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/ciFlatArray.hpp"
28 #include "ci/ciInlineKlass.hpp"
29 #include "ci/ciReplay.hpp"
30 #include "classfile/javaClasses.hpp"
31 #include "code/aotCodeCache.hpp"
32 #include "code/exceptionHandlerTable.hpp"
33 #include "code/nmethod.hpp"
34 #include "compiler/compilationFailureInfo.hpp"
35 #include "compiler/compilationMemoryStatistic.hpp"
36 #include "compiler/compileBroker.hpp"
37 #include "compiler/compileLog.hpp"
38 #include "compiler/compiler_globals.hpp"
39 #include "compiler/compilerDefinitions.hpp"
40 #include "compiler/compilerOracle.hpp"
41 #include "compiler/disassembler.hpp"
42 #include "compiler/oopMap.hpp"
43 #include "gc/shared/barrierSet.hpp"
44 #include "gc/shared/c2/barrierSetC2.hpp"
45 #include "jfr/jfrEvents.hpp"
46 #include "jvm_io.h"
47 #include "memory/allocation.hpp"
48 #include "memory/arena.hpp"
49 #include "memory/resourceArea.hpp"
50 #include "opto/addnode.hpp"
51 #include "opto/block.hpp"
52 #include "opto/c2compiler.hpp"
53 #include "opto/callGenerator.hpp"
54 #include "opto/callnode.hpp"
55 #include "opto/castnode.hpp"
56 #include "opto/cfgnode.hpp"
57 #include "opto/chaitin.hpp"
58 #include "opto/compile.hpp"
59 #include "opto/connode.hpp"
60 #include "opto/convertnode.hpp"
61 #include "opto/divnode.hpp"
62 #include "opto/escape.hpp"
63 #include "opto/idealGraphPrinter.hpp"
64 #include "opto/inlinetypenode.hpp"
65 #include "opto/locknode.hpp"
66 #include "opto/loopnode.hpp"
67 #include "opto/machnode.hpp"
68 #include "opto/macro.hpp"
69 #include "opto/matcher.hpp"
70 #include "opto/mathexactnode.hpp"
71 #include "opto/memnode.hpp"
72 #include "opto/movenode.hpp"
73 #include "opto/mulnode.hpp"
74 #include "opto/multnode.hpp"
75 #include "opto/narrowptrnode.hpp"
76 #include "opto/node.hpp"
77 #include "opto/opaquenode.hpp"
78 #include "opto/opcodes.hpp"
79 #include "opto/output.hpp"
80 #include "opto/parse.hpp"
81 #include "opto/phaseX.hpp"
82 #include "opto/rootnode.hpp"
83 #include "opto/runtime.hpp"
84 #include "opto/stringopts.hpp"
85 #include "opto/type.hpp"
86 #include "opto/vector.hpp"
87 #include "opto/vectornode.hpp"
88 #include "runtime/arguments.hpp"
89 #include "runtime/globals_extension.hpp"
90 #include "runtime/sharedRuntime.hpp"
91 #include "runtime/signature.hpp"
92 #include "runtime/stubRoutines.hpp"
93 #include "runtime/timer.hpp"
94 #include "utilities/align.hpp"
95 #include "utilities/copy.hpp"
96 #include "utilities/hashTable.hpp"
97 #include "utilities/macros.hpp"
98
99 // -------------------- Compile::mach_constant_base_node -----------------------
100 // Constant table base node singleton.
101 MachConstantBaseNode* Compile::mach_constant_base_node() {
102 if (_mach_constant_base_node == nullptr) {
103 _mach_constant_base_node = new MachConstantBaseNode();
104 _mach_constant_base_node->add_req(C->root());
105 }
106 return _mach_constant_base_node;
107 }
108
109
110 /// Support for intrinsics.
111
112 // Return the index at which m must be inserted (or already exists).
113 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
114 class IntrinsicDescPair {
115 private:
116 ciMethod* _m;
117 bool _is_virtual;
118 public:
119 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
120 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
121 ciMethod* m= elt->method();
122 ciMethod* key_m = key->_m;
123 if (key_m < m) return -1;
124 else if (key_m > m) return 1;
125 else {
126 bool is_virtual = elt->is_virtual();
127 bool key_virtual = key->_is_virtual;
128 if (key_virtual < is_virtual) return -1;
129 else if (key_virtual > is_virtual) return 1;
130 else return 0;
131 }
132 }
133 };
134 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
135 #ifdef ASSERT
136 for (int i = 1; i < _intrinsics.length(); i++) {
137 CallGenerator* cg1 = _intrinsics.at(i-1);
138 CallGenerator* cg2 = _intrinsics.at(i);
139 assert(cg1->method() != cg2->method()
140 ? cg1->method() < cg2->method()
141 : cg1->is_virtual() < cg2->is_virtual(),
142 "compiler intrinsics list must stay sorted");
143 }
144 #endif
145 IntrinsicDescPair pair(m, is_virtual);
146 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
147 }
148
149 void Compile::register_intrinsic(CallGenerator* cg) {
150 bool found = false;
151 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
152 assert(!found, "registering twice");
153 _intrinsics.insert_before(index, cg);
154 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
155 }
156
157 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
158 assert(m->is_loaded(), "don't try this on unloaded methods");
159 if (_intrinsics.length() > 0) {
160 bool found = false;
161 int index = intrinsic_insertion_index(m, is_virtual, found);
162 if (found) {
163 return _intrinsics.at(index);
164 }
165 }
166 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
167 if (m->intrinsic_id() != vmIntrinsics::_none &&
168 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
169 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
170 if (cg != nullptr) {
171 // Save it for next time:
172 register_intrinsic(cg);
173 return cg;
174 } else {
175 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
176 }
177 }
178 return nullptr;
179 }
180
181 // Compile::make_vm_intrinsic is defined in library_call.cpp.
182
183 #ifndef PRODUCT
184 // statistics gathering...
185
186 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
187 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
188
189 inline int as_int(vmIntrinsics::ID id) {
190 return vmIntrinsics::as_int(id);
191 }
192
193 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
194 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
195 int oflags = _intrinsic_hist_flags[as_int(id)];
196 assert(flags != 0, "what happened?");
197 if (is_virtual) {
198 flags |= _intrinsic_virtual;
199 }
200 bool changed = (flags != oflags);
201 if ((flags & _intrinsic_worked) != 0) {
202 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
203 if (count == 1) {
204 changed = true; // first time
205 }
206 // increment the overall count also:
207 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
208 }
209 if (changed) {
210 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
211 // Something changed about the intrinsic's virtuality.
212 if ((flags & _intrinsic_virtual) != 0) {
213 // This is the first use of this intrinsic as a virtual call.
214 if (oflags != 0) {
215 // We already saw it as a non-virtual, so note both cases.
216 flags |= _intrinsic_both;
217 }
218 } else if ((oflags & _intrinsic_both) == 0) {
219 // This is the first use of this intrinsic as a non-virtual
220 flags |= _intrinsic_both;
221 }
222 }
223 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
224 }
225 // update the overall flags also:
226 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
227 return changed;
228 }
229
230 static char* format_flags(int flags, char* buf) {
231 buf[0] = 0;
232 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
233 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
234 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
235 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
236 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
237 if (buf[0] == 0) strcat(buf, ",");
238 assert(buf[0] == ',', "must be");
239 return &buf[1];
240 }
241
242 void Compile::print_intrinsic_statistics() {
243 char flagsbuf[100];
244 ttyLocker ttyl;
245 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
246 tty->print_cr("Compiler intrinsic usage:");
247 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
248 if (total == 0) total = 1; // avoid div0 in case of no successes
249 #define PRINT_STAT_LINE(name, c, f) \
250 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
251 for (auto id : EnumRange<vmIntrinsicID>{}) {
252 int flags = _intrinsic_hist_flags[as_int(id)];
253 juint count = _intrinsic_hist_count[as_int(id)];
254 if ((flags | count) != 0) {
255 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
256 }
257 }
258 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
259 if (xtty != nullptr) xtty->tail("statistics");
260 }
261
262 void Compile::print_statistics() {
263 { ttyLocker ttyl;
264 if (xtty != nullptr) xtty->head("statistics type='opto'");
265 Parse::print_statistics();
266 PhaseStringOpts::print_statistics();
267 PhaseCCP::print_statistics();
268 PhaseRegAlloc::print_statistics();
269 PhaseOutput::print_statistics();
270 PhasePeephole::print_statistics();
271 PhaseIdealLoop::print_statistics();
272 ConnectionGraph::print_statistics();
273 PhaseMacroExpand::print_statistics();
274 if (xtty != nullptr) xtty->tail("statistics");
275 }
276 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
277 // put this under its own <statistics> element.
278 print_intrinsic_statistics();
279 }
280 }
281 #endif //PRODUCT
282
283 void Compile::gvn_replace_by(Node* n, Node* nn) {
284 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
285 Node* use = n->last_out(i);
286 bool is_in_table = initial_gvn()->hash_delete(use);
287 uint uses_found = 0;
288 for (uint j = 0; j < use->len(); j++) {
289 if (use->in(j) == n) {
290 if (j < use->req())
291 use->set_req(j, nn);
292 else
293 use->set_prec(j, nn);
294 uses_found++;
295 }
296 }
297 if (is_in_table) {
298 // reinsert into table
299 initial_gvn()->hash_find_insert(use);
300 }
301 record_for_igvn(use);
302 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
303 i -= uses_found; // we deleted 1 or more copies of this edge
304 }
305 }
306
307
308 // Identify all nodes that are reachable from below, useful.
309 // Use breadth-first pass that records state in a Unique_Node_List,
310 // recursive traversal is slower.
311 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
312 int estimated_worklist_size = live_nodes();
313 useful.map( estimated_worklist_size, nullptr ); // preallocate space
314
315 // Initialize worklist
316 if (root() != nullptr) { useful.push(root()); }
317 // If 'top' is cached, declare it useful to preserve cached node
318 if (cached_top_node()) { useful.push(cached_top_node()); }
319
320 // Push all useful nodes onto the list, breadthfirst
321 for( uint next = 0; next < useful.size(); ++next ) {
322 assert( next < unique(), "Unique useful nodes < total nodes");
323 Node *n = useful.at(next);
324 uint max = n->len();
325 for( uint i = 0; i < max; ++i ) {
326 Node *m = n->in(i);
327 if (not_a_node(m)) continue;
328 useful.push(m);
329 }
330 }
331 }
332
333 // Update dead_node_list with any missing dead nodes using useful
334 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
335 void Compile::update_dead_node_list(Unique_Node_List &useful) {
336 uint max_idx = unique();
337 VectorSet& useful_node_set = useful.member_set();
338
339 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
340 // If node with index node_idx is not in useful set,
341 // mark it as dead in dead node list.
342 if (!useful_node_set.test(node_idx)) {
343 record_dead_node(node_idx);
344 }
345 }
346 }
347
348 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
349 int shift = 0;
350 for (int i = 0; i < inlines->length(); i++) {
351 CallGenerator* cg = inlines->at(i);
352 if (useful.member(cg->call_node())) {
353 if (shift > 0) {
354 inlines->at_put(i - shift, cg);
355 }
356 } else {
357 shift++; // skip over the dead element
358 }
359 }
360 if (shift > 0) {
361 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
362 }
363 }
364
365 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
366 assert(dead != nullptr && dead->is_Call(), "sanity");
367 int found = 0;
368 for (int i = 0; i < inlines->length(); i++) {
369 if (inlines->at(i)->call_node() == dead) {
370 inlines->remove_at(i);
371 found++;
372 NOT_DEBUG( break; ) // elements are unique, so exit early
373 }
374 }
375 assert(found <= 1, "not unique");
376 }
377
378 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
379 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
380 for (int i = node_list.length() - 1; i >= 0; i--) {
381 N* node = node_list.at(i);
382 if (!useful.member(node)) {
383 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
384 }
385 }
386 }
387
388 void Compile::remove_useless_node(Node* dead) {
389 remove_modified_node(dead);
390
391 // Constant node that has no out-edges and has only one in-edge from
392 // root is usually dead. However, sometimes reshaping walk makes
393 // it reachable by adding use edges. So, we will NOT count Con nodes
394 // as dead to be conservative about the dead node count at any
395 // given time.
396 if (!dead->is_Con()) {
397 record_dead_node(dead->_idx);
398 }
399 if (dead->is_macro()) {
400 remove_macro_node(dead);
401 }
402 if (dead->is_expensive()) {
403 remove_expensive_node(dead);
404 }
405 if (dead->is_OpaqueTemplateAssertionPredicate()) {
406 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
407 }
408 if (dead->is_ParsePredicate()) {
409 remove_parse_predicate(dead->as_ParsePredicate());
410 }
411 if (dead->for_post_loop_opts_igvn()) {
412 remove_from_post_loop_opts_igvn(dead);
413 }
414 if (dead->is_InlineType()) {
415 remove_inline_type(dead);
416 }
417 if (dead->is_LoadFlat() || dead->is_StoreFlat()) {
418 remove_flat_access(dead);
419 }
420 if (dead->for_merge_stores_igvn()) {
421 remove_from_merge_stores_igvn(dead);
422 }
423 if (dead->is_Call()) {
424 remove_useless_late_inlines( &_late_inlines, dead);
425 remove_useless_late_inlines( &_string_late_inlines, dead);
426 remove_useless_late_inlines( &_boxing_late_inlines, dead);
427 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
428
429 if (dead->is_CallStaticJava()) {
430 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
431 }
432 }
433 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
434 bs->unregister_potential_barrier_node(dead);
435 }
436
437 // Disconnect all useless nodes by disconnecting those at the boundary.
438 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
439 uint next = 0;
440 while (next < useful.size()) {
441 Node *n = useful.at(next++);
442 if (n->is_SafePoint()) {
443 // We're done with a parsing phase. Replaced nodes are not valid
444 // beyond that point.
445 n->as_SafePoint()->delete_replaced_nodes();
446 }
447 // Use raw traversal of out edges since this code removes out edges
448 int max = n->outcnt();
449 for (int j = 0; j < max; ++j) {
450 Node* child = n->raw_out(j);
451 if (!useful.member(child)) {
452 assert(!child->is_top() || child != top(),
453 "If top is cached in Compile object it is in useful list");
454 // Only need to remove this out-edge to the useless node
455 n->raw_del_out(j);
456 --j;
457 --max;
458 if (child->is_data_proj_of_pure_function(n)) {
459 worklist.push(n);
460 }
461 }
462 }
463 if (n->outcnt() == 1 && n->has_special_unique_user()) {
464 assert(useful.member(n->unique_out()), "do not push a useless node");
465 worklist.push(n->unique_out());
466 }
467 if (n->outcnt() == 0) {
468 worklist.push(n);
469 }
470 }
471
472 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
473 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
474 // Remove useless Template Assertion Predicate opaque nodes
475 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
476 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
477 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
478 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
479 remove_useless_nodes(_flat_access_nodes, useful); // remove useless flat access nodes
480 #ifdef ASSERT
481 if (_modified_nodes != nullptr) {
482 _modified_nodes->remove_useless_nodes(useful.member_set());
483 }
484 #endif
485 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
486 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
487 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
488 #ifdef ASSERT
489 if (_modified_nodes != nullptr) {
490 _modified_nodes->remove_useless_nodes(useful.member_set());
491 }
492 #endif
493
494 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
495 bs->eliminate_useless_gc_barriers(useful, this);
496 // clean up the late inline lists
497 remove_useless_late_inlines( &_late_inlines, useful);
498 remove_useless_late_inlines( &_string_late_inlines, useful);
499 remove_useless_late_inlines( &_boxing_late_inlines, useful);
500 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
501 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
502 }
503
504 // ============================================================================
505 //------------------------------CompileWrapper---------------------------------
506 class CompileWrapper : public StackObj {
507 Compile *const _compile;
508 public:
509 CompileWrapper(Compile* compile);
510
511 ~CompileWrapper();
512 };
513
514 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
515 // the Compile* pointer is stored in the current ciEnv:
516 ciEnv* env = compile->env();
517 assert(env == ciEnv::current(), "must already be a ciEnv active");
518 assert(env->compiler_data() == nullptr, "compile already active?");
519 env->set_compiler_data(compile);
520 assert(compile == Compile::current(), "sanity");
521
522 compile->set_type_dict(nullptr);
523 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
524 compile->clone_map().set_clone_idx(0);
525 compile->set_type_last_size(0);
526 compile->set_last_tf(nullptr, nullptr);
527 compile->set_indexSet_arena(nullptr);
528 compile->set_indexSet_free_block_list(nullptr);
529 compile->init_type_arena();
530 Type::Initialize(compile);
531 _compile->begin_method();
532 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
533 }
534 CompileWrapper::~CompileWrapper() {
535 // simulate crash during compilation
536 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
537
538 _compile->end_method();
539 _compile->env()->set_compiler_data(nullptr);
540 }
541
542
543 //----------------------------print_compile_messages---------------------------
544 void Compile::print_compile_messages() {
545 #ifndef PRODUCT
546 // Check if recompiling
547 if (!subsume_loads() && PrintOpto) {
548 // Recompiling without allowing machine instructions to subsume loads
549 tty->print_cr("*********************************************************");
550 tty->print_cr("** Bailout: Recompile without subsuming loads **");
551 tty->print_cr("*********************************************************");
552 }
553 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
554 // Recompiling without escape analysis
555 tty->print_cr("*********************************************************");
556 tty->print_cr("** Bailout: Recompile without escape analysis **");
557 tty->print_cr("*********************************************************");
558 }
559 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
560 // Recompiling without iterative escape analysis
561 tty->print_cr("*********************************************************");
562 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
563 tty->print_cr("*********************************************************");
564 }
565 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
566 // Recompiling without reducing allocation merges
567 tty->print_cr("*********************************************************");
568 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
569 tty->print_cr("*********************************************************");
570 }
571 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
572 // Recompiling without boxing elimination
573 tty->print_cr("*********************************************************");
574 tty->print_cr("** Bailout: Recompile without boxing elimination **");
575 tty->print_cr("*********************************************************");
576 }
577 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
578 // Recompiling without locks coarsening
579 tty->print_cr("*********************************************************");
580 tty->print_cr("** Bailout: Recompile without locks coarsening **");
581 tty->print_cr("*********************************************************");
582 }
583 if (env()->break_at_compile()) {
584 // Open the debugger when compiling this method.
585 tty->print("### Breaking when compiling: ");
586 method()->print_short_name();
587 tty->cr();
588 BREAKPOINT;
589 }
590
591 if( PrintOpto ) {
592 if (is_osr_compilation()) {
593 tty->print("[OSR]%3d", _compile_id);
594 } else {
595 tty->print("%3d", _compile_id);
596 }
597 }
598 #endif
599 }
600
601 #ifndef PRODUCT
602 void Compile::print_phase(const char* phase_name) {
603 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name);
604 }
605
606 void Compile::print_ideal_ir(const char* phase_name) {
607 // keep the following output all in one block
608 // This output goes directly to the tty, not the compiler log.
609 // To enable tools to match it up with the compilation activity,
610 // be sure to tag this tty output with the compile ID.
611
612 // Node dumping can cause a safepoint, which can break the tty lock.
613 // Buffer all node dumps, so that all safepoints happen before we lock.
614 ResourceMark rm;
615 stringStream ss;
616
617 if (_output == nullptr) {
618 ss.print_cr("AFTER: %s", phase_name);
619 // Print out all nodes in ascending order of index.
620 // It is important that we traverse both inputs and outputs of nodes,
621 // so that we reach all nodes that are connected to Root.
622 root()->dump_bfs(MaxNodeLimit, nullptr, "-+S$", &ss);
623 } else {
624 // Dump the node blockwise if we have a scheduling
625 _output->print_scheduling(&ss);
626 }
627
628 // Check that the lock is not broken by a safepoint.
629 NoSafepointVerifier nsv;
630 ttyLocker ttyl;
631 if (xtty != nullptr) {
632 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
633 compile_id(),
634 is_osr_compilation() ? " compile_kind='osr'" : "",
635 phase_name);
636 }
637
638 tty->print("%s", ss.as_string());
639
640 if (xtty != nullptr) {
641 xtty->tail("ideal");
642 }
643 }
644 #endif
645
646 // ============================================================================
647 //------------------------------Compile standard-------------------------------
648
649 // Compile a method. entry_bci is -1 for normal compilations and indicates
650 // the continuation bci for on stack replacement.
651
652
653 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
654 Options options, DirectiveSet* directive)
655 : Phase(Compiler),
656 _compile_id(ci_env->compile_id()),
657 _options(options),
658 _method(target),
659 _entry_bci(osr_bci),
660 _ilt(nullptr),
661 _stub_function(nullptr),
662 _stub_name(nullptr),
663 _stub_id(StubId::NO_STUBID),
664 _stub_entry_point(nullptr),
665 _max_node_limit(MaxNodeLimit),
666 _post_loop_opts_phase(false),
667 _merge_stores_phase(false),
668 _allow_macro_nodes(true),
669 _inlining_progress(false),
670 _inlining_incrementally(false),
671 _strength_reduction(false),
672 _do_cleanup(false),
673 _has_reserved_stack_access(target->has_reserved_stack_access()),
674 _has_circular_inline_type(false),
675 #ifndef PRODUCT
676 _igv_idx(0),
677 _trace_opto_output(directive->TraceOptoOutputOption),
678 #endif
679 _clinit_barrier_on_entry(false),
680 _stress_seed(0),
681 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
682 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
683 _env(ci_env),
684 _directive(directive),
685 _log(ci_env->log()),
686 _first_failure_details(nullptr),
687 _intrinsics(comp_arena(), 0, 0, nullptr),
688 _macro_nodes(comp_arena(), 8, 0, nullptr),
689 _parse_predicates(comp_arena(), 8, 0, nullptr),
690 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
691 _expensive_nodes(comp_arena(), 8, 0, nullptr),
692 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
693 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
694 _flat_access_nodes(comp_arena(), 8, 0, nullptr),
695 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
696 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
697 _coarsened_locks(comp_arena(), 8, 0, nullptr),
698 _congraph(nullptr),
699 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
700 _unique(0),
701 _dead_node_count(0),
702 _dead_node_list(comp_arena()),
703 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
704 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
705 _node_arena(&_node_arena_one),
706 _mach_constant_base_node(nullptr),
707 _Compile_types(mtCompiler, Arena::Tag::tag_type),
708 _initial_gvn(nullptr),
709 _igvn_worklist(nullptr),
710 _types(nullptr),
711 _node_hash(nullptr),
712 _late_inlines(comp_arena(), 2, 0, nullptr),
713 _string_late_inlines(comp_arena(), 2, 0, nullptr),
714 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
715 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
716 _late_inlines_pos(0),
717 _has_mh_late_inlines(false),
718 _oom(false),
719 _replay_inline_data(nullptr),
720 _inline_printer(this),
721 _java_calls(0),
722 _inner_loops(0),
723 _FIRST_STACK_mask(comp_arena()),
724 _interpreter_frame_size(0),
725 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
726 _output(nullptr)
727 #ifndef PRODUCT
728 ,
729 _in_dump_cnt(0)
730 #endif
731 {
732 C = this;
733 CompileWrapper cw(this);
734
735 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
736 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
737
738 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
739 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
740 // We can always print a disassembly, either abstract (hex dump) or
741 // with the help of a suitable hsdis library. Thus, we should not
742 // couple print_assembly and print_opto_assembly controls.
743 // But: always print opto and regular assembly on compile command 'print'.
744 bool print_assembly = directive->PrintAssemblyOption;
745 set_print_assembly(print_opto_assembly || print_assembly);
746 #else
747 set_print_assembly(false); // must initialize.
748 #endif
749
750 #ifndef PRODUCT
751 set_parsed_irreducible_loop(false);
752 #endif
753
754 if (directive->ReplayInlineOption) {
755 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
756 }
757 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
758 set_print_intrinsics(directive->PrintIntrinsicsOption);
759 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
760
761 if (ProfileTraps) {
762 // Make sure the method being compiled gets its own MDO,
763 // so we can at least track the decompile_count().
764 method()->ensure_method_data();
765 }
766
767 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
768 StressIncrementalInlining || StressMacroExpansion ||
769 StressMacroElimination || StressUnstableIfTraps ||
770 StressBailout || StressLoopPeeling) {
771 initialize_stress_seed(directive);
772 }
773
774 Init(/*do_aliasing=*/ true);
775
776 print_compile_messages();
777
778 _ilt = InlineTree::build_inline_tree_root();
779
780 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
781 assert(num_alias_types() >= AliasIdxRaw, "");
782
783 #define MINIMUM_NODE_HASH 1023
784
785 // GVN that will be run immediately on new nodes
786 uint estimated_size = method()->code_size()*4+64;
787 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
788 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
789 _types = new (comp_arena()) Type_Array(comp_arena());
790 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
791 PhaseGVN gvn;
792 set_initial_gvn(&gvn);
793
794 { // Scope for timing the parser
795 TracePhase tp(_t_parser);
796
797 // Put top into the hash table ASAP.
