1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/macroAssembler.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "ci/ciReplay.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "code/aotCodeCache.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compilerDefinitions.hpp"
37 #include "compiler/compilerOracle.hpp"
38 #include "compiler/compiler_globals.hpp"
39 #include "compiler/disassembler.hpp"
40 #include "compiler/oopMap.hpp"
41 #include "gc/shared/barrierSet.hpp"
42 #include "gc/shared/c2/barrierSetC2.hpp"
43 #include "jfr/jfrEvents.hpp"
44 #include "jvm_io.h"
45 #include "memory/allocation.hpp"
46 #include "memory/arena.hpp"
47 #include "memory/resourceArea.hpp"
48 #include "opto/addnode.hpp"
49 #include "opto/block.hpp"
50 #include "opto/c2compiler.hpp"
51 #include "opto/callGenerator.hpp"
52 #include "opto/callnode.hpp"
53 #include "opto/castnode.hpp"
54 #include "opto/cfgnode.hpp"
55 #include "opto/chaitin.hpp"
56 #include "opto/compile.hpp"
57 #include "opto/connode.hpp"
58 #include "opto/convertnode.hpp"
59 #include "opto/divnode.hpp"
60 #include "opto/escape.hpp"
61 #include "opto/idealGraphPrinter.hpp"
62 #include "opto/inlinetypenode.hpp"
63 #include "opto/locknode.hpp"
64 #include "opto/loopnode.hpp"
65 #include "opto/machnode.hpp"
66 #include "opto/macro.hpp"
67 #include "opto/matcher.hpp"
68 #include "opto/mathexactnode.hpp"
69 #include "opto/memnode.hpp"
70 #include "opto/movenode.hpp"
71 #include "opto/mulnode.hpp"
72 #include "opto/narrowptrnode.hpp"
73 #include "opto/node.hpp"
74 #include "opto/opaquenode.hpp"
75 #include "opto/opcodes.hpp"
76 #include "opto/output.hpp"
77 #include "opto/parse.hpp"
78 #include "opto/phaseX.hpp"
79 #include "opto/rootnode.hpp"
80 #include "opto/runtime.hpp"
81 #include "opto/stringopts.hpp"
82 #include "opto/type.hpp"
83 #include "opto/vector.hpp"
84 #include "opto/vectornode.hpp"
85 #include "runtime/globals_extension.hpp"
86 #include "runtime/sharedRuntime.hpp"
87 #include "runtime/signature.hpp"
88 #include "runtime/stubRoutines.hpp"
89 #include "runtime/timer.hpp"
90 #include "utilities/align.hpp"
91 #include "utilities/copy.hpp"
92 #include "utilities/macros.hpp"
93 #include "utilities/resourceHash.hpp"
94
95 // -------------------- Compile::mach_constant_base_node -----------------------
96 // Constant table base node singleton.
97 MachConstantBaseNode* Compile::mach_constant_base_node() {
98 if (_mach_constant_base_node == nullptr) {
99 _mach_constant_base_node = new MachConstantBaseNode();
100 _mach_constant_base_node->add_req(C->root());
101 }
102 return _mach_constant_base_node;
103 }
104
105
106 /// Support for intrinsics.
107
108 // Return the index at which m must be inserted (or already exists).
109 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
110 class IntrinsicDescPair {
111 private:
112 ciMethod* _m;
113 bool _is_virtual;
114 public:
115 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
116 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
117 ciMethod* m= elt->method();
118 ciMethod* key_m = key->_m;
119 if (key_m < m) return -1;
120 else if (key_m > m) return 1;
121 else {
122 bool is_virtual = elt->is_virtual();
123 bool key_virtual = key->_is_virtual;
124 if (key_virtual < is_virtual) return -1;
125 else if (key_virtual > is_virtual) return 1;
126 else return 0;
127 }
128 }
129 };
130 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
131 #ifdef ASSERT
132 for (int i = 1; i < _intrinsics.length(); i++) {
133 CallGenerator* cg1 = _intrinsics.at(i-1);
134 CallGenerator* cg2 = _intrinsics.at(i);
135 assert(cg1->method() != cg2->method()
136 ? cg1->method() < cg2->method()
137 : cg1->is_virtual() < cg2->is_virtual(),
138 "compiler intrinsics list must stay sorted");
139 }
140 #endif
141 IntrinsicDescPair pair(m, is_virtual);
142 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
143 }
144
145 void Compile::register_intrinsic(CallGenerator* cg) {
146 bool found = false;
147 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
148 assert(!found, "registering twice");
149 _intrinsics.insert_before(index, cg);
150 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
151 }
152
153 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
154 assert(m->is_loaded(), "don't try this on unloaded methods");
155 if (_intrinsics.length() > 0) {
156 bool found = false;
157 int index = intrinsic_insertion_index(m, is_virtual, found);
158 if (found) {
159 return _intrinsics.at(index);
160 }
161 }
162 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
163 if (m->intrinsic_id() != vmIntrinsics::_none &&
164 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
165 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
166 if (cg != nullptr) {
167 // Save it for next time:
168 register_intrinsic(cg);
169 return cg;
170 } else {
171 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
172 }
173 }
174 return nullptr;
175 }
176
177 // Compile::make_vm_intrinsic is defined in library_call.cpp.
178
179 #ifndef PRODUCT
180 // statistics gathering...
181
182 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
183 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
184
185 inline int as_int(vmIntrinsics::ID id) {
186 return vmIntrinsics::as_int(id);
187 }
188
189 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
190 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
191 int oflags = _intrinsic_hist_flags[as_int(id)];
192 assert(flags != 0, "what happened?");
193 if (is_virtual) {
194 flags |= _intrinsic_virtual;
195 }
196 bool changed = (flags != oflags);
197 if ((flags & _intrinsic_worked) != 0) {
198 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
199 if (count == 1) {
200 changed = true; // first time
201 }
202 // increment the overall count also:
203 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
204 }
205 if (changed) {
206 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
207 // Something changed about the intrinsic's virtuality.
208 if ((flags & _intrinsic_virtual) != 0) {
209 // This is the first use of this intrinsic as a virtual call.
210 if (oflags != 0) {
211 // We already saw it as a non-virtual, so note both cases.
212 flags |= _intrinsic_both;
213 }
214 } else if ((oflags & _intrinsic_both) == 0) {
215 // This is the first use of this intrinsic as a non-virtual
216 flags |= _intrinsic_both;
217 }
218 }
219 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
220 }
221 // update the overall flags also:
222 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
223 return changed;
224 }
225
226 static char* format_flags(int flags, char* buf) {
227 buf[0] = 0;
228 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
229 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
230 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
231 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
232 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
233 if (buf[0] == 0) strcat(buf, ",");
234 assert(buf[0] == ',', "must be");
235 return &buf[1];
236 }
237
238 void Compile::print_intrinsic_statistics() {
239 char flagsbuf[100];
240 ttyLocker ttyl;
241 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
242 tty->print_cr("Compiler intrinsic usage:");
243 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
244 if (total == 0) total = 1; // avoid div0 in case of no successes
245 #define PRINT_STAT_LINE(name, c, f) \
246 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
247 for (auto id : EnumRange<vmIntrinsicID>{}) {
248 int flags = _intrinsic_hist_flags[as_int(id)];
249 juint count = _intrinsic_hist_count[as_int(id)];
250 if ((flags | count) != 0) {
251 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
252 }
253 }
254 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
255 if (xtty != nullptr) xtty->tail("statistics");
256 }
257
258 void Compile::print_statistics() {
259 { ttyLocker ttyl;
260 if (xtty != nullptr) xtty->head("statistics type='opto'");
261 Parse::print_statistics();
262 PhaseStringOpts::print_statistics();
263 PhaseCCP::print_statistics();
264 PhaseRegAlloc::print_statistics();
265 PhaseOutput::print_statistics();
266 PhasePeephole::print_statistics();
267 PhaseIdealLoop::print_statistics();
268 ConnectionGraph::print_statistics();
269 PhaseMacroExpand::print_statistics();
270 if (xtty != nullptr) xtty->tail("statistics");
271 }
272 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
273 // put this under its own <statistics> element.
274 print_intrinsic_statistics();
275 }
276 }
277 #endif //PRODUCT
278
279 void Compile::gvn_replace_by(Node* n, Node* nn) {
280 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
281 Node* use = n->last_out(i);
282 bool is_in_table = initial_gvn()->hash_delete(use);
283 uint uses_found = 0;
284 for (uint j = 0; j < use->len(); j++) {
285 if (use->in(j) == n) {
286 if (j < use->req())
287 use->set_req(j, nn);
288 else
289 use->set_prec(j, nn);
290 uses_found++;
291 }
292 }
293 if (is_in_table) {
294 // reinsert into table
295 initial_gvn()->hash_find_insert(use);
296 }
297 record_for_igvn(use);
298 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
299 i -= uses_found; // we deleted 1 or more copies of this edge
300 }
301 }
302
303
304 // Identify all nodes that are reachable from below, useful.
305 // Use breadth-first pass that records state in a Unique_Node_List,
306 // recursive traversal is slower.
307 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
308 int estimated_worklist_size = live_nodes();
309 useful.map( estimated_worklist_size, nullptr ); // preallocate space
310
311 // Initialize worklist
312 if (root() != nullptr) { useful.push(root()); }
313 // If 'top' is cached, declare it useful to preserve cached node
314 if (cached_top_node()) { useful.push(cached_top_node()); }
315
316 // Push all useful nodes onto the list, breadthfirst
317 for( uint next = 0; next < useful.size(); ++next ) {
318 assert( next < unique(), "Unique useful nodes < total nodes");
319 Node *n = useful.at(next);
320 uint max = n->len();
321 for( uint i = 0; i < max; ++i ) {
322 Node *m = n->in(i);
323 if (not_a_node(m)) continue;
324 useful.push(m);
325 }
326 }
327 }
328
329 // Update dead_node_list with any missing dead nodes using useful
330 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
331 void Compile::update_dead_node_list(Unique_Node_List &useful) {
332 uint max_idx = unique();
333 VectorSet& useful_node_set = useful.member_set();
334
335 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
336 // If node with index node_idx is not in useful set,
337 // mark it as dead in dead node list.
338 if (!useful_node_set.test(node_idx)) {
339 record_dead_node(node_idx);
340 }
341 }
342 }
343
344 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
345 int shift = 0;
346 for (int i = 0; i < inlines->length(); i++) {
347 CallGenerator* cg = inlines->at(i);
348 if (useful.member(cg->call_node())) {
349 if (shift > 0) {
350 inlines->at_put(i - shift, cg);
351 }
352 } else {
353 shift++; // skip over the dead element
354 }
355 }
356 if (shift > 0) {
357 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
358 }
359 }
360
361 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
362 assert(dead != nullptr && dead->is_Call(), "sanity");
363 int found = 0;
364 for (int i = 0; i < inlines->length(); i++) {
365 if (inlines->at(i)->call_node() == dead) {
366 inlines->remove_at(i);
367 found++;
368 NOT_DEBUG( break; ) // elements are unique, so exit early
369 }
370 }
371 assert(found <= 1, "not unique");
372 }
373
374 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
375 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
376 for (int i = node_list.length() - 1; i >= 0; i--) {
377 N* node = node_list.at(i);
378 if (!useful.member(node)) {
379 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
380 }
381 }
382 }
383
384 void Compile::remove_useless_node(Node* dead) {
385 remove_modified_node(dead);
386
387 // Constant node that has no out-edges and has only one in-edge from
388 // root is usually dead. However, sometimes reshaping walk makes
389 // it reachable by adding use edges. So, we will NOT count Con nodes
390 // as dead to be conservative about the dead node count at any
391 // given time.
392 if (!dead->is_Con()) {
393 record_dead_node(dead->_idx);
394 }
395 if (dead->is_macro()) {
396 remove_macro_node(dead);
397 }
398 if (dead->is_expensive()) {
399 remove_expensive_node(dead);
400 }
401 if (dead->is_OpaqueTemplateAssertionPredicate()) {
402 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
403 }
404 if (dead->is_ParsePredicate()) {
405 remove_parse_predicate(dead->as_ParsePredicate());
406 }
407 if (dead->for_post_loop_opts_igvn()) {
408 remove_from_post_loop_opts_igvn(dead);
409 }
410 if (dead->is_InlineType()) {
411 remove_inline_type(dead);
412 }
413 if (dead->for_merge_stores_igvn()) {
414 remove_from_merge_stores_igvn(dead);
415 }
416 if (dead->is_Call()) {
417 remove_useless_late_inlines( &_late_inlines, dead);
418 remove_useless_late_inlines( &_string_late_inlines, dead);
419 remove_useless_late_inlines( &_boxing_late_inlines, dead);
420 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
421
422 if (dead->is_CallStaticJava()) {
423 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
424 }
425 }
426 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
427 bs->unregister_potential_barrier_node(dead);
428 }
429
430 // Disconnect all useless nodes by disconnecting those at the boundary.
431 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
432 uint next = 0;
433 while (next < useful.size()) {
434 Node *n = useful.at(next++);
435 if (n->is_SafePoint()) {
436 // We're done with a parsing phase. Replaced nodes are not valid
437 // beyond that point.
438 n->as_SafePoint()->delete_replaced_nodes();
439 }
440 // Use raw traversal of out edges since this code removes out edges
441 int max = n->outcnt();
442 for (int j = 0; j < max; ++j) {
443 Node* child = n->raw_out(j);
444 if (!useful.member(child)) {
445 assert(!child->is_top() || child != top(),
446 "If top is cached in Compile object it is in useful list");
447 // Only need to remove this out-edge to the useless node
448 n->raw_del_out(j);
449 --j;
450 --max;
451 if (child->is_data_proj_of_pure_function(n)) {
452 worklist.push(n);
453 }
454 }
455 }
456 if (n->outcnt() == 1 && n->has_special_unique_user()) {
457 assert(useful.member(n->unique_out()), "do not push a useless node");
458 worklist.push(n->unique_out());
459 }
460 if (n->outcnt() == 0) {
461 worklist.push(n);
462 }
463 }
464
465 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
466 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
467 // Remove useless Template Assertion Predicate opaque nodes
468 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
469 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
470 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
471 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
472 #ifdef ASSERT
473 if (_modified_nodes != nullptr) {
474 _modified_nodes->remove_useless_nodes(useful.member_set());
475 }
476 #endif
477 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
478 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
479 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
480 #ifdef ASSERT
481 if (_modified_nodes != nullptr) {
482 _modified_nodes->remove_useless_nodes(useful.member_set());
483 }
484 #endif
485
486 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
487 bs->eliminate_useless_gc_barriers(useful, this);
488 // clean up the late inline lists
489 remove_useless_late_inlines( &_late_inlines, useful);
490 remove_useless_late_inlines( &_string_late_inlines, useful);
491 remove_useless_late_inlines( &_boxing_late_inlines, useful);
492 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
493 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
494 }
495
496 // ============================================================================
497 //------------------------------CompileWrapper---------------------------------
498 class CompileWrapper : public StackObj {
499 Compile *const _compile;
500 public:
501 CompileWrapper(Compile* compile);
502
503 ~CompileWrapper();
504 };
505
506 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
507 // the Compile* pointer is stored in the current ciEnv:
508 ciEnv* env = compile->env();
509 assert(env == ciEnv::current(), "must already be a ciEnv active");
510 assert(env->compiler_data() == nullptr, "compile already active?");
511 env->set_compiler_data(compile);
512 assert(compile == Compile::current(), "sanity");
513
514 compile->set_type_dict(nullptr);
515 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
516 compile->clone_map().set_clone_idx(0);
517 compile->set_type_last_size(0);
518 compile->set_last_tf(nullptr, nullptr);
519 compile->set_indexSet_arena(nullptr);
520 compile->set_indexSet_free_block_list(nullptr);
521 compile->init_type_arena();
522 Type::Initialize(compile);
523 _compile->begin_method();
524 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
525 }
526 CompileWrapper::~CompileWrapper() {
527 // simulate crash during compilation
528 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
529
530 _compile->end_method();
531 _compile->env()->set_compiler_data(nullptr);
532 }
533
534
535 //----------------------------print_compile_messages---------------------------
536 void Compile::print_compile_messages() {
537 #ifndef PRODUCT
538 // Check if recompiling
539 if (!subsume_loads() && PrintOpto) {
540 // Recompiling without allowing machine instructions to subsume loads
541 tty->print_cr("*********************************************************");
542 tty->print_cr("** Bailout: Recompile without subsuming loads **");
543 tty->print_cr("*********************************************************");
544 }
545 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
546 // Recompiling without escape analysis
547 tty->print_cr("*********************************************************");
548 tty->print_cr("** Bailout: Recompile without escape analysis **");
549 tty->print_cr("*********************************************************");
550 }
551 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
552 // Recompiling without iterative escape analysis
553 tty->print_cr("*********************************************************");
554 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
555 tty->print_cr("*********************************************************");
556 }
557 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
558 // Recompiling without reducing allocation merges
559 tty->print_cr("*********************************************************");
560 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
561 tty->print_cr("*********************************************************");
562 }
563 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
564 // Recompiling without boxing elimination
565 tty->print_cr("*********************************************************");
566 tty->print_cr("** Bailout: Recompile without boxing elimination **");
567 tty->print_cr("*********************************************************");
568 }
569 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
570 // Recompiling without locks coarsening
571 tty->print_cr("*********************************************************");
572 tty->print_cr("** Bailout: Recompile without locks coarsening **");
573 tty->print_cr("*********************************************************");
574 }
575 if (env()->break_at_compile()) {
576 // Open the debugger when compiling this method.
577 tty->print("### Breaking when compiling: ");
578 method()->print_short_name();
579 tty->cr();
580 BREAKPOINT;
581 }
582
583 if( PrintOpto ) {
584 if (is_osr_compilation()) {
585 tty->print("[OSR]%3d", _compile_id);
586 } else {
587 tty->print("%3d", _compile_id);
588 }
589 }
590 #endif
591 }
592
593 #ifndef PRODUCT
594 void Compile::print_phase(const char* phase_name) {
595 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name);
596 }
597
598 void Compile::print_ideal_ir(const char* phase_name) {
599 // keep the following output all in one block
600 // This output goes directly to the tty, not the compiler log.
601 // To enable tools to match it up with the compilation activity,
602 // be sure to tag this tty output with the compile ID.
603
604 // Node dumping can cause a safepoint, which can break the tty lock.
605 // Buffer all node dumps, so that all safepoints happen before we lock.
606 ResourceMark rm;
607 stringStream ss;
608
609 if (_output == nullptr) {
610 ss.print_cr("AFTER: %s", phase_name);
611 // Print out all nodes in ascending order of index.
612 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
613 } else {
614 // Dump the node blockwise if we have a scheduling
615 _output->print_scheduling(&ss);
616 }
617
618 // Check that the lock is not broken by a safepoint.
619 NoSafepointVerifier nsv;
620 ttyLocker ttyl;
621 if (xtty != nullptr) {
622 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
623 compile_id(),
624 is_osr_compilation() ? " compile_kind='osr'" : "",
625 phase_name);
626 }
627
628 tty->print("%s", ss.as_string());
629
630 if (xtty != nullptr) {
631 xtty->tail("ideal");
632 }
633 }
634 #endif
635
636 // ============================================================================
637 //------------------------------Compile standard-------------------------------
638
639 // Compile a method. entry_bci is -1 for normal compilations and indicates
640 // the continuation bci for on stack replacement.
641
642
643 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
644 Options options, DirectiveSet* directive)
645 : Phase(Compiler),
646 _compile_id(ci_env->compile_id()),
647 _options(options),
648 _method(target),
649 _entry_bci(osr_bci),
650 _ilt(nullptr),
651 _stub_function(nullptr),
652 _stub_name(nullptr),
653 _stub_id(-1),
654 _stub_entry_point(nullptr),
655 _max_node_limit(MaxNodeLimit),
656 _post_loop_opts_phase(false),
657 _merge_stores_phase(false),
658 _allow_macro_nodes(true),
659 _inlining_progress(false),
660 _inlining_incrementally(false),
661 _do_cleanup(false),
662 _has_reserved_stack_access(target->has_reserved_stack_access()),
663 _has_circular_inline_type(false),
664 #ifndef PRODUCT
665 _igv_idx(0),
666 _trace_opto_output(directive->TraceOptoOutputOption),
667 #endif
668 _has_method_handle_invokes(false),
669 _clinit_barrier_on_entry(false),
670 _stress_seed(0),
671 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
672 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
673 _env(ci_env),
674 _directive(directive),
675 _log(ci_env->log()),
676 _first_failure_details(nullptr),
677 _intrinsics(comp_arena(), 0, 0, nullptr),
678 _macro_nodes(comp_arena(), 8, 0, nullptr),
679 _parse_predicates(comp_arena(), 8, 0, nullptr),
680 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
681 _expensive_nodes(comp_arena(), 8, 0, nullptr),
682 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
683 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
684 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
685 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
686 _coarsened_locks(comp_arena(), 8, 0, nullptr),
687 _congraph(nullptr),
688 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
689 _unique(0),
690 _dead_node_count(0),
691 _dead_node_list(comp_arena()),
692 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
693 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
694 _node_arena(&_node_arena_one),
695 _mach_constant_base_node(nullptr),
696 _Compile_types(mtCompiler, Arena::Tag::tag_type),
697 _initial_gvn(nullptr),
698 _igvn_worklist(nullptr),
699 _types(nullptr),
700 _node_hash(nullptr),
701 _late_inlines(comp_arena(), 2, 0, nullptr),
702 _string_late_inlines(comp_arena(), 2, 0, nullptr),
703 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
704 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
705 _late_inlines_pos(0),
706 _number_of_mh_late_inlines(0),
707 _oom(false),
708 _replay_inline_data(nullptr),
709 _inline_printer(this),
710 _java_calls(0),
711 _inner_loops(0),
712 _interpreter_frame_size(0),
713 _output(nullptr)
714 #ifndef PRODUCT
715 ,
716 _in_dump_cnt(0)
717 #endif
718 {
719 C = this;
720 CompileWrapper cw(this);
721
722 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
723 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
724
725 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
726 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
727 // We can always print a disassembly, either abstract (hex dump) or
728 // with the help of a suitable hsdis library. Thus, we should not
729 // couple print_assembly and print_opto_assembly controls.
730 // But: always print opto and regular assembly on compile command 'print'.
731 bool print_assembly = directive->PrintAssemblyOption;
732 set_print_assembly(print_opto_assembly || print_assembly);
733 #else
734 set_print_assembly(false); // must initialize.
735 #endif
736
737 #ifndef PRODUCT
738 set_parsed_irreducible_loop(false);
739 #endif
740
741 if (directive->ReplayInlineOption) {
742 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
743 }
744 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
745 set_print_intrinsics(directive->PrintIntrinsicsOption);
746 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
747
748 if (ProfileTraps) {
749 // Make sure the method being compiled gets its own MDO,
750 // so we can at least track the decompile_count().
