1 /*
2 * Copyright (c) 1997, 2026, 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* compile_phase_name) const {
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", compile_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 compile_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("PrintIdeal");
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 the fixed slots
886 // for GC barriers and the original deopt pc. All these slots should be able
887 // to fit an address.
888 int reserved_gc_slots = BarrierSet::barrier_set()->barrier_set_c2()->reserved_slots();
889 int next_slot = fixed_slots() + (reserved_gc_slots + 1) * sizeof(address) / VMRegImpl::stack_slot_size;
890 set_fixed_slots(next_slot);
891
892 // Compute when to use implicit null checks. Used by matching trap based
893 // nodes and NullCheck optimization.
894 set_allowed_deopt_reasons();
895
896 // Now generate code
897 Code_Gen();
898 }
899
900 //------------------------------Compile----------------------------------------
901 // Compile a runtime stub
902 Compile::Compile(ciEnv* ci_env,
903 TypeFunc_generator generator,
904 address stub_function,
905 const char* stub_name,
906 StubId stub_id,
907 int is_fancy_jump,
908 bool pass_tls,
909 bool return_pc,
910 DirectiveSet* directive)
911 : Phase(Compiler),
912 _compile_id(0),
913 _options(Options::for_runtime_stub()),
914 _method(nullptr),
915 _entry_bci(InvocationEntryBci),
916 _stub_function(stub_function),
917 _stub_name(stub_name),
918 _stub_id(stub_id),
919 _stub_entry_point(nullptr),
920 _max_node_limit(MaxNodeLimit),
921 _post_loop_opts_phase(false),
922 _merge_stores_phase(false),
923 _allow_macro_nodes(true),
924 _inlining_progress(false),
925 _inlining_incrementally(false),
926 _has_reserved_stack_access(false),
927 #ifndef PRODUCT
928 _igv_idx(0),
929 _trace_opto_output(directive->TraceOptoOutputOption),
930 #endif
931 _clinit_barrier_on_entry(false),
932 _stress_seed(0),
933 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
934 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
935 _env(ci_env),
936 _directive(directive),
937 _log(ci_env->log()),
938 _first_failure_details(nullptr),
939 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
940 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
941 _congraph(nullptr),
942 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
943 _unique(0),
944 _dead_node_count(0),
945 _dead_node_list(comp_arena()),
946 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
947 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
948 _node_arena(&_node_arena_one),
949 _mach_constant_base_node(nullptr),
950 _Compile_types(mtCompiler, Arena::Tag::tag_type),
951 _initial_gvn(nullptr),
952 _igvn_worklist(nullptr),
953 _types(nullptr),
954 _node_hash(nullptr),
955 _has_mh_late_inlines(false),
956 _oom(false),
957 _replay_inline_data(nullptr),
958 _inline_printer(this),
959 _java_calls(0),
960 _inner_loops(0),
961 _FIRST_STACK_mask(comp_arena()),
962 _interpreter_frame_size(0),
963 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
964 _output(nullptr),
965 #ifndef PRODUCT
966 _in_dump_cnt(0),
967 #endif
968 _allowed_reasons(0) {
969 C = this;
970
971 // try to reuse an existing stub
972 {
973 BlobId blob_id = StubInfo::blob(_stub_id);
974 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
975 if (blob != nullptr) {
976 RuntimeStub* rs = blob->as_runtime_stub();
977 _stub_entry_point = rs->entry_point();
978 return;
979 }
980 }
981
982 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
983 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
984
985 #ifndef PRODUCT
986 set_print_assembly(PrintFrameConverterAssembly);
987 set_parsed_irreducible_loop(false);
988 #else
989 set_print_assembly(false); // Must initialize.
990 #endif
991 set_has_irreducible_loop(false); // no loops
992
993 CompileWrapper cw(this);
994 Init(/*do_aliasing=*/ false);
995 init_tf((*generator)());
996
997 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
998 _types = new (comp_arena()) Type_Array(comp_arena());
999 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1000
1001 if (StressLCM || StressGCM || StressBailout) {
1002 initialize_stress_seed(directive);
1003 }
1004
1005 {
1006 PhaseGVN gvn;
1007 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1008 gvn.transform(top());
1009
1010 GraphKit kit;
1011 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1012 }
1013
1014 NOT_PRODUCT( verify_graph_edges(); )
1015
1016 Code_Gen();
1017 }
1018
1019 Compile::~Compile() {
1020 delete _first_failure_details;
1021 };
1022
1023 //------------------------------Init-------------------------------------------
1024 // Prepare for a single compilation
1025 void Compile::Init(bool aliasing) {
1026 _do_aliasing = aliasing;
1027 _unique = 0;
1028 _regalloc = nullptr;
1029
1030 _tf = nullptr; // filled in later
1031 _top = nullptr; // cached later
1032 _matcher = nullptr; // filled in later
1033 _cfg = nullptr; // filled in later
1034
1035 _node_note_array = nullptr;
1036 _default_node_notes = nullptr;
1037 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1038
1039 _immutable_memory = nullptr; // filled in at first inquiry
1040
1041 #ifdef ASSERT
1042 _phase_optimize_finished = false;
1043 _phase_verify_ideal_loop = false;
1044 _exception_backedge = false;
1045 _type_verify = nullptr;
1046 #endif
1047
1048 // Globally visible Nodes
1049 // First set TOP to null to give safe behavior during creation of RootNode
1050 set_cached_top_node(nullptr);
1051 set_root(new RootNode());
1052 // Now that you have a Root to point to, create the real TOP
1053 set_cached_top_node( new ConNode(Type::TOP) );
1054 set_recent_alloc(nullptr, nullptr);
1055
1056 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1057 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1058 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1059 env()->set_dependencies(new Dependencies(env()));
1060
1061 _fixed_slots = 0;
1062 set_has_split_ifs(false);
1063 set_has_loops(false); // first approximation
1064 set_has_stringbuilder(false);
1065 set_has_boxed_value(false);
1066 _trap_can_recompile = false; // no traps emitted yet
1067 _major_progress = true; // start out assuming good things will happen
1068 set_has_unsafe_access(false);
1069 set_max_vector_size(0);
1070 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1071 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1072 set_decompile_count(0);
1073
1074 #ifndef PRODUCT
1075 _phase_counter = 0;
1076 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1077 #endif
1078
1079 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1080 _loop_opts_cnt = LoopOptsCount;
1081 set_do_inlining(Inline);
1082 set_max_inline_size(MaxInlineSize);
1083 set_freq_inline_size(FreqInlineSize);
1084 set_do_scheduling(OptoScheduling);
1085
1086 set_do_vector_loop(false);
1087 set_has_monitors(false);
1088 set_has_scoped_access(false);
1089
1090 if (AllowVectorizeOnDemand) {
1091 if (has_method() && _directive->VectorizeOption) {
1092 set_do_vector_loop(true);
1093 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());})
1094 } else if (has_method() && method()->name() != nullptr &&
1095 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1096 set_do_vector_loop(true);
1097 }
1098 }
1099 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1100 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());})
1101
1102 _max_node_limit = _directive->MaxNodeLimitOption;
1103
1104 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1105 set_clinit_barrier_on_entry(true);
1106 }
1107 if (debug_info()->recording_non_safepoints()) {
1108 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1109 (comp_arena(), 8, 0, nullptr));
1110 set_default_node_notes(Node_Notes::make(this));
1111 }
1112
1113 const int grow_ats = 16;
1114 _max_alias_types = grow_ats;
1115 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1116 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1117 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1118 {
1119 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1120 }
1121 // Initialize the first few types.
1122 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1123 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1124 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1125 _num_alias_types = AliasIdxRaw+1;
1126 // Zero out the alias type cache.
1127 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1128 // A null adr_type hits in the cache right away. Preload the right answer.
1129 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1130 }
1131
1132 #ifdef ASSERT
1133 // Verify that the current StartNode is valid.
1134 void Compile::verify_start(StartNode* s) const {
1135 assert(failing_internal() || s == start(), "should be StartNode");
1136 }
1137 #endif
1138
1139 /**
1140 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1141 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1142 * the ideal graph.
1143 */
1144 StartNode* Compile::start() const {
1145 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1146 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1147 Node* start = root()->fast_out(i);
1148 if (start->is_Start()) {
1149 return start->as_Start();
1150 }
1151 }
1152 fatal("Did not find Start node!");
1153 return nullptr;
1154 }
1155
1156 //-------------------------------immutable_memory-------------------------------------
1157 // Access immutable memory
1158 Node* Compile::immutable_memory() {
1159 if (_immutable_memory != nullptr) {
1160 return _immutable_memory;
1161 }
1162 StartNode* s = start();
1163 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1164 Node *p = s->fast_out(i);
1165 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1166 _immutable_memory = p;
1167 return _immutable_memory;
1168 }
1169 }
1170 ShouldNotReachHere();
1171 return nullptr;
1172 }
1173
1174 //----------------------set_cached_top_node------------------------------------
1175 // Install the cached top node, and make sure Node::is_top works correctly.
1176 void Compile::set_cached_top_node(Node* tn) {
1177 if (tn != nullptr) verify_top(tn);
1178 Node* old_top = _top;
1179 _top = tn;
1180 // Calling Node::setup_is_top allows the nodes the chance to adjust
1181 // their _out arrays.
1182 if (_top != nullptr) _top->setup_is_top();
1183 if (old_top != nullptr) old_top->setup_is_top();
1184 assert(_top == nullptr || top()->is_top(), "");
1185 }
1186
1187 #ifdef ASSERT
1188 uint Compile::count_live_nodes_by_graph_walk() {
1189 Unique_Node_List useful(comp_arena());
1190 // Get useful node list by walking the graph.
1191 identify_useful_nodes(useful);
1192 return useful.size();
1193 }
1194
1195 void Compile::print_missing_nodes() {
1196
1197 // Return if CompileLog is null and PrintIdealNodeCount is false.
1198 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1199 return;
1200 }
1201
1202 // This is an expensive function. It is executed only when the user
1203 // specifies VerifyIdealNodeCount option or otherwise knows the
1204 // additional work that needs to be done to identify reachable nodes
1205 // by walking the flow graph and find the missing ones using
1206 // _dead_node_list.
1207
1208 Unique_Node_List useful(comp_arena());
1209 // Get useful node list by walking the graph.
1210 identify_useful_nodes(useful);
1211
1212 uint l_nodes = C->live_nodes();
1213 uint l_nodes_by_walk = useful.size();
1214
1215 if (l_nodes != l_nodes_by_walk) {
1216 if (_log != nullptr) {
1217 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1218 _log->stamp();
1219 _log->end_head();
1220 }
1221 VectorSet& useful_member_set = useful.member_set();
1222 int last_idx = l_nodes_by_walk;
1223 for (int i = 0; i < last_idx; i++) {
1224 if (useful_member_set.test(i)) {
1225 if (_dead_node_list.test(i)) {
1226 if (_log != nullptr) {
1227 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1228 }
1229 if (PrintIdealNodeCount) {
1230 // Print the log message to tty
1231 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1232 useful.at(i)->dump();
1233 }
1234 }
1235 }
1236 else if (! _dead_node_list.test(i)) {
1237 if (_log != nullptr) {
1238 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1239 }
1240 if (PrintIdealNodeCount) {
1241 // Print the log message to tty
1242 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1243 }
1244 }
1245 }
1246 if (_log != nullptr) {
1247 _log->tail("mismatched_nodes");
1248 }
1249 }
1250 }
1251 void Compile::record_modified_node(Node* n) {
1252 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1253 _modified_nodes->push(n);
1254 }
1255 }
1256
1257 void Compile::remove_modified_node(Node* n) {
1258 if (_modified_nodes != nullptr) {
1259 _modified_nodes->remove(n);
1260 }
1261 }
1262 #endif
1263
1264 #ifndef PRODUCT
1265 void Compile::verify_top(Node* tn) const {
1266 if (tn != nullptr) {
1267 assert(tn->is_Con(), "top node must be a constant");
1268 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1269 assert(tn->in(0) != nullptr, "must have live top node");
1270 }
1271 }
1272 #endif
1273
1274
1275 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1276
1277 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1278 guarantee(arr != nullptr, "");
1279 int num_blocks = arr->length();
1280 if (grow_by < num_blocks) grow_by = num_blocks;
1281 int num_notes = grow_by * _node_notes_block_size;
1282 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1283 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1284 while (num_notes > 0) {
1285 arr->append(notes);
1286 notes += _node_notes_block_size;
1287 num_notes -= _node_notes_block_size;
1288 }
1289 assert(num_notes == 0, "exact multiple, please");
1290 }
1291
1292 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1293 if (source == nullptr || dest == nullptr) return false;
1294
1295 if (dest->is_Con())
1296 return false; // Do not push debug info onto constants.
1297
1298 #ifdef ASSERT
1299 // Leave a bread crumb trail pointing to the original node:
1300 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1301 dest->set_debug_orig(source);
1302 }
1303 #endif
1304
1305 if (node_note_array() == nullptr)
1306 return false; // Not collecting any notes now.
1307
1308 // This is a copy onto a pre-existing node, which may already have notes.
1309 // If both nodes have notes, do not overwrite any pre-existing notes.
1310 Node_Notes* source_notes = node_notes_at(source->_idx);
1311 if (source_notes == nullptr || source_notes->is_clear()) return false;
1312 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1313 if (dest_notes == nullptr || dest_notes->is_clear()) {
1314 return set_node_notes_at(dest->_idx, source_notes);
1315 }
1316
1317 Node_Notes merged_notes = (*source_notes);
1318 // The order of operations here ensures that dest notes will win...
1319 merged_notes.update_from(dest_notes);
1320 return set_node_notes_at(dest->_idx, &merged_notes);
1321 }
1322
1323
1324 //--------------------------allow_range_check_smearing-------------------------
1325 // Gating condition for coalescing similar range checks.
1326 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1327 // single covering check that is at least as strong as any of them.
1328 // If the optimization succeeds, the simplified (strengthened) range check
1329 // will always succeed. If it fails, we will deopt, and then give up
1330 // on the optimization.
1331 bool Compile::allow_range_check_smearing() const {
1332 // If this method has already thrown a range-check,
1333 // assume it was because we already tried range smearing
1334 // and it failed.
1335 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1336 return !already_trapped;
1337 }
1338
1339
1340 //------------------------------flatten_alias_type-----------------------------
1341 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1342 assert(do_aliasing(), "Aliasing should be enabled");
1343 int offset = tj->offset();
1344 TypePtr::PTR ptr = tj->ptr();
1345
1346 // Known instance (scalarizable allocation) alias only with itself.
1347 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1348 tj->is_oopptr()->is_known_instance();
1349
1350 // Process weird unsafe references.
1351 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1352 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1353 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1354 tj = TypeOopPtr::BOTTOM;
1355 ptr = tj->ptr();
1356 offset = tj->offset();
1357 }
1358
1359 // Array pointers need some flattening
1360 const TypeAryPtr* ta = tj->isa_aryptr();
1361 if (ta && ta->is_stable()) {
1362 // Erase stability property for alias analysis.
1363 tj = ta = ta->cast_to_stable(false);
1364 }
1365 if( ta && is_known_inst ) {
1366 if ( offset != Type::OffsetBot &&
1367 offset > arrayOopDesc::length_offset_in_bytes() ) {
1368 offset = Type::OffsetBot; // Flatten constant access into array body only
1369 tj = ta = ta->
1370 remove_speculative()->
1371 cast_to_ptr_type(ptr)->
1372 with_offset(offset);
1373 }
1374 } else if (ta != nullptr) {
1375 // Common slices
1376 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1377 return TypeAryPtr::RANGE;
1378 } else if (offset == oopDesc::klass_offset_in_bytes()) {
1379 return TypeInstPtr::KLASS;
1380 } else if (offset == oopDesc::mark_offset_in_bytes()) {
1381 return TypeInstPtr::MARK;
1382 }
1383
1384 // Remove size and stability
1385 const TypeAry* normalized_ary = TypeAry::make(ta->elem(), TypeInt::POS, false);
1386 // Remove ptr, const_oop, and offset
1387 if (ta->elem() == Type::BOTTOM) {
1388 // Bottom array (meet of int[] and byte[] for example), accesses to it will be done with
1389 // Unsafe. This should alias with all arrays. For now just leave it as it is (this is
1390 // incorrect, see JDK-8331133).
1391 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, false, Type::OffsetBot);
1392 } else if (ta->elem()->make_oopptr() != nullptr) {
1393 // Object arrays, all of them share the same slice
1394 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, TypeInt::POS, false);
1395 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, tary, nullptr, false, Type::OffsetBot);
1396 } else {
1397 // Primitive arrays
1398 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, ta->exact_klass(), true, Type::OffsetBot);
1399 }
1400
1401 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1402 // cannot be distinguished by bytecode alone.
1403 if (ta->elem() == TypeInt::BOOL) {
1404 tj = ta = TypeAryPtr::BYTES;
1405 }
1406 }
1407
1408 // Oop pointers need some flattening
1409 const TypeInstPtr *to = tj->isa_instptr();
1410 if (to && to != TypeOopPtr::BOTTOM) {
1411 ciInstanceKlass* ik = to->instance_klass();
1412 if( ptr == TypePtr::Constant ) {
1413 if (ik != ciEnv::current()->Class_klass() ||
1414 offset < ik->layout_helper_size_in_bytes()) {
1415 // No constant oop pointers (such as Strings); they alias with
1416 // unknown strings.
1417 assert(!is_known_inst, "not scalarizable allocation");
1418 tj = to = to->
1419 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1420 remove_speculative()->
1421 cast_to_ptr_type(TypePtr::BotPTR)->
1422 cast_to_exactness(false);
1423 }
1424 } else if( is_known_inst ) {
1425 tj = to; // Keep NotNull and klass_is_exact for instance type
1426 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1427 // During the 2nd round of IterGVN, NotNull castings are removed.
1428 // Make sure the Bottom and NotNull variants alias the same.
1429 // Also, make sure exact and non-exact variants alias the same.
1430 tj = to = to->
1431 remove_speculative()->
1432 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1433 cast_to_ptr_type(TypePtr::BotPTR)->
1434 cast_to_exactness(false);
1435 }
1436 if (to->speculative() != nullptr) {
1437 tj = to = to->remove_speculative();
1438 }
1439 // Canonicalize the holder of this field
1440 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1441 // First handle header references such as a LoadKlassNode, even if the
1442 // object's klass is unloaded at compile time (4965979).
