1 /*
2 * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/macroAssembler.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "ci/ciReplay.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "code/aotCodeCache.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compiler_globals.hpp"
37 #include "compiler/compilerDefinitions.hpp"
38 #include "compiler/compilerOracle.hpp"
39 #include "compiler/disassembler.hpp"
40 #include "compiler/oopMap.hpp"
41 #include "gc/shared/barrierSet.hpp"
42 #include "gc/shared/c2/barrierSetC2.hpp"
43 #include "jfr/jfrEvents.hpp"
44 #include "jvm_io.h"
45 #include "memory/allocation.hpp"
46 #include "memory/arena.hpp"
47 #include "memory/resourceArea.hpp"
48 #include "opto/addnode.hpp"
49 #include "opto/block.hpp"
50 #include "opto/c2compiler.hpp"
51 #include "opto/callGenerator.hpp"
52 #include "opto/callnode.hpp"
53 #include "opto/castnode.hpp"
54 #include "opto/cfgnode.hpp"
55 #include "opto/chaitin.hpp"
56 #include "opto/compile.hpp"
57 #include "opto/connode.hpp"
58 #include "opto/convertnode.hpp"
59 #include "opto/divnode.hpp"
60 #include "opto/escape.hpp"
61 #include "opto/idealGraphPrinter.hpp"
62 #include "opto/locknode.hpp"
63 #include "opto/loopnode.hpp"
64 #include "opto/machnode.hpp"
65 #include "opto/macro.hpp"
66 #include "opto/matcher.hpp"
67 #include "opto/mathexactnode.hpp"
68 #include "opto/memnode.hpp"
69 #include "opto/mulnode.hpp"
70 #include "opto/narrowptrnode.hpp"
71 #include "opto/node.hpp"
72 #include "opto/opaquenode.hpp"
73 #include "opto/opcodes.hpp"
74 #include "opto/output.hpp"
75 #include "opto/parse.hpp"
76 #include "opto/phaseX.hpp"
77 #include "opto/reachability.hpp"
78 #include "opto/rootnode.hpp"
79 #include "opto/runtime.hpp"
80 #include "opto/stringopts.hpp"
81 #include "opto/type.hpp"
82 #include "opto/vector.hpp"
83 #include "opto/vectornode.hpp"
84 #include "runtime/globals_extension.hpp"
85 #include "runtime/sharedRuntime.hpp"
86 #include "runtime/signature.hpp"
87 #include "runtime/stubRoutines.hpp"
88 #include "runtime/timer.hpp"
89 #include "utilities/align.hpp"
90 #include "utilities/copy.hpp"
91 #include "utilities/hashTable.hpp"
92 #include "utilities/macros.hpp"
93
94 // -------------------- Compile::mach_constant_base_node -----------------------
95 // Constant table base node singleton.
96 MachConstantBaseNode* Compile::mach_constant_base_node() {
97 if (_mach_constant_base_node == nullptr) {
98 _mach_constant_base_node = new MachConstantBaseNode();
99 _mach_constant_base_node->add_req(C->root());
100 }
101 return _mach_constant_base_node;
102 }
103
104
105 /// Support for intrinsics.
106
107 // Return the index at which m must be inserted (or already exists).
108 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
109 class IntrinsicDescPair {
110 private:
111 ciMethod* _m;
112 bool _is_virtual;
113 public:
114 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
115 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
116 ciMethod* m= elt->method();
117 ciMethod* key_m = key->_m;
118 if (key_m < m) return -1;
119 else if (key_m > m) return 1;
120 else {
121 bool is_virtual = elt->is_virtual();
122 bool key_virtual = key->_is_virtual;
123 if (key_virtual < is_virtual) return -1;
124 else if (key_virtual > is_virtual) return 1;
125 else return 0;
126 }
127 }
128 };
129 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
130 #ifdef ASSERT
131 for (int i = 1; i < _intrinsics.length(); i++) {
132 CallGenerator* cg1 = _intrinsics.at(i-1);
133 CallGenerator* cg2 = _intrinsics.at(i);
134 assert(cg1->method() != cg2->method()
135 ? cg1->method() < cg2->method()
136 : cg1->is_virtual() < cg2->is_virtual(),
137 "compiler intrinsics list must stay sorted");
138 }
139 #endif
140 IntrinsicDescPair pair(m, is_virtual);
141 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
142 }
143
144 void Compile::register_intrinsic(CallGenerator* cg) {
145 bool found = false;
146 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
147 assert(!found, "registering twice");
148 _intrinsics.insert_before(index, cg);
149 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
150 }
151
152 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
153 assert(m->is_loaded(), "don't try this on unloaded methods");
154 if (_intrinsics.length() > 0) {
155 bool found = false;
156 int index = intrinsic_insertion_index(m, is_virtual, found);
157 if (found) {
158 return _intrinsics.at(index);
159 }
160 }
161 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
162 if (m->intrinsic_id() != vmIntrinsics::_none &&
163 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
164 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
165 if (cg != nullptr) {
166 // Save it for next time:
167 register_intrinsic(cg);
168 return cg;
169 } else {
170 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
171 }
172 }
173 return nullptr;
174 }
175
176 // Compile::make_vm_intrinsic is defined in library_call.cpp.
177
178 #ifndef PRODUCT
179 // statistics gathering...
180
181 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
182 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
183
184 inline int as_int(vmIntrinsics::ID id) {
185 return vmIntrinsics::as_int(id);
186 }
187
188 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
189 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
190 int oflags = _intrinsic_hist_flags[as_int(id)];
191 assert(flags != 0, "what happened?");
192 if (is_virtual) {
193 flags |= _intrinsic_virtual;
194 }
195 bool changed = (flags != oflags);
196 if ((flags & _intrinsic_worked) != 0) {
197 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
198 if (count == 1) {
199 changed = true; // first time
200 }
201 // increment the overall count also:
202 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
203 }
204 if (changed) {
205 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
206 // Something changed about the intrinsic's virtuality.
207 if ((flags & _intrinsic_virtual) != 0) {
208 // This is the first use of this intrinsic as a virtual call.
209 if (oflags != 0) {
210 // We already saw it as a non-virtual, so note both cases.
211 flags |= _intrinsic_both;
212 }
213 } else if ((oflags & _intrinsic_both) == 0) {
214 // This is the first use of this intrinsic as a non-virtual
215 flags |= _intrinsic_both;
216 }
217 }
218 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
219 }
220 // update the overall flags also:
221 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
222 return changed;
223 }
224
225 static char* format_flags(int flags, char* buf) {
226 buf[0] = 0;
227 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
228 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
229 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
230 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
231 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
232 if (buf[0] == 0) strcat(buf, ",");
233 assert(buf[0] == ',', "must be");
234 return &buf[1];
235 }
236
237 void Compile::print_intrinsic_statistics() {
238 char flagsbuf[100];
239 ttyLocker ttyl;
240 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
241 tty->print_cr("Compiler intrinsic usage:");
242 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
243 if (total == 0) total = 1; // avoid div0 in case of no successes
244 #define PRINT_STAT_LINE(name, c, f) \
245 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
246 for (auto id : EnumRange<vmIntrinsicID>{}) {
247 int flags = _intrinsic_hist_flags[as_int(id)];
248 juint count = _intrinsic_hist_count[as_int(id)];
249 if ((flags | count) != 0) {
250 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
251 }
252 }
253 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
254 if (xtty != nullptr) xtty->tail("statistics");
255 }
256
257 void Compile::print_statistics() {
258 { ttyLocker ttyl;
259 if (xtty != nullptr) xtty->head("statistics type='opto'");
260 Parse::print_statistics();
261 PhaseStringOpts::print_statistics();
262 PhaseCCP::print_statistics();
263 PhaseRegAlloc::print_statistics();
264 PhaseOutput::print_statistics();
265 PhasePeephole::print_statistics();
266 PhaseIdealLoop::print_statistics();
267 ConnectionGraph::print_statistics();
268 PhaseMacroExpand::print_statistics();
269 if (xtty != nullptr) xtty->tail("statistics");
270 }
271 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
272 // put this under its own <statistics> element.
273 print_intrinsic_statistics();
274 }
275 }
276 #endif //PRODUCT
277
278 void Compile::gvn_replace_by(Node* n, Node* nn) {
279 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
280 Node* use = n->last_out(i);
281 bool is_in_table = initial_gvn()->hash_delete(use);
282 uint uses_found = 0;
283 for (uint j = 0; j < use->len(); j++) {
284 if (use->in(j) == n) {
285 if (j < use->req())
286 use->set_req(j, nn);
287 else
288 use->set_prec(j, nn);
289 uses_found++;
290 }
291 }
292 if (is_in_table) {
293 // reinsert into table
294 initial_gvn()->hash_find_insert(use);
295 }
296 record_for_igvn(use);
297 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
298 i -= uses_found; // we deleted 1 or more copies of this edge
299 }
300 }
301
302
303 // Identify all nodes that are reachable from below, useful.
304 // Use breadth-first pass that records state in a Unique_Node_List,
305 // recursive traversal is slower.
306 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
307 int estimated_worklist_size = live_nodes();
308 useful.map( estimated_worklist_size, nullptr ); // preallocate space
309
310 // Initialize worklist
311 if (root() != nullptr) { useful.push(root()); }
312 // If 'top' is cached, declare it useful to preserve cached node
313 if (cached_top_node()) { useful.push(cached_top_node()); }
314
315 // Push all useful nodes onto the list, breadthfirst
316 for( uint next = 0; next < useful.size(); ++next ) {
317 assert( next < unique(), "Unique useful nodes < total nodes");
318 Node *n = useful.at(next);
319 uint max = n->len();
320 for( uint i = 0; i < max; ++i ) {
321 Node *m = n->in(i);
322 if (not_a_node(m)) continue;
323 useful.push(m);
324 }
325 }
326 }
327
328 // Update dead_node_list with any missing dead nodes using useful
329 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
330 void Compile::update_dead_node_list(Unique_Node_List &useful) {
331 uint max_idx = unique();
332 VectorSet& useful_node_set = useful.member_set();
333
334 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
335 // If node with index node_idx is not in useful set,
336 // mark it as dead in dead node list.
337 if (!useful_node_set.test(node_idx)) {
338 record_dead_node(node_idx);
339 }
340 }
341 }
342
343 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
344 int shift = 0;
345 for (int i = 0; i < inlines->length(); i++) {
346 CallGenerator* cg = inlines->at(i);
347 if (useful.member(cg->call_node())) {
348 if (shift > 0) {
349 inlines->at_put(i - shift, cg);
350 }
351 } else {
352 shift++; // skip over the dead element
353 }
354 }
355 if (shift > 0) {
356 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
357 }
358 }
359
360 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
361 assert(dead != nullptr && dead->is_Call(), "sanity");
362 int found = 0;
363 for (int i = 0; i < inlines->length(); i++) {
364 if (inlines->at(i)->call_node() == dead) {
365 inlines->remove_at(i);
366 found++;
367 NOT_DEBUG( break; ) // elements are unique, so exit early
368 }
369 }
370 assert(found <= 1, "not unique");
371 }
372
373 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
374 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
375 for (int i = node_list.length() - 1; i >= 0; i--) {
376 N* node = node_list.at(i);
377 if (!useful.member(node)) {
378 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
379 }
380 }
381 }
382
383 void Compile::remove_useless_node(Node* dead) {
384 remove_modified_node(dead);
385
386 // Constant node that has no out-edges and has only one in-edge from
387 // root is usually dead. However, sometimes reshaping walk makes
388 // it reachable by adding use edges. So, we will NOT count Con nodes
389 // as dead to be conservative about the dead node count at any
390 // given time.
391 if (!dead->is_Con()) {
392 record_dead_node(dead->_idx);
393 }
394 if (dead->is_macro()) {
395 remove_macro_node(dead);
396 }
397 if (dead->is_expensive()) {
398 remove_expensive_node(dead);
399 }
400 if (dead->is_ReachabilityFence()) {
401 remove_reachability_fence(dead->as_ReachabilityFence());
402 }
403 if (dead->is_OpaqueTemplateAssertionPredicate()) {
404 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
405 }
406 if (dead->is_ParsePredicate()) {
407 remove_parse_predicate(dead->as_ParsePredicate());
408 }
409 if (dead->for_post_loop_opts_igvn()) {
410 remove_from_post_loop_opts_igvn(dead);
411 }
412 if (dead->for_merge_stores_igvn()) {
413 remove_from_merge_stores_igvn(dead);
414 }
415 if (dead->is_Call()) {
416 remove_useless_late_inlines( &_late_inlines, dead);
417 remove_useless_late_inlines( &_string_late_inlines, dead);
418 remove_useless_late_inlines( &_boxing_late_inlines, dead);
419 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
420
421 if (dead->is_CallStaticJava()) {
422 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
423 }
424 }
425 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
426 bs->unregister_potential_barrier_node(dead);
427 }
428
429 // Disconnect all useless nodes by disconnecting those at the boundary.
430 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
431 uint next = 0;
432 while (next < useful.size()) {
433 Node *n = useful.at(next++);
434 if (n->is_SafePoint()) {
435 // We're done with a parsing phase. Replaced nodes are not valid
436 // beyond that point.
437 n->as_SafePoint()->delete_replaced_nodes();
438 }
439 // Use raw traversal of out edges since this code removes out edges
440 int max = n->outcnt();
441 for (int j = 0; j < max; ++j) {
442 Node* child = n->raw_out(j);
443 if (!useful.member(child)) {
444 assert(!child->is_top() || child != top(),
445 "If top is cached in Compile object it is in useful list");
446 // Only need to remove this out-edge to the useless node
447 n->raw_del_out(j);
448 --j;
449 --max;
450 if (child->is_data_proj_of_pure_function(n)) {
451 worklist.push(n);
452 }
453 }
454 }
455 if (n->outcnt() == 1 && n->has_special_unique_user()) {
456 assert(useful.member(n->unique_out()), "do not push a useless node");
457 worklist.push(n->unique_out());
458 }
459 }
460
461 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
462 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
463 // Remove useless Template Assertion Predicate opaque nodes
464 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
465 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
466 remove_useless_nodes(_reachability_fences, useful); // remove useless node recorded for post loop opts IGVN pass
467 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
468 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
469 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
470 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
471 #ifdef ASSERT
472 if (_modified_nodes != nullptr) {
473 _modified_nodes->remove_useless_nodes(useful.member_set());
474 }
475 #endif
476
477 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
478 bs->eliminate_useless_gc_barriers(useful, this);
479 // clean up the late inline lists
480 remove_useless_late_inlines( &_late_inlines, useful);
481 remove_useless_late_inlines( &_string_late_inlines, useful);
482 remove_useless_late_inlines( &_boxing_late_inlines, useful);
483 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
484 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
485 }
486
487 // ============================================================================
488 //------------------------------CompileWrapper---------------------------------
489 class CompileWrapper : public StackObj {
490 Compile *const _compile;
491 public:
492 CompileWrapper(Compile* compile);
493
494 ~CompileWrapper();
495 };
496
497 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
498 // the Compile* pointer is stored in the current ciEnv:
499 ciEnv* env = compile->env();
500 assert(env == ciEnv::current(), "must already be a ciEnv active");
501 assert(env->compiler_data() == nullptr, "compile already active?");
502 env->set_compiler_data(compile);
503 assert(compile == Compile::current(), "sanity");
504
505 compile->set_type_dict(nullptr);
506 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
507 compile->clone_map().set_clone_idx(0);
508 compile->set_type_last_size(0);
509 compile->set_last_tf(nullptr, nullptr);
510 compile->set_indexSet_arena(nullptr);
511 compile->set_indexSet_free_block_list(nullptr);
512 compile->init_type_arena();
513 Type::Initialize(compile);
514 _compile->begin_method();
515 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
516 }
517 CompileWrapper::~CompileWrapper() {
518 // simulate crash during compilation
519 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
520
521 _compile->end_method();
522 _compile->env()->set_compiler_data(nullptr);
523 }
524
525
526 //----------------------------print_compile_messages---------------------------
527 void Compile::print_compile_messages() {
528 #ifndef PRODUCT
529 // Check if recompiling
530 if (!subsume_loads() && PrintOpto) {
531 // Recompiling without allowing machine instructions to subsume loads
532 tty->print_cr("*********************************************************");
533 tty->print_cr("** Bailout: Recompile without subsuming loads **");
534 tty->print_cr("*********************************************************");
535 }
536 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
537 // Recompiling without escape analysis
538 tty->print_cr("*********************************************************");
539 tty->print_cr("** Bailout: Recompile without escape analysis **");
540 tty->print_cr("*********************************************************");
541 }
542 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
543 // Recompiling without iterative escape analysis
544 tty->print_cr("*********************************************************");
545 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
546 tty->print_cr("*********************************************************");
547 }
548 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
549 // Recompiling without reducing allocation merges
550 tty->print_cr("*********************************************************");
551 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
552 tty->print_cr("*********************************************************");
553 }
554 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
555 // Recompiling without boxing elimination
556 tty->print_cr("*********************************************************");
557 tty->print_cr("** Bailout: Recompile without boxing elimination **");
558 tty->print_cr("*********************************************************");
559 }
560 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
561 // Recompiling without locks coarsening
562 tty->print_cr("*********************************************************");
563 tty->print_cr("** Bailout: Recompile without locks coarsening **");
564 tty->print_cr("*********************************************************");
565 }
566 if (env()->break_at_compile()) {
567 // Open the debugger when compiling this method.
568 tty->print("### Breaking when compiling: ");
569 method()->print_short_name();
570 tty->cr();
571 BREAKPOINT;
572 }
573
574 if( PrintOpto ) {
575 if (is_osr_compilation()) {
576 tty->print("[OSR]%3d", _compile_id);
577 } else {
578 tty->print("%3d", _compile_id);
579 }
580 }
581 #endif
582 }
583
584 #ifndef PRODUCT
585 void Compile::print_phase(const char* phase_name) {
586 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name);
587 }
588
589 void Compile::print_ideal_ir(const char* compile_phase_name) const {
590 // keep the following output all in one block
591 // This output goes directly to the tty, not the compiler log.
592 // To enable tools to match it up with the compilation activity,
593 // be sure to tag this tty output with the compile ID.
594
595 // Node dumping can cause a safepoint, which can break the tty lock.
596 // Buffer all node dumps, so that all safepoints happen before we lock.
597 ResourceMark rm;
598 stringStream ss;
599
600 if (_output == nullptr) {
601 ss.print_cr("AFTER: %s", compile_phase_name);
602 // Print out all nodes in ascending order of index.
603 // It is important that we traverse both inputs and outputs of nodes,
604 // so that we reach all nodes that are connected to Root.
605 root()->dump_bfs(MaxNodeLimit, nullptr, "-+S$", &ss);
606 } else {
607 // Dump the node blockwise if we have a scheduling
608 _output->print_scheduling(&ss);
609 }
610
611 // Check that the lock is not broken by a safepoint.
612 NoSafepointVerifier nsv;
613 ttyLocker ttyl;
614 if (xtty != nullptr) {
615 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
616 compile_id(),
617 is_osr_compilation() ? " compile_kind='osr'" : "",
618 compile_phase_name);
619 }
620
621 tty->print("%s", ss.as_string());
622
623 if (xtty != nullptr) {
624 xtty->tail("ideal");
625 }
626 }
627 #endif
628
629 // ============================================================================
630 //------------------------------Compile standard-------------------------------
631
632 // Compile a method. entry_bci is -1 for normal compilations and indicates
633 // the continuation bci for on stack replacement.
634
635
636 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
637 Options options, DirectiveSet* directive)
638 : Phase(Compiler),
639 _compile_id(ci_env->compile_id()),
640 _options(options),
641 _method(target),
642 _entry_bci(osr_bci),
643 _ilt(nullptr),
644 _stub_function(nullptr),
645 _stub_name(nullptr),
646 _stub_id(StubId::NO_STUBID),
647 _stub_entry_point(nullptr),
648 _max_node_limit(MaxNodeLimit),
649 _node_count_inlining_cutoff(NodeCountInliningCutoff),
650 _post_loop_opts_phase(false),
651 _merge_stores_phase(false),
652 _allow_macro_nodes(true),
653 _inlining_progress(false),
654 _inlining_incrementally(false),
655 _do_cleanup(false),
656 _has_reserved_stack_access(target->has_reserved_stack_access()),
657 #ifndef PRODUCT
658 _igv_idx(0),
659 _trace_opto_output(directive->TraceOptoOutputOption),
660 #endif
661 _clinit_barrier_on_entry(false),
662 _stress_seed(0),
663 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
664 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
665 _env(ci_env),
666 _directive(directive),
667 _log(ci_env->log()),
668 _first_failure_details(nullptr),
669 _intrinsics(comp_arena(), 0, 0, nullptr),
670 _macro_nodes(comp_arena(), 8, 0, nullptr),
671 _parse_predicates(comp_arena(), 8, 0, nullptr),
672 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
673 _expensive_nodes(comp_arena(), 8, 0, nullptr),
674 _reachability_fences(comp_arena(), 8, 0, nullptr),
675 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
676 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
677 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
678 _coarsened_locks(comp_arena(), 8, 0, nullptr),
679 _congraph(nullptr),
680 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
681 _unique(0),
682 _dead_node_count(0),
683 _dead_node_list(comp_arena()),
684 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
685 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
686 _node_arena(&_node_arena_one),
687 _mach_constant_base_node(nullptr),
688 _Compile_types(mtCompiler, Arena::Tag::tag_type),
689 _initial_gvn(nullptr),
690 _igvn_worklist(nullptr),
691 _types(nullptr),
692 _node_hash(nullptr),
693 _late_inlines(comp_arena(), 2, 0, nullptr),
694 _string_late_inlines(comp_arena(), 2, 0, nullptr),
695 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
696 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
697 _late_inlines_pos(0),
698 _has_mh_late_inlines(false),
699 _oom(false),
700 _replay_inline_data(nullptr),
701 _inline_printer(this),
702 _java_calls(0),
703 _inner_loops(0),
704 _FIRST_STACK_mask(comp_arena()),
705 _interpreter_frame_size(0),
706 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
707 _output(nullptr)
708 #ifndef PRODUCT
709 ,
710 _in_dump_cnt(0)
711 #endif
712 {
713 C = this;
714 CompileWrapper cw(this);
715
716 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
717 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
718
719 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
720 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
721 // We can always print a disassembly, either abstract (hex dump) or
722 // with the help of a suitable hsdis library. Thus, we should not
723 // couple print_assembly and print_opto_assembly controls.
