1 /*
2 * Copyright (c) 1997, 2024, 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 "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compiler_globals.hpp"
37 #include "compiler/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/resourceArea.hpp"
45 #include "opto/addnode.hpp"
46 #include "opto/block.hpp"
47 #include "opto/c2compiler.hpp"
48 #include "opto/callGenerator.hpp"
49 #include "opto/callnode.hpp"
50 #include "opto/castnode.hpp"
51 #include "opto/cfgnode.hpp"
52 #include "opto/chaitin.hpp"
53 #include "opto/compile.hpp"
54 #include "opto/connode.hpp"
55 #include "opto/convertnode.hpp"
56 #include "opto/divnode.hpp"
57 #include "opto/escape.hpp"
58 #include "opto/idealGraphPrinter.hpp"
59 #include "opto/locknode.hpp"
60 #include "opto/loopnode.hpp"
61 #include "opto/machnode.hpp"
62 #include "opto/macro.hpp"
63 #include "opto/matcher.hpp"
64 #include "opto/mathexactnode.hpp"
65 #include "opto/memnode.hpp"
66 #include "opto/mulnode.hpp"
67 #include "opto/narrowptrnode.hpp"
68 #include "opto/node.hpp"
69 #include "opto/opcodes.hpp"
70 #include "opto/output.hpp"
71 #include "opto/parse.hpp"
72 #include "opto/phaseX.hpp"
73 #include "opto/rootnode.hpp"
74 #include "opto/runtime.hpp"
75 #include "opto/stringopts.hpp"
76 #include "opto/type.hpp"
77 #include "opto/vector.hpp"
78 #include "opto/vectornode.hpp"
79 #include "runtime/globals_extension.hpp"
80 #include "runtime/sharedRuntime.hpp"
81 #include "runtime/signature.hpp"
82 #include "runtime/stubRoutines.hpp"
83 #include "runtime/timer.hpp"
84 #include "utilities/align.hpp"
85 #include "utilities/copy.hpp"
86 #include "utilities/macros.hpp"
87 #include "utilities/resourceHash.hpp"
88
89 // -------------------- Compile::mach_constant_base_node -----------------------
90 // Constant table base node singleton.
91 MachConstantBaseNode* Compile::mach_constant_base_node() {
92 if (_mach_constant_base_node == nullptr) {
93 _mach_constant_base_node = new MachConstantBaseNode();
94 _mach_constant_base_node->add_req(C->root());
95 }
96 return _mach_constant_base_node;
97 }
98
99
100 /// Support for intrinsics.
101
102 // Return the index at which m must be inserted (or already exists).
103 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
104 class IntrinsicDescPair {
105 private:
106 ciMethod* _m;
107 bool _is_virtual;
108 public:
109 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
110 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
111 ciMethod* m= elt->method();
112 ciMethod* key_m = key->_m;
113 if (key_m < m) return -1;
114 else if (key_m > m) return 1;
115 else {
116 bool is_virtual = elt->is_virtual();
117 bool key_virtual = key->_is_virtual;
118 if (key_virtual < is_virtual) return -1;
119 else if (key_virtual > is_virtual) return 1;
120 else return 0;
121 }
122 }
123 };
124 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
125 #ifdef ASSERT
126 for (int i = 1; i < _intrinsics.length(); i++) {
127 CallGenerator* cg1 = _intrinsics.at(i-1);
128 CallGenerator* cg2 = _intrinsics.at(i);
129 assert(cg1->method() != cg2->method()
130 ? cg1->method() < cg2->method()
131 : cg1->is_virtual() < cg2->is_virtual(),
132 "compiler intrinsics list must stay sorted");
133 }
134 #endif
135 IntrinsicDescPair pair(m, is_virtual);
136 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
137 }
138
139 void Compile::register_intrinsic(CallGenerator* cg) {
140 bool found = false;
141 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
142 assert(!found, "registering twice");
143 _intrinsics.insert_before(index, cg);
144 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
145 }
146
147 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
148 assert(m->is_loaded(), "don't try this on unloaded methods");
149 if (_intrinsics.length() > 0) {
150 bool found = false;
151 int index = intrinsic_insertion_index(m, is_virtual, found);
152 if (found) {
153 return _intrinsics.at(index);
154 }
155 }
156 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
157 if (m->intrinsic_id() != vmIntrinsics::_none &&
158 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
159 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
160 if (cg != nullptr) {
161 // Save it for next time:
162 register_intrinsic(cg);
163 return cg;
164 } else {
165 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
166 }
167 }
168 return nullptr;
169 }
170
171 // Compile::make_vm_intrinsic is defined in library_call.cpp.
172
173 #ifndef PRODUCT
174 // statistics gathering...
175
176 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
177 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
178
179 inline int as_int(vmIntrinsics::ID id) {
180 return vmIntrinsics::as_int(id);
181 }
182
183 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
184 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
185 int oflags = _intrinsic_hist_flags[as_int(id)];
186 assert(flags != 0, "what happened?");
187 if (is_virtual) {
188 flags |= _intrinsic_virtual;
189 }
190 bool changed = (flags != oflags);
191 if ((flags & _intrinsic_worked) != 0) {
192 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
193 if (count == 1) {
194 changed = true; // first time
195 }
196 // increment the overall count also:
197 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
198 }
199 if (changed) {
200 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
201 // Something changed about the intrinsic's virtuality.
202 if ((flags & _intrinsic_virtual) != 0) {
203 // This is the first use of this intrinsic as a virtual call.
204 if (oflags != 0) {
205 // We already saw it as a non-virtual, so note both cases.
206 flags |= _intrinsic_both;
207 }
208 } else if ((oflags & _intrinsic_both) == 0) {
209 // This is the first use of this intrinsic as a non-virtual
210 flags |= _intrinsic_both;
211 }
212 }
213 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
214 }
215 // update the overall flags also:
216 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
217 return changed;
218 }
219
220 static char* format_flags(int flags, char* buf) {
221 buf[0] = 0;
222 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
223 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
224 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
225 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
226 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
227 if (buf[0] == 0) strcat(buf, ",");
228 assert(buf[0] == ',', "must be");
229 return &buf[1];
230 }
231
232 void Compile::print_intrinsic_statistics() {
233 char flagsbuf[100];
234 ttyLocker ttyl;
235 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
236 tty->print_cr("Compiler intrinsic usage:");
237 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
238 if (total == 0) total = 1; // avoid div0 in case of no successes
239 #define PRINT_STAT_LINE(name, c, f) \
240 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
241 for (auto id : EnumRange<vmIntrinsicID>{}) {
242 int flags = _intrinsic_hist_flags[as_int(id)];
243 juint count = _intrinsic_hist_count[as_int(id)];
244 if ((flags | count) != 0) {
245 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
246 }
247 }
248 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
249 if (xtty != nullptr) xtty->tail("statistics");
250 }
251
252 void Compile::print_statistics() {
253 { ttyLocker ttyl;
254 if (xtty != nullptr) xtty->head("statistics type='opto'");
255 Parse::print_statistics();
256 PhaseStringOpts::print_statistics();
257 PhaseCCP::print_statistics();
258 PhaseRegAlloc::print_statistics();
259 PhaseOutput::print_statistics();
260 PhasePeephole::print_statistics();
261 PhaseIdealLoop::print_statistics();
262 ConnectionGraph::print_statistics();
263 PhaseMacroExpand::print_statistics();
264 if (xtty != nullptr) xtty->tail("statistics");
265 }
266 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
267 // put this under its own <statistics> element.
268 print_intrinsic_statistics();
269 }
270 }
271 #endif //PRODUCT
272
273 void Compile::gvn_replace_by(Node* n, Node* nn) {
274 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
275 Node* use = n->last_out(i);
276 bool is_in_table = initial_gvn()->hash_delete(use);
277 uint uses_found = 0;
278 for (uint j = 0; j < use->len(); j++) {
279 if (use->in(j) == n) {
280 if (j < use->req())
281 use->set_req(j, nn);
282 else
283 use->set_prec(j, nn);
284 uses_found++;
285 }
286 }
287 if (is_in_table) {
288 // reinsert into table
289 initial_gvn()->hash_find_insert(use);
290 }
291 record_for_igvn(use);
292 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
293 i -= uses_found; // we deleted 1 or more copies of this edge
294 }
295 }
296
297
298 // Identify all nodes that are reachable from below, useful.
299 // Use breadth-first pass that records state in a Unique_Node_List,
300 // recursive traversal is slower.
301 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
302 int estimated_worklist_size = live_nodes();
303 useful.map( estimated_worklist_size, nullptr ); // preallocate space
304
305 // Initialize worklist
306 if (root() != nullptr) { useful.push(root()); }
307 // If 'top' is cached, declare it useful to preserve cached node
308 if (cached_top_node()) { useful.push(cached_top_node()); }
309
310 // Push all useful nodes onto the list, breadthfirst
311 for( uint next = 0; next < useful.size(); ++next ) {
312 assert( next < unique(), "Unique useful nodes < total nodes");
313 Node *n = useful.at(next);
314 uint max = n->len();
315 for( uint i = 0; i < max; ++i ) {
316 Node *m = n->in(i);
317 if (not_a_node(m)) continue;
318 useful.push(m);
319 }
320 }
321 }
322
323 // Update dead_node_list with any missing dead nodes using useful
324 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
325 void Compile::update_dead_node_list(Unique_Node_List &useful) {
326 uint max_idx = unique();
327 VectorSet& useful_node_set = useful.member_set();
328
329 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
330 // If node with index node_idx is not in useful set,
331 // mark it as dead in dead node list.
332 if (!useful_node_set.test(node_idx)) {
333 record_dead_node(node_idx);
334 }
335 }
336 }
337
338 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
339 int shift = 0;
340 for (int i = 0; i < inlines->length(); i++) {
341 CallGenerator* cg = inlines->at(i);
342 if (useful.member(cg->call_node())) {
343 if (shift > 0) {
344 inlines->at_put(i - shift, cg);
345 }
346 } else {
347 shift++; // skip over the dead element
348 }
349 }
350 if (shift > 0) {
351 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
352 }
353 }
354
355 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
356 assert(dead != nullptr && dead->is_Call(), "sanity");
357 int found = 0;
358 for (int i = 0; i < inlines->length(); i++) {
359 if (inlines->at(i)->call_node() == dead) {
360 inlines->remove_at(i);
361 found++;
362 NOT_DEBUG( break; ) // elements are unique, so exit early
363 }
364 }
365 assert(found <= 1, "not unique");
366 }
367
368 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
369 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
370 for (int i = node_list.length() - 1; i >= 0; i--) {
371 N* node = node_list.at(i);
372 if (!useful.member(node)) {
373 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
374 }
375 }
376 }
377
378 void Compile::remove_useless_node(Node* dead) {
379 remove_modified_node(dead);
380
381 // Constant node that has no out-edges and has only one in-edge from
382 // root is usually dead. However, sometimes reshaping walk makes
383 // it reachable by adding use edges. So, we will NOT count Con nodes
384 // as dead to be conservative about the dead node count at any
385 // given time.
386 if (!dead->is_Con()) {
387 record_dead_node(dead->_idx);
388 }
389 if (dead->is_macro()) {
390 remove_macro_node(dead);
391 }
392 if (dead->is_expensive()) {
393 remove_expensive_node(dead);
394 }
395 if (dead->Opcode() == Op_Opaque4) {
396 remove_template_assertion_predicate_opaq(dead);
397 }
398 if (dead->is_ParsePredicate()) {
399 remove_parse_predicate(dead->as_ParsePredicate());
400 }
401 if (dead->for_post_loop_opts_igvn()) {
402 remove_from_post_loop_opts_igvn(dead);
403 }
404 if (dead->is_Call()) {
405 remove_useless_late_inlines( &_late_inlines, dead);
406 remove_useless_late_inlines( &_string_late_inlines, dead);
407 remove_useless_late_inlines( &_boxing_late_inlines, dead);
408 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
409
410 if (dead->is_CallStaticJava()) {
411 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
412 }
413 }
414 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
415 bs->unregister_potential_barrier_node(dead);
416 }
417
418 // Disconnect all useless nodes by disconnecting those at the boundary.
419 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
420 uint next = 0;
421 while (next < useful.size()) {
422 Node *n = useful.at(next++);
423 if (n->is_SafePoint()) {
424 // We're done with a parsing phase. Replaced nodes are not valid
425 // beyond that point.
426 n->as_SafePoint()->delete_replaced_nodes();
427 }
428 // Use raw traversal of out edges since this code removes out edges
429 int max = n->outcnt();
430 for (int j = 0; j < max; ++j) {
431 Node* child = n->raw_out(j);
432 if (!useful.member(child)) {
433 assert(!child->is_top() || child != top(),
434 "If top is cached in Compile object it is in useful list");
435 // Only need to remove this out-edge to the useless node
436 n->raw_del_out(j);
437 --j;
438 --max;
439 }
440 }
441 if (n->outcnt() == 1 && n->has_special_unique_user()) {
442 assert(useful.member(n->unique_out()), "do not push a useless node");
443 worklist.push(n->unique_out());
444 }
445 }
446
447 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
448 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
449 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
450 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
451 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
452 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
453 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
454 #ifdef ASSERT
455 if (_modified_nodes != nullptr) {
456 _modified_nodes->remove_useless_nodes(useful.member_set());
457 }
458 #endif
459
460 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
461 bs->eliminate_useless_gc_barriers(useful, this);
462 // clean up the late inline lists
463 remove_useless_late_inlines( &_late_inlines, useful);
464 remove_useless_late_inlines( &_string_late_inlines, useful);
465 remove_useless_late_inlines( &_boxing_late_inlines, useful);
466 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
467 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
468 }
469
470 // ============================================================================
471 //------------------------------CompileWrapper---------------------------------
472 class CompileWrapper : public StackObj {
473 Compile *const _compile;
474 public:
475 CompileWrapper(Compile* compile);
476
477 ~CompileWrapper();
478 };
479
480 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
481 // the Compile* pointer is stored in the current ciEnv:
482 ciEnv* env = compile->env();
483 assert(env == ciEnv::current(), "must already be a ciEnv active");
484 assert(env->compiler_data() == nullptr, "compile already active?");
485 env->set_compiler_data(compile);
486 assert(compile == Compile::current(), "sanity");
487
488 compile->set_type_dict(nullptr);
489 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
490 compile->clone_map().set_clone_idx(0);
491 compile->set_type_last_size(0);
492 compile->set_last_tf(nullptr, nullptr);
493 compile->set_indexSet_arena(nullptr);
494 compile->set_indexSet_free_block_list(nullptr);
495 compile->init_type_arena();
496 Type::Initialize(compile);
497 _compile->begin_method();
498 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
499 }
500 CompileWrapper::~CompileWrapper() {
501 // simulate crash during compilation
502 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
503
504 _compile->end_method();
505 _compile->env()->set_compiler_data(nullptr);
506 }
507
508
509 //----------------------------print_compile_messages---------------------------
510 void Compile::print_compile_messages() {
511 #ifndef PRODUCT
512 // Check if recompiling
513 if (!subsume_loads() && PrintOpto) {
514 // Recompiling without allowing machine instructions to subsume loads
515 tty->print_cr("*********************************************************");
516 tty->print_cr("** Bailout: Recompile without subsuming loads **");
517 tty->print_cr("*********************************************************");
518 }
519 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
520 // Recompiling without escape analysis
521 tty->print_cr("*********************************************************");
522 tty->print_cr("** Bailout: Recompile without escape analysis **");
523 tty->print_cr("*********************************************************");
524 }
525 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
526 // Recompiling without iterative escape analysis
527 tty->print_cr("*********************************************************");
528 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
529 tty->print_cr("*********************************************************");
530 }
531 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
532 // Recompiling without reducing allocation merges
533 tty->print_cr("*********************************************************");
534 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
535 tty->print_cr("*********************************************************");
536 }
537 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
538 // Recompiling without boxing elimination
539 tty->print_cr("*********************************************************");
540 tty->print_cr("** Bailout: Recompile without boxing elimination **");
541 tty->print_cr("*********************************************************");
542 }
543 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
544 // Recompiling without locks coarsening
545 tty->print_cr("*********************************************************");
546 tty->print_cr("** Bailout: Recompile without locks coarsening **");
547 tty->print_cr("*********************************************************");
548 }
549 if (env()->break_at_compile()) {
550 // Open the debugger when compiling this method.
551 tty->print("### Breaking when compiling: ");
552 method()->print_short_name();
553 tty->cr();
554 BREAKPOINT;
555 }
556
557 if( PrintOpto ) {
558 if (is_osr_compilation()) {
559 tty->print("[OSR]%3d", _compile_id);
560 } else {
561 tty->print("%3d", _compile_id);
562 }
563 }
564 #endif
565 }
566
567 #ifndef PRODUCT
568 void Compile::print_ideal_ir(const char* phase_name) {
569 // keep the following output all in one block
570 // This output goes directly to the tty, not the compiler log.
571 // To enable tools to match it up with the compilation activity,
572 // be sure to tag this tty output with the compile ID.
573
574 // Node dumping can cause a safepoint, which can break the tty lock.
575 // Buffer all node dumps, so that all safepoints happen before we lock.
576 ResourceMark rm;
577 stringStream ss;
578
579 if (_output == nullptr) {
580 ss.print_cr("AFTER: %s", phase_name);
581 // Print out all nodes in ascending order of index.
582 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
583 } else {
584 // Dump the node blockwise if we have a scheduling
585 _output->print_scheduling(&ss);
586 }
587
588 // Check that the lock is not broken by a safepoint.
589 NoSafepointVerifier nsv;
590 ttyLocker ttyl;
591 if (xtty != nullptr) {
592 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
593 compile_id(),
594 is_osr_compilation() ? " compile_kind='osr'" : "",
595 phase_name);
596 }
597
598 tty->print("%s", ss.as_string());
599
600 if (xtty != nullptr) {
601 xtty->tail("ideal");
602 }
603 }
604 #endif
605
606 // ============================================================================
607 //------------------------------Compile standard-------------------------------
608
609 // Compile a method. entry_bci is -1 for normal compilations and indicates
610 // the continuation bci for on stack replacement.
611
612
613 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
614 Options options, DirectiveSet* directive)
615 : Phase(Compiler),
616 _compile_id(ci_env->compile_id()),
617 _options(options),
618 _method(target),
619 _entry_bci(osr_bci),
620 _ilt(nullptr),
621 _stub_function(nullptr),
622 _stub_name(nullptr),
623 _stub_entry_point(nullptr),
624 _max_node_limit(MaxNodeLimit),
625 _post_loop_opts_phase(false),
626 _inlining_progress(false),
627 _inlining_incrementally(false),
628 _do_cleanup(false),
629 _has_reserved_stack_access(target->has_reserved_stack_access()),
630 #ifndef PRODUCT
631 _igv_idx(0),
632 _trace_opto_output(directive->TraceOptoOutputOption),
633 #endif
634 _has_method_handle_invokes(false),
635 _clinit_barrier_on_entry(false),
636 _stress_seed(0),
637 _comp_arena(mtCompiler),
638 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
639 _env(ci_env),
640 _directive(directive),
641 _log(ci_env->log()),
642 _first_failure_details(nullptr),
643 _intrinsics (comp_arena(), 0, 0, nullptr),
644 _macro_nodes (comp_arena(), 8, 0, nullptr),
645 _parse_predicates (comp_arena(), 8, 0, nullptr),
646 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
647 _expensive_nodes (comp_arena(), 8, 0, nullptr),
648 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
649 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
650 _coarsened_locks (comp_arena(), 8, 0, nullptr),
651 _congraph(nullptr),
652 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
653 _unique(0),
654 _dead_node_count(0),
655 _dead_node_list(comp_arena()),
656 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
657 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
658 _node_arena(&_node_arena_one),
659 _mach_constant_base_node(nullptr),
660 _Compile_types(mtCompiler),
661 _initial_gvn(nullptr),
662 _igvn_worklist(nullptr),
663 _types(nullptr),
664 _node_hash(nullptr),
665 _late_inlines(comp_arena(), 2, 0, nullptr),
666 _string_late_inlines(comp_arena(), 2, 0, nullptr),
667 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
668 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
669 _late_inlines_pos(0),
670 _number_of_mh_late_inlines(0),
671 _oom(false),
672 _print_inlining_stream(new (mtCompiler) stringStream()),
673 _print_inlining_list(nullptr),
674 _print_inlining_idx(0),
675 _print_inlining_output(nullptr),
676 _replay_inline_data(nullptr),
677 _java_calls(0),
678 _inner_loops(0),
679 _interpreter_frame_size(0),
680 _output(nullptr)
681 #ifndef PRODUCT
682 , _in_dump_cnt(0)
683 #endif
684 {
685 C = this;
686 CompileWrapper cw(this);
687
688 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
689 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
690
691 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
692 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
693 // We can always print a disassembly, either abstract (hex dump) or
694 // with the help of a suitable hsdis library. Thus, we should not
695 // couple print_assembly and print_opto_assembly controls.
696 // But: always print opto and regular assembly on compile command 'print'.
