1 /*
2 * Copyright (c) 1997, 2026, 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/aotCodeCache.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compiler_globals.hpp"
37 #include "compiler/compilerDefinitions.hpp"
38 #include "compiler/compilerOracle.hpp"
39 #include "compiler/disassembler.hpp"
40 #include "compiler/oopMap.hpp"
41 #include "gc/shared/barrierSet.hpp"
42 #include "gc/shared/c2/barrierSetC2.hpp"
43 #include "jfr/jfrEvents.hpp"
44 #include "jvm_io.h"
45 #include "memory/allocation.hpp"
46 #include "memory/arena.hpp"
47 #include "memory/resourceArea.hpp"
48 #include "opto/addnode.hpp"
49 #include "opto/block.hpp"
50 #include "opto/c2compiler.hpp"
51 #include "opto/callGenerator.hpp"
52 #include "opto/callnode.hpp"
53 #include "opto/castnode.hpp"
54 #include "opto/cfgnode.hpp"
55 #include "opto/chaitin.hpp"
56 #include "opto/compile.hpp"
57 #include "opto/connode.hpp"
58 #include "opto/convertnode.hpp"
59 #include "opto/divnode.hpp"
60 #include "opto/escape.hpp"
61 #include "opto/idealGraphPrinter.hpp"
62 #include "opto/locknode.hpp"
63 #include "opto/loopnode.hpp"
64 #include "opto/machnode.hpp"
65 #include "opto/macro.hpp"
66 #include "opto/matcher.hpp"
67 #include "opto/mathexactnode.hpp"
68 #include "opto/memnode.hpp"
69 #include "opto/mulnode.hpp"
70 #include "opto/narrowptrnode.hpp"
71 #include "opto/node.hpp"
72 #include "opto/opaquenode.hpp"
73 #include "opto/opcodes.hpp"
74 #include "opto/output.hpp"
75 #include "opto/parse.hpp"
76 #include "opto/phaseX.hpp"
77 #include "opto/rootnode.hpp"
78 #include "opto/runtime.hpp"
79 #include "opto/stringopts.hpp"
80 #include "opto/type.hpp"
81 #include "opto/vector.hpp"
82 #include "opto/vectornode.hpp"
83 #include "runtime/globals_extension.hpp"
84 #include "runtime/sharedRuntime.hpp"
85 #include "runtime/signature.hpp"
86 #include "runtime/stubRoutines.hpp"
87 #include "runtime/timer.hpp"
88 #include "utilities/align.hpp"
89 #include "utilities/copy.hpp"
90 #include "utilities/hashTable.hpp"
91 #include "utilities/macros.hpp"
92
93 // -------------------- Compile::mach_constant_base_node -----------------------
94 // Constant table base node singleton.
95 MachConstantBaseNode* Compile::mach_constant_base_node() {
96 if (_mach_constant_base_node == nullptr) {
97 _mach_constant_base_node = new MachConstantBaseNode();
98 _mach_constant_base_node->add_req(C->root());
99 }
100 return _mach_constant_base_node;
101 }
102
103
104 /// Support for intrinsics.
105
106 // Return the index at which m must be inserted (or already exists).
107 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
108 class IntrinsicDescPair {
109 private:
110 ciMethod* _m;
111 bool _is_virtual;
112 public:
113 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
114 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
115 ciMethod* m= elt->method();
116 ciMethod* key_m = key->_m;
117 if (key_m < m) return -1;
118 else if (key_m > m) return 1;
119 else {
120 bool is_virtual = elt->is_virtual();
121 bool key_virtual = key->_is_virtual;
122 if (key_virtual < is_virtual) return -1;
123 else if (key_virtual > is_virtual) return 1;
124 else return 0;
125 }
126 }
127 };
128 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
129 #ifdef ASSERT
130 for (int i = 1; i < _intrinsics.length(); i++) {
131 CallGenerator* cg1 = _intrinsics.at(i-1);
132 CallGenerator* cg2 = _intrinsics.at(i);
133 assert(cg1->method() != cg2->method()
134 ? cg1->method() < cg2->method()
135 : cg1->is_virtual() < cg2->is_virtual(),
136 "compiler intrinsics list must stay sorted");
137 }
138 #endif
139 IntrinsicDescPair pair(m, is_virtual);
140 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
141 }
142
143 void Compile::register_intrinsic(CallGenerator* cg) {
144 bool found = false;
145 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
146 assert(!found, "registering twice");
147 _intrinsics.insert_before(index, cg);
148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
149 }
150
151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
152 assert(m->is_loaded(), "don't try this on unloaded methods");
153 if (_intrinsics.length() > 0) {
154 bool found = false;
155 int index = intrinsic_insertion_index(m, is_virtual, found);
156 if (found) {
157 return _intrinsics.at(index);
158 }
159 }
160 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
161 if (m->intrinsic_id() != vmIntrinsics::_none &&
162 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
163 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
164 if (cg != nullptr) {
165 // Save it for next time:
166 register_intrinsic(cg);
167 return cg;
168 } else {
169 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
170 }
171 }
172 return nullptr;
173 }
174
175 // Compile::make_vm_intrinsic is defined in library_call.cpp.
176
177 #ifndef PRODUCT
178 // statistics gathering...
179
180 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
182
183 inline int as_int(vmIntrinsics::ID id) {
184 return vmIntrinsics::as_int(id);
185 }
186
187 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
188 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
189 int oflags = _intrinsic_hist_flags[as_int(id)];
190 assert(flags != 0, "what happened?");
191 if (is_virtual) {
192 flags |= _intrinsic_virtual;
193 }
194 bool changed = (flags != oflags);
195 if ((flags & _intrinsic_worked) != 0) {
196 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
197 if (count == 1) {
198 changed = true; // first time
199 }
200 // increment the overall count also:
201 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
202 }
203 if (changed) {
204 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
205 // Something changed about the intrinsic's virtuality.
206 if ((flags & _intrinsic_virtual) != 0) {
207 // This is the first use of this intrinsic as a virtual call.
208 if (oflags != 0) {
209 // We already saw it as a non-virtual, so note both cases.
210 flags |= _intrinsic_both;
211 }
212 } else if ((oflags & _intrinsic_both) == 0) {
213 // This is the first use of this intrinsic as a non-virtual
214 flags |= _intrinsic_both;
215 }
216 }
217 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
218 }
219 // update the overall flags also:
220 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
221 return changed;
222 }
223
224 static char* format_flags(int flags, char* buf) {
225 buf[0] = 0;
226 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
227 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
228 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
229 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
230 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
231 if (buf[0] == 0) strcat(buf, ",");
232 assert(buf[0] == ',', "must be");
233 return &buf[1];
234 }
235
236 void Compile::print_intrinsic_statistics() {
237 char flagsbuf[100];
238 ttyLocker ttyl;
239 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
240 tty->print_cr("Compiler intrinsic usage:");
241 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
242 if (total == 0) total = 1; // avoid div0 in case of no successes
243 #define PRINT_STAT_LINE(name, c, f) \
244 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
245 for (auto id : EnumRange<vmIntrinsicID>{}) {
246 int flags = _intrinsic_hist_flags[as_int(id)];
247 juint count = _intrinsic_hist_count[as_int(id)];
248 if ((flags | count) != 0) {
249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
250 }
251 }
252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
253 if (xtty != nullptr) xtty->tail("statistics");
254 }
255
256 void Compile::print_statistics() {
257 { ttyLocker ttyl;
258 if (xtty != nullptr) xtty->head("statistics type='opto'");
259 Parse::print_statistics();
260 PhaseStringOpts::print_statistics();
261 PhaseCCP::print_statistics();
262 PhaseRegAlloc::print_statistics();
263 PhaseOutput::print_statistics();
264 PhasePeephole::print_statistics();
265 PhaseIdealLoop::print_statistics();
266 ConnectionGraph::print_statistics();
267 PhaseMacroExpand::print_statistics();
268 if (xtty != nullptr) xtty->tail("statistics");
269 }
270 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
271 // put this under its own <statistics> element.
272 print_intrinsic_statistics();
273 }
274 }
275 #endif //PRODUCT
276
277 void Compile::gvn_replace_by(Node* n, Node* nn) {
278 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
279 Node* use = n->last_out(i);
280 bool is_in_table = initial_gvn()->hash_delete(use);
281 uint uses_found = 0;
282 for (uint j = 0; j < use->len(); j++) {
283 if (use->in(j) == n) {
284 if (j < use->req())
285 use->set_req(j, nn);
286 else
287 use->set_prec(j, nn);
288 uses_found++;
289 }
290 }
291 if (is_in_table) {
292 // reinsert into table
293 initial_gvn()->hash_find_insert(use);
294 }
295 record_for_igvn(use);
296 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
297 i -= uses_found; // we deleted 1 or more copies of this edge
298 }
299 }
300
301
302 // Identify all nodes that are reachable from below, useful.
303 // Use breadth-first pass that records state in a Unique_Node_List,
304 // recursive traversal is slower.
305 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
306 int estimated_worklist_size = live_nodes();
307 useful.map( estimated_worklist_size, nullptr ); // preallocate space
308
309 // Initialize worklist
310 if (root() != nullptr) { useful.push(root()); }
311 // If 'top' is cached, declare it useful to preserve cached node
312 if (cached_top_node()) { useful.push(cached_top_node()); }
313
314 // Push all useful nodes onto the list, breadthfirst
315 for( uint next = 0; next < useful.size(); ++next ) {
316 assert( next < unique(), "Unique useful nodes < total nodes");
317 Node *n = useful.at(next);
318 uint max = n->len();
319 for( uint i = 0; i < max; ++i ) {
320 Node *m = n->in(i);
321 if (not_a_node(m)) continue;
322 useful.push(m);
323 }
324 }
325 }
326
327 // Update dead_node_list with any missing dead nodes using useful
328 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
329 void Compile::update_dead_node_list(Unique_Node_List &useful) {
330 uint max_idx = unique();
331 VectorSet& useful_node_set = useful.member_set();
332
333 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
334 // If node with index node_idx is not in useful set,
335 // mark it as dead in dead node list.
336 if (!useful_node_set.test(node_idx)) {
337 record_dead_node(node_idx);
338 }
339 }
340 }
341
342 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
343 int shift = 0;
344 for (int i = 0; i < inlines->length(); i++) {
345 CallGenerator* cg = inlines->at(i);
346 if (useful.member(cg->call_node())) {
347 if (shift > 0) {
348 inlines->at_put(i - shift, cg);
349 }
350 } else {
351 shift++; // skip over the dead element
352 }
353 }
354 if (shift > 0) {
355 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
356 }
357 }
358
359 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
360 assert(dead != nullptr && dead->is_Call(), "sanity");
361 int found = 0;
362 for (int i = 0; i < inlines->length(); i++) {
363 if (inlines->at(i)->call_node() == dead) {
364 inlines->remove_at(i);
365 found++;
366 NOT_DEBUG( break; ) // elements are unique, so exit early
367 }
368 }
369 assert(found <= 1, "not unique");
370 }
371
372 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
373 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
374 for (int i = node_list.length() - 1; i >= 0; i--) {
375 N* node = node_list.at(i);
376 if (!useful.member(node)) {
377 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
378 }
379 }
380 }
381
382 void Compile::remove_useless_node(Node* dead) {
383 remove_modified_node(dead);
384
385 // Constant node that has no out-edges and has only one in-edge from
386 // root is usually dead. However, sometimes reshaping walk makes
387 // it reachable by adding use edges. So, we will NOT count Con nodes
388 // as dead to be conservative about the dead node count at any
389 // given time.
390 if (!dead->is_Con()) {
391 record_dead_node(dead->_idx);
392 }
393 if (dead->is_macro()) {
394 remove_macro_node(dead);
395 }
396 if (dead->is_expensive()) {
397 remove_expensive_node(dead);
398 }
399 if (dead->is_OpaqueTemplateAssertionPredicate()) {
400 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
401 }
402 if (dead->is_ParsePredicate()) {
403 remove_parse_predicate(dead->as_ParsePredicate());
404 }
405 if (dead->for_post_loop_opts_igvn()) {
406 remove_from_post_loop_opts_igvn(dead);
407 }
408 if (dead->for_merge_stores_igvn()) {
409 remove_from_merge_stores_igvn(dead);
410 }
411 if (dead->is_Call()) {
412 remove_useless_late_inlines( &_late_inlines, dead);
413 remove_useless_late_inlines( &_string_late_inlines, dead);
414 remove_useless_late_inlines( &_boxing_late_inlines, dead);
415 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
416
417 if (dead->is_CallStaticJava()) {
418 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
419 }
420 }
421 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
422 bs->unregister_potential_barrier_node(dead);
423 }
424
425 // Disconnect all useless nodes by disconnecting those at the boundary.
426 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
427 uint next = 0;
428 while (next < useful.size()) {
429 Node *n = useful.at(next++);
430 if (n->is_SafePoint()) {
431 // We're done with a parsing phase. Replaced nodes are not valid
432 // beyond that point.
433 n->as_SafePoint()->delete_replaced_nodes();
434 }
435 // Use raw traversal of out edges since this code removes out edges
436 int max = n->outcnt();
437 for (int j = 0; j < max; ++j) {
438 Node* child = n->raw_out(j);
439 if (!useful.member(child)) {
440 assert(!child->is_top() || child != top(),
441 "If top is cached in Compile object it is in useful list");
442 // Only need to remove this out-edge to the useless node
443 n->raw_del_out(j);
444 --j;
445 --max;
446 if (child->is_data_proj_of_pure_function(n)) {
447 worklist.push(n);
448 }
449 }
450 }
451 if (n->outcnt() == 1 && n->has_special_unique_user()) {
452 assert(useful.member(n->unique_out()), "do not push a useless node");
453 worklist.push(n->unique_out());
454 }
455 }
456
457 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
458 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
459 // Remove useless Template Assertion Predicate opaque nodes
460 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
461 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
462 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
463 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
464 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
465 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
466 #ifdef ASSERT
467 if (_modified_nodes != nullptr) {
468 _modified_nodes->remove_useless_nodes(useful.member_set());
469 }
470 #endif
471
472 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
473 bs->eliminate_useless_gc_barriers(useful, this);
474 // clean up the late inline lists
475 remove_useless_late_inlines( &_late_inlines, useful);
476 remove_useless_late_inlines( &_string_late_inlines, useful);
477 remove_useless_late_inlines( &_boxing_late_inlines, useful);
478 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
479 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
480 }
481
482 // ============================================================================
483 //------------------------------CompileWrapper---------------------------------
484 class CompileWrapper : public StackObj {
485 Compile *const _compile;
486 public:
487 CompileWrapper(Compile* compile);
488
489 ~CompileWrapper();
490 };
491
492 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
493 // the Compile* pointer is stored in the current ciEnv:
494 ciEnv* env = compile->env();
495 assert(env == ciEnv::current(), "must already be a ciEnv active");
496 assert(env->compiler_data() == nullptr, "compile already active?");
497 env->set_compiler_data(compile);
498 assert(compile == Compile::current(), "sanity");
499
500 compile->set_type_dict(nullptr);
501 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
502 compile->clone_map().set_clone_idx(0);
503 compile->set_type_last_size(0);
504 compile->set_last_tf(nullptr, nullptr);
505 compile->set_indexSet_arena(nullptr);
506 compile->set_indexSet_free_block_list(nullptr);
507 compile->init_type_arena();
508 Type::Initialize(compile);
509 _compile->begin_method();
510 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
511 }
512 CompileWrapper::~CompileWrapper() {
513 // simulate crash during compilation
514 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
515
516 _compile->end_method();
517 _compile->env()->set_compiler_data(nullptr);
518 }
519
520
521 //----------------------------print_compile_messages---------------------------
522 void Compile::print_compile_messages() {
523 #ifndef PRODUCT
524 // Check if recompiling
525 if (!subsume_loads() && PrintOpto) {
526 // Recompiling without allowing machine instructions to subsume loads
527 tty->print_cr("*********************************************************");
528 tty->print_cr("** Bailout: Recompile without subsuming loads **");
529 tty->print_cr("*********************************************************");
530 }
531 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
532 // Recompiling without escape analysis
533 tty->print_cr("*********************************************************");
534 tty->print_cr("** Bailout: Recompile without escape analysis **");
535 tty->print_cr("*********************************************************");
536 }
537 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
538 // Recompiling without iterative escape analysis
539 tty->print_cr("*********************************************************");
540 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
541 tty->print_cr("*********************************************************");
542 }
543 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
544 // Recompiling without reducing allocation merges
545 tty->print_cr("*********************************************************");
546 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
547 tty->print_cr("*********************************************************");
548 }
549 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
550 // Recompiling without boxing elimination
551 tty->print_cr("*********************************************************");
552 tty->print_cr("** Bailout: Recompile without boxing elimination **");
553 tty->print_cr("*********************************************************");
554 }
555 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
556 // Recompiling without locks coarsening
557 tty->print_cr("*********************************************************");
558 tty->print_cr("** Bailout: Recompile without locks coarsening **");
559 tty->print_cr("*********************************************************");
560 }
561 if (env()->break_at_compile()) {
562 // Open the debugger when compiling this method.
563 tty->print("### Breaking when compiling: ");
564 method()->print_short_name();
565 tty->cr();
566 BREAKPOINT;
567 }
568
569 if( PrintOpto ) {
570 if (is_osr_compilation()) {
571 tty->print("[OSR]%3d", _compile_id);
572 } else {
573 tty->print("%3d", _compile_id);
574 }
575 }
576 #endif
577 }
578
579 #ifndef PRODUCT
580 void Compile::print_phase(const char* phase_name) {
581 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name);
582 }
583
584 void Compile::print_ideal_ir(const char* compile_phase_name) const {
585 // keep the following output all in one block
586 // This output goes directly to the tty, not the compiler log.
587 // To enable tools to match it up with the compilation activity,
588 // be sure to tag this tty output with the compile ID.
589
590 // Node dumping can cause a safepoint, which can break the tty lock.
591 // Buffer all node dumps, so that all safepoints happen before we lock.
592 ResourceMark rm;
593 stringStream ss;
594
595 if (_output == nullptr) {
596 ss.print_cr("AFTER: %s", compile_phase_name);
597 // Print out all nodes in ascending order of index.
598 // It is important that we traverse both inputs and outputs of nodes,
599 // so that we reach all nodes that are connected to Root.
600 root()->dump_bfs(MaxNodeLimit, nullptr, "-+S$", &ss);
601 } else {
602 // Dump the node blockwise if we have a scheduling
603 _output->print_scheduling(&ss);
604 }
605
606 // Check that the lock is not broken by a safepoint.
607 NoSafepointVerifier nsv;
608 ttyLocker ttyl;
609 if (xtty != nullptr) {
610 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
611 compile_id(),
612 is_osr_compilation() ? " compile_kind='osr'" : "",
613 compile_phase_name);
614 }
615
616 tty->print("%s", ss.as_string());
617
618 if (xtty != nullptr) {
619 xtty->tail("ideal");
620 }
621 }
622 #endif
623
624 // ============================================================================
625 //------------------------------Compile standard-------------------------------
626
627 // Compile a method. entry_bci is -1 for normal compilations and indicates
628 // the continuation bci for on stack replacement.
629
630
631 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
632 Options options, DirectiveSet* directive)
633 : Phase(Compiler),
634 _compile_id(ci_env->compile_id()),
635 _options(options),
636 _method(target),
637 _entry_bci(osr_bci),
638 _ilt(nullptr),
639 _stub_function(nullptr),
640 _stub_name(nullptr),
641 _stub_id(StubId::NO_STUBID),
642 _stub_entry_point(nullptr),
643 _max_node_limit(MaxNodeLimit),
644 _post_loop_opts_phase(false),
645 _merge_stores_phase(false),
646 _allow_macro_nodes(true),
647 _inlining_progress(false),
648 _inlining_incrementally(false),
649 _do_cleanup(false),
650 _has_reserved_stack_access(target->has_reserved_stack_access()),
651 #ifndef PRODUCT
652 _igv_idx(0),
653 _trace_opto_output(directive->TraceOptoOutputOption),
654 #endif
655 _clinit_barrier_on_entry(false),
656 _stress_seed(0),
657 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
658 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
659 _env(ci_env),
660 _directive(directive),
661 _log(ci_env->log()),
662 _first_failure_details(nullptr),
663 _intrinsics(comp_arena(), 0, 0, nullptr),
664 _macro_nodes(comp_arena(), 8, 0, nullptr),
665 _parse_predicates(comp_arena(), 8, 0, nullptr),
666 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
667 _expensive_nodes(comp_arena(), 8, 0, nullptr),
668 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
669 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
670 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
671 _coarsened_locks(comp_arena(), 8, 0, nullptr),
672 _congraph(nullptr),
673 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
674 _unique(0),
675 _dead_node_count(0),
676 _dead_node_list(comp_arena()),
677 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
678 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
679 _node_arena(&_node_arena_one),
680 _mach_constant_base_node(nullptr),
681 _Compile_types(mtCompiler, Arena::Tag::tag_type),
682 _initial_gvn(nullptr),
683 _igvn_worklist(nullptr),
684 _types(nullptr),
685 _node_hash(nullptr),
686 _late_inlines(comp_arena(), 2, 0, nullptr),
687 _string_late_inlines(comp_arena(), 2, 0, nullptr),
688 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
689 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
690 _late_inlines_pos(0),
691 _has_mh_late_inlines(false),
692 _oom(false),
693 _replay_inline_data(nullptr),
694 _inline_printer(this),
695 _java_calls(0),
696 _inner_loops(0),
697 _FIRST_STACK_mask(comp_arena()),
698 _interpreter_frame_size(0),
699 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
700 _output(nullptr)
701 #ifndef PRODUCT
702 ,
703 _in_dump_cnt(0)
704 #endif
705 {
706 C = this;
707 CompileWrapper cw(this);
708
709 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
710 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
711
712 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
713 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
714 // We can always print a disassembly, either abstract (hex dump) or
715 // with the help of a suitable hsdis library. Thus, we should not
716 // couple print_assembly and print_opto_assembly controls.
