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/inlinetypenode.hpp"
61 #include "opto/locknode.hpp"
62 #include "opto/loopnode.hpp"
63 #include "opto/machnode.hpp"
64 #include "opto/macro.hpp"
65 #include "opto/matcher.hpp"
66 #include "opto/mathexactnode.hpp"
67 #include "opto/memnode.hpp"
68 #include "opto/mulnode.hpp"
69 #include "opto/narrowptrnode.hpp"
70 #include "opto/node.hpp"
71 #include "opto/opcodes.hpp"
72 #include "opto/output.hpp"
73 #include "opto/parse.hpp"
74 #include "opto/phaseX.hpp"
75 #include "opto/rootnode.hpp"
76 #include "opto/runtime.hpp"
77 #include "opto/stringopts.hpp"
78 #include "opto/type.hpp"
79 #include "opto/vector.hpp"
80 #include "opto/vectornode.hpp"
81 #include "runtime/globals_extension.hpp"
82 #include "runtime/sharedRuntime.hpp"
83 #include "runtime/signature.hpp"
84 #include "runtime/stubRoutines.hpp"
85 #include "runtime/timer.hpp"
86 #include "utilities/align.hpp"
87 #include "utilities/copy.hpp"
88 #include "utilities/macros.hpp"
89 #include "utilities/resourceHash.hpp"
90
91 // -------------------- Compile::mach_constant_base_node -----------------------
92 // Constant table base node singleton.
93 MachConstantBaseNode* Compile::mach_constant_base_node() {
94 if (_mach_constant_base_node == nullptr) {
95 _mach_constant_base_node = new MachConstantBaseNode();
96 _mach_constant_base_node->add_req(C->root());
97 }
98 return _mach_constant_base_node;
99 }
100
101
102 /// Support for intrinsics.
103
104 // Return the index at which m must be inserted (or already exists).
105 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
106 class IntrinsicDescPair {
107 private:
108 ciMethod* _m;
109 bool _is_virtual;
110 public:
111 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
112 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
113 ciMethod* m= elt->method();
114 ciMethod* key_m = key->_m;
115 if (key_m < m) return -1;
116 else if (key_m > m) return 1;
117 else {
118 bool is_virtual = elt->is_virtual();
119 bool key_virtual = key->_is_virtual;
120 if (key_virtual < is_virtual) return -1;
121 else if (key_virtual > is_virtual) return 1;
122 else return 0;
123 }
124 }
125 };
126 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
127 #ifdef ASSERT
128 for (int i = 1; i < _intrinsics.length(); i++) {
129 CallGenerator* cg1 = _intrinsics.at(i-1);
130 CallGenerator* cg2 = _intrinsics.at(i);
131 assert(cg1->method() != cg2->method()
132 ? cg1->method() < cg2->method()
133 : cg1->is_virtual() < cg2->is_virtual(),
134 "compiler intrinsics list must stay sorted");
135 }
136 #endif
137 IntrinsicDescPair pair(m, is_virtual);
138 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
139 }
140
141 void Compile::register_intrinsic(CallGenerator* cg) {
142 bool found = false;
143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
144 assert(!found, "registering twice");
145 _intrinsics.insert_before(index, cg);
146 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
147 }
148
149 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
150 assert(m->is_loaded(), "don't try this on unloaded methods");
151 if (_intrinsics.length() > 0) {
152 bool found = false;
153 int index = intrinsic_insertion_index(m, is_virtual, found);
154 if (found) {
155 return _intrinsics.at(index);
156 }
157 }
158 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
159 if (m->intrinsic_id() != vmIntrinsics::_none &&
160 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
161 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
162 if (cg != nullptr) {
163 // Save it for next time:
164 register_intrinsic(cg);
165 return cg;
166 } else {
167 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
168 }
169 }
170 return nullptr;
171 }
172
173 // Compile::make_vm_intrinsic is defined in library_call.cpp.
174
175 #ifndef PRODUCT
176 // statistics gathering...
177
178 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
179 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
180
181 inline int as_int(vmIntrinsics::ID id) {
182 return vmIntrinsics::as_int(id);
183 }
184
185 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
186 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
187 int oflags = _intrinsic_hist_flags[as_int(id)];
188 assert(flags != 0, "what happened?");
189 if (is_virtual) {
190 flags |= _intrinsic_virtual;
191 }
192 bool changed = (flags != oflags);
193 if ((flags & _intrinsic_worked) != 0) {
194 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
195 if (count == 1) {
196 changed = true; // first time
197 }
198 // increment the overall count also:
199 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
200 }
201 if (changed) {
202 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
203 // Something changed about the intrinsic's virtuality.
204 if ((flags & _intrinsic_virtual) != 0) {
205 // This is the first use of this intrinsic as a virtual call.
206 if (oflags != 0) {
207 // We already saw it as a non-virtual, so note both cases.
208 flags |= _intrinsic_both;
209 }
210 } else if ((oflags & _intrinsic_both) == 0) {
211 // This is the first use of this intrinsic as a non-virtual
212 flags |= _intrinsic_both;
213 }
214 }
215 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
216 }
217 // update the overall flags also:
218 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
219 return changed;
220 }
221
222 static char* format_flags(int flags, char* buf) {
223 buf[0] = 0;
224 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
225 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
226 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
227 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
228 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
229 if (buf[0] == 0) strcat(buf, ",");
230 assert(buf[0] == ',', "must be");
231 return &buf[1];
232 }
233
234 void Compile::print_intrinsic_statistics() {
235 char flagsbuf[100];
236 ttyLocker ttyl;
237 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
238 tty->print_cr("Compiler intrinsic usage:");
239 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
240 if (total == 0) total = 1; // avoid div0 in case of no successes
241 #define PRINT_STAT_LINE(name, c, f) \
242 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
243 for (auto id : EnumRange<vmIntrinsicID>{}) {
244 int flags = _intrinsic_hist_flags[as_int(id)];
245 juint count = _intrinsic_hist_count[as_int(id)];
246 if ((flags | count) != 0) {
247 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
248 }
249 }
250 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
251 if (xtty != nullptr) xtty->tail("statistics");
252 }
253
254 void Compile::print_statistics() {
255 { ttyLocker ttyl;
256 if (xtty != nullptr) xtty->head("statistics type='opto'");
257 Parse::print_statistics();
258 PhaseStringOpts::print_statistics();
259 PhaseCCP::print_statistics();
260 PhaseRegAlloc::print_statistics();
261 PhaseOutput::print_statistics();
262 PhasePeephole::print_statistics();
263 PhaseIdealLoop::print_statistics();
264 ConnectionGraph::print_statistics();
265 PhaseMacroExpand::print_statistics();
266 if (xtty != nullptr) xtty->tail("statistics");
267 }
268 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
269 // put this under its own <statistics> element.
270 print_intrinsic_statistics();
271 }
272 }
273 #endif //PRODUCT
274
275 void Compile::gvn_replace_by(Node* n, Node* nn) {
276 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
277 Node* use = n->last_out(i);
278 bool is_in_table = initial_gvn()->hash_delete(use);
279 uint uses_found = 0;
280 for (uint j = 0; j < use->len(); j++) {
281 if (use->in(j) == n) {
282 if (j < use->req())
283 use->set_req(j, nn);
284 else
285 use->set_prec(j, nn);
286 uses_found++;
287 }
288 }
289 if (is_in_table) {
290 // reinsert into table
291 initial_gvn()->hash_find_insert(use);
292 }
293 record_for_igvn(use);
294 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
295 i -= uses_found; // we deleted 1 or more copies of this edge
296 }
297 }
298
299
300 // Identify all nodes that are reachable from below, useful.
301 // Use breadth-first pass that records state in a Unique_Node_List,
302 // recursive traversal is slower.
303 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
304 int estimated_worklist_size = live_nodes();
305 useful.map( estimated_worklist_size, nullptr ); // preallocate space
306
307 // Initialize worklist
308 if (root() != nullptr) { useful.push(root()); }
309 // If 'top' is cached, declare it useful to preserve cached node
310 if (cached_top_node()) { useful.push(cached_top_node()); }
311
312 // Push all useful nodes onto the list, breadthfirst
313 for( uint next = 0; next < useful.size(); ++next ) {
314 assert( next < unique(), "Unique useful nodes < total nodes");
315 Node *n = useful.at(next);
316 uint max = n->len();
317 for( uint i = 0; i < max; ++i ) {
318 Node *m = n->in(i);
319 if (not_a_node(m)) continue;
320 useful.push(m);
321 }
322 }
323 }
324
325 // Update dead_node_list with any missing dead nodes using useful
326 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
327 void Compile::update_dead_node_list(Unique_Node_List &useful) {
328 uint max_idx = unique();
329 VectorSet& useful_node_set = useful.member_set();
330
331 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
332 // If node with index node_idx is not in useful set,
333 // mark it as dead in dead node list.
334 if (!useful_node_set.test(node_idx)) {
335 record_dead_node(node_idx);
336 }
337 }
338 }
339
340 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
341 int shift = 0;
342 for (int i = 0; i < inlines->length(); i++) {
343 CallGenerator* cg = inlines->at(i);
344 if (useful.member(cg->call_node())) {
345 if (shift > 0) {
346 inlines->at_put(i - shift, cg);
347 }
348 } else {
349 shift++; // skip over the dead element
350 }
351 }
352 if (shift > 0) {
353 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
354 }
355 }
356
357 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
358 assert(dead != nullptr && dead->is_Call(), "sanity");
359 int found = 0;
360 for (int i = 0; i < inlines->length(); i++) {
361 if (inlines->at(i)->call_node() == dead) {
362 inlines->remove_at(i);
363 found++;
364 NOT_DEBUG( break; ) // elements are unique, so exit early
365 }
366 }
367 assert(found <= 1, "not unique");
368 }
369
370 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
371 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
372 for (int i = node_list.length() - 1; i >= 0; i--) {
373 N* node = node_list.at(i);
374 if (!useful.member(node)) {
375 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
376 }
377 }
378 }
379
380 void Compile::remove_useless_node(Node* dead) {
381 remove_modified_node(dead);
382
383 // Constant node that has no out-edges and has only one in-edge from
384 // root is usually dead. However, sometimes reshaping walk makes
385 // it reachable by adding use edges. So, we will NOT count Con nodes
386 // as dead to be conservative about the dead node count at any
387 // given time.
388 if (!dead->is_Con()) {
389 record_dead_node(dead->_idx);
390 }
391 if (dead->is_macro()) {
392 remove_macro_node(dead);
393 }
394 if (dead->is_expensive()) {
395 remove_expensive_node(dead);
396 }
397 if (dead->Opcode() == Op_Opaque4) {
398 remove_template_assertion_predicate_opaq(dead);
399 }
400 if (dead->is_ParsePredicate()) {
401 remove_parse_predicate(dead->as_ParsePredicate());
402 }
403 if (dead->for_post_loop_opts_igvn()) {
404 remove_from_post_loop_opts_igvn(dead);
405 }
406 if (dead->is_InlineType()) {
407 remove_inline_type(dead);
408 }
409 if (dead->is_Call()) {
410 remove_useless_late_inlines( &_late_inlines, dead);
411 remove_useless_late_inlines( &_string_late_inlines, dead);
412 remove_useless_late_inlines( &_boxing_late_inlines, dead);
413 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
414
415 if (dead->is_CallStaticJava()) {
416 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
417 }
418 }
419 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
420 bs->unregister_potential_barrier_node(dead);
421 }
422
423 // Disconnect all useless nodes by disconnecting those at the boundary.
424 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
425 uint next = 0;
426 while (next < useful.size()) {
427 Node *n = useful.at(next++);
428 if (n->is_SafePoint()) {
429 // We're done with a parsing phase. Replaced nodes are not valid
430 // beyond that point.
431 n->as_SafePoint()->delete_replaced_nodes();
432 }
433 // Use raw traversal of out edges since this code removes out edges
434 int max = n->outcnt();
435 for (int j = 0; j < max; ++j) {
436 Node* child = n->raw_out(j);
437 if (!useful.member(child)) {
438 assert(!child->is_top() || child != top(),
439 "If top is cached in Compile object it is in useful list");
440 // Only need to remove this out-edge to the useless node
441 n->raw_del_out(j);
442 --j;
443 --max;
444 }
445 }
446 if (n->outcnt() == 1 && n->has_special_unique_user()) {
447 assert(useful.member(n->unique_out()), "do not push a useless node");
448 worklist.push(n->unique_out());
449 }
450 if (n->outcnt() == 0) {
451 worklist.push(n);
452 }
453 }
454
455 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
456 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
457 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
458 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
459 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
460 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
461 #ifdef ASSERT
462 if (_modified_nodes != nullptr) {
463 _modified_nodes->remove_useless_nodes(useful.member_set());
464 }
465 #endif
466 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
467 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
468 #ifdef ASSERT
469 if (_modified_nodes != nullptr) {
470 _modified_nodes->remove_useless_nodes(useful.member_set());
471 }
472 #endif
473
474 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
475 bs->eliminate_useless_gc_barriers(useful, this);
476 // clean up the late inline lists
477 remove_useless_late_inlines( &_late_inlines, useful);
478 remove_useless_late_inlines( &_string_late_inlines, useful);
479 remove_useless_late_inlines( &_boxing_late_inlines, useful);
480 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
481 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
482 }
483
484 // ============================================================================
485 //------------------------------CompileWrapper---------------------------------
486 class CompileWrapper : public StackObj {
487 Compile *const _compile;
488 public:
489 CompileWrapper(Compile* compile);
490
491 ~CompileWrapper();
492 };
493
494 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
495 // the Compile* pointer is stored in the current ciEnv:
496 ciEnv* env = compile->env();
497 assert(env == ciEnv::current(), "must already be a ciEnv active");
498 assert(env->compiler_data() == nullptr, "compile already active?");
499 env->set_compiler_data(compile);
500 assert(compile == Compile::current(), "sanity");
501
502 compile->set_type_dict(nullptr);
503 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
504 compile->clone_map().set_clone_idx(0);
505 compile->set_type_last_size(0);
506 compile->set_last_tf(nullptr, nullptr);
507 compile->set_indexSet_arena(nullptr);
508 compile->set_indexSet_free_block_list(nullptr);
509 compile->init_type_arena();
510 Type::Initialize(compile);
511 _compile->begin_method();
512 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
513 }
514 CompileWrapper::~CompileWrapper() {
515 // simulate crash during compilation
516 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
517
518 _compile->end_method();
519 _compile->env()->set_compiler_data(nullptr);
520 }
521
522
523 //----------------------------print_compile_messages---------------------------
524 void Compile::print_compile_messages() {
525 #ifndef PRODUCT
526 // Check if recompiling
527 if (!subsume_loads() && PrintOpto) {
528 // Recompiling without allowing machine instructions to subsume loads
529 tty->print_cr("*********************************************************");
530 tty->print_cr("** Bailout: Recompile without subsuming loads **");
531 tty->print_cr("*********************************************************");
532 }
533 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
534 // Recompiling without escape analysis
535 tty->print_cr("*********************************************************");
536 tty->print_cr("** Bailout: Recompile without escape analysis **");
537 tty->print_cr("*********************************************************");
538 }
539 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
540 // Recompiling without iterative escape analysis
541 tty->print_cr("*********************************************************");
542 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
543 tty->print_cr("*********************************************************");
544 }
545 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
546 // Recompiling without reducing allocation merges
547 tty->print_cr("*********************************************************");
548 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
549 tty->print_cr("*********************************************************");
550 }
551 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
552 // Recompiling without boxing elimination
553 tty->print_cr("*********************************************************");
554 tty->print_cr("** Bailout: Recompile without boxing elimination **");
555 tty->print_cr("*********************************************************");
556 }
557 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
558 // Recompiling without locks coarsening
559 tty->print_cr("*********************************************************");
560 tty->print_cr("** Bailout: Recompile without locks coarsening **");
561 tty->print_cr("*********************************************************");
562 }
563 if (env()->break_at_compile()) {
564 // Open the debugger when compiling this method.
565 tty->print("### Breaking when compiling: ");
566 method()->print_short_name();
567 tty->cr();
568 BREAKPOINT;
569 }
570
571 if( PrintOpto ) {
572 if (is_osr_compilation()) {
573 tty->print("[OSR]%3d", _compile_id);
574 } else {
575 tty->print("%3d", _compile_id);
576 }
577 }
578 #endif
579 }
580
581 #ifndef PRODUCT
582 void Compile::print_ideal_ir(const char* phase_name) {
583 // keep the following output all in one block
584 // This output goes directly to the tty, not the compiler log.
585 // To enable tools to match it up with the compilation activity,
586 // be sure to tag this tty output with the compile ID.
587
588 // Node dumping can cause a safepoint, which can break the tty lock.
589 // Buffer all node dumps, so that all safepoints happen before we lock.
590 ResourceMark rm;
591 stringStream ss;
592
593 if (_output == nullptr) {
594 ss.print_cr("AFTER: %s", phase_name);
595 // Print out all nodes in ascending order of index.
596 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
597 } else {
598 // Dump the node blockwise if we have a scheduling
599 _output->print_scheduling(&ss);
600 }
601
602 // Check that the lock is not broken by a safepoint.
603 NoSafepointVerifier nsv;
604 ttyLocker ttyl;
605 if (xtty != nullptr) {
606 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
607 compile_id(),
608 is_osr_compilation() ? " compile_kind='osr'" : "",
609 phase_name);
610 }
611
612 tty->print("%s", ss.as_string());
613
614 if (xtty != nullptr) {
615 xtty->tail("ideal");
616 }
617 }
618 #endif
619
620 // ============================================================================
621 //------------------------------Compile standard-------------------------------
622
623 // Compile a method. entry_bci is -1 for normal compilations and indicates
624 // the continuation bci for on stack replacement.
625
626
627 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
628 Options options, DirectiveSet* directive)
629 : Phase(Compiler),
630 _compile_id(ci_env->compile_id()),
631 _options(options),
632 _method(target),
633 _entry_bci(osr_bci),
634 _ilt(nullptr),
635 _stub_function(nullptr),
636 _stub_name(nullptr),
637 _stub_entry_point(nullptr),
638 _max_node_limit(MaxNodeLimit),
639 _post_loop_opts_phase(false),
640 _allow_macro_nodes(true),
641 _inlining_progress(false),
642 _inlining_incrementally(false),
643 _do_cleanup(false),
644 _has_reserved_stack_access(target->has_reserved_stack_access()),
645 _has_circular_inline_type(false),
646 #ifndef PRODUCT
647 _igv_idx(0),
648 _trace_opto_output(directive->TraceOptoOutputOption),
649 #endif
650 _has_method_handle_invokes(false),
651 _clinit_barrier_on_entry(false),
652 _stress_seed(0),
653 _comp_arena(mtCompiler),
654 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
655 _env(ci_env),
656 _directive(directive),
657 _log(ci_env->log()),
658 _first_failure_details(nullptr),
659 _intrinsics (comp_arena(), 0, 0, nullptr),
660 _macro_nodes (comp_arena(), 8, 0, nullptr),
661 _parse_predicates (comp_arena(), 8, 0, nullptr),
662 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
663 _expensive_nodes (comp_arena(), 8, 0, nullptr),
664 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
665 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
666 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
667 _coarsened_locks (comp_arena(), 8, 0, nullptr),
668 _congraph(nullptr),
669 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
670 _unique(0),
671 _dead_node_count(0),
672 _dead_node_list(comp_arena()),
673 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
674 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
675 _node_arena(&_node_arena_one),
676 _mach_constant_base_node(nullptr),
677 _Compile_types(mtCompiler),
678 _initial_gvn(nullptr),
679 _igvn_worklist(nullptr),
680 _types(nullptr),
681 _node_hash(nullptr),
682 _late_inlines(comp_arena(), 2, 0, nullptr),
683 _string_late_inlines(comp_arena(), 2, 0, nullptr),
684 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
685 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
686 _late_inlines_pos(0),
687 _number_of_mh_late_inlines(0),
688 _oom(false),
689 _print_inlining_stream(new (mtCompiler) stringStream()),
690 _print_inlining_list(nullptr),
691 _print_inlining_idx(0),
692 _print_inlining_output(nullptr),
693 _replay_inline_data(nullptr),
694 _java_calls(0),
695 _inner_loops(0),
696 _interpreter_frame_size(0),
697 _output(nullptr)
698 #ifndef PRODUCT
699 , _in_dump_cnt(0)
700 #endif
701 {
702 C = this;
703 CompileWrapper cw(this);
704
705 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
706 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
707
708 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
709 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
710 // We can always print a disassembly, either abstract (hex dump) or
711 // with the help of a suitable hsdis library. Thus, we should not
712 // couple print_assembly and print_opto_assembly controls.
713 // But: always print opto and regular assembly on compile command 'print'.
714 bool print_assembly = directive->PrintAssemblyOption;
715 set_print_assembly(print_opto_assembly || print_assembly);
716 #else
717 set_print_assembly(false); // must initialize.
718 #endif
719
720 #ifndef PRODUCT
721 set_parsed_irreducible_loop(false);
722 #endif
723
724 if (directive->ReplayInlineOption) {
725 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
726 }
727 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
728 set_print_intrinsics(directive->PrintIntrinsicsOption);
729 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
730
731 if (ProfileTraps) {
732 // Make sure the method being compiled gets its own MDO,
733 // so we can at least track the decompile_count().
734 method()->ensure_method_data();
735 }
736
737 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
738 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout) {
739 initialize_stress_seed(directive);
740 }
741
742 Init(/*do_aliasing=*/ true);
743
744 print_compile_messages();
745
746 _ilt = InlineTree::build_inline_tree_root();
747
748 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
749 assert(num_alias_types() >= AliasIdxRaw, "");
750
751 #define MINIMUM_NODE_HASH 1023
752
753 // GVN that will be run immediately on new nodes
754 uint estimated_size = method()->code_size()*4+64;
755 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
756 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
757 _types = new (comp_arena()) Type_Array(comp_arena());
758 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
759 PhaseGVN gvn;
760 set_initial_gvn(&gvn);
761
762 print_inlining_init();
763 { // Scope for timing the parser
764 TracePhase tp("parse", &timers[_t_parser]);
765
766 // Put top into the hash table ASAP.
767 initial_gvn()->transform(top());
768
769 // Set up tf(), start(), and find a CallGenerator.
770 CallGenerator* cg = nullptr;
771 if (is_osr_compilation()) {
772 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
773 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
774 initial_gvn()->set_type_bottom(s);
775 verify_start(s);
776 cg = CallGenerator::for_osr(method(), entry_bci());
777 } else {
778 // Normal case.
779 init_tf(TypeFunc::make(method()));
780 StartNode* s = new StartNode(root(), tf()->domain_cc());
781 initial_gvn()->set_type_bottom(s);
782 verify_start(s);
783 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
784 // With java.lang.ref.reference.get() we must go through the
785 // intrinsic - even when get() is the root
786 // method of the compile - so that, if necessary, the value in
787 // the referent field of the reference object gets recorded by
788 // the pre-barrier code.