697 bool print_assembly = directive->PrintAssemblyOption;
698 set_print_assembly(print_opto_assembly || print_assembly);
699 #else
700 set_print_assembly(false); // must initialize.
701 #endif
702
703 #ifndef PRODUCT
704 set_parsed_irreducible_loop(false);
705 #endif
706
707 if (directive->ReplayInlineOption) {
708 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
709 }
710 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
711 set_print_intrinsics(directive->PrintIntrinsicsOption);
712 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
713
714 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
715 // Make sure the method being compiled gets its own MDO,
716 // so we can at least track the decompile_count().
717 // Need MDO to record RTM code generation state.
718 method()->ensure_method_data();
719 }
720
721 Init(/*do_aliasing=*/ true);
722
723 print_compile_messages();
724
725 _ilt = InlineTree::build_inline_tree_root();
726
727 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
728 assert(num_alias_types() >= AliasIdxRaw, "");
729
730 #define MINIMUM_NODE_HASH 1023
731
732 // GVN that will be run immediately on new nodes
733 uint estimated_size = method()->code_size()*4+64;
734 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
735 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
736 _types = new (comp_arena()) Type_Array(comp_arena());
737 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
738 PhaseGVN gvn;
739 set_initial_gvn(&gvn);
740
741 print_inlining_init();
742 { // Scope for timing the parser
743 TracePhase tp("parse", &timers[_t_parser]);
744
745 // Put top into the hash table ASAP.
746 initial_gvn()->transform(top());
747
748 // Set up tf(), start(), and find a CallGenerator.
749 CallGenerator* cg = nullptr;
750 if (is_osr_compilation()) {
751 const TypeTuple *domain = StartOSRNode::osr_domain();
752 const TypeTuple *range = TypeTuple::make_range(method()->signature());
753 init_tf(TypeFunc::make(domain, range));
754 StartNode* s = new StartOSRNode(root(), domain);
755 initial_gvn()->set_type_bottom(s);
756 init_start(s);
757 cg = CallGenerator::for_osr(method(), entry_bci());
758 } else {
759 // Normal case.
760 init_tf(TypeFunc::make(method()));
761 StartNode* s = new StartNode(root(), tf()->domain());
762 initial_gvn()->set_type_bottom(s);
763 init_start(s);
764 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
765 // With java.lang.ref.reference.get() we must go through the
766 // intrinsic - even when get() is the root
767 // method of the compile - so that, if necessary, the value in
768 // the referent field of the reference object gets recorded by
769 // the pre-barrier code.
770 cg = find_intrinsic(method(), false);
771 }
772 if (cg == nullptr) {
773 float past_uses = method()->interpreter_invocation_count();
774 float expected_uses = past_uses;
775 cg = CallGenerator::for_inline(method(), expected_uses);
776 }
777 }
778 if (failing()) return;
779 if (cg == nullptr) {
780 const char* reason = InlineTree::check_can_parse(method());
781 assert(reason != nullptr, "expect reason for parse failure");
782 stringStream ss;
783 ss.print("cannot parse method: %s", reason);
784 record_method_not_compilable(ss.as_string());
785 return;
786 }
787
788 gvn.set_type(root(), root()->bottom_type());
789
790 JVMState* jvms = build_start_state(start(), tf());
791 if ((jvms = cg->generate(jvms)) == nullptr) {
792 assert(failure_reason() != nullptr, "expect reason for parse failure");
793 stringStream ss;
794 ss.print("method parse failed: %s", failure_reason());
795 record_method_not_compilable(ss.as_string());
796 return;
797 }
798 GraphKit kit(jvms);
799
800 if (!kit.stopped()) {
801 // Accept return values, and transfer control we know not where.
802 // This is done by a special, unique ReturnNode bound to root.
803 return_values(kit.jvms());
804 }
805
806 if (kit.has_exceptions()) {
807 // Any exceptions that escape from this call must be rethrown
808 // to whatever caller is dynamically above us on the stack.
809 // This is done by a special, unique RethrowNode bound to root.
810 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
811 }
812
813 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
814
815 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
816 inline_string_calls(true);
817 }
818
819 if (failing()) return;
820
821 // Remove clutter produced by parsing.
822 if (!failing()) {
823 ResourceMark rm;
824 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
825 }
826 }
827
828 // Note: Large methods are capped off in do_one_bytecode().
829 if (failing()) return;
830
831 // After parsing, node notes are no longer automagic.
832 // They must be propagated by register_new_node_with_optimizer(),
833 // clone(), or the like.
834 set_default_node_notes(nullptr);
835
836 #ifndef PRODUCT
837 if (should_print_igv(1)) {
838 _igv_printer->print_inlining();
839 }
840 #endif
841
842 if (failing()) return;
843 NOT_PRODUCT( verify_graph_edges(); )
844
845 // If any phase is randomized for stress testing, seed random number
846 // generation and log the seed for repeatability.
847 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
848 StressIncrementalInlining || StressMacroExpansion) {
849 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
850 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
851 FLAG_SET_ERGO(StressSeed, _stress_seed);
852 } else {
853 _stress_seed = StressSeed;
854 }
855 if (_log != nullptr) {
856 _log->elem("stress_test seed='%u'", _stress_seed);
857 }
858 }
859
860 // Now optimize
861 Optimize();
862 if (failing()) return;
863 NOT_PRODUCT( verify_graph_edges(); )
864
865 #ifndef PRODUCT
866 if (should_print_ideal()) {
867 print_ideal_ir("print_ideal");
868 }
869 #endif
870
871 #ifdef ASSERT
872 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
873 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
874 #endif
875
876 // Dump compilation data to replay it.
877 if (directive->DumpReplayOption) {
878 env()->dump_replay_data(_compile_id);
879 }
880 if (directive->DumpInlineOption && (ilt() != nullptr)) {
881 env()->dump_inline_data(_compile_id);
882 }
883
884 // Now that we know the size of all the monitors we can add a fixed slot
885 // for the original deopt pc.
886 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
887 set_fixed_slots(next_slot);
888
889 // Compute when to use implicit null checks. Used by matching trap based
890 // nodes and NullCheck optimization.
891 set_allowed_deopt_reasons();
892
893 // Now generate code
894 Code_Gen();
895 }
896
897 //------------------------------Compile----------------------------------------
898 // Compile a runtime stub
899 Compile::Compile( ciEnv* ci_env,
900 TypeFunc_generator generator,
901 address stub_function,
902 const char *stub_name,
903 int is_fancy_jump,
904 bool pass_tls,
905 bool return_pc,
906 DirectiveSet* directive)
907 : Phase(Compiler),
908 _compile_id(0),
909 _options(Options::for_runtime_stub()),
910 _method(nullptr),
911 _entry_bci(InvocationEntryBci),
912 _stub_function(stub_function),
913 _stub_name(stub_name),
914 _stub_entry_point(nullptr),
915 _max_node_limit(MaxNodeLimit),
916 _post_loop_opts_phase(false),
917 _inlining_progress(false),
918 _inlining_incrementally(false),
919 _has_reserved_stack_access(false),
920 #ifndef PRODUCT
921 _igv_idx(0),
922 _trace_opto_output(directive->TraceOptoOutputOption),
923 #endif
924 _has_method_handle_invokes(false),
925 _clinit_barrier_on_entry(false),
926 _stress_seed(0),
927 _comp_arena(mtCompiler),
928 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
929 _env(ci_env),
930 _directive(directive),
931 _log(ci_env->log()),
932 _first_failure_details(nullptr),
933 _congraph(nullptr),
934 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
935 _unique(0),
936 _dead_node_count(0),
937 _dead_node_list(comp_arena()),
938 _node_arena_one(mtCompiler),
939 _node_arena_two(mtCompiler),
940 _node_arena(&_node_arena_one),
941 _mach_constant_base_node(nullptr),
942 _Compile_types(mtCompiler),
943 _initial_gvn(nullptr),
944 _igvn_worklist(nullptr),
945 _types(nullptr),
946 _node_hash(nullptr),
947 _number_of_mh_late_inlines(0),
948 _oom(false),
949 _print_inlining_stream(new (mtCompiler) stringStream()),
950 _print_inlining_list(nullptr),
951 _print_inlining_idx(0),
952 _print_inlining_output(nullptr),
953 _replay_inline_data(nullptr),
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 PhaseGVN gvn;
984 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
985 gvn.transform(top());
986
987 GraphKit kit;
988 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
989 }
990
991 NOT_PRODUCT( verify_graph_edges(); )
992
993 Code_Gen();
994 }
995
996 Compile::~Compile() {
997 delete _print_inlining_stream;
998 delete _first_failure_details;
999 };
1000
1001 //------------------------------Init-------------------------------------------
1002 // Prepare for a single compilation
1003 void Compile::Init(bool aliasing) {
1004 _do_aliasing = aliasing;
1005 _unique = 0;
1006 _regalloc = nullptr;
1007
1008 _tf = nullptr; // filled in later
1009 _top = nullptr; // cached later
1010 _matcher = nullptr; // filled in later
1011 _cfg = nullptr; // filled in later
1012
1013 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1014
1015 _node_note_array = nullptr;
1016 _default_node_notes = nullptr;
1017 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1018
1019 _immutable_memory = nullptr; // filled in at first inquiry
1020
1021 #ifdef ASSERT
1022 _phase_optimize_finished = false;
1023 _exception_backedge = false;
1024 _type_verify = nullptr;
1025 #endif
1026
1027 // Globally visible Nodes
1028 // First set TOP to null to give safe behavior during creation of RootNode
1029 set_cached_top_node(nullptr);
1030 set_root(new RootNode());
1031 // Now that you have a Root to point to, create the real TOP
1032 set_cached_top_node( new ConNode(Type::TOP) );
1033 set_recent_alloc(nullptr, nullptr);
1034
1035 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1036 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1037 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1038 env()->set_dependencies(new Dependencies(env()));
1039
1040 _fixed_slots = 0;
1041 set_has_split_ifs(false);
1042 set_has_loops(false); // first approximation
1043 set_has_stringbuilder(false);
1044 set_has_boxed_value(false);
1045 _trap_can_recompile = false; // no traps emitted yet
1046 _major_progress = true; // start out assuming good things will happen
1047 set_has_unsafe_access(false);
1048 set_max_vector_size(0);
1049 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1050 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1051 set_decompile_count(0);
1052
1053 #ifndef PRODUCT
1054 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1055 #endif
1056
1057 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1058 _loop_opts_cnt = LoopOptsCount;
1059 set_do_inlining(Inline);
1060 set_max_inline_size(MaxInlineSize);
1061 set_freq_inline_size(FreqInlineSize);
1062 set_do_scheduling(OptoScheduling);
1063
1064 set_do_vector_loop(false);
1065 set_has_monitors(false);
1066
1067 if (AllowVectorizeOnDemand) {
1068 if (has_method() && _directive->VectorizeOption) {
1069 set_do_vector_loop(true);
1070 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());})
1071 } else if (has_method() && method()->name() != 0 &&
1072 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1073 set_do_vector_loop(true);
1074 }
1075 }
1076 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1077 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());})
1078
1079 set_rtm_state(NoRTM); // No RTM lock eliding by default
1080 _max_node_limit = _directive->MaxNodeLimitOption;
1081
1082 #if INCLUDE_RTM_OPT
1083 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) {
1084 int rtm_state = method()->method_data()->rtm_state();
1085 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1086 // Don't generate RTM lock eliding code.
1087 set_rtm_state(NoRTM);
1088 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1089 // Generate RTM lock eliding code without abort ratio calculation code.
1090 set_rtm_state(UseRTM);
1091 } else if (UseRTMDeopt) {
1092 // Generate RTM lock eliding code and include abort ratio calculation
1093 // code if UseRTMDeopt is on.
1094 set_rtm_state(ProfileRTM);
1095 }
1096 }
1097 #endif
1098 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1099 set_clinit_barrier_on_entry(true);
1100 }
1101 if (debug_info()->recording_non_safepoints()) {
1102 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1103 (comp_arena(), 8, 0, nullptr));
1104 set_default_node_notes(Node_Notes::make(this));
1105 }
1106
1107 const int grow_ats = 16;
1108 _max_alias_types = grow_ats;
1109 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1110 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1111 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1112 {
1113 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1114 }
1115 // Initialize the first few types.
1116 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1117 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1118 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1119 _num_alias_types = AliasIdxRaw+1;
1120 // Zero out the alias type cache.
1121 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1122 // A null adr_type hits in the cache right away. Preload the right answer.
1123 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1124 }
1125
1126 //---------------------------init_start----------------------------------------
1127 // Install the StartNode on this compile object.
1128 void Compile::init_start(StartNode* s) {
1129 if (failing())
1130 return; // already failing
1131 assert(s == start(), "");
1132 }
1133
1134 /**
1135 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1136 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1137 * the ideal graph.
1138 */
1139 StartNode* Compile::start() const {
1140 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1141 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1142 Node* start = root()->fast_out(i);
1143 if (start->is_Start()) {
1144 return start->as_Start();
1145 }
1146 }
1147 fatal("Did not find Start node!");
1148 return nullptr;
1149 }
1150
1151 //-------------------------------immutable_memory-------------------------------------
1152 // Access immutable memory
1153 Node* Compile::immutable_memory() {
1154 if (_immutable_memory != nullptr) {
1155 return _immutable_memory;
1156 }
1157 StartNode* s = start();
1158 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1159 Node *p = s->fast_out(i);
1160 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1161 _immutable_memory = p;
1162 return _immutable_memory;
1163 }
1164 }
1165 ShouldNotReachHere();
1166 return nullptr;
1167 }
1168
1169 //----------------------set_cached_top_node------------------------------------
1170 // Install the cached top node, and make sure Node::is_top works correctly.
1171 void Compile::set_cached_top_node(Node* tn) {
1172 if (tn != nullptr) verify_top(tn);
1173 Node* old_top = _top;
1174 _top = tn;
1175 // Calling Node::setup_is_top allows the nodes the chance to adjust
1176 // their _out arrays.
1177 if (_top != nullptr) _top->setup_is_top();
1178 if (old_top != nullptr) old_top->setup_is_top();
1179 assert(_top == nullptr || top()->is_top(), "");
1180 }
1181
1182 #ifdef ASSERT
1183 uint Compile::count_live_nodes_by_graph_walk() {
1184 Unique_Node_List useful(comp_arena());
1185 // Get useful node list by walking the graph.
1186 identify_useful_nodes(useful);
1187 return useful.size();
1188 }
1189
1190 void Compile::print_missing_nodes() {
1191
1192 // Return if CompileLog is null and PrintIdealNodeCount is false.
1193 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1194 return;
1195 }
1196
1197 // This is an expensive function. It is executed only when the user
1198 // specifies VerifyIdealNodeCount option or otherwise knows the
1199 // additional work that needs to be done to identify reachable nodes
1200 // by walking the flow graph and find the missing ones using
1201 // _dead_node_list.
1202
1203 Unique_Node_List useful(comp_arena());
1204 // Get useful node list by walking the graph.
1205 identify_useful_nodes(useful);
1206
1207 uint l_nodes = C->live_nodes();
1208 uint l_nodes_by_walk = useful.size();
1209
1210 if (l_nodes != l_nodes_by_walk) {
1211 if (_log != nullptr) {
1212 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1213 _log->stamp();
1214 _log->end_head();
1215 }
1216 VectorSet& useful_member_set = useful.member_set();
1217 int last_idx = l_nodes_by_walk;
1218 for (int i = 0; i < last_idx; i++) {
1219 if (useful_member_set.test(i)) {
1220 if (_dead_node_list.test(i)) {
1221 if (_log != nullptr) {
1222 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1223 }
1224 if (PrintIdealNodeCount) {
1225 // Print the log message to tty
1226 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1227 useful.at(i)->dump();
1228 }
1229 }
1230 }
1231 else if (! _dead_node_list.test(i)) {
1232 if (_log != nullptr) {
1233 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1234 }
1235 if (PrintIdealNodeCount) {
1236 // Print the log message to tty
1237 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1238 }
1239 }
1240 }
1241 if (_log != nullptr) {
1242 _log->tail("mismatched_nodes");
1243 }
1244 }
1245 }
1246 void Compile::record_modified_node(Node* n) {
1247 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1248 _modified_nodes->push(n);
1249 }
1250 }
1251
1252 void Compile::remove_modified_node(Node* n) {
1253 if (_modified_nodes != nullptr) {
1254 _modified_nodes->remove(n);
1255 }
1256 }
1257 #endif
1258
1259 #ifndef PRODUCT
1260 void Compile::verify_top(Node* tn) const {
1261 if (tn != nullptr) {
1262 assert(tn->is_Con(), "top node must be a constant");
1263 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1264 assert(tn->in(0) != nullptr, "must have live top node");
1265 }
1266 }
1267 #endif
1268
1269
1270 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1271
1272 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1273 guarantee(arr != nullptr, "");
1274 int num_blocks = arr->length();
1275 if (grow_by < num_blocks) grow_by = num_blocks;
1276 int num_notes = grow_by * _node_notes_block_size;
1277 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1278 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1279 while (num_notes > 0) {
1280 arr->append(notes);
1281 notes += _node_notes_block_size;
1282 num_notes -= _node_notes_block_size;
1283 }
1284 assert(num_notes == 0, "exact multiple, please");
1285 }
1286
1287 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1288 if (source == nullptr || dest == nullptr) return false;
1289
1290 if (dest->is_Con())
1291 return false; // Do not push debug info onto constants.
1292
1293 #ifdef ASSERT
1294 // Leave a bread crumb trail pointing to the original node:
1295 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1296 dest->set_debug_orig(source);
1297 }
1298 #endif
1299
1300 if (node_note_array() == nullptr)
1301 return false; // Not collecting any notes now.
1302
1303 // This is a copy onto a pre-existing node, which may already have notes.
1304 // If both nodes have notes, do not overwrite any pre-existing notes.
1305 Node_Notes* source_notes = node_notes_at(source->_idx);
1306 if (source_notes == nullptr || source_notes->is_clear()) return false;
1307 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1308 if (dest_notes == nullptr || dest_notes->is_clear()) {
1309 return set_node_notes_at(dest->_idx, source_notes);
1310 }
1311
1312 Node_Notes merged_notes = (*source_notes);
1313 // The order of operations here ensures that dest notes will win...
1314 merged_notes.update_from(dest_notes);
1315 return set_node_notes_at(dest->_idx, &merged_notes);
1316 }
1317
1318
1319 //--------------------------allow_range_check_smearing-------------------------
1320 // Gating condition for coalescing similar range checks.
1321 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1322 // single covering check that is at least as strong as any of them.
1323 // If the optimization succeeds, the simplified (strengthened) range check
1324 // will always succeed. If it fails, we will deopt, and then give up
1325 // on the optimization.
1326 bool Compile::allow_range_check_smearing() const {
1327 // If this method has already thrown a range-check,
1328 // assume it was because we already tried range smearing
1329 // and it failed.
1330 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1331 return !already_trapped;
1332 }
1333
1334
1335 //------------------------------flatten_alias_type-----------------------------
1336 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1337 assert(do_aliasing(), "Aliasing should be enabled");
1338 int offset = tj->offset();
1339 TypePtr::PTR ptr = tj->ptr();
1340
1341 // Known instance (scalarizable allocation) alias only with itself.
1342 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1343 tj->is_oopptr()->is_known_instance();
1344
1345 // Process weird unsafe references.
1346 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1347 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1348 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1349 tj = TypeOopPtr::BOTTOM;
1350 ptr = tj->ptr();
1351 offset = tj->offset();
1352 }
1353
1354 // Array pointers need some flattening
1355 const TypeAryPtr* ta = tj->isa_aryptr();
1356 if (ta && ta->is_stable()) {
1357 // Erase stability property for alias analysis.
1358 tj = ta = ta->cast_to_stable(false);
1359 }
1360 if( ta && is_known_inst ) {
1361 if ( offset != Type::OffsetBot &&
1362 offset > arrayOopDesc::length_offset_in_bytes() ) {
1363 offset = Type::OffsetBot; // Flatten constant access into array body only
1364 tj = ta = ta->
1365 remove_speculative()->
1366 cast_to_ptr_type(ptr)->
1367 with_offset(offset);
1368 }
1369 } else if (ta) {
1370 // For arrays indexed by constant indices, we flatten the alias
1371 // space to include all of the array body. Only the header, klass
1372 // and array length can be accessed un-aliased.
1373 if( offset != Type::OffsetBot ) {
1374 if( ta->const_oop() ) { // MethodData* or Method*
1375 offset = Type::OffsetBot; // Flatten constant access into array body
1376 tj = ta = ta->
1377 remove_speculative()->
1378 cast_to_ptr_type(ptr)->
1379 cast_to_exactness(false)->
1380 with_offset(offset);
1381 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1382 // range is OK as-is.
