1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/macroAssembler.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "ci/ciReplay.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "code/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 }
422
423 // Disconnect all useless nodes by disconnecting those at the boundary.
424 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
425 uint next = 0;
426 while (next < useful.size()) {
427 Node *n = useful.at(next++);
428 if (n->is_SafePoint()) {
429 // We're done with a parsing phase. Replaced nodes are not valid
430 // beyond that point.
431 n->as_SafePoint()->delete_replaced_nodes();
432 }
433 // Use raw traversal of out edges since this code removes out edges
434 int max = n->outcnt();
435 for (int j = 0; j < max; ++j) {
436 Node* child = n->raw_out(j);
437 if (!useful.member(child)) {
438 assert(!child->is_top() || child != top(),
439 "If top is cached in Compile object it is in useful list");
440 // Only need to remove this out-edge to the useless node
441 n->raw_del_out(j);
442 --j;
443 --max;
444 if (child->is_data_proj_of_pure_function(n)) {
445 worklist.push(n);
446 }
447 }
448 }
449 if (n->outcnt() == 1 && n->has_special_unique_user()) {
450 assert(useful.member(n->unique_out()), "do not push a useless node");
451 worklist.push(n->unique_out());
452 }
453 }
454
455 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
456 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
457 // Remove useless Template Assertion Predicate opaque nodes
458 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
459 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
460 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
461 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
462 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
463 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
464 #ifdef ASSERT
465 if (_modified_nodes != nullptr) {
466 _modified_nodes->remove_useless_nodes(useful.member_set());
467 }
468 #endif
469
470 // clean up the late inline lists
471 remove_useless_late_inlines( &_late_inlines, useful);
472 remove_useless_late_inlines( &_string_late_inlines, useful);
473 remove_useless_late_inlines( &_boxing_late_inlines, useful);
474 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
475 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
476 }
477
478 // ============================================================================
479 //------------------------------CompileWrapper---------------------------------
480 class CompileWrapper : public StackObj {
481 Compile *const _compile;
482 public:
483 CompileWrapper(Compile* compile);
484
485 ~CompileWrapper();
486 };
487
488 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
489 // the Compile* pointer is stored in the current ciEnv:
490 ciEnv* env = compile->env();
491 assert(env == ciEnv::current(), "must already be a ciEnv active");
492 assert(env->compiler_data() == nullptr, "compile already active?");
493 env->set_compiler_data(compile);
494 assert(compile == Compile::current(), "sanity");
495
496 compile->set_type_dict(nullptr);
497 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
498 compile->clone_map().set_clone_idx(0);
499 compile->set_type_last_size(0);
500 compile->set_last_tf(nullptr, nullptr);
501 compile->set_indexSet_arena(nullptr);
502 compile->set_indexSet_free_block_list(nullptr);
503 compile->init_type_arena();
504 Type::Initialize(compile);
505 _compile->begin_method();
506 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
507 }
508 CompileWrapper::~CompileWrapper() {
509 // simulate crash during compilation
510 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
511
512 _compile->end_method();
513 _compile->env()->set_compiler_data(nullptr);
514 }
515
516
517 //----------------------------print_compile_messages---------------------------
518 void Compile::print_compile_messages() {
519 #ifndef PRODUCT
520 // Check if recompiling
521 if (!subsume_loads() && PrintOpto) {
522 // Recompiling without allowing machine instructions to subsume loads
523 tty->print_cr("*********************************************************");
524 tty->print_cr("** Bailout: Recompile without subsuming loads **");
525 tty->print_cr("*********************************************************");
526 }
527 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
528 // Recompiling without escape analysis
529 tty->print_cr("*********************************************************");
530 tty->print_cr("** Bailout: Recompile without escape analysis **");
531 tty->print_cr("*********************************************************");
532 }
533 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
534 // Recompiling without iterative escape analysis
535 tty->print_cr("*********************************************************");
536 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
537 tty->print_cr("*********************************************************");
538 }
539 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
540 // Recompiling without reducing allocation merges
541 tty->print_cr("*********************************************************");
542 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
543 tty->print_cr("*********************************************************");
544 }
545 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
546 // Recompiling without boxing elimination
547 tty->print_cr("*********************************************************");
548 tty->print_cr("** Bailout: Recompile without boxing elimination **");
549 tty->print_cr("*********************************************************");
550 }
551 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
552 // Recompiling without locks coarsening
553 tty->print_cr("*********************************************************");
554 tty->print_cr("** Bailout: Recompile without locks coarsening **");
555 tty->print_cr("*********************************************************");
556 }
557 if (env()->break_at_compile()) {
558 // Open the debugger when compiling this method.
559 tty->print("### Breaking when compiling: ");
560 method()->print_short_name();
561 tty->cr();
562 BREAKPOINT;
563 }
564
565 if( PrintOpto ) {
566 if (is_osr_compilation()) {
567 tty->print("[OSR]%3d", _compile_id);
568 } else {
569 tty->print("%3d", _compile_id);
570 }
571 }
572 #endif
573 }
574
575 #ifndef PRODUCT
576 void Compile::print_phase(const char* phase_name) {
577 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name);
578 }
579
580 void Compile::print_ideal_ir(const char* phase_name) {
581 // keep the following output all in one block
582 // This output goes directly to the tty, not the compiler log.
583 // To enable tools to match it up with the compilation activity,
584 // be sure to tag this tty output with the compile ID.
585
586 // Node dumping can cause a safepoint, which can break the tty lock.
587 // Buffer all node dumps, so that all safepoints happen before we lock.
588 ResourceMark rm;
589 stringStream ss;
590
591 if (_output == nullptr) {
592 ss.print_cr("AFTER: %s", phase_name);
593 // Print out all nodes in ascending order of index.
594 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
595 } else {
596 // Dump the node blockwise if we have a scheduling
597 _output->print_scheduling(&ss);
598 }
599
600 // Check that the lock is not broken by a safepoint.
601 NoSafepointVerifier nsv;
602 ttyLocker ttyl;
603 if (xtty != nullptr) {
604 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
605 compile_id(),
606 is_osr_compilation() ? " compile_kind='osr'" : "",
607 phase_name);
608 }
609
610 tty->print("%s", ss.as_string());
611
612 if (xtty != nullptr) {
613 xtty->tail("ideal");
614 }
615 }
616 #endif
617
618 // ============================================================================
619 //------------------------------Compile standard-------------------------------
620
621 // Compile a method. entry_bci is -1 for normal compilations and indicates
622 // the continuation bci for on stack replacement.
623
624
625 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
626 Options options, DirectiveSet* directive)
627 : Phase(Compiler),
628 _compile_id(ci_env->compile_id()),
629 _options(options),
630 _method(target),
631 _entry_bci(osr_bci),
632 _ilt(nullptr),
633 _stub_function(nullptr),
634 _stub_name(nullptr),
635 _stub_id(StubId::NO_STUBID),
636 _stub_entry_point(nullptr),
637 _max_node_limit(MaxNodeLimit),
638 _post_loop_opts_phase(false),
639 _merge_stores_phase(false),
640 _allow_macro_nodes(true),
641 _inlining_progress(false),
642 _inlining_incrementally(false),
643 _do_cleanup(false),
644 _has_reserved_stack_access(target->has_reserved_stack_access()),
645 #ifndef PRODUCT
646 _igv_idx(0),
647 _trace_opto_output(directive->TraceOptoOutputOption),
648 #endif
649 _clinit_barrier_on_entry(false),
650 _stress_seed(0),
651 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
652 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
653 _env(ci_env),
654 _directive(directive),
655 _log(ci_env->log()),
656 _first_failure_details(nullptr),
657 _intrinsics(comp_arena(), 0, 0, nullptr),
658 _macro_nodes(comp_arena(), 8, 0, nullptr),
659 _parse_predicates(comp_arena(), 8, 0, nullptr),
660 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
661 _expensive_nodes(comp_arena(), 8, 0, nullptr),
662 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
663 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
664 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
665 _coarsened_locks(comp_arena(), 8, 0, nullptr),
666 _congraph(nullptr),
667 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
668 _unique(0),
669 _dead_node_count(0),
670 _dead_node_list(comp_arena()),
671 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
672 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
673 _node_arena(&_node_arena_one),
674 _mach_constant_base_node(nullptr),
675 _Compile_types(mtCompiler, Arena::Tag::tag_type),
676 _initial_gvn(nullptr),
677 _igvn_worklist(nullptr),
678 _types(nullptr),
679 _node_hash(nullptr),
680 _late_inlines(comp_arena(), 2, 0, nullptr),
681 _string_late_inlines(comp_arena(), 2, 0, nullptr),
682 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
683 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
684 _late_inlines_pos(0),
685 _has_mh_late_inlines(false),
686 _oom(false),
687 _replay_inline_data(nullptr),
688 _inline_printer(this),
689 _java_calls(0),
690 _inner_loops(0),
691 _FIRST_STACK_mask(comp_arena()),
692 _interpreter_frame_size(0),
693 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
694 _output(nullptr)
695 #ifndef PRODUCT
696 ,
697 _in_dump_cnt(0)
698 #endif
699 {
700 C = this;
701 CompileWrapper cw(this);
702
703 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
704 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
705
706 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
707 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
708 // We can always print a disassembly, either abstract (hex dump) or
709 // with the help of a suitable hsdis library. Thus, we should not
710 // couple print_assembly and print_opto_assembly controls.
711 // But: always print opto and regular assembly on compile command 'print'.
712 bool print_assembly = directive->PrintAssemblyOption;
713 set_print_assembly(print_opto_assembly || print_assembly);
714 #else
715 set_print_assembly(false); // must initialize.
716 #endif
717
718 #ifndef PRODUCT
719 set_parsed_irreducible_loop(false);
720 #endif
721
722 if (directive->ReplayInlineOption) {
723 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
724 }
725 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
726 set_print_intrinsics(directive->PrintIntrinsicsOption);
727 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
728
729 if (ProfileTraps) {
730 // Make sure the method being compiled gets its own MDO,
731 // so we can at least track the decompile_count().
732 method()->ensure_method_data();
733 }
734
735 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
736 StressIncrementalInlining || StressMacroExpansion ||
737 StressMacroElimination || StressUnstableIfTraps ||
738 StressBailout || StressLoopPeeling) {
739 initialize_stress_seed(directive);
740 }
741
742 Init(/*do_aliasing=*/ true);
743
744 print_compile_messages();
745
746 _ilt = InlineTree::build_inline_tree_root();
747
748 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
749 assert(num_alias_types() >= AliasIdxRaw, "");
750
751 #define MINIMUM_NODE_HASH 1023
752
753 // GVN that will be run immediately on new nodes
754 uint estimated_size = method()->code_size()*4+64;
755 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
756 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
757 _types = new (comp_arena()) Type_Array(comp_arena());
758 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
759 PhaseGVN gvn;
760 set_initial_gvn(&gvn);
761
762 { // Scope for timing the parser
763 TracePhase tp(_t_parser);
764
765 // Put top into the hash table ASAP.
766 initial_gvn()->transform(top());
767
768 // Set up tf(), start(), and find a CallGenerator.
769 CallGenerator* cg = nullptr;
770 if (is_osr_compilation()) {
771 const TypeTuple *domain = StartOSRNode::osr_domain();
772 const TypeTuple *range = TypeTuple::make_range(method()->signature());
773 init_tf(TypeFunc::make(domain, range));
774 StartNode* s = new StartOSRNode(root(), domain);
775 initial_gvn()->set_type_bottom(s);
776 verify_start(s);
777 cg = CallGenerator::for_osr(method(), entry_bci());
778 } else {
779 // Normal case.
780 init_tf(TypeFunc::make(method()));
781 StartNode* s = new StartNode(root(), tf()->domain());
782 initial_gvn()->set_type_bottom(s);
783 verify_start(s);
784 float past_uses = method()->interpreter_invocation_count();
785 float expected_uses = past_uses;
786 cg = CallGenerator::for_inline(method(), expected_uses);
787 }
788 if (failing()) return;
789 if (cg == nullptr) {
790 const char* reason = InlineTree::check_can_parse(method());
791 assert(reason != nullptr, "expect reason for parse failure");
792 stringStream ss;
793 ss.print("cannot parse method: %s", reason);
794 record_method_not_compilable(ss.as_string());
795 return;
796 }
797
798 gvn.set_type(root(), root()->bottom_type());
799
800 JVMState* jvms = build_start_state(start(), tf());
801 if ((jvms = cg->generate(jvms)) == nullptr) {
802 assert(failure_reason() != nullptr, "expect reason for parse failure");
803 stringStream ss;
804 ss.print("method parse failed: %s", failure_reason());
805 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
806 return;
807 }
808 GraphKit kit(jvms);
809
810 if (!kit.stopped()) {
811 // Accept return values, and transfer control we know not where.
812 // This is done by a special, unique ReturnNode bound to root.
813 return_values(kit.jvms());
814 }
815
816 if (kit.has_exceptions()) {
817 // Any exceptions that escape from this call must be rethrown
818 // to whatever caller is dynamically above us on the stack.
819 // This is done by a special, unique RethrowNode bound to root.
820 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
821 }
822
823 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
824
825 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
826 inline_string_calls(true);
827 }
828
829 if (failing()) return;
830
831 // Remove clutter produced by parsing.
832 if (!failing()) {
833 ResourceMark rm;
834 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
835 }
836 }
837
838 // Note: Large methods are capped off in do_one_bytecode().
839 if (failing()) return;
840
841 // After parsing, node notes are no longer automagic.
842 // They must be propagated by register_new_node_with_optimizer(),
843 // clone(), or the like.
844 set_default_node_notes(nullptr);
845
846 #ifndef PRODUCT
847 if (should_print_igv(1)) {
848 _igv_printer->print_inlining();
849 }
850 #endif
851
852 if (failing()) return;
853 NOT_PRODUCT( verify_graph_edges(); )
854
855 // Now optimize
856 Optimize();
857 if (failing()) return;
858 NOT_PRODUCT( verify_graph_edges(); )
859
860 #ifndef PRODUCT
861 if (should_print_ideal()) {
862 print_ideal_ir("print_ideal");
863 }
864 #endif
865
866 #ifdef ASSERT
867 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
868 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
869 #endif
870
871 // Dump compilation data to replay it.
872 if (directive->DumpReplayOption) {
873 env()->dump_replay_data(_compile_id);
874 }
875 if (directive->DumpInlineOption && (ilt() != nullptr)) {
876 env()->dump_inline_data(_compile_id);
877 }
878
879 // Now that we know the size of all the monitors we can add a fixed slot
880 // for the original deopt pc.
881 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
882 set_fixed_slots(next_slot);
883
884 // Compute when to use implicit null checks. Used by matching trap based
885 // nodes and NullCheck optimization.
886 set_allowed_deopt_reasons();
887
888 // Now generate code
889 Code_Gen();
890 }
891
892 //------------------------------Compile----------------------------------------
893 // Compile a runtime stub
894 Compile::Compile(ciEnv* ci_env,
895 TypeFunc_generator generator,
896 address stub_function,
897 const char* stub_name,
898 StubId stub_id,
899 int is_fancy_jump,
900 bool pass_tls,
901 bool return_pc,
902 DirectiveSet* directive)
903 : Phase(Compiler),
904 _compile_id(0),
905 _options(Options::for_runtime_stub()),
906 _method(nullptr),
907 _entry_bci(InvocationEntryBci),
908 _stub_function(stub_function),
909 _stub_name(stub_name),
910 _stub_id(stub_id),
911 _stub_entry_point(nullptr),
912 _max_node_limit(MaxNodeLimit),
913 _post_loop_opts_phase(false),
914 _merge_stores_phase(false),
915 _allow_macro_nodes(true),
916 _inlining_progress(false),
917 _inlining_incrementally(false),
918 _has_reserved_stack_access(false),
919 #ifndef PRODUCT
920 _igv_idx(0),
921 _trace_opto_output(directive->TraceOptoOutputOption),
922 #endif
923 _clinit_barrier_on_entry(false),
924 _stress_seed(0),
925 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
926 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
927 _env(ci_env),
928 _directive(directive),
929 _log(ci_env->log()),
930 _first_failure_details(nullptr),
931 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
932 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
933 _congraph(nullptr),
934 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
935 _unique(0),
936 _dead_node_count(0),
937 _dead_node_list(comp_arena()),
938 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
939 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
940 _node_arena(&_node_arena_one),
941 _mach_constant_base_node(nullptr),
942 _Compile_types(mtCompiler, Arena::Tag::tag_type),
943 _initial_gvn(nullptr),
944 _igvn_worklist(nullptr),
945 _types(nullptr),
946 _node_hash(nullptr),
947 _has_mh_late_inlines(false),
948 _oom(false),
949 _replay_inline_data(nullptr),
950 _inline_printer(this),
951 _java_calls(0),
952 _inner_loops(0),
953 _FIRST_STACK_mask(comp_arena()),
954 _interpreter_frame_size(0),
955 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
956 _output(nullptr),
957 #ifndef PRODUCT
958 _in_dump_cnt(0),
959 #endif
960 _allowed_reasons(0) {
961 C = this;
962
963 // try to reuse an existing stub
964 {
965 BlobId blob_id = StubInfo::blob(_stub_id);
966 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
967 if (blob != nullptr) {
968 RuntimeStub* rs = blob->as_runtime_stub();
969 _stub_entry_point = rs->entry_point();
970 return;
971 }
972 }
973
974 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
975 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
976
977 #ifndef PRODUCT
978 set_print_assembly(PrintFrameConverterAssembly);
979 set_parsed_irreducible_loop(false);
980 #else
981 set_print_assembly(false); // Must initialize.
982 #endif
983 set_has_irreducible_loop(false); // no loops
984
985 CompileWrapper cw(this);
986 Init(/*do_aliasing=*/ false);
987 init_tf((*generator)());
988
989 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
990 _types = new (comp_arena()) Type_Array(comp_arena());
991 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
992
993 if (StressLCM || StressGCM || StressBailout) {
994 initialize_stress_seed(directive);
995 }
996
997 {
998 PhaseGVN gvn;
999 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1000 gvn.transform(top());
1001
1002 GraphKit kit;
1003 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1004 }
1005
1006 NOT_PRODUCT( verify_graph_edges(); )
1007
1008 Code_Gen();
1009 }
1010
1011 Compile::~Compile() {
1012 delete _first_failure_details;
1013 };
1014
1015 //------------------------------Init-------------------------------------------
1016 // Prepare for a single compilation
1017 void Compile::Init(bool aliasing) {
1018 _do_aliasing = aliasing;
1019 _unique = 0;
1020 _regalloc = nullptr;
1021
1022 _tf = nullptr; // filled in later
1023 _top = nullptr; // cached later
1024 _matcher = nullptr; // filled in later
1025 _cfg = nullptr; // filled in later
1026
1027 _node_note_array = nullptr;
1028 _default_node_notes = nullptr;
1029 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1030
1031 _immutable_memory = nullptr; // filled in at first inquiry
1032
1033 #ifdef ASSERT
1034 _phase_optimize_finished = false;
1035 _phase_verify_ideal_loop = false;
1036 _exception_backedge = false;
1037 _type_verify = nullptr;
1038 #endif
1039
1040 // Globally visible Nodes
1041 // First set TOP to null to give safe behavior during creation of RootNode
1042 set_cached_top_node(nullptr);
1043 set_root(new RootNode());
1044 // Now that you have a Root to point to, create the real TOP
1045 set_cached_top_node( new ConNode(Type::TOP) );
1046 set_recent_alloc(nullptr, nullptr);
1047
1048 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1049 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1050 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1051 env()->set_dependencies(new Dependencies(env()));
1052
1053 _fixed_slots = 0;
1054 set_has_split_ifs(false);
1055 set_has_loops(false); // first approximation
1056 set_has_stringbuilder(false);
1057 set_has_boxed_value(false);
1058 _trap_can_recompile = false; // no traps emitted yet
1059 _major_progress = true; // start out assuming good things will happen
1060 set_has_unsafe_access(false);
1061 set_max_vector_size(0);
1062 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1063 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1064 set_decompile_count(0);
1065
1066 #ifndef PRODUCT
1067 _phase_counter = 0;
1068 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1069 #endif
1070
1071 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1072 _loop_opts_cnt = LoopOptsCount;
1073 set_do_inlining(Inline);
1074 set_max_inline_size(MaxInlineSize);
1075 set_freq_inline_size(FreqInlineSize);
1076 set_do_scheduling(OptoScheduling);
1077
1078 set_do_vector_loop(false);
1079 set_has_monitors(false);
1080 set_has_scoped_access(false);
1081
1082 if (AllowVectorizeOnDemand) {
1083 if (has_method() && _directive->VectorizeOption) {
1084 set_do_vector_loop(true);
1085 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());})
1086 } else if (has_method() && method()->name() != nullptr &&
1087 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1088 set_do_vector_loop(true);
1089 }
1090 }
1091 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1092 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());})
1093
1094 _max_node_limit = _directive->MaxNodeLimitOption;
1095
1096 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1097 set_clinit_barrier_on_entry(true);
1098 }
1099 if (debug_info()->recording_non_safepoints()) {
1100 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1101 (comp_arena(), 8, 0, nullptr));
1102 set_default_node_notes(Node_Notes::make(this));
1103 }
1104
1105 const int grow_ats = 16;
1106 _max_alias_types = grow_ats;
1107 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1108 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1109 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1110 {
1111 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1112 }
1113 // Initialize the first few types.