789 cg = find_intrinsic(method(), false);
790 }
791 if (cg == nullptr) {
792 float past_uses = method()->interpreter_invocation_count();
793 float expected_uses = past_uses;
794 cg = CallGenerator::for_inline(method(), expected_uses);
795 }
796 }
797 if (failing()) return;
798 if (cg == nullptr) {
799 const char* reason = InlineTree::check_can_parse(method());
800 assert(reason != nullptr, "expect reason for parse failure");
801 stringStream ss;
802 ss.print("cannot parse method: %s", reason);
803 record_method_not_compilable(ss.as_string());
804 return;
805 }
806
807 gvn.set_type(root(), root()->bottom_type());
808
809 JVMState* jvms = build_start_state(start(), tf());
810 if ((jvms = cg->generate(jvms)) == nullptr) {
811 assert(failure_reason() != nullptr, "expect reason for parse failure");
812 stringStream ss;
813 ss.print("method parse failed: %s", failure_reason());
814 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
815 return;
816 }
817 GraphKit kit(jvms);
818
819 if (!kit.stopped()) {
820 // Accept return values, and transfer control we know not where.
821 // This is done by a special, unique ReturnNode bound to root.
822 return_values(kit.jvms());
823 }
824
825 if (kit.has_exceptions()) {
826 // Any exceptions that escape from this call must be rethrown
827 // to whatever caller is dynamically above us on the stack.
828 // This is done by a special, unique RethrowNode bound to root.
829 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
830 }
831
832 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
833
834 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
835 inline_string_calls(true);
836 }
837
838 if (failing()) return;
839
840 // Remove clutter produced by parsing.
841 if (!failing()) {
842 ResourceMark rm;
843 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
844 }
845 }
846
847 // Note: Large methods are capped off in do_one_bytecode().
848 if (failing()) return;
849
850 // After parsing, node notes are no longer automagic.
851 // They must be propagated by register_new_node_with_optimizer(),
852 // clone(), or the like.
853 set_default_node_notes(nullptr);
854
855 #ifndef PRODUCT
856 if (should_print_igv(1)) {
857 _igv_printer->print_inlining();
858 }
859 #endif
860
861 if (failing()) return;
862 NOT_PRODUCT( verify_graph_edges(); )
863
864 // Now optimize
865 Optimize();
866 if (failing()) return;
867 NOT_PRODUCT( verify_graph_edges(); )
868
869 #ifndef PRODUCT
870 if (should_print_ideal()) {
871 print_ideal_ir("print_ideal");
872 }
873 #endif
874
875 #ifdef ASSERT
876 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
877 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
878 #endif
879
880 // Dump compilation data to replay it.
881 if (directive->DumpReplayOption) {
882 env()->dump_replay_data(_compile_id);
883 }
884 if (directive->DumpInlineOption && (ilt() != nullptr)) {
885 env()->dump_inline_data(_compile_id);
886 }
887
888 // Now that we know the size of all the monitors we can add a fixed slot
889 // for the original deopt pc.
890 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
891 if (needs_stack_repair()) {
892 // One extra slot for the special stack increment value
893 next_slot += 2;
894 }
895 // TODO 8284443 Only reserve extra slot if needed
896 if (InlineTypeReturnedAsFields) {
897 // One extra slot to hold the IsInit information for a nullable
898 // inline type return if we run out of registers.
899 next_slot += 2;
900 }
901 set_fixed_slots(next_slot);
902
903 // Compute when to use implicit null checks. Used by matching trap based
904 // nodes and NullCheck optimization.
905 set_allowed_deopt_reasons();
906
907 // Now generate code
908 Code_Gen();
909 }
910
911 //------------------------------Compile----------------------------------------
912 // Compile a runtime stub
913 Compile::Compile( ciEnv* ci_env,
914 TypeFunc_generator generator,
915 address stub_function,
916 const char *stub_name,
917 int is_fancy_jump,
918 bool pass_tls,
919 bool return_pc,
920 DirectiveSet* directive)
921 : Phase(Compiler),
922 _compile_id(0),
923 _options(Options::for_runtime_stub()),
924 _method(nullptr),
925 _entry_bci(InvocationEntryBci),
926 _stub_function(stub_function),
927 _stub_name(stub_name),
928 _stub_entry_point(nullptr),
929 _max_node_limit(MaxNodeLimit),
930 _post_loop_opts_phase(false),
931 _allow_macro_nodes(true),
932 _inlining_progress(false),
933 _inlining_incrementally(false),
934 _has_reserved_stack_access(false),
935 _has_circular_inline_type(false),
936 #ifndef PRODUCT
937 _igv_idx(0),
938 _trace_opto_output(directive->TraceOptoOutputOption),
939 #endif
940 _has_method_handle_invokes(false),
941 _clinit_barrier_on_entry(false),
942 _stress_seed(0),
943 _comp_arena(mtCompiler),
944 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
945 _env(ci_env),
946 _directive(directive),
947 _log(ci_env->log()),
948 _first_failure_details(nullptr),
949 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
950 _congraph(nullptr),
951 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
952 _unique(0),
953 _dead_node_count(0),
954 _dead_node_list(comp_arena()),
955 _node_arena_one(mtCompiler),
956 _node_arena_two(mtCompiler),
957 _node_arena(&_node_arena_one),
958 _mach_constant_base_node(nullptr),
959 _Compile_types(mtCompiler),
960 _initial_gvn(nullptr),
961 _igvn_worklist(nullptr),
962 _types(nullptr),
963 _node_hash(nullptr),
964 _number_of_mh_late_inlines(0),
965 _oom(false),
966 _print_inlining_stream(new (mtCompiler) stringStream()),
967 _print_inlining_list(nullptr),
968 _print_inlining_idx(0),
969 _print_inlining_output(nullptr),
970 _replay_inline_data(nullptr),
971 _java_calls(0),
972 _inner_loops(0),
973 _interpreter_frame_size(0),
974 _output(nullptr),
975 #ifndef PRODUCT
976 _in_dump_cnt(0),
977 #endif
978 _allowed_reasons(0) {
979 C = this;
980
981 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
982 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
983
984 #ifndef PRODUCT
985 set_print_assembly(PrintFrameConverterAssembly);
986 set_parsed_irreducible_loop(false);
987 #else
988 set_print_assembly(false); // Must initialize.
989 #endif
990 set_has_irreducible_loop(false); // no loops
991
992 CompileWrapper cw(this);
993 Init(/*do_aliasing=*/ false);
994 init_tf((*generator)());
995
996 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
997 _types = new (comp_arena()) Type_Array(comp_arena());
998 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
999
1000 if (StressLCM || StressGCM || StressBailout) {
1001 initialize_stress_seed(directive);
1002 }
1003
1004 {
1005 PhaseGVN gvn;
1006 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1007 gvn.transform(top());
1008
1009 GraphKit kit;
1010 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1011 }
1012
1013 NOT_PRODUCT( verify_graph_edges(); )
1014
1015 Code_Gen();
1016 }
1017
1018 Compile::~Compile() {
1019 delete _print_inlining_stream;
1020 delete _first_failure_details;
1021 };
1022
1023 //------------------------------Init-------------------------------------------
1024 // Prepare for a single compilation
1025 void Compile::Init(bool aliasing) {
1026 _do_aliasing = aliasing;
1027 _unique = 0;
1028 _regalloc = nullptr;
1029
1030 _tf = nullptr; // filled in later
1031 _top = nullptr; // cached later
1032 _matcher = nullptr; // filled in later
1033 _cfg = nullptr; // filled in later
1034
1035 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1036
1037 _node_note_array = nullptr;
1038 _default_node_notes = nullptr;
1039 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1040
1041 _immutable_memory = nullptr; // filled in at first inquiry
1042
1043 #ifdef ASSERT
1044 _phase_optimize_finished = false;
1045 _phase_verify_ideal_loop = false;
1046 _exception_backedge = false;
1047 _type_verify = nullptr;
1048 #endif
1049
1050 // Globally visible Nodes
1051 // First set TOP to null to give safe behavior during creation of RootNode
1052 set_cached_top_node(nullptr);
1053 set_root(new RootNode());
1054 // Now that you have a Root to point to, create the real TOP
1055 set_cached_top_node( new ConNode(Type::TOP) );
1056 set_recent_alloc(nullptr, nullptr);
1057
1058 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1059 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1060 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1061 env()->set_dependencies(new Dependencies(env()));
1062
1063 _fixed_slots = 0;
1064 set_has_split_ifs(false);
1065 set_has_loops(false); // first approximation
1066 set_has_stringbuilder(false);
1067 set_has_boxed_value(false);
1068 _trap_can_recompile = false; // no traps emitted yet
1069 _major_progress = true; // start out assuming good things will happen
1070 set_has_unsafe_access(false);
1071 set_max_vector_size(0);
1072 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1073 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1074 set_decompile_count(0);
1075
1076 #ifndef PRODUCT
1077 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1078 #endif
1079
1080 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1081 _loop_opts_cnt = LoopOptsCount;
1082 _has_flat_accesses = false;
1083 _flat_accesses_share_alias = true;
1084 _scalarize_in_safepoints = false;
1085
1086 set_do_inlining(Inline);
1087 set_max_inline_size(MaxInlineSize);
1088 set_freq_inline_size(FreqInlineSize);
1089 set_do_scheduling(OptoScheduling);
1090
1091 set_do_vector_loop(false);
1092 set_has_monitors(false);
1093 set_has_scoped_access(false);
1094
1095 if (AllowVectorizeOnDemand) {
1096 if (has_method() && _directive->VectorizeOption) {
1097 set_do_vector_loop(true);
1098 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());})
1099 } else if (has_method() && method()->name() != nullptr &&
1100 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1101 set_do_vector_loop(true);
1102 }
1103 }
1104 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1105 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());})
1106
1107 _max_node_limit = _directive->MaxNodeLimitOption;
1108
1109 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1110 set_clinit_barrier_on_entry(true);
1111 }
1112 if (debug_info()->recording_non_safepoints()) {
1113 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1114 (comp_arena(), 8, 0, nullptr));
1115 set_default_node_notes(Node_Notes::make(this));
1116 }
1117
1118 const int grow_ats = 16;
1119 _max_alias_types = grow_ats;
1120 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1121 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1122 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1123 {
1124 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1125 }
1126 // Initialize the first few types.
1127 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1128 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1129 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1130 _num_alias_types = AliasIdxRaw+1;
1131 // Zero out the alias type cache.
1132 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1133 // A null adr_type hits in the cache right away. Preload the right answer.
1134 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1135 }
1136
1137 #ifdef ASSERT
1138 // Verify that the current StartNode is valid.
1139 void Compile::verify_start(StartNode* s) const {
1140 assert(failing_internal() || s == start(), "should be StartNode");
1141 }
1142 #endif
1143
1144 /**
1145 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1146 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1147 * the ideal graph.
1148 */
1149 StartNode* Compile::start() const {
1150 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1151 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1152 Node* start = root()->fast_out(i);
1153 if (start->is_Start()) {
1154 return start->as_Start();
1155 }
1156 }
1157 fatal("Did not find Start node!");
1158 return nullptr;
1159 }
1160
1161 //-------------------------------immutable_memory-------------------------------------
1162 // Access immutable memory
1163 Node* Compile::immutable_memory() {
1164 if (_immutable_memory != nullptr) {
1165 return _immutable_memory;
1166 }
1167 StartNode* s = start();
1168 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1169 Node *p = s->fast_out(i);
1170 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1171 _immutable_memory = p;
1172 return _immutable_memory;
1173 }
1174 }
1175 ShouldNotReachHere();
1176 return nullptr;
1177 }
1178
1179 //----------------------set_cached_top_node------------------------------------
1180 // Install the cached top node, and make sure Node::is_top works correctly.
1181 void Compile::set_cached_top_node(Node* tn) {
1182 if (tn != nullptr) verify_top(tn);
1183 Node* old_top = _top;
1184 _top = tn;
1185 // Calling Node::setup_is_top allows the nodes the chance to adjust
1186 // their _out arrays.
1187 if (_top != nullptr) _top->setup_is_top();
1188 if (old_top != nullptr) old_top->setup_is_top();
1189 assert(_top == nullptr || top()->is_top(), "");
1190 }
1191
1192 #ifdef ASSERT
1193 uint Compile::count_live_nodes_by_graph_walk() {
1194 Unique_Node_List useful(comp_arena());
1195 // Get useful node list by walking the graph.
1196 identify_useful_nodes(useful);
1197 return useful.size();
1198 }
1199
1200 void Compile::print_missing_nodes() {
1201
1202 // Return if CompileLog is null and PrintIdealNodeCount is false.
1203 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1204 return;
1205 }
1206
1207 // This is an expensive function. It is executed only when the user
1208 // specifies VerifyIdealNodeCount option or otherwise knows the
1209 // additional work that needs to be done to identify reachable nodes
1210 // by walking the flow graph and find the missing ones using
1211 // _dead_node_list.
1212
1213 Unique_Node_List useful(comp_arena());
1214 // Get useful node list by walking the graph.
1215 identify_useful_nodes(useful);
1216
1217 uint l_nodes = C->live_nodes();
1218 uint l_nodes_by_walk = useful.size();
1219
1220 if (l_nodes != l_nodes_by_walk) {
1221 if (_log != nullptr) {
1222 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1223 _log->stamp();
1224 _log->end_head();
1225 }
1226 VectorSet& useful_member_set = useful.member_set();
1227 int last_idx = l_nodes_by_walk;
1228 for (int i = 0; i < last_idx; i++) {
1229 if (useful_member_set.test(i)) {
1230 if (_dead_node_list.test(i)) {
1231 if (_log != nullptr) {
1232 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1233 }
1234 if (PrintIdealNodeCount) {
1235 // Print the log message to tty
1236 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1237 useful.at(i)->dump();
1238 }
1239 }
1240 }
1241 else if (! _dead_node_list.test(i)) {
1242 if (_log != nullptr) {
1243 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1244 }
1245 if (PrintIdealNodeCount) {
1246 // Print the log message to tty
1247 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1248 }
1249 }
1250 }
1251 if (_log != nullptr) {
1252 _log->tail("mismatched_nodes");
1253 }
1254 }
1255 }
1256 void Compile::record_modified_node(Node* n) {
1257 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1258 _modified_nodes->push(n);
1259 }
1260 }
1261
1262 void Compile::remove_modified_node(Node* n) {
1263 if (_modified_nodes != nullptr) {
1264 _modified_nodes->remove(n);
1265 }
1266 }
1267 #endif
1268
1269 #ifndef PRODUCT
1270 void Compile::verify_top(Node* tn) const {
1271 if (tn != nullptr) {
1272 assert(tn->is_Con(), "top node must be a constant");
1273 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1274 assert(tn->in(0) != nullptr, "must have live top node");
1275 }
1276 }
1277 #endif
1278
1279
1280 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1281
1282 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1283 guarantee(arr != nullptr, "");
1284 int num_blocks = arr->length();
1285 if (grow_by < num_blocks) grow_by = num_blocks;
1286 int num_notes = grow_by * _node_notes_block_size;
1287 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1288 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1289 while (num_notes > 0) {
1290 arr->append(notes);
1291 notes += _node_notes_block_size;
1292 num_notes -= _node_notes_block_size;
1293 }
1294 assert(num_notes == 0, "exact multiple, please");
1295 }
1296
1297 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1298 if (source == nullptr || dest == nullptr) return false;
1299
1300 if (dest->is_Con())
1301 return false; // Do not push debug info onto constants.
1302
1303 #ifdef ASSERT
1304 // Leave a bread crumb trail pointing to the original node:
1305 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1306 dest->set_debug_orig(source);
1307 }
1308 #endif
1309
1310 if (node_note_array() == nullptr)
1311 return false; // Not collecting any notes now.
1312
1313 // This is a copy onto a pre-existing node, which may already have notes.
1314 // If both nodes have notes, do not overwrite any pre-existing notes.
1315 Node_Notes* source_notes = node_notes_at(source->_idx);
1316 if (source_notes == nullptr || source_notes->is_clear()) return false;
1317 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1318 if (dest_notes == nullptr || dest_notes->is_clear()) {
1319 return set_node_notes_at(dest->_idx, source_notes);
1320 }
1321
1322 Node_Notes merged_notes = (*source_notes);
1323 // The order of operations here ensures that dest notes will win...
1324 merged_notes.update_from(dest_notes);
1325 return set_node_notes_at(dest->_idx, &merged_notes);
1326 }
1327
1328
1329 //--------------------------allow_range_check_smearing-------------------------
1330 // Gating condition for coalescing similar range checks.
1331 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1332 // single covering check that is at least as strong as any of them.
1333 // If the optimization succeeds, the simplified (strengthened) range check
1334 // will always succeed. If it fails, we will deopt, and then give up
1335 // on the optimization.
1336 bool Compile::allow_range_check_smearing() const {
1337 // If this method has already thrown a range-check,
1338 // assume it was because we already tried range smearing
1339 // and it failed.
1340 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1341 return !already_trapped;
1342 }
1343
1344
1345 //------------------------------flatten_alias_type-----------------------------
1346 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1347 assert(do_aliasing(), "Aliasing should be enabled");
1348 int offset = tj->offset();
1349 TypePtr::PTR ptr = tj->ptr();
1350
1351 // Known instance (scalarizable allocation) alias only with itself.
1352 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1353 tj->is_oopptr()->is_known_instance();
1354
1355 // Process weird unsafe references.
1356 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1357 bool default_value_load = EnableValhalla && tj->is_instptr()->instance_klass() == ciEnv::current()->Class_klass();
1358 assert(InlineUnsafeOps || StressReflectiveCode || default_value_load, "indeterminate pointers come only from unsafe ops");
1359 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1360 tj = TypeOopPtr::BOTTOM;
1361 ptr = tj->ptr();
1362 offset = tj->offset();
1363 }
1364
1365 // Array pointers need some flattening
1366 const TypeAryPtr* ta = tj->isa_aryptr();
1367 if (ta && ta->is_stable()) {
1368 // Erase stability property for alias analysis.
1369 tj = ta = ta->cast_to_stable(false);
1370 }
1371 if (ta && ta->is_not_flat()) {
1372 // Erase not flat property for alias analysis.
1373 tj = ta = ta->cast_to_not_flat(false);
1374 }
1375 if (ta && ta->is_not_null_free()) {
1376 // Erase not null free property for alias analysis.
1377 tj = ta = ta->cast_to_not_null_free(false);
1378 }
1379
1380 if( ta && is_known_inst ) {
1381 if ( offset != Type::OffsetBot &&
1382 offset > arrayOopDesc::length_offset_in_bytes() ) {
1383 offset = Type::OffsetBot; // Flatten constant access into array body only
1384 tj = ta = ta->
1385 remove_speculative()->
1386 cast_to_ptr_type(ptr)->
1387 with_offset(offset);
1388 }
1389 } else if (ta) {
1390 // For arrays indexed by constant indices, we flatten the alias
1391 // space to include all of the array body. Only the header, klass
1392 // and array length can be accessed un-aliased.
1393 // For flat inline type array, each field has its own slice so
1394 // we must include the field offset.
1395 if( offset != Type::OffsetBot ) {
1396 if( ta->const_oop() ) { // MethodData* or Method*
1397 offset = Type::OffsetBot; // Flatten constant access into array body
1398 tj = ta = ta->
1399 remove_speculative()->
1400 cast_to_ptr_type(ptr)->
1401 cast_to_exactness(false)->
1402 with_offset(offset);
1403 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1404 // range is OK as-is.
1405 tj = ta = TypeAryPtr::RANGE;
1406 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1407 tj = TypeInstPtr::KLASS; // all klass loads look alike
1408 ta = TypeAryPtr::RANGE; // generic ignored junk
1409 ptr = TypePtr::BotPTR;
1410 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1411 tj = TypeInstPtr::MARK;
1412 ta = TypeAryPtr::RANGE; // generic ignored junk
1413 ptr = TypePtr::BotPTR;
1414 } else { // Random constant offset into array body
1415 offset = Type::OffsetBot; // Flatten constant access into array body
1416 tj = ta = ta->
1417 remove_speculative()->
1418 cast_to_ptr_type(ptr)->
1419 cast_to_exactness(false)->
1420 with_offset(offset);
1421 }
1422 }
1423 // Arrays of fixed size alias with arrays of unknown size.
1424 if (ta->size() != TypeInt::POS) {
1425 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1426 tj = ta = ta->
1427 remove_speculative()->
1428 cast_to_ptr_type(ptr)->
1429 with_ary(tary)->
1430 cast_to_exactness(false);
1431 }
1432 // Arrays of known objects become arrays of unknown objects.
1433 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1434 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1435 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1436 }
1437 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1438 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1439 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1440 }
1441 // Initially all flattened array accesses share a single slice
1442 if (ta->is_flat() && ta->elem() != TypeInstPtr::BOTTOM && _flat_accesses_share_alias) {
1443 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size(), /* stable= */ false, /* flat= */ true);
1444 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1445 }
1446 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1447 // cannot be distinguished by bytecode alone.
1448 if (ta->elem() == TypeInt::BOOL) {
1449 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1450 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1451 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1452 }
1453 // During the 2nd round of IterGVN, NotNull castings are removed.
1454 // Make sure the Bottom and NotNull variants alias the same.
1455 // Also, make sure exact and non-exact variants alias the same.
1456 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1457 tj = ta = ta->
1458 remove_speculative()->
1459 cast_to_ptr_type(TypePtr::BotPTR)->
1460 cast_to_exactness(false)->
1461 with_offset(offset);
1462 }
1463 }
1464
1465 // Oop pointers need some flattening
1466 const TypeInstPtr *to = tj->isa_instptr();
1467 if (to && to != TypeOopPtr::BOTTOM) {
1468 ciInstanceKlass* ik = to->instance_klass();
1469 if( ptr == TypePtr::Constant ) {
1470 if (ik != ciEnv::current()->Class_klass() ||
1471 offset < ik->layout_helper_size_in_bytes()) {
1472 // No constant oop pointers (such as Strings); they alias with
1473 // unknown strings.
1474 assert(!is_known_inst, "not scalarizable allocation");
1475 tj = to = to->
1476 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1477 remove_speculative()->
1478 cast_to_ptr_type(TypePtr::BotPTR)->
1479 cast_to_exactness(false);
1480 }
1481 } else if( is_known_inst ) {
1482 tj = to; // Keep NotNull and klass_is_exact for instance type
1483 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1484 // During the 2nd round of IterGVN, NotNull castings are removed.
1485 // Make sure the Bottom and NotNull variants alias the same.
1486 // Also, make sure exact and non-exact variants alias the same.
1487 tj = to = to->
1488 remove_speculative()->
1489 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1490 cast_to_ptr_type(TypePtr::BotPTR)->
1491 cast_to_exactness(false);
1492 }
1493 if (to->speculative() != nullptr) {
1494 tj = to = to->remove_speculative();
1495 }
1496 // Canonicalize the holder of this field
1497 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1498 // First handle header references such as a LoadKlassNode, even if the
1499 // object's klass is unloaded at compile time (4965979).
