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/compilerOracle.hpp"
37 #include "compiler/compiler_globals.hpp"
38 #include "compiler/disassembler.hpp"
39 #include "compiler/oopMap.hpp"
40 #include "gc/shared/barrierSet.hpp"
41 #include "gc/shared/c2/barrierSetC2.hpp"
42 #include "jfr/jfrEvents.hpp"
43 #include "jvm_io.h"
44 #include "memory/allocation.hpp"
45 #include "memory/resourceArea.hpp"
46 #include "opto/addnode.hpp"
47 #include "opto/block.hpp"
48 #include "opto/c2compiler.hpp"
49 #include "opto/callGenerator.hpp"
50 #include "opto/callnode.hpp"
51 #include "opto/castnode.hpp"
52 #include "opto/cfgnode.hpp"
53 #include "opto/chaitin.hpp"
54 #include "opto/compile.hpp"
55 #include "opto/connode.hpp"
56 #include "opto/convertnode.hpp"
57 #include "opto/divnode.hpp"
58 #include "opto/escape.hpp"
59 #include "opto/idealGraphPrinter.hpp"
60 #include "opto/locknode.hpp"
61 #include "opto/loopnode.hpp"
62 #include "opto/machnode.hpp"
63 #include "opto/macro.hpp"
64 #include "opto/matcher.hpp"
65 #include "opto/mathexactnode.hpp"
66 #include "opto/memnode.hpp"
67 #include "opto/mulnode.hpp"
68 #include "opto/narrowptrnode.hpp"
69 #include "opto/node.hpp"
70 #include "opto/opcodes.hpp"
71 #include "opto/output.hpp"
72 #include "opto/parse.hpp"
73 #include "opto/phaseX.hpp"
74 #include "opto/rootnode.hpp"
75 #include "opto/runtime.hpp"
76 #include "opto/stringopts.hpp"
77 #include "opto/type.hpp"
78 #include "opto/vector.hpp"
79 #include "opto/vectornode.hpp"
80 #include "runtime/globals_extension.hpp"
81 #include "runtime/sharedRuntime.hpp"
82 #include "runtime/signature.hpp"
83 #include "runtime/stubRoutines.hpp"
84 #include "runtime/timer.hpp"
85 #include "utilities/align.hpp"
86 #include "utilities/copy.hpp"
87 #include "utilities/macros.hpp"
88 #include "utilities/resourceHash.hpp"
89
90 // -------------------- Compile::mach_constant_base_node -----------------------
91 // Constant table base node singleton.
92 MachConstantBaseNode* Compile::mach_constant_base_node() {
93 if (_mach_constant_base_node == nullptr) {
94 _mach_constant_base_node = new MachConstantBaseNode();
95 _mach_constant_base_node->add_req(C->root());
96 }
97 return _mach_constant_base_node;
98 }
99
100
101 /// Support for intrinsics.
102
103 // Return the index at which m must be inserted (or already exists).
104 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
105 class IntrinsicDescPair {
106 private:
107 ciMethod* _m;
108 bool _is_virtual;
109 public:
110 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
111 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
112 ciMethod* m= elt->method();
113 ciMethod* key_m = key->_m;
114 if (key_m < m) return -1;
115 else if (key_m > m) return 1;
116 else {
117 bool is_virtual = elt->is_virtual();
118 bool key_virtual = key->_is_virtual;
119 if (key_virtual < is_virtual) return -1;
120 else if (key_virtual > is_virtual) return 1;
121 else return 0;
122 }
123 }
124 };
125 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
126 #ifdef ASSERT
127 for (int i = 1; i < _intrinsics.length(); i++) {
128 CallGenerator* cg1 = _intrinsics.at(i-1);
129 CallGenerator* cg2 = _intrinsics.at(i);
130 assert(cg1->method() != cg2->method()
131 ? cg1->method() < cg2->method()
132 : cg1->is_virtual() < cg2->is_virtual(),
133 "compiler intrinsics list must stay sorted");
134 }
135 #endif
136 IntrinsicDescPair pair(m, is_virtual);
137 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
138 }
139
140 void Compile::register_intrinsic(CallGenerator* cg) {
141 bool found = false;
142 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
143 assert(!found, "registering twice");
144 _intrinsics.insert_before(index, cg);
145 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
146 }
147
148 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
149 assert(m->is_loaded(), "don't try this on unloaded methods");
150 if (_intrinsics.length() > 0) {
151 bool found = false;
152 int index = intrinsic_insertion_index(m, is_virtual, found);
153 if (found) {
154 return _intrinsics.at(index);
155 }
156 }
157 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
158 if (m->intrinsic_id() != vmIntrinsics::_none &&
159 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
160 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
161 if (cg != nullptr) {
162 // Save it for next time:
163 register_intrinsic(cg);
164 return cg;
165 } else {
166 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
167 }
168 }
169 return nullptr;
170 }
171
172 // Compile::make_vm_intrinsic is defined in library_call.cpp.
173
174 #ifndef PRODUCT
175 // statistics gathering...
176
177 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
179
180 inline int as_int(vmIntrinsics::ID id) {
181 return vmIntrinsics::as_int(id);
182 }
183
184 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
185 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
186 int oflags = _intrinsic_hist_flags[as_int(id)];
187 assert(flags != 0, "what happened?");
188 if (is_virtual) {
189 flags |= _intrinsic_virtual;
190 }
191 bool changed = (flags != oflags);
192 if ((flags & _intrinsic_worked) != 0) {
193 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
194 if (count == 1) {
195 changed = true; // first time
196 }
197 // increment the overall count also:
198 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
199 }
200 if (changed) {
201 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
202 // Something changed about the intrinsic's virtuality.
203 if ((flags & _intrinsic_virtual) != 0) {
204 // This is the first use of this intrinsic as a virtual call.
205 if (oflags != 0) {
206 // We already saw it as a non-virtual, so note both cases.
207 flags |= _intrinsic_both;
208 }
209 } else if ((oflags & _intrinsic_both) == 0) {
210 // This is the first use of this intrinsic as a non-virtual
211 flags |= _intrinsic_both;
212 }
213 }
214 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
215 }
216 // update the overall flags also:
217 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
218 return changed;
219 }
220
221 static char* format_flags(int flags, char* buf) {
222 buf[0] = 0;
223 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
224 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
225 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
226 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
227 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
228 if (buf[0] == 0) strcat(buf, ",");
229 assert(buf[0] == ',', "must be");
230 return &buf[1];
231 }
232
233 void Compile::print_intrinsic_statistics() {
234 char flagsbuf[100];
235 ttyLocker ttyl;
236 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
237 tty->print_cr("Compiler intrinsic usage:");
238 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
239 if (total == 0) total = 1; // avoid div0 in case of no successes
240 #define PRINT_STAT_LINE(name, c, f) \
241 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
242 for (auto id : EnumRange<vmIntrinsicID>{}) {
243 int flags = _intrinsic_hist_flags[as_int(id)];
244 juint count = _intrinsic_hist_count[as_int(id)];
245 if ((flags | count) != 0) {
246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
247 }
248 }
249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
250 if (xtty != nullptr) xtty->tail("statistics");
251 }
252
253 void Compile::print_statistics() {
254 { ttyLocker ttyl;
255 if (xtty != nullptr) xtty->head("statistics type='opto'");
256 Parse::print_statistics();
257 PhaseStringOpts::print_statistics();
258 PhaseCCP::print_statistics();
259 PhaseRegAlloc::print_statistics();
260 PhaseOutput::print_statistics();
261 PhasePeephole::print_statistics();
262 PhaseIdealLoop::print_statistics();
263 ConnectionGraph::print_statistics();
264 PhaseMacroExpand::print_statistics();
265 if (xtty != nullptr) xtty->tail("statistics");
266 }
267 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
268 // put this under its own <statistics> element.
269 print_intrinsic_statistics();
270 }
271 }
272 #endif //PRODUCT
273
274 void Compile::gvn_replace_by(Node* n, Node* nn) {
275 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
276 Node* use = n->last_out(i);
277 bool is_in_table = initial_gvn()->hash_delete(use);
278 uint uses_found = 0;
279 for (uint j = 0; j < use->len(); j++) {
280 if (use->in(j) == n) {
281 if (j < use->req())
282 use->set_req(j, nn);
283 else
284 use->set_prec(j, nn);
285 uses_found++;
286 }
287 }
288 if (is_in_table) {
289 // reinsert into table
290 initial_gvn()->hash_find_insert(use);
291 }
292 record_for_igvn(use);
293 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
294 i -= uses_found; // we deleted 1 or more copies of this edge
295 }
296 }
297
298
299 // Identify all nodes that are reachable from below, useful.
300 // Use breadth-first pass that records state in a Unique_Node_List,
301 // recursive traversal is slower.
302 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
303 int estimated_worklist_size = live_nodes();
304 useful.map( estimated_worklist_size, nullptr ); // preallocate space
305
306 // Initialize worklist
307 if (root() != nullptr) { useful.push(root()); }
308 // If 'top' is cached, declare it useful to preserve cached node
309 if (cached_top_node()) { useful.push(cached_top_node()); }
310
311 // Push all useful nodes onto the list, breadthfirst
312 for( uint next = 0; next < useful.size(); ++next ) {
313 assert( next < unique(), "Unique useful nodes < total nodes");
314 Node *n = useful.at(next);
315 uint max = n->len();
316 for( uint i = 0; i < max; ++i ) {
317 Node *m = n->in(i);
318 if (not_a_node(m)) continue;
319 useful.push(m);
320 }
321 }
322 }
323
324 // Update dead_node_list with any missing dead nodes using useful
325 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
326 void Compile::update_dead_node_list(Unique_Node_List &useful) {
327 uint max_idx = unique();
328 VectorSet& useful_node_set = useful.member_set();
329
330 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
331 // If node with index node_idx is not in useful set,
332 // mark it as dead in dead node list.
333 if (!useful_node_set.test(node_idx)) {
334 record_dead_node(node_idx);
335 }
336 }
337 }
338
339 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
340 int shift = 0;
341 for (int i = 0; i < inlines->length(); i++) {
342 CallGenerator* cg = inlines->at(i);
343 if (useful.member(cg->call_node())) {
344 if (shift > 0) {
345 inlines->at_put(i - shift, cg);
346 }
347 } else {
348 shift++; // skip over the dead element
349 }
350 }
351 if (shift > 0) {
352 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
353 }
354 }
355
356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
357 assert(dead != nullptr && dead->is_Call(), "sanity");
358 int found = 0;
359 for (int i = 0; i < inlines->length(); i++) {
360 if (inlines->at(i)->call_node() == dead) {
361 inlines->remove_at(i);
362 found++;
363 NOT_DEBUG( break; ) // elements are unique, so exit early
364 }
365 }
366 assert(found <= 1, "not unique");
367 }
368
369 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
370 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
371 for (int i = node_list.length() - 1; i >= 0; i--) {
372 N* node = node_list.at(i);
373 if (!useful.member(node)) {
374 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
375 }
376 }
377 }
378
379 void Compile::remove_useless_node(Node* dead) {
380 remove_modified_node(dead);
381
382 // Constant node that has no out-edges and has only one in-edge from
383 // root is usually dead. However, sometimes reshaping walk makes
384 // it reachable by adding use edges. So, we will NOT count Con nodes
385 // as dead to be conservative about the dead node count at any
386 // given time.
387 if (!dead->is_Con()) {
388 record_dead_node(dead->_idx);
389 }
390 if (dead->is_macro()) {
391 remove_macro_node(dead);
392 }
393 if (dead->is_expensive()) {
394 remove_expensive_node(dead);
395 }
396 if (dead->Opcode() == Op_Opaque4) {
397 remove_template_assertion_predicate_opaq(dead);
398 }
399 if (dead->is_ParsePredicate()) {
400 remove_parse_predicate(dead->as_ParsePredicate());
401 }
402 if (dead->for_post_loop_opts_igvn()) {
403 remove_from_post_loop_opts_igvn(dead);
404 }
405 if (dead->is_Call()) {
406 remove_useless_late_inlines( &_late_inlines, dead);
407 remove_useless_late_inlines( &_string_late_inlines, dead);
408 remove_useless_late_inlines( &_boxing_late_inlines, dead);
409 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
410
411 if (dead->is_CallStaticJava()) {
412 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
413 }
414 }
415 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
416 bs->unregister_potential_barrier_node(dead);
417 }
418
419 // Disconnect all useless nodes by disconnecting those at the boundary.
420 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
421 uint next = 0;
422 while (next < useful.size()) {
423 Node *n = useful.at(next++);
424 if (n->is_SafePoint()) {
425 // We're done with a parsing phase. Replaced nodes are not valid
426 // beyond that point.
427 n->as_SafePoint()->delete_replaced_nodes();
428 }
429 // Use raw traversal of out edges since this code removes out edges
430 int max = n->outcnt();
431 for (int j = 0; j < max; ++j) {
432 Node* child = n->raw_out(j);
433 if (!useful.member(child)) {
434 assert(!child->is_top() || child != top(),
435 "If top is cached in Compile object it is in useful list");
436 // Only need to remove this out-edge to the useless node
437 n->raw_del_out(j);
438 --j;
439 --max;
440 }
441 }
442 if (n->outcnt() == 1 && n->has_special_unique_user()) {
443 assert(useful.member(n->unique_out()), "do not push a useless node");
444 worklist.push(n->unique_out());
445 }
446 }
447
448 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
449 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
450 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
451 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
452 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
453 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
454 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
455 #ifdef ASSERT
456 if (_modified_nodes != nullptr) {
457 _modified_nodes->remove_useless_nodes(useful.member_set());
458 }
459 #endif
460
461 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
462 bs->eliminate_useless_gc_barriers(useful, this);
463 // clean up the late inline lists
464 remove_useless_late_inlines( &_late_inlines, useful);
465 remove_useless_late_inlines( &_string_late_inlines, useful);
466 remove_useless_late_inlines( &_boxing_late_inlines, useful);
467 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
468 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
469 }
470
471 // ============================================================================
472 //------------------------------CompileWrapper---------------------------------
473 class CompileWrapper : public StackObj {
474 Compile *const _compile;
475 public:
476 CompileWrapper(Compile* compile);
477
478 ~CompileWrapper();
479 };
480
481 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
482 // the Compile* pointer is stored in the current ciEnv:
483 ciEnv* env = compile->env();
484 assert(env == ciEnv::current(), "must already be a ciEnv active");
485 assert(env->compiler_data() == nullptr, "compile already active?");
486 env->set_compiler_data(compile);
487 assert(compile == Compile::current(), "sanity");
488
489 compile->set_type_dict(nullptr);
490 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
491 compile->clone_map().set_clone_idx(0);
492 compile->set_type_last_size(0);
493 compile->set_last_tf(nullptr, nullptr);
494 compile->set_indexSet_arena(nullptr);
495 compile->set_indexSet_free_block_list(nullptr);
496 compile->init_type_arena();
497 Type::Initialize(compile);
498 _compile->begin_method();
499 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
500 }
501 CompileWrapper::~CompileWrapper() {
502 // simulate crash during compilation
503 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
504
505 _compile->end_method();
506 _compile->env()->set_compiler_data(nullptr);
507 }
508
509
510 //----------------------------print_compile_messages---------------------------
511 void Compile::print_compile_messages() {
512 #ifndef PRODUCT
513 // Check if recompiling
514 if (!subsume_loads() && PrintOpto) {
515 // Recompiling without allowing machine instructions to subsume loads
516 tty->print_cr("*********************************************************");
517 tty->print_cr("** Bailout: Recompile without subsuming loads **");
518 tty->print_cr("*********************************************************");
519 }
520 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
521 // Recompiling without escape analysis
522 tty->print_cr("*********************************************************");
523 tty->print_cr("** Bailout: Recompile without escape analysis **");
524 tty->print_cr("*********************************************************");
525 }
526 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
527 // Recompiling without iterative escape analysis
528 tty->print_cr("*********************************************************");
529 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
530 tty->print_cr("*********************************************************");
531 }
532 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
533 // Recompiling without reducing allocation merges
534 tty->print_cr("*********************************************************");
535 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
536 tty->print_cr("*********************************************************");
537 }
538 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
539 // Recompiling without boxing elimination
540 tty->print_cr("*********************************************************");
541 tty->print_cr("** Bailout: Recompile without boxing elimination **");
542 tty->print_cr("*********************************************************");
543 }
544 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
545 // Recompiling without locks coarsening
546 tty->print_cr("*********************************************************");
547 tty->print_cr("** Bailout: Recompile without locks coarsening **");
548 tty->print_cr("*********************************************************");
549 }
550 if (env()->break_at_compile()) {
551 // Open the debugger when compiling this method.
552 tty->print("### Breaking when compiling: ");
553 method()->print_short_name();
554 tty->cr();
555 BREAKPOINT;
556 }
557
558 if( PrintOpto ) {
559 if (is_osr_compilation()) {
560 tty->print("[OSR]%3d", _compile_id);
561 } else {
562 tty->print("%3d", _compile_id);
563 }
564 }
565 #endif
566 }
567
568 #ifndef PRODUCT
569 void Compile::print_ideal_ir(const char* phase_name) {
570 // keep the following output all in one block
571 // This output goes directly to the tty, not the compiler log.
572 // To enable tools to match it up with the compilation activity,
573 // be sure to tag this tty output with the compile ID.
574
575 // Node dumping can cause a safepoint, which can break the tty lock.
576 // Buffer all node dumps, so that all safepoints happen before we lock.
577 ResourceMark rm;
578 stringStream ss;
579
580 if (_output == nullptr) {
581 ss.print_cr("AFTER: %s", phase_name);
582 // Print out all nodes in ascending order of index.
583 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
584 } else {
585 // Dump the node blockwise if we have a scheduling
586 _output->print_scheduling(&ss);
587 }
588
589 // Check that the lock is not broken by a safepoint.
590 NoSafepointVerifier nsv;
591 ttyLocker ttyl;
592 if (xtty != nullptr) {
593 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
594 compile_id(),
595 is_osr_compilation() ? " compile_kind='osr'" : "",
596 phase_name);
597 }
598
599 tty->print("%s", ss.as_string());
600
601 if (xtty != nullptr) {
602 xtty->tail("ideal");
603 }
604 }
605 #endif
606
607 // ============================================================================
608 //------------------------------Compile standard-------------------------------
609
610 // Compile a method. entry_bci is -1 for normal compilations and indicates
611 // the continuation bci for on stack replacement.
612
613
614 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
615 Options options, DirectiveSet* directive)
616 : Phase(Compiler),
617 _compile_id(ci_env->compile_id()),
618 _options(options),
619 _method(target),
620 _entry_bci(osr_bci),
621 _ilt(nullptr),
622 _stub_function(nullptr),
623 _stub_name(nullptr),
624 _stub_entry_point(nullptr),
625 _max_node_limit(MaxNodeLimit),
626 _post_loop_opts_phase(false),
627 _allow_macro_nodes(true),
628 _inlining_progress(false),
629 _inlining_incrementally(false),
630 _do_cleanup(false),
631 _has_reserved_stack_access(target->has_reserved_stack_access()),
632 #ifndef PRODUCT
633 _igv_idx(0),
634 _trace_opto_output(directive->TraceOptoOutputOption),
635 #endif
636 _has_method_handle_invokes(false),
637 _clinit_barrier_on_entry(false),
638 _has_clinit_barriers(false),
639 _stress_seed(0),
640 _comp_arena(mtCompiler),
641 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
642 _env(ci_env),
643 _directive(directive),
644 _log(ci_env->log()),
645 _first_failure_details(nullptr),
646 _intrinsics (comp_arena(), 0, 0, nullptr),
647 _macro_nodes (comp_arena(), 8, 0, nullptr),
648 _parse_predicates (comp_arena(), 8, 0, nullptr),
649 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
650 _expensive_nodes (comp_arena(), 8, 0, nullptr),
651 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
652 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
653 _coarsened_locks (comp_arena(), 8, 0, nullptr),
654 _congraph(nullptr),
655 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
656 _unique(0),
657 _dead_node_count(0),
658 _dead_node_list(comp_arena()),
659 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
660 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
661 _node_arena(&_node_arena_one),
662 _mach_constant_base_node(nullptr),
663 _Compile_types(mtCompiler),
664 _initial_gvn(nullptr),
665 _igvn_worklist(nullptr),
666 _types(nullptr),
667 _node_hash(nullptr),
668 _late_inlines(comp_arena(), 2, 0, nullptr),
669 _string_late_inlines(comp_arena(), 2, 0, nullptr),
670 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
671 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
672 _late_inlines_pos(0),
673 _number_of_mh_late_inlines(0),
674 _oom(false),
675 _print_inlining_stream(new (mtCompiler) stringStream()),
676 _print_inlining_list(nullptr),
677 _print_inlining_idx(0),
678 _print_inlining_output(nullptr),
679 _replay_inline_data(nullptr),
680 _java_calls(0),
681 _inner_loops(0),
682 _interpreter_frame_size(0),
683 _output(nullptr)
684 #ifndef PRODUCT
685 , _in_dump_cnt(0)
686 #endif
687 {
688 C = this;
689 CompileWrapper cw(this);
690
691 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
692 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
693
694 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
695 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
696 // We can always print a disassembly, either abstract (hex dump) or
697 // with the help of a suitable hsdis library. Thus, we should not
698 // couple print_assembly and print_opto_assembly controls.