1114 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1115 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1116 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1117 _num_alias_types = AliasIdxRaw+1;
1118 // Zero out the alias type cache.
1119 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1120 // A null adr_type hits in the cache right away. Preload the right answer.
1121 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1122 }
1123
1124 #ifdef ASSERT
1125 // Verify that the current StartNode is valid.
1126 void Compile::verify_start(StartNode* s) const {
1127 assert(failing_internal() || s == start(), "should be StartNode");
1128 }
1129 #endif
1130
1131 /**
1132 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1133 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1134 * the ideal graph.
1135 */
1136 StartNode* Compile::start() const {
1137 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1138 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1139 Node* start = root()->fast_out(i);
1140 if (start->is_Start()) {
1141 return start->as_Start();
1142 }
1143 }
1144 fatal("Did not find Start node!");
1145 return nullptr;
1146 }
1147
1148 //-------------------------------immutable_memory-------------------------------------
1149 // Access immutable memory
1150 Node* Compile::immutable_memory() {
1151 if (_immutable_memory != nullptr) {
1152 return _immutable_memory;
1153 }
1154 StartNode* s = start();
1155 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1156 Node *p = s->fast_out(i);
1157 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1158 _immutable_memory = p;
1159 return _immutable_memory;
1160 }
1161 }
1162 ShouldNotReachHere();
1163 return nullptr;
1164 }
1165
1166 //----------------------set_cached_top_node------------------------------------
1167 // Install the cached top node, and make sure Node::is_top works correctly.
1168 void Compile::set_cached_top_node(Node* tn) {
1169 if (tn != nullptr) verify_top(tn);
1170 Node* old_top = _top;
1171 _top = tn;
1172 // Calling Node::setup_is_top allows the nodes the chance to adjust
1173 // their _out arrays.
1174 if (_top != nullptr) _top->setup_is_top();
1175 if (old_top != nullptr) old_top->setup_is_top();
1176 assert(_top == nullptr || top()->is_top(), "");
1177 }
1178
1179 #ifdef ASSERT
1180 uint Compile::count_live_nodes_by_graph_walk() {
1181 Unique_Node_List useful(comp_arena());
1182 // Get useful node list by walking the graph.
1183 identify_useful_nodes(useful);
1184 return useful.size();
1185 }
1186
1187 void Compile::print_missing_nodes() {
1188
1189 // Return if CompileLog is null and PrintIdealNodeCount is false.
1190 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1191 return;
1192 }
1193
1194 // This is an expensive function. It is executed only when the user
1195 // specifies VerifyIdealNodeCount option or otherwise knows the
1196 // additional work that needs to be done to identify reachable nodes
1197 // by walking the flow graph and find the missing ones using
1198 // _dead_node_list.
1199
1200 Unique_Node_List useful(comp_arena());
1201 // Get useful node list by walking the graph.
1202 identify_useful_nodes(useful);
1203
1204 uint l_nodes = C->live_nodes();
1205 uint l_nodes_by_walk = useful.size();
1206
1207 if (l_nodes != l_nodes_by_walk) {
1208 if (_log != nullptr) {
1209 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1210 _log->stamp();
1211 _log->end_head();
1212 }
1213 VectorSet& useful_member_set = useful.member_set();
1214 int last_idx = l_nodes_by_walk;
1215 for (int i = 0; i < last_idx; i++) {
1216 if (useful_member_set.test(i)) {
1217 if (_dead_node_list.test(i)) {
1218 if (_log != nullptr) {
1219 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1220 }
1221 if (PrintIdealNodeCount) {
1222 // Print the log message to tty
1223 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1224 useful.at(i)->dump();
1225 }
1226 }
1227 }
1228 else if (! _dead_node_list.test(i)) {
1229 if (_log != nullptr) {
1230 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1231 }
1232 if (PrintIdealNodeCount) {
1233 // Print the log message to tty
1234 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1235 }
1236 }
1237 }
1238 if (_log != nullptr) {
1239 _log->tail("mismatched_nodes");
1240 }
1241 }
1242 }
1243 void Compile::record_modified_node(Node* n) {
1244 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1245 _modified_nodes->push(n);
1246 }
1247 }
1248
1249 void Compile::remove_modified_node(Node* n) {
1250 if (_modified_nodes != nullptr) {
1251 _modified_nodes->remove(n);
1252 }
1253 }
1254 #endif
1255
1256 #ifndef PRODUCT
1257 void Compile::verify_top(Node* tn) const {
1258 if (tn != nullptr) {
1259 assert(tn->is_Con(), "top node must be a constant");
1260 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1261 assert(tn->in(0) != nullptr, "must have live top node");
1262 }
1263 }
1264 #endif
1265
1266
1267 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1268
1269 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1270 guarantee(arr != nullptr, "");
1271 int num_blocks = arr->length();
1272 if (grow_by < num_blocks) grow_by = num_blocks;
1273 int num_notes = grow_by * _node_notes_block_size;
1274 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1275 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1276 while (num_notes > 0) {
1277 arr->append(notes);
1278 notes += _node_notes_block_size;
1279 num_notes -= _node_notes_block_size;
1280 }
1281 assert(num_notes == 0, "exact multiple, please");
1282 }
1283
1284 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1285 if (source == nullptr || dest == nullptr) return false;
1286
1287 if (dest->is_Con())
1288 return false; // Do not push debug info onto constants.
1289
1290 #ifdef ASSERT
1291 // Leave a bread crumb trail pointing to the original node:
1292 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1293 dest->set_debug_orig(source);
1294 }
1295 #endif
1296
1297 if (node_note_array() == nullptr)
1298 return false; // Not collecting any notes now.
1299
1300 // This is a copy onto a pre-existing node, which may already have notes.
1301 // If both nodes have notes, do not overwrite any pre-existing notes.
1302 Node_Notes* source_notes = node_notes_at(source->_idx);
1303 if (source_notes == nullptr || source_notes->is_clear()) return false;
1304 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1305 if (dest_notes == nullptr || dest_notes->is_clear()) {
1306 return set_node_notes_at(dest->_idx, source_notes);
1307 }
1308
1309 Node_Notes merged_notes = (*source_notes);
1310 // The order of operations here ensures that dest notes will win...
1311 merged_notes.update_from(dest_notes);
1312 return set_node_notes_at(dest->_idx, &merged_notes);
1313 }
1314
1315
1316 //--------------------------allow_range_check_smearing-------------------------
1317 // Gating condition for coalescing similar range checks.
1318 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1319 // single covering check that is at least as strong as any of them.
1320 // If the optimization succeeds, the simplified (strengthened) range check
1321 // will always succeed. If it fails, we will deopt, and then give up
1322 // on the optimization.
1323 bool Compile::allow_range_check_smearing() const {
1324 // If this method has already thrown a range-check,
1325 // assume it was because we already tried range smearing
1326 // and it failed.
1327 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1328 return !already_trapped;
1329 }
1330
1331
1332 //------------------------------flatten_alias_type-----------------------------
1333 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1334 assert(do_aliasing(), "Aliasing should be enabled");
1335 int offset = tj->offset();
1336 TypePtr::PTR ptr = tj->ptr();
1337
1338 // Known instance (scalarizable allocation) alias only with itself.
1339 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1340 tj->is_oopptr()->is_known_instance();
1341
1342 // Process weird unsafe references.
1343 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1344 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1345 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1346 tj = TypeOopPtr::BOTTOM;
1347 ptr = tj->ptr();
1348 offset = tj->offset();
1349 }
1350
1351 // Array pointers need some flattening
1352 const TypeAryPtr* ta = tj->isa_aryptr();
1353 if (ta && ta->is_stable()) {
1354 // Erase stability property for alias analysis.
1355 tj = ta = ta->cast_to_stable(false);
1356 }
1357 if( ta && is_known_inst ) {
1358 if ( offset != Type::OffsetBot &&
1359 offset > arrayOopDesc::length_offset_in_bytes() ) {
1360 offset = Type::OffsetBot; // Flatten constant access into array body only
1361 tj = ta = ta->
1362 remove_speculative()->
1363 cast_to_ptr_type(ptr)->
1364 with_offset(offset);
1365 }
1366 } else if (ta) {
1367 // For arrays indexed by constant indices, we flatten the alias
1368 // space to include all of the array body. Only the header, klass
1369 // and array length can be accessed un-aliased.
1370 if( offset != Type::OffsetBot ) {
1371 if( ta->const_oop() ) { // MethodData* or Method*
1372 offset = Type::OffsetBot; // Flatten constant access into array body
1373 tj = ta = ta->
1374 remove_speculative()->
1375 cast_to_ptr_type(ptr)->
1376 cast_to_exactness(false)->
1377 with_offset(offset);
1378 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1379 // range is OK as-is.
1380 tj = ta = TypeAryPtr::RANGE;
1381 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1382 tj = TypeInstPtr::KLASS; // all klass loads look alike
1383 ta = TypeAryPtr::RANGE; // generic ignored junk
1384 ptr = TypePtr::BotPTR;
1385 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1386 tj = TypeInstPtr::MARK;
1387 ta = TypeAryPtr::RANGE; // generic ignored junk
1388 ptr = TypePtr::BotPTR;
1389 } else { // Random constant offset into array body
1390 offset = Type::OffsetBot; // Flatten constant access into array body
1391 tj = ta = ta->
1392 remove_speculative()->
1393 cast_to_ptr_type(ptr)->
1394 cast_to_exactness(false)->
1395 with_offset(offset);
1396 }
1397 }
1398 // Arrays of fixed size alias with arrays of unknown size.
1399 if (ta->size() != TypeInt::POS) {
1400 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1401 tj = ta = ta->
1402 remove_speculative()->
1403 cast_to_ptr_type(ptr)->
1404 with_ary(tary)->
1405 cast_to_exactness(false);
1406 }
1407 // Arrays of known objects become arrays of unknown objects.
1408 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1409 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1410 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1411 }
1412 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1413 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1414 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1415 }
1416 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1417 // cannot be distinguished by bytecode alone.
1418 if (ta->elem() == TypeInt::BOOL) {
1419 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1420 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1421 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1422 }
1423 // During the 2nd round of IterGVN, NotNull castings are removed.
1424 // Make sure the Bottom and NotNull variants alias the same.
1425 // Also, make sure exact and non-exact variants alias the same.
1426 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1427 tj = ta = ta->
1428 remove_speculative()->
1429 cast_to_ptr_type(TypePtr::BotPTR)->
1430 cast_to_exactness(false)->
1431 with_offset(offset);
1432 }
1433 }
1434
1435 // Oop pointers need some flattening
1436 const TypeInstPtr *to = tj->isa_instptr();
1437 if (to && to != TypeOopPtr::BOTTOM) {
1438 ciInstanceKlass* ik = to->instance_klass();
1439 if( ptr == TypePtr::Constant ) {
1440 if (ik != ciEnv::current()->Class_klass() ||
1441 offset < ik->layout_helper_size_in_bytes()) {
1442 // No constant oop pointers (such as Strings); they alias with
1443 // unknown strings.
1444 assert(!is_known_inst, "not scalarizable allocation");
1445 tj = to = to->
1446 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1447 remove_speculative()->
1448 cast_to_ptr_type(TypePtr::BotPTR)->
1449 cast_to_exactness(false);
1450 }
1451 } else if( is_known_inst ) {
1452 tj = to; // Keep NotNull and klass_is_exact for instance type
1453 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1454 // During the 2nd round of IterGVN, NotNull castings are removed.
1455 // Make sure the Bottom and NotNull variants alias the same.
1456 // Also, make sure exact and non-exact variants alias the same.
1457 tj = to = to->
1458 remove_speculative()->
1459 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1460 cast_to_ptr_type(TypePtr::BotPTR)->
1461 cast_to_exactness(false);
1462 }
1463 if (to->speculative() != nullptr) {
1464 tj = to = to->remove_speculative();
1465 }
1466 // Canonicalize the holder of this field
1467 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1468 // First handle header references such as a LoadKlassNode, even if the
1469 // object's klass is unloaded at compile time (4965979).
1470 if (!is_known_inst) { // Do it only for non-instance types
1471 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1472 }
1473 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1474 // Static fields are in the space above the normal instance
1475 // fields in the java.lang.Class instance.
1476 if (ik != ciEnv::current()->Class_klass()) {
1477 to = nullptr;
1478 tj = TypeOopPtr::BOTTOM;
1479 offset = tj->offset();
1480 }
1481 } else {
1482 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1483 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1484 assert(tj->offset() == offset, "no change to offset expected");
1485 bool xk = to->klass_is_exact();
1486 int instance_id = to->instance_id();
1487
1488 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1489 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1490 // its interfaces are included.
1491 if (xk && ik->equals(canonical_holder)) {
1492 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1493 } else {
1494 assert(xk || !is_known_inst, "Known instance should be exact type");
1495 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1496 }
1497 }
1498 }
1499
1500 // Klass pointers to object array klasses need some flattening
1501 const TypeKlassPtr *tk = tj->isa_klassptr();
1502 if( tk ) {
1503 // If we are referencing a field within a Klass, we need
1504 // to assume the worst case of an Object. Both exact and
1505 // inexact types must flatten to the same alias class so
1506 // use NotNull as the PTR.
1507 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1508 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1509 env()->Object_klass(),
1510 offset);
1511 }
1512
1513 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1514 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1515 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1516 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1517 } else {
1518 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1519 }
1520 }
1521
1522 // Check for precise loads from the primary supertype array and force them
1523 // to the supertype cache alias index. Check for generic array loads from
1524 // the primary supertype array and also force them to the supertype cache
1525 // alias index. Since the same load can reach both, we need to merge
1526 // these 2 disparate memories into the same alias class. Since the
1527 // primary supertype array is read-only, there's no chance of confusion
1528 // where we bypass an array load and an array store.
1529 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1530 if (offset == Type::OffsetBot ||
1531 (offset >= primary_supers_offset &&
1532 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1533 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1534 offset = in_bytes(Klass::secondary_super_cache_offset());
1535 tj = tk = tk->with_offset(offset);
1536 }
1537 }
1538
1539 // Flatten all Raw pointers together.
1540 if (tj->base() == Type::RawPtr)
1541 tj = TypeRawPtr::BOTTOM;
1542
1543 if (tj->base() == Type::AnyPtr)
1544 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1545
1546 offset = tj->offset();
1547 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1548
1549 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1550 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1551 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1552 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1553 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1554 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1555 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1556 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1557 assert( tj->ptr() != TypePtr::TopPTR &&
1558 tj->ptr() != TypePtr::AnyNull &&
1559 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1560 // assert( tj->ptr() != TypePtr::Constant ||
1561 // tj->base() == Type::RawPtr ||
1562 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1563
1564 return tj;
1565 }
1566
1567 void Compile::AliasType::Init(int i, const TypePtr* at) {
1568 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1569 _index = i;
1570 _adr_type = at;
1571 _field = nullptr;
1572 _element = nullptr;
1573 _is_rewritable = true; // default
1574 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1575 if (atoop != nullptr && atoop->is_known_instance()) {
1576 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1577 _general_index = Compile::current()->get_alias_index(gt);
1578 } else {
1579 _general_index = 0;
1580 }
1581 }
1582
1583 BasicType Compile::AliasType::basic_type() const {
1584 if (element() != nullptr) {
1585 const Type* element = adr_type()->is_aryptr()->elem();
1586 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1587 } if (field() != nullptr) {
1588 return field()->layout_type();
1589 } else {
1590 return T_ILLEGAL; // unknown
1591 }
1592 }
1593
1594 //---------------------------------print_on------------------------------------
1595 #ifndef PRODUCT
1596 void Compile::AliasType::print_on(outputStream* st) {
1597 if (index() < 10)
1598 st->print("@ <%d> ", index());
1599 else st->print("@ <%d>", index());
1600 st->print(is_rewritable() ? " " : " RO");
1601 int offset = adr_type()->offset();
1602 if (offset == Type::OffsetBot)
1603 st->print(" +any");
1604 else st->print(" +%-3d", offset);
1605 st->print(" in ");
1606 adr_type()->dump_on(st);
1607 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1608 if (field() != nullptr && tjp) {
1609 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1610 tjp->offset() != field()->offset_in_bytes()) {
1611 st->print(" != ");
1612 field()->print();
1613 st->print(" ***");
1614 }
1615 }
1616 }
1617
1618 void print_alias_types() {
1619 Compile* C = Compile::current();
1620 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1621 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1622 C->alias_type(idx)->print_on(tty);
1623 tty->cr();
1624 }
1625 }
1626 #endif
1627
1628
1629 //----------------------------probe_alias_cache--------------------------------
1630 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1631 intptr_t key = (intptr_t) adr_type;
1632 key ^= key >> logAliasCacheSize;
1633 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1634 }
1635
1636
1637 //-----------------------------grow_alias_types--------------------------------
1638 void Compile::grow_alias_types() {
1639 const int old_ats = _max_alias_types; // how many before?
1640 const int new_ats = old_ats; // how many more?
1641 const int grow_ats = old_ats+new_ats; // how many now?
1642 _max_alias_types = grow_ats;
1643 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1644 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1645 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1646 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1647 }
1648
1649
1650 //--------------------------------find_alias_type------------------------------
1651 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1652 if (!do_aliasing()) {
1653 return alias_type(AliasIdxBot);
1654 }
1655
1656 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1657 if (ace->_adr_type == adr_type) {
1658 return alias_type(ace->_index);
1659 }
1660
1661 // Handle special cases.
1662 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1663 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1664
1665 // Do it the slow way.
1666 const TypePtr* flat = flatten_alias_type(adr_type);
1667
1668 #ifdef ASSERT
1669 {
1670 ResourceMark rm;
1671 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1672 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1673 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1674 Type::str(adr_type));
1675 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1676 const TypeOopPtr* foop = flat->is_oopptr();
1677 // Scalarizable allocations have exact klass always.
1678 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1679 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1680 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1681 Type::str(foop), Type::str(xoop));
1682 }
1683 }
1684 #endif
1685
1686 int idx = AliasIdxTop;
1687 for (int i = 0; i < num_alias_types(); i++) {
1688 if (alias_type(i)->adr_type() == flat) {
1689 idx = i;
1690 break;
1691 }
1692 }
1693
1694 if (idx == AliasIdxTop) {
1695 if (no_create) return nullptr;
1696 // Grow the array if necessary.
1697 if (_num_alias_types == _max_alias_types) grow_alias_types();
1698 // Add a new alias type.
