {
3836 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3837 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3838 n->set_req(2, shift);
3839 }
3840 }
3841 if (in2->outcnt() == 0) { // Remove dead node
3842 in2->disconnect_inputs(this);
3843 }
3844 }
3845 break;
3846 case Op_MemBarStoreStore:
3847 case Op_MemBarRelease:
3848 // Break the link with AllocateNode: it is no longer useful and
3849 // confuses register allocation.
3850 if (n->req() > MemBarNode::Precedent) {
3851 n->set_req(MemBarNode::Precedent, top());
3852 }
3853 break;
3854 case Op_MemBarAcquire: {
3855 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3856 // At parse time, the trailing MemBarAcquire for a volatile load
3857 // is created with an edge to the load. After optimizations,
3858 // that input may be a chain of Phis. If those phis have no
3859 // other use, then the MemBarAcquire keeps them alive and
3860 // register allocation can be confused.
3861 dead_nodes.push(n->in(MemBarNode::Precedent));
3862 n->set_req(MemBarNode::Precedent, top());
3863 }
3864 break;
3865 }
3866 case Op_Blackhole:
3867 break;
3868 case Op_RangeCheck: {
3869 RangeCheckNode* rc = n->as_RangeCheck();
3870 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3871 n->subsume_by(iff, this);
3872 frc._tests.push(iff);
3873 break;
3874 }
3875 case Op_ConvI2L: {
3876 if (!Matcher::convi2l_type_required) {
3877 // Code generation on some platforms doesn't need accurate
3878 // ConvI2L types. Widening the type can help remove redundant
3879 // address computations.
3880 n->as_Type()->set_type(TypeLong::INT);
3881 ResourceMark rm;
3882 Unique_Node_List wq;
3883 wq.push(n);
3884 for (uint next = 0; next < wq.size(); next++) {
3885 Node *m = wq.at(next);
3886
3887 for(;;) {
3888 // Loop over all nodes with identical inputs edges as m
3889 Node* k = m->find_similar(m->Opcode());
3890 if (k == nullptr) {
3891 break;
3892 }
3893 // Push their uses so we get a chance to remove node made
3894 // redundant
3895 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3896 Node* u = k->fast_out(i);
3897 if (u->Opcode() == Op_LShiftL ||
3898 u->Opcode() == Op_AddL ||
3899 u->Opcode() == Op_SubL ||
3900 u->Opcode() == Op_AddP) {
3901 wq.push(u);
3902 }
3903 }
3904 // Replace all nodes with identical edges as m with m
3905 k->subsume_by(m, this);
3906 }
3907 }
3908 }
3909 break;
3910 }
3911 case Op_CmpUL: {
3912 if (!Matcher::has_match_rule(Op_CmpUL)) {
3913 // No support for unsigned long comparisons
3914 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3915 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3916 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3917 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3918 Node* andl = new AndLNode(orl, remove_sign_mask);
3919 Node* cmp = new CmpLNode(andl, n->in(2));
3920 n->subsume_by(cmp, this);
3921 }
3922 break;
3923 }
3924 #ifdef ASSERT
3925 case Op_ConNKlass: {
3926 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3927 ciKlass* klass = tp->is_klassptr()->exact_klass();
3928 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3929 break;
3930 }
3931 #endif
3932 default:
3933 assert(!n->is_Call(), "");
3934 assert(!n->is_Mem(), "");
3935 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3936 break;
3937 }
3938 }
3939
3940 //------------------------------final_graph_reshaping_walk---------------------
3941 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3942 // requires that the walk visits a node's inputs before visiting the node.
3943 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3944 Unique_Node_List sfpt;
3945
3946 frc._visited.set(root->_idx); // first, mark node as visited
3947 uint cnt = root->req();
3948 Node *n = root;
3949 uint i = 0;
3950 while (true) {
3951 if (i < cnt) {
3952 // Place all non-visited non-null inputs onto stack
3953 Node* m = n->in(i);
3954 ++i;
3955 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3956 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3957 // compute worst case interpreter size in case of a deoptimization
3958 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3959
3960 sfpt.push(m);
3961 }
3962 cnt = m->req();
3963 nstack.push(n, i); // put on stack parent and next input's index
3964 n = m;
3965 i = 0;
3966 }
3967 } else {
3968 // Now do post-visit work
3969 final_graph_reshaping_impl(n, frc, dead_nodes);
3970 if (nstack.is_empty())
3971 break; // finished
3972 n = nstack.node(); // Get node from stack
3973 cnt = n->req();
3974 i = nstack.index();
3975 nstack.pop(); // Shift to the next node on stack
3976 }
3977 }
3978
3979 // Skip next transformation if compressed oops are not used.
3980 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3981 (!UseCompressedOops && !UseCompressedClassPointers))
3982 return;
3983
3984 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3985 // It could be done for an uncommon traps or any safepoints/calls
3986 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3987 while (sfpt.size() > 0) {
3988 n = sfpt.pop();
3989 JVMState *jvms = n->as_SafePoint()->jvms();
3990 assert(jvms != nullptr, "sanity");
3991 int start = jvms->debug_start();
3992 int end = n->req();
3993 bool is_uncommon = (n->is_CallStaticJava() &&
3994 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3995 for (int j = start; j < end; j++) {
3996 Node* in = n->in(j);
3997 if (in->is_DecodeNarrowPtr()) {
3998 bool safe_to_skip = true;
3999 if (!is_uncommon ) {
4000 // Is it safe to skip?
4001 for (uint i = 0; i < in->outcnt(); i++) {
4002 Node* u = in->raw_out(i);
4003 if (!u->is_SafePoint() ||
4004 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4005 safe_to_skip = false;
4006 }
4007 }
4008 }
4009 if (safe_to_skip) {
4010 n->set_req(j, in->in(1));
4011 }
4012 if (in->outcnt() == 0) {
4013 in->disconnect_inputs(this);
4014 }
4015 }
4016 }
4017 }
4018 }
4019
4020 //------------------------------final_graph_reshaping--------------------------
4021 // Final Graph Reshaping.
4022 //
4023 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4024 // and not commoned up and forced early. Must come after regular
4025 // optimizations to avoid GVN undoing the cloning. Clone constant
4026 // inputs to Loop Phis; these will be split by the allocator anyways.
4027 // Remove Opaque nodes.
4028 // (2) Move last-uses by commutative operations to the left input to encourage
4029 // Intel update-in-place two-address operations and better register usage
4030 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4031 // calls canonicalizing them back.
4032 // (3) Count the number of double-precision FP ops, single-precision FP ops
4033 // and call sites. On Intel, we can get correct rounding either by
4034 // forcing singles to memory (requires extra stores and loads after each
4035 // FP bytecode) or we can set a rounding mode bit (requires setting and
4036 // clearing the mode bit around call sites). The mode bit is only used
4037 // if the relative frequency of single FP ops to calls is low enough.
