1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compilerOracle.hpp"
37 #include "compiler/compiler_globals.hpp"
38 #include "compiler/disassembler.hpp"
39 #include "compiler/oopMap.hpp"
40 #include "gc/shared/barrierSet.hpp"
41 #include "gc/shared/c2/barrierSetC2.hpp"
42 #include "jfr/jfrEvents.hpp"
43 #include "jvm_io.h"
44 #include "memory/allocation.hpp"
45 #include "memory/resourceArea.hpp"
46 #include "opto/addnode.hpp"
47 #include "opto/block.hpp"
48 #include "opto/c2compiler.hpp"
49 #include "opto/callGenerator.hpp"
50 #include "opto/callnode.hpp"
51 #include "opto/castnode.hpp"
52 #include "opto/cfgnode.hpp"
53 #include "opto/chaitin.hpp"
54 #include "opto/compile.hpp"
55 #include "opto/connode.hpp"
56 #include "opto/convertnode.hpp"
57 #include "opto/divnode.hpp"
58 #include "opto/escape.hpp"
59 #include "opto/idealGraphPrinter.hpp"
60 #include "opto/locknode.hpp"
61 #include "opto/loopnode.hpp"
62 #include "opto/machnode.hpp"
63 #include "opto/macro.hpp"
64 #include "opto/matcher.hpp"
65 #include "opto/mathexactnode.hpp"
66 #include "opto/memnode.hpp"
67 #include "opto/mulnode.hpp"
68 #include "opto/narrowptrnode.hpp"
69 #include "opto/node.hpp"
70 #include "opto/opcodes.hpp"
71 #include "opto/output.hpp"
72 #include "opto/parse.hpp"
73 #include "opto/phaseX.hpp"
74 #include "opto/rootnode.hpp"
75 #include "opto/runtime.hpp"
76 #include "opto/stringopts.hpp"
77 #include "opto/type.hpp"
78 #include "opto/vector.hpp"
79 #include "opto/vectornode.hpp"
80 #include "runtime/globals_extension.hpp"
81 #include "runtime/sharedRuntime.hpp"
82 #include "runtime/signature.hpp"
83 #include "runtime/stubRoutines.hpp"
84 #include "runtime/timer.hpp"
85 #include "utilities/align.hpp"
86 #include "utilities/copy.hpp"
87 #include "utilities/macros.hpp"
88 #include "utilities/resourceHash.hpp"
89
90 // -------------------- Compile::mach_constant_base_node -----------------------
91 // Constant table base node singleton.
92 MachConstantBaseNode* Compile::mach_constant_base_node() {
93 if (_mach_constant_base_node == nullptr) {
94 _mach_constant_base_node = new MachConstantBaseNode();
95 _mach_constant_base_node->add_req(C->root());
96 }
97 return _mach_constant_base_node;
98 }
99
100
101 /// Support for intrinsics.
102
103 // Return the index at which m must be inserted (or already exists).
104 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
105 class IntrinsicDescPair {
106 private:
107 ciMethod* _m;
108 bool _is_virtual;
109 public:
110 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
111 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
112 ciMethod* m= elt->method();
113 ciMethod* key_m = key->_m;
114 if (key_m < m) return -1;
115 else if (key_m > m) return 1;
116 else {
117 bool is_virtual = elt->is_virtual();
118 bool key_virtual = key->_is_virtual;
119 if (key_virtual < is_virtual) return -1;
120 else if (key_virtual > is_virtual) return 1;
121 else return 0;
122 }
123 }
124 };
125 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
126 #ifdef ASSERT
127 for (int i = 1; i < _intrinsics.length(); i++) {
128 CallGenerator* cg1 = _intrinsics.at(i-1);
129 CallGenerator* cg2 = _intrinsics.at(i);
130 assert(cg1->method() != cg2->method()
131 ? cg1->method() < cg2->method()
132 : cg1->is_virtual() < cg2->is_virtual(),
133 "compiler intrinsics list must stay sorted");
134 }
135 #endif
136 IntrinsicDescPair pair(m, is_virtual);
137 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
138 }
139
140 void Compile::register_intrinsic(CallGenerator* cg) {
141 bool found = false;
142 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
143 assert(!found, "registering twice");
144 _intrinsics.insert_before(index, cg);
145 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
146 }
147
148 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
149 assert(m->is_loaded(), "don't try this on unloaded methods");
150 if (_intrinsics.length() > 0) {
151 bool found = false;
152 int index = intrinsic_insertion_index(m, is_virtual, found);
153 if (found) {
154 return _intrinsics.at(index);
155 }
156 }
157 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
158 if (m->intrinsic_id() != vmIntrinsics::_none &&
159 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
160 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
161 if (cg != nullptr) {
162 // Save it for next time:
163 register_intrinsic(cg);
164 return cg;
165 } else {
166 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
167 }
168 }
169 return nullptr;
170 }
171
172 // Compile::make_vm_intrinsic is defined in library_call.cpp.
173
174 #ifndef PRODUCT
175 // statistics gathering...
176
177 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
179
180 inline int as_int(vmIntrinsics::ID id) {
181 return vmIntrinsics::as_int(id);
182 }
183
184 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
185 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
186 int oflags = _intrinsic_hist_flags[as_int(id)];
187 assert(flags != 0, "what happened?");
188 if (is_virtual) {
189 flags |= _intrinsic_virtual;
190 }
191 bool changed = (flags != oflags);
192 if ((flags & _intrinsic_worked) != 0) {
193 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
194 if (count == 1) {
195 changed = true; // first time
196 }
197 // increment the overall count also:
198 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
199 }
200 if (changed) {
201 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
202 // Something changed about the intrinsic's virtuality.
203 if ((flags & _intrinsic_virtual) != 0) {
204 // This is the first use of this intrinsic as a virtual call.
205 if (oflags != 0) {
206 // We already saw it as a non-virtual, so note both cases.
207 flags |= _intrinsic_both;
208 }
209 } else if ((oflags & _intrinsic_both) == 0) {
210 // This is the first use of this intrinsic as a non-virtual
211 flags |= _intrinsic_both;
212 }
213 }
214 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
215 }
216 // update the overall flags also:
217 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
218 return changed;
219 }
220
221 static char* format_flags(int flags, char* buf) {
222 buf[0] = 0;
223 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
224 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
225 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
226 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
227 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
228 if (buf[0] == 0) strcat(buf, ",");
229 assert(buf[0] == ',', "must be");
230 return &buf[1];
231 }
232
233 void Compile::print_intrinsic_statistics() {
234 char flagsbuf[100];
235 ttyLocker ttyl;
236 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
237 tty->print_cr("Compiler intrinsic usage:");
238 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
239 if (total == 0) total = 1; // avoid div0 in case of no successes
240 #define PRINT_STAT_LINE(name, c, f) \
241 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
242 for (auto id : EnumRange<vmIntrinsicID>{}) {
243 int flags = _intrinsic_hist_flags[as_int(id)];
244 juint count = _intrinsic_hist_count[as_int(id)];
245 if ((flags | count) != 0) {
246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
247 }
248 }
249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
250 if (xtty != nullptr) xtty->tail("statistics");
251 }
252
253 void Compile::print_statistics() {
254 { ttyLocker ttyl;
255 if (xtty != nullptr) xtty->head("statistics type='opto'");
256 Parse::print_statistics();
257 PhaseStringOpts::print_statistics();
258 PhaseCCP::print_statistics();
259 PhaseRegAlloc::print_statistics();
260 PhaseOutput::print_statistics();
261 PhasePeephole::print_statistics();
262 PhaseIdealLoop::print_statistics();
263 ConnectionGraph::print_statistics();
264 PhaseMacroExpand::print_statistics();
265 if (xtty != nullptr) xtty->tail("statistics");
266 }
267 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
268 // put this under its own <statistics> element.
269 print_intrinsic_statistics();
270 }
271 }
272 #endif //PRODUCT
273
274 void Compile::gvn_replace_by(Node* n, Node* nn) {
275 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
276 Node* use = n->last_out(i);
277 bool is_in_table = initial_gvn()->hash_delete(use);
278 uint uses_found = 0;
279 for (uint j = 0; j < use->len(); j++) {
280 if (use->in(j) == n) {
281 if (j < use->req())
282 use->set_req(j, nn);
283 else
284 use->set_prec(j, nn);
285 uses_found++;
286 }
287 }
288 if (is_in_table) {
289 // reinsert into table
290 initial_gvn()->hash_find_insert(use);
291 }
292 record_for_igvn(use);
293 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
294 i -= uses_found; // we deleted 1 or more copies of this edge
295 }
296 }
297
298
299 // Identify all nodes that are reachable from below, useful.
300 // Use breadth-first pass that records state in a Unique_Node_List,
301 // recursive traversal is slower.
302 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
303 int estimated_worklist_size = live_nodes();
304 useful.map( estimated_worklist_size, nullptr ); // preallocate space
305
306 // Initialize worklist
307 if (root() != nullptr) { useful.push(root()); }
308 // If 'top' is cached, declare it useful to preserve cached node
309 if (cached_top_node()) { useful.push(cached_top_node()); }
310
311 // Push all useful nodes onto the list, breadthfirst
312 for( uint next = 0; next < useful.size(); ++next ) {
313 assert( next < unique(), "Unique useful nodes < total nodes");
314 Node *n = useful.at(next);
315 uint max = n->len();
316 for( uint i = 0; i < max; ++i ) {
317 Node *m = n->in(i);
318 if (not_a_node(m)) continue;
319 useful.push(m);
320 }
321 }
322 }
323
324 // Update dead_node_list with any missing dead nodes using useful
325 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
326 void Compile::update_dead_node_list(Unique_Node_List &useful) {
327 uint max_idx = unique();
328 VectorSet& useful_node_set = useful.member_set();
329
330 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
331 // If node with index node_idx is not in useful set,
332 // mark it as dead in dead node list.
333 if (!useful_node_set.test(node_idx)) {
334 record_dead_node(node_idx);
335 }
336 }
337 }
338
339 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
340 int shift = 0;
341 for (int i = 0; i < inlines->length(); i++) {
342 CallGenerator* cg = inlines->at(i);
343 if (useful.member(cg->call_node())) {
344 if (shift > 0) {
345 inlines->at_put(i - shift, cg);
346 }
347 } else {
348 shift++; // skip over the dead element
349 }
350 }
351 if (shift > 0) {
352 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
353 }
354 }
355
356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
357 assert(dead != nullptr && dead->is_Call(), "sanity");
358 int found = 0;
359 for (int i = 0; i < inlines->length(); i++) {
360 if (inlines->at(i)->call_node() == dead) {
361 inlines->remove_at(i);
362 found++;
363 NOT_DEBUG( break; ) // elements are unique, so exit early
364 }
365 }
366 assert(found <= 1, "not unique");
367 }
368
369 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
370 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
371 for (int i = node_list.length() - 1; i >= 0; i--) {
372 N* node = node_list.at(i);
373 if (!useful.member(node)) {
374 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
375 }
376 }
377 }
378
379 void Compile::remove_useless_node(Node* dead) {
380 remove_modified_node(dead);
381
382 // Constant node that has no out-edges and has only one in-edge from
383 // root is usually dead. However, sometimes reshaping walk makes
384 // it reachable by adding use edges. So, we will NOT count Con nodes
385 // as dead to be conservative about the dead node count at any
386 // given time.
387 if (!dead->is_Con()) {
388 record_dead_node(dead->_idx);
389 }
390 if (dead->is_macro()) {
391 remove_macro_node(dead);
392 }
393 if (dead->is_expensive()) {
394 remove_expensive_node(dead);
395 }
396 if (dead->Opcode() == Op_Opaque4) {
397 remove_template_assertion_predicate_opaq(dead);
398 }
399 if (dead->is_ParsePredicate()) {
400 remove_parse_predicate(dead->as_ParsePredicate());
401 }
402 if (dead->for_post_loop_opts_igvn()) {
403 remove_from_post_loop_opts_igvn(dead);
404 }
405 if (dead->is_Call()) {
406 remove_useless_late_inlines( &_late_inlines, dead);
407 remove_useless_late_inlines( &_string_late_inlines, dead);
408 remove_useless_late_inlines( &_boxing_late_inlines, dead);
409 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
410
411 if (dead->is_CallStaticJava()) {
412 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
413 }
414 }
415 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
416 bs->unregister_potential_barrier_node(dead);
417 }
418
419 // Disconnect all useless nodes by disconnecting those at the boundary.
420 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
421 uint next = 0;
422 while (next < useful.size()) {
423 Node *n = useful.at(next++);
424 if (n->is_SafePoint()) {
425 // We're done with a parsing phase. Replaced nodes are not valid
426 // beyond that point.
427 n->as_SafePoint()->delete_replaced_nodes();
428 }
429 // Use raw traversal of out edges since this code removes out edges
430 int max = n->outcnt();
431 for (int j = 0; j < max; ++j) {
432 Node* child = n->raw_out(j);
433 if (!useful.member(child)) {
434 assert(!child->is_top() || child != top(),
435 "If top is cached in Compile object it is in useful list");
436 // Only need to remove this out-edge to the useless node
437 n->raw_del_out(j);
438 --j;
439 --max;
440 }
441 }
442 if (n->outcnt() == 1 && n->has_special_unique_user()) {
443 assert(useful.member(n->unique_out()), "do not push a useless node");
444 worklist.push(n->unique_out());
445 }
446 }
447
448 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
449 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
450 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
451 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
452 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
453 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
454 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
455 #ifdef ASSERT
456 if (_modified_nodes != nullptr) {
457 _modified_nodes->remove_useless_nodes(useful.member_set());
458 }
459 #endif
460
461 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
462 bs->eliminate_useless_gc_barriers(useful, this);
463 // clean up the late inline lists
464 remove_useless_late_inlines( &_late_inlines, useful);
465 remove_useless_late_inlines( &_string_late_inlines, useful);
466 remove_useless_late_inlines( &_boxing_late_inlines, useful);
467 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
468 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
469 }
470
471 // ============================================================================
472 //------------------------------CompileWrapper---------------------------------
473 class CompileWrapper : public StackObj {
474 Compile *const _compile;
475 public:
476 CompileWrapper(Compile* compile);
477
478 ~CompileWrapper();
479 };
480
481 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
482 // the Compile* pointer is stored in the current ciEnv:
483 ciEnv* env = compile->env();
484 assert(env == ciEnv::current(), "must already be a ciEnv active");
485 assert(env->compiler_data() == nullptr, "compile already active?");
486 env->set_compiler_data(compile);
487 assert(compile == Compile::current(), "sanity");
488
489 compile->set_type_dict(nullptr);
490 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
491 compile->clone_map().set_clone_idx(0);
492 compile->set_type_last_size(0);
493 compile->set_last_tf(nullptr, nullptr);
494 compile->set_indexSet_arena(nullptr);
495 compile->set_indexSet_free_block_list(nullptr);
496 compile->init_type_arena();
497 Type::Initialize(compile);
498 _compile->begin_method();
499 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
500 }
501 CompileWrapper::~CompileWrapper() {
502 // simulate crash during compilation
503 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
504
505 _compile->end_method();
506 _compile->env()->set_compiler_data(nullptr);
507 }
508
509
510 //----------------------------print_compile_messages---------------------------
511 void Compile::print_compile_messages() {
512 #ifndef PRODUCT
513 // Check if recompiling
514 if (!subsume_loads() && PrintOpto) {
515 // Recompiling without allowing machine instructions to subsume loads
516 tty->print_cr("*********************************************************");
517 tty->print_cr("** Bailout: Recompile without subsuming loads **");
518 tty->print_cr("*********************************************************");
519 }
520 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
521 // Recompiling without escape analysis
522 tty->print_cr("*********************************************************");
523 tty->print_cr("** Bailout: Recompile without escape analysis **");
524 tty->print_cr("*********************************************************");
525 }
526 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
527 // Recompiling without iterative escape analysis
528 tty->print_cr("*********************************************************");
529 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
530 tty->print_cr("*********************************************************");
531 }
532 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
533 // Recompiling without reducing allocation merges
534 tty->print_cr("*********************************************************");
535 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
536 tty->print_cr("*********************************************************");
537 }
538 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
539 // Recompiling without boxing elimination
540 tty->print_cr("*********************************************************");
541 tty->print_cr("** Bailout: Recompile without boxing elimination **");
542 tty->print_cr("*********************************************************");
543 }
544 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
545 // Recompiling without locks coarsening
546 tty->print_cr("*********************************************************");
547 tty->print_cr("** Bailout: Recompile without locks coarsening **");
548 tty->print_cr("*********************************************************");
549 }
550 if (env()->break_at_compile()) {
551 // Open the debugger when compiling this method.
552 tty->print("### Breaking when compiling: ");
553 method()->print_short_name();
554 tty->cr();
555 BREAKPOINT;
556 }
557
558 if( PrintOpto ) {
559 if (is_osr_compilation()) {
560 tty->print("[OSR]%3d", _compile_id);
561 } else {
562 tty->print("%3d", _compile_id);
563 }
564 }
565 #endif
566 }
567
568 #ifndef PRODUCT
569 void Compile::print_ideal_ir(const char* phase_name) {
570 // keep the following output all in one block
571 // This output goes directly to the tty, not the compiler log.
572 // To enable tools to match it up with the compilation activity,
573 // be sure to tag this tty output with the compile ID.
574
575 // Node dumping can cause a safepoint, which can break the tty lock.
576 // Buffer all node dumps, so that all safepoints happen before we lock.
577 ResourceMark rm;
578 stringStream ss;
579
580 if (_output == nullptr) {
581 ss.print_cr("AFTER: %s", phase_name);
582 // Print out all nodes in ascending order of index.
583 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
584 } else {
585 // Dump the node blockwise if we have a scheduling
586 _output->print_scheduling(&ss);
587 }
588
589 // Check that the lock is not broken by a safepoint.
590 NoSafepointVerifier nsv;
591 ttyLocker ttyl;
592 if (xtty != nullptr) {
593 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
594 compile_id(),
595 is_osr_compilation() ? " compile_kind='osr'" : "",
596 phase_name);
597 }
598
599 tty->print("%s", ss.as_string());
600
601 if (xtty != nullptr) {
602 xtty->tail("ideal");
603 }
604 }
605 #endif
606
607 // ============================================================================
608 //------------------------------Compile standard-------------------------------
609
610 // Compile a method. entry_bci is -1 for normal compilations and indicates
611 // the continuation bci for on stack replacement.
612
613
614 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
615 Options options, DirectiveSet* directive)
616 : Phase(Compiler),
617 _compile_id(ci_env->compile_id()),
618 _options(options),
619 _method(target),
620 _entry_bci(osr_bci),
621 _ilt(nullptr),
622 _stub_function(nullptr),
623 _stub_name(nullptr),
624 _stub_entry_point(nullptr),
625 _max_node_limit(MaxNodeLimit),
626 _post_loop_opts_phase(false),
627 _allow_macro_nodes(true),
628 _inlining_progress(false),
629 _inlining_incrementally(false),
630 _do_cleanup(false),
631 _has_reserved_stack_access(target->has_reserved_stack_access()),
632 #ifndef PRODUCT
633 _igv_idx(0),
634 _trace_opto_output(directive->TraceOptoOutputOption),
635 #endif
636 _has_method_handle_invokes(false),
637 _clinit_barrier_on_entry(false),
638 _has_clinit_barriers(false),
639 _stress_seed(0),
640 _comp_arena(mtCompiler),
641 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
642 _env(ci_env),
643 _directive(directive),
644 _log(ci_env->log()),
645 _first_failure_details(nullptr),
646 _intrinsics (comp_arena(), 0, 0, nullptr),
647 _macro_nodes (comp_arena(), 8, 0, nullptr),
648 _parse_predicates (comp_arena(), 8, 0, nullptr),
649 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
650 _expensive_nodes (comp_arena(), 8, 0, nullptr),
651 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
652 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
653 _coarsened_locks (comp_arena(), 8, 0, nullptr),
654 _congraph(nullptr),
655 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
656 _unique(0),
657 _dead_node_count(0),
658 _dead_node_list(comp_arena()),
659 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
660 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
661 _node_arena(&_node_arena_one),
662 _mach_constant_base_node(nullptr),
663 _Compile_types(mtCompiler),
664 _initial_gvn(nullptr),
665 _igvn_worklist(nullptr),
666 _types(nullptr),
667 _node_hash(nullptr),
668 _late_inlines(comp_arena(), 2, 0, nullptr),
669 _string_late_inlines(comp_arena(), 2, 0, nullptr),
670 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
671 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
672 _late_inlines_pos(0),
673 _number_of_mh_late_inlines(0),
674 _oom(false),
675 _print_inlining_stream(new (mtCompiler) stringStream()),
676 _print_inlining_list(nullptr),
677 _print_inlining_idx(0),
678 _print_inlining_output(nullptr),
679 _replay_inline_data(nullptr),
680 _java_calls(0),
681 _inner_loops(0),
682 _interpreter_frame_size(0),
683 _output(nullptr)
684 #ifndef PRODUCT
685 , _in_dump_cnt(0)
686 #endif
687 {
688 C = this;
689 CompileWrapper cw(this);
690
691 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
692 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
693
694 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
695 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
696 // We can always print a disassembly, either abstract (hex dump) or
697 // with the help of a suitable hsdis library. Thus, we should not
698 // couple print_assembly and print_opto_assembly controls.
