1 /*
2 * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSet.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "jfr/jfrEvents.hpp"
39 #include "jvm_io.h"
40 #include "memory/allocation.hpp"
41 #include "memory/resourceArea.hpp"
42 #include "opto/addnode.hpp"
43 #include "opto/block.hpp"
44 #include "opto/c2compiler.hpp"
45 #include "opto/callGenerator.hpp"
46 #include "opto/callnode.hpp"
47 #include "opto/castnode.hpp"
48 #include "opto/cfgnode.hpp"
49 #include "opto/chaitin.hpp"
50 #include "opto/compile.hpp"
51 #include "opto/connode.hpp"
52 #include "opto/convertnode.hpp"
53 #include "opto/divnode.hpp"
54 #include "opto/escape.hpp"
55 #include "opto/idealGraphPrinter.hpp"
56 #include "opto/loopnode.hpp"
57 #include "opto/machnode.hpp"
58 #include "opto/macro.hpp"
59 #include "opto/matcher.hpp"
60 #include "opto/mathexactnode.hpp"
61 #include "opto/memnode.hpp"
62 #include "opto/mulnode.hpp"
63 #include "opto/narrowptrnode.hpp"
64 #include "opto/node.hpp"
65 #include "opto/opcodes.hpp"
66 #include "opto/output.hpp"
67 #include "opto/parse.hpp"
68 #include "opto/phaseX.hpp"
69 #include "opto/rootnode.hpp"
70 #include "opto/runtime.hpp"
71 #include "opto/stringopts.hpp"
72 #include "opto/type.hpp"
73 #include "opto/vector.hpp"
74 #include "opto/vectornode.hpp"
75 #include "runtime/globals_extension.hpp"
76 #include "runtime/sharedRuntime.hpp"
77 #include "runtime/signature.hpp"
78 #include "runtime/stubRoutines.hpp"
79 #include "runtime/timer.hpp"
80 #include "utilities/align.hpp"
81 #include "utilities/copy.hpp"
82 #include "utilities/macros.hpp"
83 #include "utilities/resourceHash.hpp"
84
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == nullptr) {
89 _mach_constant_base_node = new MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
94
95
96 /// Support for intrinsics.
97
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 class IntrinsicDescPair {
101 private:
102 ciMethod* _m;
103 bool _is_virtual;
104 public:
105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
107 ciMethod* m= elt->method();
108 ciMethod* key_m = key->_m;
109 if (key_m < m) return -1;
110 else if (key_m > m) return 1;
111 else {
112 bool is_virtual = elt->is_virtual();
113 bool key_virtual = key->_is_virtual;
114 if (key_virtual < is_virtual) return -1;
115 else if (key_virtual > is_virtual) return 1;
116 else return 0;
117 }
118 }
119 };
120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
121 #ifdef ASSERT
122 for (int i = 1; i < _intrinsics.length(); i++) {
123 CallGenerator* cg1 = _intrinsics.at(i-1);
124 CallGenerator* cg2 = _intrinsics.at(i);
125 assert(cg1->method() != cg2->method()
126 ? cg1->method() < cg2->method()
127 : cg1->is_virtual() < cg2->is_virtual(),
128 "compiler intrinsics list must stay sorted");
129 }
130 #endif
131 IntrinsicDescPair pair(m, is_virtual);
132 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
133 }
134
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 bool found = false;
137 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
138 assert(!found, "registering twice");
139 _intrinsics.insert_before(index, cg);
140 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
141 }
142
143 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
144 assert(m->is_loaded(), "don't try this on unloaded methods");
145 if (_intrinsics.length() > 0) {
146 bool found = false;
147 int index = intrinsic_insertion_index(m, is_virtual, found);
148 if (found) {
149 return _intrinsics.at(index);
150 }
151 }
152 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
153 if (m->intrinsic_id() != vmIntrinsics::_none &&
154 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
155 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
156 if (cg != nullptr) {
157 // Save it for next time:
158 register_intrinsic(cg);
159 return cg;
160 } else {
161 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
162 }
163 }
164 return nullptr;
165 }
166
167 // Compile::make_vm_intrinsic is defined in library_call.cpp.
168
169 #ifndef PRODUCT
170 // statistics gathering...
171
172 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
174
175 inline int as_int(vmIntrinsics::ID id) {
176 return vmIntrinsics::as_int(id);
177 }
178
179 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
180 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
181 int oflags = _intrinsic_hist_flags[as_int(id)];
182 assert(flags != 0, "what happened?");
183 if (is_virtual) {
184 flags |= _intrinsic_virtual;
185 }
186 bool changed = (flags != oflags);
187 if ((flags & _intrinsic_worked) != 0) {
188 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
189 if (count == 1) {
190 changed = true; // first time
191 }
192 // increment the overall count also:
193 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
194 }
195 if (changed) {
196 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
197 // Something changed about the intrinsic's virtuality.
198 if ((flags & _intrinsic_virtual) != 0) {
199 // This is the first use of this intrinsic as a virtual call.
200 if (oflags != 0) {
201 // We already saw it as a non-virtual, so note both cases.
202 flags |= _intrinsic_both;
203 }
204 } else if ((oflags & _intrinsic_both) == 0) {
205 // This is the first use of this intrinsic as a non-virtual
206 flags |= _intrinsic_both;
207 }
208 }
209 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
210 }
211 // update the overall flags also:
212 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
213 return changed;
214 }
215
216 static char* format_flags(int flags, char* buf) {
217 buf[0] = 0;
218 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
219 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
220 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
221 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
222 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
223 if (buf[0] == 0) strcat(buf, ",");
224 assert(buf[0] == ',', "must be");
225 return &buf[1];
226 }
227
228 void Compile::print_intrinsic_statistics() {
229 char flagsbuf[100];
230 ttyLocker ttyl;
231 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
232 tty->print_cr("Compiler intrinsic usage:");
233 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
234 if (total == 0) total = 1; // avoid div0 in case of no successes
235 #define PRINT_STAT_LINE(name, c, f) \
236 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
237 for (auto id : EnumRange<vmIntrinsicID>{}) {
238 int flags = _intrinsic_hist_flags[as_int(id)];
239 juint count = _intrinsic_hist_count[as_int(id)];
240 if ((flags | count) != 0) {
241 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
242 }
243 }
244 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
245 if (xtty != nullptr) xtty->tail("statistics");
246 }
247
248 void Compile::print_statistics() {
249 { ttyLocker ttyl;
250 if (xtty != nullptr) xtty->head("statistics type='opto'");
251 Parse::print_statistics();
252 PhaseStringOpts::print_statistics();
253 PhaseCCP::print_statistics();
254 PhaseRegAlloc::print_statistics();
255 PhaseOutput::print_statistics();
256 PhasePeephole::print_statistics();
257 PhaseIdealLoop::print_statistics();
258 ConnectionGraph::print_statistics();
259 PhaseMacroExpand::print_statistics();
260 if (xtty != nullptr) xtty->tail("statistics");
261 }
262 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
263 // put this under its own <statistics> element.
264 print_intrinsic_statistics();
265 }
266 }
267 #endif //PRODUCT
268
269 void Compile::gvn_replace_by(Node* n, Node* nn) {
270 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
271 Node* use = n->last_out(i);
272 bool is_in_table = initial_gvn()->hash_delete(use);
273 uint uses_found = 0;
274 for (uint j = 0; j < use->len(); j++) {
275 if (use->in(j) == n) {
276 if (j < use->req())
277 use->set_req(j, nn);
278 else
279 use->set_prec(j, nn);
280 uses_found++;
281 }
282 }
283 if (is_in_table) {
284 // reinsert into table
285 initial_gvn()->hash_find_insert(use);
286 }
287 record_for_igvn(use);
288 i -= uses_found; // we deleted 1 or more copies of this edge
289 }
290 }
291
292
293 // Identify all nodes that are reachable from below, useful.
294 // Use breadth-first pass that records state in a Unique_Node_List,
295 // recursive traversal is slower.
296 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
297 int estimated_worklist_size = live_nodes();
298 useful.map( estimated_worklist_size, nullptr ); // preallocate space
299
300 // Initialize worklist
301 if (root() != nullptr) { useful.push(root()); }
302 // If 'top' is cached, declare it useful to preserve cached node
303 if (cached_top_node()) { useful.push(cached_top_node()); }
304
305 // Push all useful nodes onto the list, breadthfirst
306 for( uint next = 0; next < useful.size(); ++next ) {
307 assert( next < unique(), "Unique useful nodes < total nodes");
308 Node *n = useful.at(next);
309 uint max = n->len();
310 for( uint i = 0; i < max; ++i ) {
311 Node *m = n->in(i);
312 if (not_a_node(m)) continue;
313 useful.push(m);
314 }
315 }
316 }
317
318 // Update dead_node_list with any missing dead nodes using useful
319 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
320 void Compile::update_dead_node_list(Unique_Node_List &useful) {
321 uint max_idx = unique();
322 VectorSet& useful_node_set = useful.member_set();
323
324 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
325 // If node with index node_idx is not in useful set,
326 // mark it as dead in dead node list.
327 if (!useful_node_set.test(node_idx)) {
328 record_dead_node(node_idx);
329 }
330 }
331 }
332
333 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
334 int shift = 0;
335 for (int i = 0; i < inlines->length(); i++) {
336 CallGenerator* cg = inlines->at(i);
337 if (useful.member(cg->call_node())) {
338 if (shift > 0) {
339 inlines->at_put(i - shift, cg);
340 }
341 } else {
342 shift++; // skip over the dead element
343 }
344 }
345 if (shift > 0) {
346 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
347 }
348 }
349
350 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
351 assert(dead != nullptr && dead->is_Call(), "sanity");
352 int found = 0;
353 for (int i = 0; i < inlines->length(); i++) {
354 if (inlines->at(i)->call_node() == dead) {
355 inlines->remove_at(i);
356 found++;
357 NOT_DEBUG( break; ) // elements are unique, so exit early
358 }
359 }
360 assert(found <= 1, "not unique");
361 }
362
363 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) {
364 for (int i = node_list.length() - 1; i >= 0; i--) {
365 Node* n = node_list.at(i);
366 if (!useful.member(n)) {
367 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
368 }
369 }
370 }
371
372 void Compile::remove_useless_node(Node* dead) {
373 remove_modified_node(dead);
374
375 // Constant node that has no out-edges and has only one in-edge from
376 // root is usually dead. However, sometimes reshaping walk makes
377 // it reachable by adding use edges. So, we will NOT count Con nodes
378 // as dead to be conservative about the dead node count at any
379 // given time.
380 if (!dead->is_Con()) {
381 record_dead_node(dead->_idx);
382 }
383 if (dead->is_macro()) {
384 remove_macro_node(dead);
385 }
386 if (dead->is_expensive()) {
387 remove_expensive_node(dead);
388 }
389 if (dead->Opcode() == Op_Opaque4) {
390 remove_skeleton_predicate_opaq(dead);
391 }
392 if (dead->for_post_loop_opts_igvn()) {
393 remove_from_post_loop_opts_igvn(dead);
394 }
395 if (dead->is_Call()) {
396 remove_useless_late_inlines( &_late_inlines, dead);
397 remove_useless_late_inlines( &_string_late_inlines, dead);
398 remove_useless_late_inlines( &_boxing_late_inlines, dead);
399 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
400
401 if (dead->is_CallStaticJava()) {
402 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
403 }
404 }
405 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
406 bs->unregister_potential_barrier_node(dead);
407 }
408
409 // Disconnect all useless nodes by disconnecting those at the boundary.
410 void Compile::disconnect_useless_nodes(Unique_Node_List &useful, Unique_Node_List* worklist) {
411 uint next = 0;
412 while (next < useful.size()) {
413 Node *n = useful.at(next++);
414 if (n->is_SafePoint()) {
415 // We're done with a parsing phase. Replaced nodes are not valid
416 // beyond that point.
417 n->as_SafePoint()->delete_replaced_nodes();
418 }
419 // Use raw traversal of out edges since this code removes out edges
420 int max = n->outcnt();
421 for (int j = 0; j < max; ++j) {
422 Node* child = n->raw_out(j);
423 if (!useful.member(child)) {
424 assert(!child->is_top() || child != top(),
425 "If top is cached in Compile object it is in useful list");
426 // Only need to remove this out-edge to the useless node
427 n->raw_del_out(j);
428 --j;
429 --max;
430 }
431 }
432 if (n->outcnt() == 1 && n->has_special_unique_user()) {
433 worklist->push(n->unique_out());
434 }
435 }
436
437 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
438 remove_useless_nodes(_predicate_opaqs, useful); // remove useless predicate opaque nodes
439 remove_useless_nodes(_skeleton_predicate_opaqs, useful);
440 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
441 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
442 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
443 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
444 #ifdef ASSERT
445 if (_modified_nodes != nullptr) {
446 _modified_nodes->remove_useless_nodes(useful.member_set());
447 }
448 #endif
449
450 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
451 bs->eliminate_useless_gc_barriers(useful, this);
452 // clean up the late inline lists
453 remove_useless_late_inlines( &_late_inlines, useful);
454 remove_useless_late_inlines( &_string_late_inlines, useful);
455 remove_useless_late_inlines( &_boxing_late_inlines, useful);
456 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
457 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
458 }
459
460 // ============================================================================
461 //------------------------------CompileWrapper---------------------------------
462 class CompileWrapper : public StackObj {
463 Compile *const _compile;
464 public:
465 CompileWrapper(Compile* compile);
466
467 ~CompileWrapper();
468 };
469
470 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
471 // the Compile* pointer is stored in the current ciEnv:
472 ciEnv* env = compile->env();
473 assert(env == ciEnv::current(), "must already be a ciEnv active");
474 assert(env->compiler_data() == nullptr, "compile already active?");
475 env->set_compiler_data(compile);
476 assert(compile == Compile::current(), "sanity");
477
478 compile->set_type_dict(nullptr);
479 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
480 compile->clone_map().set_clone_idx(0);
481 compile->set_type_last_size(0);
482 compile->set_last_tf(nullptr, nullptr);
483 compile->set_indexSet_arena(nullptr);
484 compile->set_indexSet_free_block_list(nullptr);
485 compile->init_type_arena();
486 Type::Initialize(compile);
487 _compile->begin_method();
488 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
489 }
490 CompileWrapper::~CompileWrapper() {
491 // simulate crash during compilation
492 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
493
494 _compile->end_method();
495 _compile->env()->set_compiler_data(nullptr);
496 }
497
498
499 //----------------------------print_compile_messages---------------------------
500 void Compile::print_compile_messages() {
501 #ifndef PRODUCT
502 // Check if recompiling
503 if (!subsume_loads() && PrintOpto) {
504 // Recompiling without allowing machine instructions to subsume loads
505 tty->print_cr("*********************************************************");
506 tty->print_cr("** Bailout: Recompile without subsuming loads **");
507 tty->print_cr("*********************************************************");
508 }
509 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
510 // Recompiling without escape analysis
511 tty->print_cr("*********************************************************");
512 tty->print_cr("** Bailout: Recompile without escape analysis **");
513 tty->print_cr("*********************************************************");
514 }
515 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
516 // Recompiling without iterative escape analysis
517 tty->print_cr("*********************************************************");
518 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
519 tty->print_cr("*********************************************************");
520 }
521 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
522 // Recompiling without boxing elimination
523 tty->print_cr("*********************************************************");
524 tty->print_cr("** Bailout: Recompile without boxing elimination **");
525 tty->print_cr("*********************************************************");
526 }
527 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
528 // Recompiling without locks coarsening
529 tty->print_cr("*********************************************************");
530 tty->print_cr("** Bailout: Recompile without locks coarsening **");
531 tty->print_cr("*********************************************************");
532 }
533 if (env()->break_at_compile()) {
534 // Open the debugger when compiling this method.
535 tty->print("### Breaking when compiling: ");
536 method()->print_short_name();
537 tty->cr();
538 BREAKPOINT;
539 }
540
541 if( PrintOpto ) {
542 if (is_osr_compilation()) {
543 tty->print("[OSR]%3d", _compile_id);
544 } else {
545 tty->print("%3d", _compile_id);
546 }
547 }
548 #endif
549 }
550
551 #ifndef PRODUCT
552 void Compile::print_ideal_ir(const char* phase_name) {
553 ttyLocker ttyl;
554 // keep the following output all in one block
555 // This output goes directly to the tty, not the compiler log.
556 // To enable tools to match it up with the compilation activity,
557 // be sure to tag this tty output with the compile ID.
558 if (xtty != nullptr) {
559 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
560 compile_id(),
561 is_osr_compilation() ? " compile_kind='osr'" : "",
562 phase_name);
563 }
564 if (_output == nullptr) {
565 tty->print_cr("AFTER: %s", phase_name);
566 // Print out all nodes in ascending order of index.
567 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$");
568 } else {
569 // Dump the node blockwise if we have a scheduling
570 _output->print_scheduling();
571 }
572
573 if (xtty != nullptr) {
574 xtty->tail("ideal");
575 }
576 }
577 #endif
578
579 // ============================================================================
580 //------------------------------Compile standard-------------------------------
581 debug_only( int Compile::_debug_idx = 100000; )
582
583 // Compile a method. entry_bci is -1 for normal compilations and indicates
584 // the continuation bci for on stack replacement.
585
586
587 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
588 Options options, DirectiveSet* directive)
589 : Phase(Compiler),
590 _compile_id(ci_env->compile_id()),
591 _options(options),
592 _method(target),
593 _entry_bci(osr_bci),
594 _ilt(nullptr),
595 _stub_function(nullptr),
596 _stub_name(nullptr),
597 _stub_entry_point(nullptr),
598 _max_node_limit(MaxNodeLimit),
599 _post_loop_opts_phase(false),
600 _inlining_progress(false),
601 _inlining_incrementally(false),
602 _do_cleanup(false),
603 _has_reserved_stack_access(target->has_reserved_stack_access()),
604 #ifndef PRODUCT
605 _igv_idx(0),
606 _trace_opto_output(directive->TraceOptoOutputOption),
607 #endif
608 _has_method_handle_invokes(false),
609 _clinit_barrier_on_entry(false),
610 _stress_seed(0),
611 _comp_arena(mtCompiler),
612 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
613 _env(ci_env),
614 _directive(directive),
615 _log(ci_env->log()),
616 _failure_reason(nullptr),
617 _intrinsics (comp_arena(), 0, 0, nullptr),
618 _macro_nodes (comp_arena(), 8, 0, nullptr),
619 _predicate_opaqs (comp_arena(), 8, 0, nullptr),
620 _skeleton_predicate_opaqs (comp_arena(), 8, 0, nullptr),
621 _expensive_nodes (comp_arena(), 8, 0, nullptr),
622 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
623 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
624 _coarsened_locks (comp_arena(), 8, 0, nullptr),
625 _congraph(nullptr),
626 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
627 _dead_node_list(comp_arena()),
628 _dead_node_count(0),
629 _node_arena(mtCompiler),
630 _old_arena(mtCompiler),
631 _mach_constant_base_node(nullptr),
632 _Compile_types(mtCompiler),
633 _initial_gvn(nullptr),
634 _for_igvn(nullptr),
635 _late_inlines(comp_arena(), 2, 0, nullptr),
636 _string_late_inlines(comp_arena(), 2, 0, nullptr),
637 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
638 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
639 _late_inlines_pos(0),
640 _number_of_mh_late_inlines(0),
641 _print_inlining_stream(new (mtCompiler) stringStream()),
642 _print_inlining_list(nullptr),
643 _print_inlining_idx(0),
644 _print_inlining_output(nullptr),
645 _replay_inline_data(nullptr),
646 _java_calls(0),
647 _inner_loops(0),
648 _interpreter_frame_size(0),
649 _output(nullptr)
650 #ifndef PRODUCT
651 , _in_dump_cnt(0)
652 #endif
653 {
654 C = this;
655 CompileWrapper cw(this);
656
657 if (CITimeVerbose) {
658 tty->print(" ");
659 target->holder()->name()->print();
660 tty->print(".");
661 target->print_short_name();
662 tty->print(" ");
663 }
664 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
665 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
666
667 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
668 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
669 // We can always print a disassembly, either abstract (hex dump) or
670 // with the help of a suitable hsdis library. Thus, we should not
671 // couple print_assembly and print_opto_assembly controls.
672 // But: always print opto and regular assembly on compile command 'print'.
673 bool print_assembly = directive->PrintAssemblyOption;
674 set_print_assembly(print_opto_assembly || print_assembly);
675 #else
676 set_print_assembly(false); // must initialize.
