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 || StressCountedLoop) {
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. The reserved GC slots are used, for instance, by
888 // slowpath GC barrier stubs that need to save registers to the stack when
889 // making calls to GC runtime.
890 int reserved_gc_slots = BarrierSet::barrier_set()->barrier_set_c2()->reserved_slots();
891 int next_slot = fixed_slots() + (reserved_gc_slots + 1) * sizeof(address) / VMRegImpl::stack_slot_size;
892 set_fixed_slots(next_slot);
893
894 // Compute when to use implicit null checks. Used by matching trap based
895 // nodes and NullCheck optimization.
896 set_allowed_deopt_reasons();
897
898 // Now generate code
899 Code_Gen();
900 }
901
902 //------------------------------Compile----------------------------------------
903 // Compile a runtime stub
904 Compile::Compile(ciEnv* ci_env,
905 TypeFunc_generator generator,
906 address stub_function,
907 const char* stub_name,
908 StubId stub_id,
909 int is_fancy_jump,
910 bool pass_tls,
911 bool return_pc,
912 DirectiveSet* directive)
913 : Phase(Compiler),
914 _compile_id(0),
915 _options(Options::for_runtime_stub()),
916 _method(nullptr),
917 _entry_bci(InvocationEntryBci),
918 _stub_function(stub_function),
919 _stub_name(stub_name),
920 _stub_id(stub_id),
921 _stub_entry_point(nullptr),
922 _max_node_limit(MaxNodeLimit),
923 _post_loop_opts_phase(false),
924 _merge_stores_phase(false),
925 _allow_macro_nodes(true),
926 _inlining_progress(false),
927 _inlining_incrementally(false),
928 _has_reserved_stack_access(false),
929 #ifndef PRODUCT
930 _igv_idx(0),
931 _trace_opto_output(directive->TraceOptoOutputOption),
932 #endif
933 _clinit_barrier_on_entry(false),
934 _stress_seed(0),
935 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
936 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
937 _env(ci_env),
938 _directive(directive),
939 _log(ci_env->log()),
940 _first_failure_details(nullptr),
941 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
942 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
943 _congraph(nullptr),
944 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
945 _unique(0),
946 _dead_node_count(0),
947 _dead_node_list(comp_arena()),
948 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
949 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
950 _node_arena(&_node_arena_one),
951 _mach_constant_base_node(nullptr),
952 _Compile_types(mtCompiler, Arena::Tag::tag_type),
953 _initial_gvn(nullptr),
954 _igvn_worklist(nullptr),
955 _types(nullptr),
956 _node_hash(nullptr),
957 _has_mh_late_inlines(false),
958 _oom(false),
959 _replay_inline_data(nullptr),
960 _inline_printer(this),
961 _java_calls(0),
962 _inner_loops(0),
963 _FIRST_STACK_mask(comp_arena()),
964 _interpreter_frame_size(0),
965 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
966 _output(nullptr),
967 #ifndef PRODUCT
968 _in_dump_cnt(0),
969 #endif
970 _allowed_reasons(0) {
971 C = this;
972
973 // try to reuse an existing stub
974 {
975 BlobId blob_id = StubInfo::blob(_stub_id);
976 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
977 if (blob != nullptr) {
978 RuntimeStub* rs = blob->as_runtime_stub();
979 _stub_entry_point = rs->entry_point();
980 return;
981 }
982 }
983
984 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
985 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
986
987 #ifndef PRODUCT
988 set_print_assembly(PrintFrameConverterAssembly);
989 set_parsed_irreducible_loop(false);
990 #else
991 set_print_assembly(false); // Must initialize.
992 #endif
993 set_has_irreducible_loop(false); // no loops
994
995 CompileWrapper cw(this);
996 Init(/*do_aliasing=*/ false);
997 init_tf((*generator)());
998
999 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1000 _types = new (comp_arena()) Type_Array(comp_arena());
1001 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1002
1003 if (StressLCM || StressGCM || StressBailout) {
1004 initialize_stress_seed(directive);
1005 }
1006
1007 {
1008 PhaseGVN gvn;
1009 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1010 gvn.transform(top());
1011
1012 GraphKit kit;
1013 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1014 }
1015
1016 NOT_PRODUCT( verify_graph_edges(); )
1017
1018 Code_Gen();
1019 }
1020
1021 Compile::~Compile() {
1022 delete _first_failure_details;
1023 };
1024
1025 //------------------------------Init-------------------------------------------
1026 // Prepare for a single compilation
1027 void Compile::Init(bool aliasing) {
1028 _do_aliasing = aliasing;
1029 _unique = 0;
1030 _regalloc = nullptr;
1031
1032 _tf = nullptr; // filled in later
1033 _top = nullptr; // cached later
1034 _matcher = nullptr; // filled in later
1035 _cfg = nullptr; // filled in later
1036
1037 _node_note_array = nullptr;
1038 _default_node_notes = nullptr;
1039 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1040
1041 _immutable_memory = nullptr; // filled in at first inquiry
1042
1043 #ifdef ASSERT
1044 _phase_optimize_finished = false;
1045 _phase_verify_ideal_loop = false;
1046 _exception_backedge = false;
1047 _type_verify = nullptr;
1048 #endif
1049
1050 // Globally visible Nodes
1051 // First set TOP to null to give safe behavior during creation of RootNode
1052 set_cached_top_node(nullptr);
1053 set_root(new RootNode());
1054 // Now that you have a Root to point to, create the real TOP
1055 set_cached_top_node( new ConNode(Type::TOP) );
1056 set_recent_alloc(nullptr, nullptr);
1057
1058 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1059 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1060 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1061 env()->set_dependencies(new Dependencies(env()));
1062
1063 _fixed_slots = 0;
1064 set_has_split_ifs(false);
1065 set_has_loops(false); // first approximation
1066 set_has_stringbuilder(false);
1067 set_has_boxed_value(false);
1068 _trap_can_recompile = false; // no traps emitted yet
1069 _major_progress = true; // start out assuming good things will happen
1070 set_has_unsafe_access(false);
1071 set_max_vector_size(0);
1072 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1073 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1074 set_decompile_count(0);
1075
1076 #ifndef PRODUCT
1077 _phase_counter = 0;
1078 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1079 #endif
1080
1081 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1082 _loop_opts_cnt = LoopOptsCount;
1083 set_do_inlining(Inline);
1084 set_max_inline_size(MaxInlineSize);
1085 set_freq_inline_size(FreqInlineSize);
1086 set_do_scheduling(OptoScheduling);
1087
1088 set_do_vector_loop(false);
1089 set_has_monitors(false);
1090 set_has_scoped_access(false);
1091
1092 if (AllowVectorizeOnDemand) {
1093 if (has_method() && _directive->VectorizeOption) {
1094 set_do_vector_loop(true);
1095 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());})
1096 } else if (has_method() && method()->name() != nullptr &&
1097 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1098 set_do_vector_loop(true);
1099 }
1100 }
1101 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1102 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());})
1103
1104 _max_node_limit = _directive->MaxNodeLimitOption;
1105
1106 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1107 set_clinit_barrier_on_entry(true);
1108 }
1109 if (debug_info()->recording_non_safepoints()) {
1110 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1111 (comp_arena(), 8, 0, nullptr));
1112 set_default_node_notes(Node_Notes::make(this));
1113 }
1114
1115 const int grow_ats = 16;
1116 _max_alias_types = grow_ats;
1117 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1118 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1119 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1120 {
1121 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1122 }
1123 // Initialize the first few types.
1124 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1125 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1126 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1127 _num_alias_types = AliasIdxRaw+1;
1128 // Zero out the alias type cache.
1129 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1130 // A null adr_type hits in the cache right away. Preload the right answer.
1131 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1132 }
1133
1134 #ifdef ASSERT
1135 // Verify that the current StartNode is valid.
1136 void Compile::verify_start(StartNode* s) const {
1137 assert(failing_internal() || s == start(), "should be StartNode");
1138 }
1139 #endif
1140
1141 /**
1142 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1143 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1144 * the ideal graph.
1145 */
1146 StartNode* Compile::start() const {
1147 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1148 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1149 Node* start = root()->fast_out(i);
1150 if (start->is_Start()) {
1151 return start->as_Start();
1152 }
1153 }
1154 fatal("Did not find Start node!");
1155 return nullptr;
1156 }
1157
1158 //-------------------------------immutable_memory-------------------------------------
1159 // Access immutable memory
1160 Node* Compile::immutable_memory() {
1161 if (_immutable_memory != nullptr) {
1162 return _immutable_memory;
1163 }
1164 StartNode* s = start();
1165 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1166 Node *p = s->fast_out(i);
1167 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1168 _immutable_memory = p;
1169 return _immutable_memory;
1170 }
1171 }
1172 ShouldNotReachHere();
1173 return nullptr;
1174 }
1175
1176 //----------------------set_cached_top_node------------------------------------
1177 // Install the cached top node, and make sure Node::is_top works correctly.
1178 void Compile::set_cached_top_node(Node* tn) {
1179 if (tn != nullptr) verify_top(tn);
1180 Node* old_top = _top;
1181 _top = tn;
1182 // Calling Node::setup_is_top allows the nodes the chance to adjust
1183 // their _out arrays.
1184 if (_top != nullptr) _top->setup_is_top();
1185 if (old_top != nullptr) old_top->setup_is_top();
1186 assert(_top == nullptr || top()->is_top(), "");
1187 }
1188
1189 #ifdef ASSERT
1190 uint Compile::count_live_nodes_by_graph_walk() {
1191 Unique_Node_List useful(comp_arena());
1192 // Get useful node list by walking the graph.
1193 identify_useful_nodes(useful);
1194 return useful.size();
1195 }
1196
1197 void Compile::print_missing_nodes() {
1198
1199 // Return if CompileLog is null and PrintIdealNodeCount is false.
1200 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1201 return;
1202 }
1203
1204 // This is an expensive function. It is executed only when the user
1205 // specifies VerifyIdealNodeCount option or otherwise knows the
1206 // additional work that needs to be done to identify reachable nodes
1207 // by walking the flow graph and find the missing ones using
1208 // _dead_node_list.
1209
1210 Unique_Node_List useful(comp_arena());
1211 // Get useful node list by walking the graph.
1212 identify_useful_nodes(useful);
1213
1214 uint l_nodes = C->live_nodes();
1215 uint l_nodes_by_walk = useful.size();
1216
1217 if (l_nodes != l_nodes_by_walk) {
1218 if (_log != nullptr) {
1219 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1220 _log->stamp();
1221 _log->end_head();
1222 }
1223 VectorSet& useful_member_set = useful.member_set();
1224 int last_idx = l_nodes_by_walk;
1225 for (int i = 0; i < last_idx; i++) {
1226 if (useful_member_set.test(i)) {
1227 if (_dead_node_list.test(i)) {
1228 if (_log != nullptr) {
1229 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1230 }
1231 if (PrintIdealNodeCount) {
1232 // Print the log message to tty
1233 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1234 useful.at(i)->dump();
1235 }
1236 }
1237 }
1238 else if (! _dead_node_list.test(i)) {
1239 if (_log != nullptr) {
1240 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1241 }
1242 if (PrintIdealNodeCount) {
1243 // Print the log message to tty
1244 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1245 }
1246 }
1247 }
1248 if (_log != nullptr) {
1249 _log->tail("mismatched_nodes");
1250 }
1251 }
1252 }
1253 void Compile::record_modified_node(Node* n) {
1254 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1255 _modified_nodes->push(n);
1256 }
1257 }
1258
1259 void Compile::remove_modified_node(Node* n) {
1260 if (_modified_nodes != nullptr) {
1261 _modified_nodes->remove(n);
1262 }
1263 }
1264 #endif
1265
1266 #ifndef PRODUCT
1267 void Compile::verify_top(Node* tn) const {
1268 if (tn != nullptr) {
1269 assert(tn->is_Con(), "top node must be a constant");
1270 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1271 assert(tn->in(0) != nullptr, "must have live top node");
1272 }
1273 }
1274 #endif
1275
1276
1277 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1278
1279 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1280 guarantee(arr != nullptr, "");
1281 int num_blocks = arr->length();
1282 if (grow_by < num_blocks) grow_by = num_blocks;
1283 int num_notes = grow_by * _node_notes_block_size;
1284 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1285 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1286 while (num_notes > 0) {
1287 arr->append(notes);
1288 notes += _node_notes_block_size;
1289 num_notes -= _node_notes_block_size;
1290 }
1291 assert(num_notes == 0, "exact multiple, please");
1292 }
1293
1294 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1295 if (source == nullptr || dest == nullptr) return false;
1296
1297 if (dest->is_Con())
1298 return false; // Do not push debug info onto constants.
1299
1300 #ifdef ASSERT
1301 // Leave a bread crumb trail pointing to the original node:
1302 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1303 dest->set_debug_orig(source);
1304 }
1305 #endif
1306
1307 if (node_note_array() == nullptr)
1308 return false; // Not collecting any notes now.
1309
1310 // This is a copy onto a pre-existing node, which may already have notes.
1311 // If both nodes have notes, do not overwrite any pre-existing notes.
1312 Node_Notes* source_notes = node_notes_at(source->_idx);
1313 if (source_notes == nullptr || source_notes->is_clear()) return false;
1314 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1315 if (dest_notes == nullptr || dest_notes->is_clear()) {
1316 return set_node_notes_at(dest->_idx, source_notes);
1317 }
1318
1319 Node_Notes merged_notes = (*source_notes);
1320 // The order of operations here ensures that dest notes will win...
1321 merged_notes.update_from(dest_notes);
1322 return set_node_notes_at(dest->_idx, &merged_notes);
1323 }
1324
1325
1326 //--------------------------allow_range_check_smearing-------------------------
1327 // Gating condition for coalescing similar range checks.
1328 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1329 // single covering check that is at least as strong as any of them.
1330 // If the optimization succeeds, the simplified (strengthened) range check
1331 // will always succeed. If it fails, we will deopt, and then give up
1332 // on the optimization.
1333 bool Compile::allow_range_check_smearing() const {
1334 // If this method has already thrown a range-check,
1335 // assume it was because we already tried range smearing
1336 // and it failed.
1337 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1338 return !already_trapped;
1339 }
1340
1341
1342 //------------------------------flatten_alias_type-----------------------------
1343 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1344 assert(do_aliasing(), "Aliasing should be enabled");
1345 int offset = tj->offset();
1346 TypePtr::PTR ptr = tj->ptr();
1347
1348 // Known instance (scalarizable allocation) alias only with itself.
1349 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1350 tj->is_oopptr()->is_known_instance();
1351
1352 // Process weird unsafe references.
1353 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1354 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1355 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1356 tj = TypeOopPtr::BOTTOM;
1357 ptr = tj->ptr();
1358 offset = tj->offset();
1359 }
1360
1361 // Array pointers need some flattening
1362 const TypeAryPtr* ta = tj->isa_aryptr();
1363 if (ta && ta->is_stable()) {
1364 // Erase stability property for alias analysis.
1365 tj = ta = ta->cast_to_stable(false);
1366 }
1367 if( ta && is_known_inst ) {
1368 if ( offset != Type::OffsetBot &&
1369 offset > arrayOopDesc::length_offset_in_bytes() ) {
1370 offset = Type::OffsetBot; // Flatten constant access into array body only
1371 tj = ta = ta->
1372 remove_speculative()->
1373 cast_to_ptr_type(ptr)->
1374 with_offset(offset);
1375 }
1376 } else if (ta != nullptr) {
1377 // Common slices
1378 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1379 return TypeAryPtr::RANGE;
1380 } else if (offset == oopDesc::klass_offset_in_bytes()) {
1381 return TypeInstPtr::KLASS;
1382 } else if (offset == oopDesc::mark_offset_in_bytes()) {
1383 return TypeInstPtr::MARK;
1384 }
1385
1386 // Remove size and stability
1387 const TypeAry* normalized_ary = TypeAry::make(ta->elem(), TypeInt::POS, false);
1388 // Remove ptr, const_oop, and offset
1389 if (ta->elem() == Type::BOTTOM) {
1390 // Bottom array (meet of int[] and byte[] for example), accesses to it will be done with
1391 // Unsafe. This should alias with all arrays. For now just leave it as it is (this is
1392 // incorrect, see JDK-8331133).
1393 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, false, Type::OffsetBot);
1394 } else if (ta->elem()->make_oopptr() != nullptr) {
1395 // Object arrays, all of them share the same slice
1396 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, TypeInt::POS, false);
1397 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, tary, nullptr, false, Type::OffsetBot);
1398 } else {
1399 // Primitive arrays
1400 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, ta->exact_klass(), true, Type::OffsetBot);
1401 }
1402
1403 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1404 // cannot be distinguished by bytecode alone.
1405 if (ta->elem() == TypeInt::BOOL) {
1406 tj = ta = TypeAryPtr::BYTES;
1407 }
1408 }
1409
1410 // Oop pointers need some flattening
1411 const TypeInstPtr *to = tj->isa_instptr();
1412 if (to && to != TypeOopPtr::BOTTOM) {
1413 ciInstanceKlass* ik = to->instance_klass();
1414 if( ptr == TypePtr::Constant ) {
1415 if (ik != ciEnv::current()->Class_klass() ||
1416 offset < ik->layout_helper_size_in_bytes()) {
1417 // No constant oop pointers (such as Strings); they alias with
1418 // unknown strings.
1419 assert(!is_known_inst, "not scalarizable allocation");
1420 tj = to = to->
1421 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1422 remove_speculative()->
1423 cast_to_ptr_type(TypePtr::BotPTR)->
1424 cast_to_exactness(false);
1425 }
1426 } else if( is_known_inst ) {
1427 tj = to; // Keep NotNull and klass_is_exact for instance type
1428 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1429 // During the 2nd round of IterGVN, NotNull castings are removed.
1430 // Make sure the Bottom and NotNull variants alias the same.
1431 // Also, make sure exact and non-exact variants alias the same.
1432 tj = to = to->
1433 remove_speculative()->
1434 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1435 cast_to_ptr_type(TypePtr::BotPTR)->
1436 cast_to_exactness(false);
1437 }
1438 if (to->speculative() != nullptr) {
1439 tj = to = to->remove_speculative();
1440 }
1441 // Canonicalize the holder of this field
1442 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1443 // First handle header references such as a LoadKlassNode, even if the
1444 // object's klass is unloaded at compile time (4965979).