699 // But: always print opto and regular assembly on compile command 'print'.
700 bool print_assembly = directive->PrintAssemblyOption;
701 set_print_assembly(print_opto_assembly || print_assembly);
702 #else
703 set_print_assembly(false); // must initialize.
704 #endif
705
706 #ifndef PRODUCT
707 set_parsed_irreducible_loop(false);
708 #endif
709
710 if (directive->ReplayInlineOption) {
711 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
712 }
713 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
714 set_print_intrinsics(directive->PrintIntrinsicsOption);
715 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
716
717 if (ProfileTraps) {
718 // Make sure the method being compiled gets its own MDO,
719 // so we can at least track the decompile_count().
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 _has_clinit_barriers(false),
920 _stress_seed(0),
921 _comp_arena(mtCompiler),
922 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
923 _env(ci_env),
924 _directive(directive),
925 _log(ci_env->log()),
926 _first_failure_details(nullptr),
927 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
928 _congraph(nullptr),
929 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
930 _unique(0),
931 _dead_node_count(0),
932 _dead_node_list(comp_arena()),
933 _node_arena_one(mtCompiler),
934 _node_arena_two(mtCompiler),
935 _node_arena(&_node_arena_one),
936 _mach_constant_base_node(nullptr),
937 _Compile_types(mtCompiler),
938 _initial_gvn(nullptr),
939 _igvn_worklist(nullptr),
940 _types(nullptr),
941 _node_hash(nullptr),
942 _number_of_mh_late_inlines(0),
943 _oom(false),
944 _print_inlining_stream(new (mtCompiler) stringStream()),
945 _print_inlining_list(nullptr),
946 _print_inlining_idx(0),
947 _print_inlining_output(nullptr),
948 _replay_inline_data(nullptr),
949 _java_calls(0),
950 _inner_loops(0),
951 _interpreter_frame_size(0),
952 _output(nullptr),
953 #ifndef PRODUCT
954 _in_dump_cnt(0),
955 #endif
956 _allowed_reasons(0) {
957 C = this;
958
959 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
960 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
961
962 #ifndef PRODUCT
963 set_print_assembly(PrintFrameConverterAssembly);
964 set_parsed_irreducible_loop(false);
965 #else
966 set_print_assembly(false); // Must initialize.
967 #endif
968 set_has_irreducible_loop(false); // no loops
969
970 CompileWrapper cw(this);
971 Init(/*do_aliasing=*/ false);
972 init_tf((*generator)());
973
974 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
975 _types = new (comp_arena()) Type_Array(comp_arena());
976 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
977
978 if (StressLCM || StressGCM) {
979 initialize_stress_seed(directive);
980 }
981
982 {
983 PhaseGVN gvn;
984 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
985 gvn.transform(top());
986
987 GraphKit kit;
988 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
989 }
990
991 NOT_PRODUCT( verify_graph_edges(); )
992
993 Code_Gen();
994 }
995
996 Compile::~Compile() {
997 delete _print_inlining_stream;
998 delete _first_failure_details;
999 };
1000
1001 //------------------------------Init-------------------------------------------
1002 // Prepare for a single compilation
1003 void Compile::Init(bool aliasing) {
1004 _do_aliasing = aliasing;
1005 _unique = 0;
1006 _regalloc = nullptr;
1007
1008 _tf = nullptr; // filled in later
1009 _top = nullptr; // cached later
1010 _matcher = nullptr; // filled in later
1011 _cfg = nullptr; // filled in later
1012
1013 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1014
1015 _node_note_array = nullptr;
1016 _default_node_notes = nullptr;
1017 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1018
1019 _immutable_memory = nullptr; // filled in at first inquiry
1020
1021 #ifdef ASSERT
1022 _phase_optimize_finished = false;
1023 _exception_backedge = false;
1024 _type_verify = nullptr;
1025 #endif
1026
1027 // Globally visible Nodes
1028 // First set TOP to null to give safe behavior during creation of RootNode
1029 set_cached_top_node(nullptr);
1030 set_root(new RootNode());
1031 // Now that you have a Root to point to, create the real TOP
1032 set_cached_top_node( new ConNode(Type::TOP) );
1033 set_recent_alloc(nullptr, nullptr);
1034
1035 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1036 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1037 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1038 env()->set_dependencies(new Dependencies(env()));
1039
1040 _fixed_slots = 0;
1041 set_has_split_ifs(false);
1042 set_has_loops(false); // first approximation
1043 set_has_stringbuilder(false);
1044 set_has_boxed_value(false);
1045 _trap_can_recompile = false; // no traps emitted yet
1046 _major_progress = true; // start out assuming good things will happen
1047 set_has_unsafe_access(false);
1048 set_max_vector_size(0);
1049 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1050 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1051 set_decompile_count(0);
1052
1053 #ifndef PRODUCT
1054 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1055 #endif
1056
1057 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1058 _loop_opts_cnt = LoopOptsCount;
1059 set_do_inlining(Inline);
1060 set_max_inline_size(MaxInlineSize);
1061 set_freq_inline_size(FreqInlineSize);
1062 set_do_scheduling(OptoScheduling);
1063
1064 set_do_vector_loop(false);
1065 set_has_monitors(false);
1066 set_has_scoped_access(false);
1067
1068 if (AllowVectorizeOnDemand) {
1069 if (has_method() && _directive->VectorizeOption) {
1070 set_do_vector_loop(true);
1071 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());})
1072 } else if (has_method() && method()->name() != nullptr &&
1073 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1074 set_do_vector_loop(true);
1075 }
1076 }
1077 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1078 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());})
1079
1080 _max_node_limit = _directive->MaxNodeLimitOption;
1081
1082 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() &&
1083 (method()->needs_clinit_barrier() || (do_clinit_barriers() && method()->is_static()))) {
1084 set_clinit_barrier_on_entry(true);
1085 if (do_clinit_barriers()) {
1086 set_has_clinit_barriers(true); // Entry clinit barrier is in prolog code.
1087 }
1088 }
1089 if (debug_info()->recording_non_safepoints()) {
1090 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1091 (comp_arena(), 8, 0, nullptr));
1092 set_default_node_notes(Node_Notes::make(this));
1093 }
1094
1095 const int grow_ats = 16;
1096 _max_alias_types = grow_ats;
1097 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1098 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1099 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1100 {
1101 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1102 }
1103 // Initialize the first few types.
1104 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1105 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1106 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1107 _num_alias_types = AliasIdxRaw+1;
1108 // Zero out the alias type cache.
1109 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1110 // A null adr_type hits in the cache right away. Preload the right answer.
1111 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1112 }
1113
1114 //---------------------------init_start----------------------------------------
1115 // Install the StartNode on this compile object.
1116 void Compile::init_start(StartNode* s) {
1117 if (failing())
1118 return; // already failing
1119 assert(s == start(), "");
1120 }
1121
1122 /**
1123 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1124 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1125 * the ideal graph.
1126 */
1127 StartNode* Compile::start() const {
1128 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1129 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1130 Node* start = root()->fast_out(i);
1131 if (start->is_Start()) {
1132 return start->as_Start();
1133 }
1134 }
1135 fatal("Did not find Start node!");
1136 return nullptr;
1137 }
1138
1139 //-------------------------------immutable_memory-------------------------------------
1140 // Access immutable memory
1141 Node* Compile::immutable_memory() {
1142 if (_immutable_memory != nullptr) {
1143 return _immutable_memory;
1144 }
1145 StartNode* s = start();
1146 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1147 Node *p = s->fast_out(i);
1148 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1149 _immutable_memory = p;
1150 return _immutable_memory;
1151 }
1152 }
1153 ShouldNotReachHere();
1154 return nullptr;
1155 }
1156
1157 //----------------------set_cached_top_node------------------------------------
1158 // Install the cached top node, and make sure Node::is_top works correctly.
1159 void Compile::set_cached_top_node(Node* tn) {
1160 if (tn != nullptr) verify_top(tn);
1161 Node* old_top = _top;
1162 _top = tn;
1163 // Calling Node::setup_is_top allows the nodes the chance to adjust
1164 // their _out arrays.
1165 if (_top != nullptr) _top->setup_is_top();
1166 if (old_top != nullptr) old_top->setup_is_top();
1167 assert(_top == nullptr || top()->is_top(), "");
1168 }
1169
1170 #ifdef ASSERT
1171 uint Compile::count_live_nodes_by_graph_walk() {
1172 Unique_Node_List useful(comp_arena());
1173 // Get useful node list by walking the graph.
1174 identify_useful_nodes(useful);
1175 return useful.size();
1176 }
1177
1178 void Compile::print_missing_nodes() {
1179
1180 // Return if CompileLog is null and PrintIdealNodeCount is false.
1181 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1182 return;
1183 }
1184
1185 // This is an expensive function. It is executed only when the user
1186 // specifies VerifyIdealNodeCount option or otherwise knows the
1187 // additional work that needs to be done to identify reachable nodes
1188 // by walking the flow graph and find the missing ones using
1189 // _dead_node_list.
1190
1191 Unique_Node_List useful(comp_arena());
1192 // Get useful node list by walking the graph.
1193 identify_useful_nodes(useful);
1194
1195 uint l_nodes = C->live_nodes();
1196 uint l_nodes_by_walk = useful.size();
1197
1198 if (l_nodes != l_nodes_by_walk) {
1199 if (_log != nullptr) {
1200 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1201 _log->stamp();
1202 _log->end_head();
1203 }
1204 VectorSet& useful_member_set = useful.member_set();
1205 int last_idx = l_nodes_by_walk;
1206 for (int i = 0; i < last_idx; i++) {
1207 if (useful_member_set.test(i)) {
1208 if (_dead_node_list.test(i)) {
1209 if (_log != nullptr) {
1210 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1211 }
1212 if (PrintIdealNodeCount) {
1213 // Print the log message to tty
1214 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1215 useful.at(i)->dump();
1216 }
1217 }
1218 }
1219 else if (! _dead_node_list.test(i)) {
1220 if (_log != nullptr) {
1221 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1222 }
1223 if (PrintIdealNodeCount) {
1224 // Print the log message to tty
1225 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1226 }
1227 }
1228 }
1229 if (_log != nullptr) {
1230 _log->tail("mismatched_nodes");
1231 }
1232 }
1233 }
1234 void Compile::record_modified_node(Node* n) {
1235 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1236 _modified_nodes->push(n);
1237 }
1238 }
1239
1240 void Compile::remove_modified_node(Node* n) {
1241 if (_modified_nodes != nullptr) {
1242 _modified_nodes->remove(n);
1243 }
1244 }
1245 #endif
1246
1247 #ifndef PRODUCT
1248 void Compile::verify_top(Node* tn) const {
1249 if (tn != nullptr) {
1250 assert(tn->is_Con(), "top node must be a constant");
1251 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1252 assert(tn->in(0) != nullptr, "must have live top node");
1253 }
1254 }
1255 #endif
1256
1257
1258 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1259
1260 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1261 guarantee(arr != nullptr, "");
1262 int num_blocks = arr->length();
1263 if (grow_by < num_blocks) grow_by = num_blocks;
1264 int num_notes = grow_by * _node_notes_block_size;
1265 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1266 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1267 while (num_notes > 0) {
1268 arr->append(notes);
1269 notes += _node_notes_block_size;
1270 num_notes -= _node_notes_block_size;
1271 }
1272 assert(num_notes == 0, "exact multiple, please");
1273 }
1274
1275 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1276 if (source == nullptr || dest == nullptr) return false;
1277
1278 if (dest->is_Con())
1279 return false; // Do not push debug info onto constants.
1280
1281 #ifdef ASSERT
1282 // Leave a bread crumb trail pointing to the original node:
1283 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1284 dest->set_debug_orig(source);
1285 }
1286 #endif
1287
1288 if (node_note_array() == nullptr)
1289 return false; // Not collecting any notes now.
1290
1291 // This is a copy onto a pre-existing node, which may already have notes.
1292 // If both nodes have notes, do not overwrite any pre-existing notes.
1293 Node_Notes* source_notes = node_notes_at(source->_idx);
1294 if (source_notes == nullptr || source_notes->is_clear()) return false;
1295 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1296 if (dest_notes == nullptr || dest_notes->is_clear()) {
1297 return set_node_notes_at(dest->_idx, source_notes);
1298 }
1299
1300 Node_Notes merged_notes = (*source_notes);
1301 // The order of operations here ensures that dest notes will win...
1302 merged_notes.update_from(dest_notes);
1303 return set_node_notes_at(dest->_idx, &merged_notes);
1304 }
1305
1306
1307 //--------------------------allow_range_check_smearing-------------------------
1308 // Gating condition for coalescing similar range checks.
1309 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1310 // single covering check that is at least as strong as any of them.
1311 // If the optimization succeeds, the simplified (strengthened) range check
1312 // will always succeed. If it fails, we will deopt, and then give up
1313 // on the optimization.
1314 bool Compile::allow_range_check_smearing() const {
1315 // If this method has already thrown a range-check,
1316 // assume it was because we already tried range smearing
1317 // and it failed.
1318 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1319 return !already_trapped;
1320 }
1321
1322
1323 //------------------------------flatten_alias_type-----------------------------
1324 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1325 assert(do_aliasing(), "Aliasing should be enabled");
1326 int offset = tj->offset();
1327 TypePtr::PTR ptr = tj->ptr();
1328
1329 // Known instance (scalarizable allocation) alias only with itself.
1330 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1331 tj->is_oopptr()->is_known_instance();
1332
1333 // Process weird unsafe references.
1334 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1335 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1336 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1337 tj = TypeOopPtr::BOTTOM;
1338 ptr = tj->ptr();
1339 offset = tj->offset();
1340 }
1341
1342 // Array pointers need some flattening
1343 const TypeAryPtr* ta = tj->isa_aryptr();
1344 if (ta && ta->is_stable()) {
1345 // Erase stability property for alias analysis.
1346 tj = ta = ta->cast_to_stable(false);
1347 }
1348 if( ta && is_known_inst ) {
1349 if ( offset != Type::OffsetBot &&
1350 offset > arrayOopDesc::length_offset_in_bytes() ) {
1351 offset = Type::OffsetBot; // Flatten constant access into array body only
1352 tj = ta = ta->
1353 remove_speculative()->
1354 cast_to_ptr_type(ptr)->
1355 with_offset(offset);
1356 }
1357 } else if (ta) {
1358 // For arrays indexed by constant indices, we flatten the alias
1359 // space to include all of the array body. Only the header, klass
1360 // and array length can be accessed un-aliased.
1361 if( offset != Type::OffsetBot ) {
1362 if( ta->const_oop() ) { // MethodData* or Method*
1363 offset = Type::OffsetBot; // Flatten constant access into array body
1364 tj = ta = ta->
1365 remove_speculative()->
1366 cast_to_ptr_type(ptr)->
1367 cast_to_exactness(false)->
1368 with_offset(offset);
1369 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1370 // range is OK as-is.
1371 tj = ta = TypeAryPtr::RANGE;
1372 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1373 tj = TypeInstPtr::KLASS; // all klass loads look alike
1374 ta = TypeAryPtr::RANGE; // generic ignored junk
1375 ptr = TypePtr::BotPTR;
1376 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1377 tj = TypeInstPtr::MARK;
1378 ta = TypeAryPtr::RANGE; // generic ignored junk
1379 ptr = TypePtr::BotPTR;
1380 } else { // Random constant offset into array body
1381 offset = Type::OffsetBot; // Flatten constant access into array body
1382 tj = ta = ta->
1383 remove_speculative()->
1384 cast_to_ptr_type(ptr)->
1385 cast_to_exactness(false)->
1386 with_offset(offset);
1387 }
1388 }
1389 // Arrays of fixed size alias with arrays of unknown size.
1390 if (ta->size() != TypeInt::POS) {
1391 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1392 tj = ta = ta->
1393 remove_speculative()->
1394 cast_to_ptr_type(ptr)->
1395 with_ary(tary)->
1396 cast_to_exactness(false);
1397 }
1398 // Arrays of known objects become arrays of unknown objects.
1399 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1400 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1401 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1402 }
1403 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1404 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1405 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1406 }
1407 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1408 // cannot be distinguished by bytecode alone.
1409 if (ta->elem() == TypeInt::BOOL) {
1410 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1411 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1412 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1413 }
1414 // During the 2nd round of IterGVN, NotNull castings are removed.
1415 // Make sure the Bottom and NotNull variants alias the same.
1416 // Also, make sure exact and non-exact variants alias the same.
1417 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1418 tj = ta = ta->
1419 remove_speculative()->
1420 cast_to_ptr_type(TypePtr::BotPTR)->
1421 cast_to_exactness(false)->
1422 with_offset(offset);
1423 }
1424 }
1425
1426 // Oop pointers need some flattening
1427 const TypeInstPtr *to = tj->isa_instptr();
1428 if (to && to != TypeOopPtr::BOTTOM) {
1429 ciInstanceKlass* ik = to->instance_klass();
1430 if( ptr == TypePtr::Constant ) {
1431 if (ik != ciEnv::current()->Class_klass() ||
1432 offset < ik->layout_helper_size_in_bytes()) {
1433 // No constant oop pointers (such as Strings); they alias with
1434 // unknown strings.
1435 assert(!is_known_inst, "not scalarizable allocation");
1436 tj = to = to->
1437 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1438 remove_speculative()->
1439 cast_to_ptr_type(TypePtr::BotPTR)->
1440 cast_to_exactness(false);
1441 }
1442 } else if( is_known_inst ) {
1443 tj = to; // Keep NotNull and klass_is_exact for instance type
1444 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1445 // During the 2nd round of IterGVN, NotNull castings are removed.
1446 // Make sure the Bottom and NotNull variants alias the same.
1447 // Also, make sure exact and non-exact variants alias the same.
1448 tj = to = to->
1449 remove_speculative()->
1450 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1451 cast_to_ptr_type(TypePtr::BotPTR)->
1452 cast_to_exactness(false);
1453 }
1454 if (to->speculative() != nullptr) {
1455 tj = to = to->remove_speculative();
1456 }
1457 // Canonicalize the holder of this field
1458 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1459 // First handle header references such as a LoadKlassNode, even if the
1460 // object's klass is unloaded at compile time (4965979).
1461 if (!is_known_inst) { // Do it only for non-instance types
1462 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1463 }
1464 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1465 // Static fields are in the space above the normal instance
1466 // fields in the java.lang.Class instance.
1467 if (ik != ciEnv::current()->Class_klass()) {
1468 to = nullptr;
1469 tj = TypeOopPtr::BOTTOM;
1470 offset = tj->offset();
1471 }
1472 } else {
1473 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1474 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1475 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1476 if( is_known_inst ) {
1477 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id());
1478 } else {
1479 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset);
1480 }
1481 }
1482 }
1483 }
1484
1485 // Klass pointers to object array klasses need some flattening
1486 const TypeKlassPtr *tk = tj->isa_klassptr();
1487 if( tk ) {
1488 // If we are referencing a field within a Klass, we need
1489 // to assume the worst case of an Object. Both exact and
1490 // inexact types must flatten to the same alias class so
1491 // use NotNull as the PTR.