717 // But: always print opto and regular assembly on compile command 'print'.
718 bool print_assembly = directive->PrintAssemblyOption;
719 set_print_assembly(print_opto_assembly || print_assembly);
720 #else
721 set_print_assembly(false); // must initialize.
722 #endif
723
724 #ifndef PRODUCT
725 set_parsed_irreducible_loop(false);
726 #endif
727
728 if (directive->ReplayInlineOption) {
729 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
730 }
731 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
732 set_print_intrinsics(directive->PrintIntrinsicsOption);
733 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
734
735 if (ProfileTraps) {
736 // Make sure the method being compiled gets its own MDO,
737 // so we can at least track the decompile_count().
738 method()->ensure_method_data();
739 }
740
741 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
742 StressIncrementalInlining || StressMacroExpansion ||
743 StressMacroElimination || StressUnstableIfTraps ||
744 StressBailout || StressLoopPeeling) {
745 initialize_stress_seed(directive);
746 }
747
748 Init(/*do_aliasing=*/ true);
749
750 print_compile_messages();
751
752 _ilt = InlineTree::build_inline_tree_root();
753
754 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
755 assert(num_alias_types() >= AliasIdxRaw, "");
756
757 #define MINIMUM_NODE_HASH 1023
758
759 // GVN that will be run immediately on new nodes
760 uint estimated_size = method()->code_size()*4+64;
761 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
762 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
763 _types = new (comp_arena()) Type_Array(comp_arena());
764 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
765 PhaseGVN gvn;
766 set_initial_gvn(&gvn);
767
768 { // Scope for timing the parser
769 TracePhase tp(_t_parser);
770
771 // Put top into the hash table ASAP.
772 initial_gvn()->transform(top());
773
774 // Set up tf(), start(), and find a CallGenerator.
775 CallGenerator* cg = nullptr;
776 if (is_osr_compilation()) {
777 const TypeTuple *domain = StartOSRNode::osr_domain();
778 const TypeTuple *range = TypeTuple::make_range(method()->signature());
779 init_tf(TypeFunc::make(domain, range));
780 StartNode* s = new StartOSRNode(root(), domain);
781 initial_gvn()->set_type_bottom(s);
782 verify_start(s);
783 cg = CallGenerator::for_osr(method(), entry_bci());
784 } else {
785 // Normal case.
786 init_tf(TypeFunc::make(method()));
787 StartNode* s = new StartNode(root(), tf()->domain());
788 initial_gvn()->set_type_bottom(s);
789 verify_start(s);
790 float past_uses = method()->interpreter_invocation_count();
791 float expected_uses = past_uses;
792 cg = CallGenerator::for_inline(method(), expected_uses);
793 }
794 if (failing()) return;
795 if (cg == nullptr) {
796 const char* reason = InlineTree::check_can_parse(method());
797 assert(reason != nullptr, "expect reason for parse failure");
798 stringStream ss;
799 ss.print("cannot parse method: %s", reason);
800 record_method_not_compilable(ss.as_string());
801 return;
802 }
803
804 gvn.set_type(root(), root()->bottom_type());
805
806 JVMState* jvms = build_start_state(start(), tf());
807 if ((jvms = cg->generate(jvms)) == nullptr) {
808 assert(failure_reason() != nullptr, "expect reason for parse failure");
809 stringStream ss;
810 ss.print("method parse failed: %s", failure_reason());
811 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
812 return;
813 }
814 GraphKit kit(jvms);
815
816 if (!kit.stopped()) {
817 // Accept return values, and transfer control we know not where.
818 // This is done by a special, unique ReturnNode bound to root.
819 return_values(kit.jvms());
820 }
821
822 if (kit.has_exceptions()) {
823 // Any exceptions that escape from this call must be rethrown
824 // to whatever caller is dynamically above us on the stack.
825 // This is done by a special, unique RethrowNode bound to root.
826 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
827 }
828
829 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
830
831 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
832 inline_string_calls(true);
833 }
834
835 if (failing()) return;
836
837 // Remove clutter produced by parsing.
838 if (!failing()) {
839 ResourceMark rm;
840 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
841 }
842 }
843
844 // Note: Large methods are capped off in do_one_bytecode().
845 if (failing()) return;
846
847 // After parsing, node notes are no longer automagic.
848 // They must be propagated by register_new_node_with_optimizer(),
849 // clone(), or the like.
850 set_default_node_notes(nullptr);
851
852 #ifndef PRODUCT
853 if (should_print_igv(1)) {
854 _igv_printer->print_inlining();
855 }
856 #endif
857
858 if (failing()) return;
859 NOT_PRODUCT( verify_graph_edges(); )
860
861 // Now optimize
862 Optimize();
863 if (failing()) return;
864 NOT_PRODUCT( verify_graph_edges(); )
865
866 #ifndef PRODUCT
867 if (should_print_ideal()) {
868 print_ideal_ir("PrintIdeal");
869 }
870 #endif
871
872 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
873
874 #ifdef ASSERT
875 bs->verify_gc_barriers(this, BarrierSetC2::AfterOptimize);
876 #endif
877
878 bs->final_refinement(this);
879
880 #ifdef ASSERT
881 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
882 #endif
883
884 // Dump compilation data to replay it.
885 if (directive->DumpReplayOption) {
886 env()->dump_replay_data(_compile_id);
887 }
888 if (directive->DumpInlineOption && (ilt() != nullptr)) {
889 env()->dump_inline_data(_compile_id);
890 }
891
892 // Now that we know the size of all the monitors we can add a fixed slot
893 // for the original deopt pc.
894 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
895 set_fixed_slots(next_slot);
896
897 // Compute when to use implicit null checks. Used by matching trap based
898 // nodes and NullCheck optimization.
899 set_allowed_deopt_reasons();
900
901 // Now generate code
902 Code_Gen();
903 }
904
905 //------------------------------Compile----------------------------------------
906 // Compile a runtime stub
907 Compile::Compile(ciEnv* ci_env,
908 TypeFunc_generator generator,
909 address stub_function,
910 const char* stub_name,
911 StubId stub_id,
912 int is_fancy_jump,
913 bool pass_tls,
914 bool return_pc,
915 DirectiveSet* directive)
916 : Phase(Compiler),
917 _compile_id(0),
918 _options(Options::for_runtime_stub()),
919 _method(nullptr),
920 _entry_bci(InvocationEntryBci),
921 _stub_function(stub_function),
922 _stub_name(stub_name),
923 _stub_id(stub_id),
924 _stub_entry_point(nullptr),
925 _max_node_limit(MaxNodeLimit),
926 _post_loop_opts_phase(false),
927 _merge_stores_phase(false),
928 _allow_macro_nodes(true),
929 _inlining_progress(false),
930 _inlining_incrementally(false),
931 _has_reserved_stack_access(false),
932 #ifndef PRODUCT
933 _igv_idx(0),
934 _trace_opto_output(directive->TraceOptoOutputOption),
935 #endif
936 _clinit_barrier_on_entry(false),
937 _stress_seed(0),
938 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
939 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
940 _env(ci_env),
941 _directive(directive),
942 _log(ci_env->log()),
943 _first_failure_details(nullptr),
944 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
945 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
946 _congraph(nullptr),
947 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
948 _unique(0),
949 _dead_node_count(0),
950 _dead_node_list(comp_arena()),
951 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
952 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
953 _node_arena(&_node_arena_one),
954 _mach_constant_base_node(nullptr),
955 _Compile_types(mtCompiler, Arena::Tag::tag_type),
956 _initial_gvn(nullptr),
957 _igvn_worklist(nullptr),
958 _types(nullptr),
959 _node_hash(nullptr),
960 _has_mh_late_inlines(false),
961 _oom(false),
962 _replay_inline_data(nullptr),
963 _inline_printer(this),
964 _java_calls(0),
965 _inner_loops(0),
966 _FIRST_STACK_mask(comp_arena()),
967 _interpreter_frame_size(0),
968 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
969 _output(nullptr),
970 #ifndef PRODUCT
971 _in_dump_cnt(0),
972 #endif
973 _allowed_reasons(0) {
974 C = this;
975
976 // try to reuse an existing stub
977 {
978 BlobId blob_id = StubInfo::blob(_stub_id);
979 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
980 if (blob != nullptr) {
981 RuntimeStub* rs = blob->as_runtime_stub();
982 _stub_entry_point = rs->entry_point();
983 return;
984 }
985 }
986
987 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
988 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
989
990 #ifndef PRODUCT
991 set_print_assembly(PrintFrameConverterAssembly);
992 set_parsed_irreducible_loop(false);
993 #else
994 set_print_assembly(false); // Must initialize.
995 #endif
996 set_has_irreducible_loop(false); // no loops
997
998 CompileWrapper cw(this);
999 Init(/*do_aliasing=*/ false);
1000 init_tf((*generator)());
1001
1002 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1003 _types = new (comp_arena()) Type_Array(comp_arena());
1004 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1005
1006 if (StressLCM || StressGCM || StressBailout) {
1007 initialize_stress_seed(directive);
1008 }
1009
1010 {
1011 PhaseGVN gvn;
1012 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1013 gvn.transform(top());
1014
1015 GraphKit kit;
1016 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1017 }
1018
1019 NOT_PRODUCT( verify_graph_edges(); )
1020
1021 Code_Gen();
1022 }
1023
1024 Compile::~Compile() {
1025 delete _first_failure_details;
1026 };
1027
1028 //------------------------------Init-------------------------------------------
1029 // Prepare for a single compilation
1030 void Compile::Init(bool aliasing) {
1031 _do_aliasing = aliasing;
1032 _unique = 0;
1033 _regalloc = nullptr;
1034
1035 _tf = nullptr; // filled in later
1036 _top = nullptr; // cached later
1037 _matcher = nullptr; // filled in later
1038 _cfg = nullptr; // filled in later
1039
1040 _node_note_array = nullptr;
1041 _default_node_notes = nullptr;
1042 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1043
1044 _immutable_memory = nullptr; // filled in at first inquiry
1045
1046 #ifdef ASSERT
1047 _phase_optimize_finished = false;
1048 _phase_verify_ideal_loop = false;
1049 _exception_backedge = false;
1050 _type_verify = nullptr;
1051 #endif
1052
1053 // Globally visible Nodes
1054 // First set TOP to null to give safe behavior during creation of RootNode
1055 set_cached_top_node(nullptr);
1056 set_root(new RootNode());
1057 // Now that you have a Root to point to, create the real TOP
1058 set_cached_top_node( new ConNode(Type::TOP) );
1059 set_recent_alloc(nullptr, nullptr);
1060
1061 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1062 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1063 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1064 env()->set_dependencies(new Dependencies(env()));
1065
1066 _fixed_slots = 0;
1067 set_has_split_ifs(false);
1068 set_has_loops(false); // first approximation
1069 set_has_stringbuilder(false);
1070 set_has_boxed_value(false);
1071 _trap_can_recompile = false; // no traps emitted yet
1072 _major_progress = true; // start out assuming good things will happen
1073 set_has_unsafe_access(false);
1074 set_max_vector_size(0);
1075 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1076 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1077 set_decompile_count(0);
1078
1079 #ifndef PRODUCT
1080 _phase_counter = 0;
1081 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1082 #endif
1083
1084 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1085 _loop_opts_cnt = LoopOptsCount;
1086 set_do_inlining(Inline);
1087 set_max_inline_size(MaxInlineSize);
1088 set_freq_inline_size(FreqInlineSize);
1089 set_do_scheduling(OptoScheduling);
1090
1091 set_do_vector_loop(false);
1092 set_has_monitors(false);
1093 set_has_scoped_access(false);
1094
1095 if (AllowVectorizeOnDemand) {
1096 if (has_method() && _directive->VectorizeOption) {
1097 set_do_vector_loop(true);
1098 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());})
1099 } else if (has_method() && method()->name() != nullptr &&
1100 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1101 set_do_vector_loop(true);
1102 }
1103 }
1104 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1105 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());})
1106
1107 _max_node_limit = _directive->MaxNodeLimitOption;
1108
1109 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1110 set_clinit_barrier_on_entry(true);
1111 }
1112 if (debug_info()->recording_non_safepoints()) {
1113 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1114 (comp_arena(), 8, 0, nullptr));
1115 set_default_node_notes(Node_Notes::make(this));
1116 }
1117
1118 const int grow_ats = 16;
1119 _max_alias_types = grow_ats;
1120 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1121 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1122 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1123 {
1124 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1125 }
1126 // Initialize the first few types.
1127 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1128 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1129 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1130 _num_alias_types = AliasIdxRaw+1;
1131 // Zero out the alias type cache.
1132 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1133 // A null adr_type hits in the cache right away. Preload the right answer.
1134 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1135 }
1136
1137 #ifdef ASSERT
1138 // Verify that the current StartNode is valid.
1139 void Compile::verify_start(StartNode* s) const {
1140 assert(failing_internal() || s == start(), "should be StartNode");
1141 }
1142 #endif
1143
1144 /**
1145 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1146 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1147 * the ideal graph.
1148 */
1149 StartNode* Compile::start() const {
1150 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1151 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1152 Node* start = root()->fast_out(i);
1153 if (start->is_Start()) {
1154 return start->as_Start();
1155 }
1156 }
1157 fatal("Did not find Start node!");
1158 return nullptr;
1159 }
1160
1161 //-------------------------------immutable_memory-------------------------------------
1162 // Access immutable memory
1163 Node* Compile::immutable_memory() {
1164 if (_immutable_memory != nullptr) {
1165 return _immutable_memory;
1166 }
1167 StartNode* s = start();
1168 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1169 Node *p = s->fast_out(i);
1170 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1171 _immutable_memory = p;
1172 return _immutable_memory;
1173 }
1174 }
1175 ShouldNotReachHere();
1176 return nullptr;
1177 }
1178
1179 //----------------------set_cached_top_node------------------------------------
1180 // Install the cached top node, and make sure Node::is_top works correctly.
1181 void Compile::set_cached_top_node(Node* tn) {
1182 if (tn != nullptr) verify_top(tn);
1183 Node* old_top = _top;
1184 _top = tn;
1185 // Calling Node::setup_is_top allows the nodes the chance to adjust
1186 // their _out arrays.
1187 if (_top != nullptr) _top->setup_is_top();
1188 if (old_top != nullptr) old_top->setup_is_top();
1189 assert(_top == nullptr || top()->is_top(), "");
1190 }
1191
1192 #ifdef ASSERT
1193 uint Compile::count_live_nodes_by_graph_walk() {
1194 Unique_Node_List useful(comp_arena());
1195 // Get useful node list by walking the graph.
1196 identify_useful_nodes(useful);
1197 return useful.size();
1198 }
1199
1200 void Compile::print_missing_nodes() {
1201
1202 // Return if CompileLog is null and PrintIdealNodeCount is false.
1203 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1204 return;
1205 }
1206
1207 // This is an expensive function. It is executed only when the user
1208 // specifies VerifyIdealNodeCount option or otherwise knows the
1209 // additional work that needs to be done to identify reachable nodes
1210 // by walking the flow graph and find the missing ones using
1211 // _dead_node_list.
1212
1213 Unique_Node_List useful(comp_arena());
1214 // Get useful node list by walking the graph.
1215 identify_useful_nodes(useful);
1216
1217 uint l_nodes = C->live_nodes();
1218 uint l_nodes_by_walk = useful.size();
1219
1220 if (l_nodes != l_nodes_by_walk) {
1221 if (_log != nullptr) {
1222 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1223 _log->stamp();
1224 _log->end_head();
1225 }
1226 VectorSet& useful_member_set = useful.member_set();
1227 int last_idx = l_nodes_by_walk;
1228 for (int i = 0; i < last_idx; i++) {
1229 if (useful_member_set.test(i)) {
1230 if (_dead_node_list.test(i)) {
1231 if (_log != nullptr) {
1232 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1233 }
1234 if (PrintIdealNodeCount) {
1235 // Print the log message to tty
1236 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1237 useful.at(i)->dump();
1238 }
1239 }
1240 }
1241 else if (! _dead_node_list.test(i)) {
1242 if (_log != nullptr) {
1243 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1244 }
1245 if (PrintIdealNodeCount) {
1246 // Print the log message to tty
1247 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1248 }
1249 }
1250 }
1251 if (_log != nullptr) {
1252 _log->tail("mismatched_nodes");
1253 }
1254 }
1255 }
1256 void Compile::record_modified_node(Node* n) {
1257 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1258 _modified_nodes->push(n);
1259 }
1260 }
1261
1262 void Compile::remove_modified_node(Node* n) {
1263 if (_modified_nodes != nullptr) {
1264 _modified_nodes->remove(n);
1265 }
1266 }
1267 #endif
1268
1269 #ifndef PRODUCT
1270 void Compile::verify_top(Node* tn) const {
1271 if (tn != nullptr) {
1272 assert(tn->is_Con(), "top node must be a constant");
1273 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1274 assert(tn->in(0) != nullptr, "must have live top node");
1275 }
1276 }
1277 #endif
1278
1279
1280 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1281
1282 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1283 guarantee(arr != nullptr, "");
1284 int num_blocks = arr->length();
1285 if (grow_by < num_blocks) grow_by = num_blocks;
1286 int num_notes = grow_by * _node_notes_block_size;
1287 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1288 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1289 while (num_notes > 0) {
1290 arr->append(notes);
1291 notes += _node_notes_block_size;
1292 num_notes -= _node_notes_block_size;
1293 }
1294 assert(num_notes == 0, "exact multiple, please");
1295 }
1296
1297 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1298 if (source == nullptr || dest == nullptr) return false;
1299
1300 if (dest->is_Con())
1301 return false; // Do not push debug info onto constants.
1302
1303 #ifdef ASSERT
1304 // Leave a bread crumb trail pointing to the original node:
1305 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1306 dest->set_debug_orig(source);
1307 }
1308 #endif
1309
1310 if (node_note_array() == nullptr)
1311 return false; // Not collecting any notes now.
1312
1313 // This is a copy onto a pre-existing node, which may already have notes.
1314 // If both nodes have notes, do not overwrite any pre-existing notes.
1315 Node_Notes* source_notes = node_notes_at(source->_idx);
1316 if (source_notes == nullptr || source_notes->is_clear()) return false;
1317 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1318 if (dest_notes == nullptr || dest_notes->is_clear()) {
1319 return set_node_notes_at(dest->_idx, source_notes);
1320 }
1321
1322 Node_Notes merged_notes = (*source_notes);
1323 // The order of operations here ensures that dest notes will win...
1324 merged_notes.update_from(dest_notes);
1325 return set_node_notes_at(dest->_idx, &merged_notes);
1326 }
1327
1328
1329 //--------------------------allow_range_check_smearing-------------------------
1330 // Gating condition for coalescing similar range checks.
1331 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1332 // single covering check that is at least as strong as any of them.
1333 // If the optimization succeeds, the simplified (strengthened) range check
1334 // will always succeed. If it fails, we will deopt, and then give up
1335 // on the optimization.
1336 bool Compile::allow_range_check_smearing() const {
1337 // If this method has already thrown a range-check,
1338 // assume it was because we already tried range smearing
1339 // and it failed.
1340 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1341 return !already_trapped;
1342 }
1343
1344
1345 //------------------------------flatten_alias_type-----------------------------
1346 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1347 assert(do_aliasing(), "Aliasing should be enabled");
1348 int offset = tj->offset();
1349 TypePtr::PTR ptr = tj->ptr();
1350
1351 // Known instance (scalarizable allocation) alias only with itself.
1352 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1353 tj->is_oopptr()->is_known_instance();
1354
1355 // Process weird unsafe references.
1356 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1357 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1358 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1359 tj = TypeOopPtr::BOTTOM;
1360 ptr = tj->ptr();
1361 offset = tj->offset();
1362 }
1363
1364 // Array pointers need some flattening
1365 const TypeAryPtr* ta = tj->isa_aryptr();
1366 if (ta && ta->is_stable()) {
1367 // Erase stability property for alias analysis.
1368 tj = ta = ta->cast_to_stable(false);
1369 }
1370 if( ta && is_known_inst ) {
1371 if ( offset != Type::OffsetBot &&
1372 offset > arrayOopDesc::length_offset_in_bytes() ) {
1373 offset = Type::OffsetBot; // Flatten constant access into array body only
1374 tj = ta = ta->
1375 remove_speculative()->
1376 cast_to_ptr_type(ptr)->
1377 with_offset(offset);
1378 }
1379 } else if (ta) {
1380 // For arrays indexed by constant indices, we flatten the alias
1381 // space to include all of the array body. Only the header, klass
1382 // and array length can be accessed un-aliased.
1383 if( offset != Type::OffsetBot ) {
1384 if( ta->const_oop() ) { // MethodData* or Method*
1385 offset = Type::OffsetBot; // Flatten constant access into array body
1386 tj = ta = ta->
1387 remove_speculative()->
1388 cast_to_ptr_type(ptr)->
1389 cast_to_exactness(false)->
1390 with_offset(offset);
1391 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1392 // range is OK as-is.
1393 tj = ta = TypeAryPtr::RANGE;
1394 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1395 tj = TypeInstPtr::KLASS; // all klass loads look alike
1396 ta = TypeAryPtr::RANGE; // generic ignored junk
1397 ptr = TypePtr::BotPTR;
1398 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1399 tj = TypeInstPtr::MARK;
1400 ta = TypeAryPtr::RANGE; // generic ignored junk
1401 ptr = TypePtr::BotPTR;
1402 } else { // Random constant offset into array body
1403 offset = Type::OffsetBot; // Flatten constant access into array body
1404 tj = ta = ta->
1405 remove_speculative()->
1406 cast_to_ptr_type(ptr)->
1407 cast_to_exactness(false)->
1408 with_offset(offset);
1409 }
1410 }
1411 // Arrays of fixed size alias with arrays of unknown size.