1699 idx = _num_alias_types++;
1700 _alias_types[idx]->Init(idx, flat);
1701 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1702 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1703 if (flat->isa_instptr()) {
1704 if (flat->offset() == java_lang_Class::klass_offset()
1705 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1706 alias_type(idx)->set_rewritable(false);
1707 }
1708 if (flat->isa_aryptr()) {
1709 #ifdef ASSERT
1710 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1711 // (T_BYTE has the weakest alignment and size restrictions...)
1712 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1713 #endif
1714 if (flat->offset() == TypePtr::OffsetBot) {
1715 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1716 }
1717 }
1718 if (flat->isa_klassptr()) {
1719 if (UseCompactObjectHeaders) {
1720 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1721 alias_type(idx)->set_rewritable(false);
1722 }
1723 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1724 alias_type(idx)->set_rewritable(false);
1725 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1726 alias_type(idx)->set_rewritable(false);
1727 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1728 alias_type(idx)->set_rewritable(false);
1729 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1730 alias_type(idx)->set_rewritable(false);
1731 }
1732
1733 if (flat->isa_instklassptr()) {
1734 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1735 alias_type(idx)->set_rewritable(false);
1736 }
1737 }
1738 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1739 // but the base pointer type is not distinctive enough to identify
1740 // references into JavaThread.)
1741
1742 // Check for final fields.
1743 const TypeInstPtr* tinst = flat->isa_instptr();
1744 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1745 ciField* field;
1746 if (tinst->const_oop() != nullptr &&
1747 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1748 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1749 // static field
1750 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1751 field = k->get_field_by_offset(tinst->offset(), true);
1752 } else {
1753 ciInstanceKlass *k = tinst->instance_klass();
1754 field = k->get_field_by_offset(tinst->offset(), false);
1755 }
1756 assert(field == nullptr ||
1757 original_field == nullptr ||
1758 (field->holder() == original_field->holder() &&
1759 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1760 field->is_static() == original_field->is_static()), "wrong field?");
1761 // Set field() and is_rewritable() attributes.
1762 if (field != nullptr) alias_type(idx)->set_field(field);
1763 }
1764 }
1765
1766 // Fill the cache for next time.
1767 ace->_adr_type = adr_type;
1768 ace->_index = idx;
1769 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1770
1771 // Might as well try to fill the cache for the flattened version, too.
1772 AliasCacheEntry* face = probe_alias_cache(flat);
1773 if (face->_adr_type == nullptr) {
1774 face->_adr_type = flat;
1775 face->_index = idx;
1776 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1777 }
1778
1779 return alias_type(idx);
1780 }
1781
1782
1783 Compile::AliasType* Compile::alias_type(ciField* field) {
1784 const TypeOopPtr* t;
1785 if (field->is_static())
1786 t = TypeInstPtr::make(field->holder()->java_mirror());
1787 else
1788 t = TypeOopPtr::make_from_klass_raw(field->holder());
1789 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1790 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1791 return atp;
1792 }
1793
1794
1795 //------------------------------have_alias_type--------------------------------
1796 bool Compile::have_alias_type(const TypePtr* adr_type) {
1797 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1798 if (ace->_adr_type == adr_type) {
1799 return true;
1800 }
1801
1802 // Handle special cases.
1803 if (adr_type == nullptr) return true;
1804 if (adr_type == TypePtr::BOTTOM) return true;
1805
1806 return find_alias_type(adr_type, true, nullptr) != nullptr;
1807 }
1808
1809 //-----------------------------must_alias--------------------------------------
1810 // True if all values of the given address type are in the given alias category.
1811 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1812 if (alias_idx == AliasIdxBot) return true; // the universal category
1813 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1814 if (alias_idx == AliasIdxTop) return false; // the empty category
1815 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1816
1817 // the only remaining possible overlap is identity
1818 int adr_idx = get_alias_index(adr_type);
1819 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1820 assert(adr_idx == alias_idx ||
1821 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1822 && adr_type != TypeOopPtr::BOTTOM),
1823 "should not be testing for overlap with an unsafe pointer");
1824 return adr_idx == alias_idx;
1825 }
1826
1827 //------------------------------can_alias--------------------------------------
1828 // True if any values of the given address type are in the given alias category.
1829 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1830 if (alias_idx == AliasIdxTop) return false; // the empty category
1831 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1832 // Known instance doesn't alias with bottom memory
1833 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1834 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1835
1836 // the only remaining possible overlap is identity
1837 int adr_idx = get_alias_index(adr_type);
1838 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1839 return adr_idx == alias_idx;
1840 }
1841
1842 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1843 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1844 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1845 if (parse_predicate_count() == 0) {
1846 return;
1847 }
1848 for (int i = 0; i < parse_predicate_count(); i++) {
1849 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1850 parse_predicate->mark_useless(igvn);
1851 }
1852 _parse_predicates.clear();
1853 }
1854
1855 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1856 if (!n->for_post_loop_opts_igvn()) {
1857 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1858 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1859 _for_post_loop_igvn.append(n);
1860 }
1861 }
1862
1863 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1864 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1865 _for_post_loop_igvn.remove(n);
1866 }
1867
1868 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1869 // Verify that all previous optimizations produced a valid graph
1870 // at least to this point, even if no loop optimizations were done.
1871 PhaseIdealLoop::verify(igvn);
1872
1873 if (has_loops() || _loop_opts_cnt > 0) {
1874 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1875 }
1876 C->set_post_loop_opts_phase(); // no more loop opts allowed
1877
1878 assert(!C->major_progress(), "not cleared");
1879
1880 if (_for_post_loop_igvn.length() > 0) {
1881 while (_for_post_loop_igvn.length() > 0) {
1882 Node* n = _for_post_loop_igvn.pop();
1883 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1884 igvn._worklist.push(n);
1885 }
1886 igvn.optimize();
1887 if (failing()) return;
1888 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1889 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1890
1891 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1892 if (C->major_progress()) {
1893 C->clear_major_progress(); // ensure that major progress is now clear
1894 }
1895 }
1896 }
1897
1898 void Compile::record_for_merge_stores_igvn(Node* n) {
1899 if (!n->for_merge_stores_igvn()) {
1900 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1901 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1902 _for_merge_stores_igvn.append(n);
1903 }
1904 }
1905
1906 void Compile::remove_from_merge_stores_igvn(Node* n) {
1907 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1908 _for_merge_stores_igvn.remove(n);
1909 }
1910
1911 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1912 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1913 // the stores, and we merge the wrong sequence of stores.
1914 // Example:
1915 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1916 // Apply MergeStores:
1917 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1918 // Remove more RangeChecks:
1919 // StoreI [ StoreL ] StoreI
1920 // But now it would have been better to do this instead:
1921 // [ StoreL ] [ StoreL ]
1922 //
1923 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1924 // since we never unset _merge_stores_phase.
1925 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1926 C->set_merge_stores_phase();
1927
1928 if (_for_merge_stores_igvn.length() > 0) {
1929 while (_for_merge_stores_igvn.length() > 0) {
1930 Node* n = _for_merge_stores_igvn.pop();
1931 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1932 igvn._worklist.push(n);
1933 }
1934 igvn.optimize();
1935 if (failing()) return;
1936 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1937 print_method(PHASE_AFTER_MERGE_STORES, 3);
1938 }
1939 }
1940
1941 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1942 if (OptimizeUnstableIf) {
1943 _unstable_if_traps.append(trap);
1944 }
1945 }
1946
1947 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1948 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1949 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1950 Node* n = trap->uncommon_trap();
1951 if (!useful.member(n)) {
1952 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1953 }
1954 }
1955 }
1956
1957 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1958 // or fold-compares case. Return true if succeed or not found.
1959 //
1960 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1961 // false and ask fold-compares to yield.
1962 //
1963 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1964 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1965 // when deoptimization does happen.
1966 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1967 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1968 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1969 if (trap->uncommon_trap() == unc) {
1970 if (yield && trap->modified()) {
1971 return false;
1972 }
1973 _unstable_if_traps.delete_at(i);
1974 break;
1975 }
1976 }
1977 return true;
1978 }
1979
1980 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1981 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1982 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1983 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1984 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1985 CallStaticJavaNode* unc = trap->uncommon_trap();
1986 int next_bci = trap->next_bci();
1987 bool modified = trap->modified();
1988
1989 if (next_bci != -1 && !modified) {
1990 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1991 JVMState* jvms = unc->jvms();
1992 ciMethod* method = jvms->method();
1993 ciBytecodeStream iter(method);
1994
1995 iter.force_bci(jvms->bci());
1996 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1997 Bytecodes::Code c = iter.cur_bc();
1998 Node* lhs = nullptr;
1999 Node* rhs = nullptr;
2000 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2001 lhs = unc->peek_operand(0);
2002 rhs = unc->peek_operand(1);
2003 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2004 lhs = unc->peek_operand(0);
2005 }
2006
2007 ResourceMark rm;
2008 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2009 assert(live_locals.is_valid(), "broken liveness info");
2010 int len = (int)live_locals.size();
2011
2012 for (int i = 0; i < len; i++) {
2013 Node* local = unc->local(jvms, i);
2014 // kill local using the liveness of next_bci.
2015 // give up when the local looks like an operand to secure reexecution.
2016 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2017 uint idx = jvms->locoff() + i;
2018 #ifdef ASSERT
2019 if (PrintOpto && Verbose) {
2020 tty->print("[unstable_if] kill local#%d: ", idx);
2021 local->dump();
2022 tty->cr();
2023 }
2024 #endif
2025 igvn.replace_input_of(unc, idx, top());
2026 modified = true;
2027 }
2028 }
2029 }
2030
2031 // keep the mondified trap for late query
2032 if (modified) {
2033 trap->set_modified();
2034 } else {
2035 _unstable_if_traps.delete_at(i);
2036 }
2037 }
2038 igvn.optimize();
2039 }
2040
2041 // StringOpts and late inlining of string methods
2042 void Compile::inline_string_calls(bool parse_time) {
2043 {
2044 // remove useless nodes to make the usage analysis simpler
2045 ResourceMark rm;
2046 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2047 }
2048
2049 {
2050 ResourceMark rm;
2051 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2052 PhaseStringOpts pso(initial_gvn());
2053 print_method(PHASE_AFTER_STRINGOPTS, 3);
2054 }
2055
2056 // now inline anything that we skipped the first time around
2057 if (!parse_time) {
2058 _late_inlines_pos = _late_inlines.length();
2059 }
2060
2061 while (_string_late_inlines.length() > 0) {
2062 CallGenerator* cg = _string_late_inlines.pop();
2063 cg->do_late_inline();
2064 if (failing()) return;
2065 }
2066 _string_late_inlines.trunc_to(0);
2067 }
2068
2069 // Late inlining of boxing methods
2070 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2071 if (_boxing_late_inlines.length() > 0) {
2072 assert(has_boxed_value(), "inconsistent");
2073
2074 set_inlining_incrementally(true);
2075
2076 igvn_worklist()->ensure_empty(); // should be done with igvn
2077
2078 _late_inlines_pos = _late_inlines.length();
2079
2080 while (_boxing_late_inlines.length() > 0) {
2081 CallGenerator* cg = _boxing_late_inlines.pop();
2082 cg->do_late_inline();
2083 if (failing()) return;
2084 }
2085 _boxing_late_inlines.trunc_to(0);
2086
2087 inline_incrementally_cleanup(igvn);
2088
2089 set_inlining_incrementally(false);
2090 }
2091 }
2092
2093 bool Compile::inline_incrementally_one() {
2094 assert(IncrementalInline, "incremental inlining should be on");
2095
2096 TracePhase tp(_t_incrInline_inline);
2097
2098 set_inlining_progress(false);
2099 set_do_cleanup(false);
2100
2101 for (int i = 0; i < _late_inlines.length(); i++) {
2102 _late_inlines_pos = i+1;
2103 CallGenerator* cg = _late_inlines.at(i);
2104 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2105 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2106 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2107 if (should_stress_inlining()) {
2108 // randomly add repeated inline attempt if stress-inlining
2109 cg->call_node()->set_generator(cg);
2110 C->igvn_worklist()->push(cg->call_node());
2111 continue;
2112 }
2113 cg->do_late_inline();
2114 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2115 if (failing()) {
2116 return false;
2117 } else if (inlining_progress()) {
2118 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2119 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2120 break; // process one call site at a time
2121 } else {
2122 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2123 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2124 // Avoid potential infinite loop if node already in the IGVN list
2125 assert(false, "scheduled for IGVN during inlining attempt");
2126 } else {
2127 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2128 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2129 cg->call_node()->set_generator(cg);
2130 }
2131 }
2132 } else {
2133 // Ignore late inline direct calls when inlining is not allowed.
2134 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2135 }
2136 }
2137 // Remove processed elements.
2138 _late_inlines.remove_till(_late_inlines_pos);
2139 _late_inlines_pos = 0;
2140
2141 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2142
2143 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2144
2145 set_inlining_progress(false);
2146 set_do_cleanup(false);
2147
2148 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2149 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2150 }
2151
2152 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2153 {
2154 TracePhase tp(_t_incrInline_pru);
2155 ResourceMark rm;
2156 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2157 }
2158 {
2159 TracePhase tp(_t_incrInline_igvn);
2160 igvn.reset();
2161 igvn.optimize();
2162 if (failing()) return;
2163 }
2164 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2165 }
2166
2167 // Perform incremental inlining until bound on number of live nodes is reached
2168 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2169 TracePhase tp(_t_incrInline);
2170
2171 set_inlining_incrementally(true);
2172 uint low_live_nodes = 0;
2173
2174 while (_late_inlines.length() > 0) {
2175 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2176 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2177 TracePhase tp(_t_incrInline_ideal);
2178 // PhaseIdealLoop is expensive so we only try it once we are
2179 // out of live nodes and we only try it again if the previous
2180 // helped got the number of nodes down significantly
2181 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2182 if (failing()) return;
2183 low_live_nodes = live_nodes();
2184 _major_progress = true;
2185 }
2186
2187 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2188 bool do_print_inlining = print_inlining() || print_intrinsics();
2189 if (do_print_inlining || log() != nullptr) {
2190 // Print inlining message for candidates that we couldn't inline for lack of space.
2191 for (int i = 0; i < _late_inlines.length(); i++) {
2192 CallGenerator* cg = _late_inlines.at(i);
2193 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2194 if (do_print_inlining) {
2195 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2196 }
2197 log_late_inline_failure(cg, msg);
2198 }
2199 }
2200 break; // finish
2201 }
2202 }
2203
2204 igvn_worklist()->ensure_empty(); // should be done with igvn
2205
2206 while (inline_incrementally_one()) {
2207 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2208 }
2209 if (failing()) return;
2210
2211 inline_incrementally_cleanup(igvn);
2212
2213 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2214
2215 if (failing()) return;
2216
2217 if (_late_inlines.length() == 0) {
2218 break; // no more progress
2219 }
2220 }
2221
2222 igvn_worklist()->ensure_empty(); // should be done with igvn
2223
2224 if (_string_late_inlines.length() > 0) {
2225 assert(has_stringbuilder(), "inconsistent");
2226
2227 inline_string_calls(false);
2228
2229 if (failing()) return;
2230
2231 inline_incrementally_cleanup(igvn);
2232 }
2233
2234 set_inlining_incrementally(false);
2235 }
2236
2237 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2238 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2239 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2240 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2241 // as if "inlining_incrementally() == true" were set.
2242 assert(inlining_incrementally() == false, "not allowed");
2243 assert(_modified_nodes == nullptr, "not allowed");
2244 assert(_late_inlines.length() > 0, "sanity");
2245
2246 while (_late_inlines.length() > 0) {
2247 igvn_worklist()->ensure_empty(); // should be done with igvn
2248
2249 while (inline_incrementally_one()) {
2250 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2251 }
2252 if (failing()) return;
2253
2254 inline_incrementally_cleanup(igvn);
2255 }
2256 }
2257
2258 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2259 if (_loop_opts_cnt > 0) {
2260 while (major_progress() && (_loop_opts_cnt > 0)) {
2261 TracePhase tp(_t_idealLoop);
2262 PhaseIdealLoop::optimize(igvn, mode);
2263 _loop_opts_cnt--;
2264 if (failing()) return false;
2265 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2266 }
2267 }
2268 return true;
2269 }
2270
2271 // Remove edges from "root" to each SafePoint at a backward branch.
2272 // They were inserted during parsing (see add_safepoint()) to make
2273 // infinite loops without calls or exceptions visible to root, i.e.,
2274 // useful.
2275 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2276 Node *r = root();
2277 if (r != nullptr) {
2278 for (uint i = r->req(); i < r->len(); ++i) {
2279 Node *n = r->in(i);
2280 if (n != nullptr && n->is_SafePoint()) {
2281 r->rm_prec(i);
2282 if (n->outcnt() == 0) {
2283 igvn.remove_dead_node(n);
2284 }
2285 --i;
2286 }
2287 }
2288 // Parsing may have added top inputs to the root node (Path
2289 // leading to the Halt node proven dead). Make sure we get a
2290 // chance to clean them up.
2291 igvn._worklist.push(r);
2292 igvn.optimize();
2293 }
2294 }
2295
2296 //------------------------------Optimize---------------------------------------
2297 // Given a graph, optimize it.
2298 void Compile::Optimize() {
2299 TracePhase tp(_t_optimizer);
2300
2301 #ifndef PRODUCT
2302 if (env()->break_at_compile()) {
2303 BREAKPOINT;
2304 }
2305
2306 #endif
2307
2308 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2309 #ifdef ASSERT
2310 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2311 #endif
2312
2313 ResourceMark rm;
2314
2315 NOT_PRODUCT( verify_graph_edges(); )
2316
2317 print_method(PHASE_AFTER_PARSING, 1);
2318
2319 {
2320 // Iterative Global Value Numbering, including ideal transforms
2321 PhaseIterGVN igvn;
2322 #ifdef ASSERT
2323 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2324 #endif
2325 {
2326 TracePhase tp(_t_iterGVN);
2327 igvn.optimize();
2328 }
2329
2330 if (failing()) return;
2331
2332 print_method(PHASE_ITER_GVN1, 2);
2333
2334 process_for_unstable_if_traps(igvn);
2335
2336 if (failing()) return;
2337
2338 inline_incrementally(igvn);
2339
2340 print_method(PHASE_INCREMENTAL_INLINE, 2);
2341
2342 if (failing()) return;
2343
2344 if (eliminate_boxing()) {
2345 // Inline valueOf() methods now.
2346 inline_boxing_calls(igvn);
2347
2348 if (failing()) return;
2349
2350 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2351 inline_incrementally(igvn);
2352 }
2353
2354 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2355
2356 if (failing()) return;
2357 }
2358
2359 // Remove the speculative part of types and clean up the graph from
2360 // the extra CastPP nodes whose only purpose is to carry them. Do
2361 // that early so that optimizations are not disrupted by the extra
2362 // CastPP nodes.
2363 remove_speculative_types(igvn);
2364
2365 if (failing()) return;
2366
2367 // No more new expensive nodes will be added to the list from here
2368 // so keep only the actual candidates for optimizations.
2369 cleanup_expensive_nodes(igvn);
2370
2371 if (failing()) return;
2372
2373 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2374 if (EnableVectorSupport && has_vbox_nodes()) {
2375 TracePhase tp(_t_vector);
2376 PhaseVector pv(igvn);
2377 pv.optimize_vector_boxes();
2378 if (failing()) return;
2379 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2380 }
2381 assert(!has_vbox_nodes(), "sanity");
2382
2383 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2384 Compile::TracePhase tp(_t_renumberLive);
2385 igvn_worklist()->ensure_empty(); // should be done with igvn
2386 {
2387 ResourceMark rm;
2388 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2389 }
2390 igvn.reset();
2391 igvn.optimize();
2392 if (failing()) return;
2393 }
2394
2395 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2396 // safepoints
2397 remove_root_to_sfpts_edges(igvn);
2398
2399 if (failing()) return;
2400
2401 if (has_loops()) {
2402 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2403 }
2404
2405 // Perform escape analysis
2406 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2407 if (has_loops()) {
2408 // Cleanup graph (remove dead nodes).
