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 #ifdef ASSERT
880 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
881 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
882 #endif
883
884 // Dump compilation data to replay it.
885 if (directive->DumpReplayOption) {
886 env()->dump_replay_data(_compile_id);
887 }
888 if (directive->DumpInlineOption && (ilt() != nullptr)) {
889 env()->dump_inline_data(_compile_id);
890 }
891
892 // Now that we know the size of all the monitors we can add a fixed slot
893 // for the original deopt pc.
894 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
895 set_fixed_slots(next_slot);
896
897 // Compute when to use implicit null checks. Used by matching trap based
898 // nodes and NullCheck optimization.
899 set_allowed_deopt_reasons();
900
901 // Now generate code
902 Code_Gen();
903 }
904
905 // C2 uses runtime stubs serialized generation to initialize its static tables
906 // shared by all compilations, like Type::_shared_type_dict.
907 // At least one stub have to be completely generated to execute intialization
908 // before we can skip the rest stubs generation by loading AOT cached stubs.
909
910 static bool c2_do_stub_init_complete = false;
911
912 //------------------------------Compile----------------------------------------
913 // Compile a runtime stub
914 Compile::Compile(ciEnv* ci_env,
915 TypeFunc_generator generator,
916 address stub_function,
917 const char* stub_name,
918 StubId stub_id,
919 int is_fancy_jump,
920 bool pass_tls,
921 bool return_pc,
922 DirectiveSet* directive)
923 : Phase(Compiler),
924 _compile_id(0),
925 _options(Options::for_runtime_stub()),
926 _method(nullptr),
927 _entry_bci(InvocationEntryBci),
928 _stub_function(stub_function),
929 _stub_name(stub_name),
930 _stub_id(stub_id),
931 _stub_entry_point(nullptr),
932 _max_node_limit(MaxNodeLimit),
933 _node_count_inlining_cutoff(NodeCountInliningCutoff),
934 _post_loop_opts_phase(false),
935 _merge_stores_phase(false),
936 _allow_macro_nodes(true),
937 _inlining_progress(false),
938 _inlining_incrementally(false),
939 _has_reserved_stack_access(false),
940 #ifndef PRODUCT
941 _igv_idx(0),
942 _trace_opto_output(directive->TraceOptoOutputOption),
943 #endif
944 _clinit_barrier_on_entry(false),
945 _stress_seed(0),
946 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
947 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
948 _env(ci_env),
949 _directive(directive),
950 _log(ci_env->log()),
951 _first_failure_details(nullptr),
952 _reachability_fences(comp_arena(), 8, 0, nullptr),
953 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
954 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
955 _congraph(nullptr),
956 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
957 _unique(0),
958 _dead_node_count(0),
959 _dead_node_list(comp_arena()),
960 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
961 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
962 _node_arena(&_node_arena_one),
963 _mach_constant_base_node(nullptr),
964 _Compile_types(mtCompiler, Arena::Tag::tag_type),
965 _initial_gvn(nullptr),
966 _igvn_worklist(nullptr),
967 _types(nullptr),
968 _node_hash(nullptr),
969 _has_mh_late_inlines(false),
970 _oom(false),
971 _replay_inline_data(nullptr),
972 _inline_printer(this),
973 _java_calls(0),
974 _inner_loops(0),
975 _FIRST_STACK_mask(comp_arena()),
976 _interpreter_frame_size(0),
977 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
978 _output(nullptr),
979 #ifndef PRODUCT
980 _in_dump_cnt(0),
981 #endif
982 _allowed_reasons(0) {
983 C = this;
984
985 // try to reuse an existing stub
986 if (c2_do_stub_init_complete) {
987 BlobId blob_id = StubInfo::blob(_stub_id);
988 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
989 if (blob != nullptr) {
990 RuntimeStub* rs = blob->as_runtime_stub();
991 _stub_entry_point = rs->entry_point();
992 return;
993 }
994 }
995
996 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
997 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
998
999 #ifndef PRODUCT
1000 set_print_assembly(PrintFrameConverterAssembly);
1001 set_parsed_irreducible_loop(false);
1002 #else
1003 set_print_assembly(false); // Must initialize.
1004 #endif
1005 set_has_irreducible_loop(false); // no loops
1006
1007 CompileWrapper cw(this);
1008 Init(/*do_aliasing=*/ false);
1009 init_tf((*generator)());
1010
1011 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1012 _types = new (comp_arena()) Type_Array(comp_arena());
1013 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1014
1015 if (StressLCM || StressGCM || StressBailout) {
1016 initialize_stress_seed(directive);
1017 }
1018
1019 {
1020 PhaseGVN gvn;
1021 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1022 gvn.transform(top());
1023
1024 GraphKit kit;
1025 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1026 }
1027
1028 NOT_PRODUCT( verify_graph_edges(); )
1029
1030 Code_Gen();
1031
1032 // First successful stub generation will set it to `true`
1033 // and it will stay `true` after that.
1034 c2_do_stub_init_complete = c2_do_stub_init_complete || (_stub_entry_point != nullptr);
1035 }
1036
1037 Compile::~Compile() {
1038 delete _first_failure_details;
1039 };
1040
1041 //------------------------------Init-------------------------------------------
1042 // Prepare for a single compilation
1043 void Compile::Init(bool aliasing) {
1044 _do_aliasing = aliasing;
1045 _unique = 0;
1046 _regalloc = nullptr;
1047
1048 _tf = nullptr; // filled in later
1049 _top = nullptr; // cached later
1050 _matcher = nullptr; // filled in later
1051 _cfg = nullptr; // filled in later
1052
1053 _node_note_array = nullptr;
1054 _default_node_notes = nullptr;
1055 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1056
1057 _immutable_memory = nullptr; // filled in at first inquiry
1058
1059 #ifdef ASSERT
1060 _phase_optimize_finished = false;
1061 _phase_verify_ideal_loop = false;
1062 _exception_backedge = false;
1063 _type_verify = nullptr;
1064 #endif
1065
1066 // Globally visible Nodes
1067 // First set TOP to null to give safe behavior during creation of RootNode
1068 set_cached_top_node(nullptr);
1069 set_root(new RootNode());
1070 // Now that you have a Root to point to, create the real TOP
1071 set_cached_top_node( new ConNode(Type::TOP) );
1072 set_recent_alloc(nullptr, nullptr);
1073
1074 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1075 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1076 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1077 env()->set_dependencies(new Dependencies(env()));
1078
1079 _fixed_slots = 0;
1080 set_has_split_ifs(false);
1081 set_has_loops(false); // first approximation
1082 set_has_stringbuilder(false);
1083 set_has_boxed_value(false);
1084 _trap_can_recompile = false; // no traps emitted yet
1085 _major_progress = true; // start out assuming good things will happen
1086 set_has_unsafe_access(false);
1087 set_max_vector_size(0);
1088 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1089 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1090 set_decompile_count(0);
1091
1092 #ifndef PRODUCT
1093 _phase_counter = 0;
1094 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1095 #endif
1096
1097 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1098 _loop_opts_cnt = LoopOptsCount;
1099 set_do_inlining(Inline);
1100 set_max_inline_size(MaxInlineSize);
1101 set_freq_inline_size(FreqInlineSize);
1102 set_do_scheduling(OptoScheduling);
1103
1104 set_do_vector_loop(false);
1105 set_has_monitors(false);
1106 set_has_scoped_access(false);
1107
1108 if (AllowVectorizeOnDemand) {
1109 if (has_method() && _directive->VectorizeOption) {
1110 set_do_vector_loop(true);
1111 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());})
1112 } else if (has_method() && method()->name() != nullptr &&
1113 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1114 set_do_vector_loop(true);
1115 }
1116 }
1117 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1118 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());})
1119
1120 _max_node_limit = _directive->MaxNodeLimitOption;
1121
1122 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1123 set_clinit_barrier_on_entry(true);
1124 }
1125 if (debug_info()->recording_non_safepoints()) {
1126 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1127 (comp_arena(), 8, 0, nullptr));
1128 set_default_node_notes(Node_Notes::make(this));
1129 }
1130
1131 const int grow_ats = 16;
1132 _max_alias_types = grow_ats;
1133 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1134 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1135 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1136 {
1137 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1138 }
1139 // Initialize the first few types.
1140 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1141 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1142 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1143 _num_alias_types = AliasIdxRaw+1;
1144 // Zero out the alias type cache.
1145 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1146 // A null adr_type hits in the cache right away. Preload the right answer.
1147 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1148 }
1149
1150 #ifdef ASSERT
1151 // Verify that the current StartNode is valid.
1152 void Compile::verify_start(StartNode* s) const {
1153 assert(failing_internal() || s == start(), "should be StartNode");
1154 }
1155 #endif
1156
1157 /**
1158 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1159 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1160 * the ideal graph.
1161 */
1162 StartNode* Compile::start() const {
1163 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1164 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1165 Node* start = root()->fast_out(i);
1166 if (start->is_Start()) {
1167 return start->as_Start();
1168 }
1169 }
1170 fatal("Did not find Start node!");
1171 return nullptr;
1172 }
1173
1174 //-------------------------------immutable_memory-------------------------------------
1175 // Access immutable memory
1176 Node* Compile::immutable_memory() {
1177 if (_immutable_memory != nullptr) {
1178 return _immutable_memory;
1179 }
1180 StartNode* s = start();
1181 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1182 Node *p = s->fast_out(i);
1183 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1184 _immutable_memory = p;
1185 return _immutable_memory;
1186 }
1187 }
1188 ShouldNotReachHere();
1189 return nullptr;
1190 }
1191
1192 //----------------------set_cached_top_node------------------------------------
1193 // Install the cached top node, and make sure Node::is_top works correctly.
1194 void Compile::set_cached_top_node(Node* tn) {
1195 if (tn != nullptr) verify_top(tn);
1196 Node* old_top = _top;
1197 _top = tn;
1198 // Calling Node::setup_is_top allows the nodes the chance to adjust
1199 // their _out arrays.
1200 if (_top != nullptr) _top->setup_is_top();
1201 if (old_top != nullptr) old_top->setup_is_top();
1202 assert(_top == nullptr || top()->is_top(), "");
1203 }
1204
1205 #ifdef ASSERT
1206 uint Compile::count_live_nodes_by_graph_walk() {
1207 Unique_Node_List useful(comp_arena());
1208 // Get useful node list by walking the graph.
1209 identify_useful_nodes(useful);
1210 return useful.size();
1211 }
1212
1213 void Compile::print_missing_nodes() {
1214
1215 // Return if CompileLog is null and PrintIdealNodeCount is false.
1216 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1217 return;
1218 }
1219
1220 // This is an expensive function. It is executed only when the user
1221 // specifies VerifyIdealNodeCount option or otherwise knows the
1222 // additional work that needs to be done to identify reachable nodes
1223 // by walking the flow graph and find the missing ones using
1224 // _dead_node_list.
1225
1226 Unique_Node_List useful(comp_arena());
1227 // Get useful node list by walking the graph.
1228 identify_useful_nodes(useful);
1229
1230 uint l_nodes = C->live_nodes();
1231 uint l_nodes_by_walk = useful.size();
1232
1233 if (l_nodes != l_nodes_by_walk) {
1234 if (_log != nullptr) {
1235 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1236 _log->stamp();
1237 _log->end_head();
1238 }
1239 VectorSet& useful_member_set = useful.member_set();
1240 int last_idx = l_nodes_by_walk;
1241 for (int i = 0; i < last_idx; i++) {
1242 if (useful_member_set.test(i)) {
1243 if (_dead_node_list.test(i)) {
1244 if (_log != nullptr) {
1245 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1246 }
1247 if (PrintIdealNodeCount) {
1248 // Print the log message to tty
1249 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1250 useful.at(i)->dump();
1251 }
1252 }
1253 }
1254 else if (! _dead_node_list.test(i)) {
1255 if (_log != nullptr) {
1256 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1257 }
1258 if (PrintIdealNodeCount) {
1259 // Print the log message to tty
1260 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1261 }
1262 }
1263 }
1264 if (_log != nullptr) {
1265 _log->tail("mismatched_nodes");
1266 }
1267 }
1268 }
1269 void Compile::record_modified_node(Node* n) {
1270 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1271 _modified_nodes->push(n);
1272 }
1273 }
1274
1275 void Compile::remove_modified_node(Node* n) {
1276 if (_modified_nodes != nullptr) {
1277 _modified_nodes->remove(n);
1278 }
1279 }
1280 #endif
1281
1282 #ifndef PRODUCT
1283 void Compile::verify_top(Node* tn) const {
1284 if (tn != nullptr) {
1285 assert(tn->is_Con(), "top node must be a constant");
1286 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1287 assert(tn->in(0) != nullptr, "must have live top node");
1288 }
1289 }
1290 #endif
1291
1292
1293 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1294
1295 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1296 guarantee(arr != nullptr, "");
1297 int num_blocks = arr->length();
1298 if (grow_by < num_blocks) grow_by = num_blocks;
1299 int num_notes = grow_by * _node_notes_block_size;
1300 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1301 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1302 while (num_notes > 0) {
1303 arr->append(notes);
1304 notes += _node_notes_block_size;
1305 num_notes -= _node_notes_block_size;
1306 }
1307 assert(num_notes == 0, "exact multiple, please");
1308 }
1309
1310 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1311 if (source == nullptr || dest == nullptr) return false;
1312
1313 if (dest->is_Con())
1314 return false; // Do not push debug info onto constants.
1315
1316 #ifdef ASSERT
1317 // Leave a bread crumb trail pointing to the original node:
1318 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1319 dest->set_debug_orig(source);
1320 }
1321 #endif
1322
1323 if (node_note_array() == nullptr)
1324 return false; // Not collecting any notes now.
1325
1326 // This is a copy onto a pre-existing node, which may already have notes.
1327 // If both nodes have notes, do not overwrite any pre-existing notes.
1328 Node_Notes* source_notes = node_notes_at(source->_idx);
1329 if (source_notes == nullptr || source_notes->is_clear()) return false;
1330 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1331 if (dest_notes == nullptr || dest_notes->is_clear()) {
1332 return set_node_notes_at(dest->_idx, source_notes);
1333 }
1334
1335 Node_Notes merged_notes = (*source_notes);
1336 // The order of operations here ensures that dest notes will win...
1337 merged_notes.update_from(dest_notes);
1338 return set_node_notes_at(dest->_idx, &merged_notes);
1339 }
1340
1341
1342 //--------------------------allow_range_check_smearing-------------------------
1343 // Gating condition for coalescing similar range checks.
1344 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1345 // single covering check that is at least as strong as any of them.
1346 // If the optimization succeeds, the simplified (strengthened) range check
1347 // will always succeed. If it fails, we will deopt, and then give up
1348 // on the optimization.
1349 bool Compile::allow_range_check_smearing() const {
1350 // If this method has already thrown a range-check,
1351 // assume it was because we already tried range smearing
1352 // and it failed.
1353 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1354 return !already_trapped;
1355 }
1356
1357
1358 //------------------------------flatten_alias_type-----------------------------
1359 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1360 assert(do_aliasing(), "Aliasing should be enabled");
1361 int offset = tj->offset();
1362 TypePtr::PTR ptr = tj->ptr();
1363
1364 // Known instance (scalarizable allocation) alias only with itself.
1365 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1366 tj->is_oopptr()->is_known_instance();
1367
1368 // Process weird unsafe references.
1369 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1370 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1371 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1372 tj = TypeOopPtr::BOTTOM;
1373 ptr = tj->ptr();
1374 offset = tj->offset();
1375 }
1376
1377 // Array pointers need some flattening
1378 const TypeAryPtr* ta = tj->isa_aryptr();
1379 if (ta && ta->is_stable()) {
1380 // Erase stability property for alias analysis.
1381 tj = ta = ta->cast_to_stable(false);
1382 }
1383 if( ta && is_known_inst ) {
1384 if ( offset != Type::OffsetBot &&
1385 offset > arrayOopDesc::length_offset_in_bytes() ) {
1386 offset = Type::OffsetBot; // Flatten constant access into array body only
1387 tj = ta = ta->
1388 remove_speculative()->
1389 cast_to_ptr_type(ptr)->
1390 with_offset(offset);
1391 }
1392 } else if (ta != nullptr) {
1393 // Common slices
1394 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1395 return TypeAryPtr::RANGE;
1396 } else if (offset == oopDesc::klass_offset_in_bytes()) {
1397 return TypeInstPtr::KLASS;
1398 } else if (offset == oopDesc::mark_offset_in_bytes()) {
1399 return TypeInstPtr::MARK;
1400 }
1401
1402 // Remove size and stability
1403 const TypeAry* normalized_ary = TypeAry::make(ta->elem(), TypeInt::POS, false);
1404 // Remove ptr, const_oop, and offset
1405 if (ta->elem() == Type::BOTTOM) {
1406 // Bottom array (meet of int[] and byte[] for example), accesses to it will be done with
1407 // Unsafe. This should alias with all arrays. For now just leave it as it is (this is
1408 // incorrect, see JDK-8331133).
1409 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, nullptr, false, Type::OffsetBot);
1410 } else if (ta->elem()->make_oopptr() != nullptr) {
1411 // Object arrays, all of them share the same slice
1412 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, TypeInt::POS, false);
1413 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, tary, nullptr, false, Type::OffsetBot);
1414 } else {
1415 // Primitive arrays
1416 tj = ta = TypeAryPtr::make(TypePtr::BotPTR, nullptr, normalized_ary, ta->exact_klass(), true, Type::OffsetBot);
1417 }
1418
1419 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1420 // cannot be distinguished by bytecode alone.
1421 if (ta->elem() == TypeInt::BOOL) {
1422 tj = ta = TypeAryPtr::BYTES;
1423 }
1424 }
1425
1426 // Oop pointers need some flattening
1427 const TypeInstPtr *to = tj->isa_instptr();
1428 if (to && to != TypeOopPtr::BOTTOM) {
1429 ciInstanceKlass* ik = to->instance_klass();
1430 if( ptr == TypePtr::Constant ) {
1431 if (ik != ciEnv::current()->Class_klass() ||
1432 offset < ik->layout_helper_size_in_bytes()) {
1433 // No constant oop pointers (such as Strings); they alias with
1434 // unknown strings.