1445 if (!is_known_inst) { // Do it only for non-instance types
1446 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1447 }
1448 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1449 // Static fields are in the space above the normal instance
1450 // fields in the java.lang.Class instance.
1451 if (ik != ciEnv::current()->Class_klass()) {
1452 to = nullptr;
1453 tj = TypeOopPtr::BOTTOM;
1454 offset = tj->offset();
1455 }
1456 } else {
1457 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1458 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1459 assert(tj->offset() == offset, "no change to offset expected");
1460 bool xk = to->klass_is_exact();
1461 int instance_id = to->instance_id();
1462
1463 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1464 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1465 // its interfaces are included.
1466 if (xk && ik->equals(canonical_holder)) {
1467 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1468 } else {
1469 assert(xk || !is_known_inst, "Known instance should be exact type");
1470 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1471 }
1472 }
1473 }
1474
1475 // Klass pointers to object array klasses need some flattening
1476 const TypeKlassPtr *tk = tj->isa_klassptr();
1477 if( tk ) {
1478 // If we are referencing a field within a Klass, we need
1479 // to assume the worst case of an Object. Both exact and
1480 // inexact types must flatten to the same alias class so
1481 // use NotNull as the PTR.
1482 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1483 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1484 env()->Object_klass(),
1485 offset);
1486 }
1487
1488 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1489 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1490 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1491 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1492 } else {
1493 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1494 }
1495 }
1496
1497 // Check for precise loads from the primary supertype array and force them
1498 // to the supertype cache alias index. Check for generic array loads from
1499 // the primary supertype array and also force them to the supertype cache
1500 // alias index. Since the same load can reach both, we need to merge
1501 // these 2 disparate memories into the same alias class. Since the
1502 // primary supertype array is read-only, there's no chance of confusion
1503 // where we bypass an array load and an array store.
1504 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1505 if (offset == Type::OffsetBot ||
1506 (offset >= primary_supers_offset &&
1507 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1508 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1509 offset = in_bytes(Klass::secondary_super_cache_offset());
1510 tj = tk = tk->with_offset(offset);
1511 }
1512 }
1513
1514 // Flatten all Raw pointers together.
1515 if (tj->base() == Type::RawPtr)
1516 tj = TypeRawPtr::BOTTOM;
1517
1518 if (tj->base() == Type::AnyPtr)
1519 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1520
1521 offset = tj->offset();
1522 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1523
1524 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1525 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1526 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1527 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1528 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1529 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1530 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1531 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1532 assert( tj->ptr() != TypePtr::TopPTR &&
1533 tj->ptr() != TypePtr::AnyNull &&
1534 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1535 // assert( tj->ptr() != TypePtr::Constant ||
1536 // tj->base() == Type::RawPtr ||
1537 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1538
1539 return tj;
1540 }
1541
1542 void Compile::AliasType::Init(int i, const TypePtr* at) {
1543 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1544 _index = i;
1545 _adr_type = at;
1546 _field = nullptr;
1547 _element = nullptr;
1548 _is_rewritable = true; // default
1549 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1550 if (atoop != nullptr && atoop->is_known_instance()) {
1551 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1552 _general_index = Compile::current()->get_alias_index(gt);
1553 } else {
1554 _general_index = 0;
1555 }
1556 }
1557
1558 BasicType Compile::AliasType::basic_type() const {
1559 if (element() != nullptr) {
1560 const Type* element = adr_type()->is_aryptr()->elem();
1561 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1562 } if (field() != nullptr) {
1563 return field()->layout_type();
1564 } else {
1565 return T_ILLEGAL; // unknown
1566 }
1567 }
1568
1569 //---------------------------------print_on------------------------------------
1570 #ifndef PRODUCT
1571 void Compile::AliasType::print_on(outputStream* st) {
1572 if (index() < 10)
1573 st->print("@ <%d> ", index());
1574 else st->print("@ <%d>", index());
1575 st->print(is_rewritable() ? " " : " RO");
1576 int offset = adr_type()->offset();
1577 if (offset == Type::OffsetBot)
1578 st->print(" +any");
1579 else st->print(" +%-3d", offset);
1580 st->print(" in ");
1581 adr_type()->dump_on(st);
1582 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1583 if (field() != nullptr && tjp) {
1584 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1585 tjp->offset() != field()->offset_in_bytes()) {
1586 st->print(" != ");
1587 field()->print();
1588 st->print(" ***");
1589 }
1590 }
1591 }
1592
1593 void print_alias_types() {
1594 Compile* C = Compile::current();
1595 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1596 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1597 C->alias_type(idx)->print_on(tty);
1598 tty->cr();
1599 }
1600 }
1601 #endif
1602
1603
1604 //----------------------------probe_alias_cache--------------------------------
1605 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1606 intptr_t key = (intptr_t) adr_type;
1607 key ^= key >> logAliasCacheSize;
1608 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1609 }
1610
1611
1612 //-----------------------------grow_alias_types--------------------------------
1613 void Compile::grow_alias_types() {
1614 const int old_ats = _max_alias_types; // how many before?
1615 const int new_ats = old_ats; // how many more?
1616 const int grow_ats = old_ats+new_ats; // how many now?
1617 _max_alias_types = grow_ats;
1618 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1619 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1620 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1621 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1622 }
1623
1624
1625 //--------------------------------find_alias_type------------------------------
1626 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1627 if (!do_aliasing()) {
1628 return alias_type(AliasIdxBot);
1629 }
1630
1631 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1632 if (ace->_adr_type == adr_type) {
1633 return alias_type(ace->_index);
1634 }
1635
1636 // Handle special cases.
1637 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1638 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1639
1640 // Do it the slow way.
1641 const TypePtr* flat = flatten_alias_type(adr_type);
1642
1643 #ifdef ASSERT
1644 {
1645 ResourceMark rm;
1646 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1647 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1648 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1649 Type::str(adr_type));
1650 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1651 const TypeOopPtr* foop = flat->is_oopptr();
1652 // Scalarizable allocations have exact klass always.
1653 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1654 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1655 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1656 Type::str(foop), Type::str(xoop));
1657 }
1658 }
1659 #endif
1660
1661 int idx = AliasIdxTop;
1662 for (int i = 0; i < num_alias_types(); i++) {
1663 if (alias_type(i)->adr_type() == flat) {
1664 idx = i;
1665 break;
1666 }
1667 }
1668
1669 if (idx == AliasIdxTop) {
1670 if (no_create) return nullptr;
1671 // Grow the array if necessary.
1672 if (_num_alias_types == _max_alias_types) grow_alias_types();
1673 // Add a new alias type.
1674 idx = _num_alias_types++;
1675 _alias_types[idx]->Init(idx, flat);
1676 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1677 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1678 if (flat->isa_instptr()) {
1679 if (flat->offset() == java_lang_Class::klass_offset()
1680 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1681 alias_type(idx)->set_rewritable(false);
1682 }
1683 if (flat->isa_aryptr()) {
1684 #ifdef ASSERT
1685 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1686 // (T_BYTE has the weakest alignment and size restrictions...)
1687 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1688 #endif
1689 if (flat->offset() == TypePtr::OffsetBot) {
1690 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1691 }
1692 }
1693 if (flat->isa_klassptr()) {
1694 if (UseCompactObjectHeaders) {
1695 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1696 alias_type(idx)->set_rewritable(false);
1697 }
1698 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1699 alias_type(idx)->set_rewritable(false);
1700 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1701 alias_type(idx)->set_rewritable(false);
1702 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1703 alias_type(idx)->set_rewritable(false);
1704 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1705 alias_type(idx)->set_rewritable(false);
1706 }
1707
1708 if (flat->isa_instklassptr()) {
1709 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1710 alias_type(idx)->set_rewritable(false);
1711 }
1712 }
1713 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1714 // but the base pointer type is not distinctive enough to identify
1715 // references into JavaThread.)
1716
1717 // Check for final fields.
1718 const TypeInstPtr* tinst = flat->isa_instptr();
1719 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1720 ciField* field;
1721 if (tinst->const_oop() != nullptr &&
1722 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1723 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1724 // static field
1725 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1726 field = k->get_field_by_offset(tinst->offset(), true);
1727 } else {
1728 ciInstanceKlass *k = tinst->instance_klass();
1729 field = k->get_field_by_offset(tinst->offset(), false);
1730 }
1731 assert(field == nullptr ||
1732 original_field == nullptr ||
1733 (field->holder() == original_field->holder() &&
1734 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1735 field->is_static() == original_field->is_static()), "wrong field?");
1736 // Set field() and is_rewritable() attributes.
1737 if (field != nullptr) alias_type(idx)->set_field(field);
1738 }
1739 }
1740
1741 // Fill the cache for next time.
1742 ace->_adr_type = adr_type;
1743 ace->_index = idx;
1744 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1745
1746 // Might as well try to fill the cache for the flattened version, too.
1747 AliasCacheEntry* face = probe_alias_cache(flat);
1748 if (face->_adr_type == nullptr) {
1749 face->_adr_type = flat;
1750 face->_index = idx;
1751 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1752 }
1753
1754 return alias_type(idx);
1755 }
1756
1757
1758 Compile::AliasType* Compile::alias_type(ciField* field) {
1759 const TypeOopPtr* t;
1760 if (field->is_static())
1761 t = TypeInstPtr::make(field->holder()->java_mirror());
1762 else
1763 t = TypeOopPtr::make_from_klass_raw(field->holder());
1764 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1765 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1766 return atp;
1767 }
1768
1769
1770 //------------------------------have_alias_type--------------------------------
1771 bool Compile::have_alias_type(const TypePtr* adr_type) {
1772 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1773 if (ace->_adr_type == adr_type) {
1774 return true;
1775 }
1776
1777 // Handle special cases.
1778 if (adr_type == nullptr) return true;
1779 if (adr_type == TypePtr::BOTTOM) return true;
1780
1781 return find_alias_type(adr_type, true, nullptr) != nullptr;
1782 }
1783
1784 //-----------------------------must_alias--------------------------------------
1785 // True if all values of the given address type are in the given alias category.
1786 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1787 if (alias_idx == AliasIdxBot) return true; // the universal category
1788 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1789 if (alias_idx == AliasIdxTop) return false; // the empty category
1790 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1791
1792 // the only remaining possible overlap is identity
1793 int adr_idx = get_alias_index(adr_type);
1794 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1795 assert(adr_idx == alias_idx ||
1796 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1797 && adr_type != TypeOopPtr::BOTTOM),
1798 "should not be testing for overlap with an unsafe pointer");
1799 return adr_idx == alias_idx;
1800 }
1801
1802 //------------------------------can_alias--------------------------------------
1803 // True if any values of the given address type are in the given alias category.
1804 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1805 if (alias_idx == AliasIdxTop) return false; // the empty category
1806 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1807 // Known instance doesn't alias with bottom memory
1808 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1809 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1810
1811 // the only remaining possible overlap is identity
1812 int adr_idx = get_alias_index(adr_type);
1813 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1814 return adr_idx == alias_idx;
1815 }
1816
1817 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1818 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1819 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1820 if (parse_predicate_count() == 0) {
1821 return;
1822 }
1823 for (int i = 0; i < parse_predicate_count(); i++) {
1824 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1825 parse_predicate->mark_useless(igvn);
1826 }
1827 _parse_predicates.clear();
1828 }
1829
1830 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1831 if (!n->for_post_loop_opts_igvn()) {
1832 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1833 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1834 _for_post_loop_igvn.append(n);
1835 }
1836 }
1837
1838 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1839 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1840 _for_post_loop_igvn.remove(n);
1841 }
1842
1843 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1844 // Verify that all previous optimizations produced a valid graph
1845 // at least to this point, even if no loop optimizations were done.
1846 PhaseIdealLoop::verify(igvn);
1847
1848 if (_print_phase_loop_opts) {
1849 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1850 }
1851 C->set_post_loop_opts_phase(); // no more loop opts allowed
1852
1853 assert(!C->major_progress(), "not cleared");
1854
1855 if (_for_post_loop_igvn.length() > 0) {
1856 while (_for_post_loop_igvn.length() > 0) {
1857 Node* n = _for_post_loop_igvn.pop();
1858 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1859 igvn._worklist.push(n);
1860 }
1861 igvn.optimize();
1862 if (failing()) return;
1863 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1864 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1865
1866 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1867 if (C->major_progress()) {
1868 C->clear_major_progress(); // ensure that major progress is now clear
1869 }
1870 }
1871 }
1872
1873 void Compile::record_for_merge_stores_igvn(Node* n) {
1874 if (!n->for_merge_stores_igvn()) {
1875 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1876 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1877 _for_merge_stores_igvn.append(n);
1878 }
1879 }
1880
1881 void Compile::remove_from_merge_stores_igvn(Node* n) {
1882 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1883 _for_merge_stores_igvn.remove(n);
1884 }
1885
1886 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1887 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1888 // the stores, and we merge the wrong sequence of stores.
1889 // Example:
1890 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1891 // Apply MergeStores:
1892 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1893 // Remove more RangeChecks:
1894 // StoreI [ StoreL ] StoreI
1895 // But now it would have been better to do this instead:
1896 // [ StoreL ] [ StoreL ]
1897 //
1898 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1899 // since we never unset _merge_stores_phase.
1900 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1901 C->set_merge_stores_phase();
1902
1903 if (_for_merge_stores_igvn.length() > 0) {
1904 while (_for_merge_stores_igvn.length() > 0) {
1905 Node* n = _for_merge_stores_igvn.pop();
1906 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1907 igvn._worklist.push(n);
1908 }
1909 igvn.optimize();
1910 if (failing()) return;
1911 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1912 print_method(PHASE_AFTER_MERGE_STORES, 3);
1913 }
1914 }
1915
1916 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1917 if (OptimizeUnstableIf) {
1918 _unstable_if_traps.append(trap);
1919 }
1920 }
1921
1922 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1923 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1924 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1925 Node* n = trap->uncommon_trap();
1926 if (!useful.member(n)) {
1927 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1928 }
1929 }
1930 }
1931
1932 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1933 // or fold-compares case. Return true if succeed or not found.
1934 //
1935 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1936 // false and ask fold-compares to yield.
1937 //
1938 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1939 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1940 // when deoptimization does happen.
1941 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1942 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1943 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1944 if (trap->uncommon_trap() == unc) {
1945 if (yield && trap->modified()) {
1946 return false;
1947 }
1948 _unstable_if_traps.delete_at(i);
1949 break;
1950 }
1951 }
1952 return true;
1953 }
1954
1955 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1956 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1957 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1958 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1959 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1960 CallStaticJavaNode* unc = trap->uncommon_trap();
1961 int next_bci = trap->next_bci();
1962 bool modified = trap->modified();
1963
1964 if (next_bci != -1 && !modified) {
1965 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1966 JVMState* jvms = unc->jvms();
1967 ciMethod* method = jvms->method();
1968 ciBytecodeStream iter(method);
1969
1970 iter.force_bci(jvms->bci());
1971 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1972 Bytecodes::Code c = iter.cur_bc();
1973 Node* lhs = nullptr;
1974 Node* rhs = nullptr;
1975 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1976 lhs = unc->peek_operand(0);
1977 rhs = unc->peek_operand(1);
1978 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1979 lhs = unc->peek_operand(0);
1980 }
1981
1982 ResourceMark rm;
1983 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1984 assert(live_locals.is_valid(), "broken liveness info");
1985 int len = (int)live_locals.size();
1986
1987 for (int i = 0; i < len; i++) {
1988 Node* local = unc->local(jvms, i);
1989 // kill local using the liveness of next_bci.
1990 // give up when the local looks like an operand to secure reexecution.
1991 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1992 uint idx = jvms->locoff() + i;
1993 #ifdef ASSERT
1994 if (PrintOpto && Verbose) {
1995 tty->print("[unstable_if] kill local#%d: ", idx);
1996 local->dump();
1997 tty->cr();
1998 }
1999 #endif
2000 igvn.replace_input_of(unc, idx, top());
2001 modified = true;
2002 }
2003 }
2004 }
2005
2006 // keep the mondified trap for late query
2007 if (modified) {
2008 trap->set_modified();
2009 } else {
2010 _unstable_if_traps.delete_at(i);
2011 }
2012 }
2013 igvn.optimize();
2014 }
2015
2016 // StringOpts and late inlining of string methods
2017 void Compile::inline_string_calls(bool parse_time) {
2018 {
2019 // remove useless nodes to make the usage analysis simpler
2020 ResourceMark rm;
2021 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2022 }
2023
2024 {
2025 ResourceMark rm;
2026 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2027 PhaseStringOpts pso(initial_gvn());
2028 print_method(PHASE_AFTER_STRINGOPTS, 3);
2029 }
2030
2031 // now inline anything that we skipped the first time around
2032 if (!parse_time) {
2033 _late_inlines_pos = _late_inlines.length();
2034 }
2035
2036 while (_string_late_inlines.length() > 0) {
2037 CallGenerator* cg = _string_late_inlines.pop();
2038 cg->do_late_inline();
2039 if (failing()) return;
2040 }
2041 _string_late_inlines.trunc_to(0);
2042 }
2043
2044 // Late inlining of boxing methods
2045 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2046 if (_boxing_late_inlines.length() > 0) {
2047 assert(has_boxed_value(), "inconsistent");
2048
2049 set_inlining_incrementally(true);
2050
2051 igvn_worklist()->ensure_empty(); // should be done with igvn
2052
2053 _late_inlines_pos = _late_inlines.length();
2054
2055 while (_boxing_late_inlines.length() > 0) {
2056 CallGenerator* cg = _boxing_late_inlines.pop();
2057 cg->do_late_inline();
2058 if (failing()) return;
2059 }
2060 _boxing_late_inlines.trunc_to(0);
2061
2062 inline_incrementally_cleanup(igvn);
2063
2064 set_inlining_incrementally(false);
2065 }
2066 }
2067
2068 bool Compile::inline_incrementally_one() {
2069 assert(IncrementalInline, "incremental inlining should be on");
2070 assert(_late_inlines.length() > 0, "should have been checked by caller");
2071
2072 TracePhase tp(_t_incrInline_inline);
2073
2074 set_inlining_progress(false);
2075 set_do_cleanup(false);
2076
2077 for (int i = 0; i < _late_inlines.length(); i++) {
2078 _late_inlines_pos = i+1;
2079 CallGenerator* cg = _late_inlines.at(i);
2080 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2081 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2082 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2083 if (should_stress_inlining()) {
2084 // randomly add repeated inline attempt if stress-inlining
2085 cg->call_node()->set_generator(cg);
2086 C->igvn_worklist()->push(cg->call_node());
2087 continue;
2088 }
2089 cg->do_late_inline();
2090 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2091 if (failing()) {
2092 return false;
2093 } else if (inlining_progress()) {
2094 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2095 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2096 break; // process one call site at a time
2097 } else {
2098 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2099 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2100 // Avoid potential infinite loop if node already in the IGVN list
2101 assert(false, "scheduled for IGVN during inlining attempt");
2102 } else {
2103 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2104 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2105 cg->call_node()->set_generator(cg);
2106 }
2107 }
2108 } else {
2109 // Ignore late inline direct calls when inlining is not allowed.
