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 _stress_seed(0),
649 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
650 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
651 _env(ci_env),
652 _directive(directive),
653 _log(ci_env->log()),
654 _first_failure_details(nullptr),
655 _intrinsics(comp_arena(), 0, 0, nullptr),
656 _macro_nodes(comp_arena(), 8, 0, nullptr),
657 _parse_predicates(comp_arena(), 8, 0, nullptr),
658 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
659 _expensive_nodes(comp_arena(), 8, 0, nullptr),
660 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
661 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
662 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
663 _coarsened_locks(comp_arena(), 8, 0, nullptr),
664 _congraph(nullptr),
665 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
666 _unique(0),
667 _dead_node_count(0),
668 _dead_node_list(comp_arena()),
669 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
670 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
671 _node_arena(&_node_arena_one),
672 _mach_constant_base_node(nullptr),
673 _Compile_types(mtCompiler, Arena::Tag::tag_type),
674 _initial_gvn(nullptr),
675 _igvn_worklist(nullptr),
676 _types(nullptr),
677 _node_hash(nullptr),
678 _late_inlines(comp_arena(), 2, 0, nullptr),
679 _string_late_inlines(comp_arena(), 2, 0, nullptr),
680 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
681 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
682 _late_inlines_pos(0),
683 _number_of_mh_late_inlines(0),
684 _oom(false),
685 _replay_inline_data(nullptr),
686 _inline_printer(this),
687 _java_calls(0),
688 _inner_loops(0),
689 _interpreter_frame_size(0),
690 _output(nullptr)
691 #ifndef PRODUCT
692 ,
693 _in_dump_cnt(0)
694 #endif
695 {
696 C = this;
697 CompileWrapper cw(this);
698
699 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
700 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
701
702 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
703 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
704 // We can always print a disassembly, either abstract (hex dump) or
705 // with the help of a suitable hsdis library. Thus, we should not
706 // couple print_assembly and print_opto_assembly controls.
707 // But: always print opto and regular assembly on compile command 'print'.
708 bool print_assembly = directive->PrintAssemblyOption;
709 set_print_assembly(print_opto_assembly || print_assembly);
710 #else
711 set_print_assembly(false); // must initialize.
712 #endif
713
714 #ifndef PRODUCT
715 set_parsed_irreducible_loop(false);
716 #endif
717
718 if (directive->ReplayInlineOption) {
719 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
720 }
721 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
722 set_print_intrinsics(directive->PrintIntrinsicsOption);
723 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
724
725 if (ProfileTraps) {
726 // Make sure the method being compiled gets its own MDO,
727 // so we can at least track the decompile_count().
728 method()->ensure_method_data();
729 }
730
731 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
732 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout) {
733 initialize_stress_seed(directive);
734 }
735
736 Init(/*do_aliasing=*/ true);
737
738 print_compile_messages();
739
740 _ilt = InlineTree::build_inline_tree_root();
741
742 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
743 assert(num_alias_types() >= AliasIdxRaw, "");
744
745 #define MINIMUM_NODE_HASH 1023
746
747 // GVN that will be run immediately on new nodes
748 uint estimated_size = method()->code_size()*4+64;
749 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
750 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
751 _types = new (comp_arena()) Type_Array(comp_arena());
752 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
753 PhaseGVN gvn;
754 set_initial_gvn(&gvn);
755
756 { // Scope for timing the parser
757 TracePhase tp(_t_parser);
758
759 // Put top into the hash table ASAP.
760 initial_gvn()->transform(top());
761
762 // Set up tf(), start(), and find a CallGenerator.
763 CallGenerator* cg = nullptr;
764 if (is_osr_compilation()) {
765 const TypeTuple *domain = StartOSRNode::osr_domain();
766 const TypeTuple *range = TypeTuple::make_range(method()->signature());
767 init_tf(TypeFunc::make(domain, range));
768 StartNode* s = new StartOSRNode(root(), domain);
769 initial_gvn()->set_type_bottom(s);
770 verify_start(s);
771 cg = CallGenerator::for_osr(method(), entry_bci());
772 } else {
773 // Normal case.
774 init_tf(TypeFunc::make(method()));
775 StartNode* s = new StartNode(root(), tf()->domain());
776 initial_gvn()->set_type_bottom(s);
777 verify_start(s);
778 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
779 // With java.lang.ref.reference.get() we must go through the
780 // intrinsic - even when get() is the root
781 // method of the compile - so that, if necessary, the value in
782 // the referent field of the reference object gets recorded by
783 // the pre-barrier code.
784 cg = find_intrinsic(method(), false);
785 }
786 if (cg == nullptr) {
787 float past_uses = method()->interpreter_invocation_count();
788 float expected_uses = past_uses;
789 cg = CallGenerator::for_inline(method(), expected_uses);
790 }
791 }
792 if (failing()) return;
793 if (cg == nullptr) {
794 const char* reason = InlineTree::check_can_parse(method());
795 assert(reason != nullptr, "expect reason for parse failure");
796 stringStream ss;
797 ss.print("cannot parse method: %s", reason);
798 record_method_not_compilable(ss.as_string());
799 return;
800 }
801
802 gvn.set_type(root(), root()->bottom_type());
803
804 JVMState* jvms = build_start_state(start(), tf());
805 if ((jvms = cg->generate(jvms)) == nullptr) {
806 assert(failure_reason() != nullptr, "expect reason for parse failure");
807 stringStream ss;
808 ss.print("method parse failed: %s", failure_reason());
809 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
810 return;
811 }
812 GraphKit kit(jvms);
813
814 if (!kit.stopped()) {
815 // Accept return values, and transfer control we know not where.
816 // This is done by a special, unique ReturnNode bound to root.
817 return_values(kit.jvms());
818 }
819
820 if (kit.has_exceptions()) {
821 // Any exceptions that escape from this call must be rethrown
822 // to whatever caller is dynamically above us on the stack.
823 // This is done by a special, unique RethrowNode bound to root.
824 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
825 }
826
827 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
828
829 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
830 inline_string_calls(true);
831 }
832
833 if (failing()) return;
834
835 // Remove clutter produced by parsing.
836 if (!failing()) {
837 ResourceMark rm;
838 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
839 }
840 }
841
842 // Note: Large methods are capped off in do_one_bytecode().
843 if (failing()) return;
844
845 // After parsing, node notes are no longer automagic.
846 // They must be propagated by register_new_node_with_optimizer(),
847 // clone(), or the like.
848 set_default_node_notes(nullptr);
849
850 #ifndef PRODUCT
851 if (should_print_igv(1)) {
852 _igv_printer->print_inlining();
853 }
854 #endif
855
856 if (failing()) return;
857 NOT_PRODUCT( verify_graph_edges(); )
858
859 // Now optimize
860 Optimize();
861 if (failing()) return;
862 NOT_PRODUCT( verify_graph_edges(); )
863
864 #ifndef PRODUCT
865 if (should_print_ideal()) {
866 print_ideal_ir("print_ideal");
867 }
868 #endif
869
870 #ifdef ASSERT
871 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
872 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
873 #endif
874
875 // Dump compilation data to replay it.
876 if (directive->DumpReplayOption) {
877 env()->dump_replay_data(_compile_id);
878 }
879 if (directive->DumpInlineOption && (ilt() != nullptr)) {
880 env()->dump_inline_data(_compile_id);
881 }
882
883 // Now that we know the size of all the monitors we can add a fixed slot
884 // for the original deopt pc.
885 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
886 set_fixed_slots(next_slot);
887
888 // Compute when to use implicit null checks. Used by matching trap based
889 // nodes and NullCheck optimization.
890 set_allowed_deopt_reasons();
891
892 // Now generate code
893 Code_Gen();
894 }
895
896 //------------------------------Compile----------------------------------------
897 // Compile a runtime stub
898 Compile::Compile(ciEnv* ci_env,
899 TypeFunc_generator generator,
900 address stub_function,
901 const char* stub_name,
902 int is_fancy_jump,
903 bool pass_tls,
904 bool return_pc,
905 DirectiveSet* directive)
906 : Phase(Compiler),
907 _compile_id(0),
908 _options(Options::for_runtime_stub()),
909 _method(nullptr),
910 _entry_bci(InvocationEntryBci),
911 _stub_function(stub_function),
912 _stub_name(stub_name),
913 _stub_entry_point(nullptr),
914 _max_node_limit(MaxNodeLimit),
915 _post_loop_opts_phase(false),
916 _merge_stores_phase(false),
917 _allow_macro_nodes(true),
918 _inlining_progress(false),
919 _inlining_incrementally(false),
920 _has_reserved_stack_access(false),
921 #ifndef PRODUCT
922 _igv_idx(0),
923 _trace_opto_output(directive->TraceOptoOutputOption),
924 #endif
925 _has_method_handle_invokes(false),
926 _clinit_barrier_on_entry(false),
927 _stress_seed(0),
928 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
929 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
930 _env(ci_env),
931 _directive(directive),
932 _log(ci_env->log()),
933 _first_failure_details(nullptr),
934 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
935 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
936 _congraph(nullptr),
937 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
938 _unique(0),
939 _dead_node_count(0),
940 _dead_node_list(comp_arena()),
941 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
942 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
943 _node_arena(&_node_arena_one),
944 _mach_constant_base_node(nullptr),
945 _Compile_types(mtCompiler, Arena::Tag::tag_type),
946 _initial_gvn(nullptr),
947 _igvn_worklist(nullptr),
948 _types(nullptr),
949 _node_hash(nullptr),
950 _number_of_mh_late_inlines(0),
951 _oom(false),
952 _replay_inline_data(nullptr),
953 _inline_printer(this),
954 _java_calls(0),
955 _inner_loops(0),
956 _interpreter_frame_size(0),
957 _output(nullptr),
958 #ifndef PRODUCT
959 _in_dump_cnt(0),
960 #endif
961 _allowed_reasons(0) {
962 C = this;
963
964 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
965 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
966
967 #ifndef PRODUCT
968 set_print_assembly(PrintFrameConverterAssembly);
969 set_parsed_irreducible_loop(false);
970 #else
971 set_print_assembly(false); // Must initialize.
972 #endif
973 set_has_irreducible_loop(false); // no loops
974
975 CompileWrapper cw(this);
976 Init(/*do_aliasing=*/ false);
977 init_tf((*generator)());
978
979 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
980 _types = new (comp_arena()) Type_Array(comp_arena());
981 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
982
983 if (StressLCM || StressGCM || StressBailout) {
984 initialize_stress_seed(directive);
985 }
986
987 {
988 PhaseGVN gvn;
989 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
990 gvn.transform(top());
991
992 GraphKit kit;
993 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
994 }
995
996 NOT_PRODUCT( verify_graph_edges(); )
997
998 Code_Gen();
999 }
1000
1001 Compile::~Compile() {
1002 delete _first_failure_details;
1003 };
1004
1005 //------------------------------Init-------------------------------------------
1006 // Prepare for a single compilation
1007 void Compile::Init(bool aliasing) {
1008 _do_aliasing = aliasing;
1009 _unique = 0;
1010 _regalloc = nullptr;
1011
1012 _tf = nullptr; // filled in later
1013 _top = nullptr; // cached later
1014 _matcher = nullptr; // filled in later
1015 _cfg = nullptr; // filled in later
1016
1017 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
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() && method()->needs_clinit_barrier()) {
1088 set_clinit_barrier_on_entry(true);
1089 }
1090 if (debug_info()->recording_non_safepoints()) {
1091 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1092 (comp_arena(), 8, 0, nullptr));
1093 set_default_node_notes(Node_Notes::make(this));
1094 }
1095
1096 const int grow_ats = 16;
1097 _max_alias_types = grow_ats;
1098 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1099 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1100 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1101 {
1102 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1103 }
1104 // Initialize the first few types.
1105 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1106 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1107 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1108 _num_alias_types = AliasIdxRaw+1;
1109 // Zero out the alias type cache.
1110 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1111 // A null adr_type hits in the cache right away. Preload the right answer.
1112 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1113 }
1114
1115 #ifdef ASSERT
1116 // Verify that the current StartNode is valid.
1117 void Compile::verify_start(StartNode* s) const {
1118 assert(failing_internal() || s == start(), "should be StartNode");
1119 }
1120 #endif
1121
1122 /**
1123 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1124 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1125 * the ideal graph.
1126 */
1127 StartNode* Compile::start() const {
1128 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1129 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1130 Node* start = root()->fast_out(i);
1131 if (start->is_Start()) {
1132 return start->as_Start();
1133 }
1134 }
1135 fatal("Did not find Start node!");
1136 return nullptr;
1137 }
1138
1139 //-------------------------------immutable_memory-------------------------------------
1140 // Access immutable memory
1141 Node* Compile::immutable_memory() {
1142 if (_immutable_memory != nullptr) {
1143 return _immutable_memory;
1144 }
1145 StartNode* s = start();
1146 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1147 Node *p = s->fast_out(i);
1148 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1149 _immutable_memory = p;
1150 return _immutable_memory;
1151 }
1152 }
1153 ShouldNotReachHere();
1154 return nullptr;
1155 }
1156
1157 //----------------------set_cached_top_node------------------------------------
1158 // Install the cached top node, and make sure Node::is_top works correctly.
1159 void Compile::set_cached_top_node(Node* tn) {
1160 if (tn != nullptr) verify_top(tn);
1161 Node* old_top = _top;
1162 _top = tn;
1163 // Calling Node::setup_is_top allows the nodes the chance to adjust
1164 // their _out arrays.
1165 if (_top != nullptr) _top->setup_is_top();
1166 if (old_top != nullptr) old_top->setup_is_top();
1167 assert(_top == nullptr || top()->is_top(), "");
1168 }
1169
1170 #ifdef ASSERT
1171 uint Compile::count_live_nodes_by_graph_walk() {
1172 Unique_Node_List useful(comp_arena());
1173 // Get useful node list by walking the graph.
1174 identify_useful_nodes(useful);
1175 return useful.size();
1176 }
1177
1178 void Compile::print_missing_nodes() {
1179
1180 // Return if CompileLog is null and PrintIdealNodeCount is false.
1181 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1182 return;
1183 }
1184
1185 // This is an expensive function. It is executed only when the user
1186 // specifies VerifyIdealNodeCount option or otherwise knows the
1187 // additional work that needs to be done to identify reachable nodes
1188 // by walking the flow graph and find the missing ones using
1189 // _dead_node_list.
1190
1191 Unique_Node_List useful(comp_arena());
1192 // Get useful node list by walking the graph.
1193 identify_useful_nodes(useful);
1194
1195 uint l_nodes = C->live_nodes();
1196 uint l_nodes_by_walk = useful.size();
1197
1198 if (l_nodes != l_nodes_by_walk) {
1199 if (_log != nullptr) {
1200 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1201 _log->stamp();
1202 _log->end_head();
1203 }
1204 VectorSet& useful_member_set = useful.member_set();
1205 int last_idx = l_nodes_by_walk;
1206 for (int i = 0; i < last_idx; i++) {
1207 if (useful_member_set.test(i)) {
1208 if (_dead_node_list.test(i)) {
1209 if (_log != nullptr) {
1210 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1211 }
1212 if (PrintIdealNodeCount) {
1213 // Print the log message to tty
1214 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1215 useful.at(i)->dump();
1216 }
1217 }
1218 }
1219 else if (! _dead_node_list.test(i)) {
1220 if (_log != nullptr) {
1221 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1222 }
1223 if (PrintIdealNodeCount) {
1224 // Print the log message to tty
1225 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1226 }
1227 }
1228 }
1229 if (_log != nullptr) {
1230 _log->tail("mismatched_nodes");
1231 }
1232 }
1233 }
1234 void Compile::record_modified_node(Node* n) {
1235 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1236 _modified_nodes->push(n);
1237 }
1238 }
1239
1240 void Compile::remove_modified_node(Node* n) {
1241 if (_modified_nodes != nullptr) {
1242 _modified_nodes->remove(n);
1243 }
1244 }
1245 #endif
1246
1247 #ifndef PRODUCT
1248 void Compile::verify_top(Node* tn) const {
1249 if (tn != nullptr) {
1250 assert(tn->is_Con(), "top node must be a constant");
1251 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1252 assert(tn->in(0) != nullptr, "must have live top node");
1253 }
1254 }
1255 #endif
1256
1257
1258 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1259
1260 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1261 guarantee(arr != nullptr, "");
1262 int num_blocks = arr->length();
1263 if (grow_by < num_blocks) grow_by = num_blocks;
1264 int num_notes = grow_by * _node_notes_block_size;
1265 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1266 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1267 while (num_notes > 0) {
1268 arr->append(notes);
1269 notes += _node_notes_block_size;
1270 num_notes -= _node_notes_block_size;
1271 }
1272 assert(num_notes == 0, "exact multiple, please");
1273 }
1274
1275 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1276 if (source == nullptr || dest == nullptr) return false;
1277
1278 if (dest->is_Con())
1279 return false; // Do not push debug info onto constants.
1280
1281 #ifdef ASSERT
1282 // Leave a bread crumb trail pointing to the original node:
1283 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1284 dest->set_debug_orig(source);
1285 }
1286 #endif
1287
1288 if (node_note_array() == nullptr)
1289 return false; // Not collecting any notes now.
1290
1291 // This is a copy onto a pre-existing node, which may already have notes.
1292 // If both nodes have notes, do not overwrite any pre-existing notes.
1293 Node_Notes* source_notes = node_notes_at(source->_idx);
1294 if (source_notes == nullptr || source_notes->is_clear()) return false;
1295 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1296 if (dest_notes == nullptr || dest_notes->is_clear()) {
1297 return set_node_notes_at(dest->_idx, source_notes);
1298 }
1299
1300 Node_Notes merged_notes = (*source_notes);
1301 // The order of operations here ensures that dest notes will win...
1302 merged_notes.update_from(dest_notes);
1303 return set_node_notes_at(dest->_idx, &merged_notes);
1304 }
1305
1306
1307 //--------------------------allow_range_check_smearing-------------------------
1308 // Gating condition for coalescing similar range checks.