1443 if (!is_known_inst) { // Do it only for non-instance types
1444 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1445 }
1446 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1447 // Static fields are in the space above the normal instance
1448 // fields in the java.lang.Class instance.
1449 if (ik != ciEnv::current()->Class_klass()) {
1450 to = nullptr;
1451 tj = TypeOopPtr::BOTTOM;
1452 offset = tj->offset();
1453 }
1454 } else {
1455 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1456 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1457 assert(tj->offset() == offset, "no change to offset expected");
1458 bool xk = to->klass_is_exact();
1459 int instance_id = to->instance_id();
1460
1461 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1462 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1463 // its interfaces are included.
1464 if (xk && ik->equals(canonical_holder)) {
1465 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1466 } else {
1467 assert(xk || !is_known_inst, "Known instance should be exact type");
1468 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1469 }
1470 }
1471 }
1472
1473 // Klass pointers to object array klasses need some flattening
1474 const TypeKlassPtr *tk = tj->isa_klassptr();
1475 if( tk ) {
1476 // If we are referencing a field within a Klass, we need
1477 // to assume the worst case of an Object. Both exact and
1478 // inexact types must flatten to the same alias class so
1479 // use NotNull as the PTR.
1480 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1481 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1482 env()->Object_klass(),
1483 offset);
1484 }
1485
1486 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1487 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1488 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1489 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1490 } else {
1491 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1492 }
1493 }
1494
1495 // Check for precise loads from the primary supertype array and force them
1496 // to the supertype cache alias index. Check for generic array loads from
1497 // the primary supertype array and also force them to the supertype cache
1498 // alias index. Since the same load can reach both, we need to merge
1499 // these 2 disparate memories into the same alias class. Since the
1500 // primary supertype array is read-only, there's no chance of confusion
1501 // where we bypass an array load and an array store.
1502 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1503 if (offset == Type::OffsetBot ||
1504 (offset >= primary_supers_offset &&
1505 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1506 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1507 offset = in_bytes(Klass::secondary_super_cache_offset());
1508 tj = tk = tk->with_offset(offset);
1509 }
1510 }
1511
1512 // Flatten all Raw pointers together.
1513 if (tj->base() == Type::RawPtr)
1514 tj = TypeRawPtr::BOTTOM;
1515
1516 if (tj->base() == Type::AnyPtr)
1517 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1518
1519 offset = tj->offset();
1520 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1521
1522 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1523 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1524 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1525 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1526 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1527 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1528 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1529 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1530 assert( tj->ptr() != TypePtr::TopPTR &&
1531 tj->ptr() != TypePtr::AnyNull &&
1532 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1533 // assert( tj->ptr() != TypePtr::Constant ||
1534 // tj->base() == Type::RawPtr ||
1535 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1536
1537 return tj;
1538 }
1539
1540 void Compile::AliasType::Init(int i, const TypePtr* at) {
1541 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1542 _index = i;
1543 _adr_type = at;
1544 _field = nullptr;
1545 _element = nullptr;
1546 _is_rewritable = true; // default
1547 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1548 if (atoop != nullptr && atoop->is_known_instance()) {
1549 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1550 _general_index = Compile::current()->get_alias_index(gt);
1551 } else {
1552 _general_index = 0;
1553 }
1554 }
1555
1556 BasicType Compile::AliasType::basic_type() const {
1557 if (element() != nullptr) {
1558 const Type* element = adr_type()->is_aryptr()->elem();
1559 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1560 } if (field() != nullptr) {
1561 return field()->layout_type();
1562 } else {
1563 return T_ILLEGAL; // unknown
1564 }
1565 }
1566
1567 //---------------------------------print_on------------------------------------
1568 #ifndef PRODUCT
1569 void Compile::AliasType::print_on(outputStream* st) {
1570 if (index() < 10)
1571 st->print("@ <%d> ", index());
1572 else st->print("@ <%d>", index());
1573 st->print(is_rewritable() ? " " : " RO");
1574 int offset = adr_type()->offset();
1575 if (offset == Type::OffsetBot)
1576 st->print(" +any");
1577 else st->print(" +%-3d", offset);
1578 st->print(" in ");
1579 adr_type()->dump_on(st);
1580 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1581 if (field() != nullptr && tjp) {
1582 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1583 tjp->offset() != field()->offset_in_bytes()) {
1584 st->print(" != ");
1585 field()->print();
1586 st->print(" ***");
1587 }
1588 }
1589 }
1590
1591 void print_alias_types() {
1592 Compile* C = Compile::current();
1593 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1594 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1595 C->alias_type(idx)->print_on(tty);
1596 tty->cr();
1597 }
1598 }
1599 #endif
1600
1601
1602 //----------------------------probe_alias_cache--------------------------------
1603 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1604 intptr_t key = (intptr_t) adr_type;
1605 key ^= key >> logAliasCacheSize;
1606 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1607 }
1608
1609
1610 //-----------------------------grow_alias_types--------------------------------
1611 void Compile::grow_alias_types() {
1612 const int old_ats = _max_alias_types; // how many before?
1613 const int new_ats = old_ats; // how many more?
1614 const int grow_ats = old_ats+new_ats; // how many now?
1615 _max_alias_types = grow_ats;
1616 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1617 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1618 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1619 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1620 }
1621
1622
1623 //--------------------------------find_alias_type------------------------------
1624 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1625 if (!do_aliasing()) {
1626 return alias_type(AliasIdxBot);
1627 }
1628
1629 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1630 if (ace->_adr_type == adr_type) {
1631 return alias_type(ace->_index);
1632 }
1633
1634 // Handle special cases.
1635 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1636 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1637
1638 // Do it the slow way.
1639 const TypePtr* flat = flatten_alias_type(adr_type);
1640
1641 #ifdef ASSERT
1642 {
1643 ResourceMark rm;
1644 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1645 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1646 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1647 Type::str(adr_type));
1648 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1649 const TypeOopPtr* foop = flat->is_oopptr();
1650 // Scalarizable allocations have exact klass always.
1651 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1652 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1653 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1654 Type::str(foop), Type::str(xoop));
1655 }
1656 }
1657 #endif
1658
1659 int idx = AliasIdxTop;
1660 for (int i = 0; i < num_alias_types(); i++) {
1661 if (alias_type(i)->adr_type() == flat) {
1662 idx = i;
1663 break;
1664 }
1665 }
1666
1667 if (idx == AliasIdxTop) {
1668 if (no_create) return nullptr;
1669 // Grow the array if necessary.
1670 if (_num_alias_types == _max_alias_types) grow_alias_types();
1671 // Add a new alias type.
1672 idx = _num_alias_types++;
1673 _alias_types[idx]->Init(idx, flat);
1674 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1675 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1676 if (flat->isa_instptr()) {
1677 if (flat->offset() == java_lang_Class::klass_offset()
1678 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1679 alias_type(idx)->set_rewritable(false);
1680 }
1681 if (flat->isa_aryptr()) {
1682 #ifdef ASSERT
1683 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1684 // (T_BYTE has the weakest alignment and size restrictions...)
1685 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1686 #endif
1687 if (flat->offset() == TypePtr::OffsetBot) {
1688 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1689 }
1690 }
1691 if (flat->isa_klassptr()) {
1692 if (UseCompactObjectHeaders) {
1693 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1694 alias_type(idx)->set_rewritable(false);
1695 }
1696 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1697 alias_type(idx)->set_rewritable(false);
1698 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1699 alias_type(idx)->set_rewritable(false);
1700 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1701 alias_type(idx)->set_rewritable(false);
1702 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1703 alias_type(idx)->set_rewritable(false);
1704 }
1705
1706 if (flat->isa_instklassptr()) {
1707 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1708 alias_type(idx)->set_rewritable(false);
1709 }
1710 }
1711 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1712 // but the base pointer type is not distinctive enough to identify
1713 // references into JavaThread.)
1714
1715 // Check for final fields.
1716 const TypeInstPtr* tinst = flat->isa_instptr();
1717 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1718 ciField* field;
1719 if (tinst->const_oop() != nullptr &&
1720 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1721 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1722 // static field
1723 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1724 field = k->get_field_by_offset(tinst->offset(), true);
1725 } else {
1726 ciInstanceKlass *k = tinst->instance_klass();
1727 field = k->get_field_by_offset(tinst->offset(), false);
1728 }
1729 assert(field == nullptr ||
1730 original_field == nullptr ||
1731 (field->holder() == original_field->holder() &&
1732 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1733 field->is_static() == original_field->is_static()), "wrong field?");
1734 // Set field() and is_rewritable() attributes.
1735 if (field != nullptr) alias_type(idx)->set_field(field);
1736 }
1737 }
1738
1739 // Fill the cache for next time.
1740 ace->_adr_type = adr_type;
1741 ace->_index = idx;
1742 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1743
1744 // Might as well try to fill the cache for the flattened version, too.
1745 AliasCacheEntry* face = probe_alias_cache(flat);
1746 if (face->_adr_type == nullptr) {
1747 face->_adr_type = flat;
1748 face->_index = idx;
1749 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1750 }
1751
1752 return alias_type(idx);
1753 }
1754
1755
1756 Compile::AliasType* Compile::alias_type(ciField* field) {
1757 const TypeOopPtr* t;
1758 if (field->is_static())
1759 t = TypeInstPtr::make(field->holder()->java_mirror());
1760 else
1761 t = TypeOopPtr::make_from_klass_raw(field->holder());
1762 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1763 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1764 return atp;
1765 }
1766
1767
1768 //------------------------------have_alias_type--------------------------------
1769 bool Compile::have_alias_type(const TypePtr* adr_type) {
1770 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1771 if (ace->_adr_type == adr_type) {
1772 return true;
1773 }
1774
1775 // Handle special cases.
1776 if (adr_type == nullptr) return true;
1777 if (adr_type == TypePtr::BOTTOM) return true;
1778
1779 return find_alias_type(adr_type, true, nullptr) != nullptr;
1780 }
1781
1782 //-----------------------------must_alias--------------------------------------
1783 // True if all values of the given address type are in the given alias category.
1784 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1785 if (alias_idx == AliasIdxBot) return true; // the universal category
1786 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1787 if (alias_idx == AliasIdxTop) return false; // the empty category
1788 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1789
1790 // the only remaining possible overlap is identity
1791 int adr_idx = get_alias_index(adr_type);
1792 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1793 assert(adr_idx == alias_idx ||
1794 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1795 && adr_type != TypeOopPtr::BOTTOM),
1796 "should not be testing for overlap with an unsafe pointer");
1797 return adr_idx == alias_idx;
1798 }
1799
1800 //------------------------------can_alias--------------------------------------
1801 // True if any values of the given address type are in the given alias category.
1802 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1803 if (alias_idx == AliasIdxTop) return false; // the empty category
1804 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1805 // Known instance doesn't alias with bottom memory
1806 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1807 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1808
1809 // the only remaining possible overlap is identity
1810 int adr_idx = get_alias_index(adr_type);
1811 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1812 return adr_idx == alias_idx;
1813 }
1814
1815 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1816 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1817 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1818 if (parse_predicate_count() == 0) {
1819 return;
1820 }
1821 for (int i = 0; i < parse_predicate_count(); i++) {
1822 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1823 parse_predicate->mark_useless(igvn);
1824 }
1825 _parse_predicates.clear();
1826 }
1827
1828 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1829 if (!n->for_post_loop_opts_igvn()) {
1830 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1831 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1832 _for_post_loop_igvn.append(n);
1833 }
1834 }
1835
1836 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1837 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1838 _for_post_loop_igvn.remove(n);
1839 }
1840
1841 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1842 // Verify that all previous optimizations produced a valid graph
1843 // at least to this point, even if no loop optimizations were done.
1844 PhaseIdealLoop::verify(igvn);
1845
1846 if (_print_phase_loop_opts) {
1847 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1848 }
1849 C->set_post_loop_opts_phase(); // no more loop opts allowed
1850
1851 assert(!C->major_progress(), "not cleared");
1852
1853 if (_for_post_loop_igvn.length() > 0) {
1854 while (_for_post_loop_igvn.length() > 0) {
1855 Node* n = _for_post_loop_igvn.pop();
1856 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1857 igvn._worklist.push(n);
1858 }
1859 igvn.optimize();
1860 if (failing()) return;
1861 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1862 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1863
1864 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1865 if (C->major_progress()) {
1866 C->clear_major_progress(); // ensure that major progress is now clear
1867 }
1868 }
1869 }
1870
1871 void Compile::record_for_merge_stores_igvn(Node* n) {
1872 if (!n->for_merge_stores_igvn()) {
1873 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1874 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1875 _for_merge_stores_igvn.append(n);
1876 }
1877 }
1878
1879 void Compile::remove_from_merge_stores_igvn(Node* n) {
1880 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1881 _for_merge_stores_igvn.remove(n);
1882 }
1883
1884 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1885 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1886 // the stores, and we merge the wrong sequence of stores.
1887 // Example:
1888 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1889 // Apply MergeStores:
1890 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1891 // Remove more RangeChecks:
1892 // StoreI [ StoreL ] StoreI
1893 // But now it would have been better to do this instead:
1894 // [ StoreL ] [ StoreL ]
1895 //
1896 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1897 // since we never unset _merge_stores_phase.
1898 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1899 C->set_merge_stores_phase();
1900
1901 if (_for_merge_stores_igvn.length() > 0) {
1902 while (_for_merge_stores_igvn.length() > 0) {
1903 Node* n = _for_merge_stores_igvn.pop();
1904 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1905 igvn._worklist.push(n);
1906 }
1907 igvn.optimize();
1908 if (failing()) return;
1909 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1910 print_method(PHASE_AFTER_MERGE_STORES, 3);
1911 }
1912 }
1913
1914 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1915 if (OptimizeUnstableIf) {
1916 _unstable_if_traps.append(trap);
1917 }
1918 }
1919
1920 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1921 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1922 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1923 Node* n = trap->uncommon_trap();
1924 if (!useful.member(n)) {
1925 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1926 }
1927 }
1928 }
1929
1930 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1931 // or fold-compares case. Return true if succeed or not found.
1932 //
1933 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1934 // false and ask fold-compares to yield.
1935 //
1936 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1937 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1938 // when deoptimization does happen.
1939 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1940 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1941 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1942 if (trap->uncommon_trap() == unc) {
1943 if (yield && trap->modified()) {
1944 return false;
1945 }
1946 _unstable_if_traps.delete_at(i);
1947 break;
1948 }
1949 }
1950 return true;
1951 }
1952
1953 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1954 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1955 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1956 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1957 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1958 CallStaticJavaNode* unc = trap->uncommon_trap();
1959 int next_bci = trap->next_bci();
1960 bool modified = trap->modified();
1961
1962 if (next_bci != -1 && !modified) {
1963 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1964 JVMState* jvms = unc->jvms();
1965 ciMethod* method = jvms->method();
1966 ciBytecodeStream iter(method);
1967
1968 iter.force_bci(jvms->bci());
1969 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1970 Bytecodes::Code c = iter.cur_bc();
1971 Node* lhs = nullptr;
1972 Node* rhs = nullptr;
1973 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1974 lhs = unc->peek_operand(0);
1975 rhs = unc->peek_operand(1);
1976 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1977 lhs = unc->peek_operand(0);
1978 }
1979
1980 ResourceMark rm;
1981 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1982 assert(live_locals.is_valid(), "broken liveness info");
1983 int len = (int)live_locals.size();
1984
1985 for (int i = 0; i < len; i++) {
1986 Node* local = unc->local(jvms, i);
1987 // kill local using the liveness of next_bci.
1988 // give up when the local looks like an operand to secure reexecution.
1989 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1990 uint idx = jvms->locoff() + i;
1991 #ifdef ASSERT
1992 if (PrintOpto && Verbose) {
1993 tty->print("[unstable_if] kill local#%d: ", idx);
1994 local->dump();
1995 tty->cr();
1996 }
1997 #endif
1998 igvn.replace_input_of(unc, idx, top());
1999 modified = true;
2000 }
2001 }
2002 }
2003
2004 // keep the mondified trap for late query
2005 if (modified) {
2006 trap->set_modified();
2007 } else {
2008 _unstable_if_traps.delete_at(i);
2009 }
2010 }
2011 igvn.optimize();
2012 }
2013
2014 // StringOpts and late inlining of string methods
2015 void Compile::inline_string_calls(bool parse_time) {
2016 {
2017 // remove useless nodes to make the usage analysis simpler
2018 ResourceMark rm;
2019 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2020 }
2021
2022 {
2023 ResourceMark rm;
2024 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2025 PhaseStringOpts pso(initial_gvn());
2026 print_method(PHASE_AFTER_STRINGOPTS, 3);
2027 }
2028
2029 // now inline anything that we skipped the first time around
2030 if (!parse_time) {
2031 _late_inlines_pos = _late_inlines.length();
2032 }
2033
2034 while (_string_late_inlines.length() > 0) {
2035 CallGenerator* cg = _string_late_inlines.pop();
2036 cg->do_late_inline();
2037 if (failing()) return;
2038 }
2039 _string_late_inlines.trunc_to(0);
2040 }
2041
2042 // Late inlining of boxing methods
2043 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2044 if (_boxing_late_inlines.length() > 0) {
2045 assert(has_boxed_value(), "inconsistent");
2046
2047 set_inlining_incrementally(true);
2048
2049 igvn_worklist()->ensure_empty(); // should be done with igvn
2050
2051 _late_inlines_pos = _late_inlines.length();
2052
2053 while (_boxing_late_inlines.length() > 0) {
2054 CallGenerator* cg = _boxing_late_inlines.pop();
2055 cg->do_late_inline();
2056 if (failing()) return;
2057 }
2058 _boxing_late_inlines.trunc_to(0);
2059
2060 inline_incrementally_cleanup(igvn);
2061
2062 set_inlining_incrementally(false);
2063 }
2064 }
2065
2066 bool Compile::inline_incrementally_one() {
2067 assert(IncrementalInline, "incremental inlining should be on");
2068 assert(_late_inlines.length() > 0, "should have been checked by caller");
2069
2070 TracePhase tp(_t_incrInline_inline);
2071
2072 set_inlining_progress(false);
2073 set_do_cleanup(false);
2074
2075 for (int i = 0; i < _late_inlines.length(); i++) {
2076 _late_inlines_pos = i+1;
2077 CallGenerator* cg = _late_inlines.at(i);
2078 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2079 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2080 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2081 if (should_stress_inlining()) {
2082 // randomly add repeated inline attempt if stress-inlining
2083 cg->call_node()->set_generator(cg);
2084 C->igvn_worklist()->push(cg->call_node());
2085 continue;
2086 }
2087 cg->do_late_inline();
2088 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2089 if (failing()) {
2090 return false;
2091 } else if (inlining_progress()) {
2092 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2093 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2094 break; // process one call site at a time
2095 } else {
2096 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2097 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2098 // Avoid potential infinite loop if node already in the IGVN list
2099 assert(false, "scheduled for IGVN during inlining attempt");
2100 } else {
2101 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2102 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2103 cg->call_node()->set_generator(cg);
2104 }
2105 }
2106 } else {
2107 // Ignore late inline direct calls when inlining is not allowed.