1500 if (!is_known_inst) { // Do it only for non-instance types
1501 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1502 }
1503 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1504 // Static fields are in the space above the normal instance
1505 // fields in the java.lang.Class instance.
1506 if (ik != ciEnv::current()->Class_klass()) {
1507 to = nullptr;
1508 tj = TypeOopPtr::BOTTOM;
1509 offset = tj->offset();
1510 }
1511 } else {
1512 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1513 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1514 assert(tj->offset() == offset, "no change to offset expected");
1515 bool xk = to->klass_is_exact();
1516 int instance_id = to->instance_id();
1517
1518 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1519 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1520 // its interfaces are included.
1521 if (xk && ik->equals(canonical_holder)) {
1522 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id), "exact type should be canonical type");
1523 } else {
1524 assert(xk || !is_known_inst, "Known instance should be exact type");
1525 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id);
1526 }
1527 }
1528 }
1529
1530 // Klass pointers to object array klasses need some flattening
1531 const TypeKlassPtr *tk = tj->isa_klassptr();
1532 if( tk ) {
1533 // If we are referencing a field within a Klass, we need
1534 // to assume the worst case of an Object. Both exact and
1535 // inexact types must flatten to the same alias class so
1536 // use NotNull as the PTR.
1537 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1538 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1539 env()->Object_klass(),
1540 Type::Offset(offset));
1541 }
1542
1543 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1544 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1545 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1546 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset));
1547 } else {
1548 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, Type::Offset(offset), tk->is_not_flat(), tk->is_not_null_free(), tk->is_null_free());
1549 }
1550 }
1551 // Check for precise loads from the primary supertype array and force them
1552 // to the supertype cache alias index. Check for generic array loads from
1553 // the primary supertype array and also force them to the supertype cache
1554 // alias index. Since the same load can reach both, we need to merge
1555 // these 2 disparate memories into the same alias class. Since the
1556 // primary supertype array is read-only, there's no chance of confusion
1557 // where we bypass an array load and an array store.
1558 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1559 if (offset == Type::OffsetBot ||
1560 (offset >= primary_supers_offset &&
1561 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1562 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1563 offset = in_bytes(Klass::secondary_super_cache_offset());
1564 tj = tk = tk->with_offset(offset);
1565 }
1566 }
1567
1568 // Flatten all Raw pointers together.
1569 if (tj->base() == Type::RawPtr)
1570 tj = TypeRawPtr::BOTTOM;
1571
1572 if (tj->base() == Type::AnyPtr)
1573 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1574
1575 offset = tj->offset();
1576 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1577
1578 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1579 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1580 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1581 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1582 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1583 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1584 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1585 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1586 assert( tj->ptr() != TypePtr::TopPTR &&
1587 tj->ptr() != TypePtr::AnyNull &&
1588 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1589 // assert( tj->ptr() != TypePtr::Constant ||
1590 // tj->base() == Type::RawPtr ||
1591 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1592
1593 return tj;
1594 }
1595
1596 void Compile::AliasType::Init(int i, const TypePtr* at) {
1597 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1598 _index = i;
1599 _adr_type = at;
1600 _field = nullptr;
1601 _element = nullptr;
1602 _is_rewritable = true; // default
1603 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1604 if (atoop != nullptr && atoop->is_known_instance()) {
1605 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1606 _general_index = Compile::current()->get_alias_index(gt);
1607 } else {
1608 _general_index = 0;
1609 }
1610 }
1611
1612 BasicType Compile::AliasType::basic_type() const {
1613 if (element() != nullptr) {
1614 const Type* element = adr_type()->is_aryptr()->elem();
1615 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1616 } if (field() != nullptr) {
1617 return field()->layout_type();
1618 } else {
1619 return T_ILLEGAL; // unknown
1620 }
1621 }
1622
1623 //---------------------------------print_on------------------------------------
1624 #ifndef PRODUCT
1625 void Compile::AliasType::print_on(outputStream* st) {
1626 if (index() < 10)
1627 st->print("@ <%d> ", index());
1628 else st->print("@ <%d>", index());
1629 st->print(is_rewritable() ? " " : " RO");
1630 int offset = adr_type()->offset();
1631 if (offset == Type::OffsetBot)
1632 st->print(" +any");
1633 else st->print(" +%-3d", offset);
1634 st->print(" in ");
1635 adr_type()->dump_on(st);
1636 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1637 if (field() != nullptr && tjp) {
1638 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1639 tjp->offset() != field()->offset_in_bytes()) {
1640 st->print(" != ");
1641 field()->print();
1642 st->print(" ***");
1643 }
1644 }
1645 }
1646
1647 void print_alias_types() {
1648 Compile* C = Compile::current();
1649 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1650 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1651 C->alias_type(idx)->print_on(tty);
1652 tty->cr();
1653 }
1654 }
1655 #endif
1656
1657
1658 //----------------------------probe_alias_cache--------------------------------
1659 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1660 intptr_t key = (intptr_t) adr_type;
1661 key ^= key >> logAliasCacheSize;
1662 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1663 }
1664
1665
1666 //-----------------------------grow_alias_types--------------------------------
1667 void Compile::grow_alias_types() {
1668 const int old_ats = _max_alias_types; // how many before?
1669 const int new_ats = old_ats; // how many more?
1670 const int grow_ats = old_ats+new_ats; // how many now?
1671 _max_alias_types = grow_ats;
1672 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1673 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1674 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1675 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1676 }
1677
1678
1679 //--------------------------------find_alias_type------------------------------
1680 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1681 if (!do_aliasing()) {
1682 return alias_type(AliasIdxBot);
1683 }
1684
1685 AliasCacheEntry* ace = nullptr;
1686 if (!uncached) {
1687 ace = probe_alias_cache(adr_type);
1688 if (ace->_adr_type == adr_type) {
1689 return alias_type(ace->_index);
1690 }
1691 }
1692
1693 // Handle special cases.
1694 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1695 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1696
1697 // Do it the slow way.
1698 const TypePtr* flat = flatten_alias_type(adr_type);
1699
1700 #ifdef ASSERT
1701 {
1702 ResourceMark rm;
1703 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1704 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1705 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1706 Type::str(adr_type));
1707 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1708 const TypeOopPtr* foop = flat->is_oopptr();
1709 // Scalarizable allocations have exact klass always.
1710 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1711 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1712 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1713 Type::str(foop), Type::str(xoop));
1714 }
1715 }
1716 #endif
1717
1718 int idx = AliasIdxTop;
1719 for (int i = 0; i < num_alias_types(); i++) {
1720 if (alias_type(i)->adr_type() == flat) {
1721 idx = i;
1722 break;
1723 }
1724 }
1725
1726 if (idx == AliasIdxTop) {
1727 if (no_create) return nullptr;
1728 // Grow the array if necessary.
1729 if (_num_alias_types == _max_alias_types) grow_alias_types();
1730 // Add a new alias type.
1731 idx = _num_alias_types++;
1732 _alias_types[idx]->Init(idx, flat);
1733 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1734 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1735 if (flat->isa_instptr()) {
1736 if (flat->offset() == java_lang_Class::klass_offset()
1737 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1738 alias_type(idx)->set_rewritable(false);
1739 }
1740 ciField* field = nullptr;
1741 if (flat->isa_aryptr()) {
1742 #ifdef ASSERT
1743 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1744 // (T_BYTE has the weakest alignment and size restrictions...)
1745 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1746 #endif
1747 const Type* elemtype = flat->is_aryptr()->elem();
1748 if (flat->offset() == TypePtr::OffsetBot) {
1749 alias_type(idx)->set_element(elemtype);
1750 }
1751 int field_offset = flat->is_aryptr()->field_offset().get();
1752 if (flat->is_flat() &&
1753 field_offset != Type::OffsetBot) {
1754 ciInlineKlass* vk = elemtype->inline_klass();
1755 field_offset += vk->first_field_offset();
1756 field = vk->get_field_by_offset(field_offset, false);
1757 }
1758 }
1759 if (flat->isa_klassptr()) {
1760 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1761 alias_type(idx)->set_rewritable(false);
1762 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1763 alias_type(idx)->set_rewritable(false);
1764 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1765 alias_type(idx)->set_rewritable(false);
1766 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1767 alias_type(idx)->set_rewritable(false);
1768 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1769 alias_type(idx)->set_rewritable(false);
1770 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1771 alias_type(idx)->set_rewritable(false);
1772 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1773 alias_type(idx)->set_rewritable(false);
1774 }
1775 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1776 // but the base pointer type is not distinctive enough to identify
1777 // references into JavaThread.)
1778
1779 // Check for final fields.
1780 const TypeInstPtr* tinst = flat->isa_instptr();
1781 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1782 if (tinst->const_oop() != nullptr &&
1783 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1784 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1785 // static field
1786 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1787 field = k->get_field_by_offset(tinst->offset(), true);
1788 } else if (tinst->is_inlinetypeptr()) {
1789 // Inline type field
1790 ciInlineKlass* vk = tinst->inline_klass();
1791 field = vk->get_field_by_offset(tinst->offset(), false);
1792 } else {
1793 ciInstanceKlass *k = tinst->instance_klass();
1794 field = k->get_field_by_offset(tinst->offset(), false);
1795 }
1796 }
1797 assert(field == nullptr ||
1798 original_field == nullptr ||
1799 (field->holder() == original_field->holder() &&
1800 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1801 field->is_static() == original_field->is_static()), "wrong field?");
1802 // Set field() and is_rewritable() attributes.
1803 if (field != nullptr) {
1804 alias_type(idx)->set_field(field);
1805 if (flat->isa_aryptr()) {
1806 // Fields of flat arrays are rewritable although they are declared final
1807 assert(flat->is_flat(), "must be a flat array");
1808 alias_type(idx)->set_rewritable(true);
1809 }
1810 }
1811 }
1812
1813 // Fill the cache for next time.
1814 if (!uncached) {
1815 ace->_adr_type = adr_type;
1816 ace->_index = idx;
1817 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1818
1819 // Might as well try to fill the cache for the flattened version, too.
1820 AliasCacheEntry* face = probe_alias_cache(flat);
1821 if (face->_adr_type == nullptr) {
1822 face->_adr_type = flat;
1823 face->_index = idx;
1824 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1825 }
1826 }
1827
1828 return alias_type(idx);
1829 }
1830
1831
1832 Compile::AliasType* Compile::alias_type(ciField* field) {
1833 const TypeOopPtr* t;
1834 if (field->is_static())
1835 t = TypeInstPtr::make(field->holder()->java_mirror());
1836 else
1837 t = TypeOopPtr::make_from_klass_raw(field->holder());
1838 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1839 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1840 return atp;
1841 }
1842
1843
1844 //------------------------------have_alias_type--------------------------------
1845 bool Compile::have_alias_type(const TypePtr* adr_type) {
1846 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1847 if (ace->_adr_type == adr_type) {
1848 return true;
1849 }
1850
1851 // Handle special cases.
1852 if (adr_type == nullptr) return true;
1853 if (adr_type == TypePtr::BOTTOM) return true;
1854
1855 return find_alias_type(adr_type, true, nullptr) != nullptr;
1856 }
1857
1858 //-----------------------------must_alias--------------------------------------
1859 // True if all values of the given address type are in the given alias category.
1860 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1861 if (alias_idx == AliasIdxBot) return true; // the universal category
1862 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1863 if (alias_idx == AliasIdxTop) return false; // the empty category
1864 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1865
1866 // the only remaining possible overlap is identity
1867 int adr_idx = get_alias_index(adr_type);
1868 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1869 assert(adr_idx == alias_idx ||
1870 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1871 && adr_type != TypeOopPtr::BOTTOM),
1872 "should not be testing for overlap with an unsafe pointer");
1873 return adr_idx == alias_idx;
1874 }
1875
1876 //------------------------------can_alias--------------------------------------
1877 // True if any values of the given address type are in the given alias category.
1878 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1879 if (alias_idx == AliasIdxTop) return false; // the empty category
1880 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1881 // Known instance doesn't alias with bottom memory
1882 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1883 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1884
1885 // the only remaining possible overlap is identity
1886 int adr_idx = get_alias_index(adr_type);
1887 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1888 return adr_idx == alias_idx;
1889 }
1890
1891 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1892 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1893 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1894 if (parse_predicate_count() == 0) {
1895 return;
1896 }
1897 for (int i = 0; i < parse_predicate_count(); i++) {
1898 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1899 parse_predicate->mark_useless();
1900 igvn._worklist.push(parse_predicate);
1901 }
1902 _parse_predicates.clear();
1903 }
1904
1905 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1906 if (!n->for_post_loop_opts_igvn()) {
1907 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1908 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1909 _for_post_loop_igvn.append(n);
1910 }
1911 }
1912
1913 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1914 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1915 _for_post_loop_igvn.remove(n);
1916 }
1917
1918 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1919 // Verify that all previous optimizations produced a valid graph
1920 // at least to this point, even if no loop optimizations were done.
1921 PhaseIdealLoop::verify(igvn);
1922
1923 C->set_post_loop_opts_phase(); // no more loop opts allowed
1924
1925 assert(!C->major_progress(), "not cleared");
1926
1927 if (_for_post_loop_igvn.length() > 0) {
1928 while (_for_post_loop_igvn.length() > 0) {
1929 Node* n = _for_post_loop_igvn.pop();
1930 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1931 igvn._worklist.push(n);
1932 }
1933 igvn.optimize();
1934 if (failing()) return;
1935 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1936 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1937
1938 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1939 if (C->major_progress()) {
1940 C->clear_major_progress(); // ensure that major progress is now clear
1941 }
1942 }
1943 }
1944
1945 void Compile::add_inline_type(Node* n) {
1946 assert(n->is_InlineType(), "unexpected node");
1947 _inline_type_nodes.push(n);
1948 }
1949
1950 void Compile::remove_inline_type(Node* n) {
1951 assert(n->is_InlineType(), "unexpected node");
1952 if (_inline_type_nodes.contains(n)) {
1953 _inline_type_nodes.remove(n);
1954 }
1955 }
1956
1957 // Does the return value keep otherwise useless inline type allocations alive?
1958 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1959 ResourceMark rm;
1960 Unique_Node_List wq;
1961 wq.push(ret_val);
1962 bool some_allocations = false;
1963 for (uint i = 0; i < wq.size(); i++) {
1964 Node* n = wq.at(i);
1965 if (n->outcnt() > 1) {
1966 // Some other use for the allocation
1967 return false;
1968 } else if (n->is_InlineType()) {
1969 wq.push(n->in(1));
1970 } else if (n->is_Phi()) {
1971 for (uint j = 1; j < n->req(); j++) {
1972 wq.push(n->in(j));
1973 }
1974 } else if (n->is_CheckCastPP() &&
1975 n->in(1)->is_Proj() &&
1976 n->in(1)->in(0)->is_Allocate()) {
1977 some_allocations = true;
1978 } else if (n->is_CheckCastPP()) {
1979 wq.push(n->in(1));
1980 }
1981 }
1982 return some_allocations;
1983 }
1984
1985 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1986 // Make sure that the return value does not keep an otherwise unused allocation alive
1987 if (tf()->returns_inline_type_as_fields()) {
1988 Node* ret = nullptr;
1989 for (uint i = 1; i < root()->req(); i++) {
1990 Node* in = root()->in(i);
1991 if (in->Opcode() == Op_Return) {
1992 assert(ret == nullptr, "only one return");
1993 ret = in;
1994 }
1995 }
1996 if (ret != nullptr) {
1997 Node* ret_val = ret->in(TypeFunc::Parms);
1998 if (igvn.type(ret_val)->isa_oopptr() &&
1999 return_val_keeps_allocations_alive(ret_val)) {
2000 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
2001 assert(ret_val->outcnt() == 0, "should be dead now");
2002 igvn.remove_dead_node(ret_val);
2003 }
2004 }
2005 }
2006 if (_inline_type_nodes.length() == 0) {
2007 return;
2008 }
2009 // Scalarize inline types in safepoint debug info.
2010 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2011 set_scalarize_in_safepoints(true);
2012 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2013 InlineTypeNode* vt = _inline_type_nodes.at(i)->as_InlineType();
2014 vt->make_scalar_in_safepoints(&igvn);
2015 igvn.record_for_igvn(vt);
2016 }
2017 if (remove) {
2018 // Remove inline type nodes by replacing them with their oop input
2019 while (_inline_type_nodes.length() > 0) {
2020 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2021 if (vt->outcnt() == 0) {
2022 igvn.remove_dead_node(vt);
2023 continue;
2024 }
2025 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2026 DEBUG_ONLY(bool must_be_buffered = false);
2027 Node* u = vt->out(i);
2028 // Check if any users are blackholes. If so, rewrite them to use either the
2029 // allocated buffer, or individual components, instead of the inline type node
2030 // that goes away.
2031 if (u->is_Blackhole()) {
2032 BlackholeNode* bh = u->as_Blackhole();
2033
2034 // Unlink the old input
2035 int idx = bh->find_edge(vt);
2036 assert(idx != -1, "The edge should be there");
2037 bh->del_req(idx);
2038 --i;
2039
2040 if (vt->is_allocated(&igvn)) {
2041 // Already has the allocated instance, blackhole that
2042 bh->add_req(vt->get_oop());
2043 } else {
2044 // Not allocated yet, blackhole the components
2045 for (uint c = 0; c < vt->field_count(); c++) {
2046 bh->add_req(vt->field_value(c));
2047 }
2048 }
2049
2050 // Node modified, record for IGVN
2051 igvn.record_for_igvn(bh);
2052 }
2053 #ifdef ASSERT
2054 // Verify that inline type is buffered when replacing by oop
2055 else if (u->is_InlineType()) {
2056 // InlineType uses don't need buffering because they are about to be replaced as well
2057 } else if (u->is_Phi()) {
2058 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2059 } else {
2060 must_be_buffered = true;
2061 }
2062 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2063 vt->dump(0);
2064 u->dump(0);
2065 assert(false, "Should have been buffered");
2066 }
2067 #endif
2068 }
2069 igvn.replace_node(vt, vt->get_oop());
2070 }
2071 }
2072 igvn.optimize();
2073 }
2074
2075 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2076 if (!_has_flat_accesses) {
2077 return;
2078 }
2079 // Initially, all flat array accesses share the same slice to
2080 // keep dependencies with Object[] array accesses (that could be
2081 // to a flat array) correct. We're done with parsing so we
2082 // now know all flat array accesses in this compile
2083 // unit. Let's move flat array accesses to their own slice,
2084 // one per element field. This should help memory access
2085 // optimizations.
2086 ResourceMark rm;
2087 Unique_Node_List wq;
2088 wq.push(root());
2089
2090 Node_List mergememnodes;
2091 Node_List memnodes;
2092
2093 // Alias index currently shared by all flat memory accesses
2094 int index = get_alias_index(TypeAryPtr::INLINES);
2095
2096 // Find MergeMem nodes and flat array accesses
2097 for (uint i = 0; i < wq.size(); i++) {
2098 Node* n = wq.at(i);
2099 if (n->is_Mem()) {
2100 const TypePtr* adr_type = nullptr;
2101 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2102 if (adr_type == TypeAryPtr::INLINES) {
2103 memnodes.push(n);
2104 }
2105 } else if (n->is_MergeMem()) {
2106 MergeMemNode* mm = n->as_MergeMem();
2107 if (mm->memory_at(index) != mm->base_memory()) {
2108 mergememnodes.push(n);
2109 }
2110 }
2111 for (uint j = 0; j < n->req(); j++) {
2112 Node* m = n->in(j);
2113 if (m != nullptr) {
2114 wq.push(m);
2115 }
2116 }
2117 }
2118
2119 if (memnodes.size() > 0) {
2120 _flat_accesses_share_alias = false;
2121
2122 // We are going to change the slice for the flat array
2123 // accesses so we need to clear the cache entries that refer to
2124 // them.
2125 for (uint i = 0; i < AliasCacheSize; i++) {
2126 AliasCacheEntry* ace = &_alias_cache[i];
2127 if (ace->_adr_type != nullptr &&
2128 ace->_adr_type->is_flat()) {
2129 ace->_adr_type = nullptr;
2130 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2131 }
2132 }
2133
2134 // Find what aliases we are going to add
2135 int start_alias = num_alias_types()-1;
2136 int stop_alias = 0;
2137
2138 for (uint i = 0; i < memnodes.size(); i++) {
2139 Node* m = memnodes.at(i);
2140 const TypePtr* adr_type = nullptr;
2141 adr_type = m->adr_type();
2142 #ifdef ASSERT
2143 m->as_Mem()->set_adr_type(adr_type);
2144 #endif
2145 int idx = get_alias_index(adr_type);
2146 start_alias = MIN2(start_alias, idx);
2147 stop_alias = MAX2(stop_alias, idx);
2148 }
2149
2150 assert(stop_alias >= start_alias, "should have expanded aliases");
2151
2152 Node_Stack stack(0);
2153 #ifdef ASSERT
2154 VectorSet seen(Thread::current()->resource_area());
2155 #endif
2156 // Now let's fix the memory graph so each flat array access
2157 // is moved to the right slice. Start from the MergeMem nodes.
2158 uint last = unique();
2159 for (uint i = 0; i < mergememnodes.size(); i++) {
2160 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2161 Node* n = current->memory_at(index);
2162 MergeMemNode* mm = nullptr;
2163 do {
2164 // Follow memory edges through memory accesses, phis and
2165 // narrow membars and push nodes on the stack. Once we hit
2166 // bottom memory, we pop element off the stack one at a
2167 // time, in reverse order, and move them to the right slice
2168 // by changing their memory edges.
2169 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2170 assert(!seen.test_set(n->_idx), "");
2171 // Uses (a load for instance) will need to be moved to the
2172 // right slice as well and will get a new memory state
2173 // that we don't know yet. The use could also be the
2174 // backedge of a loop. We put a place holder node between
2175 // the memory node and its uses. We replace that place
2176 // holder with the correct memory state once we know it,
2177 // i.e. when nodes are popped off the stack. Using the
2178 // place holder make the logic work in the presence of
2179 // loops.
2180 if (n->outcnt() > 1) {
2181 Node* place_holder = nullptr;
2182 assert(!n->has_out_with(Op_Node), "");
2183 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2184 Node* u = n->out(k);
2185 if (u != current && u->_idx < last) {
2186 bool success = false;
2187 for (uint l = 0; l < u->req(); l++) {
2188 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2189 continue;
2190 }
2191 Node* in = u->in(l);
2192 if (in == n) {
2193 if (place_holder == nullptr) {
2194 place_holder = new Node(1);
2195 place_holder->init_req(0, n);
2196 }
2197 igvn.replace_input_of(u, l, place_holder);
2198 success = true;
2199 }
2200 }
2201 if (success) {
2202 --k;
2203 }
2204 }
2205 }
2206 }
2207 if (n->is_Phi()) {
2208 stack.push(n, 1);
2209 n = n->in(1);
2210 } else if (n->is_Mem()) {
2211 stack.push(n, n->req());
2212 n = n->in(MemNode::Memory);
2213 } else {
2214 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2215 stack.push(n, n->req());
2216 n = n->in(0)->in(TypeFunc::Memory);
2217 }
2218 } else {
2219 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2220 // Build a new MergeMem node to carry the new memory state
2221 // as we build it. IGVN should fold extraneous MergeMem
2222 // nodes.