724 // But: always print opto and regular assembly on compile command 'print'.
725 bool print_assembly = directive->PrintAssemblyOption;
726 set_print_assembly(print_opto_assembly || print_assembly);
727 #else
728 set_print_assembly(false); // must initialize.
729 #endif
730
731 #ifndef PRODUCT
732 set_parsed_irreducible_loop(false);
733 #endif
734
735 if (directive->ReplayInlineOption) {
736 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
737 }
738 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
739 set_print_intrinsics(directive->PrintIntrinsicsOption);
740 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
741
742 if (ProfileTraps) {
743 // Make sure the method being compiled gets its own MDO,
744 // so we can at least track the decompile_count().
745 method()->ensure_method_data();
746 }
747
748 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
749 StressIncrementalInlining || StressMacroExpansion ||
750 StressMacroElimination || StressUnstableIfTraps ||
751 StressBailout || StressLoopPeeling || StressCountedLoop) {
752 initialize_stress_seed(directive);
753 }
754
755 Init(/*do_aliasing=*/ true);
756
757 print_compile_messages();
758
759 _ilt = InlineTree::build_inline_tree_root();
760
761 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
762 assert(num_alias_types() >= AliasIdxRaw, "");
763
764 #define MINIMUM_NODE_HASH 1023
765
766 // GVN that will be run immediately on new nodes
767 uint estimated_size = method()->code_size()*4+64;
768 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
769 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
770 _types = new (comp_arena()) Type_Array(comp_arena());
771 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
772 PhaseGVN gvn;
773 set_initial_gvn(&gvn);
774
775 { // Scope for timing the parser
776 TracePhase tp(_t_parser);
777
778 // Put top into the hash table ASAP.
779 initial_gvn()->transform(top());
780
781 // Set up tf(), start(), and find a CallGenerator.
782 CallGenerator* cg = nullptr;
783 if (is_osr_compilation()) {
784 const TypeTuple *domain = StartOSRNode::osr_domain();
785 const TypeTuple *range = TypeTuple::make_range(method()->signature());
786 init_tf(TypeFunc::make(domain, range));
787 StartNode* s = new StartOSRNode(root(), domain);
788 initial_gvn()->set_type_bottom(s);
789 verify_start(s);
790 cg = CallGenerator::for_osr(method(), entry_bci());
791 } else {
792 // Normal case.
793 init_tf(TypeFunc::make(method()));
794 StartNode* s = new StartNode(root(), tf()->domain());
795 initial_gvn()->set_type_bottom(s);
796 verify_start(s);
797 float past_uses = method()->interpreter_invocation_count();
798 float expected_uses = past_uses;
799 cg = CallGenerator::for_inline(method(), expected_uses);
800 }
801 if (failing()) return;
802 if (cg == nullptr) {
803 const char* reason = InlineTree::check_can_parse(method());
804 assert(reason != nullptr, "expect reason for parse failure");
805 stringStream ss;
806 ss.print("cannot parse method: %s", reason);
807 record_method_not_compilable(ss.as_string());
808 return;
809 }
810
811 gvn.set_type(root(), root()->bottom_type());
812
813 JVMState* jvms = build_start_state(start(), tf());
814 if ((jvms = cg->generate(jvms)) == nullptr) {
815 assert(failure_reason() != nullptr, "expect reason for parse failure");
816 stringStream ss;
817 ss.print("method parse failed: %s", failure_reason());
818 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
819 return;
820 }
821 GraphKit kit(jvms);
822
823 if (!kit.stopped()) {
824 // Accept return values, and transfer control we know not where.
825 // This is done by a special, unique ReturnNode bound to root.
826 return_values(kit.jvms());
827 }
828
829 if (kit.has_exceptions()) {
830 // Any exceptions that escape from this call must be rethrown
831 // to whatever caller is dynamically above us on the stack.
832 // This is done by a special, unique RethrowNode bound to root.
833 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
834 }
835
836 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
837
838 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
839 inline_string_calls(true);
840 }
841
842 if (failing()) return;
843
844 // Remove clutter produced by parsing.
845 if (!failing()) {
846 ResourceMark rm;
847 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
848 }
849 }
850
851 // Note: Large methods are capped off in do_one_bytecode().
852 if (failing()) return;
853
854 // After parsing, node notes are no longer automagic.
855 // They must be propagated by register_new_node_with_optimizer(),
856 // clone(), or the like.
857 set_default_node_notes(nullptr);
858
859 #ifndef PRODUCT
860 if (should_print_igv(1)) {
861 _igv_printer->print_inlining();
862 }
863 #endif
864
865 if (failing()) return;
866 NOT_PRODUCT( verify_graph_edges(); )
867
868 // Now optimize
869 Optimize();
870 if (failing()) return;
871 NOT_PRODUCT( verify_graph_edges(); )
872
873 #ifndef PRODUCT
874 if (should_print_ideal()) {
875 print_ideal_ir("PrintIdeal");
876 }
877 #endif
878
879 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
880 bs->final_refinement(this);
881
882 #ifdef ASSERT
883 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
884 #endif
885
886 // Dump compilation data to replay it.
887 if (directive->DumpReplayOption) {
888 env()->dump_replay_data(_compile_id);
889 }
890 if (directive->DumpInlineOption && (ilt() != nullptr)) {
891 env()->dump_inline_data(_compile_id);
892 }
893
894 // Now that we know the size of all the monitors we can add a fixed slot
895 // for the original deopt pc.
896 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
897 set_fixed_slots(next_slot);
898
899 // Compute when to use implicit null checks. Used by matching trap based
900 // nodes and NullCheck optimization.
901 set_allowed_deopt_reasons();
902
903 // Now generate code
904 Code_Gen();
905 }
906
907 //------------------------------Compile----------------------------------------
908 // Compile a runtime stub
909 Compile::Compile(ciEnv* ci_env,
910 TypeFunc_generator generator,
911 address stub_function,
912 const char* stub_name,
913 StubId stub_id,
914 int is_fancy_jump,
915 bool pass_tls,
916 bool return_pc,
917 DirectiveSet* directive)
918 : Phase(Compiler),
919 _compile_id(0),
920 _options(Options::for_runtime_stub()),
921 _method(nullptr),
922 _entry_bci(InvocationEntryBci),
923 _stub_function(stub_function),
924 _stub_name(stub_name),
925 _stub_id(stub_id),
926 _stub_entry_point(nullptr),
927 _max_node_limit(MaxNodeLimit),
928 _node_count_inlining_cutoff(NodeCountInliningCutoff),
929 _post_loop_opts_phase(false),
930 _merge_stores_phase(false),
931 _allow_macro_nodes(true),
932 _inlining_progress(false),
933 _inlining_incrementally(false),
934 _has_reserved_stack_access(false),
935 #ifndef PRODUCT
936 _igv_idx(0),
937 _trace_opto_output(directive->TraceOptoOutputOption),
938 #endif
939 _clinit_barrier_on_entry(false),
940 _stress_seed(0),
941 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
942 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
943 _env(ci_env),
944 _directive(directive),
945 _log(ci_env->log()),
946 _first_failure_details(nullptr),
947 _reachability_fences(comp_arena(), 8, 0, nullptr),
948 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
949 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
950 _congraph(nullptr),
951 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
952 _unique(0),
953 _dead_node_count(0),
954 _dead_node_list(comp_arena()),
955 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
956 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
957 _node_arena(&_node_arena_one),
958 _mach_constant_base_node(nullptr),
959 _Compile_types(mtCompiler, Arena::Tag::tag_type),
960 _initial_gvn(nullptr),
961 _igvn_worklist(nullptr),
962 _types(nullptr),
963 _node_hash(nullptr),
964 _has_mh_late_inlines(false),
965 _oom(false),
966 _replay_inline_data(nullptr),
967 _inline_printer(this),
968 _java_calls(0),
969 _inner_loops(0),
970 _FIRST_STACK_mask(comp_arena()),
971 _interpreter_frame_size(0),
972 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
973 _output(nullptr),
974 #ifndef PRODUCT
975 _in_dump_cnt(0),
976 #endif
977 _allowed_reasons(0) {
978 C = this;
979
980 // try to reuse an existing stub
981 {
982 BlobId blob_id = StubInfo::blob(_stub_id);
983 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
984 if (blob != nullptr) {
985 RuntimeStub* rs = blob->as_runtime_stub();
986 _stub_entry_point = rs->entry_point();
987 return;
988 }
989 }
990
991 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
992 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
993
994 #ifndef PRODUCT
995 set_print_assembly(PrintFrameConverterAssembly);
996 set_parsed_irreducible_loop(false);
997 #else
998 set_print_assembly(false); // Must initialize.
999 #endif
1000 set_has_irreducible_loop(false); // no loops
1001
1002 CompileWrapper cw(this);
1003 Init(/*do_aliasing=*/ false);
1004 init_tf((*generator)());
1005
1006 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1007 _types = new (comp_arena()) Type_Array(comp_arena());
1008 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1009
1010 if (StressLCM || StressGCM || StressBailout) {
1011 initialize_stress_seed(directive);
1012 }
1013
1014 {
1015 PhaseGVN gvn;
1016 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1017 gvn.transform(top());
1018
1019 GraphKit kit;
1020 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1021 }
1022
1023 NOT_PRODUCT( verify_graph_edges(); )
1024
1025 Code_Gen();
1026 }
1027
1028 Compile::~Compile() {
1029 delete _first_failure_details;
1030 };
1031
1032 //------------------------------Init-------------------------------------------
1033 // Prepare for a single compilation
1034 void Compile::Init(bool aliasing) {
1035 _do_aliasing = aliasing;
1036 _unique = 0;
1037 _regalloc = nullptr;
1038
1039 _tf = nullptr; // filled in later
1040 _top = nullptr; // cached later
1041 _matcher = nullptr; // filled in later
1042 _cfg = nullptr; // filled in later
1043
1044 _node_note_array = nullptr;
1045 _default_node_notes = nullptr;
1046 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1047
1048 _immutable_memory = nullptr; // filled in at first inquiry
1049
1050 #ifdef ASSERT
1051 _phase_optimize_finished = false;
1052 _phase_verify_ideal_loop = false;
1053 _exception_backedge = false;
1054 _type_verify = nullptr;
1055 #endif
1056
1057 // Globally visible Nodes
1058 // First set TOP to null to give safe behavior during creation of RootNode
1059 set_cached_top_node(nullptr);
1060 set_root(new RootNode());
1061 // Now that you have a Root to point to, create the real TOP
1062 set_cached_top_node( new ConNode(Type::TOP) );
1063 set_recent_alloc(nullptr, nullptr);
1064
1065 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1066 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1067 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1068 env()->set_dependencies(new Dependencies(env()));
1069
1070 _fixed_slots = 0;
1071 set_has_split_ifs(false);
1072 set_has_loops(false); // first approximation
1073 set_has_stringbuilder(false);
1074 set_has_boxed_value(false);
1075 _trap_can_recompile = false; // no traps emitted yet
1076 _major_progress = true; // start out assuming good things will happen
1077 set_has_unsafe_access(false);
1078 set_max_vector_size(0);
1079 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1080 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1081 set_decompile_count(0);
1082
1083 #ifndef PRODUCT
1084 _phase_counter = 0;
1085 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1086 #endif
1087
1088 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1089 _loop_opts_cnt = LoopOptsCount;
1090 set_do_inlining(Inline);
1091 set_max_inline_size(MaxInlineSize);
1092 set_freq_inline_size(FreqInlineSize);
1093 set_do_scheduling(OptoScheduling);
1094
1095 set_do_vector_loop(false);
1096 set_has_monitors(false);
1097 set_has_scoped_access(false);
1098
1099 if (AllowVectorizeOnDemand) {
1100 if (has_method() && _directive->VectorizeOption) {
1101 set_do_vector_loop(true);
1102 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());})
1103 } else if (has_method() && method()->name() != nullptr &&
1104 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1105 set_do_vector_loop(true);
1106 }
1107 }
1108 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1109 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());})
1110
1111 _max_node_limit = _directive->MaxNodeLimitOption;
1112
1113 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1114 set_clinit_barrier_on_entry(true);
1115 }
1116 if (debug_info()->recording_non_safepoints()) {
1117 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1118 (comp_arena(), 8, 0, nullptr));
1119 set_default_node_notes(Node_Notes::make(this));
1120 }
1121
1122 const int grow_ats = 16;
1123 _max_alias_types = grow_ats;
1124 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1125 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1126 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1127 {
1128 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1129 }
1130 // Initialize the first few types.
1131 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1132 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1133 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1134 _num_alias_types = AliasIdxRaw+1;
1135 // Zero out the alias type cache.
1136 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1137 // A null adr_type hits in the cache right away. Preload the right answer.
1138 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1139 }
1140
1141 #ifdef ASSERT
1142 // Verify that the current StartNode is valid.
1143 void Compile::verify_start(StartNode* s) const {
1144 assert(failing_internal() || s == start(), "should be StartNode");
1145 }
1146 #endif
1147
1148 /**
1149 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1150 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1151 * the ideal graph.
1152 */
1153 StartNode* Compile::start() const {
1154 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1155 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1156 Node* start = root()->fast_out(i);
1157 if (start->is_Start()) {
1158 return start->as_Start();
1159 }
1160 }
1161 fatal("Did not find Start node!");
1162 return nullptr;
1163 }
1164
1165 //-------------------------------immutable_memory-------------------------------------
1166 // Access immutable memory
1167 Node* Compile::immutable_memory() {
1168 if (_immutable_memory != nullptr) {
1169 return _immutable_memory;
1170 }
1171 StartNode* s = start();
1172 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1173 Node *p = s->fast_out(i);
1174 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1175 _immutable_memory = p;
1176 return _immutable_memory;
1177 }
1178 }
1179 ShouldNotReachHere();
1180 return nullptr;
1181 }
1182
1183 //----------------------set_cached_top_node------------------------------------
1184 // Install the cached top node, and make sure Node::is_top works correctly.
1185 void Compile::set_cached_top_node(Node* tn) {
1186 if (tn != nullptr) verify_top(tn);
1187 Node* old_top = _top;
1188 _top = tn;
1189 // Calling Node::setup_is_top allows the nodes the chance to adjust
1190 // their _out arrays.
1191 if (_top != nullptr) _top->setup_is_top();
1192 if (old_top != nullptr) old_top->setup_is_top();
1193 assert(_top == nullptr || top()->is_top(), "");
1194 }
1195
1196 #ifdef ASSERT
1197 uint Compile::count_live_nodes_by_graph_walk() {
1198 Unique_Node_List useful(comp_arena());
1199 // Get useful node list by walking the graph.
1200 identify_useful_nodes(useful);
1201 return useful.size();
1202 }
1203
1204 void Compile::print_missing_nodes() {
1205
1206 // Return if CompileLog is null and PrintIdealNodeCount is false.
1207 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1208 return;
1209 }
1210
1211 // This is an expensive function. It is executed only when the user
1212 // specifies VerifyIdealNodeCount option or otherwise knows the
1213 // additional work that needs to be done to identify reachable nodes
1214 // by walking the flow graph and find the missing ones using
1215 // _dead_node_list.
1216
1217 Unique_Node_List useful(comp_arena());
1218 // Get useful node list by walking the graph.
1219 identify_useful_nodes(useful);
1220
1221 uint l_nodes = C->live_nodes();
1222 uint l_nodes_by_walk = useful.size();
1223
1224 if (l_nodes != l_nodes_by_walk) {
1225 if (_log != nullptr) {
1226 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1227 _log->stamp();
1228 _log->end_head();
1229 }
1230 VectorSet& useful_member_set = useful.member_set();
1231 int last_idx = l_nodes_by_walk;
1232 for (int i = 0; i < last_idx; i++) {
1233 if (useful_member_set.test(i)) {
1234 if (_dead_node_list.test(i)) {
1235 if (_log != nullptr) {
1236 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1237 }
1238 if (PrintIdealNodeCount) {
1239 // Print the log message to tty
1240 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1241 useful.at(i)->dump();
1242 }
1243 }
1244 }
1245 else if (! _dead_node_list.test(i)) {
1246 if (_log != nullptr) {
1247 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1248 }
1249 if (PrintIdealNodeCount) {
1250 // Print the log message to tty
1251 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1252 }
1253 }
1254 }
1255 if (_log != nullptr) {
1256 _log->tail("mismatched_nodes");
1257 }
1258 }
1259 }
1260 void Compile::record_modified_node(Node* n) {
1261 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1262 _modified_nodes->push(n);
1263 }
1264 }
1265
1266 void Compile::remove_modified_node(Node* n) {
1267 if (_modified_nodes != nullptr) {
1268 _modified_nodes->remove(n);
1269 }
1270 }
1271 #endif
1272
1273 #ifndef PRODUCT
1274 void Compile::verify_top(Node* tn) const {
1275 if (tn != nullptr) {
1276 assert(tn->is_Con(), "top node must be a constant");
1277 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1278 assert(tn->in(0) != nullptr, "must have live top node");
1279 }
1280 }
1281 #endif
1282
1283
1284 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1285
1286 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1287 guarantee(arr != nullptr, "");
1288 int num_blocks = arr->length();
1289 if (grow_by < num_blocks) grow_by = num_blocks;
1290 int num_notes = grow_by * _node_notes_block_size;
1291 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1292 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1293 while (num_notes > 0) {
1294 arr->append(notes);
1295 notes += _node_notes_block_size;
1296 num_notes -= _node_notes_block_size;
1297 }
1298 assert(num_notes == 0, "exact multiple, please");
1299 }
1300
1301 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1302 if (source == nullptr || dest == nullptr) return false;
1303
1304 if (dest->is_Con())
1305 return false; // Do not push debug info onto constants.
1306
1307 #ifdef ASSERT
1308 // Leave a bread crumb trail pointing to the original node:
1309 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1310 dest->set_debug_orig(source);
1311 }
1312 #endif
1313
1314 if (node_note_array() == nullptr)
1315 return false; // Not collecting any notes now.
1316
1317 // This is a copy onto a pre-existing node, which may already have notes.
1318 // If both nodes have notes, do not overwrite any pre-existing notes.
1319 Node_Notes* source_notes = node_notes_at(source->_idx);
1320 if (source_notes == nullptr || source_notes->is_clear()) return false;
1321 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1322 if (dest_notes == nullptr || dest_notes->is_clear()) {
1323 return set_node_notes_at(dest->_idx, source_notes);
1324 }
1325
1326 Node_Notes merged_notes = (*source_notes);
1327 // The order of operations here ensures that dest notes will win...
1328 merged_notes.update_from(dest_notes);
1329 return set_node_notes_at(dest->_idx, &merged_notes);
1330 }
1331
1332
1333 //--------------------------allow_range_check_smearing-------------------------
1334 // Gating condition for coalescing similar range checks.
1335 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1336 // single covering check that is at least as strong as any of them.
1337 // If the optimization succeeds, the simplified (strengthened) range check
1338 // will always succeed. If it fails, we will deopt, and then give up
1339 // on the optimization.
1340 bool Compile::allow_range_check_smearing() const {
1341 // If this method has already thrown a range-check,
1342 // assume it was because we already tried range smearing
1343 // and it failed.
1344 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1345 return !already_trapped;
1346 }
1347
1348
1349 //------------------------------flatten_alias_type-----------------------------
1350 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1351 assert(do_aliasing(), "Aliasing should be enabled");
1352 int offset = tj->offset();
1353 TypePtr::PTR ptr = tj->ptr();
1354
1355 // Known instance (scalarizable allocation) alias only with itself.
1356 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1357 tj->is_oopptr()->is_known_instance();
1358
1359 // Process weird unsafe references.
1360 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1361 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1362 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1363 tj = TypeOopPtr::BOTTOM;
1364 ptr = tj->ptr();
1365 offset = tj->offset();
1366 }
1367
1368 // Array pointers need some flattening
1369 const TypeAryPtr* ta = tj->isa_aryptr();
1370 if (ta && ta->is_stable()) {
1371 // Erase stability property for alias analysis.
1372 tj = ta = ta->cast_to_stable(false);
1373 }
1374 if( ta && is_known_inst ) {
1375 if ( offset != Type::OffsetBot &&
1376 offset > arrayOopDesc::length_offset_in_bytes() ) {
1377 offset = Type::OffsetBot; // Flatten constant access into array body only
1378 tj = ta = ta->
1379 remove_speculative()->
1380 cast_to_ptr_type(ptr)->
1381 with_offset(offset);
1382 }
1383 } else if (ta != nullptr) {
1384 // Common slices
1385 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1386 return TypeAryPtr::RANGE;
1387 } else if (offset == oopDesc::klass_offset_in_bytes()) {
1388 return TypeInstPtr::KLASS;
1389 } else if (offset == oopDesc::mark_offset_in_bytes()) {
1390 return TypeInstPtr::MARK;
1391 }
1392
1393 // Remove size and stability
1394 const TypeAry* normalized_ary = TypeAry::make(ta->elem(), TypeInt::POS, false);
1395 // Remove ptr, const_oop, and offset
1396 if (ta->elem() == Type::BOTTOM) {
1397 // Bottom array (meet of int[] and byte[] for example), accesses to it will be done with
1398 // Unsafe. This should alias with all arrays. For now just leave it as it is (this is
1399 // incorrect, see JDK-8331133).
1400 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, false, Type::OffsetBot);
1401 } else if (ta->elem()->make_oopptr() != nullptr) {
1402 // Object arrays, all of them share the same slice
1403 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, TypeInt::POS, false);
1404 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, tary, nullptr, false, Type::OffsetBot);
1405 } else {
1406 // Primitive arrays
1407 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, ta->exact_klass(), true, Type::OffsetBot);
1408 }
1409
1410 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1411 // cannot be distinguished by bytecode alone.
