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