1492 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1493 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1494 env()->Object_klass(),
1495 offset);
1496 }
1497
1498 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1499 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1500 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1501 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1502 } else {
1503 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1504 }
1505 }
1506
1507 // Check for precise loads from the primary supertype array and force them
1508 // to the supertype cache alias index. Check for generic array loads from
1509 // the primary supertype array and also force them to the supertype cache
1510 // alias index. Since the same load can reach both, we need to merge
1511 // these 2 disparate memories into the same alias class. Since the
1512 // primary supertype array is read-only, there's no chance of confusion
1513 // where we bypass an array load and an array store.
1514 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1515 if (offset == Type::OffsetBot ||
1516 (offset >= primary_supers_offset &&
1517 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1518 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1519 offset = in_bytes(Klass::secondary_super_cache_offset());
1520 tj = tk = tk->with_offset(offset);
1521 }
1522 }
1523
1524 // Flatten all Raw pointers together.
1525 if (tj->base() == Type::RawPtr)
1526 tj = TypeRawPtr::BOTTOM;
1527
1528 if (tj->base() == Type::AnyPtr)
1529 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1530
1531 offset = tj->offset();
1532 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1533
1534 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1535 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1536 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1537 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1538 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1539 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1540 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1541 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1542 assert( tj->ptr() != TypePtr::TopPTR &&
1543 tj->ptr() != TypePtr::AnyNull &&
1544 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1545 // assert( tj->ptr() != TypePtr::Constant ||
1546 // tj->base() == Type::RawPtr ||
1547 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1548
1549 return tj;
1550 }
1551
1552 void Compile::AliasType::Init(int i, const TypePtr* at) {
1553 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1554 _index = i;
1555 _adr_type = at;
1556 _field = nullptr;
1557 _element = nullptr;
1558 _is_rewritable = true; // default
1559 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1560 if (atoop != nullptr && atoop->is_known_instance()) {
1561 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1562 _general_index = Compile::current()->get_alias_index(gt);
1563 } else {
1564 _general_index = 0;
1565 }
1566 }
1567
1568 BasicType Compile::AliasType::basic_type() const {
1569 if (element() != nullptr) {
1570 const Type* element = adr_type()->is_aryptr()->elem();
1571 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1572 } if (field() != nullptr) {
1573 return field()->layout_type();
1574 } else {
1575 return T_ILLEGAL; // unknown
1576 }
1577 }
1578
1579 //---------------------------------print_on------------------------------------
1580 #ifndef PRODUCT
1581 void Compile::AliasType::print_on(outputStream* st) {
1582 if (index() < 10)
1583 st->print("@ <%d> ", index());
1584 else st->print("@ <%d>", index());
1585 st->print(is_rewritable() ? " " : " RO");
1586 int offset = adr_type()->offset();
1587 if (offset == Type::OffsetBot)
1588 st->print(" +any");
1589 else st->print(" +%-3d", offset);
1590 st->print(" in ");
1591 adr_type()->dump_on(st);
1592 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1593 if (field() != nullptr && tjp) {
1594 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1595 tjp->offset() != field()->offset_in_bytes()) {
1596 st->print(" != ");
1597 field()->print();
1598 st->print(" ***");
1599 }
1600 }
1601 }
1602
1603 void print_alias_types() {
1604 Compile* C = Compile::current();
1605 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1606 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1607 C->alias_type(idx)->print_on(tty);
1608 tty->cr();
1609 }
1610 }
1611 #endif
1612
1613
1614 //----------------------------probe_alias_cache--------------------------------
1615 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1616 intptr_t key = (intptr_t) adr_type;
1617 key ^= key >> logAliasCacheSize;
1618 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1619 }
1620
1621
1622 //-----------------------------grow_alias_types--------------------------------
1623 void Compile::grow_alias_types() {
1624 const int old_ats = _max_alias_types; // how many before?
1625 const int new_ats = old_ats; // how many more?
1626 const int grow_ats = old_ats+new_ats; // how many now?
1627 _max_alias_types = grow_ats;
1628 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1629 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1630 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1631 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1632 }
1633
1634
1635 //--------------------------------find_alias_type------------------------------
1636 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1637 if (!do_aliasing()) {
1638 return alias_type(AliasIdxBot);
1639 }
1640
1641 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1642 if (ace->_adr_type == adr_type) {
1643 return alias_type(ace->_index);
1644 }
1645
1646 // Handle special cases.
1647 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1648 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1649
1650 // Do it the slow way.
1651 const TypePtr* flat = flatten_alias_type(adr_type);
1652
1653 #ifdef ASSERT
1654 {
1655 ResourceMark rm;
1656 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1657 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1658 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1659 Type::str(adr_type));
1660 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1661 const TypeOopPtr* foop = flat->is_oopptr();
1662 // Scalarizable allocations have exact klass always.
1663 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1664 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1665 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1666 Type::str(foop), Type::str(xoop));
1667 }
1668 }
1669 #endif
1670
1671 int idx = AliasIdxTop;
1672 for (int i = 0; i < num_alias_types(); i++) {
1673 if (alias_type(i)->adr_type() == flat) {
1674 idx = i;
1675 break;
1676 }
1677 }
1678
1679 if (idx == AliasIdxTop) {
1680 if (no_create) return nullptr;
1681 // Grow the array if necessary.
1682 if (_num_alias_types == _max_alias_types) grow_alias_types();
1683 // Add a new alias type.
1684 idx = _num_alias_types++;
1685 _alias_types[idx]->Init(idx, flat);
1686 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1687 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1688 if (flat->isa_instptr()) {
1689 if (flat->offset() == java_lang_Class::klass_offset()
1690 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1691 alias_type(idx)->set_rewritable(false);
1692 }
1693 if (flat->isa_aryptr()) {
1694 #ifdef ASSERT
1695 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1696 // (T_BYTE has the weakest alignment and size restrictions...)
1697 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1698 #endif
1699 if (flat->offset() == TypePtr::OffsetBot) {
1700 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1701 }
1702 }
1703 if (flat->isa_klassptr()) {
1704 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1705 alias_type(idx)->set_rewritable(false);
1706 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1707 alias_type(idx)->set_rewritable(false);
1708 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1709 alias_type(idx)->set_rewritable(false);
1710 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1711 alias_type(idx)->set_rewritable(false);
1712 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1713 alias_type(idx)->set_rewritable(false);
1714 }
1715 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1716 // but the base pointer type is not distinctive enough to identify
1717 // references into JavaThread.)
1718
1719 // Check for final fields.
1720 const TypeInstPtr* tinst = flat->isa_instptr();
1721 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1722 ciField* field;
1723 if (tinst->const_oop() != nullptr &&
1724 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1725 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1726 // static field
1727 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1728 field = k->get_field_by_offset(tinst->offset(), true);
1729 } else {
1730 ciInstanceKlass *k = tinst->instance_klass();
1731 field = k->get_field_by_offset(tinst->offset(), false);
1732 }
1733 assert(field == nullptr ||
1734 original_field == nullptr ||
1735 (field->holder() == original_field->holder() &&
1736 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1737 field->is_static() == original_field->is_static()), "wrong field?");
1738 // Set field() and is_rewritable() attributes.
1739 if (field != nullptr) alias_type(idx)->set_field(field);
1740 }
1741 }
1742
1743 // Fill the cache for next time.
1744 ace->_adr_type = adr_type;
1745 ace->_index = idx;
1746 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1747
1748 // Might as well try to fill the cache for the flattened version, too.
1749 AliasCacheEntry* face = probe_alias_cache(flat);
1750 if (face->_adr_type == nullptr) {
1751 face->_adr_type = flat;
1752 face->_index = idx;
1753 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1754 }
1755
1756 return alias_type(idx);
1757 }
1758
1759
1760 Compile::AliasType* Compile::alias_type(ciField* field) {
1761 const TypeOopPtr* t;
1762 if (field->is_static())
1763 t = TypeInstPtr::make(field->holder()->java_mirror());
1764 else
1765 t = TypeOopPtr::make_from_klass_raw(field->holder());
1766 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1767 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1768 return atp;
1769 }
1770
1771
1772 //------------------------------have_alias_type--------------------------------
1773 bool Compile::have_alias_type(const TypePtr* adr_type) {
1774 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1775 if (ace->_adr_type == adr_type) {
1776 return true;
1777 }
1778
1779 // Handle special cases.
1780 if (adr_type == nullptr) return true;
1781 if (adr_type == TypePtr::BOTTOM) return true;
1782
1783 return find_alias_type(adr_type, true, nullptr) != nullptr;
1784 }
1785
1786 //-----------------------------must_alias--------------------------------------
1787 // True if all values of the given address type are in the given alias category.
1788 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1789 if (alias_idx == AliasIdxBot) return true; // the universal category
1790 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1791 if (alias_idx == AliasIdxTop) return false; // the empty category
1792 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1793
1794 // the only remaining possible overlap is identity
1795 int adr_idx = get_alias_index(adr_type);
1796 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1797 assert(adr_idx == alias_idx ||
1798 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1799 && adr_type != TypeOopPtr::BOTTOM),
1800 "should not be testing for overlap with an unsafe pointer");
1801 return adr_idx == alias_idx;
1802 }
1803
1804 //------------------------------can_alias--------------------------------------
1805 // True if any values of the given address type are in the given alias category.
1806 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1807 if (alias_idx == AliasIdxTop) return false; // the empty category
1808 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1809 // Known instance doesn't alias with bottom memory
1810 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1811 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1812
1813 // the only remaining possible overlap is identity
1814 int adr_idx = get_alias_index(adr_type);
1815 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1816 return adr_idx == alias_idx;
1817 }
1818
1819 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1820 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1821 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1822 if (parse_predicate_count() == 0) {
1823 return;
1824 }
1825 for (int i = 0; i < parse_predicate_count(); i++) {
1826 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1827 parse_predicate->mark_useless();
1828 igvn._worklist.push(parse_predicate);
1829 }
1830 _parse_predicates.clear();
1831 }
1832
1833 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1834 if (!n->for_post_loop_opts_igvn()) {
1835 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1836 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1837 _for_post_loop_igvn.append(n);
1838 }
1839 }
1840
1841 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1842 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1843 _for_post_loop_igvn.remove(n);
1844 }
1845
1846 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1847 // Verify that all previous optimizations produced a valid graph
1848 // at least to this point, even if no loop optimizations were done.
1849 PhaseIdealLoop::verify(igvn);
1850
1851 C->set_post_loop_opts_phase(); // no more loop opts allowed
1852
1853 assert(!C->major_progress(), "not cleared");
1854
1855 if (_for_post_loop_igvn.length() > 0) {
1856 while (_for_post_loop_igvn.length() > 0) {
1857 Node* n = _for_post_loop_igvn.pop();
1858 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1859 igvn._worklist.push(n);
1860 }
1861 igvn.optimize();
1862 if (failing()) return;
1863 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1864 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1865
1866 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1867 if (C->major_progress()) {
1868 C->clear_major_progress(); // ensure that major progress is now clear
1869 }
1870 }
1871 }
1872
1873 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1874 if (OptimizeUnstableIf) {
1875 _unstable_if_traps.append(trap);
1876 }
1877 }
1878
1879 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1880 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1881 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1882 Node* n = trap->uncommon_trap();
1883 if (!useful.member(n)) {
1884 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1885 }
1886 }
1887 }
1888
1889 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1890 // or fold-compares case. Return true if succeed or not found.
1891 //
1892 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1893 // false and ask fold-compares to yield.
1894 //
1895 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1896 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1897 // when deoptimization does happen.
1898 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1899 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1900 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1901 if (trap->uncommon_trap() == unc) {
1902 if (yield && trap->modified()) {
1903 return false;
1904 }
1905 _unstable_if_traps.delete_at(i);
1906 break;
1907 }
1908 }
1909 return true;
1910 }
1911
1912 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1913 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1914 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1915 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1916 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1917 CallStaticJavaNode* unc = trap->uncommon_trap();
1918 int next_bci = trap->next_bci();
1919 bool modified = trap->modified();
1920
1921 if (next_bci != -1 && !modified) {
1922 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1923 JVMState* jvms = unc->jvms();
1924 ciMethod* method = jvms->method();
1925 ciBytecodeStream iter(method);
1926
1927 iter.force_bci(jvms->bci());
1928 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1929 Bytecodes::Code c = iter.cur_bc();
1930 Node* lhs = nullptr;
1931 Node* rhs = nullptr;
1932 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1933 lhs = unc->peek_operand(0);
1934 rhs = unc->peek_operand(1);
1935 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1936 lhs = unc->peek_operand(0);
1937 }
1938
1939 ResourceMark rm;
1940 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1941 assert(live_locals.is_valid(), "broken liveness info");
1942 int len = (int)live_locals.size();
1943
1944 for (int i = 0; i < len; i++) {
1945 Node* local = unc->local(jvms, i);
1946 // kill local using the liveness of next_bci.
1947 // give up when the local looks like an operand to secure reexecution.
1948 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1949 uint idx = jvms->locoff() + i;
1950 #ifdef ASSERT
1951 if (PrintOpto && Verbose) {
1952 tty->print("[unstable_if] kill local#%d: ", idx);
1953 local->dump();
1954 tty->cr();
1955 }
1956 #endif
1957 igvn.replace_input_of(unc, idx, top());
1958 modified = true;
1959 }
1960 }
1961 }
1962
1963 // keep the mondified trap for late query
1964 if (modified) {
1965 trap->set_modified();
1966 } else {
1967 _unstable_if_traps.delete_at(i);
1968 }
1969 }
1970 igvn.optimize();
1971 }
1972
1973 // StringOpts and late inlining of string methods
1974 void Compile::inline_string_calls(bool parse_time) {
1975 {
1976 // remove useless nodes to make the usage analysis simpler
1977 ResourceMark rm;
1978 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1979 }
1980
1981 {
1982 ResourceMark rm;
1983 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1984 PhaseStringOpts pso(initial_gvn());
1985 print_method(PHASE_AFTER_STRINGOPTS, 3);
1986 }
1987
1988 // now inline anything that we skipped the first time around
1989 if (!parse_time) {
1990 _late_inlines_pos = _late_inlines.length();
1991 }
1992
1993 while (_string_late_inlines.length() > 0) {
1994 CallGenerator* cg = _string_late_inlines.pop();
1995 cg->do_late_inline();
1996 if (failing()) return;
1997 }
1998 _string_late_inlines.trunc_to(0);
1999 }
2000
2001 // Late inlining of boxing methods
2002 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2003 if (_boxing_late_inlines.length() > 0) {
2004 assert(has_boxed_value(), "inconsistent");
2005
2006 set_inlining_incrementally(true);
2007
2008 igvn_worklist()->ensure_empty(); // should be done with igvn
2009
2010 _late_inlines_pos = _late_inlines.length();
2011
2012 while (_boxing_late_inlines.length() > 0) {
2013 CallGenerator* cg = _boxing_late_inlines.pop();
2014 cg->do_late_inline();
2015 if (failing()) return;
2016 }
2017 _boxing_late_inlines.trunc_to(0);
2018
2019 inline_incrementally_cleanup(igvn);
2020
2021 set_inlining_incrementally(false);
2022 }
2023 }
2024
2025 bool Compile::inline_incrementally_one() {
2026 assert(IncrementalInline, "incremental inlining should be on");
2027
2028 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2029
2030 set_inlining_progress(false);
2031 set_do_cleanup(false);
2032
2033 for (int i = 0; i < _late_inlines.length(); i++) {
2034 _late_inlines_pos = i+1;
2035 CallGenerator* cg = _late_inlines.at(i);
2036 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2037 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2038 cg->do_late_inline();
2039 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2040 if (failing()) {
2041 return false;
2042 } else if (inlining_progress()) {
2043 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2044 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2045 break; // process one call site at a time
2046 }
2047 } else {
2048 // Ignore late inline direct calls when inlining is not allowed.
2049 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2050 }
2051 }
2052 // Remove processed elements.
2053 _late_inlines.remove_till(_late_inlines_pos);
2054 _late_inlines_pos = 0;
2055
2056 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2057
2058 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2059
2060 set_inlining_progress(false);
2061 set_do_cleanup(false);
2062
2063 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2064 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2065 }
2066
2067 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2068 {
2069 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2070 ResourceMark rm;
2071 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2072 }
2073 {
2074 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2075 igvn.reset_from_gvn(initial_gvn());
2076 igvn.optimize();
2077 if (failing()) return;
2078 }
2079 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2080 }
2081
2082 // Perform incremental inlining until bound on number of live nodes is reached
2083 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2084 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2085
2086 set_inlining_incrementally(true);
2087 uint low_live_nodes = 0;
2088
2089 while (_late_inlines.length() > 0) {
2090 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2091 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2092 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2093 // PhaseIdealLoop is expensive so we only try it once we are
2094 // out of live nodes and we only try it again if the previous
2095 // helped got the number of nodes down significantly
2096 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2097 if (failing()) return;
2098 low_live_nodes = live_nodes();
2099 _major_progress = true;
2100 }
2101
2102 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2103 bool do_print_inlining = print_inlining() || print_intrinsics();
2104 if (do_print_inlining || log() != nullptr) {
2105 // Print inlining message for candidates that we couldn't inline for lack of space.
2106 for (int i = 0; i < _late_inlines.length(); i++) {
2107 CallGenerator* cg = _late_inlines.at(i);
2108 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2109 if (do_print_inlining) {
2110 cg->print_inlining_late(InliningResult::FAILURE, msg);
2111 }
2112 log_late_inline_failure(cg, msg);
2113 }
2114 }
2115 break; // finish
2116 }
2117 }
2118
2119 igvn_worklist()->ensure_empty(); // should be done with igvn
2120
2121 while (inline_incrementally_one()) {
2122 assert(!failing(), "inconsistent");
2123 }
2124 if (failing()) return;
2125
2126 inline_incrementally_cleanup(igvn);
2127
2128 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2129
2130 if (failing()) return;
2131
2132 if (_late_inlines.length() == 0) {
2133 break; // no more progress
2134 }
2135 }
2136
2137 igvn_worklist()->ensure_empty(); // should be done with igvn
2138
2139 if (_string_late_inlines.length() > 0) {
2140 assert(has_stringbuilder(), "inconsistent");
2141
2142 inline_string_calls(false);
2143
2144 if (failing()) return;
2145
2146 inline_incrementally_cleanup(igvn);
2147 }
2148
2149 set_inlining_incrementally(false);
2150 }
2151
2152 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2153 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2154 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2155 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2156 // as if "inlining_incrementally() == true" were set.
2157 assert(inlining_incrementally() == false, "not allowed");
2158 assert(_modified_nodes == nullptr, "not allowed");
2159 assert(_late_inlines.length() > 0, "sanity");
2160
2161 while (_late_inlines.length() > 0) {
2162 igvn_worklist()->ensure_empty(); // should be done with igvn
2163
2164 while (inline_incrementally_one()) {
2165 assert(!failing(), "inconsistent");
2166 }
2167 if (failing()) return;
2168
2169 inline_incrementally_cleanup(igvn);
2170 }
2171 }
2172
2173 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2174 if (_loop_opts_cnt > 0) {
2175 while (major_progress() && (_loop_opts_cnt > 0)) {
2176 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2177 PhaseIdealLoop::optimize(igvn, mode);
2178 _loop_opts_cnt--;
2179 if (failing()) return false;
2180 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2181 }
2182 }
2183 return true;
2184 }
2185
2186 // Remove edges from "root" to each SafePoint at a backward branch.
2187 // They were inserted during parsing (see add_safepoint()) to make
2188 // infinite loops without calls or exceptions visible to root, i.e.,
2189 // useful.
2190 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2191 Node *r = root();
2192 if (r != nullptr) {
2193 for (uint i = r->req(); i < r->len(); ++i) {
2194 Node *n = r->in(i);
2195 if (n != nullptr && n->is_SafePoint()) {
2196 r->rm_prec(i);
2197 if (n->outcnt() == 0) {
2198 igvn.remove_dead_node(n);
2199 }
2200 --i;
2201 }
2202 }
2203 // Parsing may have added top inputs to the root node (Path
2204 // leading to the Halt node proven dead). Make sure we get a
2205 // chance to clean them up.
2206 igvn._worklist.push(r);
2207 igvn.optimize();
2208 }
2209 }
2210
2211 //------------------------------Optimize---------------------------------------
2212 // Given a graph, optimize it.