4038 // This is a key transform for SPEC mpeg_audio.
4039 // (4) Detect infinite loops; blobs of code reachable from above but not
4040 // below. Several of the Code_Gen algorithms fail on such code shapes,
4041 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4042 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4043 // Detection is by looking for IfNodes where only 1 projection is
4044 // reachable from below or CatchNodes missing some targets.
4045 // (5) Assert for insane oop offsets in debug mode.
4046
4047 bool Compile::final_graph_reshaping() {
4048 // an infinite loop may have been eliminated by the optimizer,
4049 // in which case the graph will be empty.
4050 if (root()->req() == 1) {
4051 // Do not compile method that is only a trivial infinite loop,
4052 // since the content of the loop may have been eliminated.
4053 record_method_not_compilable("trivial infinite loop");
4054 return true;
4055 }
4056
4057 // Expensive nodes have their control input set to prevent the GVN
4058 // from freely commoning them. There's no GVN beyond this point so
4059 // no need to keep the control input. We want the expensive nodes to
4060 // be freely moved to the least frequent code path by gcm.
4061 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4062 for (int i = 0; i < expensive_count(); i++) {
4063 _expensive_nodes.at(i)->set_req(0, nullptr);
4064 }
4065
4066 Final_Reshape_Counts frc;
4067
4068 // Visit everybody reachable!
4069 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4070 Node_Stack nstack(live_nodes() >> 1);
4071 Unique_Node_List dead_nodes;
4072 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4073
4074 // Check for unreachable (from below) code (i.e., infinite loops).
4075 for( uint i = 0; i < frc._tests.size(); i++ ) {
4076 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4077 // Get number of CFG targets.
4078 // Note that PCTables include exception targets after calls.
4079 uint required_outcnt = n->required_outcnt();
4080 if (n->outcnt() != required_outcnt) {
4081 // Check for a few special cases. Rethrow Nodes never take the
4082 // 'fall-thru' path, so expected kids is 1 less.
4083 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4084 if (n->in(0)->in(0)->is_Call()) {
4085 CallNode* call = n->in(0)->in(0)->as_Call();
4086 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4087 required_outcnt--; // Rethrow always has 1 less kid
4088 } else if (call->req() > TypeFunc::Parms &&
4089 call->is_CallDynamicJava()) {
4090 // Check for null receiver. In such case, the optimizer has
4091 // detected that the virtual call will always result in a null
4092 // pointer exception. The fall-through projection of this CatchNode
4093 // will not be populated.
4094 Node* arg0 = call->in(TypeFunc::Parms);
4095 if (arg0->is_Type() &&
4096 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4097 required_outcnt--;
4098 }
4099 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4100 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4101 // Check for illegal array length. In such case, the optimizer has
4102 // detected that the allocation attempt will always result in an
4103 // exception. There is no fall-through projection of this CatchNode .
4104 assert(call->is_CallStaticJava(), "static call expected");
4105 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4106 uint valid_length_test_input = call->req() - 1;
4107 Node* valid_length_test = call->in(valid_length_test_input);
4108 call->del_req(valid_length_test_input);
4109 if (valid_length_test->find_int_con(1) == 0) {
4110 required_outcnt--;
4111 }
4112 dead_nodes.push(valid_length_test);
4113 assert(n->outcnt() == required_outcnt, "malformed control flow");
4114 continue;
4115 }
4116 }
4117 }
4118
4119 // Recheck with a better notion of 'required_outcnt'
4120 if (n->outcnt() != required_outcnt) {
4121 record_method_not_compilable("malformed control flow");
4122 return true; // Not all targets reachable!
4123 }
4124 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4125 CallNode* call = n->in(0)->in(0)->as_Call();
4126 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4127 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4128 assert(call->is_CallStaticJava(), "static call expected");
4129 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4130 uint valid_length_test_input = call->req() - 1;
4131 dead_nodes.push(call->in(valid_length_test_input));
4132 call->del_req(valid_length_test_input); // valid length test useless now
4133 }
4134 }
4135 // Check that I actually visited all kids. Unreached kids
4136 // must be infinite loops.
4137 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4138 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4139 record_method_not_compilable("infinite loop");
4140 return true; // Found unvisited kid; must be unreach
4141 }
4142
4143 // Here so verification code in final_graph_reshaping_walk()
4144 // always see an OuterStripMinedLoopEnd
4145 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4146 IfNode* init_iff = n->as_If();
4147 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4148 n->subsume_by(iff, this);
4149 }
4150 }
4151
4152 while (dead_nodes.size() > 0) {
4153 Node* m = dead_nodes.pop();
4154 if (m->outcnt() == 0 && m != top()) {
4155 for (uint j = 0; j < m->req(); j++) {
4156 Node* in = m->in(j);
4157 if (in != nullptr) {
4158 dead_nodes.push(in);
4159 }
4160 }
4161 m->disconnect_inputs(this);
4162 }
4163 }
4164
4165 set_java_calls(frc.get_java_call_count());
4166 set_inner_loops(frc.get_inner_loop_count());
4167
4168 // No infinite loops, no reason to bail out.
4169 return false;
4170 }
4171
4172 //-----------------------------too_many_traps----------------------------------
4173 // Report if there are too many traps at the current method and bci.
4174 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4175 bool Compile::too_many_traps(ciMethod* method,
4176 int bci,
4177 Deoptimization::DeoptReason reason) {
4178 ciMethodData* md = method->method_data();
4179 if (md->is_empty()) {
4180 // Assume the trap has not occurred, or that it occurred only
4181 // because of a transient condition during start-up in the interpreter.
4182 return false;
4183 }
4184 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4185 if (md->has_trap_at(bci, m, reason) != 0) {
4186 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4187 // Also, if there are multiple reasons, or if there is no per-BCI record,
4188 // assume the worst.
4189 if (log())
4190 log()->elem("observe trap='%s' count='%d'",
4191 Deoptimization::trap_reason_name(reason),
4192 md->trap_count(reason));
4193 return true;
4194 } else {
4195 // Ignore method/bci and see if there have been too many globally.
4196 return too_many_traps(reason, md);
4197 }
4198 }
4199
4200 // Less-accurate variant which does not require a method and bci.
4201 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4202 ciMethodData* logmd) {
4203 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4204 // Too many traps globally.
4205 // Note that we use cumulative trap_count, not just md->trap_count.
4206 if (log()) {
4207 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4208 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4209 Deoptimization::trap_reason_name(reason),
4210 mcount, trap_count(reason));
4211 }
4212 return true;
4213 } else {
4214 // The coast is clear.
4215 return false;
4216 }
4217 }
4218
4219 //--------------------------too_many_recompiles--------------------------------
4220 // Report if there are too many recompiles at the current method and bci.
4221 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4222 // Is not eager to return true, since this will cause the compiler to use
4223 // Action_none for a trap point, to avoid too many recompilations.