2223 mm = MergeMemNode::make(n);
2224 igvn.register_new_node_with_optimizer(mm);
2225 while (stack.size() > 0) {
2226 Node* m = stack.node();
2227 uint idx = stack.index();
2228 if (m->is_Mem()) {
2229 // Move memory node to its new slice
2230 const TypePtr* adr_type = m->adr_type();
2231 int alias = get_alias_index(adr_type);
2232 Node* prev = mm->memory_at(alias);
2233 igvn.replace_input_of(m, MemNode::Memory, prev);
2234 mm->set_memory_at(alias, m);
2235 } else if (m->is_Phi()) {
2236 // We need as many new phis as there are new aliases
2237 igvn.replace_input_of(m, idx, mm);
2238 if (idx == m->req()-1) {
2239 Node* r = m->in(0);
2240 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2241 const TypePtr* adr_type = get_adr_type(j);
2242 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2243 continue;
2244 }
2245 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2246 igvn.register_new_node_with_optimizer(phi);
2247 for (uint k = 1; k < m->req(); k++) {
2248 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2249 }
2250 mm->set_memory_at(j, phi);
2251 }
2252 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2253 igvn.register_new_node_with_optimizer(base_phi);
2254 for (uint k = 1; k < m->req(); k++) {
2255 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2256 }
2257 mm->set_base_memory(base_phi);
2258 }
2259 } else {
2260 // This is a MemBarCPUOrder node from
2261 // Parse::array_load()/Parse::array_store(), in the
2262 // branch that handles flat arrays hidden under
2263 // an Object[] array. We also need one new membar per
2264 // new alias to keep the unknown access that the
2265 // membars protect properly ordered with accesses to
2266 // known flat array.
2267 assert(m->is_Proj(), "projection expected");
2268 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2269 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2270 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2271 const TypePtr* adr_type = get_adr_type(j);
2272 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2273 continue;
2274 }
2275 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2276 igvn.register_new_node_with_optimizer(mb);
2277 Node* mem = mm->memory_at(j);
2278 mb->init_req(TypeFunc::Control, ctrl);
2279 mb->init_req(TypeFunc::Memory, mem);
2280 ctrl = new ProjNode(mb, TypeFunc::Control);
2281 igvn.register_new_node_with_optimizer(ctrl);
2282 mem = new ProjNode(mb, TypeFunc::Memory);
2283 igvn.register_new_node_with_optimizer(mem);
2284 mm->set_memory_at(j, mem);
2285 }
2286 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2287 }
2288 if (idx < m->req()-1) {
2289 idx += 1;
2290 stack.set_index(idx);
2291 n = m->in(idx);
2292 break;
2293 }
2294 // Take care of place holder nodes
2295 if (m->has_out_with(Op_Node)) {
2296 Node* place_holder = m->find_out_with(Op_Node);
2297 if (place_holder != nullptr) {
2298 Node* mm_clone = mm->clone();
2299 igvn.register_new_node_with_optimizer(mm_clone);
2300 Node* hook = new Node(1);
2301 hook->init_req(0, mm);
2302 igvn.replace_node(place_holder, mm_clone);
2303 hook->destruct(&igvn);
2304 }
2305 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2306 }
2307 stack.pop();
2308 }
2309 }
2310 } while(stack.size() > 0);
2311 // Fix the memory state at the MergeMem we started from
2312 igvn.rehash_node_delayed(current);
2313 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2314 const TypePtr* adr_type = get_adr_type(j);
2315 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2316 continue;
2317 }
2318 current->set_memory_at(j, mm);
2319 }
2320 current->set_memory_at(index, current->base_memory());
2321 }
2322 igvn.optimize();
2323 }
2324 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2325 #ifdef ASSERT
2326 if (!_flat_accesses_share_alias) {
2327 wq.clear();
2328 wq.push(root());
2329 for (uint i = 0; i < wq.size(); i++) {
2330 Node* n = wq.at(i);
2331 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2332 for (uint j = 0; j < n->req(); j++) {
2333 Node* m = n->in(j);
2334 if (m != nullptr) {
2335 wq.push(m);
2336 }
2337 }
2338 }
2339 }
2340 #endif
2341 }
2342
2343 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2344 if (OptimizeUnstableIf) {
2345 _unstable_if_traps.append(trap);
2346 }
2347 }
2348
2349 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2350 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2351 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2352 Node* n = trap->uncommon_trap();
2353 if (!useful.member(n)) {
2354 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2355 }
2356 }
2357 }
2358
2359 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2360 // or fold-compares case. Return true if succeed or not found.
2361 //
2362 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2363 // false and ask fold-compares to yield.
2364 //
2365 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2366 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2367 // when deoptimization does happen.
2368 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2369 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2370 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2371 if (trap->uncommon_trap() == unc) {
2372 if (yield && trap->modified()) {
2373 return false;
2374 }
2375 _unstable_if_traps.delete_at(i);
2376 break;
2377 }
2378 }
2379 return true;
2380 }
2381
2382 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2383 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2384 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2385 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2386 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2387 CallStaticJavaNode* unc = trap->uncommon_trap();
2388 int next_bci = trap->next_bci();
2389 bool modified = trap->modified();
2390
2391 if (next_bci != -1 && !modified) {
2392 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2393 JVMState* jvms = unc->jvms();
2394 ciMethod* method = jvms->method();
2395 ciBytecodeStream iter(method);
2396
2397 iter.force_bci(jvms->bci());
2398 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2399 Bytecodes::Code c = iter.cur_bc();
2400 Node* lhs = nullptr;
2401 Node* rhs = nullptr;
2402 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2403 lhs = unc->peek_operand(0);
2404 rhs = unc->peek_operand(1);
2405 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2406 lhs = unc->peek_operand(0);
2407 }
2408
2409 ResourceMark rm;
2410 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2411 assert(live_locals.is_valid(), "broken liveness info");
2412 int len = (int)live_locals.size();
2413
2414 for (int i = 0; i < len; i++) {
2415 Node* local = unc->local(jvms, i);
2416 // kill local using the liveness of next_bci.
2417 // give up when the local looks like an operand to secure reexecution.
2418 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2419 uint idx = jvms->locoff() + i;
2420 #ifdef ASSERT
2421 if (PrintOpto && Verbose) {
2422 tty->print("[unstable_if] kill local#%d: ", idx);
2423 local->dump();
2424 tty->cr();
2425 }
2426 #endif
2427 igvn.replace_input_of(unc, idx, top());
2428 modified = true;
2429 }
2430 }
2431 }
2432
2433 // keep the modified trap for late query
2434 if (modified) {
2435 trap->set_modified();
2436 } else {
2437 _unstable_if_traps.delete_at(i);
2438 }
2439 }
2440 igvn.optimize();
2441 }
2442
2443 // StringOpts and late inlining of string methods
2444 void Compile::inline_string_calls(bool parse_time) {
2445 {
2446 // remove useless nodes to make the usage analysis simpler
2447 ResourceMark rm;
2448 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2449 }
2450
2451 {
2452 ResourceMark rm;
2453 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2454 PhaseStringOpts pso(initial_gvn());
2455 print_method(PHASE_AFTER_STRINGOPTS, 3);
2456 }
2457
2458 // now inline anything that we skipped the first time around
2459 if (!parse_time) {
2460 _late_inlines_pos = _late_inlines.length();
2461 }
2462
2463 while (_string_late_inlines.length() > 0) {
2464 CallGenerator* cg = _string_late_inlines.pop();
2465 cg->do_late_inline();
2466 if (failing()) return;
2467 }
2468 _string_late_inlines.trunc_to(0);
2469 }
2470
2471 // Late inlining of boxing methods
2472 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2473 if (_boxing_late_inlines.length() > 0) {
2474 assert(has_boxed_value(), "inconsistent");
2475
2476 set_inlining_incrementally(true);
2477
2478 igvn_worklist()->ensure_empty(); // should be done with igvn
2479
2480 _late_inlines_pos = _late_inlines.length();
2481
2482 while (_boxing_late_inlines.length() > 0) {
2483 CallGenerator* cg = _boxing_late_inlines.pop();
2484 cg->do_late_inline();
2485 if (failing()) return;
2486 }
2487 _boxing_late_inlines.trunc_to(0);
2488
2489 inline_incrementally_cleanup(igvn);
2490
2491 set_inlining_incrementally(false);
2492 }
2493 }
2494
2495 bool Compile::inline_incrementally_one() {
2496 assert(IncrementalInline, "incremental inlining should be on");
2497
2498 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2499
2500 set_inlining_progress(false);
2501 set_do_cleanup(false);
2502
2503 for (int i = 0; i < _late_inlines.length(); i++) {
2504 _late_inlines_pos = i+1;
2505 CallGenerator* cg = _late_inlines.at(i);
2506 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2507 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2508 cg->do_late_inline();
2509 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2510 if (failing()) {
2511 return false;
2512 } else if (inlining_progress()) {
2513 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2514 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2515 break; // process one call site at a time
2516 }
2517 } else {
2518 // Ignore late inline direct calls when inlining is not allowed.
2519 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2520 }
2521 }
2522 // Remove processed elements.
2523 _late_inlines.remove_till(_late_inlines_pos);
2524 _late_inlines_pos = 0;
2525
2526 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2527
2528 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2529
2530 set_inlining_progress(false);
2531 set_do_cleanup(false);
2532
2533 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2534 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2535 }
2536
2537 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2538 {
2539 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2540 ResourceMark rm;
2541 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2542 }
2543 {
2544 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2545 igvn.reset_from_gvn(initial_gvn());
2546 igvn.optimize();
2547 if (failing()) return;
2548 }
2549 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2550 }
2551
2552 // Perform incremental inlining until bound on number of live nodes is reached
2553 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2554 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2555
2556 set_inlining_incrementally(true);
2557 uint low_live_nodes = 0;
2558
2559 while (_late_inlines.length() > 0) {
2560 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2561 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2562 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2563 // PhaseIdealLoop is expensive so we only try it once we are
2564 // out of live nodes and we only try it again if the previous
2565 // helped got the number of nodes down significantly
2566 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2567 if (failing()) return;
2568 low_live_nodes = live_nodes();
2569 _major_progress = true;
2570 }
2571
2572 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2573 bool do_print_inlining = print_inlining() || print_intrinsics();
2574 if (do_print_inlining || log() != nullptr) {
2575 // Print inlining message for candidates that we couldn't inline for lack of space.
2576 for (int i = 0; i < _late_inlines.length(); i++) {
2577 CallGenerator* cg = _late_inlines.at(i);
2578 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2579 if (do_print_inlining) {
2580 cg->print_inlining_late(InliningResult::FAILURE, msg);
2581 }
2582 log_late_inline_failure(cg, msg);
2583 }
2584 }
2585 break; // finish
2586 }
2587 }
2588
2589 igvn_worklist()->ensure_empty(); // should be done with igvn
2590
2591 while (inline_incrementally_one()) {
2592 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2593 }
2594 if (failing()) return;
2595
2596 inline_incrementally_cleanup(igvn);
2597
2598 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2599
2600 if (failing()) return;
2601
2602 if (_late_inlines.length() == 0) {
2603 break; // no more progress
2604 }
2605 }
2606
2607 igvn_worklist()->ensure_empty(); // should be done with igvn
2608
2609 if (_string_late_inlines.length() > 0) {
2610 assert(has_stringbuilder(), "inconsistent");
2611
2612 inline_string_calls(false);
2613
2614 if (failing()) return;
2615
2616 inline_incrementally_cleanup(igvn);
2617 }
2618
2619 set_inlining_incrementally(false);
2620 }
2621
2622 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2623 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2624 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2625 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2626 // as if "inlining_incrementally() == true" were set.
2627 assert(inlining_incrementally() == false, "not allowed");
2628 #ifdef ASSERT
2629 Unique_Node_List* modified_nodes = _modified_nodes;
2630 _modified_nodes = nullptr;
2631 #endif
2632 assert(_late_inlines.length() > 0, "sanity");
2633
2634 while (_late_inlines.length() > 0) {
2635 igvn_worklist()->ensure_empty(); // should be done with igvn
2636
2637 while (inline_incrementally_one()) {
2638 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2639 }
2640 if (failing()) return;
2641
2642 inline_incrementally_cleanup(igvn);
2643 }
2644 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2645 }
2646
2647 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2648 if (_loop_opts_cnt > 0) {
2649 while (major_progress() && (_loop_opts_cnt > 0)) {
2650 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2651 PhaseIdealLoop::optimize(igvn, mode);
2652 _loop_opts_cnt--;
2653 if (failing()) return false;
2654 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2655 }
2656 }
2657 return true;
2658 }
2659
2660 // Remove edges from "root" to each SafePoint at a backward branch.
2661 // They were inserted during parsing (see add_safepoint()) to make
2662 // infinite loops without calls or exceptions visible to root, i.e.,
2663 // useful.
2664 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2665 Node *r = root();
2666 if (r != nullptr) {
2667 for (uint i = r->req(); i < r->len(); ++i) {
2668 Node *n = r->in(i);
2669 if (n != nullptr && n->is_SafePoint()) {
2670 r->rm_prec(i);
2671 if (n->outcnt() == 0) {
2672 igvn.remove_dead_node(n);
2673 }
2674 --i;
2675 }
2676 }
2677 // Parsing may have added top inputs to the root node (Path
2678 // leading to the Halt node proven dead). Make sure we get a
2679 // chance to clean them up.
2680 igvn._worklist.push(r);
2681 igvn.optimize();
2682 }
2683 }
2684
2685 //------------------------------Optimize---------------------------------------
2686 // Given a graph, optimize it.
2687 void Compile::Optimize() {
2688 TracePhase tp("optimizer", &timers[_t_optimizer]);
2689
2690 #ifndef PRODUCT
2691 if (env()->break_at_compile()) {
2692 BREAKPOINT;
2693 }
2694
2695 #endif
2696
2697 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2698 #ifdef ASSERT
2699 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2700 #endif
2701
2702 ResourceMark rm;
2703
2704 print_inlining_reinit();
2705
2706 NOT_PRODUCT( verify_graph_edges(); )
2707
2708 print_method(PHASE_AFTER_PARSING, 1);
2709
2710 {
2711 // Iterative Global Value Numbering, including ideal transforms
2712 // Initialize IterGVN with types and values from parse-time GVN
2713 PhaseIterGVN igvn(initial_gvn());
2714 #ifdef ASSERT
2715 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2716 #endif
2717 {
2718 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2719 igvn.optimize();
2720 }
2721
2722 if (failing()) return;
2723
2724 print_method(PHASE_ITER_GVN1, 2);
2725
2726 process_for_unstable_if_traps(igvn);
2727
2728 if (failing()) return;
2729
2730 inline_incrementally(igvn);
2731
2732 print_method(PHASE_INCREMENTAL_INLINE, 2);
2733
2734 if (failing()) return;
2735
2736 if (eliminate_boxing()) {
2737 // Inline valueOf() methods now.
2738 inline_boxing_calls(igvn);
2739
2740 if (failing()) return;
2741
2742 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2743 inline_incrementally(igvn);
2744 }
2745
2746 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2747
2748 if (failing()) return;
2749 }
2750
2751 // Remove the speculative part of types and clean up the graph from
2752 // the extra CastPP nodes whose only purpose is to carry them. Do
2753 // that early so that optimizations are not disrupted by the extra
2754 // CastPP nodes.
2755 remove_speculative_types(igvn);
2756
2757 if (failing()) return;
2758
2759 // No more new expensive nodes will be added to the list from here
2760 // so keep only the actual candidates for optimizations.
2761 cleanup_expensive_nodes(igvn);
2762
2763 if (failing()) return;
2764
2765 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2766 if (EnableVectorSupport && has_vbox_nodes()) {
2767 TracePhase tp("", &timers[_t_vector]);
2768 PhaseVector pv(igvn);
2769 pv.optimize_vector_boxes();
2770 if (failing()) return;
2771 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2772 }
2773 assert(!has_vbox_nodes(), "sanity");
2774
2775 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2776 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2777 igvn_worklist()->ensure_empty(); // should be done with igvn
2778 {
2779 ResourceMark rm;
2780 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2781 }
2782 igvn.reset_from_gvn(initial_gvn());
2783 igvn.optimize();
2784 if (failing()) return;
2785 }
2786
2787 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2788 // safepoints
2789 remove_root_to_sfpts_edges(igvn);
2790
2791 // Process inline type nodes now that all inlining is over
2792 process_inline_types(igvn);
2793
2794 adjust_flat_array_access_aliases(igvn);
2795
2796 if (failing()) return;
2797
2798 // Perform escape analysis
2799 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2800 if (has_loops()) {
2801 // Cleanup graph (remove dead nodes).
2802 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2803 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2804 if (failing()) return;
2805 }
2806 bool progress;
2807 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2808 do {
2809 ConnectionGraph::do_analysis(this, &igvn);
2810
2811 if (failing()) return;
2812
2813 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2814
2815 // Optimize out fields loads from scalar replaceable allocations.
2816 igvn.optimize();
2817 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2818
2819 if (failing()) return;
2820
2821 if (congraph() != nullptr && macro_count() > 0) {
2822 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2823 PhaseMacroExpand mexp(igvn);
2824 mexp.eliminate_macro_nodes();
2825 if (failing()) return;
2826
2827 igvn.set_delay_transform(false);
2828 igvn.optimize();
2829 if (failing()) return;
2830
2831 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2832 }
2833
2834 ConnectionGraph::verify_ram_nodes(this, root());
2835 if (failing()) return;
2836
2837 progress = do_iterative_escape_analysis() &&
2838 (macro_count() < mcount) &&
2839 ConnectionGraph::has_candidates(this);
2840 // Try again if candidates exist and made progress
2841 // by removing some allocations and/or locks.
2842 } while (progress);
2843 }
2844
2845 // Loop transforms on the ideal graph. Range Check Elimination,
2846 // peeling, unrolling, etc.
2847
2848 // Set loop opts counter
2849 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2850 {
2851 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2852 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2853 _loop_opts_cnt--;
2854 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2855 if (failing()) return;
2856 }
2857 // Loop opts pass if partial peeling occurred in previous pass
2858 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2859 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2860 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2861 _loop_opts_cnt--;
2862 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2863 if (failing()) return;
2864 }
2865 // Loop opts pass for loop-unrolling before CCP
2866 if(major_progress() && (_loop_opts_cnt > 0)) {
2867 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2868 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2869 _loop_opts_cnt--;
2870 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2871 }
2872 if (!failing()) {
2873 // Verify that last round of loop opts produced a valid graph
2874 PhaseIdealLoop::verify(igvn);
2875 }
2876 }
2877 if (failing()) return;
2878
2879 // Conditional Constant Propagation;
2880 print_method(PHASE_BEFORE_CCP1, 2);
2881 PhaseCCP ccp( &igvn );
2882 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2883 {
2884 TracePhase tp("ccp", &timers[_t_ccp]);
2885 ccp.do_transform();
2886 }
2887 print_method(PHASE_CCP1, 2);
2888
2889 assert( true, "Break here to ccp.dump_old2new_map()");
2890
2891 // Iterative Global Value Numbering, including ideal transforms
2892 {
2893 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2894 igvn.reset_from_igvn(&ccp);
2895 igvn.optimize();
2896 }
2897 print_method(PHASE_ITER_GVN2, 2);
2898
2899 if (failing()) return;
2900
2901 // Loop transforms on the ideal graph. Range Check Elimination,
2902 // peeling, unrolling, etc.
2903 if (!optimize_loops(igvn, LoopOptsDefault)) {
2904 return;
2905 }
2906
2907 if (failing()) return;
2908
2909 C->clear_major_progress(); // ensure that major progress is now clear
2910
2911 process_for_post_loop_opts_igvn(igvn);
2912
2913 if (failing()) return;
2914
2915 #ifdef ASSERT
2916 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2917 #endif
2918
2919 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2920
2921 if (_late_inlines.length() > 0) {
2922 // More opportunities to optimize virtual and MH calls.
2923 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2924 process_late_inline_calls_no_inline(igvn);
2925 }
2926
2927 {
2928 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2929 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2930 PhaseMacroExpand mex(igvn);
2931 if (mex.expand_macro_nodes()) {
2932 assert(failing(), "must bail out w/ explicit message");
2933 return;
2934 }
2935 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2936 }
2937
2938 // Process inline type nodes again and remove them. From here
2939 // on we don't need to keep track of field values anymore.
2940 process_inline_types(igvn, /* remove= */ true);
2941
2942 {
2943 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2944 if (bs->expand_barriers(this, igvn)) {
2945 assert(failing(), "must bail out w/ explicit message");
2946 return;
2947 }
2948 print_method(PHASE_BARRIER_EXPANSION, 2);
2949 }
2950
2951 if (C->max_vector_size() > 0) {
2952 C->optimize_logic_cones(igvn);
2953 igvn.optimize();
2954 if (failing()) return;
2955 }
2956
2957 DEBUG_ONLY( _modified_nodes = nullptr; )
2958 DEBUG_ONLY( _late_inlines.clear(); )
2959
2960 assert(igvn._worklist.size() == 0, "not empty");
2961 } // (End scope of igvn; run destructor if necessary for asserts.)
2962
2963 check_no_dead_use();
2964
2965 process_print_inlining();
2966
2967 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2968 // to remove hashes to unlock nodes for modifications.