1435 assert(!is_known_inst, "not scalarizable allocation");
1436 tj = to = to->
1437 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1438 remove_speculative()->
1439 cast_to_ptr_type(TypePtr::BotPTR)->
1440 cast_to_exactness(false);
1441 }
1442 } else if( is_known_inst ) {
1443 tj = to; // Keep NotNull and klass_is_exact for instance type
1444 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1445 // During the 2nd round of IterGVN, NotNull castings are removed.
1446 // Make sure the Bottom and NotNull variants alias the same.
1447 // Also, make sure exact and non-exact variants alias the same.
1448 tj = to = to->
1449 remove_speculative()->
1450 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1451 cast_to_ptr_type(TypePtr::BotPTR)->
1452 cast_to_exactness(false);
1453 }
1454 if (to->speculative() != nullptr) {
1455 tj = to = to->remove_speculative();
1456 }
1457 // Canonicalize the holder of this field
1458 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1459 // First handle header references such as a LoadKlassNode, even if the
1460 // object's klass is unloaded at compile time (4965979).
1461 if (!is_known_inst) { // Do it only for non-instance types
1462 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1463 }
1464 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1465 // Static fields are in the space above the normal instance
1466 // fields in the java.lang.Class instance.
1467 if (ik != ciEnv::current()->Class_klass()) {
1468 to = nullptr;
1469 tj = TypeOopPtr::BOTTOM;
1470 offset = tj->offset();
1471 }
1472 } else {
1473 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1474 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1475 assert(tj->offset() == offset, "no change to offset expected");
1476 bool xk = to->klass_is_exact();
1477 int instance_id = to->instance_id();
1478
1479 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1480 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1481 // its interfaces are included.
1482 if (xk && ik->equals(canonical_holder)) {
1483 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1484 } else {
1485 assert(xk || !is_known_inst, "Known instance should be exact type");
1486 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1487 }
1488 }
1489 }
1490
1491 // Klass pointers to object array klasses need some flattening
1492 const TypeKlassPtr *tk = tj->isa_klassptr();
1493 if( tk ) {
1494 // If we are referencing a field within a Klass, we need
1495 // to assume the worst case of an Object. Both exact and
1496 // inexact types must flatten to the same alias class so
1497 // use NotNull as the PTR.
1498 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1499 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1500 env()->Object_klass(),
1501 offset);
1502 }
1503
1504 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1505 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1506 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1507 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1508 } else {
1509 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1510 }
1511 }
1512
1513 // Check for precise loads from the primary supertype array and force them
1514 // to the supertype cache alias index. Check for generic array loads from
1515 // the primary supertype array and also force them to the supertype cache
1516 // alias index. Since the same load can reach both, we need to merge
1517 // these 2 disparate memories into the same alias class. Since the
1518 // primary supertype array is read-only, there's no chance of confusion
1519 // where we bypass an array load and an array store.
1520 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1521 if (offset == Type::OffsetBot ||
1522 (offset >= primary_supers_offset &&
1523 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1524 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1525 offset = in_bytes(Klass::secondary_super_cache_offset());
1526 tj = tk = tk->with_offset(offset);
1527 }
1528 }
1529
1530 // Flatten all Raw pointers together.
1531 if (tj->base() == Type::RawPtr)
1532 tj = TypeRawPtr::BOTTOM;
1533
1534 if (tj->base() == Type::AnyPtr)
1535 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1536
1537 offset = tj->offset();
1538 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1539
1540 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1541 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1542 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1543 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1544 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1545 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1546 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1547 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1548 assert( tj->ptr() != TypePtr::TopPTR &&
1549 tj->ptr() != TypePtr::AnyNull &&
1550 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1551 // assert( tj->ptr() != TypePtr::Constant ||
1552 // tj->base() == Type::RawPtr ||
1553 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1554
1555 return tj;
1556 }
1557
1558 void Compile::AliasType::Init(int i, const TypePtr* at) {
1559 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1560 _index = i;
1561 _adr_type = at;
1562 _field = nullptr;
1563 _element = nullptr;
1564 _is_rewritable = true; // default
1565 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1566 if (atoop != nullptr && atoop->is_known_instance()) {
1567 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1568 _general_index = Compile::current()->get_alias_index(gt);
1569 } else {
1570 _general_index = 0;
1571 }
1572 }
1573
1574 BasicType Compile::AliasType::basic_type() const {
1575 if (element() != nullptr) {
1576 const Type* element = adr_type()->is_aryptr()->elem();
1577 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1578 } if (field() != nullptr) {
1579 return field()->layout_type();
1580 } else {
1581 return T_ILLEGAL; // unknown
1582 }
1583 }
1584
1585 //---------------------------------print_on------------------------------------
1586 #ifndef PRODUCT
1587 void Compile::AliasType::print_on(outputStream* st) {
1588 if (index() < 10)
1589 st->print("@ <%d> ", index());
1590 else st->print("@ <%d>", index());
1591 st->print(is_rewritable() ? " " : " RO");
1592 int offset = adr_type()->offset();
1593 if (offset == Type::OffsetBot)
1594 st->print(" +any");
1595 else st->print(" +%-3d", offset);
1596 st->print(" in ");
1597 adr_type()->dump_on(st);
1598 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1599 if (field() != nullptr && tjp) {
1600 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1601 tjp->offset() != field()->offset_in_bytes()) {
1602 st->print(" != ");
1603 field()->print();
1604 st->print(" ***");
1605 }
1606 }
1607 }
1608
1609 void print_alias_types() {
1610 Compile* C = Compile::current();
1611 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1612 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1613 C->alias_type(idx)->print_on(tty);
1614 tty->cr();
1615 }
1616 }
1617 #endif
1618
1619
1620 //----------------------------probe_alias_cache--------------------------------
1621 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1622 intptr_t key = (intptr_t) adr_type;
1623 key ^= key >> logAliasCacheSize;
1624 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1625 }
1626
1627
1628 //-----------------------------grow_alias_types--------------------------------
1629 void Compile::grow_alias_types() {
1630 const int old_ats = _max_alias_types; // how many before?
1631 const int new_ats = old_ats; // how many more?
1632 const int grow_ats = old_ats+new_ats; // how many now?
1633 _max_alias_types = grow_ats;
1634 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), _alias_types, old_ats, grow_ats);
1635 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1636 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1637 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1638 }
1639
1640
1641 //--------------------------------find_alias_type------------------------------
1642 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1643 if (!do_aliasing()) {
1644 return alias_type(AliasIdxBot);
1645 }
1646
1647 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1648 if (ace->_adr_type == adr_type) {
1649 return alias_type(ace->_index);
1650 }
1651
1652 // Handle special cases.
1653 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1654 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1655
1656 // Do it the slow way.
1657 const TypePtr* flat = flatten_alias_type(adr_type);
1658
1659 #ifdef ASSERT
1660 {
1661 ResourceMark rm;
1662 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1663 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1664 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1665 Type::str(adr_type));
1666 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1667 const TypeOopPtr* foop = flat->is_oopptr();
1668 // Scalarizable allocations have exact klass always.
1669 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1670 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1671 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1672 Type::str(foop), Type::str(xoop));
1673 }
1674 }
1675 #endif
1676
1677 int idx = AliasIdxTop;
1678 for (int i = 0; i < num_alias_types(); i++) {
1679 if (alias_type(i)->adr_type() == flat) {
1680 idx = i;
1681 break;
1682 }
1683 }
1684
1685 if (idx == AliasIdxTop) {
1686 if (no_create) return nullptr;
1687 // Grow the array if necessary.
1688 if (_num_alias_types == _max_alias_types) grow_alias_types();
1689 // Add a new alias type.
1690 idx = _num_alias_types++;
1691 _alias_types[idx]->Init(idx, flat);
1692 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1693 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1694 if (flat->isa_instptr()) {
1695 if (flat->offset() == java_lang_Class::klass_offset()
1696 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1697 alias_type(idx)->set_rewritable(false);
1698 }
1699 if (flat->isa_aryptr()) {
1700 #ifdef ASSERT
1701 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1702 // (T_BYTE has the weakest alignment and size restrictions...)
1703 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1704 #endif
1705 if (flat->offset() == TypePtr::OffsetBot) {
1706 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1707 }
1708 }
1709 if (flat->isa_klassptr()) {
1710 if (UseCompactObjectHeaders) {
1711 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1712 alias_type(idx)->set_rewritable(false);
1713 }
1714 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1715 alias_type(idx)->set_rewritable(false);
1716 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1717 alias_type(idx)->set_rewritable(false);
1718 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1719 alias_type(idx)->set_rewritable(false);
1720 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1721 alias_type(idx)->set_rewritable(false);
1722 }
1723
1724 if (flat->isa_instklassptr()) {
1725 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1726 alias_type(idx)->set_rewritable(false);
1727 }
1728 }
1729 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1730 // but the base pointer type is not distinctive enough to identify
1731 // references into JavaThread.)
1732
1733 // Check for final fields.
1734 const TypeInstPtr* tinst = flat->isa_instptr();
1735 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1736 ciField* field;
1737 if (tinst->const_oop() != nullptr &&
1738 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1739 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1740 // static field
1741 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1742 field = k->get_field_by_offset(tinst->offset(), true);
1743 } else {
1744 ciInstanceKlass *k = tinst->instance_klass();
1745 field = k->get_field_by_offset(tinst->offset(), false);
1746 }
1747 assert(field == nullptr ||
1748 original_field == nullptr ||
1749 (field->holder() == original_field->holder() &&
1750 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1751 field->is_static() == original_field->is_static()), "wrong field?");
1752 // Set field() and is_rewritable() attributes.
1753 if (field != nullptr) alias_type(idx)->set_field(field);
1754 }
1755 }
1756
1757 // Fill the cache for next time.
1758 ace->_adr_type = adr_type;
1759 ace->_index = idx;
1760 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1761
1762 // Might as well try to fill the cache for the flattened version, too.
1763 AliasCacheEntry* face = probe_alias_cache(flat);
1764 if (face->_adr_type == nullptr) {
1765 face->_adr_type = flat;
1766 face->_index = idx;
1767 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1768 }
1769
1770 return alias_type(idx);
1771 }
1772
1773
1774 Compile::AliasType* Compile::alias_type(ciField* field) {
1775 const TypeOopPtr* t;
1776 if (field->is_static())
1777 t = TypeInstPtr::make(field->holder()->java_mirror());
1778 else
1779 t = TypeOopPtr::make_from_klass_raw(field->holder());
1780 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1781 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1782 return atp;
1783 }
1784
1785
1786 //------------------------------have_alias_type--------------------------------
1787 bool Compile::have_alias_type(const TypePtr* adr_type) {
1788 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1789 if (ace->_adr_type == adr_type) {
1790 return true;
1791 }
1792
1793 // Handle special cases.
1794 if (adr_type == nullptr) return true;
1795 if (adr_type == TypePtr::BOTTOM) return true;
1796
1797 return find_alias_type(adr_type, true, nullptr) != nullptr;
1798 }
1799
1800 //-----------------------------must_alias--------------------------------------
1801 // True if all values of the given address type are in the given alias category.
1802 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1803 if (alias_idx == AliasIdxBot) return true; // the universal category
1804 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1805 if (alias_idx == AliasIdxTop) return false; // the empty category
1806 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1807
1808 // the only remaining possible overlap is identity
1809 int adr_idx = get_alias_index(adr_type);
1810 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1811 assert(adr_idx == alias_idx ||
1812 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1813 && adr_type != TypeOopPtr::BOTTOM),
1814 "should not be testing for overlap with an unsafe pointer");
1815 return adr_idx == alias_idx;
1816 }
1817
1818 //------------------------------can_alias--------------------------------------
1819 // True if any values of the given address type are in the given alias category.
1820 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1821 if (alias_idx == AliasIdxTop) return false; // the empty category
1822 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1823 // Known instance doesn't alias with bottom memory
1824 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1825 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1826
1827 // the only remaining possible overlap is identity
1828 int adr_idx = get_alias_index(adr_type);
1829 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1830 return adr_idx == alias_idx;
1831 }
1832
1833 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1834 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1835 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1836 if (parse_predicate_count() == 0) {
1837 return;
1838 }
1839 for (int i = 0; i < parse_predicate_count(); i++) {
1840 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1841 parse_predicate->mark_useless(igvn);
1842 }
1843 _parse_predicates.clear();
1844 }
1845
1846 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1847 if (!n->for_post_loop_opts_igvn()) {
1848 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1849 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1850 _for_post_loop_igvn.append(n);
1851 }
1852 }
1853
1854 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1855 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1856 _for_post_loop_igvn.remove(n);
1857 }
1858
1859 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1860 // Verify that all previous optimizations produced a valid graph
1861 // at least to this point, even if no loop optimizations were done.
1862 PhaseIdealLoop::verify(igvn);
1863
1864 if (_print_phase_loop_opts) {
1865 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1866 }
1867 C->set_post_loop_opts_phase(); // no more loop opts allowed
1868
1869 assert(!C->major_progress(), "not cleared");
1870
1871 if (_for_post_loop_igvn.length() > 0) {
1872 while (_for_post_loop_igvn.length() > 0) {
1873 Node* n = _for_post_loop_igvn.pop();
1874 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1875 igvn._worklist.push(n);
1876 }
1877 igvn.optimize();
1878 if (failing()) return;
1879 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1880 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1881
1882 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1883 if (C->major_progress()) {
1884 C->clear_major_progress(); // ensure that major progress is now clear
1885 }
1886 }
1887 }
1888
1889 void Compile::record_for_merge_stores_igvn(Node* n) {
1890 if (!n->for_merge_stores_igvn()) {
1891 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1892 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1893 _for_merge_stores_igvn.append(n);
1894 }
1895 }
1896
1897 void Compile::remove_from_merge_stores_igvn(Node* n) {
1898 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1899 _for_merge_stores_igvn.remove(n);
1900 }
1901
1902 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1903 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1904 // the stores, and we merge the wrong sequence of stores.
1905 // Example:
1906 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1907 // Apply MergeStores:
1908 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1909 // Remove more RangeChecks:
1910 // StoreI [ StoreL ] StoreI
1911 // But now it would have been better to do this instead:
1912 // [ StoreL ] [ StoreL ]
1913 //
1914 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1915 // since we never unset _merge_stores_phase.
1916 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1917 C->set_merge_stores_phase();
1918
1919 if (_for_merge_stores_igvn.length() > 0) {
1920 while (_for_merge_stores_igvn.length() > 0) {
1921 Node* n = _for_merge_stores_igvn.pop();
1922 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1923 igvn._worklist.push(n);
1924 }
1925 igvn.optimize();
1926 if (failing()) return;
1927 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1928 print_method(PHASE_AFTER_MERGE_STORES, 3);
1929 }
1930 }
1931
1932 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1933 if (OptimizeUnstableIf) {
1934 _unstable_if_traps.append(trap);
1935 }
1936 }
1937
1938 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1939 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1940 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1941 Node* n = trap->uncommon_trap();
1942 if (!useful.member(n)) {
1943 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1944 }
1945 }
1946 }
1947
1948 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1949 // or fold-compares case. Return true if succeed or not found.
1950 //
1951 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1952 // false and ask fold-compares to yield.
1953 //
1954 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1955 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1956 // when deoptimization does happen.
1957 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1958 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1959 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1960 if (trap->uncommon_trap() == unc) {
1961 if (yield && trap->modified()) {
1962 return false;
1963 }
1964 _unstable_if_traps.delete_at(i);
1965 break;
1966 }
1967 }
1968 return true;
1969 }
1970
1971 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1972 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1973 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1974 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1975 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1976 CallStaticJavaNode* unc = trap->uncommon_trap();
1977 int next_bci = trap->next_bci();
1978 bool modified = trap->modified();
1979
1980 if (next_bci != -1 && !modified) {
1981 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1982 JVMState* jvms = unc->jvms();
1983 ciMethod* method = jvms->method();
1984 ciBytecodeStream iter(method);
1985
1986 iter.force_bci(jvms->bci());
1987 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1988 Bytecodes::Code c = iter.cur_bc();
1989 Node* lhs = nullptr;
1990 Node* rhs = nullptr;
1991 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1992 lhs = unc->peek_operand(0);
1993 rhs = unc->peek_operand(1);
1994 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1995 lhs = unc->peek_operand(0);
1996 }
1997
1998 ResourceMark rm;
1999 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2000 assert(live_locals.is_valid(), "broken liveness info");
2001 int len = (int)live_locals.size();
2002
2003 for (int i = 0; i < len; i++) {
2004 Node* local = unc->local(jvms, i);
2005 // kill local using the liveness of next_bci.
2006 // give up when the local looks like an operand to secure reexecution.