4224 bool Compile::too_many_recompiles(ciMethod* method,
4225 int bci,
4226 Deoptimization::DeoptReason reason) {
4227 ciMethodData* md = method->method_data();
4228 if (md->is_empty()) {
4229 // Assume the trap has not occurred, or that it occurred only
4230 // because of a transient condition during start-up in the interpreter.
4231 return false;
4232 }
4233 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4234 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4235 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4236 Deoptimization::DeoptReason per_bc_reason
4237 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4238 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4239 if ((per_bc_reason == Deoptimization::Reason_none
4240 || md->has_trap_at(bci, m, reason) != 0)
4241 // The trap frequency measure we care about is the recompile count:
4242 && md->trap_recompiled_at(bci, m)
4243 && md->overflow_recompile_count() >= bc_cutoff) {
4244 // Do not emit a trap here if it has already caused recompilations.
4245 // Also, if there are multiple reasons, or if there is no per-BCI record,
4246 // assume the worst.
4247 if (log())
4248 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4249 Deoptimization::trap_reason_name(reason),
4250 md->trap_count(reason),
4251 md->overflow_recompile_count());
4252 return true;
4253 } else if (trap_count(reason) != 0
4254 && decompile_count() >= m_cutoff) {
4255 // Too many recompiles globally, and we have seen this sort of trap.
4256 // Use cumulative decompile_count, not just md->decompile_count.
4257 if (log())
4258 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4259 Deoptimization::trap_reason_name(reason),
4260 md->trap_count(reason), trap_count(reason),
4261 md->decompile_count(), decompile_count());
4262 return true;
4263 } else {
4264 // The coast is clear.
4265 return false;
4266 }
4267 }
4268
4269 // Compute when not to trap. Used by matching trap based nodes and
4270 // NullCheck optimization.
4271 void Compile::set_allowed_deopt_reasons() {
4272 _allowed_reasons = 0;
4273 if (is_method_compilation()) {
4274 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4275 assert(rs < BitsPerInt, "recode bit map");
4276 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4277 _allowed_reasons |= nth_bit(rs);
4278 }
4279 }
4280 }
4281 }
4282
4283 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4284 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4285 }
4286
4287 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4288 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4289 }
4290
4291 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4292 if (holder->is_initialized()) {
4293 return false;
4294 }
4295 if (holder->is_being_initialized()) {
4296 if (accessing_method->holder() == holder) {
4297 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4298 // <init>, or a static method. In all those cases, there was an initialization
4299 // barrier on the holder klass passed.
4300 if (accessing_method->is_static_initializer() ||
4301 accessing_method->is_object_initializer() ||
4302 accessing_method->is_static()) {
4303 return false;
4304 }
4305 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4306 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4307 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4308 // child class can become fully initialized while its parent class is still being initialized.
4309 if (accessing_method->is_static_initializer()) {
4310 return false;
4311 }
4312 }
4313 ciMethod* root = method(); // the root method of compilation
4314 if (root != accessing_method) {
4315 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4316 }
4317 }
4318 return true;
4319 }
4320
4321 #ifndef PRODUCT
4322 //------------------------------verify_bidirectional_edges---------------------
4323 // For each input edge to a node (ie - for each Use-Def edge), verify that
4324 // there is a corresponding Def-Use edge.
4325 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4326 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4327 uint stack_size = live_nodes() >> 4;
4328 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4329 if (root_and_safepoints != nullptr) {
4330 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4331 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4332 Node* root_or_safepoint = root_and_safepoints->at(i);
4333 // If the node is a safepoint, let's check if it still has a control input
4334 // Lack of control input signifies that this node was killed by CCP or
4335 // recursively by remove_globally_dead_node and it shouldn't be a starting
4336 // point.
4337 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4338 nstack.push(root_or_safepoint);
4339 }
4340 }
4341 } else {
4342 nstack.push(_root);
4343 }
4344
4345 while (nstack.size() > 0) {
4346 Node* n = nstack.pop();
4347 if (visited.member(n)) {
4348 continue;
4349 }
4350 visited.push(n);
4351
4352 // Walk over all input edges, checking for correspondence
4353 uint length = n->len();
4354 for (uint i = 0; i < length; i++) {
4355 Node* in = n->in(i);
4356 if (in != nullptr && !visited.member(in)) {
4357 nstack.push(in); // Put it on stack
4358 }
4359 if (in != nullptr && !in->is_top()) {
4360 // Count instances of `next`
4361 int cnt = 0;
4362 for (uint idx = 0; idx < in->_outcnt; idx++) {
4363 if (in->_out[idx] == n) {
4364 cnt++;
4365 }
4366 }
4367 assert(cnt > 0, "Failed to find Def-Use edge.");
4368 // Check for duplicate edges
4369 // walk the input array downcounting the input edges to n
4370 for (uint j = 0; j < length; j++) {
4371 if (n->in(j) == in) {
4372 cnt--;
4373 }
4374 }
4375 assert(cnt == 0, "Mismatched edge count.");
4376 } else if (in == nullptr) {
4377 assert(i == 0 || i >= n->req() ||
4378 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4379 (n->is_Unlock() && i == (n->req() - 1)) ||
4380 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4381 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4382 } else {
4383 assert(in->is_top(), "sanity");
4384 // Nothing to check.
4385 }
4386 }
4387 }
4388 }
4389
4390 //------------------------------verify_graph_edges---------------------------
4391 // Walk the Graph and verify that there is a one-to-one correspondence
4392 // between Use-Def edges and Def-Use edges in the graph.
4393 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4394 if (VerifyGraphEdges) {
4395 Unique_Node_List visited;
4396
4397 // Call graph walk to check edges
4398 verify_bidirectional_edges(visited, root_and_safepoints);
4399 if (no_dead_code) {
4400 // Now make sure that no visited node is used by an unvisited node.
4401 bool dead_nodes = false;
4402 Unique_Node_List checked;
4403 while (visited.size() > 0) {
4404 Node* n = visited.pop();
4405 checked.push(n);
4406 for (uint i = 0; i < n->outcnt(); i++) {
4407 Node* use = n->raw_out(i);
4408 if (checked.member(use)) continue; // already checked
4409 if (visited.member(use)) continue; // already in the graph
4410 if (use->is_Con()) continue; // a dead ConNode is OK
4411 // At this point, we have found a dead node which is DU-reachable.
4412 if (!dead_nodes) {
4413 tty->print_cr("*** Dead nodes reachable via DU edges:");
4414 dead_nodes = true;
4415 }
4416 use->dump(2);
4417 tty->print_cr("---");
4418 checked.push(use); // No repeats; pretend it is now checked.
4419 }
4420 }
4421 assert(!dead_nodes, "using nodes must be reachable from root");
4422 }
4423 }
4424 }
4425 #endif
4426
4427 // The Compile object keeps track of failure reasons separately from the ciEnv.
