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_template_assertion_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 assert(useful.member(n->unique_out()), "do not push a useless node");
434 worklist.push(n->unique_out());
435 }
436 }
437
438 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
439 remove_useless_nodes(_parse_predicate_opaqs, useful); // remove useless Parse Predicate opaque nodes
440 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
441 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
442 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
443 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
444 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
445 #ifdef ASSERT
446 if (_modified_nodes != nullptr) {
447 _modified_nodes->remove_useless_nodes(useful.member_set());
448 }
449 #endif
450
451 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
452 bs->eliminate_useless_gc_barriers(useful, this);
453 // clean up the late inline lists
454 remove_useless_late_inlines( &_late_inlines, useful);
455 remove_useless_late_inlines( &_string_late_inlines, useful);
456 remove_useless_late_inlines( &_boxing_late_inlines, useful);
457 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
458 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
459 }
460
461 // ============================================================================
462 //------------------------------CompileWrapper---------------------------------
463 class CompileWrapper : public StackObj {
464 Compile *const _compile;
465 public:
466 CompileWrapper(Compile* compile);
467
468 ~CompileWrapper();
469 };
470
471 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
472 // the Compile* pointer is stored in the current ciEnv:
473 ciEnv* env = compile->env();
474 assert(env == ciEnv::current(), "must already be a ciEnv active");
475 assert(env->compiler_data() == nullptr, "compile already active?");
476 env->set_compiler_data(compile);
477 assert(compile == Compile::current(), "sanity");
478
479 compile->set_type_dict(nullptr);
480 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
481 compile->clone_map().set_clone_idx(0);
482 compile->set_type_last_size(0);
483 compile->set_last_tf(nullptr, nullptr);
484 compile->set_indexSet_arena(nullptr);
485 compile->set_indexSet_free_block_list(nullptr);
486 compile->init_type_arena();
487 Type::Initialize(compile);
488 _compile->begin_method();
489 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
490 }
491 CompileWrapper::~CompileWrapper() {
492 // simulate crash during compilation
493 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
494
495 _compile->end_method();
496 _compile->env()->set_compiler_data(nullptr);
497 }
498
499
500 //----------------------------print_compile_messages---------------------------
501 void Compile::print_compile_messages() {
502 #ifndef PRODUCT
503 // Check if recompiling
504 if (!subsume_loads() && PrintOpto) {
505 // Recompiling without allowing machine instructions to subsume loads
506 tty->print_cr("*********************************************************");
507 tty->print_cr("** Bailout: Recompile without subsuming loads **");
508 tty->print_cr("*********************************************************");
509 }
510 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
511 // Recompiling without escape analysis
512 tty->print_cr("*********************************************************");
513 tty->print_cr("** Bailout: Recompile without escape analysis **");
514 tty->print_cr("*********************************************************");
515 }
516 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
517 // Recompiling without iterative escape analysis
518 tty->print_cr("*********************************************************");
519 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
520 tty->print_cr("*********************************************************");
521 }
522 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
523 // Recompiling without boxing elimination
524 tty->print_cr("*********************************************************");
525 tty->print_cr("** Bailout: Recompile without boxing elimination **");
526 tty->print_cr("*********************************************************");
527 }
528 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
529 // Recompiling without locks coarsening
530 tty->print_cr("*********************************************************");
531 tty->print_cr("** Bailout: Recompile without locks coarsening **");
532 tty->print_cr("*********************************************************");
533 }
534 if (env()->break_at_compile()) {
535 // Open the debugger when compiling this method.
536 tty->print("### Breaking when compiling: ");
537 method()->print_short_name();
538 tty->cr();
539 BREAKPOINT;
540 }
541
542 if( PrintOpto ) {
543 if (is_osr_compilation()) {
544 tty->print("[OSR]%3d", _compile_id);
545 } else {
546 tty->print("%3d", _compile_id);
547 }
548 }
549 #endif
550 }
551
552 #ifndef PRODUCT
553 void Compile::print_ideal_ir(const char* phase_name) {
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
559 // Node dumping can cause a safepoint, which can break the tty lock.