2969 C->node_hash()->clear();
2970
2971 // A method with only infinite loops has no edges entering loops from root
2972 {
2973 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2974 if (final_graph_reshaping()) {
2975 assert(failing(), "must bail out w/ explicit message");
2976 return;
2977 }
2978 }
2979
2980 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2981 DEBUG_ONLY(set_phase_optimize_finished();)
2982 }
2983
2984 #ifdef ASSERT
2985 void Compile::check_no_dead_use() const {
2986 ResourceMark rm;
2987 Unique_Node_List wq;
2988 wq.push(root());
2989 for (uint i = 0; i < wq.size(); ++i) {
2990 Node* n = wq.at(i);
2991 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2992 Node* u = n->fast_out(j);
2993 if (u->outcnt() == 0 && !u->is_Con()) {
2994 u->dump();
2995 fatal("no reachable node should have no use");
2996 }
2997 wq.push(u);
2998 }
2999 }
3000 }
3001 #endif
3002
3003 void Compile::inline_vector_reboxing_calls() {
3004 if (C->_vector_reboxing_late_inlines.length() > 0) {
3005 _late_inlines_pos = C->_late_inlines.length();
3006 while (_vector_reboxing_late_inlines.length() > 0) {
3007 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
3008 cg->do_late_inline();
3009 if (failing()) return;
3010 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
3011 }
3012 _vector_reboxing_late_inlines.trunc_to(0);
3013 }
3014 }
3015
3016 bool Compile::has_vbox_nodes() {
3017 if (C->_vector_reboxing_late_inlines.length() > 0) {
3018 return true;
3019 }
3020 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3021 Node * n = C->macro_node(macro_idx);
3022 assert(n->is_macro(), "only macro nodes expected here");
3023 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3024 return true;
3025 }
3026 }
3027 return false;
3028 }
3029
3030 //---------------------------- Bitwise operation packing optimization ---------------------------
3031
3032 static bool is_vector_unary_bitwise_op(Node* n) {
3033 return n->Opcode() == Op_XorV &&
3034 VectorNode::is_vector_bitwise_not_pattern(n);
3035 }
3036
3037 static bool is_vector_binary_bitwise_op(Node* n) {
3038 switch (n->Opcode()) {
3039 case Op_AndV:
3040 case Op_OrV:
3041 return true;
3042
3043 case Op_XorV:
3044 return !is_vector_unary_bitwise_op(n);
3045
3046 default:
3047 return false;
3048 }
3049 }
3050
3051 static bool is_vector_ternary_bitwise_op(Node* n) {
3052 return n->Opcode() == Op_MacroLogicV;
3053 }
3054
3055 static bool is_vector_bitwise_op(Node* n) {
3056 return is_vector_unary_bitwise_op(n) ||
3057 is_vector_binary_bitwise_op(n) ||
3058 is_vector_ternary_bitwise_op(n);
3059 }
3060
3061 static bool is_vector_bitwise_cone_root(Node* n) {
3062 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
3063 return false;
3064 }
3065 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3066 if (is_vector_bitwise_op(n->fast_out(i))) {
3067 return false;
3068 }
3069 }
3070 return true;
3071 }
3072
3073 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3074 uint cnt = 0;
3075 if (is_vector_bitwise_op(n)) {
3076 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3077 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3078 for (uint i = 1; i < inp_cnt; i++) {
3079 Node* in = n->in(i);
3080 bool skip = VectorNode::is_all_ones_vector(in);
3081 if (!skip && !inputs.member(in)) {
3082 inputs.push(in);
3083 cnt++;
3084 }
3085 }
3086 assert(cnt <= 1, "not unary");
3087 } else {
3088 uint last_req = inp_cnt;
3089 if (is_vector_ternary_bitwise_op(n)) {
3090 last_req = inp_cnt - 1; // skip last input
3091 }
3092 for (uint i = 1; i < last_req; i++) {
3093 Node* def = n->in(i);
3094 if (!inputs.member(def)) {
3095 inputs.push(def);
3096 cnt++;
3097 }
3098 }
3099 }
3100 } else { // not a bitwise operations
3101 if (!inputs.member(n)) {
3102 inputs.push(n);
3103 cnt++;
3104 }
3105 }
3106 return cnt;
3107 }
3108
3109 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3110 Unique_Node_List useful_nodes;
3111 C->identify_useful_nodes(useful_nodes);
3112
3113 for (uint i = 0; i < useful_nodes.size(); i++) {
3114 Node* n = useful_nodes.at(i);
3115 if (is_vector_bitwise_cone_root(n)) {
3116 list.push(n);
3117 }
3118 }
3119 }
3120
3121 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3122 const TypeVect* vt,
3123 Unique_Node_List& partition,
3124 Unique_Node_List& inputs) {
3125 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3126 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3127 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3128
3129 Node* in1 = inputs.at(0);
3130 Node* in2 = inputs.at(1);
3131 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3132
3133 uint func = compute_truth_table(partition, inputs);
3134
3135 Node* pn = partition.at(partition.size() - 1);
3136 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3137 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3138 }
3139
3140 static uint extract_bit(uint func, uint pos) {
3141 return (func & (1 << pos)) >> pos;
3142 }
3143
3144 //
3145 // A macro logic node represents a truth table. It has 4 inputs,
3146 // First three inputs corresponds to 3 columns of a truth table
3147 // and fourth input captures the logic function.
3148 //
3149 // eg. fn = (in1 AND in2) OR in3;
3150 //
3151 // MacroNode(in1,in2,in3,fn)
3152 //
3153 // -----------------
3154 // in1 in2 in3 fn
3155 // -----------------
3156 // 0 0 0 0
3157 // 0 0 1 1
3158 // 0 1 0 0
3159 // 0 1 1 1
3160 // 1 0 0 0
3161 // 1 0 1 1
3162 // 1 1 0 1
3163 // 1 1 1 1
3164 //
3165
3166 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3167 int res = 0;
3168 for (int i = 0; i < 8; i++) {
3169 int bit1 = extract_bit(in1, i);
3170 int bit2 = extract_bit(in2, i);
3171 int bit3 = extract_bit(in3, i);
3172
3173 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3174 int func_bit = extract_bit(func, func_bit_pos);
3175
3176 res |= func_bit << i;
3177 }
3178 return res;
3179 }
3180
3181 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
3182 assert(n != nullptr, "");
3183 assert(eval_map.contains(n), "absent");
3184 return *(eval_map.get(n));
3185 }
3186
3187 static void eval_operands(Node* n,
3188 uint& func1, uint& func2, uint& func3,
3189 ResourceHashtable<Node*,uint>& eval_map) {
3190 assert(is_vector_bitwise_op(n), "");
3191
3192 if (is_vector_unary_bitwise_op(n)) {
3193 Node* opnd = n->in(1);
3194 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3195 opnd = n->in(2);
3196 }
3197 func1 = eval_operand(opnd, eval_map);
3198 } else if (is_vector_binary_bitwise_op(n)) {
3199 func1 = eval_operand(n->in(1), eval_map);
3200 func2 = eval_operand(n->in(2), eval_map);
3201 } else {
3202 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3203 func1 = eval_operand(n->in(1), eval_map);
3204 func2 = eval_operand(n->in(2), eval_map);
3205 func3 = eval_operand(n->in(3), eval_map);
3206 }
3207 }
3208
3209 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3210 assert(inputs.size() <= 3, "sanity");
3211 ResourceMark rm;
3212 uint res = 0;
3213 ResourceHashtable<Node*,uint> eval_map;
3214
3215 // Populate precomputed functions for inputs.
3216 // Each input corresponds to one column of 3 input truth-table.
3217 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3218 0xCC, // (_, b, _) -> b
3219 0xF0 }; // (a, _, _) -> a
3220 for (uint i = 0; i < inputs.size(); i++) {
3221 eval_map.put(inputs.at(i), input_funcs[2-i]);
3222 }
3223
3224 for (uint i = 0; i < partition.size(); i++) {
3225 Node* n = partition.at(i);
3226
3227 uint func1 = 0, func2 = 0, func3 = 0;
3228 eval_operands(n, func1, func2, func3, eval_map);
3229
3230 switch (n->Opcode()) {
3231 case Op_OrV:
3232 assert(func3 == 0, "not binary");
3233 res = func1 | func2;
3234 break;
3235 case Op_AndV:
3236 assert(func3 == 0, "not binary");
3237 res = func1 & func2;
3238 break;
3239 case Op_XorV:
3240 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3241 assert(func2 == 0 && func3 == 0, "not unary");
3242 res = (~func1) & 0xFF;
3243 } else {
3244 assert(func3 == 0, "not binary");
3245 res = func1 ^ func2;
3246 }
3247 break;
3248 case Op_MacroLogicV:
3249 // Ordering of inputs may change during evaluation of sub-tree
3250 // containing MacroLogic node as a child node, thus a re-evaluation
3251 // makes sure that function is evaluated in context of current
3252 // inputs.
3253 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3254 break;
3255
3256 default: assert(false, "not supported: %s", n->Name());
3257 }
3258 assert(res <= 0xFF, "invalid");
3259 eval_map.put(n, res);
3260 }
3261 return res;
3262 }
3263
3264 // Criteria under which nodes gets packed into a macro logic node:-
3265 // 1) Parent and both child nodes are all unmasked or masked with
3266 // same predicates.
3267 // 2) Masked parent can be packed with left child if it is predicated
3268 // and both have same predicates.
3269 // 3) Masked parent can be packed with right child if its un-predicated
3270 // or has matching predication condition.
3271 // 4) An unmasked parent can be packed with an unmasked child.
3272 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3273 assert(partition.size() == 0, "not empty");
3274 assert(inputs.size() == 0, "not empty");
3275 if (is_vector_ternary_bitwise_op(n)) {
3276 return false;
3277 }
3278
3279 bool is_unary_op = is_vector_unary_bitwise_op(n);
3280 if (is_unary_op) {
3281 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3282 return false; // too few inputs
3283 }
3284
3285 bool pack_left_child = true;
3286 bool pack_right_child = true;
3287
3288 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3289 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3290
3291 int left_child_input_cnt = 0;
3292 int right_child_input_cnt = 0;
3293
3294 bool parent_is_predicated = n->is_predicated_vector();
3295 bool left_child_predicated = n->in(1)->is_predicated_vector();
3296 bool right_child_predicated = n->in(2)->is_predicated_vector();
3297
3298 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
3299 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3300 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3301
3302 do {
3303 if (pack_left_child && left_child_LOP &&
3304 ((!parent_is_predicated && !left_child_predicated) ||
3305 ((parent_is_predicated && left_child_predicated &&
3306 parent_pred == left_child_pred)))) {
3307 partition.push(n->in(1));
3308 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3309 } else {
3310 inputs.push(n->in(1));
3311 left_child_input_cnt = 1;
3312 }
3313
3314 if (pack_right_child && right_child_LOP &&
3315 (!right_child_predicated ||
3316 (right_child_predicated && parent_is_predicated &&
3317 parent_pred == right_child_pred))) {
3318 partition.push(n->in(2));
3319 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3320 } else {
3321 inputs.push(n->in(2));
3322 right_child_input_cnt = 1;
3323 }
3324
3325 if (inputs.size() > 3) {
3326 assert(partition.size() > 0, "");
3327 inputs.clear();
3328 partition.clear();
3329 if (left_child_input_cnt > right_child_input_cnt) {
3330 pack_left_child = false;
3331 } else {
3332 pack_right_child = false;
3333 }
3334 } else {
3335 break;
3336 }
3337 } while(true);
3338
3339 if(partition.size()) {
3340 partition.push(n);
3341 }
3342
3343 return (partition.size() == 2 || partition.size() == 3) &&
3344 (inputs.size() == 2 || inputs.size() == 3);
3345 }
3346
3347 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3348 assert(is_vector_bitwise_op(n), "not a root");
3349
3350 visited.set(n->_idx);
3351
3352 // 1) Do a DFS walk over the logic cone.
3353 for (uint i = 1; i < n->req(); i++) {
3354 Node* in = n->in(i);
3355 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3356 process_logic_cone_root(igvn, in, visited);
3357 }
3358 }
3359
3360 // 2) Bottom up traversal: Merge node[s] with
3361 // the parent to form macro logic node.
3362 Unique_Node_List partition;
3363 Unique_Node_List inputs;
3364 if (compute_logic_cone(n, partition, inputs)) {
3365 const TypeVect* vt = n->bottom_type()->is_vect();
3366 Node* pn = partition.at(partition.size() - 1);
3367 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3368 if (mask == nullptr ||
3369 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3370 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3371 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3372 igvn.replace_node(n, macro_logic);
3373 }
3374 }
3375 }
3376
3377 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3378 ResourceMark rm;
3379 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3380 Unique_Node_List list;
3381 collect_logic_cone_roots(list);
3382
3383 while (list.size() > 0) {
3384 Node* n = list.pop();
3385 const TypeVect* vt = n->bottom_type()->is_vect();
3386 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3387 if (supported) {
3388 VectorSet visited(comp_arena());
3389 process_logic_cone_root(igvn, n, visited);
3390 }
3391 }
3392 }
3393 }
3394
3395 //------------------------------Code_Gen---------------------------------------
3396 // Given a graph, generate code for it
3397 void Compile::Code_Gen() {
3398 if (failing()) {
3399 return;
3400 }
3401
3402 // Perform instruction selection. You might think we could reclaim Matcher
3403 // memory PDQ, but actually the Matcher is used in generating spill code.
3404 // Internals of the Matcher (including some VectorSets) must remain live
3405 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3406 // set a bit in reclaimed memory.
3407
3408 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3409 // nodes. Mapping is only valid at the root of each matched subtree.
3410 NOT_PRODUCT( verify_graph_edges(); )
3411
3412 Matcher matcher;
3413 _matcher = &matcher;
3414 {
3415 TracePhase tp("matcher", &timers[_t_matcher]);
3416 matcher.match();
3417 if (failing()) {
3418 return;
3419 }
3420 }
3421 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3422 // nodes. Mapping is only valid at the root of each matched subtree.
3423 NOT_PRODUCT( verify_graph_edges(); )
3424
3425 // If you have too many nodes, or if matching has failed, bail out
3426 check_node_count(0, "out of nodes matching instructions");
3427 if (failing()) {
3428 return;
3429 }
3430
3431 print_method(PHASE_MATCHING, 2);
3432
3433 // Build a proper-looking CFG
3434 PhaseCFG cfg(node_arena(), root(), matcher);
3435 if (failing()) {
3436 return;
3437 }
3438 _cfg = &cfg;
3439 {
3440 TracePhase tp("scheduler", &timers[_t_scheduler]);
3441 bool success = cfg.do_global_code_motion();
3442 if (!success) {
3443 return;
3444 }
3445
3446 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3447 NOT_PRODUCT( verify_graph_edges(); )
3448 cfg.verify();
3449 }
3450
3451 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3452 _regalloc = ®alloc;
3453 {
3454 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
3455 // Perform register allocation. After Chaitin, use-def chains are
3456 // no longer accurate (at spill code) and so must be ignored.
3457 // Node->LRG->reg mappings are still accurate.
3458 _regalloc->Register_Allocate();
3459
3460 // Bail out if the allocator builds too many nodes
3461 if (failing()) {
3462 return;
3463 }
3464
3465 print_method(PHASE_REGISTER_ALLOCATION, 2);
3466 }
3467
3468 // Prior to register allocation we kept empty basic blocks in case the
3469 // the allocator needed a place to spill. After register allocation we
3470 // are not adding any new instructions. If any basic block is empty, we
3471 // can now safely remove it.
3472 {
3473 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
3474 cfg.remove_empty_blocks();
3475 if (do_freq_based_layout()) {
3476 PhaseBlockLayout layout(cfg);
3477 } else {
3478 cfg.set_loop_alignment();
3479 }
3480 cfg.fixup_flow();
3481 cfg.remove_unreachable_blocks();
3482 cfg.verify_dominator_tree();
3483 print_method(PHASE_BLOCK_ORDERING, 3);
3484 }
3485
3486 // Apply peephole optimizations
3487 if( OptoPeephole ) {
3488 TracePhase tp("peephole", &timers[_t_peephole]);
3489 PhasePeephole peep( _regalloc, cfg);
3490 peep.do_transform();
3491 print_method(PHASE_PEEPHOLE, 3);
3492 }
3493
3494 // Do late expand if CPU requires this.
3495 if (Matcher::require_postalloc_expand) {
3496 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3497 cfg.postalloc_expand(_regalloc);
3498 print_method(PHASE_POSTALLOC_EXPAND, 3);
3499 }
3500
3501 // Convert Nodes to instruction bits in a buffer
3502 {
3503 TracePhase tp("output", &timers[_t_output]);
3504 PhaseOutput output;
3505 output.Output();
3506 if (failing()) return;
3507 output.install();
3508 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3509 }
3510
3511 // He's dead, Jim.
3512 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3513 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3514 }
3515
3516 //------------------------------Final_Reshape_Counts---------------------------
3517 // This class defines counters to help identify when a method
3518 // may/must be executed using hardware with only 24-bit precision.
3519 struct Final_Reshape_Counts : public StackObj {
3520 int _call_count; // count non-inlined 'common' calls
3521 int _float_count; // count float ops requiring 24-bit precision
3522 int _double_count; // count double ops requiring more precision
3523 int _java_call_count; // count non-inlined 'java' calls
3524 int _inner_loop_count; // count loops which need alignment
3525 VectorSet _visited; // Visitation flags
3526 Node_List _tests; // Set of IfNodes & PCTableNodes
3527
3528 Final_Reshape_Counts() :
3529 _call_count(0), _float_count(0), _double_count(0),
3530 _java_call_count(0), _inner_loop_count(0) { }
3531
3532 void inc_call_count () { _call_count ++; }
3533 void inc_float_count () { _float_count ++; }
3534 void inc_double_count() { _double_count++; }
3535 void inc_java_call_count() { _java_call_count++; }
3536 void inc_inner_loop_count() { _inner_loop_count++; }
3537
3538 int get_call_count () const { return _call_count ; }
3539 int get_float_count () const { return _float_count ; }
3540 int get_double_count() const { return _double_count; }
3541 int get_java_call_count() const { return _java_call_count; }
3542 int get_inner_loop_count() const { return _inner_loop_count; }
3543 };
3544
3545 //------------------------------final_graph_reshaping_impl----------------------
3546 // Implement items 1-5 from final_graph_reshaping below.
3547 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3548
3549 if ( n->outcnt() == 0 ) return; // dead node
3550 uint nop = n->Opcode();
3551
3552 // Check for 2-input instruction with "last use" on right input.
3553 // Swap to left input. Implements item (2).
3554 if( n->req() == 3 && // two-input instruction
3555 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3556 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3557 n->in(2)->outcnt() == 1 &&// right use IS a last use
3558 !n->in(2)->is_Con() ) { // right use is not a constant
3559 // Check for commutative opcode
3560 switch( nop ) {
3561 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3562 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3563 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3564 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3565 case Op_AndL: case Op_XorL: case Op_OrL:
3566 case Op_AndI: case Op_XorI: case Op_OrI: {
3567 // Move "last use" input to left by swapping inputs
3568 n->swap_edges(1, 2);
3569 break;
3570 }
3571 default:
3572 break;
3573 }
3574 }
3575
3576 #ifdef ASSERT
3577 if( n->is_Mem() ) {
3578 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3579 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3580 // oop will be recorded in oop map if load crosses safepoint
3581 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3582 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3583 "raw memory operations should have control edge");
3584 }
3585 if (n->is_MemBar()) {
3586 MemBarNode* mb = n->as_MemBar();
3587 if (mb->trailing_store() || mb->trailing_load_store()) {
3588 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3589 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3590 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3591 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3592 } else if (mb->leading()) {
3593 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3594 }
3595 }
3596 #endif
3597 // Count FPU ops and common calls, implements item (3)
3598 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3599 if (!gc_handled) {
3600 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3601 }
3602
3603 // Collect CFG split points
3604 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3605 frc._tests.push(n);
3606 }
3607 }
3608
3609 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3610 if (!UseDivMod) {
3611 return;
3612 }
3613
3614 // Check if "a % b" and "a / b" both exist
3615 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3616 if (d == nullptr) {
3617 return;
3618 }
3619
3620 // Replace them with a fused divmod if supported
3621 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3622 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3623 d->subsume_by(divmod->div_proj(), this);
3624 n->subsume_by(divmod->mod_proj(), this);
3625 } else {
3626 // Replace "a % b" with "a - ((a / b) * b)"
3627 Node* mult = MulNode::make(d, d->in(2), bt);
3628 Node* sub = SubNode::make(d->in(1), mult, bt);
3629 n->subsume_by(sub, this);
3630 }
3631 }
3632
3633 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3634 switch( nop ) {
3635 // Count all float operations that may use FPU
3636 case Op_AddF:
3637 case Op_SubF:
3638 case Op_MulF:
3639 case Op_DivF:
3640 case Op_NegF:
3641 case Op_ModF:
3642 case Op_ConvI2F:
3643 case Op_ConF:
3644 case Op_CmpF:
3645 case Op_CmpF3:
3646 case Op_StoreF:
3647 case Op_LoadF:
3648 // case Op_ConvL2F: // longs are split into 32-bit halves
3649 frc.inc_float_count();
3650 break;
3651
3652 case Op_ConvF2D:
3653 case Op_ConvD2F:
3654 frc.inc_float_count();
3655 frc.inc_double_count();
3656 break;
3657
3658 // Count all double operations that may use FPU
3659 case Op_AddD:
3660 case Op_SubD:
3661 case Op_MulD:
3662 case Op_DivD:
3663 case Op_NegD:
3664 case Op_ModD:
3665 case Op_ConvI2D:
3666 case Op_ConvD2I:
3667 // case Op_ConvL2D: // handled by leaf call
3668 // case Op_ConvD2L: // handled by leaf call
3669 case Op_ConD:
3670 case Op_CmpD:
3671 case Op_CmpD3:
3672 case Op_StoreD:
3673 case Op_LoadD:
3674 case Op_LoadD_unaligned:
3675 frc.inc_double_count();
3676 break;
3677 case Op_Opaque1: // Remove Opaque Nodes before matching
3678 n->subsume_by(n->in(1), this);
3679 break;
3680 case Op_CallStaticJava:
3681 case Op_CallJava:
3682 case Op_CallDynamicJava:
3683 frc.inc_java_call_count(); // Count java call site;
3684 case Op_CallRuntime:
3685 case Op_CallLeaf:
3686 case Op_CallLeafVector:
3687 case Op_CallLeafNoFP: {
3688 assert (n->is_Call(), "");
3689 CallNode *call = n->as_Call();
3690 // Count call sites where the FP mode bit would have to be flipped.
3691 // Do not count uncommon runtime calls:
3692 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3693 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3694 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3695 frc.inc_call_count(); // Count the call site
3696 } else { // See if uncommon argument is shared
3697 Node *n = call->in(TypeFunc::Parms);
3698 int nop = n->Opcode();
3699 // Clone shared simple arguments to uncommon calls, item (1).