1412 if (ta->size() != TypeInt::POS) {
1413 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1414 tj = ta = ta->
1415 remove_speculative()->
1416 cast_to_ptr_type(ptr)->
1417 with_ary(tary)->
1418 cast_to_exactness(false);
1419 }
1420 // Arrays of known objects become arrays of unknown objects.
1421 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1422 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1423 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1424 }
1425 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1426 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1427 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1428 }
1429 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1430 // cannot be distinguished by bytecode alone.
1431 if (ta->elem() == TypeInt::BOOL) {
1432 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1433 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1434 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1435 }
1436 // During the 2nd round of IterGVN, NotNull castings are removed.
1437 // Make sure the Bottom and NotNull variants alias the same.
1438 // Also, make sure exact and non-exact variants alias the same.
1439 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1440 tj = ta = ta->
1441 remove_speculative()->
1442 cast_to_ptr_type(TypePtr::BotPTR)->
1443 cast_to_exactness(false)->
1444 with_offset(offset);
1445 }
1446 }
1447
1448 // Oop pointers need some flattening
1449 const TypeInstPtr *to = tj->isa_instptr();
1450 if (to && to != TypeOopPtr::BOTTOM) {
1451 ciInstanceKlass* ik = to->instance_klass();
1452 if( ptr == TypePtr::Constant ) {
1453 if (ik != ciEnv::current()->Class_klass() ||
1454 offset < ik->layout_helper_size_in_bytes()) {
1455 // No constant oop pointers (such as Strings); they alias with
1456 // unknown strings.
1457 assert(!is_known_inst, "not scalarizable allocation");
1458 tj = to = to->
1459 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1460 remove_speculative()->
1461 cast_to_ptr_type(TypePtr::BotPTR)->
1462 cast_to_exactness(false);
1463 }
1464 } else if( is_known_inst ) {
1465 tj = to; // Keep NotNull and klass_is_exact for instance type
1466 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1467 // During the 2nd round of IterGVN, NotNull castings are removed.
1468 // Make sure the Bottom and NotNull variants alias the same.
1469 // Also, make sure exact and non-exact variants alias the same.
1470 tj = to = to->
1471 remove_speculative()->
1472 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1473 cast_to_ptr_type(TypePtr::BotPTR)->
1474 cast_to_exactness(false);
1475 }
1476 if (to->speculative() != nullptr) {
1477 tj = to = to->remove_speculative();
1478 }
1479 // Canonicalize the holder of this field
1480 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1481 // First handle header references such as a LoadKlassNode, even if the
1482 // object's klass is unloaded at compile time (4965979).
1483 if (!is_known_inst) { // Do it only for non-instance types
1484 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1485 }
1486 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1487 // Static fields are in the space above the normal instance
1488 // fields in the java.lang.Class instance.
1489 if (ik != ciEnv::current()->Class_klass()) {
1490 to = nullptr;
1491 tj = TypeOopPtr::BOTTOM;
1492 offset = tj->offset();
1493 }
1494 } else {
1495 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1496 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1497 assert(tj->offset() == offset, "no change to offset expected");
1498 bool xk = to->klass_is_exact();
1499 int instance_id = to->instance_id();
1500
1501 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1502 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1503 // its interfaces are included.
1504 if (xk && ik->equals(canonical_holder)) {
1505 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1506 } else {
1507 assert(xk || !is_known_inst, "Known instance should be exact type");
1508 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1509 }
1510 }
1511 }
1512
1513 // Klass pointers to object array klasses need some flattening
1514 const TypeKlassPtr *tk = tj->isa_klassptr();
1515 if( tk ) {
1516 // If we are referencing a field within a Klass, we need
1517 // to assume the worst case of an Object. Both exact and
1518 // inexact types must flatten to the same alias class so
1519 // use NotNull as the PTR.
1520 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1521 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1522 env()->Object_klass(),
1523 offset);
1524 }
1525
1526 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1527 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1528 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1529 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1530 } else {
1531 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1532 }
1533 }
1534
1535 // Check for precise loads from the primary supertype array and force them
1536 // to the supertype cache alias index. Check for generic array loads from
1537 // the primary supertype array and also force them to the supertype cache
1538 // alias index. Since the same load can reach both, we need to merge
1539 // these 2 disparate memories into the same alias class. Since the
1540 // primary supertype array is read-only, there's no chance of confusion
1541 // where we bypass an array load and an array store.
1542 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1543 if (offset == Type::OffsetBot ||
1544 (offset >= primary_supers_offset &&
1545 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1546 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1547 offset = in_bytes(Klass::secondary_super_cache_offset());
1548 tj = tk = tk->with_offset(offset);
1549 }
1550 }
1551
1552 // Flatten all Raw pointers together.
1553 if (tj->base() == Type::RawPtr)
1554 tj = TypeRawPtr::BOTTOM;
1555
1556 if (tj->base() == Type::AnyPtr)
1557 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1558
1559 offset = tj->offset();
1560 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1561
1562 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1563 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1564 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1565 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1566 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1567 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1568 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1569 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1570 assert( tj->ptr() != TypePtr::TopPTR &&
1571 tj->ptr() != TypePtr::AnyNull &&
1572 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1573 // assert( tj->ptr() != TypePtr::Constant ||
1574 // tj->base() == Type::RawPtr ||
1575 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1576
1577 return tj;
1578 }
1579
1580 void Compile::AliasType::Init(int i, const TypePtr* at) {
1581 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1582 _index = i;
1583 _adr_type = at;
1584 _field = nullptr;
1585 _element = nullptr;
1586 _is_rewritable = true; // default
1587 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1588 if (atoop != nullptr && atoop->is_known_instance()) {
1589 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1590 _general_index = Compile::current()->get_alias_index(gt);
1591 } else {
1592 _general_index = 0;
1593 }
1594 }
1595
1596 BasicType Compile::AliasType::basic_type() const {
1597 if (element() != nullptr) {
1598 const Type* element = adr_type()->is_aryptr()->elem();
1599 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1600 } if (field() != nullptr) {
1601 return field()->layout_type();
1602 } else {
1603 return T_ILLEGAL; // unknown
1604 }
1605 }
1606
1607 //---------------------------------print_on------------------------------------
1608 #ifndef PRODUCT
1609 void Compile::AliasType::print_on(outputStream* st) {
1610 if (index() < 10)
1611 st->print("@ <%d> ", index());
1612 else st->print("@ <%d>", index());
1613 st->print(is_rewritable() ? " " : " RO");
1614 int offset = adr_type()->offset();
1615 if (offset == Type::OffsetBot)
1616 st->print(" +any");
1617 else st->print(" +%-3d", offset);
1618 st->print(" in ");
1619 adr_type()->dump_on(st);
1620 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1621 if (field() != nullptr && tjp) {
1622 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1623 tjp->offset() != field()->offset_in_bytes()) {
1624 st->print(" != ");
1625 field()->print();
1626 st->print(" ***");
1627 }
1628 }
1629 }
1630
1631 void print_alias_types() {
1632 Compile* C = Compile::current();
1633 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1634 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1635 C->alias_type(idx)->print_on(tty);
1636 tty->cr();
1637 }
1638 }
1639 #endif
1640
1641
1642 //----------------------------probe_alias_cache--------------------------------
1643 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1644 intptr_t key = (intptr_t) adr_type;
1645 key ^= key >> logAliasCacheSize;
1646 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1647 }
1648
1649
1650 //-----------------------------grow_alias_types--------------------------------
1651 void Compile::grow_alias_types() {
1652 const int old_ats = _max_alias_types; // how many before?
1653 const int new_ats = old_ats; // how many more?
1654 const int grow_ats = old_ats+new_ats; // how many now?
1655 _max_alias_types = grow_ats;
1656 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1657 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1658 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1659 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1660 }
1661
1662
1663 //--------------------------------find_alias_type------------------------------
1664 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1665 if (!do_aliasing()) {
1666 return alias_type(AliasIdxBot);
1667 }
1668
1669 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1670 if (ace->_adr_type == adr_type) {
1671 return alias_type(ace->_index);
1672 }
1673
1674 // Handle special cases.
1675 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1676 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1677
1678 // Do it the slow way.
1679 const TypePtr* flat = flatten_alias_type(adr_type);
1680
1681 #ifdef ASSERT
1682 {
1683 ResourceMark rm;
1684 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1685 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1686 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1687 Type::str(adr_type));
1688 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1689 const TypeOopPtr* foop = flat->is_oopptr();
1690 // Scalarizable allocations have exact klass always.
1691 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1692 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1693 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1694 Type::str(foop), Type::str(xoop));
1695 }
1696 }
1697 #endif
1698
1699 int idx = AliasIdxTop;
1700 for (int i = 0; i < num_alias_types(); i++) {
1701 if (alias_type(i)->adr_type() == flat) {
1702 idx = i;
1703 break;
1704 }
1705 }
1706
1707 if (idx == AliasIdxTop) {
1708 if (no_create) return nullptr;
1709 // Grow the array if necessary.
1710 if (_num_alias_types == _max_alias_types) grow_alias_types();
1711 // Add a new alias type.
1712 idx = _num_alias_types++;
1713 _alias_types[idx]->Init(idx, flat);
1714 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1715 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1716 if (flat->isa_instptr()) {
1717 if (flat->offset() == java_lang_Class::klass_offset()
1718 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1719 alias_type(idx)->set_rewritable(false);
1720 }
1721 if (flat->isa_aryptr()) {
1722 #ifdef ASSERT
1723 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1724 // (T_BYTE has the weakest alignment and size restrictions...)
1725 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1726 #endif
1727 if (flat->offset() == TypePtr::OffsetBot) {
1728 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1729 }
1730 }
1731 if (flat->isa_klassptr()) {
1732 if (UseCompactObjectHeaders) {
1733 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1734 alias_type(idx)->set_rewritable(false);
1735 }
1736 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1737 alias_type(idx)->set_rewritable(false);
1738 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1739 alias_type(idx)->set_rewritable(false);
1740 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1741 alias_type(idx)->set_rewritable(false);
1742 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1743 alias_type(idx)->set_rewritable(false);
1744 }
1745
1746 if (flat->isa_instklassptr()) {
1747 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1748 alias_type(idx)->set_rewritable(false);
1749 }
1750 }
1751 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1752 // but the base pointer type is not distinctive enough to identify
1753 // references into JavaThread.)
1754
1755 // Check for final fields.
1756 const TypeInstPtr* tinst = flat->isa_instptr();
1757 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1758 ciField* field;
1759 if (tinst->const_oop() != nullptr &&
1760 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1761 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1762 // static field
1763 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1764 field = k->get_field_by_offset(tinst->offset(), true);
1765 } else {
1766 ciInstanceKlass *k = tinst->instance_klass();
1767 field = k->get_field_by_offset(tinst->offset(), false);
1768 }
1769 assert(field == nullptr ||
1770 original_field == nullptr ||
1771 (field->holder() == original_field->holder() &&
1772 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1773 field->is_static() == original_field->is_static()), "wrong field?");
1774 // Set field() and is_rewritable() attributes.
1775 if (field != nullptr) alias_type(idx)->set_field(field);
1776 }
1777 }
1778
1779 // Fill the cache for next time.
1780 ace->_adr_type = adr_type;
1781 ace->_index = idx;
1782 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1783
1784 // Might as well try to fill the cache for the flattened version, too.
1785 AliasCacheEntry* face = probe_alias_cache(flat);
1786 if (face->_adr_type == nullptr) {
1787 face->_adr_type = flat;
1788 face->_index = idx;
1789 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1790 }
1791
1792 return alias_type(idx);
1793 }
1794
1795
1796 Compile::AliasType* Compile::alias_type(ciField* field) {
1797 const TypeOopPtr* t;
1798 if (field->is_static())
1799 t = TypeInstPtr::make(field->holder()->java_mirror());
1800 else
1801 t = TypeOopPtr::make_from_klass_raw(field->holder());
1802 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1803 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1804 return atp;
1805 }
1806
1807
1808 //------------------------------have_alias_type--------------------------------
1809 bool Compile::have_alias_type(const TypePtr* adr_type) {
1810 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1811 if (ace->_adr_type == adr_type) {
1812 return true;
1813 }
1814
1815 // Handle special cases.
1816 if (adr_type == nullptr) return true;
1817 if (adr_type == TypePtr::BOTTOM) return true;
1818
1819 return find_alias_type(adr_type, true, nullptr) != nullptr;
1820 }
1821
1822 //-----------------------------must_alias--------------------------------------
1823 // True if all values of the given address type are in the given alias category.
1824 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1825 if (alias_idx == AliasIdxBot) return true; // the universal category
1826 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1827 if (alias_idx == AliasIdxTop) return false; // the empty category
1828 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1829
1830 // the only remaining possible overlap is identity
1831 int adr_idx = get_alias_index(adr_type);
1832 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1833 assert(adr_idx == alias_idx ||
1834 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1835 && adr_type != TypeOopPtr::BOTTOM),
1836 "should not be testing for overlap with an unsafe pointer");
1837 return adr_idx == alias_idx;
1838 }
1839
1840 //------------------------------can_alias--------------------------------------
1841 // True if any values of the given address type are in the given alias category.
1842 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1843 if (alias_idx == AliasIdxTop) return false; // the empty category
1844 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1845 // Known instance doesn't alias with bottom memory
1846 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1847 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1848
1849 // the only remaining possible overlap is identity
1850 int adr_idx = get_alias_index(adr_type);
1851 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1852 return adr_idx == alias_idx;
1853 }
1854
1855 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1856 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1857 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1858 if (parse_predicate_count() == 0) {
1859 return;
1860 }
1861 for (int i = 0; i < parse_predicate_count(); i++) {
1862 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1863 parse_predicate->mark_useless(igvn);
1864 }
1865 _parse_predicates.clear();
1866 }
1867
1868 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1869 if (!n->for_post_loop_opts_igvn()) {
1870 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1871 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1872 _for_post_loop_igvn.append(n);
1873 }
1874 }
1875
1876 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1877 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1878 _for_post_loop_igvn.remove(n);
1879 }
1880
1881 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1882 // Verify that all previous optimizations produced a valid graph
1883 // at least to this point, even if no loop optimizations were done.
1884 PhaseIdealLoop::verify(igvn);
1885
1886 if (_print_phase_loop_opts) {
1887 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1888 }
1889 C->set_post_loop_opts_phase(); // no more loop opts allowed
1890
1891 assert(!C->major_progress(), "not cleared");
1892
1893 if (_for_post_loop_igvn.length() > 0) {
1894 while (_for_post_loop_igvn.length() > 0) {
1895 Node* n = _for_post_loop_igvn.pop();
1896 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1897 igvn._worklist.push(n);
1898 }
1899 igvn.optimize();
1900 if (failing()) return;
1901 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1902 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1903
1904 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1905 if (C->major_progress()) {
1906 C->clear_major_progress(); // ensure that major progress is now clear
1907 }
1908 }
1909 }
1910
1911 void Compile::record_for_merge_stores_igvn(Node* n) {
1912 if (!n->for_merge_stores_igvn()) {
1913 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1914 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1915 _for_merge_stores_igvn.append(n);
1916 }
1917 }
1918
1919 void Compile::remove_from_merge_stores_igvn(Node* n) {
1920 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1921 _for_merge_stores_igvn.remove(n);
1922 }
1923
1924 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1925 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1926 // the stores, and we merge the wrong sequence of stores.
1927 // Example:
1928 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1929 // Apply MergeStores:
1930 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1931 // Remove more RangeChecks:
1932 // StoreI [ StoreL ] StoreI
1933 // But now it would have been better to do this instead:
1934 // [ StoreL ] [ StoreL ]
1935 //
1936 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1937 // since we never unset _merge_stores_phase.
1938 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1939 C->set_merge_stores_phase();
1940
1941 if (_for_merge_stores_igvn.length() > 0) {
1942 while (_for_merge_stores_igvn.length() > 0) {
1943 Node* n = _for_merge_stores_igvn.pop();
1944 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1945 igvn._worklist.push(n);
1946 }
1947 igvn.optimize();
1948 if (failing()) return;
1949 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1950 print_method(PHASE_AFTER_MERGE_STORES, 3);
1951 }
1952 }
1953
1954 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1955 if (OptimizeUnstableIf) {
1956 _unstable_if_traps.append(trap);
1957 }
1958 }
1959
1960 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1961 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1962 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1963 Node* n = trap->uncommon_trap();
1964 if (!useful.member(n)) {
1965 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1966 }
1967 }
1968 }
1969
1970 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1971 // or fold-compares case. Return true if succeed or not found.
1972 //
1973 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1974 // false and ask fold-compares to yield.
1975 //
1976 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1977 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1978 // when deoptimization does happen.
1979 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1980 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1981 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1982 if (trap->uncommon_trap() == unc) {
1983 if (yield && trap->modified()) {
1984 return false;
1985 }
1986 _unstable_if_traps.delete_at(i);
1987 break;
1988 }
1989 }
1990 return true;
1991 }
1992
1993 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1994 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1995 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1996 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1997 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1998 CallStaticJavaNode* unc = trap->uncommon_trap();
1999 int next_bci = trap->next_bci();
2000 bool modified = trap->modified();
2001
2002 if (next_bci != -1 && !modified) {
2003 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2004 JVMState* jvms = unc->jvms();
2005 ciMethod* method = jvms->method();
2006 ciBytecodeStream iter(method);
2007
2008 iter.force_bci(jvms->bci());
2009 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2010 Bytecodes::Code c = iter.cur_bc();
2011 Node* lhs = nullptr;
2012 Node* rhs = nullptr;
2013 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2014 lhs = unc->peek_operand(0);
2015 rhs = unc->peek_operand(1);
2016 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2017 lhs = unc->peek_operand(0);
2018 }
2019
2020 ResourceMark rm;
2021 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2022 assert(live_locals.is_valid(), "broken liveness info");
2023 int len = (int)live_locals.size();
2024
2025 for (int i = 0; i < len; i++) {
2026 Node* local = unc->local(jvms, i);
2027 // kill local using the liveness of next_bci.
2028 // give up when the local looks like an operand to secure reexecution.
2029 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2030 uint idx = jvms->locoff() + i;
2031 #ifdef ASSERT
2032 if (PrintOpto && Verbose) {
2033 tty->print("[unstable_if] kill local#%d: ", idx);
2034 local->dump();
2035 tty->cr();
2036 }
2037 #endif
2038 igvn.replace_input_of(unc, idx, top());
2039 modified = true;
2040 }
2041 }
2042 }
2043
2044 // keep the mondified trap for late query
2045 if (modified) {
2046 trap->set_modified();
2047 } else {
2048 _unstable_if_traps.delete_at(i);
2049 }
2050 }
2051 igvn.optimize();
2052 }
2053
2054 // StringOpts and late inlining of string methods
2055 void Compile::inline_string_calls(bool parse_time) {
2056 {
2057 // remove useless nodes to make the usage analysis simpler
2058 ResourceMark rm;
2059 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2060 }
2061
2062 {
2063 ResourceMark rm;
2064 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2065 PhaseStringOpts pso(initial_gvn());
2066 print_method(PHASE_AFTER_STRINGOPTS, 3);
2067 }
2068
2069 // now inline anything that we skipped the first time around
2070 if (!parse_time) {
2071 _late_inlines_pos = _late_inlines.length();
2072 }
2073
2074 while (_string_late_inlines.length() > 0) {
2075 CallGenerator* cg = _string_late_inlines.pop();
2076 cg->do_late_inline();
2077 if (failing()) return;
2078 }
2079 _string_late_inlines.trunc_to(0);
2080 }
2081
2082 // Late inlining of boxing methods
2083 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2084 if (_boxing_late_inlines.length() > 0) {
2085 assert(has_boxed_value(), "inconsistent");
2086
2087 set_inlining_incrementally(true);
2088
2089 igvn_worklist()->ensure_empty(); // should be done with igvn
2090
2091 _late_inlines_pos = _late_inlines.length();
2092
2093 while (_boxing_late_inlines.length() > 0) {
2094 CallGenerator* cg = _boxing_late_inlines.pop();
2095 cg->do_late_inline();
2096 if (failing()) return;
2097 }
2098 _boxing_late_inlines.trunc_to(0);
2099
2100 inline_incrementally_cleanup(igvn);
2101
2102 set_inlining_incrementally(false);
2103 }
2104 }
2105
2106 bool Compile::inline_incrementally_one() {
2107 assert(IncrementalInline, "incremental inlining should be on");
2108 assert(_late_inlines.length() > 0, "should have been checked by caller");
2109
2110 TracePhase tp(_t_incrInline_inline);
2111
2112 set_inlining_progress(false);
2113 set_do_cleanup(false);
2114
2115 for (int i = 0; i < _late_inlines.length(); i++) {
2116 _late_inlines_pos = i+1;
2117 CallGenerator* cg = _late_inlines.at(i);
2118 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2119 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2120 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2121 if (should_stress_inlining()) {
2122 // randomly add repeated inline attempt if stress-inlining
2123 cg->call_node()->set_generator(cg);
2124 C->igvn_worklist()->push(cg->call_node());
2125 continue;
2126 }
2127 cg->do_late_inline();
2128 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2129 if (failing()) {
2130 return false;
2131 } else if (inlining_progress()) {
2132 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2133 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2134 break; // process one call site at a time
2135 } else {
2136 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2137 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2138 // Avoid potential infinite loop if node already in the IGVN list
2139 assert(false, "scheduled for IGVN during inlining attempt");
2140 } else {
2141 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2142 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2143 cg->call_node()->set_generator(cg);
2144 }
2145 }
2146 } else {
2147 // Ignore late inline direct calls when inlining is not allowed.