677 #endif
678
679 #ifndef PRODUCT
680 set_parsed_irreducible_loop(false);
681
682 if (directive->ReplayInlineOption) {
683 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
684 }
685 #endif
686 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
687 set_print_intrinsics(directive->PrintIntrinsicsOption);
688 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
689
690 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
691 // Make sure the method being compiled gets its own MDO,
692 // so we can at least track the decompile_count().
693 // Need MDO to record RTM code generation state.
694 method()->ensure_method_data();
695 }
696
697 Init(/*do_aliasing=*/ true);
698
699 print_compile_messages();
700
701 _ilt = InlineTree::build_inline_tree_root();
702
703 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
704 assert(num_alias_types() >= AliasIdxRaw, "");
705
706 #define MINIMUM_NODE_HASH 1023
707 // Node list that Iterative GVN will start with
708 Unique_Node_List for_igvn(comp_arena());
709 set_for_igvn(&for_igvn);
710
711 // GVN that will be run immediately on new nodes
712 uint estimated_size = method()->code_size()*4+64;
713 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
714 PhaseGVN gvn(node_arena(), estimated_size);
715 set_initial_gvn(&gvn);
716
717 print_inlining_init();
718 { // Scope for timing the parser
719 TracePhase tp("parse", &timers[_t_parser]);
720
721 // Put top into the hash table ASAP.
722 initial_gvn()->transform_no_reclaim(top());
723
724 // Set up tf(), start(), and find a CallGenerator.
725 CallGenerator* cg = nullptr;
726 if (is_osr_compilation()) {
727 const TypeTuple *domain = StartOSRNode::osr_domain();
728 const TypeTuple *range = TypeTuple::make_range(method()->signature());
729 init_tf(TypeFunc::make(domain, range));
730 StartNode* s = new StartOSRNode(root(), domain);
731 initial_gvn()->set_type_bottom(s);
732 init_start(s);
733 cg = CallGenerator::for_osr(method(), entry_bci());
734 } else {
735 // Normal case.
736 init_tf(TypeFunc::make(method()));
737 StartNode* s = new StartNode(root(), tf()->domain());
738 initial_gvn()->set_type_bottom(s);
739 init_start(s);
740 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
741 // With java.lang.ref.reference.get() we must go through the
742 // intrinsic - even when get() is the root
743 // method of the compile - so that, if necessary, the value in
744 // the referent field of the reference object gets recorded by
745 // the pre-barrier code.
746 cg = find_intrinsic(method(), false);
747 }
748 if (cg == nullptr) {
749 float past_uses = method()->interpreter_invocation_count();
750 float expected_uses = past_uses;
751 cg = CallGenerator::for_inline(method(), expected_uses);
752 }
753 }
754 if (failing()) return;
755 if (cg == nullptr) {
756 const char* reason = InlineTree::check_can_parse(method());
757 assert(reason != nullptr, "expect reason for parse failure");
758 stringStream ss;
759 ss.print("cannot parse method: %s", reason);
760 record_method_not_compilable(ss.as_string());
761 return;
762 }
763
764 gvn.set_type(root(), root()->bottom_type());
765
766 JVMState* jvms = build_start_state(start(), tf());
767 if ((jvms = cg->generate(jvms)) == nullptr) {
768 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
769 assert(failure_reason() != nullptr, "expect reason for parse failure");
770 stringStream ss;
771 ss.print("method parse failed: %s", failure_reason());
772 record_method_not_compilable(ss.as_string());
773 }
774 return;
775 }
776 GraphKit kit(jvms);
777
778 if (!kit.stopped()) {
779 // Accept return values, and transfer control we know not where.
780 // This is done by a special, unique ReturnNode bound to root.
781 return_values(kit.jvms());
782 }
783
784 if (kit.has_exceptions()) {
785 // Any exceptions that escape from this call must be rethrown
786 // to whatever caller is dynamically above us on the stack.
787 // This is done by a special, unique RethrowNode bound to root.
788 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
789 }
790
791 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
792
793 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
794 inline_string_calls(true);
795 }
796
797 if (failing()) return;
798
799 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
800
801 // Remove clutter produced by parsing.
802 if (!failing()) {
803 ResourceMark rm;
804 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
805 }
806 }
807
808 // Note: Large methods are capped off in do_one_bytecode().
809 if (failing()) return;
810
811 // After parsing, node notes are no longer automagic.
812 // They must be propagated by register_new_node_with_optimizer(),
813 // clone(), or the like.
814 set_default_node_notes(nullptr);
815
816 #ifndef PRODUCT
817 if (should_print_igv(1)) {
818 _igv_printer->print_inlining();
819 }
820 #endif
821
822 if (failing()) return;
823 NOT_PRODUCT( verify_graph_edges(); )
824
825 // If any phase is randomized for stress testing, seed random number
826 // generation and log the seed for repeatability.
827 if (StressLCM || StressGCM || StressIGVN || StressCCP) {
828 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) {
829 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
830 FLAG_SET_ERGO(StressSeed, _stress_seed);
831 } else {
832 _stress_seed = StressSeed;
833 }
834 if (_log != nullptr) {
835 _log->elem("stress_test seed='%u'", _stress_seed);
836 }
837 }
838
839 // Now optimize
840 Optimize();
841 if (failing()) return;
842 NOT_PRODUCT( verify_graph_edges(); )
843
844 #ifndef PRODUCT
845 if (should_print_ideal()) {
846 print_ideal_ir("print_ideal");
847 }
848 #endif
849
850 #ifdef ASSERT
851 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
852 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
853 #endif
854
855 // Dump compilation data to replay it.
856 if (directive->DumpReplayOption) {
857 env()->dump_replay_data(_compile_id);
858 }
859 if (directive->DumpInlineOption && (ilt() != nullptr)) {
860 env()->dump_inline_data(_compile_id);
861 }
862
863 // Now that we know the size of all the monitors we can add a fixed slot
864 // for the original deopt pc.
865 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
866 set_fixed_slots(next_slot);
867
868 // Compute when to use implicit null checks. Used by matching trap based
869 // nodes and NullCheck optimization.
870 set_allowed_deopt_reasons();
871
872 // Now generate code
873 Code_Gen();
874 }
875
876 //------------------------------Compile----------------------------------------
877 // Compile a runtime stub
878 Compile::Compile( ciEnv* ci_env,
879 TypeFunc_generator generator,
880 address stub_function,
881 const char *stub_name,
882 int is_fancy_jump,
883 bool pass_tls,
884 bool return_pc,
885 DirectiveSet* directive)
886 : Phase(Compiler),
887 _compile_id(0),
888 _options(Options::for_runtime_stub()),
889 _method(nullptr),
890 _entry_bci(InvocationEntryBci),
891 _stub_function(stub_function),
892 _stub_name(stub_name),
893 _stub_entry_point(nullptr),
894 _max_node_limit(MaxNodeLimit),
895 _post_loop_opts_phase(false),
896 _inlining_progress(false),
897 _inlining_incrementally(false),
898 _has_reserved_stack_access(false),
899 #ifndef PRODUCT
900 _igv_idx(0),
901 _trace_opto_output(directive->TraceOptoOutputOption),
902 #endif
903 _has_method_handle_invokes(false),
904 _clinit_barrier_on_entry(false),
905 _stress_seed(0),
906 _comp_arena(mtCompiler),
907 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
908 _env(ci_env),
909 _directive(directive),
910 _log(ci_env->log()),
911 _failure_reason(nullptr),
912 _congraph(nullptr),
913 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
914 _dead_node_list(comp_arena()),
915 _dead_node_count(0),
916 _node_arena(mtCompiler),
917 _old_arena(mtCompiler),
918 _mach_constant_base_node(nullptr),
919 _Compile_types(mtCompiler),
920 _initial_gvn(nullptr),
921 _for_igvn(nullptr),
922 _number_of_mh_late_inlines(0),
923 _print_inlining_stream(new (mtCompiler) stringStream()),
924 _print_inlining_list(nullptr),
925 _print_inlining_idx(0),
926 _print_inlining_output(nullptr),
927 _replay_inline_data(nullptr),
928 _java_calls(0),
929 _inner_loops(0),
930 _interpreter_frame_size(0),
931 _output(nullptr),
932 #ifndef PRODUCT
933 _in_dump_cnt(0),
934 #endif
935 _allowed_reasons(0) {
936 C = this;
937
938 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
939 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
940
941 #ifndef PRODUCT
942 set_print_assembly(PrintFrameConverterAssembly);
943 set_parsed_irreducible_loop(false);
944 #else
945 set_print_assembly(false); // Must initialize.
946 #endif
947 set_has_irreducible_loop(false); // no loops
948
949 CompileWrapper cw(this);
950 Init(/*do_aliasing=*/ false);
951 init_tf((*generator)());
952
953 {
954 // The following is a dummy for the sake of GraphKit::gen_stub
955 Unique_Node_List for_igvn(comp_arena());
956 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
957 PhaseGVN gvn(Thread::current()->resource_area(),255);
958 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
959 gvn.transform_no_reclaim(top());
960
961 GraphKit kit;
962 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
963 }
964
965 NOT_PRODUCT( verify_graph_edges(); )
966
967 Code_Gen();
968 }
969
970 //------------------------------Init-------------------------------------------
971 // Prepare for a single compilation
972 void Compile::Init(bool aliasing) {
973 _do_aliasing = aliasing;
974 _unique = 0;
975 _regalloc = nullptr;
976
977 _tf = nullptr; // filled in later
978 _top = nullptr; // cached later
979 _matcher = nullptr; // filled in later
980 _cfg = nullptr; // filled in later
981
982 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
983
984 _node_note_array = nullptr;
985 _default_node_notes = nullptr;
986 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
987
988 _immutable_memory = nullptr; // filled in at first inquiry
989
990 #ifdef ASSERT
991 _phase_optimize_finished = false;
992 _exception_backedge = false;
993 _type_verify = nullptr;
994 #endif
995
996 // Globally visible Nodes
997 // First set TOP to null to give safe behavior during creation of RootNode
998 set_cached_top_node(nullptr);
999 set_root(new RootNode());
1000 // Now that you have a Root to point to, create the real TOP
1001 set_cached_top_node( new ConNode(Type::TOP) );
1002 set_recent_alloc(nullptr, nullptr);
1003
1004 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1005 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1006 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1007 env()->set_dependencies(new Dependencies(env()));
1008
1009 _fixed_slots = 0;
1010 set_has_split_ifs(false);
1011 set_has_loops(false); // first approximation
1012 set_has_stringbuilder(false);
1013 set_has_boxed_value(false);
1014 _trap_can_recompile = false; // no traps emitted yet
1015 _major_progress = true; // start out assuming good things will happen
1016 set_has_unsafe_access(false);
1017 set_max_vector_size(0);
1018 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1019 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1020 set_decompile_count(0);
1021
1022 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1023 _loop_opts_cnt = LoopOptsCount;
1024 set_do_inlining(Inline);
1025 set_max_inline_size(MaxInlineSize);
1026 set_freq_inline_size(FreqInlineSize);
1027 set_do_scheduling(OptoScheduling);
1028
1029 set_do_vector_loop(false);
1030 set_has_monitors(false);
1031
1032 if (AllowVectorizeOnDemand) {
1033 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1034 set_do_vector_loop(true);
1035 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());})
1036 } else if (has_method() && method()->name() != 0 &&
1037 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1038 set_do_vector_loop(true);
1039 }
1040 }
1041 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1042 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());})
1043
1044 set_rtm_state(NoRTM); // No RTM lock eliding by default
1045 _max_node_limit = _directive->MaxNodeLimitOption;
1046
1047 #if INCLUDE_RTM_OPT
1048 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) {
1049 int rtm_state = method()->method_data()->rtm_state();
1050 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1051 // Don't generate RTM lock eliding code.
1052 set_rtm_state(NoRTM);
1053 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1054 // Generate RTM lock eliding code without abort ratio calculation code.
1055 set_rtm_state(UseRTM);
1056 } else if (UseRTMDeopt) {
1057 // Generate RTM lock eliding code and include abort ratio calculation
1058 // code if UseRTMDeopt is on.
1059 set_rtm_state(ProfileRTM);
1060 }
1061 }
1062 #endif
1063 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1064 set_clinit_barrier_on_entry(true);
1065 }
1066 if (debug_info()->recording_non_safepoints()) {
1067 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1068 (comp_arena(), 8, 0, nullptr));
1069 set_default_node_notes(Node_Notes::make(this));
1070 }
1071
1072 const int grow_ats = 16;
1073 _max_alias_types = grow_ats;
1074 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1075 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1076 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1077 {
1078 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1079 }
1080 // Initialize the first few types.
1081 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1082 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1083 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1084 _num_alias_types = AliasIdxRaw+1;
1085 // Zero out the alias type cache.
1086 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1087 // A null adr_type hits in the cache right away. Preload the right answer.
1088 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1089 }
1090
1091 //---------------------------init_start----------------------------------------
1092 // Install the StartNode on this compile object.
1093 void Compile::init_start(StartNode* s) {
1094 if (failing())
1095 return; // already failing
1096 assert(s == start(), "");
1097 }
1098
1099 /**
1100 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1101 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1102 * the ideal graph.
1103 */
1104 StartNode* Compile::start() const {
1105 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1106 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1107 Node* start = root()->fast_out(i);
1108 if (start->is_Start()) {
1109 return start->as_Start();
1110 }
1111 }
1112 fatal("Did not find Start node!");
1113 return nullptr;
1114 }
1115
1116 //-------------------------------immutable_memory-------------------------------------
1117 // Access immutable memory
1118 Node* Compile::immutable_memory() {
1119 if (_immutable_memory != nullptr) {
1120 return _immutable_memory;
1121 }
1122 StartNode* s = start();
1123 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1124 Node *p = s->fast_out(i);
1125 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1126 _immutable_memory = p;
1127 return _immutable_memory;
1128 }
1129 }
1130 ShouldNotReachHere();
1131 return nullptr;
1132 }
1133
1134 //----------------------set_cached_top_node------------------------------------
1135 // Install the cached top node, and make sure Node::is_top works correctly.
1136 void Compile::set_cached_top_node(Node* tn) {
1137 if (tn != nullptr) verify_top(tn);
1138 Node* old_top = _top;
1139 _top = tn;
1140 // Calling Node::setup_is_top allows the nodes the chance to adjust
1141 // their _out arrays.
1142 if (_top != nullptr) _top->setup_is_top();
1143 if (old_top != nullptr) old_top->setup_is_top();
1144 assert(_top == nullptr || top()->is_top(), "");
1145 }
1146
1147 #ifdef ASSERT
1148 uint Compile::count_live_nodes_by_graph_walk() {
1149 Unique_Node_List useful(comp_arena());
1150 // Get useful node list by walking the graph.
1151 identify_useful_nodes(useful);
1152 return useful.size();
1153 }
1154
1155 void Compile::print_missing_nodes() {
1156
1157 // Return if CompileLog is null and PrintIdealNodeCount is false.
1158 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1159 return;
1160 }
1161
1162 // This is an expensive function. It is executed only when the user
1163 // specifies VerifyIdealNodeCount option or otherwise knows the
1164 // additional work that needs to be done to identify reachable nodes
1165 // by walking the flow graph and find the missing ones using
1166 // _dead_node_list.
1167
1168 Unique_Node_List useful(comp_arena());
1169 // Get useful node list by walking the graph.
1170 identify_useful_nodes(useful);
1171
1172 uint l_nodes = C->live_nodes();
1173 uint l_nodes_by_walk = useful.size();
1174
1175 if (l_nodes != l_nodes_by_walk) {
1176 if (_log != nullptr) {
1177 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1178 _log->stamp();
1179 _log->end_head();
1180 }
1181 VectorSet& useful_member_set = useful.member_set();
1182 int last_idx = l_nodes_by_walk;
1183 for (int i = 0; i < last_idx; i++) {
1184 if (useful_member_set.test(i)) {
1185 if (_dead_node_list.test(i)) {
1186 if (_log != nullptr) {
1187 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1188 }
1189 if (PrintIdealNodeCount) {
1190 // Print the log message to tty
1191 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1192 useful.at(i)->dump();
1193 }
1194 }
1195 }
1196 else if (! _dead_node_list.test(i)) {
1197 if (_log != nullptr) {
1198 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1199 }
1200 if (PrintIdealNodeCount) {
1201 // Print the log message to tty
1202 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1203 }
1204 }
1205 }
1206 if (_log != nullptr) {
1207 _log->tail("mismatched_nodes");
1208 }
1209 }
1210 }
1211 void Compile::record_modified_node(Node* n) {
1212 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1213 _modified_nodes->push(n);
1214 }
1215 }
1216
1217 void Compile::remove_modified_node(Node* n) {
1218 if (_modified_nodes != nullptr) {
1219 _modified_nodes->remove(n);
1220 }
1221 }
1222 #endif
1223
1224 #ifndef PRODUCT
1225 void Compile::verify_top(Node* tn) const {
1226 if (tn != nullptr) {
1227 assert(tn->is_Con(), "top node must be a constant");
1228 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1229 assert(tn->in(0) != nullptr, "must have live top node");
1230 }
1231 }
1232 #endif
1233
1234
1235 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1236
1237 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1238 guarantee(arr != nullptr, "");
1239 int num_blocks = arr->length();
1240 if (grow_by < num_blocks) grow_by = num_blocks;
1241 int num_notes = grow_by * _node_notes_block_size;
1242 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1243 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1244 while (num_notes > 0) {
1245 arr->append(notes);
1246 notes += _node_notes_block_size;
1247 num_notes -= _node_notes_block_size;
1248 }
1249 assert(num_notes == 0, "exact multiple, please");
1250 }
1251
1252 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1253 if (source == nullptr || dest == nullptr) return false;
1254
1255 if (dest->is_Con())
1256 return false; // Do not push debug info onto constants.
1257
1258 #ifdef ASSERT
1259 // Leave a bread crumb trail pointing to the original node:
1260 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1261 dest->set_debug_orig(source);
1262 }
1263 #endif
1264
1265 if (node_note_array() == nullptr)
1266 return false; // Not collecting any notes now.
1267
1268 // This is a copy onto a pre-existing node, which may already have notes.
1269 // If both nodes have notes, do not overwrite any pre-existing notes.
1270 Node_Notes* source_notes = node_notes_at(source->_idx);
1271 if (source_notes == nullptr || source_notes->is_clear()) return false;
1272 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1273 if (dest_notes == nullptr || dest_notes->is_clear()) {
1274 return set_node_notes_at(dest->_idx, source_notes);
1275 }
1276
1277 Node_Notes merged_notes = (*source_notes);
1278 // The order of operations here ensures that dest notes will win...
1279 merged_notes.update_from(dest_notes);
1280 return set_node_notes_at(dest->_idx, &merged_notes);
1281 }
1282
1283
1284 //--------------------------allow_range_check_smearing-------------------------
1285 // Gating condition for coalescing similar range checks.
1286 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1287 // single covering check that is at least as strong as any of them.
1288 // If the optimization succeeds, the simplified (strengthened) range check
1289 // will always succeed. If it fails, we will deopt, and then give up
1290 // on the optimization.
1291 bool Compile::allow_range_check_smearing() const {
1292 // If this method has already thrown a range-check,
1293 // assume it was because we already tried range smearing
1294 // and it failed.
1295 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1296 return !already_trapped;
1297 }
1298
1299
1300 //------------------------------flatten_alias_type-----------------------------
1301 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1302 assert(do_aliasing(), "Aliasing should be enabled");
1303 int offset = tj->offset();
1304 TypePtr::PTR ptr = tj->ptr();
1305
1306 // Known instance (scalarizable allocation) alias only with itself.
1307 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1308 tj->is_oopptr()->is_known_instance();
1309
1310 // Process weird unsafe references.
1311 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1312 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1313 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1314 tj = TypeOopPtr::BOTTOM;
1315 ptr = tj->ptr();
1316 offset = tj->offset();
1317 }
1318
1319 // Array pointers need some flattening
1320 const TypeAryPtr* ta = tj->isa_aryptr();
1321 if (ta && ta->is_stable()) {
1322 // Erase stability property for alias analysis.
1323 tj = ta = ta->cast_to_stable(false);
1324 }
1325 if( ta && is_known_inst ) {
1326 if ( offset != Type::OffsetBot &&
1327 offset > arrayOopDesc::length_offset_in_bytes() ) {
1328 offset = Type::OffsetBot; // Flatten constant access into array body only
1329 tj = ta = ta->
1330 remove_speculative()->
1331 cast_to_ptr_type(ptr)->
1332 with_offset(offset);
1333 }
1334 } else if (ta) {
1335 // For arrays indexed by constant indices, we flatten the alias
1336 // space to include all of the array body. Only the header, klass
1337 // and array length can be accessed un-aliased.