798 initial_gvn()->transform(top());
799
800 // Set up tf(), start(), and find a CallGenerator.
801 CallGenerator* cg = nullptr;
802 if (is_osr_compilation()) {
803 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
804 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
805 initial_gvn()->set_type_bottom(s);
806 verify_start(s);
807 cg = CallGenerator::for_osr(method(), entry_bci());
808 } else {
809 // Normal case.
810 init_tf(TypeFunc::make(method()));
811 StartNode* s = new StartNode(root(), tf()->domain_cc());
812 initial_gvn()->set_type_bottom(s);
813 verify_start(s);
814 float past_uses = method()->interpreter_invocation_count();
815 float expected_uses = past_uses;
816 cg = CallGenerator::for_inline(method(), expected_uses);
817 }
818 if (failing()) return;
819 if (cg == nullptr) {
820 const char* reason = InlineTree::check_can_parse(method());
821 assert(reason != nullptr, "expect reason for parse failure");
822 stringStream ss;
823 ss.print("cannot parse method: %s", reason);
824 record_method_not_compilable(ss.as_string());
825 return;
826 }
827
828 gvn.set_type(root(), root()->bottom_type());
829
830 JVMState* jvms = build_start_state(start(), tf());
831 if ((jvms = cg->generate(jvms)) == nullptr) {
832 assert(failure_reason() != nullptr, "expect reason for parse failure");
833 stringStream ss;
834 ss.print("method parse failed: %s", failure_reason());
835 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
836 return;
837 }
838 GraphKit kit(jvms);
839
840 if (!kit.stopped()) {
841 // Accept return values, and transfer control we know not where.
842 // This is done by a special, unique ReturnNode bound to root.
843 return_values(kit.jvms());
844 }
845
846 if (kit.has_exceptions()) {
847 // Any exceptions that escape from this call must be rethrown
848 // to whatever caller is dynamically above us on the stack.
849 // This is done by a special, unique RethrowNode bound to root.
850 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
851 }
852
853 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
854
855 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
856 inline_string_calls(true);
857 }
858
859 if (failing()) return;
860
861 // Remove clutter produced by parsing.
862 if (!failing()) {
863 ResourceMark rm;
864 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
865 }
866 }
867
868 // Note: Large methods are capped off in do_one_bytecode().
869 if (failing()) return;
870
871 // After parsing, node notes are no longer automagic.
872 // They must be propagated by register_new_node_with_optimizer(),
873 // clone(), or the like.
874 set_default_node_notes(nullptr);
875
876 #ifndef PRODUCT
877 if (should_print_igv(1)) {
878 _igv_printer->print_inlining();
879 }
880 #endif
881
882 if (failing()) return;
883 NOT_PRODUCT( verify_graph_edges(); )
884
885 // Now optimize
886 Optimize();
887 if (failing()) return;
888 NOT_PRODUCT( verify_graph_edges(); )
889
890 #ifndef PRODUCT
891 if (should_print_ideal()) {
892 print_ideal_ir("print_ideal");
893 }
894 #endif
895
896 #ifdef ASSERT
897 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
898 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
899 #endif
900
901 // Dump compilation data to replay it.
902 if (directive->DumpReplayOption) {
903 env()->dump_replay_data(_compile_id);
904 }
905 if (directive->DumpInlineOption && (ilt() != nullptr)) {
906 env()->dump_inline_data(_compile_id);
907 }
908
909 // Now that we know the size of all the monitors we can add a fixed slot
910 // for the original deopt pc.
911 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
912 if (needs_stack_repair()) {
913 // One extra slot for the special stack increment value
914 next_slot += 2;
915 }
916 // TODO 8284443 Only reserve extra slot if needed
917 if (InlineTypeReturnedAsFields) {
918 // One extra slot to hold the null marker for a nullable
919 // inline type return if we run out of registers.
920 next_slot += 2;
921 }
922 set_fixed_slots(next_slot);
923
924 // Compute when to use implicit null checks. Used by matching trap based
925 // nodes and NullCheck optimization.
926 set_allowed_deopt_reasons();
927
928 // Now generate code
929 Code_Gen();
930 }
931
932 //------------------------------Compile----------------------------------------
933 // Compile a runtime stub
934 Compile::Compile(ciEnv* ci_env,
935 TypeFunc_generator generator,
936 address stub_function,
937 const char* stub_name,
938 StubId stub_id,
939 int is_fancy_jump,
940 bool pass_tls,
941 bool return_pc,
942 DirectiveSet* directive)
943 : Phase(Compiler),
944 _compile_id(0),
945 _options(Options::for_runtime_stub()),
946 _method(nullptr),
947 _entry_bci(InvocationEntryBci),
948 _stub_function(stub_function),
949 _stub_name(stub_name),
950 _stub_id(stub_id),
951 _stub_entry_point(nullptr),
952 _max_node_limit(MaxNodeLimit),
953 _post_loop_opts_phase(false),
954 _merge_stores_phase(false),
955 _allow_macro_nodes(true),
956 _inlining_progress(false),
957 _inlining_incrementally(false),
958 _has_reserved_stack_access(false),
959 _has_circular_inline_type(false),
960 #ifndef PRODUCT
961 _igv_idx(0),
962 _trace_opto_output(directive->TraceOptoOutputOption),
963 #endif
964 _clinit_barrier_on_entry(false),
965 _stress_seed(0),
966 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
967 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
968 _env(ci_env),
969 _directive(directive),
970 _log(ci_env->log()),
971 _first_failure_details(nullptr),
972 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
973 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
974 _congraph(nullptr),
975 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
976 _unique(0),
977 _dead_node_count(0),
978 _dead_node_list(comp_arena()),
979 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
980 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
981 _node_arena(&_node_arena_one),
982 _mach_constant_base_node(nullptr),
983 _Compile_types(mtCompiler, Arena::Tag::tag_type),
984 _initial_gvn(nullptr),
985 _igvn_worklist(nullptr),
986 _types(nullptr),
987 _node_hash(nullptr),
988 _has_mh_late_inlines(false),
989 _oom(false),
990 _replay_inline_data(nullptr),
991 _inline_printer(this),
992 _java_calls(0),
993 _inner_loops(0),
994 _FIRST_STACK_mask(comp_arena()),
995 _interpreter_frame_size(0),
996 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
997 _output(nullptr),
998 #ifndef PRODUCT
999 _in_dump_cnt(0),
1000 #endif
1001 _allowed_reasons(0) {
1002 C = this;
1003
1004 // try to reuse an existing stub
1005 {
1006 BlobId blob_id = StubInfo::blob(_stub_id);
1007 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
1008 if (blob != nullptr) {
1009 RuntimeStub* rs = blob->as_runtime_stub();
1010 _stub_entry_point = rs->entry_point();
1011 return;
1012 }
1013 }
1014
1015 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
1016 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
1017
1018 #ifndef PRODUCT
1019 set_print_assembly(PrintFrameConverterAssembly);
1020 set_parsed_irreducible_loop(false);
1021 #else
1022 set_print_assembly(false); // Must initialize.
1023 #endif
1024 set_has_irreducible_loop(false); // no loops
1025
1026 CompileWrapper cw(this);
1027 Init(/*do_aliasing=*/ false);
1028 init_tf((*generator)());
1029
1030 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1031 _types = new (comp_arena()) Type_Array(comp_arena());
1032 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1033
1034 if (StressLCM || StressGCM || StressBailout) {
1035 initialize_stress_seed(directive);
1036 }
1037
1038 {
1039 PhaseGVN gvn;
1040 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1041 gvn.transform(top());
1042
1043 GraphKit kit;
1044 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1045 }
1046
1047 NOT_PRODUCT( verify_graph_edges(); )
1048
1049 Code_Gen();
1050 }
1051
1052 Compile::~Compile() {
1053 delete _first_failure_details;
1054 };
1055
1056 //------------------------------Init-------------------------------------------
1057 // Prepare for a single compilation
1058 void Compile::Init(bool aliasing) {
1059 _do_aliasing = aliasing;
1060 _unique = 0;
1061 _regalloc = nullptr;
1062
1063 _tf = nullptr; // filled in later
1064 _top = nullptr; // cached later
1065 _matcher = nullptr; // filled in later
1066 _cfg = nullptr; // filled in later
1067
1068 _node_note_array = nullptr;
1069 _default_node_notes = nullptr;
1070 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1071
1072 _immutable_memory = nullptr; // filled in at first inquiry
1073
1074 #ifdef ASSERT
1075 _phase_optimize_finished = false;
1076 _phase_verify_ideal_loop = false;
1077 _exception_backedge = false;
1078 _type_verify = nullptr;
1079 #endif
1080
1081 // Globally visible Nodes
1082 // First set TOP to null to give safe behavior during creation of RootNode
1083 set_cached_top_node(nullptr);
1084 set_root(new RootNode());
1085 // Now that you have a Root to point to, create the real TOP
1086 set_cached_top_node( new ConNode(Type::TOP) );
1087 set_recent_alloc(nullptr, nullptr);
1088
1089 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1090 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1091 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1092 env()->set_dependencies(new Dependencies(env()));
1093
1094 _fixed_slots = 0;
1095 set_has_split_ifs(false);
1096 set_has_loops(false); // first approximation
1097 set_has_stringbuilder(false);
1098 set_has_boxed_value(false);
1099 _trap_can_recompile = false; // no traps emitted yet
1100 _major_progress = true; // start out assuming good things will happen
1101 set_has_unsafe_access(false);
1102 set_max_vector_size(0);
1103 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1104 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1105 set_decompile_count(0);
1106
1107 #ifndef PRODUCT
1108 _phase_counter = 0;
1109 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1110 #endif
1111
1112 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1113 _loop_opts_cnt = LoopOptsCount;
1114 _has_flat_accesses = false;
1115 _flat_accesses_share_alias = true;
1116 _scalarize_in_safepoints = false;
1117
1118 set_do_inlining(Inline);
1119 set_max_inline_size(MaxInlineSize);
1120 set_freq_inline_size(FreqInlineSize);
1121 set_do_scheduling(OptoScheduling);
1122
1123 set_do_vector_loop(false);
1124 set_has_monitors(false);
1125 set_has_scoped_access(false);
1126
1127 if (AllowVectorizeOnDemand) {
1128 if (has_method() && _directive->VectorizeOption) {
1129 set_do_vector_loop(true);
1130 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());})
1131 } else if (has_method() && method()->name() != nullptr &&
1132 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1133 set_do_vector_loop(true);
1134 }
1135 }
1136 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1137 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());})
1138
1139 _max_node_limit = _directive->MaxNodeLimitOption;
1140
1141 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1142 set_clinit_barrier_on_entry(true);
1143 }
1144 if (debug_info()->recording_non_safepoints()) {
1145 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1146 (comp_arena(), 8, 0, nullptr));
1147 set_default_node_notes(Node_Notes::make(this));
1148 }
1149
1150 const int grow_ats = 16;
1151 _max_alias_types = grow_ats;
1152 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1153 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1154 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1155 {
1156 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1157 }
1158 // Initialize the first few types.
1159 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1160 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1161 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1162 _num_alias_types = AliasIdxRaw+1;
1163 // Zero out the alias type cache.
1164 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1165 // A null adr_type hits in the cache right away. Preload the right answer.
1166 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1167 }
1168
1169 #ifdef ASSERT
1170 // Verify that the current StartNode is valid.
1171 void Compile::verify_start(StartNode* s) const {
1172 assert(failing_internal() || s == start(), "should be StartNode");
1173 }
1174 #endif
1175
1176 /**
1177 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1178 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1179 * the ideal graph.
1180 */
1181 StartNode* Compile::start() const {
1182 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1183 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1184 Node* start = root()->fast_out(i);
1185 if (start->is_Start()) {
1186 return start->as_Start();
1187 }
1188 }
1189 fatal("Did not find Start node!");
1190 return nullptr;
1191 }
1192
1193 //-------------------------------immutable_memory-------------------------------------
1194 // Access immutable memory
1195 Node* Compile::immutable_memory() {
1196 if (_immutable_memory != nullptr) {
1197 return _immutable_memory;
1198 }
1199 StartNode* s = start();
1200 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1201 Node *p = s->fast_out(i);
1202 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1203 _immutable_memory = p;
1204 return _immutable_memory;
1205 }
1206 }
1207 ShouldNotReachHere();
1208 return nullptr;
1209 }
1210
1211 //----------------------set_cached_top_node------------------------------------
1212 // Install the cached top node, and make sure Node::is_top works correctly.
1213 void Compile::set_cached_top_node(Node* tn) {
1214 if (tn != nullptr) verify_top(tn);
1215 Node* old_top = _top;
1216 _top = tn;
1217 // Calling Node::setup_is_top allows the nodes the chance to adjust
1218 // their _out arrays.
1219 if (_top != nullptr) _top->setup_is_top();
1220 if (old_top != nullptr) old_top->setup_is_top();
1221 assert(_top == nullptr || top()->is_top(), "");
1222 }
1223
1224 #ifdef ASSERT
1225 uint Compile::count_live_nodes_by_graph_walk() {
1226 Unique_Node_List useful(comp_arena());
1227 // Get useful node list by walking the graph.
1228 identify_useful_nodes(useful);
1229 return useful.size();
1230 }
1231
1232 void Compile::print_missing_nodes() {
1233
1234 // Return if CompileLog is null and PrintIdealNodeCount is false.
1235 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1236 return;
1237 }
1238
1239 // This is an expensive function. It is executed only when the user
1240 // specifies VerifyIdealNodeCount option or otherwise knows the
1241 // additional work that needs to be done to identify reachable nodes
1242 // by walking the flow graph and find the missing ones using
1243 // _dead_node_list.
1244
1245 Unique_Node_List useful(comp_arena());
1246 // Get useful node list by walking the graph.
1247 identify_useful_nodes(useful);
1248
1249 uint l_nodes = C->live_nodes();
1250 uint l_nodes_by_walk = useful.size();
1251
1252 if (l_nodes != l_nodes_by_walk) {
1253 if (_log != nullptr) {
1254 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1255 _log->stamp();
1256 _log->end_head();
1257 }
1258 VectorSet& useful_member_set = useful.member_set();
1259 int last_idx = l_nodes_by_walk;
1260 for (int i = 0; i < last_idx; i++) {
1261 if (useful_member_set.test(i)) {
1262 if (_dead_node_list.test(i)) {
1263 if (_log != nullptr) {
1264 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1265 }
1266 if (PrintIdealNodeCount) {
1267 // Print the log message to tty
1268 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1269 useful.at(i)->dump();
1270 }
1271 }
1272 }
1273 else if (! _dead_node_list.test(i)) {
1274 if (_log != nullptr) {
1275 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1276 }
1277 if (PrintIdealNodeCount) {
1278 // Print the log message to tty
1279 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1280 }
1281 }
1282 }
1283 if (_log != nullptr) {
1284 _log->tail("mismatched_nodes");
1285 }
1286 }
1287 }
1288 void Compile::record_modified_node(Node* n) {
1289 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1290 _modified_nodes->push(n);
1291 }
1292 }
1293
1294 void Compile::remove_modified_node(Node* n) {
1295 if (_modified_nodes != nullptr) {
1296 _modified_nodes->remove(n);
1297 }
1298 }
1299 #endif
1300
1301 #ifndef PRODUCT
1302 void Compile::verify_top(Node* tn) const {
1303 if (tn != nullptr) {
1304 assert(tn->is_Con(), "top node must be a constant");
1305 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1306 assert(tn->in(0) != nullptr, "must have live top node");
1307 }
1308 }
1309 #endif
1310
1311
1312 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1313
1314 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1315 guarantee(arr != nullptr, "");
1316 int num_blocks = arr->length();
1317 if (grow_by < num_blocks) grow_by = num_blocks;
1318 int num_notes = grow_by * _node_notes_block_size;
1319 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1320 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1321 while (num_notes > 0) {
1322 arr->append(notes);
1323 notes += _node_notes_block_size;
1324 num_notes -= _node_notes_block_size;
1325 }
1326 assert(num_notes == 0, "exact multiple, please");
1327 }
1328
1329 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1330 if (source == nullptr || dest == nullptr) return false;
1331
1332 if (dest->is_Con())
1333 return false; // Do not push debug info onto constants.
1334
1335 #ifdef ASSERT
1336 // Leave a bread crumb trail pointing to the original node:
1337 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1338 dest->set_debug_orig(source);
1339 }
1340 #endif
1341
1342 if (node_note_array() == nullptr)
1343 return false; // Not collecting any notes now.
1344
1345 // This is a copy onto a pre-existing node, which may already have notes.
1346 // If both nodes have notes, do not overwrite any pre-existing notes.
1347 Node_Notes* source_notes = node_notes_at(source->_idx);
1348 if (source_notes == nullptr || source_notes->is_clear()) return false;
1349 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1350 if (dest_notes == nullptr || dest_notes->is_clear()) {
1351 return set_node_notes_at(dest->_idx, source_notes);
1352 }
1353
1354 Node_Notes merged_notes = (*source_notes);
1355 // The order of operations here ensures that dest notes will win...
1356 merged_notes.update_from(dest_notes);
1357 return set_node_notes_at(dest->_idx, &merged_notes);
1358 }
1359
1360
1361 //--------------------------allow_range_check_smearing-------------------------
1362 // Gating condition for coalescing similar range checks.
1363 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1364 // single covering check that is at least as strong as any of them.
1365 // If the optimization succeeds, the simplified (strengthened) range check
1366 // will always succeed. If it fails, we will deopt, and then give up
1367 // on the optimization.
1368 bool Compile::allow_range_check_smearing() const {
1369 // If this method has already thrown a range-check,
1370 // assume it was because we already tried range smearing
1371 // and it failed.
1372 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1373 return !already_trapped;
1374 }
1375
1376
1377 //------------------------------flatten_alias_type-----------------------------
1378 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1379 assert(do_aliasing(), "Aliasing should be enabled");
1380 int offset = tj->offset();
1381 TypePtr::PTR ptr = tj->ptr();
1382
1383 // Known instance (scalarizable allocation) alias only with itself.
1384 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1385 tj->is_oopptr()->is_known_instance();
1386
1387 // Process weird unsafe references.
1388 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1389 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1390 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1391 tj = TypeOopPtr::BOTTOM;
1392 ptr = tj->ptr();
1393 offset = tj->offset();
1394 }
1395
1396 // Array pointers need some flattening
1397 const TypeAryPtr* ta = tj->isa_aryptr();
1398 if( ta && is_known_inst ) {
1399 if ( offset != Type::OffsetBot &&
1400 offset > arrayOopDesc::length_offset_in_bytes() ) {
1401 offset = Type::OffsetBot; // Flatten constant access into array body only
1402 tj = ta = ta->
1403 remove_speculative()->
1404 cast_to_ptr_type(ptr)->
1405 with_offset(offset);
1406 }
1407 } else if (ta) {
1408 // Common slices
1409 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1410 return TypeAryPtr::RANGE;
1411 } else if (offset == oopDesc::klass_offset_in_bytes()) {
1412 return TypeInstPtr::KLASS;
1413 } else if (offset == oopDesc::mark_offset_in_bytes()) {
1414 return TypeInstPtr::MARK;
1415 }
1416
1417 // Remove size and stability
1418 const TypeAry* normalized_ary = TypeAry::make(ta->elem(), TypeInt::POS, false, ta->is_flat(), ta->is_not_flat(), ta->is_not_null_free(), ta->is_atomic());
1419 // Remove ptr, const_oop, and offset
1420 if (ta->elem() == Type::BOTTOM) {
1421 // Bottom array (meet of int[] and byte[] for example), accesses to it will be done with
1422 // Unsafe. This should alias with all arrays. For now just leave it as it is (this is
1423 // incorrect, see JDK-8331133).
1424 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, false, Type::Offset::bottom);
1425 } else if (ta->elem()->make_oopptr() != nullptr) {
1426 // Object arrays, keep field_offset
1427 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, ta->klass_is_exact(), Type::Offset::bottom, Type::Offset(ta->field_offset()));
1428 } else {
1429 // Primitive arrays
1430 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, ta->exact_klass(), true, Type::Offset::bottom);
1431 }
1432
1433 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1434 // cannot be distinguished by bytecode alone.
1435 if (ta->elem() == TypeInt::BOOL) {
1436 tj = ta = TypeAryPtr::BYTES;
1437 }
1438
1439 // All arrays of references share the same slice
1440 if (!ta->is_flat() && ta->elem()->make_oopptr() != nullptr) {
1441 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, TypeInt::POS, false, false, true, true, true);
1442 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, tary, nullptr, false, Type::Offset::bottom);
1443 }
1444
1445 if (ta->is_flat()) {
1446 if (_flat_accesses_share_alias) {
1447 // Initially all flattened array accesses share a single slice
1448 tj = ta = TypeAryPtr::INLINES;
1449 } else {
1450 // Flat accesses are always exact
1451 tj = ta = ta->cast_to_exactness(true);
1452 }
1453 }
1454 }
1455
1456 // Oop pointers need some flattening
1457 const TypeInstPtr *to = tj->isa_instptr();
1458 if (to && to != TypeOopPtr::BOTTOM) {
1459 ciInstanceKlass* ik = to->instance_klass();
1460 tj = to = to->cast_to_maybe_flat_in_array(); // flatten to maybe flat in array
1461 if( ptr == TypePtr::Constant ) {
1462 if (ik != ciEnv::current()->Class_klass() ||
1463 offset < ik->layout_helper_size_in_bytes()) {
1464 // No constant oop pointers (such as Strings); they alias with
1465 // unknown strings.
1466 assert(!is_known_inst, "not scalarizable allocation");
1467 tj = to = to->
1468 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1469 remove_speculative()->
1470 cast_to_ptr_type(TypePtr::BotPTR)->
1471 cast_to_exactness(false);
1472 }
1473 } else if( is_known_inst ) {
1474 tj = to; // Keep NotNull and klass_is_exact for instance type
1475 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1476 // During the 2nd round of IterGVN, NotNull castings are removed.
1477 // Make sure the Bottom and NotNull variants alias the same.
1478 // Also, make sure exact and non-exact variants alias the same.
1479 tj = to = to->
1480 remove_speculative()->
1481 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1482 cast_to_ptr_type(TypePtr::BotPTR)->
1483 cast_to_exactness(false);
1484 }
1485 if (to->speculative() != nullptr) {
1486 tj = to = to->remove_speculative();
1487 }
1488 // Canonicalize the holder of this field
1489 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1490 // First handle header references such as a LoadKlassNode, even if the
1491 // object's klass is unloaded at compile time (4965979).
1492 if (!is_known_inst) { // Do it only for non-instance types
1493 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1494 }
1495 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1496 // Static fields are in the space above the normal instance
1497 // fields in the java.lang.Class instance.
1498 if (ik != ciEnv::current()->Class_klass()) {
1499 to = nullptr;
1500 tj = TypeOopPtr::BOTTOM;
1501 offset = tj->offset();
1502 }
1503 } else {
1504 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1505 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1506 assert(tj->offset() == offset, "no change to offset expected");
1507 bool xk = to->klass_is_exact();
1508 int instance_id = to->instance_id();
1509
1510 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1511 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1512 // its interfaces are included.
1513 if (xk && ik->equals(canonical_holder)) {
1514 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id,
1515 TypePtr::MaybeFlat), "exact type should be canonical type");
1516 } else {
1517 assert(xk || !is_known_inst, "Known instance should be exact type");
1518 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id,
1519 TypePtr::MaybeFlat);
1520 }
1521 }
1522 }
1523
1524 // Klass pointers to object array klasses need some flattening
1525 const TypeKlassPtr *tk = tj->isa_klassptr();
1526 if( tk ) {
1527 // If we are referencing a field within a Klass, we need
1528 // to assume the worst case of an Object. Both exact and
1529 // inexact types must flatten to the same alias class so
1530 // use NotNull as the PTR.
1531 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1532 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1533 env()->Object_klass(),
1534 Type::Offset(offset),
1535 TypePtr::MaybeFlat);
1536 }
1537
1538 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1539 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1540 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1541 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset), TypePtr::MaybeFlat);
1542 } else {
1543 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, Type::Offset(offset), tk->is_not_flat(), tk->is_not_null_free(), tk->is_flat(), tk->is_null_free(), tk->is_atomic(), tk->is_aryklassptr()->is_refined_type());
1544 }
1545 }
1546 // Check for precise loads from the primary supertype array and force them
1547 // to the supertype cache alias index. Check for generic array loads from
1548 // the primary supertype array and also force them to the supertype cache
1549 // alias index. Since the same load can reach both, we need to merge
1550 // these 2 disparate memories into the same alias class. Since the
1551 // primary supertype array is read-only, there's no chance of confusion
1552 // where we bypass an array load and an array store.