1309 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1310 // single covering check that is at least as strong as any of them.
1311 // If the optimization succeeds, the simplified (strengthened) range check
1312 // will always succeed. If it fails, we will deopt, and then give up
1313 // on the optimization.
1314 bool Compile::allow_range_check_smearing() const {
1315 // If this method has already thrown a range-check,
1316 // assume it was because we already tried range smearing
1317 // and it failed.
1318 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1319 return !already_trapped;
1320 }
1321
1322
1323 //------------------------------flatten_alias_type-----------------------------
1324 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1325 assert(do_aliasing(), "Aliasing should be enabled");
1326 int offset = tj->offset();
1327 TypePtr::PTR ptr = tj->ptr();
1328
1329 // Known instance (scalarizable allocation) alias only with itself.
1330 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1331 tj->is_oopptr()->is_known_instance();
1332
1333 // Process weird unsafe references.
1334 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1335 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1336 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1337 tj = TypeOopPtr::BOTTOM;
1338 ptr = tj->ptr();
1339 offset = tj->offset();
1340 }
1341
1342 // Array pointers need some flattening
1343 const TypeAryPtr* ta = tj->isa_aryptr();
1344 if (ta && ta->is_stable()) {
1345 // Erase stability property for alias analysis.
1346 tj = ta = ta->cast_to_stable(false);
1347 }
1348 if( ta && is_known_inst ) {
1349 if ( offset != Type::OffsetBot &&
1350 offset > arrayOopDesc::length_offset_in_bytes() ) {
1351 offset = Type::OffsetBot; // Flatten constant access into array body only
1352 tj = ta = ta->
1353 remove_speculative()->
1354 cast_to_ptr_type(ptr)->
1355 with_offset(offset);
1356 }
1357 } else if (ta) {
1358 // For arrays indexed by constant indices, we flatten the alias
1359 // space to include all of the array body. Only the header, klass
1360 // and array length can be accessed un-aliased.
1361 if( offset != Type::OffsetBot ) {
1362 if( ta->const_oop() ) { // MethodData* or Method*
1363 offset = Type::OffsetBot; // Flatten constant access into array body
1364 tj = ta = ta->
1365 remove_speculative()->
1366 cast_to_ptr_type(ptr)->
1367 cast_to_exactness(false)->
1368 with_offset(offset);
1369 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1370 // range is OK as-is.
1371 tj = ta = TypeAryPtr::RANGE;
1372 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1373 tj = TypeInstPtr::KLASS; // all klass loads look alike
1374 ta = TypeAryPtr::RANGE; // generic ignored junk
1375 ptr = TypePtr::BotPTR;
1376 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1377 tj = TypeInstPtr::MARK;
1378 ta = TypeAryPtr::RANGE; // generic ignored junk
1379 ptr = TypePtr::BotPTR;
1380 } else { // Random constant offset into array body
1381 offset = Type::OffsetBot; // Flatten constant access into array body
1382 tj = ta = ta->
1383 remove_speculative()->
1384 cast_to_ptr_type(ptr)->
1385 cast_to_exactness(false)->
1386 with_offset(offset);
1387 }
1388 }
1389 // Arrays of fixed size alias with arrays of unknown size.
1390 if (ta->size() != TypeInt::POS) {
1391 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1392 tj = ta = ta->
1393 remove_speculative()->
1394 cast_to_ptr_type(ptr)->
1395 with_ary(tary)->
1396 cast_to_exactness(false);
1397 }
1398 // Arrays of known objects become arrays of unknown objects.
1399 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1400 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1401 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1402 }
1403 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1404 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1405 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1406 }
1407 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1408 // cannot be distinguished by bytecode alone.
1409 if (ta->elem() == TypeInt::BOOL) {
1410 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1411 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1412 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1413 }
1414 // During the 2nd round of IterGVN, NotNull castings are removed.
1415 // Make sure the Bottom and NotNull variants alias the same.
1416 // Also, make sure exact and non-exact variants alias the same.
1417 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1418 tj = ta = ta->
1419 remove_speculative()->
1420 cast_to_ptr_type(TypePtr::BotPTR)->
1421 cast_to_exactness(false)->
1422 with_offset(offset);
1423 }
1424 }
1425
1426 // Oop pointers need some flattening
1427 const TypeInstPtr *to = tj->isa_instptr();
1428 if (to && to != TypeOopPtr::BOTTOM) {
1429 ciInstanceKlass* ik = to->instance_klass();
1430 if( ptr == TypePtr::Constant ) {
1431 if (ik != ciEnv::current()->Class_klass() ||
1432 offset < ik->layout_helper_size_in_bytes()) {
1433 // No constant oop pointers (such as Strings); they alias with
1434 // unknown strings.
1435 assert(!is_known_inst, "not scalarizable allocation");
1436 tj = to = to->
1437 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1438 remove_speculative()->
1439 cast_to_ptr_type(TypePtr::BotPTR)->
1440 cast_to_exactness(false);
1441 }
1442 } else if( is_known_inst ) {
1443 tj = to; // Keep NotNull and klass_is_exact for instance type
1444 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1445 // During the 2nd round of IterGVN, NotNull castings are removed.
1446 // Make sure the Bottom and NotNull variants alias the same.
1447 // Also, make sure exact and non-exact variants alias the same.
1448 tj = to = to->
1449 remove_speculative()->
1450 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1451 cast_to_ptr_type(TypePtr::BotPTR)->
1452 cast_to_exactness(false);
1453 }
1454 if (to->speculative() != nullptr) {
1455 tj = to = to->remove_speculative();
1456 }
1457 // Canonicalize the holder of this field
1458 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1459 // First handle header references such as a LoadKlassNode, even if the
1460 // object's klass is unloaded at compile time (4965979).
1461 if (!is_known_inst) { // Do it only for non-instance types
1462 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1463 }
1464 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1465 // Static fields are in the space above the normal instance
1466 // fields in the java.lang.Class instance.
1467 if (ik != ciEnv::current()->Class_klass()) {
1468 to = nullptr;
1469 tj = TypeOopPtr::BOTTOM;
1470 offset = tj->offset();
1471 }
1472 } else {
1473 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1474 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1475 assert(tj->offset() == offset, "no change to offset expected");
1476 bool xk = to->klass_is_exact();
1477 int instance_id = to->instance_id();
1478
1479 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1480 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1481 // its interfaces are included.
1482 if (xk && ik->equals(canonical_holder)) {
1483 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1484 } else {
1485 assert(xk || !is_known_inst, "Known instance should be exact type");
1486 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1487 }
1488 }
1489 }
1490
1491 // Klass pointers to object array klasses need some flattening
1492 const TypeKlassPtr *tk = tj->isa_klassptr();
1493 if( tk ) {
1494 // If we are referencing a field within a Klass, we need
1495 // to assume the worst case of an Object. Both exact and
1496 // inexact types must flatten to the same alias class so
1497 // use NotNull as the PTR.
1498 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1499 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1500 env()->Object_klass(),
1501 offset);
1502 }
1503
1504 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1505 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1506 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1507 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1508 } else {
1509 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1510 }
1511 }
1512
1513 // Check for precise loads from the primary supertype array and force them
1514 // to the supertype cache alias index. Check for generic array loads from
1515 // the primary supertype array and also force them to the supertype cache
1516 // alias index. Since the same load can reach both, we need to merge
1517 // these 2 disparate memories into the same alias class. Since the
1518 // primary supertype array is read-only, there's no chance of confusion
1519 // where we bypass an array load and an array store.
1520 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1521 if (offset == Type::OffsetBot ||
1522 (offset >= primary_supers_offset &&
1523 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1524 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1525 offset = in_bytes(Klass::secondary_super_cache_offset());
1526 tj = tk = tk->with_offset(offset);
1527 }
1528 }
1529
1530 // Flatten all Raw pointers together.
1531 if (tj->base() == Type::RawPtr)
1532 tj = TypeRawPtr::BOTTOM;
1533
1534 if (tj->base() == Type::AnyPtr)
1535 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1536
1537 offset = tj->offset();
1538 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1539
1540 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1541 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1542 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1543 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1544 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1545 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1546 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1547 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1548 assert( tj->ptr() != TypePtr::TopPTR &&
1549 tj->ptr() != TypePtr::AnyNull &&
1550 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1551 // assert( tj->ptr() != TypePtr::Constant ||
1552 // tj->base() == Type::RawPtr ||
1553 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1554
1555 return tj;
1556 }
1557
1558 void Compile::AliasType::Init(int i, const TypePtr* at) {
1559 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1560 _index = i;
1561 _adr_type = at;
1562 _field = nullptr;
1563 _element = nullptr;
1564 _is_rewritable = true; // default
1565 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1566 if (atoop != nullptr && atoop->is_known_instance()) {
1567 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1568 _general_index = Compile::current()->get_alias_index(gt);
1569 } else {
1570 _general_index = 0;
1571 }
1572 }
1573
1574 BasicType Compile::AliasType::basic_type() const {
1575 if (element() != nullptr) {
1576 const Type* element = adr_type()->is_aryptr()->elem();
1577 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1578 } if (field() != nullptr) {
1579 return field()->layout_type();
1580 } else {
1581 return T_ILLEGAL; // unknown
1582 }
1583 }
1584
1585 //---------------------------------print_on------------------------------------
1586 #ifndef PRODUCT
1587 void Compile::AliasType::print_on(outputStream* st) {
1588 if (index() < 10)
1589 st->print("@ <%d> ", index());
1590 else st->print("@ <%d>", index());
1591 st->print(is_rewritable() ? " " : " RO");
1592 int offset = adr_type()->offset();
1593 if (offset == Type::OffsetBot)
1594 st->print(" +any");
1595 else st->print(" +%-3d", offset);
1596 st->print(" in ");
1597 adr_type()->dump_on(st);
1598 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1599 if (field() != nullptr && tjp) {
1600 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1601 tjp->offset() != field()->offset_in_bytes()) {
1602 st->print(" != ");
1603 field()->print();
1604 st->print(" ***");
1605 }
1606 }
1607 }
1608
1609 void print_alias_types() {
1610 Compile* C = Compile::current();
1611 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1612 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1613 C->alias_type(idx)->print_on(tty);
1614 tty->cr();
1615 }
1616 }
1617 #endif
1618
1619
1620 //----------------------------probe_alias_cache--------------------------------
1621 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1622 intptr_t key = (intptr_t) adr_type;
1623 key ^= key >> logAliasCacheSize;
1624 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1625 }
1626
1627
1628 //-----------------------------grow_alias_types--------------------------------
1629 void Compile::grow_alias_types() {
1630 const int old_ats = _max_alias_types; // how many before?
1631 const int new_ats = old_ats; // how many more?
1632 const int grow_ats = old_ats+new_ats; // how many now?
1633 _max_alias_types = grow_ats;
1634 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1635 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1636 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1637 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1638 }
1639
1640
1641 //--------------------------------find_alias_type------------------------------
1642 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1643 if (!do_aliasing()) {
1644 return alias_type(AliasIdxBot);
1645 }
1646
1647 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1648 if (ace->_adr_type == adr_type) {
1649 return alias_type(ace->_index);
1650 }
1651
1652 // Handle special cases.
1653 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1654 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1655
1656 // Do it the slow way.
1657 const TypePtr* flat = flatten_alias_type(adr_type);
1658
1659 #ifdef ASSERT
1660 {
1661 ResourceMark rm;
1662 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1663 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1664 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1665 Type::str(adr_type));
1666 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1667 const TypeOopPtr* foop = flat->is_oopptr();
1668 // Scalarizable allocations have exact klass always.
1669 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1670 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1671 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1672 Type::str(foop), Type::str(xoop));
1673 }
1674 }
1675 #endif
1676
1677 int idx = AliasIdxTop;
1678 for (int i = 0; i < num_alias_types(); i++) {
1679 if (alias_type(i)->adr_type() == flat) {
1680 idx = i;
1681 break;
1682 }
1683 }
1684
1685 if (idx == AliasIdxTop) {
1686 if (no_create) return nullptr;
1687 // Grow the array if necessary.
1688 if (_num_alias_types == _max_alias_types) grow_alias_types();
1689 // Add a new alias type.
1690 idx = _num_alias_types++;
1691 _alias_types[idx]->Init(idx, flat);
1692 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1693 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1694 if (flat->isa_instptr()) {
1695 if (flat->offset() == java_lang_Class::klass_offset()
1696 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1697 alias_type(idx)->set_rewritable(false);
1698 }
1699 if (flat->isa_aryptr()) {
1700 #ifdef ASSERT
1701 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1702 // (T_BYTE has the weakest alignment and size restrictions...)
1703 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1704 #endif
1705 if (flat->offset() == TypePtr::OffsetBot) {
1706 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1707 }
1708 }
1709 if (flat->isa_klassptr()) {
1710 if (UseCompactObjectHeaders) {
1711 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1712 alias_type(idx)->set_rewritable(false);
1713 }
1714 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1715 alias_type(idx)->set_rewritable(false);
1716 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1717 alias_type(idx)->set_rewritable(false);
1718 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1719 alias_type(idx)->set_rewritable(false);
1720 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1721 alias_type(idx)->set_rewritable(false);
1722 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1723 alias_type(idx)->set_rewritable(false);
1724 }
1725 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1726 // but the base pointer type is not distinctive enough to identify
1727 // references into JavaThread.)
1728
1729 // Check for final fields.
1730 const TypeInstPtr* tinst = flat->isa_instptr();
1731 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1732 ciField* field;
1733 if (tinst->const_oop() != nullptr &&
1734 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1735 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1736 // static field
1737 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1738 field = k->get_field_by_offset(tinst->offset(), true);
1739 } else {
1740 ciInstanceKlass *k = tinst->instance_klass();
1741 field = k->get_field_by_offset(tinst->offset(), false);
1742 }
1743 assert(field == nullptr ||
1744 original_field == nullptr ||
1745 (field->holder() == original_field->holder() &&
1746 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1747 field->is_static() == original_field->is_static()), "wrong field?");
1748 // Set field() and is_rewritable() attributes.
1749 if (field != nullptr) alias_type(idx)->set_field(field);
1750 }
1751 }
1752
1753 // Fill the cache for next time.
1754 ace->_adr_type = adr_type;
1755 ace->_index = idx;
1756 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1757
1758 // Might as well try to fill the cache for the flattened version, too.
1759 AliasCacheEntry* face = probe_alias_cache(flat);
1760 if (face->_adr_type == nullptr) {
1761 face->_adr_type = flat;
1762 face->_index = idx;
1763 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1764 }
1765
1766 return alias_type(idx);
1767 }
1768
1769
1770 Compile::AliasType* Compile::alias_type(ciField* field) {
1771 const TypeOopPtr* t;
1772 if (field->is_static())
1773 t = TypeInstPtr::make(field->holder()->java_mirror());
1774 else
1775 t = TypeOopPtr::make_from_klass_raw(field->holder());
1776 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1777 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1778 return atp;
1779 }
1780
1781
1782 //------------------------------have_alias_type--------------------------------
1783 bool Compile::have_alias_type(const TypePtr* adr_type) {
1784 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1785 if (ace->_adr_type == adr_type) {
1786 return true;
1787 }
1788
1789 // Handle special cases.
1790 if (adr_type == nullptr) return true;
1791 if (adr_type == TypePtr::BOTTOM) return true;
1792
1793 return find_alias_type(adr_type, true, nullptr) != nullptr;
1794 }
1795
1796 //-----------------------------must_alias--------------------------------------
1797 // True if all values of the given address type are in the given alias category.
1798 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1799 if (alias_idx == AliasIdxBot) return true; // the universal category
1800 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1801 if (alias_idx == AliasIdxTop) return false; // the empty category
1802 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1803
1804 // the only remaining possible overlap is identity
1805 int adr_idx = get_alias_index(adr_type);
1806 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1807 assert(adr_idx == alias_idx ||
1808 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1809 && adr_type != TypeOopPtr::BOTTOM),
1810 "should not be testing for overlap with an unsafe pointer");
1811 return adr_idx == alias_idx;
1812 }
1813
1814 //------------------------------can_alias--------------------------------------
1815 // True if any values of the given address type are in the given alias category.
1816 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1817 if (alias_idx == AliasIdxTop) return false; // the empty category
1818 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1819 // Known instance doesn't alias with bottom memory
1820 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1821 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1822
1823 // the only remaining possible overlap is identity
1824 int adr_idx = get_alias_index(adr_type);
1825 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1826 return adr_idx == alias_idx;
1827 }
1828
1829 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1830 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1831 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1832 if (parse_predicate_count() == 0) {
1833 return;
1834 }
1835 for (int i = 0; i < parse_predicate_count(); i++) {
1836 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1837 parse_predicate->mark_useless(igvn);
1838 }
1839 _parse_predicates.clear();
1840 }
1841
1842 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1843 if (!n->for_post_loop_opts_igvn()) {
1844 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1845 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1846 _for_post_loop_igvn.append(n);
1847 }
1848 }
1849
1850 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1851 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1852 _for_post_loop_igvn.remove(n);
1853 }
1854
1855 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1856 // Verify that all previous optimizations produced a valid graph
1857 // at least to this point, even if no loop optimizations were done.
1858 PhaseIdealLoop::verify(igvn);
1859
1860 C->set_post_loop_opts_phase(); // no more loop opts allowed
1861
1862 assert(!C->major_progress(), "not cleared");
1863
1864 if (_for_post_loop_igvn.length() > 0) {
1865 while (_for_post_loop_igvn.length() > 0) {
1866 Node* n = _for_post_loop_igvn.pop();
1867 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1868 igvn._worklist.push(n);
1869 }
1870 igvn.optimize();
1871 if (failing()) return;
1872 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1873 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1874
1875 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1876 if (C->major_progress()) {
1877 C->clear_major_progress(); // ensure that major progress is now clear
1878 }
1879 }
1880 }
1881
1882 void Compile::record_for_merge_stores_igvn(Node* n) {
1883 if (!n->for_merge_stores_igvn()) {
1884 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1885 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1886 _for_merge_stores_igvn.append(n);
1887 }
1888 }
1889
1890 void Compile::remove_from_merge_stores_igvn(Node* n) {
1891 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1892 _for_merge_stores_igvn.remove(n);
1893 }
1894
1895 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1896 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1897 // the stores, and we merge the wrong sequence of stores.