2148 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2149 }
2150 }
2151 // Remove processed elements.
2152 _late_inlines.remove_till(_late_inlines_pos);
2153 _late_inlines_pos = 0;
2154
2155 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2156
2157 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2158
2159 set_inlining_progress(false);
2160 set_do_cleanup(false);
2161
2162 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2163 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2164 }
2165
2166 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2167 {
2168 TracePhase tp(_t_incrInline_pru);
2169 ResourceMark rm;
2170 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2171 }
2172 {
2173 TracePhase tp(_t_incrInline_igvn);
2174 igvn.reset();
2175 igvn.optimize();
2176 if (failing()) return;
2177 }
2178 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2179 }
2180
2181 template<typename E>
2182 static void shuffle_array(Compile& C, GrowableArray<E>& array) {
2183 if (array.length() < 2) {
2184 return;
2185 }
2186 for (uint i = array.length() - 1; i >= 1; i--) {
2187 uint j = C.random() % (i + 1);
2188 swap(array.at(i), array.at(j));
2189 }
2190 }
2191
2192 void Compile::shuffle_late_inlines() {
2193 shuffle_array(*C, _late_inlines);
2194 }
2195
2196 // Perform incremental inlining until bound on number of live nodes is reached
2197 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2198 TracePhase tp(_t_incrInline);
2199
2200 set_inlining_incrementally(true);
2201 uint low_live_nodes = 0;
2202
2203 if (StressIncrementalInlining) {
2204 shuffle_late_inlines();
2205 }
2206
2207 while (_late_inlines.length() > 0) {
2208 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2209 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2210 TracePhase tp(_t_incrInline_ideal);
2211 // PhaseIdealLoop is expensive so we only try it once we are
2212 // out of live nodes and we only try it again if the previous
2213 // helped got the number of nodes down significantly
2214 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2215 if (failing()) return;
2216 low_live_nodes = live_nodes();
2217 _major_progress = true;
2218 }
2219
2220 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2221 bool do_print_inlining = print_inlining() || print_intrinsics();
2222 if (do_print_inlining || log() != nullptr) {
2223 // Print inlining message for candidates that we couldn't inline for lack of space.
2224 for (int i = 0; i < _late_inlines.length(); i++) {
2225 CallGenerator* cg = _late_inlines.at(i);
2226 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2227 if (do_print_inlining) {
2228 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2229 }
2230 log_late_inline_failure(cg, msg);
2231 }
2232 }
2233 break; // finish
2234 }
2235 }
2236
2237 igvn_worklist()->ensure_empty(); // should be done with igvn
2238
2239 if (_late_inlines.length() == 0) {
2240 break; // no more progress
2241 }
2242
2243 while (inline_incrementally_one()) {
2244 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2245 }
2246 if (failing()) return;
2247
2248 inline_incrementally_cleanup(igvn);
2249
2250 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2251
2252 if (failing()) return;
2253 }
2254
2255 igvn_worklist()->ensure_empty(); // should be done with igvn
2256
2257 if (_string_late_inlines.length() > 0) {
2258 assert(has_stringbuilder(), "inconsistent");
2259
2260 inline_string_calls(false);
2261
2262 if (failing()) return;
2263
2264 inline_incrementally_cleanup(igvn);
2265 }
2266
2267 set_inlining_incrementally(false);
2268 }
2269
2270 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2271 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2272 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2273 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2274 // as if "inlining_incrementally() == true" were set.
2275 assert(inlining_incrementally() == false, "not allowed");
2276 assert(_modified_nodes == nullptr, "not allowed");
2277 assert(_late_inlines.length() > 0, "sanity");
2278
2279 if (StressIncrementalInlining) {
2280 shuffle_late_inlines();
2281 }
2282
2283 while (_late_inlines.length() > 0) {
2284 igvn_worklist()->ensure_empty(); // should be done with igvn
2285
2286 while (inline_incrementally_one()) {
2287 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2288 }
2289 if (failing()) return;
2290
2291 inline_incrementally_cleanup(igvn);
2292 }
2293 }
2294
2295 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2296 if (_loop_opts_cnt > 0) {
2297 while (major_progress() && (_loop_opts_cnt > 0)) {
2298 TracePhase tp(_t_idealLoop);
2299 PhaseIdealLoop::optimize(igvn, mode);
2300 _loop_opts_cnt--;
2301 if (failing()) return false;
2302 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2303 }
2304 }
2305 return true;
2306 }
2307
2308 // Remove edges from "root" to each SafePoint at a backward branch.
2309 // They were inserted during parsing (see add_safepoint()) to make
2310 // infinite loops without calls or exceptions visible to root, i.e.,
2311 // useful.
2312 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2313 Node *r = root();
2314 if (r != nullptr) {
2315 for (uint i = r->req(); i < r->len(); ++i) {
2316 Node *n = r->in(i);
2317 if (n != nullptr && n->is_SafePoint()) {
2318 r->rm_prec(i);
2319 if (n->outcnt() == 0) {
2320 igvn.remove_dead_node(n);
2321 }
2322 --i;
2323 }
2324 }
2325 // Parsing may have added top inputs to the root node (Path
2326 // leading to the Halt node proven dead). Make sure we get a
2327 // chance to clean them up.
2328 igvn._worklist.push(r);
2329 igvn.optimize();
2330 }
2331 }
2332
2333 //------------------------------Optimize---------------------------------------
2334 // Given a graph, optimize it.
2335 void Compile::Optimize() {
2336 TracePhase tp(_t_optimizer);
2337
2338 #ifndef PRODUCT
2339 if (env()->break_at_compile()) {
2340 BREAKPOINT;
2341 }
2342
2343 #endif
2344
2345 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2346 #ifdef ASSERT
2347 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2348 #endif
2349
2350 ResourceMark rm;
2351
2352 NOT_PRODUCT( verify_graph_edges(); )
2353
2354 print_method(PHASE_AFTER_PARSING, 1);
2355
2356 {
2357 // Iterative Global Value Numbering, including ideal transforms
2358 PhaseIterGVN igvn;
2359 #ifdef ASSERT
2360 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2361 #endif
2362 {
2363 TracePhase tp(_t_iterGVN);
2364 igvn.optimize();
2365 }
2366
2367 if (failing()) return;
2368
2369 print_method(PHASE_ITER_GVN1, 2);
2370
2371 process_for_unstable_if_traps(igvn);
2372
2373 if (failing()) return;
2374
2375 inline_incrementally(igvn);
2376
2377 print_method(PHASE_INCREMENTAL_INLINE, 2);
2378
2379 if (failing()) return;
2380
2381 if (eliminate_boxing()) {
2382 // Inline valueOf() methods now.
2383 inline_boxing_calls(igvn);
2384
2385 if (failing()) return;
2386
2387 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2388 inline_incrementally(igvn);
2389 }
2390
2391 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2392
2393 if (failing()) return;
2394 }
2395
2396 // Remove the speculative part of types and clean up the graph from
2397 // the extra CastPP nodes whose only purpose is to carry them. Do
2398 // that early so that optimizations are not disrupted by the extra
2399 // CastPP nodes.
2400 remove_speculative_types(igvn);
2401
2402 if (failing()) return;
2403
2404 // No more new expensive nodes will be added to the list from here
2405 // so keep only the actual candidates for optimizations.
2406 cleanup_expensive_nodes(igvn);
2407
2408 if (failing()) return;
2409
2410 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2411 if (EnableVectorSupport && has_vbox_nodes()) {
2412 TracePhase tp(_t_vector);
2413 PhaseVector pv(igvn);
2414 pv.optimize_vector_boxes();
2415 if (failing()) return;
2416 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2417 }
2418 assert(!has_vbox_nodes(), "sanity");
2419
2420 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2421 Compile::TracePhase tp(_t_renumberLive);
2422 igvn_worklist()->ensure_empty(); // should be done with igvn
2423 {
2424 ResourceMark rm;
2425 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2426 }
2427 igvn.reset();
2428 igvn.optimize();
2429 if (failing()) return;
2430 }
2431
2432 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2433 // safepoints
2434 remove_root_to_sfpts_edges(igvn);
2435
2436 if (failing()) return;
2437
2438 _print_phase_loop_opts = has_loops();
2439 if (_print_phase_loop_opts) {
2440 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2441 }
2442
2443 // Perform escape analysis
2444 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2445 if (has_loops()) {
2446 // Cleanup graph (remove dead nodes).
2447 TracePhase tp(_t_idealLoop);
2448 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2449 if (failing()) return;
2450 }
2451 bool progress;
2452 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2453 do {
2454 ConnectionGraph::do_analysis(this, &igvn);
2455
2456 if (failing()) return;
2457
2458 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2459
2460 // Optimize out fields loads from scalar replaceable allocations.
2461 igvn.optimize();
2462 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2463
2464 if (failing()) return;
2465
2466 if (congraph() != nullptr && macro_count() > 0) {
2467 TracePhase tp(_t_macroEliminate);
2468 PhaseMacroExpand mexp(igvn);
2469 mexp.eliminate_macro_nodes();
2470 if (failing()) return;
2471 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2472
2473 igvn.set_delay_transform(false);
2474 igvn.optimize();
2475 if (failing()) return;
2476
2477 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2478 }
2479
2480 ConnectionGraph::verify_ram_nodes(this, root());
2481 if (failing()) return;
2482
2483 progress = do_iterative_escape_analysis() &&
2484 (macro_count() < mcount) &&
2485 ConnectionGraph::has_candidates(this);
2486 // Try again if candidates exist and made progress
2487 // by removing some allocations and/or locks.
2488 } while (progress);
2489 }
2490
2491 // Loop transforms on the ideal graph. Range Check Elimination,
2492 // peeling, unrolling, etc.
2493
2494 // Set loop opts counter
2495 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2496 {
2497 TracePhase tp(_t_idealLoop);
2498 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2499 _loop_opts_cnt--;
2500 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2501 if (failing()) return;
2502 }
2503 // Loop opts pass if partial peeling occurred in previous pass
2504 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2505 TracePhase tp(_t_idealLoop);
2506 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2507 _loop_opts_cnt--;
2508 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2509 if (failing()) return;
2510 }
2511 // Loop opts pass for loop-unrolling before CCP
2512 if(major_progress() && (_loop_opts_cnt > 0)) {
2513 TracePhase tp(_t_idealLoop);
2514 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2515 _loop_opts_cnt--;
2516 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2517 }
2518 if (!failing()) {
2519 // Verify that last round of loop opts produced a valid graph
2520 PhaseIdealLoop::verify(igvn);
2521 }
2522 }
2523 if (failing()) return;
2524
2525 // Conditional Constant Propagation;
2526 print_method(PHASE_BEFORE_CCP1, 2);
2527 PhaseCCP ccp( &igvn );
2528 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2529 {
2530 TracePhase tp(_t_ccp);
2531 ccp.do_transform();
2532 }
2533 print_method(PHASE_CCP1, 2);
2534
2535 assert( true, "Break here to ccp.dump_old2new_map()");
2536
2537 // Iterative Global Value Numbering, including ideal transforms
2538 {
2539 TracePhase tp(_t_iterGVN2);
2540 igvn.reset_from_igvn(&ccp);
2541 igvn.optimize();
2542 }
2543 print_method(PHASE_ITER_GVN2, 2);
2544
2545 if (failing()) return;
2546
2547 // Loop transforms on the ideal graph. Range Check Elimination,
2548 // peeling, unrolling, etc.
2549 if (!optimize_loops(igvn, LoopOptsDefault)) {
2550 return;
2551 }
2552
2553 if (failing()) return;
2554
2555 C->clear_major_progress(); // ensure that major progress is now clear
2556
2557 process_for_post_loop_opts_igvn(igvn);
2558
2559 process_for_merge_stores_igvn(igvn);
2560
2561 if (failing()) return;
2562
2563 #ifdef ASSERT
2564 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2565 #endif
2566
2567 {
2568 TracePhase tp(_t_macroExpand);
2569 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2570 PhaseMacroExpand mex(igvn);
2571 // Do not allow new macro nodes once we start to eliminate and expand
2572 C->reset_allow_macro_nodes();
2573 // Last attempt to eliminate macro nodes before expand
2574 mex.eliminate_macro_nodes();
2575 if (failing()) {
2576 return;
2577 }
2578 mex.eliminate_opaque_looplimit_macro_nodes();
2579 if (failing()) {
2580 return;
2581 }
2582 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2583 if (mex.expand_macro_nodes()) {
2584 assert(failing(), "must bail out w/ explicit message");
2585 return;
2586 }
2587 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2588 }
2589
2590 {
2591 TracePhase tp(_t_barrierExpand);
2592 if (bs->expand_barriers(this, igvn)) {
2593 assert(failing(), "must bail out w/ explicit message");
2594 return;
2595 }
2596 print_method(PHASE_BARRIER_EXPANSION, 2);
2597 }
2598
2599 if (C->max_vector_size() > 0) {
2600 C->optimize_logic_cones(igvn);
2601 igvn.optimize();
2602 if (failing()) return;
2603 }
2604
2605 DEBUG_ONLY( _modified_nodes = nullptr; )
2606
2607 assert(igvn._worklist.size() == 0, "not empty");
2608
2609 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2610
2611 if (_late_inlines.length() > 0) {
2612 // More opportunities to optimize virtual and MH calls.
2613 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2614 process_late_inline_calls_no_inline(igvn);
2615 if (failing()) return;
2616 }
2617 } // (End scope of igvn; run destructor if necessary for asserts.)
2618
2619 check_no_dead_use();
2620
2621 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2622 // to remove hashes to unlock nodes for modifications.
2623 C->node_hash()->clear();
2624
2625 // A method with only infinite loops has no edges entering loops from root
2626 {
2627 TracePhase tp(_t_graphReshaping);
2628 if (final_graph_reshaping()) {
2629 assert(failing(), "must bail out w/ explicit message");
2630 return;
2631 }
2632 }
2633
2634 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2635 DEBUG_ONLY(set_phase_optimize_finished();)
2636 }
2637
2638 #ifdef ASSERT
2639 void Compile::check_no_dead_use() const {
2640 ResourceMark rm;
2641 Unique_Node_List wq;
2642 wq.push(root());
2643 for (uint i = 0; i < wq.size(); ++i) {
2644 Node* n = wq.at(i);
2645 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2646 Node* u = n->fast_out(j);
2647 if (u->outcnt() == 0 && !u->is_Con()) {
2648 u->dump();
2649 fatal("no reachable node should have no use");
2650 }
2651 wq.push(u);
2652 }
2653 }
2654 }
2655 #endif
2656
2657 void Compile::inline_vector_reboxing_calls() {
2658 if (C->_vector_reboxing_late_inlines.length() > 0) {
2659 _late_inlines_pos = C->_late_inlines.length();
2660 while (_vector_reboxing_late_inlines.length() > 0) {
2661 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2662 cg->do_late_inline();
2663 if (failing()) return;
2664 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2665 }
2666 _vector_reboxing_late_inlines.trunc_to(0);
2667 }
2668 }
2669
2670 bool Compile::has_vbox_nodes() {
2671 if (C->_vector_reboxing_late_inlines.length() > 0) {
2672 return true;
2673 }
2674 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2675 Node * n = C->macro_node(macro_idx);
2676 assert(n->is_macro(), "only macro nodes expected here");
2677 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2678 return true;
2679 }
2680 }
2681 return false;
2682 }
2683
2684 //---------------------------- Bitwise operation packing optimization ---------------------------
2685
2686 static bool is_vector_unary_bitwise_op(Node* n) {
2687 return n->Opcode() == Op_XorV &&
2688 VectorNode::is_vector_bitwise_not_pattern(n);
2689 }
2690
2691 static bool is_vector_binary_bitwise_op(Node* n) {
2692 switch (n->Opcode()) {
2693 case Op_AndV:
2694 case Op_OrV:
2695 return true;
2696
2697 case Op_XorV:
2698 return !is_vector_unary_bitwise_op(n);
2699
2700 default:
2701 return false;
2702 }
2703 }
2704
2705 static bool is_vector_ternary_bitwise_op(Node* n) {
2706 return n->Opcode() == Op_MacroLogicV;
2707 }
2708
2709 static bool is_vector_bitwise_op(Node* n) {
2710 return is_vector_unary_bitwise_op(n) ||
2711 is_vector_binary_bitwise_op(n) ||
2712 is_vector_ternary_bitwise_op(n);
2713 }
2714
2715 static bool is_vector_bitwise_cone_root(Node* n) {
2716 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2717 return false;
2718 }
2719 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2720 if (is_vector_bitwise_op(n->fast_out(i))) {
2721 return false;
2722 }
2723 }
2724 return true;
2725 }
2726
2727 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2728 uint cnt = 0;
2729 if (is_vector_bitwise_op(n)) {
2730 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2731 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2732 for (uint i = 1; i < inp_cnt; i++) {
2733 Node* in = n->in(i);
2734 bool skip = VectorNode::is_all_ones_vector(in);
2735 if (!skip && !inputs.member(in)) {
2736 inputs.push(in);
2737 cnt++;
2738 }
2739 }
2740 assert(cnt <= 1, "not unary");
2741 } else {
2742 uint last_req = inp_cnt;
2743 if (is_vector_ternary_bitwise_op(n)) {
2744 last_req = inp_cnt - 1; // skip last input
2745 }
2746 for (uint i = 1; i < last_req; i++) {
2747 Node* def = n->in(i);
2748 if (!inputs.member(def)) {
2749 inputs.push(def);
2750 cnt++;
2751 }
2752 }
2753 }
2754 } else { // not a bitwise operations
2755 if (!inputs.member(n)) {
2756 inputs.push(n);
2757 cnt++;
2758 }
2759 }
2760 return cnt;
2761 }
2762
2763 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2764 Unique_Node_List useful_nodes;
2765 C->identify_useful_nodes(useful_nodes);
2766
2767 for (uint i = 0; i < useful_nodes.size(); i++) {
2768 Node* n = useful_nodes.at(i);
2769 if (is_vector_bitwise_cone_root(n)) {
2770 list.push(n);
2771 }
2772 }
2773 }
2774
2775 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2776 const TypeVect* vt,
2777 Unique_Node_List& partition,
2778 Unique_Node_List& inputs) {
2779 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2780 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2781 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2782
2783 Node* in1 = inputs.at(0);
2784 Node* in2 = inputs.at(1);
2785 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2786
2787 uint func = compute_truth_table(partition, inputs);
2788
2789 Node* pn = partition.at(partition.size() - 1);
2790 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2791 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2792 }
2793
2794 static uint extract_bit(uint func, uint pos) {
2795 return (func & (1 << pos)) >> pos;
2796 }
2797
2798 //
2799 // A macro logic node represents a truth table. It has 4 inputs,
2800 // First three inputs corresponds to 3 columns of a truth table
2801 // and fourth input captures the logic function.
2802 //
2803 // eg. fn = (in1 AND in2) OR in3;
2804 //
2805 // MacroNode(in1,in2,in3,fn)
2806 //
2807 // -----------------
2808 // in1 in2 in3 fn
2809 // -----------------
2810 // 0 0 0 0
2811 // 0 0 1 1
2812 // 0 1 0 0
2813 // 0 1 1 1
2814 // 1 0 0 0
2815 // 1 0 1 1
2816 // 1 1 0 1
2817 // 1 1 1 1
2818 //
2819
2820 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2821 int res = 0;
2822 for (int i = 0; i < 8; i++) {
2823 int bit1 = extract_bit(in1, i);
2824 int bit2 = extract_bit(in2, i);
2825 int bit3 = extract_bit(in3, i);
2826
2827 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2828 int func_bit = extract_bit(func, func_bit_pos);
2829
2830 res |= func_bit << i;
2831 }
2832 return res;
2833 }
2834
2835 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
2836 assert(n != nullptr, "");
2837 assert(eval_map.contains(n), "absent");
2838 return *(eval_map.get(n));
2839 }
2840
2841 static void eval_operands(Node* n,
2842 uint& func1, uint& func2, uint& func3,
2843 HashTable<Node*,uint>& eval_map) {
2844 assert(is_vector_bitwise_op(n), "");
2845
2846 if (is_vector_unary_bitwise_op(n)) {
2847 Node* opnd = n->in(1);
2848 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2849 opnd = n->in(2);
2850 }
2851 func1 = eval_operand(opnd, eval_map);
2852 } else if (is_vector_binary_bitwise_op(n)) {
2853 func1 = eval_operand(n->in(1), eval_map);
2854 func2 = eval_operand(n->in(2), eval_map);
2855 } else {
2856 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2857 func1 = eval_operand(n->in(1), eval_map);
2858 func2 = eval_operand(n->in(2), eval_map);
2859 func3 = eval_operand(n->in(3), eval_map);
2860 }
2861 }
2862
2863 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2864 assert(inputs.size() <= 3, "sanity");
2865 ResourceMark rm;
2866 uint res = 0;
2867 HashTable<Node*,uint> eval_map;
2868
2869 // Populate precomputed functions for inputs.
2870 // Each input corresponds to one column of 3 input truth-table.