699 // But: always print opto and regular assembly on compile command 'print'.
700 bool print_assembly = directive->PrintAssemblyOption;
701 set_print_assembly(print_opto_assembly || print_assembly);
702 #else
703 set_print_assembly(false); // must initialize.
704 #endif
705
706 #ifndef PRODUCT
707 set_parsed_irreducible_loop(false);
708 #endif
709
710 if (directive->ReplayInlineOption) {
711 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
712 }
713 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
714 set_print_intrinsics(directive->PrintIntrinsicsOption);
715 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
716
717 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
718 // Make sure the method being compiled gets its own MDO,
719 // so we can at least track the decompile_count().
720 // Need MDO to record RTM code generation state.
721 method()->ensure_method_data();
722 }
723
724 Init(/*do_aliasing=*/ true);
725
726 print_compile_messages();
727
728 _ilt = InlineTree::build_inline_tree_root();
729
730 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
731 assert(num_alias_types() >= AliasIdxRaw, "");
732
733 #define MINIMUM_NODE_HASH 1023
734
735 // GVN that will be run immediately on new nodes
736 uint estimated_size = method()->code_size()*4+64;
737 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
738 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
739 _types = new (comp_arena()) Type_Array(comp_arena());
740 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
741 PhaseGVN gvn;
742 set_initial_gvn(&gvn);
743
744 print_inlining_init();
745 { // Scope for timing the parser
746 TracePhase tp("parse", &timers[_t_parser]);
747
748 // Put top into the hash table ASAP.
749 initial_gvn()->transform(top());
750
751 // Set up tf(), start(), and find a CallGenerator.
752 CallGenerator* cg = nullptr;
753 if (is_osr_compilation()) {
754 const TypeTuple *domain = StartOSRNode::osr_domain();
755 const TypeTuple *range = TypeTuple::make_range(method()->signature());
756 init_tf(TypeFunc::make(domain, range));
757 StartNode* s = new StartOSRNode(root(), domain);
758 initial_gvn()->set_type_bottom(s);
759 init_start(s);
760 cg = CallGenerator::for_osr(method(), entry_bci());
761 } else {
762 // Normal case.
763 init_tf(TypeFunc::make(method()));
764 StartNode* s = new StartNode(root(), tf()->domain());
765 initial_gvn()->set_type_bottom(s);
766 init_start(s);
767 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
768 // With java.lang.ref.reference.get() we must go through the
769 // intrinsic - even when get() is the root
770 // method of the compile - so that, if necessary, the value in
771 // the referent field of the reference object gets recorded by
772 // the pre-barrier code.
773 cg = find_intrinsic(method(), false);
774 }
775 if (cg == nullptr) {
776 float past_uses = method()->interpreter_invocation_count();
777 float expected_uses = past_uses;
778 cg = CallGenerator::for_inline(method(), expected_uses);
779 }
780 }
781 if (failing()) return;
782 if (cg == nullptr) {
783 const char* reason = InlineTree::check_can_parse(method());
784 assert(reason != nullptr, "expect reason for parse failure");
785 stringStream ss;
786 ss.print("cannot parse method: %s", reason);
787 record_method_not_compilable(ss.as_string());
788 return;
789 }
790
791 gvn.set_type(root(), root()->bottom_type());
792
793 JVMState* jvms = build_start_state(start(), tf());
794 if ((jvms = cg->generate(jvms)) == nullptr) {
795 assert(failure_reason() != nullptr, "expect reason for parse failure");
796 stringStream ss;
797 ss.print("method parse failed: %s", failure_reason());
798 record_method_not_compilable(ss.as_string());
799 return;
800 }
801 GraphKit kit(jvms);
802
803 if (!kit.stopped()) {
804 // Accept return values, and transfer control we know not where.
805 // This is done by a special, unique ReturnNode bound to root.
806 return_values(kit.jvms());
807 }
808
809 if (kit.has_exceptions()) {
810 // Any exceptions that escape from this call must be rethrown
811 // to whatever caller is dynamically above us on the stack.
812 // This is done by a special, unique RethrowNode bound to root.
813 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
814 }
815
816 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
817
818 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
819 inline_string_calls(true);
820 }
821
822 if (failing()) return;
823
824 // Remove clutter produced by parsing.
825 if (!failing()) {
826 ResourceMark rm;
827 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
828 }
829 }
830
831 // Note: Large methods are capped off in do_one_bytecode().
832 if (failing()) return;
833
834 // After parsing, node notes are no longer automagic.
835 // They must be propagated by register_new_node_with_optimizer(),
836 // clone(), or the like.
837 set_default_node_notes(nullptr);
838
839 #ifndef PRODUCT
840 if (should_print_igv(1)) {
841 _igv_printer->print_inlining();
842 }
843 #endif
844
845 if (failing()) return;
846 NOT_PRODUCT( verify_graph_edges(); )
847
848 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
849 StressIncrementalInlining || StressMacroExpansion) {
850 initialize_stress_seed(directive);
851 }
852
853 // Now optimize
854 Optimize();
855 if (failing()) return;
856 NOT_PRODUCT( verify_graph_edges(); )
857
858 #ifndef PRODUCT
859 if (should_print_ideal()) {
860 print_ideal_ir("print_ideal");
861 }
862 #endif
863
864 #ifdef ASSERT
865 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
866 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
867 #endif
868
869 // Dump compilation data to replay it.
870 if (directive->DumpReplayOption) {
871 env()->dump_replay_data(_compile_id);
872 }
873 if (directive->DumpInlineOption && (ilt() != nullptr)) {
874 env()->dump_inline_data(_compile_id);
875 }
876
877 // Now that we know the size of all the monitors we can add a fixed slot
878 // for the original deopt pc.
879 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
880 set_fixed_slots(next_slot);
881
882 // Compute when to use implicit null checks. Used by matching trap based
883 // nodes and NullCheck optimization.
884 set_allowed_deopt_reasons();
885
886 // Now generate code
887 Code_Gen();
888 }
889
890 //------------------------------Compile----------------------------------------
891 // Compile a runtime stub
892 Compile::Compile( ciEnv* ci_env,
893 TypeFunc_generator generator,
894 address stub_function,
895 const char *stub_name,
896 int is_fancy_jump,
897 bool pass_tls,
898 bool return_pc,
899 DirectiveSet* directive)
900 : Phase(Compiler),
901 _compile_id(0),
902 _options(Options::for_runtime_stub()),
903 _method(nullptr),
904 _entry_bci(InvocationEntryBci),
905 _stub_function(stub_function),
906 _stub_name(stub_name),
907 _stub_entry_point(nullptr),
908 _max_node_limit(MaxNodeLimit),
909 _post_loop_opts_phase(false),
910 _allow_macro_nodes(true),
911 _inlining_progress(false),
912 _inlining_incrementally(false),
913 _has_reserved_stack_access(false),
914 #ifndef PRODUCT
915 _igv_idx(0),
916 _trace_opto_output(directive->TraceOptoOutputOption),
917 #endif
918 _has_method_handle_invokes(false),
919 _clinit_barrier_on_entry(false),
920 _has_clinit_barriers(false),
921 _stress_seed(0),
922 _comp_arena(mtCompiler),
923 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
924 _env(ci_env),
925 _directive(directive),
926 _log(ci_env->log()),
927 _first_failure_details(nullptr),
928 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
929 _congraph(nullptr),
930 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
931 _unique(0),
932 _dead_node_count(0),
933 _dead_node_list(comp_arena()),
934 _node_arena_one(mtCompiler),
935 _node_arena_two(mtCompiler),
936 _node_arena(&_node_arena_one),
937 _mach_constant_base_node(nullptr),
938 _Compile_types(mtCompiler),
939 _initial_gvn(nullptr),
940 _igvn_worklist(nullptr),
941 _types(nullptr),
942 _node_hash(nullptr),
943 _number_of_mh_late_inlines(0),
944 _oom(false),
945 _print_inlining_stream(new (mtCompiler) stringStream()),
946 _print_inlining_list(nullptr),
947 _print_inlining_idx(0),
948 _print_inlining_output(nullptr),
949 _replay_inline_data(nullptr),
950 _java_calls(0),
951 _inner_loops(0),
952 _interpreter_frame_size(0),
953 _output(nullptr),
954 #ifndef PRODUCT
955 _in_dump_cnt(0),
956 #endif
957 _allowed_reasons(0) {
958 C = this;
959
960 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
961 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
962
963 #ifndef PRODUCT
964 set_print_assembly(PrintFrameConverterAssembly);
965 set_parsed_irreducible_loop(false);
966 #else
967 set_print_assembly(false); // Must initialize.
968 #endif
969 set_has_irreducible_loop(false); // no loops
970
971 CompileWrapper cw(this);
972 Init(/*do_aliasing=*/ false);
973 init_tf((*generator)());
974
975 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
976 _types = new (comp_arena()) Type_Array(comp_arena());
977 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
978
979 if (StressLCM || StressGCM) {
980 initialize_stress_seed(directive);
981 }
982
983 {
984 PhaseGVN gvn;
985 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
986 gvn.transform(top());
987
988 GraphKit kit;
989 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
990 }
991
992 NOT_PRODUCT( verify_graph_edges(); )
993
994 Code_Gen();
995 }
996
997 Compile::~Compile() {
998 delete _print_inlining_stream;
999 delete _first_failure_details;
1000 };
1001
1002 //------------------------------Init-------------------------------------------
1003 // Prepare for a single compilation
1004 void Compile::Init(bool aliasing) {
1005 _do_aliasing = aliasing;
1006 _unique = 0;
1007 _regalloc = nullptr;
1008
1009 _tf = nullptr; // filled in later
1010 _top = nullptr; // cached later
1011 _matcher = nullptr; // filled in later
1012 _cfg = nullptr; // filled in later
1013
1014 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1015
1016 _node_note_array = nullptr;
1017 _default_node_notes = nullptr;
1018 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1019
1020 _immutable_memory = nullptr; // filled in at first inquiry
1021
1022 #ifdef ASSERT
1023 _phase_optimize_finished = false;
1024 _exception_backedge = false;
1025 _type_verify = nullptr;
1026 #endif
1027
1028 // Globally visible Nodes
1029 // First set TOP to null to give safe behavior during creation of RootNode
1030 set_cached_top_node(nullptr);
1031 set_root(new RootNode());
1032 // Now that you have a Root to point to, create the real TOP
1033 set_cached_top_node( new ConNode(Type::TOP) );
1034 set_recent_alloc(nullptr, nullptr);
1035
1036 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1037 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1038 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1039 env()->set_dependencies(new Dependencies(env()));
1040
1041 _fixed_slots = 0;
1042 set_has_split_ifs(false);
1043 set_has_loops(false); // first approximation
1044 set_has_stringbuilder(false);
1045 set_has_boxed_value(false);
1046 _trap_can_recompile = false; // no traps emitted yet
1047 _major_progress = true; // start out assuming good things will happen
1048 set_has_unsafe_access(false);
1049 set_max_vector_size(0);
1050 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1051 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1052 set_decompile_count(0);
1053
1054 #ifndef PRODUCT
1055 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1056 #endif
1057
1058 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1059 _loop_opts_cnt = LoopOptsCount;
1060 set_do_inlining(Inline);
1061 set_max_inline_size(MaxInlineSize);
1062 set_freq_inline_size(FreqInlineSize);
1063 set_do_scheduling(OptoScheduling);
1064
1065 set_do_vector_loop(false);
1066 set_has_monitors(false);
1067
1068 if (AllowVectorizeOnDemand) {
1069 if (has_method() && _directive->VectorizeOption) {
1070 set_do_vector_loop(true);
1071 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());})
1072 } else if (has_method() && method()->name() != 0 &&
1073 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1074 set_do_vector_loop(true);
1075 }
1076 }
1077 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1078 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());})
1079
1080 set_rtm_state(NoRTM); // No RTM lock eliding by default
1081 _max_node_limit = _directive->MaxNodeLimitOption;
1082
1083 #if INCLUDE_RTM_OPT
1084 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) {
1085 int rtm_state = method()->method_data()->rtm_state();
1086 if (method_has_option(CompileCommandEnum::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1087 // Don't generate RTM lock eliding code.
1088 set_rtm_state(NoRTM);
1089 } else if (method_has_option(CompileCommandEnum::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1090 // Generate RTM lock eliding code without abort ratio calculation code.
1091 set_rtm_state(UseRTM);
1092 } else if (UseRTMDeopt) {
1093 // Generate RTM lock eliding code and include abort ratio calculation
1094 // code if UseRTMDeopt is on.
1095 set_rtm_state(ProfileRTM);
1096 }
1097 }
1098 #endif
1099 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() &&
1100 (method()->needs_clinit_barrier() || (do_clinit_barriers() && method()->is_static()))) {
1101 set_clinit_barrier_on_entry(true);
1102 if (do_clinit_barriers()) {
1103 set_has_clinit_barriers(true); // Entry clinit barrier is in prolog code.
1104 }
1105 }
1106 if (debug_info()->recording_non_safepoints()) {
1107 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1108 (comp_arena(), 8, 0, nullptr));
1109 set_default_node_notes(Node_Notes::make(this));
1110 }
1111
1112 const int grow_ats = 16;
1113 _max_alias_types = grow_ats;
1114 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1115 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1116 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1117 {
1118 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1119 }
1120 // Initialize the first few types.
1121 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1122 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1123 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1124 _num_alias_types = AliasIdxRaw+1;
1125 // Zero out the alias type cache.
1126 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1127 // A null adr_type hits in the cache right away. Preload the right answer.
1128 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1129 }
1130
1131 //---------------------------init_start----------------------------------------
1132 // Install the StartNode on this compile object.
1133 void Compile::init_start(StartNode* s) {
1134 if (failing())
1135 return; // already failing
1136 assert(s == start(), "");
1137 }
1138
1139 /**
1140 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1141 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1142 * the ideal graph.
1143 */
1144 StartNode* Compile::start() const {
1145 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1146 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1147 Node* start = root()->fast_out(i);
1148 if (start->is_Start()) {
1149 return start->as_Start();
1150 }
1151 }
1152 fatal("Did not find Start node!");
1153 return nullptr;
1154 }
1155
1156 //-------------------------------immutable_memory-------------------------------------
1157 // Access immutable memory
1158 Node* Compile::immutable_memory() {
1159 if (_immutable_memory != nullptr) {
1160 return _immutable_memory;
1161 }
1162 StartNode* s = start();
1163 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1164 Node *p = s->fast_out(i);
1165 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1166 _immutable_memory = p;
1167 return _immutable_memory;
1168 }
1169 }
1170 ShouldNotReachHere();
1171 return nullptr;
1172 }
1173
1174 //----------------------set_cached_top_node------------------------------------
1175 // Install the cached top node, and make sure Node::is_top works correctly.
1176 void Compile::set_cached_top_node(Node* tn) {
1177 if (tn != nullptr) verify_top(tn);
1178 Node* old_top = _top;
1179 _top = tn;
1180 // Calling Node::setup_is_top allows the nodes the chance to adjust
1181 // their _out arrays.
1182 if (_top != nullptr) _top->setup_is_top();
1183 if (old_top != nullptr) old_top->setup_is_top();
1184 assert(_top == nullptr || top()->is_top(), "");
1185 }
1186
1187 #ifdef ASSERT
1188 uint Compile::count_live_nodes_by_graph_walk() {
1189 Unique_Node_List useful(comp_arena());
1190 // Get useful node list by walking the graph.
1191 identify_useful_nodes(useful);
1192 return useful.size();
1193 }
1194
1195 void Compile::print_missing_nodes() {
1196
1197 // Return if CompileLog is null and PrintIdealNodeCount is false.
1198 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1199 return;
1200 }
1201
1202 // This is an expensive function. It is executed only when the user
1203 // specifies VerifyIdealNodeCount option or otherwise knows the
1204 // additional work that needs to be done to identify reachable nodes
1205 // by walking the flow graph and find the missing ones using
1206 // _dead_node_list.
1207
1208 Unique_Node_List useful(comp_arena());
1209 // Get useful node list by walking the graph.
1210 identify_useful_nodes(useful);
1211
1212 uint l_nodes = C->live_nodes();
1213 uint l_nodes_by_walk = useful.size();
1214
1215 if (l_nodes != l_nodes_by_walk) {
1216 if (_log != nullptr) {
1217 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1218 _log->stamp();
1219 _log->end_head();
1220 }
1221 VectorSet& useful_member_set = useful.member_set();
1222 int last_idx = l_nodes_by_walk;
1223 for (int i = 0; i < last_idx; i++) {
1224 if (useful_member_set.test(i)) {
1225 if (_dead_node_list.test(i)) {
1226 if (_log != nullptr) {
1227 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1228 }
1229 if (PrintIdealNodeCount) {
1230 // Print the log message to tty
1231 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1232 useful.at(i)->dump();
1233 }
1234 }
1235 }
1236 else if (! _dead_node_list.test(i)) {
1237 if (_log != nullptr) {
1238 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1239 }
1240 if (PrintIdealNodeCount) {
1241 // Print the log message to tty
1242 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1243 }
1244 }
1245 }
1246 if (_log != nullptr) {
1247 _log->tail("mismatched_nodes");
1248 }
1249 }
1250 }
1251 void Compile::record_modified_node(Node* n) {
1252 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1253 _modified_nodes->push(n);
1254 }
1255 }
1256
1257 void Compile::remove_modified_node(Node* n) {
1258 if (_modified_nodes != nullptr) {
1259 _modified_nodes->remove(n);
1260 }
1261 }
1262 #endif
1263
1264 #ifndef PRODUCT
1265 void Compile::verify_top(Node* tn) const {
1266 if (tn != nullptr) {
1267 assert(tn->is_Con(), "top node must be a constant");
1268 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1269 assert(tn->in(0) != nullptr, "must have live top node");
1270 }
1271 }
1272 #endif
1273
1274
1275 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1276
1277 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1278 guarantee(arr != nullptr, "");
1279 int num_blocks = arr->length();
1280 if (grow_by < num_blocks) grow_by = num_blocks;
1281 int num_notes = grow_by * _node_notes_block_size;
1282 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1283 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1284 while (num_notes > 0) {
1285 arr->append(notes);
1286 notes += _node_notes_block_size;
1287 num_notes -= _node_notes_block_size;
1288 }
1289 assert(num_notes == 0, "exact multiple, please");
1290 }
1291
1292 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1293 if (source == nullptr || dest == nullptr) return false;
1294
1295 if (dest->is_Con())
1296 return false; // Do not push debug info onto constants.
1297
1298 #ifdef ASSERT
1299 // Leave a bread crumb trail pointing to the original node:
1300 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1301 dest->set_debug_orig(source);
1302 }
1303 #endif
1304
1305 if (node_note_array() == nullptr)
1306 return false; // Not collecting any notes now.
1307
1308 // This is a copy onto a pre-existing node, which may already have notes.
1309 // If both nodes have notes, do not overwrite any pre-existing notes.
1310 Node_Notes* source_notes = node_notes_at(source->_idx);
1311 if (source_notes == nullptr || source_notes->is_clear()) return false;
1312 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1313 if (dest_notes == nullptr || dest_notes->is_clear()) {
1314 return set_node_notes_at(dest->_idx, source_notes);
1315 }
1316
1317 Node_Notes merged_notes = (*source_notes);
1318 // The order of operations here ensures that dest notes will win...
1319 merged_notes.update_from(dest_notes);
1320 return set_node_notes_at(dest->_idx, &merged_notes);
1321 }
1322
1323
1324 //--------------------------allow_range_check_smearing-------------------------
1325 // Gating condition for coalescing similar range checks.
1326 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1327 // single covering check that is at least as strong as any of them.
1328 // If the optimization succeeds, the simplified (strengthened) range check
1329 // will always succeed. If it fails, we will deopt, and then give up
1330 // on the optimization.
1331 bool Compile::allow_range_check_smearing() const {
1332 // If this method has already thrown a range-check,
1333 // assume it was because we already tried range smearing
1334 // and it failed.
1335 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1336 return !already_trapped;
1337 }
1338
1339
1340 //------------------------------flatten_alias_type-----------------------------
1341 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1342 assert(do_aliasing(), "Aliasing should be enabled");
1343 int offset = tj->offset();
1344 TypePtr::PTR ptr = tj->ptr();
1345
1346 // Known instance (scalarizable allocation) alias only with itself.
1347 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1348 tj->is_oopptr()->is_known_instance();
1349
1350 // Process weird unsafe references.
1351 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1352 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1353 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1354 tj = TypeOopPtr::BOTTOM;
1355 ptr = tj->ptr();
1356 offset = tj->offset();
1357 }
1358
1359 // Array pointers need some flattening
1360 const TypeAryPtr* ta = tj->isa_aryptr();
1361 if (ta && ta->is_stable()) {
1362 // Erase stability property for alias analysis.
1363 tj = ta = ta->cast_to_stable(false);
1364 }
1365 if( ta && is_known_inst ) {
1366 if ( offset != Type::OffsetBot &&
1367 offset > arrayOopDesc::length_offset_in_bytes() ) {
1368 offset = Type::OffsetBot; // Flatten constant access into array body only
1369 tj = ta = ta->
1370 remove_speculative()->
1371 cast_to_ptr_type(ptr)->
1372 with_offset(offset);
1373 }
1374 } else if (ta) {
1375 // For arrays indexed by constant indices, we flatten the alias
1376 // space to include all of the array body. Only the header, klass
1377 // and array length can be accessed un-aliased.