2213 void Compile::Optimize() {
2214 TracePhase tp("optimizer", &timers[_t_optimizer]);
2215
2216 #ifndef PRODUCT
2217 if (env()->break_at_compile()) {
2218 BREAKPOINT;
2219 }
2220
2221 #endif
2222
2223 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2224 #ifdef ASSERT
2225 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2226 #endif
2227
2228 ResourceMark rm;
2229
2230 print_inlining_reinit();
2231
2232 NOT_PRODUCT( verify_graph_edges(); )
2233
2234 print_method(PHASE_AFTER_PARSING, 1);
2235
2236 {
2237 // Iterative Global Value Numbering, including ideal transforms
2238 // Initialize IterGVN with types and values from parse-time GVN
2239 PhaseIterGVN igvn(initial_gvn());
2240 #ifdef ASSERT
2241 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2242 #endif
2243 {
2244 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2245 igvn.optimize();
2246 }
2247
2248 if (failing()) return;
2249
2250 print_method(PHASE_ITER_GVN1, 2);
2251
2252 process_for_unstable_if_traps(igvn);
2253
2254 if (failing()) return;
2255
2256 inline_incrementally(igvn);
2257
2258 print_method(PHASE_INCREMENTAL_INLINE, 2);
2259
2260 if (failing()) return;
2261
2262 if (eliminate_boxing()) {
2263 // Inline valueOf() methods now.
2264 inline_boxing_calls(igvn);
2265
2266 if (failing()) return;
2267
2268 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2269 inline_incrementally(igvn);
2270 }
2271
2272 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2273
2274 if (failing()) return;
2275 }
2276
2277 // Remove the speculative part of types and clean up the graph from
2278 // the extra CastPP nodes whose only purpose is to carry them. Do
2279 // that early so that optimizations are not disrupted by the extra
2280 // CastPP nodes.
2281 remove_speculative_types(igvn);
2282
2283 if (failing()) return;
2284
2285 // No more new expensive nodes will be added to the list from here
2286 // so keep only the actual candidates for optimizations.
2287 cleanup_expensive_nodes(igvn);
2288
2289 if (failing()) return;
2290
2291 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2292 if (EnableVectorSupport && has_vbox_nodes()) {
2293 TracePhase tp("", &timers[_t_vector]);
2294 PhaseVector pv(igvn);
2295 pv.optimize_vector_boxes();
2296 if (failing()) return;
2297 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2298 }
2299 assert(!has_vbox_nodes(), "sanity");
2300
2301 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2302 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2303 igvn_worklist()->ensure_empty(); // should be done with igvn
2304 {
2305 ResourceMark rm;
2306 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2307 }
2308 igvn.reset_from_gvn(initial_gvn());
2309 igvn.optimize();
2310 if (failing()) return;
2311 }
2312
2313 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2314 // safepoints
2315 remove_root_to_sfpts_edges(igvn);
2316
2317 if (failing()) return;
2318
2319 // Perform escape analysis
2320 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2321 if (has_loops()) {
2322 // Cleanup graph (remove dead nodes).
2323 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2324 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2325 if (failing()) return;
2326 }
2327 bool progress;
2328 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2329 do {
2330 ConnectionGraph::do_analysis(this, &igvn);
2331
2332 if (failing()) return;
2333
2334 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2335
2336 // Optimize out fields loads from scalar replaceable allocations.
2337 igvn.optimize();
2338 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2339
2340 if (failing()) return;
2341
2342 if (congraph() != nullptr && macro_count() > 0) {
2343 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2344 PhaseMacroExpand mexp(igvn);
2345 mexp.eliminate_macro_nodes();
2346 if (failing()) return;
2347
2348 igvn.set_delay_transform(false);
2349 igvn.optimize();
2350 if (failing()) return;
2351
2352 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2353 }
2354
2355 ConnectionGraph::verify_ram_nodes(this, root());
2356 if (failing()) return;
2357
2358 progress = do_iterative_escape_analysis() &&
2359 (macro_count() < mcount) &&
2360 ConnectionGraph::has_candidates(this);
2361 // Try again if candidates exist and made progress
2362 // by removing some allocations and/or locks.
2363 } while (progress);
2364 }
2365
2366 // Loop transforms on the ideal graph. Range Check Elimination,
2367 // peeling, unrolling, etc.
2368
2369 // Set loop opts counter
2370 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2371 {
2372 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2373 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2374 _loop_opts_cnt--;
2375 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2376 if (failing()) return;
2377 }
2378 // Loop opts pass if partial peeling occurred in previous pass
2379 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2380 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2381 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2382 _loop_opts_cnt--;
2383 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2384 if (failing()) return;
2385 }
2386 // Loop opts pass for loop-unrolling before CCP
2387 if(major_progress() && (_loop_opts_cnt > 0)) {
2388 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2389 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2390 _loop_opts_cnt--;
2391 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2392 }
2393 if (!failing()) {
2394 // Verify that last round of loop opts produced a valid graph
2395 PhaseIdealLoop::verify(igvn);
2396 }
2397 }
2398 if (failing()) return;
2399
2400 // Conditional Constant Propagation;
2401 print_method(PHASE_BEFORE_CCP1, 2);
2402 PhaseCCP ccp( &igvn );
2403 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2404 {
2405 TracePhase tp("ccp", &timers[_t_ccp]);
2406 ccp.do_transform();
2407 }
2408 print_method(PHASE_CCP1, 2);
2409
2410 assert( true, "Break here to ccp.dump_old2new_map()");
2411
2412 // Iterative Global Value Numbering, including ideal transforms
2413 {
2414 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2415 igvn.reset_from_igvn(&ccp);
2416 igvn.optimize();
2417 }
2418 print_method(PHASE_ITER_GVN2, 2);
2419
2420 if (failing()) return;
2421
2422 // Loop transforms on the ideal graph. Range Check Elimination,
2423 // peeling, unrolling, etc.
2424 if (!optimize_loops(igvn, LoopOptsDefault)) {
2425 return;
2426 }
2427
2428 if (failing()) return;
2429
2430 C->clear_major_progress(); // ensure that major progress is now clear
2431
2432 process_for_post_loop_opts_igvn(igvn);
2433
2434 if (failing()) return;
2435
2436 #ifdef ASSERT
2437 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2438 #endif
2439
2440 {
2441 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2442 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2443 PhaseMacroExpand mex(igvn);
2444 if (mex.expand_macro_nodes()) {
2445 assert(failing(), "must bail out w/ explicit message");
2446 return;
2447 }
2448 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2449 }
2450
2451 {
2452 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2453 if (bs->expand_barriers(this, igvn)) {
2454 assert(failing(), "must bail out w/ explicit message");
2455 return;
2456 }
2457 print_method(PHASE_BARRIER_EXPANSION, 2);
2458 }
2459
2460 if (C->max_vector_size() > 0) {
2461 C->optimize_logic_cones(igvn);
2462 igvn.optimize();
2463 if (failing()) return;
2464 }
2465
2466 DEBUG_ONLY( _modified_nodes = nullptr; )
2467
2468 assert(igvn._worklist.size() == 0, "not empty");
2469
2470 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2471
2472 if (_late_inlines.length() > 0) {
2473 // More opportunities to optimize virtual and MH calls.
2474 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2475 process_late_inline_calls_no_inline(igvn);
2476 if (failing()) return;
2477 }
2478 } // (End scope of igvn; run destructor if necessary for asserts.)
2479
2480 check_no_dead_use();
2481
2482 process_print_inlining();
2483
2484 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2485 // to remove hashes to unlock nodes for modifications.
2486 C->node_hash()->clear();
2487
2488 // A method with only infinite loops has no edges entering loops from root
2489 {
2490 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2491 if (final_graph_reshaping()) {
2492 assert(failing(), "must bail out w/ explicit message");
2493 return;
2494 }
2495 }
2496
2497 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2498 DEBUG_ONLY(set_phase_optimize_finished();)
2499 }
2500
2501 #ifdef ASSERT
2502 void Compile::check_no_dead_use() const {
2503 ResourceMark rm;
2504 Unique_Node_List wq;
2505 wq.push(root());
2506 for (uint i = 0; i < wq.size(); ++i) {
2507 Node* n = wq.at(i);
2508 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2509 Node* u = n->fast_out(j);
2510 if (u->outcnt() == 0 && !u->is_Con()) {
2511 u->dump();
2512 fatal("no reachable node should have no use");
2513 }
2514 wq.push(u);
2515 }
2516 }
2517 }
2518 #endif
2519
2520 void Compile::inline_vector_reboxing_calls() {
2521 if (C->_vector_reboxing_late_inlines.length() > 0) {
2522 _late_inlines_pos = C->_late_inlines.length();
2523 while (_vector_reboxing_late_inlines.length() > 0) {
2524 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2525 cg->do_late_inline();
2526 if (failing()) return;
2527 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2528 }
2529 _vector_reboxing_late_inlines.trunc_to(0);
2530 }
2531 }
2532
2533 bool Compile::has_vbox_nodes() {
2534 if (C->_vector_reboxing_late_inlines.length() > 0) {
2535 return true;
2536 }
2537 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2538 Node * n = C->macro_node(macro_idx);
2539 assert(n->is_macro(), "only macro nodes expected here");
2540 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2541 return true;
2542 }
2543 }
2544 return false;
2545 }
2546
2547 //---------------------------- Bitwise operation packing optimization ---------------------------
2548
2549 static bool is_vector_unary_bitwise_op(Node* n) {
2550 return n->Opcode() == Op_XorV &&
2551 VectorNode::is_vector_bitwise_not_pattern(n);
2552 }
2553
2554 static bool is_vector_binary_bitwise_op(Node* n) {
2555 switch (n->Opcode()) {
2556 case Op_AndV:
2557 case Op_OrV:
2558 return true;
2559
2560 case Op_XorV:
2561 return !is_vector_unary_bitwise_op(n);
2562
2563 default:
2564 return false;
2565 }
2566 }
2567
2568 static bool is_vector_ternary_bitwise_op(Node* n) {
2569 return n->Opcode() == Op_MacroLogicV;
2570 }
2571
2572 static bool is_vector_bitwise_op(Node* n) {
2573 return is_vector_unary_bitwise_op(n) ||
2574 is_vector_binary_bitwise_op(n) ||
2575 is_vector_ternary_bitwise_op(n);
2576 }
2577
2578 static bool is_vector_bitwise_cone_root(Node* n) {
2579 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2580 return false;
2581 }
2582 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2583 if (is_vector_bitwise_op(n->fast_out(i))) {
2584 return false;
2585 }
2586 }
2587 return true;
2588 }
2589
2590 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2591 uint cnt = 0;
2592 if (is_vector_bitwise_op(n)) {
2593 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2594 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2595 for (uint i = 1; i < inp_cnt; i++) {
2596 Node* in = n->in(i);
2597 bool skip = VectorNode::is_all_ones_vector(in);
2598 if (!skip && !inputs.member(in)) {
2599 inputs.push(in);
2600 cnt++;
2601 }
2602 }
2603 assert(cnt <= 1, "not unary");
2604 } else {
2605 uint last_req = inp_cnt;
2606 if (is_vector_ternary_bitwise_op(n)) {
2607 last_req = inp_cnt - 1; // skip last input
2608 }
2609 for (uint i = 1; i < last_req; i++) {
2610 Node* def = n->in(i);
2611 if (!inputs.member(def)) {
2612 inputs.push(def);
2613 cnt++;
2614 }
2615 }
2616 }
2617 } else { // not a bitwise operations
2618 if (!inputs.member(n)) {
2619 inputs.push(n);
2620 cnt++;
2621 }
2622 }
2623 return cnt;
2624 }
2625
2626 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2627 Unique_Node_List useful_nodes;
2628 C->identify_useful_nodes(useful_nodes);
2629
2630 for (uint i = 0; i < useful_nodes.size(); i++) {
2631 Node* n = useful_nodes.at(i);
2632 if (is_vector_bitwise_cone_root(n)) {
2633 list.push(n);
2634 }
2635 }
2636 }
2637
2638 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2639 const TypeVect* vt,
2640 Unique_Node_List& partition,
2641 Unique_Node_List& inputs) {
2642 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2643 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2644 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2645
2646 Node* in1 = inputs.at(0);
2647 Node* in2 = inputs.at(1);
2648 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2649
2650 uint func = compute_truth_table(partition, inputs);
2651
2652 Node* pn = partition.at(partition.size() - 1);
2653 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2654 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2655 }
2656
2657 static uint extract_bit(uint func, uint pos) {
2658 return (func & (1 << pos)) >> pos;
2659 }
2660
2661 //
2662 // A macro logic node represents a truth table. It has 4 inputs,
2663 // First three inputs corresponds to 3 columns of a truth table
2664 // and fourth input captures the logic function.
2665 //
2666 // eg. fn = (in1 AND in2) OR in3;
2667 //
2668 // MacroNode(in1,in2,in3,fn)
2669 //
2670 // -----------------
2671 // in1 in2 in3 fn
2672 // -----------------
2673 // 0 0 0 0
2674 // 0 0 1 1
2675 // 0 1 0 0
2676 // 0 1 1 1
2677 // 1 0 0 0
2678 // 1 0 1 1
2679 // 1 1 0 1
2680 // 1 1 1 1
2681 //
2682
2683 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2684 int res = 0;
2685 for (int i = 0; i < 8; i++) {
2686 int bit1 = extract_bit(in1, i);
2687 int bit2 = extract_bit(in2, i);
2688 int bit3 = extract_bit(in3, i);
2689
2690 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2691 int func_bit = extract_bit(func, func_bit_pos);
2692
2693 res |= func_bit << i;
2694 }
2695 return res;
2696 }
2697
2698 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2699 assert(n != nullptr, "");
2700 assert(eval_map.contains(n), "absent");
2701 return *(eval_map.get(n));
2702 }
2703
2704 static void eval_operands(Node* n,
2705 uint& func1, uint& func2, uint& func3,
2706 ResourceHashtable<Node*,uint>& eval_map) {
2707 assert(is_vector_bitwise_op(n), "");
2708
2709 if (is_vector_unary_bitwise_op(n)) {
2710 Node* opnd = n->in(1);
2711 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2712 opnd = n->in(2);
2713 }
2714 func1 = eval_operand(opnd, eval_map);
2715 } else if (is_vector_binary_bitwise_op(n)) {
2716 func1 = eval_operand(n->in(1), eval_map);
2717 func2 = eval_operand(n->in(2), eval_map);
2718 } else {
2719 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2720 func1 = eval_operand(n->in(1), eval_map);
2721 func2 = eval_operand(n->in(2), eval_map);
2722 func3 = eval_operand(n->in(3), eval_map);
2723 }
2724 }
2725
2726 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2727 assert(inputs.size() <= 3, "sanity");
2728 ResourceMark rm;
2729 uint res = 0;
2730 ResourceHashtable<Node*,uint> eval_map;
2731
2732 // Populate precomputed functions for inputs.
2733 // Each input corresponds to one column of 3 input truth-table.
2734 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2735 0xCC, // (_, b, _) -> b
2736 0xF0 }; // (a, _, _) -> a
2737 for (uint i = 0; i < inputs.size(); i++) {
2738 eval_map.put(inputs.at(i), input_funcs[2-i]);
2739 }
2740
2741 for (uint i = 0; i < partition.size(); i++) {
2742 Node* n = partition.at(i);
2743
2744 uint func1 = 0, func2 = 0, func3 = 0;
2745 eval_operands(n, func1, func2, func3, eval_map);
2746
2747 switch (n->Opcode()) {
2748 case Op_OrV:
2749 assert(func3 == 0, "not binary");
2750 res = func1 | func2;
2751 break;
2752 case Op_AndV:
2753 assert(func3 == 0, "not binary");
2754 res = func1 & func2;
2755 break;
2756 case Op_XorV:
2757 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2758 assert(func2 == 0 && func3 == 0, "not unary");
2759 res = (~func1) & 0xFF;
2760 } else {
2761 assert(func3 == 0, "not binary");
2762 res = func1 ^ func2;
2763 }
2764 break;
2765 case Op_MacroLogicV:
2766 // Ordering of inputs may change during evaluation of sub-tree
2767 // containing MacroLogic node as a child node, thus a re-evaluation
2768 // makes sure that function is evaluated in context of current
2769 // inputs.
2770 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2771 break;
2772
2773 default: assert(false, "not supported: %s", n->Name());
2774 }
2775 assert(res <= 0xFF, "invalid");
2776 eval_map.put(n, res);
2777 }
2778 return res;
2779 }
2780
2781 // Criteria under which nodes gets packed into a macro logic node:-
2782 // 1) Parent and both child nodes are all unmasked or masked with
2783 // same predicates.
2784 // 2) Masked parent can be packed with left child if it is predicated
2785 // and both have same predicates.
2786 // 3) Masked parent can be packed with right child if its un-predicated
2787 // or has matching predication condition.
2788 // 4) An unmasked parent can be packed with an unmasked child.
2789 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2790 assert(partition.size() == 0, "not empty");
2791 assert(inputs.size() == 0, "not empty");
2792 if (is_vector_ternary_bitwise_op(n)) {
2793 return false;
2794 }
2795
2796 bool is_unary_op = is_vector_unary_bitwise_op(n);
2797 if (is_unary_op) {
2798 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2799 return false; // too few inputs
2800 }
2801
2802 bool pack_left_child = true;
2803 bool pack_right_child = true;
2804
2805 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2806 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2807
2808 int left_child_input_cnt = 0;
2809 int right_child_input_cnt = 0;
2810
2811 bool parent_is_predicated = n->is_predicated_vector();
2812 bool left_child_predicated = n->in(1)->is_predicated_vector();
2813 bool right_child_predicated = n->in(2)->is_predicated_vector();
2814
2815 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2816 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2817 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2818
2819 do {
2820 if (pack_left_child && left_child_LOP &&
2821 ((!parent_is_predicated && !left_child_predicated) ||
2822 ((parent_is_predicated && left_child_predicated &&
2823 parent_pred == left_child_pred)))) {
2824 partition.push(n->in(1));
2825 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2826 } else {
2827 inputs.push(n->in(1));
2828 left_child_input_cnt = 1;
2829 }
2830
2831 if (pack_right_child && right_child_LOP &&
2832 (!right_child_predicated ||
2833 (right_child_predicated && parent_is_predicated &&
2834 parent_pred == right_child_pred))) {
2835 partition.push(n->in(2));
2836 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2837 } else {
2838 inputs.push(n->in(2));
2839 right_child_input_cnt = 1;
2840 }
2841
2842 if (inputs.size() > 3) {
2843 assert(partition.size() > 0, "");
2844 inputs.clear();
2845 partition.clear();
2846 if (left_child_input_cnt > right_child_input_cnt) {
2847 pack_left_child = false;
2848 } else {
2849 pack_right_child = false;
2850 }
2851 } else {
2852 break;
2853 }
2854 } while(true);
2855
2856 if(partition.size()) {
2857 partition.push(n);
2858 }
2859
2860 return (partition.size() == 2 || partition.size() == 3) &&
2861 (inputs.size() == 2 || inputs.size() == 3);
2862 }
2863
2864 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2865 assert(is_vector_bitwise_op(n), "not a root");
2866
2867 visited.set(n->_idx);
2868
2869 // 1) Do a DFS walk over the logic cone.
2870 for (uint i = 1; i < n->req(); i++) {
2871 Node* in = n->in(i);
2872 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2873 process_logic_cone_root(igvn, in, visited);
2874 }
2875 }
2876
2877 // 2) Bottom up traversal: Merge node[s] with
2878 // the parent to form macro logic node.
2879 Unique_Node_List partition;
2880 Unique_Node_List inputs;
2881 if (compute_logic_cone(n, partition, inputs)) {
2882 const TypeVect* vt = n->bottom_type()->is_vect();
2883 Node* pn = partition.at(partition.size() - 1);
2884 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2885 if (mask == nullptr ||
2886 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2887 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2888 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2889 igvn.replace_node(n, macro_logic);
2890 }
2891 }
2892 }
2893
2894 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2895 ResourceMark rm;
2896 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2897 Unique_Node_List list;
2898 collect_logic_cone_roots(list);
2899
2900 while (list.size() > 0) {
2901 Node* n = list.pop();
2902 const TypeVect* vt = n->bottom_type()->is_vect();
2903 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2904 if (supported) {
2905 VectorSet visited(comp_arena());
2906 process_logic_cone_root(igvn, n, visited);
2907 }
2908 }
2909 }
2910 }
2911
2912 //------------------------------Code_Gen---------------------------------------
2913 // Given a graph, generate code for it
2914 void Compile::Code_Gen() {
2915 if (failing()) {
2916 return;
2917 }
2918
2919 // Perform instruction selection. You might think we could reclaim Matcher
2920 // memory PDQ, but actually the Matcher is used in generating spill code.