751 method()->ensure_method_data();
752 }
753
754 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
755 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout ||
756 StressLoopPeeling) {
757 initialize_stress_seed(directive);
758 }
759
760 Init(/*do_aliasing=*/ true);
761
762 print_compile_messages();
763
764 _ilt = InlineTree::build_inline_tree_root();
765
766 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
767 assert(num_alias_types() >= AliasIdxRaw, "");
768
769 #define MINIMUM_NODE_HASH 1023
770
771 // GVN that will be run immediately on new nodes
772 uint estimated_size = method()->code_size()*4+64;
773 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
774 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
775 _types = new (comp_arena()) Type_Array(comp_arena());
776 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
777 PhaseGVN gvn;
778 set_initial_gvn(&gvn);
779
780 { // Scope for timing the parser
781 TracePhase tp(_t_parser);
782
783 // Put top into the hash table ASAP.
784 initial_gvn()->transform(top());
785
786 // Set up tf(), start(), and find a CallGenerator.
787 CallGenerator* cg = nullptr;
788 if (is_osr_compilation()) {
789 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
790 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
791 initial_gvn()->set_type_bottom(s);
792 verify_start(s);
793 cg = CallGenerator::for_osr(method(), entry_bci());
794 } else {
795 // Normal case.
796 init_tf(TypeFunc::make(method()));
797 StartNode* s = new StartNode(root(), tf()->domain_cc());
798 initial_gvn()->set_type_bottom(s);
799 verify_start(s);
800 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
801 // With java.lang.ref.reference.get() we must go through the
802 // intrinsic - even when get() is the root
803 // method of the compile - so that, if necessary, the value in
804 // the referent field of the reference object gets recorded by
805 // the pre-barrier code.
806 cg = find_intrinsic(method(), false);
807 }
808 if (cg == nullptr) {
809 float past_uses = method()->interpreter_invocation_count();
810 float expected_uses = past_uses;
811 cg = CallGenerator::for_inline(method(), expected_uses);
812 }
813 }
814 if (failing()) return;
815 if (cg == nullptr) {
816 const char* reason = InlineTree::check_can_parse(method());
817 assert(reason != nullptr, "expect reason for parse failure");
818 stringStream ss;
819 ss.print("cannot parse method: %s", reason);
820 record_method_not_compilable(ss.as_string());
821 return;
822 }
823
824 gvn.set_type(root(), root()->bottom_type());
825
826 JVMState* jvms = build_start_state(start(), tf());
827 if ((jvms = cg->generate(jvms)) == nullptr) {
828 assert(failure_reason() != nullptr, "expect reason for parse failure");
829 stringStream ss;
830 ss.print("method parse failed: %s", failure_reason());
831 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
832 return;
833 }
834 GraphKit kit(jvms);
835
836 if (!kit.stopped()) {
837 // Accept return values, and transfer control we know not where.
838 // This is done by a special, unique ReturnNode bound to root.
839 return_values(kit.jvms());
840 }
841
842 if (kit.has_exceptions()) {
843 // Any exceptions that escape from this call must be rethrown
844 // to whatever caller is dynamically above us on the stack.
845 // This is done by a special, unique RethrowNode bound to root.
846 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
847 }
848
849 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
850
851 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
852 inline_string_calls(true);
853 }
854
855 if (failing()) return;
856
857 // Remove clutter produced by parsing.
858 if (!failing()) {
859 ResourceMark rm;
860 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
861 }
862 }
863
864 // Note: Large methods are capped off in do_one_bytecode().
865 if (failing()) return;
866
867 // After parsing, node notes are no longer automagic.
868 // They must be propagated by register_new_node_with_optimizer(),
869 // clone(), or the like.
870 set_default_node_notes(nullptr);
871
872 #ifndef PRODUCT
873 if (should_print_igv(1)) {
874 _igv_printer->print_inlining();
875 }
876 #endif
877
878 if (failing()) return;
879 NOT_PRODUCT( verify_graph_edges(); )
880
881 // Now optimize
882 Optimize();
883 if (failing()) return;
884 NOT_PRODUCT( verify_graph_edges(); )
885
886 #ifndef PRODUCT
887 if (should_print_ideal()) {
888 print_ideal_ir("print_ideal");
889 }
890 #endif
891
892 #ifdef ASSERT
893 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
894 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
895 #endif
896
897 // Dump compilation data to replay it.
898 if (directive->DumpReplayOption) {
899 env()->dump_replay_data(_compile_id);
900 }
901 if (directive->DumpInlineOption && (ilt() != nullptr)) {
902 env()->dump_inline_data(_compile_id);
903 }
904
905 // Now that we know the size of all the monitors we can add a fixed slot
906 // for the original deopt pc.
907 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
908 if (needs_stack_repair()) {
909 // One extra slot for the special stack increment value
910 next_slot += 2;
911 }
912 // TODO 8284443 Only reserve extra slot if needed
913 if (InlineTypeReturnedAsFields) {
914 // One extra slot to hold the null marker for a nullable
915 // inline type return if we run out of registers.
916 next_slot += 2;
917 }
918 set_fixed_slots(next_slot);
919
920 // Compute when to use implicit null checks. Used by matching trap based
921 // nodes and NullCheck optimization.
922 set_allowed_deopt_reasons();
923
924 // Now generate code
925 Code_Gen();
926 }
927
928 //------------------------------Compile----------------------------------------
929 // Compile a runtime stub
930 Compile::Compile(ciEnv* ci_env,
931 TypeFunc_generator generator,
932 address stub_function,
933 const char* stub_name,
934 int stub_id,
935 int is_fancy_jump,
936 bool pass_tls,
937 bool return_pc,
938 DirectiveSet* directive)
939 : Phase(Compiler),
940 _compile_id(0),
941 _options(Options::for_runtime_stub()),
942 _method(nullptr),
943 _entry_bci(InvocationEntryBci),
944 _stub_function(stub_function),
945 _stub_name(stub_name),
946 _stub_id(stub_id),
947 _stub_entry_point(nullptr),
948 _max_node_limit(MaxNodeLimit),
949 _post_loop_opts_phase(false),
950 _merge_stores_phase(false),
951 _allow_macro_nodes(true),
952 _inlining_progress(false),
953 _inlining_incrementally(false),
954 _has_reserved_stack_access(false),
955 _has_circular_inline_type(false),
956 #ifndef PRODUCT
957 _igv_idx(0),
958 _trace_opto_output(directive->TraceOptoOutputOption),
959 #endif
960 _has_method_handle_invokes(false),
961 _clinit_barrier_on_entry(false),
962 _stress_seed(0),
963 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
964 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
965 _env(ci_env),
966 _directive(directive),
967 _log(ci_env->log()),
968 _first_failure_details(nullptr),
969 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
970 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
971 _congraph(nullptr),
972 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
973 _unique(0),
974 _dead_node_count(0),
975 _dead_node_list(comp_arena()),
976 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
977 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
978 _node_arena(&_node_arena_one),
979 _mach_constant_base_node(nullptr),
980 _Compile_types(mtCompiler, Arena::Tag::tag_type),
981 _initial_gvn(nullptr),
982 _igvn_worklist(nullptr),
983 _types(nullptr),
984 _node_hash(nullptr),
985 _number_of_mh_late_inlines(0),
986 _oom(false),
987 _replay_inline_data(nullptr),
988 _inline_printer(this),
989 _java_calls(0),
990 _inner_loops(0),
991 _interpreter_frame_size(0),
992 _output(nullptr),
993 #ifndef PRODUCT
994 _in_dump_cnt(0),
995 #endif
996 _allowed_reasons(0) {
997 C = this;
998
999 // try to reuse an existing stub
1000 {
1001 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, _stub_id, stub_name);
1002 if (blob != nullptr) {
1003 RuntimeStub* rs = blob->as_runtime_stub();
1004 _stub_entry_point = rs->entry_point();
1005 return;
1006 }
1007 }
1008
1009 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
1010 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
1011
1012 #ifndef PRODUCT
1013 set_print_assembly(PrintFrameConverterAssembly);
1014 set_parsed_irreducible_loop(false);
1015 #else
1016 set_print_assembly(false); // Must initialize.
1017 #endif
1018 set_has_irreducible_loop(false); // no loops
1019
1020 CompileWrapper cw(this);
1021 Init(/*do_aliasing=*/ false);
1022 init_tf((*generator)());
1023
1024 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1025 _types = new (comp_arena()) Type_Array(comp_arena());
1026 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1027
1028 if (StressLCM || StressGCM || StressBailout) {
1029 initialize_stress_seed(directive);
1030 }
1031
1032 {
1033 PhaseGVN gvn;
1034 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1035 gvn.transform(top());
1036
1037 GraphKit kit;
1038 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1039 }
1040
1041 NOT_PRODUCT( verify_graph_edges(); )
1042
1043 Code_Gen();
1044 }
1045
1046 Compile::~Compile() {
1047 delete _first_failure_details;
1048 };
1049
1050 //------------------------------Init-------------------------------------------
1051 // Prepare for a single compilation
1052 void Compile::Init(bool aliasing) {
1053 _do_aliasing = aliasing;
1054 _unique = 0;
1055 _regalloc = nullptr;
1056
1057 _tf = nullptr; // filled in later
1058 _top = nullptr; // cached later
1059 _matcher = nullptr; // filled in later
1060 _cfg = nullptr; // filled in later
1061
1062 _node_note_array = nullptr;
1063 _default_node_notes = nullptr;
1064 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1065
1066 _immutable_memory = nullptr; // filled in at first inquiry
1067
1068 #ifdef ASSERT
1069 _phase_optimize_finished = false;
1070 _phase_verify_ideal_loop = false;
1071 _exception_backedge = false;
1072 _type_verify = nullptr;
1073 #endif
1074
1075 // Globally visible Nodes
1076 // First set TOP to null to give safe behavior during creation of RootNode
1077 set_cached_top_node(nullptr);
1078 set_root(new RootNode());
1079 // Now that you have a Root to point to, create the real TOP
1080 set_cached_top_node( new ConNode(Type::TOP) );
1081 set_recent_alloc(nullptr, nullptr);
1082
1083 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1084 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1085 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1086 env()->set_dependencies(new Dependencies(env()));
1087
1088 _fixed_slots = 0;
1089 set_has_split_ifs(false);
1090 set_has_loops(false); // first approximation
1091 set_has_stringbuilder(false);
1092 set_has_boxed_value(false);
1093 _trap_can_recompile = false; // no traps emitted yet
1094 _major_progress = true; // start out assuming good things will happen
1095 set_has_unsafe_access(false);
1096 set_max_vector_size(0);
1097 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1098 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1099 set_decompile_count(0);
1100
1101 #ifndef PRODUCT
1102 _phase_counter = 0;
1103 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1104 #endif
1105
1106 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1107 _loop_opts_cnt = LoopOptsCount;
1108 _has_flat_accesses = false;
1109 _flat_accesses_share_alias = true;
1110 _scalarize_in_safepoints = false;
1111
1112 set_do_inlining(Inline);
1113 set_max_inline_size(MaxInlineSize);
1114 set_freq_inline_size(FreqInlineSize);
1115 set_do_scheduling(OptoScheduling);
1116
1117 set_do_vector_loop(false);
1118 set_has_monitors(false);
1119 set_has_scoped_access(false);
1120
1121 if (AllowVectorizeOnDemand) {
1122 if (has_method() && _directive->VectorizeOption) {
1123 set_do_vector_loop(true);
1124 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());})
1125 } else if (has_method() && method()->name() != nullptr &&
1126 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1127 set_do_vector_loop(true);
1128 }
1129 }
1130 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1131 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());})
1132
1133 _max_node_limit = _directive->MaxNodeLimitOption;
1134
1135 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1136 set_clinit_barrier_on_entry(true);
1137 }
1138 if (debug_info()->recording_non_safepoints()) {
1139 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1140 (comp_arena(), 8, 0, nullptr));
1141 set_default_node_notes(Node_Notes::make(this));
1142 }
1143
1144 const int grow_ats = 16;
1145 _max_alias_types = grow_ats;
1146 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1147 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1148 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1149 {
1150 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1151 }
1152 // Initialize the first few types.
1153 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1154 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1155 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1156 _num_alias_types = AliasIdxRaw+1;
1157 // Zero out the alias type cache.
1158 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1159 // A null adr_type hits in the cache right away. Preload the right answer.
1160 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1161 }
1162
1163 #ifdef ASSERT
1164 // Verify that the current StartNode is valid.
1165 void Compile::verify_start(StartNode* s) const {
1166 assert(failing_internal() || s == start(), "should be StartNode");
1167 }
1168 #endif
1169
1170 /**
1171 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1172 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1173 * the ideal graph.
1174 */
1175 StartNode* Compile::start() const {
1176 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1177 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1178 Node* start = root()->fast_out(i);
1179 if (start->is_Start()) {
1180 return start->as_Start();
1181 }
1182 }
1183 fatal("Did not find Start node!");
1184 return nullptr;
1185 }
1186
1187 //-------------------------------immutable_memory-------------------------------------
1188 // Access immutable memory
1189 Node* Compile::immutable_memory() {
1190 if (_immutable_memory != nullptr) {
1191 return _immutable_memory;
1192 }
1193 StartNode* s = start();
1194 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1195 Node *p = s->fast_out(i);
1196 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1197 _immutable_memory = p;
1198 return _immutable_memory;
1199 }
1200 }
1201 ShouldNotReachHere();
1202 return nullptr;
1203 }
1204
1205 //----------------------set_cached_top_node------------------------------------
1206 // Install the cached top node, and make sure Node::is_top works correctly.
1207 void Compile::set_cached_top_node(Node* tn) {
1208 if (tn != nullptr) verify_top(tn);
1209 Node* old_top = _top;
1210 _top = tn;
1211 // Calling Node::setup_is_top allows the nodes the chance to adjust
1212 // their _out arrays.
1213 if (_top != nullptr) _top->setup_is_top();
1214 if (old_top != nullptr) old_top->setup_is_top();
1215 assert(_top == nullptr || top()->is_top(), "");
1216 }
1217
1218 #ifdef ASSERT
1219 uint Compile::count_live_nodes_by_graph_walk() {
1220 Unique_Node_List useful(comp_arena());
1221 // Get useful node list by walking the graph.
1222 identify_useful_nodes(useful);
1223 return useful.size();
1224 }
1225
1226 void Compile::print_missing_nodes() {
1227
1228 // Return if CompileLog is null and PrintIdealNodeCount is false.
1229 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1230 return;
1231 }
1232
1233 // This is an expensive function. It is executed only when the user
1234 // specifies VerifyIdealNodeCount option or otherwise knows the
1235 // additional work that needs to be done to identify reachable nodes
1236 // by walking the flow graph and find the missing ones using
1237 // _dead_node_list.
1238
1239 Unique_Node_List useful(comp_arena());
1240 // Get useful node list by walking the graph.
1241 identify_useful_nodes(useful);
1242
1243 uint l_nodes = C->live_nodes();
1244 uint l_nodes_by_walk = useful.size();
1245
1246 if (l_nodes != l_nodes_by_walk) {
1247 if (_log != nullptr) {
1248 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1249 _log->stamp();
1250 _log->end_head();
1251 }
1252 VectorSet& useful_member_set = useful.member_set();
1253 int last_idx = l_nodes_by_walk;
1254 for (int i = 0; i < last_idx; i++) {
1255 if (useful_member_set.test(i)) {
1256 if (_dead_node_list.test(i)) {
1257 if (_log != nullptr) {
1258 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1259 }
1260 if (PrintIdealNodeCount) {
1261 // Print the log message to tty
1262 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1263 useful.at(i)->dump();
1264 }
1265 }
1266 }
1267 else if (! _dead_node_list.test(i)) {
1268 if (_log != nullptr) {
1269 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1270 }
1271 if (PrintIdealNodeCount) {
1272 // Print the log message to tty
1273 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1274 }
1275 }
1276 }
1277 if (_log != nullptr) {
1278 _log->tail("mismatched_nodes");
1279 }
1280 }
1281 }
1282 void Compile::record_modified_node(Node* n) {
1283 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1284 _modified_nodes->push(n);
1285 }
1286 }
1287
1288 void Compile::remove_modified_node(Node* n) {
1289 if (_modified_nodes != nullptr) {
1290 _modified_nodes->remove(n);
1291 }
1292 }
1293 #endif
1294
1295 #ifndef PRODUCT
1296 void Compile::verify_top(Node* tn) const {
1297 if (tn != nullptr) {
1298 assert(tn->is_Con(), "top node must be a constant");
1299 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1300 assert(tn->in(0) != nullptr, "must have live top node");
1301 }
1302 }
1303 #endif
1304
1305
1306 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1307
1308 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1309 guarantee(arr != nullptr, "");
1310 int num_blocks = arr->length();
1311 if (grow_by < num_blocks) grow_by = num_blocks;
1312 int num_notes = grow_by * _node_notes_block_size;
1313 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1314 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1315 while (num_notes > 0) {
1316 arr->append(notes);
1317 notes += _node_notes_block_size;
1318 num_notes -= _node_notes_block_size;
1319 }
1320 assert(num_notes == 0, "exact multiple, please");
1321 }
1322
1323 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1324 if (source == nullptr || dest == nullptr) return false;
1325
1326 if (dest->is_Con())
1327 return false; // Do not push debug info onto constants.
1328
1329 #ifdef ASSERT
1330 // Leave a bread crumb trail pointing to the original node:
1331 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1332 dest->set_debug_orig(source);
1333 }
1334 #endif
1335
1336 if (node_note_array() == nullptr)
1337 return false; // Not collecting any notes now.
1338
1339 // This is a copy onto a pre-existing node, which may already have notes.
1340 // If both nodes have notes, do not overwrite any pre-existing notes.
1341 Node_Notes* source_notes = node_notes_at(source->_idx);
1342 if (source_notes == nullptr || source_notes->is_clear()) return false;
1343 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1344 if (dest_notes == nullptr || dest_notes->is_clear()) {
1345 return set_node_notes_at(dest->_idx, source_notes);
1346 }
1347
1348 Node_Notes merged_notes = (*source_notes);
1349 // The order of operations here ensures that dest notes will win...
1350 merged_notes.update_from(dest_notes);
1351 return set_node_notes_at(dest->_idx, &merged_notes);
1352 }
1353
1354
1355 //--------------------------allow_range_check_smearing-------------------------
1356 // Gating condition for coalescing similar range checks.
1357 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1358 // single covering check that is at least as strong as any of them.
1359 // If the optimization succeeds, the simplified (strengthened) range check
1360 // will always succeed. If it fails, we will deopt, and then give up
1361 // on the optimization.
1362 bool Compile::allow_range_check_smearing() const {
1363 // If this method has already thrown a range-check,
1364 // assume it was because we already tried range smearing
1365 // and it failed.
1366 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1367 return !already_trapped;
1368 }
1369
1370
1371 //------------------------------flatten_alias_type-----------------------------
1372 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1373 assert(do_aliasing(), "Aliasing should be enabled");
1374 int offset = tj->offset();
1375 TypePtr::PTR ptr = tj->ptr();
1376
1377 // Known instance (scalarizable allocation) alias only with itself.
1378 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1379 tj->is_oopptr()->is_known_instance();
1380
1381 // Process weird unsafe references.
1382 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1383 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1384 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1385 tj = TypeOopPtr::BOTTOM;
1386 ptr = tj->ptr();
1387 offset = tj->offset();
1388 }
1389
1390 // Array pointers need some flattening
1391 const TypeAryPtr* ta = tj->isa_aryptr();
1392 if (ta && ta->is_stable()) {
1393 // Erase stability property for alias analysis.
1394 tj = ta = ta->cast_to_stable(false);
1395 }
1396 if (ta && ta->is_not_flat()) {
1397 // Erase not flat property for alias analysis.
1398 tj = ta = ta->cast_to_not_flat(false);
1399 }
1400 if (ta && ta->is_not_null_free()) {
1401 // Erase not null free property for alias analysis.
1402 tj = ta = ta->cast_to_not_null_free(false);
1403 }
1404
1405 if( ta && is_known_inst ) {
1406 if ( offset != Type::OffsetBot &&
1407 offset > arrayOopDesc::length_offset_in_bytes() ) {
1408 offset = Type::OffsetBot; // Flatten constant access into array body only
1409 tj = ta = ta->
1410 remove_speculative()->
1411 cast_to_ptr_type(ptr)->
1412 with_offset(offset);
1413 }
1414 } else if (ta) {
1415 // For arrays indexed by constant indices, we flatten the alias
1416 // space to include all of the array body. Only the header, klass
1417 // and array length can be accessed un-aliased.
1418 // For flat inline type array, each field has its own slice so
1419 // we must include the field offset.
1420 if( offset != Type::OffsetBot ) {
1421 if( ta->const_oop() ) { // MethodData* or Method*
1422 offset = Type::OffsetBot; // Flatten constant access into array body
1423 tj = ta = ta->
1424 remove_speculative()->
1425 cast_to_ptr_type(ptr)->
1426 cast_to_exactness(false)->
1427 with_offset(offset);
1428 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1429 // range is OK as-is.
1430 tj = ta = TypeAryPtr::RANGE;
1431 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1432 tj = TypeInstPtr::KLASS; // all klass loads look alike
1433 ta = TypeAryPtr::RANGE; // generic ignored junk
1434 ptr = TypePtr::BotPTR;
1435 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1436 tj = TypeInstPtr::MARK;
1437 ta = TypeAryPtr::RANGE; // generic ignored junk
1438 ptr = TypePtr::BotPTR;
1439 } else { // Random constant offset into array body
1440 offset = Type::OffsetBot; // Flatten constant access into array body
1441 tj = ta = ta->
1442 remove_speculative()->
1443 cast_to_ptr_type(ptr)->
1444 cast_to_exactness(false)->
1445 with_offset(offset);
1446 }
1447 }
1448 // Arrays of fixed size alias with arrays of unknown size.
1449 if (ta->size() != TypeInt::POS) {
1450 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1451 tj = ta = ta->
1452 remove_speculative()->
1453 cast_to_ptr_type(ptr)->
1454 with_ary(tary)->
1455 cast_to_exactness(false);
1456 }
1457 // Arrays of known objects become arrays of unknown objects.
1458 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1459 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1460 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1461 }
1462 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1463 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1464 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1465 }
1466 // Initially all flattened array accesses share a single slice
1467 if (ta->is_flat() && ta->elem() != TypeInstPtr::BOTTOM && _flat_accesses_share_alias) {
1468 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size(), /* stable= */ false, /* flat= */ true);
1469 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1470 }
1471 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1472 // cannot be distinguished by bytecode alone.
1473 if (ta->elem() == TypeInt::BOOL) {
1474 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1475 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1476 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1477 }
1478 // During the 2nd round of IterGVN, NotNull castings are removed.
1479 // Make sure the Bottom and NotNull variants alias the same.
1480 // Also, make sure exact and non-exact variants alias the same.
1481 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1482 tj = ta = ta->
1483 remove_speculative()->
1484 cast_to_ptr_type(TypePtr::BotPTR)->
1485 cast_to_exactness(false)->
1486 with_offset(offset);
1487 }
1488 }
1489
1490 // Oop pointers need some flattening
1491 const TypeInstPtr *to = tj->isa_instptr();
1492 if (to && to != TypeOopPtr::BOTTOM) {
1493 ciInstanceKlass* ik = to->instance_klass();
1494 if( ptr == TypePtr::Constant ) {
1495 if (ik != ciEnv::current()->Class_klass() ||
1496 offset < ik->layout_helper_size_in_bytes()) {
1497 // No constant oop pointers (such as Strings); they alias with
1498 // unknown strings.