1553 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1554 if (offset == Type::OffsetBot ||
1555 (offset >= primary_supers_offset &&
1556 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1557 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1558 offset = in_bytes(Klass::secondary_super_cache_offset());
1559 tj = tk = tk->with_offset(offset);
1560 }
1561 }
1562
1563 // Flatten all Raw pointers together.
1564 if (tj->base() == Type::RawPtr)
1565 tj = TypeRawPtr::BOTTOM;
1566
1567 if (tj->base() == Type::AnyPtr)
1568 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1569
1570 offset = tj->offset();
1571 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1572
1573 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1574 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1575 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1576 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1577 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1578 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1579 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1580 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1581 assert( tj->ptr() != TypePtr::TopPTR &&
1582 tj->ptr() != TypePtr::AnyNull &&
1583 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1584 // assert( tj->ptr() != TypePtr::Constant ||
1585 // tj->base() == Type::RawPtr ||
1586 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1587
1588 return tj;
1589 }
1590
1591 void Compile::AliasType::Init(int i, const TypePtr* at) {
1592 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1593 _index = i;
1594 _adr_type = at;
1595 _field = nullptr;
1596 _element = nullptr;
1597 _is_rewritable = true; // default
1598 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1599 if (atoop != nullptr && atoop->is_known_instance()) {
1600 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1601 _general_index = Compile::current()->get_alias_index(gt);
1602 } else {
1603 _general_index = 0;
1604 }
1605 }
1606
1607 BasicType Compile::AliasType::basic_type() const {
1608 if (element() != nullptr) {
1609 const Type* element = adr_type()->is_aryptr()->elem();
1610 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1611 } if (field() != nullptr) {
1612 return field()->layout_type();
1613 } else {
1614 return T_ILLEGAL; // unknown
1615 }
1616 }
1617
1618 //---------------------------------print_on------------------------------------
1619 #ifndef PRODUCT
1620 void Compile::AliasType::print_on(outputStream* st) {
1621 if (index() < 10)
1622 st->print("@ <%d> ", index());
1623 else st->print("@ <%d>", index());
1624 st->print(is_rewritable() ? " " : " RO");
1625 int offset = adr_type()->offset();
1626 if (offset == Type::OffsetBot)
1627 st->print(" +any");
1628 else st->print(" +%-3d", offset);
1629 st->print(" in ");
1630 adr_type()->dump_on(st);
1631 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1632 if (field() != nullptr && tjp) {
1633 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1634 tjp->offset() != field()->offset_in_bytes()) {
1635 st->print(" != ");
1636 field()->print();
1637 st->print(" ***");
1638 }
1639 }
1640 }
1641
1642 void print_alias_types() {
1643 Compile* C = Compile::current();
1644 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1645 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1646 C->alias_type(idx)->print_on(tty);
1647 tty->cr();
1648 }
1649 }
1650 #endif
1651
1652
1653 //----------------------------probe_alias_cache--------------------------------
1654 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1655 intptr_t key = (intptr_t) adr_type;
1656 key ^= key >> logAliasCacheSize;
1657 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1658 }
1659
1660
1661 //-----------------------------grow_alias_types--------------------------------
1662 void Compile::grow_alias_types() {
1663 const int old_ats = _max_alias_types; // how many before?
1664 const int new_ats = old_ats; // how many more?
1665 const int grow_ats = old_ats+new_ats; // how many now?
1666 _max_alias_types = grow_ats;
1667 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1668 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1669 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1670 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1671 }
1672
1673
1674 //--------------------------------find_alias_type------------------------------
1675 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1676 if (!do_aliasing()) {
1677 return alias_type(AliasIdxBot);
1678 }
1679
1680 AliasCacheEntry* ace = nullptr;
1681 if (!uncached) {
1682 ace = probe_alias_cache(adr_type);
1683 if (ace->_adr_type == adr_type) {
1684 return alias_type(ace->_index);
1685 }
1686 }
1687
1688 // Handle special cases.
1689 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1690 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1691
1692 // Do it the slow way.
1693 const TypePtr* flat = flatten_alias_type(adr_type);
1694
1695 #ifdef ASSERT
1696 {
1697 ResourceMark rm;
1698 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1699 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1700 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1701 Type::str(adr_type));
1702 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1703 const TypeOopPtr* foop = flat->is_oopptr();
1704 // Scalarizable allocations have exact klass always.
1705 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1706 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1707 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1708 Type::str(foop), Type::str(xoop));
1709 }
1710 }
1711 #endif
1712
1713 int idx = AliasIdxTop;
1714 for (int i = 0; i < num_alias_types(); i++) {
1715 if (alias_type(i)->adr_type() == flat) {
1716 idx = i;
1717 break;
1718 }
1719 }
1720
1721 if (idx == AliasIdxTop) {
1722 if (no_create) return nullptr;
1723 // Grow the array if necessary.
1724 if (_num_alias_types == _max_alias_types) grow_alias_types();
1725 // Add a new alias type.
1726 idx = _num_alias_types++;
1727 _alias_types[idx]->Init(idx, flat);
1728 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1729 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1730 if (flat->isa_instptr()) {
1731 if (flat->offset() == java_lang_Class::klass_offset()
1732 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1733 alias_type(idx)->set_rewritable(false);
1734 }
1735 ciField* field = nullptr;
1736 if (flat->isa_aryptr()) {
1737 #ifdef ASSERT
1738 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1739 // (T_BYTE has the weakest alignment and size restrictions...)
1740 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1741 #endif
1742 const Type* elemtype = flat->is_aryptr()->elem();
1743 if (flat->offset() == TypePtr::OffsetBot) {
1744 alias_type(idx)->set_element(elemtype);
1745 }
1746 int field_offset = flat->is_aryptr()->field_offset().get();
1747 if (flat->is_flat() &&
1748 field_offset != Type::OffsetBot) {
1749 ciInlineKlass* vk = elemtype->inline_klass();
1750 field_offset += vk->payload_offset();
1751 field = vk->get_field_by_offset(field_offset, false);
1752 }
1753 }
1754 if (flat->isa_klassptr()) {
1755 if (UseCompactObjectHeaders) {
1756 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1757 alias_type(idx)->set_rewritable(false);
1758 }
1759 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1760 alias_type(idx)->set_rewritable(false);
1761 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1762 alias_type(idx)->set_rewritable(false);
1763 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1764 alias_type(idx)->set_rewritable(false);
1765 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1766 alias_type(idx)->set_rewritable(false);
1767 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1768 alias_type(idx)->set_rewritable(false);
1769 }
1770
1771 if (flat->isa_instklassptr()) {
1772 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1773 alias_type(idx)->set_rewritable(false);
1774 }
1775 }
1776 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1777 // but the base pointer type is not distinctive enough to identify
1778 // references into JavaThread.)
1779
1780 // Check for final fields.
1781 const TypeInstPtr* tinst = flat->isa_instptr();
1782 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1783 if (tinst->const_oop() != nullptr &&
1784 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1785 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1786 // static field
1787 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1788 field = k->get_field_by_offset(tinst->offset(), true);
1789 } else if (tinst->is_inlinetypeptr()) {
1790 // Inline type field
1791 ciInlineKlass* vk = tinst->inline_klass();
1792 field = vk->get_field_by_offset(tinst->offset(), false);
1793 } else {
1794 ciInstanceKlass *k = tinst->instance_klass();
1795 field = k->get_field_by_offset(tinst->offset(), false);
1796 }
1797 }
1798 assert(field == nullptr ||
1799 original_field == nullptr ||
1800 (field->holder() == original_field->holder() &&
1801 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1802 field->is_static() == original_field->is_static()), "wrong field?");
1803 // Set field() and is_rewritable() attributes.
1804 if (field != nullptr) {
1805 alias_type(idx)->set_field(field);
1806 if (flat->isa_aryptr()) {
1807 // Fields of flat arrays are rewritable although they are declared final
1808 assert(flat->is_flat(), "must be a flat array");
1809 alias_type(idx)->set_rewritable(true);
1810 }
1811 }
1812 }
1813
1814 // Fill the cache for next time.
1815 if (!uncached) {
1816 ace->_adr_type = adr_type;
1817 ace->_index = idx;
1818 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1819
1820 // Might as well try to fill the cache for the flattened version, too.
1821 AliasCacheEntry* face = probe_alias_cache(flat);
1822 if (face->_adr_type == nullptr) {
1823 face->_adr_type = flat;
1824 face->_index = idx;
1825 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1826 }
1827 }
1828
1829 return alias_type(idx);
1830 }
1831
1832
1833 Compile::AliasType* Compile::alias_type(ciField* field) {
1834 const TypeOopPtr* t;
1835 if (field->is_static())
1836 t = TypeInstPtr::make(field->holder()->java_mirror());
1837 else
1838 t = TypeOopPtr::make_from_klass_raw(field->holder());
1839 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1840 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1841 return atp;
1842 }
1843
1844
1845 //------------------------------have_alias_type--------------------------------
1846 bool Compile::have_alias_type(const TypePtr* adr_type) {
1847 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1848 if (ace->_adr_type == adr_type) {
1849 return true;
1850 }
1851
1852 // Handle special cases.
1853 if (adr_type == nullptr) return true;
1854 if (adr_type == TypePtr::BOTTOM) return true;
1855
1856 return find_alias_type(adr_type, true, nullptr) != nullptr;
1857 }
1858
1859 //-----------------------------must_alias--------------------------------------
1860 // True if all values of the given address type are in the given alias category.
1861 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1862 if (alias_idx == AliasIdxBot) return true; // the universal category
1863 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1864 if (alias_idx == AliasIdxTop) return false; // the empty category
1865 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1866
1867 // the only remaining possible overlap is identity
1868 int adr_idx = get_alias_index(adr_type);
1869 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1870 assert(adr_idx == alias_idx ||
1871 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1872 && adr_type != TypeOopPtr::BOTTOM),
1873 "should not be testing for overlap with an unsafe pointer");
1874 return adr_idx == alias_idx;
1875 }
1876
1877 //------------------------------can_alias--------------------------------------
1878 // True if any values of the given address type are in the given alias category.
1879 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1880 if (alias_idx == AliasIdxTop) return false; // the empty category
1881 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1882 // Known instance doesn't alias with bottom memory
1883 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1884 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1885
1886 // the only remaining possible overlap is identity
1887 int adr_idx = get_alias_index(adr_type);
1888 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1889 return adr_idx == alias_idx;
1890 }
1891
1892 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1893 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1894 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1895 if (parse_predicate_count() == 0) {
1896 return;
1897 }
1898 for (int i = 0; i < parse_predicate_count(); i++) {
1899 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1900 parse_predicate->mark_useless(igvn);
1901 }
1902 _parse_predicates.clear();
1903 }
1904
1905 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1906 if (!n->for_post_loop_opts_igvn()) {
1907 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1908 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1909 _for_post_loop_igvn.append(n);
1910 }
1911 }
1912
1913 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1914 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1915 _for_post_loop_igvn.remove(n);
1916 }
1917
1918 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1919 // Verify that all previous optimizations produced a valid graph
1920 // at least to this point, even if no loop optimizations were done.
1921 PhaseIdealLoop::verify(igvn);
1922
1923 if (has_loops() || _loop_opts_cnt > 0) {
1924 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1925 }
1926 C->set_post_loop_opts_phase(); // no more loop opts allowed
1927
1928 assert(!C->major_progress(), "not cleared");
1929
1930 if (_for_post_loop_igvn.length() > 0) {
1931 while (_for_post_loop_igvn.length() > 0) {
1932 Node* n = _for_post_loop_igvn.pop();
1933 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1934 igvn._worklist.push(n);
1935 }
1936 igvn.optimize();
1937 if (failing()) return;
1938 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1939 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1940
1941 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1942 if (C->major_progress()) {
1943 C->clear_major_progress(); // ensure that major progress is now clear
1944 }
1945 }
1946 }
1947
1948 void Compile::add_inline_type(Node* n) {
1949 assert(n->is_InlineType(), "unexpected node");
1950 _inline_type_nodes.push(n);
1951 }
1952
1953 void Compile::remove_inline_type(Node* n) {
1954 assert(n->is_InlineType(), "unexpected node");
1955 if (_inline_type_nodes.contains(n)) {
1956 _inline_type_nodes.remove(n);
1957 }
1958 }
1959
1960 // Does the return value keep otherwise useless inline type allocations alive?
1961 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1962 ResourceMark rm;
1963 Unique_Node_List wq;
1964 wq.push(ret_val);
1965 bool some_allocations = false;
1966 for (uint i = 0; i < wq.size(); i++) {
1967 Node* n = wq.at(i);
1968 if (n->outcnt() > 1) {
1969 // Some other use for the allocation
1970 return false;
1971 } else if (n->is_InlineType()) {
1972 wq.push(n->in(1));
1973 } else if (n->is_Phi()) {
1974 for (uint j = 1; j < n->req(); j++) {
1975 wq.push(n->in(j));
1976 }
1977 } else if (n->is_CheckCastPP() &&
1978 n->in(1)->is_Proj() &&
1979 n->in(1)->in(0)->is_Allocate()) {
1980 some_allocations = true;
1981 } else if (n->is_CheckCastPP()) {
1982 wq.push(n->in(1));
1983 }
1984 }
1985 return some_allocations;
1986 }
1987
1988 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1989 // Make sure that the return value does not keep an otherwise unused allocation alive
1990 if (tf()->returns_inline_type_as_fields()) {
1991 Node* ret = nullptr;
1992 for (uint i = 1; i < root()->req(); i++) {
1993 Node* in = root()->in(i);
1994 if (in->Opcode() == Op_Return) {
1995 assert(ret == nullptr, "only one return");
1996 ret = in;
1997 }
1998 }
1999 if (ret != nullptr) {
2000 Node* ret_val = ret->in(TypeFunc::Parms);
2001 if (igvn.type(ret_val)->isa_oopptr() &&
2002 return_val_keeps_allocations_alive(ret_val)) {
2003 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
2004 assert(ret_val->outcnt() == 0, "should be dead now");
2005 igvn.remove_dead_node(ret_val);
2006 }
2007 }
2008 }
2009 if (_inline_type_nodes.length() == 0) {
2010 // keep the graph canonical
2011 igvn.optimize();
2012 return;
2013 }
2014 // Scalarize inline types in safepoint debug info.
2015 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2016 set_scalarize_in_safepoints(true);
2017 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2018 InlineTypeNode* vt = _inline_type_nodes.at(i)->as_InlineType();
2019 vt->make_scalar_in_safepoints(&igvn);
2020 igvn.record_for_igvn(vt);
2021 }
2022 if (remove) {
2023 // Remove inline type nodes by replacing them with their oop input
2024 while (_inline_type_nodes.length() > 0) {
2025 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2026 if (vt->outcnt() == 0) {
2027 igvn.remove_dead_node(vt);
2028 continue;
2029 }
2030 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2031 DEBUG_ONLY(bool must_be_buffered = false);
2032 Node* u = vt->out(i);
2033 // Check if any users are blackholes. If so, rewrite them to use either the
2034 // allocated buffer, or individual components, instead of the inline type node
2035 // that goes away.
2036 if (u->is_Blackhole()) {
2037 BlackholeNode* bh = u->as_Blackhole();
2038
2039 // Unlink the old input
2040 int idx = bh->find_edge(vt);
2041 assert(idx != -1, "The edge should be there");
2042 bh->del_req(idx);
2043 --i;
2044
2045 if (vt->is_allocated(&igvn)) {
2046 // Already has the allocated instance, blackhole that
2047 bh->add_req(vt->get_oop());
2048 } else {
2049 // Not allocated yet, blackhole the components
2050 for (uint c = 0; c < vt->field_count(); c++) {
2051 bh->add_req(vt->field_value(c));
2052 }
2053 }
2054
2055 // Node modified, record for IGVN
2056 igvn.record_for_igvn(bh);
2057 }
2058 #ifdef ASSERT
2059 // Verify that inline type is buffered when replacing by oop
2060 else if (u->is_InlineType()) {
2061 // InlineType uses don't need buffering because they are about to be replaced as well
2062 } else if (u->is_Phi()) {
2063 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2064 } else {
2065 must_be_buffered = true;
2066 }
2067 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2068 vt->dump(0);
2069 u->dump(0);
2070 assert(false, "Should have been buffered");
2071 }
2072 #endif
2073 }
2074 igvn.replace_node(vt, vt->get_oop());
2075 }
2076 }
2077 igvn.optimize();
2078 }
2079
2080 void Compile::add_flat_access(Node* n) {
2081 assert(n != nullptr && (n->Opcode() == Op_LoadFlat || n->Opcode() == Op_StoreFlat), "unexpected node %s", n == nullptr ? "nullptr" : n->Name());
2082 assert(!_flat_access_nodes.contains(n), "duplicate insertion");
2083 _flat_access_nodes.push(n);
2084 }
2085
2086 void Compile::remove_flat_access(Node* n) {
2087 assert(n != nullptr && (n->Opcode() == Op_LoadFlat || n->Opcode() == Op_StoreFlat), "unexpected node %s", n == nullptr ? "nullptr" : n->Name());
2088 _flat_access_nodes.remove_if_existing(n);
2089 }
2090
2091 void Compile::process_flat_accesses(PhaseIterGVN& igvn) {
2092 assert(igvn._worklist.size() == 0, "should be empty");
2093 igvn.set_delay_transform(true);
2094 for (int i = _flat_access_nodes.length() - 1; i >= 0; i--) {
2095 Node* n = _flat_access_nodes.at(i);
2096 assert(n != nullptr, "unexpected nullptr");
2097 if (n->is_LoadFlat()) {
2098 LoadFlatNode* loadn = n->as_LoadFlat();
2099 // Expending a flat load atomically means that we get a chunk of memory spanning multiple fields
2100 // that we chop with bitwise operations. That is too subtle for some optimizations, especially
2101 // constant folding when fields are constant. But if the flattened field being accessed is read-only
2102 // then no concurrent writes can happen and non-atomic loads are fine, allowing better optimizations.
2103 // A way for fields to be read-only is to be stable and already initialized. Here, we check if the
2104 // field being accessed is stable, and if the null marker of the field/array element is non-zero.
2105 // If so, we know that the stable value was initialized away from the default value (null), and
2106 // that we can assume it's read-only, so can the load can be performed non-atomically.
2107 bool non_atomic_is_fine = false;
2108 if (FoldStableValues) {
2109 const TypeOopPtr* base_type = igvn.type(loadn->base())->isa_oopptr();
2110 ciObject* oop = base_type->const_oop();
2111 ciInstance* holder = oop != nullptr && oop->is_instance() ? oop->as_instance() : nullptr;
2112 ciArray* array = oop != nullptr && oop->is_array() ? oop->as_array() : nullptr;
2113 int off = igvn.type(loadn->ptr())->isa_ptr()->offset();
2114
2115 if (holder != nullptr) {
2116 ciKlass* klass = holder->klass();
2117 ciInstanceKlass* iklass = klass->as_instance_klass();
2118 const ciField* field = iklass->get_non_flat_field_by_offset(off);
2119 ciField* nm_field = iklass->get_field_by_offset(field->null_marker_offset(), false);
2120 ciConstant cst = nm_field != nullptr ? holder->field_value(nm_field) : ciConstant() /* invalid */;
2121 non_atomic_is_fine = field->is_stable() && cst.is_valid() && cst.as_boolean();
2122 } else if (array != nullptr) {
2123 const TypeAryPtr* aryptr = base_type->is_aryptr();
2124 ciConstant elt = ((ciFlatArray*)array)->null_marker_of_element_by_offset(off);
2125 non_atomic_is_fine = aryptr->is_stable() && elt.is_valid() && !elt.is_null_or_zero();
2126 }
2127 }
2128
2129 if (non_atomic_is_fine) {
2130 loadn->expand_non_atomic(igvn);
2131 } else {
2132 loadn->expand_atomic(igvn);
2133 }
2134 } else {
2135 n->as_StoreFlat()->expand_atomic(igvn);
2136 }
2137 }
2138 _flat_access_nodes.clear_and_deallocate();
2139 igvn.set_delay_transform(false);
2140 igvn.optimize();
2141 }
2142
2143 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2144 DEBUG_ONLY(igvn.verify_empty_worklist(nullptr));
2145 if (!_has_flat_accesses) {
2146 return;
2147 }
2148 // Initially, all flat array accesses share the same slice to
2149 // keep dependencies with Object[] array accesses (that could be
2150 // to a flat array) correct. We're done with parsing so we
2151 // now know all flat array accesses in this compile
2152 // unit. Let's move flat array accesses to their own slice,
2153 // one per element field. This should help memory access
2154 // optimizations.
2155 ResourceMark rm;
2156 Unique_Node_List wq;
2157 wq.push(root());
2158
2159 Node_List mergememnodes;
2160 Node_List memnodes;
2161
2162 // Alias index currently shared by all flat memory accesses
2163 int index = get_alias_index(TypeAryPtr::INLINES);
2164
2165 // Find MergeMem nodes and flat array accesses
2166 for (uint i = 0; i < wq.size(); i++) {
2167 Node* n = wq.at(i);
2168 if (n->is_Mem()) {
2169 const TypePtr* adr_type = nullptr;
2170 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2171 if (adr_type == TypeAryPtr::INLINES) {
2172 memnodes.push(n);
2173 }
2174 } else if (n->is_MergeMem()) {
2175 MergeMemNode* mm = n->as_MergeMem();
2176 if (mm->memory_at(index) != mm->base_memory()) {
2177 mergememnodes.push(n);
2178 }
2179 }
2180 for (uint j = 0; j < n->req(); j++) {
2181 Node* m = n->in(j);
2182 if (m != nullptr) {
2183 wq.push(m);
2184 }
2185 }
2186 }
2187
2188 _flat_accesses_share_alias = false;
2189
2190 // We are going to change the slice for the flat array
2191 // accesses so we need to clear the cache entries that refer to
2192 // them.
2193 for (uint i = 0; i < AliasCacheSize; i++) {
2194 AliasCacheEntry* ace = &_alias_cache[i];
2195 if (ace->_adr_type != nullptr &&
2196 ace->_adr_type->is_flat()) {
2197 ace->_adr_type = nullptr;
2198 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2199 }
2200 }
2201
2202 #ifdef ASSERT
2203 for (uint i = 0; i < memnodes.size(); i++) {
2204 Node* m = memnodes.at(i);
2205 const TypePtr* adr_type = m->adr_type();
2206 m->as_Mem()->set_adr_type(adr_type);
2207 }
2208 #endif // ASSERT
2209
2210 int start_alias = num_alias_types(); // Start of new aliases
2211 Node_Stack stack(0);
2212 #ifdef ASSERT
2213 VectorSet seen(Thread::current()->resource_area());
2214 #endif
2215 // Now let's fix the memory graph so each flat array access
2216 // is moved to the right slice. Start from the MergeMem nodes.
2217 uint last = unique();
2218 for (uint i = 0; i < mergememnodes.size(); i++) {
2219 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2220 if (current->outcnt() == 0) {
2221 // This node is killed by a previous iteration
2222 continue;
2223 }
2224
2225 Node* n = current->memory_at(index);
2226 MergeMemNode* mm = nullptr;
2227 do {
2228 // Follow memory edges through memory accesses, phis and
2229 // narrow membars and push nodes on the stack. Once we hit
2230 // bottom memory, we pop element off the stack one at a
2231 // time, in reverse order, and move them to the right slice
2232 // by changing their memory edges.
2233 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() ||
2234 (n->adr_type() == TypeAryPtr::INLINES && !n->is_NarrowMemProj())) {
2235 assert(!seen.test_set(n->_idx), "");
2236 // Uses (a load for instance) will need to be moved to the
2237 // right slice as well and will get a new memory state
2238 // that we don't know yet. The use could also be the
2239 // backedge of a loop. We put a place holder node between
2240 // the memory node and its uses. We replace that place
2241 // holder with the correct memory state once we know it,
2242 // i.e. when nodes are popped off the stack. Using the
2243 // place holder make the logic work in the presence of
2244 // loops.
2245 if (n->outcnt() > 1) {
2246 Node* place_holder = nullptr;
2247 assert(!n->has_out_with(Op_Node), "");
2248 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2249 Node* u = n->out(k);
2250 if (u != current && u->_idx < last) {
2251 bool success = false;
2252 for (uint l = 0; l < u->req(); l++) {
2253 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2254 continue;
2255 }
2256 Node* in = u->in(l);
2257 if (in == n) {
2258 if (place_holder == nullptr) {
2259 place_holder = new Node(1);
2260 place_holder->init_req(0, n);
2261 }
2262 igvn.replace_input_of(u, l, place_holder);
2263 success = true;
2264 }
2265 }
2266 if (success) {
2267 --k;
2268 }
2269 }
2270 }
2271 }
2272 if (n->is_Phi()) {
2273 stack.push(n, 1);
2274 n = n->in(1);
2275 } else if (n->is_Mem()) {
2276 stack.push(n, n->req());
2277 n = n->in(MemNode::Memory);
2278 } else {
2279 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2280 stack.push(n, n->req());
2281 n = n->in(0)->in(TypeFunc::Memory);
2282 }
2283 } else {
2284 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || n->is_NarrowMemProj(), "");
2285 // Build a new MergeMem node to carry the new memory state
2286 // as we build it. IGVN should fold extraneous MergeMem
2287 // nodes.
