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