4428 // This is required because there is not quite a 1-1 relation between the
4429 // ciEnv and its compilation task and the Compile object. Note that one
4430 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4431 // to backtrack and retry without subsuming loads. Other than this backtracking
4432 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4433 // by the logic in C2Compiler.
4434 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4435 if (log() != nullptr) {
4436 log()->elem("failure reason='%s' phase='compile'", reason);
4437 }
4438 if (_failure_reason.get() == nullptr) {
4439 // Record the first failure reason.
4440 _failure_reason.set(reason);
4441 if (CaptureBailoutInformation) {
4442 _first_failure_details = new CompilationFailureInfo(reason);
4443 }
4444 } else {
4445 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4446 }
4447
4448 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4449 C->print_method(PHASE_FAILURE, 1);
4450 }
4451 _root = nullptr; // flush the graph, too
4452 }
4453
4454 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4455 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4456 _compile(Compile::current()),
4457 _log(nullptr),
4458 _dolog(CITimeVerbose)
4459 {
4460 assert(_compile != nullptr, "sanity check");
4461 assert(id != PhaseTraceId::_t_none, "Don't use none");
4462 if (_dolog) {
4463 _log = _compile->log();
4464 }
4465 if (_log != nullptr) {
4466 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4467 _log->stamp();
4468 _log->end_head();
4469 }
4470
4471 // Inform memory statistic, if enabled
4472 if (CompilationMemoryStatistic::enabled()) {
4473 CompilationMemoryStatistic::on_phase_start((int)id, name);
4474 }
4475 }
4476
4477 Compile::TracePhase::TracePhase(PhaseTraceId id)
4478 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4479
4480 Compile::TracePhase::~TracePhase() {
4481
4482 // Inform memory statistic, if enabled
4483 if (CompilationMemoryStatistic::enabled()) {
4484 CompilationMemoryStatistic::on_phase_end();
4485 }
4486
4487 if (_compile->failing_internal()) {
4488 if (_log != nullptr) {
4489 _log->done("phase");
4490 }
4491 return; // timing code, not stressing bailouts.
4492 }
4493 #ifdef ASSERT
4494 if (PrintIdealNodeCount) {
4495 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4496 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4497 }
4498
4499 if (VerifyIdealNodeCount) {
4500 _compile->print_missing_nodes();
4501 }
4502 #endif
4503
4504 if (_log != nullptr) {
4505 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4506 }
4507 }
4508
4509 //----------------------------static_subtype_check-----------------------------
4510 // Shortcut important common cases when superklass is exact:
4511 // (0) superklass is java.lang.Object (can occur in reflective code)
4512 // (1) subklass is already limited to a subtype of superklass => always ok
4513 // (2) subklass does not overlap with superklass => always fail
4514 // (3) superklass has NO subtypes and we can check with a simple compare.
4515 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4516 if (skip) {
4517 return SSC_full_test; // Let caller generate the general case.
4518 }
4519
4520 if (subk->is_java_subtype_of(superk)) {
4521 return SSC_always_true; // (0) and (1) this test cannot fail
4522 }
4523
4524 if (!subk->maybe_java_subtype_of(superk)) {
4525 return SSC_always_false; // (2) true path dead; no dynamic test needed
4526 }
4527
4528 const Type* superelem = superk;
4529 if (superk->isa_aryklassptr()) {
4530 int ignored;
4531 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4532 }
4533
4534 if (superelem->isa_instklassptr()) {
4535 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4536 if (!ik->has_subklass()) {
4537 if (!ik->is_final()) {
4538 // Add a dependency if there is a chance of a later subclass.
4539 dependencies()->assert_leaf_type(ik);
4540 }
4541 if (!superk->maybe_java_subtype_of(subk)) {
4542 return SSC_always_false;
4543 }
4544 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4545 }
4546 } else {
4547 // A primitive array type has no subtypes.
4548 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4549 }
4550
4551 return SSC_full_test;
4552 }
4553
4554 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4555 #ifdef _LP64
4556 // The scaled index operand to AddP must be a clean 64-bit value.
4557 // Java allows a 32-bit int to be incremented to a negative
4558 // value, which appears in a 64-bit register as a large
4559 // positive number. Using that large positive number as an
4560 // operand in pointer arithmetic has bad consequences.
4561 // On the other hand, 32-bit overflow is rare, and the possibility
4562 // can often be excluded, if we annotate the ConvI2L node with
4563 // a type assertion that its value is known to be a small positive
4564 // number. (The prior range check has ensured this.)
4565 // This assertion is used by ConvI2LNode::Ideal.
4566 int index_max = max_jint - 1; // array size is max_jint, index is one less
4567 if (sizetype != nullptr && sizetype->_hi > 0) {
4568 index_max = sizetype->_hi - 1;
4569 }
4570 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4571 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4572 #endif
4573 return idx;
4574 }
4575
4576 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4577 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4578 if (ctrl != nullptr) {
4579 // Express control dependency by a CastII node with a narrow type.
4580 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4581 // node from floating above the range check during loop optimizations. Otherwise, the
4582 // ConvI2L node may be eliminated independently of the range check, causing the data path
4583 // to become TOP while the control path is still there (although it's unreachable).
4584 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
4585 value = phase->transform(value);
4586 }
4587 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4588 return phase->transform(new ConvI2LNode(value, ltype));
4589 }
4590
4591 void Compile::dump_print_inlining() {
4592 inline_printer()->print_on(tty);
4593 }
4594
4595 void Compile::log_late_inline(CallGenerator* cg) {
4596 if (log() != nullptr) {
4597 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4598 cg->unique_id());
4599 JVMState* p = cg->call_node()->jvms();
4600 while (p != nullptr) {
4601 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4602 p = p->caller();
4603 }
4604 log()->tail("late_inline");
4605 }
4606 }
4607
4608 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4609 log_late_inline(cg);
4610 if (log() != nullptr) {
4611 log()->inline_fail(msg);
4612 }
4613 }
4614
4615 void Compile::log_inline_id(CallGenerator* cg) {
4616 if (log() != nullptr) {
4617 // The LogCompilation tool needs a unique way to identify late
4618 // inline call sites. This id must be unique for this call site in
4619 // this compilation. Try to have it unique across compilations as
4620 // well because it can be convenient when grepping through the log
4621 // file.
4622 // Distinguish OSR compilations from others in case CICountOSR is
4623 // on.
4624 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4625 cg->set_unique_id(id);
4626 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4627 }
4628 }
4629
4630 void Compile::log_inline_failure(const char* msg) {
4631 if (C->log() != nullptr) {
4632 C->log()->inline_fail(msg);
4633 }
4634 }
4635
4636
4637 // Dump inlining replay data to the stream.
4638 // Don't change thread state and acquire any locks.