2871 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2872 0xCC, // (_, b, _) -> b
2873 0xF0 }; // (a, _, _) -> a
2874 for (uint i = 0; i < inputs.size(); i++) {
2875 eval_map.put(inputs.at(i), input_funcs[2-i]);
2876 }
2877
2878 for (uint i = 0; i < partition.size(); i++) {
2879 Node* n = partition.at(i);
2880
2881 uint func1 = 0, func2 = 0, func3 = 0;
2882 eval_operands(n, func1, func2, func3, eval_map);
2883
2884 switch (n->Opcode()) {
2885 case Op_OrV:
2886 assert(func3 == 0, "not binary");
2887 res = func1 | func2;
2888 break;
2889 case Op_AndV:
2890 assert(func3 == 0, "not binary");
2891 res = func1 & func2;
2892 break;
2893 case Op_XorV:
2894 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2895 assert(func2 == 0 && func3 == 0, "not unary");
2896 res = (~func1) & 0xFF;
2897 } else {
2898 assert(func3 == 0, "not binary");
2899 res = func1 ^ func2;
2900 }
2901 break;
2902 case Op_MacroLogicV:
2903 // Ordering of inputs may change during evaluation of sub-tree
2904 // containing MacroLogic node as a child node, thus a re-evaluation
2905 // makes sure that function is evaluated in context of current
2906 // inputs.
2907 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2908 break;
2909
2910 default: assert(false, "not supported: %s", n->Name());
2911 }
2912 assert(res <= 0xFF, "invalid");
2913 eval_map.put(n, res);
2914 }
2915 return res;
2916 }
2917
2918 // Criteria under which nodes gets packed into a macro logic node:-
2919 // 1) Parent and both child nodes are all unmasked or masked with
2920 // same predicates.
2921 // 2) Masked parent can be packed with left child if it is predicated
2922 // and both have same predicates.
2923 // 3) Masked parent can be packed with right child if its un-predicated
2924 // or has matching predication condition.
2925 // 4) An unmasked parent can be packed with an unmasked child.
2926 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2927 assert(partition.size() == 0, "not empty");
2928 assert(inputs.size() == 0, "not empty");
2929 if (is_vector_ternary_bitwise_op(n)) {
2930 return false;
2931 }
2932
2933 bool is_unary_op = is_vector_unary_bitwise_op(n);
2934 if (is_unary_op) {
2935 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2936 return false; // too few inputs
2937 }
2938
2939 bool pack_left_child = true;
2940 bool pack_right_child = true;
2941
2942 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2943 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2944
2945 int left_child_input_cnt = 0;
2946 int right_child_input_cnt = 0;
2947
2948 bool parent_is_predicated = n->is_predicated_vector();
2949 bool left_child_predicated = n->in(1)->is_predicated_vector();
2950 bool right_child_predicated = n->in(2)->is_predicated_vector();
2951
2952 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2953 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2954 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2955
2956 do {
2957 if (pack_left_child && left_child_LOP &&
2958 ((!parent_is_predicated && !left_child_predicated) ||
2959 ((parent_is_predicated && left_child_predicated &&
2960 parent_pred == left_child_pred)))) {
2961 partition.push(n->in(1));
2962 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2963 } else {
2964 inputs.push(n->in(1));
2965 left_child_input_cnt = 1;
2966 }
2967
2968 if (pack_right_child && right_child_LOP &&
2969 (!right_child_predicated ||
2970 (right_child_predicated && parent_is_predicated &&
2971 parent_pred == right_child_pred))) {
2972 partition.push(n->in(2));
2973 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2974 } else {
2975 inputs.push(n->in(2));
2976 right_child_input_cnt = 1;
2977 }
2978
2979 if (inputs.size() > 3) {
2980 assert(partition.size() > 0, "");
2981 inputs.clear();
2982 partition.clear();
2983 if (left_child_input_cnt > right_child_input_cnt) {
2984 pack_left_child = false;
2985 } else {
2986 pack_right_child = false;
2987 }
2988 } else {
2989 break;
2990 }
2991 } while(true);
2992
2993 if(partition.size()) {
2994 partition.push(n);
2995 }
2996
2997 return (partition.size() == 2 || partition.size() == 3) &&
2998 (inputs.size() == 2 || inputs.size() == 3);
2999 }
3000
3001 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3002 assert(is_vector_bitwise_op(n), "not a root");
3003
3004 visited.set(n->_idx);
3005
3006 // 1) Do a DFS walk over the logic cone.
3007 for (uint i = 1; i < n->req(); i++) {
3008 Node* in = n->in(i);
3009 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3010 process_logic_cone_root(igvn, in, visited);
3011 }
3012 }
3013
3014 // 2) Bottom up traversal: Merge node[s] with
3015 // the parent to form macro logic node.
3016 Unique_Node_List partition;
3017 Unique_Node_List inputs;
3018 if (compute_logic_cone(n, partition, inputs)) {
3019 const TypeVect* vt = n->bottom_type()->is_vect();
3020 Node* pn = partition.at(partition.size() - 1);
3021 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3022 if (mask == nullptr ||
3023 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3024 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3025 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3026 igvn.replace_node(n, macro_logic);
3027 }
3028 }
3029 }
3030
3031 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3032 ResourceMark rm;
3033 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3034 Unique_Node_List list;
3035 collect_logic_cone_roots(list);
3036
3037 while (list.size() > 0) {
3038 Node* n = list.pop();
3039 const TypeVect* vt = n->bottom_type()->is_vect();
3040 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3041 if (supported) {
3042 VectorSet visited(comp_arena());
3043 process_logic_cone_root(igvn, n, visited);
3044 }
3045 }
3046 }
3047 }
3048
3049 //------------------------------Code_Gen---------------------------------------
3050 // Given a graph, generate code for it
3051 void Compile::Code_Gen() {
3052 if (failing()) {
3053 return;
3054 }
3055
3056 // Perform instruction selection. You might think we could reclaim Matcher
3057 // memory PDQ, but actually the Matcher is used in generating spill code.
3058 // Internals of the Matcher (including some VectorSets) must remain live
3059 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3060 // set a bit in reclaimed memory.
3061
3062 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3063 // nodes. Mapping is only valid at the root of each matched subtree.
3064 NOT_PRODUCT( verify_graph_edges(); )
3065
3066 Matcher matcher;
3067 _matcher = &matcher;
3068 {
3069 TracePhase tp(_t_matcher);
3070 matcher.match();
3071 if (failing()) {
3072 return;
3073 }
3074 }
3075 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3076 // nodes. Mapping is only valid at the root of each matched subtree.
3077 NOT_PRODUCT( verify_graph_edges(); )
3078
3079 // If you have too many nodes, or if matching has failed, bail out
3080 check_node_count(0, "out of nodes matching instructions");
3081 if (failing()) {
3082 return;
3083 }
3084
3085 print_method(PHASE_MATCHING, 2);
3086
3087 // Build a proper-looking CFG
3088 PhaseCFG cfg(node_arena(), root(), matcher);
3089 if (failing()) {
3090 return;
3091 }
3092 _cfg = &cfg;
3093 {
3094 TracePhase tp(_t_scheduler);
3095 bool success = cfg.do_global_code_motion();
3096 if (!success) {
3097 return;
3098 }
3099
3100 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3101 NOT_PRODUCT( verify_graph_edges(); )
3102 cfg.verify();
3103 if (failing()) {
3104 return;
3105 }
3106 }
3107
3108 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3109 _regalloc = ®alloc;
3110 {
3111 TracePhase tp(_t_registerAllocation);
3112 // Perform register allocation. After Chaitin, use-def chains are
3113 // no longer accurate (at spill code) and so must be ignored.
3114 // Node->LRG->reg mappings are still accurate.
3115 _regalloc->Register_Allocate();
3116
3117 // Bail out if the allocator builds too many nodes
3118 if (failing()) {
3119 return;
3120 }
3121
3122 print_method(PHASE_REGISTER_ALLOCATION, 2);
3123 }
3124
3125 // Prior to register allocation we kept empty basic blocks in case the
3126 // the allocator needed a place to spill. After register allocation we
3127 // are not adding any new instructions. If any basic block is empty, we
3128 // can now safely remove it.
3129 {
3130 TracePhase tp(_t_blockOrdering);
3131 cfg.remove_empty_blocks();
3132 if (do_freq_based_layout()) {
3133 PhaseBlockLayout layout(cfg);
3134 } else {
3135 cfg.set_loop_alignment();
3136 }
3137 cfg.fixup_flow();
3138 cfg.remove_unreachable_blocks();
3139 cfg.verify_dominator_tree();
3140 print_method(PHASE_BLOCK_ORDERING, 3);
3141 }
3142
3143 // Apply peephole optimizations
3144 if( OptoPeephole ) {
3145 TracePhase tp(_t_peephole);
3146 PhasePeephole peep( _regalloc, cfg);
3147 peep.do_transform();
3148 print_method(PHASE_PEEPHOLE, 3);
3149 }
3150
3151 // Do late expand if CPU requires this.
3152 if (Matcher::require_postalloc_expand) {
3153 TracePhase tp(_t_postalloc_expand);
3154 cfg.postalloc_expand(_regalloc);
3155 print_method(PHASE_POSTALLOC_EXPAND, 3);
3156 }
3157
3158 #ifdef ASSERT
3159 {
3160 CompilationMemoryStatistic::do_test_allocations();
3161 if (failing()) return;
3162 }
3163 #endif
3164
3165 // Convert Nodes to instruction bits in a buffer
3166 {
3167 TracePhase tp(_t_output);
3168 PhaseOutput output;
3169 output.Output();
3170 if (failing()) return;
3171 output.install();
3172 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3173 }
3174
3175 // He's dead, Jim.
3176 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3177 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3178 }
3179
3180 //------------------------------Final_Reshape_Counts---------------------------
3181 // This class defines counters to help identify when a method
3182 // may/must be executed using hardware with only 24-bit precision.
3183 struct Final_Reshape_Counts : public StackObj {
3184 int _call_count; // count non-inlined 'common' calls
3185 int _float_count; // count float ops requiring 24-bit precision
3186 int _double_count; // count double ops requiring more precision
3187 int _java_call_count; // count non-inlined 'java' calls
3188 int _inner_loop_count; // count loops which need alignment
3189 VectorSet _visited; // Visitation flags
3190 Node_List _tests; // Set of IfNodes & PCTableNodes
3191
3192 Final_Reshape_Counts() :
3193 _call_count(0), _float_count(0), _double_count(0),
3194 _java_call_count(0), _inner_loop_count(0) { }
3195
3196 void inc_call_count () { _call_count ++; }
3197 void inc_float_count () { _float_count ++; }
3198 void inc_double_count() { _double_count++; }
3199 void inc_java_call_count() { _java_call_count++; }
3200 void inc_inner_loop_count() { _inner_loop_count++; }
3201
3202 int get_call_count () const { return _call_count ; }
3203 int get_float_count () const { return _float_count ; }
3204 int get_double_count() const { return _double_count; }
3205 int get_java_call_count() const { return _java_call_count; }
3206 int get_inner_loop_count() const { return _inner_loop_count; }
3207 };
3208
3209 //------------------------------final_graph_reshaping_impl----------------------
3210 // Implement items 1-5 from final_graph_reshaping below.
3211 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3212
3213 if ( n->outcnt() == 0 ) return; // dead node
3214 uint nop = n->Opcode();
3215
3216 // Check for 2-input instruction with "last use" on right input.
3217 // Swap to left input. Implements item (2).
3218 if( n->req() == 3 && // two-input instruction
3219 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3220 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3221 n->in(2)->outcnt() == 1 &&// right use IS a last use
3222 !n->in(2)->is_Con() ) { // right use is not a constant
3223 // Check for commutative opcode
3224 switch( nop ) {
3225 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3226 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3227 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3228 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3229 case Op_AndL: case Op_XorL: case Op_OrL:
3230 case Op_AndI: case Op_XorI: case Op_OrI: {
3231 // Move "last use" input to left by swapping inputs
3232 n->swap_edges(1, 2);
3233 break;
3234 }
3235 default:
3236 break;
3237 }
3238 }
3239
3240 #ifdef ASSERT
3241 if( n->is_Mem() ) {
3242 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3243 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3244 // oop will be recorded in oop map if load crosses safepoint
3245 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3246 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3247 "raw memory operations should have control edge");
3248 }
3249 if (n->is_MemBar()) {
3250 MemBarNode* mb = n->as_MemBar();
3251 if (mb->trailing_store() || mb->trailing_load_store()) {
3252 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3253 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3254 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3255 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3256 } else if (mb->leading()) {
3257 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3258 }
3259 }
3260 #endif
3261 // Count FPU ops and common calls, implements item (3)
3262 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3263 if (!gc_handled) {
3264 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3265 }
3266
3267 // Collect CFG split points
3268 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3269 frc._tests.push(n);
3270 }
3271 }
3272
3273 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3274 if (!UseDivMod) {
3275 return;
3276 }
3277
3278 // Check if "a % b" and "a / b" both exist
3279 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3280 if (d == nullptr) {
3281 return;
3282 }
3283
3284 // Replace them with a fused divmod if supported
3285 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3286 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3287 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3288 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3289 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3290 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3291 // DivMod node so the dependency is not lost.
3292 divmod->add_prec_from(n);
3293 divmod->add_prec_from(d);
3294 d->subsume_by(divmod->div_proj(), this);
3295 n->subsume_by(divmod->mod_proj(), this);
3296 } else {
3297 // Replace "a % b" with "a - ((a / b) * b)"
3298 Node* mult = MulNode::make(d, d->in(2), bt);
3299 Node* sub = SubNode::make(d->in(1), mult, bt);
3300 n->subsume_by(sub, this);
3301 }
3302 }
3303
3304 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3305 switch( nop ) {
3306 // Count all float operations that may use FPU
3307 case Op_AddF:
3308 case Op_SubF:
3309 case Op_MulF:
3310 case Op_DivF:
3311 case Op_NegF:
3312 case Op_ModF:
3313 case Op_ConvI2F:
3314 case Op_ConF:
3315 case Op_CmpF:
3316 case Op_CmpF3:
3317 case Op_StoreF:
3318 case Op_LoadF:
3319 // case Op_ConvL2F: // longs are split into 32-bit halves
3320 frc.inc_float_count();
3321 break;
3322
3323 case Op_ConvF2D:
3324 case Op_ConvD2F:
3325 frc.inc_float_count();
3326 frc.inc_double_count();
3327 break;
3328
3329 // Count all double operations that may use FPU
3330 case Op_AddD:
3331 case Op_SubD:
3332 case Op_MulD:
3333 case Op_DivD:
3334 case Op_NegD:
3335 case Op_ModD:
3336 case Op_ConvI2D:
3337 case Op_ConvD2I:
3338 // case Op_ConvL2D: // handled by leaf call
3339 // case Op_ConvD2L: // handled by leaf call
3340 case Op_ConD:
3341 case Op_CmpD:
3342 case Op_CmpD3:
3343 case Op_StoreD:
3344 case Op_LoadD:
3345 case Op_LoadD_unaligned:
3346 frc.inc_double_count();
3347 break;
3348 case Op_Opaque1: // Remove Opaque Nodes before matching
3349 n->subsume_by(n->in(1), this);
3350 break;
3351 case Op_CallLeafPure: {
3352 // If the pure call is not supported, then lower to a CallLeaf.
3353 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3354 CallNode* call = n->as_Call();
3355 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3356 call->_name, TypeRawPtr::BOTTOM);
3357 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3358 new_call->init_req(TypeFunc::I_O, C->top());
3359 new_call->init_req(TypeFunc::Memory, C->top());
3360 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3361 new_call->init_req(TypeFunc::FramePtr, C->top());
3362 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
3363 new_call->init_req(i, call->in(i));
3364 }
3365 n->subsume_by(new_call, this);
3366 }
3367 frc.inc_call_count();
3368 break;
3369 }
3370 case Op_CallStaticJava:
3371 case Op_CallJava:
3372 case Op_CallDynamicJava:
3373 frc.inc_java_call_count(); // Count java call site;
3374 case Op_CallRuntime:
3375 case Op_CallLeaf:
3376 case Op_CallLeafVector:
3377 case Op_CallLeafNoFP: {
3378 assert (n->is_Call(), "");
3379 CallNode *call = n->as_Call();
3380 // Count call sites where the FP mode bit would have to be flipped.
3381 // Do not count uncommon runtime calls:
3382 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3383 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3384 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3385 frc.inc_call_count(); // Count the call site
3386 } else { // See if uncommon argument is shared
3387 Node *n = call->in(TypeFunc::Parms);
3388 int nop = n->Opcode();
3389 // Clone shared simple arguments to uncommon calls, item (1).
3390 if (n->outcnt() > 1 &&
3391 !n->is_Proj() &&
3392 nop != Op_CreateEx &&
3393 nop != Op_CheckCastPP &&
3394 nop != Op_DecodeN &&
3395 nop != Op_DecodeNKlass &&
3396 !n->is_Mem() &&
3397 !n->is_Phi()) {
3398 Node *x = n->clone();
3399 call->set_req(TypeFunc::Parms, x);
3400 }
3401 }
3402 break;
3403 }
3404 case Op_StoreB:
3405 case Op_StoreC:
3406 case Op_StoreI:
3407 case Op_StoreL:
3408 case Op_CompareAndSwapB:
3409 case Op_CompareAndSwapS:
3410 case Op_CompareAndSwapI:
3411 case Op_CompareAndSwapL:
3412 case Op_CompareAndSwapP:
3413 case Op_CompareAndSwapN:
3414 case Op_WeakCompareAndSwapB:
3415 case Op_WeakCompareAndSwapS:
3416 case Op_WeakCompareAndSwapI:
3417 case Op_WeakCompareAndSwapL:
3418 case Op_WeakCompareAndSwapP:
3419 case Op_WeakCompareAndSwapN:
3420 case Op_CompareAndExchangeB:
3421 case Op_CompareAndExchangeS:
3422 case Op_CompareAndExchangeI:
3423 case Op_CompareAndExchangeL:
3424 case Op_CompareAndExchangeP:
3425 case Op_CompareAndExchangeN:
3426 case Op_GetAndAddS:
3427 case Op_GetAndAddB:
3428 case Op_GetAndAddI:
3429 case Op_GetAndAddL:
3430 case Op_GetAndSetS:
3431 case Op_GetAndSetB:
3432 case Op_GetAndSetI:
3433 case Op_GetAndSetL:
3434 case Op_GetAndSetP:
3435 case Op_GetAndSetN:
3436 case Op_StoreP:
3437 case Op_StoreN:
3438 case Op_StoreNKlass:
3439 case Op_LoadB:
3440 case Op_LoadUB:
3441 case Op_LoadUS:
3442 case Op_LoadI:
3443 case Op_LoadKlass:
3444 case Op_LoadNKlass:
3445 case Op_LoadL:
3446 case Op_LoadL_unaligned:
3447 case Op_LoadP:
3448 case Op_LoadN:
3449 case Op_LoadRange:
3450 case Op_LoadS:
3451 break;
3452
3453 case Op_AddP: { // Assert sane base pointers
3454 Node *addp = n->in(AddPNode::Address);
3455 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
3456 #ifdef _LP64
3457 if ((UseCompressedOops || UseCompressedClassPointers) &&
3458 addp->Opcode() == Op_ConP &&
3459 addp == n->in(AddPNode::Base) &&
3460 n->in(AddPNode::Offset)->is_Con()) {
3461 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3462 // on the platform and on the compressed oops mode.
3463 // Use addressing with narrow klass to load with offset on x86.
3464 // Some platforms can use the constant pool to load ConP.
3465 // Do this transformation here since IGVN will convert ConN back to ConP.
3466 const Type* t = addp->bottom_type();
3467 bool is_oop = t->isa_oopptr() != nullptr;
3468 bool is_klass = t->isa_klassptr() != nullptr;
3469
3470 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3471 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3472 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3473 Node* nn = nullptr;
3474
3475 int op = is_oop ? Op_ConN : Op_ConNKlass;
3476
3477 // Look for existing ConN node of the same exact type.
3478 Node* r = root();
3479 uint cnt = r->outcnt();
3480 for (uint i = 0; i < cnt; i++) {
3481 Node* m = r->raw_out(i);
3482 if (m!= nullptr && m->Opcode() == op &&
3483 m->bottom_type()->make_ptr() == t) {
3484 nn = m;
3485 break;
3486 }
3487 }
3488 if (nn != nullptr) {
3489 // Decode a narrow oop to match address
3490 // [R12 + narrow_oop_reg<<3 + offset]
3491 if (is_oop) {
3492 nn = new DecodeNNode(nn, t);
3493 } else {
3494 nn = new DecodeNKlassNode(nn, t);
3495 }
3496 // Check for succeeding AddP which uses the same Base.
3497 // Otherwise we will run into the assertion above when visiting that guy.
3498 for (uint i = 0; i < n->outcnt(); ++i) {
3499 Node *out_i = n->raw_out(i);
3500 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3501 out_i->set_req(AddPNode::Base, nn);
3502 #ifdef ASSERT
3503 for (uint j = 0; j < out_i->outcnt(); ++j) {
3504 Node *out_j = out_i->raw_out(j);
3505 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3506 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3507 }
3508 #endif
3509 }
3510 }
3511 n->set_req(AddPNode::Base, nn);
3512 n->set_req(AddPNode::Address, nn);
3513 if (addp->outcnt() == 0) {
3514 addp->disconnect_inputs(this);
3515 }
3516 }
3517 }
3518 }
3519 #endif
3520 break;
3521 }
3522
3523 case Op_CastPP: {
3524 // Remove CastPP nodes to gain more freedom during scheduling but
3525 // keep the dependency they encode as control or precedence edges
3526 // (if control is set already) on memory operations. Some CastPP
3527 // nodes don't have a control (don't carry a dependency): skip
3528 // those.