1338 if( offset != Type::OffsetBot ) {
1339 if( ta->const_oop() ) { // MethodData* or Method*
1340 offset = Type::OffsetBot; // Flatten constant access into array body
1341 tj = ta = ta->
1342 remove_speculative()->
1343 cast_to_ptr_type(ptr)->
1344 cast_to_exactness(false)->
1345 with_offset(offset);
1346 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1347 // range is OK as-is.
1348 tj = ta = TypeAryPtr::RANGE;
1349 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1350 tj = TypeInstPtr::KLASS; // all klass loads look alike
1351 ta = TypeAryPtr::RANGE; // generic ignored junk
1352 ptr = TypePtr::BotPTR;
1353 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1354 tj = TypeInstPtr::MARK;
1355 ta = TypeAryPtr::RANGE; // generic ignored junk
1356 ptr = TypePtr::BotPTR;
1357 } else { // Random constant offset into array body
1358 offset = Type::OffsetBot; // Flatten constant access into array body
1359 tj = ta = ta->
1360 remove_speculative()->
1361 cast_to_ptr_type(ptr)->
1362 cast_to_exactness(false)->
1363 with_offset(offset);
1364 }
1365 }
1366 // Arrays of fixed size alias with arrays of unknown size.
1367 if (ta->size() != TypeInt::POS) {
1368 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1369 tj = ta = ta->
1370 remove_speculative()->
1371 cast_to_ptr_type(ptr)->
1372 with_ary(tary)->
1373 cast_to_exactness(false);
1374 }
1375 // Arrays of known objects become arrays of unknown objects.
1376 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1377 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1378 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1379 }
1380 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1381 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1382 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1383 }
1384 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1385 // cannot be distinguished by bytecode alone.
1386 if (ta->elem() == TypeInt::BOOL) {
1387 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1388 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1389 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1390 }
1391 // During the 2nd round of IterGVN, NotNull castings are removed.
1392 // Make sure the Bottom and NotNull variants alias the same.
1393 // Also, make sure exact and non-exact variants alias the same.
1394 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1395 tj = ta = ta->
1396 remove_speculative()->
1397 cast_to_ptr_type(TypePtr::BotPTR)->
1398 cast_to_exactness(false)->
1399 with_offset(offset);
1400 }
1401 }
1402
1403 // Oop pointers need some flattening
1404 const TypeInstPtr *to = tj->isa_instptr();
1405 if (to && to != TypeOopPtr::BOTTOM) {
1406 ciInstanceKlass* ik = to->instance_klass();
1407 if( ptr == TypePtr::Constant ) {
1408 if (ik != ciEnv::current()->Class_klass() ||
1409 offset < ik->layout_helper_size_in_bytes()) {
1410 // No constant oop pointers (such as Strings); they alias with
1411 // unknown strings.
1412 assert(!is_known_inst, "not scalarizable allocation");
1413 tj = to = to->
1414 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1415 remove_speculative()->
1416 cast_to_ptr_type(TypePtr::BotPTR)->
1417 cast_to_exactness(false);
1418 }
1419 } else if( is_known_inst ) {
1420 tj = to; // Keep NotNull and klass_is_exact for instance type
1421 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1422 // During the 2nd round of IterGVN, NotNull castings are removed.
1423 // Make sure the Bottom and NotNull variants alias the same.
1424 // Also, make sure exact and non-exact variants alias the same.
1425 tj = to = to->
1426 remove_speculative()->
1427 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1428 cast_to_ptr_type(TypePtr::BotPTR)->
1429 cast_to_exactness(false);
1430 }
1431 if (to->speculative() != nullptr) {
1432 tj = to = to->remove_speculative();
1433 }
1434 // Canonicalize the holder of this field
1435 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1436 // First handle header references such as a LoadKlassNode, even if the
1437 // object's klass is unloaded at compile time (4965979).
1438 if (!is_known_inst) { // Do it only for non-instance types
1439 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1440 }
1441 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1442 // Static fields are in the space above the normal instance
1443 // fields in the java.lang.Class instance.
1444 if (ik != ciEnv::current()->Class_klass()) {
1445 to = nullptr;
1446 tj = TypeOopPtr::BOTTOM;
1447 offset = tj->offset();
1448 }
1449 } else {
1450 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1451 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1452 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1453 if( is_known_inst ) {
1454 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id());
1455 } else {
1456 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset);
1457 }
1458 }
1459 }
1460 }
1461
1462 // Klass pointers to object array klasses need some flattening
1463 const TypeKlassPtr *tk = tj->isa_klassptr();
1464 if( tk ) {
1465 // If we are referencing a field within a Klass, we need
1466 // to assume the worst case of an Object. Both exact and
1467 // inexact types must flatten to the same alias class so
1468 // use NotNull as the PTR.
1469 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1470 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1471 env()->Object_klass(),
1472 offset);
1473 }
1474
1475 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1476 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1477 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1478 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1479 } else {
1480 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1481 }
1482 }
1483
1484 // Check for precise loads from the primary supertype array and force them
1485 // to the supertype cache alias index. Check for generic array loads from
1486 // the primary supertype array and also force them to the supertype cache
1487 // alias index. Since the same load can reach both, we need to merge
1488 // these 2 disparate memories into the same alias class. Since the
1489 // primary supertype array is read-only, there's no chance of confusion
1490 // where we bypass an array load and an array store.
1491 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1492 if (offset == Type::OffsetBot ||
1493 (offset >= primary_supers_offset &&
1494 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1495 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1496 offset = in_bytes(Klass::secondary_super_cache_offset());
1497 tj = tk = tk->with_offset(offset);
1498 }
1499 }
1500
1501 // Flatten all Raw pointers together.
1502 if (tj->base() == Type::RawPtr)
1503 tj = TypeRawPtr::BOTTOM;
1504
1505 if (tj->base() == Type::AnyPtr)
1506 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1507
1508 offset = tj->offset();
1509 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1510
1511 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1512 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1513 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1514 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1515 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1516 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1517 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1518 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1519 assert( tj->ptr() != TypePtr::TopPTR &&
1520 tj->ptr() != TypePtr::AnyNull &&
1521 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1522 // assert( tj->ptr() != TypePtr::Constant ||
1523 // tj->base() == Type::RawPtr ||
1524 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1525
1526 return tj;
1527 }
1528
1529 void Compile::AliasType::Init(int i, const TypePtr* at) {
1530 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1531 _index = i;
1532 _adr_type = at;
1533 _field = nullptr;
1534 _element = nullptr;
1535 _is_rewritable = true; // default
1536 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1537 if (atoop != nullptr && atoop->is_known_instance()) {
1538 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1539 _general_index = Compile::current()->get_alias_index(gt);
1540 } else {
1541 _general_index = 0;
1542 }
1543 }
1544
1545 BasicType Compile::AliasType::basic_type() const {
1546 if (element() != nullptr) {
1547 const Type* element = adr_type()->is_aryptr()->elem();
1548 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1549 } if (field() != nullptr) {
1550 return field()->layout_type();
1551 } else {
1552 return T_ILLEGAL; // unknown
1553 }
1554 }
1555
1556 //---------------------------------print_on------------------------------------
1557 #ifndef PRODUCT
1558 void Compile::AliasType::print_on(outputStream* st) {
1559 if (index() < 10)
1560 st->print("@ <%d> ", index());
1561 else st->print("@ <%d>", index());
1562 st->print(is_rewritable() ? " " : " RO");
1563 int offset = adr_type()->offset();
1564 if (offset == Type::OffsetBot)
1565 st->print(" +any");
1566 else st->print(" +%-3d", offset);
1567 st->print(" in ");
1568 adr_type()->dump_on(st);
1569 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1570 if (field() != nullptr && tjp) {
1571 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1572 tjp->offset() != field()->offset_in_bytes()) {
1573 st->print(" != ");
1574 field()->print();
1575 st->print(" ***");
1576 }
1577 }
1578 }
1579
1580 void print_alias_types() {
1581 Compile* C = Compile::current();
1582 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1583 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1584 C->alias_type(idx)->print_on(tty);
1585 tty->cr();
1586 }
1587 }
1588 #endif
1589
1590
1591 //----------------------------probe_alias_cache--------------------------------
1592 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1593 intptr_t key = (intptr_t) adr_type;
1594 key ^= key >> logAliasCacheSize;
1595 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1596 }
1597
1598
1599 //-----------------------------grow_alias_types--------------------------------
1600 void Compile::grow_alias_types() {
1601 const int old_ats = _max_alias_types; // how many before?
1602 const int new_ats = old_ats; // how many more?
1603 const int grow_ats = old_ats+new_ats; // how many now?
1604 _max_alias_types = grow_ats;
1605 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1606 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1607 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1608 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1609 }
1610
1611
1612 //--------------------------------find_alias_type------------------------------
1613 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1614 if (!do_aliasing()) {
1615 return alias_type(AliasIdxBot);
1616 }
1617
1618 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1619 if (ace->_adr_type == adr_type) {
1620 return alias_type(ace->_index);
1621 }
1622
1623 // Handle special cases.
1624 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1625 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1626
1627 // Do it the slow way.
1628 const TypePtr* flat = flatten_alias_type(adr_type);
1629
1630 #ifdef ASSERT
1631 {
1632 ResourceMark rm;
1633 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1634 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1635 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1636 Type::str(adr_type));
1637 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1638 const TypeOopPtr* foop = flat->is_oopptr();
1639 // Scalarizable allocations have exact klass always.
1640 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1641 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1642 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1643 Type::str(foop), Type::str(xoop));
1644 }
1645 }
1646 #endif
1647
1648 int idx = AliasIdxTop;
1649 for (int i = 0; i < num_alias_types(); i++) {
1650 if (alias_type(i)->adr_type() == flat) {
1651 idx = i;
1652 break;
1653 }
1654 }
1655
1656 if (idx == AliasIdxTop) {
1657 if (no_create) return nullptr;
1658 // Grow the array if necessary.
1659 if (_num_alias_types == _max_alias_types) grow_alias_types();
1660 // Add a new alias type.
1661 idx = _num_alias_types++;
1662 _alias_types[idx]->Init(idx, flat);
1663 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1664 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1665 if (flat->isa_instptr()) {
1666 if (flat->offset() == java_lang_Class::klass_offset()
1667 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1668 alias_type(idx)->set_rewritable(false);
1669 }
1670 if (flat->isa_aryptr()) {
1671 #ifdef ASSERT
1672 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1673 // (T_BYTE has the weakest alignment and size restrictions...)
1674 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1675 #endif
1676 if (flat->offset() == TypePtr::OffsetBot) {
1677 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1678 }
1679 }
1680 if (flat->isa_klassptr()) {
1681 if (UseCompactObjectHeaders) {
1682 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1683 alias_type(idx)->set_rewritable(false);
1684 }
1685 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1686 alias_type(idx)->set_rewritable(false);
1687 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1688 alias_type(idx)->set_rewritable(false);
1689 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1690 alias_type(idx)->set_rewritable(false);
1691 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1692 alias_type(idx)->set_rewritable(false);
1693 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1694 alias_type(idx)->set_rewritable(false);
1695 }
1696 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1697 // but the base pointer type is not distinctive enough to identify
1698 // references into JavaThread.)
1699
1700 // Check for final fields.
1701 const TypeInstPtr* tinst = flat->isa_instptr();
1702 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1703 ciField* field;
1704 if (tinst->const_oop() != nullptr &&
1705 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1706 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1707 // static field
1708 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1709 field = k->get_field_by_offset(tinst->offset(), true);
1710 } else {
1711 ciInstanceKlass *k = tinst->instance_klass();
1712 field = k->get_field_by_offset(tinst->offset(), false);
1713 }
1714 assert(field == nullptr ||
1715 original_field == nullptr ||
1716 (field->holder() == original_field->holder() &&
1717 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1718 field->is_static() == original_field->is_static()), "wrong field?");
1719 // Set field() and is_rewritable() attributes.
1720 if (field != nullptr) alias_type(idx)->set_field(field);
1721 }
1722 }
1723
1724 // Fill the cache for next time.
1725 ace->_adr_type = adr_type;
1726 ace->_index = idx;
1727 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1728
1729 // Might as well try to fill the cache for the flattened version, too.
1730 AliasCacheEntry* face = probe_alias_cache(flat);
1731 if (face->_adr_type == nullptr) {
1732 face->_adr_type = flat;
1733 face->_index = idx;
1734 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1735 }
1736
1737 return alias_type(idx);
1738 }
1739
1740
1741 Compile::AliasType* Compile::alias_type(ciField* field) {
1742 const TypeOopPtr* t;
1743 if (field->is_static())
1744 t = TypeInstPtr::make(field->holder()->java_mirror());
1745 else
1746 t = TypeOopPtr::make_from_klass_raw(field->holder());
1747 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1748 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1749 return atp;
1750 }
1751
1752
1753 //------------------------------have_alias_type--------------------------------
1754 bool Compile::have_alias_type(const TypePtr* adr_type) {
1755 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1756 if (ace->_adr_type == adr_type) {
1757 return true;
1758 }
1759
1760 // Handle special cases.
1761 if (adr_type == nullptr) return true;
1762 if (adr_type == TypePtr::BOTTOM) return true;
1763
1764 return find_alias_type(adr_type, true, nullptr) != nullptr;
1765 }
1766
1767 //-----------------------------must_alias--------------------------------------
1768 // True if all values of the given address type are in the given alias category.
1769 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1770 if (alias_idx == AliasIdxBot) return true; // the universal category
1771 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1772 if (alias_idx == AliasIdxTop) return false; // the empty category
1773 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1774
1775 // the only remaining possible overlap is identity
1776 int adr_idx = get_alias_index(adr_type);
1777 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1778 assert(adr_idx == alias_idx ||
1779 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1780 && adr_type != TypeOopPtr::BOTTOM),
1781 "should not be testing for overlap with an unsafe pointer");
1782 return adr_idx == alias_idx;
1783 }
1784
1785 //------------------------------can_alias--------------------------------------
1786 // True if any values of the given address type are in the given alias category.
1787 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1788 if (alias_idx == AliasIdxTop) return false; // the empty category
1789 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1790 // Known instance doesn't alias with bottom memory
1791 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1792 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1793
1794 // the only remaining possible overlap is identity
1795 int adr_idx = get_alias_index(adr_type);
1796 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1797 return adr_idx == alias_idx;
1798 }
1799
1800 //---------------------cleanup_loop_predicates-----------------------
1801 // Remove the opaque nodes that protect the predicates so that all unused
1802 // checks and uncommon_traps will be eliminated from the ideal graph
1803 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1804 if (predicate_count()==0) return;
1805 for (int i = predicate_count(); i > 0; i--) {
1806 Node * n = predicate_opaque1_node(i-1);
1807 assert(n->Opcode() == Op_Opaque1, "must be");
1808 igvn.replace_node(n, n->in(1));
1809 }
1810 assert(predicate_count()==0, "should be clean!");
1811 }
1812
1813 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1814 if (!n->for_post_loop_opts_igvn()) {
1815 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1816 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1817 _for_post_loop_igvn.append(n);
1818 }
1819 }
1820
1821 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1822 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1823 _for_post_loop_igvn.remove(n);
1824 }
1825
1826 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1827 // Verify that all previous optimizations produced a valid graph
1828 // at least to this point, even if no loop optimizations were done.
1829 PhaseIdealLoop::verify(igvn);
1830
1831 C->set_post_loop_opts_phase(); // no more loop opts allowed
1832
1833 assert(!C->major_progress(), "not cleared");
1834
1835 if (_for_post_loop_igvn.length() > 0) {
1836 while (_for_post_loop_igvn.length() > 0) {
1837 Node* n = _for_post_loop_igvn.pop();
1838 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1839 igvn._worklist.push(n);
1840 }
1841 igvn.optimize();
1842 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1843
1844 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1845 if (C->major_progress()) {
1846 C->clear_major_progress(); // ensure that major progress is now clear
1847 }
1848 }
1849 }
1850
1851 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1852 if (OptimizeUnstableIf) {
1853 _unstable_if_traps.append(trap);
1854 }
1855 }
1856
1857 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1858 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1859 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1860 Node* n = trap->uncommon_trap();
1861 if (!useful.member(n)) {
1862 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1863 }
1864 }
1865 }
1866
1867 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1868 // or fold-compares case. Return true if succeed or not found.
1869 //
1870 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1871 // false and ask fold-compares to yield.
1872 //
1873 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1874 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1875 // when deoptimization does happen.
1876 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1877 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1878 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1879 if (trap->uncommon_trap() == unc) {
1880 if (yield && trap->modified()) {
1881 return false;
1882 }
1883 _unstable_if_traps.delete_at(i);
1884 break;
1885 }
1886 }
1887 return true;
1888 }
1889
1890 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1891 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1892 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1893 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1894 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1895 CallStaticJavaNode* unc = trap->uncommon_trap();
1896 int next_bci = trap->next_bci();
1897 bool modified = trap->modified();
1898
1899 if (next_bci != -1 && !modified) {
1900 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1901 JVMState* jvms = unc->jvms();
1902 ciMethod* method = jvms->method();
1903 ciBytecodeStream iter(method);
1904
1905 iter.force_bci(jvms->bci());
1906 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1907 Bytecodes::Code c = iter.cur_bc();
1908 Node* lhs = nullptr;
1909 Node* rhs = nullptr;
1910 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1911 lhs = unc->peek_operand(0);
1912 rhs = unc->peek_operand(1);
1913 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1914 lhs = unc->peek_operand(0);
1915 }
1916
1917 ResourceMark rm;
1918 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1919 assert(live_locals.is_valid(), "broken liveness info");
1920 int len = (int)live_locals.size();
1921
1922 for (int i = 0; i < len; i++) {
1923 Node* local = unc->local(jvms, i);
1924 // kill local using the liveness of next_bci.
1925 // give up when the local looks like an operand to secure reexecution.
1926 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1927 uint idx = jvms->locoff() + i;
1928 #ifdef ASSERT
1929 if (Verbose) {
1930 tty->print("[unstable_if] kill local#%d: ", idx);
1931 local->dump();
1932 tty->cr();
1933 }
1934 #endif
1935 igvn.replace_input_of(unc, idx, top());
1936 modified = true;
1937 }
1938 }
1939 }
1940
1941 // keep the mondified trap for late query
1942 if (modified) {
1943 trap->set_modified();
1944 } else {
1945 _unstable_if_traps.delete_at(i);
1946 }
1947 }
1948 igvn.optimize();
1949 }
1950
1951 // StringOpts and late inlining of string methods
1952 void Compile::inline_string_calls(bool parse_time) {
1953 {
1954 // remove useless nodes to make the usage analysis simpler
1955 ResourceMark rm;
1956 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1957 }
1958
1959 {
1960 ResourceMark rm;
1961 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1962 PhaseStringOpts pso(initial_gvn());
1963 print_method(PHASE_AFTER_STRINGOPTS, 3);
1964 }
1965
1966 // now inline anything that we skipped the first time around
1967 if (!parse_time) {
1968 _late_inlines_pos = _late_inlines.length();
1969 }
1970
1971 while (_string_late_inlines.length() > 0) {
1972 CallGenerator* cg = _string_late_inlines.pop();
1973 cg->do_late_inline();
1974 if (failing()) return;
1975 }
1976 _string_late_inlines.trunc_to(0);
1977 }
1978
1979 // Late inlining of boxing methods
1980 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1981 if (_boxing_late_inlines.length() > 0) {
1982 assert(has_boxed_value(), "inconsistent");
1983
1984 PhaseGVN* gvn = initial_gvn();
1985 set_inlining_incrementally(true);
1986
1987 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1988 for_igvn()->clear();
1989 gvn->replace_with(&igvn);
1990
1991 _late_inlines_pos = _late_inlines.length();
1992
1993 while (_boxing_late_inlines.length() > 0) {
1994 CallGenerator* cg = _boxing_late_inlines.pop();
1995 cg->do_late_inline();
1996 if (failing()) return;
1997 }
1998 _boxing_late_inlines.trunc_to(0);
1999
2000 inline_incrementally_cleanup(igvn);
2001
2002 set_inlining_incrementally(false);
2003 }
2004 }
2005
2006 bool Compile::inline_incrementally_one() {
2007 assert(IncrementalInline, "incremental inlining should be on");
2008
2009 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2010
2011 set_inlining_progress(false);
2012 set_do_cleanup(false);
2013
2014 for (int i = 0; i < _late_inlines.length(); i++) {
2015 _late_inlines_pos = i+1;
2016 CallGenerator* cg = _late_inlines.at(i);
2017 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2018 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2019 cg->do_late_inline();
2020 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2021 if (failing()) {
2022 return false;
2023 } else if (inlining_progress()) {
2024 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2025 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2026 break; // process one call site at a time
2027 }
2028 } else {
2029 // Ignore late inline direct calls when inlining is not allowed.