4639 void Compile::dump_inline_data(outputStream* out) {
4640 InlineTree* inl_tree = ilt();
4641 if (inl_tree != nullptr) {
4642 out->print(" inline %d", inl_tree->count());
4643 inl_tree->dump_replay_data(out);
4644 }
4645 }
4646
4647 void Compile::dump_inline_data_reduced(outputStream* out) {
4648 assert(ReplayReduce, "");
4649
4650 InlineTree* inl_tree = ilt();
4651 if (inl_tree == nullptr) {
4652 return;
4653 }
4654 // Enable iterative replay file reduction
4655 // Output "compile" lines for depth 1 subtrees,
4656 // simulating that those trees were compiled
4657 // instead of inlined.
4658 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4659 InlineTree* sub = inl_tree->subtrees().at(i);
4660 if (sub->inline_level() != 1) {
4661 continue;
4662 }
4663
4664 ciMethod* method = sub->method();
4665 int entry_bci = -1;
4666 int comp_level = env()->task()->comp_level();
4667 out->print("compile ");
4668 method->dump_name_as_ascii(out);
4669 out->print(" %d %d", entry_bci, comp_level);
4670 out->print(" inline %d", sub->count());
4671 sub->dump_replay_data(out, -1);
4672 out->cr();
4673 }
4674 }
4675
4676 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4677 if (n1->Opcode() < n2->Opcode()) return -1;
4678 else if (n1->Opcode() > n2->Opcode()) return 1;
4679
4680 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4681 for (uint i = 1; i < n1->req(); i++) {
4682 if (n1->in(i) < n2->in(i)) return -1;
4683 else if (n1->in(i) > n2->in(i)) return 1;
4684 }
4685
4686 return 0;
4687 }
4688
4689 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4690 Node* n1 = *n1p;
4691 Node* n2 = *n2p;
4692
4693 return cmp_expensive_nodes(n1, n2);
4694 }
4695
4696 void Compile::sort_expensive_nodes() {
4697 if (!expensive_nodes_sorted()) {
4698 _expensive_nodes.sort(cmp_expensive_nodes);
4699 }
4700 }
4701
4702 bool Compile::expensive_nodes_sorted() const {
4703 for (int i = 1; i < _expensive_nodes.length(); i++) {
4704 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4705 return false;
4706 }
4707 }
4708 return true;
4709 }
4710
4711 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4712 if (_expensive_nodes.length() == 0) {
4713 return false;
4714 }
4715
4716 assert(OptimizeExpensiveOps, "optimization off?");
4717
4718 // Take this opportunity to remove dead nodes from the list
4719 int j = 0;
4720 for (int i = 0; i < _expensive_nodes.length(); i++) {
4721 Node* n = _expensive_nodes.at(i);
4722 if (!n->is_unreachable(igvn)) {
4723 assert(n->is_expensive(), "should be expensive");
4724 _expensive_nodes.at_put(j, n);
4725 j++;
4726 }
4727 }
4728 _expensive_nodes.trunc_to(j);
4729
4730 // Then sort the list so that similar nodes are next to each other
4731 // and check for at least two nodes of identical kind with same data
4732 // inputs.
4733 sort_expensive_nodes();
4734
4735 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4736 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4737 return true;
4738 }
4739 }
4740
4741 return false;
4742 }
4743
4744 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4745 if (_expensive_nodes.length() == 0) {
4746 return;
4747 }
4748
4749 assert(OptimizeExpensiveOps, "optimization off?");
4750
4751 // Sort to bring similar nodes next to each other and clear the
4752 // control input of nodes for which there's only a single copy.
4753 sort_expensive_nodes();
4754
4755 int j = 0;
4756 int identical = 0;
4757 int i = 0;
4758 bool modified = false;
4759 for (; i < _expensive_nodes.length()-1; i++) {
4760 assert(j <= i, "can't write beyond current index");
4761 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4762 identical++;
4763 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4764 continue;
4765 }
4766 if (identical > 0) {
4767 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4768 identical = 0;
4769 } else {
4770 Node* n = _expensive_nodes.at(i);
4771 igvn.replace_input_of(n, 0, nullptr);
4772 igvn.hash_insert(n);
4773 modified = true;
4774 }
4775 }
4776 if (identical > 0) {
4777 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4778 } else if (_expensive_nodes.length() >= 1) {
4779 Node* n = _expensive_nodes.at(i);
4780 igvn.replace_input_of(n, 0, nullptr);
4781 igvn.hash_insert(n);
4782 modified = true;
4783 }
4784 _expensive_nodes.trunc_to(j);
4785 if (modified) {
4786 igvn.optimize();
4787 }
4788 }
4789
4790 void Compile::add_expensive_node(Node * n) {
4791 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4792 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4793 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4794 if (OptimizeExpensiveOps) {
4795 _expensive_nodes.append(n);
4796 } else {
4797 // Clear control input and let IGVN optimize expensive nodes if
4798 // OptimizeExpensiveOps is off.
4799 n->set_req(0, nullptr);
4800 }
4801 }
4802
4803 /**
4804 * Track coarsened Lock and Unlock nodes.
4805 */
4806
4807 class Lock_List : public Node_List {
4808 uint _origin_cnt;
4809 public:
4810 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4811 uint origin_cnt() const { return _origin_cnt; }
4812 };
4813
4814 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4815 int length = locks.length();
4816 if (length > 0) {
4817 // Have to keep this list until locks elimination during Macro nodes elimination.
4818 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4819 AbstractLockNode* alock = locks.at(0);
4820 BoxLockNode* box = alock->box_node()->as_BoxLock();
4821 for (int i = 0; i < length; i++) {
4822 AbstractLockNode* lock = locks.at(i);
4823 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4824 locks_list->push(lock);
4825 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4826 if (this_box != box) {
4827 // Locking regions (BoxLock) could be Unbalanced here:
4828 // - its coarsened locks were eliminated in earlier
4829 // macro nodes elimination followed by loop unroll
4830 // - it is OSR locking region (no Lock node)
4831 // Preserve Unbalanced status in such cases.