2110 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2111 }
2112 }
2113 // Remove processed elements.
2114 _late_inlines.remove_till(_late_inlines_pos);
2115 _late_inlines_pos = 0;
2116
2117 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2118
2119 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2120
2121 set_inlining_progress(false);
2122 set_do_cleanup(false);
2123
2124 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2125 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2126 }
2127
2128 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2129 {
2130 TracePhase tp(_t_incrInline_pru);
2131 ResourceMark rm;
2132 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2133 }
2134 {
2135 TracePhase tp(_t_incrInline_igvn);
2136 igvn.reset();
2137 igvn.optimize();
2138 if (failing()) return;
2139 }
2140 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2141 }
2142
2143 template<typename E>
2144 static void shuffle_array(Compile& C, GrowableArray<E>& array) {
2145 if (array.length() < 2) {
2146 return;
2147 }
2148 for (uint i = array.length() - 1; i >= 1; i--) {
2149 uint j = C.random() % (i + 1);
2150 swap(array.at(i), array.at(j));
2151 }
2152 }
2153
2154 void Compile::shuffle_late_inlines() {
2155 shuffle_array(*C, _late_inlines);
2156 }
2157
2158 // Perform incremental inlining until bound on number of live nodes is reached
2159 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2160 TracePhase tp(_t_incrInline);
2161
2162 set_inlining_incrementally(true);
2163 uint low_live_nodes = 0;
2164
2165 if (StressIncrementalInlining) {
2166 shuffle_late_inlines();
2167 }
2168
2169 while (_late_inlines.length() > 0) {
2170 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2171 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2172 TracePhase tp(_t_incrInline_ideal);
2173 // PhaseIdealLoop is expensive so we only try it once we are
2174 // out of live nodes and we only try it again if the previous
2175 // helped got the number of nodes down significantly
2176 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2177 if (failing()) return;
2178 low_live_nodes = live_nodes();
2179 _major_progress = true;
2180 }
2181
2182 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2183 bool do_print_inlining = print_inlining() || print_intrinsics();
2184 if (do_print_inlining || log() != nullptr) {
2185 // Print inlining message for candidates that we couldn't inline for lack of space.
2186 for (int i = 0; i < _late_inlines.length(); i++) {
2187 CallGenerator* cg = _late_inlines.at(i);
2188 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2189 if (do_print_inlining) {
2190 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2191 }
2192 log_late_inline_failure(cg, msg);
2193 }
2194 }
2195 break; // finish
2196 }
2197 }
2198
2199 igvn_worklist()->ensure_empty(); // should be done with igvn
2200
2201 if (_late_inlines.length() == 0) {
2202 break; // no more progress
2203 }
2204
2205 while (inline_incrementally_one()) {
2206 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2207 }
2208 if (failing()) return;
2209
2210 inline_incrementally_cleanup(igvn);
2211
2212 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2213
2214 if (failing()) return;
2215 }
2216
2217 igvn_worklist()->ensure_empty(); // should be done with igvn
2218
2219 if (_string_late_inlines.length() > 0) {
2220 assert(has_stringbuilder(), "inconsistent");
2221
2222 inline_string_calls(false);
2223
2224 if (failing()) return;
2225
2226 inline_incrementally_cleanup(igvn);
2227 }
2228
2229 set_inlining_incrementally(false);
2230 }
2231
2232 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2233 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2234 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2235 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2236 // as if "inlining_incrementally() == true" were set.
2237 assert(inlining_incrementally() == false, "not allowed");
2238 assert(_modified_nodes == nullptr, "not allowed");
2239 assert(_late_inlines.length() > 0, "sanity");
2240
2241 if (StressIncrementalInlining) {
2242 shuffle_late_inlines();
2243 }
2244
2245 while (_late_inlines.length() > 0) {
2246 igvn_worklist()->ensure_empty(); // should be done with igvn
2247
2248 while (inline_incrementally_one()) {
2249 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2250 }
2251 if (failing()) return;
2252
2253 inline_incrementally_cleanup(igvn);
2254 }
2255 }
2256
2257 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2258 if (_loop_opts_cnt > 0) {
2259 while (major_progress() && (_loop_opts_cnt > 0)) {
2260 TracePhase tp(_t_idealLoop);
2261 PhaseIdealLoop::optimize(igvn, mode);
2262 _loop_opts_cnt--;
2263 if (failing()) return false;
2264 if (major_progress()) {
2265 print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2266 }
2267 }
2268 }
2269 return true;
2270 }
2271
2272 // Remove edges from "root" to each SafePoint at a backward branch.
2273 // They were inserted during parsing (see add_safepoint()) to make
2274 // infinite loops without calls or exceptions visible to root, i.e.,
2275 // useful.
2276 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2277 Node *r = root();
2278 if (r != nullptr) {
2279 for (uint i = r->req(); i < r->len(); ++i) {
2280 Node *n = r->in(i);
2281 if (n != nullptr && n->is_SafePoint()) {
2282 r->rm_prec(i);
2283 if (n->outcnt() == 0) {
2284 igvn.remove_dead_node(n, PhaseIterGVN::NodeOrigin::Graph);
2285 }
2286 --i;
2287 }
2288 }
2289 // Parsing may have added top inputs to the root node (Path
2290 // leading to the Halt node proven dead). Make sure we get a
2291 // chance to clean them up.
2292 igvn._worklist.push(r);
2293 igvn.optimize();
2294 }
2295 }
2296
2297 //------------------------------Optimize---------------------------------------
2298 // Given a graph, optimize it.
2299 void Compile::Optimize() {
2300 TracePhase tp(_t_optimizer);
2301
2302 #ifndef PRODUCT
2303 if (env()->break_at_compile()) {
2304 BREAKPOINT;
2305 }
2306
2307 #endif
2308
2309 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2310 #ifdef ASSERT
2311 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2312 #endif
2313
2314 ResourceMark rm;
2315
2316 NOT_PRODUCT( verify_graph_edges(); )
2317
2318 print_method(PHASE_AFTER_PARSING, 1);
2319
2320 {
2321 // Iterative Global Value Numbering, including ideal transforms
2322 PhaseIterGVN igvn;
2323 #ifdef ASSERT
2324 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2325 #endif
2326 {
2327 TracePhase tp(_t_iterGVN);
2328 igvn.optimize(true);
2329 }
2330
2331 if (failing()) return;
2332
2333 print_method(PHASE_ITER_GVN1, 2);
2334
2335 process_for_unstable_if_traps(igvn);
2336
2337 if (failing()) return;
2338
2339 inline_incrementally(igvn);
2340
2341 print_method(PHASE_INCREMENTAL_INLINE, 2);
2342
2343 if (failing()) return;
2344
2345 if (eliminate_boxing()) {
2346 // Inline valueOf() methods now.
2347 inline_boxing_calls(igvn);
2348
2349 if (failing()) return;
2350
2351 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2352 inline_incrementally(igvn);
2353 }
2354
2355 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2356
2357 if (failing()) return;
2358 }
2359
2360 // Remove the speculative part of types and clean up the graph from
2361 // the extra CastPP nodes whose only purpose is to carry them. Do
2362 // that early so that optimizations are not disrupted by the extra
2363 // CastPP nodes.
2364 remove_speculative_types(igvn);
2365
2366 if (failing()) return;
2367
2368 // No more new expensive nodes will be added to the list from here
2369 // so keep only the actual candidates for optimizations.
2370 cleanup_expensive_nodes(igvn);
2371
2372 if (failing()) return;
2373
2374 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2375 if (EnableVectorSupport && has_vbox_nodes()) {
2376 TracePhase tp(_t_vector);
2377 PhaseVector pv(igvn);
2378 pv.optimize_vector_boxes();
2379 if (failing()) return;
2380 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2381 }
2382 assert(!has_vbox_nodes(), "sanity");
2383
2384 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2385 Compile::TracePhase tp(_t_renumberLive);
2386 igvn_worklist()->ensure_empty(); // should be done with igvn
2387 {
2388 ResourceMark rm;
2389 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2390 }
2391 igvn.reset();
2392 igvn.optimize(true);
2393 if (failing()) return;
2394 }
2395
2396 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2397 // safepoints
2398 remove_root_to_sfpts_edges(igvn);
2399
2400 if (failing()) return;
2401
2402 _print_phase_loop_opts = has_loops();
2403 if (_print_phase_loop_opts) {
2404 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2405 }
2406
2407 // Perform escape analysis
2408 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2409 if (has_loops()) {
2410 // Cleanup graph (remove dead nodes).
2411 TracePhase tp(_t_idealLoop);
2412 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2413 if (failing()) return;
2414 }
2415 bool progress;
2416 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2417 do {
2418 ConnectionGraph::do_analysis(this, &igvn);
2419
2420 if (failing()) return;
2421
2422 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2423
2424 // Optimize out fields loads from scalar replaceable allocations.
2425 igvn.optimize(true);
2426 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2427
2428 if (failing()) return;
2429
2430 if (congraph() != nullptr && macro_count() > 0) {
2431 TracePhase tp(_t_macroEliminate);
2432 PhaseMacroExpand mexp(igvn);
2433 mexp.eliminate_macro_nodes();
2434 if (failing()) return;
2435 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2436
2437 igvn.set_delay_transform(false);
2438 igvn.optimize();
2439 if (failing()) return;
2440
2441 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2442 }
2443
2444 ConnectionGraph::verify_ram_nodes(this, root());
2445 if (failing()) return;
2446
2447 progress = do_iterative_escape_analysis() &&
2448 (macro_count() < mcount) &&
2449 ConnectionGraph::has_candidates(this);
2450 // Try again if candidates exist and made progress
2451 // by removing some allocations and/or locks.
2452 } while (progress);
2453 }
2454
2455 // Loop transforms on the ideal graph. Range Check Elimination,
2456 // peeling, unrolling, etc.
2457
2458 // Set loop opts counter
2459 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2460 {
2461 TracePhase tp(_t_idealLoop);
2462 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2463 _loop_opts_cnt--;
2464 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2465 if (failing()) return;
2466 }
2467 // Loop opts pass if partial peeling occurred in previous pass
2468 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2469 TracePhase tp(_t_idealLoop);
2470 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2471 _loop_opts_cnt--;
2472 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2473 if (failing()) return;
2474 }
2475 // Loop opts pass for loop-unrolling before CCP
2476 if(major_progress() && (_loop_opts_cnt > 0)) {
2477 TracePhase tp(_t_idealLoop);
2478 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2479 _loop_opts_cnt--;
2480 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2481 }
2482 if (!failing()) {
2483 // Verify that last round of loop opts produced a valid graph
2484 PhaseIdealLoop::verify(igvn);
2485 }
2486 }
2487 if (failing()) return;
2488
2489 // Conditional Constant Propagation;
2490 print_method(PHASE_BEFORE_CCP1, 2);
2491 PhaseCCP ccp( &igvn );
2492 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2493 {
2494 TracePhase tp(_t_ccp);
2495 ccp.do_transform();
2496 }
2497 print_method(PHASE_CCP1, 2);
2498
2499 assert( true, "Break here to ccp.dump_old2new_map()");
2500
2501 // Iterative Global Value Numbering, including ideal transforms
2502 {
2503 TracePhase tp(_t_iterGVN2);
2504 igvn.reset_from_igvn(&ccp);
2505 igvn.optimize(true);
2506 }
2507 print_method(PHASE_ITER_GVN2, 2);
2508
2509 if (failing()) return;
2510
2511 // Loop transforms on the ideal graph. Range Check Elimination,
2512 // peeling, unrolling, etc.
2513 if (!optimize_loops(igvn, LoopOptsDefault)) {
2514 return;
2515 }
2516
2517 if (failing()) return;
2518
2519 C->clear_major_progress(); // ensure that major progress is now clear
2520
2521 process_for_post_loop_opts_igvn(igvn);
2522
2523 process_for_merge_stores_igvn(igvn);
2524
2525 if (failing()) return;
2526
2527 #ifdef ASSERT
2528 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2529 #endif
2530
2531 {
2532 TracePhase tp(_t_macroExpand);
2533 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2534 PhaseMacroExpand mex(igvn);
2535 // Do not allow new macro nodes once we start to eliminate and expand
2536 C->reset_allow_macro_nodes();
2537 // Last attempt to eliminate macro nodes before expand
2538 mex.eliminate_macro_nodes();
2539 if (failing()) {
2540 return;
2541 }
2542 mex.eliminate_opaque_looplimit_macro_nodes();
2543 if (failing()) {
2544 return;
2545 }
2546 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2547 if (mex.expand_macro_nodes()) {
2548 assert(failing(), "must bail out w/ explicit message");
2549 return;
2550 }
2551 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2552 }
2553
2554 {
2555 TracePhase tp(_t_barrierExpand);
2556 if (bs->expand_barriers(this, igvn)) {
2557 assert(failing(), "must bail out w/ explicit message");
2558 return;
2559 }
2560 print_method(PHASE_BARRIER_EXPANSION, 2);
2561 }
2562
2563 if (C->max_vector_size() > 0) {
2564 C->optimize_logic_cones(igvn);
2565 igvn.optimize();
2566 if (failing()) return;
2567 }
2568
2569 DEBUG_ONLY( _modified_nodes = nullptr; )
2570
2571 assert(igvn._worklist.size() == 0, "not empty");
2572
2573 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2574
2575 if (_late_inlines.length() > 0) {
2576 // More opportunities to optimize virtual and MH calls.
2577 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2578 process_late_inline_calls_no_inline(igvn);
2579 if (failing()) return;
2580 }
2581 } // (End scope of igvn; run destructor if necessary for asserts.)
2582
2583 check_no_dead_use();
2584
2585 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2586 // to remove hashes to unlock nodes for modifications.
2587 C->node_hash()->clear();
2588
2589 // A method with only infinite loops has no edges entering loops from root
2590 {
2591 TracePhase tp(_t_graphReshaping);
2592 if (final_graph_reshaping()) {
2593 assert(failing(), "must bail out w/ explicit message");
2594 return;
2595 }
2596 }
2597
2598 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2599 DEBUG_ONLY(set_phase_optimize_finished();)
2600 }
2601
2602 #ifdef ASSERT
2603 void Compile::check_no_dead_use() const {
2604 ResourceMark rm;
2605 Unique_Node_List wq;
2606 wq.push(root());
2607 for (uint i = 0; i < wq.size(); ++i) {
2608 Node* n = wq.at(i);
2609 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2610 Node* u = n->fast_out(j);
2611 if (u->outcnt() == 0 && !u->is_Con()) {
2612 u->dump();
2613 fatal("no reachable node should have no use");
2614 }
2615 wq.push(u);
2616 }
2617 }
2618 }
2619 #endif
2620
2621 void Compile::inline_vector_reboxing_calls() {
2622 if (C->_vector_reboxing_late_inlines.length() > 0) {
2623 _late_inlines_pos = C->_late_inlines.length();
2624 while (_vector_reboxing_late_inlines.length() > 0) {
2625 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2626 cg->do_late_inline();
2627 if (failing()) return;
2628 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2629 }
2630 _vector_reboxing_late_inlines.trunc_to(0);
2631 }
2632 }
2633
2634 bool Compile::has_vbox_nodes() {
2635 if (C->_vector_reboxing_late_inlines.length() > 0) {
2636 return true;
2637 }
2638 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2639 Node * n = C->macro_node(macro_idx);
2640 assert(n->is_macro(), "only macro nodes expected here");
2641 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2642 return true;
2643 }
2644 }
2645 return false;
2646 }
2647
2648 //---------------------------- Bitwise operation packing optimization ---------------------------
2649
2650 static bool is_vector_unary_bitwise_op(Node* n) {
2651 return n->Opcode() == Op_XorV &&
2652 VectorNode::is_vector_bitwise_not_pattern(n);
2653 }
2654
2655 static bool is_vector_binary_bitwise_op(Node* n) {
2656 switch (n->Opcode()) {
2657 case Op_AndV:
2658 case Op_OrV:
2659 return true;
2660
2661 case Op_XorV:
2662 return !is_vector_unary_bitwise_op(n);
2663
2664 default:
2665 return false;
2666 }
2667 }
2668
2669 static bool is_vector_ternary_bitwise_op(Node* n) {
2670 return n->Opcode() == Op_MacroLogicV;
2671 }
2672
2673 static bool is_vector_bitwise_op(Node* n) {
2674 return is_vector_unary_bitwise_op(n) ||
2675 is_vector_binary_bitwise_op(n) ||
2676 is_vector_ternary_bitwise_op(n);
2677 }
2678
2679 static bool is_vector_bitwise_cone_root(Node* n) {
2680 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2681 return false;
2682 }
2683 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2684 if (is_vector_bitwise_op(n->fast_out(i))) {
2685 return false;
2686 }
2687 }
2688 return true;
2689 }
2690
2691 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2692 uint cnt = 0;
2693 if (is_vector_bitwise_op(n)) {
2694 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2695 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2696 for (uint i = 1; i < inp_cnt; i++) {
2697 Node* in = n->in(i);
2698 bool skip = VectorNode::is_all_ones_vector(in);
2699 if (!skip && !inputs.member(in)) {
2700 inputs.push(in);
2701 cnt++;
2702 }
2703 }
2704 assert(cnt <= 1, "not unary");
2705 } else {
2706 uint last_req = inp_cnt;
2707 if (is_vector_ternary_bitwise_op(n)) {
2708 last_req = inp_cnt - 1; // skip last input
2709 }
2710 for (uint i = 1; i < last_req; i++) {
2711 Node* def = n->in(i);
2712 if (!inputs.member(def)) {
2713 inputs.push(def);
2714 cnt++;
2715 }
2716 }
2717 }
2718 } else { // not a bitwise operations
2719 if (!inputs.member(n)) {
2720 inputs.push(n);
2721 cnt++;
2722 }
2723 }
2724 return cnt;
2725 }
2726
2727 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2728 Unique_Node_List useful_nodes;
2729 C->identify_useful_nodes(useful_nodes);
2730
2731 for (uint i = 0; i < useful_nodes.size(); i++) {
2732 Node* n = useful_nodes.at(i);
2733 if (is_vector_bitwise_cone_root(n)) {
2734 list.push(n);
2735 }
2736 }
2737 }
2738
2739 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2740 const TypeVect* vt,
2741 Unique_Node_List& partition,
2742 Unique_Node_List& inputs) {
2743 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2744 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2745 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2746
2747 Node* in1 = inputs.at(0);
2748 Node* in2 = inputs.at(1);
2749 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2750
2751 uint func = compute_truth_table(partition, inputs);
2752
2753 Node* pn = partition.at(partition.size() - 1);
2754 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2755 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2756 }
2757
2758 static uint extract_bit(uint func, uint pos) {
2759 return (func & (1 << pos)) >> pos;
2760 }
2761
2762 //
2763 // A macro logic node represents a truth table. It has 4 inputs,
2764 // First three inputs corresponds to 3 columns of a truth table
2765 // and fourth input captures the logic function.
