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