4832 if (!this_box->is_unbalanced()) {
4833 this_box->set_coarsened();
4834 }
4835 if (!box->is_unbalanced()) {
4836 box->set_coarsened();
4837 }
4838 }
4839 }
4840 _coarsened_locks.append(locks_list);
4841 }
4842 }
4843
4844 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4845 int count = coarsened_count();
4846 for (int i = 0; i < count; i++) {
4847 Node_List* locks_list = _coarsened_locks.at(i);
4848 for (uint j = 0; j < locks_list->size(); j++) {
4849 Node* lock = locks_list->at(j);
4850 assert(lock->is_AbstractLock(), "sanity");
4851 if (!useful.member(lock)) {
4852 locks_list->yank(lock);
4853 }
4854 }
4855 }
4856 }
4857
4858 void Compile::remove_coarsened_lock(Node* n) {
4859 if (n->is_AbstractLock()) {
4860 int count = coarsened_count();
4861 for (int i = 0; i < count; i++) {
4862 Node_List* locks_list = _coarsened_locks.at(i);
4863 locks_list->yank(n);
4864 }
4865 }
4866 }
4867
4868 bool Compile::coarsened_locks_consistent() {
4869 int count = coarsened_count();
4870 for (int i = 0; i < count; i++) {
4871 bool unbalanced = false;
4872 bool modified = false; // track locks kind modifications
4873 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4874 uint size = locks_list->size();
4875 if (size == 0) {
4876 unbalanced = false; // All locks were eliminated - good
4877 } else if (size != locks_list->origin_cnt()) {
4878 unbalanced = true; // Some locks were removed from list
4879 } else {
4880 for (uint j = 0; j < size; j++) {
4881 Node* lock = locks_list->at(j);
4882 // All nodes in group should have the same state (modified or not)
4883 if (!lock->as_AbstractLock()->is_coarsened()) {
4884 if (j == 0) {
4885 // first on list was modified, the rest should be too for consistency
4886 modified = true;
4887 } else if (!modified) {
4888 // this lock was modified but previous locks on the list were not
4889 unbalanced = true;
4890 break;
4891 }
4892 } else if (modified) {
4893 // previous locks on list were modified but not this lock
4894 unbalanced = true;
4895 break;
4896 }
4897 }
4898 }
4899 if (unbalanced) {
4900 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4901 #ifdef ASSERT
4902 if (PrintEliminateLocks) {
4903 tty->print_cr("=== unbalanced coarsened locks ===");
4904 for (uint l = 0; l < size; l++) {
4905 locks_list->at(l)->dump();
4906 }
4907 }
4908 #endif
4909 record_failure(C2Compiler::retry_no_locks_coarsening());
4910 return false;
4911 }
4912 }
4913 return true;
4914 }
4915
4916 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4917 // locks coarsening optimization removed Lock/Unlock nodes from them.
4918 // Such regions become unbalanced because coarsening only removes part
4919 // of Lock/Unlock nodes in region. As result we can't execute other
4920 // locks elimination optimizations which assume all code paths have
4921 // corresponding pair of Lock/Unlock nodes - they are balanced.
4922 void Compile::mark_unbalanced_boxes() const {
4923 int count = coarsened_count();
4924 for (int i = 0; i < count; i++) {
4925 Node_List* locks_list = _coarsened_locks.at(i);
4926 uint size = locks_list->size();
4927 if (size > 0) {
4928 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4929 BoxLockNode* box = alock->box_node()->as_BoxLock();
4930 if (alock->is_coarsened()) {
4931 // coarsened_locks_consistent(), which is called before this method, verifies
4932 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4933 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4934 for (uint j = 1; j < size; j++) {
4935 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4936 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4937 if (box != this_box) {
4938 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4939 box->set_unbalanced();
4940 this_box->set_unbalanced();
4941 }
4942 }
4943 }
4944 }
4945 }
4946 }
4947
4948 /**
4949 * Remove the speculative part of types and clean up the graph
4950 */
4951 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4952 if (UseTypeSpeculation) {
4953 Unique_Node_List worklist;
4954 worklist.push(root());
4955 int modified = 0;
4956 // Go over all type nodes that carry a speculative type, drop the
4957 // speculative part of the type and enqueue the node for an igvn
4958 // which may optimize it out.
4959 for (uint next = 0; next < worklist.size(); ++next) {
4960 Node *n = worklist.at(next);
4961 if (n->is_Type()) {
4962 TypeNode* tn = n->as_Type();
4963 const Type* t = tn->type();
4964 const Type* t_no_spec = t->remove_speculative();
4965 if (t_no_spec != t) {
4966 bool in_hash = igvn.hash_delete(n);
4967 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4968 tn->set_type(t_no_spec);
4969 igvn.hash_insert(n);
4970 igvn._worklist.push(n); // give it a chance to go away
4971 modified++;
4972 }
4973 }
4974 // Iterate over outs - endless loops is unreachable from below
4975 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4976 Node *m = n->fast_out(i);
4977 if (not_a_node(m)) {
4978 continue;
4979 }
4980 worklist.push(m);
4981 }
4982 }
4983 // Drop the speculative part of all types in the igvn's type table
4984 igvn.remove_speculative_types();
4985 if (modified > 0) {
4986 igvn.optimize();
4987 if (failing()) return;
4988 }
4989 #ifdef ASSERT
4990 // Verify that after the IGVN is over no speculative type has resurfaced
4991 worklist.clear();
4992 worklist.push(root());
4993 for (uint next = 0; next < worklist.size(); ++next) {
4994 Node *n = worklist.at(next);
4995 const Type* t = igvn.type_or_null(n);
4996 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4997 if (n->is_Type()) {
4998 t = n->as_Type()->type();
4999 assert(t == t->remove_speculative(), "no more speculative types");
5000 }
5001 // Iterate over outs - endless loops is unreachable from below
5002 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5003 Node *m = n->fast_out(i);
5004 if (not_a_node(m)) {
5005 continue;
5006 }
5007 worklist.push(m);
5008 }
5009 }
5010 igvn.check_no_speculative_types();
5011 #endif
5012 }
5013 }
5014
5015 // Auxiliary methods to support randomized stressing/fuzzing.
5016
5017 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5018 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5019 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5020 FLAG_SET_ERGO(StressSeed, _stress_seed);
5021 } else {
5022 _stress_seed = StressSeed;
5023 }
5024 if (_log != nullptr) {
5025 _log->elem("stress_test seed='%u'", _stress_seed);
5026 }
5027 }
5028
5029 int Compile::random() {
5030 _stress_seed = os::next_random(_stress_seed);
5031 return static_cast<int>(_stress_seed);
5032 }
5033
5034 // This method can be called the arbitrary number of times, with current count
5035 // as the argument. The logic allows selecting a single candidate from the
5036 // running list of candidates as follows:
5037 // int count = 0;
5038 // Cand* selected = null;
5039 // while(cand = cand->next()) {
5040 // if (randomized_select(++count)) {
5041 // selected = cand;
5042 // }
5043 // }
5044 //
5045 // Including count equalizes the chances any candidate is "selected".
5046 // This is useful when we don't have the complete list of candidates to choose
5047 // from uniformly. In this case, we need to adjust the randomicity of the
5048 // selection, or else we will end up biasing the selection towards the latter
5049 // candidates.
5050 //
5051 // Quick back-envelope calculation shows that for the list of n candidates
5052 // the equal probability for the candidate to persist as "best" can be
5053 // achieved by replacing it with "next" k-th candidate with the probability
5054 // of 1/k. It can be easily shown that by the end of the run, the
5055 // probability for any candidate is converged to 1/n, thus giving the
5056 // uniform distribution among all the candidates.