3529 if (n->in(0) != nullptr) {
3530 ResourceMark rm;
3531 Unique_Node_List wq;
3532 wq.push(n);
3533 for (uint next = 0; next < wq.size(); ++next) {
3534 Node *m = wq.at(next);
3535 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3536 Node* use = m->fast_out(i);
3537 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3538 use->ensure_control_or_add_prec(n->in(0));
3539 } else {
3540 switch(use->Opcode()) {
3541 case Op_AddP:
3542 case Op_DecodeN:
3543 case Op_DecodeNKlass:
3544 case Op_CheckCastPP:
3545 case Op_CastPP:
3546 wq.push(use);
3547 break;
3548 }
3549 }
3550 }
3551 }
3552 }
3553 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3554 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3555 Node* in1 = n->in(1);
3556 const Type* t = n->bottom_type();
3557 Node* new_in1 = in1->clone();
3558 new_in1->as_DecodeN()->set_type(t);
3559
3560 if (!Matcher::narrow_oop_use_complex_address()) {
3561 //
3562 // x86, ARM and friends can handle 2 adds in addressing mode
3563 // and Matcher can fold a DecodeN node into address by using
3564 // a narrow oop directly and do implicit null check in address:
3565 //
3566 // [R12 + narrow_oop_reg<<3 + offset]
3567 // NullCheck narrow_oop_reg
3568 //
3569 // On other platforms (Sparc) we have to keep new DecodeN node and
3570 // use it to do implicit null check in address:
3571 //
3572 // decode_not_null narrow_oop_reg, base_reg
3573 // [base_reg + offset]
3574 // NullCheck base_reg
3575 //
3576 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3577 // to keep the information to which null check the new DecodeN node
3578 // corresponds to use it as value in implicit_null_check().
3579 //
3580 new_in1->set_req(0, n->in(0));
3581 }
3582
3583 n->subsume_by(new_in1, this);
3584 if (in1->outcnt() == 0) {
3585 in1->disconnect_inputs(this);
3586 }
3587 } else {
3588 n->subsume_by(n->in(1), this);
3589 if (n->outcnt() == 0) {
3590 n->disconnect_inputs(this);
3591 }
3592 }
3593 break;
3594 }
3595 case Op_CastII: {
3596 n->as_CastII()->remove_range_check_cast(this);
3597 break;
3598 }
3599 #ifdef _LP64
3600 case Op_CmpP:
3601 // Do this transformation here to preserve CmpPNode::sub() and
3602 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3603 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3604 Node* in1 = n->in(1);
3605 Node* in2 = n->in(2);
3606 if (!in1->is_DecodeNarrowPtr()) {
3607 in2 = in1;
3608 in1 = n->in(2);
3609 }
3610 assert(in1->is_DecodeNarrowPtr(), "sanity");
3611
3612 Node* new_in2 = nullptr;
3613 if (in2->is_DecodeNarrowPtr()) {
3614 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3615 new_in2 = in2->in(1);
3616 } else if (in2->Opcode() == Op_ConP) {
3617 const Type* t = in2->bottom_type();
3618 if (t == TypePtr::NULL_PTR) {
3619 assert(in1->is_DecodeN(), "compare klass to null?");
3620 // Don't convert CmpP null check into CmpN if compressed
3621 // oops implicit null check is not generated.
3622 // This will allow to generate normal oop implicit null check.
3623 if (Matcher::gen_narrow_oop_implicit_null_checks())
3624 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3625 //
3626 // This transformation together with CastPP transformation above
3627 // will generated code for implicit null checks for compressed oops.
3628 //
3629 // The original code after Optimize()
3630 //
3631 // LoadN memory, narrow_oop_reg
3632 // decode narrow_oop_reg, base_reg
3633 // CmpP base_reg, nullptr
3634 // CastPP base_reg // NotNull
3635 // Load [base_reg + offset], val_reg
3636 //
3637 // after these transformations will be
3638 //
3639 // LoadN memory, narrow_oop_reg
3640 // CmpN narrow_oop_reg, nullptr
3641 // decode_not_null narrow_oop_reg, base_reg
3642 // Load [base_reg + offset], val_reg
3643 //
3644 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3645 // since narrow oops can be used in debug info now (see the code in
3646 // final_graph_reshaping_walk()).
3647 //
3648 // At the end the code will be matched to
3649 // on x86:
3650 //
3651 // Load_narrow_oop memory, narrow_oop_reg
3652 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3653 // NullCheck narrow_oop_reg
3654 //
3655 // and on sparc:
3656 //
3657 // Load_narrow_oop memory, narrow_oop_reg
3658 // decode_not_null narrow_oop_reg, base_reg
3659 // Load [base_reg + offset], val_reg
3660 // NullCheck base_reg
3661 //
3662 } else if (t->isa_oopptr()) {
3663 new_in2 = ConNode::make(t->make_narrowoop());
3664 } else if (t->isa_klassptr()) {
3665 ciKlass* klass = t->is_klassptr()->exact_klass();
3666 if (klass->is_in_encoding_range()) {
3667 new_in2 = ConNode::make(t->make_narrowklass());
3668 }
3669 }
3670 }
3671 if (new_in2 != nullptr) {
3672 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3673 n->subsume_by(cmpN, this);
3674 if (in1->outcnt() == 0) {
3675 in1->disconnect_inputs(this);
3676 }
3677 if (in2->outcnt() == 0) {
3678 in2->disconnect_inputs(this);
3679 }
3680 }
3681 }
3682 break;
3683
3684 case Op_DecodeN:
3685 case Op_DecodeNKlass:
3686 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3687 // DecodeN could be pinned when it can't be fold into
3688 // an address expression, see the code for Op_CastPP above.
3689 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3690 break;
3691
3692 case Op_EncodeP:
3693 case Op_EncodePKlass: {
3694 Node* in1 = n->in(1);
3695 if (in1->is_DecodeNarrowPtr()) {
3696 n->subsume_by(in1->in(1), this);
3697 } else if (in1->Opcode() == Op_ConP) {
3698 const Type* t = in1->bottom_type();
3699 if (t == TypePtr::NULL_PTR) {
3700 assert(t->isa_oopptr(), "null klass?");
3701 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3702 } else if (t->isa_oopptr()) {
3703 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3704 } else if (t->isa_klassptr()) {
3705 ciKlass* klass = t->is_klassptr()->exact_klass();
3706 if (klass->is_in_encoding_range()) {
3707 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3708 } else {
3709 assert(false, "unencodable klass in ConP -> EncodeP");
3710 C->record_failure("unencodable klass in ConP -> EncodeP");
3711 }
3712 }
3713 }
3714 if (in1->outcnt() == 0) {
3715 in1->disconnect_inputs(this);
3716 }
3717 break;
3718 }
3719
3720 case Op_Proj: {
3721 if (OptimizeStringConcat || IncrementalInline) {
3722 ProjNode* proj = n->as_Proj();
3723 if (proj->_is_io_use) {
3724 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3725 // Separate projections were used for the exception path which
3726 // are normally removed by a late inline. If it wasn't inlined
3727 // then they will hang around and should just be replaced with
3728 // the original one. Merge them.
3729 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3730 if (non_io_proj != nullptr) {
3731 proj->subsume_by(non_io_proj , this);
3732 }
3733 }
3734 }
3735 break;
3736 }
3737
3738 case Op_Phi:
3739 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3740 // The EncodeP optimization may create Phi with the same edges
3741 // for all paths. It is not handled well by Register Allocator.
3742 Node* unique_in = n->in(1);
3743 assert(unique_in != nullptr, "");
3744 uint cnt = n->req();
3745 for (uint i = 2; i < cnt; i++) {
3746 Node* m = n->in(i);
3747 assert(m != nullptr, "");
3748 if (unique_in != m)
3749 unique_in = nullptr;
3750 }
3751 if (unique_in != nullptr) {
3752 n->subsume_by(unique_in, this);
3753 }
3754 }
3755 break;
3756
3757 #endif
3758
3759 case Op_ModI:
3760 handle_div_mod_op(n, T_INT, false);
3761 break;
3762
3763 case Op_ModL:
3764 handle_div_mod_op(n, T_LONG, false);
3765 break;
3766
3767 case Op_UModI:
3768 handle_div_mod_op(n, T_INT, true);
3769 break;
3770
3771 case Op_UModL:
3772 handle_div_mod_op(n, T_LONG, true);
3773 break;
3774
3775 case Op_LoadVector:
3776 case Op_StoreVector:
3777 #ifdef ASSERT
3778 // Add VerifyVectorAlignment node between adr and load / store.
3779 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3780 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3781 n->as_StoreVector()->must_verify_alignment();
3782 if (must_verify_alignment) {
3783 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3784 n->as_StoreVector()->memory_size();
3785 // The memory access should be aligned to the vector width in bytes.
3786 // However, the underlying array is possibly less well aligned, but at least
3787 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3788 // a loop we can expect at least the following alignment:
3789 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3790 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3791 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3792 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3793 jlong mask = guaranteed_alignment - 1;
3794 Node* mask_con = ConLNode::make(mask);
3795 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3796 n->set_req(MemNode::Address, va);
3797 }
3798 }
3799 #endif
3800 break;
3801
3802 case Op_LoadVectorGather:
3803 case Op_StoreVectorScatter:
3804 case Op_LoadVectorGatherMasked:
3805 case Op_StoreVectorScatterMasked:
3806 case Op_VectorCmpMasked:
3807 case Op_VectorMaskGen:
3808 case Op_LoadVectorMasked:
3809 case Op_StoreVectorMasked:
3810 break;
3811
3812 case Op_AddReductionVI:
3813 case Op_AddReductionVL:
3814 case Op_AddReductionVF:
3815 case Op_AddReductionVD:
3816 case Op_MulReductionVI:
3817 case Op_MulReductionVL:
3818 case Op_MulReductionVF:
3819 case Op_MulReductionVD:
3820 case Op_MinReductionV:
3821 case Op_MaxReductionV:
3822 case Op_UMinReductionV:
3823 case Op_UMaxReductionV:
3824 case Op_AndReductionV:
3825 case Op_OrReductionV:
3826 case Op_XorReductionV:
3827 break;
3828
3829 case Op_PackB:
3830 case Op_PackS:
3831 case Op_PackI:
3832 case Op_PackF:
3833 case Op_PackL:
3834 case Op_PackD:
3835 if (n->req()-1 > 2) {
3836 // Replace many operand PackNodes with a binary tree for matching
3837 PackNode* p = (PackNode*) n;
3838 Node* btp = p->binary_tree_pack(1, n->req());
3839 n->subsume_by(btp, this);
3840 }
3841 break;
3842 case Op_Loop:
3843 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3844 case Op_CountedLoop:
3845 case Op_LongCountedLoop:
3846 case Op_OuterStripMinedLoop:
3847 if (n->as_Loop()->is_inner_loop()) {
3848 frc.inc_inner_loop_count();
3849 }
3850 n->as_Loop()->verify_strip_mined(0);
3851 break;
3852 case Op_LShiftI:
3853 case Op_RShiftI:
3854 case Op_URShiftI:
3855 case Op_LShiftL:
3856 case Op_RShiftL:
3857 case Op_URShiftL:
3858 if (Matcher::need_masked_shift_count) {
3859 // The cpu's shift instructions don't restrict the count to the
3860 // lower 5/6 bits. We need to do the masking ourselves.
3861 Node* in2 = n->in(2);
3862 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3863 const TypeInt* t = in2->find_int_type();
3864 if (t != nullptr && t->is_con()) {
3865 juint shift = t->get_con();
3866 if (shift > mask) { // Unsigned cmp
3867 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3868 }
3869 } else {
3870 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3871 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3872 n->set_req(2, shift);
3873 }
3874 }
3875 if (in2->outcnt() == 0) { // Remove dead node
3876 in2->disconnect_inputs(this);
3877 }
3878 }
3879 break;
3880 case Op_MemBarStoreStore:
3881 case Op_MemBarRelease:
3882 // Break the link with AllocateNode: it is no longer useful and
3883 // confuses register allocation.
3884 if (n->req() > MemBarNode::Precedent) {
3885 n->set_req(MemBarNode::Precedent, top());
3886 }
3887 break;
3888 case Op_MemBarAcquire: {
3889 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3890 // At parse time, the trailing MemBarAcquire for a volatile load
3891 // is created with an edge to the load. After optimizations,
3892 // that input may be a chain of Phis. If those phis have no
3893 // other use, then the MemBarAcquire keeps them alive and
3894 // register allocation can be confused.
3895 dead_nodes.push(n->in(MemBarNode::Precedent));
3896 n->set_req(MemBarNode::Precedent, top());
3897 }
3898 break;
3899 }
3900 case Op_Blackhole:
3901 break;
3902 case Op_RangeCheck: {
3903 RangeCheckNode* rc = n->as_RangeCheck();
3904 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3905 n->subsume_by(iff, this);
3906 frc._tests.push(iff);
3907 break;
3908 }
3909 case Op_ConvI2L: {
3910 if (!Matcher::convi2l_type_required) {
3911 // Code generation on some platforms doesn't need accurate
3912 // ConvI2L types. Widening the type can help remove redundant
3913 // address computations.
3914 n->as_Type()->set_type(TypeLong::INT);
3915 ResourceMark rm;
3916 Unique_Node_List wq;
3917 wq.push(n);
3918 for (uint next = 0; next < wq.size(); next++) {
3919 Node *m = wq.at(next);
3920
3921 for(;;) {
3922 // Loop over all nodes with identical inputs edges as m
3923 Node* k = m->find_similar(m->Opcode());
3924 if (k == nullptr) {
3925 break;
3926 }
3927 // Push their uses so we get a chance to remove node made
3928 // redundant
3929 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3930 Node* u = k->fast_out(i);
3931 if (u->Opcode() == Op_LShiftL ||
3932 u->Opcode() == Op_AddL ||
3933 u->Opcode() == Op_SubL ||
3934 u->Opcode() == Op_AddP) {
3935 wq.push(u);
3936 }
3937 }
3938 // Replace all nodes with identical edges as m with m
3939 k->subsume_by(m, this);
3940 }
3941 }
3942 }
3943 break;
3944 }
3945 case Op_CmpUL: {
3946 if (!Matcher::has_match_rule(Op_CmpUL)) {
3947 // No support for unsigned long comparisons
3948 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3949 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3950 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3951 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3952 Node* andl = new AndLNode(orl, remove_sign_mask);
3953 Node* cmp = new CmpLNode(andl, n->in(2));
3954 n->subsume_by(cmp, this);
3955 }
3956 break;
3957 }
3958 #ifdef ASSERT
3959 case Op_ConNKlass: {
3960 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3961 ciKlass* klass = tp->is_klassptr()->exact_klass();
3962 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3963 break;
3964 }
3965 #endif
3966 default:
3967 assert(!n->is_Call(), "");
3968 assert(!n->is_Mem(), "");
3969 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3970 break;
3971 }
3972 }
3973
3974 //------------------------------final_graph_reshaping_walk---------------------
3975 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3976 // requires that the walk visits a node's inputs before visiting the node.
3977 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3978 Unique_Node_List sfpt;
3979
3980 frc._visited.set(root->_idx); // first, mark node as visited
3981 uint cnt = root->req();
3982 Node *n = root;
3983 uint i = 0;
3984 while (true) {
3985 if (i < cnt) {
3986 // Place all non-visited non-null inputs onto stack
3987 Node* m = n->in(i);
3988 ++i;
3989 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3990 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3991 // compute worst case interpreter size in case of a deoptimization
3992 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3993
3994 sfpt.push(m);
3995 }
3996 cnt = m->req();
3997 nstack.push(n, i); // put on stack parent and next input's index
3998 n = m;
3999 i = 0;
4000 }
4001 } else {
4002 // Now do post-visit work
4003 final_graph_reshaping_impl(n, frc, dead_nodes);
4004 if (nstack.is_empty())
4005 break; // finished
4006 n = nstack.node(); // Get node from stack
4007 cnt = n->req();
4008 i = nstack.index();
4009 nstack.pop(); // Shift to the next node on stack
4010 }
4011 }
4012
4013 // Skip next transformation if compressed oops are not used.
4014 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4015 (!UseCompressedOops && !UseCompressedClassPointers))
4016 return;
4017
4018 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4019 // It could be done for an uncommon traps or any safepoints/calls
4020 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4021 while (sfpt.size() > 0) {
4022 n = sfpt.pop();
4023 JVMState *jvms = n->as_SafePoint()->jvms();
4024 assert(jvms != nullptr, "sanity");
4025 int start = jvms->debug_start();
4026 int end = n->req();
4027 bool is_uncommon = (n->is_CallStaticJava() &&
4028 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4029 for (int j = start; j < end; j++) {
4030 Node* in = n->in(j);
4031 if (in->is_DecodeNarrowPtr()) {
4032 bool safe_to_skip = true;
4033 if (!is_uncommon ) {
4034 // Is it safe to skip?
4035 for (uint i = 0; i < in->outcnt(); i++) {
4036 Node* u = in->raw_out(i);
4037 if (!u->is_SafePoint() ||
4038 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4039 safe_to_skip = false;
4040 }
4041 }
4042 }
4043 if (safe_to_skip) {
4044 n->set_req(j, in->in(1));
4045 }
4046 if (in->outcnt() == 0) {
4047 in->disconnect_inputs(this);
4048 }
4049 }
4050 }
4051 }
4052 }
4053
4054 //------------------------------final_graph_reshaping--------------------------
4055 // Final Graph Reshaping.
4056 //
4057 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4058 // and not commoned up and forced early. Must come after regular
4059 // optimizations to avoid GVN undoing the cloning. Clone constant
4060 // inputs to Loop Phis; these will be split by the allocator anyways.
4061 // Remove Opaque nodes.
4062 // (2) Move last-uses by commutative operations to the left input to encourage
4063 // Intel update-in-place two-address operations and better register usage
4064 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4065 // calls canonicalizing them back.
4066 // (3) Count the number of double-precision FP ops, single-precision FP ops
4067 // and call sites. On Intel, we can get correct rounding either by
4068 // forcing singles to memory (requires extra stores and loads after each
4069 // FP bytecode) or we can set a rounding mode bit (requires setting and
4070 // clearing the mode bit around call sites). The mode bit is only used
4071 // if the relative frequency of single FP ops to calls is low enough.
4072 // This is a key transform for SPEC mpeg_audio.
4073 // (4) Detect infinite loops; blobs of code reachable from above but not
4074 // below. Several of the Code_Gen algorithms fail on such code shapes,
4075 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4076 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4077 // Detection is by looking for IfNodes where only 1 projection is
4078 // reachable from below or CatchNodes missing some targets.
4079 // (5) Assert for insane oop offsets in debug mode.
4080
4081 bool Compile::final_graph_reshaping() {
4082 // an infinite loop may have been eliminated by the optimizer,
4083 // in which case the graph will be empty.
4084 if (root()->req() == 1) {
4085 // Do not compile method that is only a trivial infinite loop,
4086 // since the content of the loop may have been eliminated.
4087 record_method_not_compilable("trivial infinite loop");
4088 return true;
4089 }
4090
4091 // Expensive nodes have their control input set to prevent the GVN
4092 // from freely commoning them. There's no GVN beyond this point so
4093 // no need to keep the control input. We want the expensive nodes to
4094 // be freely moved to the least frequent code path by gcm.
4095 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4096 for (int i = 0; i < expensive_count(); i++) {
4097 _expensive_nodes.at(i)->set_req(0, nullptr);
4098 }
4099
4100 Final_Reshape_Counts frc;
4101
4102 // Visit everybody reachable!
4103 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4104 Node_Stack nstack(live_nodes() >> 1);
4105 Unique_Node_List dead_nodes;
4106 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4107
4108 // Check for unreachable (from below) code (i.e., infinite loops).
4109 for( uint i = 0; i < frc._tests.size(); i++ ) {
4110 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4111 // Get number of CFG targets.
4112 // Note that PCTables include exception targets after calls.
4113 uint required_outcnt = n->required_outcnt();
4114 if (n->outcnt() != required_outcnt) {
4115 // Check for a few special cases. Rethrow Nodes never take the
4116 // 'fall-thru' path, so expected kids is 1 less.
4117 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4118 if (n->in(0)->in(0)->is_Call()) {
4119 CallNode* call = n->in(0)->in(0)->as_Call();
4120 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4121 required_outcnt--; // Rethrow always has 1 less kid
4122 } else if (call->req() > TypeFunc::Parms &&
4123 call->is_CallDynamicJava()) {
4124 // Check for null receiver. In such case, the optimizer has
4125 // detected that the virtual call will always result in a null
4126 // pointer exception. The fall-through projection of this CatchNode
4127 // will not be populated.
4128 Node* arg0 = call->in(TypeFunc::Parms);
4129 if (arg0->is_Type() &&
4130 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4131 required_outcnt--;
4132 }
4133 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4134 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4135 // Check for illegal array length. In such case, the optimizer has
4136 // detected that the allocation attempt will always result in an
4137 // exception. There is no fall-through projection of this CatchNode .
4138 assert(call->is_CallStaticJava(), "static call expected");
4139 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4140 uint valid_length_test_input = call->req() - 1;
4141 Node* valid_length_test = call->in(valid_length_test_input);
4142 call->del_req(valid_length_test_input);
4143 if (valid_length_test->find_int_con(1) == 0) {
4144 required_outcnt--;
4145 }
4146 dead_nodes.push(valid_length_test);
4147 assert(n->outcnt() == required_outcnt, "malformed control flow");
4148 continue;
4149 }
4150 }
4151 }
4152
4153 // Recheck with a better notion of 'required_outcnt'
4154 if (n->outcnt() != required_outcnt) {
4155 record_method_not_compilable("malformed control flow");
4156 return true; // Not all targets reachable!
4157 }
4158 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4159 CallNode* call = n->in(0)->in(0)->as_Call();
4160 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4161 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4162 assert(call->is_CallStaticJava(), "static call expected");
4163 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4164 uint valid_length_test_input = call->req() - 1;
4165 dead_nodes.push(call->in(valid_length_test_input));
4166 call->del_req(valid_length_test_input); // valid length test useless now
4167 }
4168 }
4169 // Check that I actually visited all kids. Unreached kids
4170 // must be infinite loops.