1898 // Example:
1899 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1900 // Apply MergeStores:
1901 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1902 // Remove more RangeChecks:
1903 // StoreI [ StoreL ] StoreI
1904 // But now it would have been better to do this instead:
1905 // [ StoreL ] [ StoreL ]
1906 //
1907 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1908 // since we never unset _merge_stores_phase.
1909 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1910 C->set_merge_stores_phase();
1911
1912 if (_for_merge_stores_igvn.length() > 0) {
1913 while (_for_merge_stores_igvn.length() > 0) {
1914 Node* n = _for_merge_stores_igvn.pop();
1915 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1916 igvn._worklist.push(n);
1917 }
1918 igvn.optimize();
1919 if (failing()) return;
1920 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1921 print_method(PHASE_AFTER_MERGE_STORES, 3);
1922 }
1923 }
1924
1925 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1926 if (OptimizeUnstableIf) {
1927 _unstable_if_traps.append(trap);
1928 }
1929 }
1930
1931 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1932 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1933 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1934 Node* n = trap->uncommon_trap();
1935 if (!useful.member(n)) {
1936 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1937 }
1938 }
1939 }
1940
1941 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1942 // or fold-compares case. Return true if succeed or not found.
1943 //
1944 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1945 // false and ask fold-compares to yield.
1946 //
1947 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1948 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1949 // when deoptimization does happen.
1950 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1951 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1952 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1953 if (trap->uncommon_trap() == unc) {
1954 if (yield && trap->modified()) {
1955 return false;
1956 }
1957 _unstable_if_traps.delete_at(i);
1958 break;
1959 }
1960 }
1961 return true;
1962 }
1963
1964 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1965 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1966 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1967 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1968 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1969 CallStaticJavaNode* unc = trap->uncommon_trap();
1970 int next_bci = trap->next_bci();
1971 bool modified = trap->modified();
1972
1973 if (next_bci != -1 && !modified) {
1974 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1975 JVMState* jvms = unc->jvms();
1976 ciMethod* method = jvms->method();
1977 ciBytecodeStream iter(method);
1978
1979 iter.force_bci(jvms->bci());
1980 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1981 Bytecodes::Code c = iter.cur_bc();
1982 Node* lhs = nullptr;
1983 Node* rhs = nullptr;
1984 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1985 lhs = unc->peek_operand(0);
1986 rhs = unc->peek_operand(1);
1987 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1988 lhs = unc->peek_operand(0);
1989 }
1990
1991 ResourceMark rm;
1992 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1993 assert(live_locals.is_valid(), "broken liveness info");
1994 int len = (int)live_locals.size();
1995
1996 for (int i = 0; i < len; i++) {
1997 Node* local = unc->local(jvms, i);
1998 // kill local using the liveness of next_bci.
1999 // give up when the local looks like an operand to secure reexecution.
2000 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2001 uint idx = jvms->locoff() + i;
2002 #ifdef ASSERT
2003 if (PrintOpto && Verbose) {
2004 tty->print("[unstable_if] kill local#%d: ", idx);
2005 local->dump();
2006 tty->cr();
2007 }
2008 #endif
2009 igvn.replace_input_of(unc, idx, top());
2010 modified = true;
2011 }
2012 }
2013 }
2014
2015 // keep the mondified trap for late query
2016 if (modified) {
2017 trap->set_modified();
2018 } else {
2019 _unstable_if_traps.delete_at(i);
2020 }
2021 }
2022 igvn.optimize();
2023 }
2024
2025 // StringOpts and late inlining of string methods
2026 void Compile::inline_string_calls(bool parse_time) {
2027 {
2028 // remove useless nodes to make the usage analysis simpler
2029 ResourceMark rm;
2030 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2031 }
2032
2033 {
2034 ResourceMark rm;
2035 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2036 PhaseStringOpts pso(initial_gvn());
2037 print_method(PHASE_AFTER_STRINGOPTS, 3);
2038 }
2039
2040 // now inline anything that we skipped the first time around
2041 if (!parse_time) {
2042 _late_inlines_pos = _late_inlines.length();
2043 }
2044
2045 while (_string_late_inlines.length() > 0) {
2046 CallGenerator* cg = _string_late_inlines.pop();
2047 cg->do_late_inline();
2048 if (failing()) return;
2049 }
2050 _string_late_inlines.trunc_to(0);
2051 }
2052
2053 // Late inlining of boxing methods
2054 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2055 if (_boxing_late_inlines.length() > 0) {
2056 assert(has_boxed_value(), "inconsistent");
2057
2058 set_inlining_incrementally(true);
2059
2060 igvn_worklist()->ensure_empty(); // should be done with igvn
2061
2062 _late_inlines_pos = _late_inlines.length();
2063
2064 while (_boxing_late_inlines.length() > 0) {
2065 CallGenerator* cg = _boxing_late_inlines.pop();
2066 cg->do_late_inline();
2067 if (failing()) return;
2068 }
2069 _boxing_late_inlines.trunc_to(0);
2070
2071 inline_incrementally_cleanup(igvn);
2072
2073 set_inlining_incrementally(false);
2074 }
2075 }
2076
2077 bool Compile::inline_incrementally_one() {
2078 assert(IncrementalInline, "incremental inlining should be on");
2079
2080 TracePhase tp(_t_incrInline_inline);
2081
2082 set_inlining_progress(false);
2083 set_do_cleanup(false);
2084
2085 for (int i = 0; i < _late_inlines.length(); i++) {
2086 _late_inlines_pos = i+1;
2087 CallGenerator* cg = _late_inlines.at(i);
2088 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2089 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2090 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2091 cg->do_late_inline();
2092 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2093 if (failing()) {
2094 return false;
2095 } else if (inlining_progress()) {
2096 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2097 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2098 break; // process one call site at a time
2099 } else {
2100 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2101 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2102 // Avoid potential infinite loop if node already in the IGVN list
2103 assert(false, "scheduled for IGVN during inlining attempt");
2104 } else {
2105 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2106 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2107 cg->call_node()->set_generator(cg);
2108 }
2109 }
2110 } else {
2111 // Ignore late inline direct calls when inlining is not allowed.
2112 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2113 }
2114 }
2115 // Remove processed elements.
2116 _late_inlines.remove_till(_late_inlines_pos);
2117 _late_inlines_pos = 0;
2118
2119 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2120
2121 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2122
2123 set_inlining_progress(false);
2124 set_do_cleanup(false);
2125
2126 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2127 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2128 }
2129
2130 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2131 {
2132 TracePhase tp(_t_incrInline_pru);
2133 ResourceMark rm;
2134 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2135 }
2136 {
2137 TracePhase tp(_t_incrInline_igvn);
2138 igvn.reset_from_gvn(initial_gvn());
2139 igvn.optimize();
2140 if (failing()) return;
2141 }
2142 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2143 }
2144
2145 // Perform incremental inlining until bound on number of live nodes is reached
2146 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2147 TracePhase tp(_t_incrInline);
2148
2149 set_inlining_incrementally(true);
2150 uint low_live_nodes = 0;
2151
2152 while (_late_inlines.length() > 0) {
2153 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2154 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2155 TracePhase tp(_t_incrInline_ideal);
2156 // PhaseIdealLoop is expensive so we only try it once we are
2157 // out of live nodes and we only try it again if the previous
2158 // helped got the number of nodes down significantly
2159 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2160 if (failing()) return;
2161 low_live_nodes = live_nodes();
2162 _major_progress = true;
2163 }
2164
2165 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2166 bool do_print_inlining = print_inlining() || print_intrinsics();
2167 if (do_print_inlining || log() != nullptr) {
2168 // Print inlining message for candidates that we couldn't inline for lack of space.
2169 for (int i = 0; i < _late_inlines.length(); i++) {
2170 CallGenerator* cg = _late_inlines.at(i);
2171 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2172 if (do_print_inlining) {
2173 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2174 }
2175 log_late_inline_failure(cg, msg);
2176 }
2177 }
2178 break; // finish
2179 }
2180 }
2181
2182 igvn_worklist()->ensure_empty(); // should be done with igvn
2183
2184 while (inline_incrementally_one()) {
2185 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2186 }
2187 if (failing()) return;
2188
2189 inline_incrementally_cleanup(igvn);
2190
2191 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2192
2193 if (failing()) return;
2194
2195 if (_late_inlines.length() == 0) {
2196 break; // no more progress
2197 }
2198 }
2199
2200 igvn_worklist()->ensure_empty(); // should be done with igvn
2201
2202 if (_string_late_inlines.length() > 0) {
2203 assert(has_stringbuilder(), "inconsistent");
2204
2205 inline_string_calls(false);
2206
2207 if (failing()) return;
2208
2209 inline_incrementally_cleanup(igvn);
2210 }
2211
2212 set_inlining_incrementally(false);
2213 }
2214
2215 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2216 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2217 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2218 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2219 // as if "inlining_incrementally() == true" were set.
2220 assert(inlining_incrementally() == false, "not allowed");
2221 assert(_modified_nodes == nullptr, "not allowed");
2222 assert(_late_inlines.length() > 0, "sanity");
2223
2224 while (_late_inlines.length() > 0) {
2225 igvn_worklist()->ensure_empty(); // should be done with igvn
2226
2227 while (inline_incrementally_one()) {
2228 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2229 }
2230 if (failing()) return;
2231
2232 inline_incrementally_cleanup(igvn);
2233 }
2234 }
2235
2236 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2237 if (_loop_opts_cnt > 0) {
2238 while (major_progress() && (_loop_opts_cnt > 0)) {
2239 TracePhase tp(_t_idealLoop);
2240 PhaseIdealLoop::optimize(igvn, mode);
2241 _loop_opts_cnt--;
2242 if (failing()) return false;
2243 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2244 }
2245 }
2246 return true;
2247 }
2248
2249 // Remove edges from "root" to each SafePoint at a backward branch.
2250 // They were inserted during parsing (see add_safepoint()) to make
2251 // infinite loops without calls or exceptions visible to root, i.e.,
2252 // useful.
2253 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2254 Node *r = root();
2255 if (r != nullptr) {
2256 for (uint i = r->req(); i < r->len(); ++i) {
2257 Node *n = r->in(i);
2258 if (n != nullptr && n->is_SafePoint()) {
2259 r->rm_prec(i);
2260 if (n->outcnt() == 0) {
2261 igvn.remove_dead_node(n);
2262 }
2263 --i;
2264 }
2265 }
2266 // Parsing may have added top inputs to the root node (Path
2267 // leading to the Halt node proven dead). Make sure we get a
2268 // chance to clean them up.
2269 igvn._worklist.push(r);
2270 igvn.optimize();
2271 }
2272 }
2273
2274 //------------------------------Optimize---------------------------------------
2275 // Given a graph, optimize it.
2276 void Compile::Optimize() {
2277 TracePhase tp(_t_optimizer);
2278
2279 #ifndef PRODUCT
2280 if (env()->break_at_compile()) {
2281 BREAKPOINT;
2282 }
2283
2284 #endif
2285
2286 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2287 #ifdef ASSERT
2288 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2289 #endif
2290
2291 ResourceMark rm;
2292
2293 NOT_PRODUCT( verify_graph_edges(); )
2294
2295 print_method(PHASE_AFTER_PARSING, 1);
2296
2297 {
2298 // Iterative Global Value Numbering, including ideal transforms
2299 // Initialize IterGVN with types and values from parse-time GVN
2300 PhaseIterGVN igvn(initial_gvn());
2301 #ifdef ASSERT
2302 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2303 #endif
2304 {
2305 TracePhase tp(_t_iterGVN);
2306 igvn.optimize();
2307 }
2308
2309 if (failing()) return;
2310
2311 print_method(PHASE_ITER_GVN1, 2);
2312
2313 process_for_unstable_if_traps(igvn);
2314
2315 if (failing()) return;
2316
2317 inline_incrementally(igvn);
2318
2319 print_method(PHASE_INCREMENTAL_INLINE, 2);
2320
2321 if (failing()) return;
2322
2323 if (eliminate_boxing()) {
2324 // Inline valueOf() methods now.
2325 inline_boxing_calls(igvn);
2326
2327 if (failing()) return;
2328
2329 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2330 inline_incrementally(igvn);
2331 }
2332
2333 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2334
2335 if (failing()) return;
2336 }
2337
2338 // Remove the speculative part of types and clean up the graph from
2339 // the extra CastPP nodes whose only purpose is to carry them. Do
2340 // that early so that optimizations are not disrupted by the extra
2341 // CastPP nodes.
2342 remove_speculative_types(igvn);
2343
2344 if (failing()) return;
2345
2346 // No more new expensive nodes will be added to the list from here
2347 // so keep only the actual candidates for optimizations.
2348 cleanup_expensive_nodes(igvn);
2349
2350 if (failing()) return;
2351
2352 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2353 if (EnableVectorSupport && has_vbox_nodes()) {
2354 TracePhase tp(_t_vector);
2355 PhaseVector pv(igvn);
2356 pv.optimize_vector_boxes();
2357 if (failing()) return;
2358 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2359 }
2360 assert(!has_vbox_nodes(), "sanity");
2361
2362 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2363 Compile::TracePhase tp(_t_renumberLive);
2364 igvn_worklist()->ensure_empty(); // should be done with igvn
2365 {
2366 ResourceMark rm;
2367 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2368 }
2369 igvn.reset_from_gvn(initial_gvn());
2370 igvn.optimize();
2371 if (failing()) return;
2372 }
2373
2374 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2375 // safepoints
2376 remove_root_to_sfpts_edges(igvn);
2377
2378 if (failing()) return;
2379
2380 // Perform escape analysis
2381 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2382 if (has_loops()) {
2383 // Cleanup graph (remove dead nodes).
2384 TracePhase tp(_t_idealLoop);
2385 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2386 if (failing()) return;
2387 }
2388 bool progress;
2389 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2390 do {
2391 ConnectionGraph::do_analysis(this, &igvn);
2392
2393 if (failing()) return;
2394
2395 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2396
2397 // Optimize out fields loads from scalar replaceable allocations.
2398 igvn.optimize();
2399 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2400
2401 if (failing()) return;
2402
2403 if (congraph() != nullptr && macro_count() > 0) {
2404 TracePhase tp(_t_macroEliminate);
2405 PhaseMacroExpand mexp(igvn);
2406 mexp.eliminate_macro_nodes();
2407 if (failing()) return;
2408
2409 igvn.set_delay_transform(false);
2410 igvn.optimize();
2411 if (failing()) return;
2412
2413 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2414 }
2415
2416 ConnectionGraph::verify_ram_nodes(this, root());
2417 if (failing()) return;
2418
2419 progress = do_iterative_escape_analysis() &&
2420 (macro_count() < mcount) &&
2421 ConnectionGraph::has_candidates(this);
2422 // Try again if candidates exist and made progress
2423 // by removing some allocations and/or locks.
2424 } while (progress);
2425 }
2426
2427 // Loop transforms on the ideal graph. Range Check Elimination,
2428 // peeling, unrolling, etc.
2429
2430 // Set loop opts counter
2431 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2432 {
2433 TracePhase tp(_t_idealLoop);
2434 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2435 _loop_opts_cnt--;
2436 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2437 if (failing()) return;
2438 }
2439 // Loop opts pass if partial peeling occurred in previous pass
2440 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2441 TracePhase tp(_t_idealLoop);
2442 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2443 _loop_opts_cnt--;
2444 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2445 if (failing()) return;
2446 }
2447 // Loop opts pass for loop-unrolling before CCP
2448 if(major_progress() && (_loop_opts_cnt > 0)) {
2449 TracePhase tp(_t_idealLoop);
2450 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2451 _loop_opts_cnt--;
2452 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2453 }
2454 if (!failing()) {
2455 // Verify that last round of loop opts produced a valid graph
2456 PhaseIdealLoop::verify(igvn);
2457 }
2458 }
2459 if (failing()) return;
2460
2461 // Conditional Constant Propagation;
2462 print_method(PHASE_BEFORE_CCP1, 2);
2463 PhaseCCP ccp( &igvn );
2464 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2465 {
2466 TracePhase tp(_t_ccp);
2467 ccp.do_transform();
2468 }
2469 print_method(PHASE_CCP1, 2);
2470
2471 assert( true, "Break here to ccp.dump_old2new_map()");
2472
2473 // Iterative Global Value Numbering, including ideal transforms
2474 {
2475 TracePhase tp(_t_iterGVN2);
2476 igvn.reset_from_igvn(&ccp);
2477 igvn.optimize();
2478 }
2479 print_method(PHASE_ITER_GVN2, 2);
2480
2481 if (failing()) return;
2482
2483 // Loop transforms on the ideal graph. Range Check Elimination,
2484 // peeling, unrolling, etc.
2485 if (!optimize_loops(igvn, LoopOptsDefault)) {
2486 return;
2487 }
2488
2489 if (failing()) return;
2490
2491 C->clear_major_progress(); // ensure that major progress is now clear
2492
2493 process_for_post_loop_opts_igvn(igvn);
2494
2495 process_for_merge_stores_igvn(igvn);
2496
2497 if (failing()) return;
2498
2499 #ifdef ASSERT
2500 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2501 #endif
2502
2503 {
2504 TracePhase tp(_t_macroExpand);
2505 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2506 PhaseMacroExpand mex(igvn);
2507 if (mex.expand_macro_nodes()) {
2508 assert(failing(), "must bail out w/ explicit message");
2509 return;
2510 }
2511 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2512 }
2513
2514 {
2515 TracePhase tp(_t_barrierExpand);
2516 if (bs->expand_barriers(this, igvn)) {
2517 assert(failing(), "must bail out w/ explicit message");
2518 return;
2519 }
2520 print_method(PHASE_BARRIER_EXPANSION, 2);
2521 }
2522
2523 if (C->max_vector_size() > 0) {
2524 C->optimize_logic_cones(igvn);
2525 igvn.optimize();
2526 if (failing()) return;
2527 }
2528
2529 DEBUG_ONLY( _modified_nodes = nullptr; )
2530
2531 assert(igvn._worklist.size() == 0, "not empty");
2532
2533 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2534
2535 if (_late_inlines.length() > 0) {
2536 // More opportunities to optimize virtual and MH calls.