2007 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2008 uint idx = jvms->locoff() + i;
2009 #ifdef ASSERT
2010 if (PrintOpto && Verbose) {
2011 tty->print("[unstable_if] kill local#%d: ", idx);
2012 local->dump();
2013 tty->cr();
2014 }
2015 #endif
2016 igvn.replace_input_of(unc, idx, top());
2017 modified = true;
2018 }
2019 }
2020 }
2021
2022 // keep the mondified trap for late query
2023 if (modified) {
2024 trap->set_modified();
2025 } else {
2026 _unstable_if_traps.delete_at(i);
2027 }
2028 }
2029 igvn.optimize();
2030 }
2031
2032 // StringOpts and late inlining of string methods
2033 void Compile::inline_string_calls(bool parse_time) {
2034 {
2035 // remove useless nodes to make the usage analysis simpler
2036 ResourceMark rm;
2037 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2038 }
2039
2040 {
2041 ResourceMark rm;
2042 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2043 PhaseStringOpts pso(initial_gvn());
2044 print_method(PHASE_AFTER_STRINGOPTS, 3);
2045 }
2046
2047 // now inline anything that we skipped the first time around
2048 if (!parse_time) {
2049 _late_inlines_pos = _late_inlines.length();
2050 }
2051
2052 while (_string_late_inlines.length() > 0) {
2053 CallGenerator* cg = _string_late_inlines.pop();
2054 cg->do_late_inline();
2055 if (failing()) return;
2056 }
2057 _string_late_inlines.trunc_to(0);
2058 }
2059
2060 // Late inlining of boxing methods
2061 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2062 if (_boxing_late_inlines.length() > 0) {
2063 assert(has_boxed_value(), "inconsistent");
2064
2065 set_inlining_incrementally(true);
2066
2067 igvn_worklist()->ensure_empty(); // should be done with igvn
2068
2069 _late_inlines_pos = _late_inlines.length();
2070
2071 while (_boxing_late_inlines.length() > 0) {
2072 CallGenerator* cg = _boxing_late_inlines.pop();
2073 cg->do_late_inline();
2074 if (failing()) return;
2075 }
2076 _boxing_late_inlines.trunc_to(0);
2077
2078 inline_incrementally_cleanup(igvn);
2079
2080 set_inlining_incrementally(false);
2081 }
2082 }
2083
2084 bool Compile::inline_incrementally_one() {
2085 assert(IncrementalInline, "incremental inlining should be on");
2086 assert(_late_inlines.length() > 0, "should have been checked by caller");
2087
2088 TracePhase tp(_t_incrInline_inline);
2089
2090 set_inlining_progress(false);
2091 set_do_cleanup(false);
2092
2093 for (int i = 0; i < _late_inlines.length(); i++) {
2094 _late_inlines_pos = i+1;
2095 CallGenerator* cg = _late_inlines.at(i);
2096 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2097 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2098 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2099 if (should_stress_inlining()) {
2100 // randomly add repeated inline attempt if stress-inlining
2101 cg->call_node()->set_generator(cg);
2102 C->igvn_worklist()->push(cg->call_node());
2103 continue;
2104 }
2105 cg->do_late_inline();
2106 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2107 if (failing()) {
2108 return false;
2109 } else if (inlining_progress()) {
2110 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2111 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2112 break; // process one call site at a time
2113 } else {
2114 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2115 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2116 // Avoid potential infinite loop if node already in the IGVN list
2117 assert(false, "scheduled for IGVN during inlining attempt");
2118 } else {
2119 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2120 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2121 cg->call_node()->set_generator(cg);
2122 }
2123 }
2124 } else {
2125 // Ignore late inline direct calls when inlining is not allowed.
2126 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2127 }
2128 }
2129 // Remove processed elements.
2130 _late_inlines.remove_till(_late_inlines_pos);
2131 _late_inlines_pos = 0;
2132
2133 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2134
2135 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2136
2137 set_inlining_progress(false);
2138 set_do_cleanup(false);
2139
2140 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2141 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2142 }
2143
2144 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2145 {
2146 TracePhase tp(_t_incrInline_pru);
2147 ResourceMark rm;
2148 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2149 }
2150 {
2151 TracePhase tp(_t_incrInline_igvn);
2152 igvn.reset();
2153 igvn.optimize();
2154 if (failing()) return;
2155 }
2156 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2157 }
2158
2159 template<typename E>
2160 static void shuffle_array(Compile& C, GrowableArray<E>& array) {
2161 if (array.length() < 2) {
2162 return;
2163 }
2164 for (uint i = array.length() - 1; i >= 1; i--) {
2165 uint j = C.random() % (i + 1);
2166 swap(array.at(i), array.at(j));
2167 }
2168 }
2169
2170 void Compile::shuffle_late_inlines() {
2171 shuffle_array(*C, _late_inlines);
2172 }
2173
2174 // Perform incremental inlining until bound on number of live nodes is reached
2175 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2176 TracePhase tp(_t_incrInline);
2177
2178 set_inlining_incrementally(true);
2179 uint low_live_nodes = 0;
2180
2181 if (StressIncrementalInlining) {
2182 shuffle_late_inlines();
2183 }
2184
2185 while (_late_inlines.length() > 0) {
2186 if (live_nodes() > node_count_inlining_cutoff()) {
2187 if (low_live_nodes < node_count_inlining_cutoff() * 8 / 10) {
2188 TracePhase tp(_t_incrInline_ideal);
2189 // PhaseIdealLoop is expensive so we only try it once we are
2190 // out of live nodes and we only try it again if the previous
2191 // helped got the number of nodes down significantly
2192 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2193 if (failing()) return;
2194 low_live_nodes = live_nodes();
2195 _major_progress = true;
2196 }
2197
2198 if (live_nodes() > node_count_inlining_cutoff()) {
2199 bool do_print_inlining = print_inlining() || print_intrinsics();
2200 if (do_print_inlining || log() != nullptr) {
2201 // Print inlining message for candidates that we couldn't inline for lack of space.
2202 for (int i = 0; i < _late_inlines.length(); i++) {
2203 CallGenerator* cg = _late_inlines.at(i);
2204 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2205 if (do_print_inlining) {
2206 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2207 }
2208 log_late_inline_failure(cg, msg);
2209 }
2210 }
2211 break; // finish
2212 }
2213 }
2214
2215 igvn_worklist()->ensure_empty(); // should be done with igvn
2216
2217 if (_late_inlines.length() == 0) {
2218 break; // no more progress
2219 }
2220
2221 while (inline_incrementally_one()) {
2222 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2223 }
2224 if (failing()) return;
2225
2226 inline_incrementally_cleanup(igvn);
2227
2228 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2229
2230 if (failing()) return;
2231 }
2232
2233 igvn_worklist()->ensure_empty(); // should be done with igvn
2234
2235 if (_string_late_inlines.length() > 0) {
2236 assert(has_stringbuilder(), "inconsistent");
2237
2238 inline_string_calls(false);
2239
2240 if (failing()) return;
2241
2242 inline_incrementally_cleanup(igvn);
2243 }
2244
2245 set_inlining_incrementally(false);
2246 }
2247
2248 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2249 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2250 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2251 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2252 // as if "inlining_incrementally() == true" were set.
2253 assert(inlining_incrementally() == false, "not allowed");
2254 assert(_modified_nodes == nullptr, "not allowed");
2255 assert(_late_inlines.length() > 0, "sanity");
2256
2257 if (StressIncrementalInlining) {
2258 shuffle_late_inlines();
2259 }
2260
2261 while (_late_inlines.length() > 0) {
2262 igvn_worklist()->ensure_empty(); // should be done with igvn
2263
2264 while (inline_incrementally_one()) {
2265 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2266 }
2267 if (failing()) return;
2268
2269 inline_incrementally_cleanup(igvn);
2270 }
2271 }
2272
2273 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2274 if (_loop_opts_cnt > 0) {
2275 while (major_progress() && (_loop_opts_cnt > 0)) {
2276 TracePhase tp(_t_idealLoop);
2277 PhaseIdealLoop::optimize(igvn, mode);
2278 _loop_opts_cnt--;
2279 if (failing()) return false;
2280 if (major_progress()) {
2281 print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2282 }
2283 }
2284 }
2285 return true;
2286 }
2287
2288 // Remove edges from "root" to each SafePoint at a backward branch.
2289 // They were inserted during parsing (see add_safepoint()) to make
2290 // infinite loops without calls or exceptions visible to root, i.e.,
2291 // useful.
2292 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2293 Node *r = root();
2294 if (r != nullptr) {
2295 for (uint i = r->req(); i < r->len(); ++i) {
2296 Node *n = r->in(i);
2297 if (n != nullptr && n->is_SafePoint()) {
2298 r->rm_prec(i);
2299 if (n->outcnt() == 0) {
2300 igvn.remove_dead_node(n, PhaseIterGVN::NodeOrigin::Graph);
2301 }
2302 --i;
2303 }
2304 }
2305 // Parsing may have added top inputs to the root node (Path
2306 // leading to the Halt node proven dead). Make sure we get a
2307 // chance to clean them up.
2308 igvn._worklist.push(r);
2309 igvn.optimize();
2310 }
2311 }
2312
2313 //------------------------------Optimize---------------------------------------
2314 // Given a graph, optimize it.
2315 void Compile::Optimize() {
2316 TracePhase tp(_t_optimizer);
2317
2318 #ifndef PRODUCT
2319 if (env()->break_at_compile()) {
2320 BREAKPOINT;
2321 }
2322
2323 #endif
2324
2325 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2326 #ifdef ASSERT
2327 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2328 #endif
2329
2330 ResourceMark rm;
2331
2332 NOT_PRODUCT( verify_graph_edges(); )
2333
2334 print_method(PHASE_AFTER_PARSING, 1);
2335
2336 {
2337 // Iterative Global Value Numbering, including ideal transforms
2338 PhaseIterGVN igvn;
2339 #ifdef ASSERT
2340 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2341 #endif
2342 {
2343 TracePhase tp(_t_iterGVN);
2344 igvn.optimize(true);
2345 }
2346
2347 if (failing()) return;
2348
2349 print_method(PHASE_ITER_GVN1, 2);
2350
2351 process_for_unstable_if_traps(igvn);
2352
2353 if (failing()) return;
2354
2355 inline_incrementally(igvn);
2356
2357 print_method(PHASE_INCREMENTAL_INLINE, 2);
2358
2359 if (failing()) return;
2360
2361 if (eliminate_boxing()) {
2362 // Inline valueOf() methods now.
2363 inline_boxing_calls(igvn);
2364
2365 if (failing()) return;
2366
2367 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2368 inline_incrementally(igvn);
2369 }
2370
2371 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2372
2373 if (failing()) return;
2374 }
2375
2376 // Remove the speculative part of types and clean up the graph from
2377 // the extra CastPP nodes whose only purpose is to carry them. Do
2378 // that early so that optimizations are not disrupted by the extra
2379 // CastPP nodes.
2380 remove_speculative_types(igvn);
2381
2382 if (failing()) return;
2383
2384 // No more new expensive nodes will be added to the list from here
2385 // so keep only the actual candidates for optimizations.
2386 cleanup_expensive_nodes(igvn);
2387
2388 if (failing()) return;
2389
2390 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2391 if (EnableVectorSupport && has_vbox_nodes()) {
2392 TracePhase tp(_t_vector);
2393 PhaseVector pv(igvn);
2394 pv.optimize_vector_boxes();
2395 if (failing()) return;
2396 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2397 }
2398 assert(!has_vbox_nodes(), "sanity");
2399
2400 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2401 Compile::TracePhase tp(_t_renumberLive);
2402 igvn_worklist()->ensure_empty(); // should be done with igvn
2403 {
2404 ResourceMark rm;
2405 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2406 }
2407 igvn.reset();
2408 igvn.optimize(true);
2409 if (failing()) return;
2410 }
2411
2412 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2413 // safepoints
2414 remove_root_to_sfpts_edges(igvn);
2415
2416 if (failing()) return;
2417
2418 _print_phase_loop_opts = has_loops();
2419 if (_print_phase_loop_opts) {
2420 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2421 }
2422
2423 // Perform escape analysis
2424 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2425 if (has_loops()) {
2426 // Cleanup graph (remove dead nodes).
2427 TracePhase tp(_t_idealLoop);
2428 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2429 if (failing()) return;
2430 }
2431 bool progress;
2432 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2433 do {
2434 ConnectionGraph::do_analysis(this, &igvn);
2435
2436 if (failing()) return;
2437
2438 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2439
2440 // Optimize out fields loads from scalar replaceable allocations.
2441 igvn.optimize(true);
2442 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2443
2444 if (failing()) return;
2445
2446 if (congraph() != nullptr && macro_count() > 0) {
2447 TracePhase tp(_t_macroEliminate);
2448 PhaseMacroExpand mexp(igvn);
2449 mexp.eliminate_macro_nodes();
2450 if (failing()) return;
2451 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2452
2453 igvn.set_delay_transform(false);
2454 igvn.optimize();
2455 if (failing()) return;
2456
2457 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2458 }
2459
2460 ConnectionGraph::verify_ram_nodes(this, root());
2461 if (failing()) return;
2462
2463 progress = do_iterative_escape_analysis() &&
2464 (macro_count() < mcount) &&
2465 ConnectionGraph::has_candidates(this);
2466 // Try again if candidates exist and made progress
2467 // by removing some allocations and/or locks.
2468 } while (progress);
2469 }
2470
2471 // Loop transforms on the ideal graph. Range Check Elimination,
2472 // peeling, unrolling, etc.
2473
2474 // Set loop opts counter
2475 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2476 {
2477 TracePhase tp(_t_idealLoop);
2478 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2479 _loop_opts_cnt--;
2480 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2481 if (failing()) return;
2482 }
2483 // Loop opts pass if partial peeling occurred in previous pass
2484 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2485 TracePhase tp(_t_idealLoop);
2486 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2487 _loop_opts_cnt--;
2488 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2489 if (failing()) return;
2490 }
2491 // Loop opts pass for loop-unrolling before CCP
2492 if(major_progress() && (_loop_opts_cnt > 0)) {
2493 TracePhase tp(_t_idealLoop);
2494 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2495 _loop_opts_cnt--;
2496 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2497 }
2498 if (!failing()) {
2499 // Verify that last round of loop opts produced a valid graph
2500 PhaseIdealLoop::verify(igvn);
2501 }
2502 }
2503 if (failing()) return;
2504
2505 // Conditional Constant Propagation;
2506 print_method(PHASE_BEFORE_CCP1, 2);
2507 PhaseCCP ccp( &igvn );
2508 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2509 {
2510 TracePhase tp(_t_ccp);
2511 ccp.do_transform();
2512 }
2513 print_method(PHASE_CCP1, 2);
2514
2515 assert( true, "Break here to ccp.dump_old2new_map()");
2516
2517 // Iterative Global Value Numbering, including ideal transforms
2518 {
2519 TracePhase tp(_t_iterGVN2);
2520 igvn.reset_from_igvn(&ccp);
2521 igvn.optimize(true);
2522 }
2523 print_method(PHASE_ITER_GVN2, 2);
2524
2525 if (failing()) return;
2526
2527 // Loop transforms on the ideal graph. Range Check Elimination,
2528 // peeling, unrolling, etc.
2529 if (!optimize_loops(igvn, LoopOptsDefault)) {
2530 return;
2531 }
2532
2533 C->clear_major_progress(); // ensure that major progress is now clear
2534
2535 process_for_post_loop_opts_igvn(igvn);
2536
2537 if (failing()) return;
2538
2539 // Once loop optimizations are over, it is safe to get rid of all reachability fence nodes and
2540 // migrate reachability edges to safepoints.
2541 if (OptimizeReachabilityFences && _reachability_fences.length() > 0) {
2542 TracePhase tp1(_t_idealLoop);
2543 TracePhase tp2(_t_reachability);
2544 PhaseIdealLoop::optimize(igvn, PostLoopOptsExpandReachabilityFences);
2545 print_method(PHASE_EXPAND_REACHABILITY_FENCES, 2);
2546 if (failing()) return;
2547 assert(_reachability_fences.length() == 0 || PreserveReachabilityFencesOnConstants, "no RF nodes allowed");
2548 }
2549
2550 process_for_merge_stores_igvn(igvn);
2551
2552 if (failing()) return;
2553
2554 #ifdef ASSERT
2555 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2556 #endif
2557
2558 {
2559 TracePhase tp(_t_macroExpand);
2560 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2561 PhaseMacroExpand mex(igvn);
2562 // Do not allow new macro nodes once we start to eliminate and expand
2563 C->reset_allow_macro_nodes();
2564 // Last attempt to eliminate macro nodes before expand
2565 mex.eliminate_macro_nodes();
2566 if (failing()) {
2567 return;
2568 }
2569 mex.eliminate_opaque_looplimit_macro_nodes();
2570 if (failing()) {
2571 return;
2572 }
2573 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2574 if (mex.expand_macro_nodes()) {
2575 assert(failing(), "must bail out w/ explicit message");
2576 return;
2577 }
2578 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2579 }
2580
2581 {
2582 TracePhase tp(_t_barrierExpand);
2583 if (bs->expand_barriers(this, igvn)) {
2584 assert(failing(), "must bail out w/ explicit message");
2585 return;
2586 }
2587 print_method(PHASE_BARRIER_EXPANSION, 2);
2588 }
2589
2590 if (C->max_vector_size() > 0) {
2591 C->optimize_logic_cones(igvn);
2592 igvn.optimize();
2593 if (failing()) return;
2594 }
2595
2596 DEBUG_ONLY( _modified_nodes = nullptr; )
2597
2598 assert(igvn._worklist.size() == 0, "not empty");
2599
2600 if (_late_inlines.length() > 0) {
2601 // More opportunities to optimize virtual and MH calls.
2602 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2603 process_late_inline_calls_no_inline(igvn);
2604 if (failing()) return;
2605 }
2606 assert(_late_inlines.length() == 0, "late inline queue must be drained");
2607 } // (End scope of igvn; run destructor if necessary for asserts.)
2608
2609 check_no_dead_use();
2610
2611 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2612 // to remove hashes to unlock nodes for modifications.