2409 TracePhase tp(_t_idealLoop);
2410 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2411 if (failing()) return;
2412 }
2413 bool progress;
2414 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2415 do {
2416 ConnectionGraph::do_analysis(this, &igvn);
2417
2418 if (failing()) return;
2419
2420 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2421
2422 // Optimize out fields loads from scalar replaceable allocations.
2423 igvn.optimize();
2424 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2425
2426 if (failing()) return;
2427
2428 if (congraph() != nullptr && macro_count() > 0) {
2429 TracePhase tp(_t_macroEliminate);
2430 PhaseMacroExpand mexp(igvn);
2431 mexp.eliminate_macro_nodes();
2432 if (failing()) return;
2433 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2434
2435 igvn.set_delay_transform(false);
2436 igvn.optimize();
2437 if (failing()) return;
2438
2439 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2440 }
2441
2442 ConnectionGraph::verify_ram_nodes(this, root());
2443 if (failing()) return;
2444
2445 progress = do_iterative_escape_analysis() &&
2446 (macro_count() < mcount) &&
2447 ConnectionGraph::has_candidates(this);
2448 // Try again if candidates exist and made progress
2449 // by removing some allocations and/or locks.
2450 } while (progress);
2451 }
2452
2453 // Loop transforms on the ideal graph. Range Check Elimination,
2454 // peeling, unrolling, etc.
2455
2456 // Set loop opts counter
2457 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2458 {
2459 TracePhase tp(_t_idealLoop);
2460 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2461 _loop_opts_cnt--;
2462 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2463 if (failing()) return;
2464 }
2465 // Loop opts pass if partial peeling occurred in previous pass
2466 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2467 TracePhase tp(_t_idealLoop);
2468 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2469 _loop_opts_cnt--;
2470 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2471 if (failing()) return;
2472 }
2473 // Loop opts pass for loop-unrolling before CCP
2474 if(major_progress() && (_loop_opts_cnt > 0)) {
2475 TracePhase tp(_t_idealLoop);
2476 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2477 _loop_opts_cnt--;
2478 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2479 }
2480 if (!failing()) {
2481 // Verify that last round of loop opts produced a valid graph
2482 PhaseIdealLoop::verify(igvn);
2483 }
2484 }
2485 if (failing()) return;
2486
2487 // Conditional Constant Propagation;
2488 print_method(PHASE_BEFORE_CCP1, 2);
2489 PhaseCCP ccp( &igvn );
2490 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2491 {
2492 TracePhase tp(_t_ccp);
2493 ccp.do_transform();
2494 }
2495 print_method(PHASE_CCP1, 2);
2496
2497 assert( true, "Break here to ccp.dump_old2new_map()");
2498
2499 // Iterative Global Value Numbering, including ideal transforms
2500 {
2501 TracePhase tp(_t_iterGVN2);
2502 igvn.reset_from_igvn(&ccp);
2503 igvn.optimize();
2504 }
2505 print_method(PHASE_ITER_GVN2, 2);
2506
2507 if (failing()) return;
2508
2509 // Loop transforms on the ideal graph. Range Check Elimination,
2510 // peeling, unrolling, etc.
2511 if (!optimize_loops(igvn, LoopOptsDefault)) {
2512 return;
2513 }
2514
2515 if (failing()) return;
2516
2517 C->clear_major_progress(); // ensure that major progress is now clear
2518
2519 process_for_post_loop_opts_igvn(igvn);
2520
2521 process_for_merge_stores_igvn(igvn);
2522
2523 if (failing()) return;
2524
2525 #ifdef ASSERT
2526 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2527 #endif
2528
2529 {
2530 TracePhase tp(_t_macroExpand);
2531 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2532 PhaseMacroExpand mex(igvn);
2533 // Do not allow new macro nodes once we start to eliminate and expand
2534 C->reset_allow_macro_nodes();
2535 // Last attempt to eliminate macro nodes before expand
2536 mex.eliminate_macro_nodes();
2537 if (failing()) {
2538 return;
2539 }
2540 mex.eliminate_opaque_looplimit_macro_nodes();
2541 if (failing()) {
2542 return;
2543 }
2544 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2545 if (mex.expand_macro_nodes()) {
2546 assert(failing(), "must bail out w/ explicit message");
2547 return;
2548 }
2549 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2550 }
2551
2552 {
2553 TracePhase tp(_t_barrierExpand);
2554 if (bs->expand_barriers(this, igvn)) {
2555 assert(failing(), "must bail out w/ explicit message");
2556 return;
2557 }
2558 print_method(PHASE_BARRIER_EXPANSION, 2);
2559 }
2560
2561 if (C->max_vector_size() > 0) {
2562 C->optimize_logic_cones(igvn);
2563 igvn.optimize();
2564 if (failing()) return;
2565 }
2566
2567 DEBUG_ONLY( _modified_nodes = nullptr; )
2568
2569 assert(igvn._worklist.size() == 0, "not empty");
2570
2571 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2572
2573 if (_late_inlines.length() > 0) {
2574 // More opportunities to optimize virtual and MH calls.
2575 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2576 process_late_inline_calls_no_inline(igvn);
2577 if (failing()) return;
2578 }
2579 } // (End scope of igvn; run destructor if necessary for asserts.)
2580
2581 check_no_dead_use();
2582
2583 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2584 // to remove hashes to unlock nodes for modifications.
2585 C->node_hash()->clear();
2586
2587 // A method with only infinite loops has no edges entering loops from root
2588 {
2589 TracePhase tp(_t_graphReshaping);
2590 if (final_graph_reshaping()) {
2591 assert(failing(), "must bail out w/ explicit message");
2592 return;
2593 }
2594 }
2595
2596 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2597 DEBUG_ONLY(set_phase_optimize_finished();)
2598 }
2599
2600 #ifdef ASSERT
2601 void Compile::check_no_dead_use() const {
2602 ResourceMark rm;
2603 Unique_Node_List wq;
2604 wq.push(root());
2605 for (uint i = 0; i < wq.size(); ++i) {
2606 Node* n = wq.at(i);
2607 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2608 Node* u = n->fast_out(j);
2609 if (u->outcnt() == 0 && !u->is_Con()) {
2610 u->dump();
2611 fatal("no reachable node should have no use");
2612 }
2613 wq.push(u);
2614 }
2615 }
2616 }
2617 #endif
2618
2619 void Compile::inline_vector_reboxing_calls() {
2620 if (C->_vector_reboxing_late_inlines.length() > 0) {
2621 _late_inlines_pos = C->_late_inlines.length();
2622 while (_vector_reboxing_late_inlines.length() > 0) {
2623 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2624 cg->do_late_inline();
2625 if (failing()) return;
2626 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2627 }
2628 _vector_reboxing_late_inlines.trunc_to(0);
2629 }
2630 }
2631
2632 bool Compile::has_vbox_nodes() {
2633 if (C->_vector_reboxing_late_inlines.length() > 0) {
2634 return true;
2635 }
2636 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2637 Node * n = C->macro_node(macro_idx);
2638 assert(n->is_macro(), "only macro nodes expected here");
2639 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2640 return true;
2641 }
2642 }
2643 return false;
2644 }
2645
2646 //---------------------------- Bitwise operation packing optimization ---------------------------
2647
2648 static bool is_vector_unary_bitwise_op(Node* n) {
2649 return n->Opcode() == Op_XorV &&
2650 VectorNode::is_vector_bitwise_not_pattern(n);
2651 }
2652
2653 static bool is_vector_binary_bitwise_op(Node* n) {
2654 switch (n->Opcode()) {
2655 case Op_AndV:
2656 case Op_OrV:
2657 return true;
2658
2659 case Op_XorV:
2660 return !is_vector_unary_bitwise_op(n);
2661
2662 default:
2663 return false;
2664 }
2665 }
2666
2667 static bool is_vector_ternary_bitwise_op(Node* n) {
2668 return n->Opcode() == Op_MacroLogicV;
2669 }
2670
2671 static bool is_vector_bitwise_op(Node* n) {
2672 return is_vector_unary_bitwise_op(n) ||
2673 is_vector_binary_bitwise_op(n) ||
2674 is_vector_ternary_bitwise_op(n);
2675 }
2676
2677 static bool is_vector_bitwise_cone_root(Node* n) {
2678 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2679 return false;
2680 }
2681 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2682 if (is_vector_bitwise_op(n->fast_out(i))) {
2683 return false;
2684 }
2685 }
2686 return true;
2687 }
2688
2689 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2690 uint cnt = 0;
2691 if (is_vector_bitwise_op(n)) {
2692 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2693 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2694 for (uint i = 1; i < inp_cnt; i++) {
2695 Node* in = n->in(i);
2696 bool skip = VectorNode::is_all_ones_vector(in);
2697 if (!skip && !inputs.member(in)) {
2698 inputs.push(in);
2699 cnt++;
2700 }
2701 }
2702 assert(cnt <= 1, "not unary");
2703 } else {
2704 uint last_req = inp_cnt;
2705 if (is_vector_ternary_bitwise_op(n)) {
2706 last_req = inp_cnt - 1; // skip last input
2707 }
2708 for (uint i = 1; i < last_req; i++) {
2709 Node* def = n->in(i);
2710 if (!inputs.member(def)) {
2711 inputs.push(def);
2712 cnt++;
2713 }
2714 }
2715 }
2716 } else { // not a bitwise operations
2717 if (!inputs.member(n)) {
2718 inputs.push(n);
2719 cnt++;
2720 }
2721 }
2722 return cnt;
2723 }
2724
2725 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2726 Unique_Node_List useful_nodes;
2727 C->identify_useful_nodes(useful_nodes);
2728
2729 for (uint i = 0; i < useful_nodes.size(); i++) {
2730 Node* n = useful_nodes.at(i);
2731 if (is_vector_bitwise_cone_root(n)) {
2732 list.push(n);
2733 }
2734 }
2735 }
2736
2737 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2738 const TypeVect* vt,
2739 Unique_Node_List& partition,
2740 Unique_Node_List& inputs) {
2741 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2742 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2743 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2744
2745 Node* in1 = inputs.at(0);
2746 Node* in2 = inputs.at(1);
2747 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2748
2749 uint func = compute_truth_table(partition, inputs);
2750
2751 Node* pn = partition.at(partition.size() - 1);
2752 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2753 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2754 }
2755
2756 static uint extract_bit(uint func, uint pos) {
2757 return (func & (1 << pos)) >> pos;
2758 }
2759
2760 //
2761 // A macro logic node represents a truth table. It has 4 inputs,
2762 // First three inputs corresponds to 3 columns of a truth table
2763 // and fourth input captures the logic function.
2764 //
2765 // eg. fn = (in1 AND in2) OR in3;
2766 //
2767 // MacroNode(in1,in2,in3,fn)
2768 //
2769 // -----------------
2770 // in1 in2 in3 fn
2771 // -----------------
2772 // 0 0 0 0
2773 // 0 0 1 1
2774 // 0 1 0 0
2775 // 0 1 1 1
2776 // 1 0 0 0
2777 // 1 0 1 1
2778 // 1 1 0 1
2779 // 1 1 1 1
2780 //
2781
2782 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2783 int res = 0;
2784 for (int i = 0; i < 8; i++) {
2785 int bit1 = extract_bit(in1, i);
2786 int bit2 = extract_bit(in2, i);
2787 int bit3 = extract_bit(in3, i);
2788
2789 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2790 int func_bit = extract_bit(func, func_bit_pos);
2791
2792 res |= func_bit << i;
2793 }
2794 return res;
2795 }
2796
2797 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
2798 assert(n != nullptr, "");
2799 assert(eval_map.contains(n), "absent");
2800 return *(eval_map.get(n));
2801 }
2802
2803 static void eval_operands(Node* n,
2804 uint& func1, uint& func2, uint& func3,
2805 HashTable<Node*,uint>& eval_map) {
2806 assert(is_vector_bitwise_op(n), "");
2807
2808 if (is_vector_unary_bitwise_op(n)) {
2809 Node* opnd = n->in(1);
2810 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2811 opnd = n->in(2);
2812 }
2813 func1 = eval_operand(opnd, eval_map);
2814 } else if (is_vector_binary_bitwise_op(n)) {
2815 func1 = eval_operand(n->in(1), eval_map);
2816 func2 = eval_operand(n->in(2), eval_map);
2817 } else {
2818 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2819 func1 = eval_operand(n->in(1), eval_map);
2820 func2 = eval_operand(n->in(2), eval_map);
2821 func3 = eval_operand(n->in(3), eval_map);
2822 }
2823 }
2824
2825 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2826 assert(inputs.size() <= 3, "sanity");
2827 ResourceMark rm;
2828 uint res = 0;
2829 HashTable<Node*,uint> eval_map;
2830
2831 // Populate precomputed functions for inputs.
2832 // Each input corresponds to one column of 3 input truth-table.
2833 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2834 0xCC, // (_, b, _) -> b
2835 0xF0 }; // (a, _, _) -> a
2836 for (uint i = 0; i < inputs.size(); i++) {
2837 eval_map.put(inputs.at(i), input_funcs[2-i]);
2838 }
2839
2840 for (uint i = 0; i < partition.size(); i++) {
2841 Node* n = partition.at(i);
2842
2843 uint func1 = 0, func2 = 0, func3 = 0;
2844 eval_operands(n, func1, func2, func3, eval_map);
2845
2846 switch (n->Opcode()) {
2847 case Op_OrV:
2848 assert(func3 == 0, "not binary");
2849 res = func1 | func2;
2850 break;
2851 case Op_AndV:
2852 assert(func3 == 0, "not binary");
2853 res = func1 & func2;
2854 break;
2855 case Op_XorV:
2856 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2857 assert(func2 == 0 && func3 == 0, "not unary");
2858 res = (~func1) & 0xFF;
2859 } else {
2860 assert(func3 == 0, "not binary");
2861 res = func1 ^ func2;
2862 }
2863 break;
2864 case Op_MacroLogicV:
2865 // Ordering of inputs may change during evaluation of sub-tree
2866 // containing MacroLogic node as a child node, thus a re-evaluation
2867 // makes sure that function is evaluated in context of current
2868 // inputs.
2869 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2870 break;
2871
2872 default: assert(false, "not supported: %s", n->Name());
2873 }
2874 assert(res <= 0xFF, "invalid");
2875 eval_map.put(n, res);
2876 }
2877 return res;
2878 }
2879
2880 // Criteria under which nodes gets packed into a macro logic node:-
2881 // 1) Parent and both child nodes are all unmasked or masked with
2882 // same predicates.
2883 // 2) Masked parent can be packed with left child if it is predicated
2884 // and both have same predicates.
2885 // 3) Masked parent can be packed with right child if its un-predicated
2886 // or has matching predication condition.
2887 // 4) An unmasked parent can be packed with an unmasked child.
2888 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2889 assert(partition.size() == 0, "not empty");
2890 assert(inputs.size() == 0, "not empty");
2891 if (is_vector_ternary_bitwise_op(n)) {
2892 return false;
2893 }
2894
2895 bool is_unary_op = is_vector_unary_bitwise_op(n);
2896 if (is_unary_op) {
2897 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2898 return false; // too few inputs
2899 }
2900
2901 bool pack_left_child = true;
2902 bool pack_right_child = true;
2903
2904 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2905 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2906
2907 int left_child_input_cnt = 0;
2908 int right_child_input_cnt = 0;
2909
2910 bool parent_is_predicated = n->is_predicated_vector();
2911 bool left_child_predicated = n->in(1)->is_predicated_vector();
2912 bool right_child_predicated = n->in(2)->is_predicated_vector();
2913
2914 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2915 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2916 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2917
2918 do {
2919 if (pack_left_child && left_child_LOP &&
2920 ((!parent_is_predicated && !left_child_predicated) ||
2921 ((parent_is_predicated && left_child_predicated &&
2922 parent_pred == left_child_pred)))) {
2923 partition.push(n->in(1));
2924 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2925 } else {
2926 inputs.push(n->in(1));
2927 left_child_input_cnt = 1;
2928 }
2929
2930 if (pack_right_child && right_child_LOP &&
2931 (!right_child_predicated ||
2932 (right_child_predicated && parent_is_predicated &&
2933 parent_pred == right_child_pred))) {
2934 partition.push(n->in(2));
2935 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2936 } else {
2937 inputs.push(n->in(2));
2938 right_child_input_cnt = 1;
2939 }
2940
2941 if (inputs.size() > 3) {
2942 assert(partition.size() > 0, "");
2943 inputs.clear();
2944 partition.clear();
2945 if (left_child_input_cnt > right_child_input_cnt) {
2946 pack_left_child = false;
2947 } else {
2948 pack_right_child = false;
2949 }
2950 } else {
2951 break;
2952 }
2953 } while(true);
2954
2955 if(partition.size()) {
2956 partition.push(n);
2957 }
2958
2959 return (partition.size() == 2 || partition.size() == 3) &&
2960 (inputs.size() == 2 || inputs.size() == 3);
2961 }
2962
2963 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2964 assert(is_vector_bitwise_op(n), "not a root");
2965
2966 visited.set(n->_idx);
2967
2968 // 1) Do a DFS walk over the logic cone.
2969 for (uint i = 1; i < n->req(); i++) {
2970 Node* in = n->in(i);
2971 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2972 process_logic_cone_root(igvn, in, visited);
2973 }
2974 }
2975
2976 // 2) Bottom up traversal: Merge node[s] with
2977 // the parent to form macro logic node.
2978 Unique_Node_List partition;
2979 Unique_Node_List inputs;
2980 if (compute_logic_cone(n, partition, inputs)) {
2981 const TypeVect* vt = n->bottom_type()->is_vect();
2982 Node* pn = partition.at(partition.size() - 1);
2983 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2984 if (mask == nullptr ||
2985 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2986 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2987 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2988 igvn.replace_node(n, macro_logic);
2989 }
2990 }
2991 }
2992
2993 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2994 ResourceMark rm;
2995 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2996 Unique_Node_List list;
2997 collect_logic_cone_roots(list);
2998
2999 while (list.size() > 0) {
3000 Node* n = list.pop();
3001 const TypeVect* vt = n->bottom_type()->is_vect();
3002 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3003 if (supported) {
3004 VectorSet visited(comp_arena());
3005 process_logic_cone_root(igvn, n, visited);
3006 }
3007 }
3008 }
3009 }
3010
3011 //------------------------------Code_Gen---------------------------------------
3012 // Given a graph, generate code for it
3013 void Compile::Code_Gen() {
3014 if (failing()) {
3015 return;
3016 }
3017
3018 // Perform instruction selection. You might think we could reclaim Matcher
3019 // memory PDQ, but actually the Matcher is used in generating spill code.
3020 // Internals of the Matcher (including some VectorSets) must remain live
3021 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3022 // set a bit in reclaimed memory.
3023
3024 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3025 // nodes. Mapping is only valid at the root of each matched subtree.
3026 NOT_PRODUCT( verify_graph_edges(); )
3027
3028 Matcher matcher;
3029 _matcher = &matcher;
3030 {
3031 TracePhase tp(_t_matcher);
3032 matcher.match();
3033 if (failing()) {
3034 return;
3035 }
3036 }
3037 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3038 // nodes. Mapping is only valid at the root of each matched subtree.
3039 NOT_PRODUCT( verify_graph_edges(); )
3040
3041 // If you have too many nodes, or if matching has failed, bail out
3042 check_node_count(0, "out of nodes matching instructions");
3043 if (failing()) {
3044 return;
3045 }
3046
3047 print_method(PHASE_MATCHING, 2);
3048
3049 // Build a proper-looking CFG
3050 PhaseCFG cfg(node_arena(), root(), matcher);
3051 if (failing()) {
3052 return;
3053 }
3054 _cfg = &cfg;
3055 {
3056 TracePhase tp(_t_scheduler);
3057 bool success = cfg.do_global_code_motion();
3058 if (!success) {
3059 return;
3060 }
3061
3062 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3063 NOT_PRODUCT( verify_graph_edges(); )
3064 cfg.verify();
3065 if (failing()) {
3066 return;
3067 }
3068 }
3069
3070 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3071 _regalloc = ®alloc;
3072 {
3073 TracePhase tp(_t_registerAllocation);
3074 // Perform register allocation. After Chaitin, use-def chains are
3075 // no longer accurate (at spill code) and so must be ignored.