3700 if (n->outcnt() > 1 &&
3701 !n->is_Proj() &&
3702 nop != Op_CreateEx &&
3703 nop != Op_CheckCastPP &&
3704 nop != Op_DecodeN &&
3705 nop != Op_DecodeNKlass &&
3706 !n->is_Mem() &&
3707 !n->is_Phi()) {
3708 Node *x = n->clone();
3709 call->set_req(TypeFunc::Parms, x);
3710 }
3711 }
3712 break;
3713 }
3714 case Op_StoreB:
3715 case Op_StoreC:
3716 case Op_StoreI:
3717 case Op_StoreL:
3718 case Op_CompareAndSwapB:
3719 case Op_CompareAndSwapS:
3720 case Op_CompareAndSwapI:
3721 case Op_CompareAndSwapL:
3722 case Op_CompareAndSwapP:
3723 case Op_CompareAndSwapN:
3724 case Op_WeakCompareAndSwapB:
3725 case Op_WeakCompareAndSwapS:
3726 case Op_WeakCompareAndSwapI:
3727 case Op_WeakCompareAndSwapL:
3728 case Op_WeakCompareAndSwapP:
3729 case Op_WeakCompareAndSwapN:
3730 case Op_CompareAndExchangeB:
3731 case Op_CompareAndExchangeS:
3732 case Op_CompareAndExchangeI:
3733 case Op_CompareAndExchangeL:
3734 case Op_CompareAndExchangeP:
3735 case Op_CompareAndExchangeN:
3736 case Op_GetAndAddS:
3737 case Op_GetAndAddB:
3738 case Op_GetAndAddI:
3739 case Op_GetAndAddL:
3740 case Op_GetAndSetS:
3741 case Op_GetAndSetB:
3742 case Op_GetAndSetI:
3743 case Op_GetAndSetL:
3744 case Op_GetAndSetP:
3745 case Op_GetAndSetN:
3746 case Op_StoreP:
3747 case Op_StoreN:
3748 case Op_StoreNKlass:
3749 case Op_LoadB:
3750 case Op_LoadUB:
3751 case Op_LoadUS:
3752 case Op_LoadI:
3753 case Op_LoadKlass:
3754 case Op_LoadNKlass:
3755 case Op_LoadL:
3756 case Op_LoadL_unaligned:
3757 case Op_LoadP:
3758 case Op_LoadN:
3759 case Op_LoadRange:
3760 case Op_LoadS:
3761 break;
3762
3763 case Op_AddP: { // Assert sane base pointers
3764 Node *addp = n->in(AddPNode::Address);
3765 assert( !addp->is_AddP() ||
3766 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3767 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3768 "Base pointers must match (addp %u)", addp->_idx );
3769 #ifdef _LP64
3770 if ((UseCompressedOops || UseCompressedClassPointers) &&
3771 addp->Opcode() == Op_ConP &&
3772 addp == n->in(AddPNode::Base) &&
3773 n->in(AddPNode::Offset)->is_Con()) {
3774 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3775 // on the platform and on the compressed oops mode.
3776 // Use addressing with narrow klass to load with offset on x86.
3777 // Some platforms can use the constant pool to load ConP.
3778 // Do this transformation here since IGVN will convert ConN back to ConP.
3779 const Type* t = addp->bottom_type();
3780 bool is_oop = t->isa_oopptr() != nullptr;
3781 bool is_klass = t->isa_klassptr() != nullptr;
3782
3783 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3784 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3785 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3786 Node* nn = nullptr;
3787
3788 int op = is_oop ? Op_ConN : Op_ConNKlass;
3789
3790 // Look for existing ConN node of the same exact type.
3791 Node* r = root();
3792 uint cnt = r->outcnt();
3793 for (uint i = 0; i < cnt; i++) {
3794 Node* m = r->raw_out(i);
3795 if (m!= nullptr && m->Opcode() == op &&
3796 m->bottom_type()->make_ptr() == t) {
3797 nn = m;
3798 break;
3799 }
3800 }
3801 if (nn != nullptr) {
3802 // Decode a narrow oop to match address
3803 // [R12 + narrow_oop_reg<<3 + offset]
3804 if (is_oop) {
3805 nn = new DecodeNNode(nn, t);
3806 } else {
3807 nn = new DecodeNKlassNode(nn, t);
3808 }
3809 // Check for succeeding AddP which uses the same Base.
3810 // Otherwise we will run into the assertion above when visiting that guy.
3811 for (uint i = 0; i < n->outcnt(); ++i) {
3812 Node *out_i = n->raw_out(i);
3813 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3814 out_i->set_req(AddPNode::Base, nn);
3815 #ifdef ASSERT
3816 for (uint j = 0; j < out_i->outcnt(); ++j) {
3817 Node *out_j = out_i->raw_out(j);
3818 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3819 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3820 }
3821 #endif
3822 }
3823 }
3824 n->set_req(AddPNode::Base, nn);
3825 n->set_req(AddPNode::Address, nn);
3826 if (addp->outcnt() == 0) {
3827 addp->disconnect_inputs(this);
3828 }
3829 }
3830 }
3831 }
3832 #endif
3833 break;
3834 }
3835
3836 case Op_CastPP: {
3837 // Remove CastPP nodes to gain more freedom during scheduling but
3838 // keep the dependency they encode as control or precedence edges
3839 // (if control is set already) on memory operations. Some CastPP
3840 // nodes don't have a control (don't carry a dependency): skip
3841 // those.
3842 if (n->in(0) != nullptr) {
3843 ResourceMark rm;
3844 Unique_Node_List wq;
3845 wq.push(n);
3846 for (uint next = 0; next < wq.size(); ++next) {
3847 Node *m = wq.at(next);
3848 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3849 Node* use = m->fast_out(i);
3850 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3851 use->ensure_control_or_add_prec(n->in(0));
3852 } else {
3853 switch(use->Opcode()) {
3854 case Op_AddP:
3855 case Op_DecodeN:
3856 case Op_DecodeNKlass:
3857 case Op_CheckCastPP:
3858 case Op_CastPP:
3859 wq.push(use);
3860 break;
3861 }
3862 }
3863 }
3864 }
3865 }
3866 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3867 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3868 Node* in1 = n->in(1);
3869 const Type* t = n->bottom_type();
3870 Node* new_in1 = in1->clone();
3871 new_in1->as_DecodeN()->set_type(t);
3872
3873 if (!Matcher::narrow_oop_use_complex_address()) {
3874 //
3875 // x86, ARM and friends can handle 2 adds in addressing mode
3876 // and Matcher can fold a DecodeN node into address by using
3877 // a narrow oop directly and do implicit null check in address:
3878 //
3879 // [R12 + narrow_oop_reg<<3 + offset]
3880 // NullCheck narrow_oop_reg
3881 //
3882 // On other platforms (Sparc) we have to keep new DecodeN node and
3883 // use it to do implicit null check in address:
3884 //
3885 // decode_not_null narrow_oop_reg, base_reg
3886 // [base_reg + offset]
3887 // NullCheck base_reg
3888 //
3889 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3890 // to keep the information to which null check the new DecodeN node
3891 // corresponds to use it as value in implicit_null_check().
3892 //
3893 new_in1->set_req(0, n->in(0));
3894 }
3895
3896 n->subsume_by(new_in1, this);
3897 if (in1->outcnt() == 0) {
3898 in1->disconnect_inputs(this);
3899 }
3900 } else {
3901 n->subsume_by(n->in(1), this);
3902 if (n->outcnt() == 0) {
3903 n->disconnect_inputs(this);
3904 }
3905 }
3906 break;
3907 }
3908 #ifdef _LP64
3909 case Op_CmpP:
3910 // Do this transformation here to preserve CmpPNode::sub() and
3911 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3912 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3913 Node* in1 = n->in(1);
3914 Node* in2 = n->in(2);
3915 if (!in1->is_DecodeNarrowPtr()) {
3916 in2 = in1;
3917 in1 = n->in(2);
3918 }
3919 assert(in1->is_DecodeNarrowPtr(), "sanity");
3920
3921 Node* new_in2 = nullptr;
3922 if (in2->is_DecodeNarrowPtr()) {
3923 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3924 new_in2 = in2->in(1);
3925 } else if (in2->Opcode() == Op_ConP) {
3926 const Type* t = in2->bottom_type();
3927 if (t == TypePtr::NULL_PTR) {
3928 assert(in1->is_DecodeN(), "compare klass to null?");
3929 // Don't convert CmpP null check into CmpN if compressed
3930 // oops implicit null check is not generated.
3931 // This will allow to generate normal oop implicit null check.
3932 if (Matcher::gen_narrow_oop_implicit_null_checks())
3933 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3934 //
3935 // This transformation together with CastPP transformation above
3936 // will generated code for implicit null checks for compressed oops.
3937 //
3938 // The original code after Optimize()
3939 //
3940 // LoadN memory, narrow_oop_reg
3941 // decode narrow_oop_reg, base_reg
3942 // CmpP base_reg, nullptr
3943 // CastPP base_reg // NotNull
3944 // Load [base_reg + offset], val_reg
3945 //
3946 // after these transformations will be
3947 //
3948 // LoadN memory, narrow_oop_reg
3949 // CmpN narrow_oop_reg, nullptr
3950 // decode_not_null narrow_oop_reg, base_reg
3951 // Load [base_reg + offset], val_reg
3952 //
3953 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3954 // since narrow oops can be used in debug info now (see the code in
3955 // final_graph_reshaping_walk()).
3956 //
3957 // At the end the code will be matched to
3958 // on x86:
3959 //
3960 // Load_narrow_oop memory, narrow_oop_reg
3961 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3962 // NullCheck narrow_oop_reg
3963 //
3964 // and on sparc:
3965 //
3966 // Load_narrow_oop memory, narrow_oop_reg
3967 // decode_not_null narrow_oop_reg, base_reg
3968 // Load [base_reg + offset], val_reg
3969 // NullCheck base_reg
3970 //
3971 } else if (t->isa_oopptr()) {
3972 new_in2 = ConNode::make(t->make_narrowoop());
3973 } else if (t->isa_klassptr()) {
3974 new_in2 = ConNode::make(t->make_narrowklass());
3975 }
3976 }
3977 if (new_in2 != nullptr) {
3978 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3979 n->subsume_by(cmpN, this);
3980 if (in1->outcnt() == 0) {
3981 in1->disconnect_inputs(this);
3982 }
3983 if (in2->outcnt() == 0) {
3984 in2->disconnect_inputs(this);
3985 }
3986 }
3987 }
3988 break;
3989
3990 case Op_DecodeN:
3991 case Op_DecodeNKlass:
3992 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3993 // DecodeN could be pinned when it can't be fold into
3994 // an address expression, see the code for Op_CastPP above.
3995 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3996 break;
3997
3998 case Op_EncodeP:
3999 case Op_EncodePKlass: {
4000 Node* in1 = n->in(1);
4001 if (in1->is_DecodeNarrowPtr()) {
4002 n->subsume_by(in1->in(1), this);
4003 } else if (in1->Opcode() == Op_ConP) {
4004 const Type* t = in1->bottom_type();
4005 if (t == TypePtr::NULL_PTR) {
4006 assert(t->isa_oopptr(), "null klass?");
4007 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4008 } else if (t->isa_oopptr()) {
4009 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4010 } else if (t->isa_klassptr()) {
4011 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4012 }
4013 }
4014 if (in1->outcnt() == 0) {
4015 in1->disconnect_inputs(this);
4016 }
4017 break;
4018 }
4019
4020 case Op_Proj: {
4021 if (OptimizeStringConcat || IncrementalInline) {
4022 ProjNode* proj = n->as_Proj();
4023 if (proj->_is_io_use) {
4024 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4025 // Separate projections were used for the exception path which
4026 // are normally removed by a late inline. If it wasn't inlined
4027 // then they will hang around and should just be replaced with
4028 // the original one. Merge them.
4029 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4030 if (non_io_proj != nullptr) {
4031 proj->subsume_by(non_io_proj , this);
4032 }
4033 }
4034 }
4035 break;
4036 }
4037
4038 case Op_Phi:
4039 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4040 // The EncodeP optimization may create Phi with the same edges
4041 // for all paths. It is not handled well by Register Allocator.
4042 Node* unique_in = n->in(1);
4043 assert(unique_in != nullptr, "");
4044 uint cnt = n->req();
4045 for (uint i = 2; i < cnt; i++) {
4046 Node* m = n->in(i);
4047 assert(m != nullptr, "");
4048 if (unique_in != m)
4049 unique_in = nullptr;
4050 }
4051 if (unique_in != nullptr) {
4052 n->subsume_by(unique_in, this);
4053 }
4054 }
4055 break;
4056
4057 #endif
4058
4059 #ifdef ASSERT
4060 case Op_CastII:
4061 // Verify that all range check dependent CastII nodes were removed.
4062 if (n->isa_CastII()->has_range_check()) {
4063 n->dump(3);
4064 assert(false, "Range check dependent CastII node was not removed");
4065 }
4066 break;
4067 #endif
4068
4069 case Op_ModI:
4070 handle_div_mod_op(n, T_INT, false);
4071 break;
4072
4073 case Op_ModL:
4074 handle_div_mod_op(n, T_LONG, false);
4075 break;
4076
4077 case Op_UModI:
4078 handle_div_mod_op(n, T_INT, true);
4079 break;
4080
4081 case Op_UModL:
4082 handle_div_mod_op(n, T_LONG, true);
4083 break;
4084
4085 case Op_LoadVector:
4086 case Op_StoreVector:
4087 #ifdef ASSERT
4088 // Add VerifyVectorAlignment node between adr and load / store.
4089 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
4090 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
4091 n->as_StoreVector()->must_verify_alignment();
4092 if (must_verify_alignment) {
4093 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
4094 n->as_StoreVector()->memory_size();
4095 // The memory access should be aligned to the vector width in bytes.
4096 // However, the underlying array is possibly less well aligned, but at least
4097 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
4098 // a loop we can expect at least the following alignment:
4099 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
4100 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
4101 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
4102 // Create mask from alignment. e.g. 0b1000 -> 0b0111
4103 jlong mask = guaranteed_alignment - 1;
4104 Node* mask_con = ConLNode::make(mask);
4105 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
4106 n->set_req(MemNode::Address, va);
4107 }
4108 }
4109 #endif
4110 break;
4111
4112 case Op_LoadVectorGather:
4113 case Op_StoreVectorScatter:
4114 case Op_LoadVectorGatherMasked:
4115 case Op_StoreVectorScatterMasked:
4116 case Op_VectorCmpMasked:
4117 case Op_VectorMaskGen:
4118 case Op_LoadVectorMasked:
4119 case Op_StoreVectorMasked:
4120 break;
4121
4122 case Op_AddReductionVI:
4123 case Op_AddReductionVL:
4124 case Op_AddReductionVF:
4125 case Op_AddReductionVD:
4126 case Op_MulReductionVI:
4127 case Op_MulReductionVL:
4128 case Op_MulReductionVF:
4129 case Op_MulReductionVD:
4130 case Op_MinReductionV:
4131 case Op_MaxReductionV:
4132 case Op_AndReductionV:
4133 case Op_OrReductionV:
4134 case Op_XorReductionV:
4135 break;
4136
4137 case Op_PackB:
4138 case Op_PackS:
4139 case Op_PackI:
4140 case Op_PackF:
4141 case Op_PackL:
4142 case Op_PackD:
4143 if (n->req()-1 > 2) {
4144 // Replace many operand PackNodes with a binary tree for matching
4145 PackNode* p = (PackNode*) n;
4146 Node* btp = p->binary_tree_pack(1, n->req());
4147 n->subsume_by(btp, this);
4148 }
4149 break;
4150 case Op_Loop:
4151 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4152 case Op_CountedLoop:
4153 case Op_LongCountedLoop:
4154 case Op_OuterStripMinedLoop:
4155 if (n->as_Loop()->is_inner_loop()) {
4156 frc.inc_inner_loop_count();
4157 }
4158 n->as_Loop()->verify_strip_mined(0);
4159 break;
4160 case Op_LShiftI:
4161 case Op_RShiftI:
4162 case Op_URShiftI:
4163 case Op_LShiftL:
4164 case Op_RShiftL:
4165 case Op_URShiftL:
4166 if (Matcher::need_masked_shift_count) {
4167 // The cpu's shift instructions don't restrict the count to the
4168 // lower 5/6 bits. We need to do the masking ourselves.
4169 Node* in2 = n->in(2);
4170 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4171 const TypeInt* t = in2->find_int_type();
4172 if (t != nullptr && t->is_con()) {
4173 juint shift = t->get_con();
4174 if (shift > mask) { // Unsigned cmp
4175 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4176 }
4177 } else {
4178 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
4179 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4180 n->set_req(2, shift);
4181 }
4182 }
4183 if (in2->outcnt() == 0) { // Remove dead node
4184 in2->disconnect_inputs(this);
4185 }
4186 }
4187 break;
4188 case Op_MemBarStoreStore:
4189 case Op_MemBarRelease:
4190 // Break the link with AllocateNode: it is no longer useful and
4191 // confuses register allocation.
4192 if (n->req() > MemBarNode::Precedent) {
4193 n->set_req(MemBarNode::Precedent, top());
4194 }
4195 break;
4196 case Op_MemBarAcquire: {
4197 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4198 // At parse time, the trailing MemBarAcquire for a volatile load
4199 // is created with an edge to the load. After optimizations,
4200 // that input may be a chain of Phis. If those phis have no
4201 // other use, then the MemBarAcquire keeps them alive and
4202 // register allocation can be confused.
4203 dead_nodes.push(n->in(MemBarNode::Precedent));
4204 n->set_req(MemBarNode::Precedent, top());
4205 }
4206 break;
4207 }
4208 case Op_Blackhole:
4209 break;
4210 case Op_RangeCheck: {
4211 RangeCheckNode* rc = n->as_RangeCheck();
4212 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4213 n->subsume_by(iff, this);
4214 frc._tests.push(iff);
4215 break;
4216 }
4217 case Op_ConvI2L: {
4218 if (!Matcher::convi2l_type_required) {
4219 // Code generation on some platforms doesn't need accurate
4220 // ConvI2L types. Widening the type can help remove redundant
4221 // address computations.
4222 n->as_Type()->set_type(TypeLong::INT);
4223 ResourceMark rm;
4224 Unique_Node_List wq;
4225 wq.push(n);
4226 for (uint next = 0; next < wq.size(); next++) {
4227 Node *m = wq.at(next);
4228
4229 for(;;) {
4230 // Loop over all nodes with identical inputs edges as m
4231 Node* k = m->find_similar(m->Opcode());
4232 if (k == nullptr) {
4233 break;
4234 }
4235 // Push their uses so we get a chance to remove node made
4236 // redundant
4237 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4238 Node* u = k->fast_out(i);
4239 if (u->Opcode() == Op_LShiftL ||
4240 u->Opcode() == Op_AddL ||
4241 u->Opcode() == Op_SubL ||
4242 u->Opcode() == Op_AddP) {
4243 wq.push(u);
4244 }
4245 }
4246 // Replace all nodes with identical edges as m with m
4247 k->subsume_by(m, this);
4248 }
4249 }
4250 }
4251 break;
4252 }
4253 case Op_CmpUL: {
4254 if (!Matcher::has_match_rule(Op_CmpUL)) {
4255 // No support for unsigned long comparisons
4256 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4257 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4258 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4259 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4260 Node* andl = new AndLNode(orl, remove_sign_mask);
4261 Node* cmp = new CmpLNode(andl, n->in(2));
4262 n->subsume_by(cmp, this);
4263 }
4264 break;
4265 }
4266 #ifdef ASSERT
4267 case Op_InlineType: {
4268 n->dump(-1);
4269 assert(false, "inline type node was not removed");
4270 break;
4271 }
4272 #endif
4273 default:
4274 assert(!n->is_Call(), "");
4275 assert(!n->is_Mem(), "");
4276 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4277 break;
4278 }
4279 }
4280
4281 //------------------------------final_graph_reshaping_walk---------------------
4282 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4283 // requires that the walk visits a node's inputs before visiting the node.
4284 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4285 Unique_Node_List sfpt;
4286
4287 frc._visited.set(root->_idx); // first, mark node as visited
4288 uint cnt = root->req();
4289 Node *n = root;
4290 uint i = 0;
4291 while (true) {
4292 if (i < cnt) {
4293 // Place all non-visited non-null inputs onto stack
4294 Node* m = n->in(i);
4295 ++i;
4296 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
4297 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
4298 // compute worst case interpreter size in case of a deoptimization
4299 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4300
4301 sfpt.push(m);
4302 }
4303 cnt = m->req();
4304 nstack.push(n, i); // put on stack parent and next input's index
4305 n = m;
4306 i = 0;
4307 }
4308 } else {
4309 // Now do post-visit work
4310 final_graph_reshaping_impl(n, frc, dead_nodes);
4311 if (nstack.is_empty())
4312 break; // finished
4313 n = nstack.node(); // Get node from stack
4314 cnt = n->req();
4315 i = nstack.index();
4316 nstack.pop(); // Shift to the next node on stack
4317 }
4318 }
4319
4320 // Skip next transformation if compressed oops are not used.
4321 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4322 (!UseCompressedOops && !UseCompressedClassPointers))
4323 return;
4324
4325 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4326 // It could be done for an uncommon traps or any safepoints/calls
4327 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4328 while (sfpt.size() > 0) {
4329 n = sfpt.pop();
4330 JVMState *jvms = n->as_SafePoint()->jvms();
4331 assert(jvms != nullptr, "sanity");
4332 int start = jvms->debug_start();
4333 int end = n->req();
4334 bool is_uncommon = (n->is_CallStaticJava() &&
4335 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4336 for (int j = start; j < end; j++) {
4337 Node* in = n->in(j);
4338 if (in->is_DecodeNarrowPtr()) {
4339 bool safe_to_skip = true;
4340 if (!is_uncommon ) {
4341 // Is it safe to skip?
4342 for (uint i = 0; i < in->outcnt(); i++) {
4343 Node* u = in->raw_out(i);
4344 if (!u->is_SafePoint() ||
4345 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4346 safe_to_skip = false;
4347 }
4348 }
4349 }
4350 if (safe_to_skip) {
4351 n->set_req(j, in->in(1));
4352 }
4353 if (in->outcnt() == 0) {
4354 in->disconnect_inputs(this);
4355 }
4356 }
4357 }
4358 }
4359 }
4360
4361 //------------------------------final_graph_reshaping--------------------------
4362 // Final Graph Reshaping.
4363 //
4364 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4365 // and not commoned up and forced early. Must come after regular
4366 // optimizations to avoid GVN undoing the cloning. Clone constant
4367 // inputs to Loop Phis; these will be split by the allocator anyways.
4368 // Remove Opaque nodes.
4369 // (2) Move last-uses by commutative operations to the left input to encourage
4370 // Intel update-in-place two-address operations and better register usage
4371 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4372 // calls canonicalizing them back.
4373 // (3) Count the number of double-precision FP ops, single-precision FP ops
4374 // and call sites. On Intel, we can get correct rounding either by
4375 // forcing singles to memory (requires extra stores and loads after each
4376 // FP bytecode) or we can set a rounding mode bit (requires setting and
4377 // clearing the mode bit around call sites). The mode bit is only used
4378 // if the relative frequency of single FP ops to calls is low enough.
4379 // This is a key transform for SPEC mpeg_audio.
4380 // (4) Detect infinite loops; blobs of code reachable from above but not
4381 // below. Several of the Code_Gen algorithms fail on such code shapes,
4382 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4383 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4384 // Detection is by looking for IfNodes where only 1 projection is
4385 // reachable from below or CatchNodes missing some targets.
4386 // (5) Assert for insane oop offsets in debug mode.
4387
4388 bool Compile::final_graph_reshaping() {
4389 // an infinite loop may have been eliminated by the optimizer,
4390 // in which case the graph will be empty.
4391 if (root()->req() == 1) {
4392 // Do not compile method that is only a trivial infinite loop,
4393 // since the content of the loop may have been eliminated.