2766 //
2767 // eg. fn = (in1 AND in2) OR in3;
2768 //
2769 // MacroNode(in1,in2,in3,fn)
2770 //
2771 // -----------------
2772 // in1 in2 in3 fn
2773 // -----------------
2774 // 0 0 0 0
2775 // 0 0 1 1
2776 // 0 1 0 0
2777 // 0 1 1 1
2778 // 1 0 0 0
2779 // 1 0 1 1
2780 // 1 1 0 1
2781 // 1 1 1 1
2782 //
2783
2784 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2785 int res = 0;
2786 for (int i = 0; i < 8; i++) {
2787 int bit1 = extract_bit(in1, i);
2788 int bit2 = extract_bit(in2, i);
2789 int bit3 = extract_bit(in3, i);
2790
2791 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2792 int func_bit = extract_bit(func, func_bit_pos);
2793
2794 res |= func_bit << i;
2795 }
2796 return res;
2797 }
2798
2799 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
2800 assert(n != nullptr, "");
2801 assert(eval_map.contains(n), "absent");
2802 return *(eval_map.get(n));
2803 }
2804
2805 static void eval_operands(Node* n,
2806 uint& func1, uint& func2, uint& func3,
2807 HashTable<Node*,uint>& eval_map) {
2808 assert(is_vector_bitwise_op(n), "");
2809
2810 if (is_vector_unary_bitwise_op(n)) {
2811 Node* opnd = n->in(1);
2812 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2813 opnd = n->in(2);
2814 }
2815 func1 = eval_operand(opnd, eval_map);
2816 } else if (is_vector_binary_bitwise_op(n)) {
2817 func1 = eval_operand(n->in(1), eval_map);
2818 func2 = eval_operand(n->in(2), eval_map);
2819 } else {
2820 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2821 func1 = eval_operand(n->in(1), eval_map);
2822 func2 = eval_operand(n->in(2), eval_map);
2823 func3 = eval_operand(n->in(3), eval_map);
2824 }
2825 }
2826
2827 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2828 assert(inputs.size() <= 3, "sanity");
2829 ResourceMark rm;
2830 uint res = 0;
2831 HashTable<Node*,uint> eval_map;
2832
2833 // Populate precomputed functions for inputs.
2834 // Each input corresponds to one column of 3 input truth-table.
2835 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2836 0xCC, // (_, b, _) -> b
2837 0xF0 }; // (a, _, _) -> a
2838 for (uint i = 0; i < inputs.size(); i++) {
2839 eval_map.put(inputs.at(i), input_funcs[2-i]);
2840 }
2841
2842 for (uint i = 0; i < partition.size(); i++) {
2843 Node* n = partition.at(i);
2844
2845 uint func1 = 0, func2 = 0, func3 = 0;
2846 eval_operands(n, func1, func2, func3, eval_map);
2847
2848 switch (n->Opcode()) {
2849 case Op_OrV:
2850 assert(func3 == 0, "not binary");
2851 res = func1 | func2;
2852 break;
2853 case Op_AndV:
2854 assert(func3 == 0, "not binary");
2855 res = func1 & func2;
2856 break;
2857 case Op_XorV:
2858 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2859 assert(func2 == 0 && func3 == 0, "not unary");
2860 res = (~func1) & 0xFF;
2861 } else {
2862 assert(func3 == 0, "not binary");
2863 res = func1 ^ func2;
2864 }
2865 break;
2866 case Op_MacroLogicV:
2867 // Ordering of inputs may change during evaluation of sub-tree
2868 // containing MacroLogic node as a child node, thus a re-evaluation
2869 // makes sure that function is evaluated in context of current
2870 // inputs.
2871 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2872 break;
2873
2874 default: assert(false, "not supported: %s", n->Name());
2875 }
2876 assert(res <= 0xFF, "invalid");
2877 eval_map.put(n, res);
2878 }
2879 return res;
2880 }
2881
2882 // Criteria under which nodes gets packed into a macro logic node:-
2883 // 1) Parent and both child nodes are all unmasked or masked with
2884 // same predicates.
2885 // 2) Masked parent can be packed with left child if it is predicated
2886 // and both have same predicates.
2887 // 3) Masked parent can be packed with right child if its un-predicated
2888 // or has matching predication condition.
2889 // 4) An unmasked parent can be packed with an unmasked child.
2890 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2891 assert(partition.size() == 0, "not empty");
2892 assert(inputs.size() == 0, "not empty");
2893 if (is_vector_ternary_bitwise_op(n)) {
2894 return false;
2895 }
2896
2897 bool is_unary_op = is_vector_unary_bitwise_op(n);
2898 if (is_unary_op) {
2899 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2900 return false; // too few inputs
2901 }
2902
2903 bool pack_left_child = true;
2904 bool pack_right_child = true;
2905
2906 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2907 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2908
2909 int left_child_input_cnt = 0;
2910 int right_child_input_cnt = 0;
2911
2912 bool parent_is_predicated = n->is_predicated_vector();
2913 bool left_child_predicated = n->in(1)->is_predicated_vector();
2914 bool right_child_predicated = n->in(2)->is_predicated_vector();
2915
2916 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2917 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2918 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2919
2920 do {
2921 if (pack_left_child && left_child_LOP &&
2922 ((!parent_is_predicated && !left_child_predicated) ||
2923 ((parent_is_predicated && left_child_predicated &&
2924 parent_pred == left_child_pred)))) {
2925 partition.push(n->in(1));
2926 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2927 } else {
2928 inputs.push(n->in(1));
2929 left_child_input_cnt = 1;
2930 }
2931
2932 if (pack_right_child && right_child_LOP &&
2933 (!right_child_predicated ||
2934 (right_child_predicated && parent_is_predicated &&
2935 parent_pred == right_child_pred))) {
2936 partition.push(n->in(2));
2937 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2938 } else {
2939 inputs.push(n->in(2));
2940 right_child_input_cnt = 1;
2941 }
2942
2943 if (inputs.size() > 3) {
2944 assert(partition.size() > 0, "");
2945 inputs.clear();
2946 partition.clear();
2947 if (left_child_input_cnt > right_child_input_cnt) {
2948 pack_left_child = false;
2949 } else {
2950 pack_right_child = false;
2951 }
2952 } else {
2953 break;
2954 }
2955 } while(true);
2956
2957 if(partition.size()) {
2958 partition.push(n);
2959 }
2960
2961 return (partition.size() == 2 || partition.size() == 3) &&
2962 (inputs.size() == 2 || inputs.size() == 3);
2963 }
2964
2965 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2966 assert(is_vector_bitwise_op(n), "not a root");
2967
2968 visited.set(n->_idx);
2969
2970 // 1) Do a DFS walk over the logic cone.
2971 for (uint i = 1; i < n->req(); i++) {
2972 Node* in = n->in(i);
2973 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2974 process_logic_cone_root(igvn, in, visited);
2975 }
2976 }
2977
2978 // 2) Bottom up traversal: Merge node[s] with
2979 // the parent to form macro logic node.
2980 Unique_Node_List partition;
2981 Unique_Node_List inputs;
2982 if (compute_logic_cone(n, partition, inputs)) {
2983 const TypeVect* vt = n->bottom_type()->is_vect();
2984 Node* pn = partition.at(partition.size() - 1);
2985 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2986 if (mask == nullptr ||
2987 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2988 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2989 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2990 igvn.replace_node(n, macro_logic);
2991 }
2992 }
2993 }
2994
2995 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2996 ResourceMark rm;
2997 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2998 Unique_Node_List list;
2999 collect_logic_cone_roots(list);
3000
3001 while (list.size() > 0) {
3002 Node* n = list.pop();
3003 const TypeVect* vt = n->bottom_type()->is_vect();
3004 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3005 if (supported) {
3006 VectorSet visited(comp_arena());
3007 process_logic_cone_root(igvn, n, visited);
3008 }
3009 }
3010 }
3011 }
3012
3013 //------------------------------Code_Gen---------------------------------------
3014 // Given a graph, generate code for it
3015 void Compile::Code_Gen() {
3016 if (failing()) {
3017 return;
3018 }
3019
3020 // Perform instruction selection. You might think we could reclaim Matcher
3021 // memory PDQ, but actually the Matcher is used in generating spill code.
3022 // Internals of the Matcher (including some VectorSets) must remain live
3023 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3024 // set a bit in reclaimed memory.
3025
3026 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3027 // nodes. Mapping is only valid at the root of each matched subtree.
3028 NOT_PRODUCT( verify_graph_edges(); )
3029
3030 Matcher matcher;
3031 _matcher = &matcher;
3032 {
3033 TracePhase tp(_t_matcher);
3034 matcher.match();
3035 if (failing()) {
3036 return;
3037 }
3038 }
3039 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3040 // nodes. Mapping is only valid at the root of each matched subtree.
3041 NOT_PRODUCT( verify_graph_edges(); )
3042
3043 // If you have too many nodes, or if matching has failed, bail out
3044 check_node_count(0, "out of nodes matching instructions");
3045 if (failing()) {
3046 return;
3047 }
3048
3049 print_method(PHASE_MATCHING, 2);
3050
3051 // Build a proper-looking CFG
3052 PhaseCFG cfg(node_arena(), root(), matcher);
3053 if (failing()) {
3054 return;
3055 }
3056 _cfg = &cfg;
3057 {
3058 TracePhase tp(_t_scheduler);
3059 bool success = cfg.do_global_code_motion();
3060 if (!success) {
3061 return;
3062 }
3063
3064 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3065 NOT_PRODUCT( verify_graph_edges(); )
3066 cfg.verify();
3067 if (failing()) {
3068 return;
3069 }
3070 }
3071
3072 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3073 _regalloc = ®alloc;
3074 {
3075 TracePhase tp(_t_registerAllocation);
3076 // Perform register allocation. After Chaitin, use-def chains are
3077 // no longer accurate (at spill code) and so must be ignored.
3078 // Node->LRG->reg mappings are still accurate.
3079 _regalloc->Register_Allocate();
3080
3081 // Bail out if the allocator builds too many nodes
3082 if (failing()) {
3083 return;
3084 }
3085
3086 print_method(PHASE_REGISTER_ALLOCATION, 2);
3087 }
3088
3089 // Prior to register allocation we kept empty basic blocks in case the
3090 // the allocator needed a place to spill. After register allocation we
3091 // are not adding any new instructions. If any basic block is empty, we
3092 // can now safely remove it.
3093 {
3094 TracePhase tp(_t_blockOrdering);
3095 cfg.remove_empty_blocks();
3096 if (do_freq_based_layout()) {
3097 PhaseBlockLayout layout(cfg);
3098 } else {
3099 cfg.set_loop_alignment();
3100 }
3101 cfg.fixup_flow();
3102 cfg.remove_unreachable_blocks();
3103 cfg.verify_dominator_tree();
3104 print_method(PHASE_BLOCK_ORDERING, 3);
3105 }
3106
3107 // Apply peephole optimizations
3108 if( OptoPeephole ) {
3109 TracePhase tp(_t_peephole);
3110 PhasePeephole peep( _regalloc, cfg);
3111 peep.do_transform();
3112 print_method(PHASE_PEEPHOLE, 3);
3113 }
3114
3115 // Do late expand if CPU requires this.
3116 if (Matcher::require_postalloc_expand) {
3117 TracePhase tp(_t_postalloc_expand);
3118 cfg.postalloc_expand(_regalloc);
3119 print_method(PHASE_POSTALLOC_EXPAND, 3);
3120 }
3121
3122 #ifdef ASSERT
3123 {
3124 CompilationMemoryStatistic::do_test_allocations();
3125 if (failing()) return;
3126 }
3127 #endif
3128
3129 // Convert Nodes to instruction bits in a buffer
3130 {
3131 TracePhase tp(_t_output);
3132 PhaseOutput output;
3133 output.Output();
3134 if (failing()) return;
3135 output.install();
3136 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3137 }
3138
3139 // He's dead, Jim.
3140 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3141 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3142 }
3143
3144 //------------------------------Final_Reshape_Counts---------------------------
3145 // This class defines counters to help identify when a method
3146 // may/must be executed using hardware with only 24-bit precision.
3147 struct Final_Reshape_Counts : public StackObj {
3148 int _call_count; // count non-inlined 'common' calls
3149 int _float_count; // count float ops requiring 24-bit precision
3150 int _double_count; // count double ops requiring more precision
3151 int _java_call_count; // count non-inlined 'java' calls
3152 int _inner_loop_count; // count loops which need alignment
3153 VectorSet _visited; // Visitation flags
3154 Node_List _tests; // Set of IfNodes & PCTableNodes
3155
3156 Final_Reshape_Counts() :
3157 _call_count(0), _float_count(0), _double_count(0),
3158 _java_call_count(0), _inner_loop_count(0) { }
3159
3160 void inc_call_count () { _call_count ++; }
3161 void inc_float_count () { _float_count ++; }
3162 void inc_double_count() { _double_count++; }
3163 void inc_java_call_count() { _java_call_count++; }
3164 void inc_inner_loop_count() { _inner_loop_count++; }
3165
3166 int get_call_count () const { return _call_count ; }
3167 int get_float_count () const { return _float_count ; }
3168 int get_double_count() const { return _double_count; }
3169 int get_java_call_count() const { return _java_call_count; }
3170 int get_inner_loop_count() const { return _inner_loop_count; }
3171 };
3172
3173 //------------------------------final_graph_reshaping_impl----------------------
3174 // Implement items 1-5 from final_graph_reshaping below.
3175 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3176
3177 if ( n->outcnt() == 0 ) return; // dead node
3178 uint nop = n->Opcode();
3179
3180 // Check for 2-input instruction with "last use" on right input.
3181 // Swap to left input. Implements item (2).
3182 if( n->req() == 3 && // two-input instruction
3183 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3184 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3185 n->in(2)->outcnt() == 1 &&// right use IS a last use
3186 !n->in(2)->is_Con() ) { // right use is not a constant
3187 // Check for commutative opcode
3188 switch( nop ) {
3189 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3190 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3191 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3192 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3193 case Op_AndL: case Op_XorL: case Op_OrL:
3194 case Op_AndI: case Op_XorI: case Op_OrI: {
3195 // Move "last use" input to left by swapping inputs
3196 n->swap_edges(1, 2);
3197 break;
3198 }
3199 default:
3200 break;
3201 }
3202 }
3203
3204 #ifdef ASSERT
3205 if( n->is_Mem() ) {
3206 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3207 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3208 // oop will be recorded in oop map if load crosses safepoint
3209 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3210 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3211 "raw memory operations should have control edge");
3212 }
3213 if (n->is_MemBar()) {
3214 MemBarNode* mb = n->as_MemBar();
3215 if (mb->trailing_store() || mb->trailing_load_store()) {
3216 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3217 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3218 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3219 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3220 } else if (mb->leading()) {
3221 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3222 }
3223 }
3224 #endif
3225 // Count FPU ops and common calls, implements item (3)
3226 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3227 if (!gc_handled) {
3228 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3229 }
3230
3231 // Collect CFG split points
3232 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3233 frc._tests.push(n);
3234 }
3235 }
3236
3237 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3238 if (!UseDivMod) {
3239 return;
3240 }
3241
3242 // Check if "a % b" and "a / b" both exist
3243 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3244 if (d == nullptr) {
3245 return;
3246 }
3247
3248 // Replace them with a fused divmod if supported
3249 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3250 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3251 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3252 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3253 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3254 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3255 // DivMod node so the dependency is not lost.
3256 divmod->add_prec_from(n);
3257 divmod->add_prec_from(d);
3258 d->subsume_by(divmod->div_proj(), this);
3259 n->subsume_by(divmod->mod_proj(), this);
3260 } else {
3261 // Replace "a % b" with "a - ((a / b) * b)"
3262 Node* mult = MulNode::make(d, d->in(2), bt);
3263 Node* sub = SubNode::make(d->in(1), mult, bt);
3264 n->subsume_by(sub, this);
3265 }
3266 }
3267
3268 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3269 switch( nop ) {
3270 // Count all float operations that may use FPU
3271 case Op_AddF:
3272 case Op_SubF:
3273 case Op_MulF:
3274 case Op_DivF:
3275 case Op_NegF:
3276 case Op_ModF:
3277 case Op_ConvI2F:
3278 case Op_ConF:
3279 case Op_CmpF:
3280 case Op_CmpF3:
3281 case Op_StoreF:
3282 case Op_LoadF:
3283 // case Op_ConvL2F: // longs are split into 32-bit halves
3284 frc.inc_float_count();
3285 break;
3286
3287 case Op_ConvF2D:
3288 case Op_ConvD2F:
3289 frc.inc_float_count();
3290 frc.inc_double_count();
3291 break;
3292
3293 // Count all double operations that may use FPU
3294 case Op_AddD:
3295 case Op_SubD:
3296 case Op_MulD:
3297 case Op_DivD:
3298 case Op_NegD:
3299 case Op_ModD:
3300 case Op_ConvI2D:
3301 case Op_ConvD2I:
3302 // case Op_ConvL2D: // handled by leaf call
3303 // case Op_ConvD2L: // handled by leaf call
3304 case Op_ConD:
3305 case Op_CmpD:
3306 case Op_CmpD3:
3307 case Op_StoreD:
3308 case Op_LoadD:
3309 case Op_LoadD_unaligned:
3310 frc.inc_double_count();
3311 break;
3312 case Op_Opaque1: // Remove Opaque Nodes before matching
3313 n->subsume_by(n->in(1), this);
3314 break;
3315 case Op_CallLeafPure: {
3316 // If the pure call is not supported, then lower to a CallLeaf.