2030 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2031 }
2032 }
2033 // Remove processed elements.
2034 _late_inlines.remove_till(_late_inlines_pos);
2035 _late_inlines_pos = 0;
2036
2037 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2038
2039 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2040
2041 set_inlining_progress(false);
2042 set_do_cleanup(false);
2043
2044 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2045 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2046 }
2047
2048 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2049 {
2050 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2051 ResourceMark rm;
2052 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2053 }
2054 {
2055 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2056 igvn = PhaseIterGVN(initial_gvn());
2057 igvn.optimize();
2058 }
2059 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2060 }
2061
2062 // Perform incremental inlining until bound on number of live nodes is reached
2063 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2064 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2065
2066 set_inlining_incrementally(true);
2067 uint low_live_nodes = 0;
2068
2069 while (_late_inlines.length() > 0) {
2070 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2071 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2072 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2073 // PhaseIdealLoop is expensive so we only try it once we are
2074 // out of live nodes and we only try it again if the previous
2075 // helped got the number of nodes down significantly
2076 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2077 if (failing()) return;
2078 low_live_nodes = live_nodes();
2079 _major_progress = true;
2080 }
2081
2082 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2083 bool do_print_inlining = print_inlining() || print_intrinsics();
2084 if (do_print_inlining || log() != nullptr) {
2085 // Print inlining message for candidates that we couldn't inline for lack of space.
2086 for (int i = 0; i < _late_inlines.length(); i++) {
2087 CallGenerator* cg = _late_inlines.at(i);
2088 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2089 if (do_print_inlining) {
2090 cg->print_inlining_late(msg);
2091 }
2092 log_late_inline_failure(cg, msg);
2093 }
2094 }
2095 break; // finish
2096 }
2097 }
2098
2099 for_igvn()->clear();
2100 initial_gvn()->replace_with(&igvn);
2101
2102 while (inline_incrementally_one()) {
2103 assert(!failing(), "inconsistent");
2104 }
2105 if (failing()) return;
2106
2107 inline_incrementally_cleanup(igvn);
2108
2109 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2110
2111 if (failing()) return;
2112
2113 if (_late_inlines.length() == 0) {
2114 break; // no more progress
2115 }
2116 }
2117 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2118
2119 if (_string_late_inlines.length() > 0) {
2120 assert(has_stringbuilder(), "inconsistent");
2121 for_igvn()->clear();
2122 initial_gvn()->replace_with(&igvn);
2123
2124 inline_string_calls(false);
2125
2126 if (failing()) return;
2127
2128 inline_incrementally_cleanup(igvn);
2129 }
2130
2131 set_inlining_incrementally(false);
2132 }
2133
2134 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2135 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2136 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2137 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2138 // as if "inlining_incrementally() == true" were set.
2139 assert(inlining_incrementally() == false, "not allowed");
2140 assert(_modified_nodes == nullptr, "not allowed");
2141 assert(_late_inlines.length() > 0, "sanity");
2142
2143 while (_late_inlines.length() > 0) {
2144 for_igvn()->clear();
2145 initial_gvn()->replace_with(&igvn);
2146
2147 while (inline_incrementally_one()) {
2148 assert(!failing(), "inconsistent");
2149 }
2150 if (failing()) return;
2151
2152 inline_incrementally_cleanup(igvn);
2153 }
2154 }
2155
2156 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2157 if (_loop_opts_cnt > 0) {
2158 while (major_progress() && (_loop_opts_cnt > 0)) {
2159 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2160 PhaseIdealLoop::optimize(igvn, mode);
2161 _loop_opts_cnt--;
2162 if (failing()) return false;
2163 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2164 }
2165 }
2166 return true;
2167 }
2168
2169 // Remove edges from "root" to each SafePoint at a backward branch.
2170 // They were inserted during parsing (see add_safepoint()) to make
2171 // infinite loops without calls or exceptions visible to root, i.e.,
2172 // useful.
2173 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2174 Node *r = root();
2175 if (r != nullptr) {
2176 for (uint i = r->req(); i < r->len(); ++i) {
2177 Node *n = r->in(i);
2178 if (n != nullptr && n->is_SafePoint()) {
2179 r->rm_prec(i);
2180 if (n->outcnt() == 0) {
2181 igvn.remove_dead_node(n);
2182 }
2183 --i;
2184 }
2185 }
2186 // Parsing may have added top inputs to the root node (Path
2187 // leading to the Halt node proven dead). Make sure we get a
2188 // chance to clean them up.
2189 igvn._worklist.push(r);
2190 igvn.optimize();
2191 }
2192 }
2193
2194 //------------------------------Optimize---------------------------------------
2195 // Given a graph, optimize it.
2196 void Compile::Optimize() {
2197 TracePhase tp("optimizer", &timers[_t_optimizer]);
2198
2199 #ifndef PRODUCT
2200 if (env()->break_at_compile()) {
2201 BREAKPOINT;
2202 }
2203
2204 #endif
2205
2206 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2207 #ifdef ASSERT
2208 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2209 #endif
2210
2211 ResourceMark rm;
2212
2213 print_inlining_reinit();
2214
2215 NOT_PRODUCT( verify_graph_edges(); )
2216
2217 print_method(PHASE_AFTER_PARSING, 1);
2218
2219 {
2220 // Iterative Global Value Numbering, including ideal transforms
2221 // Initialize IterGVN with types and values from parse-time GVN
2222 PhaseIterGVN igvn(initial_gvn());
2223 #ifdef ASSERT
2224 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2225 #endif
2226 {
2227 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2228 igvn.optimize();
2229 }
2230
2231 if (failing()) return;
2232
2233 print_method(PHASE_ITER_GVN1, 2);
2234
2235 process_for_unstable_if_traps(igvn);
2236
2237 inline_incrementally(igvn);
2238
2239 print_method(PHASE_INCREMENTAL_INLINE, 2);
2240
2241 if (failing()) return;
2242
2243 if (eliminate_boxing()) {
2244 // Inline valueOf() methods now.
2245 inline_boxing_calls(igvn);
2246
2247 if (AlwaysIncrementalInline) {
2248 inline_incrementally(igvn);
2249 }
2250
2251 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2252
2253 if (failing()) return;
2254 }
2255
2256 // Remove the speculative part of types and clean up the graph from
2257 // the extra CastPP nodes whose only purpose is to carry them. Do
2258 // that early so that optimizations are not disrupted by the extra
2259 // CastPP nodes.
2260 remove_speculative_types(igvn);
2261
2262 // No more new expensive nodes will be added to the list from here
2263 // so keep only the actual candidates for optimizations.
2264 cleanup_expensive_nodes(igvn);
2265
2266 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2267 if (EnableVectorSupport && has_vbox_nodes()) {
2268 TracePhase tp("", &timers[_t_vector]);
2269 PhaseVector pv(igvn);
2270 pv.optimize_vector_boxes();
2271
2272 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2273 }
2274 assert(!has_vbox_nodes(), "sanity");
2275
2276 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2277 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2278 initial_gvn()->replace_with(&igvn);
2279 Unique_Node_List* old_worklist = for_igvn();
2280 old_worklist->clear();
2281 Unique_Node_List new_worklist(C->comp_arena());
2282 {
2283 ResourceMark rm;
2284 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2285 }
2286 Unique_Node_List* save_for_igvn = for_igvn();
2287 set_for_igvn(&new_worklist);
2288 igvn = PhaseIterGVN(initial_gvn());
2289 igvn.optimize();
2290 set_for_igvn(old_worklist); // new_worklist is dead beyond this point
2291 }
2292
2293 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2294 // safepoints
2295 remove_root_to_sfpts_edges(igvn);
2296
2297 // Perform escape analysis
2298 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2299 if (has_loops()) {
2300 // Cleanup graph (remove dead nodes).
2301 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2302 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2303 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2304 if (failing()) return;
2305 }
2306 bool progress;
2307 do {
2308 ConnectionGraph::do_analysis(this, &igvn);
2309
2310 if (failing()) return;
2311
2312 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2313
2314 // Optimize out fields loads from scalar replaceable allocations.
2315 igvn.optimize();
2316 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2317
2318 if (failing()) return;
2319
2320 if (congraph() != nullptr && macro_count() > 0) {
2321 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2322 PhaseMacroExpand mexp(igvn);
2323 mexp.eliminate_macro_nodes();
2324 igvn.set_delay_transform(false);
2325
2326 igvn.optimize();
2327 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2328
2329 if (failing()) return;
2330 }
2331 progress = do_iterative_escape_analysis() &&
2332 (macro_count() < mcount) &&
2333 ConnectionGraph::has_candidates(this);
2334 // Try again if candidates exist and made progress
2335 // by removing some allocations and/or locks.
2336 } while (progress);
2337 }
2338
2339 // Loop transforms on the ideal graph. Range Check Elimination,
2340 // peeling, unrolling, etc.
2341
2342 // Set loop opts counter
2343 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2344 {
2345 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2346 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2347 _loop_opts_cnt--;
2348 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2349 if (failing()) return;
2350 }
2351 // Loop opts pass if partial peeling occurred in previous pass
2352 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2353 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2354 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2355 _loop_opts_cnt--;
2356 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2357 if (failing()) return;
2358 }
2359 // Loop opts pass for loop-unrolling before CCP
2360 if(major_progress() && (_loop_opts_cnt > 0)) {
2361 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2362 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2363 _loop_opts_cnt--;
2364 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2365 }
2366 if (!failing()) {
2367 // Verify that last round of loop opts produced a valid graph
2368 PhaseIdealLoop::verify(igvn);
2369 }
2370 }
2371 if (failing()) return;
2372
2373 // Conditional Constant Propagation;
2374 PhaseCCP ccp( &igvn );
2375 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2376 {
2377 TracePhase tp("ccp", &timers[_t_ccp]);
2378 ccp.do_transform();
2379 }
2380 print_method(PHASE_CCP1, 2);
2381
2382 assert( true, "Break here to ccp.dump_old2new_map()");
2383
2384 // Iterative Global Value Numbering, including ideal transforms
2385 {
2386 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2387 igvn = ccp;
2388 igvn.optimize();
2389 }
2390 print_method(PHASE_ITER_GVN2, 2);
2391
2392 if (failing()) return;
2393
2394 // Loop transforms on the ideal graph. Range Check Elimination,
2395 // peeling, unrolling, etc.
2396 if (!optimize_loops(igvn, LoopOptsDefault)) {
2397 return;
2398 }
2399
2400 if (failing()) return;
2401
2402 C->clear_major_progress(); // ensure that major progress is now clear
2403
2404 process_for_post_loop_opts_igvn(igvn);
2405
2406 #ifdef ASSERT
2407 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2408 #endif
2409
2410 {
2411 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2412 PhaseMacroExpand mex(igvn);
2413 if (mex.expand_macro_nodes()) {
2414 assert(failing(), "must bail out w/ explicit message");
2415 return;
2416 }
2417 print_method(PHASE_MACRO_EXPANSION, 2);
2418 }
2419
2420 {
2421 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2422 if (bs->expand_barriers(this, igvn)) {
2423 assert(failing(), "must bail out w/ explicit message");
2424 return;
2425 }
2426 print_method(PHASE_BARRIER_EXPANSION, 2);
2427 }
2428
2429 if (C->max_vector_size() > 0) {
2430 C->optimize_logic_cones(igvn);
2431 igvn.optimize();
2432 }
2433
2434 DEBUG_ONLY( _modified_nodes = nullptr; )
2435
2436 assert(igvn._worklist.size() == 0, "not empty");
2437
2438 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2439
2440 if (_late_inlines.length() > 0) {
2441 // More opportunities to optimize virtual and MH calls.
2442 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2443 process_late_inline_calls_no_inline(igvn);
2444 }
2445 } // (End scope of igvn; run destructor if necessary for asserts.)
2446
2447 check_no_dead_use();
2448
2449 process_print_inlining();
2450
2451 // A method with only infinite loops has no edges entering loops from root
2452 {
2453 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2454 if (final_graph_reshaping()) {
2455 assert(failing(), "must bail out w/ explicit message");
2456 return;
2457 }
2458 }
2459
2460 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2461 DEBUG_ONLY(set_phase_optimize_finished();)
2462 }
2463
2464 #ifdef ASSERT
2465 void Compile::check_no_dead_use() const {
2466 ResourceMark rm;
2467 Unique_Node_List wq;
2468 wq.push(root());
2469 for (uint i = 0; i < wq.size(); ++i) {
2470 Node* n = wq.at(i);
2471 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2472 Node* u = n->fast_out(j);
2473 if (u->outcnt() == 0 && !u->is_Con()) {
2474 u->dump();
2475 fatal("no reachable node should have no use");
2476 }
2477 wq.push(u);
2478 }
2479 }
2480 }
2481 #endif
2482
2483 void Compile::inline_vector_reboxing_calls() {
2484 if (C->_vector_reboxing_late_inlines.length() > 0) {
2485 _late_inlines_pos = C->_late_inlines.length();
2486 while (_vector_reboxing_late_inlines.length() > 0) {
2487 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2488 cg->do_late_inline();
2489 if (failing()) return;
2490 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2491 }
2492 _vector_reboxing_late_inlines.trunc_to(0);
2493 }
2494 }
2495
2496 bool Compile::has_vbox_nodes() {
2497 if (C->_vector_reboxing_late_inlines.length() > 0) {
2498 return true;
2499 }
2500 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2501 Node * n = C->macro_node(macro_idx);
2502 assert(n->is_macro(), "only macro nodes expected here");
2503 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2504 return true;
2505 }
2506 }
2507 return false;
2508 }
2509
2510 //---------------------------- Bitwise operation packing optimization ---------------------------
2511
2512 static bool is_vector_unary_bitwise_op(Node* n) {
2513 return n->Opcode() == Op_XorV &&
2514 VectorNode::is_vector_bitwise_not_pattern(n);
2515 }
2516
2517 static bool is_vector_binary_bitwise_op(Node* n) {
2518 switch (n->Opcode()) {
2519 case Op_AndV:
2520 case Op_OrV:
2521 return true;
2522
2523 case Op_XorV:
2524 return !is_vector_unary_bitwise_op(n);
2525
2526 default:
2527 return false;
2528 }
2529 }
2530
2531 static bool is_vector_ternary_bitwise_op(Node* n) {
2532 return n->Opcode() == Op_MacroLogicV;
2533 }
2534
2535 static bool is_vector_bitwise_op(Node* n) {
2536 return is_vector_unary_bitwise_op(n) ||
2537 is_vector_binary_bitwise_op(n) ||
2538 is_vector_ternary_bitwise_op(n);
2539 }
2540
2541 static bool is_vector_bitwise_cone_root(Node* n) {
2542 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2543 return false;
2544 }
2545 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2546 if (is_vector_bitwise_op(n->fast_out(i))) {
2547 return false;
2548 }
2549 }
2550 return true;
2551 }
2552
2553 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2554 uint cnt = 0;
2555 if (is_vector_bitwise_op(n)) {
2556 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2557 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2558 for (uint i = 1; i < inp_cnt; i++) {
2559 Node* in = n->in(i);
2560 bool skip = VectorNode::is_all_ones_vector(in);
2561 if (!skip && !inputs.member(in)) {
2562 inputs.push(in);
2563 cnt++;
2564 }
2565 }
2566 assert(cnt <= 1, "not unary");
2567 } else {
2568 uint last_req = inp_cnt;
2569 if (is_vector_ternary_bitwise_op(n)) {
2570 last_req = inp_cnt - 1; // skip last input
2571 }
2572 for (uint i = 1; i < last_req; i++) {
2573 Node* def = n->in(i);
2574 if (!inputs.member(def)) {
2575 inputs.push(def);
2576 cnt++;
2577 }
2578 }
2579 }
2580 } else { // not a bitwise operations
2581 if (!inputs.member(n)) {
2582 inputs.push(n);
2583 cnt++;
2584 }
2585 }
2586 return cnt;
2587 }
2588
2589 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2590 Unique_Node_List useful_nodes;
2591 C->identify_useful_nodes(useful_nodes);
2592
2593 for (uint i = 0; i < useful_nodes.size(); i++) {
2594 Node* n = useful_nodes.at(i);
2595 if (is_vector_bitwise_cone_root(n)) {
2596 list.push(n);
2597 }
2598 }
2599 }
2600
2601 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2602 const TypeVect* vt,
2603 Unique_Node_List& partition,
2604 Unique_Node_List& inputs) {
2605 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2606 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2607 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2608
2609 Node* in1 = inputs.at(0);
2610 Node* in2 = inputs.at(1);
2611 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2612
2613 uint func = compute_truth_table(partition, inputs);
2614
2615 Node* pn = partition.at(partition.size() - 1);
2616 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2617 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2618 }
2619
2620 static uint extract_bit(uint func, uint pos) {
2621 return (func & (1 << pos)) >> pos;
2622 }
2623
2624 //
2625 // A macro logic node represents a truth table. It has 4 inputs,
2626 // First three inputs corresponds to 3 columns of a truth table
2627 // and fourth input captures the logic function.
2628 //
2629 // eg. fn = (in1 AND in2) OR in3;
2630 //
2631 // MacroNode(in1,in2,in3,fn)
2632 //
2633 // -----------------
2634 // in1 in2 in3 fn
2635 // -----------------
2636 // 0 0 0 0
2637 // 0 0 1 1
2638 // 0 1 0 0
2639 // 0 1 1 1
2640 // 1 0 0 0
2641 // 1 0 1 1
2642 // 1 1 0 1
2643 // 1 1 1 1
2644 //
2645
2646 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2647 int res = 0;
2648 for (int i = 0; i < 8; i++) {
2649 int bit1 = extract_bit(in1, i);
2650 int bit2 = extract_bit(in2, i);
2651 int bit3 = extract_bit(in3, i);
2652
2653 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2654 int func_bit = extract_bit(func, func_bit_pos);
2655
2656 res |= func_bit << i;
2657 }
2658 return res;
2659 }
2660
2661 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2662 assert(n != nullptr, "");
2663 assert(eval_map.contains(n), "absent");
2664 return *(eval_map.get(n));
2665 }
2666
2667 static void eval_operands(Node* n,
2668 uint& func1, uint& func2, uint& func3,
2669 ResourceHashtable<Node*,uint>& eval_map) {
2670 assert(is_vector_bitwise_op(n), "");
2671
2672 if (is_vector_unary_bitwise_op(n)) {
2673 Node* opnd = n->in(1);
2674 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2675 opnd = n->in(2);
2676 }
2677 func1 = eval_operand(opnd, eval_map);
2678 } else if (is_vector_binary_bitwise_op(n)) {
2679 func1 = eval_operand(n->in(1), eval_map);
2680 func2 = eval_operand(n->in(2), eval_map);
2681 } else {
2682 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2683 func1 = eval_operand(n->in(1), eval_map);
2684 func2 = eval_operand(n->in(2), eval_map);
2685 func3 = eval_operand(n->in(3), eval_map);
2686 }
2687 }
2688
2689 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2690 assert(inputs.size() <= 3, "sanity");
2691 ResourceMark rm;
2692 uint res = 0;
2693 ResourceHashtable<Node*,uint> eval_map;
2694
2695 // Populate precomputed functions for inputs.
2696 // Each input corresponds to one column of 3 input truth-table.
2697 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2698 0xCC, // (_, b, _) -> b
2699 0xF0 }; // (a, _, _) -> a
2700 for (uint i = 0; i < inputs.size(); i++) {
2701 eval_map.put(inputs.at(i), input_funcs[2-i]);
2702 }
2703
2704 for (uint i = 0; i < partition.size(); i++) {
2705 Node* n = partition.at(i);
2706
2707 uint func1 = 0, func2 = 0, func3 = 0;
2708 eval_operands(n, func1, func2, func3, eval_map);
2709
2710 switch (n->Opcode()) {
2711 case Op_OrV:
2712 assert(func3 == 0, "not binary");
2713 res = func1 | func2;
2714 break;
2715 case Op_AndV:
2716 assert(func3 == 0, "not binary");
2717 res = func1 & func2;
2718 break;
2719 case Op_XorV:
2720 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2721 assert(func2 == 0 && func3 == 0, "not unary");
2722 res = (~func1) & 0xFF;
2723 } else {
2724 assert(func3 == 0, "not binary");
2725 res = func1 ^ func2;
2726 }
2727 break;
2728 case Op_MacroLogicV:
2729 // Ordering of inputs may change during evaluation of sub-tree
2730 // containing MacroLogic node as a child node, thus a re-evaluation
2731 // makes sure that function is evaluated in context of current
2732 // inputs.