2921 // Internals of the Matcher (including some VectorSets) must remain live
2922 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2923 // set a bit in reclaimed memory.
2924
2925 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2926 // nodes. Mapping is only valid at the root of each matched subtree.
2927 NOT_PRODUCT( verify_graph_edges(); )
2928
2929 Matcher matcher;
2930 _matcher = &matcher;
2931 {
2932 TracePhase tp("matcher", &timers[_t_matcher]);
2933 matcher.match();
2934 if (failing()) {
2935 return;
2936 }
2937 }
2938 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2939 // nodes. Mapping is only valid at the root of each matched subtree.
2940 NOT_PRODUCT( verify_graph_edges(); )
2941
2942 // If you have too many nodes, or if matching has failed, bail out
2943 check_node_count(0, "out of nodes matching instructions");
2944 if (failing()) {
2945 return;
2946 }
2947
2948 print_method(PHASE_MATCHING, 2);
2949
2950 // Build a proper-looking CFG
2951 PhaseCFG cfg(node_arena(), root(), matcher);
2952 _cfg = &cfg;
2953 {
2954 TracePhase tp("scheduler", &timers[_t_scheduler]);
2955 bool success = cfg.do_global_code_motion();
2956 if (!success) {
2957 return;
2958 }
2959
2960 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2961 NOT_PRODUCT( verify_graph_edges(); )
2962 cfg.verify();
2963 }
2964
2965 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2966 _regalloc = ®alloc;
2967 {
2968 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2969 // Perform register allocation. After Chaitin, use-def chains are
2970 // no longer accurate (at spill code) and so must be ignored.
2971 // Node->LRG->reg mappings are still accurate.
2972 _regalloc->Register_Allocate();
2973
2974 // Bail out if the allocator builds too many nodes
2975 if (failing()) {
2976 return;
2977 }
2978
2979 print_method(PHASE_REGISTER_ALLOCATION, 2);
2980 }
2981
2982 // Prior to register allocation we kept empty basic blocks in case the
2983 // the allocator needed a place to spill. After register allocation we
2984 // are not adding any new instructions. If any basic block is empty, we
2985 // can now safely remove it.
2986 {
2987 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2988 cfg.remove_empty_blocks();
2989 if (do_freq_based_layout()) {
2990 PhaseBlockLayout layout(cfg);
2991 } else {
2992 cfg.set_loop_alignment();
2993 }
2994 cfg.fixup_flow();
2995 cfg.remove_unreachable_blocks();
2996 cfg.verify_dominator_tree();
2997 print_method(PHASE_BLOCK_ORDERING, 3);
2998 }
2999
3000 // Apply peephole optimizations
3001 if( OptoPeephole ) {
3002 TracePhase tp("peephole", &timers[_t_peephole]);
3003 PhasePeephole peep( _regalloc, cfg);
3004 peep.do_transform();
3005 print_method(PHASE_PEEPHOLE, 3);
3006 }
3007
3008 // Do late expand if CPU requires this.
3009 if (Matcher::require_postalloc_expand) {
3010 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3011 cfg.postalloc_expand(_regalloc);
3012 print_method(PHASE_POSTALLOC_EXPAND, 3);
3013 }
3014
3015 // Convert Nodes to instruction bits in a buffer
3016 {
3017 TracePhase tp("output", &timers[_t_output]);
3018 PhaseOutput output;
3019 output.Output();
3020 if (failing()) return;
3021 output.install();
3022 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3023 }
3024
3025 // He's dead, Jim.
3026 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3027 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3028 }
3029
3030 //------------------------------Final_Reshape_Counts---------------------------
3031 // This class defines counters to help identify when a method
3032 // may/must be executed using hardware with only 24-bit precision.
3033 struct Final_Reshape_Counts : public StackObj {
3034 int _call_count; // count non-inlined 'common' calls
3035 int _float_count; // count float ops requiring 24-bit precision
3036 int _double_count; // count double ops requiring more precision
3037 int _java_call_count; // count non-inlined 'java' calls
3038 int _inner_loop_count; // count loops which need alignment
3039 VectorSet _visited; // Visitation flags
3040 Node_List _tests; // Set of IfNodes & PCTableNodes
3041
3042 Final_Reshape_Counts() :
3043 _call_count(0), _float_count(0), _double_count(0),
3044 _java_call_count(0), _inner_loop_count(0) { }
3045
3046 void inc_call_count () { _call_count ++; }
3047 void inc_float_count () { _float_count ++; }
3048 void inc_double_count() { _double_count++; }
3049 void inc_java_call_count() { _java_call_count++; }
3050 void inc_inner_loop_count() { _inner_loop_count++; }
3051
3052 int get_call_count () const { return _call_count ; }
3053 int get_float_count () const { return _float_count ; }
3054 int get_double_count() const { return _double_count; }
3055 int get_java_call_count() const { return _java_call_count; }
3056 int get_inner_loop_count() const { return _inner_loop_count; }
3057 };
3058
3059 // Eliminate trivially redundant StoreCMs and accumulate their
3060 // precedence edges.
3061 void Compile::eliminate_redundant_card_marks(Node* n) {
3062 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3063 if (n->in(MemNode::Address)->outcnt() > 1) {
3064 // There are multiple users of the same address so it might be
3065 // possible to eliminate some of the StoreCMs
3066 Node* mem = n->in(MemNode::Memory);
3067 Node* adr = n->in(MemNode::Address);
3068 Node* val = n->in(MemNode::ValueIn);
3069 Node* prev = n;
3070 bool done = false;
3071 // Walk the chain of StoreCMs eliminating ones that match. As
3072 // long as it's a chain of single users then the optimization is
3073 // safe. Eliminating partially redundant StoreCMs would require
3074 // cloning copies down the other paths.
3075 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3076 if (adr == mem->in(MemNode::Address) &&
3077 val == mem->in(MemNode::ValueIn)) {
3078 // redundant StoreCM
3079 if (mem->req() > MemNode::OopStore) {
3080 // Hasn't been processed by this code yet.
3081 n->add_prec(mem->in(MemNode::OopStore));
3082 } else {
3083 // Already converted to precedence edge
3084 for (uint i = mem->req(); i < mem->len(); i++) {
3085 // Accumulate any precedence edges
3086 if (mem->in(i) != nullptr) {
3087 n->add_prec(mem->in(i));
3088 }
3089 }
3090 // Everything above this point has been processed.
3091 done = true;
3092 }
3093 // Eliminate the previous StoreCM
3094 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3095 assert(mem->outcnt() == 0, "should be dead");
3096 mem->disconnect_inputs(this);
3097 } else {
3098 prev = mem;
3099 }
3100 mem = prev->in(MemNode::Memory);
3101 }
3102 }
3103 }
3104
3105 //------------------------------final_graph_reshaping_impl----------------------
3106 // Implement items 1-5 from final_graph_reshaping below.
3107 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3108
3109 if ( n->outcnt() == 0 ) return; // dead node
3110 uint nop = n->Opcode();
3111
3112 // Check for 2-input instruction with "last use" on right input.
3113 // Swap to left input. Implements item (2).
3114 if( n->req() == 3 && // two-input instruction
3115 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3116 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3117 n->in(2)->outcnt() == 1 &&// right use IS a last use
3118 !n->in(2)->is_Con() ) { // right use is not a constant
3119 // Check for commutative opcode
3120 switch( nop ) {
3121 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3122 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3123 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3124 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3125 case Op_AndL: case Op_XorL: case Op_OrL:
3126 case Op_AndI: case Op_XorI: case Op_OrI: {
3127 // Move "last use" input to left by swapping inputs
3128 n->swap_edges(1, 2);
3129 break;
3130 }
3131 default:
3132 break;
3133 }
3134 }
3135
3136 #ifdef ASSERT
3137 if( n->is_Mem() ) {
3138 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3139 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3140 // oop will be recorded in oop map if load crosses safepoint
3141 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3142 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3143 "raw memory operations should have control edge");
3144 }
3145 if (n->is_MemBar()) {
3146 MemBarNode* mb = n->as_MemBar();
3147 if (mb->trailing_store() || mb->trailing_load_store()) {
3148 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3149 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3150 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3151 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3152 } else if (mb->leading()) {
3153 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3154 }
3155 }
3156 #endif
3157 // Count FPU ops and common calls, implements item (3)
3158 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3159 if (!gc_handled) {
3160 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3161 }
3162
3163 // Collect CFG split points
3164 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3165 frc._tests.push(n);
3166 }
3167 }
3168
3169 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3170 switch( nop ) {
3171 // Count all float operations that may use FPU
3172 case Op_AddF:
3173 case Op_SubF:
3174 case Op_MulF:
3175 case Op_DivF:
3176 case Op_NegF:
3177 case Op_ModF:
3178 case Op_ConvI2F:
3179 case Op_ConF:
3180 case Op_CmpF:
3181 case Op_CmpF3:
3182 case Op_StoreF:
3183 case Op_LoadF:
3184 // case Op_ConvL2F: // longs are split into 32-bit halves
3185 frc.inc_float_count();
3186 break;
3187
3188 case Op_ConvF2D:
3189 case Op_ConvD2F:
3190 frc.inc_float_count();
3191 frc.inc_double_count();
3192 break;
3193
3194 // Count all double operations that may use FPU
3195 case Op_AddD:
3196 case Op_SubD:
3197 case Op_MulD:
3198 case Op_DivD:
3199 case Op_NegD:
3200 case Op_ModD:
3201 case Op_ConvI2D:
3202 case Op_ConvD2I:
3203 // case Op_ConvL2D: // handled by leaf call
3204 // case Op_ConvD2L: // handled by leaf call
3205 case Op_ConD:
3206 case Op_CmpD:
3207 case Op_CmpD3:
3208 case Op_StoreD:
3209 case Op_LoadD:
3210 case Op_LoadD_unaligned:
3211 frc.inc_double_count();
3212 break;
3213 case Op_Opaque1: // Remove Opaque Nodes before matching
3214 n->subsume_by(n->in(1), this);
3215 break;
3216 case Op_CallStaticJava:
3217 case Op_CallJava:
3218 case Op_CallDynamicJava:
3219 frc.inc_java_call_count(); // Count java call site;
3220 case Op_CallRuntime:
3221 case Op_CallLeaf:
3222 case Op_CallLeafVector:
3223 case Op_CallLeafNoFP: {
3224 assert (n->is_Call(), "");
3225 CallNode *call = n->as_Call();
3226 // Count call sites where the FP mode bit would have to be flipped.
3227 // Do not count uncommon runtime calls:
3228 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3229 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3230 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3231 frc.inc_call_count(); // Count the call site
3232 } else { // See if uncommon argument is shared
3233 Node *n = call->in(TypeFunc::Parms);
3234 int nop = n->Opcode();
3235 // Clone shared simple arguments to uncommon calls, item (1).
3236 if (n->outcnt() > 1 &&
3237 !n->is_Proj() &&
3238 nop != Op_CreateEx &&
3239 nop != Op_CheckCastPP &&
3240 nop != Op_DecodeN &&
3241 nop != Op_DecodeNKlass &&
3242 !n->is_Mem() &&
3243 !n->is_Phi()) {
3244 Node *x = n->clone();
3245 call->set_req(TypeFunc::Parms, x);
3246 }
3247 }
3248 break;
3249 }
3250
3251 case Op_StoreCM:
3252 {
3253 // Convert OopStore dependence into precedence edge
3254 Node* prec = n->in(MemNode::OopStore);
3255 n->del_req(MemNode::OopStore);
3256 n->add_prec(prec);
3257 eliminate_redundant_card_marks(n);
3258 }
3259
3260 // fall through
3261
3262 case Op_StoreB:
3263 case Op_StoreC:
3264 case Op_StoreI:
3265 case Op_StoreL:
3266 case Op_CompareAndSwapB:
3267 case Op_CompareAndSwapS:
3268 case Op_CompareAndSwapI:
3269 case Op_CompareAndSwapL:
3270 case Op_CompareAndSwapP:
3271 case Op_CompareAndSwapN:
3272 case Op_WeakCompareAndSwapB:
3273 case Op_WeakCompareAndSwapS:
3274 case Op_WeakCompareAndSwapI:
3275 case Op_WeakCompareAndSwapL:
3276 case Op_WeakCompareAndSwapP:
3277 case Op_WeakCompareAndSwapN:
3278 case Op_CompareAndExchangeB:
3279 case Op_CompareAndExchangeS:
3280 case Op_CompareAndExchangeI:
3281 case Op_CompareAndExchangeL:
3282 case Op_CompareAndExchangeP:
3283 case Op_CompareAndExchangeN:
3284 case Op_GetAndAddS:
3285 case Op_GetAndAddB:
3286 case Op_GetAndAddI:
3287 case Op_GetAndAddL:
3288 case Op_GetAndSetS:
3289 case Op_GetAndSetB:
3290 case Op_GetAndSetI:
3291 case Op_GetAndSetL:
3292 case Op_GetAndSetP:
3293 case Op_GetAndSetN:
3294 case Op_StoreP:
3295 case Op_StoreN:
3296 case Op_StoreNKlass:
3297 case Op_LoadB:
3298 case Op_LoadUB:
3299 case Op_LoadUS:
3300 case Op_LoadI:
3301 case Op_LoadKlass:
3302 case Op_LoadNKlass:
3303 case Op_LoadL:
3304 case Op_LoadL_unaligned:
3305 case Op_LoadP:
3306 case Op_LoadN:
3307 case Op_LoadRange:
3308 case Op_LoadS:
3309 break;
3310
3311 case Op_AddP: { // Assert sane base pointers
3312 Node *addp = n->in(AddPNode::Address);
3313 assert( !addp->is_AddP() ||
3314 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3315 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3316 "Base pointers must match (addp %u)", addp->_idx );
3317 #ifdef _LP64
3318 if ((UseCompressedOops || UseCompressedClassPointers) &&
3319 addp->Opcode() == Op_ConP &&
3320 addp == n->in(AddPNode::Base) &&
3321 n->in(AddPNode::Offset)->is_Con()) {
3322 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3323 // on the platform and on the compressed oops mode.
3324 // Use addressing with narrow klass to load with offset on x86.
3325 // Some platforms can use the constant pool to load ConP.
3326 // Do this transformation here since IGVN will convert ConN back to ConP.
3327 const Type* t = addp->bottom_type();
3328 bool is_oop = t->isa_oopptr() != nullptr;
3329 bool is_klass = t->isa_klassptr() != nullptr;
3330
3331 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3332 (is_klass && Matcher::const_klass_prefer_decode())) {
3333 Node* nn = nullptr;
3334
3335 int op = is_oop ? Op_ConN : Op_ConNKlass;
3336
3337 // Look for existing ConN node of the same exact type.
3338 Node* r = root();
3339 uint cnt = r->outcnt();
3340 for (uint i = 0; i < cnt; i++) {
3341 Node* m = r->raw_out(i);
3342 if (m!= nullptr && m->Opcode() == op &&
3343 m->bottom_type()->make_ptr() == t) {
3344 nn = m;
3345 break;
3346 }
3347 }
3348 if (nn != nullptr) {
3349 // Decode a narrow oop to match address
3350 // [R12 + narrow_oop_reg<<3 + offset]
3351 if (is_oop) {
3352 nn = new DecodeNNode(nn, t);
3353 } else {
3354 nn = new DecodeNKlassNode(nn, t);
3355 }
3356 // Check for succeeding AddP which uses the same Base.
3357 // Otherwise we will run into the assertion above when visiting that guy.
3358 for (uint i = 0; i < n->outcnt(); ++i) {
3359 Node *out_i = n->raw_out(i);
3360 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3361 out_i->set_req(AddPNode::Base, nn);
3362 #ifdef ASSERT
3363 for (uint j = 0; j < out_i->outcnt(); ++j) {
3364 Node *out_j = out_i->raw_out(j);
3365 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3366 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3367 }
3368 #endif
3369 }
3370 }
3371 n->set_req(AddPNode::Base, nn);
3372 n->set_req(AddPNode::Address, nn);
3373 if (addp->outcnt() == 0) {
3374 addp->disconnect_inputs(this);
3375 }
3376 }
3377 }
3378 }
3379 #endif
3380 break;
3381 }
3382
3383 case Op_CastPP: {
3384 // Remove CastPP nodes to gain more freedom during scheduling but
3385 // keep the dependency they encode as control or precedence edges
3386 // (if control is set already) on memory operations. Some CastPP
3387 // nodes don't have a control (don't carry a dependency): skip
3388 // those.
3389 if (n->in(0) != nullptr) {
3390 ResourceMark rm;
3391 Unique_Node_List wq;
3392 wq.push(n);
3393 for (uint next = 0; next < wq.size(); ++next) {
3394 Node *m = wq.at(next);
3395 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3396 Node* use = m->fast_out(i);
3397 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3398 use->ensure_control_or_add_prec(n->in(0));
3399 } else {
3400 switch(use->Opcode()) {
3401 case Op_AddP:
3402 case Op_DecodeN:
3403 case Op_DecodeNKlass:
3404 case Op_CheckCastPP:
3405 case Op_CastPP:
3406 wq.push(use);
3407 break;
3408 }
3409 }
3410 }
3411 }
3412 }
3413 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3414 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3415 Node* in1 = n->in(1);
3416 const Type* t = n->bottom_type();
3417 Node* new_in1 = in1->clone();
3418 new_in1->as_DecodeN()->set_type(t);
3419
3420 if (!Matcher::narrow_oop_use_complex_address()) {
3421 //
3422 // x86, ARM and friends can handle 2 adds in addressing mode
3423 // and Matcher can fold a DecodeN node into address by using
3424 // a narrow oop directly and do implicit null check in address:
3425 //
3426 // [R12 + narrow_oop_reg<<3 + offset]
3427 // NullCheck narrow_oop_reg
3428 //
3429 // On other platforms (Sparc) we have to keep new DecodeN node and
3430 // use it to do implicit null check in address:
3431 //
3432 // decode_not_null narrow_oop_reg, base_reg
3433 // [base_reg + offset]
3434 // NullCheck base_reg
3435 //
3436 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3437 // to keep the information to which null check the new DecodeN node
3438 // corresponds to use it as value in implicit_null_check().
3439 //
3440 new_in1->set_req(0, n->in(0));
3441 }
3442
3443 n->subsume_by(new_in1, this);
3444 if (in1->outcnt() == 0) {
3445 in1->disconnect_inputs(this);
3446 }
3447 } else {
3448 n->subsume_by(n->in(1), this);
3449 if (n->outcnt() == 0) {
3450 n->disconnect_inputs(this);
3451 }
3452 }
3453 break;
3454 }
3455 #ifdef _LP64
3456 case Op_CmpP:
3457 // Do this transformation here to preserve CmpPNode::sub() and
3458 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3459 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3460 Node* in1 = n->in(1);
3461 Node* in2 = n->in(2);
3462 if (!in1->is_DecodeNarrowPtr()) {
3463 in2 = in1;
3464 in1 = n->in(2);
3465 }
3466 assert(in1->is_DecodeNarrowPtr(), "sanity");
3467
3468 Node* new_in2 = nullptr;
3469 if (in2->is_DecodeNarrowPtr()) {
3470 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3471 new_in2 = in2->in(1);
3472 } else if (in2->Opcode() == Op_ConP) {
3473 const Type* t = in2->bottom_type();
3474 if (t == TypePtr::NULL_PTR) {
3475 assert(in1->is_DecodeN(), "compare klass to null?");
3476 // Don't convert CmpP null check into CmpN if compressed
3477 // oops implicit null check is not generated.
3478 // This will allow to generate normal oop implicit null check.
3479 if (Matcher::gen_narrow_oop_implicit_null_checks())
3480 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3481 //
3482 // This transformation together with CastPP transformation above
3483 // will generated code for implicit null checks for compressed oops.
3484 //
3485 // The original code after Optimize()
3486 //
3487 // LoadN memory, narrow_oop_reg
3488 // decode narrow_oop_reg, base_reg
3489 // CmpP base_reg, nullptr
3490 // CastPP base_reg // NotNull
3491 // Load [base_reg + offset], val_reg
3492 //
3493 // after these transformations will be
3494 //
3495 // LoadN memory, narrow_oop_reg
3496 // CmpN narrow_oop_reg, nullptr
3497 // decode_not_null narrow_oop_reg, base_reg
3498 // Load [base_reg + offset], val_reg
3499 //
3500 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3501 // since narrow oops can be used in debug info now (see the code in
3502 // final_graph_reshaping_walk()).