1378 if( offset != Type::OffsetBot ) {
1379 if( ta->const_oop() ) { // MethodData* or Method*
1380 offset = Type::OffsetBot; // Flatten constant access into array body
1381 tj = ta = ta->
1382 remove_speculative()->
1383 cast_to_ptr_type(ptr)->
1384 cast_to_exactness(false)->
1385 with_offset(offset);
1386 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1387 // range is OK as-is.
1388 tj = ta = TypeAryPtr::RANGE;
1389 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1390 tj = TypeInstPtr::KLASS; // all klass loads look alike
1391 ta = TypeAryPtr::RANGE; // generic ignored junk
1392 ptr = TypePtr::BotPTR;
1393 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1394 tj = TypeInstPtr::MARK;
1395 ta = TypeAryPtr::RANGE; // generic ignored junk
1396 ptr = TypePtr::BotPTR;
1397 } else { // Random constant offset into array body
1398 offset = Type::OffsetBot; // Flatten constant access into array body
1399 tj = ta = ta->
1400 remove_speculative()->
1401 cast_to_ptr_type(ptr)->
1402 cast_to_exactness(false)->
1403 with_offset(offset);
1404 }
1405 }
1406 // Arrays of fixed size alias with arrays of unknown size.
1407 if (ta->size() != TypeInt::POS) {
1408 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1409 tj = ta = ta->
1410 remove_speculative()->
1411 cast_to_ptr_type(ptr)->
1412 with_ary(tary)->
1413 cast_to_exactness(false);
1414 }
1415 // Arrays of known objects become arrays of unknown objects.
1416 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1417 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1418 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1419 }
1420 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1421 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1422 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1423 }
1424 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1425 // cannot be distinguished by bytecode alone.
1426 if (ta->elem() == TypeInt::BOOL) {
1427 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1428 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1429 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1430 }
1431 // During the 2nd round of IterGVN, NotNull castings are removed.
1432 // Make sure the Bottom and NotNull variants alias the same.
1433 // Also, make sure exact and non-exact variants alias the same.
1434 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1435 tj = ta = ta->
1436 remove_speculative()->
1437 cast_to_ptr_type(TypePtr::BotPTR)->
1438 cast_to_exactness(false)->
1439 with_offset(offset);
1440 }
1441 }
1442
1443 // Oop pointers need some flattening
1444 const TypeInstPtr *to = tj->isa_instptr();
1445 if (to && to != TypeOopPtr::BOTTOM) {
1446 ciInstanceKlass* ik = to->instance_klass();
1447 if( ptr == TypePtr::Constant ) {
1448 if (ik != ciEnv::current()->Class_klass() ||
1449 offset < ik->layout_helper_size_in_bytes()) {
1450 // No constant oop pointers (such as Strings); they alias with
1451 // unknown strings.
1452 assert(!is_known_inst, "not scalarizable allocation");
1453 tj = to = to->
1454 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1455 remove_speculative()->
1456 cast_to_ptr_type(TypePtr::BotPTR)->
1457 cast_to_exactness(false);
1458 }
1459 } else if( is_known_inst ) {
1460 tj = to; // Keep NotNull and klass_is_exact for instance type
1461 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1462 // During the 2nd round of IterGVN, NotNull castings are removed.
1463 // Make sure the Bottom and NotNull variants alias the same.
1464 // Also, make sure exact and non-exact variants alias the same.
1465 tj = to = to->
1466 remove_speculative()->
1467 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1468 cast_to_ptr_type(TypePtr::BotPTR)->
1469 cast_to_exactness(false);
1470 }
1471 if (to->speculative() != nullptr) {
1472 tj = to = to->remove_speculative();
1473 }
1474 // Canonicalize the holder of this field
1475 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1476 // First handle header references such as a LoadKlassNode, even if the
1477 // object's klass is unloaded at compile time (4965979).
1478 if (!is_known_inst) { // Do it only for non-instance types
1479 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1480 }
1481 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1482 // Static fields are in the space above the normal instance
1483 // fields in the java.lang.Class instance.
1484 if (ik != ciEnv::current()->Class_klass()) {
1485 to = nullptr;
1486 tj = TypeOopPtr::BOTTOM;
1487 offset = tj->offset();
1488 }
1489 } else {
1490 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1491 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1492 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1493 if( is_known_inst ) {
1494 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id());
1495 } else {
1496 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset);
1497 }
1498 }
1499 }
1500 }
1501
1502 // Klass pointers to object array klasses need some flattening
1503 const TypeKlassPtr *tk = tj->isa_klassptr();
1504 if( tk ) {
1505 // If we are referencing a field within a Klass, we need
1506 // to assume the worst case of an Object. Both exact and
1507 // inexact types must flatten to the same alias class so
1508 // use NotNull as the PTR.
1509 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1510 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1511 env()->Object_klass(),
1512 offset);
1513 }
1514
1515 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1516 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1517 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1518 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1519 } else {
1520 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1521 }
1522 }
1523
1524 // Check for precise loads from the primary supertype array and force them
1525 // to the supertype cache alias index. Check for generic array loads from
1526 // the primary supertype array and also force them to the supertype cache
1527 // alias index. Since the same load can reach both, we need to merge
1528 // these 2 disparate memories into the same alias class. Since the
1529 // primary supertype array is read-only, there's no chance of confusion
1530 // where we bypass an array load and an array store.
1531 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1532 if (offset == Type::OffsetBot ||
1533 (offset >= primary_supers_offset &&
1534 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1535 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1536 offset = in_bytes(Klass::secondary_super_cache_offset());
1537 tj = tk = tk->with_offset(offset);
1538 }
1539 }
1540
1541 // Flatten all Raw pointers together.
1542 if (tj->base() == Type::RawPtr)
1543 tj = TypeRawPtr::BOTTOM;
1544
1545 if (tj->base() == Type::AnyPtr)
1546 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1547
1548 offset = tj->offset();
1549 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1550
1551 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1552 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1553 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1554 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1555 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1556 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1557 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1558 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1559 assert( tj->ptr() != TypePtr::TopPTR &&
1560 tj->ptr() != TypePtr::AnyNull &&
1561 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1562 // assert( tj->ptr() != TypePtr::Constant ||
1563 // tj->base() == Type::RawPtr ||
1564 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1565
1566 return tj;
1567 }
1568
1569 void Compile::AliasType::Init(int i, const TypePtr* at) {
1570 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1571 _index = i;
1572 _adr_type = at;
1573 _field = nullptr;
1574 _element = nullptr;
1575 _is_rewritable = true; // default
1576 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1577 if (atoop != nullptr && atoop->is_known_instance()) {
1578 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1579 _general_index = Compile::current()->get_alias_index(gt);
1580 } else {
1581 _general_index = 0;
1582 }
1583 }
1584
1585 BasicType Compile::AliasType::basic_type() const {
1586 if (element() != nullptr) {
1587 const Type* element = adr_type()->is_aryptr()->elem();
1588 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1589 } if (field() != nullptr) {
1590 return field()->layout_type();
1591 } else {
1592 return T_ILLEGAL; // unknown
1593 }
1594 }
1595
1596 //---------------------------------print_on------------------------------------
1597 #ifndef PRODUCT
1598 void Compile::AliasType::print_on(outputStream* st) {
1599 if (index() < 10)
1600 st->print("@ <%d> ", index());
1601 else st->print("@ <%d>", index());
1602 st->print(is_rewritable() ? " " : " RO");
1603 int offset = adr_type()->offset();
1604 if (offset == Type::OffsetBot)
1605 st->print(" +any");
1606 else st->print(" +%-3d", offset);
1607 st->print(" in ");
1608 adr_type()->dump_on(st);
1609 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1610 if (field() != nullptr && tjp) {
1611 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1612 tjp->offset() != field()->offset_in_bytes()) {
1613 st->print(" != ");
1614 field()->print();
1615 st->print(" ***");
1616 }
1617 }
1618 }
1619
1620 void print_alias_types() {
1621 Compile* C = Compile::current();
1622 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1623 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1624 C->alias_type(idx)->print_on(tty);
1625 tty->cr();
1626 }
1627 }
1628 #endif
1629
1630
1631 //----------------------------probe_alias_cache--------------------------------
1632 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1633 intptr_t key = (intptr_t) adr_type;
1634 key ^= key >> logAliasCacheSize;
1635 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1636 }
1637
1638
1639 //-----------------------------grow_alias_types--------------------------------
1640 void Compile::grow_alias_types() {
1641 const int old_ats = _max_alias_types; // how many before?
1642 const int new_ats = old_ats; // how many more?
1643 const int grow_ats = old_ats+new_ats; // how many now?
1644 _max_alias_types = grow_ats;
1645 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1646 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1647 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1648 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1649 }
1650
1651
1652 //--------------------------------find_alias_type------------------------------
1653 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1654 if (!do_aliasing()) {
1655 return alias_type(AliasIdxBot);
1656 }
1657
1658 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1659 if (ace->_adr_type == adr_type) {
1660 return alias_type(ace->_index);
1661 }
1662
1663 // Handle special cases.
1664 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1665 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1666
1667 // Do it the slow way.
1668 const TypePtr* flat = flatten_alias_type(adr_type);
1669
1670 #ifdef ASSERT
1671 {
1672 ResourceMark rm;
1673 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1674 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1675 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1676 Type::str(adr_type));
1677 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1678 const TypeOopPtr* foop = flat->is_oopptr();
1679 // Scalarizable allocations have exact klass always.
1680 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1681 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1682 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1683 Type::str(foop), Type::str(xoop));
1684 }
1685 }
1686 #endif
1687
1688 int idx = AliasIdxTop;
1689 for (int i = 0; i < num_alias_types(); i++) {
1690 if (alias_type(i)->adr_type() == flat) {
1691 idx = i;
1692 break;
1693 }
1694 }
1695
1696 if (idx == AliasIdxTop) {
1697 if (no_create) return nullptr;
1698 // Grow the array if necessary.
1699 if (_num_alias_types == _max_alias_types) grow_alias_types();
1700 // Add a new alias type.
1701 idx = _num_alias_types++;
1702 _alias_types[idx]->Init(idx, flat);
1703 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1704 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1705 if (flat->isa_instptr()) {
1706 if (flat->offset() == java_lang_Class::klass_offset()
1707 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1708 alias_type(idx)->set_rewritable(false);
1709 }
1710 if (flat->isa_aryptr()) {
1711 #ifdef ASSERT
1712 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1713 // (T_BYTE has the weakest alignment and size restrictions...)
1714 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1715 #endif
1716 if (flat->offset() == TypePtr::OffsetBot) {
1717 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1718 }
1719 }
1720 if (flat->isa_klassptr()) {
1721 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1722 alias_type(idx)->set_rewritable(false);
1723 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1724 alias_type(idx)->set_rewritable(false);
1725 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1726 alias_type(idx)->set_rewritable(false);
1727 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1728 alias_type(idx)->set_rewritable(false);
1729 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1730 alias_type(idx)->set_rewritable(false);
1731 }
1732 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1733 // but the base pointer type is not distinctive enough to identify
1734 // references into JavaThread.)
1735
1736 // Check for final fields.
1737 const TypeInstPtr* tinst = flat->isa_instptr();
1738 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1739 ciField* field;
1740 if (tinst->const_oop() != nullptr &&
1741 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1742 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1743 // static field
1744 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1745 field = k->get_field_by_offset(tinst->offset(), true);
1746 } else {
1747 ciInstanceKlass *k = tinst->instance_klass();
1748 field = k->get_field_by_offset(tinst->offset(), false);
1749 }
1750 assert(field == nullptr ||
1751 original_field == nullptr ||
1752 (field->holder() == original_field->holder() &&
1753 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1754 field->is_static() == original_field->is_static()), "wrong field?");
1755 // Set field() and is_rewritable() attributes.
1756 if (field != nullptr) alias_type(idx)->set_field(field);
1757 }
1758 }
1759
1760 // Fill the cache for next time.
1761 ace->_adr_type = adr_type;
1762 ace->_index = idx;
1763 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1764
1765 // Might as well try to fill the cache for the flattened version, too.
1766 AliasCacheEntry* face = probe_alias_cache(flat);
1767 if (face->_adr_type == nullptr) {
1768 face->_adr_type = flat;
1769 face->_index = idx;
1770 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1771 }
1772
1773 return alias_type(idx);
1774 }
1775
1776
1777 Compile::AliasType* Compile::alias_type(ciField* field) {
1778 const TypeOopPtr* t;
1779 if (field->is_static())
1780 t = TypeInstPtr::make(field->holder()->java_mirror());
1781 else
1782 t = TypeOopPtr::make_from_klass_raw(field->holder());
1783 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1784 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1785 return atp;
1786 }
1787
1788
1789 //------------------------------have_alias_type--------------------------------
1790 bool Compile::have_alias_type(const TypePtr* adr_type) {
1791 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1792 if (ace->_adr_type == adr_type) {
1793 return true;
1794 }
1795
1796 // Handle special cases.
1797 if (adr_type == nullptr) return true;
1798 if (adr_type == TypePtr::BOTTOM) return true;
1799
1800 return find_alias_type(adr_type, true, nullptr) != nullptr;
1801 }
1802
1803 //-----------------------------must_alias--------------------------------------
1804 // True if all values of the given address type are in the given alias category.
1805 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1806 if (alias_idx == AliasIdxBot) return true; // the universal category
1807 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1808 if (alias_idx == AliasIdxTop) return false; // the empty category
1809 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1810
1811 // the only remaining possible overlap is identity
1812 int adr_idx = get_alias_index(adr_type);
1813 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1814 assert(adr_idx == alias_idx ||
1815 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1816 && adr_type != TypeOopPtr::BOTTOM),
1817 "should not be testing for overlap with an unsafe pointer");
1818 return adr_idx == alias_idx;
1819 }
1820
1821 //------------------------------can_alias--------------------------------------
1822 // True if any values of the given address type are in the given alias category.
1823 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1824 if (alias_idx == AliasIdxTop) return false; // the empty category
1825 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1826 // Known instance doesn't alias with bottom memory
1827 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1828 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1829
1830 // the only remaining possible overlap is identity
1831 int adr_idx = get_alias_index(adr_type);
1832 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1833 return adr_idx == alias_idx;
1834 }
1835
1836 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1837 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1838 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1839 if (parse_predicate_count() == 0) {
1840 return;
1841 }
1842 for (int i = 0; i < parse_predicate_count(); i++) {
1843 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1844 parse_predicate->mark_useless();
1845 igvn._worklist.push(parse_predicate);
1846 }
1847 _parse_predicates.clear();
1848 }
1849
1850 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1851 if (!n->for_post_loop_opts_igvn()) {
1852 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1853 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1854 _for_post_loop_igvn.append(n);
1855 }
1856 }
1857
1858 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1859 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1860 _for_post_loop_igvn.remove(n);
1861 }
1862
1863 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1864 // Verify that all previous optimizations produced a valid graph
1865 // at least to this point, even if no loop optimizations were done.
1866 PhaseIdealLoop::verify(igvn);
1867
1868 C->set_post_loop_opts_phase(); // no more loop opts allowed
1869
1870 assert(!C->major_progress(), "not cleared");
1871
1872 if (_for_post_loop_igvn.length() > 0) {
1873 while (_for_post_loop_igvn.length() > 0) {
1874 Node* n = _for_post_loop_igvn.pop();
1875 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1876 igvn._worklist.push(n);
1877 }
1878 igvn.optimize();
1879 if (failing()) return;
1880 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1881 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1882
1883 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1884 if (C->major_progress()) {
1885 C->clear_major_progress(); // ensure that major progress is now clear
1886 }
1887 }
1888 }
1889
1890 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1891 if (OptimizeUnstableIf) {
1892 _unstable_if_traps.append(trap);
1893 }
1894 }
1895
1896 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1897 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1898 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1899 Node* n = trap->uncommon_trap();
1900 if (!useful.member(n)) {
1901 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1902 }
1903 }
1904 }
1905
1906 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1907 // or fold-compares case. Return true if succeed or not found.
1908 //
1909 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1910 // false and ask fold-compares to yield.
1911 //
1912 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1913 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1914 // when deoptimization does happen.
1915 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1916 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1917 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1918 if (trap->uncommon_trap() == unc) {
1919 if (yield && trap->modified()) {
1920 return false;
1921 }
1922 _unstable_if_traps.delete_at(i);
1923 break;
1924 }
1925 }
1926 return true;
1927 }
1928
1929 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1930 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1931 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1932 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1933 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1934 CallStaticJavaNode* unc = trap->uncommon_trap();
1935 int next_bci = trap->next_bci();
1936 bool modified = trap->modified();
1937
1938 if (next_bci != -1 && !modified) {
1939 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1940 JVMState* jvms = unc->jvms();
1941 ciMethod* method = jvms->method();
1942 ciBytecodeStream iter(method);
1943
1944 iter.force_bci(jvms->bci());
1945 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1946 Bytecodes::Code c = iter.cur_bc();
1947 Node* lhs = nullptr;
1948 Node* rhs = nullptr;
1949 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1950 lhs = unc->peek_operand(0);
1951 rhs = unc->peek_operand(1);
1952 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1953 lhs = unc->peek_operand(0);
1954 }
1955
1956 ResourceMark rm;
1957 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1958 assert(live_locals.is_valid(), "broken liveness info");
1959 int len = (int)live_locals.size();
1960
1961 for (int i = 0; i < len; i++) {
1962 Node* local = unc->local(jvms, i);
1963 // kill local using the liveness of next_bci.
1964 // give up when the local looks like an operand to secure reexecution.
1965 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1966 uint idx = jvms->locoff() + i;
1967 #ifdef ASSERT
1968 if (PrintOpto && Verbose) {
1969 tty->print("[unstable_if] kill local#%d: ", idx);
1970 local->dump();
1971 tty->cr();
1972 }
1973 #endif
1974 igvn.replace_input_of(unc, idx, top());
1975 modified = true;
1976 }
1977 }
1978 }
1979
1980 // keep the mondified trap for late query
1981 if (modified) {
1982 trap->set_modified();
1983 } else {
1984 _unstable_if_traps.delete_at(i);
1985 }
1986 }
1987 igvn.optimize();
1988 }
1989
1990 // StringOpts and late inlining of string methods
1991 void Compile::inline_string_calls(bool parse_time) {
1992 {
1993 // remove useless nodes to make the usage analysis simpler
1994 ResourceMark rm;
1995 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1996 }
1997
1998 {
1999 ResourceMark rm;
2000 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2001 PhaseStringOpts pso(initial_gvn());
2002 print_method(PHASE_AFTER_STRINGOPTS, 3);
2003 }
2004
2005 // now inline anything that we skipped the first time around
2006 if (!parse_time) {
2007 _late_inlines_pos = _late_inlines.length();
2008 }
2009
2010 while (_string_late_inlines.length() > 0) {
2011 CallGenerator* cg = _string_late_inlines.pop();
2012 cg->do_late_inline();
2013 if (failing()) return;
2014 }
2015 _string_late_inlines.trunc_to(0);
2016 }
2017
2018 // Late inlining of boxing methods
2019 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2020 if (_boxing_late_inlines.length() > 0) {
2021 assert(has_boxed_value(), "inconsistent");
2022
2023 set_inlining_incrementally(true);
2024
2025 igvn_worklist()->ensure_empty(); // should be done with igvn
2026
2027 _late_inlines_pos = _late_inlines.length();
2028
2029 while (_boxing_late_inlines.length() > 0) {
2030 CallGenerator* cg = _boxing_late_inlines.pop();
2031 cg->do_late_inline();
2032 if (failing()) return;
2033 }
2034 _boxing_late_inlines.trunc_to(0);
2035
2036 inline_incrementally_cleanup(igvn);
2037
2038 set_inlining_incrementally(false);
2039 }
2040 }
2041
2042 bool Compile::inline_incrementally_one() {
2043 assert(IncrementalInline, "incremental inlining should be on");
2044
2045 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2046
2047 set_inlining_progress(false);
2048 set_do_cleanup(false);
2049
2050 for (int i = 0; i < _late_inlines.length(); i++) {
2051 _late_inlines_pos = i+1;
2052 CallGenerator* cg = _late_inlines.at(i);
2053 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2054 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2055 cg->do_late_inline();
2056 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2057 if (failing()) {
2058 return false;
2059 } else if (inlining_progress()) {
2060 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2061 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2062 break; // process one call site at a time
2063 }
2064 } else {
2065 // Ignore late inline direct calls when inlining is not allowed.
2066 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2067 }
2068 }
2069 // Remove processed elements.