2733 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2734 break;
2735
2736 default: assert(false, "not supported: %s", n->Name());
2737 }
2738 assert(res <= 0xFF, "invalid");
2739 eval_map.put(n, res);
2740 }
2741 return res;
2742 }
2743
2744 // Criteria under which nodes gets packed into a macro logic node:-
2745 // 1) Parent and both child nodes are all unmasked or masked with
2746 // same predicates.
2747 // 2) Masked parent can be packed with left child if it is predicated
2748 // and both have same predicates.
2749 // 3) Masked parent can be packed with right child if its un-predicated
2750 // or has matching predication condition.
2751 // 4) An unmasked parent can be packed with an unmasked child.
2752 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2753 assert(partition.size() == 0, "not empty");
2754 assert(inputs.size() == 0, "not empty");
2755 if (is_vector_ternary_bitwise_op(n)) {
2756 return false;
2757 }
2758
2759 bool is_unary_op = is_vector_unary_bitwise_op(n);
2760 if (is_unary_op) {
2761 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2762 return false; // too few inputs
2763 }
2764
2765 bool pack_left_child = true;
2766 bool pack_right_child = true;
2767
2768 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2769 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2770
2771 int left_child_input_cnt = 0;
2772 int right_child_input_cnt = 0;
2773
2774 bool parent_is_predicated = n->is_predicated_vector();
2775 bool left_child_predicated = n->in(1)->is_predicated_vector();
2776 bool right_child_predicated = n->in(2)->is_predicated_vector();
2777
2778 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2779 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2780 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2781
2782 do {
2783 if (pack_left_child && left_child_LOP &&
2784 ((!parent_is_predicated && !left_child_predicated) ||
2785 ((parent_is_predicated && left_child_predicated &&
2786 parent_pred == left_child_pred)))) {
2787 partition.push(n->in(1));
2788 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2789 } else {
2790 inputs.push(n->in(1));
2791 left_child_input_cnt = 1;
2792 }
2793
2794 if (pack_right_child && right_child_LOP &&
2795 (!right_child_predicated ||
2796 (right_child_predicated && parent_is_predicated &&
2797 parent_pred == right_child_pred))) {
2798 partition.push(n->in(2));
2799 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2800 } else {
2801 inputs.push(n->in(2));
2802 right_child_input_cnt = 1;
2803 }
2804
2805 if (inputs.size() > 3) {
2806 assert(partition.size() > 0, "");
2807 inputs.clear();
2808 partition.clear();
2809 if (left_child_input_cnt > right_child_input_cnt) {
2810 pack_left_child = false;
2811 } else {
2812 pack_right_child = false;
2813 }
2814 } else {
2815 break;
2816 }
2817 } while(true);
2818
2819 if(partition.size()) {
2820 partition.push(n);
2821 }
2822
2823 return (partition.size() == 2 || partition.size() == 3) &&
2824 (inputs.size() == 2 || inputs.size() == 3);
2825 }
2826
2827 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2828 assert(is_vector_bitwise_op(n), "not a root");
2829
2830 visited.set(n->_idx);
2831
2832 // 1) Do a DFS walk over the logic cone.
2833 for (uint i = 1; i < n->req(); i++) {
2834 Node* in = n->in(i);
2835 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2836 process_logic_cone_root(igvn, in, visited);
2837 }
2838 }
2839
2840 // 2) Bottom up traversal: Merge node[s] with
2841 // the parent to form macro logic node.
2842 Unique_Node_List partition;
2843 Unique_Node_List inputs;
2844 if (compute_logic_cone(n, partition, inputs)) {
2845 const TypeVect* vt = n->bottom_type()->is_vect();
2846 Node* pn = partition.at(partition.size() - 1);
2847 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2848 if (mask == nullptr ||
2849 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2850 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2851 #ifdef ASSERT
2852 if (TraceNewVectors) {
2853 tty->print("new Vector node: ");
2854 macro_logic->dump();
2855 }
2856 #endif
2857 igvn.replace_node(n, macro_logic);
2858 }
2859 }
2860 }
2861
2862 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2863 ResourceMark rm;
2864 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2865 Unique_Node_List list;
2866 collect_logic_cone_roots(list);
2867
2868 while (list.size() > 0) {
2869 Node* n = list.pop();
2870 const TypeVect* vt = n->bottom_type()->is_vect();
2871 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2872 if (supported) {
2873 VectorSet visited(comp_arena());
2874 process_logic_cone_root(igvn, n, visited);
2875 }
2876 }
2877 }
2878 }
2879
2880 //------------------------------Code_Gen---------------------------------------
2881 // Given a graph, generate code for it
2882 void Compile::Code_Gen() {
2883 if (failing()) {
2884 return;
2885 }
2886
2887 // Perform instruction selection. You might think we could reclaim Matcher
2888 // memory PDQ, but actually the Matcher is used in generating spill code.
2889 // Internals of the Matcher (including some VectorSets) must remain live
2890 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2891 // set a bit in reclaimed memory.
2892
2893 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2894 // nodes. Mapping is only valid at the root of each matched subtree.
2895 NOT_PRODUCT( verify_graph_edges(); )
2896
2897 Matcher matcher;
2898 _matcher = &matcher;
2899 {
2900 TracePhase tp("matcher", &timers[_t_matcher]);
2901 matcher.match();
2902 if (failing()) {
2903 return;
2904 }
2905 }
2906 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2907 // nodes. Mapping is only valid at the root of each matched subtree.
2908 NOT_PRODUCT( verify_graph_edges(); )
2909
2910 // If you have too many nodes, or if matching has failed, bail out
2911 check_node_count(0, "out of nodes matching instructions");
2912 if (failing()) {
2913 return;
2914 }
2915
2916 print_method(PHASE_MATCHING, 2);
2917
2918 // Build a proper-looking CFG
2919 PhaseCFG cfg(node_arena(), root(), matcher);
2920 _cfg = &cfg;
2921 {
2922 TracePhase tp("scheduler", &timers[_t_scheduler]);
2923 bool success = cfg.do_global_code_motion();
2924 if (!success) {
2925 return;
2926 }
2927
2928 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2929 NOT_PRODUCT( verify_graph_edges(); )
2930 cfg.verify();
2931 }
2932
2933 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2934 _regalloc = ®alloc;
2935 {
2936 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2937 // Perform register allocation. After Chaitin, use-def chains are
2938 // no longer accurate (at spill code) and so must be ignored.
2939 // Node->LRG->reg mappings are still accurate.
2940 _regalloc->Register_Allocate();
2941
2942 // Bail out if the allocator builds too many nodes
2943 if (failing()) {
2944 return;
2945 }
2946 }
2947
2948 // Prior to register allocation we kept empty basic blocks in case the
2949 // the allocator needed a place to spill. After register allocation we
2950 // are not adding any new instructions. If any basic block is empty, we
2951 // can now safely remove it.
2952 {
2953 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2954 cfg.remove_empty_blocks();
2955 if (do_freq_based_layout()) {
2956 PhaseBlockLayout layout(cfg);
2957 } else {
2958 cfg.set_loop_alignment();
2959 }
2960 cfg.fixup_flow();
2961 cfg.remove_unreachable_blocks();
2962 cfg.verify_dominator_tree();
2963 }
2964
2965 // Apply peephole optimizations
2966 if( OptoPeephole ) {
2967 TracePhase tp("peephole", &timers[_t_peephole]);
2968 PhasePeephole peep( _regalloc, cfg);
2969 peep.do_transform();
2970 }
2971
2972 // Do late expand if CPU requires this.
2973 if (Matcher::require_postalloc_expand) {
2974 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2975 cfg.postalloc_expand(_regalloc);
2976 }
2977
2978 // Convert Nodes to instruction bits in a buffer
2979 {
2980 TracePhase tp("output", &timers[_t_output]);
2981 PhaseOutput output;
2982 output.Output();
2983 if (failing()) return;
2984 output.install();
2985 }
2986
2987 print_method(PHASE_FINAL_CODE, 1);
2988
2989 // He's dead, Jim.
2990 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2991 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2992 }
2993
2994 //------------------------------Final_Reshape_Counts---------------------------
2995 // This class defines counters to help identify when a method
2996 // may/must be executed using hardware with only 24-bit precision.
2997 struct Final_Reshape_Counts : public StackObj {
2998 int _call_count; // count non-inlined 'common' calls
2999 int _float_count; // count float ops requiring 24-bit precision
3000 int _double_count; // count double ops requiring more precision
3001 int _java_call_count; // count non-inlined 'java' calls
3002 int _inner_loop_count; // count loops which need alignment
3003 VectorSet _visited; // Visitation flags
3004 Node_List _tests; // Set of IfNodes & PCTableNodes
3005
3006 Final_Reshape_Counts() :
3007 _call_count(0), _float_count(0), _double_count(0),
3008 _java_call_count(0), _inner_loop_count(0) { }
3009
3010 void inc_call_count () { _call_count ++; }
3011 void inc_float_count () { _float_count ++; }
3012 void inc_double_count() { _double_count++; }
3013 void inc_java_call_count() { _java_call_count++; }
3014 void inc_inner_loop_count() { _inner_loop_count++; }
3015
3016 int get_call_count () const { return _call_count ; }
3017 int get_float_count () const { return _float_count ; }
3018 int get_double_count() const { return _double_count; }
3019 int get_java_call_count() const { return _java_call_count; }
3020 int get_inner_loop_count() const { return _inner_loop_count; }
3021 };
3022
3023 // Eliminate trivially redundant StoreCMs and accumulate their
3024 // precedence edges.
3025 void Compile::eliminate_redundant_card_marks(Node* n) {
3026 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3027 if (n->in(MemNode::Address)->outcnt() > 1) {
3028 // There are multiple users of the same address so it might be
3029 // possible to eliminate some of the StoreCMs
3030 Node* mem = n->in(MemNode::Memory);
3031 Node* adr = n->in(MemNode::Address);
3032 Node* val = n->in(MemNode::ValueIn);
3033 Node* prev = n;
3034 bool done = false;
3035 // Walk the chain of StoreCMs eliminating ones that match. As
3036 // long as it's a chain of single users then the optimization is
3037 // safe. Eliminating partially redundant StoreCMs would require
3038 // cloning copies down the other paths.
3039 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3040 if (adr == mem->in(MemNode::Address) &&
3041 val == mem->in(MemNode::ValueIn)) {
3042 // redundant StoreCM
3043 if (mem->req() > MemNode::OopStore) {
3044 // Hasn't been processed by this code yet.
3045 n->add_prec(mem->in(MemNode::OopStore));
3046 } else {
3047 // Already converted to precedence edge
3048 for (uint i = mem->req(); i < mem->len(); i++) {
3049 // Accumulate any precedence edges
3050 if (mem->in(i) != nullptr) {
3051 n->add_prec(mem->in(i));
3052 }
3053 }
3054 // Everything above this point has been processed.
3055 done = true;
3056 }
3057 // Eliminate the previous StoreCM
3058 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3059 assert(mem->outcnt() == 0, "should be dead");
3060 mem->disconnect_inputs(this);
3061 } else {
3062 prev = mem;
3063 }
3064 mem = prev->in(MemNode::Memory);
3065 }
3066 }
3067 }
3068
3069 //------------------------------final_graph_reshaping_impl----------------------
3070 // Implement items 1-5 from final_graph_reshaping below.
3071 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3072
3073 if ( n->outcnt() == 0 ) return; // dead node
3074 uint nop = n->Opcode();
3075
3076 // Check for 2-input instruction with "last use" on right input.
3077 // Swap to left input. Implements item (2).
3078 if( n->req() == 3 && // two-input instruction
3079 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3080 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3081 n->in(2)->outcnt() == 1 &&// right use IS a last use
3082 !n->in(2)->is_Con() ) { // right use is not a constant
3083 // Check for commutative opcode
3084 switch( nop ) {
3085 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3086 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3087 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3088 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3089 case Op_AndL: case Op_XorL: case Op_OrL:
3090 case Op_AndI: case Op_XorI: case Op_OrI: {
3091 // Move "last use" input to left by swapping inputs
3092 n->swap_edges(1, 2);
3093 break;
3094 }
3095 default:
3096 break;
3097 }
3098 }
3099
3100 #ifdef ASSERT
3101 if( n->is_Mem() ) {
3102 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3103 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3104 // oop will be recorded in oop map if load crosses safepoint
3105 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3106 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3107 "raw memory operations should have control edge");
3108 }
3109 if (n->is_MemBar()) {
3110 MemBarNode* mb = n->as_MemBar();
3111 if (mb->trailing_store() || mb->trailing_load_store()) {
3112 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3113 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3114 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3115 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3116 } else if (mb->leading()) {
3117 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3118 }
3119 }
3120 #endif
3121 // Count FPU ops and common calls, implements item (3)
3122 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3123 if (!gc_handled) {
3124 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3125 }
3126
3127 // Collect CFG split points
3128 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3129 frc._tests.push(n);
3130 }
3131 }
3132
3133 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3134 switch( nop ) {
3135 // Count all float operations that may use FPU
3136 case Op_AddF:
3137 case Op_SubF:
3138 case Op_MulF:
3139 case Op_DivF:
3140 case Op_NegF:
3141 case Op_ModF:
3142 case Op_ConvI2F:
3143 case Op_ConF:
3144 case Op_CmpF:
3145 case Op_CmpF3:
3146 case Op_StoreF:
3147 case Op_LoadF:
3148 // case Op_ConvL2F: // longs are split into 32-bit halves
3149 frc.inc_float_count();
3150 break;
3151
3152 case Op_ConvF2D:
3153 case Op_ConvD2F:
3154 frc.inc_float_count();
3155 frc.inc_double_count();
3156 break;
3157
3158 // Count all double operations that may use FPU
3159 case Op_AddD:
3160 case Op_SubD:
3161 case Op_MulD:
3162 case Op_DivD:
3163 case Op_NegD:
3164 case Op_ModD:
3165 case Op_ConvI2D:
3166 case Op_ConvD2I:
3167 // case Op_ConvL2D: // handled by leaf call
3168 // case Op_ConvD2L: // handled by leaf call
3169 case Op_ConD:
3170 case Op_CmpD:
3171 case Op_CmpD3:
3172 case Op_StoreD:
3173 case Op_LoadD:
3174 case Op_LoadD_unaligned:
3175 frc.inc_double_count();
3176 break;
3177 case Op_Opaque1: // Remove Opaque Nodes before matching
3178 case Op_Opaque3:
3179 n->subsume_by(n->in(1), this);
3180 break;
3181 case Op_CallStaticJava:
3182 case Op_CallJava:
3183 case Op_CallDynamicJava:
3184 frc.inc_java_call_count(); // Count java call site;
3185 case Op_CallRuntime:
3186 case Op_CallLeaf:
3187 case Op_CallLeafVector:
3188 case Op_CallLeafNoFP: {
3189 assert (n->is_Call(), "");
3190 CallNode *call = n->as_Call();
3191 // Count call sites where the FP mode bit would have to be flipped.
3192 // Do not count uncommon runtime calls:
3193 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3194 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3195 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3196 frc.inc_call_count(); // Count the call site
3197 } else { // See if uncommon argument is shared
3198 Node *n = call->in(TypeFunc::Parms);
3199 int nop = n->Opcode();
3200 // Clone shared simple arguments to uncommon calls, item (1).
3201 if (n->outcnt() > 1 &&
3202 !n->is_Proj() &&
3203 nop != Op_CreateEx &&
3204 nop != Op_CheckCastPP &&
3205 nop != Op_DecodeN &&
3206 nop != Op_DecodeNKlass &&
3207 !n->is_Mem() &&
3208 !n->is_Phi()) {
3209 Node *x = n->clone();
3210 call->set_req(TypeFunc::Parms, x);
3211 }
3212 }
3213 break;
3214 }
3215
3216 case Op_StoreCM:
3217 {
3218 // Convert OopStore dependence into precedence edge
3219 Node* prec = n->in(MemNode::OopStore);
3220 n->del_req(MemNode::OopStore);
3221 n->add_prec(prec);
3222 eliminate_redundant_card_marks(n);
3223 }
3224
3225 // fall through
3226
3227 case Op_StoreB:
3228 case Op_StoreC:
3229 case Op_StoreI:
3230 case Op_StoreL:
3231 case Op_CompareAndSwapB:
3232 case Op_CompareAndSwapS:
3233 case Op_CompareAndSwapI:
3234 case Op_CompareAndSwapL:
3235 case Op_CompareAndSwapP:
3236 case Op_CompareAndSwapN:
3237 case Op_WeakCompareAndSwapB:
3238 case Op_WeakCompareAndSwapS:
3239 case Op_WeakCompareAndSwapI:
3240 case Op_WeakCompareAndSwapL:
3241 case Op_WeakCompareAndSwapP:
3242 case Op_WeakCompareAndSwapN:
3243 case Op_CompareAndExchangeB:
3244 case Op_CompareAndExchangeS:
3245 case Op_CompareAndExchangeI:
3246 case Op_CompareAndExchangeL:
3247 case Op_CompareAndExchangeP:
3248 case Op_CompareAndExchangeN:
3249 case Op_GetAndAddS:
3250 case Op_GetAndAddB:
3251 case Op_GetAndAddI:
3252 case Op_GetAndAddL:
3253 case Op_GetAndSetS:
3254 case Op_GetAndSetB:
3255 case Op_GetAndSetI:
3256 case Op_GetAndSetL:
3257 case Op_GetAndSetP:
3258 case Op_GetAndSetN:
3259 case Op_StoreP:
3260 case Op_StoreN:
3261 case Op_StoreNKlass:
3262 case Op_LoadB:
3263 case Op_LoadUB:
3264 case Op_LoadUS:
3265 case Op_LoadI:
3266 case Op_LoadKlass:
3267 case Op_LoadNKlass:
3268 case Op_LoadL:
3269 case Op_LoadL_unaligned:
3270 case Op_LoadP:
3271 case Op_LoadN:
3272 case Op_LoadRange:
3273 case Op_LoadS:
3274 break;
3275
3276 case Op_AddP: { // Assert sane base pointers
3277 Node *addp = n->in(AddPNode::Address);
3278 assert( !addp->is_AddP() ||
3279 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3280 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3281 "Base pointers must match (addp %u)", addp->_idx );
3282 #ifdef _LP64
3283 if ((UseCompressedOops || UseCompressedClassPointers) &&
3284 addp->Opcode() == Op_ConP &&
3285 addp == n->in(AddPNode::Base) &&
3286 n->in(AddPNode::Offset)->is_Con()) {
3287 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3288 // on the platform and on the compressed oops mode.
3289 // Use addressing with narrow klass to load with offset on x86.
3290 // Some platforms can use the constant pool to load ConP.
3291 // Do this transformation here since IGVN will convert ConN back to ConP.
3292 const Type* t = addp->bottom_type();
3293 bool is_oop = t->isa_oopptr() != nullptr;
3294 bool is_klass = t->isa_klassptr() != nullptr;
3295
3296 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3297 (is_klass && Matcher::const_klass_prefer_decode())) {
3298 Node* nn = nullptr;
3299
3300 int op = is_oop ? Op_ConN : Op_ConNKlass;
3301
3302 // Look for existing ConN node of the same exact type.
3303 Node* r = root();
3304 uint cnt = r->outcnt();
3305 for (uint i = 0; i < cnt; i++) {
3306 Node* m = r->raw_out(i);
3307 if (m!= nullptr && m->Opcode() == op &&
3308 m->bottom_type()->make_ptr() == t) {
3309 nn = m;
3310 break;
3311 }
3312 }
3313 if (nn != nullptr) {
3314 // Decode a narrow oop to match address
3315 // [R12 + narrow_oop_reg<<3 + offset]
3316 if (is_oop) {
3317 nn = new DecodeNNode(nn, t);
3318 } else {
3319 nn = new DecodeNKlassNode(nn, t);
3320 }
3321 // Check for succeeding AddP which uses the same Base.
3322 // Otherwise we will run into the assertion above when visiting that guy.
3323 for (uint i = 0; i < n->outcnt(); ++i) {
3324 Node *out_i = n->raw_out(i);
3325 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3326 out_i->set_req(AddPNode::Base, nn);
3327 #ifdef ASSERT
3328 for (uint j = 0; j < out_i->outcnt(); ++j) {
3329 Node *out_j = out_i->raw_out(j);
3330 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3331 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3332 }
3333 #endif
3334 }
3335 }
3336 n->set_req(AddPNode::Base, nn);
3337 n->set_req(AddPNode::Address, nn);
3338 if (addp->outcnt() == 0) {
3339 addp->disconnect_inputs(this);
3340 }
3341 }
3342 }
3343 }
3344 #endif
3345 break;
3346 }
3347
3348 case Op_CastPP: {
3349 // Remove CastPP nodes to gain more freedom during scheduling but
3350 // keep the dependency they encode as control or precedence edges
3351 // (if control is set already) on memory operations. Some CastPP
3352 // nodes don't have a control (don't carry a dependency): skip
3353 // those.