1499 assert(!is_known_inst, "not scalarizable allocation");
1500 tj = to = to->
1501 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1502 remove_speculative()->
1503 cast_to_ptr_type(TypePtr::BotPTR)->
1504 cast_to_exactness(false);
1505 }
1506 } else if( is_known_inst ) {
1507 tj = to; // Keep NotNull and klass_is_exact for instance type
1508 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1509 // During the 2nd round of IterGVN, NotNull castings are removed.
1510 // Make sure the Bottom and NotNull variants alias the same.
1511 // Also, make sure exact and non-exact variants alias the same.
1512 tj = to = to->
1513 remove_speculative()->
1514 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1515 cast_to_ptr_type(TypePtr::BotPTR)->
1516 cast_to_exactness(false);
1517 }
1518 if (to->speculative() != nullptr) {
1519 tj = to = to->remove_speculative();
1520 }
1521 // Canonicalize the holder of this field
1522 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1523 // First handle header references such as a LoadKlassNode, even if the
1524 // object's klass is unloaded at compile time (4965979).
1525 if (!is_known_inst) { // Do it only for non-instance types
1526 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1527 }
1528 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1529 // Static fields are in the space above the normal instance
1530 // fields in the java.lang.Class instance.
1531 if (ik != ciEnv::current()->Class_klass()) {
1532 to = nullptr;
1533 tj = TypeOopPtr::BOTTOM;
1534 offset = tj->offset();
1535 }
1536 } else {
1537 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1538 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1539 assert(tj->offset() == offset, "no change to offset expected");
1540 bool xk = to->klass_is_exact();
1541 int instance_id = to->instance_id();
1542
1543 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1544 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1545 // its interfaces are included.
1546 if (xk && ik->equals(canonical_holder)) {
1547 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id), "exact type should be canonical type");
1548 } else {
1549 assert(xk || !is_known_inst, "Known instance should be exact type");
1550 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id);
1551 }
1552 }
1553 }
1554
1555 // Klass pointers to object array klasses need some flattening
1556 const TypeKlassPtr *tk = tj->isa_klassptr();
1557 if( tk ) {
1558 // If we are referencing a field within a Klass, we need
1559 // to assume the worst case of an Object. Both exact and
1560 // inexact types must flatten to the same alias class so
1561 // use NotNull as the PTR.
1562 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1563 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1564 env()->Object_klass(),
1565 Type::Offset(offset));
1566 }
1567
1568 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1569 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1570 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1571 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset));
1572 } else {
1573 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());
1574 }
1575 }
1576 // Check for precise loads from the primary supertype array and force them
1577 // to the supertype cache alias index. Check for generic array loads from
1578 // the primary supertype array and also force them to the supertype cache
1579 // alias index. Since the same load can reach both, we need to merge
1580 // these 2 disparate memories into the same alias class. Since the
1581 // primary supertype array is read-only, there's no chance of confusion
1582 // where we bypass an array load and an array store.
1583 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1584 if (offset == Type::OffsetBot ||
1585 (offset >= primary_supers_offset &&
1586 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1587 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1588 offset = in_bytes(Klass::secondary_super_cache_offset());
1589 tj = tk = tk->with_offset(offset);
1590 }
1591 }
1592
1593 // Flatten all Raw pointers together.
1594 if (tj->base() == Type::RawPtr)
1595 tj = TypeRawPtr::BOTTOM;
1596
1597 if (tj->base() == Type::AnyPtr)
1598 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1599
1600 offset = tj->offset();
1601 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1602
1603 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1604 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1605 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1606 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1607 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1608 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1609 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1610 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1611 assert( tj->ptr() != TypePtr::TopPTR &&
1612 tj->ptr() != TypePtr::AnyNull &&
1613 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1614 // assert( tj->ptr() != TypePtr::Constant ||
1615 // tj->base() == Type::RawPtr ||
1616 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1617
1618 return tj;
1619 }
1620
1621 void Compile::AliasType::Init(int i, const TypePtr* at) {
1622 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1623 _index = i;
1624 _adr_type = at;
1625 _field = nullptr;
1626 _element = nullptr;
1627 _is_rewritable = true; // default
1628 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1629 if (atoop != nullptr && atoop->is_known_instance()) {
1630 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1631 _general_index = Compile::current()->get_alias_index(gt);
1632 } else {
1633 _general_index = 0;
1634 }
1635 }
1636
1637 BasicType Compile::AliasType::basic_type() const {
1638 if (element() != nullptr) {
1639 const Type* element = adr_type()->is_aryptr()->elem();
1640 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1641 } if (field() != nullptr) {
1642 return field()->layout_type();
1643 } else {
1644 return T_ILLEGAL; // unknown
1645 }
1646 }
1647
1648 //---------------------------------print_on------------------------------------
1649 #ifndef PRODUCT
1650 void Compile::AliasType::print_on(outputStream* st) {
1651 if (index() < 10)
1652 st->print("@ <%d> ", index());
1653 else st->print("@ <%d>", index());
1654 st->print(is_rewritable() ? " " : " RO");
1655 int offset = adr_type()->offset();
1656 if (offset == Type::OffsetBot)
1657 st->print(" +any");
1658 else st->print(" +%-3d", offset);
1659 st->print(" in ");
1660 adr_type()->dump_on(st);
1661 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1662 if (field() != nullptr && tjp) {
1663 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1664 tjp->offset() != field()->offset_in_bytes()) {
1665 st->print(" != ");
1666 field()->print();
1667 st->print(" ***");
1668 }
1669 }
1670 }
1671
1672 void print_alias_types() {
1673 Compile* C = Compile::current();
1674 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1675 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1676 C->alias_type(idx)->print_on(tty);
1677 tty->cr();
1678 }
1679 }
1680 #endif
1681
1682
1683 //----------------------------probe_alias_cache--------------------------------
1684 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1685 intptr_t key = (intptr_t) adr_type;
1686 key ^= key >> logAliasCacheSize;
1687 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1688 }
1689
1690
1691 //-----------------------------grow_alias_types--------------------------------
1692 void Compile::grow_alias_types() {
1693 const int old_ats = _max_alias_types; // how many before?
1694 const int new_ats = old_ats; // how many more?
1695 const int grow_ats = old_ats+new_ats; // how many now?
1696 _max_alias_types = grow_ats;
1697 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1698 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1699 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1700 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1701 }
1702
1703
1704 //--------------------------------find_alias_type------------------------------
1705 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1706 if (!do_aliasing()) {
1707 return alias_type(AliasIdxBot);
1708 }
1709
1710 AliasCacheEntry* ace = nullptr;
1711 if (!uncached) {
1712 ace = probe_alias_cache(adr_type);
1713 if (ace->_adr_type == adr_type) {
1714 return alias_type(ace->_index);
1715 }
1716 }
1717
1718 // Handle special cases.
1719 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1720 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1721
1722 // Do it the slow way.
1723 const TypePtr* flat = flatten_alias_type(adr_type);
1724
1725 #ifdef ASSERT
1726 {
1727 ResourceMark rm;
1728 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1729 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1730 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1731 Type::str(adr_type));
1732 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1733 const TypeOopPtr* foop = flat->is_oopptr();
1734 // Scalarizable allocations have exact klass always.
1735 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1736 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1737 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1738 Type::str(foop), Type::str(xoop));
1739 }
1740 }
1741 #endif
1742
1743 int idx = AliasIdxTop;
1744 for (int i = 0; i < num_alias_types(); i++) {
1745 if (alias_type(i)->adr_type() == flat) {
1746 idx = i;
1747 break;
1748 }
1749 }
1750
1751 if (idx == AliasIdxTop) {
1752 if (no_create) return nullptr;
1753 // Grow the array if necessary.
1754 if (_num_alias_types == _max_alias_types) grow_alias_types();
1755 // Add a new alias type.
1756 idx = _num_alias_types++;
1757 _alias_types[idx]->Init(idx, flat);
1758 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1759 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1760 if (flat->isa_instptr()) {
1761 if (flat->offset() == java_lang_Class::klass_offset()
1762 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1763 alias_type(idx)->set_rewritable(false);
1764 }
1765 ciField* field = nullptr;
1766 if (flat->isa_aryptr()) {
1767 #ifdef ASSERT
1768 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1769 // (T_BYTE has the weakest alignment and size restrictions...)
1770 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1771 #endif
1772 const Type* elemtype = flat->is_aryptr()->elem();
1773 if (flat->offset() == TypePtr::OffsetBot) {
1774 alias_type(idx)->set_element(elemtype);
1775 }
1776 int field_offset = flat->is_aryptr()->field_offset().get();
1777 if (flat->is_flat() &&
1778 field_offset != Type::OffsetBot) {
1779 ciInlineKlass* vk = elemtype->inline_klass();
1780 field_offset += vk->payload_offset();
1781 field = vk->get_field_by_offset(field_offset, false);
1782 }
1783 }
1784 if (flat->isa_klassptr()) {
1785 if (UseCompactObjectHeaders) {
1786 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1787 alias_type(idx)->set_rewritable(false);
1788 }
1789 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1790 alias_type(idx)->set_rewritable(false);
1791 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1792 alias_type(idx)->set_rewritable(false);
1793 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1794 alias_type(idx)->set_rewritable(false);
1795 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1796 alias_type(idx)->set_rewritable(false);
1797 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1798 alias_type(idx)->set_rewritable(false);
1799 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1800 alias_type(idx)->set_rewritable(false);
1801 }
1802 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1803 // but the base pointer type is not distinctive enough to identify
1804 // references into JavaThread.)
1805
1806 // Check for final fields.
1807 const TypeInstPtr* tinst = flat->isa_instptr();
1808 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1809 if (tinst->const_oop() != nullptr &&
1810 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1811 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1812 // static field
1813 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1814 field = k->get_field_by_offset(tinst->offset(), true);
1815 } else if (tinst->is_inlinetypeptr()) {
1816 // Inline type field
1817 ciInlineKlass* vk = tinst->inline_klass();
1818 field = vk->get_field_by_offset(tinst->offset(), false);
1819 } else {
1820 ciInstanceKlass *k = tinst->instance_klass();
1821 field = k->get_field_by_offset(tinst->offset(), false);
1822 }
1823 }
1824 assert(field == nullptr ||
1825 original_field == nullptr ||
1826 (field->holder() == original_field->holder() &&
1827 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1828 field->is_static() == original_field->is_static()), "wrong field?");
1829 // Set field() and is_rewritable() attributes.
1830 if (field != nullptr) {
1831 alias_type(idx)->set_field(field);
1832 if (flat->isa_aryptr()) {
1833 // Fields of flat arrays are rewritable although they are declared final
1834 assert(flat->is_flat(), "must be a flat array");
1835 alias_type(idx)->set_rewritable(true);
1836 }
1837 }
1838 }
1839
1840 // Fill the cache for next time.
1841 if (!uncached) {
1842 ace->_adr_type = adr_type;
1843 ace->_index = idx;
1844 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1845
1846 // Might as well try to fill the cache for the flattened version, too.
1847 AliasCacheEntry* face = probe_alias_cache(flat);
1848 if (face->_adr_type == nullptr) {
1849 face->_adr_type = flat;
1850 face->_index = idx;
1851 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1852 }
1853 }
1854
1855 return alias_type(idx);
1856 }
1857
1858
1859 Compile::AliasType* Compile::alias_type(ciField* field) {
1860 const TypeOopPtr* t;
1861 if (field->is_static())
1862 t = TypeInstPtr::make(field->holder()->java_mirror());
1863 else
1864 t = TypeOopPtr::make_from_klass_raw(field->holder());
1865 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1866 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1867 return atp;
1868 }
1869
1870
1871 //------------------------------have_alias_type--------------------------------
1872 bool Compile::have_alias_type(const TypePtr* adr_type) {
1873 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1874 if (ace->_adr_type == adr_type) {
1875 return true;
1876 }
1877
1878 // Handle special cases.
1879 if (adr_type == nullptr) return true;
1880 if (adr_type == TypePtr::BOTTOM) return true;
1881
1882 return find_alias_type(adr_type, true, nullptr) != nullptr;
1883 }
1884
1885 //-----------------------------must_alias--------------------------------------
1886 // True if all values of the given address type are in the given alias category.
1887 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1888 if (alias_idx == AliasIdxBot) return true; // the universal category
1889 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1890 if (alias_idx == AliasIdxTop) return false; // the empty category
1891 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1892
1893 // the only remaining possible overlap is identity
1894 int adr_idx = get_alias_index(adr_type);
1895 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1896 assert(adr_idx == alias_idx ||
1897 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1898 && adr_type != TypeOopPtr::BOTTOM),
1899 "should not be testing for overlap with an unsafe pointer");
1900 return adr_idx == alias_idx;
1901 }
1902
1903 //------------------------------can_alias--------------------------------------
1904 // True if any values of the given address type are in the given alias category.
1905 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1906 if (alias_idx == AliasIdxTop) return false; // the empty category
1907 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1908 // Known instance doesn't alias with bottom memory
1909 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1910 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1911
1912 // the only remaining possible overlap is identity
1913 int adr_idx = get_alias_index(adr_type);
1914 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1915 return adr_idx == alias_idx;
1916 }
1917
1918 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1919 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1920 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1921 if (parse_predicate_count() == 0) {
1922 return;
1923 }
1924 for (int i = 0; i < parse_predicate_count(); i++) {
1925 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1926 parse_predicate->mark_useless(igvn);
1927 }
1928 _parse_predicates.clear();
1929 }
1930
1931 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1932 if (!n->for_post_loop_opts_igvn()) {
1933 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1934 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1935 _for_post_loop_igvn.append(n);
1936 }
1937 }
1938
1939 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1940 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1941 _for_post_loop_igvn.remove(n);
1942 }
1943
1944 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1945 // Verify that all previous optimizations produced a valid graph
1946 // at least to this point, even if no loop optimizations were done.
1947 PhaseIdealLoop::verify(igvn);
1948
1949 if (has_loops() || _loop_opts_cnt > 0) {
1950 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1951 }
1952 C->set_post_loop_opts_phase(); // no more loop opts allowed
1953
1954 assert(!C->major_progress(), "not cleared");
1955
1956 if (_for_post_loop_igvn.length() > 0) {
1957 while (_for_post_loop_igvn.length() > 0) {
1958 Node* n = _for_post_loop_igvn.pop();
1959 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1960 igvn._worklist.push(n);
1961 }
1962 igvn.optimize();
1963 if (failing()) return;
1964 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1965 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1966
1967 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1968 if (C->major_progress()) {
1969 C->clear_major_progress(); // ensure that major progress is now clear
1970 }
1971 }
1972 }
1973
1974 void Compile::add_inline_type(Node* n) {
1975 assert(n->is_InlineType(), "unexpected node");
1976 _inline_type_nodes.push(n);
1977 }
1978
1979 void Compile::remove_inline_type(Node* n) {
1980 assert(n->is_InlineType(), "unexpected node");
1981 if (_inline_type_nodes.contains(n)) {
1982 _inline_type_nodes.remove(n);
1983 }
1984 }
1985
1986 // Does the return value keep otherwise useless inline type allocations alive?
1987 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1988 ResourceMark rm;
1989 Unique_Node_List wq;
1990 wq.push(ret_val);
1991 bool some_allocations = false;
1992 for (uint i = 0; i < wq.size(); i++) {
1993 Node* n = wq.at(i);
1994 if (n->outcnt() > 1) {
1995 // Some other use for the allocation
1996 return false;
1997 } else if (n->is_InlineType()) {
1998 wq.push(n->in(1));
1999 } else if (n->is_Phi()) {
2000 for (uint j = 1; j < n->req(); j++) {
2001 wq.push(n->in(j));
2002 }
2003 } else if (n->is_CheckCastPP() &&
2004 n->in(1)->is_Proj() &&
2005 n->in(1)->in(0)->is_Allocate()) {
2006 some_allocations = true;
2007 } else if (n->is_CheckCastPP()) {
2008 wq.push(n->in(1));
2009 }
2010 }
2011 return some_allocations;
2012 }
2013
2014 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
2015 // Make sure that the return value does not keep an otherwise unused allocation alive
2016 if (tf()->returns_inline_type_as_fields()) {
2017 Node* ret = nullptr;
2018 for (uint i = 1; i < root()->req(); i++) {
2019 Node* in = root()->in(i);
2020 if (in->Opcode() == Op_Return) {
2021 assert(ret == nullptr, "only one return");
2022 ret = in;
2023 }
2024 }
2025 if (ret != nullptr) {
2026 Node* ret_val = ret->in(TypeFunc::Parms);
2027 if (igvn.type(ret_val)->isa_oopptr() &&
2028 return_val_keeps_allocations_alive(ret_val)) {
2029 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
2030 assert(ret_val->outcnt() == 0, "should be dead now");
2031 igvn.remove_dead_node(ret_val);
2032 }
2033 }
2034 }
2035 if (_inline_type_nodes.length() == 0) {
2036 return;
2037 }
2038 // Scalarize inline types in safepoint debug info.
2039 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2040 set_scalarize_in_safepoints(true);
2041 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2042 InlineTypeNode* vt = _inline_type_nodes.at(i)->as_InlineType();
2043 vt->make_scalar_in_safepoints(&igvn);
2044 igvn.record_for_igvn(vt);
2045 }
2046 if (remove) {
2047 // Remove inline type nodes by replacing them with their oop input
2048 while (_inline_type_nodes.length() > 0) {
2049 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2050 if (vt->outcnt() == 0) {
2051 igvn.remove_dead_node(vt);
2052 continue;
2053 }
2054 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2055 DEBUG_ONLY(bool must_be_buffered = false);
2056 Node* u = vt->out(i);
2057 // Check if any users are blackholes. If so, rewrite them to use either the
2058 // allocated buffer, or individual components, instead of the inline type node
2059 // that goes away.
2060 if (u->is_Blackhole()) {
2061 BlackholeNode* bh = u->as_Blackhole();
2062
2063 // Unlink the old input
2064 int idx = bh->find_edge(vt);
2065 assert(idx != -1, "The edge should be there");
2066 bh->del_req(idx);
2067 --i;
2068
2069 if (vt->is_allocated(&igvn)) {
2070 // Already has the allocated instance, blackhole that
2071 bh->add_req(vt->get_oop());
2072 } else {
2073 // Not allocated yet, blackhole the components
2074 for (uint c = 0; c < vt->field_count(); c++) {
2075 bh->add_req(vt->field_value(c));
2076 }
2077 }
2078
2079 // Node modified, record for IGVN
2080 igvn.record_for_igvn(bh);
2081 }
2082 #ifdef ASSERT
2083 // Verify that inline type is buffered when replacing by oop
2084 else if (u->is_InlineType()) {
2085 // InlineType uses don't need buffering because they are about to be replaced as well
2086 } else if (u->is_Phi()) {
2087 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2088 } else {
2089 must_be_buffered = true;
2090 }
2091 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2092 vt->dump(0);
2093 u->dump(0);
2094 assert(false, "Should have been buffered");
2095 }
2096 #endif
2097 }
2098 igvn.replace_node(vt, vt->get_oop());
2099 }
2100 }
2101 igvn.optimize();
2102 }
2103
2104 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2105 if (!_has_flat_accesses) {
2106 return;
2107 }
2108 // Initially, all flat array accesses share the same slice to
2109 // keep dependencies with Object[] array accesses (that could be
2110 // to a flat array) correct. We're done with parsing so we
2111 // now know all flat array accesses in this compile
2112 // unit. Let's move flat array accesses to their own slice,
2113 // one per element field. This should help memory access
2114 // optimizations.
2115 ResourceMark rm;
2116 Unique_Node_List wq;
2117 wq.push(root());
2118
2119 Node_List mergememnodes;
2120 Node_List memnodes;
2121
2122 // Alias index currently shared by all flat memory accesses
2123 int index = get_alias_index(TypeAryPtr::INLINES);
2124
2125 // Find MergeMem nodes and flat array accesses
2126 for (uint i = 0; i < wq.size(); i++) {
2127 Node* n = wq.at(i);
2128 if (n->is_Mem()) {
2129 const TypePtr* adr_type = nullptr;
2130 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2131 if (adr_type == TypeAryPtr::INLINES) {
2132 memnodes.push(n);
2133 }
2134 } else if (n->is_MergeMem()) {
2135 MergeMemNode* mm = n->as_MergeMem();
2136 if (mm->memory_at(index) != mm->base_memory()) {
2137 mergememnodes.push(n);
2138 }
2139 }
2140 for (uint j = 0; j < n->req(); j++) {
2141 Node* m = n->in(j);
2142 if (m != nullptr) {
2143 wq.push(m);
2144 }
2145 }
2146 }
2147
2148 if (memnodes.size() > 0) {
2149 _flat_accesses_share_alias = false;
2150
2151 // We are going to change the slice for the flat array
2152 // accesses so we need to clear the cache entries that refer to
2153 // them.
2154 for (uint i = 0; i < AliasCacheSize; i++) {
2155 AliasCacheEntry* ace = &_alias_cache[i];
2156 if (ace->_adr_type != nullptr &&
2157 ace->_adr_type->is_flat()) {
2158 ace->_adr_type = nullptr;
2159 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2160 }
2161 }
2162
2163 // Find what aliases we are going to add
2164 int start_alias = num_alias_types()-1;
2165 int stop_alias = 0;
2166
2167 for (uint i = 0; i < memnodes.size(); i++) {
2168 Node* m = memnodes.at(i);
2169 const TypePtr* adr_type = nullptr;
2170 adr_type = m->adr_type();
2171 #ifdef ASSERT
2172 m->as_Mem()->set_adr_type(adr_type);
2173 #endif
2174 int idx = get_alias_index(adr_type);
2175 start_alias = MIN2(start_alias, idx);
2176 stop_alias = MAX2(stop_alias, idx);
2177 }
2178
2179 assert(stop_alias >= start_alias, "should have expanded aliases");
2180
2181 Node_Stack stack(0);
2182 #ifdef ASSERT
2183 VectorSet seen(Thread::current()->resource_area());
2184 #endif
2185 // Now let's fix the memory graph so each flat array access
2186 // is moved to the right slice. Start from the MergeMem nodes.
2187 uint last = unique();
2188 for (uint i = 0; i < mergememnodes.size(); i++) {
2189 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2190 Node* n = current->memory_at(index);
2191 MergeMemNode* mm = nullptr;
2192 do {
2193 // Follow memory edges through memory accesses, phis and
2194 // narrow membars and push nodes on the stack. Once we hit
2195 // bottom memory, we pop element off the stack one at a
2196 // time, in reverse order, and move them to the right slice
2197 // by changing their memory edges.
2198 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2199 assert(!seen.test_set(n->_idx), "");
2200 // Uses (a load for instance) will need to be moved to the
2201 // right slice as well and will get a new memory state
2202 // that we don't know yet. The use could also be the
2203 // backedge of a loop. We put a place holder node between
2204 // the memory node and its uses. We replace that place
2205 // holder with the correct memory state once we know it,
2206 // i.e. when nodes are popped off the stack. Using the
2207 // place holder make the logic work in the presence of
2208 // loops.