1412 if (ta->elem() == TypeInt::BOOL) {
1413 tj = ta = TypeAryPtr::BYTES;
1414 }
1415 }
1416
1417 // Oop pointers need some flattening
1418 const TypeInstPtr *to = tj->isa_instptr();
1419 if (to && to != TypeOopPtr::BOTTOM) {
1420 ciInstanceKlass* ik = to->instance_klass();
1421 if( ptr == TypePtr::Constant ) {
1422 if (ik != ciEnv::current()->Class_klass() ||
1423 offset < ik->layout_helper_size_in_bytes()) {
1424 // No constant oop pointers (such as Strings); they alias with
1425 // unknown strings.
1426 assert(!is_known_inst, "not scalarizable allocation");
1427 tj = to = to->
1428 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1429 remove_speculative()->
1430 cast_to_ptr_type(TypePtr::BotPTR)->
1431 cast_to_exactness(false);
1432 }
1433 } else if( is_known_inst ) {
1434 tj = to; // Keep NotNull and klass_is_exact for instance type
1435 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1436 // During the 2nd round of IterGVN, NotNull castings are removed.
1437 // Make sure the Bottom and NotNull variants alias the same.
1438 // Also, make sure exact and non-exact variants alias the same.
1439 tj = to = to->
1440 remove_speculative()->
1441 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1442 cast_to_ptr_type(TypePtr::BotPTR)->
1443 cast_to_exactness(false);
1444 }
1445 if (to->speculative() != nullptr) {
1446 tj = to = to->remove_speculative();
1447 }
1448 // Canonicalize the holder of this field
1449 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1450 // First handle header references such as a LoadKlassNode, even if the
1451 // object's klass is unloaded at compile time (4965979).
1452 if (!is_known_inst) { // Do it only for non-instance types
1453 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1454 }
1455 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1456 // Static fields are in the space above the normal instance
1457 // fields in the java.lang.Class instance.
1458 if (ik != ciEnv::current()->Class_klass()) {
1459 to = nullptr;
1460 tj = TypeOopPtr::BOTTOM;
1461 offset = tj->offset();
1462 }
1463 } else {
1464 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1465 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1466 assert(tj->offset() == offset, "no change to offset expected");
1467 bool xk = to->klass_is_exact();
1468 int instance_id = to->instance_id();
1469
1470 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1471 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1472 // its interfaces are included.
1473 if (xk && ik->equals(canonical_holder)) {
1474 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1475 } else {
1476 assert(xk || !is_known_inst, "Known instance should be exact type");
1477 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1478 }
1479 }
1480 }
1481
1482 // Klass pointers to object array klasses need some flattening
1483 const TypeKlassPtr *tk = tj->isa_klassptr();
1484 if( tk ) {
1485 // If we are referencing a field within a Klass, we need
1486 // to assume the worst case of an Object. Both exact and
1487 // inexact types must flatten to the same alias class so
1488 // use NotNull as the PTR.
1489 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1490 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1491 env()->Object_klass(),
1492 offset);
1493 }
1494
1495 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1496 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1497 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1498 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1499 } else {
1500 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1501 }
1502 }
1503
1504 // Check for precise loads from the primary supertype array and force them
1505 // to the supertype cache alias index. Check for generic array loads from
1506 // the primary supertype array and also force them to the supertype cache
1507 // alias index. Since the same load can reach both, we need to merge
1508 // these 2 disparate memories into the same alias class. Since the
1509 // primary supertype array is read-only, there's no chance of confusion
1510 // where we bypass an array load and an array store.
1511 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1512 if (offset == Type::OffsetBot ||
1513 (offset >= primary_supers_offset &&
1514 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1515 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1516 offset = in_bytes(Klass::secondary_super_cache_offset());
1517 tj = tk = tk->with_offset(offset);
1518 }
1519 }
1520
1521 // Flatten all Raw pointers together.
1522 if (tj->base() == Type::RawPtr)
1523 tj = TypeRawPtr::BOTTOM;
1524
1525 if (tj->base() == Type::AnyPtr)
1526 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1527
1528 offset = tj->offset();
1529 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1530
1531 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1532 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1533 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1534 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1535 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1536 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1537 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1538 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1539 assert( tj->ptr() != TypePtr::TopPTR &&
1540 tj->ptr() != TypePtr::AnyNull &&
1541 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1542 // assert( tj->ptr() != TypePtr::Constant ||
1543 // tj->base() == Type::RawPtr ||
1544 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1545
1546 return tj;
1547 }
1548
1549 void Compile::AliasType::Init(int i, const TypePtr* at) {
1550 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1551 _index = i;
1552 _adr_type = at;
1553 _field = nullptr;
1554 _element = nullptr;
1555 _is_rewritable = true; // default
1556 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1557 if (atoop != nullptr && atoop->is_known_instance()) {
1558 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1559 _general_index = Compile::current()->get_alias_index(gt);
1560 } else {
1561 _general_index = 0;
1562 }
1563 }
1564
1565 BasicType Compile::AliasType::basic_type() const {
1566 if (element() != nullptr) {
1567 const Type* element = adr_type()->is_aryptr()->elem();
1568 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1569 } if (field() != nullptr) {
1570 return field()->layout_type();
1571 } else {
1572 return T_ILLEGAL; // unknown
1573 }
1574 }
1575
1576 //---------------------------------print_on------------------------------------
1577 #ifndef PRODUCT
1578 void Compile::AliasType::print_on(outputStream* st) {
1579 if (index() < 10)
1580 st->print("@ <%d> ", index());
1581 else st->print("@ <%d>", index());
1582 st->print(is_rewritable() ? " " : " RO");
1583 int offset = adr_type()->offset();
1584 if (offset == Type::OffsetBot)
1585 st->print(" +any");
1586 else st->print(" +%-3d", offset);
1587 st->print(" in ");
1588 adr_type()->dump_on(st);
1589 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1590 if (field() != nullptr && tjp) {
1591 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1592 tjp->offset() != field()->offset_in_bytes()) {
1593 st->print(" != ");
1594 field()->print();
1595 st->print(" ***");
1596 }
1597 }
1598 }
1599
1600 void print_alias_types() {
1601 Compile* C = Compile::current();
1602 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1603 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1604 C->alias_type(idx)->print_on(tty);
1605 tty->cr();
1606 }
1607 }
1608 #endif
1609
1610
1611 //----------------------------probe_alias_cache--------------------------------
1612 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1613 intptr_t key = (intptr_t) adr_type;
1614 key ^= key >> logAliasCacheSize;
1615 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1616 }
1617
1618
1619 //-----------------------------grow_alias_types--------------------------------
1620 void Compile::grow_alias_types() {
1621 const int old_ats = _max_alias_types; // how many before?
1622 const int new_ats = old_ats; // how many more?
1623 const int grow_ats = old_ats+new_ats; // how many now?
1624 _max_alias_types = grow_ats;
1625 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), _alias_types, old_ats, grow_ats);
1626 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1627 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1628 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1629 }
1630
1631
1632 //--------------------------------find_alias_type------------------------------
1633 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1634 if (!do_aliasing()) {
1635 return alias_type(AliasIdxBot);
1636 }
1637
1638 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1639 if (ace->_adr_type == adr_type) {
1640 return alias_type(ace->_index);
1641 }
1642
1643 // Handle special cases.
1644 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1645 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1646
1647 // Do it the slow way.
1648 const TypePtr* flat = flatten_alias_type(adr_type);
1649
1650 #ifdef ASSERT
1651 {
1652 ResourceMark rm;
1653 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1654 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1655 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1656 Type::str(adr_type));
1657 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1658 const TypeOopPtr* foop = flat->is_oopptr();
1659 // Scalarizable allocations have exact klass always.
1660 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1661 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1662 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1663 Type::str(foop), Type::str(xoop));
1664 }
1665 }
1666 #endif
1667
1668 int idx = AliasIdxTop;
1669 for (int i = 0; i < num_alias_types(); i++) {
1670 if (alias_type(i)->adr_type() == flat) {
1671 idx = i;
1672 break;
1673 }
1674 }
1675
1676 if (idx == AliasIdxTop) {
1677 if (no_create) return nullptr;
1678 // Grow the array if necessary.
1679 if (_num_alias_types == _max_alias_types) grow_alias_types();
1680 // Add a new alias type.
1681 idx = _num_alias_types++;
1682 _alias_types[idx]->Init(idx, flat);
1683 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1684 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1685 if (flat->isa_instptr()) {
1686 if (flat->offset() == java_lang_Class::klass_offset()
1687 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1688 alias_type(idx)->set_rewritable(false);
1689 }
1690 if (flat->isa_aryptr()) {
1691 #ifdef ASSERT
1692 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1693 // (T_BYTE has the weakest alignment and size restrictions...)
1694 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1695 #endif
1696 if (flat->offset() == TypePtr::OffsetBot) {
1697 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1698 }
1699 }
1700 if (flat->isa_klassptr()) {
1701 if (UseCompactObjectHeaders) {
1702 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1703 alias_type(idx)->set_rewritable(false);
1704 }
1705 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1706 alias_type(idx)->set_rewritable(false);
1707 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1708 alias_type(idx)->set_rewritable(false);
1709 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1710 alias_type(idx)->set_rewritable(false);
1711 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1712 alias_type(idx)->set_rewritable(false);
1713 }
1714
1715 if (flat->isa_instklassptr()) {
1716 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1717 alias_type(idx)->set_rewritable(false);
1718 }
1719 }
1720 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1721 // but the base pointer type is not distinctive enough to identify
1722 // references into JavaThread.)
1723
1724 // Check for final fields.
1725 const TypeInstPtr* tinst = flat->isa_instptr();
1726 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1727 ciField* field;
1728 if (tinst->const_oop() != nullptr &&
1729 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1730 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1731 // static field
1732 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1733 field = k->get_field_by_offset(tinst->offset(), true);
1734 } else {
1735 ciInstanceKlass *k = tinst->instance_klass();
1736 field = k->get_field_by_offset(tinst->offset(), false);
1737 }
1738 assert(field == nullptr ||
1739 original_field == nullptr ||
1740 (field->holder() == original_field->holder() &&
1741 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1742 field->is_static() == original_field->is_static()), "wrong field?");
1743 // Set field() and is_rewritable() attributes.
1744 if (field != nullptr) alias_type(idx)->set_field(field);
1745 }
1746 }
1747
1748 // Fill the cache for next time.
1749 ace->_adr_type = adr_type;
1750 ace->_index = idx;
1751 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1752
1753 // Might as well try to fill the cache for the flattened version, too.
1754 AliasCacheEntry* face = probe_alias_cache(flat);
1755 if (face->_adr_type == nullptr) {
1756 face->_adr_type = flat;
1757 face->_index = idx;
1758 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1759 }
1760
1761 return alias_type(idx);
1762 }
1763
1764
1765 Compile::AliasType* Compile::alias_type(ciField* field) {
1766 const TypeOopPtr* t;
1767 if (field->is_static())
1768 t = TypeInstPtr::make(field->holder()->java_mirror());
1769 else
1770 t = TypeOopPtr::make_from_klass_raw(field->holder());
1771 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1772 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1773 return atp;
1774 }
1775
1776
1777 //------------------------------have_alias_type--------------------------------
1778 bool Compile::have_alias_type(const TypePtr* adr_type) {
1779 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1780 if (ace->_adr_type == adr_type) {
1781 return true;
1782 }
1783
1784 // Handle special cases.
1785 if (adr_type == nullptr) return true;
1786 if (adr_type == TypePtr::BOTTOM) return true;
1787
1788 return find_alias_type(adr_type, true, nullptr) != nullptr;
1789 }
1790
1791 //-----------------------------must_alias--------------------------------------
1792 // True if all values of the given address type are in the given alias category.
1793 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1794 if (alias_idx == AliasIdxBot) return true; // the universal category
1795 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1796 if (alias_idx == AliasIdxTop) return false; // the empty category
1797 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1798
1799 // the only remaining possible overlap is identity
1800 int adr_idx = get_alias_index(adr_type);
1801 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1802 assert(adr_idx == alias_idx ||
1803 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1804 && adr_type != TypeOopPtr::BOTTOM),
1805 "should not be testing for overlap with an unsafe pointer");
1806 return adr_idx == alias_idx;
1807 }
1808
1809 //------------------------------can_alias--------------------------------------
1810 // True if any values of the given address type are in the given alias category.
1811 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1812 if (alias_idx == AliasIdxTop) return false; // the empty category
1813 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1814 // Known instance doesn't alias with bottom memory
1815 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1816 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1817
1818 // the only remaining possible overlap is identity
1819 int adr_idx = get_alias_index(adr_type);
1820 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1821 return adr_idx == alias_idx;
1822 }
1823
1824 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1825 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1826 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1827 if (parse_predicate_count() == 0) {
1828 return;
1829 }
1830 for (int i = 0; i < parse_predicate_count(); i++) {
1831 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1832 parse_predicate->mark_useless(igvn);
1833 }
1834 _parse_predicates.clear();
1835 }
1836
1837 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1838 if (!n->for_post_loop_opts_igvn()) {
1839 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1840 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1841 _for_post_loop_igvn.append(n);
1842 }
1843 }
1844
1845 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1846 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1847 _for_post_loop_igvn.remove(n);
1848 }
1849
1850 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1851 // Verify that all previous optimizations produced a valid graph
1852 // at least to this point, even if no loop optimizations were done.
1853 PhaseIdealLoop::verify(igvn);
1854
1855 if (_print_phase_loop_opts) {
1856 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1857 }
1858 C->set_post_loop_opts_phase(); // no more loop opts allowed
1859
1860 assert(!C->major_progress(), "not cleared");
1861
1862 if (_for_post_loop_igvn.length() > 0) {
1863 while (_for_post_loop_igvn.length() > 0) {
1864 Node* n = _for_post_loop_igvn.pop();
1865 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1866 igvn._worklist.push(n);
1867 }
1868 igvn.optimize();
1869 if (failing()) return;
1870 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1871 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1872
1873 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1874 if (C->major_progress()) {
1875 C->clear_major_progress(); // ensure that major progress is now clear
1876 }
1877 }
1878 }
1879
1880 void Compile::record_for_merge_stores_igvn(Node* n) {
1881 if (!n->for_merge_stores_igvn()) {
1882 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1883 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1884 _for_merge_stores_igvn.append(n);
1885 }
1886 }
1887
1888 void Compile::remove_from_merge_stores_igvn(Node* n) {
1889 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1890 _for_merge_stores_igvn.remove(n);
1891 }
1892
1893 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1894 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1895 // the stores, and we merge the wrong sequence of stores.
1896 // Example:
1897 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1898 // Apply MergeStores:
1899 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1900 // Remove more RangeChecks:
1901 // StoreI [ StoreL ] StoreI
1902 // But now it would have been better to do this instead:
1903 // [ StoreL ] [ StoreL ]
1904 //
1905 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1906 // since we never unset _merge_stores_phase.
1907 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1908 C->set_merge_stores_phase();
1909
1910 if (_for_merge_stores_igvn.length() > 0) {
1911 while (_for_merge_stores_igvn.length() > 0) {
1912 Node* n = _for_merge_stores_igvn.pop();
1913 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1914 igvn._worklist.push(n);
1915 }
1916 igvn.optimize();
1917 if (failing()) return;
1918 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1919 print_method(PHASE_AFTER_MERGE_STORES, 3);
1920 }
1921 }
1922
1923 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1924 if (OptimizeUnstableIf) {
1925 _unstable_if_traps.append(trap);
1926 }
1927 }
1928
1929 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1930 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1931 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1932 Node* n = trap->uncommon_trap();
1933 if (!useful.member(n)) {
1934 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1935 }
1936 }
1937 }
1938
1939 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1940 // or fold-compares case. Return true if succeed or not found.
1941 //
1942 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1943 // false and ask fold-compares to yield.
1944 //
1945 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1946 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1947 // when deoptimization does happen.
1948 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1949 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1950 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1951 if (trap->uncommon_trap() == unc) {
1952 if (yield && trap->modified()) {
1953 return false;
1954 }
1955 _unstable_if_traps.delete_at(i);
1956 break;
1957 }
1958 }
1959 return true;
1960 }
1961
1962 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1963 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1964 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1965 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1966 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1967 CallStaticJavaNode* unc = trap->uncommon_trap();
1968 int next_bci = trap->next_bci();
1969 bool modified = trap->modified();
1970
1971 if (next_bci != -1 && !modified) {
1972 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1973 JVMState* jvms = unc->jvms();
1974 ciMethod* method = jvms->method();
1975 ciBytecodeStream iter(method);
1976
1977 iter.force_bci(jvms->bci());
1978 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1979 Bytecodes::Code c = iter.cur_bc();
1980 Node* lhs = nullptr;
1981 Node* rhs = nullptr;
1982 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1983 lhs = unc->peek_operand(0);
1984 rhs = unc->peek_operand(1);
1985 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1986 lhs = unc->peek_operand(0);
1987 }
1988
1989 ResourceMark rm;
1990 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1991 assert(live_locals.is_valid(), "broken liveness info");
1992 int len = (int)live_locals.size();
1993
1994 for (int i = 0; i < len; i++) {
1995 Node* local = unc->local(jvms, i);
1996 // kill local using the liveness of next_bci.
1997 // give up when the local looks like an operand to secure reexecution.
1998 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1999 uint idx = jvms->locoff() + i;
2000 #ifdef ASSERT
2001 if (PrintOpto && Verbose) {
2002 tty->print("[unstable_if] kill local#%d: ", idx);
2003 local->dump();
2004 tty->cr();
2005 }
2006 #endif
2007 igvn.replace_input_of(unc, idx, top());
2008 modified = true;
2009 }
2010 }
2011 }
2012
2013 // keep the mondified trap for late query
2014 if (modified) {
2015 trap->set_modified();
2016 } else {
2017 _unstable_if_traps.delete_at(i);
2018 }
2019 }
2020 igvn.optimize();
2021 }
2022
2023 // StringOpts and late inlining of string methods
2024 void Compile::inline_string_calls(bool parse_time) {
2025 {
2026 // remove useless nodes to make the usage analysis simpler
2027 ResourceMark rm;
2028 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2029 }
2030
2031 {
2032 ResourceMark rm;
2033 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2034 PhaseStringOpts pso(initial_gvn());
2035 print_method(PHASE_AFTER_STRINGOPTS, 3);
2036 }
2037
2038 // now inline anything that we skipped the first time around
2039 if (!parse_time) {
2040 _late_inlines_pos = _late_inlines.length();
2041 }
2042
2043 while (_string_late_inlines.length() > 0) {
2044 CallGenerator* cg = _string_late_inlines.pop();
2045 cg->do_late_inline();
2046 if (failing()) return;
2047 }
2048 _string_late_inlines.trunc_to(0);
2049 }
2050
2051 // Late inlining of boxing methods
2052 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2053 if (_boxing_late_inlines.length() > 0) {
2054 assert(has_boxed_value(), "inconsistent");
2055
2056 set_inlining_incrementally(true);
2057
2058 igvn_worklist()->ensure_empty(); // should be done with igvn
2059
2060 _late_inlines_pos = _late_inlines.length();
2061
2062 while (_boxing_late_inlines.length() > 0) {
2063 CallGenerator* cg = _boxing_late_inlines.pop();
2064 cg->do_late_inline();
2065 if (failing()) return;
2066 }
2067 _boxing_late_inlines.trunc_to(0);
2068
2069 inline_incrementally_cleanup(igvn);
2070
2071 set_inlining_incrementally(false);
2072 }
2073 }
2074
2075 bool Compile::inline_incrementally_one() {
2076 assert(IncrementalInline, "incremental inlining should be on");
2077 assert(_late_inlines.length() > 0, "should have been checked by caller");
2078
2079 TracePhase tp(_t_incrInline_inline);
2080
2081 set_inlining_progress(false);
2082 set_do_cleanup(false);
2083
2084 for (int i = 0; i < _late_inlines.length(); i++) {
2085 _late_inlines_pos = i+1;
2086 CallGenerator* cg = _late_inlines.at(i);
2087 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2088 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2089 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2090 if (should_stress_inlining()) {
2091 // randomly add repeated inline attempt if stress-inlining
2092 cg->call_node()->set_generator(cg);
2093 C->igvn_worklist()->push(cg->call_node());
2094 continue;
2095 }
2096 cg->do_late_inline();
2097 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2098 if (failing()) {
2099 return false;
2100 } else if (inlining_progress()) {
2101 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2102 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2103 break; // process one call site at a time
2104 } else {
2105 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2106 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2107 // Avoid potential infinite loop if node already in the IGVN list
2108 assert(false, "scheduled for IGVN during inlining attempt");
2109 } else {
2110 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2111 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2112 cg->call_node()->set_generator(cg);
2113 }
2114 }
2115 } else {
2116 // Ignore late inline direct calls when inlining is not allowed.
2117 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2118 }
2119 }
2120 // Remove processed elements.