560 // Buffer all node dumps, so that all safepoints happen before we lock.
561 ResourceMark rm;
562 stringStream ss;
563
564 if (_output == nullptr) {
565 ss.print_cr("AFTER: %s", phase_name);
566 // Print out all nodes in ascending order of index.
567 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
568 } else {
569 // Dump the node blockwise if we have a scheduling
570 _output->print_scheduling(&ss);
571 }
572
573 // Check that the lock is not broken by a safepoint.
574 NoSafepointVerifier nsv;
575 ttyLocker ttyl;
576 if (xtty != nullptr) {
577 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
578 compile_id(),
579 is_osr_compilation() ? " compile_kind='osr'" : "",
580 phase_name);
581 xtty->print("%s", ss.as_string()); // print to tty would use xml escape encoding
582 xtty->tail("ideal");
583 } else {
584 tty->print("%s", ss.as_string());
585 }
586 }
587 #endif
588
589 // ============================================================================
590 //------------------------------Compile standard-------------------------------
591
592 // Compile a method. entry_bci is -1 for normal compilations and indicates
593 // the continuation bci for on stack replacement.
594
595
596 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
597 Options options, DirectiveSet* directive)
598 : Phase(Compiler),
599 _compile_id(ci_env->compile_id()),
600 _options(options),
601 _method(target),
602 _entry_bci(osr_bci),
603 _ilt(nullptr),
604 _stub_function(nullptr),
605 _stub_name(nullptr),
606 _stub_entry_point(nullptr),
607 _max_node_limit(MaxNodeLimit),
608 _post_loop_opts_phase(false),
609 _inlining_progress(false),
610 _inlining_incrementally(false),
611 _do_cleanup(false),
612 _has_reserved_stack_access(target->has_reserved_stack_access()),
613 #ifndef PRODUCT
614 _igv_idx(0),
615 _trace_opto_output(directive->TraceOptoOutputOption),
616 #endif
617 _has_method_handle_invokes(false),
618 _clinit_barrier_on_entry(false),
619 _stress_seed(0),
620 _comp_arena(mtCompiler),
621 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
622 _env(ci_env),
623 _directive(directive),
624 _log(ci_env->log()),
625 _failure_reason(nullptr),
626 _intrinsics (comp_arena(), 0, 0, nullptr),
627 _macro_nodes (comp_arena(), 8, 0, nullptr),
628 _parse_predicate_opaqs (comp_arena(), 8, 0, nullptr),
629 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
630 _expensive_nodes (comp_arena(), 8, 0, nullptr),
631 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
632 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
633 _coarsened_locks (comp_arena(), 8, 0, nullptr),
634 _congraph(nullptr),
635 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
636 _dead_node_list(comp_arena()),
637 _dead_node_count(0),
638 _node_arena(mtCompiler),
639 _old_arena(mtCompiler),
640 _mach_constant_base_node(nullptr),
641 _Compile_types(mtCompiler),
642 _initial_gvn(nullptr),
643 _igvn_worklist(nullptr),
644 _types(nullptr),
645 _node_hash(nullptr),
646 _late_inlines(comp_arena(), 2, 0, nullptr),
647 _string_late_inlines(comp_arena(), 2, 0, nullptr),
648 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
649 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
650 _late_inlines_pos(0),
651 _number_of_mh_late_inlines(0),
652 _print_inlining_stream(new (mtCompiler) stringStream()),
653 _print_inlining_list(nullptr),
654 _print_inlining_idx(0),
655 _print_inlining_output(nullptr),
656 _replay_inline_data(nullptr),
657 _java_calls(0),
658 _inner_loops(0),
659 _interpreter_frame_size(0),
660 _output(nullptr)
661 #ifndef PRODUCT
662 , _in_dump_cnt(0)
663 #endif
664 {
665 C = this;
666 CompileWrapper cw(this);
667
668 if (CITimeVerbose) {
669 tty->print(" ");
670 target->holder()->name()->print();
671 tty->print(".");
672 target->print_short_name();
673 tty->print(" ");
674 }
675 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
676 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
677
678 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
679 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
680 // We can always print a disassembly, either abstract (hex dump) or
681 // with the help of a suitable hsdis library. Thus, we should not
682 // couple print_assembly and print_opto_assembly controls.