2070 _late_inlines.remove_till(_late_inlines_pos);
2071 _late_inlines_pos = 0;
2072
2073 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2074
2075 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2076
2077 set_inlining_progress(false);
2078 set_do_cleanup(false);
2079
2080 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2081 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2082 }
2083
2084 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2085 {
2086 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2087 ResourceMark rm;
2088 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2089 }
2090 {
2091 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2092 igvn.reset_from_gvn(initial_gvn());
2093 igvn.optimize();
2094 if (failing()) return;
2095 }
2096 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2097 }
2098
2099 // Perform incremental inlining until bound on number of live nodes is reached
2100 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2101 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2102
2103 set_inlining_incrementally(true);
2104 uint low_live_nodes = 0;
2105
2106 while (_late_inlines.length() > 0) {
2107 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2108 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2109 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2110 // PhaseIdealLoop is expensive so we only try it once we are
2111 // out of live nodes and we only try it again if the previous
2112 // helped got the number of nodes down significantly
2113 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2114 if (failing()) return;
2115 low_live_nodes = live_nodes();
2116 _major_progress = true;
2117 }
2118
2119 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2120 bool do_print_inlining = print_inlining() || print_intrinsics();
2121 if (do_print_inlining || log() != nullptr) {
2122 // Print inlining message for candidates that we couldn't inline for lack of space.
2123 for (int i = 0; i < _late_inlines.length(); i++) {
2124 CallGenerator* cg = _late_inlines.at(i);
2125 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2126 if (do_print_inlining) {
2127 cg->print_inlining_late(InliningResult::FAILURE, msg);
2128 }
2129 log_late_inline_failure(cg, msg);
2130 }
2131 }
2132 break; // finish
2133 }
2134 }
2135
2136 igvn_worklist()->ensure_empty(); // should be done with igvn
2137
2138 while (inline_incrementally_one()) {
2139 assert(!failing(), "inconsistent");
2140 }
2141 if (failing()) return;
2142
2143 inline_incrementally_cleanup(igvn);
2144
2145 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2146
2147 if (failing()) return;
2148
2149 if (_late_inlines.length() == 0) {
2150 break; // no more progress
2151 }
2152 }
2153
2154 igvn_worklist()->ensure_empty(); // should be done with igvn
2155
2156 if (_string_late_inlines.length() > 0) {
2157 assert(has_stringbuilder(), "inconsistent");
2158
2159 inline_string_calls(false);
2160
2161 if (failing()) return;
2162
2163 inline_incrementally_cleanup(igvn);
2164 }
2165
2166 set_inlining_incrementally(false);
2167 }
2168
2169 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2170 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2171 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2172 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2173 // as if "inlining_incrementally() == true" were set.
2174 assert(inlining_incrementally() == false, "not allowed");
2175 assert(_modified_nodes == nullptr, "not allowed");
2176 assert(_late_inlines.length() > 0, "sanity");
2177
2178 while (_late_inlines.length() > 0) {
2179 igvn_worklist()->ensure_empty(); // should be done with igvn
2180
2181 while (inline_incrementally_one()) {
2182 assert(!failing(), "inconsistent");
2183 }
2184 if (failing()) return;
2185
2186 inline_incrementally_cleanup(igvn);
2187 }
2188 }
2189
2190 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2191 if (_loop_opts_cnt > 0) {
2192 while (major_progress() && (_loop_opts_cnt > 0)) {
2193 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2194 PhaseIdealLoop::optimize(igvn, mode);
2195 _loop_opts_cnt--;
2196 if (failing()) return false;
2197 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2198 }
2199 }
2200 return true;
2201 }
2202
2203 // Remove edges from "root" to each SafePoint at a backward branch.
2204 // They were inserted during parsing (see add_safepoint()) to make
2205 // infinite loops without calls or exceptions visible to root, i.e.,
2206 // useful.
2207 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2208 Node *r = root();
2209 if (r != nullptr) {
2210 for (uint i = r->req(); i < r->len(); ++i) {
2211 Node *n = r->in(i);
2212 if (n != nullptr && n->is_SafePoint()) {
2213 r->rm_prec(i);
2214 if (n->outcnt() == 0) {
2215 igvn.remove_dead_node(n);
2216 }
2217 --i;
2218 }
2219 }
2220 // Parsing may have added top inputs to the root node (Path
2221 // leading to the Halt node proven dead). Make sure we get a
2222 // chance to clean them up.
2223 igvn._worklist.push(r);
2224 igvn.optimize();
2225 }
2226 }
2227
2228 //------------------------------Optimize---------------------------------------
2229 // Given a graph, optimize it.
2230 void Compile::Optimize() {
2231 TracePhase tp("optimizer", &timers[_t_optimizer]);
2232
2233 #ifndef PRODUCT
2234 if (env()->break_at_compile()) {
2235 BREAKPOINT;
2236 }
2237
2238 #endif
2239
2240 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2241 #ifdef ASSERT
2242 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2243 #endif
2244
2245 ResourceMark rm;
2246
2247 print_inlining_reinit();
2248
2249 NOT_PRODUCT( verify_graph_edges(); )
2250
2251 print_method(PHASE_AFTER_PARSING, 1);
2252
2253 {
2254 // Iterative Global Value Numbering, including ideal transforms
2255 // Initialize IterGVN with types and values from parse-time GVN
2256 PhaseIterGVN igvn(initial_gvn());
2257 #ifdef ASSERT
2258 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2259 #endif
2260 {
2261 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2262 igvn.optimize();
2263 }
2264
2265 if (failing()) return;
2266
2267 print_method(PHASE_ITER_GVN1, 2);
2268
2269 process_for_unstable_if_traps(igvn);
2270
2271 if (failing()) return;
2272
2273 inline_incrementally(igvn);
2274
2275 print_method(PHASE_INCREMENTAL_INLINE, 2);
2276
2277 if (failing()) return;
2278
2279 if (eliminate_boxing()) {
2280 // Inline valueOf() methods now.
2281 inline_boxing_calls(igvn);
2282
2283 if (failing()) return;
2284
2285 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2286 inline_incrementally(igvn);
2287 }
2288
2289 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2290
2291 if (failing()) return;
2292 }
2293
2294 // Remove the speculative part of types and clean up the graph from
2295 // the extra CastPP nodes whose only purpose is to carry them. Do
2296 // that early so that optimizations are not disrupted by the extra
2297 // CastPP nodes.
2298 remove_speculative_types(igvn);
2299
2300 if (failing()) return;
2301
2302 // No more new expensive nodes will be added to the list from here
2303 // so keep only the actual candidates for optimizations.
2304 cleanup_expensive_nodes(igvn);
2305
2306 if (failing()) return;
2307
2308 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2309 if (EnableVectorSupport && has_vbox_nodes()) {
2310 TracePhase tp("", &timers[_t_vector]);
2311 PhaseVector pv(igvn);
2312 pv.optimize_vector_boxes();
2313 if (failing()) return;
2314 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2315 }
2316 assert(!has_vbox_nodes(), "sanity");
2317
2318 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2319 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2320 igvn_worklist()->ensure_empty(); // should be done with igvn
2321 {
2322 ResourceMark rm;
2323 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2324 }
2325 igvn.reset_from_gvn(initial_gvn());
2326 igvn.optimize();
2327 if (failing()) return;
2328 }
2329
2330 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2331 // safepoints
2332 remove_root_to_sfpts_edges(igvn);
2333
2334 if (failing()) return;
2335
2336 // Perform escape analysis
2337 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2338 if (has_loops()) {
2339 // Cleanup graph (remove dead nodes).
2340 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2341 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2342 if (failing()) return;
2343 }
2344 bool progress;
2345 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2346 do {
2347 ConnectionGraph::do_analysis(this, &igvn);
2348
2349 if (failing()) return;
2350
2351 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2352
2353 // Optimize out fields loads from scalar replaceable allocations.
2354 igvn.optimize();
2355 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2356
2357 if (failing()) return;
2358
2359 if (congraph() != nullptr && macro_count() > 0) {
2360 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2361 PhaseMacroExpand mexp(igvn);
2362 mexp.eliminate_macro_nodes();
2363 if (failing()) return;
2364
2365 igvn.set_delay_transform(false);
2366 igvn.optimize();
2367 if (failing()) return;
2368
2369 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2370 }
2371
2372 ConnectionGraph::verify_ram_nodes(this, root());
2373 if (failing()) return;
2374
2375 progress = do_iterative_escape_analysis() &&
2376 (macro_count() < mcount) &&
2377 ConnectionGraph::has_candidates(this);
2378 // Try again if candidates exist and made progress
2379 // by removing some allocations and/or locks.
2380 } while (progress);
2381 }
2382
2383 // Loop transforms on the ideal graph. Range Check Elimination,
2384 // peeling, unrolling, etc.
2385
2386 // Set loop opts counter
2387 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2388 {
2389 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2390 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2391 _loop_opts_cnt--;
2392 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2393 if (failing()) return;
2394 }
2395 // Loop opts pass if partial peeling occurred in previous pass
2396 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2397 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2398 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2399 _loop_opts_cnt--;
2400 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2401 if (failing()) return;
2402 }
2403 // Loop opts pass for loop-unrolling before CCP
2404 if(major_progress() && (_loop_opts_cnt > 0)) {
2405 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2406 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2407 _loop_opts_cnt--;
2408 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2409 }
2410 if (!failing()) {
2411 // Verify that last round of loop opts produced a valid graph
2412 PhaseIdealLoop::verify(igvn);
2413 }
2414 }
2415 if (failing()) return;
2416
2417 // Conditional Constant Propagation;
2418 print_method(PHASE_BEFORE_CCP1, 2);
2419 PhaseCCP ccp( &igvn );
2420 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2421 {
2422 TracePhase tp("ccp", &timers[_t_ccp]);
2423 ccp.do_transform();
2424 }
2425 print_method(PHASE_CCP1, 2);
2426
2427 assert( true, "Break here to ccp.dump_old2new_map()");
2428
2429 // Iterative Global Value Numbering, including ideal transforms
2430 {
2431 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2432 igvn.reset_from_igvn(&ccp);
2433 igvn.optimize();
2434 }
2435 print_method(PHASE_ITER_GVN2, 2);
2436
2437 if (failing()) return;
2438
2439 // Loop transforms on the ideal graph. Range Check Elimination,
2440 // peeling, unrolling, etc.
2441 if (!optimize_loops(igvn, LoopOptsDefault)) {
2442 return;
2443 }
2444
2445 if (failing()) return;
2446
2447 C->clear_major_progress(); // ensure that major progress is now clear
2448
2449 process_for_post_loop_opts_igvn(igvn);
2450
2451 if (failing()) return;
2452
2453 #ifdef ASSERT
2454 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2455 #endif
2456
2457 {
2458 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2459 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2460 PhaseMacroExpand mex(igvn);
2461 if (mex.expand_macro_nodes()) {
2462 assert(failing(), "must bail out w/ explicit message");
2463 return;
2464 }
2465 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2466 }
2467
2468 {
2469 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2470 if (bs->expand_barriers(this, igvn)) {
2471 assert(failing(), "must bail out w/ explicit message");
2472 return;
2473 }
2474 print_method(PHASE_BARRIER_EXPANSION, 2);
2475 }
2476
2477 if (C->max_vector_size() > 0) {
2478 C->optimize_logic_cones(igvn);
2479 igvn.optimize();
2480 if (failing()) return;
2481 }
2482
2483 DEBUG_ONLY( _modified_nodes = nullptr; )
2484
2485 assert(igvn._worklist.size() == 0, "not empty");
2486
2487 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2488
2489 if (_late_inlines.length() > 0) {
2490 // More opportunities to optimize virtual and MH calls.
2491 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2492 process_late_inline_calls_no_inline(igvn);
2493 if (failing()) return;
2494 }
2495 } // (End scope of igvn; run destructor if necessary for asserts.)
2496
2497 check_no_dead_use();
2498
2499 process_print_inlining();
2500
2501 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2502 // to remove hashes to unlock nodes for modifications.
2503 C->node_hash()->clear();
2504
2505 // A method with only infinite loops has no edges entering loops from root
2506 {
2507 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2508 if (final_graph_reshaping()) {
2509 assert(failing(), "must bail out w/ explicit message");
2510 return;
2511 }
2512 }
2513
2514 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2515 DEBUG_ONLY(set_phase_optimize_finished();)
2516 }
2517
2518 #ifdef ASSERT
2519 void Compile::check_no_dead_use() const {
2520 ResourceMark rm;
2521 Unique_Node_List wq;
2522 wq.push(root());
2523 for (uint i = 0; i < wq.size(); ++i) {
2524 Node* n = wq.at(i);
2525 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2526 Node* u = n->fast_out(j);
2527 if (u->outcnt() == 0 && !u->is_Con()) {
2528 u->dump();
2529 fatal("no reachable node should have no use");
2530 }
2531 wq.push(u);
2532 }
2533 }
2534 }
2535 #endif
2536
2537 void Compile::inline_vector_reboxing_calls() {
2538 if (C->_vector_reboxing_late_inlines.length() > 0) {
2539 _late_inlines_pos = C->_late_inlines.length();
2540 while (_vector_reboxing_late_inlines.length() > 0) {
2541 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2542 cg->do_late_inline();
2543 if (failing()) return;
2544 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2545 }
2546 _vector_reboxing_late_inlines.trunc_to(0);
2547 }
2548 }
2549
2550 bool Compile::has_vbox_nodes() {
2551 if (C->_vector_reboxing_late_inlines.length() > 0) {
2552 return true;
2553 }
2554 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2555 Node * n = C->macro_node(macro_idx);
2556 assert(n->is_macro(), "only macro nodes expected here");
2557 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2558 return true;
2559 }
2560 }
2561 return false;
2562 }
2563
2564 //---------------------------- Bitwise operation packing optimization ---------------------------
2565
2566 static bool is_vector_unary_bitwise_op(Node* n) {
2567 return n->Opcode() == Op_XorV &&
2568 VectorNode::is_vector_bitwise_not_pattern(n);
2569 }
2570
2571 static bool is_vector_binary_bitwise_op(Node* n) {
2572 switch (n->Opcode()) {
2573 case Op_AndV:
2574 case Op_OrV:
2575 return true;
2576
2577 case Op_XorV:
2578 return !is_vector_unary_bitwise_op(n);
2579
2580 default:
2581 return false;
2582 }
2583 }
2584
2585 static bool is_vector_ternary_bitwise_op(Node* n) {
2586 return n->Opcode() == Op_MacroLogicV;
2587 }
2588
2589 static bool is_vector_bitwise_op(Node* n) {
2590 return is_vector_unary_bitwise_op(n) ||
2591 is_vector_binary_bitwise_op(n) ||
2592 is_vector_ternary_bitwise_op(n);
2593 }
2594
2595 static bool is_vector_bitwise_cone_root(Node* n) {
2596 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2597 return false;
2598 }
2599 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2600 if (is_vector_bitwise_op(n->fast_out(i))) {
2601 return false;
2602 }
2603 }
2604 return true;
2605 }
2606
2607 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2608 uint cnt = 0;
2609 if (is_vector_bitwise_op(n)) {
2610 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2611 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2612 for (uint i = 1; i < inp_cnt; i++) {
2613 Node* in = n->in(i);
2614 bool skip = VectorNode::is_all_ones_vector(in);
2615 if (!skip && !inputs.member(in)) {
2616 inputs.push(in);
2617 cnt++;
2618 }
2619 }
2620 assert(cnt <= 1, "not unary");
2621 } else {
2622 uint last_req = inp_cnt;
2623 if (is_vector_ternary_bitwise_op(n)) {
2624 last_req = inp_cnt - 1; // skip last input
2625 }
2626 for (uint i = 1; i < last_req; i++) {
2627 Node* def = n->in(i);
2628 if (!inputs.member(def)) {
2629 inputs.push(def);
2630 cnt++;
2631 }
2632 }
2633 }
2634 } else { // not a bitwise operations
2635 if (!inputs.member(n)) {
2636 inputs.push(n);
2637 cnt++;
2638 }
2639 }
2640 return cnt;
2641 }
2642
2643 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2644 Unique_Node_List useful_nodes;
2645 C->identify_useful_nodes(useful_nodes);
2646
2647 for (uint i = 0; i < useful_nodes.size(); i++) {
2648 Node* n = useful_nodes.at(i);
2649 if (is_vector_bitwise_cone_root(n)) {
2650 list.push(n);
2651 }
2652 }
2653 }
2654
2655 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2656 const TypeVect* vt,
2657 Unique_Node_List& partition,
2658 Unique_Node_List& inputs) {
2659 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2660 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2661 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2662
2663 Node* in1 = inputs.at(0);
2664 Node* in2 = inputs.at(1);
2665 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2666
2667 uint func = compute_truth_table(partition, inputs);
2668
2669 Node* pn = partition.at(partition.size() - 1);
2670 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2671 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2672 }
2673
2674 static uint extract_bit(uint func, uint pos) {
2675 return (func & (1 << pos)) >> pos;
2676 }
2677
2678 //
2679 // A macro logic node represents a truth table. It has 4 inputs,
2680 // First three inputs corresponds to 3 columns of a truth table
2681 // and fourth input captures the logic function.
2682 //
2683 // eg. fn = (in1 AND in2) OR in3;
2684 //
2685 // MacroNode(in1,in2,in3,fn)
2686 //
2687 // -----------------
2688 // in1 in2 in3 fn
2689 // -----------------
2690 // 0 0 0 0
2691 // 0 0 1 1
2692 // 0 1 0 0
2693 // 0 1 1 1
2694 // 1 0 0 0
2695 // 1 0 1 1
2696 // 1 1 0 1
2697 // 1 1 1 1
2698 //
2699
2700 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2701 int res = 0;
2702 for (int i = 0; i < 8; i++) {
2703 int bit1 = extract_bit(in1, i);
2704 int bit2 = extract_bit(in2, i);
2705 int bit3 = extract_bit(in3, i);
2706
2707 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2708 int func_bit = extract_bit(func, func_bit_pos);
2709
2710 res |= func_bit << i;
2711 }
2712 return res;
2713 }
2714
2715 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2716 assert(n != nullptr, "");
2717 assert(eval_map.contains(n), "absent");
2718 return *(eval_map.get(n));
2719 }
2720
2721 static void eval_operands(Node* n,
2722 uint& func1, uint& func2, uint& func3,
2723 ResourceHashtable<Node*,uint>& eval_map) {
2724 assert(is_vector_bitwise_op(n), "");
2725
2726 if (is_vector_unary_bitwise_op(n)) {
2727 Node* opnd = n->in(1);
2728 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2729 opnd = n->in(2);
2730 }
2731 func1 = eval_operand(opnd, eval_map);
2732 } else if (is_vector_binary_bitwise_op(n)) {
2733 func1 = eval_operand(n->in(1), eval_map);
2734 func2 = eval_operand(n->in(2), eval_map);
2735 } else {
2736 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2737 func1 = eval_operand(n->in(1), eval_map);
2738 func2 = eval_operand(n->in(2), eval_map);
2739 func3 = eval_operand(n->in(3), eval_map);
2740 }
2741 }
2742
2743 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2744 assert(inputs.size() <= 3, "sanity");
2745 ResourceMark rm;
2746 uint res = 0;
2747 ResourceHashtable<Node*,uint> eval_map;
2748
2749 // Populate precomputed functions for inputs.
2750 // Each input corresponds to one column of 3 input truth-table.
2751 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2752 0xCC, // (_, b, _) -> b
2753 0xF0 }; // (a, _, _) -> a
2754 for (uint i = 0; i < inputs.size(); i++) {
2755 eval_map.put(inputs.at(i), input_funcs[2-i]);
2756 }
2757
2758 for (uint i = 0; i < partition.size(); i++) {
2759 Node* n = partition.at(i);
2760
2761 uint func1 = 0, func2 = 0, func3 = 0;
2762 eval_operands(n, func1, func2, func3, eval_map);
2763
2764 switch (n->Opcode()) {
2765 case Op_OrV:
2766 assert(func3 == 0, "not binary");
2767 res = func1 | func2;
2768 break;
2769 case Op_AndV:
2770 assert(func3 == 0, "not binary");
2771 res = func1 & func2;
2772 break;
2773 case Op_XorV:
2774 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2775 assert(func2 == 0 && func3 == 0, "not unary");
2776 res = (~func1) & 0xFF;
2777 } else {
2778 assert(func3 == 0, "not binary");
2779 res = func1 ^ func2;
2780 }
2781 break;
2782 case Op_MacroLogicV:
2783 // Ordering of inputs may change during evaluation of sub-tree
2784 // containing MacroLogic node as a child node, thus a re-evaluation
2785 // makes sure that function is evaluated in context of current
2786 // inputs.
2787 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2788 break;
2789
2790 default: assert(false, "not supported: %s", n->Name());
2791 }
2792 assert(res <= 0xFF, "invalid");
2793 eval_map.put(n, res);
2794 }
2795 return res;
2796 }
2797
2798 // Criteria under which nodes gets packed into a macro logic node:-
2799 // 1) Parent and both child nodes are all unmasked or masked with
2800 // same predicates.
2801 // 2) Masked parent can be packed with left child if it is predicated
2802 // and both have same predicates.
2803 // 3) Masked parent can be packed with right child if its un-predicated
2804 // or has matching predication condition.
2805 // 4) An unmasked parent can be packed with an unmasked child.