2121 _late_inlines.remove_till(_late_inlines_pos);
2122 _late_inlines_pos = 0;
2123
2124 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2125
2126 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2127
2128 set_inlining_progress(false);
2129 set_do_cleanup(false);
2130
2131 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2132 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2133 }
2134
2135 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2136 {
2137 TracePhase tp(_t_incrInline_pru);
2138 ResourceMark rm;
2139 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2140 }
2141 {
2142 TracePhase tp(_t_incrInline_igvn);
2143 igvn.reset();
2144 igvn.optimize();
2145 if (failing()) return;
2146 }
2147 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2148 }
2149
2150 template<typename E>
2151 static void shuffle_array(Compile& C, GrowableArray<E>& array) {
2152 if (array.length() < 2) {
2153 return;
2154 }
2155 for (uint i = array.length() - 1; i >= 1; i--) {
2156 uint j = C.random() % (i + 1);
2157 swap(array.at(i), array.at(j));
2158 }
2159 }
2160
2161 void Compile::shuffle_late_inlines() {
2162 shuffle_array(*C, _late_inlines);
2163 }
2164
2165 // Perform incremental inlining until bound on number of live nodes is reached
2166 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2167 TracePhase tp(_t_incrInline);
2168
2169 set_inlining_incrementally(true);
2170 uint low_live_nodes = 0;
2171
2172 if (StressIncrementalInlining) {
2173 shuffle_late_inlines();
2174 }
2175
2176 while (_late_inlines.length() > 0) {
2177 if (live_nodes() > node_count_inlining_cutoff()) {
2178 if (low_live_nodes < node_count_inlining_cutoff() * 8 / 10) {
2179 TracePhase tp(_t_incrInline_ideal);
2180 // PhaseIdealLoop is expensive so we only try it once we are
2181 // out of live nodes and we only try it again if the previous
2182 // helped got the number of nodes down significantly
2183 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2184 if (failing()) return;
2185 low_live_nodes = live_nodes();
2186 _major_progress = true;
2187 }
2188
2189 if (live_nodes() > node_count_inlining_cutoff()) {
2190 bool do_print_inlining = print_inlining() || print_intrinsics();
2191 if (do_print_inlining || log() != nullptr) {
2192 // Print inlining message for candidates that we couldn't inline for lack of space.
2193 for (int i = 0; i < _late_inlines.length(); i++) {
2194 CallGenerator* cg = _late_inlines.at(i);
2195 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2196 if (do_print_inlining) {
2197 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2198 }
2199 log_late_inline_failure(cg, msg);
2200 }
2201 }
2202 break; // finish
2203 }
2204 }
2205
2206 igvn_worklist()->ensure_empty(); // should be done with igvn
2207
2208 if (_late_inlines.length() == 0) {
2209 break; // no more progress
2210 }
2211
2212 while (inline_incrementally_one()) {
2213 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2214 }
2215 if (failing()) return;
2216
2217 inline_incrementally_cleanup(igvn);
2218
2219 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2220
2221 if (failing()) return;
2222 }
2223
2224 igvn_worklist()->ensure_empty(); // should be done with igvn
2225
2226 if (_string_late_inlines.length() > 0) {
2227 assert(has_stringbuilder(), "inconsistent");
2228
2229 inline_string_calls(false);
2230
2231 if (failing()) return;
2232
2233 inline_incrementally_cleanup(igvn);
2234 }
2235
2236 set_inlining_incrementally(false);
2237 }
2238
2239 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2240 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2241 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2242 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2243 // as if "inlining_incrementally() == true" were set.
2244 assert(inlining_incrementally() == false, "not allowed");
2245 assert(_modified_nodes == nullptr, "not allowed");
2246 assert(_late_inlines.length() > 0, "sanity");
2247
2248 if (StressIncrementalInlining) {
2249 shuffle_late_inlines();
2250 }
2251
2252 while (_late_inlines.length() > 0) {
2253 igvn_worklist()->ensure_empty(); // should be done with igvn
2254
2255 while (inline_incrementally_one()) {
2256 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2257 }
2258 if (failing()) return;
2259
2260 inline_incrementally_cleanup(igvn);
2261 }
2262 }
2263
2264 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2265 if (_loop_opts_cnt > 0) {
2266 while (major_progress() && (_loop_opts_cnt > 0)) {
2267 TracePhase tp(_t_idealLoop);
2268 PhaseIdealLoop::optimize(igvn, mode);
2269 _loop_opts_cnt--;
2270 if (failing()) return false;
2271 if (major_progress()) {
2272 print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2273 }
2274 }
2275 }
2276 return true;
2277 }
2278
2279 // Remove edges from "root" to each SafePoint at a backward branch.
2280 // They were inserted during parsing (see add_safepoint()) to make
2281 // infinite loops without calls or exceptions visible to root, i.e.,
2282 // useful.
2283 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2284 Node *r = root();
2285 if (r != nullptr) {
2286 for (uint i = r->req(); i < r->len(); ++i) {
2287 Node *n = r->in(i);
2288 if (n != nullptr && n->is_SafePoint()) {
2289 r->rm_prec(i);
2290 if (n->outcnt() == 0) {
2291 igvn.remove_dead_node(n, PhaseIterGVN::NodeOrigin::Graph);
2292 }
2293 --i;
2294 }
2295 }
2296 // Parsing may have added top inputs to the root node (Path
2297 // leading to the Halt node proven dead). Make sure we get a
2298 // chance to clean them up.
2299 igvn._worklist.push(r);
2300 igvn.optimize();
2301 }
2302 }
2303
2304 //------------------------------Optimize---------------------------------------
2305 // Given a graph, optimize it.
2306 void Compile::Optimize() {
2307 TracePhase tp(_t_optimizer);
2308
2309 #ifndef PRODUCT
2310 if (env()->break_at_compile()) {
2311 BREAKPOINT;
2312 }
2313
2314 #endif
2315
2316 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2317 #ifdef ASSERT
2318 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2319 #endif
2320
2321 ResourceMark rm;
2322
2323 NOT_PRODUCT( verify_graph_edges(); )
2324
2325 print_method(PHASE_AFTER_PARSING, 1);
2326
2327 {
2328 // Iterative Global Value Numbering, including ideal transforms
2329 PhaseIterGVN igvn;
2330 #ifdef ASSERT
2331 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2332 #endif
2333 {
2334 TracePhase tp(_t_iterGVN);
2335 igvn.optimize(true);
2336 }
2337
2338 if (failing()) return;
2339
2340 print_method(PHASE_ITER_GVN1, 2);
2341
2342 process_for_unstable_if_traps(igvn);
2343
2344 if (failing()) return;
2345
2346 inline_incrementally(igvn);
2347
2348 print_method(PHASE_INCREMENTAL_INLINE, 2);
2349
2350 if (failing()) return;
2351
2352 if (eliminate_boxing()) {
2353 // Inline valueOf() methods now.
2354 inline_boxing_calls(igvn);
2355
2356 if (failing()) return;
2357
2358 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2359 inline_incrementally(igvn);
2360 }
2361
2362 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2363
2364 if (failing()) return;
2365 }
2366
2367 // Remove the speculative part of types and clean up the graph from
2368 // the extra CastPP nodes whose only purpose is to carry them. Do
2369 // that early so that optimizations are not disrupted by the extra
2370 // CastPP nodes.
2371 remove_speculative_types(igvn);
2372
2373 if (failing()) return;
2374
2375 // No more new expensive nodes will be added to the list from here
2376 // so keep only the actual candidates for optimizations.
2377 cleanup_expensive_nodes(igvn);
2378
2379 if (failing()) return;
2380
2381 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2382 if (EnableVectorSupport && has_vbox_nodes()) {
2383 TracePhase tp(_t_vector);
2384 PhaseVector pv(igvn);
2385 pv.optimize_vector_boxes();
2386 if (failing()) return;
2387 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2388 }
2389 assert(!has_vbox_nodes(), "sanity");
2390
2391 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2392 Compile::TracePhase tp(_t_renumberLive);
2393 igvn_worklist()->ensure_empty(); // should be done with igvn
2394 {
2395 ResourceMark rm;
2396 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2397 }
2398 igvn.reset();
2399 igvn.optimize(true);
2400 if (failing()) return;
2401 }
2402
2403 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2404 // safepoints
2405 remove_root_to_sfpts_edges(igvn);
2406
2407 if (failing()) return;
2408
2409 _print_phase_loop_opts = has_loops();
2410 if (_print_phase_loop_opts) {
2411 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2412 }
2413
2414 // Perform escape analysis
2415 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2416 if (has_loops()) {
2417 // Cleanup graph (remove dead nodes).
2418 TracePhase tp(_t_idealLoop);
2419 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2420 if (failing()) return;
2421 }
2422 bool progress;
2423 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2424 do {
2425 ConnectionGraph::do_analysis(this, &igvn);
2426
2427 if (failing()) return;
2428
2429 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2430
2431 // Optimize out fields loads from scalar replaceable allocations.
2432 igvn.optimize(true);
2433 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2434
2435 if (failing()) return;
2436
2437 if (congraph() != nullptr && macro_count() > 0) {
2438 TracePhase tp(_t_macroEliminate);
2439 PhaseMacroExpand mexp(igvn);
2440 mexp.eliminate_macro_nodes();
2441 if (failing()) return;
2442 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2443
2444 igvn.set_delay_transform(false);
2445 igvn.optimize();
2446 if (failing()) return;
2447
2448 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2449 }
2450
2451 ConnectionGraph::verify_ram_nodes(this, root());
2452 if (failing()) return;
2453
2454 progress = do_iterative_escape_analysis() &&
2455 (macro_count() < mcount) &&
2456 ConnectionGraph::has_candidates(this);
2457 // Try again if candidates exist and made progress
2458 // by removing some allocations and/or locks.
2459 } while (progress);
2460 }
2461
2462 // Loop transforms on the ideal graph. Range Check Elimination,
2463 // peeling, unrolling, etc.
2464
2465 // Set loop opts counter
2466 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2467 {
2468 TracePhase tp(_t_idealLoop);
2469 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2470 _loop_opts_cnt--;
2471 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2472 if (failing()) return;
2473 }
2474 // Loop opts pass if partial peeling occurred in previous pass
2475 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2476 TracePhase tp(_t_idealLoop);
2477 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2478 _loop_opts_cnt--;
2479 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2480 if (failing()) return;
2481 }
2482 // Loop opts pass for loop-unrolling before CCP
2483 if(major_progress() && (_loop_opts_cnt > 0)) {
2484 TracePhase tp(_t_idealLoop);
2485 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2486 _loop_opts_cnt--;
2487 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2488 }
2489 if (!failing()) {
2490 // Verify that last round of loop opts produced a valid graph
2491 PhaseIdealLoop::verify(igvn);
2492 }
2493 }
2494 if (failing()) return;
2495
2496 // Conditional Constant Propagation;
2497 print_method(PHASE_BEFORE_CCP1, 2);
2498 PhaseCCP ccp( &igvn );
2499 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2500 {
2501 TracePhase tp(_t_ccp);
2502 ccp.do_transform();
2503 }
2504 print_method(PHASE_CCP1, 2);
2505
2506 assert( true, "Break here to ccp.dump_old2new_map()");
2507
2508 // Iterative Global Value Numbering, including ideal transforms
2509 {
2510 TracePhase tp(_t_iterGVN2);
2511 igvn.reset_from_igvn(&ccp);
2512 igvn.optimize(true);
2513 }
2514 print_method(PHASE_ITER_GVN2, 2);
2515
2516 if (failing()) return;
2517
2518 // Loop transforms on the ideal graph. Range Check Elimination,
2519 // peeling, unrolling, etc.
2520 if (!optimize_loops(igvn, LoopOptsDefault)) {
2521 return;
2522 }
2523
2524 C->clear_major_progress(); // ensure that major progress is now clear
2525
2526 process_for_post_loop_opts_igvn(igvn);
2527
2528 if (failing()) return;
2529
2530 // Once loop optimizations are over, it is safe to get rid of all reachability fence nodes and
2531 // migrate reachability edges to safepoints.
2532 if (OptimizeReachabilityFences && _reachability_fences.length() > 0) {
2533 TracePhase tp1(_t_idealLoop);
2534 TracePhase tp2(_t_reachability);
2535 PhaseIdealLoop::optimize(igvn, PostLoopOptsExpandReachabilityFences);
2536 print_method(PHASE_EXPAND_REACHABILITY_FENCES, 2);
2537 if (failing()) return;
2538 assert(_reachability_fences.length() == 0 || PreserveReachabilityFencesOnConstants, "no RF nodes allowed");
2539 }
2540
2541 process_for_merge_stores_igvn(igvn);
2542
2543 if (failing()) return;
2544
2545 #ifdef ASSERT
2546 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2547 #endif
2548
2549 {
2550 TracePhase tp(_t_macroExpand);
2551 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2552 PhaseMacroExpand mex(igvn);
2553 // Do not allow new macro nodes once we start to eliminate and expand
2554 C->reset_allow_macro_nodes();
2555 // Last attempt to eliminate macro nodes before expand
2556 mex.eliminate_macro_nodes();
2557 if (failing()) {
2558 return;
2559 }
2560 mex.eliminate_opaque_looplimit_macro_nodes();
2561 if (failing()) {
2562 return;
2563 }
2564 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2565 if (mex.expand_macro_nodes()) {
2566 assert(failing(), "must bail out w/ explicit message");
2567 return;
2568 }
2569 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2570 }
2571
2572 {
2573 TracePhase tp(_t_barrierExpand);
2574 if (bs->expand_barriers(this, igvn)) {
2575 assert(failing(), "must bail out w/ explicit message");
2576 return;
2577 }
2578 print_method(PHASE_BARRIER_EXPANSION, 2);
2579 }
2580
2581 if (C->max_vector_size() > 0) {
2582 C->optimize_logic_cones(igvn);
2583 igvn.optimize();
2584 if (failing()) return;
2585 }
2586
2587 DEBUG_ONLY( _modified_nodes = nullptr; )
2588
2589 assert(igvn._worklist.size() == 0, "not empty");
2590
2591 if (_late_inlines.length() > 0) {
2592 // More opportunities to optimize virtual and MH calls.
2593 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2594 process_late_inline_calls_no_inline(igvn);
2595 if (failing()) return;
2596 }
2597 assert(_late_inlines.length() == 0, "late inline queue must be drained");
2598 } // (End scope of igvn; run destructor if necessary for asserts.)
2599
2600 check_no_dead_use();
2601
2602 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2603 // to remove hashes to unlock nodes for modifications.
2604 C->node_hash()->clear();
2605
2606 // A method with only infinite loops has no edges entering loops from root
2607 {
2608 TracePhase tp(_t_graphReshaping);
2609 if (final_graph_reshaping()) {
2610 assert(failing(), "must bail out w/ explicit message");
2611 return;
2612 }
2613 }
2614
2615 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2616 DEBUG_ONLY(set_phase_optimize_finished();)
2617 }
2618
2619 #ifdef ASSERT
2620 void Compile::check_no_dead_use() const {
2621 ResourceMark rm;
2622 Unique_Node_List wq;
2623 wq.push(root());
2624 for (uint i = 0; i < wq.size(); ++i) {
2625 Node* n = wq.at(i);
2626 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2627 Node* u = n->fast_out(j);
2628 if (u->outcnt() == 0 && !u->is_Con()) {
2629 u->dump();
2630 fatal("no reachable node should have no use");
2631 }
2632 wq.push(u);
2633 }
2634 }
2635 }
2636 #endif
2637
2638 void Compile::inline_vector_reboxing_calls() {
2639 if (C->_vector_reboxing_late_inlines.length() > 0) {
2640 _late_inlines_pos = C->_late_inlines.length();
2641 while (_vector_reboxing_late_inlines.length() > 0) {
2642 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2643 cg->do_late_inline();
2644 if (failing()) return;
2645 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2646 }
2647 _vector_reboxing_late_inlines.trunc_to(0);
2648 }
2649 }
2650
2651 bool Compile::has_vbox_nodes() {
2652 if (C->_vector_reboxing_late_inlines.length() > 0) {
2653 return true;
2654 }
2655 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2656 Node * n = C->macro_node(macro_idx);
2657 assert(n->is_macro(), "only macro nodes expected here");
2658 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2659 return true;
2660 }
2661 }
2662 return false;
2663 }
2664
2665 //---------------------------- Bitwise operation packing optimization ---------------------------
2666
2667 static bool is_vector_unary_bitwise_op(Node* n) {
2668 return n->Opcode() == Op_XorV &&
2669 VectorNode::is_vector_bitwise_not_pattern(n);
2670 }
2671
2672 static bool is_vector_binary_bitwise_op(Node* n) {
2673 switch (n->Opcode()) {
2674 case Op_AndV:
2675 case Op_OrV:
2676 return true;
2677
2678 case Op_XorV:
2679 return !is_vector_unary_bitwise_op(n);
2680
2681 default:
2682 return false;
2683 }
2684 }
2685
2686 static bool is_vector_ternary_bitwise_op(Node* n) {
2687 return n->Opcode() == Op_MacroLogicV;
2688 }
2689
2690 static bool is_vector_bitwise_op(Node* n) {
2691 return is_vector_unary_bitwise_op(n) ||
2692 is_vector_binary_bitwise_op(n) ||
2693 is_vector_ternary_bitwise_op(n);
2694 }
2695
2696 static bool is_vector_bitwise_cone_root(Node* n) {
2697 if (n->bottom_type()->isa_pvectmask() || !is_vector_bitwise_op(n)) {
2698 return false;
2699 }
2700 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2701 if (is_vector_bitwise_op(n->fast_out(i))) {
2702 return false;
2703 }
2704 }
2705 return true;
2706 }
2707
2708 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2709 uint cnt = 0;
2710 if (is_vector_bitwise_op(n)) {
2711 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2712 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2713 for (uint i = 1; i < inp_cnt; i++) {
2714 Node* in = n->in(i);
2715 bool skip = VectorNode::is_all_ones_vector(in);
2716 if (!skip && !inputs.member(in)) {
2717 inputs.push(in);
2718 cnt++;
2719 }
2720 }
2721 assert(cnt <= 1, "not unary");
2722 } else {
2723 uint last_req = inp_cnt;
2724 if (is_vector_ternary_bitwise_op(n)) {
2725 last_req = inp_cnt - 1; // skip last input
2726 }
2727 for (uint i = 1; i < last_req; i++) {
2728 Node* def = n->in(i);
2729 if (!inputs.member(def)) {
2730 inputs.push(def);
2731 cnt++;
2732 }
2733 }
2734 }
2735 } else { // not a bitwise operations
2736 if (!inputs.member(n)) {
2737 inputs.push(n);
2738 cnt++;
2739 }
2740 }
2741 return cnt;
2742 }
2743
2744 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2745 Unique_Node_List useful_nodes;
2746 C->identify_useful_nodes(useful_nodes);
2747
2748 for (uint i = 0; i < useful_nodes.size(); i++) {
2749 Node* n = useful_nodes.at(i);
2750 if (is_vector_bitwise_cone_root(n)) {
2751 list.push(n);
2752 }
2753 }
2754 }
2755
2756 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2757 const TypeVect* vt,
2758 Unique_Node_List& partition,
2759 Unique_Node_List& inputs) {
2760 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2761 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2762 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2763
2764 Node* in1 = inputs.at(0);
2765 Node* in2 = inputs.at(1);
2766 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2767
2768 uint func = compute_truth_table(partition, inputs);
2769
2770 Node* pn = partition.at(partition.size() - 1);
2771 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2772 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2773 }
2774
2775 static uint extract_bit(uint func, uint pos) {
2776 return (func & (1 << pos)) >> pos;
2777 }
2778
2779 //
2780 // A macro logic node represents a truth table. It has 4 inputs,
2781 // First three inputs corresponds to 3 columns of a truth table
2782 // and fourth input captures the logic function.
2783 //
2784 // eg. fn = (in1 AND in2) OR in3;
2785 //
2786 // MacroNode(in1,in2,in3,fn)
2787 //
2788 // -----------------
2789 // in1 in2 in3 fn
2790 // -----------------
2791 // 0 0 0 0
2792 // 0 0 1 1
2793 // 0 1 0 0
2794 // 0 1 1 1
2795 // 1 0 0 0
2796 // 1 0 1 1
2797 // 1 1 0 1
2798 // 1 1 1 1
2799 //
2800
2801 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2802 int res = 0;
2803 for (int i = 0; i < 8; i++) {
2804 int bit1 = extract_bit(in1, i);
2805 int bit2 = extract_bit(in2, i);
2806 int bit3 = extract_bit(in3, i);
2807
2808 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2809 int func_bit = extract_bit(func, func_bit_pos);
2810
2811 res |= func_bit << i;
2812 }
2813 return res;
2814 }
2815
2816 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
2817 assert(n != nullptr, "");
2818 assert(eval_map.contains(n), "absent");
2819 return *(eval_map.get(n));
2820 }
2821
2822 static void eval_operands(Node* n,
2823 uint& func1, uint& func2, uint& func3,
2824 HashTable<Node*,uint>& eval_map) {
2825 assert(is_vector_bitwise_op(n), "");
2826
2827 if (is_vector_unary_bitwise_op(n)) {
2828 Node* opnd = n->in(1);
2829 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2830 opnd = n->in(2);
2831 }
2832 func1 = eval_operand(opnd, eval_map);
2833 } else if (is_vector_binary_bitwise_op(n)) {
2834 func1 = eval_operand(n->in(1), eval_map);
2835 func2 = eval_operand(n->in(2), eval_map);
2836 } else {
2837 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2838 func1 = eval_operand(n->in(1), eval_map);
2839 func2 = eval_operand(n->in(2), eval_map);
2840 func3 = eval_operand(n->in(3), eval_map);
2841 }
2842 }
2843
2844 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2845 assert(inputs.size() <= 3, "sanity");
2846 ResourceMark rm;
2847 uint res = 0;
2848 HashTable<Node*,uint> eval_map;
2849
2850 // Populate precomputed functions for inputs.
2851 // Each input corresponds to one column of 3 input truth-table.
2852 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2853 0xCC, // (_, b, _) -> b
2854 0xF0 }; // (a, _, _) -> a
2855 for (uint i = 0; i < inputs.size(); i++) {
2856 eval_map.put(inputs.at(i), input_funcs[2-i]);
2857 }
2858
2859 for (uint i = 0; i < partition.size(); i++) {
2860 Node* n = partition.at(i);
2861
2862 uint func1 = 0, func2 = 0, func3 = 0;
2863 eval_operands(n, func1, func2, func3, eval_map);
2864
2865 switch (n->Opcode()) {
2866 case Op_OrV:
2867 assert(func3 == 0, "not binary");
2868 res = func1 | func2;
2869 break;
2870 case Op_AndV:
2871 assert(func3 == 0, "not binary");
2872 res = func1 & func2;
2873 break;
2874 case Op_XorV:
2875 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2876 assert(func2 == 0 && func3 == 0, "not unary");
2877 res = (~func1) & 0xFF;
2878 } else {
2879 assert(func3 == 0, "not binary");
2880 res = func1 ^ func2;
2881 }
2882 break;
2883 case Op_MacroLogicV:
2884 // Ordering of inputs may change during evaluation of sub-tree
2885 // containing MacroLogic node as a child node, thus a re-evaluation
2886 // makes sure that function is evaluated in context of current
2887 // inputs.