2288 if (n->is_NarrowMemProj()) {
2289 // We need 1 NarrowMemProj for each slice of this array
2290 InitializeNode* init = n->in(0)->as_Initialize();
2291 AllocateNode* alloc = init->allocation();
2292 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2293 const TypeAryKlassPtr* klass_type = klass_node->bottom_type()->isa_aryklassptr();
2294 assert(klass_type != nullptr, "must be an array");
2295 assert(klass_type->klass_is_exact(), "must be an exact klass");
2296 ciArrayKlass* klass = klass_type->exact_klass()->as_array_klass();
2297 assert(klass->is_flat_array_klass(), "must be a flat array");
2298 ciInlineKlass* elem_klass = klass->element_klass()->as_inline_klass();
2299 const TypeAryPtr* oop_type = klass_type->as_instance_type()->is_aryptr();
2300 assert(oop_type->klass_is_exact(), "must be an exact klass");
2301
2302 Node* base = alloc->in(TypeFunc::Memory);
2303 assert(base->bottom_type() == Type::MEMORY, "the memory input of AllocateNode must be a memory");
2304 assert(base->adr_type() == TypePtr::BOTTOM, "the memory input of AllocateNode must be a bottom memory");
2305 // Must create a MergeMem with base as the base memory, do not clone if base is a
2306 // MergeMem because it may not be processed yet
2307 mm = MergeMemNode::make(nullptr);
2308 mm->set_base_memory(base);
2309 for (int j = 0; j < elem_klass->nof_nonstatic_fields(); j++) {
2310 int field_offset = elem_klass->nonstatic_field_at(j)->offset_in_bytes() - elem_klass->payload_offset();
2311 const TypeAryPtr* field_ptr = oop_type->with_offset(Type::OffsetBot)->with_field_offset(field_offset);
2312 int field_alias_idx = get_alias_index(field_ptr);
2313 assert(field_ptr == get_adr_type(field_alias_idx), "must match");
2314 Node* new_proj = new NarrowMemProjNode(init, field_ptr);
2315 igvn.register_new_node_with_optimizer(new_proj);
2316 mm->set_memory_at(field_alias_idx, new_proj);
2317 }
2318 if (!klass->is_elem_null_free()) {
2319 int nm_offset = elem_klass->null_marker_offset_in_payload();
2320 const TypeAryPtr* nm_ptr = oop_type->with_offset(Type::OffsetBot)->with_field_offset(nm_offset);
2321 int nm_alias_idx = get_alias_index(nm_ptr);
2322 assert(nm_ptr == get_adr_type(nm_alias_idx), "must match");
2323 Node* new_proj = new NarrowMemProjNode(init, nm_ptr);
2324 igvn.register_new_node_with_optimizer(new_proj);
2325 mm->set_memory_at(nm_alias_idx, new_proj);
2326 }
2327
2328 // Replace all uses of the old NarrowMemProj with the correct state
2329 MergeMemNode* new_n = MergeMemNode::make(mm);
2330 igvn.register_new_node_with_optimizer(new_n);
2331 igvn.replace_node(n, new_n);
2332 } else {
2333 // Must create a MergeMem with n as the base memory, do not clone if n is a MergeMem
2334 // because it may not be processed yet
2335 mm = MergeMemNode::make(nullptr);
2336 mm->set_base_memory(n);
2337 }
2338
2339 igvn.register_new_node_with_optimizer(mm);
2340 while (stack.size() > 0) {
2341 Node* m = stack.node();
2342 uint idx = stack.index();
2343 if (m->is_Mem()) {
2344 // Move memory node to its new slice
2345 const TypePtr* adr_type = m->adr_type();
2346 int alias = get_alias_index(adr_type);
2347 Node* prev = mm->memory_at(alias);
2348 igvn.replace_input_of(m, MemNode::Memory, prev);
2349 mm->set_memory_at(alias, m);
2350 } else if (m->is_Phi()) {
2351 // We need as many new phis as there are new aliases
2352 Node* new_phi_in = MergeMemNode::make(mm);
2353 igvn.register_new_node_with_optimizer(new_phi_in);
2354 igvn.replace_input_of(m, idx, new_phi_in);
2355 if (idx == m->req()-1) {
2356 Node* r = m->in(0);
2357 for (int j = start_alias; j < num_alias_types(); j++) {
2358 const TypePtr* adr_type = get_adr_type(j);
2359 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2360 continue;
2361 }
2362 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2363 igvn.register_new_node_with_optimizer(phi);
2364 for (uint k = 1; k < m->req(); k++) {
2365 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2366 }
2367 mm->set_memory_at(j, phi);
2368 }
2369 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2370 igvn.register_new_node_with_optimizer(base_phi);
2371 for (uint k = 1; k < m->req(); k++) {
2372 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2373 }
2374 mm->set_base_memory(base_phi);
2375 }
2376 } else {
2377 // This is a MemBarCPUOrder node from
2378 // Parse::array_load()/Parse::array_store(), in the
2379 // branch that handles flat arrays hidden under
2380 // an Object[] array. We also need one new membar per
2381 // new alias to keep the unknown access that the
2382 // membars protect properly ordered with accesses to
2383 // known flat array.
2384 assert(m->is_Proj(), "projection expected");
2385 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2386 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2387 for (int j = start_alias; j < num_alias_types(); j++) {
2388 const TypePtr* adr_type = get_adr_type(j);
2389 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2390 continue;
2391 }
2392 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2393 igvn.register_new_node_with_optimizer(mb);
2394 Node* mem = mm->memory_at(j);
2395 mb->init_req(TypeFunc::Control, ctrl);
2396 mb->init_req(TypeFunc::Memory, mem);
2397 ctrl = new ProjNode(mb, TypeFunc::Control);
2398 igvn.register_new_node_with_optimizer(ctrl);
2399 mem = new ProjNode(mb, TypeFunc::Memory);
2400 igvn.register_new_node_with_optimizer(mem);
2401 mm->set_memory_at(j, mem);
2402 }
2403 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2404 }
2405 if (idx < m->req()-1) {
2406 idx += 1;
2407 stack.set_index(idx);
2408 n = m->in(idx);
2409 break;
2410 }
2411 // Take care of place holder nodes
2412 if (m->has_out_with(Op_Node)) {
2413 Node* place_holder = m->find_out_with(Op_Node);
2414 if (place_holder != nullptr) {
2415 Node* mm_clone = mm->clone();
2416 igvn.register_new_node_with_optimizer(mm_clone);
2417 Node* hook = new Node(1);
2418 hook->init_req(0, mm);
2419 igvn.replace_node(place_holder, mm_clone);
2420 hook->destruct(&igvn);
2421 }
2422 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2423 }
2424 stack.pop();
2425 }
2426 }
2427 } while(stack.size() > 0);
2428 // Fix the memory state at the MergeMem we started from
2429 igvn.rehash_node_delayed(current);
2430 for (int j = start_alias; j < num_alias_types(); j++) {
2431 const TypePtr* adr_type = get_adr_type(j);
2432 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2433 continue;
2434 }
2435 current->set_memory_at(j, mm);
2436 }
2437 current->set_memory_at(index, current->base_memory());
2438 }
2439 igvn.optimize();
2440
2441 #ifdef ASSERT
2442 wq.clear();
2443 wq.push(root());
2444 for (uint i = 0; i < wq.size(); i++) {
2445 Node* n = wq.at(i);
2446 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2447 for (uint j = 0; j < n->req(); j++) {
2448 Node* m = n->in(j);
2449 if (m != nullptr) {
2450 wq.push(m);
2451 }
2452 }
2453 }
2454 #endif
2455
2456 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2457 }
2458
2459 void Compile::record_for_merge_stores_igvn(Node* n) {
2460 if (!n->for_merge_stores_igvn()) {
2461 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
2462 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2463 _for_merge_stores_igvn.append(n);
2464 }
2465 }
2466
2467 void Compile::remove_from_merge_stores_igvn(Node* n) {
2468 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2469 _for_merge_stores_igvn.remove(n);
2470 }
2471
2472 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
2473 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
2474 // the stores, and we merge the wrong sequence of stores.
2475 // Example:
2476 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
2477 // Apply MergeStores:
2478 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
2479 // Remove more RangeChecks:
2480 // StoreI [ StoreL ] StoreI
2481 // But now it would have been better to do this instead:
2482 // [ StoreL ] [ StoreL ]
2483 //
2484 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
2485 // since we never unset _merge_stores_phase.
2486 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
2487 C->set_merge_stores_phase();
2488
2489 if (_for_merge_stores_igvn.length() > 0) {
2490 while (_for_merge_stores_igvn.length() > 0) {
2491 Node* n = _for_merge_stores_igvn.pop();
2492 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2493 igvn._worklist.push(n);
2494 }
2495 igvn.optimize();
2496 if (failing()) return;
2497 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
2498 print_method(PHASE_AFTER_MERGE_STORES, 3);
2499 }
2500 }
2501
2502 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2503 if (OptimizeUnstableIf) {
2504 _unstable_if_traps.append(trap);
2505 }
2506 }
2507
2508 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2509 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2510 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2511 Node* n = trap->uncommon_trap();
2512 if (!useful.member(n)) {
2513 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2514 }
2515 }
2516 }
2517
2518 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2519 // or fold-compares case. Return true if succeed or not found.
2520 //
2521 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2522 // false and ask fold-compares to yield.
2523 //
2524 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2525 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2526 // when deoptimization does happen.
2527 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2528 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2529 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2530 if (trap->uncommon_trap() == unc) {
2531 if (yield && trap->modified()) {
2532 return false;
2533 }
2534 _unstable_if_traps.delete_at(i);
2535 break;
2536 }
2537 }
2538 return true;
2539 }
2540
2541 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2542 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2543 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2544 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2545 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2546 CallStaticJavaNode* unc = trap->uncommon_trap();
2547 int next_bci = trap->next_bci();
2548 bool modified = trap->modified();
2549
2550 if (next_bci != -1 && !modified) {
2551 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2552 JVMState* jvms = unc->jvms();
2553 ciMethod* method = jvms->method();
2554 ciBytecodeStream iter(method);
2555
2556 iter.force_bci(jvms->bci());
2557 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2558 Bytecodes::Code c = iter.cur_bc();
2559 Node* lhs = nullptr;
2560 Node* rhs = nullptr;
2561 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2562 lhs = unc->peek_operand(0);
2563 rhs = unc->peek_operand(1);
2564 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2565 lhs = unc->peek_operand(0);
2566 }
2567
2568 ResourceMark rm;
2569 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2570 assert(live_locals.is_valid(), "broken liveness info");
2571 int len = (int)live_locals.size();
2572
2573 for (int i = 0; i < len; i++) {
2574 Node* local = unc->local(jvms, i);
2575 // kill local using the liveness of next_bci.
2576 // give up when the local looks like an operand to secure reexecution.
2577 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2578 uint idx = jvms->locoff() + i;
2579 #ifdef ASSERT
2580 if (PrintOpto && Verbose) {
2581 tty->print("[unstable_if] kill local#%d: ", idx);
2582 local->dump();
2583 tty->cr();
2584 }
2585 #endif
2586 igvn.replace_input_of(unc, idx, top());
2587 modified = true;
2588 }
2589 }
2590 }
2591
2592 // keep the modified trap for late query
2593 if (modified) {
2594 trap->set_modified();
2595 } else {
2596 _unstable_if_traps.delete_at(i);
2597 }
2598 }
2599 igvn.optimize();
2600 }
2601
2602 // StringOpts and late inlining of string methods
2603 void Compile::inline_string_calls(bool parse_time) {
2604 {
2605 // remove useless nodes to make the usage analysis simpler
2606 ResourceMark rm;
2607 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2608 }
2609
2610 {
2611 ResourceMark rm;
2612 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2613 PhaseStringOpts pso(initial_gvn());
2614 print_method(PHASE_AFTER_STRINGOPTS, 3);
2615 }
2616
2617 // now inline anything that we skipped the first time around
2618 if (!parse_time) {
2619 _late_inlines_pos = _late_inlines.length();
2620 }
2621
2622 while (_string_late_inlines.length() > 0) {
2623 CallGenerator* cg = _string_late_inlines.pop();
2624 cg->do_late_inline();
2625 if (failing()) return;
2626 }
2627 _string_late_inlines.trunc_to(0);
2628 }
2629
2630 // Late inlining of boxing methods
2631 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2632 if (_boxing_late_inlines.length() > 0) {
2633 assert(has_boxed_value(), "inconsistent");
2634
2635 set_inlining_incrementally(true);
2636
2637 igvn_worklist()->ensure_empty(); // should be done with igvn
2638
2639 _late_inlines_pos = _late_inlines.length();
2640
2641 while (_boxing_late_inlines.length() > 0) {
2642 CallGenerator* cg = _boxing_late_inlines.pop();
2643 cg->do_late_inline();
2644 if (failing()) return;
2645 }
2646 _boxing_late_inlines.trunc_to(0);
2647
2648 inline_incrementally_cleanup(igvn);
2649
2650 set_inlining_incrementally(false);
2651 }
2652 }
2653
2654 bool Compile::inline_incrementally_one() {
2655 assert(IncrementalInline, "incremental inlining should be on");
2656
2657 TracePhase tp(_t_incrInline_inline);
2658
2659 set_inlining_progress(false);
2660 set_do_cleanup(false);
2661
2662 for (int i = 0; i < _late_inlines.length(); i++) {
2663 _late_inlines_pos = i+1;
2664 CallGenerator* cg = _late_inlines.at(i);
2665 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2666 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2667 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2668 if (should_stress_inlining()) {
2669 // randomly add repeated inline attempt if stress-inlining
2670 cg->call_node()->set_generator(cg);
2671 C->igvn_worklist()->push(cg->call_node());
2672 continue;
2673 }
2674 cg->do_late_inline();
2675 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2676 if (failing()) {
2677 return false;
2678 } else if (inlining_progress()) {
2679 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2680 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2681 break; // process one call site at a time
2682 } else {
2683 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2684 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2685 // Avoid potential infinite loop if node already in the IGVN list
2686 assert(false, "scheduled for IGVN during inlining attempt");
2687 } else {
2688 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2689 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2690 cg->call_node()->set_generator(cg);
2691 }
2692 }
2693 } else {
2694 // Ignore late inline direct calls when inlining is not allowed.
2695 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2696 }
2697 }
2698 // Remove processed elements.
2699 _late_inlines.remove_till(_late_inlines_pos);
2700 _late_inlines_pos = 0;
2701
2702 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2703
2704 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2705
2706 set_inlining_progress(false);
2707 set_do_cleanup(false);
2708
2709 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2710 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2711 }
2712
2713 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2714 {
2715 TracePhase tp(_t_incrInline_pru);
2716 ResourceMark rm;
2717 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2718 }
2719 {
2720 TracePhase tp(_t_incrInline_igvn);
2721 igvn.reset();
2722 igvn.optimize();
2723 if (failing()) return;
2724 }
2725 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2726 }
2727
2728 // Perform incremental inlining until bound on number of live nodes is reached
2729 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2730 TracePhase tp(_t_incrInline);
2731
2732 set_inlining_incrementally(true);
2733 uint low_live_nodes = 0;
2734
2735 while (_late_inlines.length() > 0) {
2736 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2737 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2738 TracePhase tp(_t_incrInline_ideal);
2739 // PhaseIdealLoop is expensive so we only try it once we are
2740 // out of live nodes and we only try it again if the previous
2741 // helped got the number of nodes down significantly
2742 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2743 if (failing()) return;
2744 low_live_nodes = live_nodes();
2745 _major_progress = true;
2746 }
2747
2748 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2749 bool do_print_inlining = print_inlining() || print_intrinsics();
2750 if (do_print_inlining || log() != nullptr) {
2751 // Print inlining message for candidates that we couldn't inline for lack of space.
2752 for (int i = 0; i < _late_inlines.length(); i++) {
2753 CallGenerator* cg = _late_inlines.at(i);
2754 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2755 if (do_print_inlining) {
2756 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2757 }
2758 log_late_inline_failure(cg, msg);
2759 }
2760 }
2761 break; // finish
2762 }
2763 }
2764
2765 igvn_worklist()->ensure_empty(); // should be done with igvn
2766
2767 while (inline_incrementally_one()) {
2768 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2769 }
2770 if (failing()) return;
2771
2772 inline_incrementally_cleanup(igvn);
2773
2774 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2775
2776 if (failing()) return;
2777
2778 if (_late_inlines.length() == 0) {
2779 break; // no more progress
2780 }
2781 }
2782
2783 igvn_worklist()->ensure_empty(); // should be done with igvn
2784
2785 if (_string_late_inlines.length() > 0) {
2786 assert(has_stringbuilder(), "inconsistent");
2787
2788 inline_string_calls(false);
2789
2790 if (failing()) return;
2791
2792 inline_incrementally_cleanup(igvn);
2793 }
2794
2795 set_inlining_incrementally(false);
2796 }
2797
2798 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2799 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2800 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2801 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2802 // as if "inlining_incrementally() == true" were set.
2803 assert(inlining_incrementally() == false, "not allowed");
2804 set_strength_reduction(true);
2805 #ifdef ASSERT
2806 Unique_Node_List* modified_nodes = _modified_nodes;
2807 _modified_nodes = nullptr;
2808 #endif
2809 assert(_late_inlines.length() > 0, "sanity");
2810
2811 while (_late_inlines.length() > 0) {
2812 igvn_worklist()->ensure_empty(); // should be done with igvn
2813
2814 while (inline_incrementally_one()) {
2815 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2816 }
2817 if (failing()) return;
2818
2819 inline_incrementally_cleanup(igvn);
2820 }
2821 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2822 set_strength_reduction(false);
2823 }
2824
2825 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2826 if (_loop_opts_cnt > 0) {
2827 while (major_progress() && (_loop_opts_cnt > 0)) {
2828 TracePhase tp(_t_idealLoop);
2829 PhaseIdealLoop::optimize(igvn, mode);
2830 _loop_opts_cnt--;
2831 if (failing()) return false;
2832 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2833 }
2834 }
2835 return true;
2836 }
2837
2838 // Remove edges from "root" to each SafePoint at a backward branch.
2839 // They were inserted during parsing (see add_safepoint()) to make
2840 // infinite loops without calls or exceptions visible to root, i.e.,
2841 // useful.
2842 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2843 Node *r = root();
2844 if (r != nullptr) {
2845 for (uint i = r->req(); i < r->len(); ++i) {
2846 Node *n = r->in(i);
2847 if (n != nullptr && n->is_SafePoint()) {
2848 r->rm_prec(i);
2849 if (n->outcnt() == 0) {
2850 igvn.remove_dead_node(n);
2851 }
2852 --i;
2853 }
2854 }
2855 // Parsing may have added top inputs to the root node (Path
2856 // leading to the Halt node proven dead). Make sure we get a
2857 // chance to clean them up.
2858 igvn._worklist.push(r);
2859 igvn.optimize();
2860 }
2861 }
2862
2863 //------------------------------Optimize---------------------------------------
2864 // Given a graph, optimize it.
2865 void Compile::Optimize() {
2866 TracePhase tp(_t_optimizer);
2867
2868 #ifndef PRODUCT
2869 if (env()->break_at_compile()) {
2870 BREAKPOINT;
2871 }
2872
2873 #endif
2874
2875 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2876 #ifdef ASSERT
2877 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2878 #endif
2879
2880 ResourceMark rm;
2881
2882 NOT_PRODUCT( verify_graph_edges(); )
2883
2884 print_method(PHASE_AFTER_PARSING, 1);
2885
2886 {
2887 // Iterative Global Value Numbering, including ideal transforms
2888 PhaseIterGVN igvn;
2889 #ifdef ASSERT
2890 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2891 #endif
2892 {
2893 TracePhase tp(_t_iterGVN);
2894 igvn.optimize();
2895 }
2896
2897 if (failing()) return;
2898
2899 print_method(PHASE_ITER_GVN1, 2);
2900
2901 process_for_unstable_if_traps(igvn);
2902
2903 if (failing()) return;
2904
2905 inline_incrementally(igvn);
2906
2907 print_method(PHASE_INCREMENTAL_INLINE, 2);
2908
2909 if (failing()) return;
2910
2911 if (eliminate_boxing()) {
2912 // Inline valueOf() methods now.
2913 inline_boxing_calls(igvn);
2914
2915 if (failing()) return;
2916
2917 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2918 inline_incrementally(igvn);
2919 }
2920
2921 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2922
2923 if (failing()) return;
2924 }
2925
2926 // Remove the speculative part of types and clean up the graph from
2927 // the extra CastPP nodes whose only purpose is to carry them. Do
2928 // that early so that optimizations are not disrupted by the extra
2929 // CastPP nodes.
2930 remove_speculative_types(igvn);
2931
2932 if (failing()) return;
2933
2934 // No more new expensive nodes will be added to the list from here
2935 // so keep only the actual candidates for optimizations.
2936 cleanup_expensive_nodes(igvn);
2937
2938 if (failing()) return;
2939
2940 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2941 if (EnableVectorSupport && has_vbox_nodes()) {
2942 TracePhase tp(_t_vector);
2943 PhaseVector pv(igvn);
2944 pv.optimize_vector_boxes();
2945 if (failing()) return;
2946 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2947 }
2948 assert(!has_vbox_nodes(), "sanity");
2949
2950 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2951 Compile::TracePhase tp(_t_renumberLive);
2952 igvn_worklist()->ensure_empty(); // should be done with igvn
2953 {
2954 ResourceMark rm;
2955 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2956 }
2957 igvn.reset();
2958 igvn.optimize();
2959 if (failing()) return;
2960 }
2961
2962 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2963 // safepoints
2964 remove_root_to_sfpts_edges(igvn);
2965
2966 // Process inline type nodes now that all inlining is over
2967 process_inline_types(igvn);
2968
2969 adjust_flat_array_access_aliases(igvn);
2970
2971 if (failing()) return;
2972
2973 if (C->macro_count() > 0) {
2974 // Eliminate some macro nodes before EA to reduce analysis pressure
2975 PhaseMacroExpand mexp(igvn);
2976 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
2977 if (failing()) {
2978 return;
2979 }
2980 igvn.set_delay_transform(false);
2981 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2982 }
2983
2984 if (has_loops()) {
2985 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2986 }
2987
2988 // Perform escape analysis
2989 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2990 if (has_loops()) {
2991 // Cleanup graph (remove dead nodes).
2992 TracePhase tp(_t_idealLoop);
2993 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2994 if (failing()) {
2995 return;
2996 }
2997 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2998 if (C->macro_count() > 0) {
2999 // Eliminate some macro nodes before EA to reduce analysis pressure
3000 PhaseMacroExpand mexp(igvn);
3001 mexp.eliminate_macro_nodes(/* eliminate_locks= */ false);
3002 if (failing()) {
3003 return;
3004 }
3005 igvn.set_delay_transform(false);
3006 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3007 }
3008 }
3009
3010 bool progress;
3011 do {
3012 ConnectionGraph::do_analysis(this, &igvn);
3013
3014 if (failing()) return;
3015
3016 int mcount = macro_count(); // Record number of allocations and locks before IGVN
3017
3018 // Optimize out fields loads from scalar replaceable allocations.
3019 igvn.optimize();
3020 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
3021
3022 if (failing()) return;
3023
3024 if (congraph() != nullptr && macro_count() > 0) {
3025 TracePhase tp(_t_macroEliminate);
3026 PhaseMacroExpand mexp(igvn);
3027 mexp.eliminate_macro_nodes();
3028 if (failing()) {
3029 return;
3030 }
3031 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
3032
3033 igvn.set_delay_transform(false);
3034 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
3035 }
3036
3037 ConnectionGraph::verify_ram_nodes(this, root());
3038 if (failing()) return;
3039
3040 progress = do_iterative_escape_analysis() &&
3041 (macro_count() < mcount) &&
3042 ConnectionGraph::has_candidates(this);
3043 // Try again if candidates exist and made progress
3044 // by removing some allocations and/or locks.
3045 } while (progress);
3046 }
3047
3048 process_flat_accesses(igvn);
3049 if (failing()) {
3050 return;
3051 }
3052
3053 // Loop transforms on the ideal graph. Range Check Elimination,
3054 // peeling, unrolling, etc.