2806 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2807 assert(partition.size() == 0, "not empty");
2808 assert(inputs.size() == 0, "not empty");
2809 if (is_vector_ternary_bitwise_op(n)) {
2810 return false;
2811 }
2812
2813 bool is_unary_op = is_vector_unary_bitwise_op(n);
2814 if (is_unary_op) {
2815 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2816 return false; // too few inputs
2817 }
2818
2819 bool pack_left_child = true;
2820 bool pack_right_child = true;
2821
2822 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2823 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2824
2825 int left_child_input_cnt = 0;
2826 int right_child_input_cnt = 0;
2827
2828 bool parent_is_predicated = n->is_predicated_vector();
2829 bool left_child_predicated = n->in(1)->is_predicated_vector();
2830 bool right_child_predicated = n->in(2)->is_predicated_vector();
2831
2832 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2833 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2834 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2835
2836 do {
2837 if (pack_left_child && left_child_LOP &&
2838 ((!parent_is_predicated && !left_child_predicated) ||
2839 ((parent_is_predicated && left_child_predicated &&
2840 parent_pred == left_child_pred)))) {
2841 partition.push(n->in(1));
2842 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2843 } else {
2844 inputs.push(n->in(1));
2845 left_child_input_cnt = 1;
2846 }
2847
2848 if (pack_right_child && right_child_LOP &&
2849 (!right_child_predicated ||
2850 (right_child_predicated && parent_is_predicated &&
2851 parent_pred == right_child_pred))) {
2852 partition.push(n->in(2));
2853 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2854 } else {
2855 inputs.push(n->in(2));
2856 right_child_input_cnt = 1;
2857 }
2858
2859 if (inputs.size() > 3) {
2860 assert(partition.size() > 0, "");
2861 inputs.clear();
2862 partition.clear();
2863 if (left_child_input_cnt > right_child_input_cnt) {
2864 pack_left_child = false;
2865 } else {
2866 pack_right_child = false;
2867 }
2868 } else {
2869 break;
2870 }
2871 } while(true);
2872
2873 if(partition.size()) {
2874 partition.push(n);
2875 }
2876
2877 return (partition.size() == 2 || partition.size() == 3) &&
2878 (inputs.size() == 2 || inputs.size() == 3);
2879 }
2880
2881 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2882 assert(is_vector_bitwise_op(n), "not a root");
2883
2884 visited.set(n->_idx);
2885
2886 // 1) Do a DFS walk over the logic cone.
2887 for (uint i = 1; i < n->req(); i++) {
2888 Node* in = n->in(i);
2889 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2890 process_logic_cone_root(igvn, in, visited);
2891 }
2892 }
2893
2894 // 2) Bottom up traversal: Merge node[s] with
2895 // the parent to form macro logic node.
2896 Unique_Node_List partition;
2897 Unique_Node_List inputs;
2898 if (compute_logic_cone(n, partition, inputs)) {
2899 const TypeVect* vt = n->bottom_type()->is_vect();
2900 Node* pn = partition.at(partition.size() - 1);
2901 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2902 if (mask == nullptr ||
2903 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2904 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2905 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2906 igvn.replace_node(n, macro_logic);
2907 }
2908 }
2909 }
2910
2911 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2912 ResourceMark rm;
2913 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2914 Unique_Node_List list;
2915 collect_logic_cone_roots(list);
2916
2917 while (list.size() > 0) {
2918 Node* n = list.pop();
2919 const TypeVect* vt = n->bottom_type()->is_vect();
2920 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2921 if (supported) {
2922 VectorSet visited(comp_arena());
2923 process_logic_cone_root(igvn, n, visited);
2924 }
2925 }
2926 }
2927 }
2928
2929 //------------------------------Code_Gen---------------------------------------
2930 // Given a graph, generate code for it
2931 void Compile::Code_Gen() {
2932 if (failing()) {
2933 return;
2934 }
2935
2936 // Perform instruction selection. You might think we could reclaim Matcher
2937 // memory PDQ, but actually the Matcher is used in generating spill code.
2938 // Internals of the Matcher (including some VectorSets) must remain live
2939 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2940 // set a bit in reclaimed memory.
2941
2942 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2943 // nodes. Mapping is only valid at the root of each matched subtree.
2944 NOT_PRODUCT( verify_graph_edges(); )
2945
2946 Matcher matcher;
2947 _matcher = &matcher;
2948 {
2949 TracePhase tp("matcher", &timers[_t_matcher]);
2950 matcher.match();
2951 if (failing()) {
2952 return;
2953 }
2954 }
2955 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2956 // nodes. Mapping is only valid at the root of each matched subtree.
2957 NOT_PRODUCT( verify_graph_edges(); )
2958
2959 // If you have too many nodes, or if matching has failed, bail out
2960 check_node_count(0, "out of nodes matching instructions");
2961 if (failing()) {
2962 return;
2963 }
2964
2965 print_method(PHASE_MATCHING, 2);
2966
2967 // Build a proper-looking CFG
2968 PhaseCFG cfg(node_arena(), root(), matcher);
2969 _cfg = &cfg;
2970 {
2971 TracePhase tp("scheduler", &timers[_t_scheduler]);
2972 bool success = cfg.do_global_code_motion();
2973 if (!success) {
2974 return;
2975 }
2976
2977 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2978 NOT_PRODUCT( verify_graph_edges(); )
2979 cfg.verify();
2980 }
2981
2982 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2983 _regalloc = ®alloc;
2984 {
2985 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2986 // Perform register allocation. After Chaitin, use-def chains are
2987 // no longer accurate (at spill code) and so must be ignored.
2988 // Node->LRG->reg mappings are still accurate.
2989 _regalloc->Register_Allocate();
2990
2991 // Bail out if the allocator builds too many nodes
2992 if (failing()) {
2993 return;
2994 }
2995
2996 print_method(PHASE_REGISTER_ALLOCATION, 2);
2997 }
2998
2999 // Prior to register allocation we kept empty basic blocks in case the
3000 // the allocator needed a place to spill. After register allocation we
3001 // are not adding any new instructions. If any basic block is empty, we
3002 // can now safely remove it.
3003 {
3004 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
3005 cfg.remove_empty_blocks();
3006 if (do_freq_based_layout()) {
3007 PhaseBlockLayout layout(cfg);
3008 } else {
3009 cfg.set_loop_alignment();
3010 }
3011 cfg.fixup_flow();
3012 cfg.remove_unreachable_blocks();
3013 cfg.verify_dominator_tree();
3014 print_method(PHASE_BLOCK_ORDERING, 3);
3015 }
3016
3017 // Apply peephole optimizations
3018 if( OptoPeephole ) {
3019 TracePhase tp("peephole", &timers[_t_peephole]);
3020 PhasePeephole peep( _regalloc, cfg);
3021 peep.do_transform();
3022 print_method(PHASE_PEEPHOLE, 3);
3023 }
3024
3025 // Do late expand if CPU requires this.
3026 if (Matcher::require_postalloc_expand) {
3027 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3028 cfg.postalloc_expand(_regalloc);
3029 print_method(PHASE_POSTALLOC_EXPAND, 3);
3030 }
3031
3032 // Convert Nodes to instruction bits in a buffer
3033 {
3034 TracePhase tp("output", &timers[_t_output]);
3035 PhaseOutput output;
3036 output.Output();
3037 if (failing()) return;
3038 output.install();
3039 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3040 }
3041
3042 // He's dead, Jim.
3043 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3044 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3045 }
3046
3047 //------------------------------Final_Reshape_Counts---------------------------
3048 // This class defines counters to help identify when a method
3049 // may/must be executed using hardware with only 24-bit precision.
3050 struct Final_Reshape_Counts : public StackObj {
3051 int _call_count; // count non-inlined 'common' calls
3052 int _float_count; // count float ops requiring 24-bit precision
3053 int _double_count; // count double ops requiring more precision
3054 int _java_call_count; // count non-inlined 'java' calls
3055 int _inner_loop_count; // count loops which need alignment
3056 VectorSet _visited; // Visitation flags
3057 Node_List _tests; // Set of IfNodes & PCTableNodes
3058
3059 Final_Reshape_Counts() :
3060 _call_count(0), _float_count(0), _double_count(0),
3061 _java_call_count(0), _inner_loop_count(0) { }
3062
3063 void inc_call_count () { _call_count ++; }
3064 void inc_float_count () { _float_count ++; }
3065 void inc_double_count() { _double_count++; }
3066 void inc_java_call_count() { _java_call_count++; }
3067 void inc_inner_loop_count() { _inner_loop_count++; }
3068
3069 int get_call_count () const { return _call_count ; }
3070 int get_float_count () const { return _float_count ; }
3071 int get_double_count() const { return _double_count; }
3072 int get_java_call_count() const { return _java_call_count; }
3073 int get_inner_loop_count() const { return _inner_loop_count; }
3074 };
3075
3076 // Eliminate trivially redundant StoreCMs and accumulate their
3077 // precedence edges.
3078 void Compile::eliminate_redundant_card_marks(Node* n) {
3079 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3080 if (n->in(MemNode::Address)->outcnt() > 1) {
3081 // There are multiple users of the same address so it might be
3082 // possible to eliminate some of the StoreCMs
3083 Node* mem = n->in(MemNode::Memory);
3084 Node* adr = n->in(MemNode::Address);
3085 Node* val = n->in(MemNode::ValueIn);
3086 Node* prev = n;
3087 bool done = false;
3088 // Walk the chain of StoreCMs eliminating ones that match. As
3089 // long as it's a chain of single users then the optimization is
3090 // safe. Eliminating partially redundant StoreCMs would require
3091 // cloning copies down the other paths.
3092 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3093 if (adr == mem->in(MemNode::Address) &&
3094 val == mem->in(MemNode::ValueIn)) {
3095 // redundant StoreCM
3096 if (mem->req() > MemNode::OopStore) {
3097 // Hasn't been processed by this code yet.
3098 n->add_prec(mem->in(MemNode::OopStore));
3099 } else {
3100 // Already converted to precedence edge
3101 for (uint i = mem->req(); i < mem->len(); i++) {
3102 // Accumulate any precedence edges
3103 if (mem->in(i) != nullptr) {
3104 n->add_prec(mem->in(i));
3105 }
3106 }
3107 // Everything above this point has been processed.
3108 done = true;
3109 }
3110 // Eliminate the previous StoreCM
3111 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3112 assert(mem->outcnt() == 0, "should be dead");
3113 mem->disconnect_inputs(this);
3114 } else {
3115 prev = mem;
3116 }
3117 mem = prev->in(MemNode::Memory);
3118 }
3119 }
3120 }
3121
3122 //------------------------------final_graph_reshaping_impl----------------------
3123 // Implement items 1-5 from final_graph_reshaping below.
3124 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3125
3126 if ( n->outcnt() == 0 ) return; // dead node
3127 uint nop = n->Opcode();
3128
3129 // Check for 2-input instruction with "last use" on right input.
3130 // Swap to left input. Implements item (2).
3131 if( n->req() == 3 && // two-input instruction
3132 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3133 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3134 n->in(2)->outcnt() == 1 &&// right use IS a last use
3135 !n->in(2)->is_Con() ) { // right use is not a constant
3136 // Check for commutative opcode
3137 switch( nop ) {
3138 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3139 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3140 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3141 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3142 case Op_AndL: case Op_XorL: case Op_OrL:
3143 case Op_AndI: case Op_XorI: case Op_OrI: {
3144 // Move "last use" input to left by swapping inputs
3145 n->swap_edges(1, 2);
3146 break;
3147 }
3148 default:
3149 break;
3150 }
3151 }
3152
3153 #ifdef ASSERT
3154 if( n->is_Mem() ) {
3155 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3156 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3157 // oop will be recorded in oop map if load crosses safepoint
3158 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3159 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3160 "raw memory operations should have control edge");
3161 }
3162 if (n->is_MemBar()) {
3163 MemBarNode* mb = n->as_MemBar();
3164 if (mb->trailing_store() || mb->trailing_load_store()) {
3165 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3166 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3167 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3168 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3169 } else if (mb->leading()) {
3170 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3171 }
3172 }
3173 #endif
3174 // Count FPU ops and common calls, implements item (3)
3175 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3176 if (!gc_handled) {
3177 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3178 }
3179
3180 // Collect CFG split points
3181 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3182 frc._tests.push(n);
3183 }
3184 }
3185
3186 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3187 switch( nop ) {
3188 // Count all float operations that may use FPU
3189 case Op_AddF:
3190 case Op_SubF:
3191 case Op_MulF:
3192 case Op_DivF:
3193 case Op_NegF:
3194 case Op_ModF:
3195 case Op_ConvI2F:
3196 case Op_ConF:
3197 case Op_CmpF:
3198 case Op_CmpF3:
3199 case Op_StoreF:
3200 case Op_LoadF:
3201 // case Op_ConvL2F: // longs are split into 32-bit halves
3202 frc.inc_float_count();
3203 break;
3204
3205 case Op_ConvF2D:
3206 case Op_ConvD2F:
3207 frc.inc_float_count();
3208 frc.inc_double_count();
3209 break;
3210
3211 // Count all double operations that may use FPU
3212 case Op_AddD:
3213 case Op_SubD:
3214 case Op_MulD:
3215 case Op_DivD:
3216 case Op_NegD:
3217 case Op_ModD:
3218 case Op_ConvI2D:
3219 case Op_ConvD2I:
3220 // case Op_ConvL2D: // handled by leaf call
3221 // case Op_ConvD2L: // handled by leaf call
3222 case Op_ConD:
3223 case Op_CmpD:
3224 case Op_CmpD3:
3225 case Op_StoreD:
3226 case Op_LoadD:
3227 case Op_LoadD_unaligned:
3228 frc.inc_double_count();
3229 break;
3230 case Op_Opaque1: // Remove Opaque Nodes before matching
3231 case Op_Opaque3:
3232 n->subsume_by(n->in(1), this);
3233 break;
3234 case Op_CallStaticJava:
3235 case Op_CallJava:
3236 case Op_CallDynamicJava:
3237 frc.inc_java_call_count(); // Count java call site;
3238 case Op_CallRuntime:
3239 case Op_CallLeaf:
3240 case Op_CallLeafVector:
3241 case Op_CallLeafNoFP: {
3242 assert (n->is_Call(), "");
3243 CallNode *call = n->as_Call();
3244 // Count call sites where the FP mode bit would have to be flipped.
3245 // Do not count uncommon runtime calls:
3246 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3247 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3248 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3249 frc.inc_call_count(); // Count the call site
3250 } else { // See if uncommon argument is shared
3251 Node *n = call->in(TypeFunc::Parms);
3252 int nop = n->Opcode();
3253 // Clone shared simple arguments to uncommon calls, item (1).
3254 if (n->outcnt() > 1 &&
3255 !n->is_Proj() &&
3256 nop != Op_CreateEx &&
3257 nop != Op_CheckCastPP &&
3258 nop != Op_DecodeN &&
3259 nop != Op_DecodeNKlass &&
3260 !n->is_Mem() &&
3261 !n->is_Phi()) {
3262 Node *x = n->clone();
3263 call->set_req(TypeFunc::Parms, x);
3264 }
3265 }
3266 break;
3267 }
3268
3269 case Op_StoreCM:
3270 {
3271 // Convert OopStore dependence into precedence edge
3272 Node* prec = n->in(MemNode::OopStore);
3273 n->del_req(MemNode::OopStore);
3274 n->add_prec(prec);
3275 eliminate_redundant_card_marks(n);
3276 }
3277
3278 // fall through
3279
3280 case Op_StoreB:
3281 case Op_StoreC:
3282 case Op_StoreI:
3283 case Op_StoreL:
3284 case Op_CompareAndSwapB:
3285 case Op_CompareAndSwapS:
3286 case Op_CompareAndSwapI:
3287 case Op_CompareAndSwapL:
3288 case Op_CompareAndSwapP:
3289 case Op_CompareAndSwapN:
3290 case Op_WeakCompareAndSwapB:
3291 case Op_WeakCompareAndSwapS:
3292 case Op_WeakCompareAndSwapI:
3293 case Op_WeakCompareAndSwapL:
3294 case Op_WeakCompareAndSwapP:
3295 case Op_WeakCompareAndSwapN:
3296 case Op_CompareAndExchangeB:
3297 case Op_CompareAndExchangeS:
3298 case Op_CompareAndExchangeI:
3299 case Op_CompareAndExchangeL:
3300 case Op_CompareAndExchangeP:
3301 case Op_CompareAndExchangeN:
3302 case Op_GetAndAddS:
3303 case Op_GetAndAddB:
3304 case Op_GetAndAddI:
3305 case Op_GetAndAddL:
3306 case Op_GetAndSetS:
3307 case Op_GetAndSetB:
3308 case Op_GetAndSetI:
3309 case Op_GetAndSetL:
3310 case Op_GetAndSetP:
3311 case Op_GetAndSetN:
3312 case Op_StoreP:
3313 case Op_StoreN:
3314 case Op_StoreNKlass:
3315 case Op_LoadB:
3316 case Op_LoadUB:
3317 case Op_LoadUS:
3318 case Op_LoadI:
3319 case Op_LoadKlass:
3320 case Op_LoadNKlass:
3321 case Op_LoadL:
3322 case Op_LoadL_unaligned:
3323 case Op_LoadP:
3324 case Op_LoadN:
3325 case Op_LoadRange:
3326 case Op_LoadS:
3327 break;
3328
3329 case Op_AddP: { // Assert sane base pointers
3330 Node *addp = n->in(AddPNode::Address);
3331 assert( !addp->is_AddP() ||
3332 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3333 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3334 "Base pointers must match (addp %u)", addp->_idx );
3335 #ifdef _LP64
3336 if ((UseCompressedOops || UseCompressedClassPointers) &&
3337 addp->Opcode() == Op_ConP &&
3338 addp == n->in(AddPNode::Base) &&
3339 n->in(AddPNode::Offset)->is_Con()) {
3340 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3341 // on the platform and on the compressed oops mode.
3342 // Use addressing with narrow klass to load with offset on x86.
3343 // Some platforms can use the constant pool to load ConP.
3344 // Do this transformation here since IGVN will convert ConN back to ConP.
3345 const Type* t = addp->bottom_type();
3346 bool is_oop = t->isa_oopptr() != nullptr;
3347 bool is_klass = t->isa_klassptr() != nullptr;
3348
3349 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3350 (is_klass && Matcher::const_klass_prefer_decode())) {
3351 Node* nn = nullptr;
3352
3353 int op = is_oop ? Op_ConN : Op_ConNKlass;
3354
3355 // Look for existing ConN node of the same exact type.
3356 Node* r = root();
3357 uint cnt = r->outcnt();
3358 for (uint i = 0; i < cnt; i++) {
3359 Node* m = r->raw_out(i);
3360 if (m!= nullptr && m->Opcode() == op &&
3361 m->bottom_type()->make_ptr() == t) {
3362 nn = m;
3363 break;
3364 }
3365 }
3366 if (nn != nullptr) {
3367 // Decode a narrow oop to match address
3368 // [R12 + narrow_oop_reg<<3 + offset]
3369 if (is_oop) {
3370 nn = new DecodeNNode(nn, t);
3371 } else {
3372 nn = new DecodeNKlassNode(nn, t);
3373 }
3374 // Check for succeeding AddP which uses the same Base.
3375 // Otherwise we will run into the assertion above when visiting that guy.
3376 for (uint i = 0; i < n->outcnt(); ++i) {
3377 Node *out_i = n->raw_out(i);
3378 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3379 out_i->set_req(AddPNode::Base, nn);
3380 #ifdef ASSERT
3381 for (uint j = 0; j < out_i->outcnt(); ++j) {
3382 Node *out_j = out_i->raw_out(j);
3383 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3384 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3385 }
3386 #endif
3387 }
3388 }
3389 n->set_req(AddPNode::Base, nn);
3390 n->set_req(AddPNode::Address, nn);
3391 if (addp->outcnt() == 0) {
3392 addp->disconnect_inputs(this);
3393 }
3394 }
3395 }
3396 }
3397 #endif
3398 break;
3399 }
3400
3401 case Op_CastPP: {
3402 // Remove CastPP nodes to gain more freedom during scheduling but
3403 // keep the dependency they encode as control or precedence edges
3404 // (if control is set already) on memory operations. Some CastPP
3405 // nodes don't have a control (don't carry a dependency): skip
3406 // those.
3407 if (n->in(0) != nullptr) {
3408 ResourceMark rm;
3409 Unique_Node_List wq;
3410 wq.push(n);
3411 for (uint next = 0; next < wq.size(); ++next) {
3412 Node *m = wq.at(next);
3413 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3414 Node* use = m->fast_out(i);
3415 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3416 use->ensure_control_or_add_prec(n->in(0));
3417 } else {
3418 switch(use->Opcode()) {
3419 case Op_AddP:
3420 case Op_DecodeN:
3421 case Op_DecodeNKlass:
3422 case Op_CheckCastPP:
3423 case Op_CastPP:
3424 wq.push(use);
3425 break;
3426 }
3427 }
3428 }
3429 }
3430 }
3431 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3432 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3433 Node* in1 = n->in(1);
3434 const Type* t = n->bottom_type();
3435 Node* new_in1 = in1->clone();
3436 new_in1->as_DecodeN()->set_type(t);
3437
3438 if (!Matcher::narrow_oop_use_complex_address()) {
3439 //
3440 // x86, ARM and friends can handle 2 adds in addressing mode
3441 // and Matcher can fold a DecodeN node into address by using
3442 // a narrow oop directly and do implicit null check in address:
3443 //
3444 // [R12 + narrow_oop_reg<<3 + offset]
3445 // NullCheck narrow_oop_reg
3446 //
3447 // On other platforms (Sparc) we have to keep new DecodeN node and
3448 // use it to do implicit null check in address:
3449 //
3450 // decode_not_null narrow_oop_reg, base_reg
3451 // [base_reg + offset]
3452 // NullCheck base_reg
3453 //
3454 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3455 // to keep the information to which null check the new DecodeN node
3456 // corresponds to use it as value in implicit_null_check().