2209 if (n->outcnt() > 1) {
2210 Node* place_holder = nullptr;
2211 assert(!n->has_out_with(Op_Node), "");
2212 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2213 Node* u = n->out(k);
2214 if (u != current && u->_idx < last) {
2215 bool success = false;
2216 for (uint l = 0; l < u->req(); l++) {
2217 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2218 continue;
2219 }
2220 Node* in = u->in(l);
2221 if (in == n) {
2222 if (place_holder == nullptr) {
2223 place_holder = new Node(1);
2224 place_holder->init_req(0, n);
2225 }
2226 igvn.replace_input_of(u, l, place_holder);
2227 success = true;
2228 }
2229 }
2230 if (success) {
2231 --k;
2232 }
2233 }
2234 }
2235 }
2236 if (n->is_Phi()) {
2237 stack.push(n, 1);
2238 n = n->in(1);
2239 } else if (n->is_Mem()) {
2240 stack.push(n, n->req());
2241 n = n->in(MemNode::Memory);
2242 } else {
2243 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2244 stack.push(n, n->req());
2245 n = n->in(0)->in(TypeFunc::Memory);
2246 }
2247 } else {
2248 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2249 // Build a new MergeMem node to carry the new memory state
2250 // as we build it. IGVN should fold extraneous MergeMem
2251 // nodes.
2252 mm = MergeMemNode::make(n);
2253 igvn.register_new_node_with_optimizer(mm);
2254 while (stack.size() > 0) {
2255 Node* m = stack.node();
2256 uint idx = stack.index();
2257 if (m->is_Mem()) {
2258 // Move memory node to its new slice
2259 const TypePtr* adr_type = m->adr_type();
2260 int alias = get_alias_index(adr_type);
2261 Node* prev = mm->memory_at(alias);
2262 igvn.replace_input_of(m, MemNode::Memory, prev);
2263 mm->set_memory_at(alias, m);
2264 } else if (m->is_Phi()) {
2265 // We need as many new phis as there are new aliases
2266 igvn.replace_input_of(m, idx, mm);
2267 if (idx == m->req()-1) {
2268 Node* r = m->in(0);
2269 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2270 const TypePtr* adr_type = get_adr_type(j);
2271 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2272 continue;
2273 }
2274 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2275 igvn.register_new_node_with_optimizer(phi);
2276 for (uint k = 1; k < m->req(); k++) {
2277 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2278 }
2279 mm->set_memory_at(j, phi);
2280 }
2281 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2282 igvn.register_new_node_with_optimizer(base_phi);
2283 for (uint k = 1; k < m->req(); k++) {
2284 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2285 }
2286 mm->set_base_memory(base_phi);
2287 }
2288 } else {
2289 // This is a MemBarCPUOrder node from
2290 // Parse::array_load()/Parse::array_store(), in the
2291 // branch that handles flat arrays hidden under
2292 // an Object[] array. We also need one new membar per
2293 // new alias to keep the unknown access that the
2294 // membars protect properly ordered with accesses to
2295 // known flat array.
2296 assert(m->is_Proj(), "projection expected");
2297 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2298 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2299 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2300 const TypePtr* adr_type = get_adr_type(j);
2301 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2302 continue;
2303 }
2304 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2305 igvn.register_new_node_with_optimizer(mb);
2306 Node* mem = mm->memory_at(j);
2307 mb->init_req(TypeFunc::Control, ctrl);
2308 mb->init_req(TypeFunc::Memory, mem);
2309 ctrl = new ProjNode(mb, TypeFunc::Control);
2310 igvn.register_new_node_with_optimizer(ctrl);
2311 mem = new ProjNode(mb, TypeFunc::Memory);
2312 igvn.register_new_node_with_optimizer(mem);
2313 mm->set_memory_at(j, mem);
2314 }
2315 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2316 }
2317 if (idx < m->req()-1) {
2318 idx += 1;
2319 stack.set_index(idx);
2320 n = m->in(idx);
2321 break;
2322 }
2323 // Take care of place holder nodes
2324 if (m->has_out_with(Op_Node)) {
2325 Node* place_holder = m->find_out_with(Op_Node);
2326 if (place_holder != nullptr) {
2327 Node* mm_clone = mm->clone();
2328 igvn.register_new_node_with_optimizer(mm_clone);
2329 Node* hook = new Node(1);
2330 hook->init_req(0, mm);
2331 igvn.replace_node(place_holder, mm_clone);
2332 hook->destruct(&igvn);
2333 }
2334 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2335 }
2336 stack.pop();
2337 }
2338 }
2339 } while(stack.size() > 0);
2340 // Fix the memory state at the MergeMem we started from
2341 igvn.rehash_node_delayed(current);
2342 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2343 const TypePtr* adr_type = get_adr_type(j);
2344 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2345 continue;
2346 }
2347 current->set_memory_at(j, mm);
2348 }
2349 current->set_memory_at(index, current->base_memory());
2350 }
2351 igvn.optimize();
2352 }
2353 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2354 #ifdef ASSERT
2355 if (!_flat_accesses_share_alias) {
2356 wq.clear();
2357 wq.push(root());
2358 for (uint i = 0; i < wq.size(); i++) {
2359 Node* n = wq.at(i);
2360 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2361 for (uint j = 0; j < n->req(); j++) {
2362 Node* m = n->in(j);
2363 if (m != nullptr) {
2364 wq.push(m);
2365 }
2366 }
2367 }
2368 }
2369 #endif
2370 }
2371
2372 void Compile::record_for_merge_stores_igvn(Node* n) {
2373 if (!n->for_merge_stores_igvn()) {
2374 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
2375 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2376 _for_merge_stores_igvn.append(n);
2377 }
2378 }
2379
2380 void Compile::remove_from_merge_stores_igvn(Node* n) {
2381 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2382 _for_merge_stores_igvn.remove(n);
2383 }
2384
2385 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
2386 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
2387 // the stores, and we merge the wrong sequence of stores.
2388 // Example:
2389 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
2390 // Apply MergeStores:
2391 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
2392 // Remove more RangeChecks:
2393 // StoreI [ StoreL ] StoreI
2394 // But now it would have been better to do this instead:
2395 // [ StoreL ] [ StoreL ]
2396 //
2397 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
2398 // since we never unset _merge_stores_phase.
2399 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
2400 C->set_merge_stores_phase();
2401
2402 if (_for_merge_stores_igvn.length() > 0) {
2403 while (_for_merge_stores_igvn.length() > 0) {
2404 Node* n = _for_merge_stores_igvn.pop();
2405 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2406 igvn._worklist.push(n);
2407 }
2408 igvn.optimize();
2409 if (failing()) return;
2410 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
2411 print_method(PHASE_AFTER_MERGE_STORES, 3);
2412 }
2413 }
2414
2415 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2416 if (OptimizeUnstableIf) {
2417 _unstable_if_traps.append(trap);
2418 }
2419 }
2420
2421 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2422 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2423 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2424 Node* n = trap->uncommon_trap();
2425 if (!useful.member(n)) {
2426 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2427 }
2428 }
2429 }
2430
2431 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2432 // or fold-compares case. Return true if succeed or not found.
2433 //
2434 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2435 // false and ask fold-compares to yield.
2436 //
2437 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2438 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2439 // when deoptimization does happen.
2440 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2441 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2442 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2443 if (trap->uncommon_trap() == unc) {
2444 if (yield && trap->modified()) {
2445 return false;
2446 }
2447 _unstable_if_traps.delete_at(i);
2448 break;
2449 }
2450 }
2451 return true;
2452 }
2453
2454 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2455 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2456 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2457 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2458 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2459 CallStaticJavaNode* unc = trap->uncommon_trap();
2460 int next_bci = trap->next_bci();
2461 bool modified = trap->modified();
2462
2463 if (next_bci != -1 && !modified) {
2464 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2465 JVMState* jvms = unc->jvms();
2466 ciMethod* method = jvms->method();
2467 ciBytecodeStream iter(method);
2468
2469 iter.force_bci(jvms->bci());
2470 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2471 Bytecodes::Code c = iter.cur_bc();
2472 Node* lhs = nullptr;
2473 Node* rhs = nullptr;
2474 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2475 lhs = unc->peek_operand(0);
2476 rhs = unc->peek_operand(1);
2477 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2478 lhs = unc->peek_operand(0);
2479 }
2480
2481 ResourceMark rm;
2482 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2483 assert(live_locals.is_valid(), "broken liveness info");
2484 int len = (int)live_locals.size();
2485
2486 for (int i = 0; i < len; i++) {
2487 Node* local = unc->local(jvms, i);
2488 // kill local using the liveness of next_bci.
2489 // give up when the local looks like an operand to secure reexecution.
2490 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2491 uint idx = jvms->locoff() + i;
2492 #ifdef ASSERT
2493 if (PrintOpto && Verbose) {
2494 tty->print("[unstable_if] kill local#%d: ", idx);
2495 local->dump();
2496 tty->cr();
2497 }
2498 #endif
2499 igvn.replace_input_of(unc, idx, top());
2500 modified = true;
2501 }
2502 }
2503 }
2504
2505 // keep the modified trap for late query
2506 if (modified) {
2507 trap->set_modified();
2508 } else {
2509 _unstable_if_traps.delete_at(i);
2510 }
2511 }
2512 igvn.optimize();
2513 }
2514
2515 // StringOpts and late inlining of string methods
2516 void Compile::inline_string_calls(bool parse_time) {
2517 {
2518 // remove useless nodes to make the usage analysis simpler
2519 ResourceMark rm;
2520 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2521 }
2522
2523 {
2524 ResourceMark rm;
2525 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2526 PhaseStringOpts pso(initial_gvn());
2527 print_method(PHASE_AFTER_STRINGOPTS, 3);
2528 }
2529
2530 // now inline anything that we skipped the first time around
2531 if (!parse_time) {
2532 _late_inlines_pos = _late_inlines.length();
2533 }
2534
2535 while (_string_late_inlines.length() > 0) {
2536 CallGenerator* cg = _string_late_inlines.pop();
2537 cg->do_late_inline();
2538 if (failing()) return;
2539 }
2540 _string_late_inlines.trunc_to(0);
2541 }
2542
2543 // Late inlining of boxing methods
2544 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2545 if (_boxing_late_inlines.length() > 0) {
2546 assert(has_boxed_value(), "inconsistent");
2547
2548 set_inlining_incrementally(true);
2549
2550 igvn_worklist()->ensure_empty(); // should be done with igvn
2551
2552 _late_inlines_pos = _late_inlines.length();
2553
2554 while (_boxing_late_inlines.length() > 0) {
2555 CallGenerator* cg = _boxing_late_inlines.pop();
2556 cg->do_late_inline();
2557 if (failing()) return;
2558 }
2559 _boxing_late_inlines.trunc_to(0);
2560
2561 inline_incrementally_cleanup(igvn);
2562
2563 set_inlining_incrementally(false);
2564 }
2565 }
2566
2567 bool Compile::inline_incrementally_one() {
2568 assert(IncrementalInline, "incremental inlining should be on");
2569
2570 TracePhase tp(_t_incrInline_inline);
2571
2572 set_inlining_progress(false);
2573 set_do_cleanup(false);
2574
2575 for (int i = 0; i < _late_inlines.length(); i++) {
2576 _late_inlines_pos = i+1;
2577 CallGenerator* cg = _late_inlines.at(i);
2578 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2579 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2580 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2581 cg->do_late_inline();
2582 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2583 if (failing()) {
2584 return false;
2585 } else if (inlining_progress()) {
2586 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2587 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2588 break; // process one call site at a time
2589 } else {
2590 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2591 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2592 // Avoid potential infinite loop if node already in the IGVN list
2593 assert(false, "scheduled for IGVN during inlining attempt");
2594 } else {
2595 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2596 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2597 cg->call_node()->set_generator(cg);
2598 }
2599 }
2600 } else {
2601 // Ignore late inline direct calls when inlining is not allowed.
2602 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2603 }
2604 }
2605 // Remove processed elements.
2606 _late_inlines.remove_till(_late_inlines_pos);
2607 _late_inlines_pos = 0;
2608
2609 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2610
2611 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2612
2613 set_inlining_progress(false);
2614 set_do_cleanup(false);
2615
2616 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2617 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2618 }
2619
2620 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2621 {
2622 TracePhase tp(_t_incrInline_pru);
2623 ResourceMark rm;
2624 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2625 }
2626 {
2627 TracePhase tp(_t_incrInline_igvn);
2628 igvn.reset_from_gvn(initial_gvn());
2629 igvn.optimize();
2630 if (failing()) return;
2631 }
2632 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2633 }
2634
2635 // Perform incremental inlining until bound on number of live nodes is reached
2636 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2637 TracePhase tp(_t_incrInline);
2638
2639 set_inlining_incrementally(true);
2640 uint low_live_nodes = 0;
2641
2642 while (_late_inlines.length() > 0) {
2643 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2644 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2645 TracePhase tp(_t_incrInline_ideal);
2646 // PhaseIdealLoop is expensive so we only try it once we are
2647 // out of live nodes and we only try it again if the previous
2648 // helped got the number of nodes down significantly
2649 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2650 if (failing()) return;
2651 low_live_nodes = live_nodes();
2652 _major_progress = true;
2653 }
2654
2655 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2656 bool do_print_inlining = print_inlining() || print_intrinsics();
2657 if (do_print_inlining || log() != nullptr) {
2658 // Print inlining message for candidates that we couldn't inline for lack of space.
2659 for (int i = 0; i < _late_inlines.length(); i++) {
2660 CallGenerator* cg = _late_inlines.at(i);
2661 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2662 if (do_print_inlining) {
2663 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2664 }
2665 log_late_inline_failure(cg, msg);
2666 }
2667 }
2668 break; // finish
2669 }
2670 }
2671
2672 igvn_worklist()->ensure_empty(); // should be done with igvn
2673
2674 while (inline_incrementally_one()) {
2675 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2676 }
2677 if (failing()) return;
2678
2679 inline_incrementally_cleanup(igvn);
2680
2681 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2682
2683 if (failing()) return;
2684
2685 if (_late_inlines.length() == 0) {
2686 break; // no more progress
2687 }
2688 }
2689
2690 igvn_worklist()->ensure_empty(); // should be done with igvn
2691
2692 if (_string_late_inlines.length() > 0) {
2693 assert(has_stringbuilder(), "inconsistent");
2694
2695 inline_string_calls(false);
2696
2697 if (failing()) return;
2698
2699 inline_incrementally_cleanup(igvn);
2700 }
2701
2702 set_inlining_incrementally(false);
2703 }
2704
2705 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2706 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2707 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2708 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2709 // as if "inlining_incrementally() == true" were set.
2710 assert(inlining_incrementally() == false, "not allowed");
2711 #ifdef ASSERT
2712 Unique_Node_List* modified_nodes = _modified_nodes;
2713 _modified_nodes = nullptr;
2714 #endif
2715 assert(_late_inlines.length() > 0, "sanity");
2716
2717 while (_late_inlines.length() > 0) {
2718 igvn_worklist()->ensure_empty(); // should be done with igvn
2719
2720 while (inline_incrementally_one()) {
2721 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2722 }
2723 if (failing()) return;
2724
2725 inline_incrementally_cleanup(igvn);
2726 }
2727 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2728 }
2729
2730 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2731 if (_loop_opts_cnt > 0) {
2732 while (major_progress() && (_loop_opts_cnt > 0)) {
2733 TracePhase tp(_t_idealLoop);
2734 PhaseIdealLoop::optimize(igvn, mode);
2735 _loop_opts_cnt--;
2736 if (failing()) return false;
2737 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2738 }
2739 }
2740 return true;
2741 }
2742
2743 // Remove edges from "root" to each SafePoint at a backward branch.
2744 // They were inserted during parsing (see add_safepoint()) to make
2745 // infinite loops without calls or exceptions visible to root, i.e.,
2746 // useful.
2747 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2748 Node *r = root();
2749 if (r != nullptr) {
2750 for (uint i = r->req(); i < r->len(); ++i) {
2751 Node *n = r->in(i);
2752 if (n != nullptr && n->is_SafePoint()) {
2753 r->rm_prec(i);
2754 if (n->outcnt() == 0) {
2755 igvn.remove_dead_node(n);
2756 }
2757 --i;
2758 }
2759 }
2760 // Parsing may have added top inputs to the root node (Path
2761 // leading to the Halt node proven dead). Make sure we get a
2762 // chance to clean them up.
2763 igvn._worklist.push(r);
2764 igvn.optimize();
2765 }
2766 }
2767
2768 //------------------------------Optimize---------------------------------------
2769 // Given a graph, optimize it.
2770 void Compile::Optimize() {
2771 TracePhase tp(_t_optimizer);
2772
2773 #ifndef PRODUCT
2774 if (env()->break_at_compile()) {
2775 BREAKPOINT;
2776 }
2777
2778 #endif
2779
2780 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2781 #ifdef ASSERT
2782 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2783 #endif
2784
2785 ResourceMark rm;
2786
2787 NOT_PRODUCT( verify_graph_edges(); )
2788
2789 print_method(PHASE_AFTER_PARSING, 1);
2790
2791 {
2792 // Iterative Global Value Numbering, including ideal transforms
2793 // Initialize IterGVN with types and values from parse-time GVN
2794 PhaseIterGVN igvn(initial_gvn());
2795 #ifdef ASSERT
2796 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2797 #endif
2798 {
2799 TracePhase tp(_t_iterGVN);
2800 igvn.optimize();
2801 }
2802
2803 if (failing()) return;
2804
2805 print_method(PHASE_ITER_GVN1, 2);
2806
2807 process_for_unstable_if_traps(igvn);
2808
2809 if (failing()) return;
2810
2811 inline_incrementally(igvn);
2812
2813 print_method(PHASE_INCREMENTAL_INLINE, 2);
2814
2815 if (failing()) return;
2816
2817 if (eliminate_boxing()) {
2818 // Inline valueOf() methods now.
2819 inline_boxing_calls(igvn);
2820
2821 if (failing()) return;
2822
2823 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2824 inline_incrementally(igvn);
2825 }
2826
2827 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2828
2829 if (failing()) return;
2830 }
2831
2832 // Remove the speculative part of types and clean up the graph from
2833 // the extra CastPP nodes whose only purpose is to carry them. Do
2834 // that early so that optimizations are not disrupted by the extra
2835 // CastPP nodes.
2836 remove_speculative_types(igvn);
2837
2838 if (failing()) return;
2839
2840 // No more new expensive nodes will be added to the list from here
2841 // so keep only the actual candidates for optimizations.
2842 cleanup_expensive_nodes(igvn);
2843
2844 if (failing()) return;
2845
2846 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2847 if (EnableVectorSupport && has_vbox_nodes()) {
2848 TracePhase tp(_t_vector);
2849 PhaseVector pv(igvn);
2850 pv.optimize_vector_boxes();
2851 if (failing()) return;
2852 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2853 }
2854 assert(!has_vbox_nodes(), "sanity");
2855
2856 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2857 Compile::TracePhase tp(_t_renumberLive);
2858 igvn_worklist()->ensure_empty(); // should be done with igvn
2859 {
2860 ResourceMark rm;
2861 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2862 }
2863 igvn.reset_from_gvn(initial_gvn());
2864 igvn.optimize();
2865 if (failing()) return;
2866 }
2867
2868 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2869 // safepoints
2870 remove_root_to_sfpts_edges(igvn);
2871
2872 // Process inline type nodes now that all inlining is over
2873 process_inline_types(igvn);
2874
2875 adjust_flat_array_access_aliases(igvn);
2876
2877 if (failing()) return;
2878
2879 if (C->macro_count() > 0) {
2880 // Eliminate some macro nodes before EA to reduce analysis pressure
2881 PhaseMacroExpand mexp(igvn);
2882 mexp.eliminate_macro_nodes();
2883 if (failing()) {
2884 return;
2885 }
2886 igvn.set_delay_transform(false);
2887 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2888 }
2889
2890 if (has_loops()) {
2891 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2892 }
2893
2894 // Perform escape analysis
2895 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2896 if (has_loops()) {
2897 // Cleanup graph (remove dead nodes).
2898 TracePhase tp(_t_idealLoop);
2899 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2900 if (failing()) {
2901 return;
2902 }
2903 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2904 if (C->macro_count() > 0) {
2905 // Eliminate some macro nodes before EA to reduce analysis pressure
2906 PhaseMacroExpand mexp(igvn);
2907 mexp.eliminate_macro_nodes();
2908 if (failing()) {
2909 return;
2910 }
2911 igvn.set_delay_transform(false);
2912 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2913 }
2914 }
2915
2916 bool progress;
2917 do {
2918 ConnectionGraph::do_analysis(this, &igvn);
2919
2920 if (failing()) return;
2921
2922 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2923
2924 // Optimize out fields loads from scalar replaceable allocations.
2925 igvn.optimize();
2926 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2927
2928 if (failing()) return;
2929
2930 if (congraph() != nullptr && macro_count() > 0) {
2931 TracePhase tp(_t_macroEliminate);
2932 PhaseMacroExpand mexp(igvn);
2933 mexp.eliminate_macro_nodes();
2934 if (failing()) {
2935 return;
2936 }
2937 igvn.set_delay_transform(false);
2938 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2939 }
2940
2941 ConnectionGraph::verify_ram_nodes(this, root());
2942 if (failing()) return;
2943
2944 progress = do_iterative_escape_analysis() &&
2945 (macro_count() < mcount) &&
2946 ConnectionGraph::has_candidates(this);
2947 // Try again if candidates exist and made progress
2948 // by removing some allocations and/or locks.
2949 } while (progress);
2950 }
2951
2952 // Loop transforms on the ideal graph. Range Check Elimination,
2953 // peeling, unrolling, etc.
2954
2955 // Set loop opts counter
2956 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2957 {
2958 TracePhase tp(_t_idealLoop);
2959 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2960 _loop_opts_cnt--;
2961 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2962 if (failing()) return;
2963 }
2964 // Loop opts pass if partial peeling occurred in previous pass
2965 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2966 TracePhase tp(_t_idealLoop);
2967 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2968 _loop_opts_cnt--;
2969 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2970 if (failing()) return;
2971 }
2972 // Loop opts pass for loop-unrolling before CCP
2973 if(major_progress() && (_loop_opts_cnt > 0)) {
2974 TracePhase tp(_t_idealLoop);
2975 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2976 _loop_opts_cnt--;
2977 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2978 }
2979 if (!failing()) {
2980 // Verify that last round of loop opts produced a valid graph
2981 PhaseIdealLoop::verify(igvn);
2982 }
2983 }
2984 if (failing()) return;
2985
2986 // Conditional Constant Propagation;
2987 print_method(PHASE_BEFORE_CCP1, 2);
2988 PhaseCCP ccp( &igvn );
2989 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2990 {
2991 TracePhase tp(_t_ccp);
2992 ccp.do_transform();
2993 }
2994 print_method(PHASE_CCP1, 2);
2995
2996 assert( true, "Break here to ccp.dump_old2new_map()");
2997
2998 // Iterative Global Value Numbering, including ideal transforms
2999 {
3000 TracePhase tp(_t_iterGVN2);
3001 igvn.reset_from_igvn(&ccp);
3002 igvn.optimize();
3003 }
3004 print_method(PHASE_ITER_GVN2, 2);
3005
3006 if (failing()) return;
3007
3008 // Loop transforms on the ideal graph. Range Check Elimination,
3009 // peeling, unrolling, etc.