3055
3056 // Set loop opts counter
3057 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
3058 {
3059 TracePhase tp(_t_idealLoop);
3060 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
3061 _loop_opts_cnt--;
3062 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
3063 if (failing()) return;
3064 }
3065 // Loop opts pass if partial peeling occurred in previous pass
3066 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
3067 TracePhase tp(_t_idealLoop);
3068 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
3069 _loop_opts_cnt--;
3070 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
3071 if (failing()) return;
3072 }
3073 // Loop opts pass for loop-unrolling before CCP
3074 if(major_progress() && (_loop_opts_cnt > 0)) {
3075 TracePhase tp(_t_idealLoop);
3076 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
3077 _loop_opts_cnt--;
3078 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
3079 }
3080 if (!failing()) {
3081 // Verify that last round of loop opts produced a valid graph
3082 PhaseIdealLoop::verify(igvn);
3083 }
3084 }
3085 if (failing()) return;
3086
3087 // Conditional Constant Propagation;
3088 print_method(PHASE_BEFORE_CCP1, 2);
3089 PhaseCCP ccp( &igvn );
3090 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
3091 {
3092 TracePhase tp(_t_ccp);
3093 ccp.do_transform();
3094 }
3095 print_method(PHASE_CCP1, 2);
3096
3097 assert( true, "Break here to ccp.dump_old2new_map()");
3098
3099 // Iterative Global Value Numbering, including ideal transforms
3100 {
3101 TracePhase tp(_t_iterGVN2);
3102 igvn.reset_from_igvn(&ccp);
3103 igvn.optimize();
3104 }
3105 print_method(PHASE_ITER_GVN2, 2);
3106
3107 if (failing()) return;
3108
3109 // Loop transforms on the ideal graph. Range Check Elimination,
3110 // peeling, unrolling, etc.
3111 if (!optimize_loops(igvn, LoopOptsDefault)) {
3112 return;
3113 }
3114
3115 if (failing()) return;
3116
3117 C->clear_major_progress(); // ensure that major progress is now clear
3118
3119 process_for_post_loop_opts_igvn(igvn);
3120
3121 process_for_merge_stores_igvn(igvn);
3122
3123 if (failing()) return;
3124
3125 #ifdef ASSERT
3126 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
3127 #endif
3128
3129 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
3130
3131 if (_late_inlines.length() > 0) {
3132 // More opportunities to optimize virtual and MH calls.
3133 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
3134 process_late_inline_calls_no_inline(igvn);
3135 }
3136
3137 {
3138 TracePhase tp(_t_macroExpand);
3139 PhaseMacroExpand mex(igvn);
3140 // Last attempt to eliminate macro nodes.
3141 mex.eliminate_macro_nodes();
3142 if (failing()) {
3143 return;
3144 }
3145
3146 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
3147 // Do not allow new macro nodes once we start to eliminate and expand
3148 C->reset_allow_macro_nodes();
3149 // Last attempt to eliminate macro nodes before expand
3150 mex.eliminate_macro_nodes();
3151 if (failing()) {
3152 return;
3153 }
3154 mex.eliminate_opaque_looplimit_macro_nodes();
3155 if (failing()) {
3156 return;
3157 }
3158 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
3159 if (mex.expand_macro_nodes()) {
3160 assert(failing(), "must bail out w/ explicit message");
3161 return;
3162 }
3163 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
3164 }
3165
3166 // Process inline type nodes again and remove them. From here
3167 // on we don't need to keep track of field values anymore.
3168 process_inline_types(igvn, /* remove= */ true);
3169
3170 {
3171 TracePhase tp(_t_barrierExpand);
3172 if (bs->expand_barriers(this, igvn)) {
3173 assert(failing(), "must bail out w/ explicit message");
3174 return;
3175 }
3176 print_method(PHASE_BARRIER_EXPANSION, 2);
3177 }
3178
3179 if (C->max_vector_size() > 0) {
3180 C->optimize_logic_cones(igvn);
3181 igvn.optimize();
3182 if (failing()) return;
3183 }
3184
3185 DEBUG_ONLY( _modified_nodes = nullptr; )
3186 DEBUG_ONLY( _late_inlines.clear(); )
3187
3188 assert(igvn._worklist.size() == 0, "not empty");
3189 } // (End scope of igvn; run destructor if necessary for asserts.)
3190
3191 check_no_dead_use();
3192
3193 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
3194 // to remove hashes to unlock nodes for modifications.
3195 C->node_hash()->clear();
3196
3197 // A method with only infinite loops has no edges entering loops from root
3198 {
3199 TracePhase tp(_t_graphReshaping);
3200 if (final_graph_reshaping()) {
3201 assert(failing(), "must bail out w/ explicit message");
3202 return;
3203 }
3204 }
3205
3206 print_method(PHASE_OPTIMIZE_FINISHED, 2);
3207 DEBUG_ONLY(set_phase_optimize_finished();)
3208 }
3209
3210 #ifdef ASSERT
3211 void Compile::check_no_dead_use() const {
3212 ResourceMark rm;
3213 Unique_Node_List wq;
3214 wq.push(root());
3215 for (uint i = 0; i < wq.size(); ++i) {
3216 Node* n = wq.at(i);
3217 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3218 Node* u = n->fast_out(j);
3219 if (u->outcnt() == 0 && !u->is_Con()) {
3220 u->dump();
3221 fatal("no reachable node should have no use");
3222 }
3223 wq.push(u);
3224 }
3225 }
3226 }
3227 #endif
3228
3229 void Compile::inline_vector_reboxing_calls() {
3230 if (C->_vector_reboxing_late_inlines.length() > 0) {
3231 _late_inlines_pos = C->_late_inlines.length();
3232 while (_vector_reboxing_late_inlines.length() > 0) {
3233 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
3234 cg->do_late_inline();
3235 if (failing()) return;
3236 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
3237 }
3238 _vector_reboxing_late_inlines.trunc_to(0);
3239 }
3240 }
3241
3242 bool Compile::has_vbox_nodes() {
3243 if (C->_vector_reboxing_late_inlines.length() > 0) {
3244 return true;
3245 }
3246 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3247 Node * n = C->macro_node(macro_idx);
3248 assert(n->is_macro(), "only macro nodes expected here");
3249 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3250 return true;
3251 }
3252 }
3253 return false;
3254 }
3255
3256 //---------------------------- Bitwise operation packing optimization ---------------------------
3257
3258 static bool is_vector_unary_bitwise_op(Node* n) {
3259 return n->Opcode() == Op_XorV &&
3260 VectorNode::is_vector_bitwise_not_pattern(n);
3261 }
3262
3263 static bool is_vector_binary_bitwise_op(Node* n) {
3264 switch (n->Opcode()) {
3265 case Op_AndV:
3266 case Op_OrV:
3267 return true;
3268
3269 case Op_XorV:
3270 return !is_vector_unary_bitwise_op(n);
3271
3272 default:
3273 return false;
3274 }
3275 }
3276
3277 static bool is_vector_ternary_bitwise_op(Node* n) {
3278 return n->Opcode() == Op_MacroLogicV;
3279 }
3280
3281 static bool is_vector_bitwise_op(Node* n) {
3282 return is_vector_unary_bitwise_op(n) ||
3283 is_vector_binary_bitwise_op(n) ||
3284 is_vector_ternary_bitwise_op(n);
3285 }
3286
3287 static bool is_vector_bitwise_cone_root(Node* n) {
3288 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
3289 return false;
3290 }
3291 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3292 if (is_vector_bitwise_op(n->fast_out(i))) {
3293 return false;
3294 }
3295 }
3296 return true;
3297 }
3298
3299 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3300 uint cnt = 0;
3301 if (is_vector_bitwise_op(n)) {
3302 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3303 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3304 for (uint i = 1; i < inp_cnt; i++) {
3305 Node* in = n->in(i);
3306 bool skip = VectorNode::is_all_ones_vector(in);
3307 if (!skip && !inputs.member(in)) {
3308 inputs.push(in);
3309 cnt++;
3310 }
3311 }
3312 assert(cnt <= 1, "not unary");
3313 } else {
3314 uint last_req = inp_cnt;
3315 if (is_vector_ternary_bitwise_op(n)) {
3316 last_req = inp_cnt - 1; // skip last input
3317 }
3318 for (uint i = 1; i < last_req; i++) {
3319 Node* def = n->in(i);
3320 if (!inputs.member(def)) {
3321 inputs.push(def);
3322 cnt++;
3323 }
3324 }
3325 }
3326 } else { // not a bitwise operations
3327 if (!inputs.member(n)) {
3328 inputs.push(n);
3329 cnt++;
3330 }
3331 }
3332 return cnt;
3333 }
3334
3335 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3336 Unique_Node_List useful_nodes;
3337 C->identify_useful_nodes(useful_nodes);
3338
3339 for (uint i = 0; i < useful_nodes.size(); i++) {
3340 Node* n = useful_nodes.at(i);
3341 if (is_vector_bitwise_cone_root(n)) {
3342 list.push(n);
3343 }
3344 }
3345 }
3346
3347 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3348 const TypeVect* vt,
3349 Unique_Node_List& partition,
3350 Unique_Node_List& inputs) {
3351 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3352 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3353 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3354
3355 Node* in1 = inputs.at(0);
3356 Node* in2 = inputs.at(1);
3357 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3358
3359 uint func = compute_truth_table(partition, inputs);
3360
3361 Node* pn = partition.at(partition.size() - 1);
3362 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3363 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3364 }
3365
3366 static uint extract_bit(uint func, uint pos) {
3367 return (func & (1 << pos)) >> pos;
3368 }
3369
3370 //
3371 // A macro logic node represents a truth table. It has 4 inputs,
3372 // First three inputs corresponds to 3 columns of a truth table
3373 // and fourth input captures the logic function.
3374 //
3375 // eg. fn = (in1 AND in2) OR in3;
3376 //
3377 // MacroNode(in1,in2,in3,fn)
3378 //
3379 // -----------------
3380 // in1 in2 in3 fn
3381 // -----------------
3382 // 0 0 0 0
3383 // 0 0 1 1
3384 // 0 1 0 0
3385 // 0 1 1 1
3386 // 1 0 0 0
3387 // 1 0 1 1
3388 // 1 1 0 1
3389 // 1 1 1 1
3390 //
3391
3392 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3393 int res = 0;
3394 for (int i = 0; i < 8; i++) {
3395 int bit1 = extract_bit(in1, i);
3396 int bit2 = extract_bit(in2, i);
3397 int bit3 = extract_bit(in3, i);
3398
3399 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3400 int func_bit = extract_bit(func, func_bit_pos);
3401
3402 res |= func_bit << i;
3403 }
3404 return res;
3405 }
3406
3407 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
3408 assert(n != nullptr, "");
3409 assert(eval_map.contains(n), "absent");
3410 return *(eval_map.get(n));
3411 }
3412
3413 static void eval_operands(Node* n,
3414 uint& func1, uint& func2, uint& func3,
3415 HashTable<Node*,uint>& eval_map) {
3416 assert(is_vector_bitwise_op(n), "");
3417
3418 if (is_vector_unary_bitwise_op(n)) {
3419 Node* opnd = n->in(1);
3420 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3421 opnd = n->in(2);
3422 }
3423 func1 = eval_operand(opnd, eval_map);
3424 } else if (is_vector_binary_bitwise_op(n)) {
3425 func1 = eval_operand(n->in(1), eval_map);
3426 func2 = eval_operand(n->in(2), eval_map);
3427 } else {
3428 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3429 func1 = eval_operand(n->in(1), eval_map);
3430 func2 = eval_operand(n->in(2), eval_map);
3431 func3 = eval_operand(n->in(3), eval_map);
3432 }
3433 }
3434
3435 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3436 assert(inputs.size() <= 3, "sanity");
3437 ResourceMark rm;
3438 uint res = 0;
3439 HashTable<Node*,uint> eval_map;
3440
3441 // Populate precomputed functions for inputs.
3442 // Each input corresponds to one column of 3 input truth-table.
3443 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3444 0xCC, // (_, b, _) -> b
3445 0xF0 }; // (a, _, _) -> a
3446 for (uint i = 0; i < inputs.size(); i++) {
3447 eval_map.put(inputs.at(i), input_funcs[2-i]);
3448 }
3449
3450 for (uint i = 0; i < partition.size(); i++) {
3451 Node* n = partition.at(i);
3452
3453 uint func1 = 0, func2 = 0, func3 = 0;
3454 eval_operands(n, func1, func2, func3, eval_map);
3455
3456 switch (n->Opcode()) {
3457 case Op_OrV:
3458 assert(func3 == 0, "not binary");
3459 res = func1 | func2;
3460 break;
3461 case Op_AndV:
3462 assert(func3 == 0, "not binary");
3463 res = func1 & func2;
3464 break;
3465 case Op_XorV:
3466 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3467 assert(func2 == 0 && func3 == 0, "not unary");
3468 res = (~func1) & 0xFF;
3469 } else {
3470 assert(func3 == 0, "not binary");
3471 res = func1 ^ func2;
3472 }
3473 break;
3474 case Op_MacroLogicV:
3475 // Ordering of inputs may change during evaluation of sub-tree
3476 // containing MacroLogic node as a child node, thus a re-evaluation
3477 // makes sure that function is evaluated in context of current
3478 // inputs.
3479 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3480 break;
3481
3482 default: assert(false, "not supported: %s", n->Name());
3483 }
3484 assert(res <= 0xFF, "invalid");
3485 eval_map.put(n, res);
3486 }
3487 return res;
3488 }
3489
3490 // Criteria under which nodes gets packed into a macro logic node:-
3491 // 1) Parent and both child nodes are all unmasked or masked with
3492 // same predicates.
3493 // 2) Masked parent can be packed with left child if it is predicated
3494 // and both have same predicates.
3495 // 3) Masked parent can be packed with right child if its un-predicated
3496 // or has matching predication condition.
3497 // 4) An unmasked parent can be packed with an unmasked child.
3498 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3499 assert(partition.size() == 0, "not empty");
3500 assert(inputs.size() == 0, "not empty");
3501 if (is_vector_ternary_bitwise_op(n)) {
3502 return false;
3503 }
3504
3505 bool is_unary_op = is_vector_unary_bitwise_op(n);
3506 if (is_unary_op) {
3507 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3508 return false; // too few inputs
3509 }
3510
3511 bool pack_left_child = true;
3512 bool pack_right_child = true;
3513
3514 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3515 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3516
3517 int left_child_input_cnt = 0;
3518 int right_child_input_cnt = 0;
3519
3520 bool parent_is_predicated = n->is_predicated_vector();
3521 bool left_child_predicated = n->in(1)->is_predicated_vector();
3522 bool right_child_predicated = n->in(2)->is_predicated_vector();
3523
3524 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
3525 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3526 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3527
3528 do {
3529 if (pack_left_child && left_child_LOP &&
3530 ((!parent_is_predicated && !left_child_predicated) ||
3531 ((parent_is_predicated && left_child_predicated &&
3532 parent_pred == left_child_pred)))) {
3533 partition.push(n->in(1));
3534 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3535 } else {
3536 inputs.push(n->in(1));
3537 left_child_input_cnt = 1;
3538 }
3539
3540 if (pack_right_child && right_child_LOP &&
3541 (!right_child_predicated ||
3542 (right_child_predicated && parent_is_predicated &&
3543 parent_pred == right_child_pred))) {
3544 partition.push(n->in(2));
3545 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3546 } else {
3547 inputs.push(n->in(2));
3548 right_child_input_cnt = 1;
3549 }
3550
3551 if (inputs.size() > 3) {
3552 assert(partition.size() > 0, "");
3553 inputs.clear();
3554 partition.clear();
3555 if (left_child_input_cnt > right_child_input_cnt) {
3556 pack_left_child = false;
3557 } else {
3558 pack_right_child = false;
3559 }
3560 } else {
3561 break;
3562 }
3563 } while(true);
3564
3565 if(partition.size()) {
3566 partition.push(n);
3567 }
3568
3569 return (partition.size() == 2 || partition.size() == 3) &&
3570 (inputs.size() == 2 || inputs.size() == 3);
3571 }
3572
3573 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3574 assert(is_vector_bitwise_op(n), "not a root");
3575
3576 visited.set(n->_idx);
3577
3578 // 1) Do a DFS walk over the logic cone.
3579 for (uint i = 1; i < n->req(); i++) {
3580 Node* in = n->in(i);
3581 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3582 process_logic_cone_root(igvn, in, visited);
3583 }
3584 }
3585
3586 // 2) Bottom up traversal: Merge node[s] with
3587 // the parent to form macro logic node.
3588 Unique_Node_List partition;
3589 Unique_Node_List inputs;
3590 if (compute_logic_cone(n, partition, inputs)) {
3591 const TypeVect* vt = n->bottom_type()->is_vect();
3592 Node* pn = partition.at(partition.size() - 1);
3593 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3594 if (mask == nullptr ||
3595 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3596 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3597 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3598 igvn.replace_node(n, macro_logic);
3599 }
3600 }
3601 }
3602
3603 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3604 ResourceMark rm;
3605 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3606 Unique_Node_List list;
3607 collect_logic_cone_roots(list);
3608
3609 while (list.size() > 0) {
3610 Node* n = list.pop();
3611 const TypeVect* vt = n->bottom_type()->is_vect();
3612 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3613 if (supported) {
3614 VectorSet visited(comp_arena());
3615 process_logic_cone_root(igvn, n, visited);
3616 }
3617 }
3618 }
3619 }
3620
3621 //------------------------------Code_Gen---------------------------------------
3622 // Given a graph, generate code for it
3623 void Compile::Code_Gen() {
3624 if (failing()) {
3625 return;
3626 }
3627
3628 // Perform instruction selection. You might think we could reclaim Matcher
3629 // memory PDQ, but actually the Matcher is used in generating spill code.
3630 // Internals of the Matcher (including some VectorSets) must remain live
3631 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3632 // set a bit in reclaimed memory.
3633
3634 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3635 // nodes. Mapping is only valid at the root of each matched subtree.
3636 NOT_PRODUCT( verify_graph_edges(); )
3637
3638 Matcher matcher;
3639 _matcher = &matcher;
3640 {
3641 TracePhase tp(_t_matcher);
3642 matcher.match();
3643 if (failing()) {
3644 return;
3645 }
3646 }
3647 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3648 // nodes. Mapping is only valid at the root of each matched subtree.
3649 NOT_PRODUCT( verify_graph_edges(); )
3650
3651 // If you have too many nodes, or if matching has failed, bail out
3652 check_node_count(0, "out of nodes matching instructions");
3653 if (failing()) {
3654 return;
3655 }
3656
3657 print_method(PHASE_MATCHING, 2);
3658
3659 // Build a proper-looking CFG
3660 PhaseCFG cfg(node_arena(), root(), matcher);
3661 if (failing()) {
3662 return;
3663 }
3664 _cfg = &cfg;
3665 {
3666 TracePhase tp(_t_scheduler);
3667 bool success = cfg.do_global_code_motion();
3668 if (!success) {
3669 return;
3670 }
3671
3672 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3673 NOT_PRODUCT( verify_graph_edges(); )
3674 cfg.verify();
3675 if (failing()) {
3676 return;
3677 }
3678 }
3679
3680 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3681 _regalloc = ®alloc;
3682 {
3683 TracePhase tp(_t_registerAllocation);
3684 // Perform register allocation. After Chaitin, use-def chains are
3685 // no longer accurate (at spill code) and so must be ignored.
3686 // Node->LRG->reg mappings are still accurate.
3687 _regalloc->Register_Allocate();
3688
3689 // Bail out if the allocator builds too many nodes
3690 if (failing()) {
3691 return;
3692 }
3693
3694 print_method(PHASE_REGISTER_ALLOCATION, 2);
3695 }
3696
3697 // Prior to register allocation we kept empty basic blocks in case the
3698 // the allocator needed a place to spill. After register allocation we
3699 // are not adding any new instructions. If any basic block is empty, we
3700 // can now safely remove it.
3701 {
3702 TracePhase tp(_t_blockOrdering);
3703 cfg.remove_empty_blocks();
3704 if (do_freq_based_layout()) {
3705 PhaseBlockLayout layout(cfg);
3706 } else {
3707 cfg.set_loop_alignment();
3708 }
3709 cfg.fixup_flow();
3710 cfg.remove_unreachable_blocks();
3711 cfg.verify_dominator_tree();
3712 print_method(PHASE_BLOCK_ORDERING, 3);
3713 }
3714
3715 // Apply peephole optimizations
3716 if( OptoPeephole ) {
3717 TracePhase tp(_t_peephole);
3718 PhasePeephole peep( _regalloc, cfg);
3719 peep.do_transform();
3720 print_method(PHASE_PEEPHOLE, 3);
3721 }
3722
3723 // Do late expand if CPU requires this.
3724 if (Matcher::require_postalloc_expand) {
3725 TracePhase tp(_t_postalloc_expand);
3726 cfg.postalloc_expand(_regalloc);
3727 print_method(PHASE_POSTALLOC_EXPAND, 3);
3728 }
3729
3730 #ifdef ASSERT
3731 {
3732 CompilationMemoryStatistic::do_test_allocations();
3733 if (failing()) return;
3734 }
3735 #endif
3736
3737 // Convert Nodes to instruction bits in a buffer
3738 {
3739 TracePhase tp(_t_output);
3740 PhaseOutput output;
3741 output.Output();
3742 if (failing()) return;
3743 output.install();
3744 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3745 }
3746
3747 // He's dead, Jim.
3748 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3749 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3750 }
3751
3752 //------------------------------Final_Reshape_Counts---------------------------
3753 // This class defines counters to help identify when a method
3754 // may/must be executed using hardware with only 24-bit precision.
3755 struct Final_Reshape_Counts : public StackObj {
3756 int _call_count; // count non-inlined 'common' calls
3757 int _float_count; // count float ops requiring 24-bit precision
3758 int _double_count; // count double ops requiring more precision
3759 int _java_call_count; // count non-inlined 'java' calls
3760 int _inner_loop_count; // count loops which need alignment
3761 VectorSet _visited; // Visitation flags
3762 Node_List _tests; // Set of IfNodes & PCTableNodes
3763
3764 Final_Reshape_Counts() :
3765 _call_count(0), _float_count(0), _double_count(0),
3766 _java_call_count(0), _inner_loop_count(0) { }
3767
3768 void inc_call_count () { _call_count ++; }
3769 void inc_float_count () { _float_count ++; }
3770 void inc_double_count() { _double_count++; }
3771 void inc_java_call_count() { _java_call_count++; }
3772 void inc_inner_loop_count() { _inner_loop_count++; }
3773
3774 int get_call_count () const { return _call_count ; }
3775 int get_float_count () const { return _float_count ; }
3776 int get_double_count() const { return _double_count; }
3777 int get_java_call_count() const { return _java_call_count; }
3778 int get_inner_loop_count() const { return _inner_loop_count; }
3779 };
3780
3781 //------------------------------final_graph_reshaping_impl----------------------
3782 // Implement items 1-5 from final_graph_reshaping below.
3783 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3784
3785 if ( n->outcnt() == 0 ) return; // dead node
3786 uint nop = n->Opcode();
3787
3788 // Check for 2-input instruction with "last use" on right input.
3789 // Swap to left input. Implements item (2).
3790 if( n->req() == 3 && // two-input instruction
3791 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3792 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3793 n->in(2)->outcnt() == 1 &&// right use IS a last use
3794 !n->in(2)->is_Con() ) { // right use is not a constant
3795 // Check for commutative opcode
3796 switch( nop ) {
3797 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3798 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3799 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3800 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3801 case Op_AndL: case Op_XorL: case Op_OrL:
3802 case Op_AndI: case Op_XorI: case Op_OrI: {
3803 // Move "last use" input to left by swapping inputs
3804 n->swap_edges(1, 2);
3805 break;
3806 }
3807 default:
3808 break;
3809 }
3810 }
3811
3812 #ifdef ASSERT
3813 if( n->is_Mem() ) {
3814 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3815 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3816 // oop will be recorded in oop map if load crosses safepoint
3817 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3818 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3819 "raw memory operations should have control edge");
3820 }
3821 if (n->is_MemBar()) {
3822 MemBarNode* mb = n->as_MemBar();
3823 if (mb->trailing_store() || mb->trailing_load_store()) {
3824 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3825 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3826 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3827 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3828 } else if (mb->leading()) {
3829 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3830 }
3831 }
3832 #endif
3833 // Count FPU ops and common calls, implements item (3)
3834 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3835 if (!gc_handled) {
3836 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3837 }
3838
3839 // Collect CFG split points
3840 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3841 frc._tests.push(n);
3842 }
3843 }
3844
3845 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3846 if (!UseDivMod) {
3847 return;
3848 }
3849
3850 // Check if "a % b" and "a / b" both exist
3851 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3852 if (d == nullptr) {
3853 return;
3854 }
3855
3856 // Replace them with a fused divmod if supported
3857 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3858 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3859 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3860 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3861 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3862 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3863 // DivMod node so the dependency is not lost.