3457 //
3458 new_in1->set_req(0, n->in(0));
3459 }
3460
3461 n->subsume_by(new_in1, this);
3462 if (in1->outcnt() == 0) {
3463 in1->disconnect_inputs(this);
3464 }
3465 } else {
3466 n->subsume_by(n->in(1), this);
3467 if (n->outcnt() == 0) {
3468 n->disconnect_inputs(this);
3469 }
3470 }
3471 break;
3472 }
3473 #ifdef _LP64
3474 case Op_CmpP:
3475 // Do this transformation here to preserve CmpPNode::sub() and
3476 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3477 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3478 Node* in1 = n->in(1);
3479 Node* in2 = n->in(2);
3480 if (!in1->is_DecodeNarrowPtr()) {
3481 in2 = in1;
3482 in1 = n->in(2);
3483 }
3484 assert(in1->is_DecodeNarrowPtr(), "sanity");
3485
3486 Node* new_in2 = nullptr;
3487 if (in2->is_DecodeNarrowPtr()) {
3488 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3489 new_in2 = in2->in(1);
3490 } else if (in2->Opcode() == Op_ConP) {
3491 const Type* t = in2->bottom_type();
3492 if (t == TypePtr::NULL_PTR) {
3493 assert(in1->is_DecodeN(), "compare klass to null?");
3494 // Don't convert CmpP null check into CmpN if compressed
3495 // oops implicit null check is not generated.
3496 // This will allow to generate normal oop implicit null check.
3497 if (Matcher::gen_narrow_oop_implicit_null_checks())
3498 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3499 //
3500 // This transformation together with CastPP transformation above
3501 // will generated code for implicit null checks for compressed oops.
3502 //
3503 // The original code after Optimize()
3504 //
3505 // LoadN memory, narrow_oop_reg
3506 // decode narrow_oop_reg, base_reg
3507 // CmpP base_reg, nullptr
3508 // CastPP base_reg // NotNull
3509 // Load [base_reg + offset], val_reg
3510 //
3511 // after these transformations will be
3512 //
3513 // LoadN memory, narrow_oop_reg
3514 // CmpN narrow_oop_reg, nullptr
3515 // decode_not_null narrow_oop_reg, base_reg
3516 // Load [base_reg + offset], val_reg
3517 //
3518 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3519 // since narrow oops can be used in debug info now (see the code in
3520 // final_graph_reshaping_walk()).
3521 //
3522 // At the end the code will be matched to
3523 // on x86:
3524 //
3525 // Load_narrow_oop memory, narrow_oop_reg
3526 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3527 // NullCheck narrow_oop_reg
3528 //
3529 // and on sparc:
3530 //
3531 // Load_narrow_oop memory, narrow_oop_reg
3532 // decode_not_null narrow_oop_reg, base_reg
3533 // Load [base_reg + offset], val_reg
3534 // NullCheck base_reg
3535 //
3536 } else if (t->isa_oopptr()) {
3537 new_in2 = ConNode::make(t->make_narrowoop());
3538 } else if (t->isa_klassptr()) {
3539 new_in2 = ConNode::make(t->make_narrowklass());
3540 }
3541 }
3542 if (new_in2 != nullptr) {
3543 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3544 n->subsume_by(cmpN, this);
3545 if (in1->outcnt() == 0) {
3546 in1->disconnect_inputs(this);
3547 }
3548 if (in2->outcnt() == 0) {
3549 in2->disconnect_inputs(this);
3550 }
3551 }
3552 }
3553 break;
3554
3555 case Op_DecodeN:
3556 case Op_DecodeNKlass:
3557 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3558 // DecodeN could be pinned when it can't be fold into
3559 // an address expression, see the code for Op_CastPP above.
3560 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3561 break;
3562
3563 case Op_EncodeP:
3564 case Op_EncodePKlass: {
3565 Node* in1 = n->in(1);
3566 if (in1->is_DecodeNarrowPtr()) {
3567 n->subsume_by(in1->in(1), this);
3568 } else if (in1->Opcode() == Op_ConP) {
3569 const Type* t = in1->bottom_type();
3570 if (t == TypePtr::NULL_PTR) {
3571 assert(t->isa_oopptr(), "null klass?");
3572 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3573 } else if (t->isa_oopptr()) {
3574 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3575 } else if (t->isa_klassptr()) {
3576 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3577 }
3578 }
3579 if (in1->outcnt() == 0) {
3580 in1->disconnect_inputs(this);
3581 }
3582 break;
3583 }
3584
3585 case Op_Proj: {
3586 if (OptimizeStringConcat || IncrementalInline) {
3587 ProjNode* proj = n->as_Proj();
3588 if (proj->_is_io_use) {
3589 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3590 // Separate projections were used for the exception path which
3591 // are normally removed by a late inline. If it wasn't inlined
3592 // then they will hang around and should just be replaced with
3593 // the original one. Merge them.
3594 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3595 if (non_io_proj != nullptr) {
3596 proj->subsume_by(non_io_proj , this);
3597 }
3598 }
3599 }
3600 break;
3601 }
3602
3603 case Op_Phi:
3604 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3605 // The EncodeP optimization may create Phi with the same edges
3606 // for all paths. It is not handled well by Register Allocator.
3607 Node* unique_in = n->in(1);
3608 assert(unique_in != nullptr, "");
3609 uint cnt = n->req();
3610 for (uint i = 2; i < cnt; i++) {
3611 Node* m = n->in(i);
3612 assert(m != nullptr, "");
3613 if (unique_in != m)
3614 unique_in = nullptr;
3615 }
3616 if (unique_in != nullptr) {
3617 n->subsume_by(unique_in, this);
3618 }
3619 }
3620 break;
3621
3622 #endif
3623
3624 #ifdef ASSERT
3625 case Op_CastII:
3626 // Verify that all range check dependent CastII nodes were removed.
3627 if (n->isa_CastII()->has_range_check()) {
3628 n->dump(3);
3629 assert(false, "Range check dependent CastII node was not removed");
3630 }
3631 break;
3632 #endif
3633
3634 case Op_ModI:
3635 if (UseDivMod) {
3636 // Check if a%b and a/b both exist
3637 Node* d = n->find_similar(Op_DivI);
3638 if (d) {
3639 // Replace them with a fused divmod if supported
3640 if (Matcher::has_match_rule(Op_DivModI)) {
3641 DivModINode* divmod = DivModINode::make(n);
3642 d->subsume_by(divmod->div_proj(), this);
3643 n->subsume_by(divmod->mod_proj(), this);
3644 } else {
3645 // replace a%b with a-((a/b)*b)
3646 Node* mult = new MulINode(d, d->in(2));
3647 Node* sub = new SubINode(d->in(1), mult);
3648 n->subsume_by(sub, this);
3649 }
3650 }
3651 }
3652 break;
3653
3654 case Op_ModL:
3655 if (UseDivMod) {
3656 // Check if a%b and a/b both exist
3657 Node* d = n->find_similar(Op_DivL);
3658 if (d) {
3659 // Replace them with a fused divmod if supported
3660 if (Matcher::has_match_rule(Op_DivModL)) {
3661 DivModLNode* divmod = DivModLNode::make(n);
3662 d->subsume_by(divmod->div_proj(), this);
3663 n->subsume_by(divmod->mod_proj(), this);
3664 } else {
3665 // replace a%b with a-((a/b)*b)
3666 Node* mult = new MulLNode(d, d->in(2));
3667 Node* sub = new SubLNode(d->in(1), mult);
3668 n->subsume_by(sub, this);
3669 }
3670 }
3671 }
3672 break;
3673
3674 case Op_UModI:
3675 if (UseDivMod) {
3676 // Check if a%b and a/b both exist
3677 Node* d = n->find_similar(Op_UDivI);
3678 if (d) {
3679 // Replace them with a fused unsigned divmod if supported
3680 if (Matcher::has_match_rule(Op_UDivModI)) {
3681 UDivModINode* divmod = UDivModINode::make(n);
3682 d->subsume_by(divmod->div_proj(), this);
3683 n->subsume_by(divmod->mod_proj(), this);
3684 } else {
3685 // replace a%b with a-((a/b)*b)
3686 Node* mult = new MulINode(d, d->in(2));
3687 Node* sub = new SubINode(d->in(1), mult);
3688 n->subsume_by(sub, this);
3689 }
3690 }
3691 }
3692 break;
3693
3694 case Op_UModL:
3695 if (UseDivMod) {
3696 // Check if a%b and a/b both exist
3697 Node* d = n->find_similar(Op_UDivL);
3698 if (d) {
3699 // Replace them with a fused unsigned divmod if supported
3700 if (Matcher::has_match_rule(Op_UDivModL)) {
3701 UDivModLNode* divmod = UDivModLNode::make(n);
3702 d->subsume_by(divmod->div_proj(), this);
3703 n->subsume_by(divmod->mod_proj(), this);
3704 } else {
3705 // replace a%b with a-((a/b)*b)
3706 Node* mult = new MulLNode(d, d->in(2));
3707 Node* sub = new SubLNode(d->in(1), mult);
3708 n->subsume_by(sub, this);
3709 }
3710 }
3711 }
3712 break;
3713
3714 case Op_LoadVector:
3715 case Op_StoreVector:
3716 #ifdef ASSERT
3717 // Add VerifyVectorAlignment node between adr and load / store.
3718 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3719 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3720 n->as_StoreVector()->must_verify_alignment();
3721 if (must_verify_alignment) {
3722 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3723 n->as_StoreVector()->memory_size();
3724 // The memory access should be aligned to the vector width in bytes.
3725 // However, the underlying array is possibly less well aligned, but at least
3726 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3727 // a loop we can expect at least the following alignment:
3728 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3729 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3730 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3731 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3732 jlong mask = guaranteed_alignment - 1;
3733 Node* mask_con = ConLNode::make(mask);
3734 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3735 n->set_req(MemNode::Address, va);
3736 }
3737 }
3738 #endif
3739 break;
3740
3741 case Op_LoadVectorGather:
3742 case Op_StoreVectorScatter:
3743 case Op_LoadVectorGatherMasked:
3744 case Op_StoreVectorScatterMasked:
3745 case Op_VectorCmpMasked:
3746 case Op_VectorMaskGen:
3747 case Op_LoadVectorMasked:
3748 case Op_StoreVectorMasked:
3749 break;
3750
3751 case Op_AddReductionVI:
3752 case Op_AddReductionVL:
3753 case Op_AddReductionVF:
3754 case Op_AddReductionVD:
3755 case Op_MulReductionVI:
3756 case Op_MulReductionVL:
3757 case Op_MulReductionVF:
3758 case Op_MulReductionVD:
3759 case Op_MinReductionV:
3760 case Op_MaxReductionV:
3761 case Op_AndReductionV:
3762 case Op_OrReductionV:
3763 case Op_XorReductionV:
3764 break;
3765
3766 case Op_PackB:
3767 case Op_PackS:
3768 case Op_PackI:
3769 case Op_PackF:
3770 case Op_PackL:
3771 case Op_PackD:
3772 if (n->req()-1 > 2) {
3773 // Replace many operand PackNodes with a binary tree for matching
3774 PackNode* p = (PackNode*) n;
3775 Node* btp = p->binary_tree_pack(1, n->req());
3776 n->subsume_by(btp, this);
3777 }
3778 break;
3779 case Op_Loop:
3780 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3781 case Op_CountedLoop:
3782 case Op_LongCountedLoop:
3783 case Op_OuterStripMinedLoop:
3784 if (n->as_Loop()->is_inner_loop()) {
3785 frc.inc_inner_loop_count();
3786 }
3787 n->as_Loop()->verify_strip_mined(0);
3788 break;
3789 case Op_LShiftI:
3790 case Op_RShiftI:
3791 case Op_URShiftI:
3792 case Op_LShiftL:
3793 case Op_RShiftL:
3794 case Op_URShiftL:
3795 if (Matcher::need_masked_shift_count) {
3796 // The cpu's shift instructions don't restrict the count to the
3797 // lower 5/6 bits. We need to do the masking ourselves.
3798 Node* in2 = n->in(2);
3799 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3800 const TypeInt* t = in2->find_int_type();
3801 if (t != nullptr && t->is_con()) {
3802 juint shift = t->get_con();
3803 if (shift > mask) { // Unsigned cmp
3804 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3805 }
3806 } else {
3807 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3808 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3809 n->set_req(2, shift);
3810 }
3811 }
3812 if (in2->outcnt() == 0) { // Remove dead node
3813 in2->disconnect_inputs(this);
3814 }
3815 }
3816 break;
3817 case Op_MemBarStoreStore:
3818 case Op_MemBarRelease:
3819 // Break the link with AllocateNode: it is no longer useful and
3820 // confuses register allocation.
3821 if (n->req() > MemBarNode::Precedent) {
3822 n->set_req(MemBarNode::Precedent, top());
3823 }
3824 break;
3825 case Op_MemBarAcquire: {
3826 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3827 // At parse time, the trailing MemBarAcquire for a volatile load
3828 // is created with an edge to the load. After optimizations,
3829 // that input may be a chain of Phis. If those phis have no
3830 // other use, then the MemBarAcquire keeps them alive and
3831 // register allocation can be confused.
3832 dead_nodes.push(n->in(MemBarNode::Precedent));
3833 n->set_req(MemBarNode::Precedent, top());
3834 }
3835 break;
3836 }
3837 case Op_Blackhole:
3838 break;
3839 case Op_RangeCheck: {
3840 RangeCheckNode* rc = n->as_RangeCheck();
3841 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3842 n->subsume_by(iff, this);
3843 frc._tests.push(iff);
3844 break;
3845 }
3846 case Op_ConvI2L: {
3847 if (!Matcher::convi2l_type_required) {
3848 // Code generation on some platforms doesn't need accurate
3849 // ConvI2L types. Widening the type can help remove redundant
3850 // address computations.
3851 n->as_Type()->set_type(TypeLong::INT);
3852 ResourceMark rm;
3853 Unique_Node_List wq;
3854 wq.push(n);
3855 for (uint next = 0; next < wq.size(); next++) {
3856 Node *m = wq.at(next);
3857
3858 for(;;) {
3859 // Loop over all nodes with identical inputs edges as m
3860 Node* k = m->find_similar(m->Opcode());
3861 if (k == nullptr) {
3862 break;
3863 }
3864 // Push their uses so we get a chance to remove node made
3865 // redundant
3866 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3867 Node* u = k->fast_out(i);
3868 if (u->Opcode() == Op_LShiftL ||
3869 u->Opcode() == Op_AddL ||
3870 u->Opcode() == Op_SubL ||
3871 u->Opcode() == Op_AddP) {
3872 wq.push(u);
3873 }
3874 }
3875 // Replace all nodes with identical edges as m with m
3876 k->subsume_by(m, this);
3877 }
3878 }
3879 }
3880 break;
3881 }
3882 case Op_CmpUL: {
3883 if (!Matcher::has_match_rule(Op_CmpUL)) {
3884 // No support for unsigned long comparisons
3885 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3886 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3887 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3888 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3889 Node* andl = new AndLNode(orl, remove_sign_mask);
3890 Node* cmp = new CmpLNode(andl, n->in(2));
3891 n->subsume_by(cmp, this);
3892 }
3893 break;
3894 }
3895 default:
3896 assert(!n->is_Call(), "");
3897 assert(!n->is_Mem(), "");
3898 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3899 break;
3900 }
3901 }
3902
3903 //------------------------------final_graph_reshaping_walk---------------------
3904 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3905 // requires that the walk visits a node's inputs before visiting the node.
3906 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3907 Unique_Node_List sfpt;
3908
3909 frc._visited.set(root->_idx); // first, mark node as visited
3910 uint cnt = root->req();
3911 Node *n = root;
3912 uint i = 0;
3913 while (true) {
3914 if (i < cnt) {
3915 // Place all non-visited non-null inputs onto stack
3916 Node* m = n->in(i);
3917 ++i;
3918 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3919 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3920 // compute worst case interpreter size in case of a deoptimization
3921 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3922
3923 sfpt.push(m);
3924 }
3925 cnt = m->req();
3926 nstack.push(n, i); // put on stack parent and next input's index
3927 n = m;
3928 i = 0;
3929 }
3930 } else {
3931 // Now do post-visit work
3932 final_graph_reshaping_impl(n, frc, dead_nodes);
3933 if (nstack.is_empty())
3934 break; // finished
3935 n = nstack.node(); // Get node from stack
3936 cnt = n->req();
3937 i = nstack.index();
3938 nstack.pop(); // Shift to the next node on stack
3939 }
3940 }
3941
3942 // Skip next transformation if compressed oops are not used.
3943 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3944 (!UseCompressedOops && !UseCompressedClassPointers))
3945 return;
3946
3947 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3948 // It could be done for an uncommon traps or any safepoints/calls
3949 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3950 while (sfpt.size() > 0) {
3951 n = sfpt.pop();
3952 JVMState *jvms = n->as_SafePoint()->jvms();
3953 assert(jvms != nullptr, "sanity");
3954 int start = jvms->debug_start();
3955 int end = n->req();
3956 bool is_uncommon = (n->is_CallStaticJava() &&
3957 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3958 for (int j = start; j < end; j++) {
3959 Node* in = n->in(j);
3960 if (in->is_DecodeNarrowPtr()) {
3961 bool safe_to_skip = true;
3962 if (!is_uncommon ) {
3963 // Is it safe to skip?
3964 for (uint i = 0; i < in->outcnt(); i++) {
3965 Node* u = in->raw_out(i);
3966 if (!u->is_SafePoint() ||
3967 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3968 safe_to_skip = false;
3969 }
3970 }
3971 }
3972 if (safe_to_skip) {
3973 n->set_req(j, in->in(1));
3974 }
3975 if (in->outcnt() == 0) {
3976 in->disconnect_inputs(this);
3977 }
3978 }
3979 }
3980 }
3981 }
3982
3983 //------------------------------final_graph_reshaping--------------------------
3984 // Final Graph Reshaping.
3985 //
3986 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3987 // and not commoned up and forced early. Must come after regular
3988 // optimizations to avoid GVN undoing the cloning. Clone constant
3989 // inputs to Loop Phis; these will be split by the allocator anyways.
3990 // Remove Opaque nodes.
3991 // (2) Move last-uses by commutative operations to the left input to encourage
3992 // Intel update-in-place two-address operations and better register usage
3993 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3994 // calls canonicalizing them back.
3995 // (3) Count the number of double-precision FP ops, single-precision FP ops
3996 // and call sites. On Intel, we can get correct rounding either by
3997 // forcing singles to memory (requires extra stores and loads after each
3998 // FP bytecode) or we can set a rounding mode bit (requires setting and
3999 // clearing the mode bit around call sites). The mode bit is only used
4000 // if the relative frequency of single FP ops to calls is low enough.
4001 // This is a key transform for SPEC mpeg_audio.
4002 // (4) Detect infinite loops; blobs of code reachable from above but not
4003 // below. Several of the Code_Gen algorithms fail on such code shapes,
4004 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4005 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4006 // Detection is by looking for IfNodes where only 1 projection is
4007 // reachable from below or CatchNodes missing some targets.
4008 // (5) Assert for insane oop offsets in debug mode.
4009
4010 bool Compile::final_graph_reshaping() {
4011 // an infinite loop may have been eliminated by the optimizer,
4012 // in which case the graph will be empty.
4013 if (root()->req() == 1) {
4014 // Do not compile method that is only a trivial infinite loop,
4015 // since the content of the loop may have been eliminated.
4016 record_method_not_compilable("trivial infinite loop");
4017 return true;
4018 }
4019
4020 // Expensive nodes have their control input set to prevent the GVN
4021 // from freely commoning them. There's no GVN beyond this point so
4022 // no need to keep the control input. We want the expensive nodes to
4023 // be freely moved to the least frequent code path by gcm.
4024 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4025 for (int i = 0; i < expensive_count(); i++) {
4026 _expensive_nodes.at(i)->set_req(0, nullptr);
4027 }
4028
4029 Final_Reshape_Counts frc;
4030
4031 // Visit everybody reachable!
4032 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4033 Node_Stack nstack(live_nodes() >> 1);
4034 Unique_Node_List dead_nodes;
4035 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4036
4037 // Check for unreachable (from below) code (i.e., infinite loops).
4038 for( uint i = 0; i < frc._tests.size(); i++ ) {
4039 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4040 // Get number of CFG targets.
4041 // Note that PCTables include exception targets after calls.
4042 uint required_outcnt = n->required_outcnt();
4043 if (n->outcnt() != required_outcnt) {
4044 // Check for a few special cases. Rethrow Nodes never take the
4045 // 'fall-thru' path, so expected kids is 1 less.