5057 //
5058 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5059 #define RANDOMIZED_DOMAIN_POW 29
5060 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5061 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5062 bool Compile::randomized_select(int count) {
5063 assert(count > 0, "only positive");
5064 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5065 }
5066
5067 #ifdef ASSERT
5068 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5069 bool Compile::fail_randomly() {
5070 if ((random() % StressBailoutMean) != 0) {
5071 return false;
5072 }
5073 record_failure("StressBailout");
5074 return true;
5075 }
5076
5077 bool Compile::failure_is_artificial() {
5078 return C->failure_reason_is("StressBailout");
5079 }
5080 #endif
5081
5082 CloneMap& Compile::clone_map() { return _clone_map; }
5083 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5084
5085 void NodeCloneInfo::dump_on(outputStream* st) const {
5086 st->print(" {%d:%d} ", idx(), gen());
5087 }
5088
5089 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5090 uint64_t val = value(old->_idx);
5091 NodeCloneInfo cio(val);
5092 assert(val != 0, "old node should be in the map");
5093 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5094 insert(nnn->_idx, cin.get());
5095 #ifndef PRODUCT
5096 if (is_debug()) {
5097 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5098 }
5099 #endif
5100 }
5101
5102 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5103 NodeCloneInfo cio(value(old->_idx));
5104 if (cio.get() == 0) {
5105 cio.set(old->_idx, 0);
5106 insert(old->_idx, cio.get());
5107 #ifndef PRODUCT
5108 if (is_debug()) {
5109 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5110 }
5111 #endif
5112 }
5113 clone(old, nnn, gen);
5114 }
5115
5116 int CloneMap::max_gen() const {
5117 int g = 0;
5118 DictI di(_dict);
5119 for(; di.test(); ++di) {
5120 int t = gen(di._key);
5121 if (g < t) {
5122 g = t;
5123 #ifndef PRODUCT
5124 if (is_debug()) {
5125 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5126 }
5127 #endif
5128 }
5129 }
5130 return g;
5131 }
5132
5133 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5134 uint64_t val = value(key);
5135 if (val != 0) {
5136 NodeCloneInfo ni(val);
5137 ni.dump_on(st);
5138 }
5139 }
5140
5141 void Compile::shuffle_macro_nodes() {
5142 if (_macro_nodes.length() < 2) {
5143 return;
5144 }
5145 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5146 uint j = C->random() % (i + 1);
5147 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5148 }
5149 }
5150
5151 // Move Allocate nodes to the start of the list
5152 void Compile::sort_macro_nodes() {
5153 int count = macro_count();
5154 int allocates = 0;
5155 for (int i = 0; i < count; i++) {
5156 Node* n = macro_node(i);
5157 if (n->is_Allocate()) {
5158 if (i != allocates) {
5159 Node* tmp = macro_node(allocates);
5160 _macro_nodes.at_put(allocates, n);
5161 _macro_nodes.at_put(i, tmp);
5162 }
5163 allocates++;
5164 }
5165 }
5166 }
5167
5168 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5169 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5170 EventCompilerPhase event(UNTIMED);
5171 if (event.should_commit()) {
5172 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5173 }
5174 #ifndef PRODUCT
5175 ResourceMark rm;
5176 stringStream ss;
5177 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5178 int iter = ++_igv_phase_iter[cpt];
5179 if (iter > 1) {
5180 ss.print(" %d", iter);
5181 }
5182 if (n != nullptr) {
5183 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5184 if (n->is_Call()) {
5185 CallNode* call = n->as_Call();
5186 if (call->_name != nullptr) {
5187 // E.g. uncommon traps etc.
5188 ss.print(" - %s", call->_name);
5189 } else if (call->is_CallJava()) {
5190 CallJavaNode* call_java = call->as_CallJava();
5191 if (call_java->method() != nullptr) {
5192 ss.print(" -");
5193 call_java->method()->print_short_name(&ss);
5194 }
5195 }
5196 }
5197 }
5198
5199 const char* name = ss.as_string();
5200 if (should_print_igv(level)) {
5201 _igv_printer->print_graph(name);
5202 }
5203 if (should_print_phase(level)) {
5204 print_phase(name);
5205 }
5206 if (should_print_ideal_phase(cpt)) {
5207 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5208 }
5209 #endif
5210 C->_latest_stage_start_counter.stamp();
5211 }
5212
5213 // Only used from CompileWrapper
5214 void Compile::begin_method() {
5215 #ifndef PRODUCT
5216 if (_method != nullptr && should_print_igv(1)) {
5217 _igv_printer->begin_method();
5218 }
5219 #endif
5220 C->_latest_stage_start_counter.stamp();
5221 }
5222
5223 // Only used from CompileWrapper
5224 void Compile::end_method() {
5225 EventCompilerPhase event(UNTIMED);
5226 if (event.should_commit()) {
5227 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5228 }
5229
5230 #ifndef PRODUCT
5231 if (_method != nullptr && should_print_igv(1)) {
5232 _igv_printer->end_method();
5233 }
5234 #endif
5235 }
5236
5237 #ifndef PRODUCT
5238 bool Compile::should_print_phase(const int level) const {
5239 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5240 _method != nullptr; // Do not print phases for stubs.
5241 }
5242
5243 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5244 return _directive->should_print_ideal_phase(cpt);
5245 }
5246
5247 void Compile::init_igv() {
5248 if (_igv_printer == nullptr) {
5249 _igv_printer = IdealGraphPrinter::printer();
5250 _igv_printer->set_compile(this);
5251 }
5252 }
5253
5254 bool Compile::should_print_igv(const int level) {
5255 PRODUCT_RETURN_(return false;);
5256
5257 if (PrintIdealGraphLevel < 0) { // disabled by the user
5258 return false;
5259 }
5260
5261 bool need = directive()->IGVPrintLevelOption >= level;
5262 if (need) {
5263 Compile::init_igv();
5264 }
5265 return need;
5266 }
5267
5268 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5269 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5270
5271 // Called from debugger. Prints method to the default file with the default phase name.
5272 // This works regardless of any Ideal Graph Visualizer flags set or not.
5273 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5274 void igv_print(void* sp, void* fp, void* pc) {
5275 frame fr(sp, fp, pc);
5276 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5277 }
5278
5279 // Same as igv_print() above but with a specified phase name.
5280 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5281 frame fr(sp, fp, pc);
5282 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5283 }
5284
5285 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5286 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5287 // This works regardless of any Ideal Graph Visualizer flags set or not.
5288 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5289 void igv_print(bool network, void* sp, void* fp, void* pc) {
5290 frame fr(sp, fp, pc);
5291 if (network) {
5292 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5293 } else {
5294 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5295 }
5296 }
5297
5298 // Same as igv_print(bool network, ...) above but with a specified phase name.
5299 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5300 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5301 frame fr(sp, fp, pc);
5302 if (network) {
5303 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5304 } else {
5305 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5306 }
5307 }
5308
5309 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5310 void igv_print_default() {
5311 Compile::current()->print_method(PHASE_DEBUG, 0);
5312 }
5313
5314 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5315 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5316 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5317 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5318 void igv_append(void* sp, void* fp, void* pc) {
5319 frame fr(sp, fp, pc);
5320 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5321 }
5322
5323 // Same as igv_append(...) above but with a specified phase name.