3010 if (!optimize_loops(igvn, LoopOptsDefault)) {
3011 return;
3012 }
3013
3014 if (failing()) return;
3015
3016 C->clear_major_progress(); // ensure that major progress is now clear
3017
3018 process_for_post_loop_opts_igvn(igvn);
3019
3020 process_for_merge_stores_igvn(igvn);
3021
3022 if (failing()) return;
3023
3024 #ifdef ASSERT
3025 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
3026 #endif
3027
3028 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
3029
3030 if (_late_inlines.length() > 0) {
3031 // More opportunities to optimize virtual and MH calls.
3032 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
3033 process_late_inline_calls_no_inline(igvn);
3034 }
3035
3036 {
3037 TracePhase tp(_t_macroExpand);
3038 PhaseMacroExpand mex(igvn);
3039 // Last attempt to eliminate macro nodes.
3040 mex.eliminate_macro_nodes();
3041 if (failing()) {
3042 return;
3043 }
3044
3045 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
3046 if (mex.expand_macro_nodes()) {
3047 assert(failing(), "must bail out w/ explicit message");
3048 return;
3049 }
3050 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
3051 }
3052
3053 // Process inline type nodes again and remove them. From here
3054 // on we don't need to keep track of field values anymore.
3055 process_inline_types(igvn, /* remove= */ true);
3056
3057 {
3058 TracePhase tp(_t_barrierExpand);
3059 if (bs->expand_barriers(this, igvn)) {
3060 assert(failing(), "must bail out w/ explicit message");
3061 return;
3062 }
3063 print_method(PHASE_BARRIER_EXPANSION, 2);
3064 }
3065
3066 if (C->max_vector_size() > 0) {
3067 C->optimize_logic_cones(igvn);
3068 igvn.optimize();
3069 if (failing()) return;
3070 }
3071
3072 DEBUG_ONLY( _modified_nodes = nullptr; )
3073 DEBUG_ONLY( _late_inlines.clear(); )
3074
3075 assert(igvn._worklist.size() == 0, "not empty");
3076 } // (End scope of igvn; run destructor if necessary for asserts.)
3077
3078 check_no_dead_use();
3079
3080 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
3081 // to remove hashes to unlock nodes for modifications.
3082 C->node_hash()->clear();
3083
3084 // A method with only infinite loops has no edges entering loops from root
3085 {
3086 TracePhase tp(_t_graphReshaping);
3087 if (final_graph_reshaping()) {
3088 assert(failing(), "must bail out w/ explicit message");
3089 return;
3090 }
3091 }
3092
3093 print_method(PHASE_OPTIMIZE_FINISHED, 2);
3094 DEBUG_ONLY(set_phase_optimize_finished();)
3095 }
3096
3097 #ifdef ASSERT
3098 void Compile::check_no_dead_use() const {
3099 ResourceMark rm;
3100 Unique_Node_List wq;
3101 wq.push(root());
3102 for (uint i = 0; i < wq.size(); ++i) {
3103 Node* n = wq.at(i);
3104 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3105 Node* u = n->fast_out(j);
3106 if (u->outcnt() == 0 && !u->is_Con()) {
3107 u->dump();
3108 fatal("no reachable node should have no use");
3109 }
3110 wq.push(u);
3111 }
3112 }
3113 }
3114 #endif
3115
3116 void Compile::inline_vector_reboxing_calls() {
3117 if (C->_vector_reboxing_late_inlines.length() > 0) {
3118 _late_inlines_pos = C->_late_inlines.length();
3119 while (_vector_reboxing_late_inlines.length() > 0) {
3120 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
3121 cg->do_late_inline();
3122 if (failing()) return;
3123 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
3124 }
3125 _vector_reboxing_late_inlines.trunc_to(0);
3126 }
3127 }
3128
3129 bool Compile::has_vbox_nodes() {
3130 if (C->_vector_reboxing_late_inlines.length() > 0) {
3131 return true;
3132 }
3133 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3134 Node * n = C->macro_node(macro_idx);
3135 assert(n->is_macro(), "only macro nodes expected here");
3136 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3137 return true;
3138 }
3139 }
3140 return false;
3141 }
3142
3143 //---------------------------- Bitwise operation packing optimization ---------------------------
3144
3145 static bool is_vector_unary_bitwise_op(Node* n) {
3146 return n->Opcode() == Op_XorV &&
3147 VectorNode::is_vector_bitwise_not_pattern(n);
3148 }
3149
3150 static bool is_vector_binary_bitwise_op(Node* n) {
3151 switch (n->Opcode()) {
3152 case Op_AndV:
3153 case Op_OrV:
3154 return true;
3155
3156 case Op_XorV:
3157 return !is_vector_unary_bitwise_op(n);
3158
3159 default:
3160 return false;
3161 }
3162 }
3163
3164 static bool is_vector_ternary_bitwise_op(Node* n) {
3165 return n->Opcode() == Op_MacroLogicV;
3166 }
3167
3168 static bool is_vector_bitwise_op(Node* n) {
3169 return is_vector_unary_bitwise_op(n) ||
3170 is_vector_binary_bitwise_op(n) ||
3171 is_vector_ternary_bitwise_op(n);
3172 }
3173
3174 static bool is_vector_bitwise_cone_root(Node* n) {
3175 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
3176 return false;
3177 }
3178 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3179 if (is_vector_bitwise_op(n->fast_out(i))) {
3180 return false;
3181 }
3182 }
3183 return true;
3184 }
3185
3186 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3187 uint cnt = 0;
3188 if (is_vector_bitwise_op(n)) {
3189 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3190 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3191 for (uint i = 1; i < inp_cnt; i++) {
3192 Node* in = n->in(i);
3193 bool skip = VectorNode::is_all_ones_vector(in);
3194 if (!skip && !inputs.member(in)) {
3195 inputs.push(in);
3196 cnt++;
3197 }
3198 }
3199 assert(cnt <= 1, "not unary");
3200 } else {
3201 uint last_req = inp_cnt;
3202 if (is_vector_ternary_bitwise_op(n)) {
3203 last_req = inp_cnt - 1; // skip last input
3204 }
3205 for (uint i = 1; i < last_req; i++) {
3206 Node* def = n->in(i);
3207 if (!inputs.member(def)) {
3208 inputs.push(def);
3209 cnt++;
3210 }
3211 }
3212 }
3213 } else { // not a bitwise operations
3214 if (!inputs.member(n)) {
3215 inputs.push(n);
3216 cnt++;
3217 }
3218 }
3219 return cnt;
3220 }
3221
3222 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3223 Unique_Node_List useful_nodes;
3224 C->identify_useful_nodes(useful_nodes);
3225
3226 for (uint i = 0; i < useful_nodes.size(); i++) {
3227 Node* n = useful_nodes.at(i);
3228 if (is_vector_bitwise_cone_root(n)) {
3229 list.push(n);
3230 }
3231 }
3232 }
3233
3234 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3235 const TypeVect* vt,
3236 Unique_Node_List& partition,
3237 Unique_Node_List& inputs) {
3238 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3239 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3240 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3241
3242 Node* in1 = inputs.at(0);
3243 Node* in2 = inputs.at(1);
3244 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3245
3246 uint func = compute_truth_table(partition, inputs);
3247
3248 Node* pn = partition.at(partition.size() - 1);
3249 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3250 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3251 }
3252
3253 static uint extract_bit(uint func, uint pos) {
3254 return (func & (1 << pos)) >> pos;
3255 }
3256
3257 //
3258 // A macro logic node represents a truth table. It has 4 inputs,
3259 // First three inputs corresponds to 3 columns of a truth table
3260 // and fourth input captures the logic function.
3261 //
3262 // eg. fn = (in1 AND in2) OR in3;
3263 //
3264 // MacroNode(in1,in2,in3,fn)
3265 //
3266 // -----------------
3267 // in1 in2 in3 fn
3268 // -----------------
3269 // 0 0 0 0
3270 // 0 0 1 1
3271 // 0 1 0 0
3272 // 0 1 1 1
3273 // 1 0 0 0
3274 // 1 0 1 1
3275 // 1 1 0 1
3276 // 1 1 1 1
3277 //
3278
3279 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3280 int res = 0;
3281 for (int i = 0; i < 8; i++) {
3282 int bit1 = extract_bit(in1, i);
3283 int bit2 = extract_bit(in2, i);
3284 int bit3 = extract_bit(in3, i);
3285
3286 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3287 int func_bit = extract_bit(func, func_bit_pos);
3288
3289 res |= func_bit << i;
3290 }
3291 return res;
3292 }
3293
3294 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
3295 assert(n != nullptr, "");
3296 assert(eval_map.contains(n), "absent");
3297 return *(eval_map.get(n));
3298 }
3299
3300 static void eval_operands(Node* n,
3301 uint& func1, uint& func2, uint& func3,
3302 ResourceHashtable<Node*,uint>& eval_map) {
3303 assert(is_vector_bitwise_op(n), "");
3304
3305 if (is_vector_unary_bitwise_op(n)) {
3306 Node* opnd = n->in(1);
3307 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3308 opnd = n->in(2);
3309 }
3310 func1 = eval_operand(opnd, eval_map);
3311 } else if (is_vector_binary_bitwise_op(n)) {
3312 func1 = eval_operand(n->in(1), eval_map);
3313 func2 = eval_operand(n->in(2), eval_map);
3314 } else {
3315 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3316 func1 = eval_operand(n->in(1), eval_map);
3317 func2 = eval_operand(n->in(2), eval_map);
3318 func3 = eval_operand(n->in(3), eval_map);
3319 }
3320 }
3321
3322 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3323 assert(inputs.size() <= 3, "sanity");
3324 ResourceMark rm;
3325 uint res = 0;
3326 ResourceHashtable<Node*,uint> eval_map;
3327
3328 // Populate precomputed functions for inputs.
3329 // Each input corresponds to one column of 3 input truth-table.
3330 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3331 0xCC, // (_, b, _) -> b
3332 0xF0 }; // (a, _, _) -> a
3333 for (uint i = 0; i < inputs.size(); i++) {
3334 eval_map.put(inputs.at(i), input_funcs[2-i]);
3335 }
3336
3337 for (uint i = 0; i < partition.size(); i++) {
3338 Node* n = partition.at(i);
3339
3340 uint func1 = 0, func2 = 0, func3 = 0;
3341 eval_operands(n, func1, func2, func3, eval_map);
3342
3343 switch (n->Opcode()) {
3344 case Op_OrV:
3345 assert(func3 == 0, "not binary");
3346 res = func1 | func2;
3347 break;
3348 case Op_AndV:
3349 assert(func3 == 0, "not binary");
3350 res = func1 & func2;
3351 break;
3352 case Op_XorV:
3353 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3354 assert(func2 == 0 && func3 == 0, "not unary");
3355 res = (~func1) & 0xFF;
3356 } else {
3357 assert(func3 == 0, "not binary");
3358 res = func1 ^ func2;
3359 }
3360 break;
3361 case Op_MacroLogicV:
3362 // Ordering of inputs may change during evaluation of sub-tree
3363 // containing MacroLogic node as a child node, thus a re-evaluation
3364 // makes sure that function is evaluated in context of current
3365 // inputs.
3366 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3367 break;
3368
3369 default: assert(false, "not supported: %s", n->Name());
3370 }
3371 assert(res <= 0xFF, "invalid");
3372 eval_map.put(n, res);
3373 }
3374 return res;
3375 }
3376
3377 // Criteria under which nodes gets packed into a macro logic node:-
3378 // 1) Parent and both child nodes are all unmasked or masked with
3379 // same predicates.
3380 // 2) Masked parent can be packed with left child if it is predicated
3381 // and both have same predicates.
3382 // 3) Masked parent can be packed with right child if its un-predicated
3383 // or has matching predication condition.
3384 // 4) An unmasked parent can be packed with an unmasked child.
3385 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3386 assert(partition.size() == 0, "not empty");
3387 assert(inputs.size() == 0, "not empty");
3388 if (is_vector_ternary_bitwise_op(n)) {
3389 return false;
3390 }
3391
3392 bool is_unary_op = is_vector_unary_bitwise_op(n);
3393 if (is_unary_op) {
3394 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3395 return false; // too few inputs
3396 }
3397
3398 bool pack_left_child = true;
3399 bool pack_right_child = true;
3400
3401 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3402 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3403
3404 int left_child_input_cnt = 0;
3405 int right_child_input_cnt = 0;
3406
3407 bool parent_is_predicated = n->is_predicated_vector();
3408 bool left_child_predicated = n->in(1)->is_predicated_vector();
3409 bool right_child_predicated = n->in(2)->is_predicated_vector();
3410
3411 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
3412 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3413 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3414
3415 do {
3416 if (pack_left_child && left_child_LOP &&
3417 ((!parent_is_predicated && !left_child_predicated) ||
3418 ((parent_is_predicated && left_child_predicated &&
3419 parent_pred == left_child_pred)))) {
3420 partition.push(n->in(1));
3421 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3422 } else {
3423 inputs.push(n->in(1));
3424 left_child_input_cnt = 1;
3425 }
3426
3427 if (pack_right_child && right_child_LOP &&
3428 (!right_child_predicated ||
3429 (right_child_predicated && parent_is_predicated &&
3430 parent_pred == right_child_pred))) {
3431 partition.push(n->in(2));
3432 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3433 } else {
3434 inputs.push(n->in(2));
3435 right_child_input_cnt = 1;
3436 }
3437
3438 if (inputs.size() > 3) {
3439 assert(partition.size() > 0, "");
3440 inputs.clear();
3441 partition.clear();
3442 if (left_child_input_cnt > right_child_input_cnt) {
3443 pack_left_child = false;
3444 } else {
3445 pack_right_child = false;
3446 }
3447 } else {
3448 break;
3449 }
3450 } while(true);
3451
3452 if(partition.size()) {
3453 partition.push(n);
3454 }
3455
3456 return (partition.size() == 2 || partition.size() == 3) &&
3457 (inputs.size() == 2 || inputs.size() == 3);
3458 }
3459
3460 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3461 assert(is_vector_bitwise_op(n), "not a root");
3462
3463 visited.set(n->_idx);
3464
3465 // 1) Do a DFS walk over the logic cone.
3466 for (uint i = 1; i < n->req(); i++) {
3467 Node* in = n->in(i);
3468 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3469 process_logic_cone_root(igvn, in, visited);
3470 }
3471 }
3472
3473 // 2) Bottom up traversal: Merge node[s] with
3474 // the parent to form macro logic node.
3475 Unique_Node_List partition;
3476 Unique_Node_List inputs;
3477 if (compute_logic_cone(n, partition, inputs)) {
3478 const TypeVect* vt = n->bottom_type()->is_vect();
3479 Node* pn = partition.at(partition.size() - 1);
3480 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3481 if (mask == nullptr ||
3482 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3483 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3484 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3485 igvn.replace_node(n, macro_logic);
3486 }
3487 }
3488 }
3489
3490 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3491 ResourceMark rm;
3492 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3493 Unique_Node_List list;
3494 collect_logic_cone_roots(list);
3495
3496 while (list.size() > 0) {
3497 Node* n = list.pop();
3498 const TypeVect* vt = n->bottom_type()->is_vect();
3499 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3500 if (supported) {
3501 VectorSet visited(comp_arena());
3502 process_logic_cone_root(igvn, n, visited);
3503 }
3504 }
3505 }
3506 }
3507
3508 //------------------------------Code_Gen---------------------------------------
3509 // Given a graph, generate code for it
3510 void Compile::Code_Gen() {
3511 if (failing()) {
3512 return;
3513 }
3514
3515 // Perform instruction selection. You might think we could reclaim Matcher
3516 // memory PDQ, but actually the Matcher is used in generating spill code.
3517 // Internals of the Matcher (including some VectorSets) must remain live
3518 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3519 // set a bit in reclaimed memory.
3520
3521 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3522 // nodes. Mapping is only valid at the root of each matched subtree.
3523 NOT_PRODUCT( verify_graph_edges(); )
3524
3525 Matcher matcher;
3526 _matcher = &matcher;
3527 {
3528 TracePhase tp(_t_matcher);
3529 matcher.match();
3530 if (failing()) {
3531 return;
3532 }
3533 }
3534 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3535 // nodes. Mapping is only valid at the root of each matched subtree.
3536 NOT_PRODUCT( verify_graph_edges(); )
3537
3538 // If you have too many nodes, or if matching has failed, bail out
3539 check_node_count(0, "out of nodes matching instructions");
3540 if (failing()) {
3541 return;
3542 }
3543
3544 print_method(PHASE_MATCHING, 2);
3545
3546 // Build a proper-looking CFG
3547 PhaseCFG cfg(node_arena(), root(), matcher);
3548 if (failing()) {
3549 return;
3550 }
3551 _cfg = &cfg;
3552 {
3553 TracePhase tp(_t_scheduler);
3554 bool success = cfg.do_global_code_motion();
3555 if (!success) {
3556 return;
3557 }
3558
3559 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3560 NOT_PRODUCT( verify_graph_edges(); )
3561 cfg.verify();
3562 if (failing()) {
3563 return;
3564 }
3565 }
3566
3567 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3568 _regalloc = ®alloc;
3569 {
3570 TracePhase tp(_t_registerAllocation);
3571 // Perform register allocation. After Chaitin, use-def chains are
3572 // no longer accurate (at spill code) and so must be ignored.
3573 // Node->LRG->reg mappings are still accurate.
3574 _regalloc->Register_Allocate();
3575
3576 // Bail out if the allocator builds too many nodes
3577 if (failing()) {
3578 return;
3579 }
3580
3581 print_method(PHASE_REGISTER_ALLOCATION, 2);
3582 }
3583
3584 // Prior to register allocation we kept empty basic blocks in case the
3585 // the allocator needed a place to spill. After register allocation we
3586 // are not adding any new instructions. If any basic block is empty, we
3587 // can now safely remove it.
3588 {
3589 TracePhase tp(_t_blockOrdering);
3590 cfg.remove_empty_blocks();
3591 if (do_freq_based_layout()) {
3592 PhaseBlockLayout layout(cfg);
3593 } else {
3594 cfg.set_loop_alignment();
3595 }
3596 cfg.fixup_flow();
3597 cfg.remove_unreachable_blocks();
3598 cfg.verify_dominator_tree();
3599 print_method(PHASE_BLOCK_ORDERING, 3);
3600 }
3601
3602 // Apply peephole optimizations
3603 if( OptoPeephole ) {
3604 TracePhase tp(_t_peephole);
3605 PhasePeephole peep( _regalloc, cfg);
3606 peep.do_transform();
3607 print_method(PHASE_PEEPHOLE, 3);
3608 }
3609
3610 // Do late expand if CPU requires this.
3611 if (Matcher::require_postalloc_expand) {
3612 TracePhase tp(_t_postalloc_expand);
3613 cfg.postalloc_expand(_regalloc);
3614 print_method(PHASE_POSTALLOC_EXPAND, 3);
3615 }
3616
3617 #ifdef ASSERT
3618 {
3619 CompilationMemoryStatistic::do_test_allocations();
3620 if (failing()) return;
3621 }
3622 #endif
3623
3624 // Convert Nodes to instruction bits in a buffer
3625 {
3626 TracePhase tp(_t_output);
3627 PhaseOutput output;
3628 output.Output();
3629 if (failing()) return;
3630 output.install();
3631 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3632 }
3633
3634 // He's dead, Jim.
3635 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3636 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3637 }
3638
3639 //------------------------------Final_Reshape_Counts---------------------------
3640 // This class defines counters to help identify when a method
3641 // may/must be executed using hardware with only 24-bit precision.
3642 struct Final_Reshape_Counts : public StackObj {
3643 int _call_count; // count non-inlined 'common' calls
3644 int _float_count; // count float ops requiring 24-bit precision
3645 int _double_count; // count double ops requiring more precision
3646 int _java_call_count; // count non-inlined 'java' calls
3647 int _inner_loop_count; // count loops which need alignment
3648 VectorSet _visited; // Visitation flags
3649 Node_List _tests; // Set of IfNodes & PCTableNodes
3650
3651 Final_Reshape_Counts() :
3652 _call_count(0), _float_count(0), _double_count(0),
3653 _java_call_count(0), _inner_loop_count(0) { }
3654
3655 void inc_call_count () { _call_count ++; }
3656 void inc_float_count () { _float_count ++; }
3657 void inc_double_count() { _double_count++; }
3658 void inc_java_call_count() { _java_call_count++; }
3659 void inc_inner_loop_count() { _inner_loop_count++; }
3660
3661 int get_call_count () const { return _call_count ; }
3662 int get_float_count () const { return _float_count ; }
3663 int get_double_count() const { return _double_count; }
3664 int get_java_call_count() const { return _java_call_count; }
3665 int get_inner_loop_count() const { return _inner_loop_count; }
3666 };
3667
3668 //------------------------------final_graph_reshaping_impl----------------------
3669 // Implement items 1-5 from final_graph_reshaping below.
3670 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3671
3672 if ( n->outcnt() == 0 ) return; // dead node
3673 uint nop = n->Opcode();
3674
3675 // Check for 2-input instruction with "last use" on right input.
3676 // Swap to left input. Implements item (2).
3677 if( n->req() == 3 && // two-input instruction
3678 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3679 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3680 n->in(2)->outcnt() == 1 &&// right use IS a last use
3681 !n->in(2)->is_Con() ) { // right use is not a constant
3682 // Check for commutative opcode
3683 switch( nop ) {
3684 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3685 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3686 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3687 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3688 case Op_AndL: case Op_XorL: case Op_OrL:
3689 case Op_AndI: case Op_XorI: case Op_OrI: {
3690 // Move "last use" input to left by swapping inputs
3691 n->swap_edges(1, 2);
3692 break;
3693 }
3694 default:
3695 break;
3696 }
3697 }
3698
3699 #ifdef ASSERT
3700 if( n->is_Mem() ) {
3701 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3702 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3703 // oop will be recorded in oop map if load crosses safepoint
3704 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3705 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3706 "raw memory operations should have control edge");
3707 }
3708 if (n->is_MemBar()) {
3709 MemBarNode* mb = n->as_MemBar();
3710 if (mb->trailing_store() || mb->trailing_load_store()) {
3711 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3712 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3713 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3714 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3715 } else if (mb->leading()) {
3716 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3717 }
3718 }
3719 #endif
3720 // Count FPU ops and common calls, implements item (3)
3721 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3722 if (!gc_handled) {
3723 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3724 }
3725
3726 // Collect CFG split points
3727 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3728 frc._tests.push(n);
3729 }
3730 }
3731
3732 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3733 if (!UseDivMod) {
3734 return;
3735 }
3736
3737 // Check if "a % b" and "a / b" both exist
3738 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3739 if (d == nullptr) {
3740 return;
3741 }
3742
3743 // Replace them with a fused divmod if supported
3744 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3745 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3746 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3747 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3748 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3749 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3750 // DivMod node so the dependency is not lost.