2613 C->node_hash()->clear();
2614
2615 // A method with only infinite loops has no edges entering loops from root
2616 {
2617 TracePhase tp(_t_graphReshaping);
2618 if (final_graph_reshaping()) {
2619 assert(failing(), "must bail out w/ explicit message");
2620 return;
2621 }
2622 }
2623
2624 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2625 DEBUG_ONLY(set_phase_optimize_finished();)
2626 }
2627
2628 #ifdef ASSERT
2629 void Compile::check_no_dead_use() const {
2630 ResourceMark rm;
2631 Unique_Node_List wq;
2632 wq.push(root());
2633 for (uint i = 0; i < wq.size(); ++i) {
2634 Node* n = wq.at(i);
2635 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2636 Node* u = n->fast_out(j);
2637 if (u->outcnt() == 0 && !u->is_Con()) {
2638 u->dump();
2639 fatal("no reachable node should have no use");
2640 }
2641 wq.push(u);
2642 }
2643 }
2644 }
2645 #endif
2646
2647 void Compile::inline_vector_reboxing_calls() {
2648 if (C->_vector_reboxing_late_inlines.length() > 0) {
2649 _late_inlines_pos = C->_late_inlines.length();
2650 while (_vector_reboxing_late_inlines.length() > 0) {
2651 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2652 cg->do_late_inline();
2653 if (failing()) return;
2654 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2655 }
2656 _vector_reboxing_late_inlines.trunc_to(0);
2657 }
2658 }
2659
2660 bool Compile::has_vbox_nodes() {
2661 if (C->_vector_reboxing_late_inlines.length() > 0) {
2662 return true;
2663 }
2664 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2665 Node * n = C->macro_node(macro_idx);
2666 assert(n->is_macro(), "only macro nodes expected here");
2667 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2668 return true;
2669 }
2670 }
2671 return false;
2672 }
2673
2674 //---------------------------- Bitwise operation packing optimization ---------------------------
2675
2676 static bool is_vector_unary_bitwise_op(Node* n) {
2677 return n->Opcode() == Op_XorV &&
2678 VectorNode::is_vector_bitwise_not_pattern(n);
2679 }
2680
2681 static bool is_vector_binary_bitwise_op(Node* n) {
2682 switch (n->Opcode()) {
2683 case Op_AndV:
2684 case Op_OrV:
2685 return true;
2686
2687 case Op_XorV:
2688 return !is_vector_unary_bitwise_op(n);
2689
2690 default:
2691 return false;
2692 }
2693 }
2694
2695 static bool is_vector_ternary_bitwise_op(Node* n) {
2696 return n->Opcode() == Op_MacroLogicV;
2697 }
2698
2699 static bool is_vector_bitwise_op(Node* n) {
2700 return is_vector_unary_bitwise_op(n) ||
2701 is_vector_binary_bitwise_op(n) ||
2702 is_vector_ternary_bitwise_op(n);
2703 }
2704
2705 static bool is_vector_bitwise_cone_root(Node* n) {
2706 if (n->bottom_type()->isa_pvectmask() || !is_vector_bitwise_op(n)) {
2707 return false;
2708 }
2709 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2710 if (is_vector_bitwise_op(n->fast_out(i))) {
2711 return false;
2712 }
2713 }
2714 return true;
2715 }
2716
2717 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2718 uint cnt = 0;
2719 if (is_vector_bitwise_op(n)) {
2720 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2721 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2722 for (uint i = 1; i < inp_cnt; i++) {
2723 Node* in = n->in(i);
2724 bool skip = VectorNode::is_all_ones_vector(in);
2725 if (!skip && !inputs.member(in)) {
2726 inputs.push(in);
2727 cnt++;
2728 }
2729 }
2730 assert(cnt <= 1, "not unary");
2731 } else {
2732 uint last_req = inp_cnt;
2733 if (is_vector_ternary_bitwise_op(n)) {
2734 last_req = inp_cnt - 1; // skip last input
2735 }
2736 for (uint i = 1; i < last_req; i++) {
2737 Node* def = n->in(i);
2738 if (!inputs.member(def)) {
2739 inputs.push(def);
2740 cnt++;
2741 }
2742 }
2743 }
2744 } else { // not a bitwise operations
2745 if (!inputs.member(n)) {
2746 inputs.push(n);
2747 cnt++;
2748 }
2749 }
2750 return cnt;
2751 }
2752
2753 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2754 Unique_Node_List useful_nodes;
2755 C->identify_useful_nodes(useful_nodes);
2756
2757 for (uint i = 0; i < useful_nodes.size(); i++) {
2758 Node* n = useful_nodes.at(i);
2759 if (is_vector_bitwise_cone_root(n)) {
2760 list.push(n);
2761 }
2762 }
2763 }
2764
2765 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2766 const TypeVect* vt,
2767 Unique_Node_List& partition,
2768 Unique_Node_List& inputs) {
2769 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2770 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2771 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2772
2773 Node* in1 = inputs.at(0);
2774 Node* in2 = inputs.at(1);
2775 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2776
2777 uint func = compute_truth_table(partition, inputs);
2778
2779 Node* pn = partition.at(partition.size() - 1);
2780 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2781 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2782 }
2783
2784 static uint extract_bit(uint func, uint pos) {
2785 return (func & (1 << pos)) >> pos;
2786 }
2787
2788 //
2789 // A macro logic node represents a truth table. It has 4 inputs,
2790 // First three inputs corresponds to 3 columns of a truth table
2791 // and fourth input captures the logic function.
2792 //
2793 // eg. fn = (in1 AND in2) OR in3;
2794 //
2795 // MacroNode(in1,in2,in3,fn)
2796 //
2797 // -----------------
2798 // in1 in2 in3 fn
2799 // -----------------
2800 // 0 0 0 0
2801 // 0 0 1 1
2802 // 0 1 0 0
2803 // 0 1 1 1
2804 // 1 0 0 0
2805 // 1 0 1 1
2806 // 1 1 0 1
2807 // 1 1 1 1
2808 //
2809
2810 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2811 int res = 0;
2812 for (int i = 0; i < 8; i++) {
2813 int bit1 = extract_bit(in1, i);
2814 int bit2 = extract_bit(in2, i);
2815 int bit3 = extract_bit(in3, i);
2816
2817 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2818 int func_bit = extract_bit(func, func_bit_pos);
2819
2820 res |= func_bit << i;
2821 }
2822 return res;
2823 }
2824
2825 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
2826 assert(n != nullptr, "");
2827 assert(eval_map.contains(n), "absent");
2828 return *(eval_map.get(n));
2829 }
2830
2831 static void eval_operands(Node* n,
2832 uint& func1, uint& func2, uint& func3,
2833 HashTable<Node*,uint>& eval_map) {
2834 assert(is_vector_bitwise_op(n), "");
2835
2836 if (is_vector_unary_bitwise_op(n)) {
2837 Node* opnd = n->in(1);
2838 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2839 opnd = n->in(2);
2840 }
2841 func1 = eval_operand(opnd, eval_map);
2842 } else if (is_vector_binary_bitwise_op(n)) {
2843 func1 = eval_operand(n->in(1), eval_map);
2844 func2 = eval_operand(n->in(2), eval_map);
2845 } else {
2846 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2847 func1 = eval_operand(n->in(1), eval_map);
2848 func2 = eval_operand(n->in(2), eval_map);
2849 func3 = eval_operand(n->in(3), eval_map);
2850 }
2851 }
2852
2853 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2854 assert(inputs.size() <= 3, "sanity");
2855 ResourceMark rm;
2856 uint res = 0;
2857 HashTable<Node*,uint> eval_map;
2858
2859 // Populate precomputed functions for inputs.
2860 // Each input corresponds to one column of 3 input truth-table.
2861 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2862 0xCC, // (_, b, _) -> b
2863 0xF0 }; // (a, _, _) -> a
2864 for (uint i = 0; i < inputs.size(); i++) {
2865 eval_map.put(inputs.at(i), input_funcs[2-i]);
2866 }
2867
2868 for (uint i = 0; i < partition.size(); i++) {
2869 Node* n = partition.at(i);
2870
2871 uint func1 = 0, func2 = 0, func3 = 0;
2872 eval_operands(n, func1, func2, func3, eval_map);
2873
2874 switch (n->Opcode()) {
2875 case Op_OrV:
2876 assert(func3 == 0, "not binary");
2877 res = func1 | func2;
2878 break;
2879 case Op_AndV:
2880 assert(func3 == 0, "not binary");
2881 res = func1 & func2;
2882 break;
2883 case Op_XorV:
2884 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2885 assert(func2 == 0 && func3 == 0, "not unary");
2886 res = (~func1) & 0xFF;
2887 } else {
2888 assert(func3 == 0, "not binary");
2889 res = func1 ^ func2;
2890 }
2891 break;
2892 case Op_MacroLogicV:
2893 // Ordering of inputs may change during evaluation of sub-tree
2894 // containing MacroLogic node as a child node, thus a re-evaluation
2895 // makes sure that function is evaluated in context of current
2896 // inputs.
2897 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2898 break;
2899
2900 default: assert(false, "not supported: %s", n->Name());
2901 }
2902 assert(res <= 0xFF, "invalid");
2903 eval_map.put(n, res);
2904 }
2905 return res;
2906 }
2907
2908 // Criteria under which nodes gets packed into a macro logic node:-
2909 // 1) Parent and both child nodes are all unmasked or masked with
2910 // same predicates.
2911 // 2) Masked parent can be packed with left child if it is predicated
2912 // and both have same predicates.
2913 // 3) Masked parent can be packed with right child if its un-predicated
2914 // or has matching predication condition.
2915 // 4) An unmasked parent can be packed with an unmasked child.
2916 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2917 assert(partition.size() == 0, "not empty");
2918 assert(inputs.size() == 0, "not empty");
2919 if (is_vector_ternary_bitwise_op(n)) {
2920 return false;
2921 }
2922
2923 bool is_unary_op = is_vector_unary_bitwise_op(n);
2924 if (is_unary_op) {
2925 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2926 return false; // too few inputs
2927 }
2928
2929 bool pack_left_child = true;
2930 bool pack_right_child = true;
2931
2932 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2933 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2934
2935 int left_child_input_cnt = 0;
2936 int right_child_input_cnt = 0;
2937
2938 bool parent_is_predicated = n->is_predicated_vector();
2939 bool left_child_predicated = n->in(1)->is_predicated_vector();
2940 bool right_child_predicated = n->in(2)->is_predicated_vector();
2941
2942 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2943 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2944 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2945
2946 do {
2947 if (pack_left_child && left_child_LOP &&
2948 ((!parent_is_predicated && !left_child_predicated) ||
2949 ((parent_is_predicated && left_child_predicated &&
2950 parent_pred == left_child_pred)))) {
2951 partition.push(n->in(1));
2952 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2953 } else {
2954 inputs.push(n->in(1));
2955 left_child_input_cnt = 1;
2956 }
2957
2958 if (pack_right_child && right_child_LOP &&
2959 (!right_child_predicated ||
2960 (right_child_predicated && parent_is_predicated &&
2961 parent_pred == right_child_pred))) {
2962 partition.push(n->in(2));
2963 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2964 } else {
2965 inputs.push(n->in(2));
2966 right_child_input_cnt = 1;
2967 }
2968
2969 if (inputs.size() > 3) {
2970 assert(partition.size() > 0, "");
2971 inputs.clear();
2972 partition.clear();
2973 if (left_child_input_cnt > right_child_input_cnt) {
2974 pack_left_child = false;
2975 } else {
2976 pack_right_child = false;
2977 }
2978 } else {
2979 break;
2980 }
2981 } while(true);
2982
2983 if(partition.size()) {
2984 partition.push(n);
2985 }
2986
2987 return (partition.size() == 2 || partition.size() == 3) &&
2988 (inputs.size() == 2 || inputs.size() == 3);
2989 }
2990
2991 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2992 assert(is_vector_bitwise_op(n), "not a root");
2993
2994 visited.set(n->_idx);
2995
2996 // 1) Do a DFS walk over the logic cone.
2997 for (uint i = 1; i < n->req(); i++) {
2998 Node* in = n->in(i);
2999 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3000 process_logic_cone_root(igvn, in, visited);
3001 }
3002 }
3003
3004 // 2) Bottom up traversal: Merge node[s] with
3005 // the parent to form macro logic node.
3006 Unique_Node_List partition;
3007 Unique_Node_List inputs;
3008 if (compute_logic_cone(n, partition, inputs)) {
3009 const TypeVect* vt = n->bottom_type()->is_vect();
3010 Node* pn = partition.at(partition.size() - 1);
3011 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3012 if (mask == nullptr ||
3013 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3014 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3015 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3016 igvn.replace_node(n, macro_logic);
3017 }
3018 }
3019 }
3020
3021 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3022 ResourceMark rm;
3023 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3024 Unique_Node_List list;
3025 collect_logic_cone_roots(list);
3026
3027 while (list.size() > 0) {
3028 Node* n = list.pop();
3029 const TypeVect* vt = n->bottom_type()->is_vect();
3030 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3031 if (supported) {
3032 VectorSet visited(comp_arena());
3033 process_logic_cone_root(igvn, n, visited);
3034 }
3035 }
3036 }
3037 }
3038
3039 //------------------------------Code_Gen---------------------------------------
3040 // Given a graph, generate code for it
3041 void Compile::Code_Gen() {
3042 if (failing()) {
3043 return;
3044 }
3045
3046 // Perform instruction selection. You might think we could reclaim Matcher
3047 // memory PDQ, but actually the Matcher is used in generating spill code.
3048 // Internals of the Matcher (including some VectorSets) must remain live
3049 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3050 // set a bit in reclaimed memory.
3051
3052 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3053 // nodes. Mapping is only valid at the root of each matched subtree.
3054 NOT_PRODUCT( verify_graph_edges(); )
3055
3056 Matcher matcher;
3057 _matcher = &matcher;
3058 {
3059 TracePhase tp(_t_matcher);
3060 matcher.match();
3061 if (failing()) {
3062 return;
3063 }
3064 }
3065 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3066 // nodes. Mapping is only valid at the root of each matched subtree.
3067 NOT_PRODUCT( verify_graph_edges(); )
3068
3069 // If you have too many nodes, or if matching has failed, bail out
3070 check_node_count(0, "out of nodes matching instructions");
3071 if (failing()) {
3072 return;
3073 }
3074
3075 print_method(PHASE_MATCHING, 2);
3076
3077 // Build a proper-looking CFG
3078 PhaseCFG cfg(node_arena(), root(), matcher);
3079 if (failing()) {
3080 return;
3081 }
3082 _cfg = &cfg;
3083 {
3084 TracePhase tp(_t_scheduler);
3085 bool success = cfg.do_global_code_motion();
3086 if (!success) {
3087 return;
3088 }
3089
3090 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3091 NOT_PRODUCT( verify_graph_edges(); )
3092 cfg.verify();
3093 if (failing()) {
3094 return;
3095 }
3096 }
3097
3098 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3099 _regalloc = ®alloc;
3100 {
3101 TracePhase tp(_t_registerAllocation);
3102 // Perform register allocation. After Chaitin, use-def chains are
3103 // no longer accurate (at spill code) and so must be ignored.
3104 // Node->LRG->reg mappings are still accurate.
3105 _regalloc->Register_Allocate();
3106
3107 // Bail out if the allocator builds too many nodes
3108 if (failing()) {
3109 return;
3110 }
3111
3112 print_method(PHASE_REGISTER_ALLOCATION, 2);
3113 }
3114
3115 // Prior to register allocation we kept empty basic blocks in case the
3116 // the allocator needed a place to spill. After register allocation we
3117 // are not adding any new instructions. If any basic block is empty, we
3118 // can now safely remove it.
3119 {
3120 TracePhase tp(_t_blockOrdering);
3121 cfg.remove_empty_blocks();
3122 if (do_freq_based_layout()) {
3123 PhaseBlockLayout layout(cfg);
3124 } else {
3125 cfg.set_loop_alignment();
3126 }
3127 cfg.fixup_flow();
3128 cfg.remove_unreachable_blocks();
3129 cfg.verify_dominator_tree();
3130 print_method(PHASE_BLOCK_ORDERING, 3);
3131 }
3132
3133 // Apply peephole optimizations
3134 if( OptoPeephole ) {
3135 TracePhase tp(_t_peephole);
3136 PhasePeephole peep( _regalloc, cfg);
3137 peep.do_transform();
3138 print_method(PHASE_PEEPHOLE, 3);
3139 }
3140
3141 // Do late expand if CPU requires this.
3142 if (Matcher::require_postalloc_expand) {
3143 TracePhase tp(_t_postalloc_expand);
3144 cfg.postalloc_expand(_regalloc);
3145 print_method(PHASE_POSTALLOC_EXPAND, 3);
3146 }
3147
3148 #ifdef ASSERT
3149 {
3150 CompilationMemoryStatistic::do_test_allocations();
3151 if (failing()) return;
3152 }
3153 #endif
3154
3155 // Convert Nodes to instruction bits in a buffer
3156 {
3157 TracePhase tp(_t_output);
3158 PhaseOutput output;
3159 output.Output();
3160 if (failing()) return;
3161 output.install();
3162 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3163 }
3164
3165 // He's dead, Jim.
3166 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3167 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3168 }
3169
3170 //------------------------------Final_Reshape_Counts---------------------------
3171 // This class defines counters and node lists collected during
3172 // the final graph reshaping.
3173 struct Final_Reshape_Counts : public StackObj {
3174 int _java_call_count; // count non-inlined 'java' calls
3175 int _inner_loop_count; // count loops which need alignment
3176 VectorSet _visited; // Visitation flags
3177 Node_List _tests; // Set of IfNodes & PCTableNodes
3178
3179 Final_Reshape_Counts() :
3180 _java_call_count(0), _inner_loop_count(0) { }
3181
3182 void inc_java_call_count() { _java_call_count++; }
3183 void inc_inner_loop_count() { _inner_loop_count++; }
3184
3185 int get_java_call_count() const { return _java_call_count; }
3186 int get_inner_loop_count() const { return _inner_loop_count; }
3187 };
3188
3189 //------------------------------final_graph_reshaping_impl----------------------
3190 // Implement items 1-5 from final_graph_reshaping below.
3191 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3192
3193 if ( n->outcnt() == 0 ) return; // dead node
3194 uint nop = n->Opcode();
3195
3196 // Check for 2-input instruction with "last use" on right input.
3197 // Swap to left input. Implements item (2).
3198 if( n->req() == 3 && // two-input instruction
3199 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3200 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3201 n->in(2)->outcnt() == 1 &&// right use IS a last use
3202 !n->in(2)->is_Con() ) { // right use is not a constant
3203 // Check for commutative opcode
3204 switch( nop ) {
3205 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddHF: case Op_AddL:
3206 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3207 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3208 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulHF: case Op_MulL:
3209 case Op_AndL: case Op_XorL: case Op_OrL:
3210 case Op_AndI: case Op_XorI: case Op_OrI: {
3211 // Move "last use" input to left by swapping inputs
3212 n->swap_edges(1, 2);
3213 break;
3214 }
3215 default:
3216 break;
3217 }
3218 }
3219
3220 #ifdef ASSERT
3221 if( n->is_Mem() ) {
3222 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3223 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3224 // oop will be recorded in oop map if load crosses safepoint
3225 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3226 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3227 "raw memory operations should have control edge");
3228 }
3229 if (n->is_MemBar()) {
3230 MemBarNode* mb = n->as_MemBar();
3231 if (mb->trailing_store() || mb->trailing_load_store()) {
3232 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3233 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3234 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3235 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3236 } else if (mb->leading()) {
3237 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3238 }
3239 }
3240 if (n->is_CallLeafPure()) {
3241 // A pure call whose result projection is unused should have been
3242 // eliminated by CallLeafPureNode::Ideal during IGVN.