2888 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2889 break;
2890
2891 default: assert(false, "not supported: %s", n->Name());
2892 }
2893 assert(res <= 0xFF, "invalid");
2894 eval_map.put(n, res);
2895 }
2896 return res;
2897 }
2898
2899 // Criteria under which nodes gets packed into a macro logic node:-
2900 // 1) Parent and both child nodes are all unmasked or masked with
2901 // same predicates.
2902 // 2) Masked parent can be packed with left child if it is predicated
2903 // and both have same predicates.
2904 // 3) Masked parent can be packed with right child if its un-predicated
2905 // or has matching predication condition.
2906 // 4) An unmasked parent can be packed with an unmasked child.
2907 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2908 assert(partition.size() == 0, "not empty");
2909 assert(inputs.size() == 0, "not empty");
2910 if (is_vector_ternary_bitwise_op(n)) {
2911 return false;
2912 }
2913
2914 bool is_unary_op = is_vector_unary_bitwise_op(n);
2915 if (is_unary_op) {
2916 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2917 return false; // too few inputs
2918 }
2919
2920 bool pack_left_child = true;
2921 bool pack_right_child = true;
2922
2923 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2924 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2925
2926 int left_child_input_cnt = 0;
2927 int right_child_input_cnt = 0;
2928
2929 bool parent_is_predicated = n->is_predicated_vector();
2930 bool left_child_predicated = n->in(1)->is_predicated_vector();
2931 bool right_child_predicated = n->in(2)->is_predicated_vector();
2932
2933 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2934 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2935 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2936
2937 do {
2938 if (pack_left_child && left_child_LOP &&
2939 ((!parent_is_predicated && !left_child_predicated) ||
2940 ((parent_is_predicated && left_child_predicated &&
2941 parent_pred == left_child_pred)))) {
2942 partition.push(n->in(1));
2943 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2944 } else {
2945 inputs.push(n->in(1));
2946 left_child_input_cnt = 1;
2947 }
2948
2949 if (pack_right_child && right_child_LOP &&
2950 (!right_child_predicated ||
2951 (right_child_predicated && parent_is_predicated &&
2952 parent_pred == right_child_pred))) {
2953 partition.push(n->in(2));
2954 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2955 } else {
2956 inputs.push(n->in(2));
2957 right_child_input_cnt = 1;
2958 }
2959
2960 if (inputs.size() > 3) {
2961 assert(partition.size() > 0, "");
2962 inputs.clear();
2963 partition.clear();
2964 if (left_child_input_cnt > right_child_input_cnt) {
2965 pack_left_child = false;
2966 } else {
2967 pack_right_child = false;
2968 }
2969 } else {
2970 break;
2971 }
2972 } while(true);
2973
2974 if(partition.size()) {
2975 partition.push(n);
2976 }
2977
2978 return (partition.size() == 2 || partition.size() == 3) &&
2979 (inputs.size() == 2 || inputs.size() == 3);
2980 }
2981
2982 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2983 assert(is_vector_bitwise_op(n), "not a root");
2984
2985 visited.set(n->_idx);
2986
2987 // 1) Do a DFS walk over the logic cone.
2988 for (uint i = 1; i < n->req(); i++) {
2989 Node* in = n->in(i);
2990 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2991 process_logic_cone_root(igvn, in, visited);
2992 }
2993 }
2994
2995 // 2) Bottom up traversal: Merge node[s] with
2996 // the parent to form macro logic node.
2997 Unique_Node_List partition;
2998 Unique_Node_List inputs;
2999 if (compute_logic_cone(n, partition, inputs)) {
3000 const TypeVect* vt = n->bottom_type()->is_vect();
3001 Node* pn = partition.at(partition.size() - 1);
3002 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3003 if (mask == nullptr ||
3004 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3005 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3006 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3007 igvn.replace_node(n, macro_logic);
3008 }
3009 }
3010 }
3011
3012 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3013 ResourceMark rm;
3014 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3015 Unique_Node_List list;
3016 collect_logic_cone_roots(list);
3017
3018 while (list.size() > 0) {
3019 Node* n = list.pop();
3020 const TypeVect* vt = n->bottom_type()->is_vect();
3021 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3022 if (supported) {
3023 VectorSet visited(comp_arena());
3024 process_logic_cone_root(igvn, n, visited);
3025 }
3026 }
3027 }
3028 }
3029
3030 //------------------------------Code_Gen---------------------------------------
3031 // Given a graph, generate code for it
3032 void Compile::Code_Gen() {
3033 if (failing()) {
3034 return;
3035 }
3036
3037 // Perform instruction selection. You might think we could reclaim Matcher
3038 // memory PDQ, but actually the Matcher is used in generating spill code.
3039 // Internals of the Matcher (including some VectorSets) must remain live
3040 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3041 // set a bit in reclaimed memory.
3042
3043 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3044 // nodes. Mapping is only valid at the root of each matched subtree.
3045 NOT_PRODUCT( verify_graph_edges(); )
3046
3047 Matcher matcher;
3048 _matcher = &matcher;
3049 {
3050 TracePhase tp(_t_matcher);
3051 matcher.match();
3052 if (failing()) {
3053 return;
3054 }
3055 }
3056 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3057 // nodes. Mapping is only valid at the root of each matched subtree.
3058 NOT_PRODUCT( verify_graph_edges(); )
3059
3060 // If you have too many nodes, or if matching has failed, bail out
3061 check_node_count(0, "out of nodes matching instructions");
3062 if (failing()) {
3063 return;
3064 }
3065
3066 print_method(PHASE_MATCHING, 2);
3067
3068 // Build a proper-looking CFG
3069 PhaseCFG cfg(node_arena(), root(), matcher);
3070 if (failing()) {
3071 return;
3072 }
3073 _cfg = &cfg;
3074 {
3075 TracePhase tp(_t_scheduler);
3076 bool success = cfg.do_global_code_motion();
3077 if (!success) {
3078 return;
3079 }
3080
3081 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3082 NOT_PRODUCT( verify_graph_edges(); )
3083 cfg.verify();
3084 if (failing()) {
3085 return;
3086 }
3087 }
3088
3089 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3090 _regalloc = ®alloc;
3091 {
3092 TracePhase tp(_t_registerAllocation);
3093 // Perform register allocation. After Chaitin, use-def chains are
3094 // no longer accurate (at spill code) and so must be ignored.
3095 // Node->LRG->reg mappings are still accurate.
3096 _regalloc->Register_Allocate();
3097
3098 // Bail out if the allocator builds too many nodes
3099 if (failing()) {
3100 return;
3101 }
3102
3103 print_method(PHASE_REGISTER_ALLOCATION, 2);
3104 }
3105
3106 // Prior to register allocation we kept empty basic blocks in case the
3107 // the allocator needed a place to spill. After register allocation we
3108 // are not adding any new instructions. If any basic block is empty, we
3109 // can now safely remove it.
3110 {
3111 TracePhase tp(_t_blockOrdering);
3112 cfg.remove_empty_blocks();
3113 if (do_freq_based_layout()) {
3114 PhaseBlockLayout layout(cfg);
3115 } else {
3116 cfg.set_loop_alignment();
3117 }
3118 cfg.fixup_flow();
3119 cfg.remove_unreachable_blocks();
3120 cfg.verify_dominator_tree();
3121 print_method(PHASE_BLOCK_ORDERING, 3);
3122 }
3123
3124 // Apply peephole optimizations
3125 if( OptoPeephole ) {
3126 TracePhase tp(_t_peephole);
3127 PhasePeephole peep( _regalloc, cfg);
3128 peep.do_transform();
3129 print_method(PHASE_PEEPHOLE, 3);
3130 }
3131
3132 // Do late expand if CPU requires this.
3133 if (Matcher::require_postalloc_expand) {
3134 TracePhase tp(_t_postalloc_expand);
3135 cfg.postalloc_expand(_regalloc);
3136 print_method(PHASE_POSTALLOC_EXPAND, 3);
3137 }
3138
3139 #ifdef ASSERT
3140 {
3141 CompilationMemoryStatistic::do_test_allocations();
3142 if (failing()) return;
3143 }
3144 #endif
3145
3146 // Convert Nodes to instruction bits in a buffer
3147 {
3148 TracePhase tp(_t_output);
3149 PhaseOutput output;
3150 output.Output();
3151 if (failing()) return;
3152 output.install();
3153 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3154 }
3155
3156 // He's dead, Jim.
3157 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3158 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3159 }
3160
3161 //------------------------------Final_Reshape_Counts---------------------------
3162 // This class defines counters to help identify when a method
3163 // may/must be executed using hardware with only 24-bit precision.
3164 struct Final_Reshape_Counts : public StackObj {
3165 int _call_count; // count non-inlined 'common' calls
3166 int _float_count; // count float ops requiring 24-bit precision
3167 int _double_count; // count double ops requiring more precision
3168 int _java_call_count; // count non-inlined 'java' calls
3169 int _inner_loop_count; // count loops which need alignment
3170 VectorSet _visited; // Visitation flags
3171 Node_List _tests; // Set of IfNodes & PCTableNodes
3172
3173 Final_Reshape_Counts() :
3174 _call_count(0), _float_count(0), _double_count(0),
3175 _java_call_count(0), _inner_loop_count(0) { }
3176
3177 void inc_call_count () { _call_count ++; }
3178 void inc_float_count () { _float_count ++; }
3179 void inc_double_count() { _double_count++; }
3180 void inc_java_call_count() { _java_call_count++; }
3181 void inc_inner_loop_count() { _inner_loop_count++; }
3182
3183 int get_call_count () const { return _call_count ; }
3184 int get_float_count () const { return _float_count ; }
3185 int get_double_count() const { return _double_count; }
3186 int get_java_call_count() const { return _java_call_count; }
3187 int get_inner_loop_count() const { return _inner_loop_count; }
3188 };
3189
3190 //------------------------------final_graph_reshaping_impl----------------------
3191 // Implement items 1-5 from final_graph_reshaping below.
3192 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3193
3194 if ( n->outcnt() == 0 ) return; // dead node
3195 uint nop = n->Opcode();
3196
3197 // Check for 2-input instruction with "last use" on right input.
3198 // Swap to left input. Implements item (2).
3199 if( n->req() == 3 && // two-input instruction
3200 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3201 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3202 n->in(2)->outcnt() == 1 &&// right use IS a last use
3203 !n->in(2)->is_Con() ) { // right use is not a constant
3204 // Check for commutative opcode
3205 switch( nop ) {
3206 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddHF: case Op_AddL:
3207 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3208 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3209 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulHF: case Op_MulL:
3210 case Op_AndL: case Op_XorL: case Op_OrL:
3211 case Op_AndI: case Op_XorI: case Op_OrI: {
3212 // Move "last use" input to left by swapping inputs
3213 n->swap_edges(1, 2);
3214 break;
3215 }
3216 default:
3217 break;
3218 }
3219 }
3220
3221 #ifdef ASSERT
3222 if( n->is_Mem() ) {
3223 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3224 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3225 // oop will be recorded in oop map if load crosses safepoint
3226 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3227 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3228 "raw memory operations should have control edge");
3229 }
3230 if (n->is_MemBar()) {
3231 MemBarNode* mb = n->as_MemBar();
3232 if (mb->trailing_store() || mb->trailing_load_store()) {
3233 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3234 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3235 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3236 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3237 } else if (mb->leading()) {
3238 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3239 }
3240 }
3241 if (n->is_CallLeafPure()) {
3242 // A pure call whose result projection is unused should have been
3243 // eliminated by CallLeafPureNode::Ideal during IGVN.
3244 assert(n->as_CallLeafPure()->proj_out_or_null(TypeFunc::Parms) != nullptr,
3245 "unused CallLeafPureNode should have been removed before final graph reshaping");
3246 }
3247 #endif
3248 // Count FPU ops and common calls, implements item (3)
3249 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3250 if (!gc_handled) {
3251 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3252 }
3253
3254 // Collect CFG split points
3255 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3256 frc._tests.push(n);
3257 }
3258 }
3259
3260 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3261 if (!UseDivMod) {
3262 return;
3263 }
3264
3265 // Check if "a % b" and "a / b" both exist
3266 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3267 if (d == nullptr) {
3268 return;
3269 }
3270
3271 // Replace them with a fused divmod if supported
3272 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3273 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3274 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3275 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3276 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3277 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3278 // DivMod node so the dependency is not lost.
3279 divmod->add_prec_from(n);
3280 divmod->add_prec_from(d);
3281 d->subsume_by(divmod->div_proj(), this);
3282 n->subsume_by(divmod->mod_proj(), this);
3283 } else {
3284 // Replace "a % b" with "a - ((a / b) * b)"
3285 Node* mult = MulNode::make(d, d->in(2), bt);
3286 Node* sub = SubNode::make(d->in(1), mult, bt);
3287 n->subsume_by(sub, this);
3288 }
3289 }
3290
3291 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3292 switch( nop ) {
3293 // Count all float operations that may use FPU
3294 case Op_AddHF:
3295 case Op_MulHF:
3296 case Op_AddF:
3297 case Op_SubF:
3298 case Op_MulF:
3299 case Op_DivF:
3300 case Op_NegF:
3301 case Op_ModF:
3302 case Op_ConvI2F:
3303 case Op_ConF:
3304 case Op_CmpF:
3305 case Op_CmpF3:
3306 case Op_StoreF:
3307 case Op_LoadF:
3308 // case Op_ConvL2F: // longs are split into 32-bit halves
3309 frc.inc_float_count();
3310 break;
3311
3312 case Op_ConvF2D:
3313 case Op_ConvD2F:
3314 frc.inc_float_count();
3315 frc.inc_double_count();
3316 break;
3317
3318 // Count all double operations that may use FPU
3319 case Op_AddD:
3320 case Op_SubD:
3321 case Op_MulD:
3322 case Op_DivD:
3323 case Op_NegD:
3324 case Op_ModD:
3325 case Op_ConvI2D:
3326 case Op_ConvD2I:
3327 // case Op_ConvL2D: // handled by leaf call
3328 // case Op_ConvD2L: // handled by leaf call
3329 case Op_ConD:
3330 case Op_CmpD:
3331 case Op_CmpD3:
3332 case Op_StoreD:
3333 case Op_LoadD:
3334 case Op_LoadD_unaligned:
3335 frc.inc_double_count();
3336 break;
3337 case Op_Opaque1: // Remove Opaque Nodes before matching
3338 n->subsume_by(n->in(1), this);
3339 break;
3340 case Op_CallLeafPure: {
3341 // If the pure call is not supported, then lower to a CallLeaf.
3342 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3343 CallNode* call = n->as_Call();
3344 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3345 call->_name, TypeRawPtr::BOTTOM);
3346 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3347 new_call->init_req(TypeFunc::I_O, C->top());
3348 new_call->init_req(TypeFunc::Memory, C->top());
3349 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3350 new_call->init_req(TypeFunc::FramePtr, C->top());
3351 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
3352 new_call->init_req(i, call->in(i));
3353 }
3354 n->subsume_by(new_call, this);
3355 }
3356 frc.inc_call_count();
3357 break;
3358 }
3359 case Op_CallStaticJava:
3360 case Op_CallJava:
3361 case Op_CallDynamicJava:
3362 frc.inc_java_call_count(); // Count java call site;
3363 case Op_CallRuntime:
3364 case Op_CallLeaf:
3365 case Op_CallLeafVector:
3366 case Op_CallLeafNoFP: {
3367 assert (n->is_Call(), "");
3368 CallNode *call = n->as_Call();
3369 // Count call sites where the FP mode bit would have to be flipped.
3370 // Do not count uncommon runtime calls:
3371 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3372 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3373 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3374 frc.inc_call_count(); // Count the call site
3375 } else { // See if uncommon argument is shared
3376 Node *n = call->in(TypeFunc::Parms);
3377 int nop = n->Opcode();
3378 // Clone shared simple arguments to uncommon calls, item (1).
3379 if (n->outcnt() > 1 &&
3380 !n->is_Proj() &&
3381 nop != Op_CreateEx &&
3382 nop != Op_CheckCastPP &&
3383 nop != Op_DecodeN &&
3384 nop != Op_DecodeNKlass &&
3385 !n->is_Mem() &&
3386 !n->is_Phi()) {
3387 Node *x = n->clone();
3388 call->set_req(TypeFunc::Parms, x);
3389 }
3390 }
3391 break;
3392 }
3393 case Op_StoreB:
3394 case Op_StoreC:
3395 case Op_StoreI:
3396 case Op_StoreL:
3397 case Op_CompareAndSwapB:
3398 case Op_CompareAndSwapS:
3399 case Op_CompareAndSwapI:
3400 case Op_CompareAndSwapL:
3401 case Op_CompareAndSwapP:
3402 case Op_CompareAndSwapN:
3403 case Op_WeakCompareAndSwapB:
3404 case Op_WeakCompareAndSwapS:
3405 case Op_WeakCompareAndSwapI:
3406 case Op_WeakCompareAndSwapL:
3407 case Op_WeakCompareAndSwapP:
3408 case Op_WeakCompareAndSwapN:
3409 case Op_CompareAndExchangeB:
3410 case Op_CompareAndExchangeS:
3411 case Op_CompareAndExchangeI:
3412 case Op_CompareAndExchangeL:
3413 case Op_CompareAndExchangeP:
3414 case Op_CompareAndExchangeN:
3415 case Op_GetAndAddS:
3416 case Op_GetAndAddB:
3417 case Op_GetAndAddI:
3418 case Op_GetAndAddL:
3419 case Op_GetAndSetS:
3420 case Op_GetAndSetB:
3421 case Op_GetAndSetI:
3422 case Op_GetAndSetL:
3423 case Op_GetAndSetP:
3424 case Op_GetAndSetN:
3425 case Op_StoreP:
3426 case Op_StoreN:
3427 case Op_StoreNKlass:
3428 case Op_LoadB:
3429 case Op_LoadUB:
3430 case Op_LoadUS:
3431 case Op_LoadI:
3432 case Op_LoadKlass:
3433 case Op_LoadNKlass:
3434 case Op_LoadL:
3435 case Op_LoadL_unaligned:
3436 case Op_LoadP:
3437 case Op_LoadN:
3438 case Op_LoadRange:
3439 case Op_LoadS:
3440 break;
3441
3442 case Op_AddP: { // Assert sane base pointers
3443 Node *addp = n->in(AddPNode::Address);
3444 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
3445 #ifdef _LP64
3446 if (addp->Opcode() == Op_ConP &&
3447 addp == n->in(AddPNode::Base) &&
3448 n->in(AddPNode::Offset)->is_Con()) {
3449 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3450 // on the platform and on the compressed oops mode.
3451 // Use addressing with narrow klass to load with offset on x86.
3452 // Some platforms can use the constant pool to load ConP.
3453 // Do this transformation here since IGVN will convert ConN back to ConP.
3454 const Type* t = addp->bottom_type();
3455 bool is_oop = t->isa_oopptr() != nullptr;
3456 bool is_klass = t->isa_klassptr() != nullptr;
3457
3458 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3459 (is_klass && Matcher::const_klass_prefer_decode() &&
3460 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3461 Node* nn = nullptr;
3462
3463 int op = is_oop ? Op_ConN : Op_ConNKlass;
3464
3465 // Look for existing ConN node of the same exact type.
3466 Node* r = root();
3467 uint cnt = r->outcnt();
3468 for (uint i = 0; i < cnt; i++) {
3469 Node* m = r->raw_out(i);
3470 if (m!= nullptr && m->Opcode() == op &&
3471 m->bottom_type()->make_ptr() == t) {
3472 nn = m;
3473 break;
3474 }
3475 }
3476 if (nn != nullptr) {
3477 // Decode a narrow oop to match address
3478 // [R12 + narrow_oop_reg<<3 + offset]
3479 if (is_oop) {
3480 nn = new DecodeNNode(nn, t);
3481 } else {
3482 nn = new DecodeNKlassNode(nn, t);
3483 }
3484 // Check for succeeding AddP which uses the same Base.
3485 // Otherwise we will run into the assertion above when visiting that guy.
3486 for (uint i = 0; i < n->outcnt(); ++i) {
3487 Node *out_i = n->raw_out(i);
3488 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3489 out_i->set_req(AddPNode::Base, nn);
3490 #ifdef ASSERT
3491 for (uint j = 0; j < out_i->outcnt(); ++j) {
3492 Node *out_j = out_i->raw_out(j);
3493 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3494 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3495 }
3496 #endif
3497 }
3498 }
3499 n->set_req(AddPNode::Base, nn);
3500 n->set_req(AddPNode::Address, nn);
3501 if (addp->outcnt() == 0) {
3502 addp->disconnect_inputs(this);
3503 }
3504 }
3505 }
3506 }
3507 #endif
3508 break;
3509 }
3510
3511 case Op_CastPP: {
3512 // Remove CastPP nodes to gain more freedom during scheduling but
3513 // keep the dependency they encode as control or precedence edges
3514 // (if control is set already) on memory operations. Some CastPP
3515 // nodes don't have a control (don't carry a dependency): skip
3516 // those.