2108 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2109 }
2110 }
2111 // Remove processed elements.
2112 _late_inlines.remove_till(_late_inlines_pos);
2113 _late_inlines_pos = 0;
2114
2115 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2116
2117 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2118
2119 set_inlining_progress(false);
2120 set_do_cleanup(false);
2121
2122 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2123 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2124 }
2125
2126 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2127 {
2128 TracePhase tp(_t_incrInline_pru);
2129 ResourceMark rm;
2130 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2131 }
2132 {
2133 TracePhase tp(_t_incrInline_igvn);
2134 igvn.reset();
2135 igvn.optimize();
2136 if (failing()) return;
2137 }
2138 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2139 }
2140
2141 template<typename E>
2142 static void shuffle_array(Compile& C, GrowableArray<E>& array) {
2143 if (array.length() < 2) {
2144 return;
2145 }
2146 for (uint i = array.length() - 1; i >= 1; i--) {
2147 uint j = C.random() % (i + 1);
2148 swap(array.at(i), array.at(j));
2149 }
2150 }
2151
2152 void Compile::shuffle_late_inlines() {
2153 shuffle_array(*C, _late_inlines);
2154 }
2155
2156 // Perform incremental inlining until bound on number of live nodes is reached
2157 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2158 TracePhase tp(_t_incrInline);
2159
2160 set_inlining_incrementally(true);
2161 uint low_live_nodes = 0;
2162
2163 if (StressIncrementalInlining) {
2164 shuffle_late_inlines();
2165 }
2166
2167 while (_late_inlines.length() > 0) {
2168 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2169 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2170 TracePhase tp(_t_incrInline_ideal);
2171 // PhaseIdealLoop is expensive so we only try it once we are
2172 // out of live nodes and we only try it again if the previous
2173 // helped got the number of nodes down significantly
2174 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2175 if (failing()) return;
2176 low_live_nodes = live_nodes();
2177 _major_progress = true;
2178 }
2179
2180 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2181 bool do_print_inlining = print_inlining() || print_intrinsics();
2182 if (do_print_inlining || log() != nullptr) {
2183 // Print inlining message for candidates that we couldn't inline for lack of space.
2184 for (int i = 0; i < _late_inlines.length(); i++) {
2185 CallGenerator* cg = _late_inlines.at(i);
2186 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2187 if (do_print_inlining) {
2188 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2189 }
2190 log_late_inline_failure(cg, msg);
2191 }
2192 }
2193 break; // finish
2194 }
2195 }
2196
2197 igvn_worklist()->ensure_empty(); // should be done with igvn
2198
2199 if (_late_inlines.length() == 0) {
2200 break; // no more progress
2201 }
2202
2203 while (inline_incrementally_one()) {
2204 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2205 }
2206 if (failing()) return;
2207
2208 inline_incrementally_cleanup(igvn);
2209
2210 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2211
2212 if (failing()) return;
2213 }
2214
2215 igvn_worklist()->ensure_empty(); // should be done with igvn
2216
2217 if (_string_late_inlines.length() > 0) {
2218 assert(has_stringbuilder(), "inconsistent");
2219
2220 inline_string_calls(false);
2221
2222 if (failing()) return;
2223
2224 inline_incrementally_cleanup(igvn);
2225 }
2226
2227 set_inlining_incrementally(false);
2228 }
2229
2230 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2231 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2232 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2233 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2234 // as if "inlining_incrementally() == true" were set.
2235 assert(inlining_incrementally() == false, "not allowed");
2236 assert(_modified_nodes == nullptr, "not allowed");
2237 assert(_late_inlines.length() > 0, "sanity");
2238
2239 if (StressIncrementalInlining) {
2240 shuffle_late_inlines();
2241 }
2242
2243 while (_late_inlines.length() > 0) {
2244 igvn_worklist()->ensure_empty(); // should be done with igvn
2245
2246 while (inline_incrementally_one()) {
2247 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2248 }
2249 if (failing()) return;
2250
2251 inline_incrementally_cleanup(igvn);
2252 }
2253 }
2254
2255 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2256 if (_loop_opts_cnt > 0) {
2257 while (major_progress() && (_loop_opts_cnt > 0)) {
2258 TracePhase tp(_t_idealLoop);
2259 PhaseIdealLoop::optimize(igvn, mode);
2260 _loop_opts_cnt--;
2261 if (failing()) return false;
2262 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2263 }
2264 }
2265 return true;
2266 }
2267
2268 // Remove edges from "root" to each SafePoint at a backward branch.
2269 // They were inserted during parsing (see add_safepoint()) to make
2270 // infinite loops without calls or exceptions visible to root, i.e.,
2271 // useful.
2272 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2273 Node *r = root();
2274 if (r != nullptr) {
2275 for (uint i = r->req(); i < r->len(); ++i) {
2276 Node *n = r->in(i);
2277 if (n != nullptr && n->is_SafePoint()) {
2278 r->rm_prec(i);
2279 if (n->outcnt() == 0) {
2280 igvn.remove_dead_node(n);
2281 }
2282 --i;
2283 }
2284 }
2285 // Parsing may have added top inputs to the root node (Path
2286 // leading to the Halt node proven dead). Make sure we get a
2287 // chance to clean them up.
2288 igvn._worklist.push(r);
2289 igvn.optimize();
2290 }
2291 }
2292
2293 //------------------------------Optimize---------------------------------------
2294 // Given a graph, optimize it.
2295 void Compile::Optimize() {
2296 TracePhase tp(_t_optimizer);
2297
2298 #ifndef PRODUCT
2299 if (env()->break_at_compile()) {
2300 BREAKPOINT;
2301 }
2302
2303 #endif
2304
2305 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2306 #ifdef ASSERT
2307 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2308 #endif
2309
2310 ResourceMark rm;
2311
2312 NOT_PRODUCT( verify_graph_edges(); )
2313
2314 print_method(PHASE_AFTER_PARSING, 1);
2315
2316 {
2317 // Iterative Global Value Numbering, including ideal transforms
2318 PhaseIterGVN igvn;
2319 #ifdef ASSERT
2320 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2321 #endif
2322 {
2323 TracePhase tp(_t_iterGVN);
2324 igvn.optimize();
2325 }
2326
2327 if (failing()) return;
2328
2329 print_method(PHASE_ITER_GVN1, 2);
2330
2331 process_for_unstable_if_traps(igvn);
2332
2333 if (failing()) return;
2334
2335 inline_incrementally(igvn);
2336
2337 print_method(PHASE_INCREMENTAL_INLINE, 2);
2338
2339 if (failing()) return;
2340
2341 if (eliminate_boxing()) {
2342 // Inline valueOf() methods now.
2343 inline_boxing_calls(igvn);
2344
2345 if (failing()) return;
2346
2347 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2348 inline_incrementally(igvn);
2349 }
2350
2351 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2352
2353 if (failing()) return;
2354 }
2355
2356 // Remove the speculative part of types and clean up the graph from
2357 // the extra CastPP nodes whose only purpose is to carry them. Do
2358 // that early so that optimizations are not disrupted by the extra
2359 // CastPP nodes.
2360 remove_speculative_types(igvn);
2361
2362 if (failing()) return;
2363
2364 // No more new expensive nodes will be added to the list from here
2365 // so keep only the actual candidates for optimizations.
2366 cleanup_expensive_nodes(igvn);
2367
2368 if (failing()) return;
2369
2370 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2371 if (EnableVectorSupport && has_vbox_nodes()) {
2372 TracePhase tp(_t_vector);
2373 PhaseVector pv(igvn);
2374 pv.optimize_vector_boxes();
2375 if (failing()) return;
2376 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2377 }
2378 assert(!has_vbox_nodes(), "sanity");
2379
2380 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2381 Compile::TracePhase tp(_t_renumberLive);
2382 igvn_worklist()->ensure_empty(); // should be done with igvn
2383 {
2384 ResourceMark rm;
2385 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2386 }
2387 igvn.reset();
2388 igvn.optimize();
2389 if (failing()) return;
2390 }
2391
2392 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2393 // safepoints
2394 remove_root_to_sfpts_edges(igvn);
2395
2396 if (failing()) return;
2397
2398 _print_phase_loop_opts = has_loops();
2399 if (_print_phase_loop_opts) {
2400 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2401 }
2402
2403 // Perform escape analysis
2404 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2405 if (has_loops()) {
2406 // Cleanup graph (remove dead nodes).
2407 TracePhase tp(_t_idealLoop);
2408 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2409 if (failing()) return;
2410 }
2411 bool progress;
2412 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2413 do {
2414 ConnectionGraph::do_analysis(this, &igvn);
2415
2416 if (failing()) return;
2417
2418 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2419
2420 // Optimize out fields loads from scalar replaceable allocations.
2421 igvn.optimize();
2422 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2423
2424 if (failing()) return;
2425
2426 if (congraph() != nullptr && macro_count() > 0) {
2427 TracePhase tp(_t_macroEliminate);
2428 PhaseMacroExpand mexp(igvn);
2429 mexp.eliminate_macro_nodes();
2430 if (failing()) return;
2431 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2432
2433 igvn.set_delay_transform(false);
2434 igvn.optimize();
2435 if (failing()) return;
2436
2437 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2438 }
2439
2440 ConnectionGraph::verify_ram_nodes(this, root());
2441 if (failing()) return;
2442
2443 progress = do_iterative_escape_analysis() &&
2444 (macro_count() < mcount) &&
2445 ConnectionGraph::has_candidates(this);
2446 // Try again if candidates exist and made progress
2447 // by removing some allocations and/or locks.
2448 } while (progress);
2449 }
2450
2451 // Loop transforms on the ideal graph. Range Check Elimination,
2452 // peeling, unrolling, etc.
2453
2454 // Set loop opts counter
2455 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2456 {
2457 TracePhase tp(_t_idealLoop);
2458 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2459 _loop_opts_cnt--;
2460 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2461 if (failing()) return;
2462 }
2463 // Loop opts pass if partial peeling occurred in previous pass
2464 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2465 TracePhase tp(_t_idealLoop);
2466 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2467 _loop_opts_cnt--;
2468 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2469 if (failing()) return;
2470 }
2471 // Loop opts pass for loop-unrolling before CCP
2472 if(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_PHASEIDEALLOOP3, 2);
2477 }
2478 if (!failing()) {
2479 // Verify that last round of loop opts produced a valid graph
2480 PhaseIdealLoop::verify(igvn);
2481 }
2482 }
2483 if (failing()) return;
2484
2485 // Conditional Constant Propagation;
2486 print_method(PHASE_BEFORE_CCP1, 2);
2487 PhaseCCP ccp( &igvn );
2488 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2489 {
2490 TracePhase tp(_t_ccp);
2491 ccp.do_transform();
2492 }
2493 print_method(PHASE_CCP1, 2);
2494
2495 assert( true, "Break here to ccp.dump_old2new_map()");
2496
2497 // Iterative Global Value Numbering, including ideal transforms
2498 {
2499 TracePhase tp(_t_iterGVN2);
2500 igvn.reset_from_igvn(&ccp);
2501 igvn.optimize();
2502 }
2503 print_method(PHASE_ITER_GVN2, 2);
2504
2505 if (failing()) return;
2506
2507 // Loop transforms on the ideal graph. Range Check Elimination,
2508 // peeling, unrolling, etc.
2509 if (!optimize_loops(igvn, LoopOptsDefault)) {
2510 return;
2511 }
2512
2513 if (failing()) return;
2514
2515 C->clear_major_progress(); // ensure that major progress is now clear
2516
2517 process_for_post_loop_opts_igvn(igvn);
2518
2519 process_for_merge_stores_igvn(igvn);
2520
2521 if (failing()) return;
2522
2523 #ifdef ASSERT
2524 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2525 #endif
2526
2527 {
2528 TracePhase tp(_t_macroExpand);
2529 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2530 PhaseMacroExpand mex(igvn);
2531 // Do not allow new macro nodes once we start to eliminate and expand
2532 C->reset_allow_macro_nodes();
2533 // Last attempt to eliminate macro nodes before expand
2534 mex.eliminate_macro_nodes();
2535 if (failing()) {
2536 return;
2537 }
2538 mex.eliminate_opaque_looplimit_macro_nodes();
2539 if (failing()) {
2540 return;
2541 }
2542 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2543 if (mex.expand_macro_nodes()) {
2544 assert(failing(), "must bail out w/ explicit message");
2545 return;
2546 }
2547 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2548 }
2549
2550 {
2551 TracePhase tp(_t_barrierExpand);
2552 if (bs->expand_barriers(this, igvn)) {
2553 assert(failing(), "must bail out w/ explicit message");
2554 return;
2555 }
2556 print_method(PHASE_BARRIER_EXPANSION, 2);
2557 }
2558
2559 if (C->max_vector_size() > 0) {
2560 C->optimize_logic_cones(igvn);
2561 igvn.optimize();
2562 if (failing()) return;
2563 }
2564
2565 DEBUG_ONLY( _modified_nodes = nullptr; )
2566
2567 assert(igvn._worklist.size() == 0, "not empty");
2568
2569 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2570
2571 if (_late_inlines.length() > 0) {
2572 // More opportunities to optimize virtual and MH calls.
2573 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2574 process_late_inline_calls_no_inline(igvn);
2575 if (failing()) return;
2576 }
2577 } // (End scope of igvn; run destructor if necessary for asserts.)
2578
2579 check_no_dead_use();
2580
2581 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2582 // to remove hashes to unlock nodes for modifications.
2583 C->node_hash()->clear();
2584
2585 // A method with only infinite loops has no edges entering loops from root
2586 {
2587 TracePhase tp(_t_graphReshaping);
2588 if (final_graph_reshaping()) {
2589 assert(failing(), "must bail out w/ explicit message");
2590 return;
2591 }
2592 }
2593
2594 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2595 DEBUG_ONLY(set_phase_optimize_finished();)
2596 }
2597
2598 #ifdef ASSERT
2599 void Compile::check_no_dead_use() const {
2600 ResourceMark rm;
2601 Unique_Node_List wq;
2602 wq.push(root());
2603 for (uint i = 0; i < wq.size(); ++i) {
2604 Node* n = wq.at(i);
2605 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2606 Node* u = n->fast_out(j);
2607 if (u->outcnt() == 0 && !u->is_Con()) {
2608 u->dump();
2609 fatal("no reachable node should have no use");
2610 }
2611 wq.push(u);
2612 }
2613 }
2614 }
2615 #endif
2616
2617 void Compile::inline_vector_reboxing_calls() {
2618 if (C->_vector_reboxing_late_inlines.length() > 0) {
2619 _late_inlines_pos = C->_late_inlines.length();
2620 while (_vector_reboxing_late_inlines.length() > 0) {
2621 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2622 cg->do_late_inline();
2623 if (failing()) return;
2624 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2625 }
2626 _vector_reboxing_late_inlines.trunc_to(0);
2627 }
2628 }
2629
2630 bool Compile::has_vbox_nodes() {
2631 if (C->_vector_reboxing_late_inlines.length() > 0) {
2632 return true;
2633 }
2634 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2635 Node * n = C->macro_node(macro_idx);
2636 assert(n->is_macro(), "only macro nodes expected here");
2637 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2638 return true;
2639 }
2640 }
2641 return false;
2642 }
2643
2644 //---------------------------- Bitwise operation packing optimization ---------------------------
2645
2646 static bool is_vector_unary_bitwise_op(Node* n) {
2647 return n->Opcode() == Op_XorV &&
2648 VectorNode::is_vector_bitwise_not_pattern(n);
2649 }
2650
2651 static bool is_vector_binary_bitwise_op(Node* n) {
2652 switch (n->Opcode()) {
2653 case Op_AndV:
2654 case Op_OrV:
2655 return true;
2656
2657 case Op_XorV:
2658 return !is_vector_unary_bitwise_op(n);
2659
2660 default:
2661 return false;
2662 }
2663 }
2664
2665 static bool is_vector_ternary_bitwise_op(Node* n) {
2666 return n->Opcode() == Op_MacroLogicV;
2667 }
2668
2669 static bool is_vector_bitwise_op(Node* n) {
2670 return is_vector_unary_bitwise_op(n) ||
2671 is_vector_binary_bitwise_op(n) ||
2672 is_vector_ternary_bitwise_op(n);
2673 }
2674
2675 static bool is_vector_bitwise_cone_root(Node* n) {
2676 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2677 return false;
2678 }
2679 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2680 if (is_vector_bitwise_op(n->fast_out(i))) {
2681 return false;
2682 }
2683 }
2684 return true;
2685 }
2686
2687 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2688 uint cnt = 0;
2689 if (is_vector_bitwise_op(n)) {
2690 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2691 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2692 for (uint i = 1; i < inp_cnt; i++) {
2693 Node* in = n->in(i);
2694 bool skip = VectorNode::is_all_ones_vector(in);
2695 if (!skip && !inputs.member(in)) {
2696 inputs.push(in);
2697 cnt++;
2698 }
2699 }
2700 assert(cnt <= 1, "not unary");
2701 } else {
2702 uint last_req = inp_cnt;
2703 if (is_vector_ternary_bitwise_op(n)) {
2704 last_req = inp_cnt - 1; // skip last input
2705 }
2706 for (uint i = 1; i < last_req; i++) {
2707 Node* def = n->in(i);
2708 if (!inputs.member(def)) {
2709 inputs.push(def);
2710 cnt++;
2711 }
2712 }
2713 }
2714 } else { // not a bitwise operations
2715 if (!inputs.member(n)) {
2716 inputs.push(n);
2717 cnt++;
2718 }
2719 }
2720 return cnt;
2721 }
2722
2723 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2724 Unique_Node_List useful_nodes;
2725 C->identify_useful_nodes(useful_nodes);
2726
2727 for (uint i = 0; i < useful_nodes.size(); i++) {
2728 Node* n = useful_nodes.at(i);
2729 if (is_vector_bitwise_cone_root(n)) {
2730 list.push(n);
2731 }
2732 }
2733 }
2734
2735 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2736 const TypeVect* vt,
2737 Unique_Node_List& partition,
2738 Unique_Node_List& inputs) {
2739 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2740 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2741 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2742
2743 Node* in1 = inputs.at(0);
2744 Node* in2 = inputs.at(1);
2745 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2746
2747 uint func = compute_truth_table(partition, inputs);
2748
2749 Node* pn = partition.at(partition.size() - 1);
2750 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2751 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2752 }
2753
2754 static uint extract_bit(uint func, uint pos) {
2755 return (func & (1 << pos)) >> pos;
2756 }
2757
2758 //
2759 // A macro logic node represents a truth table. It has 4 inputs,
2760 // First three inputs corresponds to 3 columns of a truth table
2761 // and fourth input captures the logic function.