1383 tj = ta = TypeAryPtr::RANGE;
1384 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1385 tj = TypeInstPtr::KLASS; // all klass loads look alike
1386 ta = TypeAryPtr::RANGE; // generic ignored junk
1387 ptr = TypePtr::BotPTR;
1388 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1389 tj = TypeInstPtr::MARK;
1390 ta = TypeAryPtr::RANGE; // generic ignored junk
1391 ptr = TypePtr::BotPTR;
1392 } else { // Random constant offset into array body
1393 offset = Type::OffsetBot; // Flatten constant access into array body
1394 tj = ta = ta->
1395 remove_speculative()->
1396 cast_to_ptr_type(ptr)->
1397 cast_to_exactness(false)->
1398 with_offset(offset);
1399 }
1400 }
1401 // Arrays of fixed size alias with arrays of unknown size.
1402 if (ta->size() != TypeInt::POS) {
1403 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1404 tj = ta = ta->
1405 remove_speculative()->
1406 cast_to_ptr_type(ptr)->
1407 with_ary(tary)->
1408 cast_to_exactness(false);
1409 }
1410 // Arrays of known objects become arrays of unknown objects.
1411 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1412 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1413 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1414 }
1415 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1416 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1417 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1418 }
1419 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1420 // cannot be distinguished by bytecode alone.
1421 if (ta->elem() == TypeInt::BOOL) {
1422 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1423 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1424 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1425 }
1426 // During the 2nd round of IterGVN, NotNull castings are removed.
1427 // Make sure the Bottom and NotNull variants alias the same.
1428 // Also, make sure exact and non-exact variants alias the same.
1429 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1430 tj = ta = ta->
1431 remove_speculative()->
1432 cast_to_ptr_type(TypePtr::BotPTR)->
1433 cast_to_exactness(false)->
1434 with_offset(offset);
1435 }
1436 }
1437
1438 // Oop pointers need some flattening
1439 const TypeInstPtr *to = tj->isa_instptr();
1440 if (to && to != TypeOopPtr::BOTTOM) {
1441 ciInstanceKlass* ik = to->instance_klass();
1442 if( ptr == TypePtr::Constant ) {
1443 if (ik != ciEnv::current()->Class_klass() ||
1444 offset < ik->layout_helper_size_in_bytes()) {
1445 // No constant oop pointers (such as Strings); they alias with
1446 // unknown strings.
1447 assert(!is_known_inst, "not scalarizable allocation");
1448 tj = to = to->
1449 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1450 remove_speculative()->
1451 cast_to_ptr_type(TypePtr::BotPTR)->
1452 cast_to_exactness(false);
1453 }
1454 } else if( is_known_inst ) {
1455 tj = to; // Keep NotNull and klass_is_exact for instance type
1456 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1457 // During the 2nd round of IterGVN, NotNull castings are removed.
1458 // Make sure the Bottom and NotNull variants alias the same.
1459 // Also, make sure exact and non-exact variants alias the same.
1460 tj = to = to->
1461 remove_speculative()->
1462 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1463 cast_to_ptr_type(TypePtr::BotPTR)->
1464 cast_to_exactness(false);
1465 }
1466 if (to->speculative() != nullptr) {
1467 tj = to = to->remove_speculative();
1468 }
1469 // Canonicalize the holder of this field
1470 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1471 // First handle header references such as a LoadKlassNode, even if the
1472 // object's klass is unloaded at compile time (4965979).
1473 if (!is_known_inst) { // Do it only for non-instance types
1474 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1475 }
1476 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1477 // Static fields are in the space above the normal instance
1478 // fields in the java.lang.Class instance.
1479 if (ik != ciEnv::current()->Class_klass()) {
1480 to = nullptr;
1481 tj = TypeOopPtr::BOTTOM;
1482 offset = tj->offset();
1483 }
1484 } else {
1485 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1486 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1487 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1488 if( is_known_inst ) {
1489 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id());
1490 } else {
1491 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset);
1492 }
1493 }
1494 }
1495 }
1496
1497 // Klass pointers to object array klasses need some flattening
1498 const TypeKlassPtr *tk = tj->isa_klassptr();
1499 if( tk ) {
1500 // If we are referencing a field within a Klass, we need
1501 // to assume the worst case of an Object. Both exact and
1502 // inexact types must flatten to the same alias class so
1503 // use NotNull as the PTR.
1504 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1505 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1506 env()->Object_klass(),
1507 offset);
1508 }
1509
1510 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1511 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1512 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1513 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1514 } else {
1515 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1516 }
1517 }
1518
1519 // Check for precise loads from the primary supertype array and force them
1520 // to the supertype cache alias index. Check for generic array loads from
1521 // the primary supertype array and also force them to the supertype cache
1522 // alias index. Since the same load can reach both, we need to merge
1523 // these 2 disparate memories into the same alias class. Since the
1524 // primary supertype array is read-only, there's no chance of confusion
1525 // where we bypass an array load and an array store.
1526 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1527 if (offset == Type::OffsetBot ||
1528 (offset >= primary_supers_offset &&
1529 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1530 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1531 offset = in_bytes(Klass::secondary_super_cache_offset());
1532 tj = tk = tk->with_offset(offset);
1533 }
1534 }
1535
1536 // Flatten all Raw pointers together.
1537 if (tj->base() == Type::RawPtr)
1538 tj = TypeRawPtr::BOTTOM;
1539
1540 if (tj->base() == Type::AnyPtr)
1541 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1542
1543 offset = tj->offset();
1544 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1545
1546 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1547 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1548 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1549 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1550 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1551 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1552 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1553 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1554 assert( tj->ptr() != TypePtr::TopPTR &&
1555 tj->ptr() != TypePtr::AnyNull &&
1556 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1557 // assert( tj->ptr() != TypePtr::Constant ||
1558 // tj->base() == Type::RawPtr ||
1559 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1560
1561 return tj;
1562 }
1563
1564 void Compile::AliasType::Init(int i, const TypePtr* at) {
1565 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1566 _index = i;
1567 _adr_type = at;
1568 _field = nullptr;
1569 _element = nullptr;
1570 _is_rewritable = true; // default
1571 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1572 if (atoop != nullptr && atoop->is_known_instance()) {
1573 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1574 _general_index = Compile::current()->get_alias_index(gt);
1575 } else {
1576 _general_index = 0;
1577 }
1578 }
1579
1580 BasicType Compile::AliasType::basic_type() const {
1581 if (element() != nullptr) {
1582 const Type* element = adr_type()->is_aryptr()->elem();
1583 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1584 } if (field() != nullptr) {
1585 return field()->layout_type();
1586 } else {
1587 return T_ILLEGAL; // unknown
1588 }
1589 }
1590
1591 //---------------------------------print_on------------------------------------
1592 #ifndef PRODUCT
1593 void Compile::AliasType::print_on(outputStream* st) {
1594 if (index() < 10)
1595 st->print("@ <%d> ", index());
1596 else st->print("@ <%d>", index());
1597 st->print(is_rewritable() ? " " : " RO");
1598 int offset = adr_type()->offset();
1599 if (offset == Type::OffsetBot)
1600 st->print(" +any");
1601 else st->print(" +%-3d", offset);
1602 st->print(" in ");
1603 adr_type()->dump_on(st);
1604 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1605 if (field() != nullptr && tjp) {
1606 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1607 tjp->offset() != field()->offset_in_bytes()) {
1608 st->print(" != ");
1609 field()->print();
1610 st->print(" ***");
1611 }
1612 }
1613 }
1614
1615 void print_alias_types() {
1616 Compile* C = Compile::current();
1617 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1618 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1619 C->alias_type(idx)->print_on(tty);
1620 tty->cr();
1621 }
1622 }
1623 #endif
1624
1625
1626 //----------------------------probe_alias_cache--------------------------------
1627 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1628 intptr_t key = (intptr_t) adr_type;
1629 key ^= key >> logAliasCacheSize;
1630 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1631 }
1632
1633
1634 //-----------------------------grow_alias_types--------------------------------
1635 void Compile::grow_alias_types() {
1636 const int old_ats = _max_alias_types; // how many before?
1637 const int new_ats = old_ats; // how many more?
1638 const int grow_ats = old_ats+new_ats; // how many now?
1639 _max_alias_types = grow_ats;
1640 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1641 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1642 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1643 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1644 }
1645
1646
1647 //--------------------------------find_alias_type------------------------------
1648 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1649 if (!do_aliasing()) {
1650 return alias_type(AliasIdxBot);
1651 }
1652
1653 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1654 if (ace->_adr_type == adr_type) {
1655 return alias_type(ace->_index);
1656 }
1657
1658 // Handle special cases.
1659 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1660 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1661
1662 // Do it the slow way.
1663 const TypePtr* flat = flatten_alias_type(adr_type);
1664
1665 #ifdef ASSERT
1666 {
1667 ResourceMark rm;
1668 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1669 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1670 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1671 Type::str(adr_type));
1672 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1673 const TypeOopPtr* foop = flat->is_oopptr();
1674 // Scalarizable allocations have exact klass always.
1675 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1676 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1677 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1678 Type::str(foop), Type::str(xoop));
1679 }
1680 }
1681 #endif
1682
1683 int idx = AliasIdxTop;
1684 for (int i = 0; i < num_alias_types(); i++) {
1685 if (alias_type(i)->adr_type() == flat) {
1686 idx = i;
1687 break;
1688 }
1689 }
1690
1691 if (idx == AliasIdxTop) {
1692 if (no_create) return nullptr;
1693 // Grow the array if necessary.
1694 if (_num_alias_types == _max_alias_types) grow_alias_types();
1695 // Add a new alias type.
1696 idx = _num_alias_types++;
1697 _alias_types[idx]->Init(idx, flat);
1698 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1699 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1700 if (flat->isa_instptr()) {
1701 if (flat->offset() == java_lang_Class::klass_offset()
1702 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1703 alias_type(idx)->set_rewritable(false);
1704 }
1705 if (flat->isa_aryptr()) {
1706 #ifdef ASSERT
1707 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1708 // (T_BYTE has the weakest alignment and size restrictions...)
1709 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1710 #endif
1711 if (flat->offset() == TypePtr::OffsetBot) {
1712 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1713 }
1714 }
1715 if (flat->isa_klassptr()) {
1716 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1717 alias_type(idx)->set_rewritable(false);
1718 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1719 alias_type(idx)->set_rewritable(false);
1720 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1721 alias_type(idx)->set_rewritable(false);
1722 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1723 alias_type(idx)->set_rewritable(false);
1724 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1725 alias_type(idx)->set_rewritable(false);
1726 }
1727 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1728 // but the base pointer type is not distinctive enough to identify
1729 // references into JavaThread.)
1730
1731 // Check for final fields.
1732 const TypeInstPtr* tinst = flat->isa_instptr();
1733 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1734 ciField* field;
1735 if (tinst->const_oop() != nullptr &&
1736 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1737 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1738 // static field
1739 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1740 field = k->get_field_by_offset(tinst->offset(), true);
1741 } else {
1742 ciInstanceKlass *k = tinst->instance_klass();
1743 field = k->get_field_by_offset(tinst->offset(), false);
1744 }
1745 assert(field == nullptr ||
1746 original_field == nullptr ||
1747 (field->holder() == original_field->holder() &&
1748 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1749 field->is_static() == original_field->is_static()), "wrong field?");
1750 // Set field() and is_rewritable() attributes.
1751 if (field != nullptr) alias_type(idx)->set_field(field);
1752 }
1753 }
1754
1755 // Fill the cache for next time.
1756 ace->_adr_type = adr_type;
1757 ace->_index = idx;
1758 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1759
1760 // Might as well try to fill the cache for the flattened version, too.
1761 AliasCacheEntry* face = probe_alias_cache(flat);
1762 if (face->_adr_type == nullptr) {
1763 face->_adr_type = flat;
1764 face->_index = idx;
1765 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1766 }
1767
1768 return alias_type(idx);
1769 }
1770
1771
1772 Compile::AliasType* Compile::alias_type(ciField* field) {
1773 const TypeOopPtr* t;
1774 if (field->is_static())
1775 t = TypeInstPtr::make(field->holder()->java_mirror());
1776 else
1777 t = TypeOopPtr::make_from_klass_raw(field->holder());
1778 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1779 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1780 return atp;
1781 }
1782
1783
1784 //------------------------------have_alias_type--------------------------------
1785 bool Compile::have_alias_type(const TypePtr* adr_type) {
1786 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1787 if (ace->_adr_type == adr_type) {
1788 return true;
1789 }
1790
1791 // Handle special cases.
1792 if (adr_type == nullptr) return true;
1793 if (adr_type == TypePtr::BOTTOM) return true;
1794
1795 return find_alias_type(adr_type, true, nullptr) != nullptr;
1796 }
1797
1798 //-----------------------------must_alias--------------------------------------
1799 // True if all values of the given address type are in the given alias category.
1800 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1801 if (alias_idx == AliasIdxBot) return true; // the universal category
1802 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1803 if (alias_idx == AliasIdxTop) return false; // the empty category
1804 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1805
1806 // the only remaining possible overlap is identity
1807 int adr_idx = get_alias_index(adr_type);
1808 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1809 assert(adr_idx == alias_idx ||
1810 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1811 && adr_type != TypeOopPtr::BOTTOM),
1812 "should not be testing for overlap with an unsafe pointer");
1813 return adr_idx == alias_idx;
1814 }
1815
1816 //------------------------------can_alias--------------------------------------
1817 // True if any values of the given address type are in the given alias category.
1818 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1819 if (alias_idx == AliasIdxTop) return false; // the empty category
1820 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1821 // Known instance doesn't alias with bottom memory
1822 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1823 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1824
1825 // the only remaining possible overlap is identity
1826 int adr_idx = get_alias_index(adr_type);
1827 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1828 return adr_idx == alias_idx;
1829 }
1830
1831 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1832 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1833 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1834 if (parse_predicate_count() == 0) {
1835 return;
1836 }
1837 for (int i = 0; i < parse_predicate_count(); i++) {
1838 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1839 parse_predicate->mark_useless();
1840 igvn._worklist.push(parse_predicate);
1841 }
1842 _parse_predicates.clear();
1843 }
1844
1845 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1846 if (!n->for_post_loop_opts_igvn()) {
1847 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1848 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1849 _for_post_loop_igvn.append(n);
1850 }
1851 }
1852
1853 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1854 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1855 _for_post_loop_igvn.remove(n);
1856 }
1857
1858 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1859 // Verify that all previous optimizations produced a valid graph
1860 // at least to this point, even if no loop optimizations were done.
1861 PhaseIdealLoop::verify(igvn);
1862
1863 C->set_post_loop_opts_phase(); // no more loop opts allowed
1864
1865 assert(!C->major_progress(), "not cleared");
1866
1867 if (_for_post_loop_igvn.length() > 0) {
1868 while (_for_post_loop_igvn.length() > 0) {
1869 Node* n = _for_post_loop_igvn.pop();
1870 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1871 igvn._worklist.push(n);
1872 }
1873 igvn.optimize();
1874 if (failing()) return;
1875 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1876 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1877
1878 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1879 if (C->major_progress()) {
1880 C->clear_major_progress(); // ensure that major progress is now clear
1881 }
1882 }
1883 }
1884
1885 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1886 if (OptimizeUnstableIf) {
1887 _unstable_if_traps.append(trap);
1888 }
1889 }
1890
1891 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1892 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1893 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1894 Node* n = trap->uncommon_trap();
1895 if (!useful.member(n)) {
1896 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1897 }
1898 }
1899 }
1900
1901 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1902 // or fold-compares case. Return true if succeed or not found.
1903 //
1904 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1905 // false and ask fold-compares to yield.
1906 //
1907 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1908 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1909 // when deoptimization does happen.
1910 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1911 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1912 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1913 if (trap->uncommon_trap() == unc) {
1914 if (yield && trap->modified()) {
1915 return false;
1916 }
1917 _unstable_if_traps.delete_at(i);
1918 break;
1919 }
1920 }
1921 return true;
1922 }
1923
1924 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1925 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1926 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1927 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1928 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1929 CallStaticJavaNode* unc = trap->uncommon_trap();
1930 int next_bci = trap->next_bci();
1931 bool modified = trap->modified();
1932
1933 if (next_bci != -1 && !modified) {
1934 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1935 JVMState* jvms = unc->jvms();
1936 ciMethod* method = jvms->method();
1937 ciBytecodeStream iter(method);
1938
1939 iter.force_bci(jvms->bci());
1940 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1941 Bytecodes::Code c = iter.cur_bc();
1942 Node* lhs = nullptr;
1943 Node* rhs = nullptr;
1944 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1945 lhs = unc->peek_operand(0);
1946 rhs = unc->peek_operand(1);
1947 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1948 lhs = unc->peek_operand(0);
1949 }
1950
1951 ResourceMark rm;
1952 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1953 assert(live_locals.is_valid(), "broken liveness info");
1954 int len = (int)live_locals.size();
1955
1956 for (int i = 0; i < len; i++) {
1957 Node* local = unc->local(jvms, i);
1958 // kill local using the liveness of next_bci.
1959 // give up when the local looks like an operand to secure reexecution.
1960 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1961 uint idx = jvms->locoff() + i;
1962 #ifdef ASSERT
1963 if (PrintOpto && Verbose) {
1964 tty->print("[unstable_if] kill local#%d: ", idx);
1965 local->dump();
1966 tty->cr();
1967 }
1968 #endif
1969 igvn.replace_input_of(unc, idx, top());
1970 modified = true;
1971 }
1972 }
1973 }
1974
1975 // keep the mondified trap for late query
1976 if (modified) {
1977 trap->set_modified();
1978 } else {
1979 _unstable_if_traps.delete_at(i);
1980 }
1981 }
1982 igvn.optimize();
1983 }
1984
1985 // StringOpts and late inlining of string methods
1986 void Compile::inline_string_calls(bool parse_time) {
1987 {
1988 // remove useless nodes to make the usage analysis simpler
1989 ResourceMark rm;
1990 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1991 }
1992
1993 {
1994 ResourceMark rm;
1995 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1996 PhaseStringOpts pso(initial_gvn());
1997 print_method(PHASE_AFTER_STRINGOPTS, 3);
1998 }
1999
2000 // now inline anything that we skipped the first time around
2001 if (!parse_time) {
2002 _late_inlines_pos = _late_inlines.length();
2003 }
2004
2005 while (_string_late_inlines.length() > 0) {
2006 CallGenerator* cg = _string_late_inlines.pop();
2007 cg->do_late_inline();
2008 if (failing()) return;
2009 }
2010 _string_late_inlines.trunc_to(0);
2011 }
2012
2013 // Late inlining of boxing methods
2014 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2015 if (_boxing_late_inlines.length() > 0) {
2016 assert(has_boxed_value(), "inconsistent");
2017
2018 set_inlining_incrementally(true);
2019
2020 igvn_worklist()->ensure_empty(); // should be done with igvn
2021
2022 _late_inlines_pos = _late_inlines.length();
2023
2024 while (_boxing_late_inlines.length() > 0) {
2025 CallGenerator* cg = _boxing_late_inlines.pop();
2026 cg->do_late_inline();
2027 if (failing()) return;
2028 }
2029 _boxing_late_inlines.trunc_to(0);
2030
2031 inline_incrementally_cleanup(igvn);
2032
2033 set_inlining_incrementally(false);
2034 }
2035 }
2036
2037 bool Compile::inline_incrementally_one() {
2038 assert(IncrementalInline, "incremental inlining should be on");
2039
2040 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2041
2042 set_inlining_progress(false);
2043 set_do_cleanup(false);
2044
2045 for (int i = 0; i < _late_inlines.length(); i++) {
2046 _late_inlines_pos = i+1;
2047 CallGenerator* cg = _late_inlines.at(i);
2048 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2049 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2050 cg->do_late_inline();
2051 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2052 if (failing()) {
2053 return false;
2054 } else if (inlining_progress()) {
2055 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2056 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2057 break; // process one call site at a time
2058 }
2059 } else {
2060 // Ignore late inline direct calls when inlining is not allowed.
2061 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2062 }
2063 }
2064 // Remove processed elements.
2065 _late_inlines.remove_till(_late_inlines_pos);
2066 _late_inlines_pos = 0;
2067
2068 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2069
2070 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2071
2072 set_inlining_progress(false);
2073 set_do_cleanup(false);
2074
2075 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2076 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2077 }
2078
2079 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2080 {
2081 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2082 ResourceMark rm;
2083 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2084 }
2085 {
2086 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2087 igvn.reset_from_gvn(initial_gvn());
2088 igvn.optimize();
2089 if (failing()) return;
2090 }
2091 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2092 }
2093
2094 // Perform incremental inlining until bound on number of live nodes is reached
2095 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2096 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2097
2098 set_inlining_incrementally(true);
2099 uint low_live_nodes = 0;
2100
2101 while (_late_inlines.length() > 0) {
2102 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2103 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2104 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2105 // PhaseIdealLoop is expensive so we only try it once we are
2106 // out of live nodes and we only try it again if the previous
2107 // helped got the number of nodes down significantly
2108 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2109 if (failing()) return;
2110 low_live_nodes = live_nodes();
2111 _major_progress = true;
2112 }
2113
2114 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2115 bool do_print_inlining = print_inlining() || print_intrinsics();
2116 if (do_print_inlining || log() != nullptr) {
2117 // Print inlining message for candidates that we couldn't inline for lack of space.