683 // But: always print opto and regular assembly on compile command 'print'.
684 bool print_assembly = directive->PrintAssemblyOption;
685 set_print_assembly(print_opto_assembly || print_assembly);
686 #else
687 set_print_assembly(false); // must initialize.
688 #endif
689
690 #ifndef PRODUCT
691 set_parsed_irreducible_loop(false);
692 #endif
693
694 if (directive->ReplayInlineOption) {
695 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
696 }
697 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
698 set_print_intrinsics(directive->PrintIntrinsicsOption);
699 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
700
701 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
702 // Make sure the method being compiled gets its own MDO,
703 // so we can at least track the decompile_count().
704 // Need MDO to record RTM code generation state.
705 method()->ensure_method_data();
706 }
707
708 Init(/*do_aliasing=*/ true);
709
710 print_compile_messages();
711
712 _ilt = InlineTree::build_inline_tree_root();
713
714 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
715 assert(num_alias_types() >= AliasIdxRaw, "");
716
717 #define MINIMUM_NODE_HASH 1023
718
719 // GVN that will be run immediately on new nodes
720 uint estimated_size = method()->code_size()*4+64;
721 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
722 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
723 _types = new (comp_arena()) Type_Array(comp_arena());
724 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
725 PhaseGVN gvn;
726 set_initial_gvn(&gvn);
727
728 print_inlining_init();
729 { // Scope for timing the parser
730 TracePhase tp("parse", &timers[_t_parser]);
731
732 // Put top into the hash table ASAP.
733 initial_gvn()->transform_no_reclaim(top());
734
735 // Set up tf(), start(), and find a CallGenerator.
736 CallGenerator* cg = nullptr;
737 if (is_osr_compilation()) {
738 const TypeTuple *domain = StartOSRNode::osr_domain();
739 const TypeTuple *range = TypeTuple::make_range(method()->signature());
740 init_tf(TypeFunc::make(domain, range));
741 StartNode* s = new StartOSRNode(root(), domain);
742 initial_gvn()->set_type_bottom(s);
743 init_start(s);
744 cg = CallGenerator::for_osr(method(), entry_bci());
745 } else {
746 // Normal case.
747 init_tf(TypeFunc::make(method()));
748 StartNode* s = new StartNode(root(), tf()->domain());
749 initial_gvn()->set_type_bottom(s);
750 init_start(s);
751 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
752 // With java.lang.ref.reference.get() we must go through the
753 // intrinsic - even when get() is the root
754 // method of the compile - so that, if necessary, the value in
755 // the referent field of the reference object gets recorded by
756 // the pre-barrier code.
757 cg = find_intrinsic(method(), false);
758 }
759 if (cg == nullptr) {
760 float past_uses = method()->interpreter_invocation_count();
761 float expected_uses = past_uses;
762 cg = CallGenerator::for_inline(method(), expected_uses);
763 }
764 }
765 if (failing()) return;
766 if (cg == nullptr) {
767 const char* reason = InlineTree::check_can_parse(method());
768 assert(reason != nullptr, "expect reason for parse failure");
769 stringStream ss;
770 ss.print("cannot parse method: %s", reason);
771 record_method_not_compilable(ss.as_string());
772 return;
773 }
774
775 gvn.set_type(root(), root()->bottom_type());
776
777 JVMState* jvms = build_start_state(start(), tf());
778 if ((jvms = cg->generate(jvms)) == nullptr) {
779 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
780 assert(failure_reason() != nullptr, "expect reason for parse failure");
781 stringStream ss;
782 ss.print("method parse failed: %s", failure_reason());
783 record_method_not_compilable(ss.as_string());
784 }
785 return;
786 }
787 GraphKit kit(jvms);
788
789 if (!kit.stopped()) {
790 // Accept return values, and transfer control we know not where.
791 // This is done by a special, unique ReturnNode bound to root.
792 return_values(kit.jvms());
793 }
794
795 if (kit.has_exceptions()) {
796 // Any exceptions that escape from this call must be rethrown
797 // to whatever caller is dynamically above us on the stack.