4171 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4172 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4173 record_method_not_compilable("infinite loop");
4174 return true; // Found unvisited kid; must be unreach
4175 }
4176
4177 // Here so verification code in final_graph_reshaping_walk()
4178 // always see an OuterStripMinedLoopEnd
4179 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4180 IfNode* init_iff = n->as_If();
4181 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4182 n->subsume_by(iff, this);
4183 }
4184 }
4185
4186 while (dead_nodes.size() > 0) {
4187 Node* m = dead_nodes.pop();
4188 if (m->outcnt() == 0 && m != top()) {
4189 for (uint j = 0; j < m->req(); j++) {
4190 Node* in = m->in(j);
4191 if (in != nullptr) {
4192 dead_nodes.push(in);
4193 }
4194 }
4195 m->disconnect_inputs(this);
4196 }
4197 }
4198
4199 set_java_calls(frc.get_java_call_count());
4200 set_inner_loops(frc.get_inner_loop_count());
4201
4202 // No infinite loops, no reason to bail out.
4203 return false;
4204 }
4205
4206 //-----------------------------too_many_traps----------------------------------
4207 // Report if there are too many traps at the current method and bci.
4208 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4209 bool Compile::too_many_traps(ciMethod* method,
4210 int bci,
4211 Deoptimization::DeoptReason reason) {
4212 ciMethodData* md = method->method_data();
4213 if (md->is_empty()) {
4214 // Assume the trap has not occurred, or that it occurred only
4215 // because of a transient condition during start-up in the interpreter.
4216 return false;
4217 }
4218 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4219 if (md->has_trap_at(bci, m, reason) != 0) {
4220 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4221 // Also, if there are multiple reasons, or if there is no per-BCI record,
4222 // assume the worst.
4223 if (log())
4224 log()->elem("observe trap='%s' count='%d'",
4225 Deoptimization::trap_reason_name(reason),
4226 md->trap_count(reason));
4227 return true;
4228 } else {
4229 // Ignore method/bci and see if there have been too many globally.
4230 return too_many_traps(reason, md);
4231 }
4232 }
4233
4234 // Less-accurate variant which does not require a method and bci.
4235 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4236 ciMethodData* logmd) {
4237 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4238 // Too many traps globally.
4239 // Note that we use cumulative trap_count, not just md->trap_count.
4240 if (log()) {
4241 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4242 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4243 Deoptimization::trap_reason_name(reason),
4244 mcount, trap_count(reason));
4245 }
4246 return true;
4247 } else {
4248 // The coast is clear.
4249 return false;
4250 }
4251 }
4252
4253 //--------------------------too_many_recompiles--------------------------------
4254 // Report if there are too many recompiles at the current method and bci.
4255 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4256 // Is not eager to return true, since this will cause the compiler to use
4257 // Action_none for a trap point, to avoid too many recompilations.
4258 bool Compile::too_many_recompiles(ciMethod* method,
4259 int bci,
4260 Deoptimization::DeoptReason reason) {
4261 ciMethodData* md = method->method_data();
4262 if (md->is_empty()) {
4263 // Assume the trap has not occurred, or that it occurred only
4264 // because of a transient condition during start-up in the interpreter.
4265 return false;
4266 }
4267 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4268 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4269 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4270 Deoptimization::DeoptReason per_bc_reason
4271 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4272 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4273 if ((per_bc_reason == Deoptimization::Reason_none
4274 || md->has_trap_at(bci, m, reason) != 0)
4275 // The trap frequency measure we care about is the recompile count:
4276 && md->trap_recompiled_at(bci, m)
4277 && md->overflow_recompile_count() >= bc_cutoff) {
4278 // Do not emit a trap here if it has already caused recompilations.
4279 // Also, if there are multiple reasons, or if there is no per-BCI record,
4280 // assume the worst.
4281 if (log())
4282 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4283 Deoptimization::trap_reason_name(reason),
4284 md->trap_count(reason),
4285 md->overflow_recompile_count());
4286 return true;
4287 } else if (trap_count(reason) != 0
4288 && decompile_count() >= m_cutoff) {
4289 // Too many recompiles globally, and we have seen this sort of trap.
4290 // Use cumulative decompile_count, not just md->decompile_count.
4291 if (log())
4292 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4293 Deoptimization::trap_reason_name(reason),
4294 md->trap_count(reason), trap_count(reason),
4295 md->decompile_count(), decompile_count());
4296 return true;
4297 } else {
4298 // The coast is clear.
4299 return false;
4300 }
4301 }
4302
4303 // Compute when not to trap. Used by matching trap based nodes and
4304 // NullCheck optimization.
4305 void Compile::set_allowed_deopt_reasons() {
4306 _allowed_reasons = 0;
4307 if (is_method_compilation()) {
4308 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4309 assert(rs < BitsPerInt, "recode bit map");
4310 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4311 _allowed_reasons |= nth_bit(rs);
4312 }
4313 }
4314 }
4315 }
4316
4317 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4318 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4319 }
4320
4321 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4322 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4323 }
4324
4325 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4326 if (holder->is_initialized()) {
4327 return false;
4328 }
4329 if (holder->is_being_initialized()) {
4330 if (accessing_method->holder() == holder) {
4331 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4332 // <init>, or a static method. In all those cases, there was an initialization
4333 // barrier on the holder klass passed.
4334 if (accessing_method->is_static_initializer() ||
4335 accessing_method->is_object_initializer() ||
4336 accessing_method->is_static()) {
4337 return false;
4338 }
4339 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4340 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4341 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4342 // child class can become fully initialized while its parent class is still being initialized.
4343 if (accessing_method->is_static_initializer()) {
4344 return false;
4345 }
4346 }
4347 ciMethod* root = method(); // the root method of compilation
4348 if (root != accessing_method) {
4349 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4350 }
4351 }
4352 return true;
4353 }
4354
4355 #ifndef PRODUCT
4356 //------------------------------verify_bidirectional_edges---------------------
4357 // For each input edge to a node (ie - for each Use-Def edge), verify that
4358 // there is a corresponding Def-Use edge.
4359 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4360 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4361 uint stack_size = live_nodes() >> 4;
4362 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4363 if (root_and_safepoints != nullptr) {
4364 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4365 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4366 Node* root_or_safepoint = root_and_safepoints->at(i);
4367 // If the node is a safepoint, let's check if it still has a control input
4368 // Lack of control input signifies that this node was killed by CCP or
4369 // recursively by remove_globally_dead_node and it shouldn't be a starting
4370 // point.
4371 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4372 nstack.push(root_or_safepoint);
4373 }
4374 }
4375 } else {
4376 nstack.push(_root);
4377 }
4378
4379 while (nstack.size() > 0) {
4380 Node* n = nstack.pop();
4381 if (visited.member(n)) {
4382 continue;
4383 }
4384 visited.push(n);
4385
4386 // Walk over all input edges, checking for correspondence
4387 uint length = n->len();
4388 for (uint i = 0; i < length; i++) {
4389 Node* in = n->in(i);
4390 if (in != nullptr && !visited.member(in)) {
4391 nstack.push(in); // Put it on stack
4392 }
4393 if (in != nullptr && !in->is_top()) {
4394 // Count instances of `next`
4395 int cnt = 0;
4396 for (uint idx = 0; idx < in->_outcnt; idx++) {
4397 if (in->_out[idx] == n) {
4398 cnt++;
4399 }
4400 }
4401 assert(cnt > 0, "Failed to find Def-Use edge.");
4402 // Check for duplicate edges
4403 // walk the input array downcounting the input edges to n
4404 for (uint j = 0; j < length; j++) {
4405 if (n->in(j) == in) {
4406 cnt--;
4407 }
4408 }
4409 assert(cnt == 0, "Mismatched edge count.");
4410 } else if (in == nullptr) {
4411 assert(i == 0 || i >= n->req() ||
4412 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4413 (n->is_Unlock() && i == (n->req() - 1)) ||
4414 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4415 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4416 } else {
4417 assert(in->is_top(), "sanity");
4418 // Nothing to check.
4419 }
4420 }
4421 }
4422 }
4423
4424 //------------------------------verify_graph_edges---------------------------
4425 // Walk the Graph and verify that there is a one-to-one correspondence
4426 // between Use-Def edges and Def-Use edges in the graph.
4427 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4428 if (VerifyGraphEdges) {
4429 Unique_Node_List visited;
4430
4431 // Call graph walk to check edges
4432 verify_bidirectional_edges(visited, root_and_safepoints);
4433 if (no_dead_code) {
4434 // Now make sure that no visited node is used by an unvisited node.
4435 bool dead_nodes = false;
4436 Unique_Node_List checked;
4437 while (visited.size() > 0) {
4438 Node* n = visited.pop();
4439 checked.push(n);
4440 for (uint i = 0; i < n->outcnt(); i++) {
4441 Node* use = n->raw_out(i);
4442 if (checked.member(use)) continue; // already checked
4443 if (visited.member(use)) continue; // already in the graph
4444 if (use->is_Con()) continue; // a dead ConNode is OK
4445 // At this point, we have found a dead node which is DU-reachable.
4446 if (!dead_nodes) {
4447 tty->print_cr("*** Dead nodes reachable via DU edges:");
4448 dead_nodes = true;
4449 }
4450 use->dump(2);
4451 tty->print_cr("---");
4452 checked.push(use); // No repeats; pretend it is now checked.
4453 }
4454 }
4455 assert(!dead_nodes, "using nodes must be reachable from root");
4456 }
4457 }
4458 }
4459 #endif
4460
4461 // The Compile object keeps track of failure reasons separately from the ciEnv.
4462 // This is required because there is not quite a 1-1 relation between the
4463 // ciEnv and its compilation task and the Compile object. Note that one
4464 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4465 // to backtrack and retry without subsuming loads. Other than this backtracking
4466 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4467 // by the logic in C2Compiler.
4468 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4469 if (log() != nullptr) {
4470 log()->elem("failure reason='%s' phase='compile'", reason);
4471 }
4472 if (_failure_reason.get() == nullptr) {
4473 // Record the first failure reason.
4474 _failure_reason.set(reason);
4475 if (CaptureBailoutInformation) {
4476 _first_failure_details = new CompilationFailureInfo(reason);
4477 }
4478 } else {
4479 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4480 }
4481
4482 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4483 C->print_method(PHASE_FAILURE, 1);
4484 }
4485 _root = nullptr; // flush the graph, too
4486 }
4487
4488 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4489 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4490 _compile(Compile::current()),
4491 _log(nullptr),
4492 _dolog(CITimeVerbose)
4493 {
4494 assert(_compile != nullptr, "sanity check");
4495 assert(id != PhaseTraceId::_t_none, "Don't use none");
4496 if (_dolog) {
4497 _log = _compile->log();
4498 }
4499 if (_log != nullptr) {
4500 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4501 _log->stamp();
4502 _log->end_head();
4503 }
4504
4505 // Inform memory statistic, if enabled
4506 if (CompilationMemoryStatistic::enabled()) {
4507 CompilationMemoryStatistic::on_phase_start((int)id, name);
4508 }
4509 }
4510
4511 Compile::TracePhase::TracePhase(PhaseTraceId id)
4512 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4513
4514 Compile::TracePhase::~TracePhase() {
4515
4516 // Inform memory statistic, if enabled
4517 if (CompilationMemoryStatistic::enabled()) {
4518 CompilationMemoryStatistic::on_phase_end();
4519 }
4520
4521 if (_compile->failing_internal()) {
4522 if (_log != nullptr) {
4523 _log->done("phase");
4524 }
4525 return; // timing code, not stressing bailouts.
4526 }
4527 #ifdef ASSERT
4528 if (PrintIdealNodeCount) {
4529 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4530 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4531 }
4532
4533 if (VerifyIdealNodeCount) {
4534 _compile->print_missing_nodes();
4535 }
4536 #endif
4537
4538 if (_log != nullptr) {
4539 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4540 }
4541 }
4542
4543 //----------------------------static_subtype_check-----------------------------
4544 // Shortcut important common cases when superklass is exact:
4545 // (0) superklass is java.lang.Object (can occur in reflective code)
4546 // (1) subklass is already limited to a subtype of superklass => always ok
4547 // (2) subklass does not overlap with superklass => always fail
4548 // (3) superklass has NO subtypes and we can check with a simple compare.
4549 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4550 if (skip) {
4551 return SSC_full_test; // Let caller generate the general case.
4552 }
4553
4554 if (subk->is_java_subtype_of(superk)) {
4555 return SSC_always_true; // (0) and (1) this test cannot fail
4556 }
4557
4558 if (!subk->maybe_java_subtype_of(superk)) {
4559 return SSC_always_false; // (2) true path dead; no dynamic test needed
4560 }
4561
4562 const Type* superelem = superk;
4563 if (superk->isa_aryklassptr()) {
4564 int ignored;
4565 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4566 }
4567
4568 if (superelem->isa_instklassptr()) {
4569 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4570 if (!ik->has_subklass()) {
4571 if (!ik->is_final()) {
4572 // Add a dependency if there is a chance of a later subclass.
4573 dependencies()->assert_leaf_type(ik);
4574 }
4575 if (!superk->maybe_java_subtype_of(subk)) {
4576 return SSC_always_false;
4577 }
4578 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4579 }
4580 } else {
4581 // A primitive array type has no subtypes.
4582 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4583 }
4584
4585 return SSC_full_test;
4586 }
4587
4588 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4589 #ifdef _LP64
4590 // The scaled index operand to AddP must be a clean 64-bit value.
4591 // Java allows a 32-bit int to be incremented to a negative
4592 // value, which appears in a 64-bit register as a large
4593 // positive number. Using that large positive number as an
4594 // operand in pointer arithmetic has bad consequences.
4595 // On the other hand, 32-bit overflow is rare, and the possibility
4596 // can often be excluded, if we annotate the ConvI2L node with
4597 // a type assertion that its value is known to be a small positive
4598 // number. (The prior range check has ensured this.)
4599 // This assertion is used by ConvI2LNode::Ideal.
4600 int index_max = max_jint - 1; // array size is max_jint, index is one less
4601 if (sizetype != nullptr && sizetype->_hi > 0) {
4602 index_max = sizetype->_hi - 1;
4603 }
4604 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4605 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4606 #endif
4607 return idx;
4608 }
4609
4610 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4611 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4612 if (ctrl != nullptr) {
4613 // Express control dependency by a CastII node with a narrow type.
4614 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4615 // node from floating above the range check during loop optimizations. Otherwise, the
4616 // ConvI2L node may be eliminated independently of the range check, causing the data path
4617 // to become TOP while the control path is still there (although it's unreachable).
4618 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
4619 value = phase->transform(value);
4620 }
4621 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4622 return phase->transform(new ConvI2LNode(value, ltype));
4623 }
4624
4625 void Compile::dump_print_inlining() {
4626 inline_printer()->print_on(tty);
4627 }
4628
4629 void Compile::log_late_inline(CallGenerator* cg) {
4630 if (log() != nullptr) {
4631 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4632 cg->unique_id());
4633 JVMState* p = cg->call_node()->jvms();
4634 while (p != nullptr) {
4635 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4636 p = p->caller();
4637 }
4638 log()->tail("late_inline");
4639 }
4640 }
4641
4642 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4643 log_late_inline(cg);
4644 if (log() != nullptr) {
4645 log()->inline_fail(msg);
4646 }
4647 }
4648
4649 void Compile::log_inline_id(CallGenerator* cg) {
4650 if (log() != nullptr) {
4651 // The LogCompilation tool needs a unique way to identify late
4652 // inline call sites. This id must be unique for this call site in
4653 // this compilation. Try to have it unique across compilations as
4654 // well because it can be convenient when grepping through the log
4655 // file.
4656 // Distinguish OSR compilations from others in case CICountOSR is
4657 // on.
4658 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4659 cg->set_unique_id(id);
4660 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4661 }
4662 }
4663
4664 void Compile::log_inline_failure(const char* msg) {
4665 if (C->log() != nullptr) {
4666 C->log()->inline_fail(msg);
4667 }
4668 }
4669
4670
4671 // Dump inlining replay data to the stream.
4672 // Don't change thread state and acquire any locks.
4673 void Compile::dump_inline_data(outputStream* out) {
4674 InlineTree* inl_tree = ilt();
4675 if (inl_tree != nullptr) {
4676 out->print(" inline %d", inl_tree->count());
4677 inl_tree->dump_replay_data(out);
4678 }
4679 }
4680
4681 void Compile::dump_inline_data_reduced(outputStream* out) {
4682 assert(ReplayReduce, "");
4683
4684 InlineTree* inl_tree = ilt();
4685 if (inl_tree == nullptr) {
4686 return;
4687 }
4688 // Enable iterative replay file reduction
4689 // Output "compile" lines for depth 1 subtrees,
4690 // simulating that those trees were compiled
4691 // instead of inlined.
4692 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4693 InlineTree* sub = inl_tree->subtrees().at(i);
4694 if (sub->inline_level() != 1) {
4695 continue;
4696 }
4697
4698 ciMethod* method = sub->method();
4699 int entry_bci = -1;
4700 int comp_level = env()->task()->comp_level();
4701 out->print("compile ");
4702 method->dump_name_as_ascii(out);
4703 out->print(" %d %d", entry_bci, comp_level);
4704 out->print(" inline %d", sub->count());
4705 sub->dump_replay_data(out, -1);
4706 out->cr();
4707 }
4708 }
4709
4710 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4711 if (n1->Opcode() < n2->Opcode()) return -1;
4712 else if (n1->Opcode() > n2->Opcode()) return 1;
4713
4714 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4715 for (uint i = 1; i < n1->req(); i++) {
4716 if (n1->in(i) < n2->in(i)) return -1;
4717 else if (n1->in(i) > n2->in(i)) return 1;
4718 }
4719
4720 return 0;
4721 }
4722
4723 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4724 Node* n1 = *n1p;
4725 Node* n2 = *n2p;
4726
4727 return cmp_expensive_nodes(n1, n2);
4728 }
4729
4730 void Compile::sort_expensive_nodes() {
4731 if (!expensive_nodes_sorted()) {
4732 _expensive_nodes.sort(cmp_expensive_nodes);
4733 }
4734 }
4735
4736 bool Compile::expensive_nodes_sorted() const {
4737 for (int i = 1; i < _expensive_nodes.length(); i++) {
4738 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4739 return false;
4740 }
4741 }
4742 return true;
4743 }
4744
4745 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4746 if (_expensive_nodes.length() == 0) {
4747 return false;
4748 }
4749
4750 assert(OptimizeExpensiveOps, "optimization off?");
4751
4752 // Take this opportunity to remove dead nodes from the list
4753 int j = 0;
4754 for (int i = 0; i < _expensive_nodes.length(); i++) {
4755 Node* n = _expensive_nodes.at(i);
4756 if (!n->is_unreachable(igvn)) {
4757 assert(n->is_expensive(), "should be expensive");
4758 _expensive_nodes.at_put(j, n);
4759 j++;
4760 }
4761 }
4762 _expensive_nodes.trunc_to(j);
4763
4764 // Then sort the list so that similar nodes are next to each other
4765 // and check for at least two nodes of identical kind with same data
4766 // inputs.
4767 sort_expensive_nodes();
4768
4769 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4770 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4771 return true;
4772 }
4773 }
4774
4775 return false;
4776 }
4777
4778 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4779 if (_expensive_nodes.length() == 0) {
4780 return;
4781 }
4782
4783 assert(OptimizeExpensiveOps, "optimization off?");
4784
4785 // Sort to bring similar nodes next to each other and clear the
4786 // control input of nodes for which there's only a single copy.
4787 sort_expensive_nodes();
4788
4789 int j = 0;
4790 int identical = 0;
4791 int i = 0;
4792 bool modified = false;
4793 for (; i < _expensive_nodes.length()-1; i++) {
4794 assert(j <= i, "can't write beyond current index");
4795 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4796 identical++;
4797 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4798 continue;
4799 }
4800 if (identical > 0) {
4801 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4802 identical = 0;
4803 } else {
4804 Node* n = _expensive_nodes.at(i);
4805 igvn.replace_input_of(n, 0, nullptr);
4806 igvn.hash_insert(n);
4807 modified = true;
4808 }
4809 }
4810 if (identical > 0) {
4811 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4812 } else if (_expensive_nodes.length() >= 1) {
4813 Node* n = _expensive_nodes.at(i);
4814 igvn.replace_input_of(n, 0, nullptr);
4815 igvn.hash_insert(n);
4816 modified = true;
4817 }
4818 _expensive_nodes.trunc_to(j);
4819 if (modified) {
4820 igvn.optimize();
4821 }
4822 }
4823
4824 void Compile::add_expensive_node(Node * n) {
4825 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4826 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4827 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4828 if (OptimizeExpensiveOps) {
4829 _expensive_nodes.append(n);
4830 } else {
4831 // Clear control input and let IGVN optimize expensive nodes if
4832 // OptimizeExpensiveOps is off.
4833 n->set_req(0, nullptr);
4834 }
4835 }
4836
4837 /**
4838 * Track coarsened Lock and Unlock nodes.
4839 */
4840
4841 class Lock_List : public Node_List {
4842 uint _origin_cnt;
4843 public:
4844 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4845 uint origin_cnt() const { return _origin_cnt; }
4846 };
4847
4848 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4849 int length = locks.length();
4850 if (length > 0) {
4851 // Have to keep this list until locks elimination during Macro nodes elimination.
4852 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4853 AbstractLockNode* alock = locks.at(0);
4854 BoxLockNode* box = alock->box_node()->as_BoxLock();
4855 for (int i = 0; i < length; i++) {
4856 AbstractLockNode* lock = locks.at(i);
4857 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4858 locks_list->push(lock);
4859 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4860 if (this_box != box) {
4861 // Locking regions (BoxLock) could be Unbalanced here:
4862 // - its coarsened locks were eliminated in earlier
4863 // macro nodes elimination followed by loop unroll
4864 // - it is OSR locking region (no Lock node)
4865 // Preserve Unbalanced status in such cases.