2118 for (int i = 0; i < _late_inlines.length(); i++) {
2119 CallGenerator* cg = _late_inlines.at(i);
2120 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2121 if (do_print_inlining) {
2122 cg->print_inlining_late(InliningResult::FAILURE, msg);
2123 }
2124 log_late_inline_failure(cg, msg);
2125 }
2126 }
2127 break; // finish
2128 }
2129 }
2130
2131 igvn_worklist()->ensure_empty(); // should be done with igvn
2132
2133 while (inline_incrementally_one()) {
2134 assert(!failing(), "inconsistent");
2135 }
2136 if (failing()) return;
2137
2138 inline_incrementally_cleanup(igvn);
2139
2140 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2141
2142 if (failing()) return;
2143
2144 if (_late_inlines.length() == 0) {
2145 break; // no more progress
2146 }
2147 }
2148
2149 igvn_worklist()->ensure_empty(); // should be done with igvn
2150
2151 if (_string_late_inlines.length() > 0) {
2152 assert(has_stringbuilder(), "inconsistent");
2153
2154 inline_string_calls(false);
2155
2156 if (failing()) return;
2157
2158 inline_incrementally_cleanup(igvn);
2159 }
2160
2161 set_inlining_incrementally(false);
2162 }
2163
2164 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2165 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2166 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2167 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2168 // as if "inlining_incrementally() == true" were set.
2169 assert(inlining_incrementally() == false, "not allowed");
2170 assert(_modified_nodes == nullptr, "not allowed");
2171 assert(_late_inlines.length() > 0, "sanity");
2172
2173 while (_late_inlines.length() > 0) {
2174 igvn_worklist()->ensure_empty(); // should be done with igvn
2175
2176 while (inline_incrementally_one()) {
2177 assert(!failing(), "inconsistent");
2178 }
2179 if (failing()) return;
2180
2181 inline_incrementally_cleanup(igvn);
2182 }
2183 }
2184
2185 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2186 if (_loop_opts_cnt > 0) {
2187 while (major_progress() && (_loop_opts_cnt > 0)) {
2188 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2189 PhaseIdealLoop::optimize(igvn, mode);
2190 _loop_opts_cnt--;
2191 if (failing()) return false;
2192 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2193 }
2194 }
2195 return true;
2196 }
2197
2198 // Remove edges from "root" to each SafePoint at a backward branch.
2199 // They were inserted during parsing (see add_safepoint()) to make
2200 // infinite loops without calls or exceptions visible to root, i.e.,
2201 // useful.
2202 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2203 Node *r = root();
2204 if (r != nullptr) {
2205 for (uint i = r->req(); i < r->len(); ++i) {
2206 Node *n = r->in(i);
2207 if (n != nullptr && n->is_SafePoint()) {
2208 r->rm_prec(i);
2209 if (n->outcnt() == 0) {
2210 igvn.remove_dead_node(n);
2211 }
2212 --i;
2213 }
2214 }
2215 // Parsing may have added top inputs to the root node (Path
2216 // leading to the Halt node proven dead). Make sure we get a
2217 // chance to clean them up.
2218 igvn._worklist.push(r);
2219 igvn.optimize();
2220 }
2221 }
2222
2223 //------------------------------Optimize---------------------------------------
2224 // Given a graph, optimize it.
2225 void Compile::Optimize() {
2226 TracePhase tp("optimizer", &timers[_t_optimizer]);
2227
2228 #ifndef PRODUCT
2229 if (env()->break_at_compile()) {
2230 BREAKPOINT;
2231 }
2232
2233 #endif
2234
2235 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2236 #ifdef ASSERT
2237 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2238 #endif
2239
2240 ResourceMark rm;
2241
2242 print_inlining_reinit();
2243
2244 NOT_PRODUCT( verify_graph_edges(); )
2245
2246 print_method(PHASE_AFTER_PARSING, 1);
2247
2248 {
2249 // Iterative Global Value Numbering, including ideal transforms
2250 // Initialize IterGVN with types and values from parse-time GVN
2251 PhaseIterGVN igvn(initial_gvn());
2252 #ifdef ASSERT
2253 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2254 #endif
2255 {
2256 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2257 igvn.optimize();
2258 }
2259
2260 if (failing()) return;
2261
2262 print_method(PHASE_ITER_GVN1, 2);
2263
2264 process_for_unstable_if_traps(igvn);
2265
2266 if (failing()) return;
2267
2268 inline_incrementally(igvn);
2269
2270 print_method(PHASE_INCREMENTAL_INLINE, 2);
2271
2272 if (failing()) return;
2273
2274 if (eliminate_boxing()) {
2275 // Inline valueOf() methods now.
2276 inline_boxing_calls(igvn);
2277
2278 if (failing()) return;
2279
2280 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2281 inline_incrementally(igvn);
2282 }
2283
2284 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2285
2286 if (failing()) return;
2287 }
2288
2289 // Remove the speculative part of types and clean up the graph from
2290 // the extra CastPP nodes whose only purpose is to carry them. Do
2291 // that early so that optimizations are not disrupted by the extra
2292 // CastPP nodes.
2293 remove_speculative_types(igvn);
2294
2295 if (failing()) return;
2296
2297 // No more new expensive nodes will be added to the list from here
2298 // so keep only the actual candidates for optimizations.
2299 cleanup_expensive_nodes(igvn);
2300
2301 if (failing()) return;
2302
2303 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2304 if (EnableVectorSupport && has_vbox_nodes()) {
2305 TracePhase tp("", &timers[_t_vector]);
2306 PhaseVector pv(igvn);
2307 pv.optimize_vector_boxes();
2308 if (failing()) return;
2309 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2310 }
2311 assert(!has_vbox_nodes(), "sanity");
2312
2313 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2314 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2315 igvn_worklist()->ensure_empty(); // should be done with igvn
2316 {
2317 ResourceMark rm;
2318 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2319 }
2320 igvn.reset_from_gvn(initial_gvn());
2321 igvn.optimize();
2322 if (failing()) return;
2323 }
2324
2325 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2326 // safepoints
2327 remove_root_to_sfpts_edges(igvn);
2328
2329 if (failing()) return;
2330
2331 // Perform escape analysis
2332 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2333 if (has_loops()) {
2334 // Cleanup graph (remove dead nodes).
2335 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2336 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2337 if (failing()) return;
2338 }
2339 bool progress;
2340 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2341 do {
2342 ConnectionGraph::do_analysis(this, &igvn);
2343
2344 if (failing()) return;
2345
2346 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2347
2348 // Optimize out fields loads from scalar replaceable allocations.
2349 igvn.optimize();
2350 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2351
2352 if (failing()) return;
2353
2354 if (congraph() != nullptr && macro_count() > 0) {
2355 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2356 PhaseMacroExpand mexp(igvn);
2357 mexp.eliminate_macro_nodes();
2358 if (failing()) return;
2359
2360 igvn.set_delay_transform(false);
2361 igvn.optimize();
2362 if (failing()) return;
2363
2364 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2365 }
2366
2367 ConnectionGraph::verify_ram_nodes(this, root());
2368 if (failing()) return;
2369
2370 progress = do_iterative_escape_analysis() &&
2371 (macro_count() < mcount) &&
2372 ConnectionGraph::has_candidates(this);
2373 // Try again if candidates exist and made progress
2374 // by removing some allocations and/or locks.
2375 } while (progress);
2376 }
2377
2378 // Loop transforms on the ideal graph. Range Check Elimination,
2379 // peeling, unrolling, etc.
2380
2381 // Set loop opts counter
2382 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2383 {
2384 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2385 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2386 _loop_opts_cnt--;
2387 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2388 if (failing()) return;
2389 }
2390 // Loop opts pass if partial peeling occurred in previous pass
2391 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2392 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2393 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2394 _loop_opts_cnt--;
2395 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2396 if (failing()) return;
2397 }
2398 // Loop opts pass for loop-unrolling before CCP
2399 if(major_progress() && (_loop_opts_cnt > 0)) {
2400 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2401 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2402 _loop_opts_cnt--;
2403 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2404 }
2405 if (!failing()) {
2406 // Verify that last round of loop opts produced a valid graph
2407 PhaseIdealLoop::verify(igvn);
2408 }
2409 }
2410 if (failing()) return;
2411
2412 // Conditional Constant Propagation;
2413 print_method(PHASE_BEFORE_CCP1, 2);
2414 PhaseCCP ccp( &igvn );
2415 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2416 {
2417 TracePhase tp("ccp", &timers[_t_ccp]);
2418 ccp.do_transform();
2419 }
2420 print_method(PHASE_CCP1, 2);
2421
2422 assert( true, "Break here to ccp.dump_old2new_map()");
2423
2424 // Iterative Global Value Numbering, including ideal transforms
2425 {
2426 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2427 igvn.reset_from_igvn(&ccp);
2428 igvn.optimize();
2429 }
2430 print_method(PHASE_ITER_GVN2, 2);
2431
2432 if (failing()) return;
2433
2434 // Loop transforms on the ideal graph. Range Check Elimination,
2435 // peeling, unrolling, etc.
2436 if (!optimize_loops(igvn, LoopOptsDefault)) {
2437 return;
2438 }
2439
2440 if (failing()) return;
2441
2442 C->clear_major_progress(); // ensure that major progress is now clear
2443
2444 process_for_post_loop_opts_igvn(igvn);
2445
2446 if (failing()) return;
2447
2448 #ifdef ASSERT
2449 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2450 #endif
2451
2452 {
2453 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2454 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2455 PhaseMacroExpand mex(igvn);
2456 if (mex.expand_macro_nodes()) {
2457 assert(failing(), "must bail out w/ explicit message");
2458 return;
2459 }
2460 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2461 }
2462
2463 {
2464 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2465 if (bs->expand_barriers(this, igvn)) {
2466 assert(failing(), "must bail out w/ explicit message");
2467 return;
2468 }
2469 print_method(PHASE_BARRIER_EXPANSION, 2);
2470 }
2471
2472 if (C->max_vector_size() > 0) {
2473 C->optimize_logic_cones(igvn);
2474 igvn.optimize();
2475 if (failing()) return;
2476 }
2477
2478 DEBUG_ONLY( _modified_nodes = nullptr; )
2479
2480 assert(igvn._worklist.size() == 0, "not empty");
2481
2482 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2483
2484 if (_late_inlines.length() > 0) {
2485 // More opportunities to optimize virtual and MH calls.
2486 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2487 process_late_inline_calls_no_inline(igvn);
2488 if (failing()) return;
2489 }
2490 } // (End scope of igvn; run destructor if necessary for asserts.)
2491
2492 check_no_dead_use();
2493
2494 process_print_inlining();
2495
2496 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2497 // to remove hashes to unlock nodes for modifications.
2498 C->node_hash()->clear();
2499
2500 // A method with only infinite loops has no edges entering loops from root
2501 {
2502 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2503 if (final_graph_reshaping()) {
2504 assert(failing(), "must bail out w/ explicit message");
2505 return;
2506 }
2507 }
2508
2509 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2510 DEBUG_ONLY(set_phase_optimize_finished();)
2511 }
2512
2513 #ifdef ASSERT
2514 void Compile::check_no_dead_use() const {
2515 ResourceMark rm;
2516 Unique_Node_List wq;
2517 wq.push(root());
2518 for (uint i = 0; i < wq.size(); ++i) {
2519 Node* n = wq.at(i);
2520 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2521 Node* u = n->fast_out(j);
2522 if (u->outcnt() == 0 && !u->is_Con()) {
2523 u->dump();
2524 fatal("no reachable node should have no use");
2525 }
2526 wq.push(u);
2527 }
2528 }
2529 }
2530 #endif
2531
2532 void Compile::inline_vector_reboxing_calls() {
2533 if (C->_vector_reboxing_late_inlines.length() > 0) {
2534 _late_inlines_pos = C->_late_inlines.length();
2535 while (_vector_reboxing_late_inlines.length() > 0) {
2536 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2537 cg->do_late_inline();
2538 if (failing()) return;
2539 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2540 }
2541 _vector_reboxing_late_inlines.trunc_to(0);
2542 }
2543 }
2544
2545 bool Compile::has_vbox_nodes() {
2546 if (C->_vector_reboxing_late_inlines.length() > 0) {
2547 return true;
2548 }
2549 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2550 Node * n = C->macro_node(macro_idx);
2551 assert(n->is_macro(), "only macro nodes expected here");
2552 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2553 return true;
2554 }
2555 }
2556 return false;
2557 }
2558
2559 //---------------------------- Bitwise operation packing optimization ---------------------------
2560
2561 static bool is_vector_unary_bitwise_op(Node* n) {
2562 return n->Opcode() == Op_XorV &&
2563 VectorNode::is_vector_bitwise_not_pattern(n);
2564 }
2565
2566 static bool is_vector_binary_bitwise_op(Node* n) {
2567 switch (n->Opcode()) {
2568 case Op_AndV:
2569 case Op_OrV:
2570 return true;
2571
2572 case Op_XorV:
2573 return !is_vector_unary_bitwise_op(n);
2574
2575 default:
2576 return false;
2577 }
2578 }
2579
2580 static bool is_vector_ternary_bitwise_op(Node* n) {
2581 return n->Opcode() == Op_MacroLogicV;
2582 }
2583
2584 static bool is_vector_bitwise_op(Node* n) {
2585 return is_vector_unary_bitwise_op(n) ||
2586 is_vector_binary_bitwise_op(n) ||
2587 is_vector_ternary_bitwise_op(n);
2588 }
2589
2590 static bool is_vector_bitwise_cone_root(Node* n) {
2591 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2592 return false;
2593 }
2594 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2595 if (is_vector_bitwise_op(n->fast_out(i))) {
2596 return false;
2597 }
2598 }
2599 return true;
2600 }
2601
2602 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2603 uint cnt = 0;
2604 if (is_vector_bitwise_op(n)) {
2605 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2606 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2607 for (uint i = 1; i < inp_cnt; i++) {
2608 Node* in = n->in(i);
2609 bool skip = VectorNode::is_all_ones_vector(in);
2610 if (!skip && !inputs.member(in)) {
2611 inputs.push(in);
2612 cnt++;
2613 }
2614 }
2615 assert(cnt <= 1, "not unary");
2616 } else {
2617 uint last_req = inp_cnt;
2618 if (is_vector_ternary_bitwise_op(n)) {
2619 last_req = inp_cnt - 1; // skip last input
2620 }
2621 for (uint i = 1; i < last_req; i++) {
2622 Node* def = n->in(i);
2623 if (!inputs.member(def)) {
2624 inputs.push(def);
2625 cnt++;
2626 }
2627 }
2628 }
2629 } else { // not a bitwise operations
2630 if (!inputs.member(n)) {
2631 inputs.push(n);
2632 cnt++;
2633 }
2634 }
2635 return cnt;
2636 }
2637
2638 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2639 Unique_Node_List useful_nodes;
2640 C->identify_useful_nodes(useful_nodes);
2641
2642 for (uint i = 0; i < useful_nodes.size(); i++) {
2643 Node* n = useful_nodes.at(i);
2644 if (is_vector_bitwise_cone_root(n)) {
2645 list.push(n);
2646 }
2647 }
2648 }
2649
2650 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2651 const TypeVect* vt,
2652 Unique_Node_List& partition,
2653 Unique_Node_List& inputs) {
2654 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2655 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2656 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2657
2658 Node* in1 = inputs.at(0);
2659 Node* in2 = inputs.at(1);
2660 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2661
2662 uint func = compute_truth_table(partition, inputs);
2663
2664 Node* pn = partition.at(partition.size() - 1);
2665 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2666 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2667 }
2668
2669 static uint extract_bit(uint func, uint pos) {
2670 return (func & (1 << pos)) >> pos;
2671 }
2672
2673 //
2674 // A macro logic node represents a truth table. It has 4 inputs,
2675 // First three inputs corresponds to 3 columns of a truth table
2676 // and fourth input captures the logic function.
2677 //
2678 // eg. fn = (in1 AND in2) OR in3;
2679 //
2680 // MacroNode(in1,in2,in3,fn)
2681 //
2682 // -----------------
2683 // in1 in2 in3 fn
2684 // -----------------
2685 // 0 0 0 0
2686 // 0 0 1 1
2687 // 0 1 0 0
2688 // 0 1 1 1
2689 // 1 0 0 0
2690 // 1 0 1 1
2691 // 1 1 0 1
2692 // 1 1 1 1
2693 //
2694
2695 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2696 int res = 0;
2697 for (int i = 0; i < 8; i++) {
2698 int bit1 = extract_bit(in1, i);
2699 int bit2 = extract_bit(in2, i);
2700 int bit3 = extract_bit(in3, i);
2701
2702 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2703 int func_bit = extract_bit(func, func_bit_pos);
2704
2705 res |= func_bit << i;
2706 }
2707 return res;
2708 }
2709
2710 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2711 assert(n != nullptr, "");
2712 assert(eval_map.contains(n), "absent");
2713 return *(eval_map.get(n));
2714 }
2715
2716 static void eval_operands(Node* n,
2717 uint& func1, uint& func2, uint& func3,
2718 ResourceHashtable<Node*,uint>& eval_map) {
2719 assert(is_vector_bitwise_op(n), "");
2720
2721 if (is_vector_unary_bitwise_op(n)) {
2722 Node* opnd = n->in(1);
2723 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2724 opnd = n->in(2);
2725 }
2726 func1 = eval_operand(opnd, eval_map);
2727 } else if (is_vector_binary_bitwise_op(n)) {
2728 func1 = eval_operand(n->in(1), eval_map);
2729 func2 = eval_operand(n->in(2), eval_map);
2730 } else {
2731 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2732 func1 = eval_operand(n->in(1), eval_map);
2733 func2 = eval_operand(n->in(2), eval_map);
2734 func3 = eval_operand(n->in(3), eval_map);
2735 }
2736 }
2737
2738 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2739 assert(inputs.size() <= 3, "sanity");
2740 ResourceMark rm;
2741 uint res = 0;
2742 ResourceHashtable<Node*,uint> eval_map;
2743
2744 // Populate precomputed functions for inputs.
2745 // Each input corresponds to one column of 3 input truth-table.
2746 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2747 0xCC, // (_, b, _) -> b
2748 0xF0 }; // (a, _, _) -> a
2749 for (uint i = 0; i < inputs.size(); i++) {
2750 eval_map.put(inputs.at(i), input_funcs[2-i]);
2751 }
2752
2753 for (uint i = 0; i < partition.size(); i++) {
2754 Node* n = partition.at(i);
2755
2756 uint func1 = 0, func2 = 0, func3 = 0;
2757 eval_operands(n, func1, func2, func3, eval_map);
2758
2759 switch (n->Opcode()) {
2760 case Op_OrV:
2761 assert(func3 == 0, "not binary");
2762 res = func1 | func2;
2763 break;
2764 case Op_AndV:
2765 assert(func3 == 0, "not binary");
2766 res = func1 & func2;
2767 break;
2768 case Op_XorV:
2769 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2770 assert(func2 == 0 && func3 == 0, "not unary");
2771 res = (~func1) & 0xFF;
2772 } else {
2773 assert(func3 == 0, "not binary");
2774 res = func1 ^ func2;
2775 }
2776 break;
2777 case Op_MacroLogicV:
2778 // Ordering of inputs may change during evaluation of sub-tree
2779 // containing MacroLogic node as a child node, thus a re-evaluation
2780 // makes sure that function is evaluated in context of current
2781 // inputs.
2782 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2783 break;
2784
2785 default: assert(false, "not supported: %s", n->Name());
2786 }
2787 assert(res <= 0xFF, "invalid");
2788 eval_map.put(n, res);
2789 }
2790 return res;
2791 }
2792
2793 // Criteria under which nodes gets packed into a macro logic node:-
2794 // 1) Parent and both child nodes are all unmasked or masked with
2795 // same predicates.
2796 // 2) Masked parent can be packed with left child if it is predicated
2797 // and both have same predicates.
2798 // 3) Masked parent can be packed with right child if its un-predicated
2799 // or has matching predication condition.
2800 // 4) An unmasked parent can be packed with an unmasked child.