2762 //
2763 // eg. fn = (in1 AND in2) OR in3;
2764 //
2765 // MacroNode(in1,in2,in3,fn)
2766 //
2767 // -----------------
2768 // in1 in2 in3 fn
2769 // -----------------
2770 // 0 0 0 0
2771 // 0 0 1 1
2772 // 0 1 0 0
2773 // 0 1 1 1
2774 // 1 0 0 0
2775 // 1 0 1 1
2776 // 1 1 0 1
2777 // 1 1 1 1
2778 //
2779
2780 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2781 int res = 0;
2782 for (int i = 0; i < 8; i++) {
2783 int bit1 = extract_bit(in1, i);
2784 int bit2 = extract_bit(in2, i);
2785 int bit3 = extract_bit(in3, i);
2786
2787 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2788 int func_bit = extract_bit(func, func_bit_pos);
2789
2790 res |= func_bit << i;
2791 }
2792 return res;
2793 }
2794
2795 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
2796 assert(n != nullptr, "");
2797 assert(eval_map.contains(n), "absent");
2798 return *(eval_map.get(n));
2799 }
2800
2801 static void eval_operands(Node* n,
2802 uint& func1, uint& func2, uint& func3,
2803 HashTable<Node*,uint>& eval_map) {
2804 assert(is_vector_bitwise_op(n), "");
2805
2806 if (is_vector_unary_bitwise_op(n)) {
2807 Node* opnd = n->in(1);
2808 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2809 opnd = n->in(2);
2810 }
2811 func1 = eval_operand(opnd, eval_map);
2812 } else if (is_vector_binary_bitwise_op(n)) {
2813 func1 = eval_operand(n->in(1), eval_map);
2814 func2 = eval_operand(n->in(2), eval_map);
2815 } else {
2816 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2817 func1 = eval_operand(n->in(1), eval_map);
2818 func2 = eval_operand(n->in(2), eval_map);
2819 func3 = eval_operand(n->in(3), eval_map);
2820 }
2821 }
2822
2823 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2824 assert(inputs.size() <= 3, "sanity");
2825 ResourceMark rm;
2826 uint res = 0;
2827 HashTable<Node*,uint> eval_map;
2828
2829 // Populate precomputed functions for inputs.
2830 // Each input corresponds to one column of 3 input truth-table.
2831 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2832 0xCC, // (_, b, _) -> b
2833 0xF0 }; // (a, _, _) -> a
2834 for (uint i = 0; i < inputs.size(); i++) {
2835 eval_map.put(inputs.at(i), input_funcs[2-i]);
2836 }
2837
2838 for (uint i = 0; i < partition.size(); i++) {
2839 Node* n = partition.at(i);
2840
2841 uint func1 = 0, func2 = 0, func3 = 0;
2842 eval_operands(n, func1, func2, func3, eval_map);
2843
2844 switch (n->Opcode()) {
2845 case Op_OrV:
2846 assert(func3 == 0, "not binary");
2847 res = func1 | func2;
2848 break;
2849 case Op_AndV:
2850 assert(func3 == 0, "not binary");
2851 res = func1 & func2;
2852 break;
2853 case Op_XorV:
2854 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2855 assert(func2 == 0 && func3 == 0, "not unary");
2856 res = (~func1) & 0xFF;
2857 } else {
2858 assert(func3 == 0, "not binary");
2859 res = func1 ^ func2;
2860 }
2861 break;
2862 case Op_MacroLogicV:
2863 // Ordering of inputs may change during evaluation of sub-tree
2864 // containing MacroLogic node as a child node, thus a re-evaluation
2865 // makes sure that function is evaluated in context of current
2866 // inputs.
2867 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2868 break;
2869
2870 default: assert(false, "not supported: %s", n->Name());
2871 }
2872 assert(res <= 0xFF, "invalid");
2873 eval_map.put(n, res);
2874 }
2875 return res;
2876 }
2877
2878 // Criteria under which nodes gets packed into a macro logic node:-
2879 // 1) Parent and both child nodes are all unmasked or masked with
2880 // same predicates.
2881 // 2) Masked parent can be packed with left child if it is predicated
2882 // and both have same predicates.
2883 // 3) Masked parent can be packed with right child if its un-predicated
2884 // or has matching predication condition.
2885 // 4) An unmasked parent can be packed with an unmasked child.
2886 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2887 assert(partition.size() == 0, "not empty");
2888 assert(inputs.size() == 0, "not empty");
2889 if (is_vector_ternary_bitwise_op(n)) {
2890 return false;
2891 }
2892
2893 bool is_unary_op = is_vector_unary_bitwise_op(n);
2894 if (is_unary_op) {
2895 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2896 return false; // too few inputs
2897 }
2898
2899 bool pack_left_child = true;
2900 bool pack_right_child = true;
2901
2902 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2903 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2904
2905 int left_child_input_cnt = 0;
2906 int right_child_input_cnt = 0;
2907
2908 bool parent_is_predicated = n->is_predicated_vector();
2909 bool left_child_predicated = n->in(1)->is_predicated_vector();
2910 bool right_child_predicated = n->in(2)->is_predicated_vector();
2911
2912 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2913 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2914 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2915
2916 do {
2917 if (pack_left_child && left_child_LOP &&
2918 ((!parent_is_predicated && !left_child_predicated) ||
2919 ((parent_is_predicated && left_child_predicated &&
2920 parent_pred == left_child_pred)))) {
2921 partition.push(n->in(1));
2922 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2923 } else {
2924 inputs.push(n->in(1));
2925 left_child_input_cnt = 1;
2926 }
2927
2928 if (pack_right_child && right_child_LOP &&
2929 (!right_child_predicated ||
2930 (right_child_predicated && parent_is_predicated &&
2931 parent_pred == right_child_pred))) {
2932 partition.push(n->in(2));
2933 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2934 } else {
2935 inputs.push(n->in(2));
2936 right_child_input_cnt = 1;
2937 }
2938
2939 if (inputs.size() > 3) {
2940 assert(partition.size() > 0, "");
2941 inputs.clear();
2942 partition.clear();
2943 if (left_child_input_cnt > right_child_input_cnt) {
2944 pack_left_child = false;
2945 } else {
2946 pack_right_child = false;
2947 }
2948 } else {
2949 break;
2950 }
2951 } while(true);
2952
2953 if(partition.size()) {
2954 partition.push(n);
2955 }
2956
2957 return (partition.size() == 2 || partition.size() == 3) &&
2958 (inputs.size() == 2 || inputs.size() == 3);
2959 }
2960
2961 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2962 assert(is_vector_bitwise_op(n), "not a root");
2963
2964 visited.set(n->_idx);
2965
2966 // 1) Do a DFS walk over the logic cone.
2967 for (uint i = 1; i < n->req(); i++) {
2968 Node* in = n->in(i);
2969 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2970 process_logic_cone_root(igvn, in, visited);
2971 }
2972 }
2973
2974 // 2) Bottom up traversal: Merge node[s] with
2975 // the parent to form macro logic node.
2976 Unique_Node_List partition;
2977 Unique_Node_List inputs;
2978 if (compute_logic_cone(n, partition, inputs)) {
2979 const TypeVect* vt = n->bottom_type()->is_vect();
2980 Node* pn = partition.at(partition.size() - 1);
2981 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2982 if (mask == nullptr ||
2983 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2984 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2985 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2986 igvn.replace_node(n, macro_logic);
2987 }
2988 }
2989 }
2990
2991 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2992 ResourceMark rm;
2993 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2994 Unique_Node_List list;
2995 collect_logic_cone_roots(list);
2996
2997 while (list.size() > 0) {
2998 Node* n = list.pop();
2999 const TypeVect* vt = n->bottom_type()->is_vect();
3000 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3001 if (supported) {
3002 VectorSet visited(comp_arena());
3003 process_logic_cone_root(igvn, n, visited);
3004 }
3005 }
3006 }
3007 }
3008
3009 //------------------------------Code_Gen---------------------------------------
3010 // Given a graph, generate code for it
3011 void Compile::Code_Gen() {
3012 if (failing()) {
3013 return;
3014 }
3015
3016 // Perform instruction selection. You might think we could reclaim Matcher
3017 // memory PDQ, but actually the Matcher is used in generating spill code.
3018 // Internals of the Matcher (including some VectorSets) must remain live
3019 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3020 // set a bit in reclaimed memory.
3021
3022 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3023 // nodes. Mapping is only valid at the root of each matched subtree.
3024 NOT_PRODUCT( verify_graph_edges(); )
3025
3026 Matcher matcher;
3027 _matcher = &matcher;
3028 {
3029 TracePhase tp(_t_matcher);
3030 matcher.match();
3031 if (failing()) {
3032 return;
3033 }
3034 }
3035 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3036 // nodes. Mapping is only valid at the root of each matched subtree.
3037 NOT_PRODUCT( verify_graph_edges(); )
3038
3039 // If you have too many nodes, or if matching has failed, bail out
3040 check_node_count(0, "out of nodes matching instructions");
3041 if (failing()) {
3042 return;
3043 }
3044
3045 print_method(PHASE_MATCHING, 2);
3046
3047 // Build a proper-looking CFG
3048 PhaseCFG cfg(node_arena(), root(), matcher);
3049 if (failing()) {
3050 return;
3051 }
3052 _cfg = &cfg;
3053 {
3054 TracePhase tp(_t_scheduler);
3055 bool success = cfg.do_global_code_motion();
3056 if (!success) {
3057 return;
3058 }
3059
3060 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3061 NOT_PRODUCT( verify_graph_edges(); )
3062 cfg.verify();
3063 if (failing()) {
3064 return;
3065 }
3066 }
3067
3068 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3069 _regalloc = ®alloc;
3070 {
3071 TracePhase tp(_t_registerAllocation);
3072 // Perform register allocation. After Chaitin, use-def chains are
3073 // no longer accurate (at spill code) and so must be ignored.
3074 // Node->LRG->reg mappings are still accurate.
3075 _regalloc->Register_Allocate();
3076
3077 // Bail out if the allocator builds too many nodes
3078 if (failing()) {
3079 return;
3080 }
3081
3082 print_method(PHASE_REGISTER_ALLOCATION, 2);
3083 }
3084
3085 // Prior to register allocation we kept empty basic blocks in case the
3086 // the allocator needed a place to spill. After register allocation we
3087 // are not adding any new instructions. If any basic block is empty, we
3088 // can now safely remove it.
3089 {
3090 TracePhase tp(_t_blockOrdering);
3091 cfg.remove_empty_blocks();
3092 if (do_freq_based_layout()) {
3093 PhaseBlockLayout layout(cfg);
3094 } else {
3095 cfg.set_loop_alignment();
3096 }
3097 cfg.fixup_flow();
3098 cfg.remove_unreachable_blocks();
3099 cfg.verify_dominator_tree();
3100 print_method(PHASE_BLOCK_ORDERING, 3);
3101 }
3102
3103 // Apply peephole optimizations
3104 if( OptoPeephole ) {
3105 TracePhase tp(_t_peephole);
3106 PhasePeephole peep( _regalloc, cfg);
3107 peep.do_transform();
3108 print_method(PHASE_PEEPHOLE, 3);
3109 }
3110
3111 // Do late expand if CPU requires this.
3112 if (Matcher::require_postalloc_expand) {
3113 TracePhase tp(_t_postalloc_expand);
3114 cfg.postalloc_expand(_regalloc);
3115 print_method(PHASE_POSTALLOC_EXPAND, 3);
3116 }
3117
3118 #ifdef ASSERT
3119 {
3120 CompilationMemoryStatistic::do_test_allocations();
3121 if (failing()) return;
3122 }
3123 #endif
3124
3125 // Convert Nodes to instruction bits in a buffer
3126 {
3127 TracePhase tp(_t_output);
3128 PhaseOutput output;
3129 output.Output();
3130 if (failing()) return;
3131 output.install();
3132 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3133 }
3134
3135 // He's dead, Jim.
3136 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3137 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3138 }
3139
3140 //------------------------------Final_Reshape_Counts---------------------------
3141 // This class defines counters to help identify when a method
3142 // may/must be executed using hardware with only 24-bit precision.
3143 struct Final_Reshape_Counts : public StackObj {
3144 int _call_count; // count non-inlined 'common' calls
3145 int _float_count; // count float ops requiring 24-bit precision
3146 int _double_count; // count double ops requiring more precision
3147 int _java_call_count; // count non-inlined 'java' calls
3148 int _inner_loop_count; // count loops which need alignment
3149 VectorSet _visited; // Visitation flags
3150 Node_List _tests; // Set of IfNodes & PCTableNodes
3151
3152 Final_Reshape_Counts() :
3153 _call_count(0), _float_count(0), _double_count(0),
3154 _java_call_count(0), _inner_loop_count(0) { }
3155
3156 void inc_call_count () { _call_count ++; }
3157 void inc_float_count () { _float_count ++; }
3158 void inc_double_count() { _double_count++; }
3159 void inc_java_call_count() { _java_call_count++; }
3160 void inc_inner_loop_count() { _inner_loop_count++; }
3161
3162 int get_call_count () const { return _call_count ; }
3163 int get_float_count () const { return _float_count ; }
3164 int get_double_count() const { return _double_count; }
3165 int get_java_call_count() const { return _java_call_count; }
3166 int get_inner_loop_count() const { return _inner_loop_count; }
3167 };
3168
3169 //------------------------------final_graph_reshaping_impl----------------------
3170 // Implement items 1-5 from final_graph_reshaping below.
3171 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3172
3173 if ( n->outcnt() == 0 ) return; // dead node
3174 uint nop = n->Opcode();
3175
3176 // Check for 2-input instruction with "last use" on right input.
3177 // Swap to left input. Implements item (2).
3178 if( n->req() == 3 && // two-input instruction
3179 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3180 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3181 n->in(2)->outcnt() == 1 &&// right use IS a last use
3182 !n->in(2)->is_Con() ) { // right use is not a constant
3183 // Check for commutative opcode
3184 switch( nop ) {
3185 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3186 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3187 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3188 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3189 case Op_AndL: case Op_XorL: case Op_OrL:
3190 case Op_AndI: case Op_XorI: case Op_OrI: {
3191 // Move "last use" input to left by swapping inputs
3192 n->swap_edges(1, 2);
3193 break;
3194 }
3195 default:
3196 break;
3197 }
3198 }
3199
3200 #ifdef ASSERT
3201 if( n->is_Mem() ) {
3202 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3203 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3204 // oop will be recorded in oop map if load crosses safepoint
3205 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3206 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3207 "raw memory operations should have control edge");
3208 }
3209 if (n->is_MemBar()) {
3210 MemBarNode* mb = n->as_MemBar();
3211 if (mb->trailing_store() || mb->trailing_load_store()) {
3212 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3213 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3214 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3215 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3216 } else if (mb->leading()) {
3217 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3218 }
3219 }
3220 #endif
3221 // Count FPU ops and common calls, implements item (3)
3222 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3223 if (!gc_handled) {
3224 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3225 }
3226
3227 // Collect CFG split points
3228 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3229 frc._tests.push(n);
3230 }
3231 }
3232
3233 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3234 if (!UseDivMod) {
3235 return;
3236 }
3237
3238 // Check if "a % b" and "a / b" both exist
3239 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3240 if (d == nullptr) {
3241 return;
3242 }
3243
3244 // Replace them with a fused divmod if supported
3245 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3246 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3247 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3248 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3249 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3250 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3251 // DivMod node so the dependency is not lost.
3252 divmod->add_prec_from(n);
3253 divmod->add_prec_from(d);
3254 d->subsume_by(divmod->div_proj(), this);
3255 n->subsume_by(divmod->mod_proj(), this);
3256 } else {
3257 // Replace "a % b" with "a - ((a / b) * b)"
3258 Node* mult = MulNode::make(d, d->in(2), bt);
3259 Node* sub = SubNode::make(d->in(1), mult, bt);
3260 n->subsume_by(sub, this);
3261 }
3262 }
3263
3264 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3265 switch( nop ) {
3266 // Count all float operations that may use FPU
3267 case Op_AddF:
3268 case Op_SubF:
3269 case Op_MulF:
3270 case Op_DivF:
3271 case Op_NegF:
3272 case Op_ModF:
3273 case Op_ConvI2F:
3274 case Op_ConF:
3275 case Op_CmpF:
3276 case Op_CmpF3:
3277 case Op_StoreF:
3278 case Op_LoadF:
3279 // case Op_ConvL2F: // longs are split into 32-bit halves
3280 frc.inc_float_count();
3281 break;
3282
3283 case Op_ConvF2D:
3284 case Op_ConvD2F:
3285 frc.inc_float_count();
3286 frc.inc_double_count();
3287 break;
3288
3289 // Count all double operations that may use FPU
3290 case Op_AddD:
3291 case Op_SubD:
3292 case Op_MulD:
3293 case Op_DivD:
3294 case Op_NegD:
3295 case Op_ModD:
3296 case Op_ConvI2D:
3297 case Op_ConvD2I:
3298 // case Op_ConvL2D: // handled by leaf call
3299 // case Op_ConvD2L: // handled by leaf call
3300 case Op_ConD:
3301 case Op_CmpD:
3302 case Op_CmpD3:
3303 case Op_StoreD:
3304 case Op_LoadD:
3305 case Op_LoadD_unaligned:
3306 frc.inc_double_count();
3307 break;
3308 case Op_Opaque1: // Remove Opaque Nodes before matching
3309 n->subsume_by(n->in(1), this);
3310 break;
3311 case Op_CallLeafPure: {
3312 // If the pure call is not supported, then lower to a CallLeaf.