3317 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3318 CallNode* call = n->as_Call();
3319 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3320 call->_name, TypeRawPtr::BOTTOM);
3321 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3322 new_call->init_req(TypeFunc::I_O, C->top());
3323 new_call->init_req(TypeFunc::Memory, C->top());
3324 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3325 new_call->init_req(TypeFunc::FramePtr, C->top());
3326 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
3327 new_call->init_req(i, call->in(i));
3328 }
3329 n->subsume_by(new_call, this);
3330 }
3331 frc.inc_call_count();
3332 break;
3333 }
3334 case Op_CallStaticJava:
3335 case Op_CallJava:
3336 case Op_CallDynamicJava:
3337 frc.inc_java_call_count(); // Count java call site;
3338 case Op_CallRuntime:
3339 case Op_CallLeaf:
3340 case Op_CallLeafVector:
3341 case Op_CallLeafNoFP: {
3342 assert (n->is_Call(), "");
3343 CallNode *call = n->as_Call();
3344 // Count call sites where the FP mode bit would have to be flipped.
3345 // Do not count uncommon runtime calls:
3346 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3347 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3348 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3349 frc.inc_call_count(); // Count the call site
3350 } else { // See if uncommon argument is shared
3351 Node *n = call->in(TypeFunc::Parms);
3352 int nop = n->Opcode();
3353 // Clone shared simple arguments to uncommon calls, item (1).
3354 if (n->outcnt() > 1 &&
3355 !n->is_Proj() &&
3356 nop != Op_CreateEx &&
3357 nop != Op_CheckCastPP &&
3358 nop != Op_DecodeN &&
3359 nop != Op_DecodeNKlass &&
3360 !n->is_Mem() &&
3361 !n->is_Phi()) {
3362 Node *x = n->clone();
3363 call->set_req(TypeFunc::Parms, x);
3364 }
3365 }
3366 break;
3367 }
3368 case Op_StoreB:
3369 case Op_StoreC:
3370 case Op_StoreI:
3371 case Op_StoreL:
3372 case Op_CompareAndSwapB:
3373 case Op_CompareAndSwapS:
3374 case Op_CompareAndSwapI:
3375 case Op_CompareAndSwapL:
3376 case Op_CompareAndSwapP:
3377 case Op_CompareAndSwapN:
3378 case Op_WeakCompareAndSwapB:
3379 case Op_WeakCompareAndSwapS:
3380 case Op_WeakCompareAndSwapI:
3381 case Op_WeakCompareAndSwapL:
3382 case Op_WeakCompareAndSwapP:
3383 case Op_WeakCompareAndSwapN:
3384 case Op_CompareAndExchangeB:
3385 case Op_CompareAndExchangeS:
3386 case Op_CompareAndExchangeI:
3387 case Op_CompareAndExchangeL:
3388 case Op_CompareAndExchangeP:
3389 case Op_CompareAndExchangeN:
3390 case Op_GetAndAddS:
3391 case Op_GetAndAddB:
3392 case Op_GetAndAddI:
3393 case Op_GetAndAddL:
3394 case Op_GetAndSetS:
3395 case Op_GetAndSetB:
3396 case Op_GetAndSetI:
3397 case Op_GetAndSetL:
3398 case Op_GetAndSetP:
3399 case Op_GetAndSetN:
3400 case Op_StoreP:
3401 case Op_StoreN:
3402 case Op_StoreNKlass:
3403 case Op_LoadB:
3404 case Op_LoadUB:
3405 case Op_LoadUS:
3406 case Op_LoadI:
3407 case Op_LoadKlass:
3408 case Op_LoadNKlass:
3409 case Op_LoadL:
3410 case Op_LoadL_unaligned:
3411 case Op_LoadP:
3412 case Op_LoadN:
3413 case Op_LoadRange:
3414 case Op_LoadS:
3415 break;
3416
3417 case Op_AddP: { // Assert sane base pointers
3418 Node *addp = n->in(AddPNode::Address);
3419 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
3420 #ifdef _LP64
3421 if (addp->Opcode() == Op_ConP &&
3422 addp == n->in(AddPNode::Base) &&
3423 n->in(AddPNode::Offset)->is_Con()) {
3424 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3425 // on the platform and on the compressed oops mode.
3426 // Use addressing with narrow klass to load with offset on x86.
3427 // Some platforms can use the constant pool to load ConP.
3428 // Do this transformation here since IGVN will convert ConN back to ConP.
3429 const Type* t = addp->bottom_type();
3430 bool is_oop = t->isa_oopptr() != nullptr;
3431 bool is_klass = t->isa_klassptr() != nullptr;
3432
3433 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3434 (is_klass && Matcher::const_klass_prefer_decode() &&
3435 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3436 Node* nn = nullptr;
3437
3438 int op = is_oop ? Op_ConN : Op_ConNKlass;
3439
3440 // Look for existing ConN node of the same exact type.
3441 Node* r = root();
3442 uint cnt = r->outcnt();
3443 for (uint i = 0; i < cnt; i++) {
3444 Node* m = r->raw_out(i);
3445 if (m!= nullptr && m->Opcode() == op &&
3446 m->bottom_type()->make_ptr() == t) {
3447 nn = m;
3448 break;
3449 }
3450 }
3451 if (nn != nullptr) {
3452 // Decode a narrow oop to match address
3453 // [R12 + narrow_oop_reg<<3 + offset]
3454 if (is_oop) {
3455 nn = new DecodeNNode(nn, t);
3456 } else {
3457 nn = new DecodeNKlassNode(nn, t);
3458 }
3459 // Check for succeeding AddP which uses the same Base.
3460 // Otherwise we will run into the assertion above when visiting that guy.
3461 for (uint i = 0; i < n->outcnt(); ++i) {
3462 Node *out_i = n->raw_out(i);
3463 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3464 out_i->set_req(AddPNode::Base, nn);
3465 #ifdef ASSERT
3466 for (uint j = 0; j < out_i->outcnt(); ++j) {
3467 Node *out_j = out_i->raw_out(j);
3468 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3469 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3470 }
3471 #endif
3472 }
3473 }
3474 n->set_req(AddPNode::Base, nn);
3475 n->set_req(AddPNode::Address, nn);
3476 if (addp->outcnt() == 0) {
3477 addp->disconnect_inputs(this);
3478 }
3479 }
3480 }
3481 }
3482 #endif
3483 break;
3484 }
3485
3486 case Op_CastPP: {
3487 // Remove CastPP nodes to gain more freedom during scheduling but
3488 // keep the dependency they encode as control or precedence edges
3489 // (if control is set already) on memory operations. Some CastPP
3490 // nodes don't have a control (don't carry a dependency): skip
3491 // those.
3492 if (n->in(0) != nullptr) {
3493 ResourceMark rm;
3494 Unique_Node_List wq;
3495 wq.push(n);
3496 for (uint next = 0; next < wq.size(); ++next) {
3497 Node *m = wq.at(next);
3498 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3499 Node* use = m->fast_out(i);
3500 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3501 use->ensure_control_or_add_prec(n->in(0));
3502 } else {
3503 switch(use->Opcode()) {
3504 case Op_AddP:
3505 case Op_DecodeN:
3506 case Op_DecodeNKlass:
3507 case Op_CheckCastPP:
3508 case Op_CastPP:
3509 wq.push(use);
3510 break;
3511 }
3512 }
3513 }
3514 }
3515 }
3516 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3517 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3518 Node* in1 = n->in(1);
3519 const Type* t = n->bottom_type();
3520 Node* new_in1 = in1->clone();
3521 new_in1->as_DecodeN()->set_type(t);
3522
3523 if (!Matcher::narrow_oop_use_complex_address()) {
3524 //
3525 // x86, ARM and friends can handle 2 adds in addressing mode
3526 // and Matcher can fold a DecodeN node into address by using
3527 // a narrow oop directly and do implicit null check in address:
3528 //
3529 // [R12 + narrow_oop_reg<<3 + offset]
3530 // NullCheck narrow_oop_reg
3531 //
3532 // On other platforms (Sparc) we have to keep new DecodeN node and
3533 // use it to do implicit null check in address:
3534 //
3535 // decode_not_null narrow_oop_reg, base_reg
3536 // [base_reg + offset]
3537 // NullCheck base_reg
3538 //
3539 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3540 // to keep the information to which null check the new DecodeN node
3541 // corresponds to use it as value in implicit_null_check().
3542 //
3543 new_in1->set_req(0, n->in(0));
3544 }
3545
3546 n->subsume_by(new_in1, this);
3547 if (in1->outcnt() == 0) {
3548 in1->disconnect_inputs(this);
3549 }
3550 } else {
3551 n->subsume_by(n->in(1), this);
3552 if (n->outcnt() == 0) {
3553 n->disconnect_inputs(this);
3554 }
3555 }
3556 break;
3557 }
3558 case Op_CastII: {
3559 n->as_CastII()->remove_range_check_cast(this);
3560 break;
3561 }
3562 #ifdef _LP64
3563 case Op_CmpP:
3564 // Do this transformation here to preserve CmpPNode::sub() and
3565 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3566 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3567 Node* in1 = n->in(1);
3568 Node* in2 = n->in(2);
3569 if (!in1->is_DecodeNarrowPtr()) {
3570 in2 = in1;
3571 in1 = n->in(2);
3572 }
3573 assert(in1->is_DecodeNarrowPtr(), "sanity");
3574
3575 Node* new_in2 = nullptr;
3576 if (in2->is_DecodeNarrowPtr()) {
3577 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3578 new_in2 = in2->in(1);
3579 } else if (in2->Opcode() == Op_ConP) {
3580 const Type* t = in2->bottom_type();
3581 if (t == TypePtr::NULL_PTR) {
3582 assert(in1->is_DecodeN(), "compare klass to null?");
3583 // Don't convert CmpP null check into CmpN if compressed
3584 // oops implicit null check is not generated.
3585 // This will allow to generate normal oop implicit null check.
3586 if (Matcher::gen_narrow_oop_implicit_null_checks())
3587 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3588 //
3589 // This transformation together with CastPP transformation above
3590 // will generated code for implicit null checks for compressed oops.
3591 //
3592 // The original code after Optimize()
3593 //
3594 // LoadN memory, narrow_oop_reg
3595 // decode narrow_oop_reg, base_reg
3596 // CmpP base_reg, nullptr
3597 // CastPP base_reg // NotNull
3598 // Load [base_reg + offset], val_reg
3599 //
3600 // after these transformations will be
3601 //
3602 // LoadN memory, narrow_oop_reg
3603 // CmpN narrow_oop_reg, nullptr
3604 // decode_not_null narrow_oop_reg, base_reg
3605 // Load [base_reg + offset], val_reg
3606 //
3607 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3608 // since narrow oops can be used in debug info now (see the code in
3609 // final_graph_reshaping_walk()).
3610 //
3611 // At the end the code will be matched to
3612 // on x86:
3613 //
3614 // Load_narrow_oop memory, narrow_oop_reg
3615 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3616 // NullCheck narrow_oop_reg
3617 //
3618 // and on sparc:
3619 //
3620 // Load_narrow_oop memory, narrow_oop_reg
3621 // decode_not_null narrow_oop_reg, base_reg
3622 // Load [base_reg + offset], val_reg
3623 // NullCheck base_reg
3624 //
3625 } else if (t->isa_oopptr()) {
3626 new_in2 = ConNode::make(t->make_narrowoop());
3627 } else if (t->isa_klassptr()) {
3628 ciKlass* klass = t->is_klassptr()->exact_klass();
3629 if (klass->is_in_encoding_range()) {
3630 new_in2 = ConNode::make(t->make_narrowklass());
3631 }
3632 }
3633 }
3634 if (new_in2 != nullptr) {
3635 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3636 n->subsume_by(cmpN, this);
3637 if (in1->outcnt() == 0) {
3638 in1->disconnect_inputs(this);
3639 }
3640 if (in2->outcnt() == 0) {
3641 in2->disconnect_inputs(this);
3642 }
3643 }
3644 }
3645 break;
3646
3647 case Op_DecodeN:
3648 case Op_DecodeNKlass:
3649 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3650 // DecodeN could be pinned when it can't be fold into
3651 // an address expression, see the code for Op_CastPP above.
3652 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3653 break;
3654
3655 case Op_EncodeP:
3656 case Op_EncodePKlass: {
3657 Node* in1 = n->in(1);
3658 if (in1->is_DecodeNarrowPtr()) {
3659 n->subsume_by(in1->in(1), this);
3660 } else if (in1->Opcode() == Op_ConP) {
3661 const Type* t = in1->bottom_type();
3662 if (t == TypePtr::NULL_PTR) {
3663 assert(t->isa_oopptr(), "null klass?");
3664 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3665 } else if (t->isa_oopptr()) {
3666 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3667 } else if (t->isa_klassptr()) {
3668 ciKlass* klass = t->is_klassptr()->exact_klass();
3669 if (klass->is_in_encoding_range()) {
3670 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3671 } else {
3672 assert(false, "unencodable klass in ConP -> EncodeP");
3673 C->record_failure("unencodable klass in ConP -> EncodeP");
3674 }
3675 }
3676 }
3677 if (in1->outcnt() == 0) {
3678 in1->disconnect_inputs(this);
3679 }
3680 break;
3681 }
3682
3683 case Op_Proj: {
3684 if (OptimizeStringConcat || IncrementalInline) {
3685 ProjNode* proj = n->as_Proj();
3686 if (proj->_is_io_use) {
3687 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3688 // Separate projections were used for the exception path which
3689 // are normally removed by a late inline. If it wasn't inlined
3690 // then they will hang around and should just be replaced with
3691 // the original one. Merge them.
3692 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3693 if (non_io_proj != nullptr) {
3694 proj->subsume_by(non_io_proj , this);
3695 }
3696 }
3697 }
3698 break;
3699 }
3700
3701 case Op_Phi:
3702 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3703 // The EncodeP optimization may create Phi with the same edges
3704 // for all paths. It is not handled well by Register Allocator.
3705 Node* unique_in = n->in(1);
3706 assert(unique_in != nullptr, "");
3707 uint cnt = n->req();
3708 for (uint i = 2; i < cnt; i++) {
3709 Node* m = n->in(i);
3710 assert(m != nullptr, "");
3711 if (unique_in != m)
3712 unique_in = nullptr;
3713 }
3714 if (unique_in != nullptr) {
3715 n->subsume_by(unique_in, this);
3716 }
3717 }
3718 break;
3719
3720 #endif
3721
3722 case Op_ModI:
3723 handle_div_mod_op(n, T_INT, false);
3724 break;
3725
3726 case Op_ModL:
3727 handle_div_mod_op(n, T_LONG, false);
3728 break;
3729
3730 case Op_UModI:
3731 handle_div_mod_op(n, T_INT, true);
3732 break;
3733
3734 case Op_UModL:
3735 handle_div_mod_op(n, T_LONG, true);
3736 break;
3737
3738 case Op_LoadVector:
3739 case Op_StoreVector:
3740 #ifdef ASSERT
3741 // Add VerifyVectorAlignment node between adr and load / store.
3742 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3743 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3744 n->as_StoreVector()->must_verify_alignment();
3745 if (must_verify_alignment) {
3746 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3747 n->as_StoreVector()->memory_size();
3748 // The memory access should be aligned to the vector width in bytes.
3749 // However, the underlying array is possibly less well aligned, but at least
3750 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3751 // a loop we can expect at least the following alignment:
3752 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3753 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3754 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3755 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3756 jlong mask = guaranteed_alignment - 1;
3757 Node* mask_con = ConLNode::make(mask);
3758 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3759 n->set_req(MemNode::Address, va);
3760 }
3761 }
3762 #endif
3763 break;
3764
3765 case Op_LoadVectorGather:
3766 case Op_StoreVectorScatter:
3767 case Op_LoadVectorGatherMasked:
3768 case Op_StoreVectorScatterMasked:
3769 case Op_VectorCmpMasked:
3770 case Op_VectorMaskGen:
3771 case Op_LoadVectorMasked:
3772 case Op_StoreVectorMasked:
3773 break;
3774
3775 case Op_AddReductionVI:
3776 case Op_AddReductionVL:
3777 case Op_AddReductionVF:
3778 case Op_AddReductionVD:
3779 case Op_MulReductionVI:
3780 case Op_MulReductionVL:
3781 case Op_MulReductionVF:
3782 case Op_MulReductionVD:
3783 case Op_MinReductionV:
3784 case Op_MaxReductionV:
3785 case Op_UMinReductionV:
3786 case Op_UMaxReductionV:
3787 case Op_AndReductionV:
3788 case Op_OrReductionV:
3789 case Op_XorReductionV:
3790 break;
3791
3792 case Op_PackB:
3793 case Op_PackS:
3794 case Op_PackI:
3795 case Op_PackF:
3796 case Op_PackL:
3797 case Op_PackD:
3798 if (n->req()-1 > 2) {
3799 // Replace many operand PackNodes with a binary tree for matching
3800 PackNode* p = (PackNode*) n;
3801 Node* btp = p->binary_tree_pack(1, n->req());
3802 n->subsume_by(btp, this);
3803 }
3804 break;
3805 case Op_Loop:
3806 // When StressCountedLoop is enabled, this loop may intentionally avoid a counted loop conversion.
3807 // This is expected behavior for the stress mode, which exercises alternative compilation paths.