4394 record_method_not_compilable("trivial infinite loop");
4395 return true;
4396 }
4397
4398 // Expensive nodes have their control input set to prevent the GVN
4399 // from freely commoning them. There's no GVN beyond this point so
4400 // no need to keep the control input. We want the expensive nodes to
4401 // be freely moved to the least frequent code path by gcm.
4402 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4403 for (int i = 0; i < expensive_count(); i++) {
4404 _expensive_nodes.at(i)->set_req(0, nullptr);
4405 }
4406
4407 Final_Reshape_Counts frc;
4408
4409 // Visit everybody reachable!
4410 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4411 Node_Stack nstack(live_nodes() >> 1);
4412 Unique_Node_List dead_nodes;
4413 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4414
4415 // Check for unreachable (from below) code (i.e., infinite loops).
4416 for( uint i = 0; i < frc._tests.size(); i++ ) {
4417 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4418 // Get number of CFG targets.
4419 // Note that PCTables include exception targets after calls.
4420 uint required_outcnt = n->required_outcnt();
4421 if (n->outcnt() != required_outcnt) {
4422 // Check for a few special cases. Rethrow Nodes never take the
4423 // 'fall-thru' path, so expected kids is 1 less.
4424 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4425 if (n->in(0)->in(0)->is_Call()) {
4426 CallNode* call = n->in(0)->in(0)->as_Call();
4427 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4428 required_outcnt--; // Rethrow always has 1 less kid
4429 } else if (call->req() > TypeFunc::Parms &&
4430 call->is_CallDynamicJava()) {
4431 // Check for null receiver. In such case, the optimizer has
4432 // detected that the virtual call will always result in a null
4433 // pointer exception. The fall-through projection of this CatchNode
4434 // will not be populated.
4435 Node* arg0 = call->in(TypeFunc::Parms);
4436 if (arg0->is_Type() &&
4437 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4438 required_outcnt--;
4439 }
4440 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4441 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4442 // Check for illegal array length. In such case, the optimizer has
4443 // detected that the allocation attempt will always result in an
4444 // exception. There is no fall-through projection of this CatchNode .
4445 assert(call->is_CallStaticJava(), "static call expected");
4446 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4447 uint valid_length_test_input = call->req() - 1;
4448 Node* valid_length_test = call->in(valid_length_test_input);
4449 call->del_req(valid_length_test_input);
4450 if (valid_length_test->find_int_con(1) == 0) {
4451 required_outcnt--;
4452 }
4453 dead_nodes.push(valid_length_test);
4454 assert(n->outcnt() == required_outcnt, "malformed control flow");
4455 continue;
4456 }
4457 }
4458 }
4459
4460 // Recheck with a better notion of 'required_outcnt'
4461 if (n->outcnt() != required_outcnt) {
4462 record_method_not_compilable("malformed control flow");
4463 return true; // Not all targets reachable!
4464 }
4465 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4466 CallNode* call = n->in(0)->in(0)->as_Call();
4467 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4468 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4469 assert(call->is_CallStaticJava(), "static call expected");
4470 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4471 uint valid_length_test_input = call->req() - 1;
4472 dead_nodes.push(call->in(valid_length_test_input));
4473 call->del_req(valid_length_test_input); // valid length test useless now
4474 }
4475 }
4476 // Check that I actually visited all kids. Unreached kids
4477 // must be infinite loops.
4478 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4479 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4480 record_method_not_compilable("infinite loop");
4481 return true; // Found unvisited kid; must be unreach
4482 }
4483
4484 // Here so verification code in final_graph_reshaping_walk()
4485 // always see an OuterStripMinedLoopEnd
4486 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4487 IfNode* init_iff = n->as_If();
4488 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4489 n->subsume_by(iff, this);
4490 }
4491 }
4492
4493 while (dead_nodes.size() > 0) {
4494 Node* m = dead_nodes.pop();
4495 if (m->outcnt() == 0 && m != top()) {
4496 for (uint j = 0; j < m->req(); j++) {
4497 Node* in = m->in(j);
4498 if (in != nullptr) {
4499 dead_nodes.push(in);
4500 }
4501 }
4502 m->disconnect_inputs(this);
4503 }
4504 }
4505
4506 #ifdef IA32
4507 // If original bytecodes contained a mixture of floats and doubles
4508 // check if the optimizer has made it homogeneous, item (3).
4509 if (UseSSE == 0 &&
4510 frc.get_float_count() > 32 &&
4511 frc.get_double_count() == 0 &&
4512 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4513 set_24_bit_selection_and_mode(false, true);
4514 }
4515 #endif // IA32
4516
4517 set_java_calls(frc.get_java_call_count());
4518 set_inner_loops(frc.get_inner_loop_count());
4519
4520 // No infinite loops, no reason to bail out.
4521 return false;
4522 }
4523
4524 //-----------------------------too_many_traps----------------------------------
4525 // Report if there are too many traps at the current method and bci.
4526 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4527 bool Compile::too_many_traps(ciMethod* method,
4528 int bci,
4529 Deoptimization::DeoptReason reason) {
4530 ciMethodData* md = method->method_data();
4531 if (md->is_empty()) {
4532 // Assume the trap has not occurred, or that it occurred only
4533 // because of a transient condition during start-up in the interpreter.
4534 return false;
4535 }
4536 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4537 if (md->has_trap_at(bci, m, reason) != 0) {
4538 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4539 // Also, if there are multiple reasons, or if there is no per-BCI record,
4540 // assume the worst.
4541 if (log())
4542 log()->elem("observe trap='%s' count='%d'",
4543 Deoptimization::trap_reason_name(reason),
4544 md->trap_count(reason));
4545 return true;
4546 } else {
4547 // Ignore method/bci and see if there have been too many globally.
4548 return too_many_traps(reason, md);
4549 }
4550 }
4551
4552 // Less-accurate variant which does not require a method and bci.
4553 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4554 ciMethodData* logmd) {
4555 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4556 // Too many traps globally.
4557 // Note that we use cumulative trap_count, not just md->trap_count.
4558 if (log()) {
4559 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4560 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4561 Deoptimization::trap_reason_name(reason),
4562 mcount, trap_count(reason));
4563 }
4564 return true;
4565 } else {
4566 // The coast is clear.
4567 return false;
4568 }
4569 }
4570
4571 //--------------------------too_many_recompiles--------------------------------
4572 // Report if there are too many recompiles at the current method and bci.
4573 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4574 // Is not eager to return true, since this will cause the compiler to use
4575 // Action_none for a trap point, to avoid too many recompilations.
4576 bool Compile::too_many_recompiles(ciMethod* method,
4577 int bci,
4578 Deoptimization::DeoptReason reason) {
4579 ciMethodData* md = method->method_data();
4580 if (md->is_empty()) {
4581 // Assume the trap has not occurred, or that it occurred only
4582 // because of a transient condition during start-up in the interpreter.
4583 return false;
4584 }
4585 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4586 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4587 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4588 Deoptimization::DeoptReason per_bc_reason
4589 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4590 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4591 if ((per_bc_reason == Deoptimization::Reason_none
4592 || md->has_trap_at(bci, m, reason) != 0)
4593 // The trap frequency measure we care about is the recompile count:
4594 && md->trap_recompiled_at(bci, m)
4595 && md->overflow_recompile_count() >= bc_cutoff) {
4596 // Do not emit a trap here if it has already caused recompilations.
4597 // Also, if there are multiple reasons, or if there is no per-BCI record,
4598 // assume the worst.
4599 if (log())
4600 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4601 Deoptimization::trap_reason_name(reason),
4602 md->trap_count(reason),
4603 md->overflow_recompile_count());
4604 return true;
4605 } else if (trap_count(reason) != 0
4606 && decompile_count() >= m_cutoff) {
4607 // Too many recompiles globally, and we have seen this sort of trap.
4608 // Use cumulative decompile_count, not just md->decompile_count.
4609 if (log())
4610 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4611 Deoptimization::trap_reason_name(reason),
4612 md->trap_count(reason), trap_count(reason),
4613 md->decompile_count(), decompile_count());
4614 return true;
4615 } else {
4616 // The coast is clear.
4617 return false;
4618 }
4619 }
4620
4621 // Compute when not to trap. Used by matching trap based nodes and
4622 // NullCheck optimization.
4623 void Compile::set_allowed_deopt_reasons() {
4624 _allowed_reasons = 0;
4625 if (is_method_compilation()) {
4626 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4627 assert(rs < BitsPerInt, "recode bit map");
4628 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4629 _allowed_reasons |= nth_bit(rs);
4630 }
4631 }
4632 }
4633 }
4634
4635 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4636 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4637 }
4638
4639 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4640 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4641 }
4642
4643 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4644 if (holder->is_initialized()) {
4645 return false;
4646 }
4647 if (holder->is_being_initialized()) {
4648 if (accessing_method->holder() == holder) {
4649 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4650 // <init>, or a static method. In all those cases, there was an initialization
4651 // barrier on the holder klass passed.
4652 if (accessing_method->is_class_initializer() ||
4653 accessing_method->is_object_constructor() ||
4654 accessing_method->is_static()) {
4655 return false;
4656 }
4657 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4658 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4659 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4660 // child class can become fully initialized while its parent class is still being initialized.
4661 if (accessing_method->is_class_initializer()) {
4662 return false;
4663 }
4664 }
4665 ciMethod* root = method(); // the root method of compilation
4666 if (root != accessing_method) {
4667 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4668 }
4669 }
4670 return true;
4671 }
4672
4673 #ifndef PRODUCT
4674 //------------------------------verify_bidirectional_edges---------------------
4675 // For each input edge to a node (ie - for each Use-Def edge), verify that
4676 // there is a corresponding Def-Use edge.
4677 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4678 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4679 uint stack_size = live_nodes() >> 4;
4680 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4681 nstack.push(_root);
4682
4683 while (nstack.size() > 0) {
4684 Node* n = nstack.pop();
4685 if (visited.member(n)) {
4686 continue;
4687 }
4688 visited.push(n);
4689
4690 // Walk over all input edges, checking for correspondence
4691 uint length = n->len();
4692 for (uint i = 0; i < length; i++) {
4693 Node* in = n->in(i);
4694 if (in != nullptr && !visited.member(in)) {
4695 nstack.push(in); // Put it on stack
4696 }
4697 if (in != nullptr && !in->is_top()) {
4698 // Count instances of `next`
4699 int cnt = 0;
4700 for (uint idx = 0; idx < in->_outcnt; idx++) {
4701 if (in->_out[idx] == n) {
4702 cnt++;
4703 }
4704 }
4705 assert(cnt > 0, "Failed to find Def-Use edge.");
4706 // Check for duplicate edges
4707 // walk the input array downcounting the input edges to n
4708 for (uint j = 0; j < length; j++) {
4709 if (n->in(j) == in) {
4710 cnt--;
4711 }
4712 }
4713 assert(cnt == 0, "Mismatched edge count.");
4714 } else if (in == nullptr) {
4715 assert(i == 0 || i >= n->req() ||
4716 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4717 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
4718 (n->is_Unlock() && i == (n->req() - 1)) ||
4719 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4720 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
4721 } else {
4722 assert(in->is_top(), "sanity");
4723 // Nothing to check.
4724 }
4725 }
4726 }
4727 }
4728
4729 //------------------------------verify_graph_edges---------------------------
4730 // Walk the Graph and verify that there is a one-to-one correspondence
4731 // between Use-Def edges and Def-Use edges in the graph.
4732 void Compile::verify_graph_edges(bool no_dead_code) {
4733 if (VerifyGraphEdges) {
4734 Unique_Node_List visited;
4735
4736 // Call graph walk to check edges
4737 verify_bidirectional_edges(visited);
4738 if (no_dead_code) {
4739 // Now make sure that no visited node is used by an unvisited node.
4740 bool dead_nodes = false;
4741 Unique_Node_List checked;
4742 while (visited.size() > 0) {
4743 Node* n = visited.pop();
4744 checked.push(n);
4745 for (uint i = 0; i < n->outcnt(); i++) {
4746 Node* use = n->raw_out(i);
4747 if (checked.member(use)) continue; // already checked
4748 if (visited.member(use)) continue; // already in the graph
4749 if (use->is_Con()) continue; // a dead ConNode is OK
4750 // At this point, we have found a dead node which is DU-reachable.
4751 if (!dead_nodes) {
4752 tty->print_cr("*** Dead nodes reachable via DU edges:");
4753 dead_nodes = true;
4754 }
4755 use->dump(2);
4756 tty->print_cr("---");
4757 checked.push(use); // No repeats; pretend it is now checked.
4758 }
4759 }
4760 assert(!dead_nodes, "using nodes must be reachable from root");
4761 }
4762 }
4763 }
4764 #endif
4765
4766 // The Compile object keeps track of failure reasons separately from the ciEnv.
4767 // This is required because there is not quite a 1-1 relation between the
4768 // ciEnv and its compilation task and the Compile object. Note that one
4769 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4770 // to backtrack and retry without subsuming loads. Other than this backtracking
4771 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4772 // by the logic in C2Compiler.
4773 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4774 if (log() != nullptr) {
4775 log()->elem("failure reason='%s' phase='compile'", reason);
4776 }
4777 if (_failure_reason.get() == nullptr) {
4778 // Record the first failure reason.
4779 _failure_reason.set(reason);
4780 if (CaptureBailoutInformation) {
4781 _first_failure_details = new CompilationFailureInfo(reason);
4782 }
4783 } else {
4784 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4785 }
4786
4787 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4788 C->print_method(PHASE_FAILURE, 1);
4789 }
4790 _root = nullptr; // flush the graph, too
4791 }
4792
4793 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4794 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4795 _compile(Compile::current()),
4796 _log(nullptr),
4797 _phase_name(name),
4798 _dolog(CITimeVerbose)
4799 {
4800 assert(_compile != nullptr, "sanity check");
4801 if (_dolog) {
4802 _log = _compile->log();
4803 }
4804 if (_log != nullptr) {
4805 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4806 _log->stamp();
4807 _log->end_head();
4808 }
4809 }
4810
4811 Compile::TracePhase::~TracePhase() {
4812 if (_compile->failing_internal()) {
4813 return; // timing code, not stressing bailouts.
4814 }
4815 #ifdef ASSERT
4816 if (PrintIdealNodeCount) {
4817 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4818 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4819 }
4820
4821 if (VerifyIdealNodeCount) {
4822 _compile->print_missing_nodes();
4823 }
4824 #endif
4825
4826 if (_log != nullptr) {
4827 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4828 }
4829 }
4830
4831 //----------------------------static_subtype_check-----------------------------
4832 // Shortcut important common cases when superklass is exact:
4833 // (0) superklass is java.lang.Object (can occur in reflective code)
4834 // (1) subklass is already limited to a subtype of superklass => always ok
4835 // (2) subklass does not overlap with superklass => always fail
4836 // (3) superklass has NO subtypes and we can check with a simple compare.
4837 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4838 if (skip) {
4839 return SSC_full_test; // Let caller generate the general case.
4840 }
4841
4842 if (subk->is_java_subtype_of(superk)) {
4843 return SSC_always_true; // (0) and (1) this test cannot fail
4844 }
4845
4846 if (!subk->maybe_java_subtype_of(superk)) {
4847 return SSC_always_false; // (2) true path dead; no dynamic test needed
4848 }
4849
4850 const Type* superelem = superk;
4851 if (superk->isa_aryklassptr()) {
4852 int ignored;
4853 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4854
4855 // Do not fold the subtype check to an array klass pointer comparison for null-able inline type arrays
4856 // because null-free [LMyValue <: null-able [LMyValue but the klasses are different. Perform a full test.
4857 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
4858 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
4859 return SSC_full_test;
4860 }
4861 }
4862
4863 if (superelem->isa_instklassptr()) {
4864 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4865 if (!ik->has_subklass()) {
4866 if (!ik->is_final()) {
4867 // Add a dependency if there is a chance of a later subclass.
4868 dependencies()->assert_leaf_type(ik);
4869 }
4870 if (!superk->maybe_java_subtype_of(subk)) {
4871 return SSC_always_false;
4872 }
4873 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4874 }
4875 } else {
4876 // A primitive array type has no subtypes.
4877 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4878 }
4879
4880 return SSC_full_test;
4881 }
4882
4883 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4884 #ifdef _LP64
4885 // The scaled index operand to AddP must be a clean 64-bit value.
4886 // Java allows a 32-bit int to be incremented to a negative
4887 // value, which appears in a 64-bit register as a large
4888 // positive number. Using that large positive number as an
4889 // operand in pointer arithmetic has bad consequences.
4890 // On the other hand, 32-bit overflow is rare, and the possibility
4891 // can often be excluded, if we annotate the ConvI2L node with
4892 // a type assertion that its value is known to be a small positive
4893 // number. (The prior range check has ensured this.)
4894 // This assertion is used by ConvI2LNode::Ideal.
4895 int index_max = max_jint - 1; // array size is max_jint, index is one less
4896 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4897 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4898 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4899 #endif
4900 return idx;
4901 }
4902
4903 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4904 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4905 if (ctrl != nullptr) {
4906 // Express control dependency by a CastII node with a narrow type.
4907 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4908 // node from floating above the range check during loop optimizations. Otherwise, the
4909 // ConvI2L node may be eliminated independently of the range check, causing the data path
4910 // to become TOP while the control path is still there (although it's unreachable).
4911 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4912 value = phase->transform(value);
4913 }
4914 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4915 return phase->transform(new ConvI2LNode(value, ltype));
4916 }
4917
4918 // The message about the current inlining is accumulated in
4919 // _print_inlining_stream and transferred into the _print_inlining_list
4920 // once we know whether inlining succeeds or not. For regular
4921 // inlining, messages are appended to the buffer pointed by
4922 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4923 // a new buffer is added after _print_inlining_idx in the list. This
4924 // way we can update the inlining message for late inlining call site
4925 // when the inlining is attempted again.
4926 void Compile::print_inlining_init() {
4927 if (print_inlining() || print_intrinsics()) {
4928 // print_inlining_init is actually called several times.
4929 print_inlining_reset();
4930 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4931 }
4932 }
4933
4934 void Compile::print_inlining_reinit() {
4935 if (print_inlining() || print_intrinsics()) {
4936 print_inlining_reset();
4937 }
4938 }
4939
4940 void Compile::print_inlining_reset() {
4941 _print_inlining_stream->reset();
4942 }
4943
4944 void Compile::print_inlining_commit() {
4945 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4946 // Transfer the message from _print_inlining_stream to the current
4947 // _print_inlining_list buffer and clear _print_inlining_stream.
4948 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4949 print_inlining_reset();
4950 }
4951
4952 void Compile::print_inlining_push() {
4953 // Add new buffer to the _print_inlining_list at current position
4954 _print_inlining_idx++;
4955 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4956 }
4957
4958 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4959 return _print_inlining_list->at(_print_inlining_idx);
4960 }
4961
4962 void Compile::print_inlining_update(CallGenerator* cg) {
4963 if (print_inlining() || print_intrinsics()) {
4964 if (cg->is_late_inline()) {
4965 if (print_inlining_current()->cg() != cg &&
4966 (print_inlining_current()->cg() != nullptr ||
4967 print_inlining_current()->ss()->size() != 0)) {
4968 print_inlining_push();
4969 }
4970 print_inlining_commit();
4971 print_inlining_current()->set_cg(cg);
4972 } else {
4973 if (print_inlining_current()->cg() != nullptr) {
4974 print_inlining_push();
4975 }
4976 print_inlining_commit();
4977 }
4978 }
4979 }
4980
4981 void Compile::print_inlining_move_to(CallGenerator* cg) {
4982 // We resume inlining at a late inlining call site. Locate the
4983 // corresponding inlining buffer so that we can update it.
4984 if (print_inlining() || print_intrinsics()) {
4985 for (int i = 0; i < _print_inlining_list->length(); i++) {
4986 if (_print_inlining_list->at(i)->cg() == cg) {
4987 _print_inlining_idx = i;
4988 return;
4989 }
4990 }
4991 ShouldNotReachHere();
4992 }
4993 }
4994
4995 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4996 if (print_inlining() || print_intrinsics()) {
4997 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4998 assert(print_inlining_current()->cg() == cg, "wrong entry");
4999 // replace message with new message
5000 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
5001 print_inlining_commit();
5002 print_inlining_current()->set_cg(cg);
5003 }
5004 }
5005
5006 void Compile::print_inlining_assert_ready() {
5007 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
5008 }
5009
5010 void Compile::process_print_inlining() {
5011 assert(_late_inlines.length() == 0, "not drained yet");
5012 if (print_inlining() || print_intrinsics()) {
5013 ResourceMark rm;
5014 stringStream ss;
5015 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
5016 for (int i = 0; i < _print_inlining_list->length(); i++) {
5017 PrintInliningBuffer* pib = _print_inlining_list->at(i);
5018 ss.print("%s", pib->ss()->freeze());
5019 delete pib;
5020 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
5021 }
5022 // Reset _print_inlining_list, it only contains destructed objects.
5023 // It is on the arena, so it will be freed when the arena is reset.
5024 _print_inlining_list = nullptr;
5025 // _print_inlining_stream won't be used anymore, either.
5026 print_inlining_reset();
5027 size_t end = ss.size();
5028 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
5029 strncpy(_print_inlining_output, ss.freeze(), end+1);
5030 _print_inlining_output[end] = 0;
5031 }
5032 }
5033
5034 void Compile::dump_print_inlining() {
5035 if (_print_inlining_output != nullptr) {
5036 tty->print_raw(_print_inlining_output);
5037 }
5038 }
5039
5040 void Compile::log_late_inline(CallGenerator* cg) {
5041 if (log() != nullptr) {
5042 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5043 cg->unique_id());
5044 JVMState* p = cg->call_node()->jvms();
5045 while (p != nullptr) {
5046 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5047 p = p->caller();
5048 }
5049 log()->tail("late_inline");
5050 }
5051 }
5052
5053 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5054 log_late_inline(cg);
5055 if (log() != nullptr) {
5056 log()->inline_fail(msg);
5057 }
5058 }
5059
5060 void Compile::log_inline_id(CallGenerator* cg) {
5061 if (log() != nullptr) {
5062 // The LogCompilation tool needs a unique way to identify late
5063 // inline call sites. This id must be unique for this call site in
5064 // this compilation. Try to have it unique across compilations as
5065 // well because it can be convenient when grepping through the log
5066 // file.
5067 // Distinguish OSR compilations from others in case CICountOSR is
5068 // on.
5069 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5070 cg->set_unique_id(id);
5071 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5072 }
5073 }
5074
5075 void Compile::log_inline_failure(const char* msg) {
5076 if (C->log() != nullptr) {
5077 C->log()->inline_fail(msg);
5078 }
5079 }
5080
5081
5082 // Dump inlining replay data to the stream.
5083 // Don't change thread state and acquire any locks.