3864 divmod->add_prec_from(n);
3865 divmod->add_prec_from(d);
3866 d->subsume_by(divmod->div_proj(), this);
3867 n->subsume_by(divmod->mod_proj(), this);
3868 } else {
3869 // Replace "a % b" with "a - ((a / b) * b)"
3870 Node* mult = MulNode::make(d, d->in(2), bt);
3871 Node* sub = SubNode::make(d->in(1), mult, bt);
3872 n->subsume_by(sub, this);
3873 }
3874 }
3875
3876 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3877 switch( nop ) {
3878 // Count all float operations that may use FPU
3879 case Op_AddF:
3880 case Op_SubF:
3881 case Op_MulF:
3882 case Op_DivF:
3883 case Op_NegF:
3884 case Op_ModF:
3885 case Op_ConvI2F:
3886 case Op_ConF:
3887 case Op_CmpF:
3888 case Op_CmpF3:
3889 case Op_StoreF:
3890 case Op_LoadF:
3891 // case Op_ConvL2F: // longs are split into 32-bit halves
3892 frc.inc_float_count();
3893 break;
3894
3895 case Op_ConvF2D:
3896 case Op_ConvD2F:
3897 frc.inc_float_count();
3898 frc.inc_double_count();
3899 break;
3900
3901 // Count all double operations that may use FPU
3902 case Op_AddD:
3903 case Op_SubD:
3904 case Op_MulD:
3905 case Op_DivD:
3906 case Op_NegD:
3907 case Op_ModD:
3908 case Op_ConvI2D:
3909 case Op_ConvD2I:
3910 // case Op_ConvL2D: // handled by leaf call
3911 // case Op_ConvD2L: // handled by leaf call
3912 case Op_ConD:
3913 case Op_CmpD:
3914 case Op_CmpD3:
3915 case Op_StoreD:
3916 case Op_LoadD:
3917 case Op_LoadD_unaligned:
3918 frc.inc_double_count();
3919 break;
3920 case Op_Opaque1: // Remove Opaque Nodes before matching
3921 n->subsume_by(n->in(1), this);
3922 break;
3923 case Op_CallLeafPure: {
3924 // If the pure call is not supported, then lower to a CallLeaf.
3925 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3926 CallNode* call = n->as_Call();
3927 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3928 call->_name, TypeRawPtr::BOTTOM);
3929 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3930 new_call->init_req(TypeFunc::I_O, C->top());
3931 new_call->init_req(TypeFunc::Memory, C->top());
3932 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3933 new_call->init_req(TypeFunc::FramePtr, C->top());
3934 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain_sig()->cnt(); i++) {
3935 new_call->init_req(i, call->in(i));
3936 }
3937 n->subsume_by(new_call, this);
3938 }
3939 frc.inc_call_count();
3940 break;
3941 }
3942 case Op_CallStaticJava:
3943 case Op_CallJava:
3944 case Op_CallDynamicJava:
3945 frc.inc_java_call_count(); // Count java call site;
3946 case Op_CallRuntime:
3947 case Op_CallLeaf:
3948 case Op_CallLeafVector:
3949 case Op_CallLeafNoFP: {
3950 assert (n->is_Call(), "");
3951 CallNode *call = n->as_Call();
3952 // Count call sites where the FP mode bit would have to be flipped.
3953 // Do not count uncommon runtime calls:
3954 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3955 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3956 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3957 frc.inc_call_count(); // Count the call site
3958 } else { // See if uncommon argument is shared
3959 Node *n = call->in(TypeFunc::Parms);
3960 int nop = n->Opcode();
3961 // Clone shared simple arguments to uncommon calls, item (1).
3962 if (n->outcnt() > 1 &&
3963 !n->is_Proj() &&
3964 nop != Op_CreateEx &&
3965 nop != Op_CheckCastPP &&
3966 nop != Op_DecodeN &&
3967 nop != Op_DecodeNKlass &&
3968 !n->is_Mem() &&
3969 !n->is_Phi()) {
3970 Node *x = n->clone();
3971 call->set_req(TypeFunc::Parms, x);
3972 }
3973 }
3974 break;
3975 }
3976 case Op_StoreB:
3977 case Op_StoreC:
3978 case Op_StoreI:
3979 case Op_StoreL:
3980 case Op_StoreLSpecial:
3981 case Op_CompareAndSwapB:
3982 case Op_CompareAndSwapS:
3983 case Op_CompareAndSwapI:
3984 case Op_CompareAndSwapL:
3985 case Op_CompareAndSwapP:
3986 case Op_CompareAndSwapN:
3987 case Op_WeakCompareAndSwapB:
3988 case Op_WeakCompareAndSwapS:
3989 case Op_WeakCompareAndSwapI:
3990 case Op_WeakCompareAndSwapL:
3991 case Op_WeakCompareAndSwapP:
3992 case Op_WeakCompareAndSwapN:
3993 case Op_CompareAndExchangeB:
3994 case Op_CompareAndExchangeS:
3995 case Op_CompareAndExchangeI:
3996 case Op_CompareAndExchangeL:
3997 case Op_CompareAndExchangeP:
3998 case Op_CompareAndExchangeN:
3999 case Op_GetAndAddS:
4000 case Op_GetAndAddB:
4001 case Op_GetAndAddI:
4002 case Op_GetAndAddL:
4003 case Op_GetAndSetS:
4004 case Op_GetAndSetB:
4005 case Op_GetAndSetI:
4006 case Op_GetAndSetL:
4007 case Op_GetAndSetP:
4008 case Op_GetAndSetN:
4009 case Op_StoreP:
4010 case Op_StoreN:
4011 case Op_StoreNKlass:
4012 case Op_LoadB:
4013 case Op_LoadUB:
4014 case Op_LoadUS:
4015 case Op_LoadI:
4016 case Op_LoadKlass:
4017 case Op_LoadNKlass:
4018 case Op_LoadL:
4019 case Op_LoadL_unaligned:
4020 case Op_LoadP:
4021 case Op_LoadN:
4022 case Op_LoadRange:
4023 case Op_LoadS:
4024 break;
4025
4026 case Op_AddP: { // Assert sane base pointers
4027 Node *addp = n->in(AddPNode::Address);
4028 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
4029 #ifdef _LP64
4030 if ((UseCompressedOops || UseCompressedClassPointers) &&
4031 addp->Opcode() == Op_ConP &&
4032 addp == n->in(AddPNode::Base) &&
4033 n->in(AddPNode::Offset)->is_Con()) {
4034 // If the transformation of ConP to ConN+DecodeN is beneficial depends
4035 // on the platform and on the compressed oops mode.
4036 // Use addressing with narrow klass to load with offset on x86.
4037 // Some platforms can use the constant pool to load ConP.
4038 // Do this transformation here since IGVN will convert ConN back to ConP.
4039 const Type* t = addp->bottom_type();
4040 bool is_oop = t->isa_oopptr() != nullptr;
4041 bool is_klass = t->isa_klassptr() != nullptr;
4042
4043 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
4044 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
4045 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
4046 Node* nn = nullptr;
4047
4048 int op = is_oop ? Op_ConN : Op_ConNKlass;
4049
4050 // Look for existing ConN node of the same exact type.
4051 Node* r = root();
4052 uint cnt = r->outcnt();
4053 for (uint i = 0; i < cnt; i++) {
4054 Node* m = r->raw_out(i);
4055 if (m!= nullptr && m->Opcode() == op &&
4056 m->bottom_type()->make_ptr() == t) {
4057 nn = m;
4058 break;
4059 }
4060 }
4061 if (nn != nullptr) {
4062 // Decode a narrow oop to match address
4063 // [R12 + narrow_oop_reg<<3 + offset]
4064 if (is_oop) {
4065 nn = new DecodeNNode(nn, t);
4066 } else {
4067 nn = new DecodeNKlassNode(nn, t);
4068 }
4069 // Check for succeeding AddP which uses the same Base.
4070 // Otherwise we will run into the assertion above when visiting that guy.
4071 for (uint i = 0; i < n->outcnt(); ++i) {
4072 Node *out_i = n->raw_out(i);
4073 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
4074 out_i->set_req(AddPNode::Base, nn);
4075 #ifdef ASSERT
4076 for (uint j = 0; j < out_i->outcnt(); ++j) {
4077 Node *out_j = out_i->raw_out(j);
4078 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
4079 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
4080 }
4081 #endif
4082 }
4083 }
4084 n->set_req(AddPNode::Base, nn);
4085 n->set_req(AddPNode::Address, nn);
4086 if (addp->outcnt() == 0) {
4087 addp->disconnect_inputs(this);
4088 }
4089 }
4090 }
4091 }
4092 #endif
4093 break;
4094 }
4095
4096 case Op_CastPP: {
4097 // Remove CastPP nodes to gain more freedom during scheduling but
4098 // keep the dependency they encode as control or precedence edges
4099 // (if control is set already) on memory operations. Some CastPP
4100 // nodes don't have a control (don't carry a dependency): skip
4101 // those.
4102 if (n->in(0) != nullptr) {
4103 ResourceMark rm;
4104 Unique_Node_List wq;
4105 wq.push(n);
4106 for (uint next = 0; next < wq.size(); ++next) {
4107 Node *m = wq.at(next);
4108 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
4109 Node* use = m->fast_out(i);
4110 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
4111 use->ensure_control_or_add_prec(n->in(0));
4112 } else {
4113 switch(use->Opcode()) {
4114 case Op_AddP:
4115 case Op_DecodeN:
4116 case Op_DecodeNKlass:
4117 case Op_CheckCastPP:
4118 case Op_CastPP:
4119 wq.push(use);
4120 break;
4121 }
4122 }
4123 }
4124 }
4125 }
4126 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
4127 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
4128 Node* in1 = n->in(1);
4129 const Type* t = n->bottom_type();
4130 Node* new_in1 = in1->clone();
4131 new_in1->as_DecodeN()->set_type(t);
4132
4133 if (!Matcher::narrow_oop_use_complex_address()) {
4134 //
4135 // x86, ARM and friends can handle 2 adds in addressing mode
4136 // and Matcher can fold a DecodeN node into address by using
4137 // a narrow oop directly and do implicit null check in address:
4138 //
4139 // [R12 + narrow_oop_reg<<3 + offset]
4140 // NullCheck narrow_oop_reg
4141 //
4142 // On other platforms (Sparc) we have to keep new DecodeN node and
4143 // use it to do implicit null check in address:
4144 //
4145 // decode_not_null narrow_oop_reg, base_reg
4146 // [base_reg + offset]
4147 // NullCheck base_reg
4148 //
4149 // Pin the new DecodeN node to non-null path on these platform (Sparc)
4150 // to keep the information to which null check the new DecodeN node
4151 // corresponds to use it as value in implicit_null_check().
4152 //
4153 new_in1->set_req(0, n->in(0));
4154 }
4155
4156 n->subsume_by(new_in1, this);
4157 if (in1->outcnt() == 0) {
4158 in1->disconnect_inputs(this);
4159 }
4160 } else {
4161 n->subsume_by(n->in(1), this);
4162 if (n->outcnt() == 0) {
4163 n->disconnect_inputs(this);
4164 }
4165 }
4166 break;
4167 }
4168 case Op_CastII: {
4169 n->as_CastII()->remove_range_check_cast(this);
4170 break;
4171 }
4172 #ifdef _LP64
4173 case Op_CmpP:
4174 // Do this transformation here to preserve CmpPNode::sub() and
4175 // other TypePtr related Ideal optimizations (for example, ptr nullness).
4176 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
4177 Node* in1 = n->in(1);
4178 Node* in2 = n->in(2);
4179 if (!in1->is_DecodeNarrowPtr()) {
4180 in2 = in1;
4181 in1 = n->in(2);
4182 }
4183 assert(in1->is_DecodeNarrowPtr(), "sanity");
4184
4185 Node* new_in2 = nullptr;
4186 if (in2->is_DecodeNarrowPtr()) {
4187 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
4188 new_in2 = in2->in(1);
4189 } else if (in2->Opcode() == Op_ConP) {
4190 const Type* t = in2->bottom_type();
4191 if (t == TypePtr::NULL_PTR) {
4192 assert(in1->is_DecodeN(), "compare klass to null?");
4193 // Don't convert CmpP null check into CmpN if compressed
4194 // oops implicit null check is not generated.
4195 // This will allow to generate normal oop implicit null check.
4196 if (Matcher::gen_narrow_oop_implicit_null_checks())
4197 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
4198 //
4199 // This transformation together with CastPP transformation above
4200 // will generated code for implicit null checks for compressed oops.
4201 //
4202 // The original code after Optimize()
4203 //
4204 // LoadN memory, narrow_oop_reg
4205 // decode narrow_oop_reg, base_reg
4206 // CmpP base_reg, nullptr
4207 // CastPP base_reg // NotNull
4208 // Load [base_reg + offset], val_reg
4209 //
4210 // after these transformations will be
4211 //
4212 // LoadN memory, narrow_oop_reg
4213 // CmpN narrow_oop_reg, nullptr
4214 // decode_not_null narrow_oop_reg, base_reg
4215 // Load [base_reg + offset], val_reg
4216 //
4217 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
4218 // since narrow oops can be used in debug info now (see the code in
4219 // final_graph_reshaping_walk()).
4220 //
4221 // At the end the code will be matched to
4222 // on x86:
4223 //
4224 // Load_narrow_oop memory, narrow_oop_reg
4225 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
4226 // NullCheck narrow_oop_reg
4227 //
4228 // and on sparc:
4229 //
4230 // Load_narrow_oop memory, narrow_oop_reg
4231 // decode_not_null narrow_oop_reg, base_reg
4232 // Load [base_reg + offset], val_reg
4233 // NullCheck base_reg
4234 //
4235 } else if (t->isa_oopptr()) {
4236 new_in2 = ConNode::make(t->make_narrowoop());
4237 } else if (t->isa_klassptr()) {
4238 ciKlass* klass = t->is_klassptr()->exact_klass();
4239 if (klass->is_in_encoding_range()) {
4240 new_in2 = ConNode::make(t->make_narrowklass());
4241 }
4242 }
4243 }
4244 if (new_in2 != nullptr) {
4245 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
4246 n->subsume_by(cmpN, this);
4247 if (in1->outcnt() == 0) {
4248 in1->disconnect_inputs(this);
4249 }
4250 if (in2->outcnt() == 0) {
4251 in2->disconnect_inputs(this);
4252 }
4253 }
4254 }
4255 break;
4256
4257 case Op_DecodeN:
4258 case Op_DecodeNKlass:
4259 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4260 // DecodeN could be pinned when it can't be fold into
4261 // an address expression, see the code for Op_CastPP above.
4262 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4263 break;
4264
4265 case Op_EncodeP:
4266 case Op_EncodePKlass: {
4267 Node* in1 = n->in(1);
4268 if (in1->is_DecodeNarrowPtr()) {
4269 n->subsume_by(in1->in(1), this);
4270 } else if (in1->Opcode() == Op_ConP) {
4271 const Type* t = in1->bottom_type();
4272 if (t == TypePtr::NULL_PTR) {
4273 assert(t->isa_oopptr(), "null klass?");
4274 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4275 } else if (t->isa_oopptr()) {
4276 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4277 } else if (t->isa_klassptr()) {
4278 ciKlass* klass = t->is_klassptr()->exact_klass();
4279 if (klass->is_in_encoding_range()) {
4280 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4281 } else {
4282 assert(false, "unencodable klass in ConP -> EncodeP");
4283 C->record_failure("unencodable klass in ConP -> EncodeP");
4284 }
4285 }
4286 }
4287 if (in1->outcnt() == 0) {
4288 in1->disconnect_inputs(this);
4289 }
4290 break;
4291 }
4292
4293 case Op_Proj: {
4294 if (OptimizeStringConcat || IncrementalInline) {
4295 ProjNode* proj = n->as_Proj();
4296 if (proj->_is_io_use) {
4297 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4298 // Separate projections were used for the exception path which
4299 // are normally removed by a late inline. If it wasn't inlined
4300 // then they will hang around and should just be replaced with
4301 // the original one. Merge them.
4302 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4303 if (non_io_proj != nullptr) {
4304 proj->subsume_by(non_io_proj , this);
4305 }
4306 }
4307 }
4308 break;
4309 }
4310
4311 case Op_Phi:
4312 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4313 // The EncodeP optimization may create Phi with the same edges
4314 // for all paths. It is not handled well by Register Allocator.
4315 Node* unique_in = n->in(1);
4316 assert(unique_in != nullptr, "");
4317 uint cnt = n->req();
4318 for (uint i = 2; i < cnt; i++) {
4319 Node* m = n->in(i);
4320 assert(m != nullptr, "");
4321 if (unique_in != m)
4322 unique_in = nullptr;
4323 }
4324 if (unique_in != nullptr) {
4325 n->subsume_by(unique_in, this);
4326 }
4327 }
4328 break;
4329
4330 #endif
4331
4332 case Op_ModI:
4333 handle_div_mod_op(n, T_INT, false);
4334 break;
4335
4336 case Op_ModL:
4337 handle_div_mod_op(n, T_LONG, false);
4338 break;
4339
4340 case Op_UModI:
4341 handle_div_mod_op(n, T_INT, true);
4342 break;
4343
4344 case Op_UModL:
4345 handle_div_mod_op(n, T_LONG, true);
4346 break;
4347
4348 case Op_LoadVector:
4349 case Op_StoreVector:
4350 #ifdef ASSERT
4351 // Add VerifyVectorAlignment node between adr and load / store.
4352 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
4353 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
4354 n->as_StoreVector()->must_verify_alignment();
4355 if (must_verify_alignment) {
4356 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
4357 n->as_StoreVector()->memory_size();
4358 // The memory access should be aligned to the vector width in bytes.
4359 // However, the underlying array is possibly less well aligned, but at least
4360 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
4361 // a loop we can expect at least the following alignment:
4362 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
4363 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
4364 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
4365 // Create mask from alignment. e.g. 0b1000 -> 0b0111
4366 jlong mask = guaranteed_alignment - 1;
4367 Node* mask_con = ConLNode::make(mask);
4368 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
4369 n->set_req(MemNode::Address, va);
4370 }
4371 }
4372 #endif
4373 break;
4374
4375 case Op_LoadVectorGather:
4376 case Op_StoreVectorScatter:
4377 case Op_LoadVectorGatherMasked:
4378 case Op_StoreVectorScatterMasked:
4379 case Op_VectorCmpMasked:
4380 case Op_VectorMaskGen:
4381 case Op_LoadVectorMasked:
4382 case Op_StoreVectorMasked:
4383 break;
4384
4385 case Op_AddReductionVI:
4386 case Op_AddReductionVL:
4387 case Op_AddReductionVF:
4388 case Op_AddReductionVD:
4389 case Op_MulReductionVI:
4390 case Op_MulReductionVL:
4391 case Op_MulReductionVF:
4392 case Op_MulReductionVD:
4393 case Op_MinReductionV:
4394 case Op_MaxReductionV:
4395 case Op_AndReductionV:
4396 case Op_OrReductionV:
4397 case Op_XorReductionV:
4398 break;
4399
4400 case Op_PackB:
4401 case Op_PackS:
4402 case Op_PackI:
4403 case Op_PackF:
4404 case Op_PackL:
4405 case Op_PackD:
4406 if (n->req()-1 > 2) {
4407 // Replace many operand PackNodes with a binary tree for matching
4408 PackNode* p = (PackNode*) n;
4409 Node* btp = p->binary_tree_pack(1, n->req());
4410 n->subsume_by(btp, this);
4411 }
4412 break;
4413 case Op_Loop:
4414 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4415 case Op_CountedLoop:
4416 case Op_LongCountedLoop:
4417 case Op_OuterStripMinedLoop:
4418 if (n->as_Loop()->is_inner_loop()) {
4419 frc.inc_inner_loop_count();
4420 }
4421 n->as_Loop()->verify_strip_mined(0);
4422 break;
4423 case Op_LShiftI:
4424 case Op_RShiftI:
4425 case Op_URShiftI:
4426 case Op_LShiftL:
4427 case Op_RShiftL:
4428 case Op_URShiftL:
4429 if (Matcher::need_masked_shift_count) {
4430 // The cpu's shift instructions don't restrict the count to the
4431 // lower 5/6 bits. We need to do the masking ourselves.
4432 Node* in2 = n->in(2);
4433 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4434 const TypeInt* t = in2->find_int_type();
4435 if (t != nullptr && t->is_con()) {
4436 juint shift = t->get_con();
4437 if (shift > mask) { // Unsigned cmp
4438 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4439 }
4440 } else {
4441 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
4442 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4443 n->set_req(2, shift);
4444 }
4445 }
4446 if (in2->outcnt() == 0) { // Remove dead node
4447 in2->disconnect_inputs(this);
4448 }
4449 }
4450 break;
4451 case Op_MemBarStoreStore:
4452 case Op_MemBarRelease:
4453 // Break the link with AllocateNode: it is no longer useful and
4454 // confuses register allocation.
4455 if (n->req() > MemBarNode::Precedent) {
4456 n->set_req(MemBarNode::Precedent, top());
4457 }
4458 break;
4459 case Op_MemBarAcquire: {
4460 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4461 // At parse time, the trailing MemBarAcquire for a volatile load
4462 // is created with an edge to the load. After optimizations,
4463 // that input may be a chain of Phis. If those phis have no
4464 // other use, then the MemBarAcquire keeps them alive and
4465 // register allocation can be confused.
4466 dead_nodes.push(n->in(MemBarNode::Precedent));
4467 n->set_req(MemBarNode::Precedent, top());
4468 }
4469 break;
4470 }
4471 case Op_Blackhole:
4472 break;
4473 case Op_RangeCheck: {
4474 RangeCheckNode* rc = n->as_RangeCheck();
4475 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4476 n->subsume_by(iff, this);
4477 frc._tests.push(iff);
4478 break;
4479 }
4480 case Op_ConvI2L: {
4481 if (!Matcher::convi2l_type_required) {
4482 // Code generation on some platforms doesn't need accurate
4483 // ConvI2L types. Widening the type can help remove redundant
4484 // address computations.
4485 n->as_Type()->set_type(TypeLong::INT);
4486 ResourceMark rm;
4487 Unique_Node_List wq;
4488 wq.push(n);
4489 for (uint next = 0; next < wq.size(); next++) {
4490 Node *m = wq.at(next);
4491
4492 for(;;) {
4493 // Loop over all nodes with identical inputs edges as m
4494 Node* k = m->find_similar(m->Opcode());
4495 if (k == nullptr) {
4496 break;
4497 }
4498 // Push their uses so we get a chance to remove node made
4499 // redundant
4500 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4501 Node* u = k->fast_out(i);
4502 if (u->Opcode() == Op_LShiftL ||
4503 u->Opcode() == Op_AddL ||
4504 u->Opcode() == Op_SubL ||
4505 u->Opcode() == Op_AddP) {
4506 wq.push(u);
4507 }
4508 }
4509 // Replace all nodes with identical edges as m with m
4510 k->subsume_by(m, this);
4511 }
4512 }
4513 }
4514 break;
4515 }
4516 case Op_CmpUL: {
4517 if (!Matcher::has_match_rule(Op_CmpUL)) {
4518 // No support for unsigned long comparisons
4519 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4520 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4521 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4522 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4523 Node* andl = new AndLNode(orl, remove_sign_mask);
4524 Node* cmp = new CmpLNode(andl, n->in(2));
4525 n->subsume_by(cmp, this);
4526 }
4527 break;
4528 }
4529 #ifdef ASSERT
4530 case Op_InlineType: {
4531 n->dump(-1);
4532 assert(false, "inline type node was not removed");
4533 break;
4534 }
4535 case Op_ConNKlass: {
4536 const TypePtr* tp = n->as_Type()->type()->make_ptr();
4537 ciKlass* klass = tp->is_klassptr()->exact_klass();
4538 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
4539 break;
4540 }
4541 #endif
4542 default:
4543 assert(!n->is_Call(), "");
4544 assert(!n->is_Mem(), "");
4545 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4546 break;
4547 }
4548 }
4549
4550 //------------------------------final_graph_reshaping_walk---------------------
4551 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4552 // requires that the walk visits a node's inputs before visiting the node.
4553 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4554 Unique_Node_List sfpt;
4555
4556 frc._visited.set(root->_idx); // first, mark node as visited
4557 uint cnt = root->req();
4558 Node *n = root;
4559 uint i = 0;
4560 while (true) {
4561 if (i < cnt) {
4562 // Place all non-visited non-null inputs onto stack
4563 Node* m = n->in(i);
4564 ++i;
4565 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
4566 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
4567 // compute worst case interpreter size in case of a deoptimization
4568 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4569
4570 sfpt.push(m);
4571 }
4572 cnt = m->req();
4573 nstack.push(n, i); // put on stack parent and next input's index
4574 n = m;
4575 i = 0;
4576 }
4577 } else {
4578 // Now do post-visit work
4579 final_graph_reshaping_impl(n, frc, dead_nodes);
4580 if (nstack.is_empty())
4581 break; // finished
4582 n = nstack.node(); // Get node from stack
4583 cnt = n->req();
4584 i = nstack.index();
4585 nstack.pop(); // Shift to the next node on stack
4586 }
4587 }
4588
4589 // Skip next transformation if compressed oops are not used.