4046 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4047 if (n->in(0)->in(0)->is_Call()) {
4048 CallNode* call = n->in(0)->in(0)->as_Call();
4049 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4050 required_outcnt--; // Rethrow always has 1 less kid
4051 } else if (call->req() > TypeFunc::Parms &&
4052 call->is_CallDynamicJava()) {
4053 // Check for null receiver. In such case, the optimizer has
4054 // detected that the virtual call will always result in a null
4055 // pointer exception. The fall-through projection of this CatchNode
4056 // will not be populated.
4057 Node* arg0 = call->in(TypeFunc::Parms);
4058 if (arg0->is_Type() &&
4059 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4060 required_outcnt--;
4061 }
4062 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4063 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4064 // Check for illegal array length. In such case, the optimizer has
4065 // detected that the allocation attempt will always result in an
4066 // exception. There is no fall-through projection of this CatchNode .
4067 assert(call->is_CallStaticJava(), "static call expected");
4068 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4069 uint valid_length_test_input = call->req() - 1;
4070 Node* valid_length_test = call->in(valid_length_test_input);
4071 call->del_req(valid_length_test_input);
4072 if (valid_length_test->find_int_con(1) == 0) {
4073 required_outcnt--;
4074 }
4075 dead_nodes.push(valid_length_test);
4076 assert(n->outcnt() == required_outcnt, "malformed control flow");
4077 continue;
4078 }
4079 }
4080 }
4081
4082 // Recheck with a better notion of 'required_outcnt'
4083 if (n->outcnt() != required_outcnt) {
4084 record_method_not_compilable("malformed control flow");
4085 return true; // Not all targets reachable!
4086 }
4087 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4088 CallNode* call = n->in(0)->in(0)->as_Call();
4089 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4090 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4091 assert(call->is_CallStaticJava(), "static call expected");
4092 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4093 uint valid_length_test_input = call->req() - 1;
4094 dead_nodes.push(call->in(valid_length_test_input));
4095 call->del_req(valid_length_test_input); // valid length test useless now
4096 }
4097 }
4098 // Check that I actually visited all kids. Unreached kids
4099 // must be infinite loops.
4100 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4101 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4102 record_method_not_compilable("infinite loop");
4103 return true; // Found unvisited kid; must be unreach
4104 }
4105
4106 // Here so verification code in final_graph_reshaping_walk()
4107 // always see an OuterStripMinedLoopEnd
4108 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4109 IfNode* init_iff = n->as_If();
4110 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4111 n->subsume_by(iff, this);
4112 }
4113 }
4114
4115 while (dead_nodes.size() > 0) {
4116 Node* m = dead_nodes.pop();
4117 if (m->outcnt() == 0 && m != top()) {
4118 for (uint j = 0; j < m->req(); j++) {
4119 Node* in = m->in(j);
4120 if (in != nullptr) {
4121 dead_nodes.push(in);
4122 }
4123 }
4124 m->disconnect_inputs(this);
4125 }
4126 }
4127
4128 #ifdef IA32
4129 // If original bytecodes contained a mixture of floats and doubles
4130 // check if the optimizer has made it homogeneous, item (3).
4131 if (UseSSE == 0 &&
4132 frc.get_float_count() > 32 &&
4133 frc.get_double_count() == 0 &&
4134 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4135 set_24_bit_selection_and_mode(false, true);
4136 }
4137 #endif // IA32
4138
4139 set_java_calls(frc.get_java_call_count());
4140 set_inner_loops(frc.get_inner_loop_count());
4141
4142 // No infinite loops, no reason to bail out.
4143 return false;
4144 }
4145
4146 //-----------------------------too_many_traps----------------------------------
4147 // Report if there are too many traps at the current method and bci.
4148 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4149 bool Compile::too_many_traps(ciMethod* method,
4150 int bci,
4151 Deoptimization::DeoptReason reason) {
4152 if (method->has_trap_at(bci)) {
4153 return true;
4154 }
4155 ciMethodData* md = method->method_data();
4156 if (md->is_empty()) {
4157 // Assume the trap has not occurred, or that it occurred only
4158 // because of a transient condition during start-up in the interpreter.
4159 return false;
4160 }
4161 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4162 if (md->has_trap_at(bci, m, reason) != 0) {
4163 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4164 // Also, if there are multiple reasons, or if there is no per-BCI record,
4165 // assume the worst.
4166 if (log())
4167 log()->elem("observe trap='%s' count='%d'",
4168 Deoptimization::trap_reason_name(reason),
4169 md->trap_count(reason));
4170 return true;
4171 } else {
4172 // Ignore method/bci and see if there have been too many globally.
4173 return too_many_traps(reason, md);
4174 }
4175 }
4176
4177 // Less-accurate variant which does not require a method and bci.
4178 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4179 ciMethodData* logmd) {
4180 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4181 // Too many traps globally.
4182 // Note that we use cumulative trap_count, not just md->trap_count.
4183 if (log()) {
4184 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4185 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4186 Deoptimization::trap_reason_name(reason),
4187 mcount, trap_count(reason));
4188 }
4189 return true;
4190 } else {
4191 // The coast is clear.
4192 return false;
4193 }
4194 }
4195
4196 //--------------------------too_many_recompiles--------------------------------
4197 // Report if there are too many recompiles at the current method and bci.
4198 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4199 // Is not eager to return true, since this will cause the compiler to use
4200 // Action_none for a trap point, to avoid too many recompilations.
4201 bool Compile::too_many_recompiles(ciMethod* method,
4202 int bci,
4203 Deoptimization::DeoptReason reason) {
4204 ciMethodData* md = method->method_data();
4205 if (md->is_empty()) {
4206 // Assume the trap has not occurred, or that it occurred only
4207 // because of a transient condition during start-up in the interpreter.
4208 return false;
4209 }
4210 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4211 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4212 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4213 Deoptimization::DeoptReason per_bc_reason
4214 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4215 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4216 if ((per_bc_reason == Deoptimization::Reason_none
4217 || md->has_trap_at(bci, m, reason) != 0)
4218 // The trap frequency measure we care about is the recompile count:
4219 && md->trap_recompiled_at(bci, m)
4220 && md->overflow_recompile_count() >= bc_cutoff) {
4221 // Do not emit a trap here if it has already caused recompilations.
4222 // Also, if there are multiple reasons, or if there is no per-BCI record,
4223 // assume the worst.
4224 if (log())
4225 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4226 Deoptimization::trap_reason_name(reason),
4227 md->trap_count(reason),
4228 md->overflow_recompile_count());
4229 return true;
4230 } else if (trap_count(reason) != 0
4231 && decompile_count() >= m_cutoff) {
4232 // Too many recompiles globally, and we have seen this sort of trap.
4233 // Use cumulative decompile_count, not just md->decompile_count.
4234 if (log())
4235 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4236 Deoptimization::trap_reason_name(reason),
4237 md->trap_count(reason), trap_count(reason),
4238 md->decompile_count(), decompile_count());
4239 return true;
4240 } else {
4241 // The coast is clear.
4242 return false;
4243 }
4244 }
4245
4246 // Compute when not to trap. Used by matching trap based nodes and
4247 // NullCheck optimization.
4248 void Compile::set_allowed_deopt_reasons() {
4249 _allowed_reasons = 0;
4250 if (is_method_compilation()) {
4251 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4252 assert(rs < BitsPerInt, "recode bit map");
4253 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4254 _allowed_reasons |= nth_bit(rs);
4255 }
4256 }
4257 }
4258 }
4259
4260 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4261 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4262 }
4263
4264 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4265 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4266 }
4267
4268 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4269 if (holder->is_initialized() && !do_clinit_barriers()) {
4270 return false;
4271 }
4272 if (holder->is_being_initialized() || do_clinit_barriers()) {
4273 if (accessing_method->holder() == holder) {
4274 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4275 // <init>, or a static method. In all those cases, there was an initialization
4276 // barrier on the holder klass passed.
4277 if (accessing_method->is_static_initializer() ||
4278 accessing_method->is_object_initializer() ||
4279 accessing_method->is_static()) {
4280 return false;
4281 }
4282 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4283 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4284 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4285 // child class can become fully initialized while its parent class is still being initialized.
4286 if (accessing_method->is_static_initializer()) {
4287 return false;
4288 }
4289 }
4290 ciMethod* root = method(); // the root method of compilation
4291 if (root != accessing_method) {
4292 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4293 }
4294 }
4295 return true;
4296 }
4297
4298 #ifndef PRODUCT
4299 //------------------------------verify_bidirectional_edges---------------------
4300 // For each input edge to a node (ie - for each Use-Def edge), verify that
4301 // there is a corresponding Def-Use edge.
4302 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4303 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4304 uint stack_size = live_nodes() >> 4;
4305 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4306 nstack.push(_root);
4307
4308 while (nstack.size() > 0) {
4309 Node* n = nstack.pop();
4310 if (visited.member(n)) {
4311 continue;
4312 }
4313 visited.push(n);
4314
4315 // Walk over all input edges, checking for correspondence
4316 uint length = n->len();
4317 for (uint i = 0; i < length; i++) {
4318 Node* in = n->in(i);
4319 if (in != nullptr && !visited.member(in)) {
4320 nstack.push(in); // Put it on stack
4321 }
4322 if (in != nullptr && !in->is_top()) {
4323 // Count instances of `next`
4324 int cnt = 0;
4325 for (uint idx = 0; idx < in->_outcnt; idx++) {
4326 if (in->_out[idx] == n) {
4327 cnt++;
4328 }
4329 }
4330 assert(cnt > 0, "Failed to find Def-Use edge.");
4331 // Check for duplicate edges
4332 // walk the input array downcounting the input edges to n
4333 for (uint j = 0; j < length; j++) {
4334 if (n->in(j) == in) {
4335 cnt--;
4336 }
4337 }
4338 assert(cnt == 0, "Mismatched edge count.");
4339 } else if (in == nullptr) {
4340 assert(i == 0 || i >= n->req() ||
4341 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4342 (n->is_Unlock() && i == (n->req() - 1)) ||
4343 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4344 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4345 } else {
4346 assert(in->is_top(), "sanity");
4347 // Nothing to check.
4348 }
4349 }
4350 }
4351 }
4352
4353 //------------------------------verify_graph_edges---------------------------
4354 // Walk the Graph and verify that there is a one-to-one correspondence
4355 // between Use-Def edges and Def-Use edges in the graph.
4356 void Compile::verify_graph_edges(bool no_dead_code) {
4357 if (VerifyGraphEdges) {
4358 Unique_Node_List visited;
4359
4360 // Call graph walk to check edges
4361 verify_bidirectional_edges(visited);
4362 if (no_dead_code) {
4363 // Now make sure that no visited node is used by an unvisited node.
4364 bool dead_nodes = false;
4365 Unique_Node_List checked;
4366 while (visited.size() > 0) {
4367 Node* n = visited.pop();
4368 checked.push(n);
4369 for (uint i = 0; i < n->outcnt(); i++) {
4370 Node* use = n->raw_out(i);
4371 if (checked.member(use)) continue; // already checked
4372 if (visited.member(use)) continue; // already in the graph
4373 if (use->is_Con()) continue; // a dead ConNode is OK
4374 // At this point, we have found a dead node which is DU-reachable.
4375 if (!dead_nodes) {
4376 tty->print_cr("*** Dead nodes reachable via DU edges:");
4377 dead_nodes = true;
4378 }
4379 use->dump(2);
4380 tty->print_cr("---");
4381 checked.push(use); // No repeats; pretend it is now checked.
4382 }
4383 }
4384 assert(!dead_nodes, "using nodes must be reachable from root");
4385 }
4386 }
4387 }
4388 #endif
4389
4390 // The Compile object keeps track of failure reasons separately from the ciEnv.
4391 // This is required because there is not quite a 1-1 relation between the
4392 // ciEnv and its compilation task and the Compile object. Note that one
4393 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4394 // to backtrack and retry without subsuming loads. Other than this backtracking
4395 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4396 // by the logic in C2Compiler.
4397 void Compile::record_failure(const char* reason) {
4398 if (log() != nullptr) {
4399 log()->elem("failure reason='%s' phase='compile'", reason);
4400 }
4401 if (_failure_reason.get() == nullptr) {
4402 // Record the first failure reason.
4403 _failure_reason.set(reason);
4404 if (CaptureBailoutInformation) {
4405 _first_failure_details = new CompilationFailureInfo(reason);
4406 }
4407 }
4408
4409 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4410 C->print_method(PHASE_FAILURE, 1);
4411 }
4412 _root = nullptr; // flush the graph, too
4413 }
4414
4415 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4416 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4417 _compile(Compile::current()),
4418 _log(nullptr),
4419 _phase_name(name),
4420 _dolog(CITimeVerbose)
4421 {
4422 assert(_compile != nullptr, "sanity check");
4423 if (_dolog) {
4424 _log = _compile->log();
4425 }
4426 if (_log != nullptr) {
4427 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4428 _log->stamp();
4429 _log->end_head();
4430 }
4431 }
4432
4433 Compile::TracePhase::~TracePhase() {
4434 if (_compile->failing()) return;
4435 #ifdef ASSERT
4436 if (PrintIdealNodeCount) {
4437 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4438 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4439 }
4440
4441 if (VerifyIdealNodeCount) {
4442 _compile->print_missing_nodes();
4443 }
4444 #endif
4445
4446 if (_log != nullptr) {
4447 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4448 }
4449 }
4450
4451 //----------------------------static_subtype_check-----------------------------
4452 // Shortcut important common cases when superklass is exact:
4453 // (0) superklass is java.lang.Object (can occur in reflective code)
4454 // (1) subklass is already limited to a subtype of superklass => always ok
4455 // (2) subklass does not overlap with superklass => always fail
4456 // (3) superklass has NO subtypes and we can check with a simple compare.
4457 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4458 if (skip) {
4459 return SSC_full_test; // Let caller generate the general case.
4460 }
4461
4462 if (subk->is_java_subtype_of(superk)) {
4463 return SSC_always_true; // (0) and (1) this test cannot fail
4464 }
4465
4466 if (!subk->maybe_java_subtype_of(superk)) {
4467 return SSC_always_false; // (2) true path dead; no dynamic test needed
4468 }
4469
4470 const Type* superelem = superk;
4471 if (superk->isa_aryklassptr()) {
4472 int ignored;
4473 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4474 }
4475
4476 if (superelem->isa_instklassptr()) {
4477 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4478 if (!ik->has_subklass()) {
4479 if (!ik->is_final()) {
4480 // Add a dependency if there is a chance of a later subclass.
4481 dependencies()->assert_leaf_type(ik);
4482 }
4483 if (!superk->maybe_java_subtype_of(subk)) {
4484 return SSC_always_false;
4485 }
4486 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4487 }
4488 } else {
4489 // A primitive array type has no subtypes.
4490 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4491 }
4492
4493 return SSC_full_test;
4494 }
4495
4496 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4497 #ifdef _LP64
4498 // The scaled index operand to AddP must be a clean 64-bit value.
4499 // Java allows a 32-bit int to be incremented to a negative
4500 // value, which appears in a 64-bit register as a large
4501 // positive number. Using that large positive number as an
4502 // operand in pointer arithmetic has bad consequences.
4503 // On the other hand, 32-bit overflow is rare, and the possibility
4504 // can often be excluded, if we annotate the ConvI2L node with
4505 // a type assertion that its value is known to be a small positive
4506 // number. (The prior range check has ensured this.)
4507 // This assertion is used by ConvI2LNode::Ideal.
4508 int index_max = max_jint - 1; // array size is max_jint, index is one less
4509 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4510 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4511 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4512 #endif
4513 return idx;
4514 }
4515
4516 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4517 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4518 if (ctrl != nullptr) {
4519 // Express control dependency by a CastII node with a narrow type.
4520 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4521 // node from floating above the range check during loop optimizations. Otherwise, the
4522 // ConvI2L node may be eliminated independently of the range check, causing the data path
4523 // to become TOP while the control path is still there (although it's unreachable).
4524 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4525 value = phase->transform(value);
4526 }
4527 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4528 return phase->transform(new ConvI2LNode(value, ltype));
4529 }
4530
4531 // The message about the current inlining is accumulated in
4532 // _print_inlining_stream and transferred into the _print_inlining_list
4533 // once we know whether inlining succeeds or not. For regular
4534 // inlining, messages are appended to the buffer pointed by
4535 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4536 // a new buffer is added after _print_inlining_idx in the list. This
4537 // way we can update the inlining message for late inlining call site
4538 // when the inlining is attempted again.
4539 void Compile::print_inlining_init() {
4540 if (print_inlining() || print_intrinsics()) {
4541 // print_inlining_init is actually called several times.
4542 print_inlining_reset();
4543 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4544 }
4545 }
4546
4547 void Compile::print_inlining_reinit() {
4548 if (print_inlining() || print_intrinsics()) {
4549 print_inlining_reset();
4550 }
4551 }
4552
4553 void Compile::print_inlining_reset() {
4554 _print_inlining_stream->reset();
4555 }
4556
4557 void Compile::print_inlining_commit() {
4558 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4559 // Transfer the message from _print_inlining_stream to the current
4560 // _print_inlining_list buffer and clear _print_inlining_stream.
4561 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4562 print_inlining_reset();
4563 }
4564
4565 void Compile::print_inlining_push() {
4566 // Add new buffer to the _print_inlining_list at current position
4567 _print_inlining_idx++;
4568 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4569 }
4570
4571 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4572 return _print_inlining_list->at(_print_inlining_idx);
4573 }
4574
4575 void Compile::print_inlining_update(CallGenerator* cg) {
4576 if (print_inlining() || print_intrinsics()) {
4577 if (cg->is_late_inline()) {
4578 if (print_inlining_current()->cg() != cg &&
4579 (print_inlining_current()->cg() != nullptr ||
4580 print_inlining_current()->ss()->size() != 0)) {
4581 print_inlining_push();
4582 }
4583 print_inlining_commit();
4584 print_inlining_current()->set_cg(cg);
4585 } else {
4586 if (print_inlining_current()->cg() != nullptr) {
4587 print_inlining_push();
4588 }
4589 print_inlining_commit();
4590 }
4591 }
4592 }
4593
4594 void Compile::print_inlining_move_to(CallGenerator* cg) {
4595 // We resume inlining at a late inlining call site. Locate the
4596 // corresponding inlining buffer so that we can update it.
4597 if (print_inlining() || print_intrinsics()) {
4598 for (int i = 0; i < _print_inlining_list->length(); i++) {
4599 if (_print_inlining_list->at(i)->cg() == cg) {
4600 _print_inlining_idx = i;
4601 return;
4602 }
4603 }
4604 ShouldNotReachHere();
4605 }
4606 }
4607
4608 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4609 if (print_inlining() || print_intrinsics()) {
4610 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4611 assert(print_inlining_current()->cg() == cg, "wrong entry");
4612 // replace message with new message
4613 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4614 print_inlining_commit();
4615 print_inlining_current()->set_cg(cg);
4616 }
4617 }
4618
4619 void Compile::print_inlining_assert_ready() {
4620 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4621 }
4622
4623 void Compile::process_print_inlining() {
4624 assert(_late_inlines.length() == 0, "not drained yet");
4625 if (print_inlining() || print_intrinsics()) {
4626 ResourceMark rm;
4627 stringStream ss;
4628 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4629 for (int i = 0; i < _print_inlining_list->length(); i++) {
4630 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4631 ss.print("%s", pib->ss()->freeze());
4632 delete pib;
4633 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4634 }
4635 // Reset _print_inlining_list, it only contains destructed objects.
4636 // It is on the arena, so it will be freed when the arena is reset.
4637 _print_inlining_list = nullptr;
4638 // _print_inlining_stream won't be used anymore, either.
4639 print_inlining_reset();
4640 size_t end = ss.size();
4641 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4642 strncpy(_print_inlining_output, ss.freeze(), end+1);
4643 _print_inlining_output[end] = 0;
4644 }
4645 }
4646
4647 void Compile::dump_print_inlining() {
4648 if (_print_inlining_output != nullptr) {
4649 tty->print_raw(_print_inlining_output);
4650 }
4651 }
4652
4653 void Compile::log_late_inline(CallGenerator* cg) {
4654 if (log() != nullptr) {
4655 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4656 cg->unique_id());
4657 JVMState* p = cg->call_node()->jvms();
4658 while (p != nullptr) {
4659 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4660 p = p->caller();
4661 }
4662 log()->tail("late_inline");
4663 }
4664 }
4665
4666 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4667 log_late_inline(cg);
4668 if (log() != nullptr) {
4669 log()->inline_fail(msg);
4670 }
4671 }
4672
4673 void Compile::log_inline_id(CallGenerator* cg) {
4674 if (log() != nullptr) {
4675 // The LogCompilation tool needs a unique way to identify late
4676 // inline call sites. This id must be unique for this call site in
4677 // this compilation. Try to have it unique across compilations as
4678 // well because it can be convenient when grepping through the log
4679 // file.
4680 // Distinguish OSR compilations from others in case CICountOSR is
4681 // on.
4682 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4683 cg->set_unique_id(id);
4684 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4685 }
4686 }
4687
4688 void Compile::log_inline_failure(const char* msg) {
4689 if (C->log() != nullptr) {
4690 C->log()->inline_fail(msg);
4691 }
4692 }
4693
4694
4695 // Dump inlining replay data to the stream.
4696 // Don't change thread state and acquire any locks.