3313 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3314 CallNode* call = n->as_Call();
3315 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3316 call->_name, TypeRawPtr::BOTTOM);
3317 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3318 new_call->init_req(TypeFunc::I_O, C->top());
3319 new_call->init_req(TypeFunc::Memory, C->top());
3320 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3321 new_call->init_req(TypeFunc::FramePtr, C->top());
3322 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
3323 new_call->init_req(i, call->in(i));
3324 }
3325 n->subsume_by(new_call, this);
3326 }
3327 frc.inc_call_count();
3328 break;
3329 }
3330 case Op_CallStaticJava:
3331 case Op_CallJava:
3332 case Op_CallDynamicJava:
3333 frc.inc_java_call_count(); // Count java call site;
3334 case Op_CallRuntime:
3335 case Op_CallLeaf:
3336 case Op_CallLeafVector:
3337 case Op_CallLeafNoFP: {
3338 assert (n->is_Call(), "");
3339 CallNode *call = n->as_Call();
3340 // Count call sites where the FP mode bit would have to be flipped.
3341 // Do not count uncommon runtime calls:
3342 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3343 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3344 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3345 frc.inc_call_count(); // Count the call site
3346 } else { // See if uncommon argument is shared
3347 Node *n = call->in(TypeFunc::Parms);
3348 int nop = n->Opcode();
3349 // Clone shared simple arguments to uncommon calls, item (1).
3350 if (n->outcnt() > 1 &&
3351 !n->is_Proj() &&
3352 nop != Op_CreateEx &&
3353 nop != Op_CheckCastPP &&
3354 nop != Op_DecodeN &&
3355 nop != Op_DecodeNKlass &&
3356 !n->is_Mem() &&
3357 !n->is_Phi()) {
3358 Node *x = n->clone();
3359 call->set_req(TypeFunc::Parms, x);
3360 }
3361 }
3362 break;
3363 }
3364 case Op_StoreB:
3365 case Op_StoreC:
3366 case Op_StoreI:
3367 case Op_StoreL:
3368 case Op_CompareAndSwapB:
3369 case Op_CompareAndSwapS:
3370 case Op_CompareAndSwapI:
3371 case Op_CompareAndSwapL:
3372 case Op_CompareAndSwapP:
3373 case Op_CompareAndSwapN:
3374 case Op_WeakCompareAndSwapB:
3375 case Op_WeakCompareAndSwapS:
3376 case Op_WeakCompareAndSwapI:
3377 case Op_WeakCompareAndSwapL:
3378 case Op_WeakCompareAndSwapP:
3379 case Op_WeakCompareAndSwapN:
3380 case Op_CompareAndExchangeB:
3381 case Op_CompareAndExchangeS:
3382 case Op_CompareAndExchangeI:
3383 case Op_CompareAndExchangeL:
3384 case Op_CompareAndExchangeP:
3385 case Op_CompareAndExchangeN:
3386 case Op_GetAndAddS:
3387 case Op_GetAndAddB:
3388 case Op_GetAndAddI:
3389 case Op_GetAndAddL:
3390 case Op_GetAndSetS:
3391 case Op_GetAndSetB:
3392 case Op_GetAndSetI:
3393 case Op_GetAndSetL:
3394 case Op_GetAndSetP:
3395 case Op_GetAndSetN:
3396 case Op_StoreP:
3397 case Op_StoreN:
3398 case Op_StoreNKlass:
3399 case Op_LoadB:
3400 case Op_LoadUB:
3401 case Op_LoadUS:
3402 case Op_LoadI:
3403 case Op_LoadKlass:
3404 case Op_LoadNKlass:
3405 case Op_LoadL:
3406 case Op_LoadL_unaligned:
3407 case Op_LoadP:
3408 case Op_LoadN:
3409 case Op_LoadRange:
3410 case Op_LoadS:
3411 break;
3412
3413 case Op_AddP: { // Assert sane base pointers
3414 Node *addp = n->in(AddPNode::Address);
3415 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
3416 #ifdef _LP64
3417 if ((UseCompressedOops || UseCompressedClassPointers) &&
3418 addp->Opcode() == Op_ConP &&
3419 addp == n->in(AddPNode::Base) &&
3420 n->in(AddPNode::Offset)->is_Con()) {
3421 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3422 // on the platform and on the compressed oops mode.
3423 // Use addressing with narrow klass to load with offset on x86.
3424 // Some platforms can use the constant pool to load ConP.
3425 // Do this transformation here since IGVN will convert ConN back to ConP.
3426 const Type* t = addp->bottom_type();
3427 bool is_oop = t->isa_oopptr() != nullptr;
3428 bool is_klass = t->isa_klassptr() != nullptr;
3429
3430 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3431 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3432 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3433 Node* nn = nullptr;
3434
3435 int op = is_oop ? Op_ConN : Op_ConNKlass;
3436
3437 // Look for existing ConN node of the same exact type.
3438 Node* r = root();
3439 uint cnt = r->outcnt();
3440 for (uint i = 0; i < cnt; i++) {
3441 Node* m = r->raw_out(i);
3442 if (m!= nullptr && m->Opcode() == op &&
3443 m->bottom_type()->make_ptr() == t) {
3444 nn = m;
3445 break;
3446 }
3447 }
3448 if (nn != nullptr) {
3449 // Decode a narrow oop to match address
3450 // [R12 + narrow_oop_reg<<3 + offset]
3451 if (is_oop) {
3452 nn = new DecodeNNode(nn, t);
3453 } else {
3454 nn = new DecodeNKlassNode(nn, t);
3455 }
3456 // Check for succeeding AddP which uses the same Base.
3457 // Otherwise we will run into the assertion above when visiting that guy.
3458 for (uint i = 0; i < n->outcnt(); ++i) {
3459 Node *out_i = n->raw_out(i);
3460 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3461 out_i->set_req(AddPNode::Base, nn);
3462 #ifdef ASSERT
3463 for (uint j = 0; j < out_i->outcnt(); ++j) {
3464 Node *out_j = out_i->raw_out(j);
3465 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3466 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3467 }
3468 #endif
3469 }
3470 }
3471 n->set_req(AddPNode::Base, nn);
3472 n->set_req(AddPNode::Address, nn);
3473 if (addp->outcnt() == 0) {
3474 addp->disconnect_inputs(this);
3475 }
3476 }
3477 }
3478 }
3479 #endif
3480 break;
3481 }
3482
3483 case Op_CastPP: {
3484 // Remove CastPP nodes to gain more freedom during scheduling but
3485 // keep the dependency they encode as control or precedence edges
3486 // (if control is set already) on memory operations. Some CastPP
3487 // nodes don't have a control (don't carry a dependency): skip
3488 // those.
3489 if (n->in(0) != nullptr) {
3490 ResourceMark rm;
3491 Unique_Node_List wq;
3492 wq.push(n);
3493 for (uint next = 0; next < wq.size(); ++next) {
3494 Node *m = wq.at(next);
3495 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3496 Node* use = m->fast_out(i);
3497 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3498 use->ensure_control_or_add_prec(n->in(0));
3499 } else {
3500 switch(use->Opcode()) {
3501 case Op_AddP:
3502 case Op_DecodeN:
3503 case Op_DecodeNKlass:
3504 case Op_CheckCastPP:
3505 case Op_CastPP:
3506 wq.push(use);
3507 break;
3508 }
3509 }
3510 }
3511 }
3512 }
3513 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3514 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3515 Node* in1 = n->in(1);
3516 const Type* t = n->bottom_type();
3517 Node* new_in1 = in1->clone();
3518 new_in1->as_DecodeN()->set_type(t);
3519
3520 if (!Matcher::narrow_oop_use_complex_address()) {
3521 //
3522 // x86, ARM and friends can handle 2 adds in addressing mode
3523 // and Matcher can fold a DecodeN node into address by using
3524 // a narrow oop directly and do implicit null check in address:
3525 //
3526 // [R12 + narrow_oop_reg<<3 + offset]
3527 // NullCheck narrow_oop_reg
3528 //
3529 // On other platforms (Sparc) we have to keep new DecodeN node and
3530 // use it to do implicit null check in address:
3531 //
3532 // decode_not_null narrow_oop_reg, base_reg
3533 // [base_reg + offset]
3534 // NullCheck base_reg
3535 //
3536 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3537 // to keep the information to which null check the new DecodeN node
3538 // corresponds to use it as value in implicit_null_check().
3539 //
3540 new_in1->set_req(0, n->in(0));
3541 }
3542
3543 n->subsume_by(new_in1, this);
3544 if (in1->outcnt() == 0) {
3545 in1->disconnect_inputs(this);
3546 }
3547 } else {
3548 n->subsume_by(n->in(1), this);
3549 if (n->outcnt() == 0) {
3550 n->disconnect_inputs(this);
3551 }
3552 }
3553 break;
3554 }
3555 case Op_CastII: {
3556 n->as_CastII()->remove_range_check_cast(this);
3557 break;
3558 }
3559 #ifdef _LP64
3560 case Op_CmpP:
3561 // Do this transformation here to preserve CmpPNode::sub() and
3562 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3563 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3564 Node* in1 = n->in(1);
3565 Node* in2 = n->in(2);
3566 if (!in1->is_DecodeNarrowPtr()) {
3567 in2 = in1;
3568 in1 = n->in(2);
3569 }
3570 assert(in1->is_DecodeNarrowPtr(), "sanity");
3571
3572 Node* new_in2 = nullptr;
3573 if (in2->is_DecodeNarrowPtr()) {
3574 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3575 new_in2 = in2->in(1);
3576 } else if (in2->Opcode() == Op_ConP) {
3577 const Type* t = in2->bottom_type();
3578 if (t == TypePtr::NULL_PTR) {
3579 assert(in1->is_DecodeN(), "compare klass to null?");
3580 // Don't convert CmpP null check into CmpN if compressed
3581 // oops implicit null check is not generated.
3582 // This will allow to generate normal oop implicit null check.
3583 if (Matcher::gen_narrow_oop_implicit_null_checks())
3584 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3585 //
3586 // This transformation together with CastPP transformation above
3587 // will generated code for implicit null checks for compressed oops.
3588 //
3589 // The original code after Optimize()
3590 //
3591 // LoadN memory, narrow_oop_reg
3592 // decode narrow_oop_reg, base_reg
3593 // CmpP base_reg, nullptr
3594 // CastPP base_reg // NotNull
3595 // Load [base_reg + offset], val_reg
3596 //
3597 // after these transformations will be
3598 //
3599 // LoadN memory, narrow_oop_reg
3600 // CmpN narrow_oop_reg, nullptr
3601 // decode_not_null narrow_oop_reg, base_reg
3602 // Load [base_reg + offset], val_reg
3603 //
3604 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3605 // since narrow oops can be used in debug info now (see the code in
3606 // final_graph_reshaping_walk()).
3607 //
3608 // At the end the code will be matched to
3609 // on x86:
3610 //
3611 // Load_narrow_oop memory, narrow_oop_reg
3612 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3613 // NullCheck narrow_oop_reg
3614 //
3615 // and on sparc:
3616 //
3617 // Load_narrow_oop memory, narrow_oop_reg
3618 // decode_not_null narrow_oop_reg, base_reg
3619 // Load [base_reg + offset], val_reg
3620 // NullCheck base_reg
3621 //
3622 } else if (t->isa_oopptr()) {
3623 new_in2 = ConNode::make(t->make_narrowoop());
3624 } else if (t->isa_klassptr()) {
3625 ciKlass* klass = t->is_klassptr()->exact_klass();
3626 if (klass->is_in_encoding_range()) {
3627 new_in2 = ConNode::make(t->make_narrowklass());
3628 }
3629 }
3630 }
3631 if (new_in2 != nullptr) {
3632 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3633 n->subsume_by(cmpN, this);
3634 if (in1->outcnt() == 0) {
3635 in1->disconnect_inputs(this);
3636 }
3637 if (in2->outcnt() == 0) {
3638 in2->disconnect_inputs(this);
3639 }
3640 }
3641 }
3642 break;
3643
3644 case Op_DecodeN:
3645 case Op_DecodeNKlass:
3646 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3647 // DecodeN could be pinned when it can't be fold into
3648 // an address expression, see the code for Op_CastPP above.
3649 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3650 break;
3651
3652 case Op_EncodeP:
3653 case Op_EncodePKlass: {
3654 Node* in1 = n->in(1);
3655 if (in1->is_DecodeNarrowPtr()) {
3656 n->subsume_by(in1->in(1), this);
3657 } else if (in1->Opcode() == Op_ConP) {
3658 const Type* t = in1->bottom_type();
3659 if (t == TypePtr::NULL_PTR) {
3660 assert(t->isa_oopptr(), "null klass?");
3661 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3662 } else if (t->isa_oopptr()) {
3663 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3664 } else if (t->isa_klassptr()) {
3665 ciKlass* klass = t->is_klassptr()->exact_klass();
3666 if (klass->is_in_encoding_range()) {
3667 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3668 } else {
3669 assert(false, "unencodable klass in ConP -> EncodeP");
3670 C->record_failure("unencodable klass in ConP -> EncodeP");
3671 }
3672 }
3673 }
3674 if (in1->outcnt() == 0) {
3675 in1->disconnect_inputs(this);
3676 }
3677 break;
3678 }
3679
3680 case Op_Proj: {
3681 if (OptimizeStringConcat || IncrementalInline) {
3682 ProjNode* proj = n->as_Proj();
3683 if (proj->_is_io_use) {
3684 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3685 // Separate projections were used for the exception path which
3686 // are normally removed by a late inline. If it wasn't inlined
3687 // then they will hang around and should just be replaced with
3688 // the original one. Merge them.
3689 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3690 if (non_io_proj != nullptr) {
3691 proj->subsume_by(non_io_proj , this);
3692 }
3693 }
3694 }
3695 break;
3696 }
3697
3698 case Op_Phi:
3699 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3700 // The EncodeP optimization may create Phi with the same edges
3701 // for all paths. It is not handled well by Register Allocator.
3702 Node* unique_in = n->in(1);
3703 assert(unique_in != nullptr, "");
3704 uint cnt = n->req();
3705 for (uint i = 2; i < cnt; i++) {
3706 Node* m = n->in(i);
3707 assert(m != nullptr, "");
3708 if (unique_in != m)
3709 unique_in = nullptr;
3710 }
3711 if (unique_in != nullptr) {
3712 n->subsume_by(unique_in, this);
3713 }
3714 }
3715 break;
3716
3717 #endif
3718
3719 case Op_ModI:
3720 handle_div_mod_op(n, T_INT, false);
3721 break;
3722
3723 case Op_ModL:
3724 handle_div_mod_op(n, T_LONG, false);
3725 break;
3726
3727 case Op_UModI:
3728 handle_div_mod_op(n, T_INT, true);
3729 break;
3730
3731 case Op_UModL:
3732 handle_div_mod_op(n, T_LONG, true);
3733 break;
3734
3735 case Op_LoadVector:
3736 case Op_StoreVector:
3737 #ifdef ASSERT
3738 // Add VerifyVectorAlignment node between adr and load / store.
3739 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3740 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3741 n->as_StoreVector()->must_verify_alignment();
3742 if (must_verify_alignment) {
3743 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3744 n->as_StoreVector()->memory_size();
3745 // The memory access should be aligned to the vector width in bytes.
3746 // However, the underlying array is possibly less well aligned, but at least
3747 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3748 // a loop we can expect at least the following alignment:
3749 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3750 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3751 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3752 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3753 jlong mask = guaranteed_alignment - 1;
3754 Node* mask_con = ConLNode::make(mask);
3755 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3756 n->set_req(MemNode::Address, va);
3757 }
3758 }
3759 #endif
3760 break;
3761
3762 case Op_LoadVectorGather:
3763 case Op_StoreVectorScatter:
3764 case Op_LoadVectorGatherMasked:
3765 case Op_StoreVectorScatterMasked:
3766 case Op_VectorCmpMasked:
3767 case Op_VectorMaskGen:
3768 case Op_LoadVectorMasked:
3769 case Op_StoreVectorMasked:
3770 break;
3771
3772 case Op_AddReductionVI:
3773 case Op_AddReductionVL:
3774 case Op_AddReductionVF:
3775 case Op_AddReductionVD:
3776 case Op_MulReductionVI:
3777 case Op_MulReductionVL:
3778 case Op_MulReductionVF:
3779 case Op_MulReductionVD:
3780 case Op_MinReductionV:
3781 case Op_MaxReductionV:
3782 case Op_UMinReductionV:
3783 case Op_UMaxReductionV:
3784 case Op_AndReductionV:
3785 case Op_OrReductionV:
3786 case Op_XorReductionV:
3787 break;
3788
3789 case Op_PackB:
3790 case Op_PackS:
3791 case Op_PackI:
3792 case Op_PackF:
3793 case Op_PackL:
3794 case Op_PackD:
3795 if (n->req()-1 > 2) {
3796 // Replace many operand PackNodes with a binary tree for matching
3797 PackNode* p = (PackNode*) n;
3798 Node* btp = p->binary_tree_pack(1, n->req());
3799 n->subsume_by(btp, this);
3800 }
3801 break;
3802 case Op_Loop:
3803 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3804 case Op_CountedLoop:
3805 case Op_LongCountedLoop:
3806 case Op_OuterStripMinedLoop:
3807 if (n->as_Loop()->is_inner_loop()) {
3808 frc.inc_inner_loop_count();
3809 }
3810 n->as_Loop()->verify_strip_mined(0);
3811 break;
3812 case Op_LShiftI:
3813 case Op_RShiftI:
3814 case Op_URShiftI:
3815 case Op_LShiftL:
3816 case Op_RShiftL:
3817 case Op_URShiftL:
3818 if (Matcher::need_masked_shift_count) {
3819 // The cpu's shift instructions don't restrict the count to the
3820 // lower 5/6 bits. We need to do the masking ourselves.