798 // This is done by a special, unique RethrowNode bound to root.
799 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
800 }
801
802 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
803
804 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
805 inline_string_calls(true);
806 }
807
808 if (failing()) return;
809
810 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
811
812 // Remove clutter produced by parsing.
813 if (!failing()) {
814 ResourceMark rm;
815 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
816 }
817 }
818
819 // Note: Large methods are capped off in do_one_bytecode().
820 if (failing()) return;
821
822 // After parsing, node notes are no longer automagic.
823 // They must be propagated by register_new_node_with_optimizer(),
824 // clone(), or the like.
825 set_default_node_notes(nullptr);
826
827 #ifndef PRODUCT
828 if (should_print_igv(1)) {
829 _igv_printer->print_inlining();
830 }
831 #endif
832
833 if (failing()) return;
834 NOT_PRODUCT( verify_graph_edges(); )
835
836 // If any phase is randomized for stress testing, seed random number
837 // generation and log the seed for repeatability.
838 if (StressLCM || StressGCM || StressIGVN || StressCCP || StressIncrementalInlining) {
839 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
840 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
841 FLAG_SET_ERGO(StressSeed, _stress_seed);
842 } else {
843 _stress_seed = StressSeed;
844 }
845 if (_log != nullptr) {
846 _log->elem("stress_test seed='%u'", _stress_seed);
847 }
848 }
849
850 // Now optimize
851 Optimize();
852 if (failing()) return;
853 NOT_PRODUCT( verify_graph_edges(); )
854
855 #ifndef PRODUCT
856 if (should_print_ideal()) {
857 print_ideal_ir("print_ideal");
858 }
859 #endif
860
861 #ifdef ASSERT
862 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
863 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
864 #endif
865
866 // Dump compilation data to replay it.
867 if (directive->DumpReplayOption) {
868 env()->dump_replay_data(_compile_id);
869 }
870 if (directive->DumpInlineOption && (ilt() != nullptr)) {
871 env()->dump_inline_data(_compile_id);
872 }
873
874 // Now that we know the size of all the monitors we can add a fixed slot
875 // for the original deopt pc.
876 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
877 set_fixed_slots(next_slot);
878
879 // Compute when to use implicit null checks. Used by matching trap based
880 // nodes and NullCheck optimization.
881 set_allowed_deopt_reasons();
882
883 // Now generate code
884 Code_Gen();
885 }
886
887 //------------------------------Compile----------------------------------------
888 // Compile a runtime stub
889 Compile::Compile( ciEnv* ci_env,
890 TypeFunc_generator generator,
891 address stub_function,
892 const char *stub_name,
893 int is_fancy_jump,
894 bool pass_tls,
895 bool return_pc,
896 DirectiveSet* directive)
897 : Phase(Compiler),
898 _compile_id(0),
899 _options(Options::for_runtime_stub()),
900 _method(nullptr),
901 _entry_bci(InvocationEntryBci),
902 _stub_function(stub_function),
903 _stub_name(stub_name),
904 _stub_entry_point(nullptr),
905 _max_node_limit(MaxNodeLimit),
906 _post_loop_opts_phase(false),
907 _inlining_progress(false),
908 _inlining_incrementally(false),
909 _has_reserved_stack_access(false),
910 #ifndef PRODUCT
911 _igv_idx(0),
912 _trace_opto_output(directive->TraceOptoOutputOption),
913 #endif
914 _has_method_handle_invokes(false),
915 _clinit_barrier_on_entry(false),
916 _stress_seed(0),
917 _comp_arena(mtCompiler),
918 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
919 _env(ci_env),
920 _directive(directive),
921 _log(ci_env->log()),
922 _failure_reason(nullptr),
923 _congraph(nullptr),
924 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
925 _dead_node_list(comp_arena()),
926 _dead_node_count(0),
927 _node_arena(mtCompiler),
928 _old_arena(mtCompiler),
929 _mach_constant_base_node(nullptr),
930 _Compile_types(mtCompiler),
931 _initial_gvn(nullptr),
932 _igvn_worklist(nullptr),
933 _types(nullptr),
934 _node_hash(nullptr),
935 _number_of_mh_late_inlines(0),
936 _print_inlining_stream(new (mtCompiler) stringStream()),
937 _print_inlining_list(nullptr),
938 _print_inlining_idx(0),
939 _print_inlining_output(nullptr),
940 _replay_inline_data(nullptr),
941 _java_calls(0),
942 _inner_loops(0),
943 _interpreter_frame_size(0),
944 _output(nullptr),
945 #ifndef PRODUCT
946 _in_dump_cnt(0),
947 #endif
948 _allowed_reasons(0) {
949 C = this;
950
951 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
952 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
953
954 #ifndef PRODUCT
955 set_print_assembly(PrintFrameConverterAssembly);
956 set_parsed_irreducible_loop(false);
957 #else
958 set_print_assembly(false); // Must initialize.