3503 //
3504 // At the end the code will be matched to
3505 // on x86:
3506 //
3507 // Load_narrow_oop memory, narrow_oop_reg
3508 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3509 // NullCheck narrow_oop_reg
3510 //
3511 // and on sparc:
3512 //
3513 // Load_narrow_oop memory, narrow_oop_reg
3514 // decode_not_null narrow_oop_reg, base_reg
3515 // Load [base_reg + offset], val_reg
3516 // NullCheck base_reg
3517 //
3518 } else if (t->isa_oopptr()) {
3519 new_in2 = ConNode::make(t->make_narrowoop());
3520 } else if (t->isa_klassptr()) {
3521 new_in2 = ConNode::make(t->make_narrowklass());
3522 }
3523 }
3524 if (new_in2 != nullptr) {
3525 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3526 n->subsume_by(cmpN, this);
3527 if (in1->outcnt() == 0) {
3528 in1->disconnect_inputs(this);
3529 }
3530 if (in2->outcnt() == 0) {
3531 in2->disconnect_inputs(this);
3532 }
3533 }
3534 }
3535 break;
3536
3537 case Op_DecodeN:
3538 case Op_DecodeNKlass:
3539 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3540 // DecodeN could be pinned when it can't be fold into
3541 // an address expression, see the code for Op_CastPP above.
3542 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3543 break;
3544
3545 case Op_EncodeP:
3546 case Op_EncodePKlass: {
3547 Node* in1 = n->in(1);
3548 if (in1->is_DecodeNarrowPtr()) {
3549 n->subsume_by(in1->in(1), this);
3550 } else if (in1->Opcode() == Op_ConP) {
3551 const Type* t = in1->bottom_type();
3552 if (t == TypePtr::NULL_PTR) {
3553 assert(t->isa_oopptr(), "null klass?");
3554 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3555 } else if (t->isa_oopptr()) {
3556 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3557 } else if (t->isa_klassptr()) {
3558 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3559 }
3560 }
3561 if (in1->outcnt() == 0) {
3562 in1->disconnect_inputs(this);
3563 }
3564 break;
3565 }
3566
3567 case Op_Proj: {
3568 if (OptimizeStringConcat || IncrementalInline) {
3569 ProjNode* proj = n->as_Proj();
3570 if (proj->_is_io_use) {
3571 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3572 // Separate projections were used for the exception path which
3573 // are normally removed by a late inline. If it wasn't inlined
3574 // then they will hang around and should just be replaced with
3575 // the original one. Merge them.
3576 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3577 if (non_io_proj != nullptr) {
3578 proj->subsume_by(non_io_proj , this);
3579 }
3580 }
3581 }
3582 break;
3583 }
3584
3585 case Op_Phi:
3586 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3587 // The EncodeP optimization may create Phi with the same edges
3588 // for all paths. It is not handled well by Register Allocator.
3589 Node* unique_in = n->in(1);
3590 assert(unique_in != nullptr, "");
3591 uint cnt = n->req();
3592 for (uint i = 2; i < cnt; i++) {
3593 Node* m = n->in(i);
3594 assert(m != nullptr, "");
3595 if (unique_in != m)
3596 unique_in = nullptr;
3597 }
3598 if (unique_in != nullptr) {
3599 n->subsume_by(unique_in, this);
3600 }
3601 }
3602 break;
3603
3604 #endif
3605
3606 #ifdef ASSERT
3607 case Op_CastII:
3608 // Verify that all range check dependent CastII nodes were removed.
3609 if (n->isa_CastII()->has_range_check()) {
3610 n->dump(3);
3611 assert(false, "Range check dependent CastII node was not removed");
3612 }
3613 break;
3614 #endif
3615
3616 case Op_ModI:
3617 if (UseDivMod) {
3618 // Check if a%b and a/b both exist
3619 Node* d = n->find_similar(Op_DivI);
3620 if (d) {
3621 // Replace them with a fused divmod if supported
3622 if (Matcher::has_match_rule(Op_DivModI)) {
3623 DivModINode* divmod = DivModINode::make(n);
3624 d->subsume_by(divmod->div_proj(), this);
3625 n->subsume_by(divmod->mod_proj(), this);
3626 } else {
3627 // replace a%b with a-((a/b)*b)
3628 Node* mult = new MulINode(d, d->in(2));
3629 Node* sub = new SubINode(d->in(1), mult);
3630 n->subsume_by(sub, this);
3631 }
3632 }
3633 }
3634 break;
3635
3636 case Op_ModL:
3637 if (UseDivMod) {
3638 // Check if a%b and a/b both exist
3639 Node* d = n->find_similar(Op_DivL);
3640 if (d) {
3641 // Replace them with a fused divmod if supported
3642 if (Matcher::has_match_rule(Op_DivModL)) {
3643 DivModLNode* divmod = DivModLNode::make(n);
3644 d->subsume_by(divmod->div_proj(), this);
3645 n->subsume_by(divmod->mod_proj(), this);
3646 } else {
3647 // replace a%b with a-((a/b)*b)
3648 Node* mult = new MulLNode(d, d->in(2));
3649 Node* sub = new SubLNode(d->in(1), mult);
3650 n->subsume_by(sub, this);
3651 }
3652 }
3653 }
3654 break;
3655
3656 case Op_UModI:
3657 if (UseDivMod) {
3658 // Check if a%b and a/b both exist
3659 Node* d = n->find_similar(Op_UDivI);
3660 if (d) {
3661 // Replace them with a fused unsigned divmod if supported
3662 if (Matcher::has_match_rule(Op_UDivModI)) {
3663 UDivModINode* divmod = UDivModINode::make(n);
3664 d->subsume_by(divmod->div_proj(), this);
3665 n->subsume_by(divmod->mod_proj(), this);
3666 } else {
3667 // replace a%b with a-((a/b)*b)
3668 Node* mult = new MulINode(d, d->in(2));
3669 Node* sub = new SubINode(d->in(1), mult);
3670 n->subsume_by(sub, this);
3671 }
3672 }
3673 }
3674 break;
3675
3676 case Op_UModL:
3677 if (UseDivMod) {
3678 // Check if a%b and a/b both exist
3679 Node* d = n->find_similar(Op_UDivL);
3680 if (d) {
3681 // Replace them with a fused unsigned divmod if supported
3682 if (Matcher::has_match_rule(Op_UDivModL)) {
3683 UDivModLNode* divmod = UDivModLNode::make(n);
3684 d->subsume_by(divmod->div_proj(), this);
3685 n->subsume_by(divmod->mod_proj(), this);
3686 } else {
3687 // replace a%b with a-((a/b)*b)
3688 Node* mult = new MulLNode(d, d->in(2));
3689 Node* sub = new SubLNode(d->in(1), mult);
3690 n->subsume_by(sub, this);
3691 }
3692 }
3693 }
3694 break;
3695
3696 case Op_LoadVector:
3697 case Op_StoreVector:
3698 #ifdef ASSERT
3699 // Add VerifyVectorAlignment node between adr and load / store.
3700 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3701 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3702 n->as_StoreVector()->must_verify_alignment();
3703 if (must_verify_alignment) {
3704 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3705 n->as_StoreVector()->memory_size();
3706 // The memory access should be aligned to the vector width in bytes.
3707 // However, the underlying array is possibly less well aligned, but at least
3708 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3709 // a loop we can expect at least the following alignment:
3710 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3711 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3712 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3713 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3714 jlong mask = guaranteed_alignment - 1;
3715 Node* mask_con = ConLNode::make(mask);
3716 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3717 n->set_req(MemNode::Address, va);
3718 }
3719 }
3720 #endif
3721 break;
3722
3723 case Op_LoadVectorGather:
3724 case Op_StoreVectorScatter:
3725 case Op_LoadVectorGatherMasked:
3726 case Op_StoreVectorScatterMasked:
3727 case Op_VectorCmpMasked:
3728 case Op_VectorMaskGen:
3729 case Op_LoadVectorMasked:
3730 case Op_StoreVectorMasked:
3731 break;
3732
3733 case Op_AddReductionVI:
3734 case Op_AddReductionVL:
3735 case Op_AddReductionVF:
3736 case Op_AddReductionVD:
3737 case Op_MulReductionVI:
3738 case Op_MulReductionVL:
3739 case Op_MulReductionVF:
3740 case Op_MulReductionVD:
3741 case Op_MinReductionV:
3742 case Op_MaxReductionV:
3743 case Op_AndReductionV:
3744 case Op_OrReductionV:
3745 case Op_XorReductionV:
3746 break;
3747
3748 case Op_PackB:
3749 case Op_PackS:
3750 case Op_PackI:
3751 case Op_PackF:
3752 case Op_PackL:
3753 case Op_PackD:
3754 if (n->req()-1 > 2) {
3755 // Replace many operand PackNodes with a binary tree for matching
3756 PackNode* p = (PackNode*) n;
3757 Node* btp = p->binary_tree_pack(1, n->req());
3758 n->subsume_by(btp, this);
3759 }
3760 break;
3761 case Op_Loop:
3762 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3763 case Op_CountedLoop:
3764 case Op_LongCountedLoop:
3765 case Op_OuterStripMinedLoop:
3766 if (n->as_Loop()->is_inner_loop()) {
3767 frc.inc_inner_loop_count();
3768 }
3769 n->as_Loop()->verify_strip_mined(0);
3770 break;
3771 case Op_LShiftI:
3772 case Op_RShiftI:
3773 case Op_URShiftI:
3774 case Op_LShiftL:
3775 case Op_RShiftL:
3776 case Op_URShiftL:
3777 if (Matcher::need_masked_shift_count) {
3778 // The cpu's shift instructions don't restrict the count to the
3779 // lower 5/6 bits. We need to do the masking ourselves.
3780 Node* in2 = n->in(2);
3781 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3782 const TypeInt* t = in2->find_int_type();
3783 if (t != nullptr && t->is_con()) {
3784 juint shift = t->get_con();
3785 if (shift > mask) { // Unsigned cmp
3786 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3787 }
3788 } else {
3789 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3790 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3791 n->set_req(2, shift);
3792 }
3793 }
3794 if (in2->outcnt() == 0) { // Remove dead node
3795 in2->disconnect_inputs(this);
3796 }
3797 }
3798 break;
3799 case Op_MemBarStoreStore:
3800 case Op_MemBarRelease:
3801 // Break the link with AllocateNode: it is no longer useful and
3802 // confuses register allocation.
3803 if (n->req() > MemBarNode::Precedent) {
3804 n->set_req(MemBarNode::Precedent, top());
3805 }
3806 break;
3807 case Op_MemBarAcquire: {
3808 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3809 // At parse time, the trailing MemBarAcquire for a volatile load
3810 // is created with an edge to the load. After optimizations,
3811 // that input may be a chain of Phis. If those phis have no
3812 // other use, then the MemBarAcquire keeps them alive and
3813 // register allocation can be confused.
3814 dead_nodes.push(n->in(MemBarNode::Precedent));
3815 n->set_req(MemBarNode::Precedent, top());
3816 }
3817 break;
3818 }
3819 case Op_Blackhole:
3820 break;
3821 case Op_RangeCheck: {
3822 RangeCheckNode* rc = n->as_RangeCheck();
3823 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3824 n->subsume_by(iff, this);
3825 frc._tests.push(iff);
3826 break;
3827 }
3828 case Op_ConvI2L: {
3829 if (!Matcher::convi2l_type_required) {
3830 // Code generation on some platforms doesn't need accurate
3831 // ConvI2L types. Widening the type can help remove redundant
3832 // address computations.
3833 n->as_Type()->set_type(TypeLong::INT);
3834 ResourceMark rm;
3835 Unique_Node_List wq;
3836 wq.push(n);
3837 for (uint next = 0; next < wq.size(); next++) {
3838 Node *m = wq.at(next);
3839
3840 for(;;) {
3841 // Loop over all nodes with identical inputs edges as m
3842 Node* k = m->find_similar(m->Opcode());
3843 if (k == nullptr) {
3844 break;
3845 }
3846 // Push their uses so we get a chance to remove node made
3847 // redundant
3848 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3849 Node* u = k->fast_out(i);
3850 if (u->Opcode() == Op_LShiftL ||
3851 u->Opcode() == Op_AddL ||
3852 u->Opcode() == Op_SubL ||
3853 u->Opcode() == Op_AddP) {
3854 wq.push(u);
3855 }
3856 }
3857 // Replace all nodes with identical edges as m with m
3858 k->subsume_by(m, this);
3859 }
3860 }
3861 }
3862 break;
3863 }
3864 case Op_CmpUL: {
3865 if (!Matcher::has_match_rule(Op_CmpUL)) {
3866 // No support for unsigned long comparisons
3867 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3868 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3869 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3870 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3871 Node* andl = new AndLNode(orl, remove_sign_mask);
3872 Node* cmp = new CmpLNode(andl, n->in(2));
3873 n->subsume_by(cmp, this);
3874 }
3875 break;
3876 }
3877 default:
3878 assert(!n->is_Call(), "");
3879 assert(!n->is_Mem(), "");
3880 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3881 break;
3882 }
3883 }
3884
3885 //------------------------------final_graph_reshaping_walk---------------------
3886 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3887 // requires that the walk visits a node's inputs before visiting the node.
3888 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3889 Unique_Node_List sfpt;
3890
3891 frc._visited.set(root->_idx); // first, mark node as visited
3892 uint cnt = root->req();
3893 Node *n = root;
3894 uint i = 0;
3895 while (true) {
3896 if (i < cnt) {
3897 // Place all non-visited non-null inputs onto stack
3898 Node* m = n->in(i);
3899 ++i;
3900 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3901 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3902 // compute worst case interpreter size in case of a deoptimization
3903 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3904
3905 sfpt.push(m);
3906 }
3907 cnt = m->req();
3908 nstack.push(n, i); // put on stack parent and next input's index
3909 n = m;
3910 i = 0;
3911 }
3912 } else {
3913 // Now do post-visit work
3914 final_graph_reshaping_impl(n, frc, dead_nodes);
3915 if (nstack.is_empty())
3916 break; // finished
3917 n = nstack.node(); // Get node from stack
3918 cnt = n->req();
3919 i = nstack.index();
3920 nstack.pop(); // Shift to the next node on stack
3921 }
3922 }
3923
3924 // Skip next transformation if compressed oops are not used.
3925 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3926 (!UseCompressedOops && !UseCompressedClassPointers))
3927 return;
3928
3929 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3930 // It could be done for an uncommon traps or any safepoints/calls
3931 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3932 while (sfpt.size() > 0) {
3933 n = sfpt.pop();
3934 JVMState *jvms = n->as_SafePoint()->jvms();
3935 assert(jvms != nullptr, "sanity");
3936 int start = jvms->debug_start();
3937 int end = n->req();
3938 bool is_uncommon = (n->is_CallStaticJava() &&
3939 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3940 for (int j = start; j < end; j++) {
3941 Node* in = n->in(j);
3942 if (in->is_DecodeNarrowPtr()) {
3943 bool safe_to_skip = true;
3944 if (!is_uncommon ) {
3945 // Is it safe to skip?
3946 for (uint i = 0; i < in->outcnt(); i++) {
3947 Node* u = in->raw_out(i);
3948 if (!u->is_SafePoint() ||
3949 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3950 safe_to_skip = false;
3951 }
3952 }
3953 }
3954 if (safe_to_skip) {
3955 n->set_req(j, in->in(1));
3956 }
3957 if (in->outcnt() == 0) {
3958 in->disconnect_inputs(this);
3959 }
3960 }
3961 }
3962 }
3963 }
3964
3965 //------------------------------final_graph_reshaping--------------------------
3966 // Final Graph Reshaping.
3967 //
3968 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3969 // and not commoned up and forced early. Must come after regular
3970 // optimizations to avoid GVN undoing the cloning. Clone constant
3971 // inputs to Loop Phis; these will be split by the allocator anyways.
3972 // Remove Opaque nodes.
3973 // (2) Move last-uses by commutative operations to the left input to encourage
3974 // Intel update-in-place two-address operations and better register usage
3975 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3976 // calls canonicalizing them back.
3977 // (3) Count the number of double-precision FP ops, single-precision FP ops
3978 // and call sites. On Intel, we can get correct rounding either by
3979 // forcing singles to memory (requires extra stores and loads after each
3980 // FP bytecode) or we can set a rounding mode bit (requires setting and
3981 // clearing the mode bit around call sites). The mode bit is only used
3982 // if the relative frequency of single FP ops to calls is low enough.
3983 // This is a key transform for SPEC mpeg_audio.
3984 // (4) Detect infinite loops; blobs of code reachable from above but not
3985 // below. Several of the Code_Gen algorithms fail on such code shapes,
3986 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3987 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3988 // Detection is by looking for IfNodes where only 1 projection is
3989 // reachable from below or CatchNodes missing some targets.
3990 // (5) Assert for insane oop offsets in debug mode.
3991
3992 bool Compile::final_graph_reshaping() {
3993 // an infinite loop may have been eliminated by the optimizer,
3994 // in which case the graph will be empty.
3995 if (root()->req() == 1) {
3996 // Do not compile method that is only a trivial infinite loop,
3997 // since the content of the loop may have been eliminated.
3998 record_method_not_compilable("trivial infinite loop");
3999 return true;
4000 }
4001
4002 // Expensive nodes have their control input set to prevent the GVN
4003 // from freely commoning them. There's no GVN beyond this point so
4004 // no need to keep the control input. We want the expensive nodes to
4005 // be freely moved to the least frequent code path by gcm.
4006 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4007 for (int i = 0; i < expensive_count(); i++) {
4008 _expensive_nodes.at(i)->set_req(0, nullptr);
4009 }
4010
4011 Final_Reshape_Counts frc;
4012
4013 // Visit everybody reachable!
4014 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4015 Node_Stack nstack(live_nodes() >> 1);
4016 Unique_Node_List dead_nodes;
4017 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4018
4019 // Check for unreachable (from below) code (i.e., infinite loops).
4020 for( uint i = 0; i < frc._tests.size(); i++ ) {
4021 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4022 // Get number of CFG targets.
4023 // Note that PCTables include exception targets after calls.
4024 uint required_outcnt = n->required_outcnt();
4025 if (n->outcnt() != required_outcnt) {
4026 // Check for a few special cases. Rethrow Nodes never take the
4027 // 'fall-thru' path, so expected kids is 1 less.
4028 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4029 if (n->in(0)->in(0)->is_Call()) {
4030 CallNode* call = n->in(0)->in(0)->as_Call();
4031 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4032 required_outcnt--; // Rethrow always has 1 less kid
4033 } else if (call->req() > TypeFunc::Parms &&
4034 call->is_CallDynamicJava()) {
4035 // Check for null receiver. In such case, the optimizer has
4036 // detected that the virtual call will always result in a null
4037 // pointer exception. The fall-through projection of this CatchNode
4038 // will not be populated.
4039 Node* arg0 = call->in(TypeFunc::Parms);
4040 if (arg0->is_Type() &&
4041 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4042 required_outcnt--;
4043 }
4044 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4045 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4046 // Check for illegal array length. In such case, the optimizer has
4047 // detected that the allocation attempt will always result in an
4048 // exception. There is no fall-through projection of this CatchNode .
4049 assert(call->is_CallStaticJava(), "static call expected");
4050 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4051 uint valid_length_test_input = call->req() - 1;
4052 Node* valid_length_test = call->in(valid_length_test_input);
4053 call->del_req(valid_length_test_input);
4054 if (valid_length_test->find_int_con(1) == 0) {
4055 required_outcnt--;
4056 }
4057 dead_nodes.push(valid_length_test);
4058 assert(n->outcnt() == required_outcnt, "malformed control flow");
4059 continue;
4060 }
4061 }
4062 }
4063
4064 // Recheck with a better notion of 'required_outcnt'
4065 if (n->outcnt() != required_outcnt) {
4066 record_method_not_compilable("malformed control flow");
4067 return true; // Not all targets reachable!