3517 if (n->in(0) != nullptr) {
3518 ResourceMark rm;
3519 Unique_Node_List wq;
3520 wq.push(n);
3521 for (uint next = 0; next < wq.size(); ++next) {
3522 Node *m = wq.at(next);
3523 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3524 Node* use = m->fast_out(i);
3525 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3526 use->ensure_control_or_add_prec(n->in(0));
3527 } else {
3528 switch(use->Opcode()) {
3529 case Op_AddP:
3530 case Op_DecodeN:
3531 case Op_DecodeNKlass:
3532 case Op_CheckCastPP:
3533 case Op_CastPP:
3534 wq.push(use);
3535 break;
3536 }
3537 }
3538 }
3539 }
3540 }
3541 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3542 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3543 Node* in1 = n->in(1);
3544 const Type* t = n->bottom_type();
3545 Node* new_in1 = in1->clone();
3546 new_in1->as_DecodeN()->set_type(t);
3547
3548 if (!Matcher::narrow_oop_use_complex_address()) {
3549 //
3550 // x86, ARM and friends can handle 2 adds in addressing mode
3551 // and Matcher can fold a DecodeN node into address by using
3552 // a narrow oop directly and do implicit null check in address:
3553 //
3554 // [R12 + narrow_oop_reg<<3 + offset]
3555 // NullCheck narrow_oop_reg
3556 //
3557 // On other platforms (Sparc) we have to keep new DecodeN node and
3558 // use it to do implicit null check in address:
3559 //
3560 // decode_not_null narrow_oop_reg, base_reg
3561 // [base_reg + offset]
3562 // NullCheck base_reg
3563 //
3564 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3565 // to keep the information to which null check the new DecodeN node
3566 // corresponds to use it as value in implicit_null_check().
3567 //
3568 new_in1->set_req(0, n->in(0));
3569 }
3570
3571 n->subsume_by(new_in1, this);
3572 if (in1->outcnt() == 0) {
3573 in1->disconnect_inputs(this);
3574 }
3575 } else {
3576 n->subsume_by(n->in(1), this);
3577 if (n->outcnt() == 0) {
3578 n->disconnect_inputs(this);
3579 }
3580 }
3581 break;
3582 }
3583 case Op_CastII: {
3584 n->as_CastII()->remove_range_check_cast(this);
3585 break;
3586 }
3587 #ifdef _LP64
3588 case Op_CmpP:
3589 // Do this transformation here to preserve CmpPNode::sub() and
3590 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3591 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3592 Node* in1 = n->in(1);
3593 Node* in2 = n->in(2);
3594 if (!in1->is_DecodeNarrowPtr()) {
3595 in2 = in1;
3596 in1 = n->in(2);
3597 }
3598 assert(in1->is_DecodeNarrowPtr(), "sanity");
3599
3600 Node* new_in2 = nullptr;
3601 if (in2->is_DecodeNarrowPtr()) {
3602 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3603 new_in2 = in2->in(1);
3604 } else if (in2->Opcode() == Op_ConP) {
3605 const Type* t = in2->bottom_type();
3606 if (t == TypePtr::NULL_PTR) {
3607 assert(in1->is_DecodeN(), "compare klass to null?");
3608 // Don't convert CmpP null check into CmpN if compressed
3609 // oops implicit null check is not generated.
3610 // This will allow to generate normal oop implicit null check.
3611 if (Matcher::gen_narrow_oop_implicit_null_checks())
3612 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3613 //
3614 // This transformation together with CastPP transformation above
3615 // will generated code for implicit null checks for compressed oops.
3616 //
3617 // The original code after Optimize()
3618 //
3619 // LoadN memory, narrow_oop_reg
3620 // decode narrow_oop_reg, base_reg
3621 // CmpP base_reg, nullptr
3622 // CastPP base_reg // NotNull
3623 // Load [base_reg + offset], val_reg
3624 //
3625 // after these transformations will be
3626 //
3627 // LoadN memory, narrow_oop_reg
3628 // CmpN narrow_oop_reg, nullptr
3629 // decode_not_null narrow_oop_reg, base_reg
3630 // Load [base_reg + offset], val_reg
3631 //
3632 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3633 // since narrow oops can be used in debug info now (see the code in
3634 // final_graph_reshaping_walk()).
3635 //
3636 // At the end the code will be matched to
3637 // on x86:
3638 //
3639 // Load_narrow_oop memory, narrow_oop_reg
3640 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3641 // NullCheck narrow_oop_reg
3642 //
3643 // and on sparc:
3644 //
3645 // Load_narrow_oop memory, narrow_oop_reg
3646 // decode_not_null narrow_oop_reg, base_reg
3647 // Load [base_reg + offset], val_reg
3648 // NullCheck base_reg
3649 //
3650 } else if (t->isa_oopptr()) {
3651 new_in2 = ConNode::make(t->make_narrowoop());
3652 } else if (t->isa_klassptr()) {
3653 ciKlass* klass = t->is_klassptr()->exact_klass();
3654 if (klass->is_in_encoding_range()) {
3655 new_in2 = ConNode::make(t->make_narrowklass());
3656 }
3657 }
3658 }
3659 if (new_in2 != nullptr) {
3660 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3661 n->subsume_by(cmpN, this);
3662 if (in1->outcnt() == 0) {
3663 in1->disconnect_inputs(this);
3664 }
3665 if (in2->outcnt() == 0) {
3666 in2->disconnect_inputs(this);
3667 }
3668 }
3669 }
3670 break;
3671
3672 case Op_DecodeN:
3673 case Op_DecodeNKlass:
3674 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3675 // DecodeN could be pinned when it can't be fold into
3676 // an address expression, see the code for Op_CastPP above.
3677 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3678 break;
3679
3680 case Op_EncodeP:
3681 case Op_EncodePKlass: {
3682 Node* in1 = n->in(1);
3683 if (in1->is_DecodeNarrowPtr()) {
3684 n->subsume_by(in1->in(1), this);
3685 } else if (in1->Opcode() == Op_ConP) {
3686 const Type* t = in1->bottom_type();
3687 if (t == TypePtr::NULL_PTR) {
3688 assert(t->isa_oopptr(), "null klass?");
3689 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3690 } else if (t->isa_oopptr()) {
3691 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3692 } else if (t->isa_klassptr()) {
3693 ciKlass* klass = t->is_klassptr()->exact_klass();
3694 if (klass->is_in_encoding_range()) {
3695 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3696 } else {
3697 assert(false, "unencodable klass in ConP -> EncodeP");
3698 C->record_failure("unencodable klass in ConP -> EncodeP");
3699 }
3700 }
3701 }
3702 if (in1->outcnt() == 0) {
3703 in1->disconnect_inputs(this);
3704 }
3705 break;
3706 }
3707
3708 case Op_Proj: {
3709 if (OptimizeStringConcat || IncrementalInline) {
3710 ProjNode* proj = n->as_Proj();
3711 if (proj->_is_io_use) {
3712 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3713 // Separate projections were used for the exception path which
3714 // are normally removed by a late inline. If it wasn't inlined
3715 // then they will hang around and should just be replaced with
3716 // the original one. Merge them.
3717 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3718 if (non_io_proj != nullptr) {
3719 proj->subsume_by(non_io_proj , this);
3720 }
3721 }
3722 }
3723 break;
3724 }
3725
3726 case Op_Phi:
3727 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3728 // The EncodeP optimization may create Phi with the same edges
3729 // for all paths. It is not handled well by Register Allocator.
3730 Node* unique_in = n->in(1);
3731 assert(unique_in != nullptr, "");
3732 uint cnt = n->req();
3733 for (uint i = 2; i < cnt; i++) {
3734 Node* m = n->in(i);
3735 assert(m != nullptr, "");
3736 if (unique_in != m)
3737 unique_in = nullptr;
3738 }
3739 if (unique_in != nullptr) {
3740 n->subsume_by(unique_in, this);
3741 }
3742 }
3743 break;
3744
3745 #endif
3746
3747 case Op_ModI:
3748 handle_div_mod_op(n, T_INT, false);
3749 break;
3750
3751 case Op_ModL:
3752 handle_div_mod_op(n, T_LONG, false);
3753 break;
3754
3755 case Op_UModI:
3756 handle_div_mod_op(n, T_INT, true);
3757 break;
3758
3759 case Op_UModL:
3760 handle_div_mod_op(n, T_LONG, true);
3761 break;
3762
3763 case Op_LoadVector:
3764 case Op_StoreVector:
3765 #ifdef ASSERT
3766 // Add VerifyVectorAlignment node between adr and load / store.
3767 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3768 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3769 n->as_StoreVector()->must_verify_alignment();
3770 if (must_verify_alignment) {
3771 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3772 n->as_StoreVector()->memory_size();
3773 // The memory access should be aligned to the vector width in bytes.
3774 // However, the underlying array is possibly less well aligned, but at least
3775 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3776 // a loop we can expect at least the following alignment:
3777 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3778 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3779 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3780 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3781 jlong mask = guaranteed_alignment - 1;
3782 Node* mask_con = ConLNode::make(mask);
3783 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3784 n->set_req(MemNode::Address, va);
3785 }
3786 }
3787 #endif
3788 break;
3789
3790 case Op_LoadVectorGather:
3791 case Op_StoreVectorScatter:
3792 case Op_LoadVectorGatherMasked:
3793 case Op_StoreVectorScatterMasked:
3794 case Op_VectorCmpMasked:
3795 case Op_VectorMaskGen:
3796 case Op_LoadVectorMasked:
3797 case Op_StoreVectorMasked:
3798 break;
3799
3800 case Op_AddReductionVI:
3801 case Op_AddReductionVL:
3802 case Op_AddReductionVHF:
3803 case Op_AddReductionVF:
3804 case Op_AddReductionVD:
3805 case Op_MulReductionVI:
3806 case Op_MulReductionVL:
3807 case Op_MulReductionVHF:
3808 case Op_MulReductionVF:
3809 case Op_MulReductionVD:
3810 case Op_MinReductionV:
3811 case Op_MaxReductionV:
3812 case Op_UMinReductionV:
3813 case Op_UMaxReductionV:
3814 case Op_AndReductionV:
3815 case Op_OrReductionV:
3816 case Op_XorReductionV:
3817 break;
3818
3819 case Op_PackB:
3820 case Op_PackS:
3821 case Op_PackI:
3822 case Op_PackF:
3823 case Op_PackL:
3824 case Op_PackD:
3825 if (n->req()-1 > 2) {
3826 // Replace many operand PackNodes with a binary tree for matching
3827 PackNode* p = (PackNode*) n;
3828 Node* btp = p->binary_tree_pack(1, n->req());
3829 n->subsume_by(btp, this);
3830 }
3831 break;
3832 case Op_Loop:
3833 // When StressCountedLoop is enabled, this loop may intentionally avoid a counted loop conversion.
3834 // This is expected behavior for the stress mode, which exercises alternative compilation paths.
3835 if (!StressCountedLoop) {
3836 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3837 }
3838 case Op_CountedLoop:
3839 case Op_LongCountedLoop:
3840 case Op_OuterStripMinedLoop:
3841 if (n->as_Loop()->is_inner_loop()) {
3842 frc.inc_inner_loop_count();
3843 }
3844 n->as_Loop()->verify_strip_mined(0);
3845 break;
3846 case Op_LShiftI:
3847 case Op_RShiftI:
3848 case Op_URShiftI:
3849 case Op_LShiftL:
3850 case Op_RShiftL:
3851 case Op_URShiftL:
3852 if (Matcher::need_masked_shift_count) {
3853 // The cpu's shift instructions don't restrict the count to the
3854 // lower 5/6 bits. We need to do the masking ourselves.
3855 Node* in2 = n->in(2);
3856 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3857 const TypeInt* t = in2->find_int_type();
3858 if (t != nullptr && t->is_con()) {
3859 juint shift = t->get_con();
3860 if (shift > mask) { // Unsigned cmp
3861 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3862 }
3863 } else {
3864 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3865 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3866 n->set_req(2, shift);
3867 }
3868 }
3869 if (in2->outcnt() == 0) { // Remove dead node
3870 in2->disconnect_inputs(this);
3871 }
3872 }
3873 break;
3874 case Op_MemBarStoreStore:
3875 case Op_MemBarRelease:
3876 // Break the link with AllocateNode: it is no longer useful and
3877 // confuses register allocation.
3878 if (n->req() > MemBarNode::Precedent) {
3879 n->set_req(MemBarNode::Precedent, top());
3880 }
3881 break;
3882 case Op_MemBarAcquire: {
3883 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3884 // At parse time, the trailing MemBarAcquire for a volatile load
3885 // is created with an edge to the load. After optimizations,
3886 // that input may be a chain of Phis. If those phis have no
3887 // other use, then the MemBarAcquire keeps them alive and
3888 // register allocation can be confused.
3889 dead_nodes.push(n->in(MemBarNode::Precedent));
3890 n->set_req(MemBarNode::Precedent, top());
3891 }
3892 break;
3893 }
3894 case Op_Blackhole:
3895 break;
3896 case Op_RangeCheck: {
3897 RangeCheckNode* rc = n->as_RangeCheck();
3898 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3899 n->subsume_by(iff, this);
3900 frc._tests.push(iff);
3901 break;
3902 }
3903 case Op_ConvI2L: {
3904 if (!Matcher::convi2l_type_required) {
3905 // Code generation on some platforms doesn't need accurate
3906 // ConvI2L types. Widening the type can help remove redundant
3907 // address computations.
3908 n->as_Type()->set_type(TypeLong::INT);
3909 ResourceMark rm;
3910 Unique_Node_List wq;
3911 wq.push(n);
3912 for (uint next = 0; next < wq.size(); next++) {
3913 Node *m = wq.at(next);
3914
3915 for(;;) {
3916 // Loop over all nodes with identical inputs edges as m
3917 Node* k = m->find_similar(m->Opcode());
3918 if (k == nullptr) {
3919 break;
3920 }
3921 // Push their uses so we get a chance to remove node made
3922 // redundant
3923 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3924 Node* u = k->fast_out(i);
3925 if (u->Opcode() == Op_LShiftL ||
3926 u->Opcode() == Op_AddL ||
3927 u->Opcode() == Op_SubL ||
3928 u->Opcode() == Op_AddP) {
3929 wq.push(u);
3930 }
3931 }
3932 // Replace all nodes with identical edges as m with m
3933 k->subsume_by(m, this);
3934 }
3935 }
3936 }
3937 break;
3938 }
3939 case Op_CmpUL: {
3940 if (!Matcher::has_match_rule(Op_CmpUL)) {
3941 // No support for unsigned long comparisons
3942 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3943 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3944 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3945 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3946 Node* andl = new AndLNode(orl, remove_sign_mask);
3947 Node* cmp = new CmpLNode(andl, n->in(2));
3948 n->subsume_by(cmp, this);
3949 }
3950 break;
3951 }
3952 #ifdef ASSERT
3953 case Op_ConNKlass: {
3954 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3955 ciKlass* klass = tp->is_klassptr()->exact_klass();
3956 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3957 break;
3958 }
3959 #endif
3960 default:
3961 assert(!n->is_Call(), "");
3962 assert(!n->is_Mem(), "");
3963 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3964 break;
3965 }
3966 }
3967
3968 //------------------------------final_graph_reshaping_walk---------------------
3969 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3970 // requires that the walk visits a node's inputs before visiting the node.
3971 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3972 Unique_Node_List sfpt;
3973
3974 frc._visited.set(root->_idx); // first, mark node as visited
3975 uint cnt = root->req();
3976 Node *n = root;
3977 uint i = 0;
3978 while (true) {
3979 if (i < cnt) {
3980 // Place all non-visited non-null inputs onto stack
3981 Node* m = n->in(i);
3982 ++i;
3983 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3984 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3985 // compute worst case interpreter size in case of a deoptimization
3986 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3987
3988 sfpt.push(m);
3989 }
3990 cnt = m->req();
3991 nstack.push(n, i); // put on stack parent and next input's index
3992 n = m;
3993 i = 0;
3994 }
3995 } else {
3996 // Now do post-visit work
3997 final_graph_reshaping_impl(n, frc, dead_nodes);
3998 if (nstack.is_empty())
3999 break; // finished
4000 n = nstack.node(); // Get node from stack
4001 cnt = n->req();
4002 i = nstack.index();
4003 nstack.pop(); // Shift to the next node on stack
4004 }
4005 }
4006
4007 expand_reachability_edges(sfpt);
4008
4009 // Skip next transformation if compressed oops are not used.
4010 if (UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks())
4011 return;
4012
4013 // Go over ReachabilityFence nodes to skip DecodeN nodes for referents.
4014 // The sole purpose of RF node is to keep the referent oop alive and
4015 // decoding the oop for that is not needed.
4016 for (int i = 0; i < C->reachability_fences_count(); i++) {
4017 ReachabilityFenceNode* rf = C->reachability_fence(i);
4018 DecodeNNode* dn = rf->in(1)->isa_DecodeN();
4019 if (dn != nullptr) {
4020 if (!dn->has_non_debug_uses() || Matcher::narrow_oop_use_complex_address()) {
4021 rf->set_req(1, dn->in(1));
4022 if (dn->outcnt() == 0) {
4023 dn->disconnect_inputs(this);
4024 }
4025 }
4026 }
4027 }
4028
4029 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4030 // It could be done for an uncommon traps or any safepoints/calls
4031 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4032 while (sfpt.size() > 0) {
4033 n = sfpt.pop();
4034 JVMState *jvms = n->as_SafePoint()->jvms();
4035 assert(jvms != nullptr, "sanity");
4036 int start = jvms->debug_start();
4037 int end = n->req();
4038 bool is_uncommon = (n->is_CallStaticJava() &&
4039 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4040 for (int j = start; j < end; j++) {
4041 Node* in = n->in(j);
4042 if (in->is_DecodeNarrowPtr() && (is_uncommon || !in->has_non_debug_uses())) {
4043 n->set_req(j, in->in(1));
4044 if (in->outcnt() == 0) {
4045 in->disconnect_inputs(this);
4046 }
4047 }
4048 }
4049 }
4050 }
4051
4052 //------------------------------final_graph_reshaping--------------------------
4053 // Final Graph Reshaping.
4054 //
4055 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4056 // and not commoned up and forced early. Must come after regular
4057 // optimizations to avoid GVN undoing the cloning. Clone constant
4058 // inputs to Loop Phis; these will be split by the allocator anyways.
4059 // Remove Opaque nodes.
4060 // (2) Move last-uses by commutative operations to the left input to encourage
4061 // Intel update-in-place two-address operations and better register usage
4062 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4063 // calls canonicalizing them back.
4064 // (3) Count the number of double-precision FP ops, single-precision FP ops
4065 // and call sites. On Intel, we can get correct rounding either by
4066 // forcing singles to memory (requires extra stores and loads after each
4067 // FP bytecode) or we can set a rounding mode bit (requires setting and
4068 // clearing the mode bit around call sites). The mode bit is only used
4069 // if the relative frequency of single FP ops to calls is low enough.
4070 // This is a key transform for SPEC mpeg_audio.
4071 // (4) Detect infinite loops; blobs of code reachable from above but not
4072 // below. Several of the Code_Gen algorithms fail on such code shapes,
4073 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4074 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4075 // Detection is by looking for IfNodes where only 1 projection is
4076 // reachable from below or CatchNodes missing some targets.
4077 // (5) Assert for insane oop offsets in debug mode.
4078
4079 bool Compile::final_graph_reshaping() {
4080 // an infinite loop may have been eliminated by the optimizer,
4081 // in which case the graph will be empty.
4082 if (root()->req() == 1) {
4083 // Do not compile method that is only a trivial infinite loop,
4084 // since the content of the loop may have been eliminated.
4085 record_method_not_compilable("trivial infinite loop");
4086 return true;
4087 }
4088
4089 // Expensive nodes have their control input set to prevent the GVN
4090 // from freely commoning them. There's no GVN beyond this point so
4091 // no need to keep the control input. We want the expensive nodes to
4092 // be freely moved to the least frequent code path by gcm.
4093 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4094 for (int i = 0; i < expensive_count(); i++) {
4095 _expensive_nodes.at(i)->set_req(0, nullptr);
4096 }
4097
4098 Final_Reshape_Counts frc;
4099
4100 // Visit everybody reachable!
4101 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4102 Node_Stack nstack(live_nodes() >> 1);
4103 Unique_Node_List dead_nodes;
4104 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4105
4106 // Check for unreachable (from below) code (i.e., infinite loops).
4107 for( uint i = 0; i < frc._tests.size(); i++ ) {
4108 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4109 // Get number of CFG targets.
4110 // Note that PCTables include exception targets after calls.
4111 uint required_outcnt = n->required_outcnt();
4112 if (n->outcnt() != required_outcnt) {
4113 // Check for a few special cases. Rethrow Nodes never take the
4114 // 'fall-thru' path, so expected kids is 1 less.
4115 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4116 if (n->in(0)->in(0)->is_Call()) {
4117 CallNode* call = n->in(0)->in(0)->as_Call();
4118 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4119 required_outcnt--; // Rethrow always has 1 less kid
4120 } else if (call->req() > TypeFunc::Parms &&
4121 call->is_CallDynamicJava()) {
4122 // Check for null receiver. In such case, the optimizer has
4123 // detected that the virtual call will always result in a null
4124 // pointer exception. The fall-through projection of this CatchNode
4125 // will not be populated.
4126 Node* arg0 = call->in(TypeFunc::Parms);
4127 if (arg0->is_Type() &&
4128 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4129 required_outcnt--;
4130 }
4131 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4132 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4133 // Check for illegal array length. In such case, the optimizer has
4134 // detected that the allocation attempt will always result in an
4135 // exception. There is no fall-through projection of this CatchNode .
4136 assert(call->is_CallStaticJava(), "static call expected");
4137 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4138 uint valid_length_test_input = call->req() - 1;
4139 Node* valid_length_test = call->in(valid_length_test_input);
4140 call->del_req(valid_length_test_input);
4141 if (valid_length_test->find_int_con(1) == 0) {
4142 required_outcnt--;
4143 }
4144 dead_nodes.push(valid_length_test);
4145 assert(n->outcnt() == required_outcnt, "malformed control flow");
4146 continue;
4147 }
4148 }
4149 }
4150
4151 // Recheck with a better notion of 'required_outcnt'
4152 if (n->outcnt() != required_outcnt) {
4153 record_method_not_compilable("malformed control flow");
4154 return true; // Not all targets reachable!