2537 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2538 process_late_inline_calls_no_inline(igvn);
2539 if (failing()) return;
2540 }
2541 } // (End scope of igvn; run destructor if necessary for asserts.)
2542
2543 check_no_dead_use();
2544
2545 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2546 // to remove hashes to unlock nodes for modifications.
2547 C->node_hash()->clear();
2548
2549 // A method with only infinite loops has no edges entering loops from root
2550 {
2551 TracePhase tp(_t_graphReshaping);
2552 if (final_graph_reshaping()) {
2553 assert(failing(), "must bail out w/ explicit message");
2554 return;
2555 }
2556 }
2557
2558 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2559 DEBUG_ONLY(set_phase_optimize_finished();)
2560 }
2561
2562 #ifdef ASSERT
2563 void Compile::check_no_dead_use() const {
2564 ResourceMark rm;
2565 Unique_Node_List wq;
2566 wq.push(root());
2567 for (uint i = 0; i < wq.size(); ++i) {
2568 Node* n = wq.at(i);
2569 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2570 Node* u = n->fast_out(j);
2571 if (u->outcnt() == 0 && !u->is_Con()) {
2572 u->dump();
2573 fatal("no reachable node should have no use");
2574 }
2575 wq.push(u);
2576 }
2577 }
2578 }
2579 #endif
2580
2581 void Compile::inline_vector_reboxing_calls() {
2582 if (C->_vector_reboxing_late_inlines.length() > 0) {
2583 _late_inlines_pos = C->_late_inlines.length();
2584 while (_vector_reboxing_late_inlines.length() > 0) {
2585 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2586 cg->do_late_inline();
2587 if (failing()) return;
2588 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2589 }
2590 _vector_reboxing_late_inlines.trunc_to(0);
2591 }
2592 }
2593
2594 bool Compile::has_vbox_nodes() {
2595 if (C->_vector_reboxing_late_inlines.length() > 0) {
2596 return true;
2597 }
2598 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2599 Node * n = C->macro_node(macro_idx);
2600 assert(n->is_macro(), "only macro nodes expected here");
2601 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2602 return true;
2603 }
2604 }
2605 return false;
2606 }
2607
2608 //---------------------------- Bitwise operation packing optimization ---------------------------
2609
2610 static bool is_vector_unary_bitwise_op(Node* n) {
2611 return n->Opcode() == Op_XorV &&
2612 VectorNode::is_vector_bitwise_not_pattern(n);
2613 }
2614
2615 static bool is_vector_binary_bitwise_op(Node* n) {
2616 switch (n->Opcode()) {
2617 case Op_AndV:
2618 case Op_OrV:
2619 return true;
2620
2621 case Op_XorV:
2622 return !is_vector_unary_bitwise_op(n);
2623
2624 default:
2625 return false;
2626 }
2627 }
2628
2629 static bool is_vector_ternary_bitwise_op(Node* n) {
2630 return n->Opcode() == Op_MacroLogicV;
2631 }
2632
2633 static bool is_vector_bitwise_op(Node* n) {
2634 return is_vector_unary_bitwise_op(n) ||
2635 is_vector_binary_bitwise_op(n) ||
2636 is_vector_ternary_bitwise_op(n);
2637 }
2638
2639 static bool is_vector_bitwise_cone_root(Node* n) {
2640 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2641 return false;
2642 }
2643 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2644 if (is_vector_bitwise_op(n->fast_out(i))) {
2645 return false;
2646 }
2647 }
2648 return true;
2649 }
2650
2651 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2652 uint cnt = 0;
2653 if (is_vector_bitwise_op(n)) {
2654 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2655 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2656 for (uint i = 1; i < inp_cnt; i++) {
2657 Node* in = n->in(i);
2658 bool skip = VectorNode::is_all_ones_vector(in);
2659 if (!skip && !inputs.member(in)) {
2660 inputs.push(in);
2661 cnt++;
2662 }
2663 }
2664 assert(cnt <= 1, "not unary");
2665 } else {
2666 uint last_req = inp_cnt;
2667 if (is_vector_ternary_bitwise_op(n)) {
2668 last_req = inp_cnt - 1; // skip last input
2669 }
2670 for (uint i = 1; i < last_req; i++) {
2671 Node* def = n->in(i);
2672 if (!inputs.member(def)) {
2673 inputs.push(def);
2674 cnt++;
2675 }
2676 }
2677 }
2678 } else { // not a bitwise operations
2679 if (!inputs.member(n)) {
2680 inputs.push(n);
2681 cnt++;
2682 }
2683 }
2684 return cnt;
2685 }
2686
2687 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2688 Unique_Node_List useful_nodes;
2689 C->identify_useful_nodes(useful_nodes);
2690
2691 for (uint i = 0; i < useful_nodes.size(); i++) {
2692 Node* n = useful_nodes.at(i);
2693 if (is_vector_bitwise_cone_root(n)) {
2694 list.push(n);
2695 }
2696 }
2697 }
2698
2699 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2700 const TypeVect* vt,
2701 Unique_Node_List& partition,
2702 Unique_Node_List& inputs) {
2703 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2704 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2705 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2706
2707 Node* in1 = inputs.at(0);
2708 Node* in2 = inputs.at(1);
2709 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2710
2711 uint func = compute_truth_table(partition, inputs);
2712
2713 Node* pn = partition.at(partition.size() - 1);
2714 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2715 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2716 }
2717
2718 static uint extract_bit(uint func, uint pos) {
2719 return (func & (1 << pos)) >> pos;
2720 }
2721
2722 //
2723 // A macro logic node represents a truth table. It has 4 inputs,
2724 // First three inputs corresponds to 3 columns of a truth table
2725 // and fourth input captures the logic function.
2726 //
2727 // eg. fn = (in1 AND in2) OR in3;
2728 //
2729 // MacroNode(in1,in2,in3,fn)
2730 //
2731 // -----------------
2732 // in1 in2 in3 fn
2733 // -----------------
2734 // 0 0 0 0
2735 // 0 0 1 1
2736 // 0 1 0 0
2737 // 0 1 1 1
2738 // 1 0 0 0
2739 // 1 0 1 1
2740 // 1 1 0 1
2741 // 1 1 1 1
2742 //
2743
2744 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2745 int res = 0;
2746 for (int i = 0; i < 8; i++) {
2747 int bit1 = extract_bit(in1, i);
2748 int bit2 = extract_bit(in2, i);
2749 int bit3 = extract_bit(in3, i);
2750
2751 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2752 int func_bit = extract_bit(func, func_bit_pos);
2753
2754 res |= func_bit << i;
2755 }
2756 return res;
2757 }
2758
2759 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2760 assert(n != nullptr, "");
2761 assert(eval_map.contains(n), "absent");
2762 return *(eval_map.get(n));
2763 }
2764
2765 static void eval_operands(Node* n,
2766 uint& func1, uint& func2, uint& func3,
2767 ResourceHashtable<Node*,uint>& eval_map) {
2768 assert(is_vector_bitwise_op(n), "");
2769
2770 if (is_vector_unary_bitwise_op(n)) {
2771 Node* opnd = n->in(1);
2772 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2773 opnd = n->in(2);
2774 }
2775 func1 = eval_operand(opnd, eval_map);
2776 } else if (is_vector_binary_bitwise_op(n)) {
2777 func1 = eval_operand(n->in(1), eval_map);
2778 func2 = eval_operand(n->in(2), eval_map);
2779 } else {
2780 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2781 func1 = eval_operand(n->in(1), eval_map);
2782 func2 = eval_operand(n->in(2), eval_map);
2783 func3 = eval_operand(n->in(3), eval_map);
2784 }
2785 }
2786
2787 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2788 assert(inputs.size() <= 3, "sanity");
2789 ResourceMark rm;
2790 uint res = 0;
2791 ResourceHashtable<Node*,uint> eval_map;
2792
2793 // Populate precomputed functions for inputs.
2794 // Each input corresponds to one column of 3 input truth-table.
2795 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2796 0xCC, // (_, b, _) -> b
2797 0xF0 }; // (a, _, _) -> a
2798 for (uint i = 0; i < inputs.size(); i++) {
2799 eval_map.put(inputs.at(i), input_funcs[2-i]);
2800 }
2801
2802 for (uint i = 0; i < partition.size(); i++) {
2803 Node* n = partition.at(i);
2804
2805 uint func1 = 0, func2 = 0, func3 = 0;
2806 eval_operands(n, func1, func2, func3, eval_map);
2807
2808 switch (n->Opcode()) {
2809 case Op_OrV:
2810 assert(func3 == 0, "not binary");
2811 res = func1 | func2;
2812 break;
2813 case Op_AndV:
2814 assert(func3 == 0, "not binary");
2815 res = func1 & func2;
2816 break;
2817 case Op_XorV:
2818 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2819 assert(func2 == 0 && func3 == 0, "not unary");
2820 res = (~func1) & 0xFF;
2821 } else {
2822 assert(func3 == 0, "not binary");
2823 res = func1 ^ func2;
2824 }
2825 break;
2826 case Op_MacroLogicV:
2827 // Ordering of inputs may change during evaluation of sub-tree
2828 // containing MacroLogic node as a child node, thus a re-evaluation
2829 // makes sure that function is evaluated in context of current
2830 // inputs.
2831 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2832 break;
2833
2834 default: assert(false, "not supported: %s", n->Name());
2835 }
2836 assert(res <= 0xFF, "invalid");
2837 eval_map.put(n, res);
2838 }
2839 return res;
2840 }
2841
2842 // Criteria under which nodes gets packed into a macro logic node:-
2843 // 1) Parent and both child nodes are all unmasked or masked with
2844 // same predicates.
2845 // 2) Masked parent can be packed with left child if it is predicated
2846 // and both have same predicates.
2847 // 3) Masked parent can be packed with right child if its un-predicated
2848 // or has matching predication condition.
2849 // 4) An unmasked parent can be packed with an unmasked child.
2850 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2851 assert(partition.size() == 0, "not empty");
2852 assert(inputs.size() == 0, "not empty");
2853 if (is_vector_ternary_bitwise_op(n)) {
2854 return false;
2855 }
2856
2857 bool is_unary_op = is_vector_unary_bitwise_op(n);
2858 if (is_unary_op) {
2859 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2860 return false; // too few inputs
2861 }
2862
2863 bool pack_left_child = true;
2864 bool pack_right_child = true;
2865
2866 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2867 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2868
2869 int left_child_input_cnt = 0;
2870 int right_child_input_cnt = 0;
2871
2872 bool parent_is_predicated = n->is_predicated_vector();
2873 bool left_child_predicated = n->in(1)->is_predicated_vector();
2874 bool right_child_predicated = n->in(2)->is_predicated_vector();
2875
2876 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2877 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2878 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2879
2880 do {
2881 if (pack_left_child && left_child_LOP &&
2882 ((!parent_is_predicated && !left_child_predicated) ||
2883 ((parent_is_predicated && left_child_predicated &&
2884 parent_pred == left_child_pred)))) {
2885 partition.push(n->in(1));
2886 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2887 } else {
2888 inputs.push(n->in(1));
2889 left_child_input_cnt = 1;
2890 }
2891
2892 if (pack_right_child && right_child_LOP &&
2893 (!right_child_predicated ||
2894 (right_child_predicated && parent_is_predicated &&
2895 parent_pred == right_child_pred))) {
2896 partition.push(n->in(2));
2897 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2898 } else {
2899 inputs.push(n->in(2));
2900 right_child_input_cnt = 1;
2901 }
2902
2903 if (inputs.size() > 3) {
2904 assert(partition.size() > 0, "");
2905 inputs.clear();
2906 partition.clear();
2907 if (left_child_input_cnt > right_child_input_cnt) {
2908 pack_left_child = false;
2909 } else {
2910 pack_right_child = false;
2911 }
2912 } else {
2913 break;
2914 }
2915 } while(true);
2916
2917 if(partition.size()) {
2918 partition.push(n);
2919 }
2920
2921 return (partition.size() == 2 || partition.size() == 3) &&
2922 (inputs.size() == 2 || inputs.size() == 3);
2923 }
2924
2925 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2926 assert(is_vector_bitwise_op(n), "not a root");
2927
2928 visited.set(n->_idx);
2929
2930 // 1) Do a DFS walk over the logic cone.
2931 for (uint i = 1; i < n->req(); i++) {
2932 Node* in = n->in(i);
2933 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2934 process_logic_cone_root(igvn, in, visited);
2935 }
2936 }
2937
2938 // 2) Bottom up traversal: Merge node[s] with
2939 // the parent to form macro logic node.
2940 Unique_Node_List partition;
2941 Unique_Node_List inputs;
2942 if (compute_logic_cone(n, partition, inputs)) {
2943 const TypeVect* vt = n->bottom_type()->is_vect();
2944 Node* pn = partition.at(partition.size() - 1);
2945 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2946 if (mask == nullptr ||
2947 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2948 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2949 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2950 igvn.replace_node(n, macro_logic);
2951 }
2952 }
2953 }
2954
2955 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2956 ResourceMark rm;
2957 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2958 Unique_Node_List list;
2959 collect_logic_cone_roots(list);
2960
2961 while (list.size() > 0) {
2962 Node* n = list.pop();
2963 const TypeVect* vt = n->bottom_type()->is_vect();
2964 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2965 if (supported) {
2966 VectorSet visited(comp_arena());
2967 process_logic_cone_root(igvn, n, visited);
2968 }
2969 }
2970 }
2971 }
2972
2973 //------------------------------Code_Gen---------------------------------------
2974 // Given a graph, generate code for it
2975 void Compile::Code_Gen() {
2976 if (failing()) {
2977 return;
2978 }
2979
2980 // Perform instruction selection. You might think we could reclaim Matcher
2981 // memory PDQ, but actually the Matcher is used in generating spill code.
2982 // Internals of the Matcher (including some VectorSets) must remain live
2983 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2984 // set a bit in reclaimed memory.
2985
2986 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2987 // nodes. Mapping is only valid at the root of each matched subtree.
2988 NOT_PRODUCT( verify_graph_edges(); )
2989
2990 Matcher matcher;
2991 _matcher = &matcher;
2992 {
2993 TracePhase tp(_t_matcher);
2994 matcher.match();
2995 if (failing()) {
2996 return;
2997 }
2998 }
2999 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3000 // nodes. Mapping is only valid at the root of each matched subtree.
3001 NOT_PRODUCT( verify_graph_edges(); )
3002
3003 // If you have too many nodes, or if matching has failed, bail out
3004 check_node_count(0, "out of nodes matching instructions");
3005 if (failing()) {
3006 return;
3007 }
3008
3009 print_method(PHASE_MATCHING, 2);
3010
3011 // Build a proper-looking CFG
3012 PhaseCFG cfg(node_arena(), root(), matcher);
3013 if (failing()) {
3014 return;
3015 }
3016 _cfg = &cfg;
3017 {
3018 TracePhase tp(_t_scheduler);
3019 bool success = cfg.do_global_code_motion();
3020 if (!success) {
3021 return;
3022 }
3023
3024 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3025 NOT_PRODUCT( verify_graph_edges(); )
3026 cfg.verify();
3027 if (failing()) {
3028 return;
3029 }
3030 }
3031
3032 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3033 _regalloc = ®alloc;
3034 {
3035 TracePhase tp(_t_registerAllocation);
3036 // Perform register allocation. After Chaitin, use-def chains are
3037 // no longer accurate (at spill code) and so must be ignored.
3038 // Node->LRG->reg mappings are still accurate.
3039 _regalloc->Register_Allocate();
3040
3041 // Bail out if the allocator builds too many nodes
3042 if (failing()) {
3043 return;
3044 }
3045
3046 print_method(PHASE_REGISTER_ALLOCATION, 2);
3047 }
3048
3049 // Prior to register allocation we kept empty basic blocks in case the
3050 // the allocator needed a place to spill. After register allocation we
3051 // are not adding any new instructions. If any basic block is empty, we
3052 // can now safely remove it.
3053 {
3054 TracePhase tp(_t_blockOrdering);
3055 cfg.remove_empty_blocks();
3056 if (do_freq_based_layout()) {
3057 PhaseBlockLayout layout(cfg);
3058 } else {
3059 cfg.set_loop_alignment();
3060 }
3061 cfg.fixup_flow();
3062 cfg.remove_unreachable_blocks();
3063 cfg.verify_dominator_tree();
3064 print_method(PHASE_BLOCK_ORDERING, 3);
3065 }
3066
3067 // Apply peephole optimizations
3068 if( OptoPeephole ) {
3069 TracePhase tp(_t_peephole);
3070 PhasePeephole peep( _regalloc, cfg);
3071 peep.do_transform();
3072 print_method(PHASE_PEEPHOLE, 3);
3073 }
3074
3075 // Do late expand if CPU requires this.
3076 if (Matcher::require_postalloc_expand) {
3077 TracePhase tp(_t_postalloc_expand);
3078 cfg.postalloc_expand(_regalloc);
3079 print_method(PHASE_POSTALLOC_EXPAND, 3);
3080 }
3081
3082 #ifdef ASSERT
3083 {
3084 CompilationMemoryStatistic::do_test_allocations();
3085 if (failing()) return;
3086 }
3087 #endif
3088
3089 // Convert Nodes to instruction bits in a buffer
3090 {
3091 TracePhase tp(_t_output);
3092 PhaseOutput output;
3093 output.Output();
3094 if (failing()) return;
3095 output.install();
3096 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3097 }
3098
3099 // He's dead, Jim.