959 #endif
960 set_has_irreducible_loop(false); // no loops
961
962 CompileWrapper cw(this);
963 Init(/*do_aliasing=*/ false);
964 init_tf((*generator)());
965
966 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
967 _types = new (comp_arena()) Type_Array(comp_arena());
968 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
969 {
970 PhaseGVN gvn;
971 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
972 gvn.transform_no_reclaim(top());
973
974 GraphKit kit;
975 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
976 }
977
978 NOT_PRODUCT( verify_graph_edges(); )
979
980 Code_Gen();
981 }
982
983 //------------------------------Init-------------------------------------------
984 // Prepare for a single compilation
985 void Compile::Init(bool aliasing) {
986 _do_aliasing = aliasing;
987 _unique = 0;
988 _regalloc = nullptr;
989
990 _tf = nullptr; // filled in later
991 _top = nullptr; // cached later
992 _matcher = nullptr; // filled in later
993 _cfg = nullptr; // filled in later
994
995 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
996
997 _node_note_array = nullptr;
998 _default_node_notes = nullptr;
999 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1000
1001 _immutable_memory = nullptr; // filled in at first inquiry
1002
1003 #ifdef ASSERT
1004 _phase_optimize_finished = false;
1005 _exception_backedge = false;
1006 _type_verify = nullptr;
1007 #endif
1008
1009 // Globally visible Nodes
1010 // First set TOP to null to give safe behavior during creation of RootNode
1011 set_cached_top_node(nullptr);
1012 set_root(new RootNode());
1013 // Now that you have a Root to point to, create the real TOP
1014 set_cached_top_node( new ConNode(Type::TOP) );
1015 set_recent_alloc(nullptr, nullptr);
1016
1017 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1018 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1019 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1020 env()->set_dependencies(new Dependencies(env()));
1021
1022 _fixed_slots = 0;
1023 set_has_split_ifs(false);
1024 set_has_loops(false); // first approximation
1025 set_has_stringbuilder(false);
1026 set_has_boxed_value(false);
1027 _trap_can_recompile = false; // no traps emitted yet
1028 _major_progress = true; // start out assuming good things will happen
1029 set_has_unsafe_access(false);
1030 set_max_vector_size(0);
1031 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1032 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1033 set_decompile_count(0);
1034
1035 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1036 _loop_opts_cnt = LoopOptsCount;
1037 set_do_inlining(Inline);
1038 set_max_inline_size(MaxInlineSize);
1039 set_freq_inline_size(FreqInlineSize);
1040 set_do_scheduling(OptoScheduling);
1041
1042 set_do_vector_loop(false);
1043 set_has_monitors(false);
1044
1045 if (AllowVectorizeOnDemand) {
1046 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1047 set_do_vector_loop(true);
1048 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());})
1049 } else if (has_method() && method()->name() != 0 &&
1050 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1051 set_do_vector_loop(true);
1052 }
1053 }
1054 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1055 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());})
1056
1057 set_rtm_state(NoRTM); // No RTM lock eliding by default
1058 _max_node_limit = _directive->MaxNodeLimitOption;
1059
1060 #if INCLUDE_RTM_OPT
1061 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) {
1062 int rtm_state = method()->method_data()->rtm_state();
1063 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1064 // Don't generate RTM lock eliding code.