3243 assert(n->as_CallLeafPure()->proj_out_or_null(TypeFunc::Parms) != nullptr,
3244 "unused CallLeafPureNode should have been removed before final graph reshaping");
3245 }
3246 #endif
3247 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3248 if (!gc_handled) {
3249 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3250 }
3251
3252 // Collect CFG split points
3253 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3254 frc._tests.push(n);
3255 }
3256 }
3257
3258 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3259 if (!UseDivMod) {
3260 return;
3261 }
3262
3263 // Check if "a % b" and "a / b" both exist
3264 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3265 if (d == nullptr) {
3266 return;
3267 }
3268
3269 // Replace them with a fused divmod if supported
3270 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3271 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3272 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3273 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3274 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3275 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3276 // DivMod node so the dependency is not lost.
3277 divmod->add_prec_from(n);
3278 divmod->add_prec_from(d);
3279 d->subsume_by(divmod->div_proj(), this);
3280 n->subsume_by(divmod->mod_proj(), this);
3281 } else {
3282 // Replace "a % b" with "a - ((a / b) * b)"
3283 Node* mult = MulNode::make(d, d->in(2), bt);
3284 Node* sub = SubNode::make(d->in(1), mult, bt);
3285 n->subsume_by(sub, this);
3286 }
3287 }
3288
3289 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3290 switch( nop ) {
3291 case Op_Opaque1: // Remove Opaque Nodes before matching
3292 n->subsume_by(n->in(1), this);
3293 break;
3294 case Op_CallLeafPure: {
3295 // If the pure call is not supported, then lower to a CallLeaf.
3296 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3297 CallNode* call = n->as_Call();
3298 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3299 call->_name, TypeRawPtr::BOTTOM);
3300 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3301 new_call->init_req(TypeFunc::I_O, C->top());
3302 new_call->init_req(TypeFunc::Memory, C->top());
3303 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3304 new_call->init_req(TypeFunc::FramePtr, C->top());
3305 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
3306 new_call->init_req(i, call->in(i));
3307 }
3308 n->subsume_by(new_call, this);
3309 }
3310 break;
3311 }
3312 case Op_CallStaticJava:
3313 case Op_CallJava:
3314 case Op_CallDynamicJava:
3315 frc.inc_java_call_count(); // Count java call site;
3316 case Op_CallRuntime:
3317 case Op_CallLeaf:
3318 case Op_CallLeafVector:
3319 case Op_CallLeafNoFP: {
3320 assert (n->is_Call(), "");
3321 CallNode *call = n->as_Call();
3322 // See if uncommon argument is shared
3323 if (call->is_CallStaticJava() && call->as_CallStaticJava()->_name) {
3324 Node *n = call->in(TypeFunc::Parms);
3325 int nop = n->Opcode();
3326 // Clone shared simple arguments to uncommon calls, item (1).
3327 if (n->outcnt() > 1 &&
3328 !n->is_Proj() &&
3329 nop != Op_CreateEx &&
3330 nop != Op_CheckCastPP &&
3331 nop != Op_DecodeN &&
3332 nop != Op_DecodeNKlass &&
3333 !n->is_Mem() &&
3334 !n->is_Phi()) {
3335 Node *x = n->clone();
3336 call->set_req(TypeFunc::Parms, x);
3337 }
3338 }
3339 break;
3340 }
3341
3342 // Mem nodes need explicit cases to satisfy assert(!n->is_Mem()) in default.
3343 case Op_StoreF:
3344 case Op_LoadF:
3345 case Op_StoreD:
3346 case Op_LoadD:
3347 case Op_LoadD_unaligned:
3348 case Op_StoreB:
3349 case Op_StoreC:
3350 case Op_StoreI:
3351 case Op_StoreL:
3352 case Op_CompareAndSwapB:
3353 case Op_CompareAndSwapS:
3354 case Op_CompareAndSwapI:
3355 case Op_CompareAndSwapL:
3356 case Op_CompareAndSwapP:
3357 case Op_CompareAndSwapN:
3358 case Op_WeakCompareAndSwapB:
3359 case Op_WeakCompareAndSwapS:
3360 case Op_WeakCompareAndSwapI:
3361 case Op_WeakCompareAndSwapL:
3362 case Op_WeakCompareAndSwapP:
3363 case Op_WeakCompareAndSwapN:
3364 case Op_CompareAndExchangeB:
3365 case Op_CompareAndExchangeS:
3366 case Op_CompareAndExchangeI:
3367 case Op_CompareAndExchangeL:
3368 case Op_CompareAndExchangeP:
3369 case Op_CompareAndExchangeN:
3370 case Op_GetAndAddS:
3371 case Op_GetAndAddB:
3372 case Op_GetAndAddI:
3373 case Op_GetAndAddL:
3374 case Op_GetAndSetS:
3375 case Op_GetAndSetB:
3376 case Op_GetAndSetI:
3377 case Op_GetAndSetL:
3378 case Op_GetAndSetP:
3379 case Op_GetAndSetN:
3380 case Op_StoreP:
3381 case Op_StoreN:
3382 case Op_StoreNKlass:
3383 case Op_LoadB:
3384 case Op_LoadUB:
3385 case Op_LoadUS:
3386 case Op_LoadI:
3387 case Op_LoadKlass:
3388 case Op_LoadNKlass:
3389 case Op_LoadL:
3390 case Op_LoadL_unaligned:
3391 case Op_LoadP:
3392 case Op_LoadN:
3393 case Op_LoadRange:
3394 case Op_LoadS:
3395 case Op_LoadVectorGather:
3396 case Op_StoreVectorScatter:
3397 case Op_LoadVectorGatherMasked:
3398 case Op_StoreVectorScatterMasked:
3399 case Op_LoadVectorMasked:
3400 case Op_StoreVectorMasked:
3401 break;
3402
3403 case Op_AddP: { // Assert sane base pointers
3404 Node *addp = n->in(AddPNode::Address);
3405 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
3406 #ifdef _LP64
3407 if (addp->Opcode() == Op_ConP &&
3408 addp == n->in(AddPNode::Base) &&
3409 n->in(AddPNode::Offset)->is_Con()) {
3410 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3411 // on the platform and on the compressed oops mode.
3412 // Use addressing with narrow klass to load with offset on x86.
3413 // Some platforms can use the constant pool to load ConP.
3414 // Do this transformation here since IGVN will convert ConN back to ConP.
3415 const Type* t = addp->bottom_type();
3416 bool is_oop = t->isa_oopptr() != nullptr;
3417 bool is_klass = t->isa_klassptr() != nullptr;
3418
3419 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3420 (is_klass && Matcher::const_klass_prefer_decode() &&
3421 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3422 Node* nn = nullptr;
3423
3424 int op = is_oop ? Op_ConN : Op_ConNKlass;
3425
3426 // Look for existing ConN node of the same exact type.
3427 Node* r = root();
3428 uint cnt = r->outcnt();
3429 for (uint i = 0; i < cnt; i++) {
3430 Node* m = r->raw_out(i);
3431 if (m!= nullptr && m->Opcode() == op &&
3432 m->bottom_type()->make_ptr() == t) {
3433 nn = m;
3434 break;
3435 }
3436 }
3437 if (nn != nullptr) {
3438 // Decode a narrow oop to match address
3439 // [R12 + narrow_oop_reg<<3 + offset]
3440 if (is_oop) {
3441 nn = new DecodeNNode(nn, t);
3442 } else {
3443 nn = new DecodeNKlassNode(nn, t);
3444 }
3445 // Check for succeeding AddP which uses the same Base.
3446 // Otherwise we will run into the assertion above when visiting that guy.
3447 for (uint i = 0; i < n->outcnt(); ++i) {
3448 Node *out_i = n->raw_out(i);
3449 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3450 out_i->set_req(AddPNode::Base, nn);
3451 #ifdef ASSERT
3452 for (uint j = 0; j < out_i->outcnt(); ++j) {
3453 Node *out_j = out_i->raw_out(j);
3454 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3455 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3456 }
3457 #endif
3458 }
3459 }
3460 n->set_req(AddPNode::Base, nn);
3461 n->set_req(AddPNode::Address, nn);
3462 if (addp->outcnt() == 0) {
3463 addp->disconnect_inputs(this);
3464 }
3465 }
3466 }
3467 }
3468 #endif
3469 break;
3470 }
3471
3472 case Op_CastPP: {
3473 // Remove CastPP nodes to gain more freedom during scheduling but
3474 // keep the dependency they encode as control or precedence edges
3475 // (if control is set already) on memory operations. Some CastPP
3476 // nodes don't have a control (don't carry a dependency): skip
3477 // those.
3478 if (n->in(0) != nullptr) {
3479 ResourceMark rm;
3480 Unique_Node_List wq;
3481 wq.push(n);
3482 for (uint next = 0; next < wq.size(); ++next) {
3483 Node *m = wq.at(next);
3484 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3485 Node* use = m->fast_out(i);
3486 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3487 use->ensure_control_or_add_prec(n->in(0));
3488 } else {
3489 switch(use->Opcode()) {
3490 case Op_AddP:
3491 case Op_DecodeN:
3492 case Op_DecodeNKlass:
3493 case Op_CheckCastPP:
3494 case Op_CastPP:
3495 wq.push(use);
3496 break;
3497 }
3498 }
3499 }
3500 }
3501 }
3502 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3503 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3504 Node* in1 = n->in(1);
3505 const Type* t = n->bottom_type();
3506 Node* new_in1 = in1->clone();
3507 new_in1->as_DecodeN()->set_type(t);
3508
3509 if (!Matcher::narrow_oop_use_complex_address()) {
3510 //
3511 // x86, ARM and friends can handle 2 adds in addressing mode
3512 // and Matcher can fold a DecodeN node into address by using
3513 // a narrow oop directly and do implicit null check in address:
3514 //
3515 // [R12 + narrow_oop_reg<<3 + offset]
3516 // NullCheck narrow_oop_reg
3517 //
3518 // On other platforms (Sparc) we have to keep new DecodeN node and
3519 // use it to do implicit null check in address:
3520 //
3521 // decode_not_null narrow_oop_reg, base_reg
3522 // [base_reg + offset]
3523 // NullCheck base_reg
3524 //
3525 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3526 // to keep the information to which null check the new DecodeN node
3527 // corresponds to use it as value in implicit_null_check().
3528 //
3529 new_in1->set_req(0, n->in(0));
3530 }
3531
3532 n->subsume_by(new_in1, this);
3533 if (in1->outcnt() == 0) {
3534 in1->disconnect_inputs(this);
3535 }
3536 } else {
3537 n->subsume_by(n->in(1), this);
3538 if (n->outcnt() == 0) {
3539 n->disconnect_inputs(this);
3540 }
3541 }
3542 break;
3543 }
3544 case Op_CastII: {
3545 n->as_CastII()->remove_range_check_cast(this);
3546 break;
3547 }
3548 #ifdef _LP64
3549 case Op_CmpP:
3550 // Do this transformation here to preserve CmpPNode::sub() and
3551 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3552 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3553 Node* in1 = n->in(1);
3554 Node* in2 = n->in(2);
3555 if (!in1->is_DecodeNarrowPtr()) {
3556 in2 = in1;
3557 in1 = n->in(2);
3558 }
3559 assert(in1->is_DecodeNarrowPtr(), "sanity");
3560
3561 Node* new_in2 = nullptr;
3562 if (in2->is_DecodeNarrowPtr()) {
3563 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3564 new_in2 = in2->in(1);
3565 } else if (in2->Opcode() == Op_ConP) {
3566 const Type* t = in2->bottom_type();
3567 if (t == TypePtr::NULL_PTR) {
3568 assert(in1->is_DecodeN(), "compare klass to null?");
3569 // Don't convert CmpP null check into CmpN if compressed
3570 // oops implicit null check is not generated.
3571 // This will allow to generate normal oop implicit null check.
3572 if (Matcher::gen_narrow_oop_implicit_null_checks())
3573 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3574 //
3575 // This transformation together with CastPP transformation above
3576 // will generated code for implicit null checks for compressed oops.
3577 //
3578 // The original code after Optimize()
3579 //
3580 // LoadN memory, narrow_oop_reg
3581 // decode narrow_oop_reg, base_reg
3582 // CmpP base_reg, nullptr
3583 // CastPP base_reg // NotNull
3584 // Load [base_reg + offset], val_reg
3585 //
3586 // after these transformations will be
3587 //
3588 // LoadN memory, narrow_oop_reg
3589 // CmpN narrow_oop_reg, nullptr
3590 // decode_not_null narrow_oop_reg, base_reg
3591 // Load [base_reg + offset], val_reg
3592 //
3593 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3594 // since narrow oops can be used in debug info now (see the code in
3595 // final_graph_reshaping_walk()).
3596 //
3597 // At the end the code will be matched to
3598 // on x86:
3599 //
3600 // Load_narrow_oop memory, narrow_oop_reg
3601 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3602 // NullCheck narrow_oop_reg
3603 //
3604 // and on sparc:
3605 //
3606 // Load_narrow_oop memory, narrow_oop_reg
3607 // decode_not_null narrow_oop_reg, base_reg
3608 // Load [base_reg + offset], val_reg
3609 // NullCheck base_reg
3610 //
3611 } else if (t->isa_oopptr()) {
3612 new_in2 = ConNode::make(t->make_narrowoop());
3613 } else if (t->isa_klassptr()) {
3614 ciKlass* klass = t->is_klassptr()->exact_klass();
3615 if (klass->is_in_encoding_range()) {
3616 new_in2 = ConNode::make(t->make_narrowklass());
3617 }
3618 }
3619 }
3620 if (new_in2 != nullptr) {
3621 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3622 n->subsume_by(cmpN, this);
3623 if (in1->outcnt() == 0) {
3624 in1->disconnect_inputs(this);
3625 }
3626 if (in2->outcnt() == 0) {
3627 in2->disconnect_inputs(this);
3628 }
3629 }
3630 }
3631 break;
3632
3633 case Op_DecodeN:
3634 case Op_DecodeNKlass:
3635 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3636 // DecodeN could be pinned when it can't be fold into
3637 // an address expression, see the code for Op_CastPP above.
3638 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3639 break;
3640
3641 case Op_EncodeP:
3642 case Op_EncodePKlass: {
3643 Node* in1 = n->in(1);
3644 if (in1->is_DecodeNarrowPtr()) {
3645 n->subsume_by(in1->in(1), this);
3646 } else if (in1->Opcode() == Op_ConP) {
3647 const Type* t = in1->bottom_type();
3648 if (t == TypePtr::NULL_PTR) {
3649 assert(t->isa_oopptr(), "null klass?");
3650 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3651 } else if (t->isa_oopptr()) {
3652 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3653 } else if (t->isa_klassptr()) {
3654 ciKlass* klass = t->is_klassptr()->exact_klass();
3655 if (klass->is_in_encoding_range()) {
3656 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3657 } else {
3658 assert(false, "unencodable klass in ConP -> EncodeP");
3659 C->record_failure("unencodable klass in ConP -> EncodeP");
3660 }
3661 }
3662 }
3663 if (in1->outcnt() == 0) {
3664 in1->disconnect_inputs(this);
3665 }
3666 break;
3667 }
3668
3669 case Op_Proj: {
3670 if (OptimizeStringConcat || IncrementalInline) {
3671 ProjNode* proj = n->as_Proj();
3672 if (proj->_is_io_use) {
3673 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3674 // Separate projections were used for the exception path which
3675 // are normally removed by a late inline. If it wasn't inlined
3676 // then they will hang around and should just be replaced with
3677 // the original one. Merge them.
3678 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3679 if (non_io_proj != nullptr) {
3680 proj->subsume_by(non_io_proj , this);
3681 }
3682 }
3683 }
3684 break;
3685 }
3686
3687 case Op_Phi:
3688 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3689 // The EncodeP optimization may create Phi with the same edges
3690 // for all paths. It is not handled well by Register Allocator.
3691 Node* unique_in = n->in(1);
3692 assert(unique_in != nullptr, "");
3693 uint cnt = n->req();
3694 for (uint i = 2; i < cnt; i++) {
3695 Node* m = n->in(i);
3696 assert(m != nullptr, "");
3697 if (unique_in != m)
3698 unique_in = nullptr;
3699 }
3700 if (unique_in != nullptr) {
3701 n->subsume_by(unique_in, this);
3702 }
3703 }
3704 break;
3705
3706 #endif
3707
3708 case Op_ModI:
3709 handle_div_mod_op(n, T_INT, false);
3710 break;
3711
3712 case Op_ModL:
3713 handle_div_mod_op(n, T_LONG, false);
3714 break;
3715
3716 case Op_UModI:
3717 handle_div_mod_op(n, T_INT, true);
3718 break;
3719
3720 case Op_UModL:
3721 handle_div_mod_op(n, T_LONG, true);
3722 break;
3723
3724 case Op_LoadVector:
3725 case Op_StoreVector:
3726 #ifdef ASSERT
3727 // Add VerifyVectorAlignment node between adr and load / store.
3728 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3729 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3730 n->as_StoreVector()->must_verify_alignment();
3731 if (must_verify_alignment) {
3732 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3733 n->as_StoreVector()->memory_size();
3734 // The memory access should be aligned to the vector width in bytes.
3735 // However, the underlying array is possibly less well aligned, but at least
3736 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3737 // a loop we can expect at least the following alignment:
3738 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3739 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3740 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3741 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3742 jlong mask = guaranteed_alignment - 1;
3743 Node* mask_con = ConLNode::make(mask);
3744 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3745 n->set_req(MemNode::Address, va);
3746 }
3747 }
3748 #endif
3749 break;
3750
3751 case Op_PackB:
3752 case Op_PackS:
3753 case Op_PackI:
3754 case Op_PackF:
3755 case Op_PackL:
3756 case Op_PackD:
3757 if (n->req()-1 > 2) {
3758 // Replace many operand PackNodes with a binary tree for matching
3759 PackNode* p = (PackNode*) n;
3760 Node* btp = p->binary_tree_pack(1, n->req());
3761 n->subsume_by(btp, this);
3762 }
3763 break;
3764 case Op_Loop:
3765 // When StressCountedLoop is enabled, this loop may intentionally avoid a counted loop conversion.
3766 // This is expected behavior for the stress mode, which exercises alternative compilation paths.
3767 if (!StressCountedLoop) {
3768 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3769 }
3770 case Op_CountedLoop:
3771 case Op_LongCountedLoop:
3772 case Op_OuterStripMinedLoop:
3773 if (n->as_Loop()->is_inner_loop()) {
3774 frc.inc_inner_loop_count();
3775 }
3776 n->as_Loop()->verify_strip_mined(0);
3777 break;
3778 case Op_LShiftI:
3779 case Op_RShiftI:
3780 case Op_URShiftI:
3781 case Op_LShiftL:
3782 case Op_RShiftL:
3783 case Op_URShiftL:
3784 if (Matcher::need_masked_shift_count) {
3785 // The cpu's shift instructions don't restrict the count to the
3786 // lower 5/6 bits. We need to do the masking ourselves.