3751 divmod->add_prec_from(n);
3752 divmod->add_prec_from(d);
3753 d->subsume_by(divmod->div_proj(), this);
3754 n->subsume_by(divmod->mod_proj(), this);
3755 } else {
3756 // Replace "a % b" with "a - ((a / b) * b)"
3757 Node* mult = MulNode::make(d, d->in(2), bt);
3758 Node* sub = SubNode::make(d->in(1), mult, bt);
3759 n->subsume_by(sub, this);
3760 }
3761 }
3762
3763 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3764 switch( nop ) {
3765 // Count all float operations that may use FPU
3766 case Op_AddF:
3767 case Op_SubF:
3768 case Op_MulF:
3769 case Op_DivF:
3770 case Op_NegF:
3771 case Op_ModF:
3772 case Op_ConvI2F:
3773 case Op_ConF:
3774 case Op_CmpF:
3775 case Op_CmpF3:
3776 case Op_StoreF:
3777 case Op_LoadF:
3778 // case Op_ConvL2F: // longs are split into 32-bit halves
3779 frc.inc_float_count();
3780 break;
3781
3782 case Op_ConvF2D:
3783 case Op_ConvD2F:
3784 frc.inc_float_count();
3785 frc.inc_double_count();
3786 break;
3787
3788 // Count all double operations that may use FPU
3789 case Op_AddD:
3790 case Op_SubD:
3791 case Op_MulD:
3792 case Op_DivD:
3793 case Op_NegD:
3794 case Op_ModD:
3795 case Op_ConvI2D:
3796 case Op_ConvD2I:
3797 // case Op_ConvL2D: // handled by leaf call
3798 // case Op_ConvD2L: // handled by leaf call
3799 case Op_ConD:
3800 case Op_CmpD:
3801 case Op_CmpD3:
3802 case Op_StoreD:
3803 case Op_LoadD:
3804 case Op_LoadD_unaligned:
3805 frc.inc_double_count();
3806 break;
3807 case Op_Opaque1: // Remove Opaque Nodes before matching
3808 n->subsume_by(n->in(1), this);
3809 break;
3810 case Op_CallStaticJava:
3811 case Op_CallJava:
3812 case Op_CallDynamicJava:
3813 frc.inc_java_call_count(); // Count java call site;
3814 case Op_CallRuntime:
3815 case Op_CallLeaf:
3816 case Op_CallLeafVector:
3817 case Op_CallLeafNoFP: {
3818 assert (n->is_Call(), "");
3819 CallNode *call = n->as_Call();
3820 // Count call sites where the FP mode bit would have to be flipped.
3821 // Do not count uncommon runtime calls:
3822 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3823 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3824 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3825 frc.inc_call_count(); // Count the call site
3826 } else { // See if uncommon argument is shared
3827 Node *n = call->in(TypeFunc::Parms);
3828 int nop = n->Opcode();
3829 // Clone shared simple arguments to uncommon calls, item (1).
3830 if (n->outcnt() > 1 &&
3831 !n->is_Proj() &&
3832 nop != Op_CreateEx &&
3833 nop != Op_CheckCastPP &&
3834 nop != Op_DecodeN &&
3835 nop != Op_DecodeNKlass &&
3836 !n->is_Mem() &&
3837 !n->is_Phi()) {
3838 Node *x = n->clone();
3839 call->set_req(TypeFunc::Parms, x);
3840 }
3841 }
3842 break;
3843 }
3844 case Op_StoreB:
3845 case Op_StoreC:
3846 case Op_StoreI:
3847 case Op_StoreL:
3848 case Op_StoreLSpecial:
3849 case Op_CompareAndSwapB:
3850 case Op_CompareAndSwapS:
3851 case Op_CompareAndSwapI:
3852 case Op_CompareAndSwapL:
3853 case Op_CompareAndSwapP:
3854 case Op_CompareAndSwapN:
3855 case Op_WeakCompareAndSwapB:
3856 case Op_WeakCompareAndSwapS:
3857 case Op_WeakCompareAndSwapI:
3858 case Op_WeakCompareAndSwapL:
3859 case Op_WeakCompareAndSwapP:
3860 case Op_WeakCompareAndSwapN:
3861 case Op_CompareAndExchangeB:
3862 case Op_CompareAndExchangeS:
3863 case Op_CompareAndExchangeI:
3864 case Op_CompareAndExchangeL:
3865 case Op_CompareAndExchangeP:
3866 case Op_CompareAndExchangeN:
3867 case Op_GetAndAddS:
3868 case Op_GetAndAddB:
3869 case Op_GetAndAddI:
3870 case Op_GetAndAddL:
3871 case Op_GetAndSetS:
3872 case Op_GetAndSetB:
3873 case Op_GetAndSetI:
3874 case Op_GetAndSetL:
3875 case Op_GetAndSetP:
3876 case Op_GetAndSetN:
3877 case Op_StoreP:
3878 case Op_StoreN:
3879 case Op_StoreNKlass:
3880 case Op_LoadB:
3881 case Op_LoadUB:
3882 case Op_LoadUS:
3883 case Op_LoadI:
3884 case Op_LoadKlass:
3885 case Op_LoadNKlass:
3886 case Op_LoadL:
3887 case Op_LoadL_unaligned:
3888 case Op_LoadP:
3889 case Op_LoadN:
3890 case Op_LoadRange:
3891 case Op_LoadS:
3892 break;
3893
3894 case Op_AddP: { // Assert sane base pointers
3895 Node *addp = n->in(AddPNode::Address);
3896 assert( !addp->is_AddP() ||
3897 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3898 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3899 "Base pointers must match (addp %u)", addp->_idx );
3900 #ifdef _LP64
3901 if ((UseCompressedOops || UseCompressedClassPointers) &&
3902 addp->Opcode() == Op_ConP &&
3903 addp == n->in(AddPNode::Base) &&
3904 n->in(AddPNode::Offset)->is_Con()) {
3905 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3906 // on the platform and on the compressed oops mode.
3907 // Use addressing with narrow klass to load with offset on x86.
3908 // Some platforms can use the constant pool to load ConP.
3909 // Do this transformation here since IGVN will convert ConN back to ConP.
3910 const Type* t = addp->bottom_type();
3911 bool is_oop = t->isa_oopptr() != nullptr;
3912 bool is_klass = t->isa_klassptr() != nullptr;
3913
3914 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3915 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3916 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3917 Node* nn = nullptr;
3918
3919 int op = is_oop ? Op_ConN : Op_ConNKlass;
3920
3921 // Look for existing ConN node of the same exact type.
3922 Node* r = root();
3923 uint cnt = r->outcnt();
3924 for (uint i = 0; i < cnt; i++) {
3925 Node* m = r->raw_out(i);
3926 if (m!= nullptr && m->Opcode() == op &&
3927 m->bottom_type()->make_ptr() == t) {
3928 nn = m;
3929 break;
3930 }
3931 }
3932 if (nn != nullptr) {
3933 // Decode a narrow oop to match address
3934 // [R12 + narrow_oop_reg<<3 + offset]
3935 if (is_oop) {
3936 nn = new DecodeNNode(nn, t);
3937 } else {
3938 nn = new DecodeNKlassNode(nn, t);
3939 }
3940 // Check for succeeding AddP which uses the same Base.
3941 // Otherwise we will run into the assertion above when visiting that guy.
3942 for (uint i = 0; i < n->outcnt(); ++i) {
3943 Node *out_i = n->raw_out(i);
3944 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3945 out_i->set_req(AddPNode::Base, nn);
3946 #ifdef ASSERT
3947 for (uint j = 0; j < out_i->outcnt(); ++j) {
3948 Node *out_j = out_i->raw_out(j);
3949 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3950 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3951 }
3952 #endif
3953 }
3954 }
3955 n->set_req(AddPNode::Base, nn);
3956 n->set_req(AddPNode::Address, nn);
3957 if (addp->outcnt() == 0) {
3958 addp->disconnect_inputs(this);
3959 }
3960 }
3961 }
3962 }
3963 #endif
3964 break;
3965 }
3966
3967 case Op_CastPP: {
3968 // Remove CastPP nodes to gain more freedom during scheduling but
3969 // keep the dependency they encode as control or precedence edges
3970 // (if control is set already) on memory operations. Some CastPP
3971 // nodes don't have a control (don't carry a dependency): skip
3972 // those.
3973 if (n->in(0) != nullptr) {
3974 ResourceMark rm;
3975 Unique_Node_List wq;
3976 wq.push(n);
3977 for (uint next = 0; next < wq.size(); ++next) {
3978 Node *m = wq.at(next);
3979 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3980 Node* use = m->fast_out(i);
3981 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3982 use->ensure_control_or_add_prec(n->in(0));
3983 } else {
3984 switch(use->Opcode()) {
3985 case Op_AddP:
3986 case Op_DecodeN:
3987 case Op_DecodeNKlass:
3988 case Op_CheckCastPP:
3989 case Op_CastPP:
3990 wq.push(use);
3991 break;
3992 }
3993 }
3994 }
3995 }
3996 }
3997 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3998 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3999 Node* in1 = n->in(1);
4000 const Type* t = n->bottom_type();
4001 Node* new_in1 = in1->clone();
4002 new_in1->as_DecodeN()->set_type(t);
4003
4004 if (!Matcher::narrow_oop_use_complex_address()) {
4005 //
4006 // x86, ARM and friends can handle 2 adds in addressing mode
4007 // and Matcher can fold a DecodeN node into address by using
4008 // a narrow oop directly and do implicit null check in address:
4009 //
4010 // [R12 + narrow_oop_reg<<3 + offset]
4011 // NullCheck narrow_oop_reg
4012 //
4013 // On other platforms (Sparc) we have to keep new DecodeN node and
4014 // use it to do implicit null check in address:
4015 //
4016 // decode_not_null narrow_oop_reg, base_reg
4017 // [base_reg + offset]
4018 // NullCheck base_reg
4019 //
4020 // Pin the new DecodeN node to non-null path on these platform (Sparc)
4021 // to keep the information to which null check the new DecodeN node
4022 // corresponds to use it as value in implicit_null_check().
4023 //
4024 new_in1->set_req(0, n->in(0));
4025 }
4026
4027 n->subsume_by(new_in1, this);
4028 if (in1->outcnt() == 0) {
4029 in1->disconnect_inputs(this);
4030 }
4031 } else {
4032 n->subsume_by(n->in(1), this);
4033 if (n->outcnt() == 0) {
4034 n->disconnect_inputs(this);
4035 }
4036 }
4037 break;
4038 }
4039 case Op_CastII: {
4040 n->as_CastII()->remove_range_check_cast(this);
4041 break;
4042 }
4043 #ifdef _LP64
4044 case Op_CmpP:
4045 // Do this transformation here to preserve CmpPNode::sub() and
4046 // other TypePtr related Ideal optimizations (for example, ptr nullness).
4047 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
4048 Node* in1 = n->in(1);
4049 Node* in2 = n->in(2);
4050 if (!in1->is_DecodeNarrowPtr()) {
4051 in2 = in1;
4052 in1 = n->in(2);
4053 }
4054 assert(in1->is_DecodeNarrowPtr(), "sanity");
4055
4056 Node* new_in2 = nullptr;
4057 if (in2->is_DecodeNarrowPtr()) {
4058 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
4059 new_in2 = in2->in(1);
4060 } else if (in2->Opcode() == Op_ConP) {
4061 const Type* t = in2->bottom_type();
4062 if (t == TypePtr::NULL_PTR) {
4063 assert(in1->is_DecodeN(), "compare klass to null?");
4064 // Don't convert CmpP null check into CmpN if compressed
4065 // oops implicit null check is not generated.
4066 // This will allow to generate normal oop implicit null check.
4067 if (Matcher::gen_narrow_oop_implicit_null_checks())
4068 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
4069 //
4070 // This transformation together with CastPP transformation above
4071 // will generated code for implicit null checks for compressed oops.
4072 //
4073 // The original code after Optimize()
4074 //
4075 // LoadN memory, narrow_oop_reg
4076 // decode narrow_oop_reg, base_reg
4077 // CmpP base_reg, nullptr
4078 // CastPP base_reg // NotNull
4079 // Load [base_reg + offset], val_reg
4080 //
4081 // after these transformations will be
4082 //
4083 // LoadN memory, narrow_oop_reg
4084 // CmpN narrow_oop_reg, nullptr
4085 // decode_not_null narrow_oop_reg, base_reg
4086 // Load [base_reg + offset], val_reg
4087 //
4088 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
4089 // since narrow oops can be used in debug info now (see the code in
4090 // final_graph_reshaping_walk()).
4091 //
4092 // At the end the code will be matched to
4093 // on x86:
4094 //
4095 // Load_narrow_oop memory, narrow_oop_reg
4096 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
4097 // NullCheck narrow_oop_reg
4098 //
4099 // and on sparc:
4100 //
4101 // Load_narrow_oop memory, narrow_oop_reg
4102 // decode_not_null narrow_oop_reg, base_reg
4103 // Load [base_reg + offset], val_reg
4104 // NullCheck base_reg
4105 //
4106 } else if (t->isa_oopptr()) {
4107 new_in2 = ConNode::make(t->make_narrowoop());
4108 } else if (t->isa_klassptr()) {
4109 new_in2 = ConNode::make(t->make_narrowklass());
4110 }
4111 }
4112 if (new_in2 != nullptr) {
4113 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
4114 n->subsume_by(cmpN, this);
4115 if (in1->outcnt() == 0) {
4116 in1->disconnect_inputs(this);
4117 }
4118 if (in2->outcnt() == 0) {
4119 in2->disconnect_inputs(this);
4120 }
4121 }
4122 }
4123 break;
4124
4125 case Op_DecodeN:
4126 case Op_DecodeNKlass:
4127 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4128 // DecodeN could be pinned when it can't be fold into
4129 // an address expression, see the code for Op_CastPP above.
4130 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4131 break;
4132
4133 case Op_EncodeP:
4134 case Op_EncodePKlass: {
4135 Node* in1 = n->in(1);
4136 if (in1->is_DecodeNarrowPtr()) {
4137 n->subsume_by(in1->in(1), this);
4138 } else if (in1->Opcode() == Op_ConP) {
4139 const Type* t = in1->bottom_type();
4140 if (t == TypePtr::NULL_PTR) {
4141 assert(t->isa_oopptr(), "null klass?");
4142 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4143 } else if (t->isa_oopptr()) {
4144 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4145 } else if (t->isa_klassptr()) {
4146 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4147 }
4148 }
4149 if (in1->outcnt() == 0) {
4150 in1->disconnect_inputs(this);
4151 }
4152 break;
4153 }
4154
4155 case Op_Proj: {
4156 if (OptimizeStringConcat || IncrementalInline) {
4157 ProjNode* proj = n->as_Proj();
4158 if (proj->_is_io_use) {
4159 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4160 // Separate projections were used for the exception path which
4161 // are normally removed by a late inline. If it wasn't inlined
4162 // then they will hang around and should just be replaced with
4163 // the original one. Merge them.
4164 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4165 if (non_io_proj != nullptr) {
4166 proj->subsume_by(non_io_proj , this);
4167 }
4168 }
4169 }
4170 break;
4171 }
4172
4173 case Op_Phi:
4174 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4175 // The EncodeP optimization may create Phi with the same edges
4176 // for all paths. It is not handled well by Register Allocator.
4177 Node* unique_in = n->in(1);
4178 assert(unique_in != nullptr, "");
4179 uint cnt = n->req();
4180 for (uint i = 2; i < cnt; i++) {
4181 Node* m = n->in(i);
4182 assert(m != nullptr, "");
4183 if (unique_in != m)
4184 unique_in = nullptr;
4185 }
4186 if (unique_in != nullptr) {
4187 n->subsume_by(unique_in, this);
4188 }
4189 }
4190 break;
4191
4192 #endif
4193
4194 case Op_ModI:
4195 handle_div_mod_op(n, T_INT, false);
4196 break;
4197
4198 case Op_ModL:
4199 handle_div_mod_op(n, T_LONG, false);
4200 break;
4201
4202 case Op_UModI:
4203 handle_div_mod_op(n, T_INT, true);
4204 break;
4205
4206 case Op_UModL:
4207 handle_div_mod_op(n, T_LONG, true);
4208 break;
4209
4210 case Op_LoadVector:
4211 case Op_StoreVector:
4212 #ifdef ASSERT
4213 // Add VerifyVectorAlignment node between adr and load / store.
4214 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
4215 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
4216 n->as_StoreVector()->must_verify_alignment();
4217 if (must_verify_alignment) {
4218 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
4219 n->as_StoreVector()->memory_size();
4220 // The memory access should be aligned to the vector width in bytes.
4221 // However, the underlying array is possibly less well aligned, but at least
4222 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
4223 // a loop we can expect at least the following alignment:
4224 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
4225 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
4226 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
4227 // Create mask from alignment. e.g. 0b1000 -> 0b0111
4228 jlong mask = guaranteed_alignment - 1;
4229 Node* mask_con = ConLNode::make(mask);
4230 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
4231 n->set_req(MemNode::Address, va);
4232 }
4233 }
4234 #endif
4235 break;
4236
4237 case Op_LoadVectorGather:
4238 case Op_StoreVectorScatter:
4239 case Op_LoadVectorGatherMasked:
4240 case Op_StoreVectorScatterMasked:
4241 case Op_VectorCmpMasked:
4242 case Op_VectorMaskGen:
4243 case Op_LoadVectorMasked:
4244 case Op_StoreVectorMasked:
4245 break;
4246
4247 case Op_AddReductionVI:
4248 case Op_AddReductionVL:
4249 case Op_AddReductionVF:
4250 case Op_AddReductionVD:
4251 case Op_MulReductionVI:
4252 case Op_MulReductionVL:
4253 case Op_MulReductionVF:
4254 case Op_MulReductionVD:
4255 case Op_MinReductionV:
4256 case Op_MaxReductionV:
4257 case Op_AndReductionV:
4258 case Op_OrReductionV:
4259 case Op_XorReductionV:
4260 break;
4261
4262 case Op_PackB:
4263 case Op_PackS:
4264 case Op_PackI:
4265 case Op_PackF:
4266 case Op_PackL:
4267 case Op_PackD:
4268 if (n->req()-1 > 2) {
4269 // Replace many operand PackNodes with a binary tree for matching
4270 PackNode* p = (PackNode*) n;
4271 Node* btp = p->binary_tree_pack(1, n->req());
4272 n->subsume_by(btp, this);
4273 }
4274 break;
4275 case Op_Loop:
4276 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4277 case Op_CountedLoop:
4278 case Op_LongCountedLoop:
4279 case Op_OuterStripMinedLoop:
4280 if (n->as_Loop()->is_inner_loop()) {
4281 frc.inc_inner_loop_count();
4282 }
4283 n->as_Loop()->verify_strip_mined(0);
4284 break;
4285 case Op_LShiftI:
4286 case Op_RShiftI:
4287 case Op_URShiftI:
4288 case Op_LShiftL:
4289 case Op_RShiftL:
4290 case Op_URShiftL:
4291 if (Matcher::need_masked_shift_count) {
4292 // The cpu's shift instructions don't restrict the count to the
4293 // lower 5/6 bits. We need to do the masking ourselves.
4294 Node* in2 = n->in(2);
4295 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4296 const TypeInt* t = in2->find_int_type();
4297 if (t != nullptr && t->is_con()) {
4298 juint shift = t->get_con();
4299 if (shift > mask) { // Unsigned cmp
4300 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4301 }
4302 } else {
4303 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
4304 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4305 n->set_req(2, shift);
4306 }
4307 }
4308 if (in2->outcnt() == 0) { // Remove dead node
4309 in2->disconnect_inputs(this);
4310 }
4311 }
4312 break;
4313 case Op_MemBarStoreStore:
4314 case Op_MemBarRelease:
4315 // Break the link with AllocateNode: it is no longer useful and
4316 // confuses register allocation.
4317 if (n->req() > MemBarNode::Precedent) {
4318 n->set_req(MemBarNode::Precedent, top());
4319 }
4320 break;
4321 case Op_MemBarAcquire: {
4322 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4323 // At parse time, the trailing MemBarAcquire for a volatile load
4324 // is created with an edge to the load. After optimizations,
4325 // that input may be a chain of Phis. If those phis have no
4326 // other use, then the MemBarAcquire keeps them alive and
4327 // register allocation can be confused.
4328 dead_nodes.push(n->in(MemBarNode::Precedent));
4329 n->set_req(MemBarNode::Precedent, top());
4330 }
4331 break;
4332 }
4333 case Op_Blackhole:
4334 break;
4335 case Op_RangeCheck: {
4336 RangeCheckNode* rc = n->as_RangeCheck();
4337 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4338 n->subsume_by(iff, this);
4339 frc._tests.push(iff);
4340 break;
4341 }
4342 case Op_ConvI2L: {
4343 if (!Matcher::convi2l_type_required) {
4344 // Code generation on some platforms doesn't need accurate
4345 // ConvI2L types. Widening the type can help remove redundant
4346 // address computations.
4347 n->as_Type()->set_type(TypeLong::INT);
4348 ResourceMark rm;
4349 Unique_Node_List wq;
4350 wq.push(n);
4351 for (uint next = 0; next < wq.size(); next++) {
4352 Node *m = wq.at(next);
4353
4354 for(;;) {
4355 // Loop over all nodes with identical inputs edges as m
4356 Node* k = m->find_similar(m->Opcode());
4357 if (k == nullptr) {
4358 break;
4359 }
4360 // Push their uses so we get a chance to remove node made
4361 // redundant
4362 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4363 Node* u = k->fast_out(i);
4364 if (u->Opcode() == Op_LShiftL ||
4365 u->Opcode() == Op_AddL ||
4366 u->Opcode() == Op_SubL ||
4367 u->Opcode() == Op_AddP) {
4368 wq.push(u);
4369 }
4370 }
4371 // Replace all nodes with identical edges as m with m
4372 k->subsume_by(m, this);
4373 }
4374 }
4375 }
4376 break;
4377 }
4378 case Op_CmpUL: {
4379 if (!Matcher::has_match_rule(Op_CmpUL)) {
4380 // No support for unsigned long comparisons
4381 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4382 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4383 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4384 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4385 Node* andl = new AndLNode(orl, remove_sign_mask);
4386 Node* cmp = new CmpLNode(andl, n->in(2));
4387 n->subsume_by(cmp, this);
4388 }
4389 break;
4390 }
4391 #ifdef ASSERT
4392 case Op_InlineType: {
4393 n->dump(-1);
4394 assert(false, "inline type node was not removed");
4395 break;
4396 }
4397 case Op_ConNKlass: {
4398 const TypePtr* tp = n->as_Type()->type()->make_ptr();
4399 ciKlass* klass = tp->is_klassptr()->exact_klass();
4400 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
4401 break;
4402 }
4403 #endif
4404 default:
4405 assert(!n->is_Call(), "");
4406 assert(!n->is_Mem(), "");
4407 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4408 break;
4409 }
4410 }
4411
4412 //------------------------------final_graph_reshaping_walk---------------------
4413 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4414 // requires that the walk visits a node's inputs before visiting the node.