3100 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3101 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3102 }
3103
3104 //------------------------------Final_Reshape_Counts---------------------------
3105 // This class defines counters to help identify when a method
3106 // may/must be executed using hardware with only 24-bit precision.
3107 struct Final_Reshape_Counts : public StackObj {
3108 int _call_count; // count non-inlined 'common' calls
3109 int _float_count; // count float ops requiring 24-bit precision
3110 int _double_count; // count double ops requiring more precision
3111 int _java_call_count; // count non-inlined 'java' calls
3112 int _inner_loop_count; // count loops which need alignment
3113 VectorSet _visited; // Visitation flags
3114 Node_List _tests; // Set of IfNodes & PCTableNodes
3115
3116 Final_Reshape_Counts() :
3117 _call_count(0), _float_count(0), _double_count(0),
3118 _java_call_count(0), _inner_loop_count(0) { }
3119
3120 void inc_call_count () { _call_count ++; }
3121 void inc_float_count () { _float_count ++; }
3122 void inc_double_count() { _double_count++; }
3123 void inc_java_call_count() { _java_call_count++; }
3124 void inc_inner_loop_count() { _inner_loop_count++; }
3125
3126 int get_call_count () const { return _call_count ; }
3127 int get_float_count () const { return _float_count ; }
3128 int get_double_count() const { return _double_count; }
3129 int get_java_call_count() const { return _java_call_count; }
3130 int get_inner_loop_count() const { return _inner_loop_count; }
3131 };
3132
3133 //------------------------------final_graph_reshaping_impl----------------------
3134 // Implement items 1-5 from final_graph_reshaping below.
3135 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3136
3137 if ( n->outcnt() == 0 ) return; // dead node
3138 uint nop = n->Opcode();
3139
3140 // Check for 2-input instruction with "last use" on right input.
3141 // Swap to left input. Implements item (2).
3142 if( n->req() == 3 && // two-input instruction
3143 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3144 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3145 n->in(2)->outcnt() == 1 &&// right use IS a last use
3146 !n->in(2)->is_Con() ) { // right use is not a constant
3147 // Check for commutative opcode
3148 switch( nop ) {
3149 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3150 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3151 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3152 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3153 case Op_AndL: case Op_XorL: case Op_OrL:
3154 case Op_AndI: case Op_XorI: case Op_OrI: {
3155 // Move "last use" input to left by swapping inputs
3156 n->swap_edges(1, 2);
3157 break;
3158 }
3159 default:
3160 break;
3161 }
3162 }
3163
3164 #ifdef ASSERT
3165 if( n->is_Mem() ) {
3166 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3167 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3168 // oop will be recorded in oop map if load crosses safepoint
3169 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3170 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3171 "raw memory operations should have control edge");
3172 }
3173 if (n->is_MemBar()) {
3174 MemBarNode* mb = n->as_MemBar();
3175 if (mb->trailing_store() || mb->trailing_load_store()) {
3176 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3177 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3178 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3179 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3180 } else if (mb->leading()) {
3181 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3182 }
3183 }
3184 #endif
3185 // Count FPU ops and common calls, implements item (3)
3186 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3187 if (!gc_handled) {
3188 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3189 }
3190
3191 // Collect CFG split points
3192 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3193 frc._tests.push(n);
3194 }
3195 }
3196
3197 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3198 if (!UseDivMod) {
3199 return;
3200 }
3201
3202 // Check if "a % b" and "a / b" both exist
3203 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3204 if (d == nullptr) {
3205 return;
3206 }
3207
3208 // Replace them with a fused divmod if supported
3209 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3210 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3211 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3212 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3213 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3214 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3215 // DivMod node so the dependency is not lost.
3216 divmod->add_prec_from(n);
3217 divmod->add_prec_from(d);
3218 d->subsume_by(divmod->div_proj(), this);
3219 n->subsume_by(divmod->mod_proj(), this);
3220 } else {
3221 // Replace "a % b" with "a - ((a / b) * b)"
3222 Node* mult = MulNode::make(d, d->in(2), bt);
3223 Node* sub = SubNode::make(d->in(1), mult, bt);
3224 n->subsume_by(sub, this);
3225 }
3226 }
3227
3228 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3229 switch( nop ) {
3230 // Count all float operations that may use FPU
3231 case Op_AddF:
3232 case Op_SubF:
3233 case Op_MulF:
3234 case Op_DivF:
3235 case Op_NegF:
3236 case Op_ModF:
3237 case Op_ConvI2F:
3238 case Op_ConF:
3239 case Op_CmpF:
3240 case Op_CmpF3:
3241 case Op_StoreF:
3242 case Op_LoadF:
3243 // case Op_ConvL2F: // longs are split into 32-bit halves
3244 frc.inc_float_count();
3245 break;
3246
3247 case Op_ConvF2D:
3248 case Op_ConvD2F:
3249 frc.inc_float_count();
3250 frc.inc_double_count();
3251 break;
3252
3253 // Count all double operations that may use FPU
3254 case Op_AddD:
3255 case Op_SubD:
3256 case Op_MulD:
3257 case Op_DivD:
3258 case Op_NegD:
3259 case Op_ModD:
3260 case Op_ConvI2D:
3261 case Op_ConvD2I:
3262 // case Op_ConvL2D: // handled by leaf call
3263 // case Op_ConvD2L: // handled by leaf call
3264 case Op_ConD:
3265 case Op_CmpD:
3266 case Op_CmpD3:
3267 case Op_StoreD:
3268 case Op_LoadD:
3269 case Op_LoadD_unaligned:
3270 frc.inc_double_count();
3271 break;
3272 case Op_Opaque1: // Remove Opaque Nodes before matching
3273 n->subsume_by(n->in(1), this);
3274 break;
3275 case Op_CallStaticJava:
3276 case Op_CallJava:
3277 case Op_CallDynamicJava:
3278 frc.inc_java_call_count(); // Count java call site;
3279 case Op_CallRuntime:
3280 case Op_CallLeaf:
3281 case Op_CallLeafVector:
3282 case Op_CallLeafNoFP: {
3283 assert (n->is_Call(), "");
3284 CallNode *call = n->as_Call();
3285 // Count call sites where the FP mode bit would have to be flipped.
3286 // Do not count uncommon runtime calls:
3287 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3288 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3289 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3290 frc.inc_call_count(); // Count the call site
3291 } else { // See if uncommon argument is shared
3292 Node *n = call->in(TypeFunc::Parms);
3293 int nop = n->Opcode();
3294 // Clone shared simple arguments to uncommon calls, item (1).
3295 if (n->outcnt() > 1 &&
3296 !n->is_Proj() &&
3297 nop != Op_CreateEx &&
3298 nop != Op_CheckCastPP &&
3299 nop != Op_DecodeN &&
3300 nop != Op_DecodeNKlass &&
3301 !n->is_Mem() &&
3302 !n->is_Phi()) {
3303 Node *x = n->clone();
3304 call->set_req(TypeFunc::Parms, x);
3305 }
3306 }
3307 break;
3308 }
3309 case Op_StoreB:
3310 case Op_StoreC:
3311 case Op_StoreI:
3312 case Op_StoreL:
3313 case Op_CompareAndSwapB:
3314 case Op_CompareAndSwapS:
3315 case Op_CompareAndSwapI:
3316 case Op_CompareAndSwapL:
3317 case Op_CompareAndSwapP:
3318 case Op_CompareAndSwapN:
3319 case Op_WeakCompareAndSwapB:
3320 case Op_WeakCompareAndSwapS:
3321 case Op_WeakCompareAndSwapI:
3322 case Op_WeakCompareAndSwapL:
3323 case Op_WeakCompareAndSwapP:
3324 case Op_WeakCompareAndSwapN:
3325 case Op_CompareAndExchangeB:
3326 case Op_CompareAndExchangeS:
3327 case Op_CompareAndExchangeI:
3328 case Op_CompareAndExchangeL:
3329 case Op_CompareAndExchangeP:
3330 case Op_CompareAndExchangeN:
3331 case Op_GetAndAddS:
3332 case Op_GetAndAddB:
3333 case Op_GetAndAddI:
3334 case Op_GetAndAddL:
3335 case Op_GetAndSetS:
3336 case Op_GetAndSetB:
3337 case Op_GetAndSetI:
3338 case Op_GetAndSetL:
3339 case Op_GetAndSetP:
3340 case Op_GetAndSetN:
3341 case Op_StoreP:
3342 case Op_StoreN:
3343 case Op_StoreNKlass:
3344 case Op_LoadB:
3345 case Op_LoadUB:
3346 case Op_LoadUS:
3347 case Op_LoadI:
3348 case Op_LoadKlass:
3349 case Op_LoadNKlass:
3350 case Op_LoadL:
3351 case Op_LoadL_unaligned:
3352 case Op_LoadP:
3353 case Op_LoadN:
3354 case Op_LoadRange:
3355 case Op_LoadS:
3356 break;
3357
3358 case Op_AddP: { // Assert sane base pointers
3359 Node *addp = n->in(AddPNode::Address);
3360 assert( !addp->is_AddP() ||
3361 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3362 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3363 "Base pointers must match (addp %u)", addp->_idx );
3364 #ifdef _LP64
3365 if ((UseCompressedOops || UseCompressedClassPointers) &&
3366 addp->Opcode() == Op_ConP &&
3367 addp == n->in(AddPNode::Base) &&
3368 n->in(AddPNode::Offset)->is_Con()) {
3369 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3370 // on the platform and on the compressed oops mode.
3371 // Use addressing with narrow klass to load with offset on x86.
3372 // Some platforms can use the constant pool to load ConP.
3373 // Do this transformation here since IGVN will convert ConN back to ConP.
3374 const Type* t = addp->bottom_type();
3375 bool is_oop = t->isa_oopptr() != nullptr;
3376 bool is_klass = t->isa_klassptr() != nullptr;
3377
3378 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3379 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3380 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3381 Node* nn = nullptr;
3382
3383 int op = is_oop ? Op_ConN : Op_ConNKlass;
3384
3385 // Look for existing ConN node of the same exact type.
3386 Node* r = root();
3387 uint cnt = r->outcnt();
3388 for (uint i = 0; i < cnt; i++) {
3389 Node* m = r->raw_out(i);
3390 if (m!= nullptr && m->Opcode() == op &&
3391 m->bottom_type()->make_ptr() == t) {
3392 nn = m;
3393 break;
3394 }
3395 }
3396 if (nn != nullptr) {
3397 // Decode a narrow oop to match address
3398 // [R12 + narrow_oop_reg<<3 + offset]
3399 if (is_oop) {
3400 nn = new DecodeNNode(nn, t);
3401 } else {
3402 nn = new DecodeNKlassNode(nn, t);
3403 }
3404 // Check for succeeding AddP which uses the same Base.
3405 // Otherwise we will run into the assertion above when visiting that guy.
3406 for (uint i = 0; i < n->outcnt(); ++i) {
3407 Node *out_i = n->raw_out(i);
3408 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3409 out_i->set_req(AddPNode::Base, nn);
3410 #ifdef ASSERT
3411 for (uint j = 0; j < out_i->outcnt(); ++j) {
3412 Node *out_j = out_i->raw_out(j);
3413 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3414 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3415 }
3416 #endif
3417 }
3418 }
3419 n->set_req(AddPNode::Base, nn);
3420 n->set_req(AddPNode::Address, nn);
3421 if (addp->outcnt() == 0) {
3422 addp->disconnect_inputs(this);
3423 }
3424 }
3425 }
3426 }
3427 #endif
3428 break;
3429 }
3430
3431 case Op_CastPP: {
3432 // Remove CastPP nodes to gain more freedom during scheduling but
3433 // keep the dependency they encode as control or precedence edges
3434 // (if control is set already) on memory operations. Some CastPP
3435 // nodes don't have a control (don't carry a dependency): skip
3436 // those.
3437 if (n->in(0) != nullptr) {
3438 ResourceMark rm;
3439 Unique_Node_List wq;
3440 wq.push(n);
3441 for (uint next = 0; next < wq.size(); ++next) {
3442 Node *m = wq.at(next);
3443 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3444 Node* use = m->fast_out(i);
3445 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3446 use->ensure_control_or_add_prec(n->in(0));
3447 } else {
3448 switch(use->Opcode()) {
3449 case Op_AddP:
3450 case Op_DecodeN:
3451 case Op_DecodeNKlass:
3452 case Op_CheckCastPP:
3453 case Op_CastPP:
3454 wq.push(use);
3455 break;
3456 }
3457 }
3458 }
3459 }
3460 }
3461 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3462 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3463 Node* in1 = n->in(1);
3464 const Type* t = n->bottom_type();
3465 Node* new_in1 = in1->clone();
3466 new_in1->as_DecodeN()->set_type(t);
3467
3468 if (!Matcher::narrow_oop_use_complex_address()) {
3469 //
3470 // x86, ARM and friends can handle 2 adds in addressing mode
3471 // and Matcher can fold a DecodeN node into address by using
3472 // a narrow oop directly and do implicit null check in address:
3473 //
3474 // [R12 + narrow_oop_reg<<3 + offset]
3475 // NullCheck narrow_oop_reg
3476 //
3477 // On other platforms (Sparc) we have to keep new DecodeN node and
3478 // use it to do implicit null check in address:
3479 //
3480 // decode_not_null narrow_oop_reg, base_reg
3481 // [base_reg + offset]
3482 // NullCheck base_reg
3483 //
3484 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3485 // to keep the information to which null check the new DecodeN node
3486 // corresponds to use it as value in implicit_null_check().
3487 //
3488 new_in1->set_req(0, n->in(0));
3489 }
3490
3491 n->subsume_by(new_in1, this);
3492 if (in1->outcnt() == 0) {
3493 in1->disconnect_inputs(this);
3494 }
3495 } else {
3496 n->subsume_by(n->in(1), this);
3497 if (n->outcnt() == 0) {
3498 n->disconnect_inputs(this);
3499 }
3500 }
3501 break;
3502 }
3503 case Op_CastII: {
3504 n->as_CastII()->remove_range_check_cast(this);
3505 break;
3506 }
3507 #ifdef _LP64
3508 case Op_CmpP:
3509 // Do this transformation here to preserve CmpPNode::sub() and
3510 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3511 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3512 Node* in1 = n->in(1);
3513 Node* in2 = n->in(2);
3514 if (!in1->is_DecodeNarrowPtr()) {
3515 in2 = in1;
3516 in1 = n->in(2);
3517 }
3518 assert(in1->is_DecodeNarrowPtr(), "sanity");
3519
3520 Node* new_in2 = nullptr;
3521 if (in2->is_DecodeNarrowPtr()) {
3522 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3523 new_in2 = in2->in(1);
3524 } else if (in2->Opcode() == Op_ConP) {
3525 const Type* t = in2->bottom_type();
3526 if (t == TypePtr::NULL_PTR) {
3527 assert(in1->is_DecodeN(), "compare klass to null?");
3528 // Don't convert CmpP null check into CmpN if compressed
3529 // oops implicit null check is not generated.
3530 // This will allow to generate normal oop implicit null check.
3531 if (Matcher::gen_narrow_oop_implicit_null_checks())
3532 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3533 //
3534 // This transformation together with CastPP transformation above
3535 // will generated code for implicit null checks for compressed oops.
3536 //
3537 // The original code after Optimize()
3538 //
3539 // LoadN memory, narrow_oop_reg
3540 // decode narrow_oop_reg, base_reg
3541 // CmpP base_reg, nullptr
3542 // CastPP base_reg // NotNull
3543 // Load [base_reg + offset], val_reg
3544 //
3545 // after these transformations will be
3546 //
3547 // LoadN memory, narrow_oop_reg
3548 // CmpN narrow_oop_reg, nullptr
3549 // decode_not_null narrow_oop_reg, base_reg
3550 // Load [base_reg + offset], val_reg
3551 //
3552 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3553 // since narrow oops can be used in debug info now (see the code in
3554 // final_graph_reshaping_walk()).
3555 //
3556 // At the end the code will be matched to
3557 // on x86:
3558 //
3559 // Load_narrow_oop memory, narrow_oop_reg
3560 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3561 // NullCheck narrow_oop_reg
3562 //
3563 // and on sparc:
3564 //
3565 // Load_narrow_oop memory, narrow_oop_reg
3566 // decode_not_null narrow_oop_reg, base_reg
3567 // Load [base_reg + offset], val_reg
3568 // NullCheck base_reg
3569 //
3570 } else if (t->isa_oopptr()) {
3571 new_in2 = ConNode::make(t->make_narrowoop());
3572 } else if (t->isa_klassptr()) {
3573 new_in2 = ConNode::make(t->make_narrowklass());
3574 }
3575 }
3576 if (new_in2 != nullptr) {
3577 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3578 n->subsume_by(cmpN, this);
3579 if (in1->outcnt() == 0) {
3580 in1->disconnect_inputs(this);
3581 }
3582 if (in2->outcnt() == 0) {
3583 in2->disconnect_inputs(this);
3584 }
3585 }
3586 }
3587 break;
3588
3589 case Op_DecodeN:
3590 case Op_DecodeNKlass:
3591 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3592 // DecodeN could be pinned when it can't be fold into
3593 // an address expression, see the code for Op_CastPP above.
3594 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3595 break;
3596
3597 case Op_EncodeP:
3598 case Op_EncodePKlass: {
3599 Node* in1 = n->in(1);
3600 if (in1->is_DecodeNarrowPtr()) {
3601 n->subsume_by(in1->in(1), this);
3602 } else if (in1->Opcode() == Op_ConP) {
3603 const Type* t = in1->bottom_type();
3604 if (t == TypePtr::NULL_PTR) {
3605 assert(t->isa_oopptr(), "null klass?");
3606 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3607 } else if (t->isa_oopptr()) {
3608 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3609 } else if (t->isa_klassptr()) {
3610 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3611 }
3612 }
3613 if (in1->outcnt() == 0) {
3614 in1->disconnect_inputs(this);
3615 }
3616 break;
3617 }
3618
3619 case Op_Proj: {
3620 if (OptimizeStringConcat || IncrementalInline) {
3621 ProjNode* proj = n->as_Proj();
3622 if (proj->_is_io_use) {
3623 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3624 // Separate projections were used for the exception path which
3625 // are normally removed by a late inline. If it wasn't inlined
3626 // then they will hang around and should just be replaced with
3627 // the original one. Merge them.