3787 Node* in2 = n->in(2);
3788 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3789 const TypeInt* t = in2->find_int_type();
3790 if (t != nullptr && t->is_con()) {
3791 juint shift = t->get_con();
3792 if (shift > mask) { // Unsigned cmp
3793 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3794 }
3795 } else {
3796 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3797 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3798 n->set_req(2, shift);
3799 }
3800 }
3801 if (in2->outcnt() == 0) { // Remove dead node
3802 in2->disconnect_inputs(this);
3803 }
3804 }
3805 break;
3806 case Op_MemBarStoreStore:
3807 case Op_MemBarRelease:
3808 // Break the link with AllocateNode: it is no longer useful and
3809 // confuses register allocation.
3810 if (n->req() > MemBarNode::Precedent) {
3811 n->set_req(MemBarNode::Precedent, top());
3812 }
3813 break;
3814 case Op_MemBarAcquire: {
3815 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3816 // At parse time, the trailing MemBarAcquire for a volatile load
3817 // is created with an edge to the load. After optimizations,
3818 // that input may be a chain of Phis. If those phis have no
3819 // other use, then the MemBarAcquire keeps them alive and
3820 // register allocation can be confused.
3821 dead_nodes.push(n->in(MemBarNode::Precedent));
3822 n->set_req(MemBarNode::Precedent, top());
3823 }
3824 break;
3825 }
3826 case Op_RangeCheck: {
3827 RangeCheckNode* rc = n->as_RangeCheck();
3828 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3829 n->subsume_by(iff, this);
3830 frc._tests.push(iff);
3831 break;
3832 }
3833 case Op_ConvI2L: {
3834 if (!Matcher::convi2l_type_required) {
3835 // Code generation on some platforms doesn't need accurate
3836 // ConvI2L types. Widening the type can help remove redundant
3837 // address computations.
3838 n->as_Type()->set_type(TypeLong::INT);
3839 ResourceMark rm;
3840 Unique_Node_List wq;
3841 wq.push(n);
3842 for (uint next = 0; next < wq.size(); next++) {
3843 Node *m = wq.at(next);
3844
3845 for(;;) {
3846 // Loop over all nodes with identical inputs edges as m
3847 Node* k = m->find_similar(m->Opcode());
3848 if (k == nullptr) {
3849 break;
3850 }
3851 // Push their uses so we get a chance to remove node made
3852 // redundant
3853 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3854 Node* u = k->fast_out(i);
3855 if (u->Opcode() == Op_LShiftL ||
3856 u->Opcode() == Op_AddL ||
3857 u->Opcode() == Op_SubL ||
3858 u->Opcode() == Op_AddP) {
3859 wq.push(u);
3860 }
3861 }
3862 // Replace all nodes with identical edges as m with m
3863 k->subsume_by(m, this);
3864 }
3865 }
3866 }
3867 break;
3868 }
3869 case Op_CmpUL: {
3870 if (!Matcher::has_match_rule(Op_CmpUL)) {
3871 // No support for unsigned long comparisons
3872 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3873 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3874 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3875 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3876 Node* andl = new AndLNode(orl, remove_sign_mask);
3877 Node* cmp = new CmpLNode(andl, n->in(2));
3878 n->subsume_by(cmp, this);
3879 }
3880 break;
3881 }
3882 #ifdef ASSERT
3883 case Op_ConNKlass: {
3884 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3885 ciKlass* klass = tp->is_klassptr()->exact_klass();
3886 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3887 break;
3888 }
3889 #endif
3890 default:
3891 assert(!n->is_Call(), "");
3892 assert(!n->is_Mem(), "");
3893 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3894 break;
3895 }
3896 }
3897
3898 //------------------------------final_graph_reshaping_walk---------------------
3899 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3900 // requires that the walk visits a node's inputs before visiting the node.
3901 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3902 Unique_Node_List sfpt;
3903
3904 frc._visited.set(root->_idx); // first, mark node as visited
3905 uint cnt = root->req();
3906 Node *n = root;
3907 uint i = 0;
3908 while (true) {
3909 if (i < cnt) {
3910 // Place all non-visited non-null inputs onto stack
3911 Node* m = n->in(i);
3912 ++i;
3913 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3914 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3915 // compute worst case interpreter size in case of a deoptimization
3916 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3917
3918 sfpt.push(m);
3919 }
3920 cnt = m->req();
3921 nstack.push(n, i); // put on stack parent and next input's index
3922 n = m;
3923 i = 0;
3924 }
3925 } else {
3926 // Now do post-visit work
3927 final_graph_reshaping_impl(n, frc, dead_nodes);
3928 if (nstack.is_empty())
3929 break; // finished
3930 n = nstack.node(); // Get node from stack
3931 cnt = n->req();
3932 i = nstack.index();
3933 nstack.pop(); // Shift to the next node on stack
3934 }
3935 }
3936
3937 expand_reachability_edges(sfpt);
3938
3939 // Skip next transformation if compressed oops are not used.
3940 if (UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks())
3941 return;
3942
3943 // Go over ReachabilityFence nodes to skip DecodeN nodes for referents.
3944 // The sole purpose of RF node is to keep the referent oop alive and
3945 // decoding the oop for that is not needed.
3946 for (int i = 0; i < C->reachability_fences_count(); i++) {
3947 ReachabilityFenceNode* rf = C->reachability_fence(i);
3948 DecodeNNode* dn = rf->in(1)->isa_DecodeN();
3949 if (dn != nullptr) {
3950 if (!dn->has_non_debug_uses() || Matcher::narrow_oop_use_complex_address()) {
3951 rf->set_req(1, dn->in(1));
3952 if (dn->outcnt() == 0) {
3953 dn->disconnect_inputs(this);
3954 }
3955 }
3956 }
3957 }
3958
3959 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3960 // It could be done for an uncommon traps or any safepoints/calls
3961 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3962 while (sfpt.size() > 0) {
3963 n = sfpt.pop();
3964 JVMState *jvms = n->as_SafePoint()->jvms();
3965 assert(jvms != nullptr, "sanity");
3966 int start = jvms->debug_start();
3967 int end = n->req();
3968 bool is_uncommon = (n->is_CallStaticJava() &&
3969 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3970 for (int j = start; j < end; j++) {
3971 Node* in = n->in(j);
3972 if (in->is_DecodeNarrowPtr() && (is_uncommon || !in->has_non_debug_uses())) {
3973 n->set_req(j, in->in(1));
3974 if (in->outcnt() == 0) {
3975 in->disconnect_inputs(this);
3976 }
3977 }
3978 }
3979 }
3980 }
3981
3982 //------------------------------final_graph_reshaping--------------------------
3983 // Final Graph Reshaping.
3984 //
3985 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3986 // and not commoned up and forced early. Must come after regular
3987 // optimizations to avoid GVN undoing the cloning. Clone constant
3988 // inputs to Loop Phis; these will be split by the allocator anyways.
3989 // Remove Opaque nodes.
3990 // (2) Move last-uses by commutative operations to the left input to encourage
3991 // Intel update-in-place two-address operations and better register usage
3992 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3993 // calls canonicalizing them back.
3994 // (3) Detect infinite loops; blobs of code reachable from above but not
3995 // below. Several of the Code_Gen algorithms fail on such code shapes,
3996 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3997 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3998 // Detection is by looking for IfNodes where only 1 projection is
3999 // reachable from below or CatchNodes missing some targets.
4000 // (4) Assert for insane oop offsets in debug mode.
4001
4002 bool Compile::final_graph_reshaping() {
4003 // an infinite loop may have been eliminated by the optimizer,
4004 // in which case the graph will be empty.
4005 if (root()->req() == 1) {
4006 // Do not compile method that is only a trivial infinite loop,
4007 // since the content of the loop may have been eliminated.
4008 record_method_not_compilable("trivial infinite loop");
4009 return true;
4010 }
4011
4012 // Expensive nodes have their control input set to prevent the GVN
4013 // from freely commoning them. There's no GVN beyond this point so
4014 // no need to keep the control input. We want the expensive nodes to
4015 // be freely moved to the least frequent code path by gcm.
4016 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4017 for (int i = 0; i < expensive_count(); i++) {
4018 _expensive_nodes.at(i)->set_req(0, nullptr);
4019 }
4020
4021 Final_Reshape_Counts frc;
4022
4023 // Visit everybody reachable!
4024 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4025 Node_Stack nstack(live_nodes() >> 1);
4026 Unique_Node_List dead_nodes;
4027 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4028
4029 // Check for unreachable (from below) code (i.e., infinite loops).
4030 for( uint i = 0; i < frc._tests.size(); i++ ) {
4031 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4032 // Get number of CFG targets.
4033 // Note that PCTables include exception targets after calls.
4034 uint required_outcnt = n->required_outcnt();
4035 if (n->outcnt() != required_outcnt) {
4036 // Check for a few special cases. Rethrow Nodes never take the
4037 // 'fall-thru' path, so expected kids is 1 less.
4038 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4039 if (n->in(0)->in(0)->is_Call()) {
4040 CallNode* call = n->in(0)->in(0)->as_Call();
4041 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4042 required_outcnt--; // Rethrow always has 1 less kid
4043 } else if (call->req() > TypeFunc::Parms &&
4044 call->is_CallDynamicJava()) {
4045 // Check for null receiver. In such case, the optimizer has
4046 // detected that the virtual call will always result in a null
4047 // pointer exception. The fall-through projection of this CatchNode
4048 // will not be populated.
4049 Node* arg0 = call->in(TypeFunc::Parms);
4050 if (arg0->is_Type() &&
4051 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4052 required_outcnt--;
4053 }
4054 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4055 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4056 // Check for illegal array length. In such case, the optimizer has
4057 // detected that the allocation attempt will always result in an
4058 // exception. There is no fall-through projection of this CatchNode .
4059 assert(call->is_CallStaticJava(), "static call expected");
4060 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4061 uint valid_length_test_input = call->req() - 1;
4062 Node* valid_length_test = call->in(valid_length_test_input);
4063 call->del_req(valid_length_test_input);
4064 if (valid_length_test->find_int_con(1) == 0) {
4065 required_outcnt--;
4066 }
4067 dead_nodes.push(valid_length_test);
4068 assert(n->outcnt() == required_outcnt, "malformed control flow");
4069 continue;
4070 }
4071 }
4072 }
4073
4074 // Recheck with a better notion of 'required_outcnt'
4075 if (n->outcnt() != required_outcnt) {
4076 record_method_not_compilable("malformed control flow");
4077 return true; // Not all targets reachable!
4078 }
4079 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4080 CallNode* call = n->in(0)->in(0)->as_Call();
4081 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4082 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4083 assert(call->is_CallStaticJava(), "static call expected");
4084 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4085 uint valid_length_test_input = call->req() - 1;
4086 dead_nodes.push(call->in(valid_length_test_input));
4087 call->del_req(valid_length_test_input); // valid length test useless now
4088 }
4089 }
4090 // Check that I actually visited all kids. Unreached kids
4091 // must be infinite loops.
4092 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4093 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4094 record_method_not_compilable("infinite loop");
4095 return true; // Found unvisited kid; must be unreach
4096 }
4097
4098 // Here so verification code in final_graph_reshaping_walk()
4099 // always see an OuterStripMinedLoopEnd
4100 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4101 IfNode* init_iff = n->as_If();
4102 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4103 n->subsume_by(iff, this);
4104 }
4105 }
4106
4107 while (dead_nodes.size() > 0) {
4108 Node* m = dead_nodes.pop();
4109 if (m->outcnt() == 0 && m != top()) {
4110 for (uint j = 0; j < m->req(); j++) {
4111 Node* in = m->in(j);
4112 if (in != nullptr) {
4113 dead_nodes.push(in);
4114 }
4115 }
4116 m->disconnect_inputs(this);
4117 }
4118 }
4119
4120 set_java_calls(frc.get_java_call_count());
4121 set_inner_loops(frc.get_inner_loop_count());
4122
4123 // No infinite loops, no reason to bail out.
4124 return false;
4125 }
4126
4127 //-----------------------------too_many_traps----------------------------------
4128 // Report if there are too many traps at the current method and bci.
4129 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4130 bool Compile::too_many_traps(ciMethod* method,
4131 int bci,
4132 Deoptimization::DeoptReason reason) {
4133 ciMethodData* md = method->method_data();
4134 if (md->is_empty()) {
4135 // Assume the trap has not occurred, or that it occurred only
4136 // because of a transient condition during start-up in the interpreter.
4137 return false;
4138 }
4139 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4140 if (md->has_trap_at(bci, m, reason) != 0) {
4141 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4142 // Also, if there are multiple reasons, or if there is no per-BCI record,
4143 // assume the worst.
4144 if (log())
4145 log()->elem("observe trap='%s' count='%d'",
4146 Deoptimization::trap_reason_name(reason),
4147 md->trap_count(reason));
4148 return true;
4149 } else {
4150 // Ignore method/bci and see if there have been too many globally.
4151 return too_many_traps(reason, md);
4152 }
4153 }
4154
4155 // Less-accurate variant which does not require a method and bci.
4156 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4157 ciMethodData* logmd) {
4158 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4159 // Too many traps globally.
4160 // Note that we use cumulative trap_count, not just md->trap_count.
4161 if (log()) {
4162 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4163 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4164 Deoptimization::trap_reason_name(reason),
4165 mcount, trap_count(reason));
4166 }
4167 return true;
4168 } else {
4169 // The coast is clear.
4170 return false;
4171 }
4172 }
4173
4174 //--------------------------too_many_recompiles--------------------------------
4175 // Report if there are too many recompiles at the current method and bci.
4176 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4177 // Is not eager to return true, since this will cause the compiler to use
4178 // Action_none for a trap point, to avoid too many recompilations.
4179 bool Compile::too_many_recompiles(ciMethod* method,
4180 int bci,
4181 Deoptimization::DeoptReason reason) {
4182 ciMethodData* md = method->method_data();
4183 if (md->is_empty()) {
4184 // Assume the trap has not occurred, or that it occurred only
4185 // because of a transient condition during start-up in the interpreter.
4186 return false;
4187 }
4188 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4189 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4190 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4191 Deoptimization::DeoptReason per_bc_reason
4192 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4193 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4194 if ((per_bc_reason == Deoptimization::Reason_none
4195 || md->has_trap_at(bci, m, reason) != 0)
4196 // The trap frequency measure we care about is the recompile count:
4197 && md->trap_recompiled_at(bci, m)
4198 && md->overflow_recompile_count() >= bc_cutoff) {
4199 // Do not emit a trap here if it has already caused recompilations.
4200 // Also, if there are multiple reasons, or if there is no per-BCI record,
4201 // assume the worst.
4202 if (log())
4203 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4204 Deoptimization::trap_reason_name(reason),
4205 md->trap_count(reason),
4206 md->overflow_recompile_count());
4207 return true;
4208 } else if (trap_count(reason) != 0
4209 && decompile_count() >= m_cutoff) {
4210 // Too many recompiles globally, and we have seen this sort of trap.
4211 // Use cumulative decompile_count, not just md->decompile_count.
4212 if (log())
4213 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4214 Deoptimization::trap_reason_name(reason),
4215 md->trap_count(reason), trap_count(reason),
4216 md->decompile_count(), decompile_count());
4217 return true;
4218 } else {
4219 // The coast is clear.
4220 return false;
4221 }
4222 }
4223
4224 // Compute when not to trap. Used by matching trap based nodes and
4225 // NullCheck optimization.
4226 void Compile::set_allowed_deopt_reasons() {
4227 _allowed_reasons = 0;
4228 if (is_method_compilation()) {
4229 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4230 assert(rs < BitsPerInt, "recode bit map");
4231 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4232 _allowed_reasons |= nth_bit(rs);
4233 }
4234 }
4235 }
4236 }
4237
4238 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4239 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4240 }
4241
4242 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4243 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4244 }
4245
4246 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4247 if (holder->is_initialized()) {
4248 return false;
4249 }
4250 if (holder->is_being_initialized()) {
4251 if (accessing_method->holder() == holder) {
4252 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4253 // <init>, or a static method. In all those cases, there was an initialization
4254 // barrier on the holder klass passed.
4255 if (accessing_method->is_static_initializer() ||
4256 accessing_method->is_object_initializer() ||
4257 accessing_method->is_static()) {
4258 return false;
4259 }
4260 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4261 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4262 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4263 // child class can become fully initialized while its parent class is still being initialized.
4264 if (accessing_method->is_static_initializer()) {
4265 return false;
4266 }
4267 }
4268 ciMethod* root = method(); // the root method of compilation
4269 if (root != accessing_method) {
4270 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4271 }
4272 }
4273 return true;
4274 }
4275
4276 #ifndef PRODUCT
4277 //------------------------------verify_bidirectional_edges---------------------
4278 // For each input edge to a node (ie - for each Use-Def edge), verify that
4279 // there is a corresponding Def-Use edge.
4280 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4281 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4282 uint stack_size = live_nodes() >> 4;
4283 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4284 if (root_and_safepoints != nullptr) {
4285 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4286 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4287 Node* root_or_safepoint = root_and_safepoints->at(i);
4288 // If the node is a safepoint, let's check if it still has a control input
4289 // Lack of control input signifies that this node was killed by CCP or
4290 // recursively by remove_globally_dead_node and it shouldn't be a starting
4291 // point.