3808 if (!StressCountedLoop) {
3809 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3810 }
3811 case Op_CountedLoop:
3812 case Op_LongCountedLoop:
3813 case Op_OuterStripMinedLoop:
3814 if (n->as_Loop()->is_inner_loop()) {
3815 frc.inc_inner_loop_count();
3816 }
3817 n->as_Loop()->verify_strip_mined(0);
3818 break;
3819 case Op_LShiftI:
3820 case Op_RShiftI:
3821 case Op_URShiftI:
3822 case Op_LShiftL:
3823 case Op_RShiftL:
3824 case Op_URShiftL:
3825 if (Matcher::need_masked_shift_count) {
3826 // The cpu's shift instructions don't restrict the count to the
3827 // lower 5/6 bits. We need to do the masking ourselves.
3828 Node* in2 = n->in(2);
3829 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3830 const TypeInt* t = in2->find_int_type();
3831 if (t != nullptr && t->is_con()) {
3832 juint shift = t->get_con();
3833 if (shift > mask) { // Unsigned cmp
3834 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3835 }
3836 } else {
3837 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3838 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3839 n->set_req(2, shift);
3840 }
3841 }
3842 if (in2->outcnt() == 0) { // Remove dead node
3843 in2->disconnect_inputs(this);
3844 }
3845 }
3846 break;
3847 case Op_MemBarStoreStore:
3848 case Op_MemBarRelease:
3849 // Break the link with AllocateNode: it is no longer useful and
3850 // confuses register allocation.
3851 if (n->req() > MemBarNode::Precedent) {
3852 n->set_req(MemBarNode::Precedent, top());
3853 }
3854 break;
3855 case Op_MemBarAcquire: {
3856 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3857 // At parse time, the trailing MemBarAcquire for a volatile load
3858 // is created with an edge to the load. After optimizations,
3859 // that input may be a chain of Phis. If those phis have no
3860 // other use, then the MemBarAcquire keeps them alive and
3861 // register allocation can be confused.
3862 dead_nodes.push(n->in(MemBarNode::Precedent));
3863 n->set_req(MemBarNode::Precedent, top());
3864 }
3865 break;
3866 }
3867 case Op_Blackhole:
3868 break;
3869 case Op_RangeCheck: {
3870 RangeCheckNode* rc = n->as_RangeCheck();
3871 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3872 n->subsume_by(iff, this);
3873 frc._tests.push(iff);
3874 break;
3875 }
3876 case Op_ConvI2L: {
3877 if (!Matcher::convi2l_type_required) {
3878 // Code generation on some platforms doesn't need accurate
3879 // ConvI2L types. Widening the type can help remove redundant
3880 // address computations.
3881 n->as_Type()->set_type(TypeLong::INT);
3882 ResourceMark rm;
3883 Unique_Node_List wq;
3884 wq.push(n);
3885 for (uint next = 0; next < wq.size(); next++) {
3886 Node *m = wq.at(next);
3887
3888 for(;;) {
3889 // Loop over all nodes with identical inputs edges as m
3890 Node* k = m->find_similar(m->Opcode());
3891 if (k == nullptr) {
3892 break;
3893 }
3894 // Push their uses so we get a chance to remove node made
3895 // redundant
3896 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3897 Node* u = k->fast_out(i);
3898 if (u->Opcode() == Op_LShiftL ||
3899 u->Opcode() == Op_AddL ||
3900 u->Opcode() == Op_SubL ||
3901 u->Opcode() == Op_AddP) {
3902 wq.push(u);
3903 }
3904 }
3905 // Replace all nodes with identical edges as m with m
3906 k->subsume_by(m, this);
3907 }
3908 }
3909 }
3910 break;
3911 }
3912 case Op_CmpUL: {
3913 if (!Matcher::has_match_rule(Op_CmpUL)) {
3914 // No support for unsigned long comparisons
3915 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3916 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3917 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3918 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3919 Node* andl = new AndLNode(orl, remove_sign_mask);
3920 Node* cmp = new CmpLNode(andl, n->in(2));
3921 n->subsume_by(cmp, this);
3922 }
3923 break;
3924 }
3925 #ifdef ASSERT
3926 case Op_ConNKlass: {
3927 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3928 ciKlass* klass = tp->is_klassptr()->exact_klass();
3929 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3930 break;
3931 }
3932 #endif
3933 default:
3934 assert(!n->is_Call(), "");
3935 assert(!n->is_Mem(), "");
3936 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3937 break;
3938 }
3939 }
3940
3941 //------------------------------final_graph_reshaping_walk---------------------
3942 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3943 // requires that the walk visits a node's inputs before visiting the node.
3944 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3945 Unique_Node_List sfpt;
3946
3947 frc._visited.set(root->_idx); // first, mark node as visited
3948 uint cnt = root->req();
3949 Node *n = root;
3950 uint i = 0;
3951 while (true) {
3952 if (i < cnt) {
3953 // Place all non-visited non-null inputs onto stack
3954 Node* m = n->in(i);
3955 ++i;
3956 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3957 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3958 // compute worst case interpreter size in case of a deoptimization
3959 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3960
3961 sfpt.push(m);
3962 }
3963 cnt = m->req();
3964 nstack.push(n, i); // put on stack parent and next input's index
3965 n = m;
3966 i = 0;
3967 }
3968 } else {
3969 // Now do post-visit work
3970 final_graph_reshaping_impl(n, frc, dead_nodes);
3971 if (nstack.is_empty())
3972 break; // finished
3973 n = nstack.node(); // Get node from stack
3974 cnt = n->req();
3975 i = nstack.index();
3976 nstack.pop(); // Shift to the next node on stack
3977 }
3978 }
3979
3980 // Skip next transformation if compressed oops are not used.
3981 if (UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks())
3982 return;
3983
3984 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3985 // It could be done for an uncommon traps or any safepoints/calls
3986 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3987 while (sfpt.size() > 0) {
3988 n = sfpt.pop();
3989 JVMState *jvms = n->as_SafePoint()->jvms();
3990 assert(jvms != nullptr, "sanity");
3991 int start = jvms->debug_start();
3992 int end = n->req();
3993 bool is_uncommon = (n->is_CallStaticJava() &&
3994 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3995 for (int j = start; j < end; j++) {
3996 Node* in = n->in(j);
3997 if (in->is_DecodeNarrowPtr()) {
3998 bool safe_to_skip = true;
3999 if (!is_uncommon ) {
4000 // Is it safe to skip?
4001 for (uint i = 0; i < in->outcnt(); i++) {
4002 Node* u = in->raw_out(i);
4003 if (!u->is_SafePoint() ||
4004 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4005 safe_to_skip = false;
4006 }
4007 }
4008 }
4009 if (safe_to_skip) {
4010 n->set_req(j, in->in(1));
4011 }
4012 if (in->outcnt() == 0) {
4013 in->disconnect_inputs(this);
4014 }
4015 }
4016 }
4017 }
4018 }
4019
4020 //------------------------------final_graph_reshaping--------------------------
4021 // Final Graph Reshaping.
4022 //
4023 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4024 // and not commoned up and forced early. Must come after regular
4025 // optimizations to avoid GVN undoing the cloning. Clone constant
4026 // inputs to Loop Phis; these will be split by the allocator anyways.
4027 // Remove Opaque nodes.
4028 // (2) Move last-uses by commutative operations to the left input to encourage
4029 // Intel update-in-place two-address operations and better register usage
4030 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4031 // calls canonicalizing them back.
4032 // (3) Count the number of double-precision FP ops, single-precision FP ops
4033 // and call sites. On Intel, we can get correct rounding either by
4034 // forcing singles to memory (requires extra stores and loads after each
4035 // FP bytecode) or we can set a rounding mode bit (requires setting and
4036 // clearing the mode bit around call sites). The mode bit is only used
4037 // if the relative frequency of single FP ops to calls is low enough.
4038 // This is a key transform for SPEC mpeg_audio.
4039 // (4) Detect infinite loops; blobs of code reachable from above but not
4040 // below. Several of the Code_Gen algorithms fail on such code shapes,
4041 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4042 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4043 // Detection is by looking for IfNodes where only 1 projection is
4044 // reachable from below or CatchNodes missing some targets.
4045 // (5) Assert for insane oop offsets in debug mode.
4046
4047 bool Compile::final_graph_reshaping() {
4048 // an infinite loop may have been eliminated by the optimizer,
4049 // in which case the graph will be empty.
4050 if (root()->req() == 1) {
4051 // Do not compile method that is only a trivial infinite loop,
4052 // since the content of the loop may have been eliminated.
4053 record_method_not_compilable("trivial infinite loop");
4054 return true;
4055 }
4056
4057 // Expensive nodes have their control input set to prevent the GVN
4058 // from freely commoning them. There's no GVN beyond this point so
4059 // no need to keep the control input. We want the expensive nodes to
4060 // be freely moved to the least frequent code path by gcm.
4061 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4062 for (int i = 0; i < expensive_count(); i++) {
4063 _expensive_nodes.at(i)->set_req(0, nullptr);
4064 }
4065
4066 Final_Reshape_Counts frc;
4067
4068 // Visit everybody reachable!
4069 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4070 Node_Stack nstack(live_nodes() >> 1);
4071 Unique_Node_List dead_nodes;
4072 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4073
4074 // Check for unreachable (from below) code (i.e., infinite loops).
4075 for( uint i = 0; i < frc._tests.size(); i++ ) {
4076 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4077 // Get number of CFG targets.
4078 // Note that PCTables include exception targets after calls.
4079 uint required_outcnt = n->required_outcnt();
4080 if (n->outcnt() != required_outcnt) {
4081 // Check for a few special cases. Rethrow Nodes never take the
4082 // 'fall-thru' path, so expected kids is 1 less.
4083 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4084 if (n->in(0)->in(0)->is_Call()) {
4085 CallNode* call = n->in(0)->in(0)->as_Call();
4086 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4087 required_outcnt--; // Rethrow always has 1 less kid
4088 } else if (call->req() > TypeFunc::Parms &&
4089 call->is_CallDynamicJava()) {
4090 // Check for null receiver. In such case, the optimizer has
4091 // detected that the virtual call will always result in a null
4092 // pointer exception. The fall-through projection of this CatchNode
4093 // will not be populated.
4094 Node* arg0 = call->in(TypeFunc::Parms);
4095 if (arg0->is_Type() &&
4096 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4097 required_outcnt--;
4098 }
4099 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4100 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4101 // Check for illegal array length. In such case, the optimizer has
4102 // detected that the allocation attempt will always result in an
4103 // exception. There is no fall-through projection of this CatchNode .
4104 assert(call->is_CallStaticJava(), "static call expected");
4105 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4106 uint valid_length_test_input = call->req() - 1;
4107 Node* valid_length_test = call->in(valid_length_test_input);
4108 call->del_req(valid_length_test_input);
4109 if (valid_length_test->find_int_con(1) == 0) {
4110 required_outcnt--;
4111 }
4112 dead_nodes.push(valid_length_test);
4113 assert(n->outcnt() == required_outcnt, "malformed control flow");
4114 continue;
4115 }
4116 }
4117 }
4118
4119 // Recheck with a better notion of 'required_outcnt'
4120 if (n->outcnt() != required_outcnt) {
4121 record_method_not_compilable("malformed control flow");
4122 return true; // Not all targets reachable!
4123 }
4124 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4125 CallNode* call = n->in(0)->in(0)->as_Call();
4126 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4127 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4128 assert(call->is_CallStaticJava(), "static call expected");
4129 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4130 uint valid_length_test_input = call->req() - 1;
4131 dead_nodes.push(call->in(valid_length_test_input));
4132 call->del_req(valid_length_test_input); // valid length test useless now
4133 }
4134 }
4135 // Check that I actually visited all kids. Unreached kids
4136 // must be infinite loops.
4137 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4138 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4139 record_method_not_compilable("infinite loop");
4140 return true; // Found unvisited kid; must be unreach
4141 }
4142
4143 // Here so verification code in final_graph_reshaping_walk()
4144 // always see an OuterStripMinedLoopEnd
4145 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4146 IfNode* init_iff = n->as_If();
4147 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4148 n->subsume_by(iff, this);
4149 }
4150 }
4151
4152 while (dead_nodes.size() > 0) {
4153 Node* m = dead_nodes.pop();
4154 if (m->outcnt() == 0 && m != top()) {
4155 for (uint j = 0; j < m->req(); j++) {
4156 Node* in = m->in(j);
4157 if (in != nullptr) {
4158 dead_nodes.push(in);
4159 }
4160 }
4161 m->disconnect_inputs(this);
4162 }
4163 }
4164
4165 set_java_calls(frc.get_java_call_count());
4166 set_inner_loops(frc.get_inner_loop_count());
4167
4168 // No infinite loops, no reason to bail out.
4169 return false;
4170 }
4171
4172 //-----------------------------too_many_traps----------------------------------
4173 // Report if there are too many traps at the current method and bci.
4174 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4175 bool Compile::too_many_traps(ciMethod* method,
4176 int bci,
4177 Deoptimization::DeoptReason reason) {
4178 ciMethodData* md = method->method_data();
4179 if (md->is_empty()) {
4180 // Assume the trap has not occurred, or that it occurred only
4181 // because of a transient condition during start-up in the interpreter.
4182 return false;
4183 }
4184 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4185 if (md->has_trap_at(bci, m, reason) != 0) {
4186 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4187 // Also, if there are multiple reasons, or if there is no per-BCI record,
4188 // assume the worst.
4189 if (log())
4190 log()->elem("observe trap='%s' count='%d'",
4191 Deoptimization::trap_reason_name(reason),
4192 md->trap_count(reason));
4193 return true;
4194 } else {
4195 // Ignore method/bci and see if there have been too many globally.
4196 return too_many_traps(reason, md);
4197 }
4198 }
4199
4200 // Less-accurate variant which does not require a method and bci.
4201 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4202 ciMethodData* logmd) {
4203 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4204 // Too many traps globally.
4205 // Note that we use cumulative trap_count, not just md->trap_count.
4206 if (log()) {
4207 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4208 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4209 Deoptimization::trap_reason_name(reason),
4210 mcount, trap_count(reason));
4211 }
4212 return true;
4213 } else {
4214 // The coast is clear.
4215 return false;
4216 }
4217 }
4218
4219 //--------------------------too_many_recompiles--------------------------------
4220 // Report if there are too many recompiles at the current method and bci.
4221 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4222 // Is not eager to return true, since this will cause the compiler to use
4223 // Action_none for a trap point, to avoid too many recompilations.
4224 bool Compile::too_many_recompiles(ciMethod* method,
4225 int bci,
4226 Deoptimization::DeoptReason reason) {
4227 ciMethodData* md = method->method_data();
4228 if (md->is_empty()) {
4229 // Assume the trap has not occurred, or that it occurred only
4230 // because of a transient condition during start-up in the interpreter.
4231 return false;
4232 }
4233 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4234 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4235 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4236 Deoptimization::DeoptReason per_bc_reason
4237 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4238 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4239 if ((per_bc_reason == Deoptimization::Reason_none
4240 || md->has_trap_at(bci, m, reason) != 0)
4241 // The trap frequency measure we care about is the recompile count:
4242 && md->trap_recompiled_at(bci, m)
4243 && md->overflow_recompile_count() >= bc_cutoff) {
4244 // Do not emit a trap here if it has already caused recompilations.
4245 // Also, if there are multiple reasons, or if there is no per-BCI record,
4246 // assume the worst.
4247 if (log())
4248 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4249 Deoptimization::trap_reason_name(reason),
4250 md->trap_count(reason),
4251 md->overflow_recompile_count());
4252 return true;
4253 } else if (trap_count(reason) != 0
4254 && decompile_count() >= m_cutoff) {
4255 // Too many recompiles globally, and we have seen this sort of trap.
4256 // Use cumulative decompile_count, not just md->decompile_count.
4257 if (log())
4258 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4259 Deoptimization::trap_reason_name(reason),
4260 md->trap_count(reason), trap_count(reason),
4261 md->decompile_count(), decompile_count());
4262 return true;
4263 } else {
4264 // The coast is clear.
4265 return false;
4266 }
4267 }
4268
4269 // Compute when not to trap. Used by matching trap based nodes and
4270 // NullCheck optimization.
4271 void Compile::set_allowed_deopt_reasons() {
4272 _allowed_reasons = 0;
4273 if (is_method_compilation()) {
4274 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4275 assert(rs < BitsPerInt, "recode bit map");
4276 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4277 _allowed_reasons |= nth_bit(rs);
4278 }
4279 }
4280 }
4281 }
4282
4283 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4284 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4285 }
4286
4287 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4288 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4289 }
4290
4291 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4292 if (holder->is_initialized()) {
4293 return false;
4294 }
4295 if (holder->is_being_initialized()) {
4296 if (accessing_method->holder() == holder) {
4297 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4298 // <init>, or a static method. In all those cases, there was an initialization
4299 // barrier on the holder klass passed.
4300 if (accessing_method->is_static_initializer() ||
4301 accessing_method->is_object_initializer() ||
4302 accessing_method->is_static()) {
4303 return false;
4304 }
4305 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4306 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4307 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4308 // child class can become fully initialized while its parent class is still being initialized.
4309 if (accessing_method->is_static_initializer()) {
4310 return false;
4311 }
4312 }
4313 ciMethod* root = method(); // the root method of compilation
4314 if (root != accessing_method) {
4315 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4316 }
4317 }
4318 return true;
4319 }
4320
4321 #ifndef PRODUCT
4322 //------------------------------verify_bidirectional_edges---------------------
4323 // For each input edge to a node (ie - for each Use-Def edge), verify that
4324 // there is a corresponding Def-Use edge.
4325 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4326 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4327 uint stack_size = live_nodes() >> 4;
4328 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4329 if (root_and_safepoints != nullptr) {
4330 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4331 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4332 Node* root_or_safepoint = root_and_safepoints->at(i);
4333 // If the node is a safepoint, let's check if it still has a control input
4334 // Lack of control input signifies that this node was killed by CCP or
4335 // recursively by remove_globally_dead_node and it shouldn't be a starting
4336 // point.