3628 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3629 if (non_io_proj != nullptr) {
3630 proj->subsume_by(non_io_proj , this);
3631 }
3632 }
3633 }
3634 break;
3635 }
3636
3637 case Op_Phi:
3638 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3639 // The EncodeP optimization may create Phi with the same edges
3640 // for all paths. It is not handled well by Register Allocator.
3641 Node* unique_in = n->in(1);
3642 assert(unique_in != nullptr, "");
3643 uint cnt = n->req();
3644 for (uint i = 2; i < cnt; i++) {
3645 Node* m = n->in(i);
3646 assert(m != nullptr, "");
3647 if (unique_in != m)
3648 unique_in = nullptr;
3649 }
3650 if (unique_in != nullptr) {
3651 n->subsume_by(unique_in, this);
3652 }
3653 }
3654 break;
3655
3656 #endif
3657
3658 case Op_ModI:
3659 handle_div_mod_op(n, T_INT, false);
3660 break;
3661
3662 case Op_ModL:
3663 handle_div_mod_op(n, T_LONG, false);
3664 break;
3665
3666 case Op_UModI:
3667 handle_div_mod_op(n, T_INT, true);
3668 break;
3669
3670 case Op_UModL:
3671 handle_div_mod_op(n, T_LONG, true);
3672 break;
3673
3674 case Op_LoadVector:
3675 case Op_StoreVector:
3676 #ifdef ASSERT
3677 // Add VerifyVectorAlignment node between adr and load / store.
3678 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3679 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3680 n->as_StoreVector()->must_verify_alignment();
3681 if (must_verify_alignment) {
3682 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3683 n->as_StoreVector()->memory_size();
3684 // The memory access should be aligned to the vector width in bytes.
3685 // However, the underlying array is possibly less well aligned, but at least
3686 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3687 // a loop we can expect at least the following alignment:
3688 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3689 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3690 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3691 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3692 jlong mask = guaranteed_alignment - 1;
3693 Node* mask_con = ConLNode::make(mask);
3694 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3695 n->set_req(MemNode::Address, va);
3696 }
3697 }
3698 #endif
3699 break;
3700
3701 case Op_LoadVectorGather:
3702 case Op_StoreVectorScatter:
3703 case Op_LoadVectorGatherMasked:
3704 case Op_StoreVectorScatterMasked:
3705 case Op_VectorCmpMasked:
3706 case Op_VectorMaskGen:
3707 case Op_LoadVectorMasked:
3708 case Op_StoreVectorMasked:
3709 break;
3710
3711 case Op_AddReductionVI:
3712 case Op_AddReductionVL:
3713 case Op_AddReductionVF:
3714 case Op_AddReductionVD:
3715 case Op_MulReductionVI:
3716 case Op_MulReductionVL:
3717 case Op_MulReductionVF:
3718 case Op_MulReductionVD:
3719 case Op_MinReductionV:
3720 case Op_MaxReductionV:
3721 case Op_AndReductionV:
3722 case Op_OrReductionV:
3723 case Op_XorReductionV:
3724 break;
3725
3726 case Op_PackB:
3727 case Op_PackS:
3728 case Op_PackI:
3729 case Op_PackF:
3730 case Op_PackL:
3731 case Op_PackD:
3732 if (n->req()-1 > 2) {
3733 // Replace many operand PackNodes with a binary tree for matching
3734 PackNode* p = (PackNode*) n;
3735 Node* btp = p->binary_tree_pack(1, n->req());
3736 n->subsume_by(btp, this);
3737 }
3738 break;
3739 case Op_Loop:
3740 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3741 case Op_CountedLoop:
3742 case Op_LongCountedLoop:
3743 case Op_OuterStripMinedLoop:
3744 if (n->as_Loop()->is_inner_loop()) {
3745 frc.inc_inner_loop_count();
3746 }
3747 n->as_Loop()->verify_strip_mined(0);
3748 break;
3749 case Op_LShiftI:
3750 case Op_RShiftI:
3751 case Op_URShiftI:
3752 case Op_LShiftL:
3753 case Op_RShiftL:
3754 case Op_URShiftL:
3755 if (Matcher::need_masked_shift_count) {
3756 // The cpu's shift instructions don't restrict the count to the
3757 // lower 5/6 bits. We need to do the masking ourselves.
3758 Node* in2 = n->in(2);
3759 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3760 const TypeInt* t = in2->find_int_type();
3761 if (t != nullptr && t->is_con()) {
3762 juint shift = t->get_con();
3763 if (shift > mask) { // Unsigned cmp
3764 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3765 }
3766 } else {
3767 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3768 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3769 n->set_req(2, shift);
3770 }
3771 }
3772 if (in2->outcnt() == 0) { // Remove dead node
3773 in2->disconnect_inputs(this);
3774 }
3775 }
3776 break;
3777 case Op_MemBarStoreStore:
3778 case Op_MemBarRelease:
3779 // Break the link with AllocateNode: it is no longer useful and
3780 // confuses register allocation.
3781 if (n->req() > MemBarNode::Precedent) {
3782 n->set_req(MemBarNode::Precedent, top());
3783 }
3784 break;
3785 case Op_MemBarAcquire: {
3786 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3787 // At parse time, the trailing MemBarAcquire for a volatile load
3788 // is created with an edge to the load. After optimizations,
3789 // that input may be a chain of Phis. If those phis have no
3790 // other use, then the MemBarAcquire keeps them alive and
3791 // register allocation can be confused.
3792 dead_nodes.push(n->in(MemBarNode::Precedent));
3793 n->set_req(MemBarNode::Precedent, top());
3794 }
3795 break;
3796 }
3797 case Op_Blackhole:
3798 break;
3799 case Op_RangeCheck: {
3800 RangeCheckNode* rc = n->as_RangeCheck();
3801 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3802 n->subsume_by(iff, this);
3803 frc._tests.push(iff);
3804 break;
3805 }
3806 case Op_ConvI2L: {
3807 if (!Matcher::convi2l_type_required) {
3808 // Code generation on some platforms doesn't need accurate
3809 // ConvI2L types. Widening the type can help remove redundant
3810 // address computations.
3811 n->as_Type()->set_type(TypeLong::INT);
3812 ResourceMark rm;
3813 Unique_Node_List wq;
3814 wq.push(n);
3815 for (uint next = 0; next < wq.size(); next++) {
3816 Node *m = wq.at(next);
3817
3818 for(;;) {
3819 // Loop over all nodes with identical inputs edges as m
3820 Node* k = m->find_similar(m->Opcode());
3821 if (k == nullptr) {
3822 break;
3823 }
3824 // Push their uses so we get a chance to remove node made
3825 // redundant
3826 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3827 Node* u = k->fast_out(i);
3828 if (u->Opcode() == Op_LShiftL ||
3829 u->Opcode() == Op_AddL ||
3830 u->Opcode() == Op_SubL ||
3831 u->Opcode() == Op_AddP) {
3832 wq.push(u);
3833 }
3834 }
3835 // Replace all nodes with identical edges as m with m
3836 k->subsume_by(m, this);
3837 }
3838 }
3839 }
3840 break;
3841 }
3842 case Op_CmpUL: {
3843 if (!Matcher::has_match_rule(Op_CmpUL)) {
3844 // No support for unsigned long comparisons
3845 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3846 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3847 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3848 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3849 Node* andl = new AndLNode(orl, remove_sign_mask);
3850 Node* cmp = new CmpLNode(andl, n->in(2));
3851 n->subsume_by(cmp, this);
3852 }
3853 break;
3854 }
3855 #ifdef ASSERT
3856 case Op_ConNKlass: {
3857 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3858 ciKlass* klass = tp->is_klassptr()->exact_klass();
3859 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3860 break;
3861 }
3862 #endif
3863 default:
3864 assert(!n->is_Call(), "");
3865 assert(!n->is_Mem(), "");
3866 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3867 break;
3868 }
3869 }
3870
3871 //------------------------------final_graph_reshaping_walk---------------------
3872 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3873 // requires that the walk visits a node's inputs before visiting the node.
3874 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3875 Unique_Node_List sfpt;
3876
3877 frc._visited.set(root->_idx); // first, mark node as visited
3878 uint cnt = root->req();
3879 Node *n = root;
3880 uint i = 0;
3881 while (true) {
3882 if (i < cnt) {
3883 // Place all non-visited non-null inputs onto stack
3884 Node* m = n->in(i);
3885 ++i;
3886 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3887 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3888 // compute worst case interpreter size in case of a deoptimization
3889 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3890
3891 sfpt.push(m);
3892 }
3893 cnt = m->req();
3894 nstack.push(n, i); // put on stack parent and next input's index
3895 n = m;
3896 i = 0;
3897 }
3898 } else {
3899 // Now do post-visit work
3900 final_graph_reshaping_impl(n, frc, dead_nodes);
3901 if (nstack.is_empty())
3902 break; // finished
3903 n = nstack.node(); // Get node from stack
3904 cnt = n->req();
3905 i = nstack.index();
3906 nstack.pop(); // Shift to the next node on stack
3907 }
3908 }
3909
3910 // Skip next transformation if compressed oops are not used.
3911 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3912 (!UseCompressedOops && !UseCompressedClassPointers))
3913 return;
3914
3915 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3916 // It could be done for an uncommon traps or any safepoints/calls
3917 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3918 while (sfpt.size() > 0) {
3919 n = sfpt.pop();
3920 JVMState *jvms = n->as_SafePoint()->jvms();
3921 assert(jvms != nullptr, "sanity");
3922 int start = jvms->debug_start();
3923 int end = n->req();
3924 bool is_uncommon = (n->is_CallStaticJava() &&
3925 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3926 for (int j = start; j < end; j++) {
3927 Node* in = n->in(j);
3928 if (in->is_DecodeNarrowPtr()) {
3929 bool safe_to_skip = true;
3930 if (!is_uncommon ) {
3931 // Is it safe to skip?
3932 for (uint i = 0; i < in->outcnt(); i++) {
3933 Node* u = in->raw_out(i);
3934 if (!u->is_SafePoint() ||
3935 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3936 safe_to_skip = false;
3937 }
3938 }
3939 }
3940 if (safe_to_skip) {
3941 n->set_req(j, in->in(1));
3942 }
3943 if (in->outcnt() == 0) {
3944 in->disconnect_inputs(this);
3945 }
3946 }
3947 }
3948 }
3949 }
3950
3951 //------------------------------final_graph_reshaping--------------------------
3952 // Final Graph Reshaping.
3953 //
3954 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3955 // and not commoned up and forced early. Must come after regular
3956 // optimizations to avoid GVN undoing the cloning. Clone constant
3957 // inputs to Loop Phis; these will be split by the allocator anyways.
3958 // Remove Opaque nodes.
3959 // (2) Move last-uses by commutative operations to the left input to encourage
3960 // Intel update-in-place two-address operations and better register usage
3961 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3962 // calls canonicalizing them back.
3963 // (3) Count the number of double-precision FP ops, single-precision FP ops
3964 // and call sites. On Intel, we can get correct rounding either by
3965 // forcing singles to memory (requires extra stores and loads after each
3966 // FP bytecode) or we can set a rounding mode bit (requires setting and
3967 // clearing the mode bit around call sites). The mode bit is only used
3968 // if the relative frequency of single FP ops to calls is low enough.
3969 // This is a key transform for SPEC mpeg_audio.
3970 // (4) Detect infinite loops; blobs of code reachable from above but not
3971 // below. Several of the Code_Gen algorithms fail on such code shapes,
3972 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3973 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3974 // Detection is by looking for IfNodes where only 1 projection is
3975 // reachable from below or CatchNodes missing some targets.
3976 // (5) Assert for insane oop offsets in debug mode.
3977
3978 bool Compile::final_graph_reshaping() {
3979 // an infinite loop may have been eliminated by the optimizer,
3980 // in which case the graph will be empty.
3981 if (root()->req() == 1) {
3982 // Do not compile method that is only a trivial infinite loop,
3983 // since the content of the loop may have been eliminated.
3984 record_method_not_compilable("trivial infinite loop");
3985 return true;
3986 }
3987
3988 // Expensive nodes have their control input set to prevent the GVN
3989 // from freely commoning them. There's no GVN beyond this point so
3990 // no need to keep the control input. We want the expensive nodes to
3991 // be freely moved to the least frequent code path by gcm.
3992 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3993 for (int i = 0; i < expensive_count(); i++) {
3994 _expensive_nodes.at(i)->set_req(0, nullptr);
3995 }
3996
3997 Final_Reshape_Counts frc;
3998
3999 // Visit everybody reachable!
4000 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4001 Node_Stack nstack(live_nodes() >> 1);
4002 Unique_Node_List dead_nodes;
4003 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4004
4005 // Check for unreachable (from below) code (i.e., infinite loops).
4006 for( uint i = 0; i < frc._tests.size(); i++ ) {
4007 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4008 // Get number of CFG targets.
4009 // Note that PCTables include exception targets after calls.
4010 uint required_outcnt = n->required_outcnt();
4011 if (n->outcnt() != required_outcnt) {
4012 // Check for a few special cases. Rethrow Nodes never take the
4013 // 'fall-thru' path, so expected kids is 1 less.
4014 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4015 if (n->in(0)->in(0)->is_Call()) {
4016 CallNode* call = n->in(0)->in(0)->as_Call();
4017 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4018 required_outcnt--; // Rethrow always has 1 less kid
4019 } else if (call->req() > TypeFunc::Parms &&
4020 call->is_CallDynamicJava()) {
4021 // Check for null receiver. In such case, the optimizer has
4022 // detected that the virtual call will always result in a null
4023 // pointer exception. The fall-through projection of this CatchNode
4024 // will not be populated.
4025 Node* arg0 = call->in(TypeFunc::Parms);
4026 if (arg0->is_Type() &&
4027 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4028 required_outcnt--;
4029 }
4030 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4031 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4032 // Check for illegal array length. In such case, the optimizer has
4033 // detected that the allocation attempt will always result in an
4034 // exception. There is no fall-through projection of this CatchNode .
4035 assert(call->is_CallStaticJava(), "static call expected");
4036 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4037 uint valid_length_test_input = call->req() - 1;
4038 Node* valid_length_test = call->in(valid_length_test_input);
4039 call->del_req(valid_length_test_input);
4040 if (valid_length_test->find_int_con(1) == 0) {
4041 required_outcnt--;
4042 }
4043 dead_nodes.push(valid_length_test);
4044 assert(n->outcnt() == required_outcnt, "malformed control flow");
4045 continue;
4046 }
4047 }
4048 }
4049
4050 // Recheck with a better notion of 'required_outcnt'
4051 if (n->outcnt() != required_outcnt) {
4052 record_method_not_compilable("malformed control flow");
4053 return true; // Not all targets reachable!
4054 }
4055 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4056 CallNode* call = n->in(0)->in(0)->as_Call();
4057 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4058 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4059 assert(call->is_CallStaticJava(), "static call expected");
4060 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4061 uint valid_length_test_input = call->req() - 1;
4062 dead_nodes.push(call->in(valid_length_test_input));
4063 call->del_req(valid_length_test_input); // valid length test useless now
4064 }
4065 }
4066 // Check that I actually visited all kids. Unreached kids
4067 // must be infinite loops.
4068 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4069 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4070 record_method_not_compilable("infinite loop");
4071 return true; // Found unvisited kid; must be unreach
4072 }
4073
4074 // Here so verification code in final_graph_reshaping_walk()
4075 // always see an OuterStripMinedLoopEnd
4076 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4077 IfNode* init_iff = n->as_If();
4078 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4079 n->subsume_by(iff, this);
4080 }
4081 }
4082
4083 while (dead_nodes.size() > 0) {
4084 Node* m = dead_nodes.pop();
4085 if (m->outcnt() == 0 && m != top()) {
4086 for (uint j = 0; j < m->req(); j++) {
4087 Node* in = m->in(j);
4088 if (in != nullptr) {
4089 dead_nodes.push(in);
4090 }
4091 }
4092 m->disconnect_inputs(this);
4093 }
4094 }
4095
4096 #ifdef IA32
4097 // If original bytecodes contained a mixture of floats and doubles
4098 // check if the optimizer has made it homogeneous, item (3).
4099 if (UseSSE == 0 &&
4100 frc.get_float_count() > 32 &&
4101 frc.get_double_count() == 0 &&
4102 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4103 set_24_bit_selection_and_mode(false, true);
4104 }
4105 #endif // IA32
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 ciMethodData* md = method->method_data();
4121 if (md->is_empty()) {
4122 // Assume the trap has not occurred, or that it occurred only
4123 // because of a transient condition during start-up in the interpreter.
4124 return false;
4125 }
4126 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4127 if (md->has_trap_at(bci, m, reason) != 0) {
4128 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4129 // Also, if there are multiple reasons, or if there is no per-BCI record,
4130 // assume the worst.
4131 if (log())
4132 log()->elem("observe trap='%s' count='%d'",
4133 Deoptimization::trap_reason_name(reason),
4134 md->trap_count(reason));
4135 return true;
4136 } else {
4137 // Ignore method/bci and see if there have been too many globally.
4138 return too_many_traps(reason, md);
4139 }
4140 }
4141
4142 // Less-accurate variant which does not require a method and bci.
4143 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4144 ciMethodData* logmd) {
4145 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4146 // Too many traps globally.
4147 // Note that we use cumulative trap_count, not just md->trap_count.
4148 if (log()) {
4149 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4150 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4151 Deoptimization::trap_reason_name(reason),
4152 mcount, trap_count(reason));
4153 }
4154 return true;
4155 } else {
4156 // The coast is clear.
4157 return false;
4158 }
4159 }
4160
4161 //--------------------------too_many_recompiles--------------------------------
4162 // Report if there are too many recompiles at the current method and bci.