4590 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4591 (!UseCompressedOops && !UseCompressedClassPointers))
4592 return;
4593
4594 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4595 // It could be done for an uncommon traps or any safepoints/calls
4596 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4597 while (sfpt.size() > 0) {
4598 n = sfpt.pop();
4599 JVMState *jvms = n->as_SafePoint()->jvms();
4600 assert(jvms != nullptr, "sanity");
4601 int start = jvms->debug_start();
4602 int end = n->req();
4603 bool is_uncommon = (n->is_CallStaticJava() &&
4604 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4605 for (int j = start; j < end; j++) {
4606 Node* in = n->in(j);
4607 if (in->is_DecodeNarrowPtr()) {
4608 bool safe_to_skip = true;
4609 if (!is_uncommon ) {
4610 // Is it safe to skip?
4611 for (uint i = 0; i < in->outcnt(); i++) {
4612 Node* u = in->raw_out(i);
4613 if (!u->is_SafePoint() ||
4614 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4615 safe_to_skip = false;
4616 }
4617 }
4618 }
4619 if (safe_to_skip) {
4620 n->set_req(j, in->in(1));
4621 }
4622 if (in->outcnt() == 0) {
4623 in->disconnect_inputs(this);
4624 }
4625 }
4626 }
4627 }
4628 }
4629
4630 //------------------------------final_graph_reshaping--------------------------
4631 // Final Graph Reshaping.
4632 //
4633 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4634 // and not commoned up and forced early. Must come after regular
4635 // optimizations to avoid GVN undoing the cloning. Clone constant
4636 // inputs to Loop Phis; these will be split by the allocator anyways.
4637 // Remove Opaque nodes.
4638 // (2) Move last-uses by commutative operations to the left input to encourage
4639 // Intel update-in-place two-address operations and better register usage
4640 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4641 // calls canonicalizing them back.
4642 // (3) Count the number of double-precision FP ops, single-precision FP ops
4643 // and call sites. On Intel, we can get correct rounding either by
4644 // forcing singles to memory (requires extra stores and loads after each
4645 // FP bytecode) or we can set a rounding mode bit (requires setting and
4646 // clearing the mode bit around call sites). The mode bit is only used
4647 // if the relative frequency of single FP ops to calls is low enough.
4648 // This is a key transform for SPEC mpeg_audio.
4649 // (4) Detect infinite loops; blobs of code reachable from above but not
4650 // below. Several of the Code_Gen algorithms fail on such code shapes,
4651 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4652 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4653 // Detection is by looking for IfNodes where only 1 projection is
4654 // reachable from below or CatchNodes missing some targets.
4655 // (5) Assert for insane oop offsets in debug mode.
4656
4657 bool Compile::final_graph_reshaping() {
4658 // an infinite loop may have been eliminated by the optimizer,
4659 // in which case the graph will be empty.
4660 if (root()->req() == 1) {
4661 // Do not compile method that is only a trivial infinite loop,
4662 // since the content of the loop may have been eliminated.
4663 record_method_not_compilable("trivial infinite loop");
4664 return true;
4665 }
4666
4667 // Expensive nodes have their control input set to prevent the GVN
4668 // from freely commoning them. There's no GVN beyond this point so
4669 // no need to keep the control input. We want the expensive nodes to
4670 // be freely moved to the least frequent code path by gcm.
4671 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4672 for (int i = 0; i < expensive_count(); i++) {
4673 _expensive_nodes.at(i)->set_req(0, nullptr);
4674 }
4675
4676 Final_Reshape_Counts frc;
4677
4678 // Visit everybody reachable!
4679 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4680 Node_Stack nstack(live_nodes() >> 1);
4681 Unique_Node_List dead_nodes;
4682 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4683
4684 // Check for unreachable (from below) code (i.e., infinite loops).
4685 for( uint i = 0; i < frc._tests.size(); i++ ) {
4686 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4687 // Get number of CFG targets.
4688 // Note that PCTables include exception targets after calls.
4689 uint required_outcnt = n->required_outcnt();
4690 if (n->outcnt() != required_outcnt) {
4691 // Check for a few special cases. Rethrow Nodes never take the
4692 // 'fall-thru' path, so expected kids is 1 less.
4693 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4694 if (n->in(0)->in(0)->is_Call()) {
4695 CallNode* call = n->in(0)->in(0)->as_Call();
4696 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4697 required_outcnt--; // Rethrow always has 1 less kid
4698 } else if (call->req() > TypeFunc::Parms &&
4699 call->is_CallDynamicJava()) {
4700 // Check for null receiver. In such case, the optimizer has
4701 // detected that the virtual call will always result in a null
4702 // pointer exception. The fall-through projection of this CatchNode
4703 // will not be populated.
4704 Node* arg0 = call->in(TypeFunc::Parms);
4705 if (arg0->is_Type() &&
4706 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4707 required_outcnt--;
4708 }
4709 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4710 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4711 // Check for illegal array length. In such case, the optimizer has
4712 // detected that the allocation attempt will always result in an
4713 // exception. There is no fall-through projection of this CatchNode .
4714 assert(call->is_CallStaticJava(), "static call expected");
4715 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4716 uint valid_length_test_input = call->req() - 1;
4717 Node* valid_length_test = call->in(valid_length_test_input);
4718 call->del_req(valid_length_test_input);
4719 if (valid_length_test->find_int_con(1) == 0) {
4720 required_outcnt--;
4721 }
4722 dead_nodes.push(valid_length_test);
4723 assert(n->outcnt() == required_outcnt, "malformed control flow");
4724 continue;
4725 }
4726 }
4727 }
4728
4729 // Recheck with a better notion of 'required_outcnt'
4730 if (n->outcnt() != required_outcnt) {
4731 record_method_not_compilable("malformed control flow");
4732 return true; // Not all targets reachable!
4733 }
4734 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4735 CallNode* call = n->in(0)->in(0)->as_Call();
4736 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4737 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4738 assert(call->is_CallStaticJava(), "static call expected");
4739 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4740 uint valid_length_test_input = call->req() - 1;
4741 dead_nodes.push(call->in(valid_length_test_input));
4742 call->del_req(valid_length_test_input); // valid length test useless now
4743 }
4744 }
4745 // Check that I actually visited all kids. Unreached kids
4746 // must be infinite loops.
4747 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4748 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4749 record_method_not_compilable("infinite loop");
4750 return true; // Found unvisited kid; must be unreach
4751 }
4752
4753 // Here so verification code in final_graph_reshaping_walk()
4754 // always see an OuterStripMinedLoopEnd
4755 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4756 IfNode* init_iff = n->as_If();
4757 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4758 n->subsume_by(iff, this);
4759 }
4760 }
4761
4762 while (dead_nodes.size() > 0) {
4763 Node* m = dead_nodes.pop();
4764 if (m->outcnt() == 0 && m != top()) {
4765 for (uint j = 0; j < m->req(); j++) {
4766 Node* in = m->in(j);
4767 if (in != nullptr) {
4768 dead_nodes.push(in);
4769 }
4770 }
4771 m->disconnect_inputs(this);
4772 }
4773 }
4774
4775 set_java_calls(frc.get_java_call_count());
4776 set_inner_loops(frc.get_inner_loop_count());
4777
4778 // No infinite loops, no reason to bail out.
4779 return false;
4780 }
4781
4782 //-----------------------------too_many_traps----------------------------------
4783 // Report if there are too many traps at the current method and bci.
4784 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4785 bool Compile::too_many_traps(ciMethod* method,
4786 int bci,
4787 Deoptimization::DeoptReason reason) {
4788 ciMethodData* md = method->method_data();
4789 if (md->is_empty()) {
4790 // Assume the trap has not occurred, or that it occurred only
4791 // because of a transient condition during start-up in the interpreter.
4792 return false;
4793 }
4794 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4795 if (md->has_trap_at(bci, m, reason) != 0) {
4796 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4797 // Also, if there are multiple reasons, or if there is no per-BCI record,
4798 // assume the worst.
4799 if (log())
4800 log()->elem("observe trap='%s' count='%d'",
4801 Deoptimization::trap_reason_name(reason),
4802 md->trap_count(reason));
4803 return true;
4804 } else {
4805 // Ignore method/bci and see if there have been too many globally.
4806 return too_many_traps(reason, md);
4807 }
4808 }
4809
4810 // Less-accurate variant which does not require a method and bci.
4811 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4812 ciMethodData* logmd) {
4813 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4814 // Too many traps globally.
4815 // Note that we use cumulative trap_count, not just md->trap_count.
4816 if (log()) {
4817 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4818 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4819 Deoptimization::trap_reason_name(reason),
4820 mcount, trap_count(reason));
4821 }
4822 return true;
4823 } else {
4824 // The coast is clear.
4825 return false;
4826 }
4827 }
4828
4829 //--------------------------too_many_recompiles--------------------------------
4830 // Report if there are too many recompiles at the current method and bci.
4831 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4832 // Is not eager to return true, since this will cause the compiler to use
4833 // Action_none for a trap point, to avoid too many recompilations.
4834 bool Compile::too_many_recompiles(ciMethod* method,
4835 int bci,
4836 Deoptimization::DeoptReason reason) {
4837 ciMethodData* md = method->method_data();
4838 if (md->is_empty()) {
4839 // Assume the trap has not occurred, or that it occurred only
4840 // because of a transient condition during start-up in the interpreter.
4841 return false;
4842 }
4843 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4844 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4845 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4846 Deoptimization::DeoptReason per_bc_reason
4847 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4848 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4849 if ((per_bc_reason == Deoptimization::Reason_none
4850 || md->has_trap_at(bci, m, reason) != 0)
4851 // The trap frequency measure we care about is the recompile count:
4852 && md->trap_recompiled_at(bci, m)
4853 && md->overflow_recompile_count() >= bc_cutoff) {
4854 // Do not emit a trap here if it has already caused recompilations.
4855 // Also, if there are multiple reasons, or if there is no per-BCI record,
4856 // assume the worst.
4857 if (log())
4858 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4859 Deoptimization::trap_reason_name(reason),
4860 md->trap_count(reason),
4861 md->overflow_recompile_count());
4862 return true;
4863 } else if (trap_count(reason) != 0
4864 && decompile_count() >= m_cutoff) {
4865 // Too many recompiles globally, and we have seen this sort of trap.
4866 // Use cumulative decompile_count, not just md->decompile_count.
4867 if (log())
4868 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4869 Deoptimization::trap_reason_name(reason),
4870 md->trap_count(reason), trap_count(reason),
4871 md->decompile_count(), decompile_count());
4872 return true;
4873 } else {
4874 // The coast is clear.
4875 return false;
4876 }
4877 }
4878
4879 // Compute when not to trap. Used by matching trap based nodes and
4880 // NullCheck optimization.
4881 void Compile::set_allowed_deopt_reasons() {
4882 _allowed_reasons = 0;
4883 if (is_method_compilation()) {
4884 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4885 assert(rs < BitsPerInt, "recode bit map");
4886 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4887 _allowed_reasons |= nth_bit(rs);
4888 }
4889 }
4890 }
4891 }
4892
4893 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4894 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4895 }
4896
4897 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4898 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4899 }
4900
4901 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4902 if (holder->is_initialized()) {
4903 return false;
4904 }
4905 if (holder->is_being_initialized()) {
4906 if (accessing_method->holder() == holder) {
4907 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4908 // <init>, or a static method. In all those cases, there was an initialization
4909 // barrier on the holder klass passed.
4910 if (accessing_method->is_class_initializer() ||
4911 accessing_method->is_object_constructor() ||
4912 accessing_method->is_static()) {
4913 return false;
4914 }
4915 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4916 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4917 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4918 // child class can become fully initialized while its parent class is still being initialized.
4919 if (accessing_method->is_class_initializer()) {
4920 return false;
4921 }
4922 }
4923 ciMethod* root = method(); // the root method of compilation
4924 if (root != accessing_method) {
4925 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4926 }
4927 }
4928 return true;
4929 }
4930
4931 #ifndef PRODUCT
4932 //------------------------------verify_bidirectional_edges---------------------
4933 // For each input edge to a node (ie - for each Use-Def edge), verify that
4934 // there is a corresponding Def-Use edge.
4935 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4936 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4937 uint stack_size = live_nodes() >> 4;
4938 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4939 if (root_and_safepoints != nullptr) {
4940 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4941 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4942 Node* root_or_safepoint = root_and_safepoints->at(i);
4943 // If the node is a safepoint, let's check if it still has a control input
4944 // Lack of control input signifies that this node was killed by CCP or
4945 // recursively by remove_globally_dead_node and it shouldn't be a starting
4946 // point.
4947 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4948 nstack.push(root_or_safepoint);
4949 }
4950 }
4951 } else {
4952 nstack.push(_root);
4953 }
4954
4955 while (nstack.size() > 0) {
4956 Node* n = nstack.pop();
4957 if (visited.member(n)) {
4958 continue;
4959 }
4960 visited.push(n);
4961
4962 // Walk over all input edges, checking for correspondence
4963 uint length = n->len();
4964 for (uint i = 0; i < length; i++) {
4965 Node* in = n->in(i);
4966 if (in != nullptr && !visited.member(in)) {
4967 nstack.push(in); // Put it on stack
4968 }
4969 if (in != nullptr && !in->is_top()) {
4970 // Count instances of `next`
4971 int cnt = 0;
4972 for (uint idx = 0; idx < in->_outcnt; idx++) {
4973 if (in->_out[idx] == n) {
4974 cnt++;
4975 }
4976 }
4977 assert(cnt > 0, "Failed to find Def-Use edge.");
4978 // Check for duplicate edges
4979 // walk the input array downcounting the input edges to n
4980 for (uint j = 0; j < length; j++) {
4981 if (n->in(j) == in) {
4982 cnt--;
4983 }
4984 }
4985 assert(cnt == 0, "Mismatched edge count.");
4986 } else if (in == nullptr) {
4987 assert(i == 0 || i >= n->req() ||
4988 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4989 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
4990 (n->is_Unlock() && i == (n->req() - 1)) ||
4991 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4992 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
4993 } else {
4994 assert(in->is_top(), "sanity");
4995 // Nothing to check.
4996 }
4997 }
4998 }
4999 }
5000
5001 //------------------------------verify_graph_edges---------------------------
5002 // Walk the Graph and verify that there is a one-to-one correspondence
5003 // between Use-Def edges and Def-Use edges in the graph.
5004 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
5005 if (VerifyGraphEdges) {
5006 Unique_Node_List visited;
5007
5008 // Call graph walk to check edges
5009 verify_bidirectional_edges(visited, root_and_safepoints);
5010 if (no_dead_code) {
5011 // Now make sure that no visited node is used by an unvisited node.
5012 bool dead_nodes = false;
5013 Unique_Node_List checked;
5014 while (visited.size() > 0) {
5015 Node* n = visited.pop();
5016 checked.push(n);
5017 for (uint i = 0; i < n->outcnt(); i++) {
5018 Node* use = n->raw_out(i);
5019 if (checked.member(use)) continue; // already checked
5020 if (visited.member(use)) continue; // already in the graph
5021 if (use->is_Con()) continue; // a dead ConNode is OK
5022 // At this point, we have found a dead node which is DU-reachable.
5023 if (!dead_nodes) {
5024 tty->print_cr("*** Dead nodes reachable via DU edges:");
5025 dead_nodes = true;
5026 }
5027 use->dump(2);
5028 tty->print_cr("---");
5029 checked.push(use); // No repeats; pretend it is now checked.
5030 }
5031 }
5032 assert(!dead_nodes, "using nodes must be reachable from root");
5033 }
5034 }
5035 }
5036 #endif
5037
5038 // The Compile object keeps track of failure reasons separately from the ciEnv.
5039 // This is required because there is not quite a 1-1 relation between the
5040 // ciEnv and its compilation task and the Compile object. Note that one
5041 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
5042 // to backtrack and retry without subsuming loads. Other than this backtracking
5043 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
5044 // by the logic in C2Compiler.
5045 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
5046 if (log() != nullptr) {
5047 log()->elem("failure reason='%s' phase='compile'", reason);
5048 }
5049 if (_failure_reason.get() == nullptr) {
5050 // Record the first failure reason.
5051 _failure_reason.set(reason);
5052 if (CaptureBailoutInformation) {
5053 _first_failure_details = new CompilationFailureInfo(reason);
5054 }
5055 } else {
5056 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
5057 }
5058
5059 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
5060 C->print_method(PHASE_FAILURE, 1);
5061 }
5062 _root = nullptr; // flush the graph, too
5063 }
5064
5065 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
5066 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
5067 _compile(Compile::current()),
5068 _log(nullptr),
5069 _dolog(CITimeVerbose)
5070 {
5071 assert(_compile != nullptr, "sanity check");
5072 assert(id != PhaseTraceId::_t_none, "Don't use none");
5073 if (_dolog) {
5074 _log = _compile->log();
5075 }
5076 if (_log != nullptr) {
5077 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
5078 _log->stamp();
5079 _log->end_head();
5080 }
5081
5082 // Inform memory statistic, if enabled
5083 if (CompilationMemoryStatistic::enabled()) {
5084 CompilationMemoryStatistic::on_phase_start((int)id, name);
5085 }
5086 }
5087
5088 Compile::TracePhase::TracePhase(PhaseTraceId id)
5089 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
5090
5091 Compile::TracePhase::~TracePhase() {
5092
5093 // Inform memory statistic, if enabled
5094 if (CompilationMemoryStatistic::enabled()) {
5095 CompilationMemoryStatistic::on_phase_end();
5096 }
5097
5098 if (_compile->failing_internal()) {
5099 if (_log != nullptr) {
5100 _log->done("phase");
5101 }
5102 return; // timing code, not stressing bailouts.
5103 }
5104 #ifdef ASSERT
5105 if (PrintIdealNodeCount) {
5106 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
5107 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
5108 }
5109
5110 if (VerifyIdealNodeCount) {
5111 _compile->print_missing_nodes();
5112 }
5113 #endif
5114
5115 if (_log != nullptr) {
5116 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
5117 }
5118 }
5119
5120 //----------------------------static_subtype_check-----------------------------
5121 // Shortcut important common cases when superklass is exact:
5122 // (0) superklass is java.lang.Object (can occur in reflective code)
5123 // (1) subklass is already limited to a subtype of superklass => always ok
5124 // (2) subklass does not overlap with superklass => always fail
5125 // (3) superklass has NO subtypes and we can check with a simple compare.
5126 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
5127 if (skip) {
5128 return SSC_full_test; // Let caller generate the general case.
5129 }
5130
5131 if (subk->is_java_subtype_of(superk)) {
5132 return SSC_always_true; // (0) and (1) this test cannot fail
5133 }
5134
5135 if (!subk->maybe_java_subtype_of(superk)) {
5136 return SSC_always_false; // (2) true path dead; no dynamic test needed
5137 }
5138
5139 const Type* superelem = superk;
5140 if (superk->isa_aryklassptr()) {
5141 int ignored;
5142 superelem = superk->is_aryklassptr()->base_element_type(ignored);
5143
5144 // Do not fold the subtype check to an array klass pointer comparison for null-able inline type arrays
5145 // because null-free [LMyValue <: null-able [LMyValue but the klasses are different. Perform a full test.
5146 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
5147 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
5148 return SSC_full_test;
5149 }
5150 }
5151
5152 if (superelem->isa_instklassptr()) {
5153 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
5154 if (!ik->has_subklass()) {
5155 if (!ik->is_final()) {
5156 // Add a dependency if there is a chance of a later subclass.
5157 dependencies()->assert_leaf_type(ik);
5158 }
5159 if (!superk->maybe_java_subtype_of(subk)) {
5160 return SSC_always_false;
5161 }
5162 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5163 }
5164 } else {
5165 // A primitive array type has no subtypes.
5166 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5167 }
5168
5169 return SSC_full_test;
5170 }
5171
5172 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
5173 #ifdef _LP64
5174 // The scaled index operand to AddP must be a clean 64-bit value.
5175 // Java allows a 32-bit int to be incremented to a negative
5176 // value, which appears in a 64-bit register as a large
5177 // positive number. Using that large positive number as an
5178 // operand in pointer arithmetic has bad consequences.
5179 // On the other hand, 32-bit overflow is rare, and the possibility
5180 // can often be excluded, if we annotate the ConvI2L node with
5181 // a type assertion that its value is known to be a small positive
5182 // number. (The prior range check has ensured this.)
5183 // This assertion is used by ConvI2LNode::Ideal.
5184 int index_max = max_jint - 1; // array size is max_jint, index is one less
5185 if (sizetype != nullptr && sizetype->_hi > 0) {
5186 index_max = sizetype->_hi - 1;
5187 }
5188 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
5189 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
5190 #endif
5191 return idx;
5192 }
5193
5194 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
5195 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
5196 if (ctrl != nullptr) {
5197 // Express control dependency by a CastII node with a narrow type.
5198 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
5199 // node from floating above the range check during loop optimizations. Otherwise, the
5200 // ConvI2L node may be eliminated independently of the range check, causing the data path
5201 // to become TOP while the control path is still there (although it's unreachable).
5202 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
5203 value = phase->transform(value);
5204 }
5205 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
5206 return phase->transform(new ConvI2LNode(value, ltype));
5207 }
5208
5209 void Compile::dump_print_inlining() {
5210 inline_printer()->print_on(tty);
5211 }
5212
5213 void Compile::log_late_inline(CallGenerator* cg) {
5214 if (log() != nullptr) {
5215 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5216 cg->unique_id());
5217 JVMState* p = cg->call_node()->jvms();
5218 while (p != nullptr) {
5219 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5220 p = p->caller();
5221 }
5222 log()->tail("late_inline");
5223 }
5224 }
5225
5226 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5227 log_late_inline(cg);
5228 if (log() != nullptr) {
5229 log()->inline_fail(msg);
5230 }
5231 }
5232
5233 void Compile::log_inline_id(CallGenerator* cg) {
5234 if (log() != nullptr) {
5235 // The LogCompilation tool needs a unique way to identify late
5236 // inline call sites. This id must be unique for this call site in
5237 // this compilation. Try to have it unique across compilations as
5238 // well because it can be convenient when grepping through the log
5239 // file.
5240 // Distinguish OSR compilations from others in case CICountOSR is
5241 // on.
5242 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5243 cg->set_unique_id(id);
5244 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5245 }
5246 }
5247
5248 void Compile::log_inline_failure(const char* msg) {
5249 if (C->log() != nullptr) {
5250 C->log()->inline_fail(msg);
5251 }
5252 }
5253
5254
5255 // Dump inlining replay data to the stream.
5256 // Don't change thread state and acquire any locks.
5257 void Compile::dump_inline_data(outputStream* out) {
5258 InlineTree* inl_tree = ilt();
5259 if (inl_tree != nullptr) {
5260 out->print(" inline %d", inl_tree->count());
5261 inl_tree->dump_replay_data(out);
5262 }
5263 }
5264
5265 void Compile::dump_inline_data_reduced(outputStream* out) {
5266 assert(ReplayReduce, "");
5267
5268 InlineTree* inl_tree = ilt();
5269 if (inl_tree == nullptr) {
5270 return;
5271 }
5272 // Enable iterative replay file reduction
5273 // Output "compile" lines for depth 1 subtrees,
5274 // simulating that those trees were compiled
5275 // instead of inlined.
5276 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5277 InlineTree* sub = inl_tree->subtrees().at(i);
5278 if (sub->inline_level() != 1) {
5279 continue;
5280 }
5281
5282 ciMethod* method = sub->method();
5283 int entry_bci = -1;
5284 int comp_level = env()->task()->comp_level();
5285 out->print("compile ");
5286 method->dump_name_as_ascii(out);
5287 out->print(" %d %d", entry_bci, comp_level);
5288 out->print(" inline %d", sub->count());
5289 sub->dump_replay_data(out, -1);
5290 out->cr();
5291 }
5292 }
5293
5294 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5295 if (n1->Opcode() < n2->Opcode()) return -1;
5296 else if (n1->Opcode() > n2->Opcode()) return 1;
5297
5298 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5299 for (uint i = 1; i < n1->req(); i++) {
5300 if (n1->in(i) < n2->in(i)) return -1;
5301 else if (n1->in(i) > n2->in(i)) return 1;
5302 }
5303
5304 return 0;
5305 }
5306
5307 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5308 Node* n1 = *n1p;
5309 Node* n2 = *n2p;
5310
5311 return cmp_expensive_nodes(n1, n2);
5312 }
5313
5314 void Compile::sort_expensive_nodes() {
5315 if (!expensive_nodes_sorted()) {
5316 _expensive_nodes.sort(cmp_expensive_nodes);
5317 }
5318 }
5319
5320 bool Compile::expensive_nodes_sorted() const {
5321 for (int i = 1; i < _expensive_nodes.length(); i++) {
5322 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5323 return false;
5324 }
5325 }
5326 return true;
5327 }
5328
5329 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5330 if (_expensive_nodes.length() == 0) {
5331 return false;
5332 }
5333
5334 assert(OptimizeExpensiveOps, "optimization off?");
5335
5336 // Take this opportunity to remove dead nodes from the list
5337 int j = 0;
5338 for (int i = 0; i < _expensive_nodes.length(); i++) {
5339 Node* n = _expensive_nodes.at(i);
5340 if (!n->is_unreachable(igvn)) {
5341 assert(n->is_expensive(), "should be expensive");
5342 _expensive_nodes.at_put(j, n);
5343 j++;
5344 }
5345 }
5346 _expensive_nodes.trunc_to(j);
5347
5348 // Then sort the list so that similar nodes are next to each other
5349 // and check for at least two nodes of identical kind with same data
5350 // inputs.
