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