3076 // Node->LRG->reg mappings are still accurate.
3077 _regalloc->Register_Allocate();
3078
3079 // Bail out if the allocator builds too many nodes
3080 if (failing()) {
3081 return;
3082 }
3083
3084 print_method(PHASE_REGISTER_ALLOCATION, 2);
3085 }
3086
3087 // Prior to register allocation we kept empty basic blocks in case the
3088 // the allocator needed a place to spill. After register allocation we
3089 // are not adding any new instructions. If any basic block is empty, we
3090 // can now safely remove it.
3091 {
3092 TracePhase tp(_t_blockOrdering);
3093 cfg.remove_empty_blocks();
3094 if (do_freq_based_layout()) {
3095 PhaseBlockLayout layout(cfg);
3096 } else {
3097 cfg.set_loop_alignment();
3098 }
3099 cfg.fixup_flow();
3100 cfg.remove_unreachable_blocks();
3101 cfg.verify_dominator_tree();
3102 print_method(PHASE_BLOCK_ORDERING, 3);
3103 }
3104
3105 // Apply peephole optimizations
3106 if( OptoPeephole ) {
3107 TracePhase tp(_t_peephole);
3108 PhasePeephole peep( _regalloc, cfg);
3109 peep.do_transform();
3110 print_method(PHASE_PEEPHOLE, 3);
3111 }
3112
3113 // Do late expand if CPU requires this.
3114 if (Matcher::require_postalloc_expand) {
3115 TracePhase tp(_t_postalloc_expand);
3116 cfg.postalloc_expand(_regalloc);
3117 print_method(PHASE_POSTALLOC_EXPAND, 3);
3118 }
3119
3120 #ifdef ASSERT
3121 {
3122 CompilationMemoryStatistic::do_test_allocations();
3123 if (failing()) return;
3124 }
3125 #endif
3126
3127 // Convert Nodes to instruction bits in a buffer
3128 {
3129 TracePhase tp(_t_output);
3130 PhaseOutput output;
3131 output.Output();
3132 if (failing()) return;
3133 output.install();
3134 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3135 }
3136
3137 // He's dead, Jim.
3138 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3139 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3140 }
3141
3142 //------------------------------Final_Reshape_Counts---------------------------
3143 // This class defines counters to help identify when a method
3144 // may/must be executed using hardware with only 24-bit precision.
3145 struct Final_Reshape_Counts : public StackObj {
3146 int _call_count; // count non-inlined 'common' calls
3147 int _float_count; // count float ops requiring 24-bit precision
3148 int _double_count; // count double ops requiring more precision
3149 int _java_call_count; // count non-inlined 'java' calls
3150 int _inner_loop_count; // count loops which need alignment
3151 VectorSet _visited; // Visitation flags
3152 Node_List _tests; // Set of IfNodes & PCTableNodes
3153
3154 Final_Reshape_Counts() :
3155 _call_count(0), _float_count(0), _double_count(0),
3156 _java_call_count(0), _inner_loop_count(0) { }
3157
3158 void inc_call_count () { _call_count ++; }
3159 void inc_float_count () { _float_count ++; }
3160 void inc_double_count() { _double_count++; }
3161 void inc_java_call_count() { _java_call_count++; }
3162 void inc_inner_loop_count() { _inner_loop_count++; }
3163
3164 int get_call_count () const { return _call_count ; }
3165 int get_float_count () const { return _float_count ; }
3166 int get_double_count() const { return _double_count; }
3167 int get_java_call_count() const { return _java_call_count; }
3168 int get_inner_loop_count() const { return _inner_loop_count; }
3169 };
3170
3171 //------------------------------final_graph_reshaping_impl----------------------
3172 // Implement items 1-5 from final_graph_reshaping below.
3173 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3174
3175 if ( n->outcnt() == 0 ) return; // dead node
3176 uint nop = n->Opcode();
3177
3178 // Check for 2-input instruction with "last use" on right input.
3179 // Swap to left input. Implements item (2).
3180 if( n->req() == 3 && // two-input instruction
3181 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3182 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3183 n->in(2)->outcnt() == 1 &&// right use IS a last use
3184 !n->in(2)->is_Con() ) { // right use is not a constant
3185 // Check for commutative opcode
3186 switch( nop ) {
3187 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3188 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3189 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3190 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3191 case Op_AndL: case Op_XorL: case Op_OrL:
3192 case Op_AndI: case Op_XorI: case Op_OrI: {
3193 // Move "last use" input to left by swapping inputs
3194 n->swap_edges(1, 2);
3195 break;
3196 }
3197 default:
3198 break;
3199 }
3200 }
3201
3202 #ifdef ASSERT
3203 if( n->is_Mem() ) {
3204 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3205 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3206 // oop will be recorded in oop map if load crosses safepoint
3207 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3208 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3209 "raw memory operations should have control edge");
3210 }
3211 if (n->is_MemBar()) {
3212 MemBarNode* mb = n->as_MemBar();
3213 if (mb->trailing_store() || mb->trailing_load_store()) {
3214 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3215 Node* mem = mb->in(MemBarNode::Precedent);
3216 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3217 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3218 } else if (mb->leading()) {
3219 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3220 }
3221 }
3222 #endif
3223 // Count FPU ops and common calls, implements item (3)
3224 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3225
3226 // Collect CFG split points
3227 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3228 frc._tests.push(n);
3229 }
3230 }
3231
3232 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3233 if (!UseDivMod) {
3234 return;
3235 }
3236
3237 // Check if "a % b" and "a / b" both exist
3238 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3239 if (d == nullptr) {
3240 return;
3241 }
3242
3243 // Replace them with a fused divmod if supported
3244 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3245 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3246 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3247 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3248 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3249 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3250 // DivMod node so the dependency is not lost.
3251 divmod->add_prec_from(n);
3252 divmod->add_prec_from(d);
3253 d->subsume_by(divmod->div_proj(), this);
3254 n->subsume_by(divmod->mod_proj(), this);
3255 } else {
3256 // Replace "a % b" with "a - ((a / b) * b)"
3257 Node* mult = MulNode::make(d, d->in(2), bt);
3258 Node* sub = SubNode::make(d->in(1), mult, bt);
3259 n->subsume_by(sub, this);
3260 }
3261 }
3262
3263 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3264 switch( nop ) {
3265 // Count all float operations that may use FPU
3266 case Op_AddF:
3267 case Op_SubF:
3268 case Op_MulF:
3269 case Op_DivF:
3270 case Op_NegF:
3271 case Op_ModF:
3272 case Op_ConvI2F:
3273 case Op_ConF:
3274 case Op_CmpF:
3275 case Op_CmpF3:
3276 case Op_StoreF:
3277 case Op_LoadF:
3278 // case Op_ConvL2F: // longs are split into 32-bit halves
3279 frc.inc_float_count();
3280 break;
3281
3282 case Op_ConvF2D:
3283 case Op_ConvD2F:
3284 frc.inc_float_count();
3285 frc.inc_double_count();
3286 break;
3287
3288 // Count all double operations that may use FPU
3289 case Op_AddD:
3290 case Op_SubD:
3291 case Op_MulD:
3292 case Op_DivD:
3293 case Op_NegD:
3294 case Op_ModD:
3295 case Op_ConvI2D:
3296 case Op_ConvD2I:
3297 // case Op_ConvL2D: // handled by leaf call
3298 // case Op_ConvD2L: // handled by leaf call
3299 case Op_ConD:
3300 case Op_CmpD:
3301 case Op_CmpD3:
3302 case Op_StoreD:
3303 case Op_LoadD:
3304 case Op_LoadD_unaligned:
3305 frc.inc_double_count();
3306 break;
3307 case Op_Opaque1: // Remove Opaque Nodes before matching
3308 n->subsume_by(n->in(1), this);
3309 break;
3310 case Op_CallLeafPure: {
3311 // If the pure call is not supported, then lower to a CallLeaf.
3312 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3313 CallNode* call = n->as_Call();
3314 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3315 call->_name, TypeRawPtr::BOTTOM);
3316 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3317 new_call->init_req(TypeFunc::I_O, C->top());
3318 new_call->init_req(TypeFunc::Memory, C->top());
3319 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3320 new_call->init_req(TypeFunc::FramePtr, C->top());
3321 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
3322 new_call->init_req(i, call->in(i));
3323 }
3324 n->subsume_by(new_call, this);
3325 }
3326 frc.inc_call_count();
3327 break;
3328 }
3329 case Op_CallStaticJava:
3330 case Op_CallJava:
3331 case Op_CallDynamicJava:
3332 frc.inc_java_call_count(); // Count java call site;
3333 case Op_CallRuntime:
3334 case Op_CallLeaf:
3335 case Op_CallLeafVector:
3336 case Op_CallLeafNoFP: {
3337 assert (n->is_Call(), "");
3338 CallNode *call = n->as_Call();
3339 // Count call sites where the FP mode bit would have to be flipped.
3340 // Do not count uncommon runtime calls:
3341 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3342 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3343 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3344 frc.inc_call_count(); // Count the call site
3345 } else { // See if uncommon argument is shared
3346 Node *n = call->in(TypeFunc::Parms);
3347 int nop = n->Opcode();
3348 // Clone shared simple arguments to uncommon calls, item (1).
3349 if (n->outcnt() > 1 &&
3350 !n->is_Proj() &&
3351 nop != Op_CreateEx &&
3352 nop != Op_CheckCastPP &&
3353 nop != Op_DecodeN &&
3354 nop != Op_DecodeNKlass &&
3355 !n->is_Mem() &&
3356 !n->is_Phi()) {
3357 Node *x = n->clone();
3358 call->set_req(TypeFunc::Parms, x);
3359 }
3360 }
3361 break;
3362 }
3363 case Op_StoreB:
3364 case Op_StoreC:
3365 case Op_StoreI:
3366 case Op_StoreL:
3367 case Op_CompareAndSwapB:
3368 case Op_CompareAndSwapS:
3369 case Op_CompareAndSwapI:
3370 case Op_CompareAndSwapL:
3371 case Op_CompareAndSwapP:
3372 case Op_CompareAndSwapN:
3373 case Op_WeakCompareAndSwapB:
3374 case Op_WeakCompareAndSwapS:
3375 case Op_WeakCompareAndSwapI:
3376 case Op_WeakCompareAndSwapL:
3377 case Op_WeakCompareAndSwapP:
3378 case Op_WeakCompareAndSwapN:
3379 case Op_CompareAndExchangeB:
3380 case Op_CompareAndExchangeS:
3381 case Op_CompareAndExchangeI:
3382 case Op_CompareAndExchangeL:
3383 case Op_CompareAndExchangeP:
3384 case Op_CompareAndExchangeN:
3385 case Op_GetAndAddS:
3386 case Op_GetAndAddB:
3387 case Op_GetAndAddI:
3388 case Op_GetAndAddL:
3389 case Op_GetAndSetS:
3390 case Op_GetAndSetB:
3391 case Op_GetAndSetI:
3392 case Op_GetAndSetL:
3393 case Op_GetAndSetP:
3394 case Op_GetAndSetN:
3395 case Op_StoreP:
3396 case Op_StoreN:
3397 case Op_StoreNKlass:
3398 case Op_LoadB:
3399 case Op_LoadUB:
3400 case Op_LoadUS:
3401 case Op_LoadI:
3402 case Op_LoadKlass:
3403 case Op_LoadNKlass:
3404 case Op_LoadL:
3405 case Op_LoadL_unaligned:
3406 case Op_LoadP:
3407 case Op_LoadN:
3408 case Op_LoadRange:
3409 case Op_LoadS:
3410 break;
3411
3412 case Op_AddP: { // Assert sane base pointers
3413 Node *addp = n->in(AddPNode::Address);
3414 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
3415 #ifdef _LP64
3416 if ((UseCompressedOops || UseCompressedClassPointers) &&
3417 addp->Opcode() == Op_ConP &&
3418 addp == n->in(AddPNode::Base) &&
3419 n->in(AddPNode::Offset)->is_Con()) {
3420 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3421 // on the platform and on the compressed oops mode.
3422 // Use addressing with narrow klass to load with offset on x86.
3423 // Some platforms can use the constant pool to load ConP.
3424 // Do this transformation here since IGVN will convert ConN back to ConP.
3425 const Type* t = addp->bottom_type();
3426 bool is_oop = t->isa_oopptr() != nullptr;
3427 bool is_klass = t->isa_klassptr() != nullptr;
3428
3429 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3430 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3431 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3432 Node* nn = nullptr;
3433
3434 int op = is_oop ? Op_ConN : Op_ConNKlass;
3435
3436 // Look for existing ConN node of the same exact type.
3437 Node* r = root();
3438 uint cnt = r->outcnt();
3439 for (uint i = 0; i < cnt; i++) {
3440 Node* m = r->raw_out(i);
3441 if (m!= nullptr && m->Opcode() == op &&
3442 m->bottom_type()->make_ptr() == t) {
3443 nn = m;
3444 break;
3445 }
3446 }
3447 if (nn != nullptr) {
3448 // Decode a narrow oop to match address
3449 // [R12 + narrow_oop_reg<<3 + offset]
3450 if (is_oop) {
3451 nn = new DecodeNNode(nn, t);
3452 } else {
3453 nn = new DecodeNKlassNode(nn, t);
3454 }
3455 // Check for succeeding AddP which uses the same Base.
3456 // Otherwise we will run into the assertion above when visiting that guy.
3457 for (uint i = 0; i < n->outcnt(); ++i) {
3458 Node *out_i = n->raw_out(i);
3459 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3460 out_i->set_req(AddPNode::Base, nn);
3461 #ifdef ASSERT
3462 for (uint j = 0; j < out_i->outcnt(); ++j) {
3463 Node *out_j = out_i->raw_out(j);
3464 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3465 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3466 }
3467 #endif
3468 }
3469 }
3470 n->set_req(AddPNode::Base, nn);
3471 n->set_req(AddPNode::Address, nn);
3472 if (addp->outcnt() == 0) {
3473 addp->disconnect_inputs(this);
3474 }
3475 }
3476 }
3477 }
3478 #endif
3479 break;
3480 }
3481
3482 case Op_CastPP: {
3483 // Remove CastPP nodes to gain more freedom during scheduling but
3484 // keep the dependency they encode as control or precedence edges
3485 // (if control is set already) on memory operations. Some CastPP
3486 // nodes don't have a control (don't carry a dependency): skip
3487 // those.
3488 if (n->in(0) != nullptr) {
3489 ResourceMark rm;
3490 Unique_Node_List wq;
3491 wq.push(n);
3492 for (uint next = 0; next < wq.size(); ++next) {
3493 Node *m = wq.at(next);
3494 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3495 Node* use = m->fast_out(i);
3496 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3497 use->ensure_control_or_add_prec(n->in(0));
3498 } else {
3499 switch(use->Opcode()) {
3500 case Op_AddP:
3501 case Op_DecodeN:
3502 case Op_DecodeNKlass:
3503 case Op_CheckCastPP:
3504 case Op_CastPP:
3505 wq.push(use);
3506 break;
3507 }
3508 }
3509 }
3510 }
3511 }
3512 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3513 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3514 Node* in1 = n->in(1);
3515 const Type* t = n->bottom_type();
3516 Node* new_in1 = in1->clone();
3517 new_in1->as_DecodeN()->set_type(t);
3518
3519 if (!Matcher::narrow_oop_use_complex_address()) {
3520 //
3521 // x86, ARM and friends can handle 2 adds in addressing mode
3522 // and Matcher can fold a DecodeN node into address by using
3523 // a narrow oop directly and do implicit null check in address:
3524 //
3525 // [R12 + narrow_oop_reg<<3 + offset]
3526 // NullCheck narrow_oop_reg
3527 //
3528 // On other platforms (Sparc) we have to keep new DecodeN node and
3529 // use it to do implicit null check in address:
3530 //
3531 // decode_not_null narrow_oop_reg, base_reg
3532 // [base_reg + offset]
3533 // NullCheck base_reg
3534 //
3535 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3536 // to keep the information to which null check the new DecodeN node
3537 // corresponds to use it as value in implicit_null_check().
3538 //
3539 new_in1->set_req(0, n->in(0));
3540 }
3541
3542 n->subsume_by(new_in1, this);
3543 if (in1->outcnt() == 0) {
3544 in1->disconnect_inputs(this);
3545 }
3546 } else {
3547 n->subsume_by(n->in(1), this);
3548 if (n->outcnt() == 0) {
3549 n->disconnect_inputs(this);
3550 }
3551 }
3552 break;
3553 }
3554 case Op_CastII: {
3555 n->as_CastII()->remove_range_check_cast(this);
3556 break;
3557 }
3558 #ifdef _LP64
3559 case Op_CmpP:
3560 // Do this transformation here to preserve CmpPNode::sub() and
3561 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3562 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3563 Node* in1 = n->in(1);
3564 Node* in2 = n->in(2);
3565 if (!in1->is_DecodeNarrowPtr()) {
3566 in2 = in1;
3567 in1 = n->in(2);
3568 }
3569 assert(in1->is_DecodeNarrowPtr(), "sanity");
3570
3571 Node* new_in2 = nullptr;
3572 if (in2->is_DecodeNarrowPtr()) {
3573 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3574 new_in2 = in2->in(1);
3575 } else if (in2->Opcode() == Op_ConP) {
3576 const Type* t = in2->bottom_type();
3577 if (t == TypePtr::NULL_PTR) {
3578 assert(in1->is_DecodeN(), "compare klass to null?");
3579 // Don't convert CmpP null check into CmpN if compressed
3580 // oops implicit null check is not generated.
3581 // This will allow to generate normal oop implicit null check.
3582 if (Matcher::gen_narrow_oop_implicit_null_checks())
3583 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3584 //
3585 // This transformation together with CastPP transformation above
3586 // will generated code for implicit null checks for compressed oops.
3587 //
3588 // The original code after Optimize()
3589 //
3590 // LoadN memory, narrow_oop_reg
3591 // decode narrow_oop_reg, base_reg
3592 // CmpP base_reg, nullptr
3593 // CastPP base_reg // NotNull
3594 // Load [base_reg + offset], val_reg
3595 //
3596 // after these transformations will be
3597 //
3598 // LoadN memory, narrow_oop_reg
3599 // CmpN narrow_oop_reg, nullptr
3600 // decode_not_null narrow_oop_reg, base_reg
3601 // Load [base_reg + offset], val_reg
3602 //
3603 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3604 // since narrow oops can be used in debug info now (see the code in
3605 // final_graph_reshaping_walk()).
3606 //
3607 // At the end the code will be matched to
3608 // on x86:
3609 //
3610 // Load_narrow_oop memory, narrow_oop_reg
3611 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3612 // NullCheck narrow_oop_reg
3613 //
3614 // and on sparc:
3615 //
3616 // Load_narrow_oop memory, narrow_oop_reg
3617 // decode_not_null narrow_oop_reg, base_reg
3618 // Load [base_reg + offset], val_reg
3619 // NullCheck base_reg
3620 //
3621 } else if (t->isa_oopptr()) {
3622 new_in2 = ConNode::make(t->make_narrowoop());
3623 } else if (t->isa_klassptr()) {
3624 ciKlass* klass = t->is_klassptr()->exact_klass();
3625 if (klass->is_in_encoding_range()) {
3626 new_in2 = ConNode::make(t->make_narrowklass());
3627 }
3628 }
3629 }
3630 if (new_in2 != nullptr) {
3631 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3632 n->subsume_by(cmpN, this);
3633 if (in1->outcnt() == 0) {
3634 in1->disconnect_inputs(this);
3635 }
3636 if (in2->outcnt() == 0) {
3637 in2->disconnect_inputs(this);
3638 }
3639 }
3640 }
3641 break;
3642
3643 case Op_DecodeN:
3644 case Op_DecodeNKlass:
3645 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3646 // DecodeN could be pinned when it can't be fold into
3647 // an address expression, see the code for Op_CastPP above.