5351 sort_expensive_nodes();
5352
5353 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5354 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5355 return true;
5356 }
5357 }
5358
5359 return false;
5360 }
5361
5362 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5363 if (_expensive_nodes.length() == 0) {
5364 return;
5365 }
5366
5367 assert(OptimizeExpensiveOps, "optimization off?");
5368
5369 // Sort to bring similar nodes next to each other and clear the
5370 // control input of nodes for which there's only a single copy.
5371 sort_expensive_nodes();
5372
5373 int j = 0;
5374 int identical = 0;
5375 int i = 0;
5376 bool modified = false;
5377 for (; i < _expensive_nodes.length()-1; i++) {
5378 assert(j <= i, "can't write beyond current index");
5379 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5380 identical++;
5381 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5382 continue;
5383 }
5384 if (identical > 0) {
5385 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5386 identical = 0;
5387 } else {
5388 Node* n = _expensive_nodes.at(i);
5389 igvn.replace_input_of(n, 0, nullptr);
5390 igvn.hash_insert(n);
5391 modified = true;
5392 }
5393 }
5394 if (identical > 0) {
5395 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5396 } else if (_expensive_nodes.length() >= 1) {
5397 Node* n = _expensive_nodes.at(i);
5398 igvn.replace_input_of(n, 0, nullptr);
5399 igvn.hash_insert(n);
5400 modified = true;
5401 }
5402 _expensive_nodes.trunc_to(j);
5403 if (modified) {
5404 igvn.optimize();
5405 }
5406 }
5407
5408 void Compile::add_expensive_node(Node * n) {
5409 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5410 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5411 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5412 if (OptimizeExpensiveOps) {
5413 _expensive_nodes.append(n);
5414 } else {
5415 // Clear control input and let IGVN optimize expensive nodes if
5416 // OptimizeExpensiveOps is off.
5417 n->set_req(0, nullptr);
5418 }
5419 }
5420
5421 /**
5422 * Track coarsened Lock and Unlock nodes.
5423 */
5424
5425 class Lock_List : public Node_List {
5426 uint _origin_cnt;
5427 public:
5428 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5429 uint origin_cnt() const { return _origin_cnt; }
5430 };
5431
5432 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5433 int length = locks.length();
5434 if (length > 0) {
5435 // Have to keep this list until locks elimination during Macro nodes elimination.
5436 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5437 AbstractLockNode* alock = locks.at(0);
5438 BoxLockNode* box = alock->box_node()->as_BoxLock();
5439 for (int i = 0; i < length; i++) {
5440 AbstractLockNode* lock = locks.at(i);
5441 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5442 locks_list->push(lock);
5443 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
5444 if (this_box != box) {
5445 // Locking regions (BoxLock) could be Unbalanced here:
5446 // - its coarsened locks were eliminated in earlier
5447 // macro nodes elimination followed by loop unroll
5448 // - it is OSR locking region (no Lock node)
5449 // Preserve Unbalanced status in such cases.
5450 if (!this_box->is_unbalanced()) {
5451 this_box->set_coarsened();
5452 }
5453 if (!box->is_unbalanced()) {
5454 box->set_coarsened();
5455 }
5456 }
5457 }
5458 _coarsened_locks.append(locks_list);
5459 }
5460 }
5461
5462 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5463 int count = coarsened_count();
5464 for (int i = 0; i < count; i++) {
5465 Node_List* locks_list = _coarsened_locks.at(i);
5466 for (uint j = 0; j < locks_list->size(); j++) {
5467 Node* lock = locks_list->at(j);
5468 assert(lock->is_AbstractLock(), "sanity");
5469 if (!useful.member(lock)) {
5470 locks_list->yank(lock);
5471 }
5472 }
5473 }
5474 }
5475
5476 void Compile::remove_coarsened_lock(Node* n) {
5477 if (n->is_AbstractLock()) {
5478 int count = coarsened_count();
5479 for (int i = 0; i < count; i++) {
5480 Node_List* locks_list = _coarsened_locks.at(i);
5481 locks_list->yank(n);
5482 }
5483 }
5484 }
5485
5486 bool Compile::coarsened_locks_consistent() {
5487 int count = coarsened_count();
5488 for (int i = 0; i < count; i++) {
5489 bool unbalanced = false;
5490 bool modified = false; // track locks kind modifications
5491 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5492 uint size = locks_list->size();
5493 if (size == 0) {
5494 unbalanced = false; // All locks were eliminated - good
5495 } else if (size != locks_list->origin_cnt()) {
5496 unbalanced = true; // Some locks were removed from list
5497 } else {
5498 for (uint j = 0; j < size; j++) {
5499 Node* lock = locks_list->at(j);
5500 // All nodes in group should have the same state (modified or not)
5501 if (!lock->as_AbstractLock()->is_coarsened()) {
5502 if (j == 0) {
5503 // first on list was modified, the rest should be too for consistency
5504 modified = true;
5505 } else if (!modified) {
5506 // this lock was modified but previous locks on the list were not
5507 unbalanced = true;
5508 break;
5509 }
5510 } else if (modified) {
5511 // previous locks on list were modified but not this lock
5512 unbalanced = true;
5513 break;
5514 }
5515 }
5516 }
5517 if (unbalanced) {
5518 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5519 #ifdef ASSERT
5520 if (PrintEliminateLocks) {
5521 tty->print_cr("=== unbalanced coarsened locks ===");
5522 for (uint l = 0; l < size; l++) {
5523 locks_list->at(l)->dump();
5524 }
5525 }
5526 #endif
5527 record_failure(C2Compiler::retry_no_locks_coarsening());
5528 return false;
5529 }
5530 }
5531 return true;
5532 }
5533
5534 // Mark locking regions (identified by BoxLockNode) as unbalanced if
5535 // locks coarsening optimization removed Lock/Unlock nodes from them.
5536 // Such regions become unbalanced because coarsening only removes part
5537 // of Lock/Unlock nodes in region. As result we can't execute other
5538 // locks elimination optimizations which assume all code paths have
5539 // corresponding pair of Lock/Unlock nodes - they are balanced.
5540 void Compile::mark_unbalanced_boxes() const {
5541 int count = coarsened_count();
5542 for (int i = 0; i < count; i++) {
5543 Node_List* locks_list = _coarsened_locks.at(i);
5544 uint size = locks_list->size();
5545 if (size > 0) {
5546 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
5547 BoxLockNode* box = alock->box_node()->as_BoxLock();
5548 if (alock->is_coarsened()) {
5549 // coarsened_locks_consistent(), which is called before this method, verifies
5550 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
5551 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5552 for (uint j = 1; j < size; j++) {
5553 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
5554 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
5555 if (box != this_box) {
5556 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5557 box->set_unbalanced();
5558 this_box->set_unbalanced();
5559 }
5560 }
5561 }
5562 }
5563 }
5564 }
5565
5566 /**
5567 * Remove the speculative part of types and clean up the graph
5568 */
5569 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5570 if (UseTypeSpeculation) {
5571 Unique_Node_List worklist;
5572 worklist.push(root());
5573 int modified = 0;
5574 // Go over all type nodes that carry a speculative type, drop the
5575 // speculative part of the type and enqueue the node for an igvn
5576 // which may optimize it out.
5577 for (uint next = 0; next < worklist.size(); ++next) {
5578 Node *n = worklist.at(next);
5579 if (n->is_Type()) {
5580 TypeNode* tn = n->as_Type();
5581 const Type* t = tn->type();
5582 const Type* t_no_spec = t->remove_speculative();
5583 if (t_no_spec != t) {
5584 bool in_hash = igvn.hash_delete(n);
5585 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5586 tn->set_type(t_no_spec);
5587 igvn.hash_insert(n);
5588 igvn._worklist.push(n); // give it a chance to go away
5589 modified++;
5590 }
5591 }
5592 // Iterate over outs - endless loops is unreachable from below
5593 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5594 Node *m = n->fast_out(i);
5595 if (not_a_node(m)) {
5596 continue;
5597 }
5598 worklist.push(m);
5599 }
5600 }
5601 // Drop the speculative part of all types in the igvn's type table
5602 igvn.remove_speculative_types();
5603 if (modified > 0) {
5604 igvn.optimize();
5605 if (failing()) return;
5606 }
5607 #ifdef ASSERT
5608 // Verify that after the IGVN is over no speculative type has resurfaced
5609 worklist.clear();
5610 worklist.push(root());
5611 for (uint next = 0; next < worklist.size(); ++next) {
5612 Node *n = worklist.at(next);
5613 const Type* t = igvn.type_or_null(n);
5614 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5615 if (n->is_Type()) {
5616 t = n->as_Type()->type();
5617 assert(t == t->remove_speculative(), "no more speculative types");
5618 }
5619 // Iterate over outs - endless loops is unreachable from below
5620 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5621 Node *m = n->fast_out(i);
5622 if (not_a_node(m)) {
5623 continue;
5624 }
5625 worklist.push(m);
5626 }
5627 }
5628 igvn.check_no_speculative_types();
5629 #endif
5630 }
5631 }
5632
5633 // Auxiliary methods to support randomized stressing/fuzzing.
5634
5635 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5636 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5637 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5638 FLAG_SET_ERGO(StressSeed, _stress_seed);
5639 } else {
5640 _stress_seed = StressSeed;
5641 }
5642 if (_log != nullptr) {
5643 _log->elem("stress_test seed='%u'", _stress_seed);
5644 }
5645 }
5646
5647 int Compile::random() {
5648 _stress_seed = os::next_random(_stress_seed);
5649 return static_cast<int>(_stress_seed);
5650 }
5651
5652 // This method can be called the arbitrary number of times, with current count
5653 // as the argument. The logic allows selecting a single candidate from the
5654 // running list of candidates as follows:
5655 // int count = 0;
5656 // Cand* selected = null;
5657 // while(cand = cand->next()) {
5658 // if (randomized_select(++count)) {
5659 // selected = cand;
5660 // }
5661 // }
5662 //
5663 // Including count equalizes the chances any candidate is "selected".
5664 // This is useful when we don't have the complete list of candidates to choose
5665 // from uniformly. In this case, we need to adjust the randomicity of the
5666 // selection, or else we will end up biasing the selection towards the latter
5667 // candidates.
5668 //
5669 // Quick back-envelope calculation shows that for the list of n candidates
5670 // the equal probability for the candidate to persist as "best" can be
5671 // achieved by replacing it with "next" k-th candidate with the probability
5672 // of 1/k. It can be easily shown that by the end of the run, the
5673 // probability for any candidate is converged to 1/n, thus giving the
5674 // uniform distribution among all the candidates.
5675 //
5676 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5677 #define RANDOMIZED_DOMAIN_POW 29
5678 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5679 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5680 bool Compile::randomized_select(int count) {
5681 assert(count > 0, "only positive");
5682 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5683 }
5684
5685 #ifdef ASSERT
5686 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5687 bool Compile::fail_randomly() {
5688 if ((random() % StressBailoutMean) != 0) {
5689 return false;
5690 }
5691 record_failure("StressBailout");
5692 return true;
5693 }
5694
5695 bool Compile::failure_is_artificial() {
5696 return C->failure_reason_is("StressBailout");
5697 }
5698 #endif
5699
5700 CloneMap& Compile::clone_map() { return _clone_map; }
5701 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5702
5703 void NodeCloneInfo::dump_on(outputStream* st) const {
5704 st->print(" {%d:%d} ", idx(), gen());
5705 }
5706
5707 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5708 uint64_t val = value(old->_idx);
5709 NodeCloneInfo cio(val);
5710 assert(val != 0, "old node should be in the map");
5711 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5712 insert(nnn->_idx, cin.get());
5713 #ifndef PRODUCT
5714 if (is_debug()) {
5715 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5716 }
5717 #endif
5718 }
5719
5720 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5721 NodeCloneInfo cio(value(old->_idx));
5722 if (cio.get() == 0) {
5723 cio.set(old->_idx, 0);
5724 insert(old->_idx, cio.get());
5725 #ifndef PRODUCT
5726 if (is_debug()) {
5727 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5728 }
5729 #endif
5730 }
5731 clone(old, nnn, gen);
5732 }
5733
5734 int CloneMap::max_gen() const {
5735 int g = 0;
5736 DictI di(_dict);
5737 for(; di.test(); ++di) {
5738 int t = gen(di._key);
5739 if (g < t) {
5740 g = t;
5741 #ifndef PRODUCT
5742 if (is_debug()) {
5743 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5744 }
5745 #endif
5746 }
5747 }
5748 return g;
5749 }
5750
5751 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5752 uint64_t val = value(key);
5753 if (val != 0) {
5754 NodeCloneInfo ni(val);
5755 ni.dump_on(st);
5756 }
5757 }
5758
5759 void Compile::shuffle_macro_nodes() {
5760 if (_macro_nodes.length() < 2) {
5761 return;
5762 }
5763 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5764 uint j = C->random() % (i + 1);
5765 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5766 }
5767 }
5768
5769 // Move Allocate nodes to the start of the list
5770 void Compile::sort_macro_nodes() {
5771 int count = macro_count();
5772 int allocates = 0;
5773 for (int i = 0; i < count; i++) {
5774 Node* n = macro_node(i);
5775 if (n->is_Allocate()) {
5776 if (i != allocates) {
5777 Node* tmp = macro_node(allocates);
5778 _macro_nodes.at_put(allocates, n);
5779 _macro_nodes.at_put(i, tmp);
5780 }
5781 allocates++;
5782 }
5783 }
5784 }
5785
5786 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5787 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5788 EventCompilerPhase event(UNTIMED);
5789 if (event.should_commit()) {
5790 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5791 }
5792 #ifndef PRODUCT
5793 ResourceMark rm;
5794 stringStream ss;
5795 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5796 int iter = ++_igv_phase_iter[cpt];
5797 if (iter > 1) {
5798 ss.print(" %d", iter);
5799 }
5800 if (n != nullptr) {
5801 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5802 if (n->is_Call()) {
5803 CallNode* call = n->as_Call();
5804 if (call->_name != nullptr) {
5805 // E.g. uncommon traps etc.
5806 ss.print(" - %s", call->_name);
5807 } else if (call->is_CallJava()) {
5808 CallJavaNode* call_java = call->as_CallJava();
5809 if (call_java->method() != nullptr) {
5810 ss.print(" -");
5811 call_java->method()->print_short_name(&ss);
5812 }
5813 }
5814 }
5815 }
5816
5817 const char* name = ss.as_string();
5818 if (should_print_igv(level)) {
5819 _igv_printer->print_graph(name);
5820 }
5821 if (should_print_phase(level)) {
5822 print_phase(name);
5823 }
5824 if (should_print_ideal_phase(cpt)) {
5825 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5826 }
5827 #endif
5828 C->_latest_stage_start_counter.stamp();
5829 }
5830
5831 // Only used from CompileWrapper
5832 void Compile::begin_method() {
5833 #ifndef PRODUCT
5834 if (_method != nullptr && should_print_igv(1)) {
5835 _igv_printer->begin_method();
5836 }
5837 #endif
5838 C->_latest_stage_start_counter.stamp();
5839 }
5840
5841 // Only used from CompileWrapper
5842 void Compile::end_method() {
5843 EventCompilerPhase event(UNTIMED);
5844 if (event.should_commit()) {
5845 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5846 }
5847
5848 #ifndef PRODUCT
5849 if (_method != nullptr && should_print_igv(1)) {
5850 _igv_printer->end_method();
5851 }
5852 #endif
5853 }
5854
5855 #ifndef PRODUCT
5856 bool Compile::should_print_phase(const int level) const {
5857 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5858 _method != nullptr; // Do not print phases for stubs.
5859 }
5860
5861 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5862 return _directive->should_print_ideal_phase(cpt);
5863 }
5864
5865 void Compile::init_igv() {
5866 if (_igv_printer == nullptr) {
5867 _igv_printer = IdealGraphPrinter::printer();
5868 _igv_printer->set_compile(this);
5869 }
5870 }
5871
5872 bool Compile::should_print_igv(const int level) {
5873 PRODUCT_RETURN_(return false;);
5874
5875 if (PrintIdealGraphLevel < 0) { // disabled by the user
5876 return false;
5877 }
5878
5879 bool need = directive()->IGVPrintLevelOption >= level;
5880 if (need) {
5881 Compile::init_igv();
5882 }
5883 return need;
5884 }
5885
5886 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5887 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5888
5889 // Called from debugger. Prints method to the default file with the default phase name.
5890 // This works regardless of any Ideal Graph Visualizer flags set or not.
5891 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5892 void igv_print(void* sp, void* fp, void* pc) {
5893 frame fr(sp, fp, pc);
5894 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5895 }
5896
5897 // Same as igv_print() above but with a specified phase name.
5898 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5899 frame fr(sp, fp, pc);
5900 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5901 }
5902
5903 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5904 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5905 // This works regardless of any Ideal Graph Visualizer flags set or not.
5906 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5907 void igv_print(bool network, void* sp, void* fp, void* pc) {
5908 frame fr(sp, fp, pc);
5909 if (network) {
5910 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5911 } else {
5912 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5913 }
5914 }
5915
5916 // Same as igv_print(bool network, ...) above but with a specified phase name.
5917 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5918 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5919 frame fr(sp, fp, pc);
5920 if (network) {
5921 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5922 } else {
5923 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5924 }
5925 }
5926
5927 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5928 void igv_print_default() {
5929 Compile::current()->print_method(PHASE_DEBUG, 0);
5930 }
5931
5932 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5933 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5934 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5935 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5936 void igv_append(void* sp, void* fp, void* pc) {
5937 frame fr(sp, fp, pc);
5938 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5939 }
5940
5941 // Same as igv_append(...) above but with a specified phase name.
5942 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5943 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5944 frame fr(sp, fp, pc);
5945 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5946 }
5947
5948 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5949 const char* file_name = "custom_debug.xml";
5950 if (_debug_file_printer == nullptr) {
5951 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5952 } else {
5953 _debug_file_printer->update_compiled_method(C->method());
5954 }
5955 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5956 _debug_file_printer->print_graph(phase_name, fr);
5957 }
5958
5959 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5960 ResourceMark rm;
5961 GrowableArray<const Node*> empty_list;
5962 igv_print_graph_to_network(phase_name, empty_list, fr);
5963 }
5964
5965 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5966 if (_debug_network_printer == nullptr) {
5967 _debug_network_printer = new IdealGraphPrinter(C);
5968 } else {
5969 _debug_network_printer->update_compiled_method(C->method());
5970 }
5971 tty->print_cr("Method printed over network stream to IGV");
5972 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5973 }
5974 #endif // !PRODUCT
5975
5976 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5977 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5978 return value;
5979 }
5980 Node* result = nullptr;
5981 if (bt == T_BYTE) {
5982 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5983 result = new RShiftINode(result, phase->intcon(24));
5984 } else if (bt == T_BOOLEAN) {
5985 result = new AndINode(value, phase->intcon(0xFF));
5986 } else if (bt == T_CHAR) {
5987 result = new AndINode(value,phase->intcon(0xFFFF));
5988 } else if (bt == T_FLOAT) {
5989 result = new MoveI2FNode(value);
5990 } else {
5991 assert(bt == T_SHORT, "unexpected narrow type");
5992 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5993 result = new RShiftINode(result, phase->intcon(16));
5994 }
5995 if (transform_res) {
5996 result = phase->transform(result);
5997 }
5998 return result;
5999 }
6000
6001 void Compile::record_method_not_compilable_oom() {
6002 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
6003 }
6004
6005 #ifndef PRODUCT
6006 // Collects all the control inputs from nodes on the worklist and from their data dependencies
6007 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
6008 // Follow non-control edges until we reach CFG nodes
6009 for (uint i = 0; i < worklist.size(); i++) {
6010 const Node* n = worklist.at(i);
6011 for (uint j = 0; j < n->req(); j++) {
6012 Node* in = n->in(j);
6013 if (in == nullptr || in->is_Root()) {
6014 continue;
6015 }
6016 if (in->is_CFG()) {
6017 if (in->is_Call()) {
6018 // The return value of a call is only available if the call did not result in an exception
6019 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
6020 if (control_proj_use->is_Catch()) {
6021 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
6022 candidates.push(fall_through);
6023 continue;
6024 }
6025 }
6026
6027 if (in->is_Multi()) {
6028 // We got here by following data inputs so we should only have one control use
6029 // (no IfNode, etc)
6030 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
6031 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
6032 } else {
6033 candidates.push(in);
6034 }
6035 } else {
6036 worklist.push(in);
6037 }
6038 }
6039 }
6040 }
6041
6042 // Returns the candidate node that is a descendant to all the other candidates
6043 static Node* pick_control(Unique_Node_List& candidates) {
6044 Unique_Node_List worklist;
6045 worklist.copy(candidates);
6046
6047 // Traverse backwards through the CFG
6048 for (uint i = 0; i < worklist.size(); i++) {
6049 const Node* n = worklist.at(i);
6050 if (n->is_Root()) {
6051 continue;
6052 }
6053 for (uint j = 0; j < n->req(); j++) {
6054 // Skip backedge of loops to avoid cycles
6055 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
6056 continue;
6057 }
6058
6059 Node* pred = n->in(j);
6060 if (pred != nullptr && pred != n && pred->is_CFG()) {
6061 worklist.push(pred);
6062 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
6063 candidates.remove(pred);
6064 }
6065 }
6066 }
6067
6068 assert(candidates.size() == 1, "unexpected control flow");
6069 return candidates.at(0);
6070 }
6071
6072 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
6073 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
6074 call->init_req((*pos)++, parm);
6075 const BasicType bt = parm->bottom_type()->basic_type();
6076 if (bt == T_LONG || bt == T_DOUBLE) {
6077 call->init_req((*pos)++, half);
6078 }
6079 }
6080
6081 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
6082 Node* parm0, Node* parm1,
6083 Node* parm2, Node* parm3,
6084 Node* parm4, Node* parm5,
6085 Node* parm6) const {
6086 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
6087 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
6088 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
6089
6090 // find the most suitable control input
6091 Unique_Node_List worklist, candidates;
6092 if (parm0 != nullptr) { worklist.push(parm0);
6093 if (parm1 != nullptr) { worklist.push(parm1);
6094 if (parm2 != nullptr) { worklist.push(parm2);
6095 if (parm3 != nullptr) { worklist.push(parm3);
6096 if (parm4 != nullptr) { worklist.push(parm4);
6097 if (parm5 != nullptr) { worklist.push(parm5);
6098 if (parm6 != nullptr) { worklist.push(parm6);
6099 /* close each nested if ===> */ } } } } } } }
6100 find_candidate_control_inputs(worklist, candidates);
6101 Node* control = nullptr;
6102 if (candidates.size() == 0) {
6103 control = C->start()->proj_out(TypeFunc::Control);
6104 } else {
6105 control = pick_control(candidates);
6106 }
6107
6108 // find all the previous users of the control we picked
6109 GrowableArray<Node*> users_of_control;
6110 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
6111 Node* use = control->fast_out(i);
6112 if (use->is_CFG() && use != control) {
6113 users_of_control.push(use);
6114 }
6115 }
6116
6117 // we do not actually care about IO and memory as it uses neither
6118 call->init_req(TypeFunc::Control, control);
6119 call->init_req(TypeFunc::I_O, top());
6120 call->init_req(TypeFunc::Memory, top());
6121 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
6122 call->init_req(TypeFunc::ReturnAdr, top());
6123
6124 int pos = TypeFunc::Parms;
6125 call->init_req(pos++, str_node);
6126 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
6127 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
6128 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
6129 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
6130 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
6131 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
6132 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
6133 /* close each nested if ===> */ } } } } } } }
6134 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
6135
6136 call = gvn->transform(call);
6137 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
6138
6139 // rewire previous users to have the new call as control instead
6140 PhaseIterGVN* igvn = gvn->is_IterGVN();
6141 for (int i = 0; i < users_of_control.length(); i++) {
6142 Node* use = users_of_control.at(i);
6143 for (uint j = 0; j < use->req(); j++) {
6144 if (use->in(j) == control) {
6145 if (igvn != nullptr) {
6146 igvn->replace_input_of(use, j, call_control_proj);
6147 } else {
6148 gvn->hash_delete(use);
6149 use->set_req(j, call_control_proj);
6150 gvn->hash_insert(use);
6151 }
6152 }
6153 }
6154 }
6155
6156 return call;
6157 }
6158 #endif // !PRODUCT