3821 Node* in2 = n->in(2);
3822 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3823 const TypeInt* t = in2->find_int_type();
3824 if (t != nullptr && t->is_con()) {
3825 juint shift = t->get_con();
3826 if (shift > mask) { // Unsigned cmp
3827 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3828 }
3829 } else {
3830 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3831 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3832 n->set_req(2, shift);
3833 }
3834 }
3835 if (in2->outcnt() == 0) { // Remove dead node
3836 in2->disconnect_inputs(this);
3837 }
3838 }
3839 break;
3840 case Op_MemBarStoreStore:
3841 case Op_MemBarRelease:
3842 // Break the link with AllocateNode: it is no longer useful and
3843 // confuses register allocation.
3844 if (n->req() > MemBarNode::Precedent) {
3845 n->set_req(MemBarNode::Precedent, top());
3846 }
3847 break;
3848 case Op_MemBarAcquire: {
3849 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3850 // At parse time, the trailing MemBarAcquire for a volatile load
3851 // is created with an edge to the load. After optimizations,
3852 // that input may be a chain of Phis. If those phis have no
3853 // other use, then the MemBarAcquire keeps them alive and
3854 // register allocation can be confused.
3855 dead_nodes.push(n->in(MemBarNode::Precedent));
3856 n->set_req(MemBarNode::Precedent, top());
3857 }
3858 break;
3859 }
3860 case Op_Blackhole:
3861 break;
3862 case Op_RangeCheck: {
3863 RangeCheckNode* rc = n->as_RangeCheck();
3864 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3865 n->subsume_by(iff, this);
3866 frc._tests.push(iff);
3867 break;
3868 }
3869 case Op_ConvI2L: {
3870 if (!Matcher::convi2l_type_required) {
3871 // Code generation on some platforms doesn't need accurate
3872 // ConvI2L types. Widening the type can help remove redundant
3873 // address computations.
3874 n->as_Type()->set_type(TypeLong::INT);
3875 ResourceMark rm;
3876 Unique_Node_List wq;
3877 wq.push(n);
3878 for (uint next = 0; next < wq.size(); next++) {
3879 Node *m = wq.at(next);
3880
3881 for(;;) {
3882 // Loop over all nodes with identical inputs edges as m
3883 Node* k = m->find_similar(m->Opcode());
3884 if (k == nullptr) {
3885 break;
3886 }
3887 // Push their uses so we get a chance to remove node made
3888 // redundant
3889 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3890 Node* u = k->fast_out(i);
3891 if (u->Opcode() == Op_LShiftL ||
3892 u->Opcode() == Op_AddL ||
3893 u->Opcode() == Op_SubL ||
3894 u->Opcode() == Op_AddP) {
3895 wq.push(u);
3896 }
3897 }
3898 // Replace all nodes with identical edges as m with m
3899 k->subsume_by(m, this);
3900 }
3901 }
3902 }
3903 break;
3904 }
3905 case Op_CmpUL: {
3906 if (!Matcher::has_match_rule(Op_CmpUL)) {
3907 // No support for unsigned long comparisons
3908 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3909 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3910 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3911 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3912 Node* andl = new AndLNode(orl, remove_sign_mask);
3913 Node* cmp = new CmpLNode(andl, n->in(2));
3914 n->subsume_by(cmp, this);
3915 }
3916 break;
3917 }
3918 #ifdef ASSERT
3919 case Op_ConNKlass: {
3920 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3921 ciKlass* klass = tp->is_klassptr()->exact_klass();
3922 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3923 break;
3924 }
3925 #endif
3926 default:
3927 assert(!n->is_Call(), "");
3928 assert(!n->is_Mem(), "");
3929 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3930 break;
3931 }
3932 }
3933
3934 //------------------------------final_graph_reshaping_walk---------------------
3935 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3936 // requires that the walk visits a node's inputs before visiting the node.
3937 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3938 Unique_Node_List sfpt;
3939
3940 frc._visited.set(root->_idx); // first, mark node as visited
3941 uint cnt = root->req();
3942 Node *n = root;
3943 uint i = 0;
3944 while (true) {
3945 if (i < cnt) {
3946 // Place all non-visited non-null inputs onto stack
3947 Node* m = n->in(i);
3948 ++i;
3949 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3950 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3951 // compute worst case interpreter size in case of a deoptimization
3952 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3953
3954 sfpt.push(m);
3955 }
3956 cnt = m->req();
3957 nstack.push(n, i); // put on stack parent and next input's index
3958 n = m;
3959 i = 0;
3960 }
3961 } else {
3962 // Now do post-visit work
3963 final_graph_reshaping_impl(n, frc, dead_nodes);
3964 if (nstack.is_empty())
3965 break; // finished
3966 n = nstack.node(); // Get node from stack
3967 cnt = n->req();
3968 i = nstack.index();
3969 nstack.pop(); // Shift to the next node on stack
3970 }
3971 }
3972
3973 // Skip next transformation if compressed oops are not used.
3974 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3975 (!UseCompressedOops && !UseCompressedClassPointers))
3976 return;
3977
3978 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3979 // It could be done for an uncommon traps or any safepoints/calls
3980 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3981 while (sfpt.size() > 0) {
3982 n = sfpt.pop();
3983 JVMState *jvms = n->as_SafePoint()->jvms();
3984 assert(jvms != nullptr, "sanity");
3985 int start = jvms->debug_start();
3986 int end = n->req();
3987 bool is_uncommon = (n->is_CallStaticJava() &&
3988 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3989 for (int j = start; j < end; j++) {
3990 Node* in = n->in(j);
3991 if (in->is_DecodeNarrowPtr()) {
3992 bool safe_to_skip = true;
3993 if (!is_uncommon ) {
3994 // Is it safe to skip?
3995 for (uint i = 0; i < in->outcnt(); i++) {
3996 Node* u = in->raw_out(i);
3997 if (!u->is_SafePoint() ||
3998 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3999 safe_to_skip = false;
4000 }
4001 }
4002 }
4003 if (safe_to_skip) {
4004 n->set_req(j, in->in(1));
4005 }
4006 if (in->outcnt() == 0) {
4007 in->disconnect_inputs(this);
4008 }
4009 }
4010 }
4011 }
4012 }
4013
4014 //------------------------------final_graph_reshaping--------------------------
4015 // Final Graph Reshaping.
4016 //
4017 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4018 // and not commoned up and forced early. Must come after regular
4019 // optimizations to avoid GVN undoing the cloning. Clone constant
4020 // inputs to Loop Phis; these will be split by the allocator anyways.
4021 // Remove Opaque nodes.
4022 // (2) Move last-uses by commutative operations to the left input to encourage
4023 // Intel update-in-place two-address operations and better register usage
4024 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4025 // calls canonicalizing them back.
4026 // (3) Count the number of double-precision FP ops, single-precision FP ops
4027 // and call sites. On Intel, we can get correct rounding either by
4028 // forcing singles to memory (requires extra stores and loads after each
4029 // FP bytecode) or we can set a rounding mode bit (requires setting and
4030 // clearing the mode bit around call sites). The mode bit is only used
4031 // if the relative frequency of single FP ops to calls is low enough.
4032 // This is a key transform for SPEC mpeg_audio.
4033 // (4) Detect infinite loops; blobs of code reachable from above but not
4034 // below. Several of the Code_Gen algorithms fail on such code shapes,
4035 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4036 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4037 // Detection is by looking for IfNodes where only 1 projection is
4038 // reachable from below or CatchNodes missing some targets.
4039 // (5) Assert for insane oop offsets in debug mode.
4040
4041 bool Compile::final_graph_reshaping() {
4042 // an infinite loop may have been eliminated by the optimizer,
4043 // in which case the graph will be empty.
4044 if (root()->req() == 1) {
4045 // Do not compile method that is only a trivial infinite loop,
4046 // since the content of the loop may have been eliminated.
4047 record_method_not_compilable("trivial infinite loop");
4048 return true;
4049 }
4050
4051 // Expensive nodes have their control input set to prevent the GVN
4052 // from freely commoning them. There's no GVN beyond this point so
4053 // no need to keep the control input. We want the expensive nodes to
4054 // be freely moved to the least frequent code path by gcm.
4055 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4056 for (int i = 0; i < expensive_count(); i++) {
4057 _expensive_nodes.at(i)->set_req(0, nullptr);
4058 }
4059
4060 Final_Reshape_Counts frc;
4061
4062 // Visit everybody reachable!
4063 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4064 Node_Stack nstack(live_nodes() >> 1);
4065 Unique_Node_List dead_nodes;
4066 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4067
4068 // Check for unreachable (from below) code (i.e., infinite loops).
4069 for( uint i = 0; i < frc._tests.size(); i++ ) {
4070 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4071 // Get number of CFG targets.
4072 // Note that PCTables include exception targets after calls.
4073 uint required_outcnt = n->required_outcnt();
4074 if (n->outcnt() != required_outcnt) {
4075 // Check for a few special cases. Rethrow Nodes never take the
4076 // 'fall-thru' path, so expected kids is 1 less.
4077 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4078 if (n->in(0)->in(0)->is_Call()) {
4079 CallNode* call = n->in(0)->in(0)->as_Call();
4080 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4081 required_outcnt--; // Rethrow always has 1 less kid
4082 } else if (call->req() > TypeFunc::Parms &&
4083 call->is_CallDynamicJava()) {
4084 // Check for null receiver. In such case, the optimizer has
4085 // detected that the virtual call will always result in a null
4086 // pointer exception. The fall-through projection of this CatchNode
4087 // will not be populated.
4088 Node* arg0 = call->in(TypeFunc::Parms);
4089 if (arg0->is_Type() &&
4090 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4091 required_outcnt--;
4092 }
4093 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4094 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4095 // Check for illegal array length. In such case, the optimizer has
4096 // detected that the allocation attempt will always result in an
4097 // exception. There is no fall-through projection of this CatchNode .
4098 assert(call->is_CallStaticJava(), "static call expected");
4099 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4100 uint valid_length_test_input = call->req() - 1;
4101 Node* valid_length_test = call->in(valid_length_test_input);
4102 call->del_req(valid_length_test_input);
4103 if (valid_length_test->find_int_con(1) == 0) {
4104 required_outcnt--;
4105 }
4106 dead_nodes.push(valid_length_test);
4107 assert(n->outcnt() == required_outcnt, "malformed control flow");
4108 continue;
4109 }
4110 }
4111 }
4112
4113 // Recheck with a better notion of 'required_outcnt'
4114 if (n->outcnt() != required_outcnt) {
4115 record_method_not_compilable("malformed control flow");
4116 return true; // Not all targets reachable!
4117 }
4118 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4119 CallNode* call = n->in(0)->in(0)->as_Call();
4120 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4121 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4122 assert(call->is_CallStaticJava(), "static call expected");
4123 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4124 uint valid_length_test_input = call->req() - 1;
4125 dead_nodes.push(call->in(valid_length_test_input));
4126 call->del_req(valid_length_test_input); // valid length test useless now
4127 }
4128 }
4129 // Check that I actually visited all kids. Unreached kids
4130 // must be infinite loops.
4131 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4132 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4133 record_method_not_compilable("infinite loop");
4134 return true; // Found unvisited kid; must be unreach
4135 }
4136
4137 // Here so verification code in final_graph_reshaping_walk()
4138 // always see an OuterStripMinedLoopEnd
4139 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4140 IfNode* init_iff = n->as_If();
4141 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4142 n->subsume_by(iff, this);
4143 }
4144 }
4145
4146 while (dead_nodes.size() > 0) {
4147 Node* m = dead_nodes.pop();
4148 if (m->outcnt() == 0 && m != top()) {
4149 for (uint j = 0; j < m->req(); j++) {
4150 Node* in = m->in(j);
4151 if (in != nullptr) {
4152 dead_nodes.push(in);
4153 }
4154 }
4155 m->disconnect_inputs(this);
4156 }
4157 }
4158
4159 set_java_calls(frc.get_java_call_count());
4160 set_inner_loops(frc.get_inner_loop_count());
4161
4162 // No infinite loops, no reason to bail out.
4163 return false;
4164 }
4165
4166 //-----------------------------too_many_traps----------------------------------
4167 // Report if there are too many traps at the current method and bci.
4168 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4169 bool Compile::too_many_traps(ciMethod* method,
4170 int bci,
4171 Deoptimization::DeoptReason reason) {
4172 ciMethodData* md = method->method_data();
4173 if (md->is_empty()) {
4174 // Assume the trap has not occurred, or that it occurred only
4175 // because of a transient condition during start-up in the interpreter.
4176 return false;
4177 }
4178 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4179 if (md->has_trap_at(bci, m, reason) != 0) {
4180 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4181 // Also, if there are multiple reasons, or if there is no per-BCI record,
4182 // assume the worst.
4183 if (log())
4184 log()->elem("observe trap='%s' count='%d'",
4185 Deoptimization::trap_reason_name(reason),
4186 md->trap_count(reason));
4187 return true;
4188 } else {
4189 // Ignore method/bci and see if there have been too many globally.
4190 return too_many_traps(reason, md);
4191 }
4192 }
4193
4194 // Less-accurate variant which does not require a method and bci.
4195 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4196 ciMethodData* logmd) {
4197 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4198 // Too many traps globally.
4199 // Note that we use cumulative trap_count, not just md->trap_count.
4200 if (log()) {
4201 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4202 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4203 Deoptimization::trap_reason_name(reason),
4204 mcount, trap_count(reason));
4205 }
4206 return true;
4207 } else {
4208 // The coast is clear.
4209 return false;
4210 }
4211 }
4212
4213 //--------------------------too_many_recompiles--------------------------------
4214 // Report if there are too many recompiles at the current method and bci.
4215 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4216 // Is not eager to return true, since this will cause the compiler to use
4217 // Action_none for a trap point, to avoid too many recompilations.
4218 bool Compile::too_many_recompiles(ciMethod* method,
4219 int bci,
4220 Deoptimization::DeoptReason reason) {
4221 ciMethodData* md = method->method_data();
4222 if (md->is_empty()) {
4223 // Assume the trap has not occurred, or that it occurred only
4224 // because of a transient condition during start-up in the interpreter.
4225 return false;
4226 }
4227 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4228 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4229 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4230 Deoptimization::DeoptReason per_bc_reason
4231 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4232 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4233 if ((per_bc_reason == Deoptimization::Reason_none
4234 || md->has_trap_at(bci, m, reason) != 0)
4235 // The trap frequency measure we care about is the recompile count:
4236 && md->trap_recompiled_at(bci, m)
4237 && md->overflow_recompile_count() >= bc_cutoff) {
4238 // Do not emit a trap here if it has already caused recompilations.
4239 // Also, if there are multiple reasons, or if there is no per-BCI record,
4240 // assume the worst.
4241 if (log())
4242 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4243 Deoptimization::trap_reason_name(reason),
4244 md->trap_count(reason),
4245 md->overflow_recompile_count());
4246 return true;
4247 } else if (trap_count(reason) != 0
4248 && decompile_count() >= m_cutoff) {
4249 // Too many recompiles globally, and we have seen this sort of trap.
4250 // Use cumulative decompile_count, not just md->decompile_count.
4251 if (log())
4252 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4253 Deoptimization::trap_reason_name(reason),
4254 md->trap_count(reason), trap_count(reason),
4255 md->decompile_count(), decompile_count());
4256 return true;
4257 } else {
4258 // The coast is clear.
4259 return false;
4260 }
4261 }
4262
4263 // Compute when not to trap. Used by matching trap based nodes and
4264 // NullCheck optimization.
4265 void Compile::set_allowed_deopt_reasons() {
4266 _allowed_reasons = 0;
4267 if (is_method_compilation()) {
4268 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4269 assert(rs < BitsPerInt, "recode bit map");
4270 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4271 _allowed_reasons |= nth_bit(rs);
4272 }
4273 }
4274 }
4275 }
4276
4277 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4278 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4279 }
4280
4281 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4282 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4283 }
4284
4285 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4286 if (holder->is_initialized()) {
4287 return false;
4288 }
4289 if (holder->is_being_initialized()) {
4290 if (accessing_method->holder() == holder) {
4291 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4292 // <init>, or a static method. In all those cases, there was an initialization
4293 // barrier on the holder klass passed.
4294 if (accessing_method->is_static_initializer() ||
4295 accessing_method->is_object_initializer() ||
4296 accessing_method->is_static()) {
4297 return false;
4298 }
4299 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4300 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4301 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4302 // child class can become fully initialized while its parent class is still being initialized.
4303 if (accessing_method->is_static_initializer()) {
4304 return false;
4305 }
4306 }
4307 ciMethod* root = method(); // the root method of compilation
4308 if (root != accessing_method) {
4309 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4310 }
4311 }
4312 return true;
4313 }
4314
4315 #ifndef PRODUCT
4316 //------------------------------verify_bidirectional_edges---------------------
4317 // For each input edge to a node (ie - for each Use-Def edge), verify that
4318 // there is a corresponding Def-Use edge.
4319 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4320 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4321 uint stack_size = live_nodes() >> 4;
4322 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4323 if (root_and_safepoints != nullptr) {
4324 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4325 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4326 Node* root_or_safepoint = root_and_safepoints->at(i);
4327 // If the node is a safepoint, let's check if it still has a control input
4328 // Lack of control input signifies that this node was killed by CCP or
4329 // recursively by remove_globally_dead_node and it shouldn't be a starting
4330 // point.