4415 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4416 Unique_Node_List sfpt;
4417
4418 frc._visited.set(root->_idx); // first, mark node as visited
4419 uint cnt = root->req();
4420 Node *n = root;
4421 uint i = 0;
4422 while (true) {
4423 if (i < cnt) {
4424 // Place all non-visited non-null inputs onto stack
4425 Node* m = n->in(i);
4426 ++i;
4427 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
4428 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
4429 // compute worst case interpreter size in case of a deoptimization
4430 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4431
4432 sfpt.push(m);
4433 }
4434 cnt = m->req();
4435 nstack.push(n, i); // put on stack parent and next input's index
4436 n = m;
4437 i = 0;
4438 }
4439 } else {
4440 // Now do post-visit work
4441 final_graph_reshaping_impl(n, frc, dead_nodes);
4442 if (nstack.is_empty())
4443 break; // finished
4444 n = nstack.node(); // Get node from stack
4445 cnt = n->req();
4446 i = nstack.index();
4447 nstack.pop(); // Shift to the next node on stack
4448 }
4449 }
4450
4451 // Skip next transformation if compressed oops are not used.
4452 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4453 (!UseCompressedOops && !UseCompressedClassPointers))
4454 return;
4455
4456 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4457 // It could be done for an uncommon traps or any safepoints/calls
4458 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4459 while (sfpt.size() > 0) {
4460 n = sfpt.pop();
4461 JVMState *jvms = n->as_SafePoint()->jvms();
4462 assert(jvms != nullptr, "sanity");
4463 int start = jvms->debug_start();
4464 int end = n->req();
4465 bool is_uncommon = (n->is_CallStaticJava() &&
4466 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4467 for (int j = start; j < end; j++) {
4468 Node* in = n->in(j);
4469 if (in->is_DecodeNarrowPtr()) {
4470 bool safe_to_skip = true;
4471 if (!is_uncommon ) {
4472 // Is it safe to skip?
4473 for (uint i = 0; i < in->outcnt(); i++) {
4474 Node* u = in->raw_out(i);
4475 if (!u->is_SafePoint() ||
4476 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4477 safe_to_skip = false;
4478 }
4479 }
4480 }
4481 if (safe_to_skip) {
4482 n->set_req(j, in->in(1));
4483 }
4484 if (in->outcnt() == 0) {
4485 in->disconnect_inputs(this);
4486 }
4487 }
4488 }
4489 }
4490 }
4491
4492 //------------------------------final_graph_reshaping--------------------------
4493 // Final Graph Reshaping.
4494 //
4495 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4496 // and not commoned up and forced early. Must come after regular
4497 // optimizations to avoid GVN undoing the cloning. Clone constant
4498 // inputs to Loop Phis; these will be split by the allocator anyways.
4499 // Remove Opaque nodes.
4500 // (2) Move last-uses by commutative operations to the left input to encourage
4501 // Intel update-in-place two-address operations and better register usage
4502 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4503 // calls canonicalizing them back.
4504 // (3) Count the number of double-precision FP ops, single-precision FP ops
4505 // and call sites. On Intel, we can get correct rounding either by
4506 // forcing singles to memory (requires extra stores and loads after each
4507 // FP bytecode) or we can set a rounding mode bit (requires setting and
4508 // clearing the mode bit around call sites). The mode bit is only used
4509 // if the relative frequency of single FP ops to calls is low enough.
4510 // This is a key transform for SPEC mpeg_audio.
4511 // (4) Detect infinite loops; blobs of code reachable from above but not
4512 // below. Several of the Code_Gen algorithms fail on such code shapes,
4513 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4514 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4515 // Detection is by looking for IfNodes where only 1 projection is
4516 // reachable from below or CatchNodes missing some targets.
4517 // (5) Assert for insane oop offsets in debug mode.
4518
4519 bool Compile::final_graph_reshaping() {
4520 // an infinite loop may have been eliminated by the optimizer,
4521 // in which case the graph will be empty.
4522 if (root()->req() == 1) {
4523 // Do not compile method that is only a trivial infinite loop,
4524 // since the content of the loop may have been eliminated.
4525 record_method_not_compilable("trivial infinite loop");
4526 return true;
4527 }
4528
4529 // Expensive nodes have their control input set to prevent the GVN
4530 // from freely commoning them. There's no GVN beyond this point so
4531 // no need to keep the control input. We want the expensive nodes to
4532 // be freely moved to the least frequent code path by gcm.
4533 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4534 for (int i = 0; i < expensive_count(); i++) {
4535 _expensive_nodes.at(i)->set_req(0, nullptr);
4536 }
4537
4538 Final_Reshape_Counts frc;
4539
4540 // Visit everybody reachable!
4541 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4542 Node_Stack nstack(live_nodes() >> 1);
4543 Unique_Node_List dead_nodes;
4544 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4545
4546 // Check for unreachable (from below) code (i.e., infinite loops).
4547 for( uint i = 0; i < frc._tests.size(); i++ ) {
4548 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4549 // Get number of CFG targets.
4550 // Note that PCTables include exception targets after calls.
4551 uint required_outcnt = n->required_outcnt();
4552 if (n->outcnt() != required_outcnt) {
4553 // Check for a few special cases. Rethrow Nodes never take the
4554 // 'fall-thru' path, so expected kids is 1 less.
4555 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4556 if (n->in(0)->in(0)->is_Call()) {
4557 CallNode* call = n->in(0)->in(0)->as_Call();
4558 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4559 required_outcnt--; // Rethrow always has 1 less kid
4560 } else if (call->req() > TypeFunc::Parms &&
4561 call->is_CallDynamicJava()) {
4562 // Check for null receiver. In such case, the optimizer has
4563 // detected that the virtual call will always result in a null
4564 // pointer exception. The fall-through projection of this CatchNode
4565 // will not be populated.
4566 Node* arg0 = call->in(TypeFunc::Parms);
4567 if (arg0->is_Type() &&
4568 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4569 required_outcnt--;
4570 }
4571 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4572 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4573 // Check for illegal array length. In such case, the optimizer has
4574 // detected that the allocation attempt will always result in an
4575 // exception. There is no fall-through projection of this CatchNode .
4576 assert(call->is_CallStaticJava(), "static call expected");
4577 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4578 uint valid_length_test_input = call->req() - 1;
4579 Node* valid_length_test = call->in(valid_length_test_input);
4580 call->del_req(valid_length_test_input);
4581 if (valid_length_test->find_int_con(1) == 0) {
4582 required_outcnt--;
4583 }
4584 dead_nodes.push(valid_length_test);
4585 assert(n->outcnt() == required_outcnt, "malformed control flow");
4586 continue;
4587 }
4588 }
4589 }
4590
4591 // Recheck with a better notion of 'required_outcnt'
4592 if (n->outcnt() != required_outcnt) {
4593 record_method_not_compilable("malformed control flow");
4594 return true; // Not all targets reachable!
4595 }
4596 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4597 CallNode* call = n->in(0)->in(0)->as_Call();
4598 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4599 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4600 assert(call->is_CallStaticJava(), "static call expected");
4601 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4602 uint valid_length_test_input = call->req() - 1;
4603 dead_nodes.push(call->in(valid_length_test_input));
4604 call->del_req(valid_length_test_input); // valid length test useless now
4605 }
4606 }
4607 // Check that I actually visited all kids. Unreached kids
4608 // must be infinite loops.
4609 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4610 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4611 record_method_not_compilable("infinite loop");
4612 return true; // Found unvisited kid; must be unreach
4613 }
4614
4615 // Here so verification code in final_graph_reshaping_walk()
4616 // always see an OuterStripMinedLoopEnd
4617 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4618 IfNode* init_iff = n->as_If();
4619 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4620 n->subsume_by(iff, this);
4621 }
4622 }
4623
4624 while (dead_nodes.size() > 0) {
4625 Node* m = dead_nodes.pop();
4626 if (m->outcnt() == 0 && m != top()) {
4627 for (uint j = 0; j < m->req(); j++) {
4628 Node* in = m->in(j);
4629 if (in != nullptr) {
4630 dead_nodes.push(in);
4631 }
4632 }
4633 m->disconnect_inputs(this);
4634 }
4635 }
4636
4637 set_java_calls(frc.get_java_call_count());
4638 set_inner_loops(frc.get_inner_loop_count());
4639
4640 // No infinite loops, no reason to bail out.
4641 return false;
4642 }
4643
4644 //-----------------------------too_many_traps----------------------------------
4645 // Report if there are too many traps at the current method and bci.
4646 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4647 bool Compile::too_many_traps(ciMethod* method,
4648 int bci,
4649 Deoptimization::DeoptReason reason) {
4650 ciMethodData* md = method->method_data();
4651 if (md->is_empty()) {
4652 // Assume the trap has not occurred, or that it occurred only
4653 // because of a transient condition during start-up in the interpreter.
4654 return false;
4655 }
4656 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4657 if (md->has_trap_at(bci, m, reason) != 0) {
4658 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4659 // Also, if there are multiple reasons, or if there is no per-BCI record,
4660 // assume the worst.
4661 if (log())
4662 log()->elem("observe trap='%s' count='%d'",
4663 Deoptimization::trap_reason_name(reason),
4664 md->trap_count(reason));
4665 return true;
4666 } else {
4667 // Ignore method/bci and see if there have been too many globally.
4668 return too_many_traps(reason, md);
4669 }
4670 }
4671
4672 // Less-accurate variant which does not require a method and bci.
4673 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4674 ciMethodData* logmd) {
4675 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4676 // Too many traps globally.
4677 // Note that we use cumulative trap_count, not just md->trap_count.
4678 if (log()) {
4679 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4680 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4681 Deoptimization::trap_reason_name(reason),
4682 mcount, trap_count(reason));
4683 }
4684 return true;
4685 } else {
4686 // The coast is clear.
4687 return false;
4688 }
4689 }
4690
4691 //--------------------------too_many_recompiles--------------------------------
4692 // Report if there are too many recompiles at the current method and bci.
4693 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4694 // Is not eager to return true, since this will cause the compiler to use
4695 // Action_none for a trap point, to avoid too many recompilations.
4696 bool Compile::too_many_recompiles(ciMethod* method,
4697 int bci,
4698 Deoptimization::DeoptReason reason) {
4699 ciMethodData* md = method->method_data();
4700 if (md->is_empty()) {
4701 // Assume the trap has not occurred, or that it occurred only
4702 // because of a transient condition during start-up in the interpreter.
4703 return false;
4704 }
4705 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4706 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4707 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4708 Deoptimization::DeoptReason per_bc_reason
4709 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4710 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4711 if ((per_bc_reason == Deoptimization::Reason_none
4712 || md->has_trap_at(bci, m, reason) != 0)
4713 // The trap frequency measure we care about is the recompile count:
4714 && md->trap_recompiled_at(bci, m)
4715 && md->overflow_recompile_count() >= bc_cutoff) {
4716 // Do not emit a trap here if it has already caused recompilations.
4717 // Also, if there are multiple reasons, or if there is no per-BCI record,
4718 // assume the worst.
4719 if (log())
4720 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4721 Deoptimization::trap_reason_name(reason),
4722 md->trap_count(reason),
4723 md->overflow_recompile_count());
4724 return true;
4725 } else if (trap_count(reason) != 0
4726 && decompile_count() >= m_cutoff) {
4727 // Too many recompiles globally, and we have seen this sort of trap.
4728 // Use cumulative decompile_count, not just md->decompile_count.
4729 if (log())
4730 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4731 Deoptimization::trap_reason_name(reason),
4732 md->trap_count(reason), trap_count(reason),
4733 md->decompile_count(), decompile_count());
4734 return true;
4735 } else {
4736 // The coast is clear.
4737 return false;
4738 }
4739 }
4740
4741 // Compute when not to trap. Used by matching trap based nodes and
4742 // NullCheck optimization.
4743 void Compile::set_allowed_deopt_reasons() {
4744 _allowed_reasons = 0;
4745 if (is_method_compilation()) {
4746 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4747 assert(rs < BitsPerInt, "recode bit map");
4748 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4749 _allowed_reasons |= nth_bit(rs);
4750 }
4751 }
4752 }
4753 }
4754
4755 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4756 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4757 }
4758
4759 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4760 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4761 }
4762
4763 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4764 if (holder->is_initialized()) {
4765 return false;
4766 }
4767 if (holder->is_being_initialized()) {
4768 if (accessing_method->holder() == holder) {
4769 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4770 // <init>, or a static method. In all those cases, there was an initialization
4771 // barrier on the holder klass passed.
4772 if (accessing_method->is_class_initializer() ||
4773 accessing_method->is_object_constructor() ||
4774 accessing_method->is_static()) {
4775 return false;
4776 }
4777 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4778 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4779 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4780 // child class can become fully initialized while its parent class is still being initialized.
4781 if (accessing_method->is_class_initializer()) {
4782 return false;
4783 }
4784 }
4785 ciMethod* root = method(); // the root method of compilation
4786 if (root != accessing_method) {
4787 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4788 }
4789 }
4790 return true;
4791 }
4792
4793 #ifndef PRODUCT
4794 //------------------------------verify_bidirectional_edges---------------------
4795 // For each input edge to a node (ie - for each Use-Def edge), verify that
4796 // there is a corresponding Def-Use edge.
4797 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4798 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4799 uint stack_size = live_nodes() >> 4;
4800 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4801 if (root_and_safepoints != nullptr) {
4802 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4803 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4804 Node* root_or_safepoint = root_and_safepoints->at(i);
4805 // If the node is a safepoint, let's check if it still has a control input
4806 // Lack of control input signifies that this node was killed by CCP or
4807 // recursively by remove_globally_dead_node and it shouldn't be a starting
4808 // point.
4809 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4810 nstack.push(root_or_safepoint);
4811 }
4812 }
4813 } else {
4814 nstack.push(_root);
4815 }
4816
4817 while (nstack.size() > 0) {
4818 Node* n = nstack.pop();
4819 if (visited.member(n)) {
4820 continue;
4821 }
4822 visited.push(n);
4823
4824 // Walk over all input edges, checking for correspondence
4825 uint length = n->len();
4826 for (uint i = 0; i < length; i++) {
4827 Node* in = n->in(i);
4828 if (in != nullptr && !visited.member(in)) {
4829 nstack.push(in); // Put it on stack
4830 }
4831 if (in != nullptr && !in->is_top()) {
4832 // Count instances of `next`
4833 int cnt = 0;
4834 for (uint idx = 0; idx < in->_outcnt; idx++) {
4835 if (in->_out[idx] == n) {
4836 cnt++;
4837 }
4838 }
4839 assert(cnt > 0, "Failed to find Def-Use edge.");
4840 // Check for duplicate edges
4841 // walk the input array downcounting the input edges to n
4842 for (uint j = 0; j < length; j++) {
4843 if (n->in(j) == in) {
4844 cnt--;
4845 }
4846 }
4847 assert(cnt == 0, "Mismatched edge count.");
4848 } else if (in == nullptr) {
4849 assert(i == 0 || i >= n->req() ||
4850 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4851 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
4852 (n->is_Unlock() && i == (n->req() - 1)) ||
4853 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4854 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
4855 } else {
4856 assert(in->is_top(), "sanity");
4857 // Nothing to check.
4858 }
4859 }
4860 }
4861 }
4862
4863 //------------------------------verify_graph_edges---------------------------
4864 // Walk the Graph and verify that there is a one-to-one correspondence
4865 // between Use-Def edges and Def-Use edges in the graph.
4866 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4867 if (VerifyGraphEdges) {
4868 Unique_Node_List visited;
4869
4870 // Call graph walk to check edges
4871 verify_bidirectional_edges(visited, root_and_safepoints);
4872 if (no_dead_code) {
4873 // Now make sure that no visited node is used by an unvisited node.
4874 bool dead_nodes = false;
4875 Unique_Node_List checked;
4876 while (visited.size() > 0) {
4877 Node* n = visited.pop();
4878 checked.push(n);
4879 for (uint i = 0; i < n->outcnt(); i++) {
4880 Node* use = n->raw_out(i);
4881 if (checked.member(use)) continue; // already checked
4882 if (visited.member(use)) continue; // already in the graph
4883 if (use->is_Con()) continue; // a dead ConNode is OK
4884 // At this point, we have found a dead node which is DU-reachable.
4885 if (!dead_nodes) {
4886 tty->print_cr("*** Dead nodes reachable via DU edges:");
4887 dead_nodes = true;
4888 }
4889 use->dump(2);
4890 tty->print_cr("---");
4891 checked.push(use); // No repeats; pretend it is now checked.
4892 }
4893 }
4894 assert(!dead_nodes, "using nodes must be reachable from root");
4895 }
4896 }
4897 }
4898 #endif
4899
4900 // The Compile object keeps track of failure reasons separately from the ciEnv.
4901 // This is required because there is not quite a 1-1 relation between the
4902 // ciEnv and its compilation task and the Compile object. Note that one
4903 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4904 // to backtrack and retry without subsuming loads. Other than this backtracking
4905 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4906 // by the logic in C2Compiler.
4907 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4908 if (log() != nullptr) {
4909 log()->elem("failure reason='%s' phase='compile'", reason);
4910 }
4911 if (_failure_reason.get() == nullptr) {
4912 // Record the first failure reason.
4913 _failure_reason.set(reason);
4914 if (CaptureBailoutInformation) {
4915 _first_failure_details = new CompilationFailureInfo(reason);
4916 }
4917 } else {
4918 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4919 }
4920
4921 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4922 C->print_method(PHASE_FAILURE, 1);
4923 }
4924 _root = nullptr; // flush the graph, too
4925 }
4926
4927 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4928 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4929 _compile(Compile::current()),
4930 _log(nullptr),
4931 _dolog(CITimeVerbose)
4932 {
4933 assert(_compile != nullptr, "sanity check");
4934 assert(id != PhaseTraceId::_t_none, "Don't use none");
4935 if (_dolog) {
4936 _log = _compile->log();
4937 }
4938 if (_log != nullptr) {
4939 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4940 _log->stamp();
4941 _log->end_head();
4942 }
4943
4944 // Inform memory statistic, if enabled
4945 if (CompilationMemoryStatistic::enabled()) {
4946 CompilationMemoryStatistic::on_phase_start((int)id, name);
4947 }
4948 }
4949
4950 Compile::TracePhase::TracePhase(PhaseTraceId id)
4951 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4952
4953 Compile::TracePhase::~TracePhase() {
4954
4955 // Inform memory statistic, if enabled
4956 if (CompilationMemoryStatistic::enabled()) {
4957 CompilationMemoryStatistic::on_phase_end();
4958 }
4959
4960 if (_compile->failing_internal()) {
4961 if (_log != nullptr) {
4962 _log->done("phase");
4963 }
4964 return; // timing code, not stressing bailouts.
4965 }
4966 #ifdef ASSERT
4967 if (PrintIdealNodeCount) {
4968 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4969 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4970 }
4971
4972 if (VerifyIdealNodeCount) {
4973 _compile->print_missing_nodes();
4974 }
4975 #endif
4976
4977 if (_log != nullptr) {
4978 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4979 }
4980 }
4981
4982 //----------------------------static_subtype_check-----------------------------
4983 // Shortcut important common cases when superklass is exact:
4984 // (0) superklass is java.lang.Object (can occur in reflective code)
4985 // (1) subklass is already limited to a subtype of superklass => always ok
4986 // (2) subklass does not overlap with superklass => always fail
4987 // (3) superklass has NO subtypes and we can check with a simple compare.
4988 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4989 if (skip) {
4990 return SSC_full_test; // Let caller generate the general case.
4991 }
4992
4993 if (subk->is_java_subtype_of(superk)) {
4994 return SSC_always_true; // (0) and (1) this test cannot fail
4995 }
4996
4997 if (!subk->maybe_java_subtype_of(superk)) {
4998 return SSC_always_false; // (2) true path dead; no dynamic test needed
4999 }
5000
5001 const Type* superelem = superk;
5002 if (superk->isa_aryklassptr()) {
5003 int ignored;
5004 superelem = superk->is_aryklassptr()->base_element_type(ignored);
5005
5006 // Do not fold the subtype check to an array klass pointer comparison for null-able inline type arrays
5007 // because null-free [LMyValue <: null-able [LMyValue but the klasses are different. Perform a full test.
5008 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
5009 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
5010 return SSC_full_test;
5011 }
5012 }
5013
5014 if (superelem->isa_instklassptr()) {
5015 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
5016 if (!ik->has_subklass()) {
5017 if (!ik->is_final()) {
5018 // Add a dependency if there is a chance of a later subclass.
5019 dependencies()->assert_leaf_type(ik);
5020 }
5021 if (!superk->maybe_java_subtype_of(subk)) {
5022 return SSC_always_false;
5023 }
5024 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5025 }
5026 } else {
5027 // A primitive array type has no subtypes.
5028 return SSC_easy_test; // (3) caller can do a simple ptr comparison
5029 }
5030
5031 return SSC_full_test;
5032 }
5033
5034 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
5035 #ifdef _LP64
5036 // The scaled index operand to AddP must be a clean 64-bit value.
5037 // Java allows a 32-bit int to be incremented to a negative
5038 // value, which appears in a 64-bit register as a large
5039 // positive number. Using that large positive number as an
5040 // operand in pointer arithmetic has bad consequences.
5041 // On the other hand, 32-bit overflow is rare, and the possibility
5042 // can often be excluded, if we annotate the ConvI2L node with
5043 // a type assertion that its value is known to be a small positive
5044 // number. (The prior range check has ensured this.)
5045 // This assertion is used by ConvI2LNode::Ideal.
5046 int index_max = max_jint - 1; // array size is max_jint, index is one less
5047 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
5048 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
5049 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
5050 #endif
5051 return idx;
5052 }
5053
5054 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
5055 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
5056 if (ctrl != nullptr) {
5057 // Express control dependency by a CastII node with a narrow type.
5058 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
5059 // node from floating above the range check during loop optimizations. Otherwise, the
5060 // ConvI2L node may be eliminated independently of the range check, causing the data path
5061 // to become TOP while the control path is still there (although it's unreachable).
5062 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
5063 value = phase->transform(value);
5064 }
5065 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
5066 return phase->transform(new ConvI2LNode(value, ltype));
5067 }
5068
5069 void Compile::dump_print_inlining() {
5070 inline_printer()->print_on(tty);
5071 }
5072
5073 void Compile::log_late_inline(CallGenerator* cg) {
5074 if (log() != nullptr) {
5075 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5076 cg->unique_id());
5077 JVMState* p = cg->call_node()->jvms();
5078 while (p != nullptr) {
5079 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5080 p = p->caller();
5081 }
5082 log()->tail("late_inline");
5083 }
5084 }
5085
5086 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5087 log_late_inline(cg);
5088 if (log() != nullptr) {
5089 log()->inline_fail(msg);
5090 }
5091 }
5092
5093 void Compile::log_inline_id(CallGenerator* cg) {
5094 if (log() != nullptr) {
5095 // The LogCompilation tool needs a unique way to identify late
5096 // inline call sites. This id must be unique for this call site in
5097 // this compilation. Try to have it unique across compilations as
5098 // well because it can be convenient when grepping through the log
5099 // file.
5100 // Distinguish OSR compilations from others in case CICountOSR is
5101 // on.
5102 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5103 cg->set_unique_id(id);
5104 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5105 }
5106 }
5107
5108 void Compile::log_inline_failure(const char* msg) {
5109 if (C->log() != nullptr) {
5110 C->log()->inline_fail(msg);
5111 }
5112 }
5113
5114
5115 // Dump inlining replay data to the stream.
5116 // Don't change thread state and acquire any locks.