2801 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2802 assert(partition.size() == 0, "not empty");
2803 assert(inputs.size() == 0, "not empty");
2804 if (is_vector_ternary_bitwise_op(n)) {
2805 return false;
2806 }
2807
2808 bool is_unary_op = is_vector_unary_bitwise_op(n);
2809 if (is_unary_op) {
2810 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2811 return false; // too few inputs
2812 }
2813
2814 bool pack_left_child = true;
2815 bool pack_right_child = true;
2816
2817 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2818 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2819
2820 int left_child_input_cnt = 0;
2821 int right_child_input_cnt = 0;
2822
2823 bool parent_is_predicated = n->is_predicated_vector();
2824 bool left_child_predicated = n->in(1)->is_predicated_vector();
2825 bool right_child_predicated = n->in(2)->is_predicated_vector();
2826
2827 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2828 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2829 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2830
2831 do {
2832 if (pack_left_child && left_child_LOP &&
2833 ((!parent_is_predicated && !left_child_predicated) ||
2834 ((parent_is_predicated && left_child_predicated &&
2835 parent_pred == left_child_pred)))) {
2836 partition.push(n->in(1));
2837 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2838 } else {
2839 inputs.push(n->in(1));
2840 left_child_input_cnt = 1;
2841 }
2842
2843 if (pack_right_child && right_child_LOP &&
2844 (!right_child_predicated ||
2845 (right_child_predicated && parent_is_predicated &&
2846 parent_pred == right_child_pred))) {
2847 partition.push(n->in(2));
2848 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2849 } else {
2850 inputs.push(n->in(2));
2851 right_child_input_cnt = 1;
2852 }
2853
2854 if (inputs.size() > 3) {
2855 assert(partition.size() > 0, "");
2856 inputs.clear();
2857 partition.clear();
2858 if (left_child_input_cnt > right_child_input_cnt) {
2859 pack_left_child = false;
2860 } else {
2861 pack_right_child = false;
2862 }
2863 } else {
2864 break;
2865 }
2866 } while(true);
2867
2868 if(partition.size()) {
2869 partition.push(n);
2870 }
2871
2872 return (partition.size() == 2 || partition.size() == 3) &&
2873 (inputs.size() == 2 || inputs.size() == 3);
2874 }
2875
2876 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2877 assert(is_vector_bitwise_op(n), "not a root");
2878
2879 visited.set(n->_idx);
2880
2881 // 1) Do a DFS walk over the logic cone.
2882 for (uint i = 1; i < n->req(); i++) {
2883 Node* in = n->in(i);
2884 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2885 process_logic_cone_root(igvn, in, visited);
2886 }
2887 }
2888
2889 // 2) Bottom up traversal: Merge node[s] with
2890 // the parent to form macro logic node.
2891 Unique_Node_List partition;
2892 Unique_Node_List inputs;
2893 if (compute_logic_cone(n, partition, inputs)) {
2894 const TypeVect* vt = n->bottom_type()->is_vect();
2895 Node* pn = partition.at(partition.size() - 1);
2896 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2897 if (mask == nullptr ||
2898 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2899 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2900 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2901 igvn.replace_node(n, macro_logic);
2902 }
2903 }
2904 }
2905
2906 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2907 ResourceMark rm;
2908 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2909 Unique_Node_List list;
2910 collect_logic_cone_roots(list);
2911
2912 while (list.size() > 0) {
2913 Node* n = list.pop();
2914 const TypeVect* vt = n->bottom_type()->is_vect();
2915 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2916 if (supported) {
2917 VectorSet visited(comp_arena());
2918 process_logic_cone_root(igvn, n, visited);
2919 }
2920 }
2921 }
2922 }
2923
2924 //------------------------------Code_Gen---------------------------------------
2925 // Given a graph, generate code for it
2926 void Compile::Code_Gen() {
2927 if (failing()) {
2928 return;
2929 }
2930
2931 // Perform instruction selection. You might think we could reclaim Matcher
2932 // memory PDQ, but actually the Matcher is used in generating spill code.
2933 // Internals of the Matcher (including some VectorSets) must remain live
2934 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2935 // set a bit in reclaimed memory.
2936
2937 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2938 // nodes. Mapping is only valid at the root of each matched subtree.
2939 NOT_PRODUCT( verify_graph_edges(); )
2940
2941 Matcher matcher;
2942 _matcher = &matcher;
2943 {
2944 TracePhase tp("matcher", &timers[_t_matcher]);
2945 matcher.match();
2946 if (failing()) {
2947 return;
2948 }
2949 }
2950 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2951 // nodes. Mapping is only valid at the root of each matched subtree.
2952 NOT_PRODUCT( verify_graph_edges(); )
2953
2954 // If you have too many nodes, or if matching has failed, bail out
2955 check_node_count(0, "out of nodes matching instructions");
2956 if (failing()) {
2957 return;
2958 }
2959
2960 print_method(PHASE_MATCHING, 2);
2961
2962 // Build a proper-looking CFG
2963 PhaseCFG cfg(node_arena(), root(), matcher);
2964 _cfg = &cfg;
2965 {
2966 TracePhase tp("scheduler", &timers[_t_scheduler]);
2967 bool success = cfg.do_global_code_motion();
2968 if (!success) {
2969 return;
2970 }
2971
2972 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2973 NOT_PRODUCT( verify_graph_edges(); )
2974 cfg.verify();
2975 }
2976
2977 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2978 _regalloc = ®alloc;
2979 {
2980 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2981 // Perform register allocation. After Chaitin, use-def chains are
2982 // no longer accurate (at spill code) and so must be ignored.
2983 // Node->LRG->reg mappings are still accurate.
2984 _regalloc->Register_Allocate();
2985
2986 // Bail out if the allocator builds too many nodes
2987 if (failing()) {
2988 return;
2989 }
2990
2991 print_method(PHASE_REGISTER_ALLOCATION, 2);
2992 }
2993
2994 // Prior to register allocation we kept empty basic blocks in case the
2995 // the allocator needed a place to spill. After register allocation we
2996 // are not adding any new instructions. If any basic block is empty, we
2997 // can now safely remove it.
2998 {
2999 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
3000 cfg.remove_empty_blocks();
3001 if (do_freq_based_layout()) {
3002 PhaseBlockLayout layout(cfg);
3003 } else {
3004 cfg.set_loop_alignment();
3005 }
3006 cfg.fixup_flow();
3007 cfg.remove_unreachable_blocks();
3008 cfg.verify_dominator_tree();
3009 print_method(PHASE_BLOCK_ORDERING, 3);
3010 }
3011
3012 // Apply peephole optimizations
3013 if( OptoPeephole ) {
3014 TracePhase tp("peephole", &timers[_t_peephole]);
3015 PhasePeephole peep( _regalloc, cfg);
3016 peep.do_transform();
3017 print_method(PHASE_PEEPHOLE, 3);
3018 }
3019
3020 // Do late expand if CPU requires this.
3021 if (Matcher::require_postalloc_expand) {
3022 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3023 cfg.postalloc_expand(_regalloc);
3024 print_method(PHASE_POSTALLOC_EXPAND, 3);
3025 }
3026
3027 // Convert Nodes to instruction bits in a buffer
3028 {
3029 TracePhase tp("output", &timers[_t_output]);
3030 PhaseOutput output;
3031 output.Output();
3032 if (failing()) return;
3033 output.install();
3034 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3035 }
3036
3037 // He's dead, Jim.
3038 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3039 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3040 }
3041
3042 //------------------------------Final_Reshape_Counts---------------------------
3043 // This class defines counters to help identify when a method
3044 // may/must be executed using hardware with only 24-bit precision.
3045 struct Final_Reshape_Counts : public StackObj {
3046 int _call_count; // count non-inlined 'common' calls
3047 int _float_count; // count float ops requiring 24-bit precision
3048 int _double_count; // count double ops requiring more precision
3049 int _java_call_count; // count non-inlined 'java' calls
3050 int _inner_loop_count; // count loops which need alignment
3051 VectorSet _visited; // Visitation flags
3052 Node_List _tests; // Set of IfNodes & PCTableNodes
3053
3054 Final_Reshape_Counts() :
3055 _call_count(0), _float_count(0), _double_count(0),
3056 _java_call_count(0), _inner_loop_count(0) { }
3057
3058 void inc_call_count () { _call_count ++; }
3059 void inc_float_count () { _float_count ++; }
3060 void inc_double_count() { _double_count++; }
3061 void inc_java_call_count() { _java_call_count++; }
3062 void inc_inner_loop_count() { _inner_loop_count++; }
3063
3064 int get_call_count () const { return _call_count ; }
3065 int get_float_count () const { return _float_count ; }
3066 int get_double_count() const { return _double_count; }
3067 int get_java_call_count() const { return _java_call_count; }
3068 int get_inner_loop_count() const { return _inner_loop_count; }
3069 };
3070
3071 // Eliminate trivially redundant StoreCMs and accumulate their
3072 // precedence edges.
3073 void Compile::eliminate_redundant_card_marks(Node* n) {
3074 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3075 if (n->in(MemNode::Address)->outcnt() > 1) {
3076 // There are multiple users of the same address so it might be
3077 // possible to eliminate some of the StoreCMs
3078 Node* mem = n->in(MemNode::Memory);
3079 Node* adr = n->in(MemNode::Address);
3080 Node* val = n->in(MemNode::ValueIn);
3081 Node* prev = n;
3082 bool done = false;
3083 // Walk the chain of StoreCMs eliminating ones that match. As
3084 // long as it's a chain of single users then the optimization is
3085 // safe. Eliminating partially redundant StoreCMs would require
3086 // cloning copies down the other paths.
3087 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3088 if (adr == mem->in(MemNode::Address) &&
3089 val == mem->in(MemNode::ValueIn)) {
3090 // redundant StoreCM
3091 if (mem->req() > MemNode::OopStore) {
3092 // Hasn't been processed by this code yet.
3093 n->add_prec(mem->in(MemNode::OopStore));
3094 } else {
3095 // Already converted to precedence edge
3096 for (uint i = mem->req(); i < mem->len(); i++) {
3097 // Accumulate any precedence edges
3098 if (mem->in(i) != nullptr) {
3099 n->add_prec(mem->in(i));
3100 }
3101 }
3102 // Everything above this point has been processed.
3103 done = true;
3104 }
3105 // Eliminate the previous StoreCM
3106 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3107 assert(mem->outcnt() == 0, "should be dead");
3108 mem->disconnect_inputs(this);
3109 } else {
3110 prev = mem;
3111 }
3112 mem = prev->in(MemNode::Memory);
3113 }
3114 }
3115 }
3116
3117 //------------------------------final_graph_reshaping_impl----------------------
3118 // Implement items 1-5 from final_graph_reshaping below.
3119 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3120
3121 if ( n->outcnt() == 0 ) return; // dead node
3122 uint nop = n->Opcode();
3123
3124 // Check for 2-input instruction with "last use" on right input.
3125 // Swap to left input. Implements item (2).
3126 if( n->req() == 3 && // two-input instruction
3127 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3128 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3129 n->in(2)->outcnt() == 1 &&// right use IS a last use
3130 !n->in(2)->is_Con() ) { // right use is not a constant
3131 // Check for commutative opcode
3132 switch( nop ) {
3133 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3134 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3135 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3136 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3137 case Op_AndL: case Op_XorL: case Op_OrL:
3138 case Op_AndI: case Op_XorI: case Op_OrI: {
3139 // Move "last use" input to left by swapping inputs
3140 n->swap_edges(1, 2);
3141 break;
3142 }
3143 default:
3144 break;
3145 }
3146 }
3147
3148 #ifdef ASSERT
3149 if( n->is_Mem() ) {
3150 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3151 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3152 // oop will be recorded in oop map if load crosses safepoint
3153 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3154 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3155 "raw memory operations should have control edge");
3156 }
3157 if (n->is_MemBar()) {
3158 MemBarNode* mb = n->as_MemBar();
3159 if (mb->trailing_store() || mb->trailing_load_store()) {
3160 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3161 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3162 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3163 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3164 } else if (mb->leading()) {
3165 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3166 }
3167 }
3168 #endif
3169 // Count FPU ops and common calls, implements item (3)
3170 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3171 if (!gc_handled) {
3172 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3173 }
3174
3175 // Collect CFG split points
3176 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3177 frc._tests.push(n);
3178 }
3179 }
3180
3181 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3182 switch( nop ) {
3183 // Count all float operations that may use FPU
3184 case Op_AddF:
3185 case Op_SubF:
3186 case Op_MulF:
3187 case Op_DivF:
3188 case Op_NegF:
3189 case Op_ModF:
3190 case Op_ConvI2F:
3191 case Op_ConF:
3192 case Op_CmpF:
3193 case Op_CmpF3:
3194 case Op_StoreF:
3195 case Op_LoadF:
3196 // case Op_ConvL2F: // longs are split into 32-bit halves
3197 frc.inc_float_count();
3198 break;
3199
3200 case Op_ConvF2D:
3201 case Op_ConvD2F:
3202 frc.inc_float_count();
3203 frc.inc_double_count();
3204 break;
3205
3206 // Count all double operations that may use FPU
3207 case Op_AddD:
3208 case Op_SubD:
3209 case Op_MulD:
3210 case Op_DivD:
3211 case Op_NegD:
3212 case Op_ModD:
3213 case Op_ConvI2D:
3214 case Op_ConvD2I:
3215 // case Op_ConvL2D: // handled by leaf call
3216 // case Op_ConvD2L: // handled by leaf call
3217 case Op_ConD:
3218 case Op_CmpD:
3219 case Op_CmpD3:
3220 case Op_StoreD:
3221 case Op_LoadD:
3222 case Op_LoadD_unaligned:
3223 frc.inc_double_count();
3224 break;
3225 case Op_Opaque1: // Remove Opaque Nodes before matching
3226 case Op_Opaque3:
3227 n->subsume_by(n->in(1), this);
3228 break;
3229 case Op_CallStaticJava:
3230 case Op_CallJava:
3231 case Op_CallDynamicJava:
3232 frc.inc_java_call_count(); // Count java call site;
3233 case Op_CallRuntime:
3234 case Op_CallLeaf:
3235 case Op_CallLeafVector:
3236 case Op_CallLeafNoFP: {
3237 assert (n->is_Call(), "");
3238 CallNode *call = n->as_Call();
3239 // Count call sites where the FP mode bit would have to be flipped.
3240 // Do not count uncommon runtime calls:
3241 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3242 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3243 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3244 frc.inc_call_count(); // Count the call site
3245 } else { // See if uncommon argument is shared
3246 Node *n = call->in(TypeFunc::Parms);
3247 int nop = n->Opcode();
3248 // Clone shared simple arguments to uncommon calls, item (1).
3249 if (n->outcnt() > 1 &&
3250 !n->is_Proj() &&
3251 nop != Op_CreateEx &&
3252 nop != Op_CheckCastPP &&
3253 nop != Op_DecodeN &&
3254 nop != Op_DecodeNKlass &&
3255 !n->is_Mem() &&
3256 !n->is_Phi()) {
3257 Node *x = n->clone();
3258 call->set_req(TypeFunc::Parms, x);
3259 }
3260 }
3261 break;
3262 }
3263
3264 case Op_StoreCM:
3265 {
3266 // Convert OopStore dependence into precedence edge
3267 Node* prec = n->in(MemNode::OopStore);
3268 n->del_req(MemNode::OopStore);
3269 n->add_prec(prec);
3270 eliminate_redundant_card_marks(n);
3271 }
3272
3273 // fall through
3274
3275 case Op_StoreB:
3276 case Op_StoreC:
3277 case Op_StoreI:
3278 case Op_StoreL:
3279 case Op_CompareAndSwapB:
3280 case Op_CompareAndSwapS:
3281 case Op_CompareAndSwapI:
3282 case Op_CompareAndSwapL:
3283 case Op_CompareAndSwapP:
3284 case Op_CompareAndSwapN:
3285 case Op_WeakCompareAndSwapB:
3286 case Op_WeakCompareAndSwapS:
3287 case Op_WeakCompareAndSwapI:
3288 case Op_WeakCompareAndSwapL:
3289 case Op_WeakCompareAndSwapP:
3290 case Op_WeakCompareAndSwapN:
3291 case Op_CompareAndExchangeB:
3292 case Op_CompareAndExchangeS:
3293 case Op_CompareAndExchangeI:
3294 case Op_CompareAndExchangeL:
3295 case Op_CompareAndExchangeP:
3296 case Op_CompareAndExchangeN:
3297 case Op_GetAndAddS:
3298 case Op_GetAndAddB:
3299 case Op_GetAndAddI:
3300 case Op_GetAndAddL:
3301 case Op_GetAndSetS:
3302 case Op_GetAndSetB:
3303 case Op_GetAndSetI:
3304 case Op_GetAndSetL:
3305 case Op_GetAndSetP:
3306 case Op_GetAndSetN:
3307 case Op_StoreP:
3308 case Op_StoreN:
3309 case Op_StoreNKlass:
3310 case Op_LoadB:
3311 case Op_LoadUB:
3312 case Op_LoadUS:
3313 case Op_LoadI:
3314 case Op_LoadKlass:
3315 case Op_LoadNKlass:
3316 case Op_LoadL:
3317 case Op_LoadL_unaligned:
3318 case Op_LoadP:
3319 case Op_LoadN:
3320 case Op_LoadRange:
3321 case Op_LoadS:
3322 break;
3323
3324 case Op_AddP: { // Assert sane base pointers
3325 Node *addp = n->in(AddPNode::Address);
3326 assert( !addp->is_AddP() ||
3327 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3328 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3329 "Base pointers must match (addp %u)", addp->_idx );
3330 #ifdef _LP64
3331 if ((UseCompressedOops || UseCompressedClassPointers) &&
3332 addp->Opcode() == Op_ConP &&
3333 addp == n->in(AddPNode::Base) &&
3334 n->in(AddPNode::Offset)->is_Con()) {
3335 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3336 // on the platform and on the compressed oops mode.
3337 // Use addressing with narrow klass to load with offset on x86.
3338 // Some platforms can use the constant pool to load ConP.
3339 // Do this transformation here since IGVN will convert ConN back to ConP.
3340 const Type* t = addp->bottom_type();
3341 bool is_oop = t->isa_oopptr() != nullptr;
3342 bool is_klass = t->isa_klassptr() != nullptr;
3343
3344 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3345 (is_klass && Matcher::const_klass_prefer_decode())) {
3346 Node* nn = nullptr;
3347
3348 int op = is_oop ? Op_ConN : Op_ConNKlass;
3349
3350 // Look for existing ConN node of the same exact type.
3351 Node* r = root();
3352 uint cnt = r->outcnt();
3353 for (uint i = 0; i < cnt; i++) {
3354 Node* m = r->raw_out(i);
3355 if (m!= nullptr && m->Opcode() == op &&
3356 m->bottom_type()->make_ptr() == t) {
3357 nn = m;
3358 break;
3359 }
3360 }
3361 if (nn != nullptr) {
3362 // Decode a narrow oop to match address
3363 // [R12 + narrow_oop_reg<<3 + offset]
3364 if (is_oop) {
3365 nn = new DecodeNNode(nn, t);
3366 } else {
3367 nn = new DecodeNKlassNode(nn, t);
3368 }
3369 // Check for succeeding AddP which uses the same Base.
3370 // Otherwise we will run into the assertion above when visiting that guy.
3371 for (uint i = 0; i < n->outcnt(); ++i) {
3372 Node *out_i = n->raw_out(i);
3373 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3374 out_i->set_req(AddPNode::Base, nn);
3375 #ifdef ASSERT
3376 for (uint j = 0; j < out_i->outcnt(); ++j) {
3377 Node *out_j = out_i->raw_out(j);
3378 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3379 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3380 }
3381 #endif
3382 }
3383 }
3384 n->set_req(AddPNode::Base, nn);
3385 n->set_req(AddPNode::Address, nn);
3386 if (addp->outcnt() == 0) {
3387 addp->disconnect_inputs(this);
3388 }
3389 }
3390 }
3391 }
3392 #endif
3393 break;
3394 }
3395
3396 case Op_CastPP: {
3397 // Remove CastPP nodes to gain more freedom during scheduling but
3398 // keep the dependency they encode as control or precedence edges
3399 // (if control is set already) on memory operations. Some CastPP
3400 // nodes don't have a control (don't carry a dependency): skip
3401 // those.
3402 if (n->in(0) != nullptr) {
3403 ResourceMark rm;
3404 Unique_Node_List wq;
3405 wq.push(n);
3406 for (uint next = 0; next < wq.size(); ++next) {
3407 Node *m = wq.at(next);
3408 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3409 Node* use = m->fast_out(i);
3410 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3411 use->ensure_control_or_add_prec(n->in(0));
3412 } else {
3413 switch(use->Opcode()) {
3414 case Op_AddP:
3415 case Op_DecodeN:
3416 case Op_DecodeNKlass:
3417 case Op_CheckCastPP:
3418 case Op_CastPP:
3419 wq.push(use);
3420 break;
3421 }
3422 }
3423 }
3424 }
3425 }
3426 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3427 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3428 Node* in1 = n->in(1);
3429 const Type* t = n->bottom_type();
3430 Node* new_in1 = in1->clone();
3431 new_in1->as_DecodeN()->set_type(t);
3432
3433 if (!Matcher::narrow_oop_use_complex_address()) {
3434 //
3435 // x86, ARM and friends can handle 2 adds in addressing mode
3436 // and Matcher can fold a DecodeN node into address by using
3437 // a narrow oop directly and do implicit null check in address:
3438 //
3439 // [R12 + narrow_oop_reg<<3 + offset]
3440 // NullCheck narrow_oop_reg
3441 //
3442 // On other platforms (Sparc) we have to keep new DecodeN node and
3443 // use it to do implicit null check in address:
3444 //
3445 // decode_not_null narrow_oop_reg, base_reg
3446 // [base_reg + offset]
3447 // NullCheck base_reg
3448 //
3449 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3450 // to keep the information to which null check the new DecodeN node
3451 // corresponds to use it as value in implicit_null_check().