4331 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4332 nstack.push(root_or_safepoint);
4333 }
4334 }
4335 } else {
4336 nstack.push(_root);
4337 }
4338
4339 while (nstack.size() > 0) {
4340 Node* n = nstack.pop();
4341 if (visited.member(n)) {
4342 continue;
4343 }
4344 visited.push(n);
4345
4346 // Walk over all input edges, checking for correspondence
4347 uint length = n->len();
4348 for (uint i = 0; i < length; i++) {
4349 Node* in = n->in(i);
4350 if (in != nullptr && !visited.member(in)) {
4351 nstack.push(in); // Put it on stack
4352 }
4353 if (in != nullptr && !in->is_top()) {
4354 // Count instances of `next`
4355 int cnt = 0;
4356 for (uint idx = 0; idx < in->_outcnt; idx++) {
4357 if (in->_out[idx] == n) {
4358 cnt++;
4359 }
4360 }
4361 assert(cnt > 0, "Failed to find Def-Use edge.");
4362 // Check for duplicate edges
4363 // walk the input array downcounting the input edges to n
4364 for (uint j = 0; j < length; j++) {
4365 if (n->in(j) == in) {
4366 cnt--;
4367 }
4368 }
4369 assert(cnt == 0, "Mismatched edge count.");
4370 } else if (in == nullptr) {
4371 assert(i == 0 || i >= n->req() ||
4372 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4373 (n->is_Unlock() && i == (n->req() - 1)) ||
4374 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4375 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4376 } else {
4377 assert(in->is_top(), "sanity");
4378 // Nothing to check.
4379 }
4380 }
4381 }
4382 }
4383
4384 //------------------------------verify_graph_edges---------------------------
4385 // Walk the Graph and verify that there is a one-to-one correspondence
4386 // between Use-Def edges and Def-Use edges in the graph.
4387 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4388 if (VerifyGraphEdges) {
4389 Unique_Node_List visited;
4390
4391 // Call graph walk to check edges
4392 verify_bidirectional_edges(visited, root_and_safepoints);
4393 if (no_dead_code) {
4394 // Now make sure that no visited node is used by an unvisited node.
4395 bool dead_nodes = false;
4396 Unique_Node_List checked;
4397 while (visited.size() > 0) {
4398 Node* n = visited.pop();
4399 checked.push(n);
4400 for (uint i = 0; i < n->outcnt(); i++) {
4401 Node* use = n->raw_out(i);
4402 if (checked.member(use)) continue; // already checked
4403 if (visited.member(use)) continue; // already in the graph
4404 if (use->is_Con()) continue; // a dead ConNode is OK
4405 // At this point, we have found a dead node which is DU-reachable.
4406 if (!dead_nodes) {
4407 tty->print_cr("*** Dead nodes reachable via DU edges:");
4408 dead_nodes = true;
4409 }
4410 use->dump(2);
4411 tty->print_cr("---");
4412 checked.push(use); // No repeats; pretend it is now checked.
4413 }
4414 }
4415 assert(!dead_nodes, "using nodes must be reachable from root");
4416 }
4417 }
4418 }
4419 #endif
4420
4421 // The Compile object keeps track of failure reasons separately from the ciEnv.
4422 // This is required because there is not quite a 1-1 relation between the
4423 // ciEnv and its compilation task and the Compile object. Note that one
4424 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4425 // to backtrack and retry without subsuming loads. Other than this backtracking
4426 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4427 // by the logic in C2Compiler.
4428 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4429 if (log() != nullptr) {
4430 log()->elem("failure reason='%s' phase='compile'", reason);
4431 }
4432 if (_failure_reason.get() == nullptr) {
4433 // Record the first failure reason.
4434 _failure_reason.set(reason);
4435 if (CaptureBailoutInformation) {
4436 _first_failure_details = new CompilationFailureInfo(reason);
4437 }
4438 } else {
4439 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4440 }
4441
4442 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4443 C->print_method(PHASE_FAILURE, 1);
4444 }
4445 _root = nullptr; // flush the graph, too
4446 }
4447
4448 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4449 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4450 _compile(Compile::current()),
4451 _log(nullptr),
4452 _dolog(CITimeVerbose)
4453 {
4454 assert(_compile != nullptr, "sanity check");
4455 assert(id != PhaseTraceId::_t_none, "Don't use none");
4456 if (_dolog) {
4457 _log = _compile->log();
4458 }
4459 if (_log != nullptr) {
4460 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4461 _log->stamp();
4462 _log->end_head();
4463 }
4464
4465 // Inform memory statistic, if enabled
4466 if (CompilationMemoryStatistic::enabled()) {
4467 CompilationMemoryStatistic::on_phase_start((int)id, name);
4468 }
4469 }
4470
4471 Compile::TracePhase::TracePhase(PhaseTraceId id)
4472 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4473
4474 Compile::TracePhase::~TracePhase() {
4475
4476 // Inform memory statistic, if enabled
4477 if (CompilationMemoryStatistic::enabled()) {
4478 CompilationMemoryStatistic::on_phase_end();
4479 }
4480
4481 if (_compile->failing_internal()) {
4482 if (_log != nullptr) {
4483 _log->done("phase");
4484 }
4485 return; // timing code, not stressing bailouts.
4486 }
4487 #ifdef ASSERT
4488 if (PrintIdealNodeCount) {
4489 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4490 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4491 }
4492
4493 if (VerifyIdealNodeCount) {
4494 _compile->print_missing_nodes();
4495 }
4496 #endif
4497
4498 if (_log != nullptr) {
4499 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4500 }
4501 }
4502
4503 //----------------------------static_subtype_check-----------------------------
4504 // Shortcut important common cases when superklass is exact:
4505 // (0) superklass is java.lang.Object (can occur in reflective code)
4506 // (1) subklass is already limited to a subtype of superklass => always ok
4507 // (2) subklass does not overlap with superklass => always fail
4508 // (3) superklass has NO subtypes and we can check with a simple compare.
4509 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4510 if (skip) {
4511 return SSC_full_test; // Let caller generate the general case.
4512 }
4513
4514 if (subk->is_java_subtype_of(superk)) {
4515 return SSC_always_true; // (0) and (1) this test cannot fail
4516 }
4517
4518 if (!subk->maybe_java_subtype_of(superk)) {
4519 return SSC_always_false; // (2) true path dead; no dynamic test needed
4520 }
4521
4522 const Type* superelem = superk;
4523 if (superk->isa_aryklassptr()) {
4524 int ignored;
4525 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4526 }
4527
4528 if (superelem->isa_instklassptr()) {
4529 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4530 if (!ik->has_subklass()) {
4531 if (!ik->is_final()) {
4532 // Add a dependency if there is a chance of a later subclass.
4533 dependencies()->assert_leaf_type(ik);
4534 }
4535 if (!superk->maybe_java_subtype_of(subk)) {
4536 return SSC_always_false;
4537 }
4538 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4539 }
4540 } else {
4541 // A primitive array type has no subtypes.
4542 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4543 }
4544
4545 return SSC_full_test;
4546 }
4547
4548 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4549 #ifdef _LP64
4550 // The scaled index operand to AddP must be a clean 64-bit value.
4551 // Java allows a 32-bit int to be incremented to a negative
4552 // value, which appears in a 64-bit register as a large
4553 // positive number. Using that large positive number as an
4554 // operand in pointer arithmetic has bad consequences.
4555 // On the other hand, 32-bit overflow is rare, and the possibility
4556 // can often be excluded, if we annotate the ConvI2L node with
4557 // a type assertion that its value is known to be a small positive
4558 // number. (The prior range check has ensured this.)
4559 // This assertion is used by ConvI2LNode::Ideal.
4560 int index_max = max_jint - 1; // array size is max_jint, index is one less
4561 if (sizetype != nullptr && sizetype->_hi > 0) {
4562 index_max = sizetype->_hi - 1;
4563 }
4564 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4565 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4566 #endif
4567 return idx;
4568 }
4569
4570 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4571 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4572 if (ctrl != nullptr) {
4573 // Express control dependency by a CastII node with a narrow type.
4574 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4575 // node from floating above the range check during loop optimizations. Otherwise, the
4576 // ConvI2L node may be eliminated independently of the range check, causing the data path
4577 // to become TOP while the control path is still there (although it's unreachable).
4578 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
4579 value = phase->transform(value);
4580 }
4581 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4582 return phase->transform(new ConvI2LNode(value, ltype));
4583 }
4584
4585 void Compile::dump_print_inlining() {
4586 inline_printer()->print_on(tty);
4587 }
4588
4589 void Compile::log_late_inline(CallGenerator* cg) {
4590 if (log() != nullptr) {
4591 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4592 cg->unique_id());
4593 JVMState* p = cg->call_node()->jvms();
4594 while (p != nullptr) {
4595 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4596 p = p->caller();
4597 }
4598 log()->tail("late_inline");
4599 }
4600 }
4601
4602 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4603 log_late_inline(cg);
4604 if (log() != nullptr) {
4605 log()->inline_fail(msg);
4606 }
4607 }
4608
4609 void Compile::log_inline_id(CallGenerator* cg) {
4610 if (log() != nullptr) {
4611 // The LogCompilation tool needs a unique way to identify late
4612 // inline call sites. This id must be unique for this call site in
4613 // this compilation. Try to have it unique across compilations as
4614 // well because it can be convenient when grepping through the log
4615 // file.
4616 // Distinguish OSR compilations from others in case CICountOSR is
4617 // on.
4618 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4619 cg->set_unique_id(id);
4620 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4621 }
4622 }
4623
4624 void Compile::log_inline_failure(const char* msg) {
4625 if (C->log() != nullptr) {
4626 C->log()->inline_fail(msg);
4627 }
4628 }
4629
4630
4631 // Dump inlining replay data to the stream.
4632 // Don't change thread state and acquire any locks.
4633 void Compile::dump_inline_data(outputStream* out) {
4634 InlineTree* inl_tree = ilt();
4635 if (inl_tree != nullptr) {
4636 out->print(" inline %d", inl_tree->count());
4637 inl_tree->dump_replay_data(out);
4638 }
4639 }
4640
4641 void Compile::dump_inline_data_reduced(outputStream* out) {
4642 assert(ReplayReduce, "");
4643
4644 InlineTree* inl_tree = ilt();
4645 if (inl_tree == nullptr) {
4646 return;
4647 }
4648 // Enable iterative replay file reduction
4649 // Output "compile" lines for depth 1 subtrees,
4650 // simulating that those trees were compiled
4651 // instead of inlined.
4652 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4653 InlineTree* sub = inl_tree->subtrees().at(i);
4654 if (sub->inline_level() != 1) {
4655 continue;
4656 }
4657
4658 ciMethod* method = sub->method();
4659 int entry_bci = -1;
4660 int comp_level = env()->task()->comp_level();
4661 out->print("compile ");
4662 method->dump_name_as_ascii(out);
4663 out->print(" %d %d", entry_bci, comp_level);
4664 out->print(" inline %d", sub->count());
4665 sub->dump_replay_data(out, -1);
4666 out->cr();
4667 }
4668 }
4669
4670 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4671 if (n1->Opcode() < n2->Opcode()) return -1;
4672 else if (n1->Opcode() > n2->Opcode()) return 1;
4673
4674 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4675 for (uint i = 1; i < n1->req(); i++) {
4676 if (n1->in(i) < n2->in(i)) return -1;
4677 else if (n1->in(i) > n2->in(i)) return 1;
4678 }
4679
4680 return 0;
4681 }
4682
4683 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4684 Node* n1 = *n1p;
4685 Node* n2 = *n2p;
4686
4687 return cmp_expensive_nodes(n1, n2);
4688 }
4689
4690 void Compile::sort_expensive_nodes() {
4691 if (!expensive_nodes_sorted()) {
4692 _expensive_nodes.sort(cmp_expensive_nodes);
4693 }
4694 }
4695
4696 bool Compile::expensive_nodes_sorted() const {
4697 for (int i = 1; i < _expensive_nodes.length(); i++) {
4698 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4699 return false;
4700 }
4701 }
4702 return true;
4703 }
4704
4705 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4706 if (_expensive_nodes.length() == 0) {
4707 return false;
4708 }
4709
4710 assert(OptimizeExpensiveOps, "optimization off?");
4711
4712 // Take this opportunity to remove dead nodes from the list
4713 int j = 0;
4714 for (int i = 0; i < _expensive_nodes.length(); i++) {
4715 Node* n = _expensive_nodes.at(i);
4716 if (!n->is_unreachable(igvn)) {
4717 assert(n->is_expensive(), "should be expensive");
4718 _expensive_nodes.at_put(j, n);
4719 j++;
4720 }
4721 }
4722 _expensive_nodes.trunc_to(j);
4723
4724 // Then sort the list so that similar nodes are next to each other
4725 // and check for at least two nodes of identical kind with same data
4726 // inputs.
4727 sort_expensive_nodes();
4728
4729 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4730 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4731 return true;
4732 }
4733 }
4734
4735 return false;
4736 }
4737
4738 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4739 if (_expensive_nodes.length() == 0) {
4740 return;
4741 }
4742
4743 assert(OptimizeExpensiveOps, "optimization off?");
4744
4745 // Sort to bring similar nodes next to each other and clear the
4746 // control input of nodes for which there's only a single copy.
4747 sort_expensive_nodes();
4748
4749 int j = 0;
4750 int identical = 0;
4751 int i = 0;
4752 bool modified = false;
4753 for (; i < _expensive_nodes.length()-1; i++) {
4754 assert(j <= i, "can't write beyond current index");
4755 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4756 identical++;
4757 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4758 continue;
4759 }
4760 if (identical > 0) {
4761 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4762 identical = 0;
4763 } else {
4764 Node* n = _expensive_nodes.at(i);
4765 igvn.replace_input_of(n, 0, nullptr);
4766 igvn.hash_insert(n);
4767 modified = true;
4768 }
4769 }
4770 if (identical > 0) {
4771 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4772 } else if (_expensive_nodes.length() >= 1) {
4773 Node* n = _expensive_nodes.at(i);
4774 igvn.replace_input_of(n, 0, nullptr);
4775 igvn.hash_insert(n);
4776 modified = true;
4777 }
4778 _expensive_nodes.trunc_to(j);
4779 if (modified) {
4780 igvn.optimize();
4781 }
4782 }
4783
4784 void Compile::add_expensive_node(Node * n) {
4785 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4786 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4787 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4788 if (OptimizeExpensiveOps) {
4789 _expensive_nodes.append(n);
4790 } else {
4791 // Clear control input and let IGVN optimize expensive nodes if
4792 // OptimizeExpensiveOps is off.
4793 n->set_req(0, nullptr);
4794 }
4795 }
4796
4797 /**
4798 * Track coarsened Lock and Unlock nodes.
4799 */
4800
4801 class Lock_List : public Node_List {
4802 uint _origin_cnt;
4803 public:
4804 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4805 uint origin_cnt() const { return _origin_cnt; }
4806 };
4807
4808 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4809 int length = locks.length();
4810 if (length > 0) {
4811 // Have to keep this list until locks elimination during Macro nodes elimination.
4812 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4813 AbstractLockNode* alock = locks.at(0);
4814 BoxLockNode* box = alock->box_node()->as_BoxLock();
4815 for (int i = 0; i < length; i++) {
4816 AbstractLockNode* lock = locks.at(i);
4817 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4818 locks_list->push(lock);
4819 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4820 if (this_box != box) {
4821 // Locking regions (BoxLock) could be Unbalanced here:
4822 // - its coarsened locks were eliminated in earlier
4823 // macro nodes elimination followed by loop unroll
4824 // - it is OSR locking region (no Lock node)
4825 // Preserve Unbalanced status in such cases.
4826 if (!this_box->is_unbalanced()) {
4827 this_box->set_coarsened();
4828 }
4829 if (!box->is_unbalanced()) {
4830 box->set_coarsened();
4831 }
4832 }
4833 }
4834 _coarsened_locks.append(locks_list);
4835 }
4836 }
4837
4838 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4839 int count = coarsened_count();
4840 for (int i = 0; i < count; i++) {
4841 Node_List* locks_list = _coarsened_locks.at(i);
4842 for (uint j = 0; j < locks_list->size(); j++) {
4843 Node* lock = locks_list->at(j);
4844 assert(lock->is_AbstractLock(), "sanity");
4845 if (!useful.member(lock)) {
4846 locks_list->yank(lock);
4847 }
4848 }
4849 }
4850 }
4851
4852 void Compile::remove_coarsened_lock(Node* n) {
4853 if (n->is_AbstractLock()) {
4854 int count = coarsened_count();
4855 for (int i = 0; i < count; i++) {
4856 Node_List* locks_list = _coarsened_locks.at(i);
4857 locks_list->yank(n);
4858 }
4859 }
4860 }
4861
4862 bool Compile::coarsened_locks_consistent() {
4863 int count = coarsened_count();
4864 for (int i = 0; i < count; i++) {
4865 bool unbalanced = false;
4866 bool modified = false; // track locks kind modifications
4867 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4868 uint size = locks_list->size();
4869 if (size == 0) {
4870 unbalanced = false; // All locks were eliminated - good
4871 } else if (size != locks_list->origin_cnt()) {
4872 unbalanced = true; // Some locks were removed from list
4873 } else {
4874 for (uint j = 0; j < size; j++) {
4875 Node* lock = locks_list->at(j);
4876 // All nodes in group should have the same state (modified or not)
4877 if (!lock->as_AbstractLock()->is_coarsened()) {
4878 if (j == 0) {
4879 // first on list was modified, the rest should be too for consistency
4880 modified = true;
4881 } else if (!modified) {
4882 // this lock was modified but previous locks on the list were not
4883 unbalanced = true;
4884 break;
4885 }
4886 } else if (modified) {
4887 // previous locks on list were modified but not this lock
4888 unbalanced = true;
4889 break;
4890 }
4891 }
4892 }
4893 if (unbalanced) {
4894 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4895 #ifdef ASSERT
4896 if (PrintEliminateLocks) {
4897 tty->print_cr("=== unbalanced coarsened locks ===");
4898 for (uint l = 0; l < size; l++) {
4899 locks_list->at(l)->dump();
4900 }
4901 }
4902 #endif
4903 record_failure(C2Compiler::retry_no_locks_coarsening());
4904 return false;
4905 }
4906 }
4907 return true;
4908 }
4909
4910 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4911 // locks coarsening optimization removed Lock/Unlock nodes from them.
4912 // Such regions become unbalanced because coarsening only removes part
4913 // of Lock/Unlock nodes in region. As result we can't execute other
4914 // locks elimination optimizations which assume all code paths have
4915 // corresponding pair of Lock/Unlock nodes - they are balanced.