4337 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4338 nstack.push(root_or_safepoint);
4339 }
4340 }
4341 } else {
4342 nstack.push(_root);
4343 }
4344
4345 while (nstack.size() > 0) {
4346 Node* n = nstack.pop();
4347 if (visited.member(n)) {
4348 continue;
4349 }
4350 visited.push(n);
4351
4352 // Walk over all input edges, checking for correspondence
4353 uint length = n->len();
4354 for (uint i = 0; i < length; i++) {
4355 Node* in = n->in(i);
4356 if (in != nullptr && !visited.member(in)) {
4357 nstack.push(in); // Put it on stack
4358 }
4359 if (in != nullptr && !in->is_top()) {
4360 // Count instances of `next`
4361 int cnt = 0;
4362 for (uint idx = 0; idx < in->_outcnt; idx++) {
4363 if (in->_out[idx] == n) {
4364 cnt++;
4365 }
4366 }
4367 assert(cnt > 0, "Failed to find Def-Use edge.");
4368 // Check for duplicate edges
4369 // walk the input array downcounting the input edges to n
4370 for (uint j = 0; j < length; j++) {
4371 if (n->in(j) == in) {
4372 cnt--;
4373 }
4374 }
4375 assert(cnt == 0, "Mismatched edge count.");
4376 } else if (in == nullptr) {
4377 assert(i == 0 || i >= n->req() ||
4378 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4379 (n->is_Unlock() && i == (n->req() - 1)) ||
4380 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4381 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4382 } else {
4383 assert(in->is_top(), "sanity");
4384 // Nothing to check.
4385 }
4386 }
4387 }
4388 }
4389
4390 //------------------------------verify_graph_edges---------------------------
4391 // Walk the Graph and verify that there is a one-to-one correspondence
4392 // between Use-Def edges and Def-Use edges in the graph.
4393 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4394 if (VerifyGraphEdges) {
4395 Unique_Node_List visited;
4396
4397 // Call graph walk to check edges
4398 verify_bidirectional_edges(visited, root_and_safepoints);
4399 if (no_dead_code) {
4400 // Now make sure that no visited node is used by an unvisited node.
4401 bool dead_nodes = false;
4402 Unique_Node_List checked;
4403 while (visited.size() > 0) {
4404 Node* n = visited.pop();
4405 checked.push(n);
4406 for (uint i = 0; i < n->outcnt(); i++) {
4407 Node* use = n->raw_out(i);
4408 if (checked.member(use)) continue; // already checked
4409 if (visited.member(use)) continue; // already in the graph
4410 if (use->is_Con()) continue; // a dead ConNode is OK
4411 // At this point, we have found a dead node which is DU-reachable.
4412 if (!dead_nodes) {
4413 tty->print_cr("*** Dead nodes reachable via DU edges:");
4414 dead_nodes = true;
4415 }
4416 use->dump(2);
4417 tty->print_cr("---");
4418 checked.push(use); // No repeats; pretend it is now checked.
4419 }
4420 }
4421 assert(!dead_nodes, "using nodes must be reachable from root");
4422 }
4423 }
4424 }
4425 #endif
4426
4427 // The Compile object keeps track of failure reasons separately from the ciEnv.
4428 // This is required because there is not quite a 1-1 relation between the
4429 // ciEnv and its compilation task and the Compile object. Note that one
4430 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4431 // to backtrack and retry without subsuming loads. Other than this backtracking
4432 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4433 // by the logic in C2Compiler.
4434 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4435 if (log() != nullptr) {
4436 log()->elem("failure reason='%s' phase='compile'", reason);
4437 }
4438 if (_failure_reason.get() == nullptr) {
4439 // Record the first failure reason.
4440 _failure_reason.set(reason);
4441 if (CaptureBailoutInformation) {
4442 _first_failure_details = new CompilationFailureInfo(reason);
4443 }
4444 } else {
4445 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4446 }
4447
4448 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4449 C->print_method(PHASE_FAILURE, 1);
4450 }
4451 _root = nullptr; // flush the graph, too
4452 }
4453
4454 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4455 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4456 _compile(Compile::current()),
4457 _log(nullptr),
4458 _dolog(CITimeVerbose)
4459 {
4460 assert(_compile != nullptr, "sanity check");
4461 assert(id != PhaseTraceId::_t_none, "Don't use none");
4462 if (_dolog) {
4463 _log = _compile->log();
4464 }
4465 if (_log != nullptr) {
4466 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4467 _log->stamp();
4468 _log->end_head();
4469 }
4470
4471 // Inform memory statistic, if enabled
4472 if (CompilationMemoryStatistic::enabled()) {
4473 CompilationMemoryStatistic::on_phase_start((int)id, name);
4474 }
4475 }
4476
4477 Compile::TracePhase::TracePhase(PhaseTraceId id)
4478 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4479
4480 Compile::TracePhase::~TracePhase() {
4481
4482 // Inform memory statistic, if enabled
4483 if (CompilationMemoryStatistic::enabled()) {
4484 CompilationMemoryStatistic::on_phase_end();
4485 }
4486
4487 if (_compile->failing_internal()) {
4488 if (_log != nullptr) {
4489 _log->done("phase");
4490 }
4491 return; // timing code, not stressing bailouts.
4492 }
4493 #ifdef ASSERT
4494 if (PrintIdealNodeCount) {
4495 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4496 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4497 }
4498
4499 if (VerifyIdealNodeCount) {
4500 _compile->print_missing_nodes();
4501 }
4502 #endif
4503
4504 if (_log != nullptr) {
4505 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4506 }
4507 }
4508
4509 //----------------------------static_subtype_check-----------------------------
4510 // Shortcut important common cases when superklass is exact:
4511 // (0) superklass is java.lang.Object (can occur in reflective code)
4512 // (1) subklass is already limited to a subtype of superklass => always ok
4513 // (2) subklass does not overlap with superklass => always fail
4514 // (3) superklass has NO subtypes and we can check with a simple compare.
4515 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4516 if (skip) {
4517 return SSC_full_test; // Let caller generate the general case.
4518 }
4519
4520 if (subk->is_java_subtype_of(superk)) {
4521 return SSC_always_true; // (0) and (1) this test cannot fail
4522 }
4523
4524 if (!subk->maybe_java_subtype_of(superk)) {
4525 return SSC_always_false; // (2) true path dead; no dynamic test needed
4526 }
4527
4528 const Type* superelem = superk;
4529 if (superk->isa_aryklassptr()) {
4530 int ignored;
4531 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4532 }
4533
4534 if (superelem->isa_instklassptr()) {
4535 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4536 if (!ik->has_subklass()) {
4537 if (!ik->is_final()) {
4538 // Add a dependency if there is a chance of a later subclass.
4539 dependencies()->assert_leaf_type(ik);
4540 }
4541 if (!superk->maybe_java_subtype_of(subk)) {
4542 return SSC_always_false;
4543 }
4544 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4545 }
4546 } else {
4547 // A primitive array type has no subtypes.
4548 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4549 }
4550
4551 return SSC_full_test;
4552 }
4553
4554 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4555 #ifdef _LP64
4556 // The scaled index operand to AddP must be a clean 64-bit value.
4557 // Java allows a 32-bit int to be incremented to a negative
4558 // value, which appears in a 64-bit register as a large
4559 // positive number. Using that large positive number as an
4560 // operand in pointer arithmetic has bad consequences.
4561 // On the other hand, 32-bit overflow is rare, and the possibility
4562 // can often be excluded, if we annotate the ConvI2L node with
4563 // a type assertion that its value is known to be a small positive
4564 // number. (The prior range check has ensured this.)
4565 // This assertion is used by ConvI2LNode::Ideal.
4566 int index_max = max_jint - 1; // array size is max_jint, index is one less
4567 if (sizetype != nullptr && sizetype->_hi > 0) {
4568 index_max = sizetype->_hi - 1;
4569 }
4570 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4571 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4572 #endif
4573 return idx;
4574 }
4575
4576 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4577 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4578 if (ctrl != nullptr) {
4579 // Express control dependency by a CastII node with a narrow type.
4580 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4581 // node from floating above the range check during loop optimizations. Otherwise, the
4582 // ConvI2L node may be eliminated independently of the range check, causing the data path
4583 // to become TOP while the control path is still there (although it's unreachable).
4584 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
4585 value = phase->transform(value);
4586 }
4587 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4588 return phase->transform(new ConvI2LNode(value, ltype));
4589 }
4590
4591 void Compile::dump_print_inlining() {
4592 inline_printer()->print_on(tty);
4593 }
4594
4595 void Compile::log_late_inline(CallGenerator* cg) {
4596 if (log() != nullptr) {
4597 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4598 cg->unique_id());
4599 JVMState* p = cg->call_node()->jvms();
4600 while (p != nullptr) {
4601 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4602 p = p->caller();
4603 }
4604 log()->tail("late_inline");
4605 }
4606 }
4607
4608 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4609 log_late_inline(cg);
4610 if (log() != nullptr) {
4611 log()->inline_fail(msg);
4612 }
4613 }
4614
4615 void Compile::log_inline_id(CallGenerator* cg) {
4616 if (log() != nullptr) {
4617 // The LogCompilation tool needs a unique way to identify late
4618 // inline call sites. This id must be unique for this call site in
4619 // this compilation. Try to have it unique across compilations as
4620 // well because it can be convenient when grepping through the log
4621 // file.
4622 // Distinguish OSR compilations from others in case CICountOSR is
4623 // on.
4624 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4625 cg->set_unique_id(id);
4626 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4627 }
4628 }
4629
4630 void Compile::log_inline_failure(const char* msg) {
4631 if (C->log() != nullptr) {
4632 C->log()->inline_fail(msg);
4633 }
4634 }
4635
4636
4637 // Dump inlining replay data to the stream.
4638 // Don't change thread state and acquire any locks.
4639 void Compile::dump_inline_data(outputStream* out) {
4640 InlineTree* inl_tree = ilt();
4641 if (inl_tree != nullptr) {
4642 out->print(" inline %d", inl_tree->count());
4643 inl_tree->dump_replay_data(out);
4644 }
4645 }
4646
4647 void Compile::dump_inline_data_reduced(outputStream* out) {
4648 assert(ReplayReduce, "");
4649
4650 InlineTree* inl_tree = ilt();
4651 if (inl_tree == nullptr) {
4652 return;
4653 }
4654 // Enable iterative replay file reduction
4655 // Output "compile" lines for depth 1 subtrees,
4656 // simulating that those trees were compiled
4657 // instead of inlined.
4658 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4659 InlineTree* sub = inl_tree->subtrees().at(i);
4660 if (sub->inline_level() != 1) {
4661 continue;
4662 }
4663
4664 ciMethod* method = sub->method();
4665 int entry_bci = -1;
4666 int comp_level = env()->task()->comp_level();
4667 out->print("compile ");
4668 method->dump_name_as_ascii(out);
4669 out->print(" %d %d", entry_bci, comp_level);
4670 out->print(" inline %d", sub->count());
4671 sub->dump_replay_data(out, -1);
4672 out->cr();
4673 }
4674 }
4675
4676 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4677 if (n1->Opcode() < n2->Opcode()) return -1;
4678 else if (n1->Opcode() > n2->Opcode()) return 1;
4679
4680 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4681 for (uint i = 1; i < n1->req(); i++) {
4682 if (n1->in(i) < n2->in(i)) return -1;
4683 else if (n1->in(i) > n2->in(i)) return 1;
4684 }
4685
4686 return 0;
4687 }
4688
4689 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4690 Node* n1 = *n1p;
4691 Node* n2 = *n2p;
4692
4693 return cmp_expensive_nodes(n1, n2);
4694 }
4695
4696 void Compile::sort_expensive_nodes() {
4697 if (!expensive_nodes_sorted()) {
4698 _expensive_nodes.sort(cmp_expensive_nodes);
4699 }
4700 }
4701
4702 bool Compile::expensive_nodes_sorted() const {
4703 for (int i = 1; i < _expensive_nodes.length(); i++) {
4704 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4705 return false;
4706 }
4707 }
4708 return true;
4709 }
4710
4711 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4712 if (_expensive_nodes.length() == 0) {
4713 return false;
4714 }
4715
4716 assert(OptimizeExpensiveOps, "optimization off?");
4717
4718 // Take this opportunity to remove dead nodes from the list
4719 int j = 0;
4720 for (int i = 0; i < _expensive_nodes.length(); i++) {
4721 Node* n = _expensive_nodes.at(i);
4722 if (!n->is_unreachable(igvn)) {
4723 assert(n->is_expensive(), "should be expensive");
4724 _expensive_nodes.at_put(j, n);
4725 j++;
4726 }
4727 }
4728 _expensive_nodes.trunc_to(j);
4729
4730 // Then sort the list so that similar nodes are next to each other
4731 // and check for at least two nodes of identical kind with same data
4732 // inputs.
4733 sort_expensive_nodes();
4734
4735 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4736 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4737 return true;
4738 }
4739 }
4740
4741 return false;
4742 }
4743
4744 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4745 if (_expensive_nodes.length() == 0) {
4746 return;
4747 }
4748
4749 assert(OptimizeExpensiveOps, "optimization off?");
4750
4751 // Sort to bring similar nodes next to each other and clear the
4752 // control input of nodes for which there's only a single copy.
4753 sort_expensive_nodes();
4754
4755 int j = 0;
4756 int identical = 0;
4757 int i = 0;
4758 bool modified = false;
4759 for (; i < _expensive_nodes.length()-1; i++) {
4760 assert(j <= i, "can't write beyond current index");
4761 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4762 identical++;
4763 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4764 continue;
4765 }
4766 if (identical > 0) {
4767 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4768 identical = 0;
4769 } else {
4770 Node* n = _expensive_nodes.at(i);
4771 igvn.replace_input_of(n, 0, nullptr);
4772 igvn.hash_insert(n);
4773 modified = true;
4774 }
4775 }
4776 if (identical > 0) {
4777 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4778 } else if (_expensive_nodes.length() >= 1) {
4779 Node* n = _expensive_nodes.at(i);
4780 igvn.replace_input_of(n, 0, nullptr);
4781 igvn.hash_insert(n);
4782 modified = true;
4783 }
4784 _expensive_nodes.trunc_to(j);
4785 if (modified) {
4786 igvn.optimize();
4787 }
4788 }
4789
4790 void Compile::add_expensive_node(Node * n) {
4791 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4792 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4793 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4794 if (OptimizeExpensiveOps) {
4795 _expensive_nodes.append(n);
4796 } else {
4797 // Clear control input and let IGVN optimize expensive nodes if
4798 // OptimizeExpensiveOps is off.
4799 n->set_req(0, nullptr);
4800 }
4801 }
4802
4803 /**
4804 * Track coarsened Lock and Unlock nodes.
4805 */
4806
4807 class Lock_List : public Node_List {
4808 uint _origin_cnt;
4809 public:
4810 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4811 uint origin_cnt() const { return _origin_cnt; }
4812 };
4813
4814 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4815 int length = locks.length();
4816 if (length > 0) {
4817 // Have to keep this list until locks elimination during Macro nodes elimination.
4818 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4819 AbstractLockNode* alock = locks.at(0);
4820 BoxLockNode* box = alock->box_node()->as_BoxLock();
4821 for (int i = 0; i < length; i++) {
4822 AbstractLockNode* lock = locks.at(i);
4823 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4824 locks_list->push(lock);
4825 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4826 if (this_box != box) {
4827 // Locking regions (BoxLock) could be Unbalanced here:
4828 // - its coarsened locks were eliminated in earlier
4829 // macro nodes elimination followed by loop unroll
4830 // - it is OSR locking region (no Lock node)
4831 // Preserve Unbalanced status in such cases.
4832 if (!this_box->is_unbalanced()) {
4833 this_box->set_coarsened();
4834 }
4835 if (!box->is_unbalanced()) {
4836 box->set_coarsened();
4837 }
4838 }
4839 }
4840 _coarsened_locks.append(locks_list);
4841 }
4842 }
4843
4844 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4845 int count = coarsened_count();
4846 for (int i = 0; i < count; i++) {
4847 Node_List* locks_list = _coarsened_locks.at(i);
4848 for (uint j = 0; j < locks_list->size(); j++) {
4849 Node* lock = locks_list->at(j);
4850 assert(lock->is_AbstractLock(), "sanity");
4851 if (!useful.member(lock)) {
4852 locks_list->yank(lock);
4853 }
4854 }
4855 }
4856 }
4857
4858 void Compile::remove_coarsened_lock(Node* n) {
4859 if (n->is_AbstractLock()) {
4860 int count = coarsened_count();
4861 for (int i = 0; i < count; i++) {
4862 Node_List* locks_list = _coarsened_locks.at(i);
4863 locks_list->yank(n);
4864 }
4865 }
4866 }
4867
4868 bool Compile::coarsened_locks_consistent() {
4869 int count = coarsened_count();
4870 for (int i = 0; i < count; i++) {
4871 bool unbalanced = false;
4872 bool modified = false; // track locks kind modifications
4873 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4874 uint size = locks_list->size();
4875 if (size == 0) {
4876 unbalanced = false; // All locks were eliminated - good
4877 } else if (size != locks_list->origin_cnt()) {
4878 unbalanced = true; // Some locks were removed from list
4879 } else {
4880 for (uint j = 0; j < size; j++) {
4881 Node* lock = locks_list->at(j);
4882 // All nodes in group should have the same state (modified or not)
4883 if (!lock->as_AbstractLock()->is_coarsened()) {
4884 if (j == 0) {
4885 // first on list was modified, the rest should be too for consistency
4886 modified = true;
4887 } else if (!modified) {
4888 // this lock was modified but previous locks on the list were not
4889 unbalanced = true;
4890 break;
4891 }
4892 } else if (modified) {
4893 // previous locks on list were modified but not this lock
4894 unbalanced = true;
4895 break;
4896 }
4897 }
4898 }
4899 if (unbalanced) {
4900 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4901 #ifdef ASSERT
4902 if (PrintEliminateLocks) {
4903 tty->print_cr("=== unbalanced coarsened locks ===");
4904 for (uint l = 0; l < size; l++) {
4905 locks_list->at(l)->dump();
4906 }
4907 }
4908 #endif
4909 record_failure(C2Compiler::retry_no_locks_coarsening());
4910 return false;
4911 }
4912 }
4913 return true;
4914 }
4915
4916 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4917 // locks coarsening optimization removed Lock/Unlock nodes from them.
4918 // Such regions become unbalanced because coarsening only removes part
4919 // of Lock/Unlock nodes in region. As result we can't execute other
4920 // locks elimination optimizations which assume all code paths have
4921 // corresponding pair of Lock/Unlock nodes - they are balanced.