4292 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4293 nstack.push(root_or_safepoint);
4294 }
4295 }
4296 } else {
4297 nstack.push(_root);
4298 }
4299
4300 while (nstack.size() > 0) {
4301 Node* n = nstack.pop();
4302 if (visited.member(n)) {
4303 continue;
4304 }
4305 visited.push(n);
4306
4307 // Walk over all input edges, checking for correspondence
4308 uint length = n->len();
4309 for (uint i = 0; i < length; i++) {
4310 Node* in = n->in(i);
4311 if (in != nullptr && !visited.member(in)) {
4312 nstack.push(in); // Put it on stack
4313 }
4314 if (in != nullptr && !in->is_top()) {
4315 // Count instances of `next`
4316 int cnt = 0;
4317 for (uint idx = 0; idx < in->_outcnt; idx++) {
4318 if (in->_out[idx] == n) {
4319 cnt++;
4320 }
4321 }
4322 assert(cnt > 0, "Failed to find Def-Use edge.");
4323 // Check for duplicate edges
4324 // walk the input array downcounting the input edges to n
4325 for (uint j = 0; j < length; j++) {
4326 if (n->in(j) == in) {
4327 cnt--;
4328 }
4329 }
4330 assert(cnt == 0, "Mismatched edge count.");
4331 } else if (in == nullptr) {
4332 assert(i == 0 || i >= n->req() ||
4333 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4334 (n->is_Unlock() && i == (n->req() - 1)) ||
4335 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4336 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4337 } else {
4338 assert(in->is_top(), "sanity");
4339 // Nothing to check.
4340 }
4341 }
4342 }
4343 }
4344
4345 //------------------------------verify_graph_edges---------------------------
4346 // Walk the Graph and verify that there is a one-to-one correspondence
4347 // between Use-Def edges and Def-Use edges in the graph.
4348 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4349 if (VerifyGraphEdges) {
4350 Unique_Node_List visited;
4351
4352 // Call graph walk to check edges
4353 verify_bidirectional_edges(visited, root_and_safepoints);
4354 if (no_dead_code) {
4355 // Now make sure that no visited node is used by an unvisited node.
4356 bool dead_nodes = false;
4357 Unique_Node_List checked;
4358 while (visited.size() > 0) {
4359 Node* n = visited.pop();
4360 checked.push(n);
4361 for (uint i = 0; i < n->outcnt(); i++) {
4362 Node* use = n->raw_out(i);
4363 if (checked.member(use)) continue; // already checked
4364 if (visited.member(use)) continue; // already in the graph
4365 if (use->is_Con()) continue; // a dead ConNode is OK
4366 // At this point, we have found a dead node which is DU-reachable.
4367 if (!dead_nodes) {
4368 tty->print_cr("*** Dead nodes reachable via DU edges:");
4369 dead_nodes = true;
4370 }
4371 use->dump(2);
4372 tty->print_cr("---");
4373 checked.push(use); // No repeats; pretend it is now checked.
4374 }
4375 }
4376 assert(!dead_nodes, "using nodes must be reachable from root");
4377 }
4378 }
4379 }
4380 #endif
4381
4382 // The Compile object keeps track of failure reasons separately from the ciEnv.
4383 // This is required because there is not quite a 1-1 relation between the
4384 // ciEnv and its compilation task and the Compile object. Note that one
4385 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4386 // to backtrack and retry without subsuming loads. Other than this backtracking
4387 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4388 // by the logic in C2Compiler.
4389 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4390 if (log() != nullptr) {
4391 log()->elem("failure reason='%s' phase='compile'", reason);
4392 }
4393 if (_failure_reason.get() == nullptr) {
4394 // Record the first failure reason.
4395 _failure_reason.set(reason);
4396 if (CaptureBailoutInformation) {
4397 _first_failure_details = new CompilationFailureInfo(reason);
4398 }
4399 } else {
4400 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4401 }
4402
4403 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4404 C->print_method(PHASE_FAILURE, 1);
4405 }
4406 _root = nullptr; // flush the graph, too
4407 }
4408
4409 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4410 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4411 _compile(Compile::current()),
4412 _log(nullptr),
4413 _dolog(CITimeVerbose)
4414 {
4415 assert(_compile != nullptr, "sanity check");
4416 assert(id != PhaseTraceId::_t_none, "Don't use none");
4417 if (_dolog) {
4418 _log = _compile->log();
4419 }
4420 if (_log != nullptr) {
4421 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4422 _log->stamp();
4423 _log->end_head();
4424 }
4425
4426 // Inform memory statistic, if enabled
4427 if (CompilationMemoryStatistic::enabled()) {
4428 CompilationMemoryStatistic::on_phase_start((int)id, name);
4429 }
4430 }
4431
4432 Compile::TracePhase::TracePhase(PhaseTraceId id)
4433 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4434
4435 Compile::TracePhase::~TracePhase() {
4436
4437 // Inform memory statistic, if enabled
4438 if (CompilationMemoryStatistic::enabled()) {
4439 CompilationMemoryStatistic::on_phase_end();
4440 }
4441
4442 if (_compile->failing_internal()) {
4443 if (_log != nullptr) {
4444 _log->done("phase");
4445 }
4446 return; // timing code, not stressing bailouts.
4447 }
4448 #ifdef ASSERT
4449 if (PrintIdealNodeCount) {
4450 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4451 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4452 }
4453
4454 if (VerifyIdealNodeCount) {
4455 _compile->print_missing_nodes();
4456 }
4457 #endif
4458
4459 if (_log != nullptr) {
4460 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4461 }
4462 }
4463
4464 //----------------------------static_subtype_check-----------------------------
4465 // Shortcut important common cases when superklass is exact:
4466 // (0) superklass is java.lang.Object (can occur in reflective code)
4467 // (1) subklass is already limited to a subtype of superklass => always ok
4468 // (2) subklass does not overlap with superklass => always fail
4469 // (3) superklass has NO subtypes and we can check with a simple compare.
4470 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4471 if (skip) {
4472 return SSC_full_test; // Let caller generate the general case.
4473 }
4474
4475 if (subk->is_java_subtype_of(superk)) {
4476 return SSC_always_true; // (0) and (1) this test cannot fail
4477 }
4478
4479 if (!subk->maybe_java_subtype_of(superk)) {
4480 return SSC_always_false; // (2) true path dead; no dynamic test needed
4481 }
4482
4483 const Type* superelem = superk;
4484 if (superk->isa_aryklassptr()) {
4485 int ignored;
4486 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4487 }
4488
4489 if (superelem->isa_instklassptr()) {
4490 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4491 if (!ik->has_subklass()) {
4492 if (!ik->is_final()) {
4493 // Add a dependency if there is a chance of a later subclass.
4494 dependencies()->assert_leaf_type(ik);
4495 }
4496 if (!superk->maybe_java_subtype_of(subk)) {
4497 return SSC_always_false;
4498 }
4499 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4500 }
4501 } else {
4502 // A primitive array type has no subtypes.
4503 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4504 }
4505
4506 return SSC_full_test;
4507 }
4508
4509 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4510 #ifdef _LP64
4511 // The scaled index operand to AddP must be a clean 64-bit value.
4512 // Java allows a 32-bit int to be incremented to a negative
4513 // value, which appears in a 64-bit register as a large
4514 // positive number. Using that large positive number as an
4515 // operand in pointer arithmetic has bad consequences.
4516 // On the other hand, 32-bit overflow is rare, and the possibility
4517 // can often be excluded, if we annotate the ConvI2L node with
4518 // a type assertion that its value is known to be a small positive
4519 // number. (The prior range check has ensured this.)
4520 // This assertion is used by ConvI2LNode::Ideal.
4521 int index_max = max_jint - 1; // array size is max_jint, index is one less
4522 if (sizetype != nullptr && sizetype->_hi > 0) {
4523 index_max = sizetype->_hi - 1;
4524 }
4525 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4526 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4527 #endif
4528 return idx;
4529 }
4530
4531 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4532 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4533 if (ctrl != nullptr) {
4534 // Express control dependency by a CastII node with a narrow type.
4535 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4536 // node from floating above the range check during loop optimizations. Otherwise, the
4537 // ConvI2L node may be eliminated independently of the range check, causing the data path
4538 // to become TOP while the control path is still there (although it's unreachable).
4539 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
4540 value = phase->transform(value);
4541 }
4542 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4543 return phase->transform(new ConvI2LNode(value, ltype));
4544 }
4545
4546 void Compile::dump_print_inlining() {
4547 inline_printer()->print_on(tty);
4548 }
4549
4550 void Compile::log_late_inline(CallGenerator* cg) {
4551 if (log() != nullptr) {
4552 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4553 cg->unique_id());
4554 JVMState* p = cg->call_node()->jvms();
4555 while (p != nullptr) {
4556 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4557 p = p->caller();
4558 }
4559 log()->tail("late_inline");
4560 }
4561 }
4562
4563 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4564 log_late_inline(cg);
4565 if (log() != nullptr) {
4566 log()->inline_fail(msg);
4567 }
4568 }
4569
4570 void Compile::log_inline_id(CallGenerator* cg) {
4571 if (log() != nullptr) {
4572 // The LogCompilation tool needs a unique way to identify late
4573 // inline call sites. This id must be unique for this call site in
4574 // this compilation. Try to have it unique across compilations as
4575 // well because it can be convenient when grepping through the log
4576 // file.
4577 // Distinguish OSR compilations from others in case CICountOSR is
4578 // on.
4579 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4580 cg->set_unique_id(id);
4581 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4582 }
4583 }
4584
4585 void Compile::log_inline_failure(const char* msg) {
4586 if (C->log() != nullptr) {
4587 C->log()->inline_fail(msg);
4588 }
4589 }
4590
4591
4592 // Dump inlining replay data to the stream.
4593 // Don't change thread state and acquire any locks.
4594 void Compile::dump_inline_data(outputStream* out) {
4595 InlineTree* inl_tree = ilt();
4596 if (inl_tree != nullptr) {
4597 out->print(" inline %d", inl_tree->count());
4598 inl_tree->dump_replay_data(out);
4599 }
4600 }
4601
4602 void Compile::dump_inline_data_reduced(outputStream* out) {
4603 assert(ReplayReduce, "");
4604
4605 InlineTree* inl_tree = ilt();
4606 if (inl_tree == nullptr) {
4607 return;
4608 }
4609 // Enable iterative replay file reduction
4610 // Output "compile" lines for depth 1 subtrees,
4611 // simulating that those trees were compiled
4612 // instead of inlined.
4613 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4614 InlineTree* sub = inl_tree->subtrees().at(i);
4615 if (sub->inline_level() != 1) {
4616 continue;
4617 }
4618
4619 ciMethod* method = sub->method();
4620 int entry_bci = -1;
4621 int comp_level = env()->task()->comp_level();
4622 out->print("compile ");
4623 method->dump_name_as_ascii(out);
4624 out->print(" %d %d", entry_bci, comp_level);
4625 out->print(" inline %d", sub->count());
4626 sub->dump_replay_data(out, -1);
4627 out->cr();
4628 }
4629 }
4630
4631 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4632 if (n1->Opcode() < n2->Opcode()) return -1;
4633 else if (n1->Opcode() > n2->Opcode()) return 1;
4634
4635 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4636 for (uint i = 1; i < n1->req(); i++) {
4637 if (n1->in(i) < n2->in(i)) return -1;
4638 else if (n1->in(i) > n2->in(i)) return 1;
4639 }
4640
4641 return 0;
4642 }
4643
4644 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4645 Node* n1 = *n1p;
4646 Node* n2 = *n2p;
4647
4648 return cmp_expensive_nodes(n1, n2);
4649 }
4650
4651 void Compile::sort_expensive_nodes() {
4652 if (!expensive_nodes_sorted()) {
4653 _expensive_nodes.sort(cmp_expensive_nodes);
4654 }
4655 }
4656
4657 bool Compile::expensive_nodes_sorted() const {
4658 for (int i = 1; i < _expensive_nodes.length(); i++) {
4659 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4660 return false;
4661 }
4662 }
4663 return true;
4664 }
4665
4666 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4667 if (_expensive_nodes.length() == 0) {
4668 return false;
4669 }
4670
4671 assert(OptimizeExpensiveOps, "optimization off?");
4672
4673 // Take this opportunity to remove dead nodes from the list
4674 int j = 0;
4675 for (int i = 0; i < _expensive_nodes.length(); i++) {
4676 Node* n = _expensive_nodes.at(i);
4677 if (!n->is_unreachable(igvn)) {
4678 assert(n->is_expensive(), "should be expensive");
4679 _expensive_nodes.at_put(j, n);
4680 j++;
4681 }
4682 }
4683 _expensive_nodes.trunc_to(j);
4684
4685 // Then sort the list so that similar nodes are next to each other
4686 // and check for at least two nodes of identical kind with same data
4687 // inputs.
4688 sort_expensive_nodes();
4689
4690 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4691 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4692 return true;
4693 }
4694 }
4695
4696 return false;
4697 }
4698
4699 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4700 if (_expensive_nodes.length() == 0) {
4701 return;
4702 }
4703
4704 assert(OptimizeExpensiveOps, "optimization off?");
4705
4706 // Sort to bring similar nodes next to each other and clear the
4707 // control input of nodes for which there's only a single copy.
4708 sort_expensive_nodes();
4709
4710 int j = 0;
4711 int identical = 0;
4712 int i = 0;
4713 bool modified = false;
4714 for (; i < _expensive_nodes.length()-1; i++) {
4715 assert(j <= i, "can't write beyond current index");
4716 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4717 identical++;
4718 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4719 continue;
4720 }
4721 if (identical > 0) {
4722 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4723 identical = 0;
4724 } else {
4725 Node* n = _expensive_nodes.at(i);
4726 igvn.replace_input_of(n, 0, nullptr);
4727 igvn.hash_insert(n);
4728 modified = true;
4729 }
4730 }
4731 if (identical > 0) {
4732 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4733 } else if (_expensive_nodes.length() >= 1) {
4734 Node* n = _expensive_nodes.at(i);
4735 igvn.replace_input_of(n, 0, nullptr);
4736 igvn.hash_insert(n);
4737 modified = true;
4738 }
4739 _expensive_nodes.trunc_to(j);
4740 if (modified) {
4741 igvn.optimize();
4742 }
4743 }
4744
4745 void Compile::add_expensive_node(Node * n) {
4746 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4747 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4748 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4749 if (OptimizeExpensiveOps) {
4750 _expensive_nodes.append(n);
4751 } else {
4752 // Clear control input and let IGVN optimize expensive nodes if
4753 // OptimizeExpensiveOps is off.
4754 n->set_req(0, nullptr);
4755 }
4756 }
4757
4758 /**
4759 * Track coarsened Lock and Unlock nodes.
4760 */
4761
4762 class Lock_List : public Node_List {
4763 uint _origin_cnt;
4764 public:
4765 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4766 uint origin_cnt() const { return _origin_cnt; }
4767 };
4768
4769 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4770 int length = locks.length();
4771 if (length > 0) {
4772 // Have to keep this list until locks elimination during Macro nodes elimination.
4773 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4774 AbstractLockNode* alock = locks.at(0);
4775 BoxLockNode* box = alock->box_node()->as_BoxLock();
4776 for (int i = 0; i < length; i++) {
4777 AbstractLockNode* lock = locks.at(i);
4778 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4779 locks_list->push(lock);
4780 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4781 if (this_box != box) {
4782 // Locking regions (BoxLock) could be Unbalanced here:
4783 // - its coarsened locks were eliminated in earlier
4784 // macro nodes elimination followed by loop unroll
4785 // - it is OSR locking region (no Lock node)
4786 // Preserve Unbalanced status in such cases.
4787 if (!this_box->is_unbalanced()) {
4788 this_box->set_coarsened();
4789 }
4790 if (!box->is_unbalanced()) {
4791 box->set_coarsened();
4792 }
4793 }
4794 }
4795 _coarsened_locks.append(locks_list);
4796 }
4797 }
4798
4799 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4800 int count = coarsened_count();
4801 for (int i = 0; i < count; i++) {
4802 Node_List* locks_list = _coarsened_locks.at(i);
4803 for (uint j = 0; j < locks_list->size(); j++) {
4804 Node* lock = locks_list->at(j);
4805 assert(lock->is_AbstractLock(), "sanity");
4806 if (!useful.member(lock)) {
4807 locks_list->yank(lock);
4808 }
4809 }
4810 }
4811 }
4812
4813 void Compile::remove_coarsened_lock(Node* n) {
4814 if (n->is_AbstractLock()) {
4815 int count = coarsened_count();
4816 for (int i = 0; i < count; i++) {
4817 Node_List* locks_list = _coarsened_locks.at(i);
4818 locks_list->yank(n);
4819 }
4820 }
4821 }
4822
4823 bool Compile::coarsened_locks_consistent() {
4824 int count = coarsened_count();
4825 for (int i = 0; i < count; i++) {
4826 bool unbalanced = false;
4827 bool modified = false; // track locks kind modifications
4828 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4829 uint size = locks_list->size();
4830 if (size == 0) {
4831 unbalanced = false; // All locks were eliminated - good
4832 } else if (size != locks_list->origin_cnt()) {
4833 unbalanced = true; // Some locks were removed from list
4834 } else {
4835 for (uint j = 0; j < size; j++) {
4836 Node* lock = locks_list->at(j);
4837 // All nodes in group should have the same state (modified or not)
4838 if (!lock->as_AbstractLock()->is_coarsened()) {
4839 if (j == 0) {
4840 // first on list was modified, the rest should be too for consistency
4841 modified = true;
4842 } else if (!modified) {
4843 // this lock was modified but previous locks on the list were not
4844 unbalanced = true;
4845 break;
4846 }
4847 } else if (modified) {
4848 // previous locks on list were modified but not this lock
4849 unbalanced = true;
4850 break;
4851 }
4852 }
4853 }
4854 if (unbalanced) {
4855 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4856 #ifdef ASSERT
4857 if (PrintEliminateLocks) {
4858 tty->print_cr("=== unbalanced coarsened locks ===");
4859 for (uint l = 0; l < size; l++) {
4860 locks_list->at(l)->dump();
4861 }
4862 }
4863 #endif
4864 record_failure(C2Compiler::retry_no_locks_coarsening());
4865 return false;
4866 }
4867 }
4868 return true;
4869 }
4870
4871 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4872 // locks coarsening optimization removed Lock/Unlock nodes from them.
4873 // Such regions become unbalanced because coarsening only removes part
4874 // of Lock/Unlock nodes in region. As result we can't execute other
4875 // locks elimination optimizations which assume all code paths have
4876 // corresponding pair of Lock/Unlock nodes - they are balanced.