4068 }
4069 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4070 CallNode* call = n->in(0)->in(0)->as_Call();
4071 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4072 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4073 assert(call->is_CallStaticJava(), "static call expected");
4074 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4075 uint valid_length_test_input = call->req() - 1;
4076 dead_nodes.push(call->in(valid_length_test_input));
4077 call->del_req(valid_length_test_input); // valid length test useless now
4078 }
4079 }
4080 // Check that I actually visited all kids. Unreached kids
4081 // must be infinite loops.
4082 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4083 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4084 record_method_not_compilable("infinite loop");
4085 return true; // Found unvisited kid; must be unreach
4086 }
4087
4088 // Here so verification code in final_graph_reshaping_walk()
4089 // always see an OuterStripMinedLoopEnd
4090 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4091 IfNode* init_iff = n->as_If();
4092 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4093 n->subsume_by(iff, this);
4094 }
4095 }
4096
4097 while (dead_nodes.size() > 0) {
4098 Node* m = dead_nodes.pop();
4099 if (m->outcnt() == 0 && m != top()) {
4100 for (uint j = 0; j < m->req(); j++) {
4101 Node* in = m->in(j);
4102 if (in != nullptr) {
4103 dead_nodes.push(in);
4104 }
4105 }
4106 m->disconnect_inputs(this);
4107 }
4108 }
4109
4110 #ifdef IA32
4111 // If original bytecodes contained a mixture of floats and doubles
4112 // check if the optimizer has made it homogeneous, item (3).
4113 if (UseSSE == 0 &&
4114 frc.get_float_count() > 32 &&
4115 frc.get_double_count() == 0 &&
4116 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4117 set_24_bit_selection_and_mode(false, true);
4118 }
4119 #endif // IA32
4120
4121 set_java_calls(frc.get_java_call_count());
4122 set_inner_loops(frc.get_inner_loop_count());
4123
4124 // No infinite loops, no reason to bail out.
4125 return false;
4126 }
4127
4128 //-----------------------------too_many_traps----------------------------------
4129 // Report if there are too many traps at the current method and bci.
4130 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4131 bool Compile::too_many_traps(ciMethod* method,
4132 int bci,
4133 Deoptimization::DeoptReason reason) {
4134 if (method->has_trap_at(bci)) {
4135 return true;
4136 }
4137 ciMethodData* md = method->method_data();
4138 if (md->is_empty()) {
4139 // Assume the trap has not occurred, or that it occurred only
4140 // because of a transient condition during start-up in the interpreter.
4141 return false;
4142 }
4143 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4144 if (md->has_trap_at(bci, m, reason) != 0) {
4145 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4146 // Also, if there are multiple reasons, or if there is no per-BCI record,
4147 // assume the worst.
4148 if (log())
4149 log()->elem("observe trap='%s' count='%d'",
4150 Deoptimization::trap_reason_name(reason),
4151 md->trap_count(reason));
4152 return true;
4153 } else {
4154 // Ignore method/bci and see if there have been too many globally.
4155 return too_many_traps(reason, md);
4156 }
4157 }
4158
4159 // Less-accurate variant which does not require a method and bci.
4160 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4161 ciMethodData* logmd) {
4162 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4163 // Too many traps globally.
4164 // Note that we use cumulative trap_count, not just md->trap_count.
4165 if (log()) {
4166 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4167 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4168 Deoptimization::trap_reason_name(reason),
4169 mcount, trap_count(reason));
4170 }
4171 return true;
4172 } else {
4173 // The coast is clear.
4174 return false;
4175 }
4176 }
4177
4178 //--------------------------too_many_recompiles--------------------------------
4179 // Report if there are too many recompiles at the current method and bci.
4180 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4181 // Is not eager to return true, since this will cause the compiler to use
4182 // Action_none for a trap point, to avoid too many recompilations.
4183 bool Compile::too_many_recompiles(ciMethod* method,
4184 int bci,
4185 Deoptimization::DeoptReason reason) {
4186 ciMethodData* md = method->method_data();
4187 if (md->is_empty()) {
4188 // Assume the trap has not occurred, or that it occurred only
4189 // because of a transient condition during start-up in the interpreter.
4190 return false;
4191 }
4192 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4193 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4194 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4195 Deoptimization::DeoptReason per_bc_reason
4196 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4197 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4198 if ((per_bc_reason == Deoptimization::Reason_none
4199 || md->has_trap_at(bci, m, reason) != 0)
4200 // The trap frequency measure we care about is the recompile count:
4201 && md->trap_recompiled_at(bci, m)
4202 && md->overflow_recompile_count() >= bc_cutoff) {
4203 // Do not emit a trap here if it has already caused recompilations.
4204 // Also, if there are multiple reasons, or if there is no per-BCI record,
4205 // assume the worst.
4206 if (log())
4207 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4208 Deoptimization::trap_reason_name(reason),
4209 md->trap_count(reason),
4210 md->overflow_recompile_count());
4211 return true;
4212 } else if (trap_count(reason) != 0
4213 && decompile_count() >= m_cutoff) {
4214 // Too many recompiles globally, and we have seen this sort of trap.
4215 // Use cumulative decompile_count, not just md->decompile_count.
4216 if (log())
4217 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4218 Deoptimization::trap_reason_name(reason),
4219 md->trap_count(reason), trap_count(reason),
4220 md->decompile_count(), decompile_count());
4221 return true;
4222 } else {
4223 // The coast is clear.
4224 return false;
4225 }
4226 }
4227
4228 // Compute when not to trap. Used by matching trap based nodes and
4229 // NullCheck optimization.
4230 void Compile::set_allowed_deopt_reasons() {
4231 _allowed_reasons = 0;
4232 if (is_method_compilation()) {
4233 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4234 assert(rs < BitsPerInt, "recode bit map");
4235 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4236 _allowed_reasons |= nth_bit(rs);
4237 }
4238 }
4239 }
4240 }
4241
4242 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4243 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4244 }
4245
4246 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4247 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4248 }
4249
4250 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4251 if (holder->is_initialized() && !do_clinit_barriers()) {
4252 return false;
4253 }
4254 if (holder->is_being_initialized() || do_clinit_barriers()) {
4255 if (accessing_method->holder() == holder) {
4256 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4257 // <init>, or a static method. In all those cases, there was an initialization
4258 // barrier on the holder klass passed.
4259 if (accessing_method->is_static_initializer() ||
4260 accessing_method->is_object_initializer() ||
4261 accessing_method->is_static()) {
4262 return false;
4263 }
4264 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4265 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4266 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4267 // child class can become fully initialized while its parent class is still being initialized.
4268 if (accessing_method->is_static_initializer()) {
4269 return false;
4270 }
4271 }
4272 ciMethod* root = method(); // the root method of compilation
4273 if (root != accessing_method) {
4274 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4275 }
4276 }
4277 return true;
4278 }
4279
4280 #ifndef PRODUCT
4281 //------------------------------verify_bidirectional_edges---------------------
4282 // For each input edge to a node (ie - for each Use-Def edge), verify that
4283 // there is a corresponding Def-Use edge.
4284 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4285 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4286 uint stack_size = live_nodes() >> 4;
4287 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4288 nstack.push(_root);
4289
4290 while (nstack.size() > 0) {
4291 Node* n = nstack.pop();
4292 if (visited.member(n)) {
4293 continue;
4294 }
4295 visited.push(n);
4296
4297 // Walk over all input edges, checking for correspondence
4298 uint length = n->len();
4299 for (uint i = 0; i < length; i++) {
4300 Node* in = n->in(i);
4301 if (in != nullptr && !visited.member(in)) {
4302 nstack.push(in); // Put it on stack
4303 }
4304 if (in != nullptr && !in->is_top()) {
4305 // Count instances of `next`
4306 int cnt = 0;
4307 for (uint idx = 0; idx < in->_outcnt; idx++) {
4308 if (in->_out[idx] == n) {
4309 cnt++;
4310 }
4311 }
4312 assert(cnt > 0, "Failed to find Def-Use edge.");
4313 // Check for duplicate edges
4314 // walk the input array downcounting the input edges to n
4315 for (uint j = 0; j < length; j++) {
4316 if (n->in(j) == in) {
4317 cnt--;
4318 }
4319 }
4320 assert(cnt == 0, "Mismatched edge count.");
4321 } else if (in == nullptr) {
4322 assert(i == 0 || i >= n->req() ||
4323 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4324 (n->is_Unlock() && i == (n->req() - 1)) ||
4325 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4326 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4327 } else {
4328 assert(in->is_top(), "sanity");
4329 // Nothing to check.
4330 }
4331 }
4332 }
4333 }
4334
4335 //------------------------------verify_graph_edges---------------------------
4336 // Walk the Graph and verify that there is a one-to-one correspondence
4337 // between Use-Def edges and Def-Use edges in the graph.
4338 void Compile::verify_graph_edges(bool no_dead_code) {
4339 if (VerifyGraphEdges) {
4340 Unique_Node_List visited;
4341
4342 // Call graph walk to check edges
4343 verify_bidirectional_edges(visited);
4344 if (no_dead_code) {
4345 // Now make sure that no visited node is used by an unvisited node.
4346 bool dead_nodes = false;
4347 Unique_Node_List checked;
4348 while (visited.size() > 0) {
4349 Node* n = visited.pop();
4350 checked.push(n);
4351 for (uint i = 0; i < n->outcnt(); i++) {
4352 Node* use = n->raw_out(i);
4353 if (checked.member(use)) continue; // already checked
4354 if (visited.member(use)) continue; // already in the graph
4355 if (use->is_Con()) continue; // a dead ConNode is OK
4356 // At this point, we have found a dead node which is DU-reachable.
4357 if (!dead_nodes) {
4358 tty->print_cr("*** Dead nodes reachable via DU edges:");
4359 dead_nodes = true;
4360 }
4361 use->dump(2);
4362 tty->print_cr("---");
4363 checked.push(use); // No repeats; pretend it is now checked.
4364 }
4365 }
4366 assert(!dead_nodes, "using nodes must be reachable from root");
4367 }
4368 }
4369 }
4370 #endif
4371
4372 // The Compile object keeps track of failure reasons separately from the ciEnv.
4373 // This is required because there is not quite a 1-1 relation between the
4374 // ciEnv and its compilation task and the Compile object. Note that one
4375 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4376 // to backtrack and retry without subsuming loads. Other than this backtracking
4377 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4378 // by the logic in C2Compiler.
4379 void Compile::record_failure(const char* reason) {
4380 if (log() != nullptr) {
4381 log()->elem("failure reason='%s' phase='compile'", reason);
4382 }
4383 if (_failure_reason.get() == nullptr) {
4384 // Record the first failure reason.
4385 _failure_reason.set(reason);
4386 if (CaptureBailoutInformation) {
4387 _first_failure_details = new CompilationFailureInfo(reason);
4388 }
4389 }
4390
4391 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4392 C->print_method(PHASE_FAILURE, 1);
4393 }
4394 _root = nullptr; // flush the graph, too
4395 }
4396
4397 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4398 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4399 _compile(Compile::current()),
4400 _log(nullptr),
4401 _phase_name(name),
4402 _dolog(CITimeVerbose)
4403 {
4404 assert(_compile != nullptr, "sanity check");
4405 if (_dolog) {
4406 _log = _compile->log();
4407 }
4408 if (_log != nullptr) {
4409 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4410 _log->stamp();
4411 _log->end_head();
4412 }
4413 }
4414
4415 Compile::TracePhase::~TracePhase() {
4416 if (_compile->failing()) return;
4417 #ifdef ASSERT
4418 if (PrintIdealNodeCount) {
4419 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4420 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4421 }
4422
4423 if (VerifyIdealNodeCount) {
4424 _compile->print_missing_nodes();
4425 }
4426 #endif
4427
4428 if (_log != nullptr) {
4429 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4430 }
4431 }
4432
4433 //----------------------------static_subtype_check-----------------------------
4434 // Shortcut important common cases when superklass is exact:
4435 // (0) superklass is java.lang.Object (can occur in reflective code)
4436 // (1) subklass is already limited to a subtype of superklass => always ok
4437 // (2) subklass does not overlap with superklass => always fail
4438 // (3) superklass has NO subtypes and we can check with a simple compare.
4439 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4440 if (skip) {
4441 return SSC_full_test; // Let caller generate the general case.
4442 }
4443
4444 if (subk->is_java_subtype_of(superk)) {
4445 return SSC_always_true; // (0) and (1) this test cannot fail
4446 }
4447
4448 if (!subk->maybe_java_subtype_of(superk)) {
4449 return SSC_always_false; // (2) true path dead; no dynamic test needed
4450 }
4451
4452 const Type* superelem = superk;
4453 if (superk->isa_aryklassptr()) {
4454 int ignored;
4455 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4456 }
4457
4458 if (superelem->isa_instklassptr()) {
4459 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4460 if (!ik->has_subklass()) {
4461 if (!ik->is_final()) {
4462 // Add a dependency if there is a chance of a later subclass.
4463 dependencies()->assert_leaf_type(ik);
4464 }
4465 if (!superk->maybe_java_subtype_of(subk)) {
4466 return SSC_always_false;
4467 }
4468 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4469 }
4470 } else {
4471 // A primitive array type has no subtypes.
4472 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4473 }
4474
4475 return SSC_full_test;
4476 }
4477
4478 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4479 #ifdef _LP64
4480 // The scaled index operand to AddP must be a clean 64-bit value.
4481 // Java allows a 32-bit int to be incremented to a negative
4482 // value, which appears in a 64-bit register as a large
4483 // positive number. Using that large positive number as an
4484 // operand in pointer arithmetic has bad consequences.
4485 // On the other hand, 32-bit overflow is rare, and the possibility
4486 // can often be excluded, if we annotate the ConvI2L node with
4487 // a type assertion that its value is known to be a small positive
4488 // number. (The prior range check has ensured this.)
4489 // This assertion is used by ConvI2LNode::Ideal.
4490 int index_max = max_jint - 1; // array size is max_jint, index is one less
4491 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4492 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4493 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4494 #endif
4495 return idx;
4496 }
4497
4498 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4499 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4500 if (ctrl != nullptr) {
4501 // Express control dependency by a CastII node with a narrow type.
4502 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4503 // node from floating above the range check during loop optimizations. Otherwise, the
4504 // ConvI2L node may be eliminated independently of the range check, causing the data path
4505 // to become TOP while the control path is still there (although it's unreachable).
4506 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4507 value = phase->transform(value);
4508 }
4509 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4510 return phase->transform(new ConvI2LNode(value, ltype));
4511 }
4512
4513 // The message about the current inlining is accumulated in
4514 // _print_inlining_stream and transferred into the _print_inlining_list
4515 // once we know whether inlining succeeds or not. For regular
4516 // inlining, messages are appended to the buffer pointed by
4517 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4518 // a new buffer is added after _print_inlining_idx in the list. This
4519 // way we can update the inlining message for late inlining call site
4520 // when the inlining is attempted again.
4521 void Compile::print_inlining_init() {
4522 if (print_inlining() || print_intrinsics()) {
4523 // print_inlining_init is actually called several times.
4524 print_inlining_reset();
4525 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4526 }
4527 }
4528
4529 void Compile::print_inlining_reinit() {
4530 if (print_inlining() || print_intrinsics()) {
4531 print_inlining_reset();
4532 }
4533 }
4534
4535 void Compile::print_inlining_reset() {
4536 _print_inlining_stream->reset();
4537 }
4538
4539 void Compile::print_inlining_commit() {
4540 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4541 // Transfer the message from _print_inlining_stream to the current
4542 // _print_inlining_list buffer and clear _print_inlining_stream.
4543 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4544 print_inlining_reset();
4545 }
4546
4547 void Compile::print_inlining_push() {
4548 // Add new buffer to the _print_inlining_list at current position
4549 _print_inlining_idx++;
4550 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4551 }
4552
4553 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4554 return _print_inlining_list->at(_print_inlining_idx);
4555 }
4556
4557 void Compile::print_inlining_update(CallGenerator* cg) {
4558 if (print_inlining() || print_intrinsics()) {
4559 if (cg->is_late_inline()) {
4560 if (print_inlining_current()->cg() != cg &&
4561 (print_inlining_current()->cg() != nullptr ||
4562 print_inlining_current()->ss()->size() != 0)) {
4563 print_inlining_push();
4564 }
4565 print_inlining_commit();
4566 print_inlining_current()->set_cg(cg);
4567 } else {
4568 if (print_inlining_current()->cg() != nullptr) {
4569 print_inlining_push();
4570 }
4571 print_inlining_commit();
4572 }
4573 }
4574 }
4575
4576 void Compile::print_inlining_move_to(CallGenerator* cg) {
4577 // We resume inlining at a late inlining call site. Locate the
4578 // corresponding inlining buffer so that we can update it.
4579 if (print_inlining() || print_intrinsics()) {
4580 for (int i = 0; i < _print_inlining_list->length(); i++) {
4581 if (_print_inlining_list->at(i)->cg() == cg) {
4582 _print_inlining_idx = i;
4583 return;
4584 }
4585 }
4586 ShouldNotReachHere();
4587 }
4588 }
4589
4590 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4591 if (print_inlining() || print_intrinsics()) {
4592 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4593 assert(print_inlining_current()->cg() == cg, "wrong entry");
4594 // replace message with new message
4595 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4596 print_inlining_commit();
4597 print_inlining_current()->set_cg(cg);
4598 }
4599 }
4600
4601 void Compile::print_inlining_assert_ready() {
4602 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4603 }
4604
4605 void Compile::process_print_inlining() {
4606 assert(_late_inlines.length() == 0, "not drained yet");
4607 if (print_inlining() || print_intrinsics()) {
4608 ResourceMark rm;
4609 stringStream ss;
4610 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4611 for (int i = 0; i < _print_inlining_list->length(); i++) {
4612 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4613 ss.print("%s", pib->ss()->freeze());
4614 delete pib;
4615 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4616 }
4617 // Reset _print_inlining_list, it only contains destructed objects.
4618 // It is on the arena, so it will be freed when the arena is reset.
4619 _print_inlining_list = nullptr;
4620 // _print_inlining_stream won't be used anymore, either.
4621 print_inlining_reset();
4622 size_t end = ss.size();
4623 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4624 strncpy(_print_inlining_output, ss.freeze(), end+1);
4625 _print_inlining_output[end] = 0;
4626 }
4627 }
4628
4629 void Compile::dump_print_inlining() {
4630 if (_print_inlining_output != nullptr) {
4631 tty->print_raw(_print_inlining_output);
4632 }
4633 }
4634
4635 void Compile::log_late_inline(CallGenerator* cg) {
4636 if (log() != nullptr) {
4637 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4638 cg->unique_id());
4639 JVMState* p = cg->call_node()->jvms();
4640 while (p != nullptr) {
4641 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4642 p = p->caller();
4643 }
4644 log()->tail("late_inline");
4645 }
4646 }
4647
4648 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4649 log_late_inline(cg);
4650 if (log() != nullptr) {
4651 log()->inline_fail(msg);
4652 }
4653 }
4654
4655 void Compile::log_inline_id(CallGenerator* cg) {
4656 if (log() != nullptr) {
4657 // The LogCompilation tool needs a unique way to identify late
4658 // inline call sites. This id must be unique for this call site in
4659 // this compilation. Try to have it unique across compilations as
4660 // well because it can be convenient when grepping through the log
4661 // file.
4662 // Distinguish OSR compilations from others in case CICountOSR is
4663 // on.
4664 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4665 cg->set_unique_id(id);
4666 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4667 }
4668 }
4669
4670 void Compile::log_inline_failure(const char* msg) {
4671 if (C->log() != nullptr) {
4672 C->log()->inline_fail(msg);
4673 }
4674 }
4675
4676
4677 // Dump inlining replay data to the stream.
4678 // Don't change thread state and acquire any locks.
4679 void Compile::dump_inline_data(outputStream* out) {
4680 InlineTree* inl_tree = ilt();
4681 if (inl_tree != nullptr) {
4682 out->print(" inline %d", inl_tree->count());
4683 inl_tree->dump_replay_data(out);
4684 }
4685 }
4686
4687 void Compile::dump_inline_data_reduced(outputStream* out) {
4688 assert(ReplayReduce, "");
4689
4690 InlineTree* inl_tree = ilt();
4691 if (inl_tree == nullptr) {
4692 return;
4693 }
4694 // Enable iterative replay file reduction
4695 // Output "compile" lines for depth 1 subtrees,
4696 // simulating that those trees were compiled
4697 // instead of inlined.