4922 void Compile::mark_unbalanced_boxes() const {
4923 int count = coarsened_count();
4924 for (int i = 0; i < count; i++) {
4925 Node_List* locks_list = _coarsened_locks.at(i);
4926 uint size = locks_list->size();
4927 if (size > 0) {
4928 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4929 BoxLockNode* box = alock->box_node()->as_BoxLock();
4930 if (alock->is_coarsened()) {
4931 // coarsened_locks_consistent(), which is called before this method, verifies
4932 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4933 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4934 for (uint j = 1; j < size; j++) {
4935 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4936 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4937 if (box != this_box) {
4938 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4939 box->set_unbalanced();
4940 this_box->set_unbalanced();
4941 }
4942 }
4943 }
4944 }
4945 }
4946 }
4947
4948 /**
4949 * Remove the speculative part of types and clean up the graph
4950 */
4951 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4952 if (UseTypeSpeculation) {
4953 Unique_Node_List worklist;
4954 worklist.push(root());
4955 int modified = 0;
4956 // Go over all type nodes that carry a speculative type, drop the
4957 // speculative part of the type and enqueue the node for an igvn
4958 // which may optimize it out.
4959 for (uint next = 0; next < worklist.size(); ++next) {
4960 Node *n = worklist.at(next);
4961 if (n->is_Type()) {
4962 TypeNode* tn = n->as_Type();
4963 const Type* t = tn->type();
4964 const Type* t_no_spec = t->remove_speculative();
4965 if (t_no_spec != t) {
4966 bool in_hash = igvn.hash_delete(n);
4967 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4968 tn->set_type(t_no_spec);
4969 igvn.hash_insert(n);
4970 igvn._worklist.push(n); // give it a chance to go away
4971 modified++;
4972 }
4973 }
4974 // Iterate over outs - endless loops is unreachable from below
4975 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4976 Node *m = n->fast_out(i);
4977 if (not_a_node(m)) {
4978 continue;
4979 }
4980 worklist.push(m);
4981 }
4982 }
4983 // Drop the speculative part of all types in the igvn's type table
4984 igvn.remove_speculative_types();
4985 if (modified > 0) {
4986 igvn.optimize();
4987 if (failing()) return;
4988 }
4989 #ifdef ASSERT
4990 // Verify that after the IGVN is over no speculative type has resurfaced
4991 worklist.clear();
4992 worklist.push(root());
4993 for (uint next = 0; next < worklist.size(); ++next) {
4994 Node *n = worklist.at(next);
4995 const Type* t = igvn.type_or_null(n);
4996 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4997 if (n->is_Type()) {
4998 t = n->as_Type()->type();
4999 assert(t == t->remove_speculative(), "no more speculative types");
5000 }
5001 // Iterate over outs - endless loops is unreachable from below
5002 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5003 Node *m = n->fast_out(i);
5004 if (not_a_node(m)) {
5005 continue;
5006 }
5007 worklist.push(m);
5008 }
5009 }
5010 igvn.check_no_speculative_types();
5011 #endif
5012 }
5013 }
5014
5015 // Auxiliary methods to support randomized stressing/fuzzing.
5016
5017 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5018 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5019 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5020 FLAG_SET_ERGO(StressSeed, _stress_seed);
5021 } else {
5022 _stress_seed = StressSeed;
5023 }
5024 if (_log != nullptr) {
5025 _log->elem("stress_test seed='%u'", _stress_seed);
5026 }
5027 }
5028
5029 int Compile::random() {
5030 _stress_seed = os::next_random(_stress_seed);
5031 return static_cast<int>(_stress_seed);
5032 }
5033
5034 // This method can be called the arbitrary number of times, with current count
5035 // as the argument. The logic allows selecting a single candidate from the
5036 // running list of candidates as follows:
5037 // int count = 0;
5038 // Cand* selected = null;
5039 // while(cand = cand->next()) {
5040 // if (randomized_select(++count)) {
5041 // selected = cand;
5042 // }
5043 // }
5044 //
5045 // Including count equalizes the chances any candidate is "selected".
5046 // This is useful when we don't have the complete list of candidates to choose
5047 // from uniformly. In this case, we need to adjust the randomicity of the
5048 // selection, or else we will end up biasing the selection towards the latter
5049 // candidates.
5050 //
5051 // Quick back-envelope calculation shows that for the list of n candidates
5052 // the equal probability for the candidate to persist as "best" can be
5053 // achieved by replacing it with "next" k-th candidate with the probability
5054 // of 1/k. It can be easily shown that by the end of the run, the
5055 // probability for any candidate is converged to 1/n, thus giving the
5056 // uniform distribution among all the candidates.
5057 //
5058 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5059 #define RANDOMIZED_DOMAIN_POW 29
5060 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5061 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5062 bool Compile::randomized_select(int count) {
5063 assert(count > 0, "only positive");
5064 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5065 }
5066
5067 #ifdef ASSERT
5068 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5069 bool Compile::fail_randomly() {
5070 if ((random() % StressBailoutMean) != 0) {
5071 return false;
5072 }
5073 record_failure("StressBailout");
5074 return true;
5075 }
5076
5077 bool Compile::failure_is_artificial() {
5078 return C->failure_reason_is("StressBailout");
5079 }
5080 #endif
5081
5082 CloneMap& Compile::clone_map() { return _clone_map; }
5083 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5084
5085 void NodeCloneInfo::dump_on(outputStream* st) const {
5086 st->print(" {%d:%d} ", idx(), gen());
5087 }
5088
5089 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5090 uint64_t val = value(old->_idx);
5091 NodeCloneInfo cio(val);
5092 assert(val != 0, "old node should be in the map");
5093 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5094 insert(nnn->_idx, cin.get());
5095 #ifndef PRODUCT
5096 if (is_debug()) {
5097 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5098 }
5099 #endif
5100 }
5101
5102 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5103 NodeCloneInfo cio(value(old->_idx));
5104 if (cio.get() == 0) {
5105 cio.set(old->_idx, 0);
5106 insert(old->_idx, cio.get());
5107 #ifndef PRODUCT
5108 if (is_debug()) {
5109 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5110 }
5111 #endif
5112 }
5113 clone(old, nnn, gen);
5114 }
5115
5116 int CloneMap::max_gen() const {
5117 int g = 0;
5118 DictI di(_dict);
5119 for(; di.test(); ++di) {
5120 int t = gen(di._key);
5121 if (g < t) {
5122 g = t;
5123 #ifndef PRODUCT
5124 if (is_debug()) {
5125 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5126 }
5127 #endif
5128 }
5129 }
5130 return g;
5131 }
5132
5133 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5134 uint64_t val = value(key);
5135 if (val != 0) {
5136 NodeCloneInfo ni(val);
5137 ni.dump_on(st);
5138 }
5139 }
5140
5141 void Compile::shuffle_macro_nodes() {
5142 shuffle_array(*C, _macro_nodes);
5143 }
5144
5145 // Move Allocate nodes to the start of the list
5146 void Compile::sort_macro_nodes() {
5147 int count = macro_count();
5148 int allocates = 0;
5149 for (int i = 0; i < count; i++) {
5150 Node* n = macro_node(i);
5151 if (n->is_Allocate()) {
5152 if (i != allocates) {
5153 Node* tmp = macro_node(allocates);
5154 _macro_nodes.at_put(allocates, n);
5155 _macro_nodes.at_put(i, tmp);
5156 }
5157 allocates++;
5158 }
5159 }
5160 }
5161
5162 void Compile::print_method(CompilerPhaseType compile_phase, int level, Node* n) {
5163 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5164 EventCompilerPhase event(UNTIMED);
5165 if (event.should_commit()) {
5166 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, compile_phase, C->_compile_id, level);
5167 }
5168 #ifndef PRODUCT
5169 ResourceMark rm;
5170 stringStream ss;
5171 ss.print_raw(CompilerPhaseTypeHelper::to_description(compile_phase));
5172 int iter = ++_igv_phase_iter[compile_phase];
5173 if (iter > 1) {
5174 ss.print(" %d", iter);
5175 }
5176 if (n != nullptr) {
5177 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5178 if (n->is_Call()) {
5179 CallNode* call = n->as_Call();
5180 if (call->_name != nullptr) {
5181 // E.g. uncommon traps etc.
5182 ss.print(" - %s", call->_name);
5183 } else if (call->is_CallJava()) {
5184 CallJavaNode* call_java = call->as_CallJava();
5185 if (call_java->method() != nullptr) {
5186 ss.print(" -");
5187 call_java->method()->print_short_name(&ss);
5188 }
5189 }
5190 }
5191 }
5192
5193 const char* name = ss.as_string();
5194 if (should_print_igv(level)) {
5195 _igv_printer->print_graph(name);
5196 }
5197 if (should_print_phase(level)) {
5198 print_phase(name);
5199 }
5200 if (should_print_ideal_phase(compile_phase)) {
5201 print_ideal_ir(CompilerPhaseTypeHelper::to_name(compile_phase));
5202 }
5203 #endif
5204 C->_latest_stage_start_counter.stamp();
5205 }
5206
5207 // Only used from CompileWrapper
5208 void Compile::begin_method() {
5209 #ifndef PRODUCT
5210 if (_method != nullptr && should_print_igv(1)) {
5211 _igv_printer->begin_method();
5212 }
5213 #endif
5214 C->_latest_stage_start_counter.stamp();
5215 }
5216
5217 // Only used from CompileWrapper
5218 void Compile::end_method() {
5219 EventCompilerPhase event(UNTIMED);
5220 if (event.should_commit()) {
5221 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5222 }
5223
5224 #ifndef PRODUCT
5225 if (_method != nullptr && should_print_igv(1)) {
5226 _igv_printer->end_method();
5227 }
5228 #endif
5229 }
5230
5231 #ifndef PRODUCT
5232 bool Compile::should_print_phase(const int level) const {
5233 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5234 _method != nullptr; // Do not print phases for stubs.
5235 }
5236
5237 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5238 return _directive->should_print_ideal_phase(cpt);
5239 }
5240
5241 void Compile::init_igv() {
5242 if (_igv_printer == nullptr) {
5243 _igv_printer = IdealGraphPrinter::printer();
5244 _igv_printer->set_compile(this);
5245 }
5246 }
5247
5248 bool Compile::should_print_igv(const int level) {
5249 PRODUCT_RETURN_(return false;);
5250
5251 if (PrintIdealGraphLevel < 0) { // disabled by the user
5252 return false;
5253 }
5254
5255 bool need = directive()->IGVPrintLevelOption >= level;
5256 if (need) {
5257 Compile::init_igv();
5258 }
5259 return need;
5260 }
5261
5262 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5263 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5264
5265 // Called from debugger. Prints method to the default file with the default phase name.
5266 // This works regardless of any Ideal Graph Visualizer flags set or not.
5267 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5268 void igv_print(void* sp, void* fp, void* pc) {
5269 frame fr(sp, fp, pc);
5270 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5271 }
5272
5273 // Same as igv_print() above but with a specified phase name.
5274 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5275 frame fr(sp, fp, pc);
5276 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5277 }
5278
5279 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5280 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5281 // This works regardless of any Ideal Graph Visualizer flags set or not.
5282 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5283 void igv_print(bool network, void* sp, void* fp, void* pc) {
5284 frame fr(sp, fp, pc);
5285 if (network) {
5286 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5287 } else {
5288 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5289 }
5290 }
5291
5292 // Same as igv_print(bool network, ...) above but with a specified phase name.
5293 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5294 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5295 frame fr(sp, fp, pc);
5296 if (network) {
5297 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5298 } else {
5299 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5300 }
5301 }
5302
5303 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5304 void igv_print_default() {
5305 Compile::current()->print_method(PHASE_DEBUG, 0);
5306 }
5307
5308 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5309 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5310 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5311 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5312 void igv_append(void* sp, void* fp, void* pc) {
5313 frame fr(sp, fp, pc);
5314 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5315 }
5316
5317 // Same as igv_append(...) above but with a specified phase name.
5318 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5319 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5320 frame fr(sp, fp, pc);
5321 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5322 }
5323
5324 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5325 const char* file_name = "custom_debug.xml";
5326 if (_debug_file_printer == nullptr) {
5327 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5328 } else {
5329 _debug_file_printer->update_compiled_method(C->method());
5330 }
5331 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5332 _debug_file_printer->print_graph(phase_name, fr);
5333 }
5334
5335 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5336 ResourceMark rm;
5337 GrowableArray<const Node*> empty_list;
5338 igv_print_graph_to_network(phase_name, empty_list, fr);
5339 }
5340
5341 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5342 if (_debug_network_printer == nullptr) {
5343 _debug_network_printer = new IdealGraphPrinter(C);
5344 } else {
5345 _debug_network_printer->update_compiled_method(C->method());
5346 }
5347 tty->print_cr("Method printed over network stream to IGV");
5348 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5349 }
5350 #endif // !PRODUCT
5351
5352 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5353 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5354 return value;
5355 }
5356 Node* result = nullptr;
5357 if (bt == T_BYTE) {
5358 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5359 result = new RShiftINode(result, phase->intcon(24));
5360 } else if (bt == T_BOOLEAN) {
5361 result = new AndINode(value, phase->intcon(0xFF));
5362 } else if (bt == T_CHAR) {
5363 result = new AndINode(value,phase->intcon(0xFFFF));
5364 } else {
5365 assert(bt == T_SHORT, "unexpected narrow type");
5366 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5367 result = new RShiftINode(result, phase->intcon(16));
5368 }
5369 if (transform_res) {
5370 result = phase->transform(result);
5371 }
5372 return result;
5373 }
5374
5375 void Compile::record_method_not_compilable_oom() {
5376 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5377 }
5378
5379 #ifndef PRODUCT
5380 // Collects all the control inputs from nodes on the worklist and from their data dependencies
5381 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
5382 // Follow non-control edges until we reach CFG nodes
5383 for (uint i = 0; i < worklist.size(); i++) {
5384 const Node* n = worklist.at(i);
5385 for (uint j = 0; j < n->req(); j++) {
5386 Node* in = n->in(j);
5387 if (in == nullptr || in->is_Root()) {
5388 continue;
5389 }
5390 if (in->is_CFG()) {
5391 if (in->is_Call()) {
5392 // The return value of a call is only available if the call did not result in an exception
5393 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
5394 if (control_proj_use->is_Catch()) {
5395 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
5396 candidates.push(fall_through);
5397 continue;
5398 }
5399 }
5400
5401 if (in->is_Multi()) {
5402 // We got here by following data inputs so we should only have one control use
5403 // (no IfNode, etc)
5404 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
5405 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
5406 } else {
5407 candidates.push(in);
5408 }
5409 } else {
5410 worklist.push(in);
5411 }
5412 }
5413 }
5414 }
5415
5416 // Returns the candidate node that is a descendant to all the other candidates
5417 static Node* pick_control(Unique_Node_List& candidates) {
5418 Unique_Node_List worklist;
5419 worklist.copy(candidates);
5420
5421 // Traverse backwards through the CFG
5422 for (uint i = 0; i < worklist.size(); i++) {
5423 const Node* n = worklist.at(i);
5424 if (n->is_Root()) {
5425 continue;
5426 }
5427 for (uint j = 0; j < n->req(); j++) {
5428 // Skip backedge of loops to avoid cycles
5429 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
5430 continue;
5431 }
5432
5433 Node* pred = n->in(j);
5434 if (pred != nullptr && pred != n && pred->is_CFG()) {
5435 worklist.push(pred);
5436 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
5437 candidates.remove(pred);
5438 }
5439 }
5440 }
5441
5442 assert(candidates.size() == 1, "unexpected control flow");
5443 return candidates.at(0);
5444 }
5445
5446 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
5447 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
5448 call->init_req((*pos)++, parm);
5449 const BasicType bt = parm->bottom_type()->basic_type();
5450 if (bt == T_LONG || bt == T_DOUBLE) {
5451 call->init_req((*pos)++, half);
5452 }
5453 }
5454
5455 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
5456 Node* parm0, Node* parm1,
5457 Node* parm2, Node* parm3,
5458 Node* parm4, Node* parm5,
5459 Node* parm6) const {
5460 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
5461 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
5462 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
5463
5464 // find the most suitable control input
5465 Unique_Node_List worklist, candidates;
5466 if (parm0 != nullptr) { worklist.push(parm0);
5467 if (parm1 != nullptr) { worklist.push(parm1);
5468 if (parm2 != nullptr) { worklist.push(parm2);
5469 if (parm3 != nullptr) { worklist.push(parm3);
5470 if (parm4 != nullptr) { worklist.push(parm4);
5471 if (parm5 != nullptr) { worklist.push(parm5);
5472 if (parm6 != nullptr) { worklist.push(parm6);
5473 /* close each nested if ===> */ } } } } } } }
5474 find_candidate_control_inputs(worklist, candidates);
5475 Node* control = nullptr;
5476 if (candidates.size() == 0) {
5477 control = C->start()->proj_out(TypeFunc::Control);
5478 } else {
5479 control = pick_control(candidates);
5480 }
5481
5482 // find all the previous users of the control we picked
5483 GrowableArray<Node*> users_of_control;
5484 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
5485 Node* use = control->fast_out(i);
5486 if (use->is_CFG() && use != control) {
5487 users_of_control.push(use);
5488 }
5489 }
5490
5491 // we do not actually care about IO and memory as it uses neither
5492 call->init_req(TypeFunc::Control, control);
5493 call->init_req(TypeFunc::I_O, top());
5494 call->init_req(TypeFunc::Memory, top());
5495 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
5496 call->init_req(TypeFunc::ReturnAdr, top());
5497
5498 int pos = TypeFunc::Parms;
5499 call->init_req(pos++, str_node);
5500 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
5501 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
5502 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
5503 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
5504 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
5505 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
5506 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
5507 /* close each nested if ===> */ } } } } } } }
5508 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
5509
5510 call = gvn->transform(call);
5511 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
5512
5513 // rewire previous users to have the new call as control instead
5514 PhaseIterGVN* igvn = gvn->is_IterGVN();
5515 for (int i = 0; i < users_of_control.length(); i++) {
5516 Node* use = users_of_control.at(i);
5517 for (uint j = 0; j < use->req(); j++) {
5518 if (use->in(j) == control) {
5519 if (igvn != nullptr) {
5520 igvn->replace_input_of(use, j, call_control_proj);
5521 } else {
5522 gvn->hash_delete(use);
5523 use->set_req(j, call_control_proj);
5524 gvn->hash_insert(use);
5525 }
5526 }
5527 }
5528 }
5529
5530 return call;
5531 }
5532 #endif // !PRODUCT