4877 void Compile::mark_unbalanced_boxes() const {
4878 int count = coarsened_count();
4879 for (int i = 0; i < count; i++) {
4880 Node_List* locks_list = _coarsened_locks.at(i);
4881 uint size = locks_list->size();
4882 if (size > 0) {
4883 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4884 BoxLockNode* box = alock->box_node()->as_BoxLock();
4885 if (alock->is_coarsened()) {
4886 // coarsened_locks_consistent(), which is called before this method, verifies
4887 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4888 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4889 for (uint j = 1; j < size; j++) {
4890 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4891 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4892 if (box != this_box) {
4893 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4894 box->set_unbalanced();
4895 this_box->set_unbalanced();
4896 }
4897 }
4898 }
4899 }
4900 }
4901 }
4902
4903 /**
4904 * Remove the speculative part of types and clean up the graph
4905 */
4906 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4907 if (UseTypeSpeculation) {
4908 Unique_Node_List worklist;
4909 worklist.push(root());
4910 int modified = 0;
4911 // Go over all type nodes that carry a speculative type, drop the
4912 // speculative part of the type and enqueue the node for an igvn
4913 // which may optimize it out.
4914 for (uint next = 0; next < worklist.size(); ++next) {
4915 Node *n = worklist.at(next);
4916 if (n->is_Type()) {
4917 TypeNode* tn = n->as_Type();
4918 const Type* t = tn->type();
4919 const Type* t_no_spec = t->remove_speculative();
4920 if (t_no_spec != t) {
4921 bool in_hash = igvn.hash_delete(n);
4922 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4923 tn->set_type(t_no_spec);
4924 igvn.hash_insert(n);
4925 igvn._worklist.push(n); // give it a chance to go away
4926 modified++;
4927 }
4928 }
4929 // Iterate over outs - endless loops is unreachable from below
4930 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4931 Node *m = n->fast_out(i);
4932 if (not_a_node(m)) {
4933 continue;
4934 }
4935 worklist.push(m);
4936 }
4937 }
4938 // Drop the speculative part of all types in the igvn's type table
4939 igvn.remove_speculative_types();
4940 if (modified > 0) {
4941 igvn.optimize();
4942 if (failing()) return;
4943 }
4944 #ifdef ASSERT
4945 // Verify that after the IGVN is over no speculative type has resurfaced
4946 worklist.clear();
4947 worklist.push(root());
4948 for (uint next = 0; next < worklist.size(); ++next) {
4949 Node *n = worklist.at(next);
4950 const Type* t = igvn.type_or_null(n);
4951 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4952 if (n->is_Type()) {
4953 t = n->as_Type()->type();
4954 assert(t == t->remove_speculative(), "no more speculative types");
4955 }
4956 // Iterate over outs - endless loops is unreachable from below
4957 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4958 Node *m = n->fast_out(i);
4959 if (not_a_node(m)) {
4960 continue;
4961 }
4962 worklist.push(m);
4963 }
4964 }
4965 igvn.check_no_speculative_types();
4966 #endif
4967 }
4968 }
4969
4970 // Auxiliary methods to support randomized stressing/fuzzing.
4971
4972 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
4973 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
4974 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
4975 FLAG_SET_ERGO(StressSeed, _stress_seed);
4976 } else {
4977 _stress_seed = StressSeed;
4978 }
4979 if (_log != nullptr) {
4980 _log->elem("stress_test seed='%u'", _stress_seed);
4981 }
4982 }
4983
4984 int Compile::random() {
4985 _stress_seed = os::next_random(_stress_seed);
4986 return static_cast<int>(_stress_seed);
4987 }
4988
4989 // This method can be called the arbitrary number of times, with current count
4990 // as the argument. The logic allows selecting a single candidate from the
4991 // running list of candidates as follows:
4992 // int count = 0;
4993 // Cand* selected = null;
4994 // while(cand = cand->next()) {
4995 // if (randomized_select(++count)) {
4996 // selected = cand;
4997 // }
4998 // }
4999 //
5000 // Including count equalizes the chances any candidate is "selected".
5001 // This is useful when we don't have the complete list of candidates to choose
5002 // from uniformly. In this case, we need to adjust the randomicity of the
5003 // selection, or else we will end up biasing the selection towards the latter
5004 // candidates.
5005 //
5006 // Quick back-envelope calculation shows that for the list of n candidates
5007 // the equal probability for the candidate to persist as "best" can be
5008 // achieved by replacing it with "next" k-th candidate with the probability
5009 // of 1/k. It can be easily shown that by the end of the run, the
5010 // probability for any candidate is converged to 1/n, thus giving the
5011 // uniform distribution among all the candidates.
5012 //
5013 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5014 #define RANDOMIZED_DOMAIN_POW 29
5015 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5016 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5017 bool Compile::randomized_select(int count) {
5018 assert(count > 0, "only positive");
5019 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5020 }
5021
5022 #ifdef ASSERT
5023 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5024 bool Compile::fail_randomly() {
5025 if ((random() % StressBailoutMean) != 0) {
5026 return false;
5027 }
5028 record_failure("StressBailout");
5029 return true;
5030 }
5031
5032 bool Compile::failure_is_artificial() {
5033 return C->failure_reason_is("StressBailout");
5034 }
5035 #endif
5036
5037 CloneMap& Compile::clone_map() { return _clone_map; }
5038 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5039
5040 void NodeCloneInfo::dump_on(outputStream* st) const {
5041 st->print(" {%d:%d} ", idx(), gen());
5042 }
5043
5044 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5045 uint64_t val = value(old->_idx);
5046 NodeCloneInfo cio(val);
5047 assert(val != 0, "old node should be in the map");
5048 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5049 insert(nnn->_idx, cin.get());
5050 #ifndef PRODUCT
5051 if (is_debug()) {
5052 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5053 }
5054 #endif
5055 }
5056
5057 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5058 NodeCloneInfo cio(value(old->_idx));
5059 if (cio.get() == 0) {
5060 cio.set(old->_idx, 0);
5061 insert(old->_idx, cio.get());
5062 #ifndef PRODUCT
5063 if (is_debug()) {
5064 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5065 }
5066 #endif
5067 }
5068 clone(old, nnn, gen);
5069 }
5070
5071 int CloneMap::max_gen() const {
5072 int g = 0;
5073 DictI di(_dict);
5074 for(; di.test(); ++di) {
5075 int t = gen(di._key);
5076 if (g < t) {
5077 g = t;
5078 #ifndef PRODUCT
5079 if (is_debug()) {
5080 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5081 }
5082 #endif
5083 }
5084 }
5085 return g;
5086 }
5087
5088 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5089 uint64_t val = value(key);
5090 if (val != 0) {
5091 NodeCloneInfo ni(val);
5092 ni.dump_on(st);
5093 }
5094 }
5095
5096 void Compile::shuffle_macro_nodes() {
5097 shuffle_array(*C, _macro_nodes);
5098 }
5099
5100 // Move Allocate nodes to the start of the list
5101 void Compile::sort_macro_nodes() {
5102 int count = macro_count();
5103 int allocates = 0;
5104 for (int i = 0; i < count; i++) {
5105 Node* n = macro_node(i);
5106 if (n->is_Allocate()) {
5107 if (i != allocates) {
5108 Node* tmp = macro_node(allocates);
5109 _macro_nodes.at_put(allocates, n);
5110 _macro_nodes.at_put(i, tmp);
5111 }
5112 allocates++;
5113 }
5114 }
5115 }
5116
5117 void Compile::print_method(CompilerPhaseType compile_phase, int level, Node* n) {
5118 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5119 EventCompilerPhase event(UNTIMED);
5120 if (event.should_commit()) {
5121 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, compile_phase, C->_compile_id, level);
5122 }
5123 #ifndef PRODUCT
5124 ResourceMark rm;
5125 stringStream ss;
5126 ss.print_raw(CompilerPhaseTypeHelper::to_description(compile_phase));
5127 int iter = ++_igv_phase_iter[compile_phase];
5128 if (iter > 1) {
5129 ss.print(" %d", iter);
5130 }
5131 if (n != nullptr) {
5132 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5133 if (n->is_Call()) {
5134 CallNode* call = n->as_Call();
5135 if (call->_name != nullptr) {
5136 // E.g. uncommon traps etc.
5137 ss.print(" - %s", call->_name);
5138 } else if (call->is_CallJava()) {
5139 CallJavaNode* call_java = call->as_CallJava();
5140 if (call_java->method() != nullptr) {
5141 ss.print(" -");
5142 call_java->method()->print_short_name(&ss);
5143 }
5144 }
5145 }
5146 }
5147
5148 const char* name = ss.as_string();
5149 if (should_print_igv(level)) {
5150 _igv_printer->print_graph(name);
5151 }
5152 if (should_print_phase(level)) {
5153 print_phase(name);
5154 }
5155 if (should_print_ideal_phase(compile_phase)) {
5156 print_ideal_ir(CompilerPhaseTypeHelper::to_name(compile_phase));
5157 }
5158 #endif
5159 C->_latest_stage_start_counter.stamp();
5160 }
5161
5162 // Only used from CompileWrapper
5163 void Compile::begin_method() {
5164 #ifndef PRODUCT
5165 if (_method != nullptr && should_print_igv(1)) {
5166 _igv_printer->begin_method();
5167 }
5168 #endif
5169 C->_latest_stage_start_counter.stamp();
5170 }
5171
5172 // Only used from CompileWrapper
5173 void Compile::end_method() {
5174 EventCompilerPhase event(UNTIMED);
5175 if (event.should_commit()) {
5176 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5177 }
5178
5179 #ifndef PRODUCT
5180 if (_method != nullptr && should_print_igv(1)) {
5181 _igv_printer->end_method();
5182 }
5183 #endif
5184 }
5185
5186 #ifndef PRODUCT
5187 bool Compile::should_print_phase(const int level) const {
5188 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5189 _method != nullptr; // Do not print phases for stubs.
5190 }
5191
5192 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5193 return _directive->should_print_ideal_phase(cpt);
5194 }
5195
5196 void Compile::init_igv() {
5197 if (_igv_printer == nullptr) {
5198 _igv_printer = IdealGraphPrinter::printer();
5199 _igv_printer->set_compile(this);
5200 }
5201 }
5202
5203 bool Compile::should_print_igv(const int level) {
5204 PRODUCT_RETURN_(return false;);
5205
5206 if (PrintIdealGraphLevel < 0) { // disabled by the user
5207 return false;
5208 }
5209
5210 bool need = directive()->IGVPrintLevelOption >= level;
5211 if (need) {
5212 Compile::init_igv();
5213 }
5214 return need;
5215 }
5216
5217 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5218 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5219
5220 // Called from debugger. Prints method to the default file with the default phase name.
5221 // This works regardless of any Ideal Graph Visualizer flags set or not.
5222 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5223 void igv_print(void* sp, void* fp, void* pc) {
5224 frame fr(sp, fp, pc);
5225 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5226 }
5227
5228 // Same as igv_print() above but with a specified phase name.
5229 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5230 frame fr(sp, fp, pc);
5231 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5232 }
5233
5234 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5235 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5236 // This works regardless of any Ideal Graph Visualizer flags set or not.
5237 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5238 void igv_print(bool network, void* sp, void* fp, void* pc) {
5239 frame fr(sp, fp, pc);
5240 if (network) {
5241 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5242 } else {
5243 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5244 }
5245 }
5246
5247 // Same as igv_print(bool network, ...) above but with a specified phase name.
5248 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5249 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5250 frame fr(sp, fp, pc);
5251 if (network) {
5252 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5253 } else {
5254 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5255 }
5256 }
5257
5258 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5259 void igv_print_default() {
5260 Compile::current()->print_method(PHASE_DEBUG, 0);
5261 }
5262
5263 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5264 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5265 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5266 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5267 void igv_append(void* sp, void* fp, void* pc) {
5268 frame fr(sp, fp, pc);
5269 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5270 }
5271
5272 // Same as igv_append(...) above but with a specified phase name.
5273 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5274 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5275 frame fr(sp, fp, pc);
5276 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5277 }
5278
5279 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5280 const char* file_name = "custom_debug.xml";
5281 if (_debug_file_printer == nullptr) {
5282 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5283 } else {
5284 _debug_file_printer->update_compiled_method(C->method());
5285 }
5286 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5287 _debug_file_printer->print_graph(phase_name, fr);
5288 }
5289
5290 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5291 ResourceMark rm;
5292 GrowableArray<const Node*> empty_list;
5293 igv_print_graph_to_network(phase_name, empty_list, fr);
5294 }
5295
5296 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5297 if (_debug_network_printer == nullptr) {
5298 _debug_network_printer = new IdealGraphPrinter(C);
5299 } else {
5300 _debug_network_printer->update_compiled_method(C->method());
5301 }
5302 tty->print_cr("Method printed over network stream to IGV");
5303 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5304 }
5305 #endif // !PRODUCT
5306
5307 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5308 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5309 return value;
5310 }
5311 Node* result = nullptr;
5312 if (bt == T_BYTE) {
5313 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5314 result = new RShiftINode(result, phase->intcon(24));
5315 } else if (bt == T_BOOLEAN) {
5316 result = new AndINode(value, phase->intcon(0xFF));
5317 } else if (bt == T_CHAR) {
5318 result = new AndINode(value,phase->intcon(0xFFFF));
5319 } else {
5320 assert(bt == T_SHORT, "unexpected narrow type");
5321 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5322 result = new RShiftINode(result, phase->intcon(16));
5323 }
5324 if (transform_res) {
5325 result = phase->transform(result);
5326 }
5327 return result;
5328 }
5329
5330 void Compile::record_method_not_compilable_oom() {
5331 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5332 }
5333
5334 #ifndef PRODUCT
5335 // Collects all the control inputs from nodes on the worklist and from their data dependencies
5336 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
5337 // Follow non-control edges until we reach CFG nodes
5338 for (uint i = 0; i < worklist.size(); i++) {
5339 const Node* n = worklist.at(i);
5340 for (uint j = 0; j < n->req(); j++) {
5341 Node* in = n->in(j);
5342 if (in == nullptr || in->is_Root()) {
5343 continue;
5344 }
5345 if (in->is_CFG()) {
5346 if (in->is_Call()) {
5347 // The return value of a call is only available if the call did not result in an exception
5348 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
5349 if (control_proj_use->is_Catch()) {
5350 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
5351 candidates.push(fall_through);
5352 continue;
5353 }
5354 }
5355
5356 if (in->is_Multi()) {
5357 // We got here by following data inputs so we should only have one control use
5358 // (no IfNode, etc)
5359 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
5360 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
5361 } else {
5362 candidates.push(in);
5363 }
5364 } else {
5365 worklist.push(in);
5366 }
5367 }
5368 }
5369 }
5370
5371 // Returns the candidate node that is a descendant to all the other candidates
5372 static Node* pick_control(Unique_Node_List& candidates) {
5373 Unique_Node_List worklist;
5374 worklist.copy(candidates);
5375
5376 // Traverse backwards through the CFG
5377 for (uint i = 0; i < worklist.size(); i++) {
5378 const Node* n = worklist.at(i);
5379 if (n->is_Root()) {
5380 continue;
5381 }
5382 for (uint j = 0; j < n->req(); j++) {
5383 // Skip backedge of loops to avoid cycles
5384 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
5385 continue;
5386 }
5387
5388 Node* pred = n->in(j);
5389 if (pred != nullptr && pred != n && pred->is_CFG()) {
5390 worklist.push(pred);
5391 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
5392 candidates.remove(pred);
5393 }
5394 }
5395 }
5396
5397 assert(candidates.size() == 1, "unexpected control flow");
5398 return candidates.at(0);
5399 }
5400
5401 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
5402 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
5403 call->init_req((*pos)++, parm);
5404 const BasicType bt = parm->bottom_type()->basic_type();
5405 if (bt == T_LONG || bt == T_DOUBLE) {
5406 call->init_req((*pos)++, half);
5407 }
5408 }
5409
5410 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
5411 Node* parm0, Node* parm1,
5412 Node* parm2, Node* parm3,
5413 Node* parm4, Node* parm5,
5414 Node* parm6) const {
5415 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
5416 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
5417 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
5418
5419 // find the most suitable control input
5420 Unique_Node_List worklist, candidates;
5421 if (parm0 != nullptr) { worklist.push(parm0);
5422 if (parm1 != nullptr) { worklist.push(parm1);
5423 if (parm2 != nullptr) { worklist.push(parm2);
5424 if (parm3 != nullptr) { worklist.push(parm3);
5425 if (parm4 != nullptr) { worklist.push(parm4);
5426 if (parm5 != nullptr) { worklist.push(parm5);
5427 if (parm6 != nullptr) { worklist.push(parm6);
5428 /* close each nested if ===> */ } } } } } } }
5429 find_candidate_control_inputs(worklist, candidates);
5430 Node* control = nullptr;
5431 if (candidates.size() == 0) {
5432 control = C->start()->proj_out(TypeFunc::Control);
5433 } else {
5434 control = pick_control(candidates);
5435 }
5436
5437 // find all the previous users of the control we picked
5438 GrowableArray<Node*> users_of_control;
5439 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
5440 Node* use = control->fast_out(i);
5441 if (use->is_CFG() && use != control) {
5442 users_of_control.push(use);
5443 }
5444 }
5445
5446 // we do not actually care about IO and memory as it uses neither
5447 call->init_req(TypeFunc::Control, control);
5448 call->init_req(TypeFunc::I_O, top());
5449 call->init_req(TypeFunc::Memory, top());
5450 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
5451 call->init_req(TypeFunc::ReturnAdr, top());
5452
5453 int pos = TypeFunc::Parms;
5454 call->init_req(pos++, str_node);
5455 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
5456 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
5457 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
5458 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
5459 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
5460 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
5461 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
5462 /* close each nested if ===> */ } } } } } } }
5463 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
5464
5465 call = gvn->transform(call);
5466 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
5467
5468 // rewire previous users to have the new call as control instead
5469 PhaseIterGVN* igvn = gvn->is_IterGVN();
5470 for (int i = 0; i < users_of_control.length(); i++) {
5471 Node* use = users_of_control.at(i);
5472 for (uint j = 0; j < use->req(); j++) {
5473 if (use->in(j) == control) {
5474 if (igvn != nullptr) {
5475 igvn->replace_input_of(use, j, call_control_proj);
5476 } else {
5477 gvn->hash_delete(use);
5478 use->set_req(j, call_control_proj);
5479 gvn->hash_insert(use);
5480 }
5481 }
5482 }
5483 }
5484
5485 return call;
5486 }
5487 #endif // !PRODUCT