3648 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3649 break;
3650
3651 case Op_EncodeP:
3652 case Op_EncodePKlass: {
3653 Node* in1 = n->in(1);
3654 if (in1->is_DecodeNarrowPtr()) {
3655 n->subsume_by(in1->in(1), this);
3656 } else if (in1->Opcode() == Op_ConP) {
3657 const Type* t = in1->bottom_type();
3658 if (t == TypePtr::NULL_PTR) {
3659 assert(t->isa_oopptr(), "null klass?");
3660 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3661 } else if (t->isa_oopptr()) {
3662 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3663 } else if (t->isa_klassptr()) {
3664 ciKlass* klass = t->is_klassptr()->exact_klass();
3665 if (klass->is_in_encoding_range()) {
3666 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3667 } else {
3668 assert(false, "unencodable klass in ConP -> EncodeP");
3669 C->record_failure("unencodable klass in ConP -> EncodeP");
3670 }
3671 }
3672 }
3673 if (in1->outcnt() == 0) {
3674 in1->disconnect_inputs(this);
3675 }
3676 break;
3677 }
3678
3679 case Op_Proj: {
3680 if (OptimizeStringConcat || IncrementalInline) {
3681 ProjNode* proj = n->as_Proj();
3682 if (proj->_is_io_use) {
3683 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3684 // Separate projections were used for the exception path which
3685 // are normally removed by a late inline. If it wasn't inlined
3686 // then they will hang around and should just be replaced with
3687 // the original one. Merge them.
3688 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3689 if (non_io_proj != nullptr) {
3690 proj->subsume_by(non_io_proj , this);
3691 }
3692 }
3693 }
3694 break;
3695 }
3696
3697 case Op_Phi:
3698 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3699 // The EncodeP optimization may create Phi with the same edges
3700 // for all paths. It is not handled well by Register Allocator.
3701 Node* unique_in = n->in(1);
3702 assert(unique_in != nullptr, "");
3703 uint cnt = n->req();
3704 for (uint i = 2; i < cnt; i++) {
3705 Node* m = n->in(i);
3706 assert(m != nullptr, "");
3707 if (unique_in != m)
3708 unique_in = nullptr;
3709 }
3710 if (unique_in != nullptr) {
3711 n->subsume_by(unique_in, this);
3712 }
3713 }
3714 break;
3715
3716 #endif
3717
3718 case Op_ModI:
3719 handle_div_mod_op(n, T_INT, false);
3720 break;
3721
3722 case Op_ModL:
3723 handle_div_mod_op(n, T_LONG, false);
3724 break;
3725
3726 case Op_UModI:
3727 handle_div_mod_op(n, T_INT, true);
3728 break;
3729
3730 case Op_UModL:
3731 handle_div_mod_op(n, T_LONG, true);
3732 break;
3733
3734 case Op_LoadVector:
3735 case Op_StoreVector:
3736 #ifdef ASSERT
3737 // Add VerifyVectorAlignment node between adr and load / store.
3738 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3739 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3740 n->as_StoreVector()->must_verify_alignment();
3741 if (must_verify_alignment) {
3742 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3743 n->as_StoreVector()->memory_size();
3744 // The memory access should be aligned to the vector width in bytes.
3745 // However, the underlying array is possibly less well aligned, but at least
3746 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3747 // a loop we can expect at least the following alignment:
3748 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3749 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3750 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3751 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3752 jlong mask = guaranteed_alignment - 1;
3753 Node* mask_con = ConLNode::make(mask);
3754 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3755 n->set_req(MemNode::Address, va);
3756 }
3757 }
3758 #endif
3759 break;
3760
3761 case Op_LoadVectorGather:
3762 case Op_StoreVectorScatter:
3763 case Op_LoadVectorGatherMasked:
3764 case Op_StoreVectorScatterMasked:
3765 case Op_VectorCmpMasked:
3766 case Op_VectorMaskGen:
3767 case Op_LoadVectorMasked:
3768 case Op_StoreVectorMasked:
3769 break;
3770
3771 case Op_AddReductionVI:
3772 case Op_AddReductionVL:
3773 case Op_AddReductionVF:
3774 case Op_AddReductionVD:
3775 case Op_MulReductionVI:
3776 case Op_MulReductionVL:
3777 case Op_MulReductionVF:
3778 case Op_MulReductionVD:
3779 case Op_MinReductionV:
3780 case Op_MaxReductionV:
3781 case Op_AndReductionV:
3782 case Op_OrReductionV:
3783 case Op_XorReductionV:
3784 break;
3785
3786 case Op_PackB:
3787 case Op_PackS:
3788 case Op_PackI:
3789 case Op_PackF:
3790 case Op_PackL:
3791 case Op_PackD:
3792 if (n->req()-1 > 2) {
3793 // Replace many operand PackNodes with a binary tree for matching
3794 PackNode* p = (PackNode*) n;
3795 Node* btp = p->binary_tree_pack(1, n->req());
3796 n->subsume_by(btp, this);
3797 }
3798 break;
3799 case Op_Loop:
3800 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3801 case Op_CountedLoop:
3802 case Op_LongCountedLoop:
3803 case Op_OuterStripMinedLoop:
3804 if (n->as_Loop()->is_inner_loop()) {
3805 frc.inc_inner_loop_count();
3806 }
3807 n->as_Loop()->verify_strip_mined(0);
3808 break;
3809 case Op_LShiftI:
3810 case Op_RShiftI:
3811 case Op_URShiftI:
3812 case Op_LShiftL:
3813 case Op_RShiftL:
3814 case Op_URShiftL:
3815 if (Matcher::need_masked_shift_count) {
3816 // The cpu's shift instructions don't restrict the count to the
3817 // lower 5/6 bits. We need to do the masking ourselves.
3818 Node* in2 = n->in(2);
3819 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3820 const TypeInt* t = in2->find_int_type();
3821 if (t != nullptr && t->is_con()) {
3822 juint shift = t->get_con();
3823 if (shift > mask) { // Unsigned cmp
3824 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3825 }
3826 } else {
3827 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3828 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3829 n->set_req(2, shift);
3830 }
3831 }
3832 if (in2->outcnt() == 0) { // Remove dead node
3833 in2->disconnect_inputs(this);
3834 }
3835 }
3836 break;
3837 case Op_MemBarStoreStore:
3838 case Op_MemBarRelease:
3839 // Break the link with AllocateNode: it is no longer useful and
3840 // confuses register allocation.
3841 if (n->req() > MemBarNode::Precedent) {
3842 n->set_req(MemBarNode::Precedent, top());
3843 }
3844 break;
3845 case Op_MemBarAcquire: {
3846 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3847 // At parse time, the trailing MemBarAcquire for a volatile load
3848 // is created with an edge to the load. After optimizations,
3849 // that input may be a chain of Phis. If those phis have no
3850 // other use, then the MemBarAcquire keeps them alive and
3851 // register allocation can be confused.
3852 dead_nodes.push(n->in(MemBarNode::Precedent));
3853 n->set_req(MemBarNode::Precedent, top());
3854 }
3855 break;
3856 }
3857 case Op_Blackhole:
3858 break;
3859 case Op_RangeCheck: {
3860 RangeCheckNode* rc = n->as_RangeCheck();
3861 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3862 n->subsume_by(iff, this);
3863 frc._tests.push(iff);
3864 break;
3865 }
3866 case Op_ConvI2L: {
3867 if (!Matcher::convi2l_type_required) {
3868 // Code generation on some platforms doesn't need accurate
3869 // ConvI2L types. Widening the type can help remove redundant
3870 // address computations.
3871 n->as_Type()->set_type(TypeLong::INT);
3872 ResourceMark rm;
3873 Unique_Node_List wq;
3874 wq.push(n);
3875 for (uint next = 0; next < wq.size(); next++) {
3876 Node *m = wq.at(next);
3877
3878 for(;;) {
3879 // Loop over all nodes with identical inputs edges as m
3880 Node* k = m->find_similar(m->Opcode());
3881 if (k == nullptr) {
3882 break;
3883 }
3884 // Push their uses so we get a chance to remove node made
3885 // redundant
3886 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3887 Node* u = k->fast_out(i);
3888 if (u->Opcode() == Op_LShiftL ||
3889 u->Opcode() == Op_AddL ||
3890 u->Opcode() == Op_SubL ||
3891 u->Opcode() == Op_AddP) {
3892 wq.push(u);
3893 }
3894 }
3895 // Replace all nodes with identical edges as m with m
3896 k->subsume_by(m, this);
3897 }
3898 }
3899 }
3900 break;
3901 }
3902 case Op_CmpUL: {
3903 if (!Matcher::has_match_rule(Op_CmpUL)) {
3904 // No support for unsigned long comparisons
3905 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3906 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3907 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3908 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3909 Node* andl = new AndLNode(orl, remove_sign_mask);
3910 Node* cmp = new CmpLNode(andl, n->in(2));
3911 n->subsume_by(cmp, this);
3912 }
3913 break;
3914 }
3915 #ifdef ASSERT
3916 case Op_ConNKlass: {
3917 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3918 ciKlass* klass = tp->is_klassptr()->exact_klass();
3919 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3920 break;
3921 }
3922 #endif
3923 default:
3924 assert(!n->is_Call(), "");
3925 assert(!n->is_Mem(), "");
3926 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3927 break;
3928 }
3929 }
3930
3931 //------------------------------final_graph_reshaping_walk---------------------
3932 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3933 // requires that the walk visits a node's inputs before visiting the node.
3934 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3935 Unique_Node_List sfpt;
3936
3937 frc._visited.set(root->_idx); // first, mark node as visited
3938 uint cnt = root->req();
3939 Node *n = root;
3940 uint i = 0;
3941 while (true) {
3942 if (i < cnt) {
3943 // Place all non-visited non-null inputs onto stack
3944 Node* m = n->in(i);
3945 ++i;
3946 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3947 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3948 // compute worst case interpreter size in case of a deoptimization
3949 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3950
3951 sfpt.push(m);
3952 }
3953 cnt = m->req();
3954 nstack.push(n, i); // put on stack parent and next input's index
3955 n = m;
3956 i = 0;
3957 }
3958 } else {
3959 // Now do post-visit work
3960 final_graph_reshaping_impl(n, frc, dead_nodes);
3961 if (nstack.is_empty())
3962 break; // finished
3963 n = nstack.node(); // Get node from stack
3964 cnt = n->req();
3965 i = nstack.index();
3966 nstack.pop(); // Shift to the next node on stack
3967 }
3968 }
3969
3970 // Skip next transformation if compressed oops are not used.
3971 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3972 (!UseCompressedOops && !UseCompressedClassPointers))
3973 return;
3974
3975 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3976 // It could be done for an uncommon traps or any safepoints/calls
3977 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3978 while (sfpt.size() > 0) {
3979 n = sfpt.pop();
3980 JVMState *jvms = n->as_SafePoint()->jvms();
3981 assert(jvms != nullptr, "sanity");
3982 int start = jvms->debug_start();
3983 int end = n->req();
3984 bool is_uncommon = (n->is_CallStaticJava() &&
3985 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3986 for (int j = start; j < end; j++) {
3987 Node* in = n->in(j);
3988 if (in->is_DecodeNarrowPtr()) {
3989 bool safe_to_skip = true;
3990 if (!is_uncommon ) {
3991 // Is it safe to skip?
3992 for (uint i = 0; i < in->outcnt(); i++) {
3993 Node* u = in->raw_out(i);
3994 if (!u->is_SafePoint() ||
3995 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3996 safe_to_skip = false;
3997 }
3998 }
3999 }
4000 if (safe_to_skip) {
4001 n->set_req(j, in->in(1));
4002 }
4003 if (in->outcnt() == 0) {
4004 in->disconnect_inputs(this);
4005 }
4006 }
4007 }
4008 }
4009 }
4010
4011 //------------------------------final_graph_reshaping--------------------------
4012 // Final Graph Reshaping.
4013 //
4014 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4015 // and not commoned up and forced early. Must come after regular
4016 // optimizations to avoid GVN undoing the cloning. Clone constant
4017 // inputs to Loop Phis; these will be split by the allocator anyways.
4018 // Remove Opaque nodes.
4019 // (2) Move last-uses by commutative operations to the left input to encourage
4020 // Intel update-in-place two-address operations and better register usage
4021 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4022 // calls canonicalizing them back.
4023 // (3) Count the number of double-precision FP ops, single-precision FP ops
4024 // and call sites. On Intel, we can get correct rounding either by
4025 // forcing singles to memory (requires extra stores and loads after each
4026 // FP bytecode) or we can set a rounding mode bit (requires setting and
4027 // clearing the mode bit around call sites). The mode bit is only used
4028 // if the relative frequency of single FP ops to calls is low enough.
4029 // This is a key transform for SPEC mpeg_audio.
4030 // (4) Detect infinite loops; blobs of code reachable from above but not
4031 // below. Several of the Code_Gen algorithms fail on such code shapes,
4032 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4033 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4034 // Detection is by looking for IfNodes where only 1 projection is
4035 // reachable from below or CatchNodes missing some targets.
4036 // (5) Assert for insane oop offsets in debug mode.
4037
4038 bool Compile::final_graph_reshaping() {
4039 // an infinite loop may have been eliminated by the optimizer,
4040 // in which case the graph will be empty.
4041 if (root()->req() == 1) {
4042 // Do not compile method that is only a trivial infinite loop,
4043 // since the content of the loop may have been eliminated.
4044 record_method_not_compilable("trivial infinite loop");
4045 return true;
4046 }
4047
4048 // Expensive nodes have their control input set to prevent the GVN
4049 // from freely commoning them. There's no GVN beyond this point so
4050 // no need to keep the control input. We want the expensive nodes to
4051 // be freely moved to the least frequent code path by gcm.
4052 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4053 for (int i = 0; i < expensive_count(); i++) {
4054 _expensive_nodes.at(i)->set_req(0, nullptr);
4055 }
4056
4057 Final_Reshape_Counts frc;
4058
4059 // Visit everybody reachable!
4060 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4061 Node_Stack nstack(live_nodes() >> 1);
4062 Unique_Node_List dead_nodes;
4063 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4064
4065 // Check for unreachable (from below) code (i.e., infinite loops).
4066 for( uint i = 0; i < frc._tests.size(); i++ ) {
4067 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4068 // Get number of CFG targets.
4069 // Note that PCTables include exception targets after calls.
4070 uint required_outcnt = n->required_outcnt();
4071 if (n->outcnt() != required_outcnt) {
4072 // Check for a few special cases. Rethrow Nodes never take the
4073 // 'fall-thru' path, so expected kids is 1 less.
4074 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4075 if (n->in(0)->in(0)->is_Call()) {
4076 CallNode* call = n->in(0)->in(0)->as_Call();
4077 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4078 required_outcnt--; // Rethrow always has 1 less kid
4079 } else if (call->req() > TypeFunc::Parms &&
4080 call->is_CallDynamicJava()) {
4081 // Check for null receiver. In such case, the optimizer has
4082 // detected that the virtual call will always result in a null
4083 // pointer exception. The fall-through projection of this CatchNode
4084 // will not be populated.
4085 Node* arg0 = call->in(TypeFunc::Parms);
4086 if (arg0->is_Type() &&
4087 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4088 required_outcnt--;
4089 }
4090 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4091 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4092 // Check for illegal array length. In such case, the optimizer has
4093 // detected that the allocation attempt will always result in an
4094 // exception. There is no fall-through projection of this CatchNode .
4095 assert(call->is_CallStaticJava(), "static call expected");
4096 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4097 uint valid_length_test_input = call->req() - 1;
4098 Node* valid_length_test = call->in(valid_length_test_input);
4099 call->del_req(valid_length_test_input);
4100 if (valid_length_test->find_int_con(1) == 0) {
4101 required_outcnt--;
4102 }
4103 dead_nodes.push(valid_length_test);
4104 assert(n->outcnt() == required_outcnt, "malformed control flow");
4105 continue;
4106 }
4107 }
4108 }
4109
4110 // Recheck with a better notion of 'required_outcnt'
4111 if (n->outcnt() != required_outcnt) {
4112 record_method_not_compilable("malformed control flow");
4113 return true; // Not all targets reachable!
4114 }
4115 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4116 CallNode* call = n->in(0)->in(0)->as_Call();
4117 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4118 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4119 assert(call->is_CallStaticJava(), "static call expected");
4120 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4121 uint valid_length_test_input = call->req() - 1;
4122 dead_nodes.push(call->in(valid_length_test_input));
4123 call->del_req(valid_length_test_input); // valid length test useless now
4124 }
4125 }
4126 // Check that I actually visited all kids. Unreached kids
4127 // must be infinite loops.
4128 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4129 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4130 record_method_not_compilable("infinite loop");
4131 return true; // Found unvisited kid; must be unreach
4132 }
4133
4134 // Here so verification code in final_graph_reshaping_walk()
4135 // always see an OuterStripMinedLoopEnd
4136 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4137 IfNode* init_iff = n->as_If();
4138 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4139 n->subsume_by(iff, this);
4140 }
4141 }
4142
4143 while (dead_nodes.size() > 0) {
4144 Node* m = dead_nodes.pop();
4145 if (m->outcnt() == 0 && m != top()) {
4146 for (uint j = 0; j < m->req(); j++) {
4147 Node* in = m->in(j);
4148 if (in != nullptr) {
4149 dead_nodes.push(in);
4150 }
4151 }
4152 m->disconnect_inputs(this);
4153 }
4154 }
4155
4156 set_java_calls(frc.get_java_call_count());
4157 set_inner_loops(frc.get_inner_loop_count());
4158
4159 // No infinite loops, no reason to bail out.
4160 return false;
4161 }
4162
4163 //-----------------------------too_many_traps----------------------------------
4164 // Report if there are too many traps at the current method and bci.
4165 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4166 bool Compile::too_many_traps(ciMethod* method,
4167 int bci,
4168 Deoptimization::DeoptReason reason) {
4169 ciMethodData* md = method->method_data();
4170 if (md->is_empty()) {
4171 // Assume the trap has not occurred, or that it occurred only
4172 // because of a transient condition during start-up in the interpreter.
4173 return false;
4174 }
4175 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4176 if (md->has_trap_at(bci, m, reason) != 0) {
4177 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4178 // Also, if there are multiple reasons, or if there is no per-BCI record,
4179 // assume the worst.
4180 if (log())
4181 log()->elem("observe trap='%s' count='%d'",
4182 Deoptimization::trap_reason_name(reason),
4183 md->trap_count(reason));
4184 return true;
4185 } else {
4186 // Ignore method/bci and see if there have been too many globally.
4187 return too_many_traps(reason, md);
4188 }
4189 }
4190
4191 // Less-accurate variant which does not require a method and bci.
4192 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4193 ciMethodData* logmd) {
4194 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4195 // Too many traps globally.
4196 // Note that we use cumulative trap_count, not just md->trap_count.
4197 if (log()) {
4198 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4199 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4200 Deoptimization::trap_reason_name(reason),
4201 mcount, trap_count(reason));
4202 }
4203 return true;
4204 } else {
4205 // The coast is clear.
4206 return false;
4207 }
4208 }
4209
4210 //--------------------------too_many_recompiles--------------------------------
4211 // Report if there are too many recompiles at the current method and bci.
4212 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4213 // Is not eager to return true, since this will cause the compiler to use
4214 // Action_none for a trap point, to avoid too many recompilations.
4215 bool Compile::too_many_recompiles(ciMethod* method,
4216 int bci,
4217 Deoptimization::DeoptReason reason) {
4218 ciMethodData* md = method->method_data();
4219 if (md->is_empty()) {
4220 // Assume the trap has not occurred, or that it occurred only
4221 // because of a transient condition during start-up in the interpreter.
4222 return false;
4223 }
4224 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4225 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4226 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4227 Deoptimization::DeoptReason per_bc_reason
4228 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4229 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4230 if ((per_bc_reason == Deoptimization::Reason_none
4231 || md->has_trap_at(bci, m, reason) != 0)
4232 // The trap frequency measure we care about is the recompile count:
4233 && md->trap_recompiled_at(bci, m)
4234 && md->overflow_recompile_count() >= bc_cutoff) {
4235 // Do not emit a trap here if it has already caused recompilations.