5117 void Compile::dump_inline_data(outputStream* out) {
5118 InlineTree* inl_tree = ilt();
5119 if (inl_tree != nullptr) {
5120 out->print(" inline %d", inl_tree->count());
5121 inl_tree->dump_replay_data(out);
5122 }
5123 }
5124
5125 void Compile::dump_inline_data_reduced(outputStream* out) {
5126 assert(ReplayReduce, "");
5127
5128 InlineTree* inl_tree = ilt();
5129 if (inl_tree == nullptr) {
5130 return;
5131 }
5132 // Enable iterative replay file reduction
5133 // Output "compile" lines for depth 1 subtrees,
5134 // simulating that those trees were compiled
5135 // instead of inlined.
5136 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5137 InlineTree* sub = inl_tree->subtrees().at(i);
5138 if (sub->inline_level() != 1) {
5139 continue;
5140 }
5141
5142 ciMethod* method = sub->method();
5143 int entry_bci = -1;
5144 int comp_level = env()->task()->comp_level();
5145 out->print("compile ");
5146 method->dump_name_as_ascii(out);
5147 out->print(" %d %d", entry_bci, comp_level);
5148 out->print(" inline %d", sub->count());
5149 sub->dump_replay_data(out, -1);
5150 out->cr();
5151 }
5152 }
5153
5154 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5155 if (n1->Opcode() < n2->Opcode()) return -1;
5156 else if (n1->Opcode() > n2->Opcode()) return 1;
5157
5158 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5159 for (uint i = 1; i < n1->req(); i++) {
5160 if (n1->in(i) < n2->in(i)) return -1;
5161 else if (n1->in(i) > n2->in(i)) return 1;
5162 }
5163
5164 return 0;
5165 }
5166
5167 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5168 Node* n1 = *n1p;
5169 Node* n2 = *n2p;
5170
5171 return cmp_expensive_nodes(n1, n2);
5172 }
5173
5174 void Compile::sort_expensive_nodes() {
5175 if (!expensive_nodes_sorted()) {
5176 _expensive_nodes.sort(cmp_expensive_nodes);
5177 }
5178 }
5179
5180 bool Compile::expensive_nodes_sorted() const {
5181 for (int i = 1; i < _expensive_nodes.length(); i++) {
5182 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5183 return false;
5184 }
5185 }
5186 return true;
5187 }
5188
5189 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5190 if (_expensive_nodes.length() == 0) {
5191 return false;
5192 }
5193
5194 assert(OptimizeExpensiveOps, "optimization off?");
5195
5196 // Take this opportunity to remove dead nodes from the list
5197 int j = 0;
5198 for (int i = 0; i < _expensive_nodes.length(); i++) {
5199 Node* n = _expensive_nodes.at(i);
5200 if (!n->is_unreachable(igvn)) {
5201 assert(n->is_expensive(), "should be expensive");
5202 _expensive_nodes.at_put(j, n);
5203 j++;
5204 }
5205 }
5206 _expensive_nodes.trunc_to(j);
5207
5208 // Then sort the list so that similar nodes are next to each other
5209 // and check for at least two nodes of identical kind with same data
5210 // inputs.
5211 sort_expensive_nodes();
5212
5213 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5214 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5215 return true;
5216 }
5217 }
5218
5219 return false;
5220 }
5221
5222 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5223 if (_expensive_nodes.length() == 0) {
5224 return;
5225 }
5226
5227 assert(OptimizeExpensiveOps, "optimization off?");
5228
5229 // Sort to bring similar nodes next to each other and clear the
5230 // control input of nodes for which there's only a single copy.
5231 sort_expensive_nodes();
5232
5233 int j = 0;
5234 int identical = 0;
5235 int i = 0;
5236 bool modified = false;
5237 for (; i < _expensive_nodes.length()-1; i++) {
5238 assert(j <= i, "can't write beyond current index");
5239 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5240 identical++;
5241 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5242 continue;
5243 }
5244 if (identical > 0) {
5245 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5246 identical = 0;
5247 } else {
5248 Node* n = _expensive_nodes.at(i);
5249 igvn.replace_input_of(n, 0, nullptr);
5250 igvn.hash_insert(n);
5251 modified = true;
5252 }
5253 }
5254 if (identical > 0) {
5255 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5256 } else if (_expensive_nodes.length() >= 1) {
5257 Node* n = _expensive_nodes.at(i);
5258 igvn.replace_input_of(n, 0, nullptr);
5259 igvn.hash_insert(n);
5260 modified = true;
5261 }
5262 _expensive_nodes.trunc_to(j);
5263 if (modified) {
5264 igvn.optimize();
5265 }
5266 }
5267
5268 void Compile::add_expensive_node(Node * n) {
5269 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5270 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5271 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5272 if (OptimizeExpensiveOps) {
5273 _expensive_nodes.append(n);
5274 } else {
5275 // Clear control input and let IGVN optimize expensive nodes if
5276 // OptimizeExpensiveOps is off.
5277 n->set_req(0, nullptr);
5278 }
5279 }
5280
5281 /**
5282 * Track coarsened Lock and Unlock nodes.
5283 */
5284
5285 class Lock_List : public Node_List {
5286 uint _origin_cnt;
5287 public:
5288 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5289 uint origin_cnt() const { return _origin_cnt; }
5290 };
5291
5292 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5293 int length = locks.length();
5294 if (length > 0) {
5295 // Have to keep this list until locks elimination during Macro nodes elimination.
5296 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5297 AbstractLockNode* alock = locks.at(0);
5298 BoxLockNode* box = alock->box_node()->as_BoxLock();
5299 for (int i = 0; i < length; i++) {
5300 AbstractLockNode* lock = locks.at(i);
5301 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5302 locks_list->push(lock);
5303 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
5304 if (this_box != box) {
5305 // Locking regions (BoxLock) could be Unbalanced here:
5306 // - its coarsened locks were eliminated in earlier
5307 // macro nodes elimination followed by loop unroll
5308 // - it is OSR locking region (no Lock node)
5309 // Preserve Unbalanced status in such cases.
5310 if (!this_box->is_unbalanced()) {
5311 this_box->set_coarsened();
5312 }
5313 if (!box->is_unbalanced()) {
5314 box->set_coarsened();
5315 }
5316 }
5317 }
5318 _coarsened_locks.append(locks_list);
5319 }
5320 }
5321
5322 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5323 int count = coarsened_count();
5324 for (int i = 0; i < count; i++) {
5325 Node_List* locks_list = _coarsened_locks.at(i);
5326 for (uint j = 0; j < locks_list->size(); j++) {
5327 Node* lock = locks_list->at(j);
5328 assert(lock->is_AbstractLock(), "sanity");
5329 if (!useful.member(lock)) {
5330 locks_list->yank(lock);
5331 }
5332 }
5333 }
5334 }
5335
5336 void Compile::remove_coarsened_lock(Node* n) {
5337 if (n->is_AbstractLock()) {
5338 int count = coarsened_count();
5339 for (int i = 0; i < count; i++) {
5340 Node_List* locks_list = _coarsened_locks.at(i);
5341 locks_list->yank(n);
5342 }
5343 }
5344 }
5345
5346 bool Compile::coarsened_locks_consistent() {
5347 int count = coarsened_count();
5348 for (int i = 0; i < count; i++) {
5349 bool unbalanced = false;
5350 bool modified = false; // track locks kind modifications
5351 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5352 uint size = locks_list->size();
5353 if (size == 0) {
5354 unbalanced = false; // All locks were eliminated - good
5355 } else if (size != locks_list->origin_cnt()) {
5356 unbalanced = true; // Some locks were removed from list
5357 } else {
5358 for (uint j = 0; j < size; j++) {
5359 Node* lock = locks_list->at(j);
5360 // All nodes in group should have the same state (modified or not)
5361 if (!lock->as_AbstractLock()->is_coarsened()) {
5362 if (j == 0) {
5363 // first on list was modified, the rest should be too for consistency
5364 modified = true;
5365 } else if (!modified) {
5366 // this lock was modified but previous locks on the list were not
5367 unbalanced = true;
5368 break;
5369 }
5370 } else if (modified) {
5371 // previous locks on list were modified but not this lock
5372 unbalanced = true;
5373 break;
5374 }
5375 }
5376 }
5377 if (unbalanced) {
5378 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5379 #ifdef ASSERT
5380 if (PrintEliminateLocks) {
5381 tty->print_cr("=== unbalanced coarsened locks ===");
5382 for (uint l = 0; l < size; l++) {
5383 locks_list->at(l)->dump();
5384 }
5385 }
5386 #endif
5387 record_failure(C2Compiler::retry_no_locks_coarsening());
5388 return false;
5389 }
5390 }
5391 return true;
5392 }
5393
5394 // Mark locking regions (identified by BoxLockNode) as unbalanced if
5395 // locks coarsening optimization removed Lock/Unlock nodes from them.
5396 // Such regions become unbalanced because coarsening only removes part
5397 // of Lock/Unlock nodes in region. As result we can't execute other
5398 // locks elimination optimizations which assume all code paths have
5399 // corresponding pair of Lock/Unlock nodes - they are balanced.
5400 void Compile::mark_unbalanced_boxes() const {
5401 int count = coarsened_count();
5402 for (int i = 0; i < count; i++) {
5403 Node_List* locks_list = _coarsened_locks.at(i);
5404 uint size = locks_list->size();
5405 if (size > 0) {
5406 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
5407 BoxLockNode* box = alock->box_node()->as_BoxLock();
5408 if (alock->is_coarsened()) {
5409 // coarsened_locks_consistent(), which is called before this method, verifies
5410 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
5411 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5412 for (uint j = 1; j < size; j++) {
5413 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
5414 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
5415 if (box != this_box) {
5416 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5417 box->set_unbalanced();
5418 this_box->set_unbalanced();
5419 }
5420 }
5421 }
5422 }
5423 }
5424 }
5425
5426 /**
5427 * Remove the speculative part of types and clean up the graph
5428 */
5429 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5430 if (UseTypeSpeculation) {
5431 Unique_Node_List worklist;
5432 worklist.push(root());
5433 int modified = 0;
5434 // Go over all type nodes that carry a speculative type, drop the
5435 // speculative part of the type and enqueue the node for an igvn
5436 // which may optimize it out.
5437 for (uint next = 0; next < worklist.size(); ++next) {
5438 Node *n = worklist.at(next);
5439 if (n->is_Type()) {
5440 TypeNode* tn = n->as_Type();
5441 const Type* t = tn->type();
5442 const Type* t_no_spec = t->remove_speculative();
5443 if (t_no_spec != t) {
5444 bool in_hash = igvn.hash_delete(n);
5445 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5446 tn->set_type(t_no_spec);
5447 igvn.hash_insert(n);
5448 igvn._worklist.push(n); // give it a chance to go away
5449 modified++;
5450 }
5451 }
5452 // Iterate over outs - endless loops is unreachable from below
5453 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5454 Node *m = n->fast_out(i);
5455 if (not_a_node(m)) {
5456 continue;
5457 }
5458 worklist.push(m);
5459 }
5460 }
5461 // Drop the speculative part of all types in the igvn's type table
5462 igvn.remove_speculative_types();
5463 if (modified > 0) {
5464 igvn.optimize();
5465 if (failing()) return;
5466 }
5467 #ifdef ASSERT
5468 // Verify that after the IGVN is over no speculative type has resurfaced
5469 worklist.clear();
5470 worklist.push(root());
5471 for (uint next = 0; next < worklist.size(); ++next) {
5472 Node *n = worklist.at(next);
5473 const Type* t = igvn.type_or_null(n);
5474 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5475 if (n->is_Type()) {
5476 t = n->as_Type()->type();
5477 assert(t == t->remove_speculative(), "no more speculative types");
5478 }
5479 // Iterate over outs - endless loops is unreachable from below
5480 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5481 Node *m = n->fast_out(i);
5482 if (not_a_node(m)) {
5483 continue;
5484 }
5485 worklist.push(m);
5486 }
5487 }
5488 igvn.check_no_speculative_types();
5489 #endif
5490 }
5491 }
5492
5493 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5494 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5495 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5496 if (!EnableValhalla || ta == nullptr || tb == nullptr ||
5497 ta->is_zero_type() || tb->is_zero_type() ||
5498 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5499 // Use old acmp if one operand is null or not an inline type
5500 return new CmpPNode(a, b);
5501 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5502 // We know that one operand is an inline type. Therefore,
5503 // new acmp will only return true if both operands are nullptr.
5504 // Check if both operands are null by or'ing the oops.
5505 a = phase->transform(new CastP2XNode(nullptr, a));
5506 b = phase->transform(new CastP2XNode(nullptr, b));
5507 a = phase->transform(new OrXNode(a, b));
5508 return new CmpXNode(a, phase->MakeConX(0));
5509 }
5510 // Use new acmp
5511 return nullptr;
5512 }
5513
5514 // Auxiliary methods to support randomized stressing/fuzzing.
5515
5516 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5517 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5518 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5519 FLAG_SET_ERGO(StressSeed, _stress_seed);
5520 } else {
5521 _stress_seed = StressSeed;
5522 }
5523 if (_log != nullptr) {
5524 _log->elem("stress_test seed='%u'", _stress_seed);
5525 }
5526 }
5527
5528 int Compile::random() {
5529 _stress_seed = os::next_random(_stress_seed);
5530 return static_cast<int>(_stress_seed);
5531 }
5532
5533 // This method can be called the arbitrary number of times, with current count
5534 // as the argument. The logic allows selecting a single candidate from the
5535 // running list of candidates as follows:
5536 // int count = 0;
5537 // Cand* selected = null;
5538 // while(cand = cand->next()) {
5539 // if (randomized_select(++count)) {
5540 // selected = cand;
5541 // }
5542 // }
5543 //
5544 // Including count equalizes the chances any candidate is "selected".
5545 // This is useful when we don't have the complete list of candidates to choose
5546 // from uniformly. In this case, we need to adjust the randomicity of the
5547 // selection, or else we will end up biasing the selection towards the latter
5548 // candidates.
5549 //
5550 // Quick back-envelope calculation shows that for the list of n candidates
5551 // the equal probability for the candidate to persist as "best" can be
5552 // achieved by replacing it with "next" k-th candidate with the probability
5553 // of 1/k. It can be easily shown that by the end of the run, the
5554 // probability for any candidate is converged to 1/n, thus giving the
5555 // uniform distribution among all the candidates.
5556 //
5557 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5558 #define RANDOMIZED_DOMAIN_POW 29
5559 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5560 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5561 bool Compile::randomized_select(int count) {
5562 assert(count > 0, "only positive");
5563 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5564 }
5565
5566 #ifdef ASSERT
5567 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5568 bool Compile::fail_randomly() {
5569 if ((random() % StressBailoutMean) != 0) {
5570 return false;
5571 }
5572 record_failure("StressBailout");
5573 return true;
5574 }
5575
5576 bool Compile::failure_is_artificial() {
5577 return C->failure_reason_is("StressBailout");
5578 }
5579 #endif
5580
5581 CloneMap& Compile::clone_map() { return _clone_map; }
5582 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5583
5584 void NodeCloneInfo::dump_on(outputStream* st) const {
5585 st->print(" {%d:%d} ", idx(), gen());
5586 }
5587
5588 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5589 uint64_t val = value(old->_idx);
5590 NodeCloneInfo cio(val);
5591 assert(val != 0, "old node should be in the map");
5592 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5593 insert(nnn->_idx, cin.get());
5594 #ifndef PRODUCT
5595 if (is_debug()) {
5596 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5597 }
5598 #endif
5599 }
5600
5601 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5602 NodeCloneInfo cio(value(old->_idx));
5603 if (cio.get() == 0) {
5604 cio.set(old->_idx, 0);
5605 insert(old->_idx, cio.get());
5606 #ifndef PRODUCT
5607 if (is_debug()) {
5608 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5609 }
5610 #endif
5611 }
5612 clone(old, nnn, gen);
5613 }
5614
5615 int CloneMap::max_gen() const {
5616 int g = 0;
5617 DictI di(_dict);
5618 for(; di.test(); ++di) {
5619 int t = gen(di._key);
5620 if (g < t) {
5621 g = t;
5622 #ifndef PRODUCT
5623 if (is_debug()) {
5624 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5625 }
5626 #endif
5627 }
5628 }
5629 return g;
5630 }
5631
5632 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5633 uint64_t val = value(key);
5634 if (val != 0) {
5635 NodeCloneInfo ni(val);
5636 ni.dump_on(st);
5637 }
5638 }
5639
5640 void Compile::shuffle_macro_nodes() {
5641 if (_macro_nodes.length() < 2) {
5642 return;
5643 }
5644 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5645 uint j = C->random() % (i + 1);
5646 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5647 }
5648 }
5649
5650 // Move Allocate nodes to the start of the list
5651 void Compile::sort_macro_nodes() {
5652 int count = macro_count();
5653 int allocates = 0;
5654 for (int i = 0; i < count; i++) {
5655 Node* n = macro_node(i);
5656 if (n->is_Allocate()) {
5657 if (i != allocates) {
5658 Node* tmp = macro_node(allocates);
5659 _macro_nodes.at_put(allocates, n);
5660 _macro_nodes.at_put(i, tmp);
5661 }
5662 allocates++;
5663 }
5664 }
5665 }
5666
5667 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5668 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5669 EventCompilerPhase event(UNTIMED);
5670 if (event.should_commit()) {
5671 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5672 }
5673 #ifndef PRODUCT
5674 ResourceMark rm;
5675 stringStream ss;
5676 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5677 int iter = ++_igv_phase_iter[cpt];
5678 if (iter > 1) {
5679 ss.print(" %d", iter);
5680 }
5681 if (n != nullptr) {
5682 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5683 if (n->is_Call()) {
5684 CallNode* call = n->as_Call();
5685 if (call->_name != nullptr) {
5686 // E.g. uncommon traps etc.
5687 ss.print(" - %s", call->_name);
5688 } else if (call->is_CallJava()) {
5689 CallJavaNode* call_java = call->as_CallJava();
5690 if (call_java->method() != nullptr) {
5691 ss.print(" -");
5692 call_java->method()->print_short_name(&ss);
5693 }
5694 }
5695 }
5696 }
5697
5698 const char* name = ss.as_string();
5699 if (should_print_igv(level)) {
5700 _igv_printer->print_graph(name);
5701 }
5702 if (should_print_phase(level)) {
5703 print_phase(name);
5704 }
5705 if (should_print_ideal_phase(cpt)) {
5706 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5707 }
5708 #endif
5709 C->_latest_stage_start_counter.stamp();
5710 }
5711
5712 // Only used from CompileWrapper
5713 void Compile::begin_method() {
5714 #ifndef PRODUCT
5715 if (_method != nullptr && should_print_igv(1)) {
5716 _igv_printer->begin_method();
5717 }
5718 #endif
5719 C->_latest_stage_start_counter.stamp();
5720 }
5721
5722 // Only used from CompileWrapper
5723 void Compile::end_method() {
5724 EventCompilerPhase event(UNTIMED);
5725 if (event.should_commit()) {
5726 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5727 }
5728
5729 #ifndef PRODUCT
5730 if (_method != nullptr && should_print_igv(1)) {
5731 _igv_printer->end_method();
5732 }
5733 #endif
5734 }
5735
5736 #ifndef PRODUCT
5737 bool Compile::should_print_phase(const int level) const {
5738 return PrintPhaseLevel > 0 && directive()->PhasePrintLevelOption >= level &&
5739 _method != nullptr; // Do not print phases for stubs.
5740 }
5741
5742 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5743 return _directive->should_print_ideal_phase(cpt);
5744 }
5745
5746 void Compile::init_igv() {
5747 if (_igv_printer == nullptr) {
5748 _igv_printer = IdealGraphPrinter::printer();
5749 _igv_printer->set_compile(this);
5750 }
5751 }
5752
5753 bool Compile::should_print_igv(const int level) {
5754 PRODUCT_RETURN_(return false;);
5755
5756 if (PrintIdealGraphLevel < 0) { // disabled by the user
5757 return false;
5758 }
5759
5760 bool need = directive()->IGVPrintLevelOption >= level;
5761 if (need) {
5762 Compile::init_igv();
5763 }
5764 return need;
5765 }
5766
5767 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5768 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5769
5770 // Called from debugger. Prints method to the default file with the default phase name.
5771 // This works regardless of any Ideal Graph Visualizer flags set or not.
5772 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5773 void igv_print(void* sp, void* fp, void* pc) {
5774 frame fr(sp, fp, pc);
5775 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5776 }
5777
5778 // Same as igv_print() above but with a specified phase name.
5779 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5780 frame fr(sp, fp, pc);
5781 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5782 }
5783
5784 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5785 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5786 // This works regardless of any Ideal Graph Visualizer flags set or not.
5787 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5788 void igv_print(bool network, void* sp, void* fp, void* pc) {
5789 frame fr(sp, fp, pc);
5790 if (network) {
5791 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5792 } else {
5793 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5794 }
5795 }
5796
5797 // Same as igv_print(bool network, ...) above but with a specified phase name.
5798 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5799 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5800 frame fr(sp, fp, pc);
5801 if (network) {
5802 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5803 } else {
5804 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5805 }
5806 }
5807
5808 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5809 void igv_print_default() {
5810 Compile::current()->print_method(PHASE_DEBUG, 0);
5811 }
5812
5813 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5814 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5815 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5816 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5817 void igv_append(void* sp, void* fp, void* pc) {
5818 frame fr(sp, fp, pc);
5819 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5820 }
5821
5822 // Same as igv_append(...) above but with a specified phase name.
5823 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5824 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5825 frame fr(sp, fp, pc);
5826 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5827 }
5828
5829 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5830 const char* file_name = "custom_debug.xml";
5831 if (_debug_file_printer == nullptr) {
5832 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5833 } else {
5834 _debug_file_printer->update_compiled_method(C->method());
5835 }
5836 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5837 _debug_file_printer->print_graph(phase_name, fr);
5838 }
5839
5840 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5841 ResourceMark rm;
5842 GrowableArray<const Node*> empty_list;
5843 igv_print_graph_to_network(phase_name, empty_list, fr);
5844 }
5845
5846 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5847 if (_debug_network_printer == nullptr) {
5848 _debug_network_printer = new IdealGraphPrinter(C);
5849 } else {
5850 _debug_network_printer->update_compiled_method(C->method());
5851 }
5852 tty->print_cr("Method printed over network stream to IGV");
5853 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5854 }
5855 #endif // !PRODUCT
5856
5857 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5858 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5859 return value;
5860 }
5861 Node* result = nullptr;
5862 if (bt == T_BYTE) {
5863 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5864 result = new RShiftINode(result, phase->intcon(24));
5865 } else if (bt == T_BOOLEAN) {
5866 result = new AndINode(value, phase->intcon(0xFF));
5867 } else if (bt == T_CHAR) {
5868 result = new AndINode(value,phase->intcon(0xFFFF));
5869 } else if (bt == T_FLOAT) {
5870 result = new MoveI2FNode(value);
5871 } else {
5872 assert(bt == T_SHORT, "unexpected narrow type");
5873 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5874 result = new RShiftINode(result, phase->intcon(16));
5875 }
5876 if (transform_res) {
5877 result = phase->transform(result);
5878 }
5879 return result;
5880 }
5881
5882 void Compile::record_method_not_compilable_oom() {
5883 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5884 }