4155 }
4156 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4157 CallNode* call = n->in(0)->in(0)->as_Call();
4158 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4159 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4160 assert(call->is_CallStaticJava(), "static call expected");
4161 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4162 uint valid_length_test_input = call->req() - 1;
4163 dead_nodes.push(call->in(valid_length_test_input));
4164 call->del_req(valid_length_test_input); // valid length test useless now
4165 }
4166 }
4167 // Check that I actually visited all kids. Unreached kids
4168 // must be infinite loops.
4169 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4170 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4171 record_method_not_compilable("infinite loop");
4172 return true; // Found unvisited kid; must be unreach
4173 }
4174
4175 // Here so verification code in final_graph_reshaping_walk()
4176 // always see an OuterStripMinedLoopEnd
4177 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4178 IfNode* init_iff = n->as_If();
4179 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4180 n->subsume_by(iff, this);
4181 }
4182 }
4183
4184 while (dead_nodes.size() > 0) {
4185 Node* m = dead_nodes.pop();
4186 if (m->outcnt() == 0 && m != top()) {
4187 for (uint j = 0; j < m->req(); j++) {
4188 Node* in = m->in(j);
4189 if (in != nullptr) {
4190 dead_nodes.push(in);
4191 }
4192 }
4193 m->disconnect_inputs(this);
4194 }
4195 }
4196
4197 set_java_calls(frc.get_java_call_count());
4198 set_inner_loops(frc.get_inner_loop_count());
4199
4200 // No infinite loops, no reason to bail out.
4201 return false;
4202 }
4203
4204 //-----------------------------too_many_traps----------------------------------
4205 // Report if there are too many traps at the current method and bci.
4206 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4207 bool Compile::too_many_traps(ciMethod* method,
4208 int bci,
4209 Deoptimization::DeoptReason reason) {
4210 ciMethodData* md = method->method_data();
4211 if (md->is_empty()) {
4212 // Assume the trap has not occurred, or that it occurred only
4213 // because of a transient condition during start-up in the interpreter.
4214 return false;
4215 }
4216 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4217 if (md->has_trap_at(bci, m, reason) != 0) {
4218 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4219 // Also, if there are multiple reasons, or if there is no per-BCI record,
4220 // assume the worst.
4221 if (log())
4222 log()->elem("observe trap='%s' count='%d'",
4223 Deoptimization::trap_reason_name(reason),
4224 md->trap_count(reason));
4225 return true;
4226 } else {
4227 // Ignore method/bci and see if there have been too many globally.
4228 return too_many_traps(reason, md);
4229 }
4230 }
4231
4232 // Less-accurate variant which does not require a method and bci.
4233 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4234 ciMethodData* logmd) {
4235 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4236 // Too many traps globally.
4237 // Note that we use cumulative trap_count, not just md->trap_count.
4238 if (log()) {
4239 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4240 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4241 Deoptimization::trap_reason_name(reason),
4242 mcount, trap_count(reason));
4243 }
4244 return true;
4245 } else {
4246 // The coast is clear.
4247 return false;
4248 }
4249 }
4250
4251 //--------------------------too_many_recompiles--------------------------------
4252 // Report if there are too many recompiles at the current method and bci.
4253 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4254 // Is not eager to return true, since this will cause the compiler to use
4255 // Action_none for a trap point, to avoid too many recompilations.
4256 bool Compile::too_many_recompiles(ciMethod* method,
4257 int bci,
4258 Deoptimization::DeoptReason reason) {
4259 ciMethodData* md = method->method_data();
4260 if (md->is_empty()) {
4261 // Assume the trap has not occurred, or that it occurred only
4262 // because of a transient condition during start-up in the interpreter.
4263 return false;
4264 }
4265 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4266 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4267 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4268 Deoptimization::DeoptReason per_bc_reason
4269 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4270 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4271 if ((per_bc_reason == Deoptimization::Reason_none
4272 || md->has_trap_at(bci, m, reason) != 0)
4273 // The trap frequency measure we care about is the recompile count:
4274 && md->trap_recompiled_at(bci, m)
4275 && md->overflow_recompile_count() >= bc_cutoff) {
4276 // Do not emit a trap here if it has already caused recompilations.
4277 // Also, if there are multiple reasons, or if there is no per-BCI record,
4278 // assume the worst.
4279 if (log())
4280 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4281 Deoptimization::trap_reason_name(reason),
4282 md->trap_count(reason),
4283 md->overflow_recompile_count());
4284 return true;
4285 } else if (trap_count(reason) != 0
4286 && decompile_count() >= m_cutoff) {
4287 // Too many recompiles globally, and we have seen this sort of trap.
4288 // Use cumulative decompile_count, not just md->decompile_count.
4289 if (log())
4290 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4291 Deoptimization::trap_reason_name(reason),
4292 md->trap_count(reason), trap_count(reason),
4293 md->decompile_count(), decompile_count());
4294 return true;
4295 } else {
4296 // The coast is clear.
4297 return false;
4298 }
4299 }
4300
4301 // Compute when not to trap. Used by matching trap based nodes and
4302 // NullCheck optimization.
4303 void Compile::set_allowed_deopt_reasons() {
4304 _allowed_reasons = 0;
4305 if (is_method_compilation()) {
4306 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4307 assert(rs < BitsPerInt, "recode bit map");
4308 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4309 _allowed_reasons |= nth_bit(rs);
4310 }
4311 }
4312 }
4313 }
4314
4315 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4316 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4317 }
4318
4319 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4320 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4321 }
4322
4323 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4324 if (holder->is_initialized()) {
4325 return false;
4326 }
4327 if (holder->is_being_initialized()) {
4328 if (accessing_method->holder() == holder) {
4329 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4330 // <init>, or a static method. In all those cases, there was an initialization
4331 // barrier on the holder klass passed.
4332 if (accessing_method->is_static_initializer() ||
4333 accessing_method->is_object_initializer() ||
4334 accessing_method->is_static()) {
4335 return false;
4336 }
4337 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4338 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4339 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4340 // child class can become fully initialized while its parent class is still being initialized.
4341 if (accessing_method->is_static_initializer()) {
4342 return false;
4343 }
4344 }
4345 ciMethod* root = method(); // the root method of compilation
4346 if (root != accessing_method) {
4347 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4348 }
4349 }
4350 return true;
4351 }
4352
4353 #ifndef PRODUCT
4354 //------------------------------verify_bidirectional_edges---------------------
4355 // For each input edge to a node (ie - for each Use-Def edge), verify that
4356 // there is a corresponding Def-Use edge.
4357 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4358 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4359 uint stack_size = live_nodes() >> 4;
4360 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4361 if (root_and_safepoints != nullptr) {
4362 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4363 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4364 Node* root_or_safepoint = root_and_safepoints->at(i);
4365 // If the node is a safepoint, let's check if it still has a control input
4366 // Lack of control input signifies that this node was killed by CCP or
4367 // recursively by remove_globally_dead_node and it shouldn't be a starting
4368 // point.
4369 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4370 nstack.push(root_or_safepoint);
4371 }
4372 }
4373 } else {
4374 nstack.push(_root);
4375 }
4376
4377 while (nstack.size() > 0) {
4378 Node* n = nstack.pop();
4379 if (visited.member(n)) {
4380 continue;
4381 }
4382 visited.push(n);
4383
4384 // Walk over all input edges, checking for correspondence
4385 uint length = n->len();
4386 for (uint i = 0; i < length; i++) {
4387 Node* in = n->in(i);
4388 if (in != nullptr && !visited.member(in)) {
4389 nstack.push(in); // Put it on stack
4390 }
4391 if (in != nullptr && !in->is_top()) {
4392 // Count instances of `next`
4393 int cnt = 0;
4394 for (uint idx = 0; idx < in->_outcnt; idx++) {
4395 if (in->_out[idx] == n) {
4396 cnt++;
4397 }
4398 }
4399 assert(cnt > 0, "Failed to find Def-Use edge.");
4400 // Check for duplicate edges
4401 // walk the input array downcounting the input edges to n
4402 for (uint j = 0; j < length; j++) {
4403 if (n->in(j) == in) {
4404 cnt--;
4405 }
4406 }
4407 assert(cnt == 0, "Mismatched edge count.");
4408 } else if (in == nullptr) {
4409 assert(i == 0 || i >= n->req() ||
4410 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4411 (n->is_Unlock() && i == (n->req() - 1)) ||
4412 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4413 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4414 } else {
4415 assert(in->is_top(), "sanity");
4416 // Nothing to check.
4417 }
4418 }
4419 }
4420 }
4421
4422 //------------------------------verify_graph_edges---------------------------
4423 // Walk the Graph and verify that there is a one-to-one correspondence
4424 // between Use-Def edges and Def-Use edges in the graph.
4425 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4426 if (VerifyGraphEdges) {
4427 Unique_Node_List visited;
4428
4429 // Call graph walk to check edges
4430 verify_bidirectional_edges(visited, root_and_safepoints);
4431 if (no_dead_code) {
4432 // Now make sure that no visited node is used by an unvisited node.
4433 bool dead_nodes = false;
4434 Unique_Node_List checked;
4435 while (visited.size() > 0) {
4436 Node* n = visited.pop();
4437 checked.push(n);
4438 for (uint i = 0; i < n->outcnt(); i++) {
4439 Node* use = n->raw_out(i);
4440 if (checked.member(use)) continue; // already checked
4441 if (visited.member(use)) continue; // already in the graph
4442 if (use->is_Con()) continue; // a dead ConNode is OK
4443 // At this point, we have found a dead node which is DU-reachable.
4444 if (!dead_nodes) {
4445 tty->print_cr("*** Dead nodes reachable via DU edges:");
4446 dead_nodes = true;
4447 }
4448 use->dump(2);
4449 tty->print_cr("---");
4450 checked.push(use); // No repeats; pretend it is now checked.
4451 }
4452 }
4453 assert(!dead_nodes, "using nodes must be reachable from root");
4454 }
4455 }
4456 }
4457 #endif
4458
4459 // The Compile object keeps track of failure reasons separately from the ciEnv.
4460 // This is required because there is not quite a 1-1 relation between the
4461 // ciEnv and its compilation task and the Compile object. Note that one
4462 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4463 // to backtrack and retry without subsuming loads. Other than this backtracking
4464 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4465 // by the logic in C2Compiler.
4466 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4467 if (log() != nullptr) {
4468 log()->elem("failure reason='%s' phase='compile'", reason);
4469 }
4470 if (_failure_reason.get() == nullptr) {
4471 // Record the first failure reason.
4472 _failure_reason.set(reason);
4473 if (CaptureBailoutInformation) {
4474 _first_failure_details = new CompilationFailureInfo(reason);
4475 }
4476 } else {
4477 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4478 }
4479
4480 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4481 C->print_method(PHASE_FAILURE, 1);
4482 }
4483 _root = nullptr; // flush the graph, too
4484 }
4485
4486 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4487 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4488 _compile(Compile::current()),
4489 _log(nullptr),
4490 _dolog(CITimeVerbose)
4491 {
4492 assert(_compile != nullptr, "sanity check");
4493 assert(id != PhaseTraceId::_t_none, "Don't use none");
4494 if (_dolog) {
4495 _log = _compile->log();
4496 }
4497 if (_log != nullptr) {
4498 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4499 _log->stamp();
4500 _log->end_head();
4501 }
4502
4503 // Inform memory statistic, if enabled
4504 if (CompilationMemoryStatistic::enabled()) {
4505 CompilationMemoryStatistic::on_phase_start((int)id, name);
4506 }
4507 }
4508
4509 Compile::TracePhase::TracePhase(PhaseTraceId id)
4510 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4511
4512 Compile::TracePhase::~TracePhase() {
4513
4514 // Inform memory statistic, if enabled
4515 if (CompilationMemoryStatistic::enabled()) {
4516 CompilationMemoryStatistic::on_phase_end();
4517 }
4518
4519 if (_compile->failing_internal()) {
4520 if (_log != nullptr) {
4521 _log->done("phase");
4522 }
4523 return; // timing code, not stressing bailouts.
4524 }
4525 #ifdef ASSERT
4526 if (PrintIdealNodeCount) {
4527 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4528 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4529 }
4530
4531 if (VerifyIdealNodeCount) {
4532 _compile->print_missing_nodes();
4533 }
4534 #endif
4535
4536 if (_log != nullptr) {
4537 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4538 }
4539 }
4540
4541 //----------------------------static_subtype_check-----------------------------
4542 // Shortcut important common cases when superklass is exact:
4543 // (0) superklass is java.lang.Object (can occur in reflective code)
4544 // (1) subklass is already limited to a subtype of superklass => always ok
4545 // (2) subklass does not overlap with superklass => always fail
4546 // (3) superklass has NO subtypes and we can check with a simple compare.
4547 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4548 if (skip) {
4549 return SSC_full_test; // Let caller generate the general case.
4550 }
4551
4552 if (subk->is_java_subtype_of(superk)) {
4553 return SSC_always_true; // (0) and (1) this test cannot fail
4554 }
4555
4556 if (!subk->maybe_java_subtype_of(superk)) {
4557 return SSC_always_false; // (2) true path dead; no dynamic test needed
4558 }
4559
4560 const Type* superelem = superk;
4561 if (superk->isa_aryklassptr()) {
4562 int ignored;
4563 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4564 }
4565
4566 if (superelem->isa_instklassptr()) {
4567 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4568 if (!ik->has_subklass()) {
4569 if (!ik->is_final()) {
4570 // Add a dependency if there is a chance of a later subclass.
4571 dependencies()->assert_leaf_type(ik);
4572 }
4573 if (!superk->maybe_java_subtype_of(subk)) {
4574 return SSC_always_false;
4575 }
4576 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4577 }
4578 } else {
4579 // A primitive array type has no subtypes.
4580 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4581 }
4582
4583 return SSC_full_test;
4584 }
4585
4586 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4587 #ifdef _LP64
4588 // The scaled index operand to AddP must be a clean 64-bit value.
4589 // Java allows a 32-bit int to be incremented to a negative
4590 // value, which appears in a 64-bit register as a large
4591 // positive number. Using that large positive number as an
4592 // operand in pointer arithmetic has bad consequences.
4593 // On the other hand, 32-bit overflow is rare, and the possibility
4594 // can often be excluded, if we annotate the ConvI2L node with
4595 // a type assertion that its value is known to be a small positive
4596 // number. (The prior range check has ensured this.)
4597 // This assertion is used by ConvI2LNode::Ideal.
4598 int index_max = max_jint - 1; // array size is max_jint, index is one less
4599 if (sizetype != nullptr && sizetype->_hi > 0) {
4600 index_max = sizetype->_hi - 1;
4601 }
4602 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4603 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4604 #endif
4605 return idx;
4606 }
4607
4608 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4609 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4610 if (ctrl != nullptr) {
4611 // Express control dependency by a CastII node with a narrow type.
4612 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4613 // node from floating above the range check during loop optimizations. Otherwise, the
4614 // ConvI2L node may be eliminated independently of the range check, causing the data path
4615 // to become TOP while the control path is still there (although it's unreachable).
4616 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
4617 value = phase->transform(value);
4618 }
4619 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4620 return phase->transform(new ConvI2LNode(value, ltype));
4621 }
4622
4623 void Compile::dump_print_inlining() {
4624 inline_printer()->print_on(tty);
4625 }
4626
4627 void Compile::log_late_inline(CallGenerator* cg) {
4628 if (log() != nullptr) {
4629 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4630 cg->unique_id());
4631 JVMState* p = cg->call_node()->jvms();
4632 while (p != nullptr) {
4633 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4634 p = p->caller();
4635 }
4636 log()->tail("late_inline");
4637 }
4638 }
4639
4640 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4641 log_late_inline(cg);
4642 if (log() != nullptr) {
4643 log()->inline_fail(msg);
4644 }
4645 }
4646
4647 void Compile::log_inline_id(CallGenerator* cg) {
4648 if (log() != nullptr) {
4649 // The LogCompilation tool needs a unique way to identify late
4650 // inline call sites. This id must be unique for this call site in
4651 // this compilation. Try to have it unique across compilations as
4652 // well because it can be convenient when grepping through the log
4653 // file.
4654 // Distinguish OSR compilations from others in case CICountOSR is
4655 // on.
4656 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4657 cg->set_unique_id(id);
4658 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4659 }
4660 }
4661
4662 void Compile::log_inline_failure(const char* msg) {
4663 if (C->log() != nullptr) {
4664 C->log()->inline_fail(msg);
4665 }
4666 }
4667
4668
4669 // Dump inlining replay data to the stream.
4670 // Don't change thread state and acquire any locks.
4671 void Compile::dump_inline_data(outputStream* out) {
4672 InlineTree* inl_tree = ilt();
4673 if (inl_tree != nullptr) {
4674 out->print(" inline %d", inl_tree->count());
4675 inl_tree->dump_replay_data(out);
4676 }
4677 }
4678
4679 void Compile::dump_inline_data_reduced(outputStream* out) {
4680 assert(ReplayReduce, "");
4681
4682 InlineTree* inl_tree = ilt();
4683 if (inl_tree == nullptr) {
4684 return;
4685 }
4686 // Enable iterative replay file reduction
4687 // Output "compile" lines for depth 1 subtrees,
4688 // simulating that those trees were compiled
4689 // instead of inlined.
4690 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4691 InlineTree* sub = inl_tree->subtrees().at(i);
4692 if (sub->inline_level() != 1) {
4693 continue;
4694 }
4695
4696 ciMethod* method = sub->method();
4697 int entry_bci = -1;
4698 int comp_level = env()->task()->comp_level();
4699 out->print("compile ");
4700 method->dump_name_as_ascii(out);
4701 out->print(" %d %d", entry_bci, comp_level);
4702 out->print(" inline %d", sub->count());
4703 sub->dump_replay_data(out, -1);
4704 out->cr();
4705 }
4706 }
4707
4708 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4709 if (n1->Opcode() < n2->Opcode()) return -1;
4710 else if (n1->Opcode() > n2->Opcode()) return 1;
4711
4712 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4713 for (uint i = 1; i < n1->req(); i++) {
4714 if (n1->in(i) < n2->in(i)) return -1;
4715 else if (n1->in(i) > n2->in(i)) return 1;
4716 }
4717
4718 return 0;
4719 }
4720
4721 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4722 Node* n1 = *n1p;
4723 Node* n2 = *n2p;
4724
4725 return cmp_expensive_nodes(n1, n2);
4726 }
4727
4728 void Compile::sort_expensive_nodes() {
4729 if (!expensive_nodes_sorted()) {
4730 _expensive_nodes.sort(cmp_expensive_nodes);
4731 }
4732 }
4733
4734 bool Compile::expensive_nodes_sorted() const {
4735 for (int i = 1; i < _expensive_nodes.length(); i++) {
4736 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4737 return false;
4738 }
4739 }
4740 return true;
4741 }
4742
4743 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4744 if (_expensive_nodes.length() == 0) {
4745 return false;
4746 }
4747
4748 assert(OptimizeExpensiveOps, "optimization off?");
4749
4750 // Take this opportunity to remove dead nodes from the list
4751 int j = 0;
4752 for (int i = 0; i < _expensive_nodes.length(); i++) {
4753 Node* n = _expensive_nodes.at(i);
4754 if (!n->is_unreachable(igvn)) {
4755 assert(n->is_expensive(), "should be expensive");
4756 _expensive_nodes.at_put(j, n);
4757 j++;
4758 }
4759 }
4760 _expensive_nodes.trunc_to(j);
4761
4762 // Then sort the list so that similar nodes are next to each other
4763 // and check for at least two nodes of identical kind with same data
4764 // inputs.
4765 sort_expensive_nodes();
4766
4767 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4768 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4769 return true;
4770 }
4771 }
4772
4773 return false;
4774 }
4775
4776 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4777 if (_expensive_nodes.length() == 0) {
4778 return;
4779 }
4780
4781 assert(OptimizeExpensiveOps, "optimization off?");
4782
4783 // Sort to bring similar nodes next to each other and clear the
4784 // control input of nodes for which there's only a single copy.
4785 sort_expensive_nodes();
4786
4787 int j = 0;
4788 int identical = 0;
4789 int i = 0;
4790 bool modified = false;
4791 for (; i < _expensive_nodes.length()-1; i++) {
4792 assert(j <= i, "can't write beyond current index");
4793 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4794 identical++;
4795 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4796 continue;
4797 }
4798 if (identical > 0) {
4799 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4800 identical = 0;
4801 } else {
4802 Node* n = _expensive_nodes.at(i);
4803 igvn.replace_input_of(n, 0, nullptr);
4804 igvn.hash_insert(n);
4805 modified = true;
4806 }
4807 }
4808 if (identical > 0) {
4809 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4810 } else if (_expensive_nodes.length() >= 1) {
4811 Node* n = _expensive_nodes.at(i);
4812 igvn.replace_input_of(n, 0, nullptr);
4813 igvn.hash_insert(n);
4814 modified = true;
4815 }
4816 _expensive_nodes.trunc_to(j);
4817 if (modified) {
4818 igvn.optimize();
4819 }
4820 }
4821
4822 void Compile::add_expensive_node(Node * n) {
4823 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4824 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4825 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4826 if (OptimizeExpensiveOps) {
4827 _expensive_nodes.append(n);
4828 } else {
4829 // Clear control input and let IGVN optimize expensive nodes if
4830 // OptimizeExpensiveOps is off.
4831 n->set_req(0, nullptr);
4832 }
4833 }
4834
4835 /**
4836 * Track coarsened Lock and Unlock nodes.
4837 */
4838
4839 class Lock_List : public Node_List {
4840 uint _origin_cnt;
4841 public:
4842 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4843 uint origin_cnt() const { return _origin_cnt; }
4844 };
4845
4846 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4847 int length = locks.length();
4848 if (length > 0) {
4849 // Have to keep this list until locks elimination during Macro nodes elimination.
