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