3354 if (n->in(0) != nullptr) {
3355 ResourceMark rm;
3356 Unique_Node_List wq;
3357 wq.push(n);
3358 for (uint next = 0; next < wq.size(); ++next) {
3359 Node *m = wq.at(next);
3360 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3361 Node* use = m->fast_out(i);
3362 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3363 use->ensure_control_or_add_prec(n->in(0));
3364 } else {
3365 switch(use->Opcode()) {
3366 case Op_AddP:
3367 case Op_DecodeN:
3368 case Op_DecodeNKlass:
3369 case Op_CheckCastPP:
3370 case Op_CastPP:
3371 wq.push(use);
3372 break;
3373 }
3374 }
3375 }
3376 }
3377 }
3378 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3379 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3380 Node* in1 = n->in(1);
3381 const Type* t = n->bottom_type();
3382 Node* new_in1 = in1->clone();
3383 new_in1->as_DecodeN()->set_type(t);
3384
3385 if (!Matcher::narrow_oop_use_complex_address()) {
3386 //
3387 // x86, ARM and friends can handle 2 adds in addressing mode
3388 // and Matcher can fold a DecodeN node into address by using
3389 // a narrow oop directly and do implicit null check in address:
3390 //
3391 // [R12 + narrow_oop_reg<<3 + offset]
3392 // NullCheck narrow_oop_reg
3393 //
3394 // On other platforms (Sparc) we have to keep new DecodeN node and
3395 // use it to do implicit null check in address:
3396 //
3397 // decode_not_null narrow_oop_reg, base_reg
3398 // [base_reg + offset]
3399 // NullCheck base_reg
3400 //
3401 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3402 // to keep the information to which null check the new DecodeN node
3403 // corresponds to use it as value in implicit_null_check().
3404 //
3405 new_in1->set_req(0, n->in(0));
3406 }
3407
3408 n->subsume_by(new_in1, this);
3409 if (in1->outcnt() == 0) {
3410 in1->disconnect_inputs(this);
3411 }
3412 } else {
3413 n->subsume_by(n->in(1), this);
3414 if (n->outcnt() == 0) {
3415 n->disconnect_inputs(this);
3416 }
3417 }
3418 break;
3419 }
3420 #ifdef _LP64
3421 case Op_CmpP:
3422 // Do this transformation here to preserve CmpPNode::sub() and
3423 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3424 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3425 Node* in1 = n->in(1);
3426 Node* in2 = n->in(2);
3427 if (!in1->is_DecodeNarrowPtr()) {
3428 in2 = in1;
3429 in1 = n->in(2);
3430 }
3431 assert(in1->is_DecodeNarrowPtr(), "sanity");
3432
3433 Node* new_in2 = nullptr;
3434 if (in2->is_DecodeNarrowPtr()) {
3435 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3436 new_in2 = in2->in(1);
3437 } else if (in2->Opcode() == Op_ConP) {
3438 const Type* t = in2->bottom_type();
3439 if (t == TypePtr::NULL_PTR) {
3440 assert(in1->is_DecodeN(), "compare klass to null?");
3441 // Don't convert CmpP null check into CmpN if compressed
3442 // oops implicit null check is not generated.
3443 // This will allow to generate normal oop implicit null check.
3444 if (Matcher::gen_narrow_oop_implicit_null_checks())
3445 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3446 //
3447 // This transformation together with CastPP transformation above
3448 // will generated code for implicit null checks for compressed oops.
3449 //
3450 // The original code after Optimize()
3451 //
3452 // LoadN memory, narrow_oop_reg
3453 // decode narrow_oop_reg, base_reg
3454 // CmpP base_reg, nullptr
3455 // CastPP base_reg // NotNull
3456 // Load [base_reg + offset], val_reg
3457 //
3458 // after these transformations will be
3459 //
3460 // LoadN memory, narrow_oop_reg
3461 // CmpN narrow_oop_reg, nullptr
3462 // decode_not_null narrow_oop_reg, base_reg
3463 // Load [base_reg + offset], val_reg
3464 //
3465 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3466 // since narrow oops can be used in debug info now (see the code in
3467 // final_graph_reshaping_walk()).
3468 //
3469 // At the end the code will be matched to
3470 // on x86:
3471 //
3472 // Load_narrow_oop memory, narrow_oop_reg
3473 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3474 // NullCheck narrow_oop_reg
3475 //
3476 // and on sparc:
3477 //
3478 // Load_narrow_oop memory, narrow_oop_reg
3479 // decode_not_null narrow_oop_reg, base_reg
3480 // Load [base_reg + offset], val_reg
3481 // NullCheck base_reg
3482 //
3483 } else if (t->isa_oopptr()) {
3484 new_in2 = ConNode::make(t->make_narrowoop());
3485 } else if (t->isa_klassptr()) {
3486 new_in2 = ConNode::make(t->make_narrowklass());
3487 }
3488 }
3489 if (new_in2 != nullptr) {
3490 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3491 n->subsume_by(cmpN, this);
3492 if (in1->outcnt() == 0) {
3493 in1->disconnect_inputs(this);
3494 }
3495 if (in2->outcnt() == 0) {
3496 in2->disconnect_inputs(this);
3497 }
3498 }
3499 }
3500 break;
3501
3502 case Op_DecodeN:
3503 case Op_DecodeNKlass:
3504 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3505 // DecodeN could be pinned when it can't be fold into
3506 // an address expression, see the code for Op_CastPP above.
3507 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3508 break;
3509
3510 case Op_EncodeP:
3511 case Op_EncodePKlass: {
3512 Node* in1 = n->in(1);
3513 if (in1->is_DecodeNarrowPtr()) {
3514 n->subsume_by(in1->in(1), this);
3515 } else if (in1->Opcode() == Op_ConP) {
3516 const Type* t = in1->bottom_type();
3517 if (t == TypePtr::NULL_PTR) {
3518 assert(t->isa_oopptr(), "null klass?");
3519 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3520 } else if (t->isa_oopptr()) {
3521 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3522 } else if (t->isa_klassptr()) {
3523 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3524 }
3525 }
3526 if (in1->outcnt() == 0) {
3527 in1->disconnect_inputs(this);
3528 }
3529 break;
3530 }
3531
3532 case Op_Proj: {
3533 if (OptimizeStringConcat || IncrementalInline) {
3534 ProjNode* proj = n->as_Proj();
3535 if (proj->_is_io_use) {
3536 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3537 // Separate projections were used for the exception path which
3538 // are normally removed by a late inline. If it wasn't inlined
3539 // then they will hang around and should just be replaced with
3540 // the original one. Merge them.
3541 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3542 if (non_io_proj != nullptr) {
3543 proj->subsume_by(non_io_proj , this);
3544 }
3545 }
3546 }
3547 break;
3548 }
3549
3550 case Op_Phi:
3551 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3552 // The EncodeP optimization may create Phi with the same edges
3553 // for all paths. It is not handled well by Register Allocator.
3554 Node* unique_in = n->in(1);
3555 assert(unique_in != nullptr, "");
3556 uint cnt = n->req();
3557 for (uint i = 2; i < cnt; i++) {
3558 Node* m = n->in(i);
3559 assert(m != nullptr, "");
3560 if (unique_in != m)
3561 unique_in = nullptr;
3562 }
3563 if (unique_in != nullptr) {
3564 n->subsume_by(unique_in, this);
3565 }
3566 }
3567 break;
3568
3569 #endif
3570
3571 #ifdef ASSERT
3572 case Op_CastII:
3573 // Verify that all range check dependent CastII nodes were removed.
3574 if (n->isa_CastII()->has_range_check()) {
3575 n->dump(3);
3576 assert(false, "Range check dependent CastII node was not removed");
3577 }
3578 break;
3579 #endif
3580
3581 case Op_ModI:
3582 if (UseDivMod) {
3583 // Check if a%b and a/b both exist
3584 Node* d = n->find_similar(Op_DivI);
3585 if (d) {
3586 // Replace them with a fused divmod if supported
3587 if (Matcher::has_match_rule(Op_DivModI)) {
3588 DivModINode* divmod = DivModINode::make(n);
3589 d->subsume_by(divmod->div_proj(), this);
3590 n->subsume_by(divmod->mod_proj(), this);
3591 } else {
3592 // replace a%b with a-((a/b)*b)
3593 Node* mult = new MulINode(d, d->in(2));
3594 Node* sub = new SubINode(d->in(1), mult);
3595 n->subsume_by(sub, this);
3596 }
3597 }
3598 }
3599 break;
3600
3601 case Op_ModL:
3602 if (UseDivMod) {
3603 // Check if a%b and a/b both exist
3604 Node* d = n->find_similar(Op_DivL);
3605 if (d) {
3606 // Replace them with a fused divmod if supported
3607 if (Matcher::has_match_rule(Op_DivModL)) {
3608 DivModLNode* divmod = DivModLNode::make(n);
3609 d->subsume_by(divmod->div_proj(), this);
3610 n->subsume_by(divmod->mod_proj(), this);
3611 } else {
3612 // replace a%b with a-((a/b)*b)
3613 Node* mult = new MulLNode(d, d->in(2));
3614 Node* sub = new SubLNode(d->in(1), mult);
3615 n->subsume_by(sub, this);
3616 }
3617 }
3618 }
3619 break;
3620
3621 case Op_UModI:
3622 if (UseDivMod) {
3623 // Check if a%b and a/b both exist
3624 Node* d = n->find_similar(Op_UDivI);
3625 if (d) {
3626 // Replace them with a fused unsigned divmod if supported
3627 if (Matcher::has_match_rule(Op_UDivModI)) {
3628 UDivModINode* divmod = UDivModINode::make(n);
3629 d->subsume_by(divmod->div_proj(), this);
3630 n->subsume_by(divmod->mod_proj(), this);
3631 } else {
3632 // replace a%b with a-((a/b)*b)
3633 Node* mult = new MulINode(d, d->in(2));
3634 Node* sub = new SubINode(d->in(1), mult);
3635 n->subsume_by(sub, this);
3636 }
3637 }
3638 }
3639 break;
3640
3641 case Op_UModL:
3642 if (UseDivMod) {
3643 // Check if a%b and a/b both exist
3644 Node* d = n->find_similar(Op_UDivL);
3645 if (d) {
3646 // Replace them with a fused unsigned divmod if supported
3647 if (Matcher::has_match_rule(Op_UDivModL)) {
3648 UDivModLNode* divmod = UDivModLNode::make(n);
3649 d->subsume_by(divmod->div_proj(), this);
3650 n->subsume_by(divmod->mod_proj(), this);
3651 } else {
3652 // replace a%b with a-((a/b)*b)
3653 Node* mult = new MulLNode(d, d->in(2));
3654 Node* sub = new SubLNode(d->in(1), mult);
3655 n->subsume_by(sub, this);
3656 }
3657 }
3658 }
3659 break;
3660
3661 case Op_LoadVector:
3662 case Op_StoreVector:
3663 case Op_LoadVectorGather:
3664 case Op_StoreVectorScatter:
3665 case Op_LoadVectorGatherMasked:
3666 case Op_StoreVectorScatterMasked:
3667 case Op_VectorCmpMasked:
3668 case Op_VectorMaskGen:
3669 case Op_LoadVectorMasked:
3670 case Op_StoreVectorMasked:
3671 break;
3672
3673 case Op_AddReductionVI:
3674 case Op_AddReductionVL:
3675 case Op_AddReductionVF:
3676 case Op_AddReductionVD:
3677 case Op_MulReductionVI:
3678 case Op_MulReductionVL:
3679 case Op_MulReductionVF:
3680 case Op_MulReductionVD:
3681 case Op_MinReductionV:
3682 case Op_MaxReductionV:
3683 case Op_AndReductionV:
3684 case Op_OrReductionV:
3685 case Op_XorReductionV:
3686 break;
3687
3688 case Op_PackB:
3689 case Op_PackS:
3690 case Op_PackI:
3691 case Op_PackF:
3692 case Op_PackL:
3693 case Op_PackD:
3694 if (n->req()-1 > 2) {
3695 // Replace many operand PackNodes with a binary tree for matching
3696 PackNode* p = (PackNode*) n;
3697 Node* btp = p->binary_tree_pack(1, n->req());
3698 n->subsume_by(btp, this);
3699 }
3700 break;
3701 case Op_Loop:
3702 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3703 case Op_CountedLoop:
3704 case Op_LongCountedLoop:
3705 case Op_OuterStripMinedLoop:
3706 if (n->as_Loop()->is_inner_loop()) {
3707 frc.inc_inner_loop_count();
3708 }
3709 n->as_Loop()->verify_strip_mined(0);
3710 break;
3711 case Op_LShiftI:
3712 case Op_RShiftI:
3713 case Op_URShiftI:
3714 case Op_LShiftL:
3715 case Op_RShiftL:
3716 case Op_URShiftL:
3717 if (Matcher::need_masked_shift_count) {
3718 // The cpu's shift instructions don't restrict the count to the
3719 // lower 5/6 bits. We need to do the masking ourselves.
3720 Node* in2 = n->in(2);
3721 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3722 const TypeInt* t = in2->find_int_type();
3723 if (t != nullptr && t->is_con()) {
3724 juint shift = t->get_con();
3725 if (shift > mask) { // Unsigned cmp
3726 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3727 }
3728 } else {
3729 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3730 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3731 n->set_req(2, shift);
3732 }
3733 }
3734 if (in2->outcnt() == 0) { // Remove dead node
3735 in2->disconnect_inputs(this);
3736 }
3737 }
3738 break;
3739 case Op_MemBarStoreStore:
3740 case Op_MemBarRelease:
3741 // Break the link with AllocateNode: it is no longer useful and
3742 // confuses register allocation.
3743 if (n->req() > MemBarNode::Precedent) {
3744 n->set_req(MemBarNode::Precedent, top());
3745 }
3746 break;
3747 case Op_MemBarAcquire: {
3748 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3749 // At parse time, the trailing MemBarAcquire for a volatile load
3750 // is created with an edge to the load. After optimizations,
3751 // that input may be a chain of Phis. If those phis have no
3752 // other use, then the MemBarAcquire keeps them alive and
3753 // register allocation can be confused.
3754 dead_nodes.push(n->in(MemBarNode::Precedent));
3755 n->set_req(MemBarNode::Precedent, top());
3756 }
3757 break;
3758 }
3759 case Op_Blackhole:
3760 break;
3761 case Op_RangeCheck: {
3762 RangeCheckNode* rc = n->as_RangeCheck();
3763 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3764 n->subsume_by(iff, this);
3765 frc._tests.push(iff);
3766 break;
3767 }
3768 case Op_ConvI2L: {
3769 if (!Matcher::convi2l_type_required) {
3770 // Code generation on some platforms doesn't need accurate
3771 // ConvI2L types. Widening the type can help remove redundant
3772 // address computations.
3773 n->as_Type()->set_type(TypeLong::INT);
3774 ResourceMark rm;
3775 Unique_Node_List wq;
3776 wq.push(n);
3777 for (uint next = 0; next < wq.size(); next++) {
3778 Node *m = wq.at(next);
3779
3780 for(;;) {
3781 // Loop over all nodes with identical inputs edges as m
3782 Node* k = m->find_similar(m->Opcode());
3783 if (k == nullptr) {
3784 break;
3785 }
3786 // Push their uses so we get a chance to remove node made
3787 // redundant
3788 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3789 Node* u = k->fast_out(i);
3790 if (u->Opcode() == Op_LShiftL ||
3791 u->Opcode() == Op_AddL ||
3792 u->Opcode() == Op_SubL ||
3793 u->Opcode() == Op_AddP) {
3794 wq.push(u);
3795 }
3796 }
3797 // Replace all nodes with identical edges as m with m
3798 k->subsume_by(m, this);
3799 }
3800 }
3801 }
3802 break;
3803 }
3804 case Op_CmpUL: {
3805 if (!Matcher::has_match_rule(Op_CmpUL)) {
3806 // No support for unsigned long comparisons
3807 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3808 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3809 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3810 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3811 Node* andl = new AndLNode(orl, remove_sign_mask);
3812 Node* cmp = new CmpLNode(andl, n->in(2));
3813 n->subsume_by(cmp, this);
3814 }
3815 break;
3816 }
3817 default:
3818 assert(!n->is_Call(), "");
3819 assert(!n->is_Mem(), "");
3820 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3821 break;
3822 }
3823 }
3824
3825 //------------------------------final_graph_reshaping_walk---------------------
3826 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3827 // requires that the walk visits a node's inputs before visiting the node.
3828 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3829 Unique_Node_List sfpt;
3830
3831 frc._visited.set(root->_idx); // first, mark node as visited
3832 uint cnt = root->req();
3833 Node *n = root;
3834 uint i = 0;
3835 while (true) {
3836 if (i < cnt) {
3837 // Place all non-visited non-null inputs onto stack
3838 Node* m = n->in(i);
3839 ++i;
3840 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3841 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3842 // compute worst case interpreter size in case of a deoptimization
3843 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3844
3845 sfpt.push(m);
3846 }
3847 cnt = m->req();
3848 nstack.push(n, i); // put on stack parent and next input's index
3849 n = m;
3850 i = 0;
3851 }
3852 } else {
3853 // Now do post-visit work
3854 final_graph_reshaping_impl(n, frc, dead_nodes);
3855 if (nstack.is_empty())
3856 break; // finished
3857 n = nstack.node(); // Get node from stack
3858 cnt = n->req();
3859 i = nstack.index();
3860 nstack.pop(); // Shift to the next node on stack
3861 }
3862 }
3863
3864 // Skip next transformation if compressed oops are not used.
3865 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3866 (!UseCompressedOops && !UseCompressedClassPointers))
3867 return;
3868
3869 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3870 // It could be done for an uncommon traps or any safepoints/calls
3871 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3872 while (sfpt.size() > 0) {
3873 n = sfpt.pop();
3874 JVMState *jvms = n->as_SafePoint()->jvms();
3875 assert(jvms != nullptr, "sanity");
3876 int start = jvms->debug_start();
3877 int end = n->req();
3878 bool is_uncommon = (n->is_CallStaticJava() &&
3879 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3880 for (int j = start; j < end; j++) {
3881 Node* in = n->in(j);
3882 if (in->is_DecodeNarrowPtr()) {
3883 bool safe_to_skip = true;
3884 if (!is_uncommon ) {
3885 // Is it safe to skip?
3886 for (uint i = 0; i < in->outcnt(); i++) {
3887 Node* u = in->raw_out(i);
3888 if (!u->is_SafePoint() ||
3889 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3890 safe_to_skip = false;
3891 }
3892 }
3893 }
3894 if (safe_to_skip) {
3895 n->set_req(j, in->in(1));
3896 }
3897 if (in->outcnt() == 0) {
3898 in->disconnect_inputs(this);
3899 }
3900 }
3901 }
3902 }
3903 }
3904
3905 //------------------------------final_graph_reshaping--------------------------
3906 // Final Graph Reshaping.
3907 //
3908 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3909 // and not commoned up and forced early. Must come after regular
3910 // optimizations to avoid GVN undoing the cloning. Clone constant
3911 // inputs to Loop Phis; these will be split by the allocator anyways.
3912 // Remove Opaque nodes.
3913 // (2) Move last-uses by commutative operations to the left input to encourage
3914 // Intel update-in-place two-address operations and better register usage
3915 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3916 // calls canonicalizing them back.
3917 // (3) Count the number of double-precision FP ops, single-precision FP ops
3918 // and call sites. On Intel, we can get correct rounding either by
3919 // forcing singles to memory (requires extra stores and loads after each
3920 // FP bytecode) or we can set a rounding mode bit (requires setting and
3921 // clearing the mode bit around call sites). The mode bit is only used
3922 // if the relative frequency of single FP ops to calls is low enough.
3923 // This is a key transform for SPEC mpeg_audio.
3924 // (4) Detect infinite loops; blobs of code reachable from above but not
3925 // below. Several of the Code_Gen algorithms fail on such code shapes,
3926 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3927 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3928 // Detection is by looking for IfNodes where only 1 projection is
3929 // reachable from below or CatchNodes missing some targets.
3930 // (5) Assert for insane oop offsets in debug mode.
3931
3932 bool Compile::final_graph_reshaping() {
3933 // an infinite loop may have been eliminated by the optimizer,
3934 // in which case the graph will be empty.
3935 if (root()->req() == 1) {
3936 // Do not compile method that is only a trivial infinite loop,
3937 // since the content of the loop may have been eliminated.