3452 //
3453 new_in1->set_req(0, n->in(0));
3454 }
3455
3456 n->subsume_by(new_in1, this);
3457 if (in1->outcnt() == 0) {
3458 in1->disconnect_inputs(this);
3459 }
3460 } else {
3461 n->subsume_by(n->in(1), this);
3462 if (n->outcnt() == 0) {
3463 n->disconnect_inputs(this);
3464 }
3465 }
3466 break;
3467 }
3468 #ifdef _LP64
3469 case Op_CmpP:
3470 // Do this transformation here to preserve CmpPNode::sub() and
3471 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3472 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3473 Node* in1 = n->in(1);
3474 Node* in2 = n->in(2);
3475 if (!in1->is_DecodeNarrowPtr()) {
3476 in2 = in1;
3477 in1 = n->in(2);
3478 }
3479 assert(in1->is_DecodeNarrowPtr(), "sanity");
3480
3481 Node* new_in2 = nullptr;
3482 if (in2->is_DecodeNarrowPtr()) {
3483 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3484 new_in2 = in2->in(1);
3485 } else if (in2->Opcode() == Op_ConP) {
3486 const Type* t = in2->bottom_type();
3487 if (t == TypePtr::NULL_PTR) {
3488 assert(in1->is_DecodeN(), "compare klass to null?");
3489 // Don't convert CmpP null check into CmpN if compressed
3490 // oops implicit null check is not generated.
3491 // This will allow to generate normal oop implicit null check.
3492 if (Matcher::gen_narrow_oop_implicit_null_checks())
3493 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3494 //
3495 // This transformation together with CastPP transformation above
3496 // will generated code for implicit null checks for compressed oops.
3497 //
3498 // The original code after Optimize()
3499 //
3500 // LoadN memory, narrow_oop_reg
3501 // decode narrow_oop_reg, base_reg
3502 // CmpP base_reg, nullptr
3503 // CastPP base_reg // NotNull
3504 // Load [base_reg + offset], val_reg
3505 //
3506 // after these transformations will be
3507 //
3508 // LoadN memory, narrow_oop_reg
3509 // CmpN narrow_oop_reg, nullptr
3510 // decode_not_null narrow_oop_reg, base_reg
3511 // Load [base_reg + offset], val_reg
3512 //
3513 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3514 // since narrow oops can be used in debug info now (see the code in
3515 // final_graph_reshaping_walk()).
3516 //
3517 // At the end the code will be matched to
3518 // on x86:
3519 //
3520 // Load_narrow_oop memory, narrow_oop_reg
3521 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3522 // NullCheck narrow_oop_reg
3523 //
3524 // and on sparc:
3525 //
3526 // Load_narrow_oop memory, narrow_oop_reg
3527 // decode_not_null narrow_oop_reg, base_reg
3528 // Load [base_reg + offset], val_reg
3529 // NullCheck base_reg
3530 //
3531 } else if (t->isa_oopptr()) {
3532 new_in2 = ConNode::make(t->make_narrowoop());
3533 } else if (t->isa_klassptr()) {
3534 new_in2 = ConNode::make(t->make_narrowklass());
3535 }
3536 }
3537 if (new_in2 != nullptr) {
3538 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3539 n->subsume_by(cmpN, this);
3540 if (in1->outcnt() == 0) {
3541 in1->disconnect_inputs(this);
3542 }
3543 if (in2->outcnt() == 0) {
3544 in2->disconnect_inputs(this);
3545 }
3546 }
3547 }
3548 break;
3549
3550 case Op_DecodeN:
3551 case Op_DecodeNKlass:
3552 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3553 // DecodeN could be pinned when it can't be fold into
3554 // an address expression, see the code for Op_CastPP above.
3555 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3556 break;
3557
3558 case Op_EncodeP:
3559 case Op_EncodePKlass: {
3560 Node* in1 = n->in(1);
3561 if (in1->is_DecodeNarrowPtr()) {
3562 n->subsume_by(in1->in(1), this);
3563 } else if (in1->Opcode() == Op_ConP) {
3564 const Type* t = in1->bottom_type();
3565 if (t == TypePtr::NULL_PTR) {
3566 assert(t->isa_oopptr(), "null klass?");
3567 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3568 } else if (t->isa_oopptr()) {
3569 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3570 } else if (t->isa_klassptr()) {
3571 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3572 }
3573 }
3574 if (in1->outcnt() == 0) {
3575 in1->disconnect_inputs(this);
3576 }
3577 break;
3578 }
3579
3580 case Op_Proj: {
3581 if (OptimizeStringConcat || IncrementalInline) {
3582 ProjNode* proj = n->as_Proj();
3583 if (proj->_is_io_use) {
3584 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3585 // Separate projections were used for the exception path which
3586 // are normally removed by a late inline. If it wasn't inlined
3587 // then they will hang around and should just be replaced with
3588 // the original one. Merge them.
3589 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3590 if (non_io_proj != nullptr) {
3591 proj->subsume_by(non_io_proj , this);
3592 }
3593 }
3594 }
3595 break;
3596 }
3597
3598 case Op_Phi:
3599 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3600 // The EncodeP optimization may create Phi with the same edges
3601 // for all paths. It is not handled well by Register Allocator.
3602 Node* unique_in = n->in(1);
3603 assert(unique_in != nullptr, "");
3604 uint cnt = n->req();
3605 for (uint i = 2; i < cnt; i++) {
3606 Node* m = n->in(i);
3607 assert(m != nullptr, "");
3608 if (unique_in != m)
3609 unique_in = nullptr;
3610 }
3611 if (unique_in != nullptr) {
3612 n->subsume_by(unique_in, this);
3613 }
3614 }
3615 break;
3616
3617 #endif
3618
3619 #ifdef ASSERT
3620 case Op_CastII:
3621 // Verify that all range check dependent CastII nodes were removed.
3622 if (n->isa_CastII()->has_range_check()) {
3623 n->dump(3);
3624 assert(false, "Range check dependent CastII node was not removed");
3625 }
3626 break;
3627 #endif
3628
3629 case Op_ModI:
3630 if (UseDivMod) {
3631 // Check if a%b and a/b both exist
3632 Node* d = n->find_similar(Op_DivI);
3633 if (d) {
3634 // Replace them with a fused divmod if supported
3635 if (Matcher::has_match_rule(Op_DivModI)) {
3636 DivModINode* divmod = DivModINode::make(n);
3637 d->subsume_by(divmod->div_proj(), this);
3638 n->subsume_by(divmod->mod_proj(), this);
3639 } else {
3640 // replace a%b with a-((a/b)*b)
3641 Node* mult = new MulINode(d, d->in(2));
3642 Node* sub = new SubINode(d->in(1), mult);
3643 n->subsume_by(sub, this);
3644 }
3645 }
3646 }
3647 break;
3648
3649 case Op_ModL:
3650 if (UseDivMod) {
3651 // Check if a%b and a/b both exist
3652 Node* d = n->find_similar(Op_DivL);
3653 if (d) {
3654 // Replace them with a fused divmod if supported
3655 if (Matcher::has_match_rule(Op_DivModL)) {
3656 DivModLNode* divmod = DivModLNode::make(n);
3657 d->subsume_by(divmod->div_proj(), this);
3658 n->subsume_by(divmod->mod_proj(), this);
3659 } else {
3660 // replace a%b with a-((a/b)*b)
3661 Node* mult = new MulLNode(d, d->in(2));
3662 Node* sub = new SubLNode(d->in(1), mult);
3663 n->subsume_by(sub, this);
3664 }
3665 }
3666 }
3667 break;
3668
3669 case Op_UModI:
3670 if (UseDivMod) {
3671 // Check if a%b and a/b both exist
3672 Node* d = n->find_similar(Op_UDivI);
3673 if (d) {
3674 // Replace them with a fused unsigned divmod if supported
3675 if (Matcher::has_match_rule(Op_UDivModI)) {
3676 UDivModINode* divmod = UDivModINode::make(n);
3677 d->subsume_by(divmod->div_proj(), this);
3678 n->subsume_by(divmod->mod_proj(), this);
3679 } else {
3680 // replace a%b with a-((a/b)*b)
3681 Node* mult = new MulINode(d, d->in(2));
3682 Node* sub = new SubINode(d->in(1), mult);
3683 n->subsume_by(sub, this);
3684 }
3685 }
3686 }
3687 break;
3688
3689 case Op_UModL:
3690 if (UseDivMod) {
3691 // Check if a%b and a/b both exist
3692 Node* d = n->find_similar(Op_UDivL);
3693 if (d) {
3694 // Replace them with a fused unsigned divmod if supported
3695 if (Matcher::has_match_rule(Op_UDivModL)) {
3696 UDivModLNode* divmod = UDivModLNode::make(n);
3697 d->subsume_by(divmod->div_proj(), this);
3698 n->subsume_by(divmod->mod_proj(), this);
3699 } else {
3700 // replace a%b with a-((a/b)*b)
3701 Node* mult = new MulLNode(d, d->in(2));
3702 Node* sub = new SubLNode(d->in(1), mult);
3703 n->subsume_by(sub, this);
3704 }
3705 }
3706 }
3707 break;
3708
3709 case Op_LoadVector:
3710 case Op_StoreVector:
3711 #ifdef ASSERT
3712 // Add VerifyVectorAlignment node between adr and load / store.
3713 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3714 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3715 n->as_StoreVector()->must_verify_alignment();
3716 if (must_verify_alignment) {
3717 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3718 n->as_StoreVector()->memory_size();
3719 // The memory access should be aligned to the vector width in bytes.
3720 // However, the underlying array is possibly less well aligned, but at least
3721 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3722 // a loop we can expect at least the following alignment:
3723 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3724 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3725 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3726 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3727 jlong mask = guaranteed_alignment - 1;
3728 Node* mask_con = ConLNode::make(mask);
3729 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3730 n->set_req(MemNode::Address, va);
3731 }
3732 }
3733 #endif
3734 break;
3735
3736 case Op_LoadVectorGather:
3737 case Op_StoreVectorScatter:
3738 case Op_LoadVectorGatherMasked:
3739 case Op_StoreVectorScatterMasked:
3740 case Op_VectorCmpMasked:
3741 case Op_VectorMaskGen:
3742 case Op_LoadVectorMasked:
3743 case Op_StoreVectorMasked:
3744 break;
3745
3746 case Op_AddReductionVI:
3747 case Op_AddReductionVL:
3748 case Op_AddReductionVF:
3749 case Op_AddReductionVD:
3750 case Op_MulReductionVI:
3751 case Op_MulReductionVL:
3752 case Op_MulReductionVF:
3753 case Op_MulReductionVD:
3754 case Op_MinReductionV:
3755 case Op_MaxReductionV:
3756 case Op_AndReductionV:
3757 case Op_OrReductionV:
3758 case Op_XorReductionV:
3759 break;
3760
3761 case Op_PackB:
3762 case Op_PackS:
3763 case Op_PackI:
3764 case Op_PackF:
3765 case Op_PackL:
3766 case Op_PackD:
3767 if (n->req()-1 > 2) {
3768 // Replace many operand PackNodes with a binary tree for matching
3769 PackNode* p = (PackNode*) n;
3770 Node* btp = p->binary_tree_pack(1, n->req());
3771 n->subsume_by(btp, this);
3772 }
3773 break;
3774 case Op_Loop:
3775 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3776 case Op_CountedLoop:
3777 case Op_LongCountedLoop:
3778 case Op_OuterStripMinedLoop:
3779 if (n->as_Loop()->is_inner_loop()) {
3780 frc.inc_inner_loop_count();
3781 }
3782 n->as_Loop()->verify_strip_mined(0);
3783 break;
3784 case Op_LShiftI:
3785 case Op_RShiftI:
3786 case Op_URShiftI:
3787 case Op_LShiftL:
3788 case Op_RShiftL:
3789 case Op_URShiftL:
3790 if (Matcher::need_masked_shift_count) {
3791 // The cpu's shift instructions don't restrict the count to the
3792 // lower 5/6 bits. We need to do the masking ourselves.
3793 Node* in2 = n->in(2);
3794 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3795 const TypeInt* t = in2->find_int_type();
3796 if (t != nullptr && t->is_con()) {
3797 juint shift = t->get_con();
3798 if (shift > mask) { // Unsigned cmp
3799 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3800 }
3801 } else {
3802 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3803 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3804 n->set_req(2, shift);
3805 }
3806 }
3807 if (in2->outcnt() == 0) { // Remove dead node
3808 in2->disconnect_inputs(this);
3809 }
3810 }
3811 break;
3812 case Op_MemBarStoreStore:
3813 case Op_MemBarRelease:
3814 // Break the link with AllocateNode: it is no longer useful and
3815 // confuses register allocation.
3816 if (n->req() > MemBarNode::Precedent) {
3817 n->set_req(MemBarNode::Precedent, top());
3818 }
3819 break;
3820 case Op_MemBarAcquire: {
3821 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3822 // At parse time, the trailing MemBarAcquire for a volatile load
3823 // is created with an edge to the load. After optimizations,
3824 // that input may be a chain of Phis. If those phis have no
3825 // other use, then the MemBarAcquire keeps them alive and
3826 // register allocation can be confused.
3827 dead_nodes.push(n->in(MemBarNode::Precedent));
3828 n->set_req(MemBarNode::Precedent, top());
3829 }
3830 break;
3831 }
3832 case Op_Blackhole:
3833 break;
3834 case Op_RangeCheck: {
3835 RangeCheckNode* rc = n->as_RangeCheck();
3836 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3837 n->subsume_by(iff, this);
3838 frc._tests.push(iff);
3839 break;
3840 }
3841 case Op_ConvI2L: {
3842 if (!Matcher::convi2l_type_required) {
3843 // Code generation on some platforms doesn't need accurate
3844 // ConvI2L types. Widening the type can help remove redundant
3845 // address computations.
3846 n->as_Type()->set_type(TypeLong::INT);
3847 ResourceMark rm;
3848 Unique_Node_List wq;
3849 wq.push(n);
3850 for (uint next = 0; next < wq.size(); next++) {
3851 Node *m = wq.at(next);
3852
3853 for(;;) {
3854 // Loop over all nodes with identical inputs edges as m
3855 Node* k = m->find_similar(m->Opcode());
3856 if (k == nullptr) {
3857 break;
3858 }
3859 // Push their uses so we get a chance to remove node made
3860 // redundant
3861 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3862 Node* u = k->fast_out(i);
3863 if (u->Opcode() == Op_LShiftL ||
3864 u->Opcode() == Op_AddL ||
3865 u->Opcode() == Op_SubL ||
3866 u->Opcode() == Op_AddP) {
3867 wq.push(u);
3868 }
3869 }
3870 // Replace all nodes with identical edges as m with m
3871 k->subsume_by(m, this);
3872 }
3873 }
3874 }
3875 break;
3876 }
3877 case Op_CmpUL: {
3878 if (!Matcher::has_match_rule(Op_CmpUL)) {
3879 // No support for unsigned long comparisons
3880 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3881 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3882 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3883 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3884 Node* andl = new AndLNode(orl, remove_sign_mask);
3885 Node* cmp = new CmpLNode(andl, n->in(2));
3886 n->subsume_by(cmp, this);
3887 }
3888 break;
3889 }
3890 default:
3891 assert(!n->is_Call(), "");
3892 assert(!n->is_Mem(), "");
3893 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3894 break;
3895 }
3896 }
3897
3898 //------------------------------final_graph_reshaping_walk---------------------
3899 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3900 // requires that the walk visits a node's inputs before visiting the node.
3901 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3902 Unique_Node_List sfpt;
3903
3904 frc._visited.set(root->_idx); // first, mark node as visited
3905 uint cnt = root->req();
3906 Node *n = root;
3907 uint i = 0;
3908 while (true) {
3909 if (i < cnt) {
3910 // Place all non-visited non-null inputs onto stack
3911 Node* m = n->in(i);
3912 ++i;
3913 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3914 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3915 // compute worst case interpreter size in case of a deoptimization
3916 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3917
3918 sfpt.push(m);
3919 }
3920 cnt = m->req();
3921 nstack.push(n, i); // put on stack parent and next input's index
3922 n = m;
3923 i = 0;
3924 }
3925 } else {
3926 // Now do post-visit work
3927 final_graph_reshaping_impl(n, frc, dead_nodes);
3928 if (nstack.is_empty())
3929 break; // finished
3930 n = nstack.node(); // Get node from stack
3931 cnt = n->req();
3932 i = nstack.index();
3933 nstack.pop(); // Shift to the next node on stack
3934 }
3935 }
3936
3937 // Skip next transformation if compressed oops are not used.
3938 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3939 (!UseCompressedOops && !UseCompressedClassPointers))
3940 return;
3941
3942 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3943 // It could be done for an uncommon traps or any safepoints/calls
3944 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3945 while (sfpt.size() > 0) {
3946 n = sfpt.pop();
3947 JVMState *jvms = n->as_SafePoint()->jvms();
3948 assert(jvms != nullptr, "sanity");
3949 int start = jvms->debug_start();
3950 int end = n->req();
3951 bool is_uncommon = (n->is_CallStaticJava() &&
3952 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3953 for (int j = start; j < end; j++) {
3954 Node* in = n->in(j);
3955 if (in->is_DecodeNarrowPtr()) {
3956 bool safe_to_skip = true;
3957 if (!is_uncommon ) {
3958 // Is it safe to skip?
3959 for (uint i = 0; i < in->outcnt(); i++) {
3960 Node* u = in->raw_out(i);
3961 if (!u->is_SafePoint() ||
3962 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3963 safe_to_skip = false;
3964 }
3965 }
3966 }
3967 if (safe_to_skip) {
3968 n->set_req(j, in->in(1));
3969 }
3970 if (in->outcnt() == 0) {
3971 in->disconnect_inputs(this);
3972 }
3973 }
3974 }
3975 }
3976 }
3977
3978 //------------------------------final_graph_reshaping--------------------------
3979 // Final Graph Reshaping.
3980 //
3981 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3982 // and not commoned up and forced early. Must come after regular
3983 // optimizations to avoid GVN undoing the cloning. Clone constant
3984 // inputs to Loop Phis; these will be split by the allocator anyways.
3985 // Remove Opaque nodes.
3986 // (2) Move last-uses by commutative operations to the left input to encourage
3987 // Intel update-in-place two-address operations and better register usage
3988 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3989 // calls canonicalizing them back.
3990 // (3) Count the number of double-precision FP ops, single-precision FP ops
3991 // and call sites. On Intel, we can get correct rounding either by
3992 // forcing singles to memory (requires extra stores and loads after each
3993 // FP bytecode) or we can set a rounding mode bit (requires setting and
3994 // clearing the mode bit around call sites). The mode bit is only used
3995 // if the relative frequency of single FP ops to calls is low enough.
3996 // This is a key transform for SPEC mpeg_audio.
3997 // (4) Detect infinite loops; blobs of code reachable from above but not
3998 // below. Several of the Code_Gen algorithms fail on such code shapes,
3999 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4000 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4001 // Detection is by looking for IfNodes where only 1 projection is
4002 // reachable from below or CatchNodes missing some targets.
4003 // (5) Assert for insane oop offsets in debug mode.
4004
4005 bool Compile::final_graph_reshaping() {
4006 // an infinite loop may have been eliminated by the optimizer,
4007 // in which case the graph will be empty.
4008 if (root()->req() == 1) {
4009 // Do not compile method that is only a trivial infinite loop,
4010 // since the content of the loop may have been eliminated.
4011 record_method_not_compilable("trivial infinite loop");
4012 return true;
4013 }
4014
4015 // Expensive nodes have their control input set to prevent the GVN
4016 // from freely commoning them. There's no GVN beyond this point so
4017 // no need to keep the control input. We want the expensive nodes to
4018 // be freely moved to the least frequent code path by gcm.
4019 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4020 for (int i = 0; i < expensive_count(); i++) {
4021 _expensive_nodes.at(i)->set_req(0, nullptr);
4022 }
4023
4024 Final_Reshape_Counts frc;
4025
4026 // Visit everybody reachable!
4027 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4028 Node_Stack nstack(live_nodes() >> 1);
4029 Unique_Node_List dead_nodes;
4030 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4031
4032 // Check for unreachable (from below) code (i.e., infinite loops).
4033 for( uint i = 0; i < frc._tests.size(); i++ ) {
4034 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4035 // Get number of CFG targets.
4036 // Note that PCTables include exception targets after calls.
4037 uint required_outcnt = n->required_outcnt();
4038 if (n->outcnt() != required_outcnt) {
4039 // Check for a few special cases. Rethrow Nodes never take the
4040 // 'fall-thru' path, so expected kids is 1 less.