5324 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5325 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5326 frame fr(sp, fp, pc);
5327 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5328 }
5329
5330 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5331 const char* file_name = "custom_debug.xml";
5332 if (_debug_file_printer == nullptr) {
5333 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5334 } else {
5335 _debug_file_printer->update_compiled_method(C->method());
5336 }
5337 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5338 _debug_file_printer->print_graph(phase_name, fr);
5339 }
5340
5341 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5342 ResourceMark rm;
5343 GrowableArray<const Node*> empty_list;
5344 igv_print_graph_to_network(phase_name, empty_list, fr);
5345 }
5346
5347 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5348 if (_debug_network_printer == nullptr) {
5349 _debug_network_printer = new IdealGraphPrinter(C);
5350 } else {
5351 _debug_network_printer->update_compiled_method(C->method());
5352 }
5353 tty->print_cr("Method printed over network stream to IGV");
5354 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5355 }
5356 #endif // !PRODUCT
5357
5358 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5359 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5360 return value;
5361 }
5362 Node* result = nullptr;
5363 if (bt == T_BYTE) {
5364 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5365 result = new RShiftINode(result, phase->intcon(24));
5366 } else if (bt == T_BOOLEAN) {
5367 result = new AndINode(value, phase->intcon(0xFF));
5368 } else if (bt == T_CHAR) {
5369 result = new AndINode(value,phase->intcon(0xFFFF));
5370 } else {
5371 assert(bt == T_SHORT, "unexpected narrow type");
5372 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5373 result = new RShiftINode(result, phase->intcon(16));
5374 }
5375 if (transform_res) {
5376 result = phase->transform(result);
5377 }
5378 return result;
5379 }
5380
5381 void Compile::record_method_not_compilable_oom() {
5382 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5383 }
5384
5385 #ifndef PRODUCT
5386 // Collects all the control inputs from nodes on the worklist and from their data dependencies
5387 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
5388 // Follow non-control edges until we reach CFG nodes
5389 for (uint i = 0; i < worklist.size(); i++) {
5390 const Node* n = worklist.at(i);
5391 for (uint j = 0; j < n->req(); j++) {
5392 Node* in = n->in(j);
5393 if (in == nullptr || in->is_Root()) {
5394 continue;
5395 }
5396 if (in->is_CFG()) {
5397 if (in->is_Call()) {
5398 // The return value of a call is only available if the call did not result in an exception
5399 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
5400 if (control_proj_use->is_Catch()) {
5401 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
5402 candidates.push(fall_through);
5403 continue;
5404 }
5405 }
5406
5407 if (in->is_Multi()) {
5408 // We got here by following data inputs so we should only have one control use
5409 // (no IfNode, etc)
5410 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
5411 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
5412 } else {
5413 candidates.push(in);
5414 }
5415 } else {
5416 worklist.push(in);
5417 }
5418 }
5419 }
5420 }
5421
5422 // Returns the candidate node that is a descendant to all the other candidates
5423 static Node* pick_control(Unique_Node_List& candidates) {
5424 Unique_Node_List worklist;
5425 worklist.copy(candidates);
5426
5427 // Traverse backwards through the CFG
5428 for (uint i = 0; i < worklist.size(); i++) {
5429 const Node* n = worklist.at(i);
5430 if (n->is_Root()) {
5431 continue;
5432 }
5433 for (uint j = 0; j < n->req(); j++) {
5434 // Skip backedge of loops to avoid cycles
5435 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
5436 continue;
5437 }
5438
5439 Node* pred = n->in(j);
5440 if (pred != nullptr && pred != n && pred->is_CFG()) {
5441 worklist.push(pred);
5442 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
5443 candidates.remove(pred);
5444 }
5445 }
5446 }
5447
5448 assert(candidates.size() == 1, "unexpected control flow");
5449 return candidates.at(0);
5450 }
5451
5452 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
5453 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
5454 call->init_req((*pos)++, parm);
5455 const BasicType bt = parm->bottom_type()->basic_type();
5456 if (bt == T_LONG || bt == T_DOUBLE) {
5457 call->init_req((*pos)++, half);
5458 }
5459 }
5460
5461 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
5462 Node* parm0, Node* parm1,
5463 Node* parm2, Node* parm3,
5464 Node* parm4, Node* parm5,
5465 Node* parm6) const {
5466 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
5467 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
5468 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
5469
5470 // find the most suitable control input
5471 Unique_Node_List worklist, candidates;
5472 if (parm0 != nullptr) { worklist.push(parm0);
5473 if (parm1 != nullptr) { worklist.push(parm1);
5474 if (parm2 != nullptr) { worklist.push(parm2);
5475 if (parm3 != nullptr) { worklist.push(parm3);
5476 if (parm4 != nullptr) { worklist.push(parm4);
5477 if (parm5 != nullptr) { worklist.push(parm5);
5478 if (parm6 != nullptr) { worklist.push(parm6);
5479 /* close each nested if ===> */ } } } } } } }
5480 find_candidate_control_inputs(worklist, candidates);
5481 Node* control = nullptr;
5482 if (candidates.size() == 0) {
5483 control = C->start()->proj_out(TypeFunc::Control);
5484 } else {
5485 control = pick_control(candidates);
5486 }
5487
5488 // find all the previous users of the control we picked
5489 GrowableArray<Node*> users_of_control;
5490 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
5491 Node* use = control->fast_out(i);
5492 if (use->is_CFG() && use != control) {
5493 users_of_control.push(use);
5494 }
5495 }
5496
5497 // we do not actually care about IO and memory as it uses neither
5498 call->init_req(TypeFunc::Control, control);
5499 call->init_req(TypeFunc::I_O, top());
5500 call->init_req(TypeFunc::Memory, top());
5501 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
5502 call->init_req(TypeFunc::ReturnAdr, top());
5503
5504 int pos = TypeFunc::Parms;
5505 call->init_req(pos++, str_node);
5506 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
5507 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
5508 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
5509 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
5510 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
5511 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
5512 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
5513 /* close each nested if ===> */ } } } } } } }
5514 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
5515
5516 call = gvn->transform(call);
5517 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
5518
5519 // rewire previous users to have the new call as control instead
5520 PhaseIterGVN* igvn = gvn->is_IterGVN();
5521 for (int i = 0; i < users_of_control.length(); i++) {
5522 Node* use = users_of_control.at(i);
5523 for (uint j = 0; j < use->req(); j++) {
5524 if (use->in(j) == control) {
5525 if (igvn != nullptr) {
5526 igvn->replace_input_of(use, j, call_control_proj);
5527 } else {
5528 gvn->hash_delete(use);
5529 use->set_req(j, call_control_proj);
5530 gvn->hash_insert(use);
5531 }
5532 }
5533 }
5534 }
5535
5536 return call;
5537 }
5538 #endif // !PRODUCT