4163 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4164 // Is not eager to return true, since this will cause the compiler to use
4165 // Action_none for a trap point, to avoid too many recompilations.
4166 bool Compile::too_many_recompiles(ciMethod* method,
4167 int bci,
4168 Deoptimization::DeoptReason reason) {
4169 ciMethodData* md = method->method_data();
4170 if (md->is_empty()) {
4171 // Assume the trap has not occurred, or that it occurred only
4172 // because of a transient condition during start-up in the interpreter.
4173 return false;
4174 }
4175 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4176 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4177 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4178 Deoptimization::DeoptReason per_bc_reason
4179 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4180 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4181 if ((per_bc_reason == Deoptimization::Reason_none
4182 || md->has_trap_at(bci, m, reason) != 0)
4183 // The trap frequency measure we care about is the recompile count:
4184 && md->trap_recompiled_at(bci, m)
4185 && md->overflow_recompile_count() >= bc_cutoff) {
4186 // Do not emit a trap here if it has already caused recompilations.
4187 // Also, if there are multiple reasons, or if there is no per-BCI record,
4188 // assume the worst.
4189 if (log())
4190 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4191 Deoptimization::trap_reason_name(reason),
4192 md->trap_count(reason),
4193 md->overflow_recompile_count());
4194 return true;
4195 } else if (trap_count(reason) != 0
4196 && decompile_count() >= m_cutoff) {
4197 // Too many recompiles globally, and we have seen this sort of trap.
4198 // Use cumulative decompile_count, not just md->decompile_count.
4199 if (log())
4200 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4201 Deoptimization::trap_reason_name(reason),
4202 md->trap_count(reason), trap_count(reason),
4203 md->decompile_count(), decompile_count());
4204 return true;
4205 } else {
4206 // The coast is clear.
4207 return false;
4208 }
4209 }
4210
4211 // Compute when not to trap. Used by matching trap based nodes and
4212 // NullCheck optimization.
4213 void Compile::set_allowed_deopt_reasons() {
4214 _allowed_reasons = 0;
4215 if (is_method_compilation()) {
4216 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4217 assert(rs < BitsPerInt, "recode bit map");
4218 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4219 _allowed_reasons |= nth_bit(rs);
4220 }
4221 }
4222 }
4223 }
4224
4225 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4226 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4227 }
4228
4229 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4230 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4231 }
4232
4233 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4234 if (holder->is_initialized()) {
4235 return false;
4236 }
4237 if (holder->is_being_initialized()) {
4238 if (accessing_method->holder() == holder) {
4239 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4240 // <init>, or a static method. In all those cases, there was an initialization
4241 // barrier on the holder klass passed.
4242 if (accessing_method->is_static_initializer() ||
4243 accessing_method->is_object_initializer() ||
4244 accessing_method->is_static()) {
4245 return false;
4246 }
4247 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4248 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4249 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4250 // child class can become fully initialized while its parent class is still being initialized.
4251 if (accessing_method->is_static_initializer()) {
4252 return false;
4253 }
4254 }
4255 ciMethod* root = method(); // the root method of compilation
4256 if (root != accessing_method) {
4257 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4258 }
4259 }
4260 return true;
4261 }
4262
4263 #ifndef PRODUCT
4264 //------------------------------verify_bidirectional_edges---------------------
4265 // For each input edge to a node (ie - for each Use-Def edge), verify that
4266 // there is a corresponding Def-Use edge.
4267 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4268 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4269 uint stack_size = live_nodes() >> 4;
4270 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4271 if (root_and_safepoints != nullptr) {
4272 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4273 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4274 Node* root_or_safepoint = root_and_safepoints->at(i);
4275 // If the node is a safepoint, let's check if it still has a control input
4276 // Lack of control input signifies that this node was killed by CCP or
4277 // recursively by remove_globally_dead_node and it shouldn't be a starting
4278 // point.
4279 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4280 nstack.push(root_or_safepoint);
4281 }
4282 }
4283 } else {
4284 nstack.push(_root);
4285 }
4286
4287 while (nstack.size() > 0) {
4288 Node* n = nstack.pop();
4289 if (visited.member(n)) {
4290 continue;
4291 }
4292 visited.push(n);
4293
4294 // Walk over all input edges, checking for correspondence
4295 uint length = n->len();
4296 for (uint i = 0; i < length; i++) {
4297 Node* in = n->in(i);
4298 if (in != nullptr && !visited.member(in)) {
4299 nstack.push(in); // Put it on stack
4300 }
4301 if (in != nullptr && !in->is_top()) {
4302 // Count instances of `next`
4303 int cnt = 0;
4304 for (uint idx = 0; idx < in->_outcnt; idx++) {
4305 if (in->_out[idx] == n) {
4306 cnt++;
4307 }
4308 }
4309 assert(cnt > 0, "Failed to find Def-Use edge.");
4310 // Check for duplicate edges
4311 // walk the input array downcounting the input edges to n
4312 for (uint j = 0; j < length; j++) {
4313 if (n->in(j) == in) {
4314 cnt--;
4315 }
4316 }
4317 assert(cnt == 0, "Mismatched edge count.");
4318 } else if (in == nullptr) {
4319 assert(i == 0 || i >= n->req() ||
4320 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4321 (n->is_Unlock() && i == (n->req() - 1)) ||
4322 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4323 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4324 } else {
4325 assert(in->is_top(), "sanity");
4326 // Nothing to check.
4327 }
4328 }
4329 }
4330 }
4331
4332 //------------------------------verify_graph_edges---------------------------
4333 // Walk the Graph and verify that there is a one-to-one correspondence
4334 // between Use-Def edges and Def-Use edges in the graph.
4335 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4336 if (VerifyGraphEdges) {
4337 Unique_Node_List visited;
4338
4339 // Call graph walk to check edges
4340 verify_bidirectional_edges(visited, root_and_safepoints);
4341 if (no_dead_code) {
4342 // Now make sure that no visited node is used by an unvisited node.
4343 bool dead_nodes = false;
4344 Unique_Node_List checked;
4345 while (visited.size() > 0) {
4346 Node* n = visited.pop();
4347 checked.push(n);
4348 for (uint i = 0; i < n->outcnt(); i++) {
4349 Node* use = n->raw_out(i);
4350 if (checked.member(use)) continue; // already checked
4351 if (visited.member(use)) continue; // already in the graph
4352 if (use->is_Con()) continue; // a dead ConNode is OK
4353 // At this point, we have found a dead node which is DU-reachable.
4354 if (!dead_nodes) {
4355 tty->print_cr("*** Dead nodes reachable via DU edges:");
4356 dead_nodes = true;
4357 }
4358 use->dump(2);
4359 tty->print_cr("---");
4360 checked.push(use); // No repeats; pretend it is now checked.
4361 }
4362 }
4363 assert(!dead_nodes, "using nodes must be reachable from root");
4364 }
4365 }
4366 }
4367 #endif
4368
4369 // The Compile object keeps track of failure reasons separately from the ciEnv.
4370 // This is required because there is not quite a 1-1 relation between the
4371 // ciEnv and its compilation task and the Compile object. Note that one
4372 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4373 // to backtrack and retry without subsuming loads. Other than this backtracking
4374 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4375 // by the logic in C2Compiler.
4376 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4377 if (log() != nullptr) {
4378 log()->elem("failure reason='%s' phase='compile'", reason);
4379 }
4380 if (_failure_reason.get() == nullptr) {
4381 // Record the first failure reason.
4382 _failure_reason.set(reason);
4383 if (CaptureBailoutInformation) {
4384 _first_failure_details = new CompilationFailureInfo(reason);
4385 }
4386 } else {
4387 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4388 }
4389
4390 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4391 C->print_method(PHASE_FAILURE, 1);
4392 }
4393 _root = nullptr; // flush the graph, too
4394 }
4395
4396 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4397 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4398 _compile(Compile::current()),
4399 _log(nullptr),
4400 _dolog(CITimeVerbose)
4401 {
4402 assert(_compile != nullptr, "sanity check");
4403 assert(id != PhaseTraceId::_t_none, "Don't use none");
4404 if (_dolog) {
4405 _log = _compile->log();
4406 }
4407 if (_log != nullptr) {
4408 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4409 _log->stamp();
4410 _log->end_head();
4411 }
4412
4413 // Inform memory statistic, if enabled
4414 if (CompilationMemoryStatistic::enabled()) {
4415 CompilationMemoryStatistic::on_phase_start((int)id, name);
4416 }
4417 }
4418
4419 Compile::TracePhase::TracePhase(PhaseTraceId id)
4420 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4421
4422 Compile::TracePhase::~TracePhase() {
4423
4424 // Inform memory statistic, if enabled
4425 if (CompilationMemoryStatistic::enabled()) {
4426 CompilationMemoryStatistic::on_phase_end();
4427 }
4428
4429 if (_compile->failing_internal()) {
4430 if (_log != nullptr) {
4431 _log->done("phase");
4432 }
4433 return; // timing code, not stressing bailouts.
4434 }
4435 #ifdef ASSERT
4436 if (PrintIdealNodeCount) {
4437 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4438 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4439 }
4440
4441 if (VerifyIdealNodeCount) {
4442 _compile->print_missing_nodes();
4443 }
4444 #endif
4445
4446 if (_log != nullptr) {
4447 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4448 }
4449 }
4450
4451 //----------------------------static_subtype_check-----------------------------
4452 // Shortcut important common cases when superklass is exact:
4453 // (0) superklass is java.lang.Object (can occur in reflective code)
4454 // (1) subklass is already limited to a subtype of superklass => always ok
4455 // (2) subklass does not overlap with superklass => always fail
4456 // (3) superklass has NO subtypes and we can check with a simple compare.
4457 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4458 if (skip) {
4459 return SSC_full_test; // Let caller generate the general case.
4460 }
4461
4462 if (subk->is_java_subtype_of(superk)) {
4463 return SSC_always_true; // (0) and (1) this test cannot fail
4464 }
4465
4466 if (!subk->maybe_java_subtype_of(superk)) {
4467 return SSC_always_false; // (2) true path dead; no dynamic test needed
4468 }
4469
4470 const Type* superelem = superk;
4471 if (superk->isa_aryklassptr()) {
4472 int ignored;
4473 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4474 }
4475
4476 if (superelem->isa_instklassptr()) {
4477 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4478 if (!ik->has_subklass()) {
4479 if (!ik->is_final()) {
4480 // Add a dependency if there is a chance of a later subclass.
4481 dependencies()->assert_leaf_type(ik);
4482 }
4483 if (!superk->maybe_java_subtype_of(subk)) {
4484 return SSC_always_false;
4485 }
4486 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4487 }
4488 } else {
4489 // A primitive array type has no subtypes.
4490 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4491 }
4492
4493 return SSC_full_test;
4494 }
4495
4496 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4497 #ifdef _LP64
4498 // The scaled index operand to AddP must be a clean 64-bit value.
4499 // Java allows a 32-bit int to be incremented to a negative
4500 // value, which appears in a 64-bit register as a large
4501 // positive number. Using that large positive number as an
4502 // operand in pointer arithmetic has bad consequences.
4503 // On the other hand, 32-bit overflow is rare, and the possibility
4504 // can often be excluded, if we annotate the ConvI2L node with
4505 // a type assertion that its value is known to be a small positive
4506 // number. (The prior range check has ensured this.)
4507 // This assertion is used by ConvI2LNode::Ideal.
4508 int index_max = max_jint - 1; // array size is max_jint, index is one less
4509 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4510 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4511 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4512 #endif
4513 return idx;
4514 }
4515
4516 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4517 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4518 if (ctrl != nullptr) {
4519 // Express control dependency by a CastII node with a narrow type.
4520 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4521 // node from floating above the range check during loop optimizations. Otherwise, the
4522 // ConvI2L node may be eliminated independently of the range check, causing the data path
4523 // to become TOP while the control path is still there (although it's unreachable).
4524 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4525 value = phase->transform(value);
4526 }
4527 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4528 return phase->transform(new ConvI2LNode(value, ltype));
4529 }
4530
4531 void Compile::dump_print_inlining() {
4532 inline_printer()->print_on(tty);
4533 }
4534
4535 void Compile::log_late_inline(CallGenerator* cg) {
4536 if (log() != nullptr) {
4537 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4538 cg->unique_id());
4539 JVMState* p = cg->call_node()->jvms();
4540 while (p != nullptr) {
4541 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4542 p = p->caller();
4543 }
4544 log()->tail("late_inline");
4545 }
4546 }
4547
4548 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4549 log_late_inline(cg);
4550 if (log() != nullptr) {
4551 log()->inline_fail(msg);
4552 }
4553 }
4554
4555 void Compile::log_inline_id(CallGenerator* cg) {
4556 if (log() != nullptr) {
4557 // The LogCompilation tool needs a unique way to identify late
4558 // inline call sites. This id must be unique for this call site in
4559 // this compilation. Try to have it unique across compilations as
4560 // well because it can be convenient when grepping through the log
4561 // file.
4562 // Distinguish OSR compilations from others in case CICountOSR is
4563 // on.
4564 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4565 cg->set_unique_id(id);
4566 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4567 }
4568 }
4569
4570 void Compile::log_inline_failure(const char* msg) {
4571 if (C->log() != nullptr) {
4572 C->log()->inline_fail(msg);
4573 }
4574 }
4575
4576
4577 // Dump inlining replay data to the stream.
4578 // Don't change thread state and acquire any locks.
4579 void Compile::dump_inline_data(outputStream* out) {
4580 InlineTree* inl_tree = ilt();
4581 if (inl_tree != nullptr) {
4582 out->print(" inline %d", inl_tree->count());
4583 inl_tree->dump_replay_data(out);
4584 }
4585 }
4586
4587 void Compile::dump_inline_data_reduced(outputStream* out) {
4588 assert(ReplayReduce, "");
4589
4590 InlineTree* inl_tree = ilt();
4591 if (inl_tree == nullptr) {
4592 return;
4593 }
4594 // Enable iterative replay file reduction
4595 // Output "compile" lines for depth 1 subtrees,
4596 // simulating that those trees were compiled
4597 // instead of inlined.
4598 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4599 InlineTree* sub = inl_tree->subtrees().at(i);
4600 if (sub->inline_level() != 1) {
4601 continue;
4602 }
4603
4604 ciMethod* method = sub->method();
4605 int entry_bci = -1;
4606 int comp_level = env()->task()->comp_level();
4607 out->print("compile ");
4608 method->dump_name_as_ascii(out);
4609 out->print(" %d %d", entry_bci, comp_level);
4610 out->print(" inline %d", sub->count());
4611 sub->dump_replay_data(out, -1);
4612 out->cr();
4613 }
4614 }
4615
4616 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4617 if (n1->Opcode() < n2->Opcode()) return -1;
4618 else if (n1->Opcode() > n2->Opcode()) return 1;
4619
4620 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4621 for (uint i = 1; i < n1->req(); i++) {
4622 if (n1->in(i) < n2->in(i)) return -1;
4623 else if (n1->in(i) > n2->in(i)) return 1;
4624 }
4625
4626 return 0;
4627 }
4628
4629 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4630 Node* n1 = *n1p;
4631 Node* n2 = *n2p;
4632
4633 return cmp_expensive_nodes(n1, n2);
4634 }
4635
4636 void Compile::sort_expensive_nodes() {
4637 if (!expensive_nodes_sorted()) {
4638 _expensive_nodes.sort(cmp_expensive_nodes);
4639 }
4640 }
4641
4642 bool Compile::expensive_nodes_sorted() const {
4643 for (int i = 1; i < _expensive_nodes.length(); i++) {
4644 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4645 return false;
4646 }
4647 }
4648 return true;
4649 }
4650
4651 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4652 if (_expensive_nodes.length() == 0) {
4653 return false;
4654 }
4655
4656 assert(OptimizeExpensiveOps, "optimization off?");
4657
4658 // Take this opportunity to remove dead nodes from the list
4659 int j = 0;
4660 for (int i = 0; i < _expensive_nodes.length(); i++) {
4661 Node* n = _expensive_nodes.at(i);
4662 if (!n->is_unreachable(igvn)) {
4663 assert(n->is_expensive(), "should be expensive");
4664 _expensive_nodes.at_put(j, n);
4665 j++;
4666 }
4667 }
4668 _expensive_nodes.trunc_to(j);
4669
4670 // Then sort the list so that similar nodes are next to each other
4671 // and check for at least two nodes of identical kind with same data
4672 // inputs.
4673 sort_expensive_nodes();
4674
4675 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4676 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4677 return true;
4678 }
4679 }
4680
4681 return false;
4682 }
4683
4684 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4685 if (_expensive_nodes.length() == 0) {
4686 return;
4687 }
4688
4689 assert(OptimizeExpensiveOps, "optimization off?");
4690
4691 // Sort to bring similar nodes next to each other and clear the
4692 // control input of nodes for which there's only a single copy.
4693 sort_expensive_nodes();
4694
4695 int j = 0;
4696 int identical = 0;
4697 int i = 0;
4698 bool modified = false;
4699 for (; i < _expensive_nodes.length()-1; i++) {
4700 assert(j <= i, "can't write beyond current index");
4701 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4702 identical++;
4703 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4704 continue;
4705 }
4706 if (identical > 0) {
4707 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4708 identical = 0;
4709 } else {
4710 Node* n = _expensive_nodes.at(i);
4711 igvn.replace_input_of(n, 0, nullptr);
4712 igvn.hash_insert(n);
4713 modified = true;
4714 }
4715 }
4716 if (identical > 0) {
4717 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4718 } else if (_expensive_nodes.length() >= 1) {
4719 Node* n = _expensive_nodes.at(i);
4720 igvn.replace_input_of(n, 0, nullptr);
4721 igvn.hash_insert(n);
4722 modified = true;
4723 }
4724 _expensive_nodes.trunc_to(j);
4725 if (modified) {
4726 igvn.optimize();
4727 }
4728 }
4729
4730 void Compile::add_expensive_node(Node * n) {
4731 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4732 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4733 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4734 if (OptimizeExpensiveOps) {
4735 _expensive_nodes.append(n);
4736 } else {
4737 // Clear control input and let IGVN optimize expensive nodes if
4738 // OptimizeExpensiveOps is off.