4041 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4042 if (n->in(0)->in(0)->is_Call()) {
4043 CallNode* call = n->in(0)->in(0)->as_Call();
4044 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4045 required_outcnt--; // Rethrow always has 1 less kid
4046 } else if (call->req() > TypeFunc::Parms &&
4047 call->is_CallDynamicJava()) {
4048 // Check for null receiver. In such case, the optimizer has
4049 // detected that the virtual call will always result in a null
4050 // pointer exception. The fall-through projection of this CatchNode
4051 // will not be populated.
4052 Node* arg0 = call->in(TypeFunc::Parms);
4053 if (arg0->is_Type() &&
4054 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4055 required_outcnt--;
4056 }
4057 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4058 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4059 // Check for illegal array length. In such case, the optimizer has
4060 // detected that the allocation attempt will always result in an
4061 // exception. There is no fall-through projection of this CatchNode .
4062 assert(call->is_CallStaticJava(), "static call expected");
4063 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4064 uint valid_length_test_input = call->req() - 1;
4065 Node* valid_length_test = call->in(valid_length_test_input);
4066 call->del_req(valid_length_test_input);
4067 if (valid_length_test->find_int_con(1) == 0) {
4068 required_outcnt--;
4069 }
4070 dead_nodes.push(valid_length_test);
4071 assert(n->outcnt() == required_outcnt, "malformed control flow");
4072 continue;
4073 }
4074 }
4075 }
4076
4077 // Recheck with a better notion of 'required_outcnt'
4078 if (n->outcnt() != required_outcnt) {
4079 record_method_not_compilable("malformed control flow");
4080 return true; // Not all targets reachable!
4081 }
4082 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4083 CallNode* call = n->in(0)->in(0)->as_Call();
4084 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4085 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4086 assert(call->is_CallStaticJava(), "static call expected");
4087 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4088 uint valid_length_test_input = call->req() - 1;
4089 dead_nodes.push(call->in(valid_length_test_input));
4090 call->del_req(valid_length_test_input); // valid length test useless now
4091 }
4092 }
4093 // Check that I actually visited all kids. Unreached kids
4094 // must be infinite loops.
4095 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4096 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4097 record_method_not_compilable("infinite loop");
4098 return true; // Found unvisited kid; must be unreach
4099 }
4100
4101 // Here so verification code in final_graph_reshaping_walk()
4102 // always see an OuterStripMinedLoopEnd
4103 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4104 IfNode* init_iff = n->as_If();
4105 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4106 n->subsume_by(iff, this);
4107 }
4108 }
4109
4110 while (dead_nodes.size() > 0) {
4111 Node* m = dead_nodes.pop();
4112 if (m->outcnt() == 0 && m != top()) {
4113 for (uint j = 0; j < m->req(); j++) {
4114 Node* in = m->in(j);
4115 if (in != nullptr) {
4116 dead_nodes.push(in);
4117 }
4118 }
4119 m->disconnect_inputs(this);
4120 }
4121 }
4122
4123 #ifdef IA32
4124 // If original bytecodes contained a mixture of floats and doubles
4125 // check if the optimizer has made it homogeneous, item (3).
4126 if (UseSSE == 0 &&
4127 frc.get_float_count() > 32 &&
4128 frc.get_double_count() == 0 &&
4129 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4130 set_24_bit_selection_and_mode(false, true);
4131 }
4132 #endif // IA32
4133
4134 set_java_calls(frc.get_java_call_count());
4135 set_inner_loops(frc.get_inner_loop_count());
4136
4137 // No infinite loops, no reason to bail out.
4138 return false;
4139 }
4140
4141 //-----------------------------too_many_traps----------------------------------
4142 // Report if there are too many traps at the current method and bci.
4143 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4144 bool Compile::too_many_traps(ciMethod* method,
4145 int bci,
4146 Deoptimization::DeoptReason reason) {
4147 ciMethodData* md = method->method_data();
4148 if (md->is_empty()) {
4149 // Assume the trap has not occurred, or that it occurred only
4150 // because of a transient condition during start-up in the interpreter.
4151 return false;
4152 }
4153 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4154 if (md->has_trap_at(bci, m, reason) != 0) {
4155 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4156 // Also, if there are multiple reasons, or if there is no per-BCI record,
4157 // assume the worst.
4158 if (log())
4159 log()->elem("observe trap='%s' count='%d'",
4160 Deoptimization::trap_reason_name(reason),
4161 md->trap_count(reason));
4162 return true;
4163 } else {
4164 // Ignore method/bci and see if there have been too many globally.
4165 return too_many_traps(reason, md);
4166 }
4167 }
4168
4169 // Less-accurate variant which does not require a method and bci.
4170 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4171 ciMethodData* logmd) {
4172 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4173 // Too many traps globally.
4174 // Note that we use cumulative trap_count, not just md->trap_count.
4175 if (log()) {
4176 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4177 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4178 Deoptimization::trap_reason_name(reason),
4179 mcount, trap_count(reason));
4180 }
4181 return true;
4182 } else {
4183 // The coast is clear.
4184 return false;
4185 }
4186 }
4187
4188 //--------------------------too_many_recompiles--------------------------------
4189 // Report if there are too many recompiles at the current method and bci.
4190 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4191 // Is not eager to return true, since this will cause the compiler to use
4192 // Action_none for a trap point, to avoid too many recompilations.
4193 bool Compile::too_many_recompiles(ciMethod* method,
4194 int bci,
4195 Deoptimization::DeoptReason reason) {
4196 ciMethodData* md = method->method_data();
4197 if (md->is_empty()) {
4198 // Assume the trap has not occurred, or that it occurred only
4199 // because of a transient condition during start-up in the interpreter.
4200 return false;
4201 }
4202 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4203 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4204 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4205 Deoptimization::DeoptReason per_bc_reason
4206 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4207 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4208 if ((per_bc_reason == Deoptimization::Reason_none
4209 || md->has_trap_at(bci, m, reason) != 0)
4210 // The trap frequency measure we care about is the recompile count:
4211 && md->trap_recompiled_at(bci, m)
4212 && md->overflow_recompile_count() >= bc_cutoff) {
4213 // Do not emit a trap here if it has already caused recompilations.
4214 // Also, if there are multiple reasons, or if there is no per-BCI record,
4215 // assume the worst.
4216 if (log())
4217 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4218 Deoptimization::trap_reason_name(reason),
4219 md->trap_count(reason),
4220 md->overflow_recompile_count());
4221 return true;
4222 } else if (trap_count(reason) != 0
4223 && decompile_count() >= m_cutoff) {
4224 // Too many recompiles globally, and we have seen this sort of trap.
4225 // Use cumulative decompile_count, not just md->decompile_count.
4226 if (log())
4227 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4228 Deoptimization::trap_reason_name(reason),
4229 md->trap_count(reason), trap_count(reason),
4230 md->decompile_count(), decompile_count());
4231 return true;
4232 } else {
4233 // The coast is clear.
4234 return false;
4235 }
4236 }
4237
4238 // Compute when not to trap. Used by matching trap based nodes and
4239 // NullCheck optimization.
4240 void Compile::set_allowed_deopt_reasons() {
4241 _allowed_reasons = 0;
4242 if (is_method_compilation()) {
4243 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4244 assert(rs < BitsPerInt, "recode bit map");
4245 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4246 _allowed_reasons |= nth_bit(rs);
4247 }
4248 }
4249 }
4250 }
4251
4252 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4253 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4254 }
4255
4256 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4257 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4258 }
4259
4260 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4261 if (holder->is_initialized()) {
4262 return false;
4263 }
4264 if (holder->is_being_initialized()) {
4265 if (accessing_method->holder() == holder) {
4266 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4267 // <init>, or a static method. In all those cases, there was an initialization
4268 // barrier on the holder klass passed.
4269 if (accessing_method->is_static_initializer() ||
4270 accessing_method->is_object_initializer() ||
4271 accessing_method->is_static()) {
4272 return false;
4273 }
4274 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4275 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4276 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4277 // child class can become fully initialized while its parent class is still being initialized.
4278 if (accessing_method->is_static_initializer()) {
4279 return false;
4280 }
4281 }
4282 ciMethod* root = method(); // the root method of compilation
4283 if (root != accessing_method) {
4284 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4285 }
4286 }
4287 return true;
4288 }
4289
4290 #ifndef PRODUCT
4291 //------------------------------verify_bidirectional_edges---------------------
4292 // For each input edge to a node (ie - for each Use-Def edge), verify that
4293 // there is a corresponding Def-Use edge.
4294 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4295 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4296 uint stack_size = live_nodes() >> 4;
4297 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4298 nstack.push(_root);
4299
4300 while (nstack.size() > 0) {
4301 Node* n = nstack.pop();
4302 if (visited.member(n)) {
4303 continue;
4304 }
4305 visited.push(n);
4306
4307 // Walk over all input edges, checking for correspondence
4308 uint length = n->len();
4309 for (uint i = 0; i < length; i++) {
4310 Node* in = n->in(i);
4311 if (in != nullptr && !visited.member(in)) {
4312 nstack.push(in); // Put it on stack
4313 }
4314 if (in != nullptr && !in->is_top()) {
4315 // Count instances of `next`
4316 int cnt = 0;
4317 for (uint idx = 0; idx < in->_outcnt; idx++) {
4318 if (in->_out[idx] == n) {
4319 cnt++;
4320 }
4321 }
4322 assert(cnt > 0, "Failed to find Def-Use edge.");
4323 // Check for duplicate edges
4324 // walk the input array downcounting the input edges to n
4325 for (uint j = 0; j < length; j++) {
4326 if (n->in(j) == in) {
4327 cnt--;
4328 }
4329 }
4330 assert(cnt == 0, "Mismatched edge count.");
4331 } else if (in == nullptr) {
4332 assert(i == 0 || i >= n->req() ||
4333 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4334 (n->is_Unlock() && i == (n->req() - 1)) ||
4335 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4336 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4337 } else {
4338 assert(in->is_top(), "sanity");
4339 // Nothing to check.
4340 }
4341 }
4342 }
4343 }
4344
4345 //------------------------------verify_graph_edges---------------------------
4346 // Walk the Graph and verify that there is a one-to-one correspondence
4347 // between Use-Def edges and Def-Use edges in the graph.
4348 void Compile::verify_graph_edges(bool no_dead_code) {
4349 if (VerifyGraphEdges) {
4350 Unique_Node_List visited;
4351
4352 // Call graph walk to check edges
4353 verify_bidirectional_edges(visited);
4354 if (no_dead_code) {
4355 // Now make sure that no visited node is used by an unvisited node.
4356 bool dead_nodes = false;
4357 Unique_Node_List checked;
4358 while (visited.size() > 0) {
4359 Node* n = visited.pop();
4360 checked.push(n);
4361 for (uint i = 0; i < n->outcnt(); i++) {
4362 Node* use = n->raw_out(i);
4363 if (checked.member(use)) continue; // already checked
4364 if (visited.member(use)) continue; // already in the graph
4365 if (use->is_Con()) continue; // a dead ConNode is OK
4366 // At this point, we have found a dead node which is DU-reachable.
4367 if (!dead_nodes) {
4368 tty->print_cr("*** Dead nodes reachable via DU edges:");
4369 dead_nodes = true;
4370 }
4371 use->dump(2);
4372 tty->print_cr("---");
4373 checked.push(use); // No repeats; pretend it is now checked.
4374 }
4375 }
4376 assert(!dead_nodes, "using nodes must be reachable from root");
4377 }
4378 }
4379 }
4380 #endif
4381
4382 // The Compile object keeps track of failure reasons separately from the ciEnv.
4383 // This is required because there is not quite a 1-1 relation between the
4384 // ciEnv and its compilation task and the Compile object. Note that one
4385 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4386 // to backtrack and retry without subsuming loads. Other than this backtracking
4387 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4388 // by the logic in C2Compiler.
4389 void Compile::record_failure(const char* reason) {
4390 if (log() != nullptr) {
4391 log()->elem("failure reason='%s' phase='compile'", reason);
4392 }
4393 if (_failure_reason.get() == nullptr) {
4394 // Record the first failure reason.
4395 _failure_reason.set(reason);
4396 if (CaptureBailoutInformation) {
4397 _first_failure_details = new CompilationFailureInfo(reason);
4398 }
4399 }
4400
4401 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4402 C->print_method(PHASE_FAILURE, 1);
4403 }
4404 _root = nullptr; // flush the graph, too
4405 }
4406
4407 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4408 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4409 _compile(Compile::current()),
4410 _log(nullptr),
4411 _phase_name(name),
4412 _dolog(CITimeVerbose)
4413 {
4414 assert(_compile != nullptr, "sanity check");
4415 if (_dolog) {
4416 _log = _compile->log();
4417 }
4418 if (_log != nullptr) {
4419 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4420 _log->stamp();
4421 _log->end_head();
4422 }
4423 }
4424
4425 Compile::TracePhase::~TracePhase() {
4426 if (_compile->failing()) return;
4427 #ifdef ASSERT
4428 if (PrintIdealNodeCount) {
4429 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4430 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4431 }
4432
4433 if (VerifyIdealNodeCount) {
4434 _compile->print_missing_nodes();
4435 }
4436 #endif
4437
4438 if (_log != nullptr) {
4439 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4440 }
4441 }
4442
4443 //----------------------------static_subtype_check-----------------------------
4444 // Shortcut important common cases when superklass is exact:
4445 // (0) superklass is java.lang.Object (can occur in reflective code)
4446 // (1) subklass is already limited to a subtype of superklass => always ok
4447 // (2) subklass does not overlap with superklass => always fail
4448 // (3) superklass has NO subtypes and we can check with a simple compare.
4449 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4450 if (skip) {
4451 return SSC_full_test; // Let caller generate the general case.
4452 }
4453
4454 if (subk->is_java_subtype_of(superk)) {
4455 return SSC_always_true; // (0) and (1) this test cannot fail
4456 }
4457
4458 if (!subk->maybe_java_subtype_of(superk)) {
4459 return SSC_always_false; // (2) true path dead; no dynamic test needed
4460 }
4461
4462 const Type* superelem = superk;
4463 if (superk->isa_aryklassptr()) {
4464 int ignored;
4465 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4466 }
4467
4468 if (superelem->isa_instklassptr()) {
4469 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4470 if (!ik->has_subklass()) {
4471 if (!ik->is_final()) {
4472 // Add a dependency if there is a chance of a later subclass.
4473 dependencies()->assert_leaf_type(ik);
4474 }
4475 if (!superk->maybe_java_subtype_of(subk)) {
4476 return SSC_always_false;
4477 }
4478 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4479 }
4480 } else {
4481 // A primitive array type has no subtypes.
4482 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4483 }
4484
4485 return SSC_full_test;
4486 }
4487
4488 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4489 #ifdef _LP64
4490 // The scaled index operand to AddP must be a clean 64-bit value.
4491 // Java allows a 32-bit int to be incremented to a negative
4492 // value, which appears in a 64-bit register as a large
4493 // positive number. Using that large positive number as an
4494 // operand in pointer arithmetic has bad consequences.
4495 // On the other hand, 32-bit overflow is rare, and the possibility
4496 // can often be excluded, if we annotate the ConvI2L node with
4497 // a type assertion that its value is known to be a small positive
4498 // number. (The prior range check has ensured this.)
4499 // This assertion is used by ConvI2LNode::Ideal.
4500 int index_max = max_jint - 1; // array size is max_jint, index is one less
4501 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4502 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4503 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4504 #endif
4505 return idx;
4506 }
4507
4508 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4509 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4510 if (ctrl != nullptr) {
4511 // Express control dependency by a CastII node with a narrow type.
4512 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4513 // node from floating above the range check during loop optimizations. Otherwise, the
4514 // ConvI2L node may be eliminated independently of the range check, causing the data path
4515 // to become TOP while the control path is still there (although it's unreachable).
4516 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4517 value = phase->transform(value);
4518 }
4519 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4520 return phase->transform(new ConvI2LNode(value, ltype));
4521 }
4522
4523 // The message about the current inlining is accumulated in
4524 // _print_inlining_stream and transferred into the _print_inlining_list
4525 // once we know whether inlining succeeds or not. For regular
4526 // inlining, messages are appended to the buffer pointed by
4527 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4528 // a new buffer is added after _print_inlining_idx in the list. This
4529 // way we can update the inlining message for late inlining call site
4530 // when the inlining is attempted again.
4531 void Compile::print_inlining_init() {
4532 if (print_inlining() || print_intrinsics()) {
4533 // print_inlining_init is actually called several times.
4534 print_inlining_reset();
4535 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4536 }
4537 }
4538
4539 void Compile::print_inlining_reinit() {
4540 if (print_inlining() || print_intrinsics()) {
4541 print_inlining_reset();
4542 }
4543 }
4544
4545 void Compile::print_inlining_reset() {
4546 _print_inlining_stream->reset();
4547 }
4548
4549 void Compile::print_inlining_commit() {
4550 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4551 // Transfer the message from _print_inlining_stream to the current
4552 // _print_inlining_list buffer and clear _print_inlining_stream.
4553 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4554 print_inlining_reset();
4555 }
4556
4557 void Compile::print_inlining_push() {
4558 // Add new buffer to the _print_inlining_list at current position
4559 _print_inlining_idx++;
4560 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4561 }
4562
4563 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4564 return _print_inlining_list->at(_print_inlining_idx);
4565 }
4566
4567 void Compile::print_inlining_update(CallGenerator* cg) {
4568 if (print_inlining() || print_intrinsics()) {
4569 if (cg->is_late_inline()) {
4570 if (print_inlining_current()->cg() != cg &&
4571 (print_inlining_current()->cg() != nullptr ||
4572 print_inlining_current()->ss()->size() != 0)) {
4573 print_inlining_push();
4574 }
4575 print_inlining_commit();
4576 print_inlining_current()->set_cg(cg);
4577 } else {
4578 if (print_inlining_current()->cg() != nullptr) {
4579 print_inlining_push();
4580 }
4581 print_inlining_commit();
4582 }
4583 }
4584 }
4585
4586 void Compile::print_inlining_move_to(CallGenerator* cg) {
4587 // We resume inlining at a late inlining call site. Locate the
4588 // corresponding inlining buffer so that we can update it.
4589 if (print_inlining() || print_intrinsics()) {
4590 for (int i = 0; i < _print_inlining_list->length(); i++) {
4591 if (_print_inlining_list->at(i)->cg() == cg) {
4592 _print_inlining_idx = i;
4593 return;
4594 }
4595 }
4596 ShouldNotReachHere();
4597 }
4598 }
4599
4600 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4601 if (print_inlining() || print_intrinsics()) {
4602 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4603 assert(print_inlining_current()->cg() == cg, "wrong entry");
4604 // replace message with new message
4605 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4606 print_inlining_commit();
4607 print_inlining_current()->set_cg(cg);
4608 }
4609 }
4610
4611 void Compile::print_inlining_assert_ready() {
4612 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4613 }
4614
4615 void Compile::process_print_inlining() {
4616 assert(_late_inlines.length() == 0, "not drained yet");
4617 if (print_inlining() || print_intrinsics()) {
4618 ResourceMark rm;
4619 stringStream ss;
4620 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4621 for (int i = 0; i < _print_inlining_list->length(); i++) {
4622 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4623 ss.print("%s", pib->ss()->freeze());
4624 delete pib;
4625 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4626 }
4627 // Reset _print_inlining_list, it only contains destructed objects.
4628 // It is on the arena, so it will be freed when the arena is reset.
4629 _print_inlining_list = nullptr;
4630 // _print_inlining_stream won't be used anymore, either.
4631 print_inlining_reset();
4632 size_t end = ss.size();
4633 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4634 strncpy(_print_inlining_output, ss.freeze(), end+1);
4635 _print_inlining_output[end] = 0;
4636 }
4637 }
4638
4639 void Compile::dump_print_inlining() {
4640 if (_print_inlining_output != nullptr) {
4641 tty->print_raw(_print_inlining_output);
4642 }
4643 }
4644
4645 void Compile::log_late_inline(CallGenerator* cg) {
4646 if (log() != nullptr) {
4647 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4648 cg->unique_id());
4649 JVMState* p = cg->call_node()->jvms();
4650 while (p != nullptr) {
4651 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4652 p = p->caller();
4653 }
4654 log()->tail("late_inline");
4655 }
4656 }
4657
4658 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4659 log_late_inline(cg);
4660 if (log() != nullptr) {
4661 log()->inline_fail(msg);
4662 }
4663 }
4664
4665 void Compile::log_inline_id(CallGenerator* cg) {
4666 if (log() != nullptr) {
4667 // The LogCompilation tool needs a unique way to identify late
4668 // inline call sites. This id must be unique for this call site in
4669 // this compilation. Try to have it unique across compilations as
4670 // well because it can be convenient when grepping through the log
4671 // file.
4672 // Distinguish OSR compilations from others in case CICountOSR is
4673 // on.
4674 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4675 cg->set_unique_id(id);
4676 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4677 }
4678 }
4679
4680 void Compile::log_inline_failure(const char* msg) {
4681 if (C->log() != nullptr) {
4682 C->log()->inline_fail(msg);
4683 }
4684 }
4685
4686
4687 // Dump inlining replay data to the stream.