4866 if (!this_box->is_unbalanced()) {
4867 this_box->set_coarsened();
4868 }
4869 if (!box->is_unbalanced()) {
4870 box->set_coarsened();
4871 }
4872 }
4873 }
4874 _coarsened_locks.append(locks_list);
4875 }
4876 }
4877
4878 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4879 int count = coarsened_count();
4880 for (int i = 0; i < count; i++) {
4881 Node_List* locks_list = _coarsened_locks.at(i);
4882 for (uint j = 0; j < locks_list->size(); j++) {
4883 Node* lock = locks_list->at(j);
4884 assert(lock->is_AbstractLock(), "sanity");
4885 if (!useful.member(lock)) {
4886 locks_list->yank(lock);
4887 }
4888 }
4889 }
4890 }
4891
4892 void Compile::remove_coarsened_lock(Node* n) {
4893 if (n->is_AbstractLock()) {
4894 int count = coarsened_count();
4895 for (int i = 0; i < count; i++) {
4896 Node_List* locks_list = _coarsened_locks.at(i);
4897 locks_list->yank(n);
4898 }
4899 }
4900 }
4901
4902 bool Compile::coarsened_locks_consistent() {
4903 int count = coarsened_count();
4904 for (int i = 0; i < count; i++) {
4905 bool unbalanced = false;
4906 bool modified = false; // track locks kind modifications
4907 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4908 uint size = locks_list->size();
4909 if (size == 0) {
4910 unbalanced = false; // All locks were eliminated - good
4911 } else if (size != locks_list->origin_cnt()) {
4912 unbalanced = true; // Some locks were removed from list
4913 } else {
4914 for (uint j = 0; j < size; j++) {
4915 Node* lock = locks_list->at(j);
4916 // All nodes in group should have the same state (modified or not)
4917 if (!lock->as_AbstractLock()->is_coarsened()) {
4918 if (j == 0) {
4919 // first on list was modified, the rest should be too for consistency
4920 modified = true;
4921 } else if (!modified) {
4922 // this lock was modified but previous locks on the list were not
4923 unbalanced = true;
4924 break;
4925 }
4926 } else if (modified) {
4927 // previous locks on list were modified but not this lock
4928 unbalanced = true;
4929 break;
4930 }
4931 }
4932 }
4933 if (unbalanced) {
4934 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4935 #ifdef ASSERT
4936 if (PrintEliminateLocks) {
4937 tty->print_cr("=== unbalanced coarsened locks ===");
4938 for (uint l = 0; l < size; l++) {
4939 locks_list->at(l)->dump();
4940 }
4941 }
4942 #endif
4943 record_failure(C2Compiler::retry_no_locks_coarsening());
4944 return false;
4945 }
4946 }
4947 return true;
4948 }
4949
4950 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4951 // locks coarsening optimization removed Lock/Unlock nodes from them.
4952 // Such regions become unbalanced because coarsening only removes part
4953 // of Lock/Unlock nodes in region. As result we can't execute other
4954 // locks elimination optimizations which assume all code paths have
4955 // corresponding pair of Lock/Unlock nodes - they are balanced.
4956 void Compile::mark_unbalanced_boxes() const {
4957 int count = coarsened_count();
4958 for (int i = 0; i < count; i++) {
4959 Node_List* locks_list = _coarsened_locks.at(i);
4960 uint size = locks_list->size();
4961 if (size > 0) {
4962 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4963 BoxLockNode* box = alock->box_node()->as_BoxLock();
4964 if (alock->is_coarsened()) {
4965 // coarsened_locks_consistent(), which is called before this method, verifies
4966 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4967 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4968 for (uint j = 1; j < size; j++) {
4969 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4970 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4971 if (box != this_box) {
4972 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4973 box->set_unbalanced();
4974 this_box->set_unbalanced();
4975 }
4976 }
4977 }
4978 }
4979 }
4980 }
4981
4982 /**
4983 * Remove the speculative part of types and clean up the graph
4984 */
4985 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4986 if (UseTypeSpeculation) {
4987 Unique_Node_List worklist;
4988 worklist.push(root());
4989 int modified = 0;
4990 // Go over all type nodes that carry a speculative type, drop the
4991 // speculative part of the type and enqueue the node for an igvn
4992 // which may optimize it out.
4993 for (uint next = 0; next < worklist.size(); ++next) {
4994 Node *n = worklist.at(next);
4995 if (n->is_Type()) {
4996 TypeNode* tn = n->as_Type();
4997 const Type* t = tn->type();
4998 const Type* t_no_spec = t->remove_speculative();
4999 if (t_no_spec != t) {
5000 bool in_hash = igvn.hash_delete(n);
5001 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5002 tn->set_type(t_no_spec);
5003 igvn.hash_insert(n);
5004 igvn._worklist.push(n); // give it a chance to go away
5005 modified++;
5006 }
5007 }
5008 // Iterate over outs - endless loops is unreachable from below
5009 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5010 Node *m = n->fast_out(i);
5011 if (not_a_node(m)) {
5012 continue;
5013 }
5014 worklist.push(m);
5015 }
5016 }
5017 // Drop the speculative part of all types in the igvn's type table
5018 igvn.remove_speculative_types();
5019 if (modified > 0) {
5020 igvn.optimize();
5021 if (failing()) return;
5022 }
5023 #ifdef ASSERT
5024 // Verify that after the IGVN is over no speculative type has resurfaced
5025 worklist.clear();
5026 worklist.push(root());
5027 for (uint next = 0; next < worklist.size(); ++next) {
5028 Node *n = worklist.at(next);
5029 const Type* t = igvn.type_or_null(n);
5030 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5031 if (n->is_Type()) {
5032 t = n->as_Type()->type();
5033 assert(t == t->remove_speculative(), "no more speculative types");
5034 }
5035 // Iterate over outs - endless loops is unreachable from below
5036 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5037 Node *m = n->fast_out(i);
5038 if (not_a_node(m)) {
5039 continue;
5040 }
5041 worklist.push(m);
5042 }
5043 }
5044 igvn.check_no_speculative_types();
5045 #endif
5046 }
5047 }
5048
5049 // Auxiliary methods to support randomized stressing/fuzzing.
5050
5051 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5052 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5053 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5054 FLAG_SET_ERGO(StressSeed, _stress_seed);
5055 } else {
5056 _stress_seed = StressSeed;
5057 }
5058 if (_log != nullptr) {
5059 _log->elem("stress_test seed='%u'", _stress_seed);
5060 }
5061 }
5062
5063 int Compile::random() {
5064 _stress_seed = os::next_random(_stress_seed);
5065 return static_cast<int>(_stress_seed);
5066 }
5067
5068 // This method can be called the arbitrary number of times, with current count
5069 // as the argument. The logic allows selecting a single candidate from the
5070 // running list of candidates as follows:
5071 // int count = 0;
5072 // Cand* selected = null;
5073 // while(cand = cand->next()) {
5074 // if (randomized_select(++count)) {
5075 // selected = cand;
5076 // }
5077 // }
5078 //
5079 // Including count equalizes the chances any candidate is "selected".
5080 // This is useful when we don't have the complete list of candidates to choose
5081 // from uniformly. In this case, we need to adjust the randomicity of the
5082 // selection, or else we will end up biasing the selection towards the latter
5083 // candidates.
5084 //
5085 // Quick back-envelope calculation shows that for the list of n candidates
5086 // the equal probability for the candidate to persist as "best" can be
5087 // achieved by replacing it with "next" k-th candidate with the probability
5088 // of 1/k. It can be easily shown that by the end of the run, the
5089 // probability for any candidate is converged to 1/n, thus giving the
5090 // uniform distribution among all the candidates.
5091 //
5092 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5093 #define RANDOMIZED_DOMAIN_POW 29
5094 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5095 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5096 bool Compile::randomized_select(int count) {
5097 assert(count > 0, "only positive");
5098 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5099 }
5100
5101 #ifdef ASSERT
5102 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5103 bool Compile::fail_randomly() {
5104 if ((random() % StressBailoutMean) != 0) {
5105 return false;
5106 }
5107 record_failure("StressBailout");
5108 return true;
5109 }
5110
5111 bool Compile::failure_is_artificial() {
5112 return C->failure_reason_is("StressBailout");
5113 }
5114 #endif
5115
5116 CloneMap& Compile::clone_map() { return _clone_map; }
5117 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5118
5119 void NodeCloneInfo::dump_on(outputStream* st) const {
5120 st->print(" {%d:%d} ", idx(), gen());
5121 }
5122
5123 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5124 uint64_t val = value(old->_idx);
5125 NodeCloneInfo cio(val);
5126 assert(val != 0, "old node should be in the map");
5127 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5128 insert(nnn->_idx, cin.get());
5129 #ifndef PRODUCT
5130 if (is_debug()) {
5131 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5132 }
5133 #endif
5134 }
5135
5136 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5137 NodeCloneInfo cio(value(old->_idx));
5138 if (cio.get() == 0) {
5139 cio.set(old->_idx, 0);
5140 insert(old->_idx, cio.get());
5141 #ifndef PRODUCT
5142 if (is_debug()) {
5143 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5144 }
5145 #endif
5146 }
5147 clone(old, nnn, gen);
5148 }
5149
5150 int CloneMap::max_gen() const {
5151 int g = 0;
5152 DictI di(_dict);
5153 for(; di.test(); ++di) {
5154 int t = gen(di._key);
5155 if (g < t) {
5156 g = t;
5157 #ifndef PRODUCT
5158 if (is_debug()) {
5159 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5160 }
5161 #endif
5162 }
5163 }
5164 return g;
5165 }
5166
5167 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5168 uint64_t val = value(key);
5169 if (val != 0) {
5170 NodeCloneInfo ni(val);
5171 ni.dump_on(st);
5172 }
5173 }
5174
5175 void Compile::shuffle_macro_nodes() {
5176 shuffle_array(*C, _macro_nodes);
5177 }
5178
5179 // Move Allocate nodes to the start of the list
5180 void Compile::sort_macro_nodes() {
5181 int count = macro_count();
5182 int allocates = 0;
5183 for (int i = 0; i < count; i++) {
5184 Node* n = macro_node(i);
5185 if (n->is_Allocate()) {
5186 if (i != allocates) {
5187 Node* tmp = macro_node(allocates);
5188 _macro_nodes.at_put(allocates, n);
5189 _macro_nodes.at_put(i, tmp);
5190 }
5191 allocates++;
5192 }
5193 }
5194 }
5195
5196 void Compile::print_method(CompilerPhaseType compile_phase, int level, Node* n) {
5197 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5198 EventCompilerPhase event(UNTIMED);
5199 if (event.should_commit()) {
5200 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, compile_phase, C->_compile_id, level);
5201 }
5202 #ifndef PRODUCT
5203 ResourceMark rm;
5204 stringStream ss;
5205 ss.print_raw(CompilerPhaseTypeHelper::to_description(compile_phase));
5206 int iter = ++_igv_phase_iter[compile_phase];
5207 if (iter > 1) {
5208 ss.print(" %d", iter);
5209 }
5210 if (n != nullptr) {
5211 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5212 if (n->is_Call()) {
5213 CallNode* call = n->as_Call();
5214 if (call->_name != nullptr) {
5215 // E.g. uncommon traps etc.
5216 ss.print(" - %s", call->_name);
5217 } else if (call->is_CallJava()) {
5218 CallJavaNode* call_java = call->as_CallJava();
5219 if (call_java->method() != nullptr) {
5220 ss.print(" -");
5221 call_java->method()->print_short_name(&ss);
5222 }
5223 }
5224 }
5225 }
5226
5227 const char* name = ss.as_string();
5228 if (should_print_igv(level)) {
5229 _igv_printer->print_graph(name);
5230 }
5231 if (should_print_phase(level)) {
5232 print_phase(name);
5233 }
5234 if (should_print_ideal_phase(compile_phase)) {
5235 print_ideal_ir(CompilerPhaseTypeHelper::to_name(compile_phase));
5236 }
5237 #endif
5238 C->_latest_stage_start_counter.stamp();
5239 }
5240
5241 // Only used from CompileWrapper
5242 void Compile::begin_method() {
5243 #ifndef PRODUCT
5244 if (_method != nullptr && should_print_igv(1)) {
5245 _igv_printer->begin_method();
5246 }
5247 #endif
5248 C->_latest_stage_start_counter.stamp();
5249 }
5250
5251 // Only used from CompileWrapper
5252 void Compile::end_method() {
5253 EventCompilerPhase event(UNTIMED);
5254 if (event.should_commit()) {
5255 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5256 }
5257
5258 #ifndef PRODUCT
5259 if (_method != nullptr && should_print_igv(1)) {
5260 _igv_printer->end_method();
5261 }
5262 #endif
5263 }
5264
5265 #ifndef PRODUCT
5266 bool Compile::should_print_phase(const int level) const {
5267 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5268 _method != nullptr; // Do not print phases for stubs.
5269 }
5270
5271 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5272 return _directive->should_print_ideal_phase(cpt);
5273 }
5274
5275 void Compile::init_igv() {
5276 if (_igv_printer == nullptr) {
5277 _igv_printer = IdealGraphPrinter::printer();
5278 _igv_printer->set_compile(this);
5279 }
5280 }
5281
5282 bool Compile::should_print_igv(const int level) {
5283 PRODUCT_RETURN_(return false;);
5284
5285 if (PrintIdealGraphLevel < 0) { // disabled by the user
5286 return false;
5287 }
5288
5289 bool need = directive()->IGVPrintLevelOption >= level;
5290 if (need) {
5291 Compile::init_igv();
5292 }
5293 return need;
5294 }
5295
5296 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5297 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5298
5299 // Called from debugger. Prints method to the default file with the default phase name.
5300 // This works regardless of any Ideal Graph Visualizer flags set or not.
5301 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5302 void igv_print(void* sp, void* fp, void* pc) {
5303 frame fr(sp, fp, pc);
5304 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5305 }
5306
5307 // Same as igv_print() above but with a specified phase name.
5308 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5309 frame fr(sp, fp, pc);
5310 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5311 }
5312
5313 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5314 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5315 // This works regardless of any Ideal Graph Visualizer flags set or not.
5316 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5317 void igv_print(bool network, void* sp, void* fp, void* pc) {
5318 frame fr(sp, fp, pc);
5319 if (network) {
5320 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5321 } else {
5322 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5323 }
5324 }
5325
5326 // Same as igv_print(bool network, ...) above but with a specified phase name.
5327 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5328 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5329 frame fr(sp, fp, pc);
5330 if (network) {
5331 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5332 } else {
5333 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5334 }
5335 }
5336
5337 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5338 void igv_print_default() {
5339 Compile::current()->print_method(PHASE_DEBUG, 0);
5340 }
5341
5342 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5343 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5344 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5345 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5346 void igv_append(void* sp, void* fp, void* pc) {
5347 frame fr(sp, fp, pc);
5348 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5349 }
5350
5351 // Same as igv_append(...) above but with a specified phase name.
5352 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5353 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5354 frame fr(sp, fp, pc);
5355 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5356 }
5357
5358 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5359 const char* file_name = "custom_debug.xml";
5360 if (_debug_file_printer == nullptr) {
5361 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5362 } else {
5363 _debug_file_printer->update_compiled_method(C->method());
5364 }
5365 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5366 _debug_file_printer->print_graph(phase_name, fr);
5367 }
5368
5369 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5370 ResourceMark rm;
5371 GrowableArray<const Node*> empty_list;
5372 igv_print_graph_to_network(phase_name, empty_list, fr);
5373 }
5374
5375 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5376 if (_debug_network_printer == nullptr) {
5377 _debug_network_printer = new IdealGraphPrinter(C);
5378 } else {
5379 _debug_network_printer->update_compiled_method(C->method());
5380 }
5381 tty->print_cr("Method printed over network stream to IGV");
5382 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5383 }
5384 #endif // !PRODUCT
5385
5386 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5387 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5388 return value;
5389 }
5390 Node* result = nullptr;
5391 if (bt == T_BYTE) {
5392 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5393 result = new RShiftINode(result, phase->intcon(24));
5394 } else if (bt == T_BOOLEAN) {
5395 result = new AndINode(value, phase->intcon(0xFF));
5396 } else if (bt == T_CHAR) {
5397 result = new AndINode(value,phase->intcon(0xFFFF));
5398 } else {
5399 assert(bt == T_SHORT, "unexpected narrow type");
5400 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5401 result = new RShiftINode(result, phase->intcon(16));
5402 }
5403 if (transform_res) {
5404 result = phase->transform(result);
5405 }
5406 return result;
5407 }
5408
5409 void Compile::record_method_not_compilable_oom() {
5410 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5411 }
5412
5413 #ifndef PRODUCT
5414 // Collects all the control inputs from nodes on the worklist and from their data dependencies
5415 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
5416 // Follow non-control edges until we reach CFG nodes
5417 for (uint i = 0; i < worklist.size(); i++) {
5418 const Node* n = worklist.at(i);
5419 for (uint j = 0; j < n->req(); j++) {
5420 Node* in = n->in(j);
5421 if (in == nullptr || in->is_Root()) {
5422 continue;
5423 }
5424 if (in->is_CFG()) {
5425 if (in->is_Call()) {
5426 // The return value of a call is only available if the call did not result in an exception
5427 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
5428 if (control_proj_use->is_Catch()) {
5429 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
5430 candidates.push(fall_through);
5431 continue;
5432 }
5433 }
5434
5435 if (in->is_Multi()) {
5436 // We got here by following data inputs so we should only have one control use
5437 // (no IfNode, etc)
5438 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
5439 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
5440 } else {
5441 candidates.push(in);
5442 }
5443 } else {
5444 worklist.push(in);
5445 }
5446 }
5447 }
5448 }
5449
5450 // Returns the candidate node that is a descendant to all the other candidates
5451 static Node* pick_control(Unique_Node_List& candidates) {
5452 Unique_Node_List worklist;
5453 worklist.copy(candidates);
5454
5455 // Traverse backwards through the CFG
5456 for (uint i = 0; i < worklist.size(); i++) {
5457 const Node* n = worklist.at(i);
5458 if (n->is_Root()) {
5459 continue;
5460 }
5461 for (uint j = 0; j < n->req(); j++) {
5462 // Skip backedge of loops to avoid cycles
5463 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
5464 continue;
5465 }
5466
5467 Node* pred = n->in(j);
5468 if (pred != nullptr && pred != n && pred->is_CFG()) {
5469 worklist.push(pred);
5470 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
5471 candidates.remove(pred);
5472 }
5473 }
5474 }
5475
5476 assert(candidates.size() == 1, "unexpected control flow");
5477 return candidates.at(0);
5478 }
5479
5480 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
5481 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
5482 call->init_req((*pos)++, parm);
5483 const BasicType bt = parm->bottom_type()->basic_type();
5484 if (bt == T_LONG || bt == T_DOUBLE) {
5485 call->init_req((*pos)++, half);
5486 }
5487 }
5488
5489 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
5490 Node* parm0, Node* parm1,
5491 Node* parm2, Node* parm3,
5492 Node* parm4, Node* parm5,
5493 Node* parm6) const {
5494 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
5495 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
5496 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
5497
5498 // find the most suitable control input
5499 Unique_Node_List worklist, candidates;
5500 if (parm0 != nullptr) { worklist.push(parm0);
5501 if (parm1 != nullptr) { worklist.push(parm1);
5502 if (parm2 != nullptr) { worklist.push(parm2);
5503 if (parm3 != nullptr) { worklist.push(parm3);
5504 if (parm4 != nullptr) { worklist.push(parm4);
5505 if (parm5 != nullptr) { worklist.push(parm5);
5506 if (parm6 != nullptr) { worklist.push(parm6);
5507 /* close each nested if ===> */ } } } } } } }
5508 find_candidate_control_inputs(worklist, candidates);
5509 Node* control = nullptr;
5510 if (candidates.size() == 0) {
5511 control = C->start()->proj_out(TypeFunc::Control);
5512 } else {
5513 control = pick_control(candidates);
5514 }
5515
5516 // find all the previous users of the control we picked
5517 GrowableArray<Node*> users_of_control;
5518 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
5519 Node* use = control->fast_out(i);
5520 if (use->is_CFG() && use != control) {
5521 users_of_control.push(use);
5522 }
5523 }
5524
5525 // we do not actually care about IO and memory as it uses neither
5526 call->init_req(TypeFunc::Control, control);
5527 call->init_req(TypeFunc::I_O, top());
5528 call->init_req(TypeFunc::Memory, top());
5529 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
5530 call->init_req(TypeFunc::ReturnAdr, top());
5531
5532 int pos = TypeFunc::Parms;
5533 call->init_req(pos++, str_node);
5534 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
5535 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
5536 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
5537 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
5538 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
5539 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
5540 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
5541 /* close each nested if ===> */ } } } } } } }
5542 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
5543
5544 call = gvn->transform(call);
5545 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
5546
5547 // rewire previous users to have the new call as control instead
5548 PhaseIterGVN* igvn = gvn->is_IterGVN();
5549 for (int i = 0; i < users_of_control.length(); i++) {
5550 Node* use = users_of_control.at(i);
5551 for (uint j = 0; j < use->req(); j++) {
5552 if (use->in(j) == control) {
5553 if (igvn != nullptr) {
5554 igvn->replace_input_of(use, j, call_control_proj);
5555 } else {
5556 gvn->hash_delete(use);
5557 use->set_req(j, call_control_proj);
5558 gvn->hash_insert(use);
5559 }
5560 }
5561 }
5562 }
5563
5564 return call;
5565 }
5566 #endif // !PRODUCT