4916 void Compile::mark_unbalanced_boxes() const {
4917 int count = coarsened_count();
4918 for (int i = 0; i < count; i++) {
4919 Node_List* locks_list = _coarsened_locks.at(i);
4920 uint size = locks_list->size();
4921 if (size > 0) {
4922 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4923 BoxLockNode* box = alock->box_node()->as_BoxLock();
4924 if (alock->is_coarsened()) {
4925 // coarsened_locks_consistent(), which is called before this method, verifies
4926 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4927 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4928 for (uint j = 1; j < size; j++) {
4929 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4930 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4931 if (box != this_box) {
4932 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4933 box->set_unbalanced();
4934 this_box->set_unbalanced();
4935 }
4936 }
4937 }
4938 }
4939 }
4940 }
4941
4942 /**
4943 * Remove the speculative part of types and clean up the graph
4944 */
4945 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4946 if (UseTypeSpeculation) {
4947 Unique_Node_List worklist;
4948 worklist.push(root());
4949 int modified = 0;
4950 // Go over all type nodes that carry a speculative type, drop the
4951 // speculative part of the type and enqueue the node for an igvn
4952 // which may optimize it out.
4953 for (uint next = 0; next < worklist.size(); ++next) {
4954 Node *n = worklist.at(next);
4955 if (n->is_Type()) {
4956 TypeNode* tn = n->as_Type();
4957 const Type* t = tn->type();
4958 const Type* t_no_spec = t->remove_speculative();
4959 if (t_no_spec != t) {
4960 bool in_hash = igvn.hash_delete(n);
4961 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4962 tn->set_type(t_no_spec);
4963 igvn.hash_insert(n);
4964 igvn._worklist.push(n); // give it a chance to go away
4965 modified++;
4966 }
4967 }
4968 // Iterate over outs - endless loops is unreachable from below
4969 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4970 Node *m = n->fast_out(i);
4971 if (not_a_node(m)) {
4972 continue;
4973 }
4974 worklist.push(m);
4975 }
4976 }
4977 // Drop the speculative part of all types in the igvn's type table
4978 igvn.remove_speculative_types();
4979 if (modified > 0) {
4980 igvn.optimize();
4981 if (failing()) return;
4982 }
4983 #ifdef ASSERT
4984 // Verify that after the IGVN is over no speculative type has resurfaced
4985 worklist.clear();
4986 worklist.push(root());
4987 for (uint next = 0; next < worklist.size(); ++next) {
4988 Node *n = worklist.at(next);
4989 const Type* t = igvn.type_or_null(n);
4990 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4991 if (n->is_Type()) {
4992 t = n->as_Type()->type();
4993 assert(t == t->remove_speculative(), "no more speculative types");
4994 }
4995 // Iterate over outs - endless loops is unreachable from below
4996 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4997 Node *m = n->fast_out(i);
4998 if (not_a_node(m)) {
4999 continue;
5000 }
5001 worklist.push(m);
5002 }
5003 }
5004 igvn.check_no_speculative_types();
5005 #endif
5006 }
5007 }
5008
5009 // Auxiliary methods to support randomized stressing/fuzzing.
5010
5011 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5012 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5013 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5014 FLAG_SET_ERGO(StressSeed, _stress_seed);
5015 } else {
5016 _stress_seed = StressSeed;
5017 }
5018 if (_log != nullptr) {
5019 _log->elem("stress_test seed='%u'", _stress_seed);
5020 }
5021 }
5022
5023 int Compile::random() {
5024 _stress_seed = os::next_random(_stress_seed);
5025 return static_cast<int>(_stress_seed);
5026 }
5027
5028 // This method can be called the arbitrary number of times, with current count
5029 // as the argument. The logic allows selecting a single candidate from the
5030 // running list of candidates as follows:
5031 // int count = 0;
5032 // Cand* selected = null;
5033 // while(cand = cand->next()) {
5034 // if (randomized_select(++count)) {
5035 // selected = cand;
5036 // }
5037 // }
5038 //
5039 // Including count equalizes the chances any candidate is "selected".
5040 // This is useful when we don't have the complete list of candidates to choose
5041 // from uniformly. In this case, we need to adjust the randomicity of the
5042 // selection, or else we will end up biasing the selection towards the latter
5043 // candidates.
5044 //
5045 // Quick back-envelope calculation shows that for the list of n candidates
5046 // the equal probability for the candidate to persist as "best" can be
5047 // achieved by replacing it with "next" k-th candidate with the probability
5048 // of 1/k. It can be easily shown that by the end of the run, the
5049 // probability for any candidate is converged to 1/n, thus giving the
5050 // uniform distribution among all the candidates.
5051 //
5052 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5053 #define RANDOMIZED_DOMAIN_POW 29
5054 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5055 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5056 bool Compile::randomized_select(int count) {
5057 assert(count > 0, "only positive");
5058 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5059 }
5060
5061 #ifdef ASSERT
5062 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5063 bool Compile::fail_randomly() {
5064 if ((random() % StressBailoutMean) != 0) {
5065 return false;
5066 }
5067 record_failure("StressBailout");
5068 return true;
5069 }
5070
5071 bool Compile::failure_is_artificial() {
5072 return C->failure_reason_is("StressBailout");
5073 }
5074 #endif
5075
5076 CloneMap& Compile::clone_map() { return _clone_map; }
5077 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5078
5079 void NodeCloneInfo::dump_on(outputStream* st) const {
5080 st->print(" {%d:%d} ", idx(), gen());
5081 }
5082
5083 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5084 uint64_t val = value(old->_idx);
5085 NodeCloneInfo cio(val);
5086 assert(val != 0, "old node should be in the map");
5087 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5088 insert(nnn->_idx, cin.get());
5089 #ifndef PRODUCT
5090 if (is_debug()) {
5091 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5092 }
5093 #endif
5094 }
5095
5096 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5097 NodeCloneInfo cio(value(old->_idx));
5098 if (cio.get() == 0) {
5099 cio.set(old->_idx, 0);
5100 insert(old->_idx, cio.get());
5101 #ifndef PRODUCT
5102 if (is_debug()) {
5103 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5104 }
5105 #endif
5106 }
5107 clone(old, nnn, gen);
5108 }
5109
5110 int CloneMap::max_gen() const {
5111 int g = 0;
5112 DictI di(_dict);
5113 for(; di.test(); ++di) {
5114 int t = gen(di._key);
5115 if (g < t) {
5116 g = t;
5117 #ifndef PRODUCT
5118 if (is_debug()) {
5119 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5120 }
5121 #endif
5122 }
5123 }
5124 return g;
5125 }
5126
5127 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5128 uint64_t val = value(key);
5129 if (val != 0) {
5130 NodeCloneInfo ni(val);
5131 ni.dump_on(st);
5132 }
5133 }
5134
5135 void Compile::shuffle_macro_nodes() {
5136 shuffle_array(*C, _macro_nodes);
5137 }
5138
5139 // Move Allocate nodes to the start of the list
5140 void Compile::sort_macro_nodes() {
5141 int count = macro_count();
5142 int allocates = 0;
5143 for (int i = 0; i < count; i++) {
5144 Node* n = macro_node(i);
5145 if (n->is_Allocate()) {
5146 if (i != allocates) {
5147 Node* tmp = macro_node(allocates);
5148 _macro_nodes.at_put(allocates, n);
5149 _macro_nodes.at_put(i, tmp);
5150 }
5151 allocates++;
5152 }
5153 }
5154 }
5155
5156 void Compile::print_method(CompilerPhaseType compile_phase, int level, Node* n) {
5157 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5158 EventCompilerPhase event(UNTIMED);
5159 if (event.should_commit()) {
5160 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, compile_phase, C->_compile_id, level);
5161 }
5162 #ifndef PRODUCT
5163 ResourceMark rm;
5164 stringStream ss;
5165 ss.print_raw(CompilerPhaseTypeHelper::to_description(compile_phase));
5166 int iter = ++_igv_phase_iter[compile_phase];
5167 if (iter > 1) {
5168 ss.print(" %d", iter);
5169 }
5170 if (n != nullptr) {
5171 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5172 if (n->is_Call()) {
5173 CallNode* call = n->as_Call();
5174 if (call->_name != nullptr) {
5175 // E.g. uncommon traps etc.
5176 ss.print(" - %s", call->_name);
5177 } else if (call->is_CallJava()) {
5178 CallJavaNode* call_java = call->as_CallJava();
5179 if (call_java->method() != nullptr) {
5180 ss.print(" -");
5181 call_java->method()->print_short_name(&ss);
5182 }
5183 }
5184 }
5185 }
5186
5187 const char* name = ss.as_string();
5188 if (should_print_igv(level)) {
5189 _igv_printer->print_graph(name);
5190 }
5191 if (should_print_phase(level)) {
5192 print_phase(name);
5193 }
5194 if (should_print_ideal_phase(compile_phase)) {
5195 print_ideal_ir(CompilerPhaseTypeHelper::to_name(compile_phase));
5196 }
5197 #endif
5198 C->_latest_stage_start_counter.stamp();
5199 }
5200
5201 // Only used from CompileWrapper
5202 void Compile::begin_method() {
5203 #ifndef PRODUCT
5204 if (_method != nullptr && should_print_igv(1)) {
5205 _igv_printer->begin_method();
5206 }
5207 #endif
5208 C->_latest_stage_start_counter.stamp();
5209 }
5210
5211 // Only used from CompileWrapper
5212 void Compile::end_method() {
5213 EventCompilerPhase event(UNTIMED);
5214 if (event.should_commit()) {
5215 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5216 }
5217
5218 #ifndef PRODUCT
5219 if (_method != nullptr && should_print_igv(1)) {
5220 _igv_printer->end_method();
5221 }
5222 #endif
5223 }
5224
5225 #ifndef PRODUCT
5226 bool Compile::should_print_phase(const int level) const {
5227 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5228 _method != nullptr; // Do not print phases for stubs.
5229 }
5230
5231 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5232 return _directive->should_print_ideal_phase(cpt);
5233 }
5234
5235 void Compile::init_igv() {
5236 if (_igv_printer == nullptr) {
5237 _igv_printer = IdealGraphPrinter::printer();
5238 _igv_printer->set_compile(this);
5239 }
5240 }
5241
5242 bool Compile::should_print_igv(const int level) {
5243 PRODUCT_RETURN_(return false;);
5244
5245 if (PrintIdealGraphLevel < 0) { // disabled by the user
5246 return false;
5247 }
5248
5249 bool need = directive()->IGVPrintLevelOption >= level;
5250 if (need) {
5251 Compile::init_igv();
5252 }
5253 return need;
5254 }
5255
5256 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5257 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5258
5259 // Called from debugger. Prints method to the default file with the default phase name.
5260 // This works regardless of any Ideal Graph Visualizer flags set or not.
5261 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5262 void igv_print(void* sp, void* fp, void* pc) {
5263 frame fr(sp, fp, pc);
5264 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5265 }
5266
5267 // Same as igv_print() above but with a specified phase name.
5268 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5269 frame fr(sp, fp, pc);
5270 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5271 }
5272
5273 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5274 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5275 // This works regardless of any Ideal Graph Visualizer flags set or not.
5276 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5277 void igv_print(bool network, void* sp, void* fp, void* pc) {
5278 frame fr(sp, fp, pc);
5279 if (network) {
5280 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5281 } else {
5282 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5283 }
5284 }
5285
5286 // Same as igv_print(bool network, ...) above but with a specified phase name.
5287 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5288 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5289 frame fr(sp, fp, pc);
5290 if (network) {
5291 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5292 } else {
5293 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5294 }
5295 }
5296
5297 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5298 void igv_print_default() {
5299 Compile::current()->print_method(PHASE_DEBUG, 0);
5300 }
5301
5302 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5303 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5304 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5305 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5306 void igv_append(void* sp, void* fp, void* pc) {
5307 frame fr(sp, fp, pc);
5308 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5309 }
5310
5311 // Same as igv_append(...) above but with a specified phase name.
5312 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5313 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5314 frame fr(sp, fp, pc);
5315 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5316 }
5317
5318 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5319 const char* file_name = "custom_debug.xml";
5320 if (_debug_file_printer == nullptr) {
5321 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5322 } else {
5323 _debug_file_printer->update_compiled_method(C->method());
5324 }
5325 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5326 _debug_file_printer->print_graph(phase_name, fr);
5327 }
5328
5329 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5330 ResourceMark rm;
5331 GrowableArray<const Node*> empty_list;
5332 igv_print_graph_to_network(phase_name, empty_list, fr);
5333 }
5334
5335 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5336 if (_debug_network_printer == nullptr) {
5337 _debug_network_printer = new IdealGraphPrinter(C);
5338 } else {
5339 _debug_network_printer->update_compiled_method(C->method());
5340 }
5341 tty->print_cr("Method printed over network stream to IGV");
5342 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5343 }
5344 #endif // !PRODUCT
5345
5346 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5347 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5348 return value;
5349 }
5350 Node* result = nullptr;
5351 if (bt == T_BYTE) {
5352 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5353 result = new RShiftINode(result, phase->intcon(24));
5354 } else if (bt == T_BOOLEAN) {
5355 result = new AndINode(value, phase->intcon(0xFF));
5356 } else if (bt == T_CHAR) {
5357 result = new AndINode(value,phase->intcon(0xFFFF));
5358 } else {
5359 assert(bt == T_SHORT, "unexpected narrow type");
5360 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5361 result = new RShiftINode(result, phase->intcon(16));
5362 }
5363 if (transform_res) {
5364 result = phase->transform(result);
5365 }
5366 return result;
5367 }
5368
5369 void Compile::record_method_not_compilable_oom() {
5370 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5371 }
5372
5373 #ifndef PRODUCT
5374 // Collects all the control inputs from nodes on the worklist and from their data dependencies
5375 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
5376 // Follow non-control edges until we reach CFG nodes
5377 for (uint i = 0; i < worklist.size(); i++) {
5378 const Node* n = worklist.at(i);
5379 for (uint j = 0; j < n->req(); j++) {
5380 Node* in = n->in(j);
5381 if (in == nullptr || in->is_Root()) {
5382 continue;
5383 }
5384 if (in->is_CFG()) {
5385 if (in->is_Call()) {
5386 // The return value of a call is only available if the call did not result in an exception
5387 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
5388 if (control_proj_use->is_Catch()) {
5389 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
5390 candidates.push(fall_through);
5391 continue;
5392 }
5393 }
5394
5395 if (in->is_Multi()) {
5396 // We got here by following data inputs so we should only have one control use
5397 // (no IfNode, etc)
5398 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
5399 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
5400 } else {
5401 candidates.push(in);
5402 }
5403 } else {
5404 worklist.push(in);
5405 }
5406 }
5407 }
5408 }
5409
5410 // Returns the candidate node that is a descendant to all the other candidates
5411 static Node* pick_control(Unique_Node_List& candidates) {
5412 Unique_Node_List worklist;
5413 worklist.copy(candidates);
5414
5415 // Traverse backwards through the CFG
5416 for (uint i = 0; i < worklist.size(); i++) {
5417 const Node* n = worklist.at(i);
5418 if (n->is_Root()) {
5419 continue;
5420 }
5421 for (uint j = 0; j < n->req(); j++) {
5422 // Skip backedge of loops to avoid cycles
5423 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
5424 continue;
5425 }
5426
5427 Node* pred = n->in(j);
5428 if (pred != nullptr && pred != n && pred->is_CFG()) {
5429 worklist.push(pred);
5430 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
5431 candidates.remove(pred);
5432 }
5433 }
5434 }
5435
5436 assert(candidates.size() == 1, "unexpected control flow");
5437 return candidates.at(0);
5438 }
5439
5440 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
5441 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
5442 call->init_req((*pos)++, parm);
5443 const BasicType bt = parm->bottom_type()->basic_type();
5444 if (bt == T_LONG || bt == T_DOUBLE) {
5445 call->init_req((*pos)++, half);
5446 }
5447 }
5448
5449 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
5450 Node* parm0, Node* parm1,
5451 Node* parm2, Node* parm3,
5452 Node* parm4, Node* parm5,
5453 Node* parm6) const {
5454 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
5455 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
5456 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
5457
5458 // find the most suitable control input
5459 Unique_Node_List worklist, candidates;
5460 if (parm0 != nullptr) { worklist.push(parm0);
5461 if (parm1 != nullptr) { worklist.push(parm1);
5462 if (parm2 != nullptr) { worklist.push(parm2);
5463 if (parm3 != nullptr) { worklist.push(parm3);
5464 if (parm4 != nullptr) { worklist.push(parm4);
5465 if (parm5 != nullptr) { worklist.push(parm5);
5466 if (parm6 != nullptr) { worklist.push(parm6);
5467 /* close each nested if ===> */ } } } } } } }
5468 find_candidate_control_inputs(worklist, candidates);
5469 Node* control = nullptr;
5470 if (candidates.size() == 0) {
5471 control = C->start()->proj_out(TypeFunc::Control);
5472 } else {
5473 control = pick_control(candidates);
5474 }
5475
5476 // find all the previous users of the control we picked
5477 GrowableArray<Node*> users_of_control;
5478 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
5479 Node* use = control->fast_out(i);
5480 if (use->is_CFG() && use != control) {
5481 users_of_control.push(use);
5482 }
5483 }
5484
5485 // we do not actually care about IO and memory as it uses neither
5486 call->init_req(TypeFunc::Control, control);
5487 call->init_req(TypeFunc::I_O, top());
5488 call->init_req(TypeFunc::Memory, top());
5489 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
5490 call->init_req(TypeFunc::ReturnAdr, top());
5491
5492 int pos = TypeFunc::Parms;
5493 call->init_req(pos++, str_node);
5494 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
5495 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
5496 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
5497 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
5498 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
5499 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
5500 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
5501 /* close each nested if ===> */ } } } } } } }
5502 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
5503
5504 call = gvn->transform(call);
5505 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
5506
5507 // rewire previous users to have the new call as control instead
5508 PhaseIterGVN* igvn = gvn->is_IterGVN();
5509 for (int i = 0; i < users_of_control.length(); i++) {
5510 Node* use = users_of_control.at(i);
5511 for (uint j = 0; j < use->req(); j++) {
5512 if (use->in(j) == control) {
5513 if (igvn != nullptr) {
5514 igvn->replace_input_of(use, j, call_control_proj);
5515 } else {
5516 gvn->hash_delete(use);
5517 use->set_req(j, call_control_proj);
5518 gvn->hash_insert(use);
5519 }
5520 }
5521 }
5522 }
5523
5524 return call;
5525 }
5526 #endif // !PRODUCT