4688 // Don't change thread state and acquire any locks.
4689 void Compile::dump_inline_data(outputStream* out) {
4690 InlineTree* inl_tree = ilt();
4691 if (inl_tree != nullptr) {
4692 out->print(" inline %d", inl_tree->count());
4693 inl_tree->dump_replay_data(out);
4694 }
4695 }
4696
4697 void Compile::dump_inline_data_reduced(outputStream* out) {
4698 assert(ReplayReduce, "");
4699
4700 InlineTree* inl_tree = ilt();
4701 if (inl_tree == nullptr) {
4702 return;
4703 }
4704 // Enable iterative replay file reduction
4705 // Output "compile" lines for depth 1 subtrees,
4706 // simulating that those trees were compiled
4707 // instead of inlined.
4708 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4709 InlineTree* sub = inl_tree->subtrees().at(i);
4710 if (sub->inline_level() != 1) {
4711 continue;
4712 }
4713
4714 ciMethod* method = sub->method();
4715 int entry_bci = -1;
4716 int comp_level = env()->task()->comp_level();
4717 out->print("compile ");
4718 method->dump_name_as_ascii(out);
4719 out->print(" %d %d", entry_bci, comp_level);
4720 out->print(" inline %d", sub->count());
4721 sub->dump_replay_data(out, -1);
4722 out->cr();
4723 }
4724 }
4725
4726 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4727 if (n1->Opcode() < n2->Opcode()) return -1;
4728 else if (n1->Opcode() > n2->Opcode()) return 1;
4729
4730 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4731 for (uint i = 1; i < n1->req(); i++) {
4732 if (n1->in(i) < n2->in(i)) return -1;
4733 else if (n1->in(i) > n2->in(i)) return 1;
4734 }
4735
4736 return 0;
4737 }
4738
4739 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4740 Node* n1 = *n1p;
4741 Node* n2 = *n2p;
4742
4743 return cmp_expensive_nodes(n1, n2);
4744 }
4745
4746 void Compile::sort_expensive_nodes() {
4747 if (!expensive_nodes_sorted()) {
4748 _expensive_nodes.sort(cmp_expensive_nodes);
4749 }
4750 }
4751
4752 bool Compile::expensive_nodes_sorted() const {
4753 for (int i = 1; i < _expensive_nodes.length(); i++) {
4754 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4755 return false;
4756 }
4757 }
4758 return true;
4759 }
4760
4761 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4762 if (_expensive_nodes.length() == 0) {
4763 return false;
4764 }
4765
4766 assert(OptimizeExpensiveOps, "optimization off?");
4767
4768 // Take this opportunity to remove dead nodes from the list
4769 int j = 0;
4770 for (int i = 0; i < _expensive_nodes.length(); i++) {
4771 Node* n = _expensive_nodes.at(i);
4772 if (!n->is_unreachable(igvn)) {
4773 assert(n->is_expensive(), "should be expensive");
4774 _expensive_nodes.at_put(j, n);
4775 j++;
4776 }
4777 }
4778 _expensive_nodes.trunc_to(j);
4779
4780 // Then sort the list so that similar nodes are next to each other
4781 // and check for at least two nodes of identical kind with same data
4782 // inputs.
4783 sort_expensive_nodes();
4784
4785 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4786 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4787 return true;
4788 }
4789 }
4790
4791 return false;
4792 }
4793
4794 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4795 if (_expensive_nodes.length() == 0) {
4796 return;
4797 }
4798
4799 assert(OptimizeExpensiveOps, "optimization off?");
4800
4801 // Sort to bring similar nodes next to each other and clear the
4802 // control input of nodes for which there's only a single copy.
4803 sort_expensive_nodes();
4804
4805 int j = 0;
4806 int identical = 0;
4807 int i = 0;
4808 bool modified = false;
4809 for (; i < _expensive_nodes.length()-1; i++) {
4810 assert(j <= i, "can't write beyond current index");
4811 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4812 identical++;
4813 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4814 continue;
4815 }
4816 if (identical > 0) {
4817 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4818 identical = 0;
4819 } else {
4820 Node* n = _expensive_nodes.at(i);
4821 igvn.replace_input_of(n, 0, nullptr);
4822 igvn.hash_insert(n);
4823 modified = true;
4824 }
4825 }
4826 if (identical > 0) {
4827 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4828 } else if (_expensive_nodes.length() >= 1) {
4829 Node* n = _expensive_nodes.at(i);
4830 igvn.replace_input_of(n, 0, nullptr);
4831 igvn.hash_insert(n);
4832 modified = true;
4833 }
4834 _expensive_nodes.trunc_to(j);
4835 if (modified) {
4836 igvn.optimize();
4837 }
4838 }
4839
4840 void Compile::add_expensive_node(Node * n) {
4841 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4842 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4843 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4844 if (OptimizeExpensiveOps) {
4845 _expensive_nodes.append(n);
4846 } else {
4847 // Clear control input and let IGVN optimize expensive nodes if
4848 // OptimizeExpensiveOps is off.
4849 n->set_req(0, nullptr);
4850 }
4851 }
4852
4853 /**
4854 * Track coarsened Lock and Unlock nodes.
4855 */
4856
4857 class Lock_List : public Node_List {
4858 uint _origin_cnt;
4859 public:
4860 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4861 uint origin_cnt() const { return _origin_cnt; }
4862 };
4863
4864 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4865 int length = locks.length();
4866 if (length > 0) {
4867 // Have to keep this list until locks elimination during Macro nodes elimination.
4868 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4869 AbstractLockNode* alock = locks.at(0);
4870 BoxLockNode* box = alock->box_node()->as_BoxLock();
4871 for (int i = 0; i < length; i++) {
4872 AbstractLockNode* lock = locks.at(i);
4873 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4874 locks_list->push(lock);
4875 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4876 if (this_box != box) {
4877 // Locking regions (BoxLock) could be Unbalanced here:
4878 // - its coarsened locks were eliminated in earlier
4879 // macro nodes elimination followed by loop unroll
4880 // - it is OSR locking region (no Lock node)
4881 // Preserve Unbalanced status in such cases.
4882 if (!this_box->is_unbalanced()) {
4883 this_box->set_coarsened();
4884 }
4885 if (!box->is_unbalanced()) {
4886 box->set_coarsened();
4887 }
4888 }
4889 }
4890 _coarsened_locks.append(locks_list);
4891 }
4892 }
4893
4894 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4895 int count = coarsened_count();
4896 for (int i = 0; i < count; i++) {
4897 Node_List* locks_list = _coarsened_locks.at(i);
4898 for (uint j = 0; j < locks_list->size(); j++) {
4899 Node* lock = locks_list->at(j);
4900 assert(lock->is_AbstractLock(), "sanity");
4901 if (!useful.member(lock)) {
4902 locks_list->yank(lock);
4903 }
4904 }
4905 }
4906 }
4907
4908 void Compile::remove_coarsened_lock(Node* n) {
4909 if (n->is_AbstractLock()) {
4910 int count = coarsened_count();
4911 for (int i = 0; i < count; i++) {
4912 Node_List* locks_list = _coarsened_locks.at(i);
4913 locks_list->yank(n);
4914 }
4915 }
4916 }
4917
4918 bool Compile::coarsened_locks_consistent() {
4919 int count = coarsened_count();
4920 for (int i = 0; i < count; i++) {
4921 bool unbalanced = false;
4922 bool modified = false; // track locks kind modifications
4923 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4924 uint size = locks_list->size();
4925 if (size == 0) {
4926 unbalanced = false; // All locks were eliminated - good
4927 } else if (size != locks_list->origin_cnt()) {
4928 unbalanced = true; // Some locks were removed from list
4929 } else {
4930 for (uint j = 0; j < size; j++) {
4931 Node* lock = locks_list->at(j);
4932 // All nodes in group should have the same state (modified or not)
4933 if (!lock->as_AbstractLock()->is_coarsened()) {
4934 if (j == 0) {
4935 // first on list was modified, the rest should be too for consistency
4936 modified = true;
4937 } else if (!modified) {
4938 // this lock was modified but previous locks on the list were not
4939 unbalanced = true;
4940 break;
4941 }
4942 } else if (modified) {
4943 // previous locks on list were modified but not this lock
4944 unbalanced = true;
4945 break;
4946 }
4947 }
4948 }
4949 if (unbalanced) {
4950 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4951 #ifdef ASSERT
4952 if (PrintEliminateLocks) {
4953 tty->print_cr("=== unbalanced coarsened locks ===");
4954 for (uint l = 0; l < size; l++) {
4955 locks_list->at(l)->dump();
4956 }
4957 }
4958 #endif
4959 record_failure(C2Compiler::retry_no_locks_coarsening());
4960 return false;
4961 }
4962 }
4963 return true;
4964 }
4965
4966 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4967 // locks coarsening optimization removed Lock/Unlock nodes from them.
4968 // Such regions become unbalanced because coarsening only removes part
4969 // of Lock/Unlock nodes in region. As result we can't execute other
4970 // locks elimination optimizations which assume all code paths have
4971 // corresponding pair of Lock/Unlock nodes - they are balanced.
4972 void Compile::mark_unbalanced_boxes() const {
4973 int count = coarsened_count();
4974 for (int i = 0; i < count; i++) {
4975 Node_List* locks_list = _coarsened_locks.at(i);
4976 uint size = locks_list->size();
4977 if (size > 0) {
4978 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4979 BoxLockNode* box = alock->box_node()->as_BoxLock();
4980 if (alock->is_coarsened()) {
4981 // coarsened_locks_consistent(), which is called before this method, verifies
4982 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4983 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4984 for (uint j = 1; j < size; j++) {
4985 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4986 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4987 if (box != this_box) {
4988 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4989 box->set_unbalanced();
4990 this_box->set_unbalanced();
4991 }
4992 }
4993 }
4994 }
4995 }
4996 }
4997
4998 /**
4999 * Remove the speculative part of types and clean up the graph
5000 */
5001 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5002 if (UseTypeSpeculation) {
5003 Unique_Node_List worklist;
5004 worklist.push(root());
5005 int modified = 0;
5006 // Go over all type nodes that carry a speculative type, drop the
5007 // speculative part of the type and enqueue the node for an igvn
5008 // which may optimize it out.
5009 for (uint next = 0; next < worklist.size(); ++next) {
5010 Node *n = worklist.at(next);
5011 if (n->is_Type()) {
5012 TypeNode* tn = n->as_Type();
5013 const Type* t = tn->type();
5014 const Type* t_no_spec = t->remove_speculative();
5015 if (t_no_spec != t) {
5016 bool in_hash = igvn.hash_delete(n);
5017 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5018 tn->set_type(t_no_spec);
5019 igvn.hash_insert(n);
5020 igvn._worklist.push(n); // give it a chance to go away
5021 modified++;
5022 }
5023 }
5024 // Iterate over outs - endless loops is unreachable from below
5025 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5026 Node *m = n->fast_out(i);
5027 if (not_a_node(m)) {
5028 continue;
5029 }
5030 worklist.push(m);
5031 }
5032 }
5033 // Drop the speculative part of all types in the igvn's type table
5034 igvn.remove_speculative_types();
5035 if (modified > 0) {
5036 igvn.optimize();
5037 if (failing()) return;
5038 }
5039 #ifdef ASSERT
5040 // Verify that after the IGVN is over no speculative type has resurfaced
5041 worklist.clear();
5042 worklist.push(root());
5043 for (uint next = 0; next < worklist.size(); ++next) {
5044 Node *n = worklist.at(next);
5045 const Type* t = igvn.type_or_null(n);
5046 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5047 if (n->is_Type()) {
5048 t = n->as_Type()->type();
5049 assert(t == t->remove_speculative(), "no more speculative types");
5050 }
5051 // Iterate over outs - endless loops is unreachable from below
5052 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5053 Node *m = n->fast_out(i);
5054 if (not_a_node(m)) {
5055 continue;
5056 }
5057 worklist.push(m);
5058 }
5059 }
5060 igvn.check_no_speculative_types();
5061 #endif
5062 }
5063 }
5064
5065 // Auxiliary methods to support randomized stressing/fuzzing.
5066
5067 int Compile::random() {
5068 _stress_seed = os::next_random(_stress_seed);
5069 return static_cast<int>(_stress_seed);
5070 }
5071
5072 // This method can be called the arbitrary number of times, with current count
5073 // as the argument. The logic allows selecting a single candidate from the
5074 // running list of candidates as follows:
5075 // int count = 0;
5076 // Cand* selected = null;
5077 // while(cand = cand->next()) {
5078 // if (randomized_select(++count)) {
5079 // selected = cand;
5080 // }
5081 // }
5082 //
5083 // Including count equalizes the chances any candidate is "selected".
5084 // This is useful when we don't have the complete list of candidates to choose
5085 // from uniformly. In this case, we need to adjust the randomicity of the
5086 // selection, or else we will end up biasing the selection towards the latter
5087 // candidates.
5088 //
5089 // Quick back-envelope calculation shows that for the list of n candidates
5090 // the equal probability for the candidate to persist as "best" can be
5091 // achieved by replacing it with "next" k-th candidate with the probability
5092 // of 1/k. It can be easily shown that by the end of the run, the
5093 // probability for any candidate is converged to 1/n, thus giving the
5094 // uniform distribution among all the candidates.
5095 //
5096 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5097 #define RANDOMIZED_DOMAIN_POW 29
5098 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5099 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5100 bool Compile::randomized_select(int count) {
5101 assert(count > 0, "only positive");
5102 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5103 }
5104
5105 CloneMap& Compile::clone_map() { return _clone_map; }
5106 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5107
5108 void NodeCloneInfo::dump_on(outputStream* st) const {
5109 st->print(" {%d:%d} ", idx(), gen());
5110 }
5111
5112 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5113 uint64_t val = value(old->_idx);
5114 NodeCloneInfo cio(val);
5115 assert(val != 0, "old node should be in the map");
5116 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5117 insert(nnn->_idx, cin.get());
5118 #ifndef PRODUCT
5119 if (is_debug()) {
5120 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5121 }
5122 #endif
5123 }
5124
5125 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5126 NodeCloneInfo cio(value(old->_idx));
5127 if (cio.get() == 0) {
5128 cio.set(old->_idx, 0);
5129 insert(old->_idx, cio.get());
5130 #ifndef PRODUCT
5131 if (is_debug()) {
5132 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5133 }
5134 #endif
5135 }
5136 clone(old, nnn, gen);
5137 }
5138
5139 int CloneMap::max_gen() const {
5140 int g = 0;
5141 DictI di(_dict);
5142 for(; di.test(); ++di) {
5143 int t = gen(di._key);
5144 if (g < t) {
5145 g = t;
5146 #ifndef PRODUCT
5147 if (is_debug()) {
5148 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5149 }
5150 #endif
5151 }
5152 }
5153 return g;
5154 }
5155
5156 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5157 uint64_t val = value(key);
5158 if (val != 0) {
5159 NodeCloneInfo ni(val);
5160 ni.dump_on(st);
5161 }
5162 }
5163
5164 void Compile::shuffle_macro_nodes() {
5165 if (_macro_nodes.length() < 2) {
5166 return;
5167 }
5168 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5169 uint j = C->random() % (i + 1);
5170 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5171 }
5172 }
5173
5174 // Move Allocate nodes to the start of the list
5175 void Compile::sort_macro_nodes() {
5176 int count = macro_count();
5177 int allocates = 0;
5178 for (int i = 0; i < count; i++) {
5179 Node* n = macro_node(i);
5180 if (n->is_Allocate()) {
5181 if (i != allocates) {
5182 Node* tmp = macro_node(allocates);
5183 _macro_nodes.at_put(allocates, n);
5184 _macro_nodes.at_put(i, tmp);
5185 }
5186 allocates++;
5187 }
5188 }
5189 }
5190
5191 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5192 if (failing()) { return; }
5193 EventCompilerPhase event(UNTIMED);
5194 if (event.should_commit()) {
5195 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5196 }
5197 #ifndef PRODUCT
5198 ResourceMark rm;
5199 stringStream ss;
5200 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5201 int iter = ++_igv_phase_iter[cpt];
5202 if (iter > 1) {
5203 ss.print(" %d", iter);
5204 }
5205 if (n != nullptr) {
5206 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5207 }
5208
5209 const char* name = ss.as_string();
5210 if (should_print_igv(level)) {
5211 _igv_printer->print_method(name, level);
5212 }
5213 if (should_print_phase(cpt)) {
5214 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5215 }
5216 #endif
5217 C->_latest_stage_start_counter.stamp();
5218 }
5219
5220 // Only used from CompileWrapper
5221 void Compile::begin_method() {
5222 #ifndef PRODUCT
5223 if (_method != nullptr && should_print_igv(1)) {
5224 _igv_printer->begin_method();
5225 }
5226 #endif
5227 C->_latest_stage_start_counter.stamp();
5228 }
5229
5230 // Only used from CompileWrapper
5231 void Compile::end_method() {
5232 EventCompilerPhase event(UNTIMED);
5233 if (event.should_commit()) {
5234 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5235 }
5236
5237 #ifndef PRODUCT
5238 if (_method != nullptr && should_print_igv(1)) {
5239 _igv_printer->end_method();
5240 }
5241 #endif
5242 }
5243
5244 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5245 #ifndef PRODUCT
5246 if (_directive->should_print_phase(cpt)) {
5247 return true;
5248 }
5249 #endif
5250 return false;
5251 }
5252
5253 bool Compile::should_print_igv(const int level) {
5254 #ifndef PRODUCT
5255 if (PrintIdealGraphLevel < 0) { // disabled by the user
5256 return false;
5257 }
5258
5259 bool need = directive()->IGVPrintLevelOption >= level;
5260 if (need && !_igv_printer) {
5261 _igv_printer = IdealGraphPrinter::printer();
5262 _igv_printer->set_compile(this);
5263 }
5264 return need;
5265 #else
5266 return false;
5267 #endif
5268 }
5269
5270 #ifndef PRODUCT
5271 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5272 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5273
5274 // Called from debugger. Prints method to the default file with the default phase name.
5275 // This works regardless of any Ideal Graph Visualizer flags set or not.
5276 void igv_print() {
5277 Compile::current()->igv_print_method_to_file();
5278 }
5279
5280 // Same as igv_print() above but with a specified phase name.
5281 void igv_print(const char* phase_name) {
5282 Compile::current()->igv_print_method_to_file(phase_name);
5283 }
5284
5285 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5286 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5287 // This works regardless of any Ideal Graph Visualizer flags set or not.
5288 void igv_print(bool network) {
5289 if (network) {
5290 Compile::current()->igv_print_method_to_network();
5291 } else {
5292 Compile::current()->igv_print_method_to_file();
5293 }
5294 }
5295
5296 // Same as igv_print(bool network) above but with a specified phase name.
5297 void igv_print(bool network, const char* phase_name) {
5298 if (network) {
5299 Compile::current()->igv_print_method_to_network(phase_name);
5300 } else {
5301 Compile::current()->igv_print_method_to_file(phase_name);
5302 }
5303 }
5304
5305 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5306 void igv_print_default() {
5307 Compile::current()->print_method(PHASE_DEBUG, 0);
5308 }
5309
5310 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5311 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5312 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5313 void igv_append() {
5314 Compile::current()->igv_print_method_to_file("Debug", true);
5315 }
5316
5317 // Same as igv_append() above but with a specified phase name.
5318 void igv_append(const char* phase_name) {
5319 Compile::current()->igv_print_method_to_file(phase_name, true);
5320 }
5321
5322 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5323 const char* file_name = "custom_debug.xml";
5324 if (_debug_file_printer == nullptr) {
5325 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5326 } else {
5327 _debug_file_printer->update_compiled_method(C->method());
5328 }
5329 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5330 _debug_file_printer->print(phase_name, (Node*)C->root());
5331 }
5332
5333 void Compile::igv_print_method_to_network(const char* phase_name) {
5334 if (_debug_network_printer == nullptr) {
5335 _debug_network_printer = new IdealGraphPrinter(C);
5336 } else {
5337 _debug_network_printer->update_compiled_method(C->method());
5338 }
5339 tty->print_cr("Method printed over network stream to IGV");
5340 _debug_network_printer->print(phase_name, (Node*)C->root());
5341 }
5342 #endif
5343
5344 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5345 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5346 return value;
5347 }
5348 Node* result = nullptr;
5349 if (bt == T_BYTE) {
5350 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5351 result = new RShiftINode(result, phase->intcon(24));
5352 } else if (bt == T_BOOLEAN) {
5353 result = new AndINode(value, phase->intcon(0xFF));
5354 } else if (bt == T_CHAR) {
5355 result = new AndINode(value,phase->intcon(0xFFFF));
5356 } else {
5357 assert(bt == T_SHORT, "unexpected narrow type");
5358 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5359 result = new RShiftINode(result, phase->intcon(16));
5360 }
5361 if (transform_res) {
5362 result = phase->transform(result);
5363 }
5364 return result;
5365 }
5366
5367 void Compile::record_method_not_compilable_oom() {
5368 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5369 }