4697 void Compile::dump_inline_data(outputStream* out) {
4698 InlineTree* inl_tree = ilt();
4699 if (inl_tree != nullptr) {
4700 out->print(" inline %d", inl_tree->count());
4701 inl_tree->dump_replay_data(out);
4702 }
4703 }
4704
4705 void Compile::dump_inline_data_reduced(outputStream* out) {
4706 assert(ReplayReduce, "");
4707
4708 InlineTree* inl_tree = ilt();
4709 if (inl_tree == nullptr) {
4710 return;
4711 }
4712 // Enable iterative replay file reduction
4713 // Output "compile" lines for depth 1 subtrees,
4714 // simulating that those trees were compiled
4715 // instead of inlined.
4716 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4717 InlineTree* sub = inl_tree->subtrees().at(i);
4718 if (sub->inline_level() != 1) {
4719 continue;
4720 }
4721
4722 ciMethod* method = sub->method();
4723 int entry_bci = -1;
4724 int comp_level = env()->task()->comp_level();
4725 out->print("compile ");
4726 method->dump_name_as_ascii(out);
4727 out->print(" %d %d", entry_bci, comp_level);
4728 out->print(" inline %d", sub->count());
4729 sub->dump_replay_data(out, -1);
4730 out->cr();
4731 }
4732 }
4733
4734 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4735 if (n1->Opcode() < n2->Opcode()) return -1;
4736 else if (n1->Opcode() > n2->Opcode()) return 1;
4737
4738 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4739 for (uint i = 1; i < n1->req(); i++) {
4740 if (n1->in(i) < n2->in(i)) return -1;
4741 else if (n1->in(i) > n2->in(i)) return 1;
4742 }
4743
4744 return 0;
4745 }
4746
4747 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4748 Node* n1 = *n1p;
4749 Node* n2 = *n2p;
4750
4751 return cmp_expensive_nodes(n1, n2);
4752 }
4753
4754 void Compile::sort_expensive_nodes() {
4755 if (!expensive_nodes_sorted()) {
4756 _expensive_nodes.sort(cmp_expensive_nodes);
4757 }
4758 }
4759
4760 bool Compile::expensive_nodes_sorted() const {
4761 for (int i = 1; i < _expensive_nodes.length(); i++) {
4762 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4763 return false;
4764 }
4765 }
4766 return true;
4767 }
4768
4769 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4770 if (_expensive_nodes.length() == 0) {
4771 return false;
4772 }
4773
4774 assert(OptimizeExpensiveOps, "optimization off?");
4775
4776 // Take this opportunity to remove dead nodes from the list
4777 int j = 0;
4778 for (int i = 0; i < _expensive_nodes.length(); i++) {
4779 Node* n = _expensive_nodes.at(i);
4780 if (!n->is_unreachable(igvn)) {
4781 assert(n->is_expensive(), "should be expensive");
4782 _expensive_nodes.at_put(j, n);
4783 j++;
4784 }
4785 }
4786 _expensive_nodes.trunc_to(j);
4787
4788 // Then sort the list so that similar nodes are next to each other
4789 // and check for at least two nodes of identical kind with same data
4790 // inputs.
4791 sort_expensive_nodes();
4792
4793 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4794 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4795 return true;
4796 }
4797 }
4798
4799 return false;
4800 }
4801
4802 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4803 if (_expensive_nodes.length() == 0) {
4804 return;
4805 }
4806
4807 assert(OptimizeExpensiveOps, "optimization off?");
4808
4809 // Sort to bring similar nodes next to each other and clear the
4810 // control input of nodes for which there's only a single copy.
4811 sort_expensive_nodes();
4812
4813 int j = 0;
4814 int identical = 0;
4815 int i = 0;
4816 bool modified = false;
4817 for (; i < _expensive_nodes.length()-1; i++) {
4818 assert(j <= i, "can't write beyond current index");
4819 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4820 identical++;
4821 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4822 continue;
4823 }
4824 if (identical > 0) {
4825 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4826 identical = 0;
4827 } else {
4828 Node* n = _expensive_nodes.at(i);
4829 igvn.replace_input_of(n, 0, nullptr);
4830 igvn.hash_insert(n);
4831 modified = true;
4832 }
4833 }
4834 if (identical > 0) {
4835 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4836 } else if (_expensive_nodes.length() >= 1) {
4837 Node* n = _expensive_nodes.at(i);
4838 igvn.replace_input_of(n, 0, nullptr);
4839 igvn.hash_insert(n);
4840 modified = true;
4841 }
4842 _expensive_nodes.trunc_to(j);
4843 if (modified) {
4844 igvn.optimize();
4845 }
4846 }
4847
4848 void Compile::add_expensive_node(Node * n) {
4849 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4850 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4851 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4852 if (OptimizeExpensiveOps) {
4853 _expensive_nodes.append(n);
4854 } else {
4855 // Clear control input and let IGVN optimize expensive nodes if
4856 // OptimizeExpensiveOps is off.
4857 n->set_req(0, nullptr);
4858 }
4859 }
4860
4861 /**
4862 * Track coarsened Lock and Unlock nodes.
4863 */
4864
4865 class Lock_List : public Node_List {
4866 uint _origin_cnt;
4867 public:
4868 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4869 uint origin_cnt() const { return _origin_cnt; }
4870 };
4871
4872 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4873 int length = locks.length();
4874 if (length > 0) {
4875 // Have to keep this list until locks elimination during Macro nodes elimination.
4876 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4877 AbstractLockNode* alock = locks.at(0);
4878 BoxLockNode* box = alock->box_node()->as_BoxLock();
4879 for (int i = 0; i < length; i++) {
4880 AbstractLockNode* lock = locks.at(i);
4881 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4882 locks_list->push(lock);
4883 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4884 if (this_box != box) {
4885 // Locking regions (BoxLock) could be Unbalanced here:
4886 // - its coarsened locks were eliminated in earlier
4887 // macro nodes elimination followed by loop unroll
4888 // - it is OSR locking region (no Lock node)
4889 // Preserve Unbalanced status in such cases.
4890 if (!this_box->is_unbalanced()) {
4891 this_box->set_coarsened();
4892 }
4893 if (!box->is_unbalanced()) {
4894 box->set_coarsened();
4895 }
4896 }
4897 }
4898 _coarsened_locks.append(locks_list);
4899 }
4900 }
4901
4902 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4903 int count = coarsened_count();
4904 for (int i = 0; i < count; i++) {
4905 Node_List* locks_list = _coarsened_locks.at(i);
4906 for (uint j = 0; j < locks_list->size(); j++) {
4907 Node* lock = locks_list->at(j);
4908 assert(lock->is_AbstractLock(), "sanity");
4909 if (!useful.member(lock)) {
4910 locks_list->yank(lock);
4911 }
4912 }
4913 }
4914 }
4915
4916 void Compile::remove_coarsened_lock(Node* n) {
4917 if (n->is_AbstractLock()) {
4918 int count = coarsened_count();
4919 for (int i = 0; i < count; i++) {
4920 Node_List* locks_list = _coarsened_locks.at(i);
4921 locks_list->yank(n);
4922 }
4923 }
4924 }
4925
4926 bool Compile::coarsened_locks_consistent() {
4927 int count = coarsened_count();
4928 for (int i = 0; i < count; i++) {
4929 bool unbalanced = false;
4930 bool modified = false; // track locks kind modifications
4931 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4932 uint size = locks_list->size();
4933 if (size == 0) {
4934 unbalanced = false; // All locks were eliminated - good
4935 } else if (size != locks_list->origin_cnt()) {
4936 unbalanced = true; // Some locks were removed from list
4937 } else {
4938 for (uint j = 0; j < size; j++) {
4939 Node* lock = locks_list->at(j);
4940 // All nodes in group should have the same state (modified or not)
4941 if (!lock->as_AbstractLock()->is_coarsened()) {
4942 if (j == 0) {
4943 // first on list was modified, the rest should be too for consistency
4944 modified = true;
4945 } else if (!modified) {
4946 // this lock was modified but previous locks on the list were not
4947 unbalanced = true;
4948 break;
4949 }
4950 } else if (modified) {
4951 // previous locks on list were modified but not this lock
4952 unbalanced = true;
4953 break;
4954 }
4955 }
4956 }
4957 if (unbalanced) {
4958 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4959 #ifdef ASSERT
4960 if (PrintEliminateLocks) {
4961 tty->print_cr("=== unbalanced coarsened locks ===");
4962 for (uint l = 0; l < size; l++) {
4963 locks_list->at(l)->dump();
4964 }
4965 }
4966 #endif
4967 record_failure(C2Compiler::retry_no_locks_coarsening());
4968 return false;
4969 }
4970 }
4971 return true;
4972 }
4973
4974 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4975 // locks coarsening optimization removed Lock/Unlock nodes from them.
4976 // Such regions become unbalanced because coarsening only removes part
4977 // of Lock/Unlock nodes in region. As result we can't execute other
4978 // locks elimination optimizations which assume all code paths have
4979 // corresponding pair of Lock/Unlock nodes - they are balanced.
4980 void Compile::mark_unbalanced_boxes() const {
4981 int count = coarsened_count();
4982 for (int i = 0; i < count; i++) {
4983 Node_List* locks_list = _coarsened_locks.at(i);
4984 uint size = locks_list->size();
4985 if (size > 0) {
4986 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4987 BoxLockNode* box = alock->box_node()->as_BoxLock();
4988 if (alock->is_coarsened()) {
4989 // coarsened_locks_consistent(), which is called before this method, verifies
4990 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4991 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4992 for (uint j = 1; j < size; j++) {
4993 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4994 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4995 if (box != this_box) {
4996 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4997 box->set_unbalanced();
4998 this_box->set_unbalanced();
4999 }
5000 }
5001 }
5002 }
5003 }
5004 }
5005
5006 /**
5007 * Remove the speculative part of types and clean up the graph
5008 */
5009 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5010 if (UseTypeSpeculation) {
5011 Unique_Node_List worklist;
5012 worklist.push(root());
5013 int modified = 0;
5014 // Go over all type nodes that carry a speculative type, drop the
5015 // speculative part of the type and enqueue the node for an igvn
5016 // which may optimize it out.
5017 for (uint next = 0; next < worklist.size(); ++next) {
5018 Node *n = worklist.at(next);
5019 if (n->is_Type()) {
5020 TypeNode* tn = n->as_Type();
5021 const Type* t = tn->type();
5022 const Type* t_no_spec = t->remove_speculative();
5023 if (t_no_spec != t) {
5024 bool in_hash = igvn.hash_delete(n);
5025 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5026 tn->set_type(t_no_spec);
5027 igvn.hash_insert(n);
5028 igvn._worklist.push(n); // give it a chance to go away
5029 modified++;
5030 }
5031 }
5032 // Iterate over outs - endless loops is unreachable from below
5033 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5034 Node *m = n->fast_out(i);
5035 if (not_a_node(m)) {
5036 continue;
5037 }
5038 worklist.push(m);
5039 }
5040 }
5041 // Drop the speculative part of all types in the igvn's type table
5042 igvn.remove_speculative_types();
5043 if (modified > 0) {
5044 igvn.optimize();
5045 if (failing()) return;
5046 }
5047 #ifdef ASSERT
5048 // Verify that after the IGVN is over no speculative type has resurfaced
5049 worklist.clear();
5050 worklist.push(root());
5051 for (uint next = 0; next < worklist.size(); ++next) {
5052 Node *n = worklist.at(next);
5053 const Type* t = igvn.type_or_null(n);
5054 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5055 if (n->is_Type()) {
5056 t = n->as_Type()->type();
5057 assert(t == t->remove_speculative(), "no more speculative types");
5058 }
5059 // Iterate over outs - endless loops is unreachable from below
5060 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5061 Node *m = n->fast_out(i);
5062 if (not_a_node(m)) {
5063 continue;
5064 }
5065 worklist.push(m);
5066 }
5067 }
5068 igvn.check_no_speculative_types();
5069 #endif
5070 }
5071 }
5072
5073 // Auxiliary methods to support randomized stressing/fuzzing.
5074
5075 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5076 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5077 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5078 FLAG_SET_ERGO(StressSeed, _stress_seed);
5079 } else {
5080 _stress_seed = StressSeed;
5081 }
5082 if (_log != nullptr) {
5083 _log->elem("stress_test seed='%u'", _stress_seed);
5084 }
5085 }
5086
5087 int Compile::random() {
5088 _stress_seed = os::next_random(_stress_seed);
5089 return static_cast<int>(_stress_seed);
5090 }
5091
5092 // This method can be called the arbitrary number of times, with current count
5093 // as the argument. The logic allows selecting a single candidate from the
5094 // running list of candidates as follows:
5095 // int count = 0;
5096 // Cand* selected = null;
5097 // while(cand = cand->next()) {
5098 // if (randomized_select(++count)) {
5099 // selected = cand;
5100 // }
5101 // }
5102 //
5103 // Including count equalizes the chances any candidate is "selected".
5104 // This is useful when we don't have the complete list of candidates to choose
5105 // from uniformly. In this case, we need to adjust the randomicity of the
5106 // selection, or else we will end up biasing the selection towards the latter
5107 // candidates.
5108 //
5109 // Quick back-envelope calculation shows that for the list of n candidates
5110 // the equal probability for the candidate to persist as "best" can be
5111 // achieved by replacing it with "next" k-th candidate with the probability
5112 // of 1/k. It can be easily shown that by the end of the run, the
5113 // probability for any candidate is converged to 1/n, thus giving the
5114 // uniform distribution among all the candidates.
5115 //
5116 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5117 #define RANDOMIZED_DOMAIN_POW 29
5118 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5119 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5120 bool Compile::randomized_select(int count) {
5121 assert(count > 0, "only positive");
5122 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5123 }
5124
5125 CloneMap& Compile::clone_map() { return _clone_map; }
5126 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5127
5128 void NodeCloneInfo::dump_on(outputStream* st) const {
5129 st->print(" {%d:%d} ", idx(), gen());
5130 }
5131
5132 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5133 uint64_t val = value(old->_idx);
5134 NodeCloneInfo cio(val);
5135 assert(val != 0, "old node should be in the map");
5136 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5137 insert(nnn->_idx, cin.get());
5138 #ifndef PRODUCT
5139 if (is_debug()) {
5140 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5141 }
5142 #endif
5143 }
5144
5145 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5146 NodeCloneInfo cio(value(old->_idx));
5147 if (cio.get() == 0) {
5148 cio.set(old->_idx, 0);
5149 insert(old->_idx, cio.get());
5150 #ifndef PRODUCT
5151 if (is_debug()) {
5152 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5153 }
5154 #endif
5155 }
5156 clone(old, nnn, gen);
5157 }
5158
5159 int CloneMap::max_gen() const {
5160 int g = 0;
5161 DictI di(_dict);
5162 for(; di.test(); ++di) {
5163 int t = gen(di._key);
5164 if (g < t) {
5165 g = t;
5166 #ifndef PRODUCT
5167 if (is_debug()) {
5168 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5169 }
5170 #endif
5171 }
5172 }
5173 return g;
5174 }
5175
5176 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5177 uint64_t val = value(key);
5178 if (val != 0) {
5179 NodeCloneInfo ni(val);
5180 ni.dump_on(st);
5181 }
5182 }
5183
5184 void Compile::shuffle_macro_nodes() {
5185 if (_macro_nodes.length() < 2) {
5186 return;
5187 }
5188 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5189 uint j = C->random() % (i + 1);
5190 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5191 }
5192 }
5193
5194 // Move Allocate nodes to the start of the list
5195 void Compile::sort_macro_nodes() {
5196 int count = macro_count();
5197 int allocates = 0;
5198 for (int i = 0; i < count; i++) {
5199 Node* n = macro_node(i);
5200 if (n->is_Allocate()) {
5201 if (i != allocates) {
5202 Node* tmp = macro_node(allocates);
5203 _macro_nodes.at_put(allocates, n);
5204 _macro_nodes.at_put(i, tmp);
5205 }
5206 allocates++;
5207 }
5208 }
5209 }
5210
5211 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5212 if (failing()) { return; }
5213 EventCompilerPhase event(UNTIMED);
5214 if (event.should_commit()) {
5215 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5216 }
5217 #ifndef PRODUCT
5218 ResourceMark rm;
5219 stringStream ss;
5220 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5221 int iter = ++_igv_phase_iter[cpt];
5222 if (iter > 1) {
5223 ss.print(" %d", iter);
5224 }
5225 if (n != nullptr) {
5226 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5227 }
5228
5229 const char* name = ss.as_string();
5230 if (should_print_igv(level)) {
5231 _igv_printer->print_method(name, level);
5232 }
5233 if (should_print_phase(cpt)) {
5234 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5235 }
5236 #endif
5237 C->_latest_stage_start_counter.stamp();
5238 }
5239
5240 // Only used from CompileWrapper
5241 void Compile::begin_method() {
5242 #ifndef PRODUCT
5243 if (_method != nullptr && should_print_igv(1)) {
5244 _igv_printer->begin_method();
5245 }
5246 #endif
5247 C->_latest_stage_start_counter.stamp();
5248 }
5249
5250 // Only used from CompileWrapper
5251 void Compile::end_method() {
5252 EventCompilerPhase event(UNTIMED);
5253 if (event.should_commit()) {
5254 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5255 }
5256
5257 #ifndef PRODUCT
5258 if (_method != nullptr && should_print_igv(1)) {
5259 _igv_printer->end_method();
5260 }
5261 #endif
5262 }
5263
5264 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5265 #ifndef PRODUCT
5266 if (_directive->should_print_phase(cpt)) {
5267 return true;
5268 }
5269 #endif
5270 return false;
5271 }
5272
5273 bool Compile::should_print_igv(const int level) {
5274 #ifndef PRODUCT
5275 if (PrintIdealGraphLevel < 0) { // disabled by the user
5276 return false;
5277 }
5278
5279 bool need = directive()->IGVPrintLevelOption >= level;
5280 if (need && !_igv_printer) {
5281 _igv_printer = IdealGraphPrinter::printer();
5282 _igv_printer->set_compile(this);
5283 }
5284 return need;
5285 #else
5286 return false;
5287 #endif
5288 }
5289
5290 #ifndef PRODUCT
5291 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5292 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5293
5294 // Called from debugger. Prints method to the default file with the default phase name.
5295 // This works regardless of any Ideal Graph Visualizer flags set or not.
5296 void igv_print() {
5297 Compile::current()->igv_print_method_to_file();
5298 }
5299
5300 // Same as igv_print() above but with a specified phase name.
5301 void igv_print(const char* phase_name) {
5302 Compile::current()->igv_print_method_to_file(phase_name);
5303 }
5304
5305 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5306 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5307 // This works regardless of any Ideal Graph Visualizer flags set or not.
5308 void igv_print(bool network) {
5309 if (network) {
5310 Compile::current()->igv_print_method_to_network();
5311 } else {
5312 Compile::current()->igv_print_method_to_file();
5313 }
5314 }
5315
5316 // Same as igv_print(bool network) above but with a specified phase name.
5317 void igv_print(bool network, const char* phase_name) {
5318 if (network) {
5319 Compile::current()->igv_print_method_to_network(phase_name);
5320 } else {
5321 Compile::current()->igv_print_method_to_file(phase_name);
5322 }
5323 }
5324
5325 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5326 void igv_print_default() {
5327 Compile::current()->print_method(PHASE_DEBUG, 0);
5328 }
5329
5330 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5331 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5332 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5333 void igv_append() {
5334 Compile::current()->igv_print_method_to_file("Debug", true);
5335 }
5336
5337 // Same as igv_append() above but with a specified phase name.
5338 void igv_append(const char* phase_name) {
5339 Compile::current()->igv_print_method_to_file(phase_name, true);
5340 }
5341
5342 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5343 const char* file_name = "custom_debug.xml";
5344 if (_debug_file_printer == nullptr) {
5345 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5346 } else {
5347 _debug_file_printer->update_compiled_method(C->method());
5348 }
5349 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5350 _debug_file_printer->print(phase_name, (Node*)C->root());
5351 }
5352
5353 void Compile::igv_print_method_to_network(const char* phase_name) {
5354 if (_debug_network_printer == nullptr) {
5355 _debug_network_printer = new IdealGraphPrinter(C);
5356 } else {
5357 _debug_network_printer->update_compiled_method(C->method());
5358 }
5359 tty->print_cr("Method printed over network stream to IGV");
5360 _debug_network_printer->print(phase_name, (Node*)C->root());
5361 }
5362 #endif
5363
5364 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5365 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5366 return value;
5367 }
5368 Node* result = nullptr;
5369 if (bt == T_BYTE) {
5370 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5371 result = new RShiftINode(result, phase->intcon(24));
5372 } else if (bt == T_BOOLEAN) {
5373 result = new AndINode(value, phase->intcon(0xFF));
5374 } else if (bt == T_CHAR) {
5375 result = new AndINode(value,phase->intcon(0xFFFF));
5376 } else {
5377 assert(bt == T_SHORT, "unexpected narrow type");
5378 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5379 result = new RShiftINode(result, phase->intcon(16));
5380 }
5381 if (transform_res) {
5382 result = phase->transform(result);
5383 }
5384 return result;
5385 }
5386
5387 void Compile::record_method_not_compilable_oom() {
5388 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5389 }