4739 n->set_req(0, nullptr);
4740 }
4741 }
4742
4743 /**
4744 * Track coarsened Lock and Unlock nodes.
4745 */
4746
4747 class Lock_List : public Node_List {
4748 uint _origin_cnt;
4749 public:
4750 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4751 uint origin_cnt() const { return _origin_cnt; }
4752 };
4753
4754 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4755 int length = locks.length();
4756 if (length > 0) {
4757 // Have to keep this list until locks elimination during Macro nodes elimination.
4758 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4759 AbstractLockNode* alock = locks.at(0);
4760 BoxLockNode* box = alock->box_node()->as_BoxLock();
4761 for (int i = 0; i < length; i++) {
4762 AbstractLockNode* lock = locks.at(i);
4763 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4764 locks_list->push(lock);
4765 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4766 if (this_box != box) {
4767 // Locking regions (BoxLock) could be Unbalanced here:
4768 // - its coarsened locks were eliminated in earlier
4769 // macro nodes elimination followed by loop unroll
4770 // - it is OSR locking region (no Lock node)
4771 // Preserve Unbalanced status in such cases.
4772 if (!this_box->is_unbalanced()) {
4773 this_box->set_coarsened();
4774 }
4775 if (!box->is_unbalanced()) {
4776 box->set_coarsened();
4777 }
4778 }
4779 }
4780 _coarsened_locks.append(locks_list);
4781 }
4782 }
4783
4784 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4785 int count = coarsened_count();
4786 for (int i = 0; i < count; i++) {
4787 Node_List* locks_list = _coarsened_locks.at(i);
4788 for (uint j = 0; j < locks_list->size(); j++) {
4789 Node* lock = locks_list->at(j);
4790 assert(lock->is_AbstractLock(), "sanity");
4791 if (!useful.member(lock)) {
4792 locks_list->yank(lock);
4793 }
4794 }
4795 }
4796 }
4797
4798 void Compile::remove_coarsened_lock(Node* n) {
4799 if (n->is_AbstractLock()) {
4800 int count = coarsened_count();
4801 for (int i = 0; i < count; i++) {
4802 Node_List* locks_list = _coarsened_locks.at(i);
4803 locks_list->yank(n);
4804 }
4805 }
4806 }
4807
4808 bool Compile::coarsened_locks_consistent() {
4809 int count = coarsened_count();
4810 for (int i = 0; i < count; i++) {
4811 bool unbalanced = false;
4812 bool modified = false; // track locks kind modifications
4813 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4814 uint size = locks_list->size();
4815 if (size == 0) {
4816 unbalanced = false; // All locks were eliminated - good
4817 } else if (size != locks_list->origin_cnt()) {
4818 unbalanced = true; // Some locks were removed from list
4819 } else {
4820 for (uint j = 0; j < size; j++) {
4821 Node* lock = locks_list->at(j);
4822 // All nodes in group should have the same state (modified or not)
4823 if (!lock->as_AbstractLock()->is_coarsened()) {
4824 if (j == 0) {
4825 // first on list was modified, the rest should be too for consistency
4826 modified = true;
4827 } else if (!modified) {
4828 // this lock was modified but previous locks on the list were not
4829 unbalanced = true;
4830 break;
4831 }
4832 } else if (modified) {
4833 // previous locks on list were modified but not this lock
4834 unbalanced = true;
4835 break;
4836 }
4837 }
4838 }
4839 if (unbalanced) {
4840 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4841 #ifdef ASSERT
4842 if (PrintEliminateLocks) {
4843 tty->print_cr("=== unbalanced coarsened locks ===");
4844 for (uint l = 0; l < size; l++) {
4845 locks_list->at(l)->dump();
4846 }
4847 }
4848 #endif
4849 record_failure(C2Compiler::retry_no_locks_coarsening());
4850 return false;
4851 }
4852 }
4853 return true;
4854 }
4855
4856 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4857 // locks coarsening optimization removed Lock/Unlock nodes from them.
4858 // Such regions become unbalanced because coarsening only removes part
4859 // of Lock/Unlock nodes in region. As result we can't execute other
4860 // locks elimination optimizations which assume all code paths have
4861 // corresponding pair of Lock/Unlock nodes - they are balanced.
4862 void Compile::mark_unbalanced_boxes() const {
4863 int count = coarsened_count();
4864 for (int i = 0; i < count; i++) {
4865 Node_List* locks_list = _coarsened_locks.at(i);
4866 uint size = locks_list->size();
4867 if (size > 0) {
4868 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4869 BoxLockNode* box = alock->box_node()->as_BoxLock();
4870 if (alock->is_coarsened()) {
4871 // coarsened_locks_consistent(), which is called before this method, verifies
4872 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4873 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4874 for (uint j = 1; j < size; j++) {
4875 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4876 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4877 if (box != this_box) {
4878 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4879 box->set_unbalanced();
4880 this_box->set_unbalanced();
4881 }
4882 }
4883 }
4884 }
4885 }
4886 }
4887
4888 /**
4889 * Remove the speculative part of types and clean up the graph
4890 */
4891 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4892 if (UseTypeSpeculation) {
4893 Unique_Node_List worklist;
4894 worklist.push(root());
4895 int modified = 0;
4896 // Go over all type nodes that carry a speculative type, drop the
4897 // speculative part of the type and enqueue the node for an igvn
4898 // which may optimize it out.
4899 for (uint next = 0; next < worklist.size(); ++next) {
4900 Node *n = worklist.at(next);
4901 if (n->is_Type()) {
4902 TypeNode* tn = n->as_Type();
4903 const Type* t = tn->type();
4904 const Type* t_no_spec = t->remove_speculative();
4905 if (t_no_spec != t) {
4906 bool in_hash = igvn.hash_delete(n);
4907 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4908 tn->set_type(t_no_spec);
4909 igvn.hash_insert(n);
4910 igvn._worklist.push(n); // give it a chance to go away
4911 modified++;
4912 }
4913 }
4914 // Iterate over outs - endless loops is unreachable from below
4915 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4916 Node *m = n->fast_out(i);
4917 if (not_a_node(m)) {
4918 continue;
4919 }
4920 worklist.push(m);
4921 }
4922 }
4923 // Drop the speculative part of all types in the igvn's type table
4924 igvn.remove_speculative_types();
4925 if (modified > 0) {
4926 igvn.optimize();
4927 if (failing()) return;
4928 }
4929 #ifdef ASSERT
4930 // Verify that after the IGVN is over no speculative type has resurfaced
4931 worklist.clear();
4932 worklist.push(root());
4933 for (uint next = 0; next < worklist.size(); ++next) {
4934 Node *n = worklist.at(next);
4935 const Type* t = igvn.type_or_null(n);
4936 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4937 if (n->is_Type()) {
4938 t = n->as_Type()->type();
4939 assert(t == t->remove_speculative(), "no more speculative types");
4940 }
4941 // Iterate over outs - endless loops is unreachable from below
4942 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4943 Node *m = n->fast_out(i);
4944 if (not_a_node(m)) {
4945 continue;
4946 }
4947 worklist.push(m);
4948 }
4949 }
4950 igvn.check_no_speculative_types();
4951 #endif
4952 }
4953 }
4954
4955 // Auxiliary methods to support randomized stressing/fuzzing.
4956
4957 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
4958 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
4959 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
4960 FLAG_SET_ERGO(StressSeed, _stress_seed);
4961 } else {
4962 _stress_seed = StressSeed;
4963 }
4964 if (_log != nullptr) {
4965 _log->elem("stress_test seed='%u'", _stress_seed);
4966 }
4967 }
4968
4969 int Compile::random() {
4970 _stress_seed = os::next_random(_stress_seed);
4971 return static_cast<int>(_stress_seed);
4972 }
4973
4974 // This method can be called the arbitrary number of times, with current count
4975 // as the argument. The logic allows selecting a single candidate from the
4976 // running list of candidates as follows:
4977 // int count = 0;
4978 // Cand* selected = null;
4979 // while(cand = cand->next()) {
4980 // if (randomized_select(++count)) {
4981 // selected = cand;
4982 // }
4983 // }
4984 //
4985 // Including count equalizes the chances any candidate is "selected".
4986 // This is useful when we don't have the complete list of candidates to choose
4987 // from uniformly. In this case, we need to adjust the randomicity of the
4988 // selection, or else we will end up biasing the selection towards the latter
4989 // candidates.
4990 //
4991 // Quick back-envelope calculation shows that for the list of n candidates
4992 // the equal probability for the candidate to persist as "best" can be
4993 // achieved by replacing it with "next" k-th candidate with the probability
4994 // of 1/k. It can be easily shown that by the end of the run, the
4995 // probability for any candidate is converged to 1/n, thus giving the
4996 // uniform distribution among all the candidates.
4997 //
4998 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4999 #define RANDOMIZED_DOMAIN_POW 29
5000 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5001 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5002 bool Compile::randomized_select(int count) {
5003 assert(count > 0, "only positive");
5004 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5005 }
5006
5007 #ifdef ASSERT
5008 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5009 bool Compile::fail_randomly() {
5010 if ((random() % StressBailoutMean) != 0) {
5011 return false;
5012 }
5013 record_failure("StressBailout");
5014 return true;
5015 }
5016
5017 bool Compile::failure_is_artificial() {
5018 return C->failure_reason_is("StressBailout");
5019 }
5020 #endif
5021
5022 CloneMap& Compile::clone_map() { return _clone_map; }
5023 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5024
5025 void NodeCloneInfo::dump_on(outputStream* st) const {
5026 st->print(" {%d:%d} ", idx(), gen());
5027 }
5028
5029 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5030 uint64_t val = value(old->_idx);
5031 NodeCloneInfo cio(val);
5032 assert(val != 0, "old node should be in the map");
5033 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5034 insert(nnn->_idx, cin.get());
5035 #ifndef PRODUCT
5036 if (is_debug()) {
5037 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5038 }
5039 #endif
5040 }
5041
5042 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5043 NodeCloneInfo cio(value(old->_idx));
5044 if (cio.get() == 0) {
5045 cio.set(old->_idx, 0);
5046 insert(old->_idx, cio.get());
5047 #ifndef PRODUCT
5048 if (is_debug()) {
5049 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5050 }
5051 #endif
5052 }
5053 clone(old, nnn, gen);
5054 }
5055
5056 int CloneMap::max_gen() const {
5057 int g = 0;
5058 DictI di(_dict);
5059 for(; di.test(); ++di) {
5060 int t = gen(di._key);
5061 if (g < t) {
5062 g = t;
5063 #ifndef PRODUCT
5064 if (is_debug()) {
5065 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5066 }
5067 #endif
5068 }
5069 }
5070 return g;
5071 }
5072
5073 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5074 uint64_t val = value(key);
5075 if (val != 0) {
5076 NodeCloneInfo ni(val);
5077 ni.dump_on(st);
5078 }
5079 }
5080
5081 void Compile::shuffle_macro_nodes() {
5082 if (_macro_nodes.length() < 2) {
5083 return;
5084 }
5085 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5086 uint j = C->random() % (i + 1);
5087 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5088 }
5089 }
5090
5091 // Move Allocate nodes to the start of the list
5092 void Compile::sort_macro_nodes() {
5093 int count = macro_count();
5094 int allocates = 0;
5095 for (int i = 0; i < count; i++) {
5096 Node* n = macro_node(i);
5097 if (n->is_Allocate()) {
5098 if (i != allocates) {
5099 Node* tmp = macro_node(allocates);
5100 _macro_nodes.at_put(allocates, n);
5101 _macro_nodes.at_put(i, tmp);
5102 }
5103 allocates++;
5104 }
5105 }
5106 }
5107
5108 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5109 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5110 EventCompilerPhase event(UNTIMED);
5111 if (event.should_commit()) {
5112 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5113 }
5114 #ifndef PRODUCT
5115 ResourceMark rm;
5116 stringStream ss;
5117 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5118 int iter = ++_igv_phase_iter[cpt];
5119 if (iter > 1) {
5120 ss.print(" %d", iter);
5121 }
5122 if (n != nullptr) {
5123 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5124 if (n->is_Call()) {
5125 CallNode* call = n->as_Call();
5126 if (call->_name != nullptr) {
5127 // E.g. uncommon traps etc.
5128 ss.print(" - %s", call->_name);
5129 } else if (call->is_CallJava()) {
5130 CallJavaNode* call_java = call->as_CallJava();
5131 if (call_java->method() != nullptr) {
5132 ss.print(" -");
5133 call_java->method()->print_short_name(&ss);
5134 }
5135 }
5136 }
5137 }
5138
5139 const char* name = ss.as_string();
5140 if (should_print_igv(level)) {
5141 _igv_printer->print_graph(name);
5142 }
5143 if (should_print_phase(cpt)) {
5144 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5145 }
5146 #endif
5147 C->_latest_stage_start_counter.stamp();
5148 }
5149
5150 // Only used from CompileWrapper
5151 void Compile::begin_method() {
5152 #ifndef PRODUCT
5153 if (_method != nullptr && should_print_igv(1)) {
5154 _igv_printer->begin_method();
5155 }
5156 #endif
5157 C->_latest_stage_start_counter.stamp();
5158 }
5159
5160 // Only used from CompileWrapper
5161 void Compile::end_method() {
5162 EventCompilerPhase event(UNTIMED);
5163 if (event.should_commit()) {
5164 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5165 }
5166
5167 #ifndef PRODUCT
5168 if (_method != nullptr && should_print_igv(1)) {
5169 _igv_printer->end_method();
5170 }
5171 #endif
5172 }
5173
5174 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5175 #ifndef PRODUCT
5176 if (_directive->should_print_phase(cpt)) {
5177 return true;
5178 }
5179 #endif
5180 return false;
5181 }
5182
5183 #ifndef PRODUCT
5184 void Compile::init_igv() {
5185 if (_igv_printer == nullptr) {
5186 _igv_printer = IdealGraphPrinter::printer();
5187 _igv_printer->set_compile(this);
5188 }
5189 }
5190 #endif
5191
5192 bool Compile::should_print_igv(const int level) {
5193 #ifndef PRODUCT
5194 if (PrintIdealGraphLevel < 0) { // disabled by the user
5195 return false;
5196 }
5197
5198 bool need = directive()->IGVPrintLevelOption >= level;
5199 if (need) {
5200 Compile::init_igv();
5201 }
5202 return need;
5203 #else
5204 return false;
5205 #endif
5206 }
5207
5208 #ifndef PRODUCT
5209 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5210 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5211
5212 // Called from debugger. Prints method to the default file with the default phase name.
5213 // This works regardless of any Ideal Graph Visualizer flags set or not.
5214 void igv_print() {
5215 Compile::current()->igv_print_method_to_file();
5216 }
5217
5218 // Same as igv_print() above but with a specified phase name.
5219 void igv_print(const char* phase_name) {
5220 Compile::current()->igv_print_method_to_file(phase_name);
5221 }
5222
5223 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5224 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5225 // This works regardless of any Ideal Graph Visualizer flags set or not.
5226 void igv_print(bool network) {
5227 if (network) {
5228 Compile::current()->igv_print_method_to_network();
5229 } else {
5230 Compile::current()->igv_print_method_to_file();
5231 }
5232 }
5233
5234 // Same as igv_print(bool network) above but with a specified phase name.
5235 void igv_print(bool network, const char* phase_name) {
5236 if (network) {
5237 Compile::current()->igv_print_method_to_network(phase_name);
5238 } else {
5239 Compile::current()->igv_print_method_to_file(phase_name);
5240 }
5241 }
5242
5243 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5244 void igv_print_default() {
5245 Compile::current()->print_method(PHASE_DEBUG, 0);
5246 }
5247
5248 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5249 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5250 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5251 void igv_append() {
5252 Compile::current()->igv_print_method_to_file("Debug", true);
5253 }
5254
5255 // Same as igv_append() above but with a specified phase name.
5256 void igv_append(const char* phase_name) {
5257 Compile::current()->igv_print_method_to_file(phase_name, true);
5258 }
5259
5260 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5261 const char* file_name = "custom_debug.xml";
5262 if (_debug_file_printer == nullptr) {
5263 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5264 } else {
5265 _debug_file_printer->update_compiled_method(C->method());
5266 }
5267 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5268 _debug_file_printer->print_graph(phase_name);
5269 }
5270
5271 void Compile::igv_print_method_to_network(const char* phase_name) {
5272 ResourceMark rm;
5273 GrowableArray<const Node*> empty_list;
5274 igv_print_graph_to_network(phase_name, (Node*) C->root(), empty_list);
5275 }
5276
5277 void Compile::igv_print_graph_to_network(const char* name, Node* node, GrowableArray<const Node*>& visible_nodes) {
5278 if (_debug_network_printer == nullptr) {
5279 _debug_network_printer = new IdealGraphPrinter(C);
5280 } else {
5281 _debug_network_printer->update_compiled_method(C->method());
5282 }
5283 tty->print_cr("Method printed over network stream to IGV");
5284 _debug_network_printer->print(name, C->root(), visible_nodes);
5285 }
5286 #endif
5287
5288 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5289 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5290 return value;
5291 }
5292 Node* result = nullptr;
5293 if (bt == T_BYTE) {
5294 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5295 result = new RShiftINode(result, phase->intcon(24));
5296 } else if (bt == T_BOOLEAN) {
5297 result = new AndINode(value, phase->intcon(0xFF));
5298 } else if (bt == T_CHAR) {
5299 result = new AndINode(value,phase->intcon(0xFFFF));
5300 } else {
5301 assert(bt == T_SHORT, "unexpected narrow type");
5302 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5303 result = new RShiftINode(result, phase->intcon(16));
5304 }
5305 if (transform_res) {
5306 result = phase->transform(result);
5307 }
5308 return result;
5309 }
5310
5311 void Compile::record_method_not_compilable_oom() {
5312 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5313 }