3938 record_method_not_compilable("trivial infinite loop");
3939 return true;
3940 }
3941
3942 // Expensive nodes have their control input set to prevent the GVN
3943 // from freely commoning them. There's no GVN beyond this point so
3944 // no need to keep the control input. We want the expensive nodes to
3945 // be freely moved to the least frequent code path by gcm.
3946 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3947 for (int i = 0; i < expensive_count(); i++) {
3948 _expensive_nodes.at(i)->set_req(0, nullptr);
3949 }
3950
3951 Final_Reshape_Counts frc;
3952
3953 // Visit everybody reachable!
3954 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3955 Node_Stack nstack(live_nodes() >> 1);
3956 Unique_Node_List dead_nodes;
3957 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
3958
3959 // Check for unreachable (from below) code (i.e., infinite loops).
3960 for( uint i = 0; i < frc._tests.size(); i++ ) {
3961 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3962 // Get number of CFG targets.
3963 // Note that PCTables include exception targets after calls.
3964 uint required_outcnt = n->required_outcnt();
3965 if (n->outcnt() != required_outcnt) {
3966 // Check for a few special cases. Rethrow Nodes never take the
3967 // 'fall-thru' path, so expected kids is 1 less.
3968 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3969 if (n->in(0)->in(0)->is_Call()) {
3970 CallNode* call = n->in(0)->in(0)->as_Call();
3971 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3972 required_outcnt--; // Rethrow always has 1 less kid
3973 } else if (call->req() > TypeFunc::Parms &&
3974 call->is_CallDynamicJava()) {
3975 // Check for null receiver. In such case, the optimizer has
3976 // detected that the virtual call will always result in a null
3977 // pointer exception. The fall-through projection of this CatchNode
3978 // will not be populated.
3979 Node* arg0 = call->in(TypeFunc::Parms);
3980 if (arg0->is_Type() &&
3981 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3982 required_outcnt--;
3983 }
3984 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
3985 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3986 // Check for illegal array length. In such case, the optimizer has
3987 // detected that the allocation attempt will always result in an
3988 // exception. There is no fall-through projection of this CatchNode .
3989 assert(call->is_CallStaticJava(), "static call expected");
3990 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
3991 uint valid_length_test_input = call->req() - 1;
3992 Node* valid_length_test = call->in(valid_length_test_input);
3993 call->del_req(valid_length_test_input);
3994 if (valid_length_test->find_int_con(1) == 0) {
3995 required_outcnt--;
3996 }
3997 dead_nodes.push(valid_length_test);
3998 assert(n->outcnt() == required_outcnt, "malformed control flow");
3999 continue;
4000 }
4001 }
4002 }
4003
4004 // Recheck with a better notion of 'required_outcnt'
4005 if (n->outcnt() != required_outcnt) {
4006 DEBUG_ONLY( n->dump_bfs(1, 0, "-"); );
4007 assert(false, "malformed control flow");
4008 record_method_not_compilable("malformed control flow");
4009 return true; // Not all targets reachable!
4010 }
4011 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4012 CallNode* call = n->in(0)->in(0)->as_Call();
4013 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4014 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4015 assert(call->is_CallStaticJava(), "static call expected");
4016 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4017 uint valid_length_test_input = call->req() - 1;
4018 dead_nodes.push(call->in(valid_length_test_input));
4019 call->del_req(valid_length_test_input); // valid length test useless now
4020 }
4021 }
4022 // Check that I actually visited all kids. Unreached kids
4023 // must be infinite loops.
4024 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4025 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4026 record_method_not_compilable("infinite loop");
4027 return true; // Found unvisited kid; must be unreach
4028 }
4029
4030 // Here so verification code in final_graph_reshaping_walk()
4031 // always see an OuterStripMinedLoopEnd
4032 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4033 IfNode* init_iff = n->as_If();
4034 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4035 n->subsume_by(iff, this);
4036 }
4037 }
4038
4039 while (dead_nodes.size() > 0) {
4040 Node* m = dead_nodes.pop();
4041 if (m->outcnt() == 0 && m != top()) {
4042 for (uint j = 0; j < m->req(); j++) {
4043 Node* in = m->in(j);
4044 if (in != nullptr) {
4045 dead_nodes.push(in);
4046 }
4047 }
4048 m->disconnect_inputs(this);
4049 }
4050 }
4051
4052 #ifdef IA32
4053 // If original bytecodes contained a mixture of floats and doubles
4054 // check if the optimizer has made it homogeneous, item (3).
4055 if (UseSSE == 0 &&
4056 frc.get_float_count() > 32 &&
4057 frc.get_double_count() == 0 &&
4058 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4059 set_24_bit_selection_and_mode(false, true);
4060 }
4061 #endif // IA32
4062
4063 set_java_calls(frc.get_java_call_count());
4064 set_inner_loops(frc.get_inner_loop_count());
4065
4066 // No infinite loops, no reason to bail out.
4067 return false;
4068 }
4069
4070 //-----------------------------too_many_traps----------------------------------
4071 // Report if there are too many traps at the current method and bci.
4072 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4073 bool Compile::too_many_traps(ciMethod* method,
4074 int bci,
4075 Deoptimization::DeoptReason reason) {
4076 ciMethodData* md = method->method_data();
4077 if (md->is_empty()) {
4078 // Assume the trap has not occurred, or that it occurred only
4079 // because of a transient condition during start-up in the interpreter.
4080 return false;
4081 }
4082 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4083 if (md->has_trap_at(bci, m, reason) != 0) {
4084 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4085 // Also, if there are multiple reasons, or if there is no per-BCI record,
4086 // assume the worst.
4087 if (log())
4088 log()->elem("observe trap='%s' count='%d'",
4089 Deoptimization::trap_reason_name(reason),
4090 md->trap_count(reason));
4091 return true;
4092 } else {
4093 // Ignore method/bci and see if there have been too many globally.
4094 return too_many_traps(reason, md);
4095 }
4096 }
4097
4098 // Less-accurate variant which does not require a method and bci.
4099 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4100 ciMethodData* logmd) {
4101 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4102 // Too many traps globally.
4103 // Note that we use cumulative trap_count, not just md->trap_count.
4104 if (log()) {
4105 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4106 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4107 Deoptimization::trap_reason_name(reason),
4108 mcount, trap_count(reason));
4109 }
4110 return true;
4111 } else {
4112 // The coast is clear.
4113 return false;
4114 }
4115 }
4116
4117 //--------------------------too_many_recompiles--------------------------------
4118 // Report if there are too many recompiles at the current method and bci.
4119 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4120 // Is not eager to return true, since this will cause the compiler to use
4121 // Action_none for a trap point, to avoid too many recompilations.
4122 bool Compile::too_many_recompiles(ciMethod* method,
4123 int bci,
4124 Deoptimization::DeoptReason reason) {
4125 ciMethodData* md = method->method_data();
4126 if (md->is_empty()) {
4127 // Assume the trap has not occurred, or that it occurred only
4128 // because of a transient condition during start-up in the interpreter.
4129 return false;
4130 }
4131 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4132 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4133 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4134 Deoptimization::DeoptReason per_bc_reason
4135 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4136 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4137 if ((per_bc_reason == Deoptimization::Reason_none
4138 || md->has_trap_at(bci, m, reason) != 0)
4139 // The trap frequency measure we care about is the recompile count:
4140 && md->trap_recompiled_at(bci, m)
4141 && md->overflow_recompile_count() >= bc_cutoff) {
4142 // Do not emit a trap here if it has already caused recompilations.
4143 // Also, if there are multiple reasons, or if there is no per-BCI record,
4144 // assume the worst.
4145 if (log())
4146 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4147 Deoptimization::trap_reason_name(reason),
4148 md->trap_count(reason),
4149 md->overflow_recompile_count());
4150 return true;
4151 } else if (trap_count(reason) != 0
4152 && decompile_count() >= m_cutoff) {
4153 // Too many recompiles globally, and we have seen this sort of trap.
4154 // Use cumulative decompile_count, not just md->decompile_count.
4155 if (log())
4156 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4157 Deoptimization::trap_reason_name(reason),
4158 md->trap_count(reason), trap_count(reason),
4159 md->decompile_count(), decompile_count());
4160 return true;
4161 } else {
4162 // The coast is clear.
4163 return false;
4164 }
4165 }
4166
4167 // Compute when not to trap. Used by matching trap based nodes and
4168 // NullCheck optimization.
4169 void Compile::set_allowed_deopt_reasons() {
4170 _allowed_reasons = 0;
4171 if (is_method_compilation()) {
4172 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4173 assert(rs < BitsPerInt, "recode bit map");
4174 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4175 _allowed_reasons |= nth_bit(rs);
4176 }
4177 }
4178 }
4179 }
4180
4181 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4182 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4183 }
4184
4185 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4186 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4187 }
4188
4189 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4190 if (holder->is_initialized()) {
4191 return false;
4192 }
4193 if (holder->is_being_initialized()) {
4194 if (accessing_method->holder() == holder) {
4195 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4196 // <init>, or a static method. In all those cases, there was an initialization
4197 // barrier on the holder klass passed.
4198 if (accessing_method->is_static_initializer() ||
4199 accessing_method->is_object_initializer() ||
4200 accessing_method->is_static()) {
4201 return false;
4202 }
4203 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4204 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4205 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4206 // child class can become fully initialized while its parent class is still being initialized.
4207 if (accessing_method->is_static_initializer()) {
4208 return false;
4209 }
4210 }
4211 ciMethod* root = method(); // the root method of compilation
4212 if (root != accessing_method) {
4213 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4214 }
4215 }
4216 return true;
4217 }
4218
4219 #ifndef PRODUCT
4220 //------------------------------verify_bidirectional_edges---------------------
4221 // For each input edge to a node (ie - for each Use-Def edge), verify that
4222 // there is a corresponding Def-Use edge.
4223 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4224 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4225 uint stack_size = live_nodes() >> 4;
4226 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4227 nstack.push(_root);
4228
4229 while (nstack.size() > 0) {
4230 Node* n = nstack.pop();
4231 if (visited.member(n)) {
4232 continue;
4233 }
4234 visited.push(n);
4235
4236 // Walk over all input edges, checking for correspondence
4237 uint length = n->len();
4238 for (uint i = 0; i < length; i++) {
4239 Node* in = n->in(i);
4240 if (in != nullptr && !visited.member(in)) {
4241 nstack.push(in); // Put it on stack
4242 }
4243 if (in != nullptr && !in->is_top()) {
4244 // Count instances of `next`
4245 int cnt = 0;
4246 for (uint idx = 0; idx < in->_outcnt; idx++) {
4247 if (in->_out[idx] == n) {
4248 cnt++;
4249 }
4250 }
4251 assert(cnt > 0, "Failed to find Def-Use edge.");
4252 // Check for duplicate edges
4253 // walk the input array downcounting the input edges to n
4254 for (uint j = 0; j < length; j++) {
4255 if (n->in(j) == in) {
4256 cnt--;
4257 }
4258 }
4259 assert(cnt == 0, "Mismatched edge count.");
4260 } else if (in == nullptr) {
4261 assert(i == 0 || i >= n->req() ||
4262 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4263 (n->is_Unlock() && i == (n->req() - 1)) ||
4264 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4265 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4266 } else {
4267 assert(in->is_top(), "sanity");
4268 // Nothing to check.
4269 }
4270 }
4271 }
4272 }
4273
4274 //------------------------------verify_graph_edges---------------------------
4275 // Walk the Graph and verify that there is a one-to-one correspondence
4276 // between Use-Def edges and Def-Use edges in the graph.
4277 void Compile::verify_graph_edges(bool no_dead_code) {
4278 if (VerifyGraphEdges) {
4279 Unique_Node_List visited;
4280
4281 // Call graph walk to check edges
4282 verify_bidirectional_edges(visited);
4283 if (no_dead_code) {
4284 // Now make sure that no visited node is used by an unvisited node.
4285 bool dead_nodes = false;
4286 Unique_Node_List checked;
4287 while (visited.size() > 0) {
4288 Node* n = visited.pop();
4289 checked.push(n);
4290 for (uint i = 0; i < n->outcnt(); i++) {
4291 Node* use = n->raw_out(i);
4292 if (checked.member(use)) continue; // already checked
4293 if (visited.member(use)) continue; // already in the graph
4294 if (use->is_Con()) continue; // a dead ConNode is OK
4295 // At this point, we have found a dead node which is DU-reachable.
4296 if (!dead_nodes) {
4297 tty->print_cr("*** Dead nodes reachable via DU edges:");
4298 dead_nodes = true;
4299 }
4300 use->dump(2);
4301 tty->print_cr("---");
4302 checked.push(use); // No repeats; pretend it is now checked.
4303 }
4304 }
4305 assert(!dead_nodes, "using nodes must be reachable from root");
4306 }
4307 }
4308 }
4309 #endif
4310
4311 // The Compile object keeps track of failure reasons separately from the ciEnv.
4312 // This is required because there is not quite a 1-1 relation between the
4313 // ciEnv and its compilation task and the Compile object. Note that one
4314 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4315 // to backtrack and retry without subsuming loads. Other than this backtracking
4316 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4317 // by the logic in C2Compiler.
4318 void Compile::record_failure(const char* reason) {
4319 if (log() != nullptr) {
4320 log()->elem("failure reason='%s' phase='compile'", reason);
4321 }
4322 if (_failure_reason == nullptr) {
4323 // Record the first failure reason.
4324 _failure_reason = reason;
4325 }
4326
4327 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4328 C->print_method(PHASE_FAILURE, 1);
4329 }
4330 _root = nullptr; // flush the graph, too
4331 }
4332
4333 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4334 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4335 _phase_name(name), _dolog(CITimeVerbose)
4336 {
4337 if (_dolog) {
4338 C = Compile::current();
4339 _log = C->log();
4340 } else {
4341 C = nullptr;
4342 _log = nullptr;
4343 }
4344 if (_log != nullptr) {
4345 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4346 _log->stamp();
4347 _log->end_head();
4348 }
4349 }
4350
4351 Compile::TracePhase::~TracePhase() {
4352
4353 C = Compile::current();
4354 if (_dolog) {
4355 _log = C->log();
4356 } else {
4357 _log = nullptr;
4358 }
4359
4360 #ifdef ASSERT
4361 if (PrintIdealNodeCount) {
4362 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4363 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4364 }
4365
4366 if (VerifyIdealNodeCount) {
4367 Compile::current()->print_missing_nodes();
4368 }
4369 #endif
4370
4371 if (_log != nullptr) {
4372 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4373 }
4374 }
4375
4376 //----------------------------static_subtype_check-----------------------------
4377 // Shortcut important common cases when superklass is exact:
4378 // (0) superklass is java.lang.Object (can occur in reflective code)
4379 // (1) subklass is already limited to a subtype of superklass => always ok
4380 // (2) subklass does not overlap with superklass => always fail
4381 // (3) superklass has NO subtypes and we can check with a simple compare.
4382 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4383 if (skip) {
4384 return SSC_full_test; // Let caller generate the general case.
4385 }
4386
4387 if (subk->is_java_subtype_of(superk)) {
4388 return SSC_always_true; // (0) and (1) this test cannot fail
4389 }
4390
4391 if (!subk->maybe_java_subtype_of(superk)) {
4392 return SSC_always_false; // (2) true path dead; no dynamic test needed
4393 }
4394
4395 const Type* superelem = superk;
4396 if (superk->isa_aryklassptr()) {
4397 int ignored;
4398 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4399 }
4400
4401 if (superelem->isa_instklassptr()) {
4402 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4403 if (!ik->has_subklass()) {
4404 if (!ik->is_final()) {
4405 // Add a dependency if there is a chance of a later subclass.
4406 dependencies()->assert_leaf_type(ik);
4407 }
4408 if (!superk->maybe_java_subtype_of(subk)) {
4409 return SSC_always_false;
4410 }
4411 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4412 }
4413 } else {
4414 // A primitive array type has no subtypes.
4415 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4416 }
4417
4418 return SSC_full_test;
4419 }
4420
4421 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4422 #ifdef _LP64
4423 // The scaled index operand to AddP must be a clean 64-bit value.
4424 // Java allows a 32-bit int to be incremented to a negative
4425 // value, which appears in a 64-bit register as a large
4426 // positive number. Using that large positive number as an
4427 // operand in pointer arithmetic has bad consequences.
4428 // On the other hand, 32-bit overflow is rare, and the possibility
4429 // can often be excluded, if we annotate the ConvI2L node with
4430 // a type assertion that its value is known to be a small positive
4431 // number. (The prior range check has ensured this.)
4432 // This assertion is used by ConvI2LNode::Ideal.
4433 int index_max = max_jint - 1; // array size is max_jint, index is one less
4434 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4435 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4436 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4437 #endif
4438 return idx;
4439 }
4440
4441 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4442 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4443 if (ctrl != nullptr) {
4444 // Express control dependency by a CastII node with a narrow type.
4445 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4446 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4447 // node from floating above the range check during loop optimizations. Otherwise, the
4448 // ConvI2L node may be eliminated independently of the range check, causing the data path
4449 // to become TOP while the control path is still there (although it's unreachable).
4450 value->set_req(0, ctrl);
4451 value = phase->transform(value);
4452 }
4453 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4454 return phase->transform(new ConvI2LNode(value, ltype));
4455 }
4456
4457 // The message about the current inlining is accumulated in
4458 // _print_inlining_stream and transferred into the _print_inlining_list
4459 // once we know whether inlining succeeds or not. For regular
4460 // inlining, messages are appended to the buffer pointed by
4461 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4462 // a new buffer is added after _print_inlining_idx in the list. This
4463 // way we can update the inlining message for late inlining call site
4464 // when the inlining is attempted again.
4465 void Compile::print_inlining_init() {
4466 if (print_inlining() || print_intrinsics()) {
4467 // print_inlining_init is actually called several times.
4468 print_inlining_reset();
4469 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4470 }
4471 }
4472
4473 void Compile::print_inlining_reinit() {
4474 if (print_inlining() || print_intrinsics()) {
4475 print_inlining_reset();
4476 }
4477 }
4478
4479 void Compile::print_inlining_reset() {
4480 _print_inlining_stream->reset();
4481 }
4482
4483 void Compile::print_inlining_commit() {
4484 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4485 // Transfer the message from _print_inlining_stream to the current
4486 // _print_inlining_list buffer and clear _print_inlining_stream.
4487 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4488 print_inlining_reset();
4489 }
4490
4491 void Compile::print_inlining_push() {
4492 // Add new buffer to the _print_inlining_list at current position
4493 _print_inlining_idx++;
4494 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4495 }
4496
4497 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4498 return _print_inlining_list->at(_print_inlining_idx);
4499 }
4500
4501 void Compile::print_inlining_update(CallGenerator* cg) {
4502 if (print_inlining() || print_intrinsics()) {
4503 if (cg->is_late_inline()) {
4504 if (print_inlining_current()->cg() != cg &&
4505 (print_inlining_current()->cg() != nullptr ||
4506 print_inlining_current()->ss()->size() != 0)) {
4507 print_inlining_push();
4508 }
4509 print_inlining_commit();
4510 print_inlining_current()->set_cg(cg);
4511 } else {
4512 if (print_inlining_current()->cg() != nullptr) {
4513 print_inlining_push();
4514 }
4515 print_inlining_commit();
4516 }
4517 }
4518 }
4519
4520 void Compile::print_inlining_move_to(CallGenerator* cg) {
4521 // We resume inlining at a late inlining call site. Locate the
4522 // corresponding inlining buffer so that we can update it.
4523 if (print_inlining() || print_intrinsics()) {
4524 for (int i = 0; i < _print_inlining_list->length(); i++) {
4525 if (_print_inlining_list->at(i)->cg() == cg) {
4526 _print_inlining_idx = i;
4527 return;
4528 }
4529 }
4530 ShouldNotReachHere();
4531 }
4532 }
4533
4534 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4535 if (print_inlining() || print_intrinsics()) {
4536 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4537 assert(print_inlining_current()->cg() == cg, "wrong entry");
4538 // replace message with new message
4539 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4540 print_inlining_commit();
4541 print_inlining_current()->set_cg(cg);
4542 }
4543 }
4544
4545 void Compile::print_inlining_assert_ready() {
4546 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4547 }
4548
4549 void Compile::process_print_inlining() {
4550 assert(_late_inlines.length() == 0, "not drained yet");
4551 if (print_inlining() || print_intrinsics()) {
4552 ResourceMark rm;
4553 stringStream ss;
4554 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4555 for (int i = 0; i < _print_inlining_list->length(); i++) {
4556 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4557 ss.print("%s", pib->ss()->freeze());
4558 delete pib;
4559 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4560 }
4561 // Reset _print_inlining_list, it only contains destructed objects.