4850 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4851 AbstractLockNode* alock = locks.at(0);
4852 BoxLockNode* box = alock->box_node()->as_BoxLock();
4853 for (int i = 0; i < length; i++) {
4854 AbstractLockNode* lock = locks.at(i);
4855 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4856 locks_list->push(lock);
4857 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4858 if (this_box != box) {
4859 // Locking regions (BoxLock) could be Unbalanced here:
4860 // - its coarsened locks were eliminated in earlier
4861 // macro nodes elimination followed by loop unroll
4862 // - it is OSR locking region (no Lock node)
4863 // Preserve Unbalanced status in such cases.
4864 if (!this_box->is_unbalanced()) {
4865 this_box->set_coarsened();
4866 }
4867 if (!box->is_unbalanced()) {
4868 box->set_coarsened();
4869 }
4870 }
4871 }
4872 _coarsened_locks.append(locks_list);
4873 }
4874 }
4875
4876 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4877 int count = coarsened_count();
4878 for (int i = 0; i < count; i++) {
4879 Node_List* locks_list = _coarsened_locks.at(i);
4880 for (uint j = 0; j < locks_list->size(); j++) {
4881 Node* lock = locks_list->at(j);
4882 assert(lock->is_AbstractLock(), "sanity");
4883 if (!useful.member(lock)) {
4884 locks_list->yank(lock);
4885 }
4886 }
4887 }
4888 }
4889
4890 void Compile::remove_coarsened_lock(Node* n) {
4891 if (n->is_AbstractLock()) {
4892 int count = coarsened_count();
4893 for (int i = 0; i < count; i++) {
4894 Node_List* locks_list = _coarsened_locks.at(i);
4895 locks_list->yank(n);
4896 }
4897 }
4898 }
4899
4900 bool Compile::coarsened_locks_consistent() {
4901 int count = coarsened_count();
4902 for (int i = 0; i < count; i++) {
4903 bool unbalanced = false;
4904 bool modified = false; // track locks kind modifications
4905 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4906 uint size = locks_list->size();
4907 if (size == 0) {
4908 unbalanced = false; // All locks were eliminated - good
4909 } else if (size != locks_list->origin_cnt()) {
4910 unbalanced = true; // Some locks were removed from list
4911 } else {
4912 for (uint j = 0; j < size; j++) {
4913 Node* lock = locks_list->at(j);
4914 // All nodes in group should have the same state (modified or not)
4915 if (!lock->as_AbstractLock()->is_coarsened()) {
4916 if (j == 0) {
4917 // first on list was modified, the rest should be too for consistency
4918 modified = true;
4919 } else if (!modified) {
4920 // this lock was modified but previous locks on the list were not
4921 unbalanced = true;
4922 break;
4923 }
4924 } else if (modified) {
4925 // previous locks on list were modified but not this lock
4926 unbalanced = true;
4927 break;
4928 }
4929 }
4930 }
4931 if (unbalanced) {
4932 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4933 #ifdef ASSERT
4934 if (PrintEliminateLocks) {
4935 tty->print_cr("=== unbalanced coarsened locks ===");
4936 for (uint l = 0; l < size; l++) {
4937 locks_list->at(l)->dump();
4938 }
4939 }
4940 #endif
4941 record_failure(C2Compiler::retry_no_locks_coarsening());
4942 return false;
4943 }
4944 }
4945 return true;
4946 }
4947
4948 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4949 // locks coarsening optimization removed Lock/Unlock nodes from them.
4950 // Such regions become unbalanced because coarsening only removes part
4951 // of Lock/Unlock nodes in region. As result we can't execute other
4952 // locks elimination optimizations which assume all code paths have
4953 // corresponding pair of Lock/Unlock nodes - they are balanced.
4954 void Compile::mark_unbalanced_boxes() const {
4955 int count = coarsened_count();
4956 for (int i = 0; i < count; i++) {
4957 Node_List* locks_list = _coarsened_locks.at(i);
4958 uint size = locks_list->size();
4959 if (size > 0) {
4960 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4961 BoxLockNode* box = alock->box_node()->as_BoxLock();
4962 if (alock->is_coarsened()) {
4963 // coarsened_locks_consistent(), which is called before this method, verifies
4964 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4965 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4966 for (uint j = 1; j < size; j++) {
4967 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4968 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4969 if (box != this_box) {
4970 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4971 box->set_unbalanced();
4972 this_box->set_unbalanced();
4973 }
4974 }
4975 }
4976 }
4977 }
4978 }
4979
4980 /**
4981 * Remove the speculative part of types and clean up the graph
4982 */
4983 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4984 if (UseTypeSpeculation) {
4985 Unique_Node_List worklist;
4986 worklist.push(root());
4987 int modified = 0;
4988 // Go over all type nodes that carry a speculative type, drop the
4989 // speculative part of the type and enqueue the node for an igvn
4990 // which may optimize it out.
4991 for (uint next = 0; next < worklist.size(); ++next) {
4992 Node *n = worklist.at(next);
4993 if (n->is_Type()) {
4994 TypeNode* tn = n->as_Type();
4995 const Type* t = tn->type();
4996 const Type* t_no_spec = t->remove_speculative();
4997 if (t_no_spec != t) {
4998 bool in_hash = igvn.hash_delete(n);
4999 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5000 tn->set_type(t_no_spec);
5001 igvn.hash_insert(n);
5002 igvn._worklist.push(n); // give it a chance to go away
5003 modified++;
5004 }
5005 }
5006 // Iterate over outs - endless loops is unreachable from below
5007 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5008 Node *m = n->fast_out(i);
5009 if (not_a_node(m)) {
5010 continue;
5011 }
5012 worklist.push(m);
5013 }
5014 }
5015 // Drop the speculative part of all types in the igvn's type table
5016 igvn.remove_speculative_types();
5017 if (modified > 0) {
5018 igvn.optimize();
5019 if (failing()) return;
5020 }
5021 #ifdef ASSERT
5022 // Verify that after the IGVN is over no speculative type has resurfaced
5023 worklist.clear();
5024 worklist.push(root());
5025 for (uint next = 0; next < worklist.size(); ++next) {
5026 Node *n = worklist.at(next);
5027 const Type* t = igvn.type_or_null(n);
5028 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5029 if (n->is_Type()) {
5030 t = n->as_Type()->type();
5031 assert(t == t->remove_speculative(), "no more speculative types");
5032 }
5033 // Iterate over outs - endless loops is unreachable from below
5034 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5035 Node *m = n->fast_out(i);
5036 if (not_a_node(m)) {
5037 continue;
5038 }
5039 worklist.push(m);
5040 }
5041 }
5042 igvn.check_no_speculative_types();
5043 #endif
5044 }
5045 }
5046
5047 // Auxiliary methods to support randomized stressing/fuzzing.
5048
5049 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5050 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5051 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5052 FLAG_SET_ERGO(StressSeed, _stress_seed);
5053 } else {
5054 _stress_seed = StressSeed;
5055 }
5056 if (_log != nullptr) {
5057 _log->elem("stress_test seed='%u'", _stress_seed);
5058 }
5059 }
5060
5061 int Compile::random() {
5062 _stress_seed = os::next_random(_stress_seed);
5063 return static_cast<int>(_stress_seed);
5064 }
5065
5066 // This method can be called the arbitrary number of times, with current count
5067 // as the argument. The logic allows selecting a single candidate from the
5068 // running list of candidates as follows:
5069 // int count = 0;
5070 // Cand* selected = null;
5071 // while(cand = cand->next()) {
5072 // if (randomized_select(++count)) {
5073 // selected = cand;
5074 // }
5075 // }
5076 //
5077 // Including count equalizes the chances any candidate is "selected".
5078 // This is useful when we don't have the complete list of candidates to choose
5079 // from uniformly. In this case, we need to adjust the randomicity of the
5080 // selection, or else we will end up biasing the selection towards the latter
5081 // candidates.
5082 //
5083 // Quick back-envelope calculation shows that for the list of n candidates
5084 // the equal probability for the candidate to persist as "best" can be
5085 // achieved by replacing it with "next" k-th candidate with the probability
5086 // of 1/k. It can be easily shown that by the end of the run, the
5087 // probability for any candidate is converged to 1/n, thus giving the
5088 // uniform distribution among all the candidates.
5089 //
5090 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5091 #define RANDOMIZED_DOMAIN_POW 29
5092 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5093 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5094 bool Compile::randomized_select(int count) {
5095 assert(count > 0, "only positive");
5096 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5097 }
5098
5099 #ifdef ASSERT
5100 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5101 bool Compile::fail_randomly() {
5102 if ((random() % StressBailoutMean) != 0) {
5103 return false;
5104 }
5105 record_failure("StressBailout");
5106 return true;
5107 }
5108
5109 bool Compile::failure_is_artificial() {
5110 return C->failure_reason_is("StressBailout");
5111 }
5112 #endif
5113
5114 CloneMap& Compile::clone_map() { return _clone_map; }
5115 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5116
5117 void NodeCloneInfo::dump_on(outputStream* st) const {
5118 st->print(" {%d:%d} ", idx(), gen());
5119 }
5120
5121 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5122 uint64_t val = value(old->_idx);
5123 NodeCloneInfo cio(val);
5124 assert(val != 0, "old node should be in the map");
5125 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5126 insert(nnn->_idx, cin.get());
5127 #ifndef PRODUCT
5128 if (is_debug()) {
5129 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5130 }
5131 #endif
5132 }
5133
5134 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5135 NodeCloneInfo cio(value(old->_idx));
5136 if (cio.get() == 0) {
5137 cio.set(old->_idx, 0);
5138 insert(old->_idx, cio.get());
5139 #ifndef PRODUCT
5140 if (is_debug()) {
5141 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5142 }
5143 #endif
5144 }
5145 clone(old, nnn, gen);
5146 }
5147
5148 int CloneMap::max_gen() const {
5149 int g = 0;
5150 DictI di(_dict);
5151 for(; di.test(); ++di) {
5152 int t = gen(di._key);
5153 if (g < t) {
5154 g = t;
5155 #ifndef PRODUCT
5156 if (is_debug()) {
5157 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5158 }
5159 #endif
5160 }
5161 }
5162 return g;
5163 }
5164
5165 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5166 uint64_t val = value(key);
5167 if (val != 0) {
5168 NodeCloneInfo ni(val);
5169 ni.dump_on(st);
5170 }
5171 }
5172
5173 void Compile::shuffle_macro_nodes() {
5174 shuffle_array(*C, _macro_nodes);
5175 }
5176
5177 // Move Allocate nodes to the start of the list
5178 void Compile::sort_macro_nodes() {
5179 int count = macro_count();
5180 int allocates = 0;
5181 for (int i = 0; i < count; i++) {
5182 Node* n = macro_node(i);
5183 if (n->is_Allocate()) {
5184 if (i != allocates) {
5185 Node* tmp = macro_node(allocates);
5186 _macro_nodes.at_put(allocates, n);
5187 _macro_nodes.at_put(i, tmp);
5188 }
5189 allocates++;
5190 }
5191 }
5192 }
5193
5194 void Compile::print_method(CompilerPhaseType compile_phase, int level, Node* n) {
5195 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5196 EventCompilerPhase event(UNTIMED);
5197 if (event.should_commit()) {
5198 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, compile_phase, C->_compile_id, level);
5199 }
5200 #ifndef PRODUCT
5201 ResourceMark rm;
5202 stringStream ss;
5203 ss.print_raw(CompilerPhaseTypeHelper::to_description(compile_phase));
5204 int iter = ++_igv_phase_iter[compile_phase];
5205 if (iter > 1) {
5206 ss.print(" %d", iter);
5207 }
5208 if (n != nullptr) {
5209 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5210 if (n->is_Call()) {
5211 CallNode* call = n->as_Call();
5212 if (call->_name != nullptr) {
5213 // E.g. uncommon traps etc.
5214 ss.print(" - %s", call->_name);
5215 } else if (call->is_CallJava()) {
5216 CallJavaNode* call_java = call->as_CallJava();
5217 if (call_java->method() != nullptr) {
5218 ss.print(" -");
5219 call_java->method()->print_short_name(&ss);
5220 }
5221 }
5222 }
5223 }
5224
5225 const char* name = ss.as_string();
5226 if (should_print_igv(level)) {
5227 _igv_printer->print_graph(name);
5228 }
5229 if (should_print_phase(level)) {
5230 print_phase(name);
5231 }
5232 if (should_print_ideal_phase(compile_phase)) {
5233 print_ideal_ir(CompilerPhaseTypeHelper::to_name(compile_phase));
5234 }
5235 #endif
5236 C->_latest_stage_start_counter.stamp();
5237 }
5238
5239 // Only used from CompileWrapper
5240 void Compile::begin_method() {
5241 #ifndef PRODUCT
5242 if (_method != nullptr && should_print_igv(1)) {
5243 _igv_printer->begin_method();
5244 }
5245 #endif
5246 C->_latest_stage_start_counter.stamp();
5247 }
5248
5249 // Only used from CompileWrapper
5250 void Compile::end_method() {
5251 EventCompilerPhase event(UNTIMED);
5252 if (event.should_commit()) {
5253 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5254 }
5255
5256 #ifndef PRODUCT
5257 if (_method != nullptr && should_print_igv(1)) {
5258 _igv_printer->end_method();
5259 }
5260 #endif
5261 }
5262
5263 #ifndef PRODUCT
5264 bool Compile::should_print_phase(const int level) const {
5265 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5266 _method != nullptr; // Do not print phases for stubs.
5267 }
5268
5269 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5270 return _directive->should_print_ideal_phase(cpt);
5271 }
5272
5273 void Compile::init_igv() {
5274 if (_igv_printer == nullptr) {
5275 _igv_printer = IdealGraphPrinter::printer();
5276 _igv_printer->set_compile(this);
5277 }
5278 }
5279
5280 bool Compile::should_print_igv(const int level) {
5281 PRODUCT_RETURN_(return false;);
5282
5283 if (PrintIdealGraphLevel < 0) { // disabled by the user
5284 return false;
5285 }
5286
5287 bool need = directive()->IGVPrintLevelOption >= level;
5288 if (need) {
5289 Compile::init_igv();
5290 }
5291 return need;
5292 }
5293
5294 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5295 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5296
5297 // Called from debugger. Prints method to the default file with the default phase name.
5298 // This works regardless of any Ideal Graph Visualizer flags set or not.
5299 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5300 void igv_print(void* sp, void* fp, void* pc) {
5301 frame fr(sp, fp, pc);
5302 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5303 }
5304
5305 // Same as igv_print() above but with a specified phase name.
5306 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5307 frame fr(sp, fp, pc);
5308 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5309 }
5310
5311 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5312 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5313 // This works regardless of any Ideal Graph Visualizer flags set or not.
5314 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5315 void igv_print(bool network, void* sp, void* fp, void* pc) {
5316 frame fr(sp, fp, pc);
5317 if (network) {
5318 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5319 } else {
5320 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5321 }
5322 }
5323
5324 // Same as igv_print(bool network, ...) above but with a specified phase name.
5325 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5326 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5327 frame fr(sp, fp, pc);
5328 if (network) {
5329 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5330 } else {
5331 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5332 }
5333 }
5334
5335 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5336 void igv_print_default() {
5337 Compile::current()->print_method(PHASE_DEBUG, 0);
5338 }
5339
5340 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5341 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5342 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5343 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5344 void igv_append(void* sp, void* fp, void* pc) {
5345 frame fr(sp, fp, pc);
5346 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5347 }
5348
5349 // Same as igv_append(...) above but with a specified phase name.
5350 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5351 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5352 frame fr(sp, fp, pc);
5353 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5354 }
5355
5356 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5357 const char* file_name = "custom_debug.xml";
5358 if (_debug_file_printer == nullptr) {
5359 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5360 } else {
5361 _debug_file_printer->update_compiled_method(C->method());
5362 }
5363 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5364 _debug_file_printer->print_graph(phase_name, fr);
5365 }
5366
5367 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5368 ResourceMark rm;
5369 GrowableArray<const Node*> empty_list;
5370 igv_print_graph_to_network(phase_name, empty_list, fr);
5371 }
5372
5373 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5374 if (_debug_network_printer == nullptr) {
5375 _debug_network_printer = new IdealGraphPrinter(C);
5376 } else {
5377 _debug_network_printer->update_compiled_method(C->method());
5378 }
5379 tty->print_cr("Method printed over network stream to IGV");
5380 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5381 }
5382 #endif // !PRODUCT
5383
5384 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5385 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5386 return value;
5387 }
5388 Node* result = nullptr;
5389 if (bt == T_BYTE) {
5390 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5391 result = new RShiftINode(result, phase->intcon(24));
5392 } else if (bt == T_BOOLEAN) {
5393 result = new AndINode(value, phase->intcon(0xFF));
5394 } else if (bt == T_CHAR) {
5395 result = new AndINode(value,phase->intcon(0xFFFF));
5396 } else {
5397 assert(bt == T_SHORT, "unexpected narrow type");
5398 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5399 result = new RShiftINode(result, phase->intcon(16));
5400 }
5401 if (transform_res) {
5402 result = phase->transform(result);
5403 }
5404 return result;
5405 }
5406
5407 void Compile::record_method_not_compilable_oom() {
5408 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5409 }
5410
5411 #ifndef PRODUCT
5412 // Collects all the control inputs from nodes on the worklist and from their data dependencies
5413 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
5414 // Follow non-control edges until we reach CFG nodes
5415 for (uint i = 0; i < worklist.size(); i++) {
5416 const Node* n = worklist.at(i);
5417 for (uint j = 0; j < n->req(); j++) {
5418 Node* in = n->in(j);
5419 if (in == nullptr || in->is_Root()) {
5420 continue;
5421 }
5422 if (in->is_CFG()) {
5423 if (in->is_Call()) {
5424 // The return value of a call is only available if the call did not result in an exception
5425 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
5426 if (control_proj_use->is_Catch()) {
5427 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
5428 candidates.push(fall_through);
5429 continue;
5430 }
5431 }
5432
5433 if (in->is_Multi()) {
5434 // We got here by following data inputs so we should only have one control use
5435 // (no IfNode, etc)
5436 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
5437 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
5438 } else {
5439 candidates.push(in);
5440 }
5441 } else {
5442 worklist.push(in);
5443 }
5444 }
5445 }
5446 }
5447
5448 // Returns the candidate node that is a descendant to all the other candidates
5449 static Node* pick_control(Unique_Node_List& candidates) {
5450 Unique_Node_List worklist;
5451 worklist.copy(candidates);
5452
5453 // Traverse backwards through the CFG
5454 for (uint i = 0; i < worklist.size(); i++) {
5455 const Node* n = worklist.at(i);
5456 if (n->is_Root()) {
5457 continue;
5458 }
5459 for (uint j = 0; j < n->req(); j++) {
5460 // Skip backedge of loops to avoid cycles
5461 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
5462 continue;
5463 }
5464
5465 Node* pred = n->in(j);
5466 if (pred != nullptr && pred != n && pred->is_CFG()) {
5467 worklist.push(pred);
5468 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
5469 candidates.remove(pred);
5470 }
5471 }
5472 }
5473
5474 assert(candidates.size() == 1, "unexpected control flow");
5475 return candidates.at(0);
5476 }
5477
5478 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
5479 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
5480 call->init_req((*pos)++, parm);
5481 const BasicType bt = parm->bottom_type()->basic_type();
5482 if (bt == T_LONG || bt == T_DOUBLE) {
5483 call->init_req((*pos)++, half);
5484 }
5485 }
5486
5487 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
5488 Node* parm0, Node* parm1,
5489 Node* parm2, Node* parm3,
5490 Node* parm4, Node* parm5,
5491 Node* parm6) const {
5492 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
5493 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
5494 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
5495
5496 // find the most suitable control input
5497 Unique_Node_List worklist, candidates;
5498 if (parm0 != nullptr) { worklist.push(parm0);
5499 if (parm1 != nullptr) { worklist.push(parm1);
5500 if (parm2 != nullptr) { worklist.push(parm2);
5501 if (parm3 != nullptr) { worklist.push(parm3);
5502 if (parm4 != nullptr) { worklist.push(parm4);
5503 if (parm5 != nullptr) { worklist.push(parm5);
5504 if (parm6 != nullptr) { worklist.push(parm6);
5505 /* close each nested if ===> */ } } } } } } }
5506 find_candidate_control_inputs(worklist, candidates);
5507 Node* control = nullptr;
5508 if (candidates.size() == 0) {
5509 control = C->start()->proj_out(TypeFunc::Control);
5510 } else {
5511 control = pick_control(candidates);
5512 }
5513
5514 // find all the previous users of the control we picked
5515 GrowableArray<Node*> users_of_control;
5516 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
5517 Node* use = control->fast_out(i);
5518 if (use->is_CFG() && use != control) {
5519 users_of_control.push(use);
5520 }
5521 }
5522
5523 // we do not actually care about IO and memory as it uses neither
5524 call->init_req(TypeFunc::Control, control);
5525 call->init_req(TypeFunc::I_O, top());
5526 call->init_req(TypeFunc::Memory, top());
5527 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
5528 call->init_req(TypeFunc::ReturnAdr, top());
5529
5530 int pos = TypeFunc::Parms;
5531 call->init_req(pos++, str_node);
5532 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
5533 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
5534 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
5535 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
5536 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
5537 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
5538 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
5539 /* close each nested if ===> */ } } } } } } }
5540 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
5541
5542 call = gvn->transform(call);
5543 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
5544
5545 // rewire previous users to have the new call as control instead
5546 PhaseIterGVN* igvn = gvn->is_IterGVN();
5547 for (int i = 0; i < users_of_control.length(); i++) {
5548 Node* use = users_of_control.at(i);
5549 for (uint j = 0; j < use->req(); j++) {
5550 if (use->in(j) == control) {
5551 if (igvn != nullptr) {
5552 igvn->replace_input_of(use, j, call_control_proj);
5553 } else {
5554 gvn->hash_delete(use);
5555 use->set_req(j, call_control_proj);
5556 gvn->hash_insert(use);
5557 }
5558 }
5559 }
5560 }
5561
5562 return call;
5563 }
5564 #endif // !PRODUCT