4236 // Also, if there are multiple reasons, or if there is no per-BCI record,
4237 // assume the worst.
4238 if (log())
4239 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4240 Deoptimization::trap_reason_name(reason),
4241 md->trap_count(reason),
4242 md->overflow_recompile_count());
4243 return true;
4244 } else if (trap_count(reason) != 0
4245 && decompile_count() >= m_cutoff) {
4246 // Too many recompiles globally, and we have seen this sort of trap.
4247 // Use cumulative decompile_count, not just md->decompile_count.
4248 if (log())
4249 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4250 Deoptimization::trap_reason_name(reason),
4251 md->trap_count(reason), trap_count(reason),
4252 md->decompile_count(), decompile_count());
4253 return true;
4254 } else {
4255 // The coast is clear.
4256 return false;
4257 }
4258 }
4259
4260 // Compute when not to trap. Used by matching trap based nodes and
4261 // NullCheck optimization.
4262 void Compile::set_allowed_deopt_reasons() {
4263 _allowed_reasons = 0;
4264 if (is_method_compilation()) {
4265 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4266 assert(rs < BitsPerInt, "recode bit map");
4267 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4268 _allowed_reasons |= nth_bit(rs);
4269 }
4270 }
4271 }
4272 }
4273
4274 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4275 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4276 }
4277
4278 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4279 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4280 }
4281
4282 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4283 if (holder->is_initialized()) {
4284 return false;
4285 }
4286 if (holder->is_being_initialized()) {
4287 if (accessing_method->holder() == holder) {
4288 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4289 // <init>, or a static method. In all those cases, there was an initialization
4290 // barrier on the holder klass passed.
4291 if (accessing_method->is_static_initializer() ||
4292 accessing_method->is_object_initializer() ||
4293 accessing_method->is_static()) {
4294 return false;
4295 }
4296 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4297 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4298 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4299 // child class can become fully initialized while its parent class is still being initialized.
4300 if (accessing_method->is_static_initializer()) {
4301 return false;
4302 }
4303 }
4304 ciMethod* root = method(); // the root method of compilation
4305 if (root != accessing_method) {
4306 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4307 }
4308 }
4309 return true;
4310 }
4311
4312 #ifndef PRODUCT
4313 //------------------------------verify_bidirectional_edges---------------------
4314 // For each input edge to a node (ie - for each Use-Def edge), verify that
4315 // there is a corresponding Def-Use edge.
4316 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4317 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4318 uint stack_size = live_nodes() >> 4;
4319 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4320 if (root_and_safepoints != nullptr) {
4321 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4322 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4323 Node* root_or_safepoint = root_and_safepoints->at(i);
4324 // If the node is a safepoint, let's check if it still has a control input
4325 // Lack of control input signifies that this node was killed by CCP or
4326 // recursively by remove_globally_dead_node and it shouldn't be a starting
4327 // point.
4328 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4329 nstack.push(root_or_safepoint);
4330 }
4331 }
4332 } else {
4333 nstack.push(_root);
4334 }
4335
4336 while (nstack.size() > 0) {
4337 Node* n = nstack.pop();
4338 if (visited.member(n)) {
4339 continue;
4340 }
4341 visited.push(n);
4342
4343 // Walk over all input edges, checking for correspondence
4344 uint length = n->len();
4345 for (uint i = 0; i < length; i++) {
4346 Node* in = n->in(i);
4347 if (in != nullptr && !visited.member(in)) {
4348 nstack.push(in); // Put it on stack
4349 }
4350 if (in != nullptr && !in->is_top()) {
4351 // Count instances of `next`
4352 int cnt = 0;
4353 for (uint idx = 0; idx < in->_outcnt; idx++) {
4354 if (in->_out[idx] == n) {
4355 cnt++;
4356 }
4357 }
4358 assert(cnt > 0, "Failed to find Def-Use edge.");
4359 // Check for duplicate edges
4360 // walk the input array downcounting the input edges to n
4361 for (uint j = 0; j < length; j++) {
4362 if (n->in(j) == in) {
4363 cnt--;
4364 }
4365 }
4366 assert(cnt == 0, "Mismatched edge count.");
4367 } else if (in == nullptr) {
4368 assert(i == 0 || i >= n->req() ||
4369 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4370 (n->is_Unlock() && i == (n->req() - 1)) ||
4371 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4372 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4373 } else {
4374 assert(in->is_top(), "sanity");
4375 // Nothing to check.
4376 }
4377 }
4378 }
4379 }
4380
4381 //------------------------------verify_graph_edges---------------------------
4382 // Walk the Graph and verify that there is a one-to-one correspondence
4383 // between Use-Def edges and Def-Use edges in the graph.
4384 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4385 if (VerifyGraphEdges) {
4386 Unique_Node_List visited;
4387
4388 // Call graph walk to check edges
4389 verify_bidirectional_edges(visited, root_and_safepoints);
4390 if (no_dead_code) {
4391 // Now make sure that no visited node is used by an unvisited node.
4392 bool dead_nodes = false;
4393 Unique_Node_List checked;
4394 while (visited.size() > 0) {
4395 Node* n = visited.pop();
4396 checked.push(n);
4397 for (uint i = 0; i < n->outcnt(); i++) {
4398 Node* use = n->raw_out(i);
4399 if (checked.member(use)) continue; // already checked
4400 if (visited.member(use)) continue; // already in the graph
4401 if (use->is_Con()) continue; // a dead ConNode is OK
4402 // At this point, we have found a dead node which is DU-reachable.
4403 if (!dead_nodes) {
4404 tty->print_cr("*** Dead nodes reachable via DU edges:");
4405 dead_nodes = true;
4406 }
4407 use->dump(2);
4408 tty->print_cr("---");
4409 checked.push(use); // No repeats; pretend it is now checked.
4410 }
4411 }
4412 assert(!dead_nodes, "using nodes must be reachable from root");
4413 }
4414 }
4415 }
4416 #endif
4417
4418 // The Compile object keeps track of failure reasons separately from the ciEnv.
4419 // This is required because there is not quite a 1-1 relation between the
4420 // ciEnv and its compilation task and the Compile object. Note that one
4421 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4422 // to backtrack and retry without subsuming loads. Other than this backtracking
4423 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4424 // by the logic in C2Compiler.
4425 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4426 if (log() != nullptr) {
4427 log()->elem("failure reason='%s' phase='compile'", reason);
4428 }
4429 if (_failure_reason.get() == nullptr) {
4430 // Record the first failure reason.
4431 _failure_reason.set(reason);
4432 if (CaptureBailoutInformation) {
4433 _first_failure_details = new CompilationFailureInfo(reason);
4434 }
4435 } else {
4436 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4437 }
4438
4439 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4440 C->print_method(PHASE_FAILURE, 1);
4441 }
4442 _root = nullptr; // flush the graph, too
4443 }
4444
4445 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4446 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4447 _compile(Compile::current()),
4448 _log(nullptr),
4449 _dolog(CITimeVerbose)
4450 {
4451 assert(_compile != nullptr, "sanity check");
4452 assert(id != PhaseTraceId::_t_none, "Don't use none");
4453 if (_dolog) {
4454 _log = _compile->log();
4455 }
4456 if (_log != nullptr) {
4457 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4458 _log->stamp();
4459 _log->end_head();
4460 }
4461
4462 // Inform memory statistic, if enabled
4463 if (CompilationMemoryStatistic::enabled()) {
4464 CompilationMemoryStatistic::on_phase_start((int)id, name);
4465 }
4466 }
4467
4468 Compile::TracePhase::TracePhase(PhaseTraceId id)
4469 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4470
4471 Compile::TracePhase::~TracePhase() {
4472
4473 // Inform memory statistic, if enabled
4474 if (CompilationMemoryStatistic::enabled()) {
4475 CompilationMemoryStatistic::on_phase_end();
4476 }
4477
4478 if (_compile->failing_internal()) {
4479 if (_log != nullptr) {
4480 _log->done("phase");
4481 }
4482 return; // timing code, not stressing bailouts.
4483 }
4484 #ifdef ASSERT
4485 if (PrintIdealNodeCount) {
4486 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4487 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4488 }
4489
4490 if (VerifyIdealNodeCount) {
4491 _compile->print_missing_nodes();
4492 }
4493 #endif
4494
4495 if (_log != nullptr) {
4496 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4497 }
4498 }
4499
4500 //----------------------------static_subtype_check-----------------------------
4501 // Shortcut important common cases when superklass is exact:
4502 // (0) superklass is java.lang.Object (can occur in reflective code)
4503 // (1) subklass is already limited to a subtype of superklass => always ok
4504 // (2) subklass does not overlap with superklass => always fail
4505 // (3) superklass has NO subtypes and we can check with a simple compare.
4506 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4507 if (skip) {
4508 return SSC_full_test; // Let caller generate the general case.
4509 }
4510
4511 if (subk->is_java_subtype_of(superk)) {
4512 return SSC_always_true; // (0) and (1) this test cannot fail
4513 }
4514
4515 if (!subk->maybe_java_subtype_of(superk)) {
4516 return SSC_always_false; // (2) true path dead; no dynamic test needed
4517 }
4518
4519 const Type* superelem = superk;
4520 if (superk->isa_aryklassptr()) {
4521 int ignored;
4522 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4523 }
4524
4525 if (superelem->isa_instklassptr()) {
4526 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4527 if (!ik->has_subklass()) {
4528 if (!ik->is_final()) {
4529 // Add a dependency if there is a chance of a later subclass.
4530 dependencies()->assert_leaf_type(ik);
4531 }
4532 if (!superk->maybe_java_subtype_of(subk)) {
4533 return SSC_always_false;
4534 }
4535 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4536 }
4537 } else {
4538 // A primitive array type has no subtypes.
4539 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4540 }
4541
4542 return SSC_full_test;
4543 }
4544
4545 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4546 #ifdef _LP64
4547 // The scaled index operand to AddP must be a clean 64-bit value.
4548 // Java allows a 32-bit int to be incremented to a negative
4549 // value, which appears in a 64-bit register as a large
4550 // positive number. Using that large positive number as an
4551 // operand in pointer arithmetic has bad consequences.
4552 // On the other hand, 32-bit overflow is rare, and the possibility
4553 // can often be excluded, if we annotate the ConvI2L node with
4554 // a type assertion that its value is known to be a small positive
4555 // number. (The prior range check has ensured this.)
4556 // This assertion is used by ConvI2LNode::Ideal.
4557 int index_max = max_jint - 1; // array size is max_jint, index is one less
4558 if (sizetype != nullptr && sizetype->_hi > 0) {
4559 index_max = sizetype->_hi - 1;
4560 }
4561 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4562 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4563 #endif
4564 return idx;
4565 }
4566
4567 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4568 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4569 if (ctrl != nullptr) {
4570 // Express control dependency by a CastII node with a narrow type.
4571 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4572 // node from floating above the range check during loop optimizations. Otherwise, the
4573 // ConvI2L node may be eliminated independently of the range check, causing the data path
4574 // to become TOP while the control path is still there (although it's unreachable).
4575 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
4576 value = phase->transform(value);
4577 }
4578 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4579 return phase->transform(new ConvI2LNode(value, ltype));
4580 }
4581
4582 void Compile::dump_print_inlining() {
4583 inline_printer()->print_on(tty);
4584 }
4585
4586 void Compile::log_late_inline(CallGenerator* cg) {
4587 if (log() != nullptr) {
4588 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4589 cg->unique_id());
4590 JVMState* p = cg->call_node()->jvms();
4591 while (p != nullptr) {
4592 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4593 p = p->caller();
4594 }
4595 log()->tail("late_inline");
4596 }
4597 }
4598
4599 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4600 log_late_inline(cg);
4601 if (log() != nullptr) {
4602 log()->inline_fail(msg);
4603 }
4604 }
4605
4606 void Compile::log_inline_id(CallGenerator* cg) {
4607 if (log() != nullptr) {
4608 // The LogCompilation tool needs a unique way to identify late
4609 // inline call sites. This id must be unique for this call site in
4610 // this compilation. Try to have it unique across compilations as
4611 // well because it can be convenient when grepping through the log
4612 // file.
4613 // Distinguish OSR compilations from others in case CICountOSR is
4614 // on.
4615 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4616 cg->set_unique_id(id);
4617 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4618 }
4619 }
4620
4621 void Compile::log_inline_failure(const char* msg) {
4622 if (C->log() != nullptr) {
4623 C->log()->inline_fail(msg);
4624 }
4625 }
4626
4627
4628 // Dump inlining replay data to the stream.
4629 // Don't change thread state and acquire any locks.
4630 void Compile::dump_inline_data(outputStream* out) {
4631 InlineTree* inl_tree = ilt();
4632 if (inl_tree != nullptr) {
4633 out->print(" inline %d", inl_tree->count());
4634 inl_tree->dump_replay_data(out);
4635 }
4636 }
4637
4638 void Compile::dump_inline_data_reduced(outputStream* out) {
4639 assert(ReplayReduce, "");
4640
4641 InlineTree* inl_tree = ilt();
4642 if (inl_tree == nullptr) {
4643 return;
4644 }
4645 // Enable iterative replay file reduction
4646 // Output "compile" lines for depth 1 subtrees,
4647 // simulating that those trees were compiled
4648 // instead of inlined.
4649 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4650 InlineTree* sub = inl_tree->subtrees().at(i);
4651 if (sub->inline_level() != 1) {
4652 continue;
4653 }
4654
4655 ciMethod* method = sub->method();
4656 int entry_bci = -1;
4657 int comp_level = env()->task()->comp_level();
4658 out->print("compile ");
4659 method->dump_name_as_ascii(out);
4660 out->print(" %d %d", entry_bci, comp_level);
4661 out->print(" inline %d", sub->count());
4662 sub->dump_replay_data(out, -1);
4663 out->cr();
4664 }
4665 }
4666
4667 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4668 if (n1->Opcode() < n2->Opcode()) return -1;
4669 else if (n1->Opcode() > n2->Opcode()) return 1;
4670
4671 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4672 for (uint i = 1; i < n1->req(); i++) {
4673 if (n1->in(i) < n2->in(i)) return -1;
4674 else if (n1->in(i) > n2->in(i)) return 1;
4675 }
4676
4677 return 0;
4678 }
4679
4680 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4681 Node* n1 = *n1p;
4682 Node* n2 = *n2p;
4683
4684 return cmp_expensive_nodes(n1, n2);
4685 }
4686
4687 void Compile::sort_expensive_nodes() {
4688 if (!expensive_nodes_sorted()) {
4689 _expensive_nodes.sort(cmp_expensive_nodes);
4690 }
4691 }
4692
4693 bool Compile::expensive_nodes_sorted() const {
4694 for (int i = 1; i < _expensive_nodes.length(); i++) {
4695 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4696 return false;
4697 }
4698 }
4699 return true;
4700 }
4701
4702 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4703 if (_expensive_nodes.length() == 0) {
4704 return false;
4705 }
4706
4707 assert(OptimizeExpensiveOps, "optimization off?");
4708
4709 // Take this opportunity to remove dead nodes from the list
4710 int j = 0;
4711 for (int i = 0; i < _expensive_nodes.length(); i++) {
4712 Node* n = _expensive_nodes.at(i);
4713 if (!n->is_unreachable(igvn)) {
4714 assert(n->is_expensive(), "should be expensive");
4715 _expensive_nodes.at_put(j, n);
4716 j++;
4717 }
4718 }
4719 _expensive_nodes.trunc_to(j);
4720
4721 // Then sort the list so that similar nodes are next to each other
4722 // and check for at least two nodes of identical kind with same data
4723 // inputs.
4724 sort_expensive_nodes();
4725
4726 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4727 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4728 return true;
4729 }
4730 }
4731
4732 return false;
4733 }
4734
4735 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4736 if (_expensive_nodes.length() == 0) {
4737 return;
4738 }
4739
4740 assert(OptimizeExpensiveOps, "optimization off?");
4741
4742 // Sort to bring similar nodes next to each other and clear the
4743 // control input of nodes for which there's only a single copy.
4744 sort_expensive_nodes();
4745
4746 int j = 0;
4747 int identical = 0;
4748 int i = 0;
4749 bool modified = false;
4750 for (; i < _expensive_nodes.length()-1; i++) {
4751 assert(j <= i, "can't write beyond current index");
4752 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4753 identical++;
4754 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4755 continue;
4756 }
4757 if (identical > 0) {
4758 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4759 identical = 0;
4760 } else {
4761 Node* n = _expensive_nodes.at(i);
4762 igvn.replace_input_of(n, 0, nullptr);
4763 igvn.hash_insert(n);
4764 modified = true;
4765 }
4766 }
4767 if (identical > 0) {
4768 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4769 } else if (_expensive_nodes.length() >= 1) {
4770 Node* n = _expensive_nodes.at(i);
4771 igvn.replace_input_of(n, 0, nullptr);
4772 igvn.hash_insert(n);
4773 modified = true;
4774 }
4775 _expensive_nodes.trunc_to(j);
4776 if (modified) {
4777 igvn.optimize();
4778 }
4779 }
4780
4781 void Compile::add_expensive_node(Node * n) {
4782 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4783 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4784 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4785 if (OptimizeExpensiveOps) {
4786 _expensive_nodes.append(n);
4787 } else {
4788 // Clear control input and let IGVN optimize expensive nodes if
4789 // OptimizeExpensiveOps is off.
4790 n->set_req(0, nullptr);
4791 }
4792 }
4793
4794 /**
4795 * Track coarsened Lock and Unlock nodes.
4796 */
4797
4798 class Lock_List : public Node_List {
4799 uint _origin_cnt;
4800 public:
4801 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4802 uint origin_cnt() const { return _origin_cnt; }
4803 };
4804
4805 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4806 int length = locks.length();
4807 if (length > 0) {
4808 // Have to keep this list until locks elimination during Macro nodes elimination.
4809 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4810 AbstractLockNode* alock = locks.at(0);
4811 BoxLockNode* box = alock->box_node()->as_BoxLock();
4812 for (int i = 0; i < length; i++) {
4813 AbstractLockNode* lock = locks.at(i);
4814 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4815 locks_list->push(lock);
4816 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4817 if (this_box != box) {
4818 // Locking regions (BoxLock) could be Unbalanced here:
4819 // - its coarsened locks were eliminated in earlier
4820 // macro nodes elimination followed by loop unroll
4821 // - it is OSR locking region (no Lock node)
4822 // Preserve Unbalanced status in such cases.