4698 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4699 InlineTree* sub = inl_tree->subtrees().at(i);
4700 if (sub->inline_level() != 1) {
4701 continue;
4702 }
4703
4704 ciMethod* method = sub->method();
4705 int entry_bci = -1;
4706 int comp_level = env()->task()->comp_level();
4707 out->print("compile ");
4708 method->dump_name_as_ascii(out);
4709 out->print(" %d %d", entry_bci, comp_level);
4710 out->print(" inline %d", sub->count());
4711 sub->dump_replay_data(out, -1);
4712 out->cr();
4713 }
4714 }
4715
4716 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4717 if (n1->Opcode() < n2->Opcode()) return -1;
4718 else if (n1->Opcode() > n2->Opcode()) return 1;
4719
4720 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4721 for (uint i = 1; i < n1->req(); i++) {
4722 if (n1->in(i) < n2->in(i)) return -1;
4723 else if (n1->in(i) > n2->in(i)) return 1;
4724 }
4725
4726 return 0;
4727 }
4728
4729 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4730 Node* n1 = *n1p;
4731 Node* n2 = *n2p;
4732
4733 return cmp_expensive_nodes(n1, n2);
4734 }
4735
4736 void Compile::sort_expensive_nodes() {
4737 if (!expensive_nodes_sorted()) {
4738 _expensive_nodes.sort(cmp_expensive_nodes);
4739 }
4740 }
4741
4742 bool Compile::expensive_nodes_sorted() const {
4743 for (int i = 1; i < _expensive_nodes.length(); i++) {
4744 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4745 return false;
4746 }
4747 }
4748 return true;
4749 }
4750
4751 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4752 if (_expensive_nodes.length() == 0) {
4753 return false;
4754 }
4755
4756 assert(OptimizeExpensiveOps, "optimization off?");
4757
4758 // Take this opportunity to remove dead nodes from the list
4759 int j = 0;
4760 for (int i = 0; i < _expensive_nodes.length(); i++) {
4761 Node* n = _expensive_nodes.at(i);
4762 if (!n->is_unreachable(igvn)) {
4763 assert(n->is_expensive(), "should be expensive");
4764 _expensive_nodes.at_put(j, n);
4765 j++;
4766 }
4767 }
4768 _expensive_nodes.trunc_to(j);
4769
4770 // Then sort the list so that similar nodes are next to each other
4771 // and check for at least two nodes of identical kind with same data
4772 // inputs.
4773 sort_expensive_nodes();
4774
4775 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4776 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4777 return true;
4778 }
4779 }
4780
4781 return false;
4782 }
4783
4784 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4785 if (_expensive_nodes.length() == 0) {
4786 return;
4787 }
4788
4789 assert(OptimizeExpensiveOps, "optimization off?");
4790
4791 // Sort to bring similar nodes next to each other and clear the
4792 // control input of nodes for which there's only a single copy.
4793 sort_expensive_nodes();
4794
4795 int j = 0;
4796 int identical = 0;
4797 int i = 0;
4798 bool modified = false;
4799 for (; i < _expensive_nodes.length()-1; i++) {
4800 assert(j <= i, "can't write beyond current index");
4801 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4802 identical++;
4803 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4804 continue;
4805 }
4806 if (identical > 0) {
4807 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4808 identical = 0;
4809 } else {
4810 Node* n = _expensive_nodes.at(i);
4811 igvn.replace_input_of(n, 0, nullptr);
4812 igvn.hash_insert(n);
4813 modified = true;
4814 }
4815 }
4816 if (identical > 0) {
4817 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4818 } else if (_expensive_nodes.length() >= 1) {
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 _expensive_nodes.trunc_to(j);
4825 if (modified) {
4826 igvn.optimize();
4827 }
4828 }
4829
4830 void Compile::add_expensive_node(Node * n) {
4831 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4832 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4833 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4834 if (OptimizeExpensiveOps) {
4835 _expensive_nodes.append(n);
4836 } else {
4837 // Clear control input and let IGVN optimize expensive nodes if
4838 // OptimizeExpensiveOps is off.
4839 n->set_req(0, nullptr);
4840 }
4841 }
4842
4843 /**
4844 * Track coarsened Lock and Unlock nodes.
4845 */
4846
4847 class Lock_List : public Node_List {
4848 uint _origin_cnt;
4849 public:
4850 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4851 uint origin_cnt() const { return _origin_cnt; }
4852 };
4853
4854 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4855 int length = locks.length();
4856 if (length > 0) {
4857 // Have to keep this list until locks elimination during Macro nodes elimination.
4858 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4859 AbstractLockNode* alock = locks.at(0);
4860 BoxLockNode* box = alock->box_node()->as_BoxLock();
4861 for (int i = 0; i < length; i++) {
4862 AbstractLockNode* lock = locks.at(i);
4863 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4864 locks_list->push(lock);
4865 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4866 if (this_box != box) {
4867 // Locking regions (BoxLock) could be Unbalanced here:
4868 // - its coarsened locks were eliminated in earlier
4869 // macro nodes elimination followed by loop unroll
4870 // - it is OSR locking region (no Lock node)
4871 // Preserve Unbalanced status in such cases.
4872 if (!this_box->is_unbalanced()) {
4873 this_box->set_coarsened();
4874 }
4875 if (!box->is_unbalanced()) {
4876 box->set_coarsened();
4877 }
4878 }
4879 }
4880 _coarsened_locks.append(locks_list);
4881 }
4882 }
4883
4884 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4885 int count = coarsened_count();
4886 for (int i = 0; i < count; i++) {
4887 Node_List* locks_list = _coarsened_locks.at(i);
4888 for (uint j = 0; j < locks_list->size(); j++) {
4889 Node* lock = locks_list->at(j);
4890 assert(lock->is_AbstractLock(), "sanity");
4891 if (!useful.member(lock)) {
4892 locks_list->yank(lock);
4893 }
4894 }
4895 }
4896 }
4897
4898 void Compile::remove_coarsened_lock(Node* n) {
4899 if (n->is_AbstractLock()) {
4900 int count = coarsened_count();
4901 for (int i = 0; i < count; i++) {
4902 Node_List* locks_list = _coarsened_locks.at(i);
4903 locks_list->yank(n);
4904 }
4905 }
4906 }
4907
4908 bool Compile::coarsened_locks_consistent() {
4909 int count = coarsened_count();
4910 for (int i = 0; i < count; i++) {
4911 bool unbalanced = false;
4912 bool modified = false; // track locks kind modifications
4913 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4914 uint size = locks_list->size();
4915 if (size == 0) {
4916 unbalanced = false; // All locks were eliminated - good
4917 } else if (size != locks_list->origin_cnt()) {
4918 unbalanced = true; // Some locks were removed from list
4919 } else {
4920 for (uint j = 0; j < size; j++) {
4921 Node* lock = locks_list->at(j);
4922 // All nodes in group should have the same state (modified or not)
4923 if (!lock->as_AbstractLock()->is_coarsened()) {
4924 if (j == 0) {
4925 // first on list was modified, the rest should be too for consistency
4926 modified = true;
4927 } else if (!modified) {
4928 // this lock was modified but previous locks on the list were not
4929 unbalanced = true;
4930 break;
4931 }
4932 } else if (modified) {
4933 // previous locks on list were modified but not this lock
4934 unbalanced = true;
4935 break;
4936 }
4937 }
4938 }
4939 if (unbalanced) {
4940 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4941 #ifdef ASSERT
4942 if (PrintEliminateLocks) {
4943 tty->print_cr("=== unbalanced coarsened locks ===");
4944 for (uint l = 0; l < size; l++) {
4945 locks_list->at(l)->dump();
4946 }
4947 }
4948 #endif
4949 record_failure(C2Compiler::retry_no_locks_coarsening());
4950 return false;
4951 }
4952 }
4953 return true;
4954 }
4955
4956 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4957 // locks coarsening optimization removed Lock/Unlock nodes from them.
4958 // Such regions become unbalanced because coarsening only removes part
4959 // of Lock/Unlock nodes in region. As result we can't execute other
4960 // locks elimination optimizations which assume all code paths have
4961 // corresponding pair of Lock/Unlock nodes - they are balanced.
4962 void Compile::mark_unbalanced_boxes() const {
4963 int count = coarsened_count();
4964 for (int i = 0; i < count; i++) {
4965 Node_List* locks_list = _coarsened_locks.at(i);
4966 uint size = locks_list->size();
4967 if (size > 0) {
4968 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4969 BoxLockNode* box = alock->box_node()->as_BoxLock();
4970 if (alock->is_coarsened()) {
4971 // coarsened_locks_consistent(), which is called before this method, verifies
4972 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4973 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4974 for (uint j = 1; j < size; j++) {
4975 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4976 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4977 if (box != this_box) {
4978 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4979 box->set_unbalanced();
4980 this_box->set_unbalanced();
4981 }
4982 }
4983 }
4984 }
4985 }
4986 }
4987
4988 /**
4989 * Remove the speculative part of types and clean up the graph
4990 */
4991 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4992 if (UseTypeSpeculation) {
4993 Unique_Node_List worklist;
4994 worklist.push(root());
4995 int modified = 0;
4996 // Go over all type nodes that carry a speculative type, drop the
4997 // speculative part of the type and enqueue the node for an igvn
4998 // which may optimize it out.
4999 for (uint next = 0; next < worklist.size(); ++next) {
5000 Node *n = worklist.at(next);
5001 if (n->is_Type()) {
5002 TypeNode* tn = n->as_Type();
5003 const Type* t = tn->type();
5004 const Type* t_no_spec = t->remove_speculative();
5005 if (t_no_spec != t) {
5006 bool in_hash = igvn.hash_delete(n);
5007 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5008 tn->set_type(t_no_spec);
5009 igvn.hash_insert(n);
5010 igvn._worklist.push(n); // give it a chance to go away
5011 modified++;
5012 }
5013 }
5014 // Iterate over outs - endless loops is unreachable from below
5015 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5016 Node *m = n->fast_out(i);
5017 if (not_a_node(m)) {
5018 continue;
5019 }
5020 worklist.push(m);
5021 }
5022 }
5023 // Drop the speculative part of all types in the igvn's type table
5024 igvn.remove_speculative_types();
5025 if (modified > 0) {
5026 igvn.optimize();
5027 if (failing()) return;
5028 }
5029 #ifdef ASSERT
5030 // Verify that after the IGVN is over no speculative type has resurfaced
5031 worklist.clear();
5032 worklist.push(root());
5033 for (uint next = 0; next < worklist.size(); ++next) {
5034 Node *n = worklist.at(next);
5035 const Type* t = igvn.type_or_null(n);
5036 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5037 if (n->is_Type()) {
5038 t = n->as_Type()->type();
5039 assert(t == t->remove_speculative(), "no more speculative types");
5040 }
5041 // Iterate over outs - endless loops is unreachable from below
5042 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5043 Node *m = n->fast_out(i);
5044 if (not_a_node(m)) {
5045 continue;
5046 }
5047 worklist.push(m);
5048 }
5049 }
5050 igvn.check_no_speculative_types();
5051 #endif
5052 }
5053 }
5054
5055 // Auxiliary methods to support randomized stressing/fuzzing.
5056
5057 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5058 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5059 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5060 FLAG_SET_ERGO(StressSeed, _stress_seed);
5061 } else {
5062 _stress_seed = StressSeed;
5063 }
5064 if (_log != nullptr) {
5065 _log->elem("stress_test seed='%u'", _stress_seed);
5066 }
5067 }
5068
5069 int Compile::random() {
5070 _stress_seed = os::next_random(_stress_seed);
5071 return static_cast<int>(_stress_seed);
5072 }
5073
5074 // This method can be called the arbitrary number of times, with current count
5075 // as the argument. The logic allows selecting a single candidate from the
5076 // running list of candidates as follows:
5077 // int count = 0;
5078 // Cand* selected = null;
5079 // while(cand = cand->next()) {
5080 // if (randomized_select(++count)) {
5081 // selected = cand;
5082 // }
5083 // }
5084 //
5085 // Including count equalizes the chances any candidate is "selected".
5086 // This is useful when we don't have the complete list of candidates to choose
5087 // from uniformly. In this case, we need to adjust the randomicity of the
5088 // selection, or else we will end up biasing the selection towards the latter
5089 // candidates.
5090 //
5091 // Quick back-envelope calculation shows that for the list of n candidates
5092 // the equal probability for the candidate to persist as "best" can be
5093 // achieved by replacing it with "next" k-th candidate with the probability
5094 // of 1/k. It can be easily shown that by the end of the run, the
5095 // probability for any candidate is converged to 1/n, thus giving the
5096 // uniform distribution among all the candidates.
5097 //
5098 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5099 #define RANDOMIZED_DOMAIN_POW 29
5100 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5101 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5102 bool Compile::randomized_select(int count) {
5103 assert(count > 0, "only positive");
5104 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5105 }
5106
5107 CloneMap& Compile::clone_map() { return _clone_map; }
5108 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5109
5110 void NodeCloneInfo::dump_on(outputStream* st) const {
5111 st->print(" {%d:%d} ", idx(), gen());
5112 }
5113
5114 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5115 uint64_t val = value(old->_idx);
5116 NodeCloneInfo cio(val);
5117 assert(val != 0, "old node should be in the map");
5118 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5119 insert(nnn->_idx, cin.get());
5120 #ifndef PRODUCT
5121 if (is_debug()) {
5122 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5123 }
5124 #endif
5125 }
5126
5127 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5128 NodeCloneInfo cio(value(old->_idx));
5129 if (cio.get() == 0) {
5130 cio.set(old->_idx, 0);
5131 insert(old->_idx, cio.get());
5132 #ifndef PRODUCT
5133 if (is_debug()) {
5134 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5135 }
5136 #endif
5137 }
5138 clone(old, nnn, gen);
5139 }
5140
5141 int CloneMap::max_gen() const {
5142 int g = 0;
5143 DictI di(_dict);
5144 for(; di.test(); ++di) {
5145 int t = gen(di._key);
5146 if (g < t) {
5147 g = t;
5148 #ifndef PRODUCT
5149 if (is_debug()) {
5150 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5151 }
5152 #endif
5153 }
5154 }
5155 return g;
5156 }
5157
5158 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5159 uint64_t val = value(key);
5160 if (val != 0) {
5161 NodeCloneInfo ni(val);
5162 ni.dump_on(st);
5163 }
5164 }
5165
5166 void Compile::shuffle_macro_nodes() {
5167 if (_macro_nodes.length() < 2) {
5168 return;
5169 }
5170 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5171 uint j = C->random() % (i + 1);
5172 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5173 }
5174 }
5175
5176 // Move Allocate nodes to the start of the list
5177 void Compile::sort_macro_nodes() {
5178 int count = macro_count();
5179 int allocates = 0;
5180 for (int i = 0; i < count; i++) {
5181 Node* n = macro_node(i);
5182 if (n->is_Allocate()) {
5183 if (i != allocates) {
5184 Node* tmp = macro_node(allocates);
5185 _macro_nodes.at_put(allocates, n);
5186 _macro_nodes.at_put(i, tmp);
5187 }
5188 allocates++;
5189 }
5190 }
5191 }
5192
5193 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5194 if (failing()) { return; }
5195 EventCompilerPhase event(UNTIMED);
5196 if (event.should_commit()) {
5197 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5198 }
5199 #ifndef PRODUCT
5200 ResourceMark rm;
5201 stringStream ss;
5202 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5203 int iter = ++_igv_phase_iter[cpt];
5204 if (iter > 1) {
5205 ss.print(" %d", iter);
5206 }
5207 if (n != nullptr) {
5208 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5209 }
5210
5211 const char* name = ss.as_string();
5212 if (should_print_igv(level)) {
5213 _igv_printer->print_method(name, level);
5214 }
5215 if (should_print_phase(cpt)) {
5216 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5217 }
5218 #endif
5219 C->_latest_stage_start_counter.stamp();
5220 }
5221
5222 // Only used from CompileWrapper
5223 void Compile::begin_method() {
5224 #ifndef PRODUCT
5225 if (_method != nullptr && should_print_igv(1)) {
5226 _igv_printer->begin_method();
5227 }
5228 #endif
5229 C->_latest_stage_start_counter.stamp();
5230 }
5231
5232 // Only used from CompileWrapper
5233 void Compile::end_method() {
5234 EventCompilerPhase event(UNTIMED);
5235 if (event.should_commit()) {
5236 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5237 }
5238
5239 #ifndef PRODUCT
5240 if (_method != nullptr && should_print_igv(1)) {
5241 _igv_printer->end_method();
5242 }
5243 #endif
5244 }
5245
5246 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5247 #ifndef PRODUCT
5248 if (_directive->should_print_phase(cpt)) {
5249 return true;
5250 }
5251 #endif
5252 return false;
5253 }
5254
5255 bool Compile::should_print_igv(const int level) {
5256 #ifndef PRODUCT
5257 if (PrintIdealGraphLevel < 0) { // disabled by the user
5258 return false;
5259 }
5260
5261 bool need = directive()->IGVPrintLevelOption >= level;
5262 if (need && !_igv_printer) {
5263 _igv_printer = IdealGraphPrinter::printer();
5264 _igv_printer->set_compile(this);
5265 }
5266 return need;
5267 #else
5268 return false;
5269 #endif
5270 }
5271
5272 #ifndef PRODUCT
5273 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5274 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5275
5276 // Called from debugger. Prints method to the default file with the default phase name.
5277 // This works regardless of any Ideal Graph Visualizer flags set or not.
5278 void igv_print() {
5279 Compile::current()->igv_print_method_to_file();
5280 }
5281
5282 // Same as igv_print() above but with a specified phase name.
5283 void igv_print(const char* phase_name) {
5284 Compile::current()->igv_print_method_to_file(phase_name);
5285 }
5286
5287 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5288 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5289 // This works regardless of any Ideal Graph Visualizer flags set or not.
5290 void igv_print(bool network) {
5291 if (network) {
5292 Compile::current()->igv_print_method_to_network();
5293 } else {
5294 Compile::current()->igv_print_method_to_file();
5295 }
5296 }
5297
5298 // Same as igv_print(bool network) above but with a specified phase name.
5299 void igv_print(bool network, const char* phase_name) {
5300 if (network) {
5301 Compile::current()->igv_print_method_to_network(phase_name);
5302 } else {
5303 Compile::current()->igv_print_method_to_file(phase_name);
5304 }
5305 }
5306
5307 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5308 void igv_print_default() {
5309 Compile::current()->print_method(PHASE_DEBUG, 0);
5310 }
5311
5312 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5313 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5314 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5315 void igv_append() {
5316 Compile::current()->igv_print_method_to_file("Debug", true);
5317 }
5318
5319 // Same as igv_append() above but with a specified phase name.
5320 void igv_append(const char* phase_name) {
5321 Compile::current()->igv_print_method_to_file(phase_name, true);
5322 }
5323
5324 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5325 const char* file_name = "custom_debug.xml";
5326 if (_debug_file_printer == nullptr) {
5327 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5328 } else {
5329 _debug_file_printer->update_compiled_method(C->method());
5330 }
5331 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5332 _debug_file_printer->print(phase_name, (Node*)C->root());
5333 }
5334
5335 void Compile::igv_print_method_to_network(const char* phase_name) {
5336 if (_debug_network_printer == nullptr) {
5337 _debug_network_printer = new IdealGraphPrinter(C);
5338 } else {
5339 _debug_network_printer->update_compiled_method(C->method());
5340 }
5341 tty->print_cr("Method printed over network stream to IGV");
5342 _debug_network_printer->print(phase_name, (Node*)C->root());
5343 }
5344 #endif
5345
5346 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5347 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5348 return value;
5349 }
5350 Node* result = nullptr;
5351 if (bt == T_BYTE) {
5352 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5353 result = new RShiftINode(result, phase->intcon(24));
5354 } else if (bt == T_BOOLEAN) {
5355 result = new AndINode(value, phase->intcon(0xFF));
5356 } else if (bt == T_CHAR) {
5357 result = new AndINode(value,phase->intcon(0xFFFF));
5358 } else {
5359 assert(bt == T_SHORT, "unexpected narrow type");
5360 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5361 result = new RShiftINode(result, phase->intcon(16));
5362 }
5363 if (transform_res) {
5364 result = phase->transform(result);
5365 }
5366 return result;
5367 }
5368
5369 void Compile::record_method_not_compilable_oom() {
5370 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5371 }