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