4823 if (!this_box->is_unbalanced()) {
4824 this_box->set_coarsened();
4825 }
4826 if (!box->is_unbalanced()) {
4827 box->set_coarsened();
4828 }
4829 }
4830 }
4831 _coarsened_locks.append(locks_list);
4832 }
4833 }
4834
4835 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4836 int count = coarsened_count();
4837 for (int i = 0; i < count; i++) {
4838 Node_List* locks_list = _coarsened_locks.at(i);
4839 for (uint j = 0; j < locks_list->size(); j++) {
4840 Node* lock = locks_list->at(j);
4841 assert(lock->is_AbstractLock(), "sanity");
4842 if (!useful.member(lock)) {
4843 locks_list->yank(lock);
4844 }
4845 }
4846 }
4847 }
4848
4849 void Compile::remove_coarsened_lock(Node* n) {
4850 if (n->is_AbstractLock()) {
4851 int count = coarsened_count();
4852 for (int i = 0; i < count; i++) {
4853 Node_List* locks_list = _coarsened_locks.at(i);
4854 locks_list->yank(n);
4855 }
4856 }
4857 }
4858
4859 bool Compile::coarsened_locks_consistent() {
4860 int count = coarsened_count();
4861 for (int i = 0; i < count; i++) {
4862 bool unbalanced = false;
4863 bool modified = false; // track locks kind modifications
4864 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4865 uint size = locks_list->size();
4866 if (size == 0) {
4867 unbalanced = false; // All locks were eliminated - good
4868 } else if (size != locks_list->origin_cnt()) {
4869 unbalanced = true; // Some locks were removed from list
4870 } else {
4871 for (uint j = 0; j < size; j++) {
4872 Node* lock = locks_list->at(j);
4873 // All nodes in group should have the same state (modified or not)
4874 if (!lock->as_AbstractLock()->is_coarsened()) {
4875 if (j == 0) {
4876 // first on list was modified, the rest should be too for consistency
4877 modified = true;
4878 } else if (!modified) {
4879 // this lock was modified but previous locks on the list were not
4880 unbalanced = true;
4881 break;
4882 }
4883 } else if (modified) {
4884 // previous locks on list were modified but not this lock
4885 unbalanced = true;
4886 break;
4887 }
4888 }
4889 }
4890 if (unbalanced) {
4891 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4892 #ifdef ASSERT
4893 if (PrintEliminateLocks) {
4894 tty->print_cr("=== unbalanced coarsened locks ===");
4895 for (uint l = 0; l < size; l++) {
4896 locks_list->at(l)->dump();
4897 }
4898 }
4899 #endif
4900 record_failure(C2Compiler::retry_no_locks_coarsening());
4901 return false;
4902 }
4903 }
4904 return true;
4905 }
4906
4907 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4908 // locks coarsening optimization removed Lock/Unlock nodes from them.
4909 // Such regions become unbalanced because coarsening only removes part
4910 // of Lock/Unlock nodes in region. As result we can't execute other
4911 // locks elimination optimizations which assume all code paths have
4912 // corresponding pair of Lock/Unlock nodes - they are balanced.
4913 void Compile::mark_unbalanced_boxes() const {
4914 int count = coarsened_count();
4915 for (int i = 0; i < count; i++) {
4916 Node_List* locks_list = _coarsened_locks.at(i);
4917 uint size = locks_list->size();
4918 if (size > 0) {
4919 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4920 BoxLockNode* box = alock->box_node()->as_BoxLock();
4921 if (alock->is_coarsened()) {
4922 // coarsened_locks_consistent(), which is called before this method, verifies
4923 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4924 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4925 for (uint j = 1; j < size; j++) {
4926 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4927 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4928 if (box != this_box) {
4929 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4930 box->set_unbalanced();
4931 this_box->set_unbalanced();
4932 }
4933 }
4934 }
4935 }
4936 }
4937 }
4938
4939 /**
4940 * Remove the speculative part of types and clean up the graph
4941 */
4942 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4943 if (UseTypeSpeculation) {
4944 Unique_Node_List worklist;
4945 worklist.push(root());
4946 int modified = 0;
4947 // Go over all type nodes that carry a speculative type, drop the
4948 // speculative part of the type and enqueue the node for an igvn
4949 // which may optimize it out.
4950 for (uint next = 0; next < worklist.size(); ++next) {
4951 Node *n = worklist.at(next);
4952 if (n->is_Type()) {
4953 TypeNode* tn = n->as_Type();
4954 const Type* t = tn->type();
4955 const Type* t_no_spec = t->remove_speculative();
4956 if (t_no_spec != t) {
4957 bool in_hash = igvn.hash_delete(n);
4958 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4959 tn->set_type(t_no_spec);
4960 igvn.hash_insert(n);
4961 igvn._worklist.push(n); // give it a chance to go away
4962 modified++;
4963 }
4964 }
4965 // Iterate over outs - endless loops is unreachable from below
4966 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4967 Node *m = n->fast_out(i);
4968 if (not_a_node(m)) {
4969 continue;
4970 }
4971 worklist.push(m);
4972 }
4973 }
4974 // Drop the speculative part of all types in the igvn's type table
4975 igvn.remove_speculative_types();
4976 if (modified > 0) {
4977 igvn.optimize();
4978 if (failing()) return;
4979 }
4980 #ifdef ASSERT
4981 // Verify that after the IGVN is over no speculative type has resurfaced
4982 worklist.clear();
4983 worklist.push(root());
4984 for (uint next = 0; next < worklist.size(); ++next) {
4985 Node *n = worklist.at(next);
4986 const Type* t = igvn.type_or_null(n);
4987 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4988 if (n->is_Type()) {
4989 t = n->as_Type()->type();
4990 assert(t == t->remove_speculative(), "no more speculative types");
4991 }
4992 // Iterate over outs - endless loops is unreachable from below
4993 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4994 Node *m = n->fast_out(i);
4995 if (not_a_node(m)) {
4996 continue;
4997 }
4998 worklist.push(m);
4999 }
5000 }
5001 igvn.check_no_speculative_types();
5002 #endif
5003 }
5004 }
5005
5006 // Auxiliary methods to support randomized stressing/fuzzing.
5007
5008 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5009 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5010 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5011 FLAG_SET_ERGO(StressSeed, _stress_seed);
5012 } else {
5013 _stress_seed = StressSeed;
5014 }
5015 if (_log != nullptr) {
5016 _log->elem("stress_test seed='%u'", _stress_seed);
5017 }
5018 }
5019
5020 int Compile::random() {
5021 _stress_seed = os::next_random(_stress_seed);
5022 return static_cast<int>(_stress_seed);
5023 }
5024
5025 // This method can be called the arbitrary number of times, with current count
5026 // as the argument. The logic allows selecting a single candidate from the
5027 // running list of candidates as follows:
5028 // int count = 0;
5029 // Cand* selected = null;
5030 // while(cand = cand->next()) {
5031 // if (randomized_select(++count)) {
5032 // selected = cand;
5033 // }
5034 // }
5035 //
5036 // Including count equalizes the chances any candidate is "selected".
5037 // This is useful when we don't have the complete list of candidates to choose
5038 // from uniformly. In this case, we need to adjust the randomicity of the
5039 // selection, or else we will end up biasing the selection towards the latter
5040 // candidates.
5041 //
5042 // Quick back-envelope calculation shows that for the list of n candidates
5043 // the equal probability for the candidate to persist as "best" can be
5044 // achieved by replacing it with "next" k-th candidate with the probability
5045 // of 1/k. It can be easily shown that by the end of the run, the
5046 // probability for any candidate is converged to 1/n, thus giving the
5047 // uniform distribution among all the candidates.
5048 //
5049 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5050 #define RANDOMIZED_DOMAIN_POW 29
5051 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5052 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5053 bool Compile::randomized_select(int count) {
5054 assert(count > 0, "only positive");
5055 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5056 }
5057
5058 #ifdef ASSERT
5059 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5060 bool Compile::fail_randomly() {
5061 if ((random() % StressBailoutMean) != 0) {
5062 return false;
5063 }
5064 record_failure("StressBailout");
5065 return true;
5066 }
5067
5068 bool Compile::failure_is_artificial() {
5069 return C->failure_reason_is("StressBailout");
5070 }
5071 #endif
5072
5073 CloneMap& Compile::clone_map() { return _clone_map; }
5074 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5075
5076 void NodeCloneInfo::dump_on(outputStream* st) const {
5077 st->print(" {%d:%d} ", idx(), gen());
5078 }
5079
5080 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5081 uint64_t val = value(old->_idx);
5082 NodeCloneInfo cio(val);
5083 assert(val != 0, "old node should be in the map");
5084 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5085 insert(nnn->_idx, cin.get());
5086 #ifndef PRODUCT
5087 if (is_debug()) {
5088 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5089 }
5090 #endif
5091 }
5092
5093 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5094 NodeCloneInfo cio(value(old->_idx));
5095 if (cio.get() == 0) {
5096 cio.set(old->_idx, 0);
5097 insert(old->_idx, cio.get());
5098 #ifndef PRODUCT
5099 if (is_debug()) {
5100 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5101 }
5102 #endif
5103 }
5104 clone(old, nnn, gen);
5105 }
5106
5107 int CloneMap::max_gen() const {
5108 int g = 0;
5109 DictI di(_dict);
5110 for(; di.test(); ++di) {
5111 int t = gen(di._key);
5112 if (g < t) {
5113 g = t;
5114 #ifndef PRODUCT
5115 if (is_debug()) {
5116 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5117 }
5118 #endif
5119 }
5120 }
5121 return g;
5122 }
5123
5124 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5125 uint64_t val = value(key);
5126 if (val != 0) {
5127 NodeCloneInfo ni(val);
5128 ni.dump_on(st);
5129 }
5130 }
5131
5132 void Compile::shuffle_macro_nodes() {
5133 if (_macro_nodes.length() < 2) {
5134 return;
5135 }
5136 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5137 uint j = C->random() % (i + 1);
5138 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5139 }
5140 }
5141
5142 // Move Allocate nodes to the start of the list
5143 void Compile::sort_macro_nodes() {
5144 int count = macro_count();
5145 int allocates = 0;
5146 for (int i = 0; i < count; i++) {
5147 Node* n = macro_node(i);
5148 if (n->is_Allocate()) {
5149 if (i != allocates) {
5150 Node* tmp = macro_node(allocates);
5151 _macro_nodes.at_put(allocates, n);
5152 _macro_nodes.at_put(i, tmp);
5153 }
5154 allocates++;
5155 }
5156 }
5157 }
5158
5159 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5160 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5161 EventCompilerPhase event(UNTIMED);
5162 if (event.should_commit()) {
5163 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5164 }
5165 #ifndef PRODUCT
5166 ResourceMark rm;
5167 stringStream ss;
5168 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5169 int iter = ++_igv_phase_iter[cpt];
5170 if (iter > 1) {
5171 ss.print(" %d", iter);
5172 }
5173 if (n != nullptr) {
5174 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5175 if (n->is_Call()) {
5176 CallNode* call = n->as_Call();
5177 if (call->_name != nullptr) {
5178 // E.g. uncommon traps etc.
5179 ss.print(" - %s", call->_name);
5180 } else if (call->is_CallJava()) {
5181 CallJavaNode* call_java = call->as_CallJava();
5182 if (call_java->method() != nullptr) {
5183 ss.print(" -");
5184 call_java->method()->print_short_name(&ss);
5185 }
5186 }
5187 }
5188 }
5189
5190 const char* name = ss.as_string();
5191 if (should_print_igv(level)) {
5192 _igv_printer->print_graph(name);
5193 }
5194 if (should_print_phase(level)) {
5195 print_phase(name);
5196 }
5197 if (should_print_ideal_phase(cpt)) {
5198 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5199 }
5200 #endif
5201 C->_latest_stage_start_counter.stamp();
5202 }
5203
5204 // Only used from CompileWrapper
5205 void Compile::begin_method() {
5206 #ifndef PRODUCT
5207 if (_method != nullptr && should_print_igv(1)) {
5208 _igv_printer->begin_method();
5209 }
5210 #endif
5211 C->_latest_stage_start_counter.stamp();
5212 }
5213
5214 // Only used from CompileWrapper
5215 void Compile::end_method() {
5216 EventCompilerPhase event(UNTIMED);
5217 if (event.should_commit()) {
5218 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5219 }
5220
5221 #ifndef PRODUCT
5222 if (_method != nullptr && should_print_igv(1)) {
5223 _igv_printer->end_method();
5224 }
5225 #endif
5226 }
5227
5228 #ifndef PRODUCT
5229 bool Compile::should_print_phase(const int level) const {
5230 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5231 _method != nullptr; // Do not print phases for stubs.
5232 }
5233
5234 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5235 return _directive->should_print_ideal_phase(cpt);
5236 }
5237
5238 void Compile::init_igv() {
5239 if (_igv_printer == nullptr) {
5240 _igv_printer = IdealGraphPrinter::printer();
5241 _igv_printer->set_compile(this);
5242 }
5243 }
5244
5245 bool Compile::should_print_igv(const int level) {
5246 PRODUCT_RETURN_(return false;);
5247
5248 if (PrintIdealGraphLevel < 0) { // disabled by the user
5249 return false;
5250 }
5251
5252 bool need = directive()->IGVPrintLevelOption >= level;
5253 if (need) {
5254 Compile::init_igv();
5255 }
5256 return need;
5257 }
5258
5259 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5260 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5261
5262 // Called from debugger. Prints method to the default file with the default phase name.
5263 // This works regardless of any Ideal Graph Visualizer flags set or not.
5264 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5265 void igv_print(void* sp, void* fp, void* pc) {
5266 frame fr(sp, fp, pc);
5267 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5268 }
5269
5270 // Same as igv_print() above but with a specified phase name.
5271 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5272 frame fr(sp, fp, pc);
5273 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5274 }
5275
5276 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5277 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5278 // This works regardless of any Ideal Graph Visualizer flags set or not.
5279 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5280 void igv_print(bool network, void* sp, void* fp, void* pc) {
5281 frame fr(sp, fp, pc);
5282 if (network) {
5283 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5284 } else {
5285 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5286 }
5287 }
5288
5289 // Same as igv_print(bool network, ...) above but with a specified phase name.
5290 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5291 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5292 frame fr(sp, fp, pc);
5293 if (network) {
5294 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5295 } else {
5296 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5297 }
5298 }
5299
5300 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5301 void igv_print_default() {
5302 Compile::current()->print_method(PHASE_DEBUG, 0);
5303 }
5304
5305 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5306 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5307 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5308 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5309 void igv_append(void* sp, void* fp, void* pc) {
5310 frame fr(sp, fp, pc);
5311 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5312 }
5313
5314 // Same as igv_append(...) above but with a specified phase name.
5315 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5316 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5317 frame fr(sp, fp, pc);
5318 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5319 }
5320
5321 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5322 const char* file_name = "custom_debug.xml";
5323 if (_debug_file_printer == nullptr) {
5324 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5325 } else {
5326 _debug_file_printer->update_compiled_method(C->method());
5327 }
5328 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5329 _debug_file_printer->print_graph(phase_name, fr);
5330 }
5331
5332 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5333 ResourceMark rm;
5334 GrowableArray<const Node*> empty_list;
5335 igv_print_graph_to_network(phase_name, empty_list, fr);
5336 }
5337
5338 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5339 if (_debug_network_printer == nullptr) {
5340 _debug_network_printer = new IdealGraphPrinter(C);
5341 } else {
5342 _debug_network_printer->update_compiled_method(C->method());
5343 }
5344 tty->print_cr("Method printed over network stream to IGV");
5345 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5346 }
5347 #endif // !PRODUCT
5348
5349 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5350 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5351 return value;
5352 }
5353 Node* result = nullptr;
5354 if (bt == T_BYTE) {
5355 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5356 result = new RShiftINode(result, phase->intcon(24));
5357 } else if (bt == T_BOOLEAN) {
5358 result = new AndINode(value, phase->intcon(0xFF));
5359 } else if (bt == T_CHAR) {
5360 result = new AndINode(value,phase->intcon(0xFFFF));
5361 } else {
5362 assert(bt == T_SHORT, "unexpected narrow type");
5363 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5364 result = new RShiftINode(result, phase->intcon(16));
5365 }
5366 if (transform_res) {
5367 result = phase->transform(result);
5368 }
5369 return result;
5370 }
5371
5372 void Compile::record_method_not_compilable_oom() {
5373 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5374 }
5375
5376 #ifndef PRODUCT
5377 // Collects all the control inputs from nodes on the worklist and from their data dependencies
5378 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
5379 // Follow non-control edges until we reach CFG nodes
5380 for (uint i = 0; i < worklist.size(); i++) {
5381 const Node* n = worklist.at(i);
5382 for (uint j = 0; j < n->req(); j++) {
5383 Node* in = n->in(j);
5384 if (in == nullptr || in->is_Root()) {
5385 continue;
5386 }
5387 if (in->is_CFG()) {
5388 if (in->is_Call()) {
5389 // The return value of a call is only available if the call did not result in an exception
5390 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
5391 if (control_proj_use->is_Catch()) {
5392 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
5393 candidates.push(fall_through);
5394 continue;
5395 }
5396 }
5397
5398 if (in->is_Multi()) {
5399 // We got here by following data inputs so we should only have one control use
5400 // (no IfNode, etc)
5401 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
5402 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
5403 } else {
5404 candidates.push(in);
5405 }
5406 } else {
5407 worklist.push(in);
5408 }
5409 }
5410 }
5411 }
5412
5413 // Returns the candidate node that is a descendant to all the other candidates
5414 static Node* pick_control(Unique_Node_List& candidates) {
5415 Unique_Node_List worklist;
5416 worklist.copy(candidates);
5417
5418 // Traverse backwards through the CFG
5419 for (uint i = 0; i < worklist.size(); i++) {
5420 const Node* n = worklist.at(i);
5421 if (n->is_Root()) {
5422 continue;
5423 }
5424 for (uint j = 0; j < n->req(); j++) {
5425 // Skip backedge of loops to avoid cycles
5426 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
5427 continue;
5428 }
5429
5430 Node* pred = n->in(j);
5431 if (pred != nullptr && pred != n && pred->is_CFG()) {
5432 worklist.push(pred);
5433 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
5434 candidates.remove(pred);
5435 }
5436 }
5437 }
5438
5439 assert(candidates.size() == 1, "unexpected control flow");
5440 return candidates.at(0);
5441 }
5442
5443 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
5444 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
5445 call->init_req((*pos)++, parm);
5446 const BasicType bt = parm->bottom_type()->basic_type();
5447 if (bt == T_LONG || bt == T_DOUBLE) {
5448 call->init_req((*pos)++, half);
5449 }
5450 }
5451
5452 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
5453 Node* parm0, Node* parm1,
5454 Node* parm2, Node* parm3,
5455 Node* parm4, Node* parm5,
5456 Node* parm6) const {
5457 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
5458 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
5459 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
5460
5461 // find the most suitable control input
5462 Unique_Node_List worklist, candidates;
5463 if (parm0 != nullptr) { worklist.push(parm0);
5464 if (parm1 != nullptr) { worklist.push(parm1);
5465 if (parm2 != nullptr) { worklist.push(parm2);
5466 if (parm3 != nullptr) { worklist.push(parm3);
5467 if (parm4 != nullptr) { worklist.push(parm4);
5468 if (parm5 != nullptr) { worklist.push(parm5);
5469 if (parm6 != nullptr) { worklist.push(parm6);
5470 /* close each nested if ===> */ } } } } } } }
5471 find_candidate_control_inputs(worklist, candidates);
5472 Node* control = nullptr;
5473 if (candidates.size() == 0) {
5474 control = C->start()->proj_out(TypeFunc::Control);
5475 } else {
5476 control = pick_control(candidates);
5477 }
5478
5479 // find all the previous users of the control we picked
5480 GrowableArray<Node*> users_of_control;
5481 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
5482 Node* use = control->fast_out(i);
5483 if (use->is_CFG() && use != control) {
5484 users_of_control.push(use);
5485 }
5486 }
5487
5488 // we do not actually care about IO and memory as it uses neither
5489 call->init_req(TypeFunc::Control, control);
5490 call->init_req(TypeFunc::I_O, top());
5491 call->init_req(TypeFunc::Memory, top());
5492 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
5493 call->init_req(TypeFunc::ReturnAdr, top());
5494
5495 int pos = TypeFunc::Parms;
5496 call->init_req(pos++, str_node);
5497 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
5498 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
5499 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
5500 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
5501 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
5502 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
5503 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
5504 /* close each nested if ===> */ } } } } } } }
5505 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
5506
5507 call = gvn->transform(call);
5508 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
5509
5510 // rewire previous users to have the new call as control instead
5511 PhaseIterGVN* igvn = gvn->is_IterGVN();
5512 for (int i = 0; i < users_of_control.length(); i++) {
5513 Node* use = users_of_control.at(i);
5514 for (uint j = 0; j < use->req(); j++) {
5515 if (use->in(j) == control) {
5516 if (igvn != nullptr) {
5517 igvn->replace_input_of(use, j, call_control_proj);
5518 } else {
5519 gvn->hash_delete(use);
5520 use->set_req(j, call_control_proj);
5521 gvn->hash_insert(use);
5522 }
5523 }
5524 }
5525 }
5526
5527 return call;
5528 }
5529 #endif // !PRODUCT