5084 void Compile::dump_inline_data(outputStream* out) {
5085 InlineTree* inl_tree = ilt();
5086 if (inl_tree != nullptr) {
5087 out->print(" inline %d", inl_tree->count());
5088 inl_tree->dump_replay_data(out);
5089 }
5090 }
5091
5092 void Compile::dump_inline_data_reduced(outputStream* out) {
5093 assert(ReplayReduce, "");
5094
5095 InlineTree* inl_tree = ilt();
5096 if (inl_tree == nullptr) {
5097 return;
5098 }
5099 // Enable iterative replay file reduction
5100 // Output "compile" lines for depth 1 subtrees,
5101 // simulating that those trees were compiled
5102 // instead of inlined.
5103 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5104 InlineTree* sub = inl_tree->subtrees().at(i);
5105 if (sub->inline_level() != 1) {
5106 continue;
5107 }
5108
5109 ciMethod* method = sub->method();
5110 int entry_bci = -1;
5111 int comp_level = env()->task()->comp_level();
5112 out->print("compile ");
5113 method->dump_name_as_ascii(out);
5114 out->print(" %d %d", entry_bci, comp_level);
5115 out->print(" inline %d", sub->count());
5116 sub->dump_replay_data(out, -1);
5117 out->cr();
5118 }
5119 }
5120
5121 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5122 if (n1->Opcode() < n2->Opcode()) return -1;
5123 else if (n1->Opcode() > n2->Opcode()) return 1;
5124
5125 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5126 for (uint i = 1; i < n1->req(); i++) {
5127 if (n1->in(i) < n2->in(i)) return -1;
5128 else if (n1->in(i) > n2->in(i)) return 1;
5129 }
5130
5131 return 0;
5132 }
5133
5134 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5135 Node* n1 = *n1p;
5136 Node* n2 = *n2p;
5137
5138 return cmp_expensive_nodes(n1, n2);
5139 }
5140
5141 void Compile::sort_expensive_nodes() {
5142 if (!expensive_nodes_sorted()) {
5143 _expensive_nodes.sort(cmp_expensive_nodes);
5144 }
5145 }
5146
5147 bool Compile::expensive_nodes_sorted() const {
5148 for (int i = 1; i < _expensive_nodes.length(); i++) {
5149 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5150 return false;
5151 }
5152 }
5153 return true;
5154 }
5155
5156 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5157 if (_expensive_nodes.length() == 0) {
5158 return false;
5159 }
5160
5161 assert(OptimizeExpensiveOps, "optimization off?");
5162
5163 // Take this opportunity to remove dead nodes from the list
5164 int j = 0;
5165 for (int i = 0; i < _expensive_nodes.length(); i++) {
5166 Node* n = _expensive_nodes.at(i);
5167 if (!n->is_unreachable(igvn)) {
5168 assert(n->is_expensive(), "should be expensive");
5169 _expensive_nodes.at_put(j, n);
5170 j++;
5171 }
5172 }
5173 _expensive_nodes.trunc_to(j);
5174
5175 // Then sort the list so that similar nodes are next to each other
5176 // and check for at least two nodes of identical kind with same data
5177 // inputs.
5178 sort_expensive_nodes();
5179
5180 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5181 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5182 return true;
5183 }
5184 }
5185
5186 return false;
5187 }
5188
5189 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5190 if (_expensive_nodes.length() == 0) {
5191 return;
5192 }
5193
5194 assert(OptimizeExpensiveOps, "optimization off?");
5195
5196 // Sort to bring similar nodes next to each other and clear the
5197 // control input of nodes for which there's only a single copy.
5198 sort_expensive_nodes();
5199
5200 int j = 0;
5201 int identical = 0;
5202 int i = 0;
5203 bool modified = false;
5204 for (; i < _expensive_nodes.length()-1; i++) {
5205 assert(j <= i, "can't write beyond current index");
5206 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5207 identical++;
5208 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5209 continue;
5210 }
5211 if (identical > 0) {
5212 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5213 identical = 0;
5214 } else {
5215 Node* n = _expensive_nodes.at(i);
5216 igvn.replace_input_of(n, 0, nullptr);
5217 igvn.hash_insert(n);
5218 modified = true;
5219 }
5220 }
5221 if (identical > 0) {
5222 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5223 } else if (_expensive_nodes.length() >= 1) {
5224 Node* n = _expensive_nodes.at(i);
5225 igvn.replace_input_of(n, 0, nullptr);
5226 igvn.hash_insert(n);
5227 modified = true;
5228 }
5229 _expensive_nodes.trunc_to(j);
5230 if (modified) {
5231 igvn.optimize();
5232 }
5233 }
5234
5235 void Compile::add_expensive_node(Node * n) {
5236 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5237 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5238 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5239 if (OptimizeExpensiveOps) {
5240 _expensive_nodes.append(n);
5241 } else {
5242 // Clear control input and let IGVN optimize expensive nodes if
5243 // OptimizeExpensiveOps is off.
5244 n->set_req(0, nullptr);
5245 }
5246 }
5247
5248 /**
5249 * Track coarsened Lock and Unlock nodes.
5250 */
5251
5252 class Lock_List : public Node_List {
5253 uint _origin_cnt;
5254 public:
5255 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5256 uint origin_cnt() const { return _origin_cnt; }
5257 };
5258
5259 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5260 int length = locks.length();
5261 if (length > 0) {
5262 // Have to keep this list until locks elimination during Macro nodes elimination.
5263 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5264 AbstractLockNode* alock = locks.at(0);
5265 BoxLockNode* box = alock->box_node()->as_BoxLock();
5266 for (int i = 0; i < length; i++) {
5267 AbstractLockNode* lock = locks.at(i);
5268 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5269 locks_list->push(lock);
5270 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
5271 if (this_box != box) {
5272 // Locking regions (BoxLock) could be Unbalanced here:
5273 // - its coarsened locks were eliminated in earlier
5274 // macro nodes elimination followed by loop unroll
5275 // - it is OSR locking region (no Lock node)
5276 // Preserve Unbalanced status in such cases.
5277 if (!this_box->is_unbalanced()) {
5278 this_box->set_coarsened();
5279 }
5280 if (!box->is_unbalanced()) {
5281 box->set_coarsened();
5282 }
5283 }
5284 }
5285 _coarsened_locks.append(locks_list);
5286 }
5287 }
5288
5289 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5290 int count = coarsened_count();
5291 for (int i = 0; i < count; i++) {
5292 Node_List* locks_list = _coarsened_locks.at(i);
5293 for (uint j = 0; j < locks_list->size(); j++) {
5294 Node* lock = locks_list->at(j);
5295 assert(lock->is_AbstractLock(), "sanity");
5296 if (!useful.member(lock)) {
5297 locks_list->yank(lock);
5298 }
5299 }
5300 }
5301 }
5302
5303 void Compile::remove_coarsened_lock(Node* n) {
5304 if (n->is_AbstractLock()) {
5305 int count = coarsened_count();
5306 for (int i = 0; i < count; i++) {
5307 Node_List* locks_list = _coarsened_locks.at(i);
5308 locks_list->yank(n);
5309 }
5310 }
5311 }
5312
5313 bool Compile::coarsened_locks_consistent() {
5314 int count = coarsened_count();
5315 for (int i = 0; i < count; i++) {
5316 bool unbalanced = false;
5317 bool modified = false; // track locks kind modifications
5318 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5319 uint size = locks_list->size();
5320 if (size == 0) {
5321 unbalanced = false; // All locks were eliminated - good
5322 } else if (size != locks_list->origin_cnt()) {
5323 unbalanced = true; // Some locks were removed from list
5324 } else {
5325 for (uint j = 0; j < size; j++) {
5326 Node* lock = locks_list->at(j);
5327 // All nodes in group should have the same state (modified or not)
5328 if (!lock->as_AbstractLock()->is_coarsened()) {
5329 if (j == 0) {
5330 // first on list was modified, the rest should be too for consistency
5331 modified = true;
5332 } else if (!modified) {
5333 // this lock was modified but previous locks on the list were not
5334 unbalanced = true;
5335 break;
5336 }
5337 } else if (modified) {
5338 // previous locks on list were modified but not this lock
5339 unbalanced = true;
5340 break;
5341 }
5342 }
5343 }
5344 if (unbalanced) {
5345 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5346 #ifdef ASSERT
5347 if (PrintEliminateLocks) {
5348 tty->print_cr("=== unbalanced coarsened locks ===");
5349 for (uint l = 0; l < size; l++) {
5350 locks_list->at(l)->dump();
5351 }
5352 }
5353 #endif
5354 record_failure(C2Compiler::retry_no_locks_coarsening());
5355 return false;
5356 }
5357 }
5358 return true;
5359 }
5360
5361 // Mark locking regions (identified by BoxLockNode) as unbalanced if
5362 // locks coarsening optimization removed Lock/Unlock nodes from them.
5363 // Such regions become unbalanced because coarsening only removes part
5364 // of Lock/Unlock nodes in region. As result we can't execute other
5365 // locks elimination optimizations which assume all code paths have
5366 // corresponding pair of Lock/Unlock nodes - they are balanced.
5367 void Compile::mark_unbalanced_boxes() const {
5368 int count = coarsened_count();
5369 for (int i = 0; i < count; i++) {
5370 Node_List* locks_list = _coarsened_locks.at(i);
5371 uint size = locks_list->size();
5372 if (size > 0) {
5373 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
5374 BoxLockNode* box = alock->box_node()->as_BoxLock();
5375 if (alock->is_coarsened()) {
5376 // coarsened_locks_consistent(), which is called before this method, verifies
5377 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
5378 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5379 for (uint j = 1; j < size; j++) {
5380 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
5381 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
5382 if (box != this_box) {
5383 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5384 box->set_unbalanced();
5385 this_box->set_unbalanced();
5386 }
5387 }
5388 }
5389 }
5390 }
5391 }
5392
5393 /**
5394 * Remove the speculative part of types and clean up the graph
5395 */
5396 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5397 if (UseTypeSpeculation) {
5398 Unique_Node_List worklist;
5399 worklist.push(root());
5400 int modified = 0;
5401 // Go over all type nodes that carry a speculative type, drop the
5402 // speculative part of the type and enqueue the node for an igvn
5403 // which may optimize it out.
5404 for (uint next = 0; next < worklist.size(); ++next) {
5405 Node *n = worklist.at(next);
5406 if (n->is_Type()) {
5407 TypeNode* tn = n->as_Type();
5408 const Type* t = tn->type();
5409 const Type* t_no_spec = t->remove_speculative();
5410 if (t_no_spec != t) {
5411 bool in_hash = igvn.hash_delete(n);
5412 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5413 tn->set_type(t_no_spec);
5414 igvn.hash_insert(n);
5415 igvn._worklist.push(n); // give it a chance to go away
5416 modified++;
5417 }
5418 }
5419 // Iterate over outs - endless loops is unreachable from below
5420 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5421 Node *m = n->fast_out(i);
5422 if (not_a_node(m)) {
5423 continue;
5424 }
5425 worklist.push(m);
5426 }
5427 }
5428 // Drop the speculative part of all types in the igvn's type table
5429 igvn.remove_speculative_types();
5430 if (modified > 0) {
5431 igvn.optimize();
5432 if (failing()) return;
5433 }
5434 #ifdef ASSERT
5435 // Verify that after the IGVN is over no speculative type has resurfaced
5436 worklist.clear();
5437 worklist.push(root());
5438 for (uint next = 0; next < worklist.size(); ++next) {
5439 Node *n = worklist.at(next);
5440 const Type* t = igvn.type_or_null(n);
5441 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5442 if (n->is_Type()) {
5443 t = n->as_Type()->type();
5444 assert(t == t->remove_speculative(), "no more speculative types");
5445 }
5446 // Iterate over outs - endless loops is unreachable from below
5447 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5448 Node *m = n->fast_out(i);
5449 if (not_a_node(m)) {
5450 continue;
5451 }
5452 worklist.push(m);
5453 }
5454 }
5455 igvn.check_no_speculative_types();
5456 #endif
5457 }
5458 }
5459
5460 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5461 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5462 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5463 if (!EnableValhalla || ta == nullptr || tb == nullptr ||
5464 ta->is_zero_type() || tb->is_zero_type() ||
5465 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5466 // Use old acmp if one operand is null or not an inline type
5467 return new CmpPNode(a, b);
5468 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5469 // We know that one operand is an inline type. Therefore,
5470 // new acmp will only return true if both operands are nullptr.
5471 // Check if both operands are null by or'ing the oops.
5472 a = phase->transform(new CastP2XNode(nullptr, a));
5473 b = phase->transform(new CastP2XNode(nullptr, b));
5474 a = phase->transform(new OrXNode(a, b));
5475 return new CmpXNode(a, phase->MakeConX(0));
5476 }
5477 // Use new acmp
5478 return nullptr;
5479 }
5480
5481 // Auxiliary methods to support randomized stressing/fuzzing.
5482
5483 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5484 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5485 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5486 FLAG_SET_ERGO(StressSeed, _stress_seed);
5487 } else {
5488 _stress_seed = StressSeed;
5489 }
5490 if (_log != nullptr) {
5491 _log->elem("stress_test seed='%u'", _stress_seed);
5492 }
5493 }
5494
5495 int Compile::random() {
5496 _stress_seed = os::next_random(_stress_seed);
5497 return static_cast<int>(_stress_seed);
5498 }
5499
5500 // This method can be called the arbitrary number of times, with current count
5501 // as the argument. The logic allows selecting a single candidate from the
5502 // running list of candidates as follows:
5503 // int count = 0;
5504 // Cand* selected = null;
5505 // while(cand = cand->next()) {
5506 // if (randomized_select(++count)) {
5507 // selected = cand;
5508 // }
5509 // }
5510 //
5511 // Including count equalizes the chances any candidate is "selected".
5512 // This is useful when we don't have the complete list of candidates to choose
5513 // from uniformly. In this case, we need to adjust the randomicity of the
5514 // selection, or else we will end up biasing the selection towards the latter
5515 // candidates.
5516 //
5517 // Quick back-envelope calculation shows that for the list of n candidates
5518 // the equal probability for the candidate to persist as "best" can be
5519 // achieved by replacing it with "next" k-th candidate with the probability
5520 // of 1/k. It can be easily shown that by the end of the run, the
5521 // probability for any candidate is converged to 1/n, thus giving the
5522 // uniform distribution among all the candidates.
5523 //
5524 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5525 #define RANDOMIZED_DOMAIN_POW 29
5526 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5527 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5528 bool Compile::randomized_select(int count) {
5529 assert(count > 0, "only positive");
5530 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5531 }
5532
5533 #ifdef ASSERT
5534 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5535 bool Compile::fail_randomly() {
5536 if ((random() % StressBailoutMean) != 0) {
5537 return false;
5538 }
5539 record_failure("StressBailout");
5540 return true;
5541 }
5542
5543 bool Compile::failure_is_artificial() {
5544 assert(failing_internal(), "should be failing");
5545 return C->failure_reason_is("StressBailout");
5546 }
5547 #endif
5548
5549 CloneMap& Compile::clone_map() { return _clone_map; }
5550 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5551
5552 void NodeCloneInfo::dump_on(outputStream* st) const {
5553 st->print(" {%d:%d} ", idx(), gen());
5554 }
5555
5556 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5557 uint64_t val = value(old->_idx);
5558 NodeCloneInfo cio(val);
5559 assert(val != 0, "old node should be in the map");
5560 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5561 insert(nnn->_idx, cin.get());
5562 #ifndef PRODUCT
5563 if (is_debug()) {
5564 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5565 }
5566 #endif
5567 }
5568
5569 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5570 NodeCloneInfo cio(value(old->_idx));
5571 if (cio.get() == 0) {
5572 cio.set(old->_idx, 0);
5573 insert(old->_idx, cio.get());
5574 #ifndef PRODUCT
5575 if (is_debug()) {
5576 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5577 }
5578 #endif
5579 }
5580 clone(old, nnn, gen);
5581 }
5582
5583 int CloneMap::max_gen() const {
5584 int g = 0;
5585 DictI di(_dict);
5586 for(; di.test(); ++di) {
5587 int t = gen(di._key);
5588 if (g < t) {
5589 g = t;
5590 #ifndef PRODUCT
5591 if (is_debug()) {
5592 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5593 }
5594 #endif
5595 }
5596 }
5597 return g;
5598 }
5599
5600 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5601 uint64_t val = value(key);
5602 if (val != 0) {
5603 NodeCloneInfo ni(val);
5604 ni.dump_on(st);
5605 }
5606 }
5607
5608 void Compile::shuffle_macro_nodes() {
5609 if (_macro_nodes.length() < 2) {
5610 return;
5611 }
5612 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5613 uint j = C->random() % (i + 1);
5614 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5615 }
5616 }
5617
5618 // Move Allocate nodes to the start of the list
5619 void Compile::sort_macro_nodes() {
5620 int count = macro_count();
5621 int allocates = 0;
5622 for (int i = 0; i < count; i++) {
5623 Node* n = macro_node(i);
5624 if (n->is_Allocate()) {
5625 if (i != allocates) {
5626 Node* tmp = macro_node(allocates);
5627 _macro_nodes.at_put(allocates, n);
5628 _macro_nodes.at_put(i, tmp);
5629 }
5630 allocates++;
5631 }
5632 }
5633 }
5634
5635 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5636 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5637 EventCompilerPhase event(UNTIMED);
5638 if (event.should_commit()) {
5639 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5640 }
5641 #ifndef PRODUCT
5642 ResourceMark rm;
5643 stringStream ss;
5644 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5645 int iter = ++_igv_phase_iter[cpt];
5646 if (iter > 1) {
5647 ss.print(" %d", iter);
5648 }
5649 if (n != nullptr) {
5650 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5651 if (n->is_Call()) {
5652 CallNode* call = n->as_Call();
5653 if (call->_name != nullptr) {
5654 // E.g. uncommon traps etc.
5655 ss.print(" - %s", call->_name);
5656 } else if (call->is_CallJava()) {
5657 CallJavaNode* call_java = call->as_CallJava();
5658 if (call_java->method() != nullptr) {
5659 ss.print(" -");
5660 call_java->method()->print_short_name(&ss);
5661 }
5662 }
5663 }
5664 }
5665
5666 const char* name = ss.as_string();
5667 if (should_print_igv(level)) {
5668 _igv_printer->print_method(name, level);
5669 }
5670 if (should_print_phase(cpt)) {
5671 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5672 }
5673 #endif
5674 C->_latest_stage_start_counter.stamp();
5675 }
5676
5677 // Only used from CompileWrapper
5678 void Compile::begin_method() {
5679 #ifndef PRODUCT
5680 if (_method != nullptr && should_print_igv(1)) {
5681 _igv_printer->begin_method();
5682 }
5683 #endif
5684 C->_latest_stage_start_counter.stamp();
5685 }
5686
5687 // Only used from CompileWrapper
5688 void Compile::end_method() {
5689 EventCompilerPhase event(UNTIMED);
5690 if (event.should_commit()) {
5691 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5692 }
5693
5694 #ifndef PRODUCT
5695 if (_method != nullptr && should_print_igv(1)) {
5696 _igv_printer->end_method();
5697 }
5698 #endif
5699 }
5700
5701 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5702 #ifndef PRODUCT
5703 if (_directive->should_print_phase(cpt)) {
5704 return true;
5705 }
5706 #endif
5707 return false;
5708 }
5709
5710 bool Compile::should_print_igv(const int level) {
5711 #ifndef PRODUCT
5712 if (PrintIdealGraphLevel < 0) { // disabled by the user
5713 return false;
5714 }
5715
5716 bool need = directive()->IGVPrintLevelOption >= level;
5717 if (need && !_igv_printer) {
5718 _igv_printer = IdealGraphPrinter::printer();
5719 _igv_printer->set_compile(this);
5720 }
5721 return need;
5722 #else
5723 return false;
5724 #endif
5725 }
5726
5727 #ifndef PRODUCT
5728 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5729 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5730
5731 // Called from debugger. Prints method to the default file with the default phase name.
5732 // This works regardless of any Ideal Graph Visualizer flags set or not.
5733 void igv_print() {
5734 Compile::current()->igv_print_method_to_file();
5735 }
5736
5737 // Same as igv_print() above but with a specified phase name.
5738 void igv_print(const char* phase_name) {
5739 Compile::current()->igv_print_method_to_file(phase_name);
5740 }
5741
5742 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5743 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5744 // This works regardless of any Ideal Graph Visualizer flags set or not.
5745 void igv_print(bool network) {
5746 if (network) {
5747 Compile::current()->igv_print_method_to_network();
5748 } else {
5749 Compile::current()->igv_print_method_to_file();
5750 }
5751 }
5752
5753 // Same as igv_print(bool network) above but with a specified phase name.
5754 void igv_print(bool network, const char* phase_name) {
5755 if (network) {
5756 Compile::current()->igv_print_method_to_network(phase_name);
5757 } else {
5758 Compile::current()->igv_print_method_to_file(phase_name);
5759 }
5760 }
5761
5762 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5763 void igv_print_default() {
5764 Compile::current()->print_method(PHASE_DEBUG, 0);
5765 }
5766
5767 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5768 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5769 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5770 void igv_append() {
5771 Compile::current()->igv_print_method_to_file("Debug", true);
5772 }
5773
5774 // Same as igv_append() above but with a specified phase name.
5775 void igv_append(const char* phase_name) {
5776 Compile::current()->igv_print_method_to_file(phase_name, true);
5777 }
5778
5779 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5780 const char* file_name = "custom_debug.xml";
5781 if (_debug_file_printer == nullptr) {
5782 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5783 } else {
5784 _debug_file_printer->update_compiled_method(C->method());
5785 }
5786 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5787 _debug_file_printer->print(phase_name, (Node*)C->root());
5788 }
5789
5790 void Compile::igv_print_method_to_network(const char* phase_name) {
5791 if (_debug_network_printer == nullptr) {
5792 _debug_network_printer = new IdealGraphPrinter(C);
5793 } else {
5794 _debug_network_printer->update_compiled_method(C->method());
5795 }
5796 tty->print_cr("Method printed over network stream to IGV");
5797 _debug_network_printer->print(phase_name, (Node*)C->root());
5798 }
5799 #endif
5800
5801 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5802 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5803 return value;
5804 }
5805 Node* result = nullptr;
5806 if (bt == T_BYTE) {
5807 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5808 result = new RShiftINode(result, phase->intcon(24));
5809 } else if (bt == T_BOOLEAN) {
5810 result = new AndINode(value, phase->intcon(0xFF));
5811 } else if (bt == T_CHAR) {
5812 result = new AndINode(value,phase->intcon(0xFFFF));
5813 } else {
5814 assert(bt == T_SHORT, "unexpected narrow type");
5815 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5816 result = new RShiftINode(result, phase->intcon(16));
5817 }
5818 if (transform_res) {
5819 result = phase->transform(result);
5820 }
5821 return result;
5822 }
5823
5824 void Compile::record_method_not_compilable_oom() {
5825 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5826 }