4562 // It is on the arena, so it will be freed when the arena is reset.
4563 _print_inlining_list = nullptr;
4564 // _print_inlining_stream won't be used anymore, either.
4565 print_inlining_reset();
4566 size_t end = ss.size();
4567 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4568 strncpy(_print_inlining_output, ss.freeze(), end+1);
4569 _print_inlining_output[end] = 0;
4570 }
4571 }
4572
4573 void Compile::dump_print_inlining() {
4574 if (_print_inlining_output != nullptr) {
4575 tty->print_raw(_print_inlining_output);
4576 }
4577 }
4578
4579 void Compile::log_late_inline(CallGenerator* cg) {
4580 if (log() != nullptr) {
4581 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4582 cg->unique_id());
4583 JVMState* p = cg->call_node()->jvms();
4584 while (p != nullptr) {
4585 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4586 p = p->caller();
4587 }
4588 log()->tail("late_inline");
4589 }
4590 }
4591
4592 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4593 log_late_inline(cg);
4594 if (log() != nullptr) {
4595 log()->inline_fail(msg);
4596 }
4597 }
4598
4599 void Compile::log_inline_id(CallGenerator* cg) {
4600 if (log() != nullptr) {
4601 // The LogCompilation tool needs a unique way to identify late
4602 // inline call sites. This id must be unique for this call site in
4603 // this compilation. Try to have it unique across compilations as
4604 // well because it can be convenient when grepping through the log
4605 // file.
4606 // Distinguish OSR compilations from others in case CICountOSR is
4607 // on.
4608 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4609 cg->set_unique_id(id);
4610 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4611 }
4612 }
4613
4614 void Compile::log_inline_failure(const char* msg) {
4615 if (C->log() != nullptr) {
4616 C->log()->inline_fail(msg);
4617 }
4618 }
4619
4620
4621 // Dump inlining replay data to the stream.
4622 // Don't change thread state and acquire any locks.
4623 void Compile::dump_inline_data(outputStream* out) {
4624 InlineTree* inl_tree = ilt();
4625 if (inl_tree != nullptr) {
4626 out->print(" inline %d", inl_tree->count());
4627 inl_tree->dump_replay_data(out);
4628 }
4629 }
4630
4631 void Compile::dump_inline_data_reduced(outputStream* out) {
4632 assert(ReplayReduce, "");
4633
4634 InlineTree* inl_tree = ilt();
4635 if (inl_tree == nullptr) {
4636 return;
4637 }
4638 // Enable iterative replay file reduction
4639 // Output "compile" lines for depth 1 subtrees,
4640 // simulating that those trees were compiled
4641 // instead of inlined.
4642 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4643 InlineTree* sub = inl_tree->subtrees().at(i);
4644 if (sub->inline_level() != 1) {
4645 continue;
4646 }
4647
4648 ciMethod* method = sub->method();
4649 int entry_bci = -1;
4650 int comp_level = env()->task()->comp_level();
4651 out->print("compile ");
4652 method->dump_name_as_ascii(out);
4653 out->print(" %d %d", entry_bci, comp_level);
4654 out->print(" inline %d", sub->count());
4655 sub->dump_replay_data(out, -1);
4656 out->cr();
4657 }
4658 }
4659
4660 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4661 if (n1->Opcode() < n2->Opcode()) return -1;
4662 else if (n1->Opcode() > n2->Opcode()) return 1;
4663
4664 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4665 for (uint i = 1; i < n1->req(); i++) {
4666 if (n1->in(i) < n2->in(i)) return -1;
4667 else if (n1->in(i) > n2->in(i)) return 1;
4668 }
4669
4670 return 0;
4671 }
4672
4673 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4674 Node* n1 = *n1p;
4675 Node* n2 = *n2p;
4676
4677 return cmp_expensive_nodes(n1, n2);
4678 }
4679
4680 void Compile::sort_expensive_nodes() {
4681 if (!expensive_nodes_sorted()) {
4682 _expensive_nodes.sort(cmp_expensive_nodes);
4683 }
4684 }
4685
4686 bool Compile::expensive_nodes_sorted() const {
4687 for (int i = 1; i < _expensive_nodes.length(); i++) {
4688 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4689 return false;
4690 }
4691 }
4692 return true;
4693 }
4694
4695 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4696 if (_expensive_nodes.length() == 0) {
4697 return false;
4698 }
4699
4700 assert(OptimizeExpensiveOps, "optimization off?");
4701
4702 // Take this opportunity to remove dead nodes from the list
4703 int j = 0;
4704 for (int i = 0; i < _expensive_nodes.length(); i++) {
4705 Node* n = _expensive_nodes.at(i);
4706 if (!n->is_unreachable(igvn)) {
4707 assert(n->is_expensive(), "should be expensive");
4708 _expensive_nodes.at_put(j, n);
4709 j++;
4710 }
4711 }
4712 _expensive_nodes.trunc_to(j);
4713
4714 // Then sort the list so that similar nodes are next to each other
4715 // and check for at least two nodes of identical kind with same data
4716 // inputs.
4717 sort_expensive_nodes();
4718
4719 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4720 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4721 return true;
4722 }
4723 }
4724
4725 return false;
4726 }
4727
4728 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4729 if (_expensive_nodes.length() == 0) {
4730 return;
4731 }
4732
4733 assert(OptimizeExpensiveOps, "optimization off?");
4734
4735 // Sort to bring similar nodes next to each other and clear the
4736 // control input of nodes for which there's only a single copy.
4737 sort_expensive_nodes();
4738
4739 int j = 0;
4740 int identical = 0;
4741 int i = 0;
4742 bool modified = false;
4743 for (; i < _expensive_nodes.length()-1; i++) {
4744 assert(j <= i, "can't write beyond current index");
4745 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4746 identical++;
4747 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4748 continue;
4749 }
4750 if (identical > 0) {
4751 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4752 identical = 0;
4753 } else {
4754 Node* n = _expensive_nodes.at(i);
4755 igvn.replace_input_of(n, 0, nullptr);
4756 igvn.hash_insert(n);
4757 modified = true;
4758 }
4759 }
4760 if (identical > 0) {
4761 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4762 } else if (_expensive_nodes.length() >= 1) {
4763 Node* n = _expensive_nodes.at(i);
4764 igvn.replace_input_of(n, 0, nullptr);
4765 igvn.hash_insert(n);
4766 modified = true;
4767 }
4768 _expensive_nodes.trunc_to(j);
4769 if (modified) {
4770 igvn.optimize();
4771 }
4772 }
4773
4774 void Compile::add_expensive_node(Node * n) {
4775 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4776 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4777 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4778 if (OptimizeExpensiveOps) {
4779 _expensive_nodes.append(n);
4780 } else {
4781 // Clear control input and let IGVN optimize expensive nodes if
4782 // OptimizeExpensiveOps is off.
4783 n->set_req(0, nullptr);
4784 }
4785 }
4786
4787 /**
4788 * Track coarsened Lock and Unlock nodes.
4789 */
4790
4791 class Lock_List : public Node_List {
4792 uint _origin_cnt;
4793 public:
4794 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4795 uint origin_cnt() const { return _origin_cnt; }
4796 };
4797
4798 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4799 int length = locks.length();
4800 if (length > 0) {
4801 // Have to keep this list until locks elimination during Macro nodes elimination.
4802 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4803 for (int i = 0; i < length; i++) {
4804 AbstractLockNode* lock = locks.at(i);
4805 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4806 locks_list->push(lock);
4807 }
4808 _coarsened_locks.append(locks_list);
4809 }
4810 }
4811
4812 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4813 int count = coarsened_count();
4814 for (int i = 0; i < count; i++) {
4815 Node_List* locks_list = _coarsened_locks.at(i);
4816 for (uint j = 0; j < locks_list->size(); j++) {
4817 Node* lock = locks_list->at(j);
4818 assert(lock->is_AbstractLock(), "sanity");
4819 if (!useful.member(lock)) {
4820 locks_list->yank(lock);
4821 }
4822 }
4823 }
4824 }
4825
4826 void Compile::remove_coarsened_lock(Node* n) {
4827 if (n->is_AbstractLock()) {
4828 int count = coarsened_count();
4829 for (int i = 0; i < count; i++) {
4830 Node_List* locks_list = _coarsened_locks.at(i);
4831 locks_list->yank(n);
4832 }
4833 }
4834 }
4835
4836 bool Compile::coarsened_locks_consistent() {
4837 int count = coarsened_count();
4838 for (int i = 0; i < count; i++) {
4839 bool unbalanced = false;
4840 bool modified = false; // track locks kind modifications
4841 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4842 uint size = locks_list->size();
4843 if (size == 0) {
4844 unbalanced = false; // All locks were eliminated - good
4845 } else if (size != locks_list->origin_cnt()) {
4846 unbalanced = true; // Some locks were removed from list
4847 } else {
4848 for (uint j = 0; j < size; j++) {
4849 Node* lock = locks_list->at(j);
4850 // All nodes in group should have the same state (modified or not)
4851 if (!lock->as_AbstractLock()->is_coarsened()) {
4852 if (j == 0) {
4853 // first on list was modified, the rest should be too for consistency
4854 modified = true;
4855 } else if (!modified) {
4856 // this lock was modified but previous locks on the list were not
4857 unbalanced = true;
4858 break;
4859 }
4860 } else if (modified) {
4861 // previous locks on list were modified but not this lock
4862 unbalanced = true;
4863 break;
4864 }
4865 }
4866 }
4867 if (unbalanced) {
4868 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4869 #ifdef ASSERT
4870 if (PrintEliminateLocks) {
4871 tty->print_cr("=== unbalanced coarsened locks ===");
4872 for (uint l = 0; l < size; l++) {
4873 locks_list->at(l)->dump();
4874 }
4875 }
4876 #endif
4877 record_failure(C2Compiler::retry_no_locks_coarsening());
4878 return false;
4879 }
4880 }
4881 return true;
4882 }
4883
4884 /**
4885 * Remove the speculative part of types and clean up the graph
4886 */
4887 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4888 if (UseTypeSpeculation) {
4889 Unique_Node_List worklist;
4890 worklist.push(root());
4891 int modified = 0;
4892 // Go over all type nodes that carry a speculative type, drop the
4893 // speculative part of the type and enqueue the node for an igvn
4894 // which may optimize it out.
4895 for (uint next = 0; next < worklist.size(); ++next) {
4896 Node *n = worklist.at(next);
4897 if (n->is_Type()) {
4898 TypeNode* tn = n->as_Type();
4899 const Type* t = tn->type();
4900 const Type* t_no_spec = t->remove_speculative();
4901 if (t_no_spec != t) {
4902 bool in_hash = igvn.hash_delete(n);
4903 assert(in_hash, "node should be in igvn hash table");
4904 tn->set_type(t_no_spec);
4905 igvn.hash_insert(n);
4906 igvn._worklist.push(n); // give it a chance to go away
4907 modified++;
4908 }
4909 }
4910 // Iterate over outs - endless loops is unreachable from below
4911 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4912 Node *m = n->fast_out(i);
4913 if (not_a_node(m)) {
4914 continue;
4915 }
4916 worklist.push(m);
4917 }
4918 }
4919 // Drop the speculative part of all types in the igvn's type table
4920 igvn.remove_speculative_types();
4921 if (modified > 0) {
4922 igvn.optimize();
4923 }
4924 #ifdef ASSERT
4925 // Verify that after the IGVN is over no speculative type has resurfaced
4926 worklist.clear();
4927 worklist.push(root());
4928 for (uint next = 0; next < worklist.size(); ++next) {
4929 Node *n = worklist.at(next);
4930 const Type* t = igvn.type_or_null(n);
4931 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4932 if (n->is_Type()) {
4933 t = n->as_Type()->type();
4934 assert(t == t->remove_speculative(), "no more speculative types");
4935 }
4936 // Iterate over outs - endless loops is unreachable from below
4937 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4938 Node *m = n->fast_out(i);
4939 if (not_a_node(m)) {
4940 continue;
4941 }
4942 worklist.push(m);
4943 }
4944 }
4945 igvn.check_no_speculative_types();
4946 #endif
4947 }
4948 }
4949
4950 // Auxiliary methods to support randomized stressing/fuzzing.
4951
4952 int Compile::random() {
4953 _stress_seed = os::next_random(_stress_seed);
4954 return static_cast<int>(_stress_seed);
4955 }
4956
4957 // This method can be called the arbitrary number of times, with current count
4958 // as the argument. The logic allows selecting a single candidate from the
4959 // running list of candidates as follows:
4960 // int count = 0;
4961 // Cand* selected = null;
4962 // while(cand = cand->next()) {
4963 // if (randomized_select(++count)) {
4964 // selected = cand;
4965 // }
4966 // }
4967 //
4968 // Including count equalizes the chances any candidate is "selected".
4969 // This is useful when we don't have the complete list of candidates to choose
4970 // from uniformly. In this case, we need to adjust the randomicity of the
4971 // selection, or else we will end up biasing the selection towards the latter
4972 // candidates.
4973 //
4974 // Quick back-envelope calculation shows that for the list of n candidates
4975 // the equal probability for the candidate to persist as "best" can be
4976 // achieved by replacing it with "next" k-th candidate with the probability
4977 // of 1/k. It can be easily shown that by the end of the run, the
4978 // probability for any candidate is converged to 1/n, thus giving the
4979 // uniform distribution among all the candidates.
4980 //
4981 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4982 #define RANDOMIZED_DOMAIN_POW 29
4983 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4984 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4985 bool Compile::randomized_select(int count) {
4986 assert(count > 0, "only positive");
4987 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4988 }
4989
4990 CloneMap& Compile::clone_map() { return _clone_map; }
4991 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4992
4993 void NodeCloneInfo::dump() const {
4994 tty->print(" {%d:%d} ", idx(), gen());
4995 }
4996
4997 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4998 uint64_t val = value(old->_idx);
4999 NodeCloneInfo cio(val);
5000 assert(val != 0, "old node should be in the map");
5001 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5002 insert(nnn->_idx, cin.get());
5003 #ifndef PRODUCT
5004 if (is_debug()) {
5005 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5006 }
5007 #endif
5008 }
5009
5010 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5011 NodeCloneInfo cio(value(old->_idx));
5012 if (cio.get() == 0) {
5013 cio.set(old->_idx, 0);
5014 insert(old->_idx, cio.get());
5015 #ifndef PRODUCT
5016 if (is_debug()) {
5017 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5018 }
5019 #endif
5020 }
5021 clone(old, nnn, gen);
5022 }
5023
5024 int CloneMap::max_gen() const {
5025 int g = 0;
5026 DictI di(_dict);
5027 for(; di.test(); ++di) {
5028 int t = gen(di._key);
5029 if (g < t) {
5030 g = t;
5031 #ifndef PRODUCT
5032 if (is_debug()) {
5033 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5034 }
5035 #endif
5036 }
5037 }
5038 return g;
5039 }
5040
5041 void CloneMap::dump(node_idx_t key) const {
5042 uint64_t val = value(key);
5043 if (val != 0) {
5044 NodeCloneInfo ni(val);
5045 ni.dump();
5046 }
5047 }
5048
5049 // Move Allocate nodes to the start of the list
5050 void Compile::sort_macro_nodes() {
5051 int count = macro_count();
5052 int allocates = 0;
5053 for (int i = 0; i < count; i++) {
5054 Node* n = macro_node(i);
5055 if (n->is_Allocate()) {
5056 if (i != allocates) {
5057 Node* tmp = macro_node(allocates);
5058 _macro_nodes.at_put(allocates, n);
5059 _macro_nodes.at_put(i, tmp);
5060 }
5061 allocates++;
5062 }
5063 }
5064 }
5065
5066 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5067 EventCompilerPhase event;
5068 if (event.should_commit()) {
5069 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5070 }
5071 #ifndef PRODUCT
5072 ResourceMark rm;
5073 stringStream ss;
5074 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5075 if (n != nullptr) {
5076 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5077 }
5078
5079 const char* name = ss.as_string();
5080 if (should_print_igv(level)) {
5081 _igv_printer->print_method(name, level);
5082 }
5083 if (should_print_phase(cpt)) {
5084 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5085 }
5086 #endif
5087 C->_latest_stage_start_counter.stamp();
5088 }
5089
5090 // Only used from CompileWrapper
5091 void Compile::begin_method() {
5092 #ifndef PRODUCT
5093 if (_method != nullptr && should_print_igv(1)) {
5094 _igv_printer->begin_method();
5095 }
5096 #endif
5097 C->_latest_stage_start_counter.stamp();
5098 }
5099
5100 // Only used from CompileWrapper
5101 void Compile::end_method() {
5102 EventCompilerPhase event;
5103 if (event.should_commit()) {
5104 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5105 }
5106
5107 #ifndef PRODUCT
5108 if (_method != nullptr && should_print_igv(1)) {
5109 _igv_printer->end_method();
5110 }
5111 #endif
5112 }
5113
5114 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5115 #ifndef PRODUCT
5116 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) {
5117 return true;
5118 }
5119 #endif
5120 return false;
5121 }
5122
5123 bool Compile::should_print_igv(int level) {
5124 #ifndef PRODUCT
5125 if (PrintIdealGraphLevel < 0) { // disabled by the user
5126 return false;
5127 }
5128
5129 bool need = directive()->IGVPrintLevelOption >= level;
5130 if (need && !_igv_printer) {
5131 _igv_printer = IdealGraphPrinter::printer();
5132 _igv_printer->set_compile(this);
5133 }
5134 return need;
5135 #else
5136 return false;
5137 #endif
5138 }
5139
5140 #ifndef PRODUCT
5141 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5142 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5143
5144 // Called from debugger. Prints method to the default file with the default phase name.
5145 // This works regardless of any Ideal Graph Visualizer flags set or not.
5146 void igv_print() {
5147 Compile::current()->igv_print_method_to_file();
5148 }
5149
5150 // Same as igv_print() above but with a specified phase name.
5151 void igv_print(const char* phase_name) {
5152 Compile::current()->igv_print_method_to_file(phase_name);
5153 }
5154
5155 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5156 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5157 // This works regardless of any Ideal Graph Visualizer flags set or not.
5158 void igv_print(bool network) {
5159 if (network) {
5160 Compile::current()->igv_print_method_to_network();
5161 } else {
5162 Compile::current()->igv_print_method_to_file();
5163 }
5164 }
5165
5166 // Same as igv_print(bool network) above but with a specified phase name.
5167 void igv_print(bool network, const char* phase_name) {
5168 if (network) {
5169 Compile::current()->igv_print_method_to_network(phase_name);
5170 } else {
5171 Compile::current()->igv_print_method_to_file(phase_name);
5172 }
5173 }
5174
5175 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5176 void igv_print_default() {
5177 Compile::current()->print_method(PHASE_DEBUG, 0);
5178 }
5179
5180 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5181 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5182 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5183 void igv_append() {
5184 Compile::current()->igv_print_method_to_file("Debug", true);
5185 }
5186
5187 // Same as igv_append() above but with a specified phase name.
5188 void igv_append(const char* phase_name) {
5189 Compile::current()->igv_print_method_to_file(phase_name, true);
5190 }
5191
5192 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5193 const char* file_name = "custom_debug.xml";
5194 if (_debug_file_printer == nullptr) {
5195 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5196 } else {
5197 _debug_file_printer->update_compiled_method(C->method());
5198 }
5199 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5200 _debug_file_printer->print(phase_name, (Node*)C->root());
5201 }
5202
5203 void Compile::igv_print_method_to_network(const char* phase_name) {
5204 if (_debug_network_printer == nullptr) {
5205 _debug_network_printer = new IdealGraphPrinter(C);
5206 } else {
5207 _debug_network_printer->update_compiled_method(C->method());
5208 }
5209 tty->print_cr("Method printed over network stream to IGV");
5210 _debug_network_printer->print(phase_name, (Node*)C->root());
5211 }
5212 #endif
5213
5214 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5215 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5216 return value;
5217 }
5218 Node* result = nullptr;
5219 if (bt == T_BYTE) {
5220 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5221 result = new RShiftINode(result, phase->intcon(24));
5222 } else if (bt == T_BOOLEAN) {
5223 result = new AndINode(value, phase->intcon(0xFF));
5224 } else if (bt == T_CHAR) {
5225 result = new AndINode(value,phase->intcon(0xFFFF));
5226 } else {
5227 assert(bt == T_SHORT, "unexpected narrow type");
5228 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5229 result = new RShiftINode(result, phase->intcon(16));
5230 }
5231 if (transform_res) {
5232 result = phase->transform(result);
5233 }
5234 return result;
5235 }
5236