1065 set_rtm_state(NoRTM);
1066 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1067 // Generate RTM lock eliding code without abort ratio calculation code.
1068 set_rtm_state(UseRTM);
1069 } else if (UseRTMDeopt) {
1070 // Generate RTM lock eliding code and include abort ratio calculation
1071 // code if UseRTMDeopt is on.
1072 set_rtm_state(ProfileRTM);
1073 }
1074 }
1075 #endif
1076 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1077 set_clinit_barrier_on_entry(true);
1078 }
1079 if (debug_info()->recording_non_safepoints()) {
1080 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1081 (comp_arena(), 8, 0, nullptr));
1082 set_default_node_notes(Node_Notes::make(this));
1083 }
1084
1085 const int grow_ats = 16;
1086 _max_alias_types = grow_ats;
1087 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1088 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1089 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1090 {
1091 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1092 }
1093 // Initialize the first few types.
1094 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1095 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1096 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1097 _num_alias_types = AliasIdxRaw+1;
1098 // Zero out the alias type cache.
1099 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1100 // A null adr_type hits in the cache right away. Preload the right answer.
1101 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1102 }
1103
1104 //---------------------------init_start----------------------------------------
1105 // Install the StartNode on this compile object.
1106 void Compile::init_start(StartNode* s) {
1107 if (failing())
1108 return; // already failing
1109 assert(s == start(), "");
1110 }
1111
1112 /**
1113 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1114 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1115 * the ideal graph.
1116 */
1117 StartNode* Compile::start() const {
1118 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1119 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1120 Node* start = root()->fast_out(i);
1121 if (start->is_Start()) {
1122 return start->as_Start();
1123 }
1124 }
1125 fatal("Did not find Start node!");
1126 return nullptr;
1127 }
1128
1129 //-------------------------------immutable_memory-------------------------------------
1130 // Access immutable memory
1131 Node* Compile::immutable_memory() {
1132 if (_immutable_memory != nullptr) {
1133 return _immutable_memory;
1134 }
1135 StartNode* s = start();
1136 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1137 Node *p = s->fast_out(i);
1138 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1139 _immutable_memory = p;
1140 return _immutable_memory;
1141 }
1142 }
1143 ShouldNotReachHere();
1144 return nullptr;
1145 }
1146
1147 //----------------------set_cached_top_node------------------------------------
1148 // Install the cached top node, and make sure Node::is_top works correctly.
1149 void Compile::set_cached_top_node(Node* tn) {
1150 if (tn != nullptr) verify_top(tn);
1151 Node* old_top = _top;
1152 _top = tn;
1153 // Calling Node::setup_is_top allows the nodes the chance to adjust
1154 // their _out arrays.
1155 if (_top != nullptr) _top->setup_is_top();
1156 if (old_top != nullptr) old_top->setup_is_top();
1157 assert(_top == nullptr || top()->is_top(), "");
1158 }
1159
1160 #ifdef ASSERT
1161 uint Compile::count_live_nodes_by_graph_walk() {
1162 Unique_Node_List useful(comp_arena());
1163 // Get useful node list by walking the graph.
1164 identify_useful_nodes(useful);
1165 return useful.size();
1166 }
1167
1168 void Compile::print_missing_nodes() {
1169
1170 // Return if CompileLog is null and PrintIdealNodeCount is false.
1171 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1172 return;
1173 }
1174
1175 // This is an expensive function. It is executed only when the user
1176 // specifies VerifyIdealNodeCount option or otherwise knows the
1177 // additional work that needs to be done to identify reachable nodes
1178 // by walking the flow graph and find the missing ones using
1179 // _dead_node_list.
1180
1181 Unique_Node_List useful(comp_arena());
1182 // Get useful node list by walking the graph.
1183 identify_useful_nodes(useful);
1184
1185 uint l_nodes = C->live_nodes();
1186 uint l_nodes_by_walk = useful.size();
1187
1188 if (l_nodes != l_nodes_by_walk) {
1189 if (_log != nullptr) {
1190 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1191 _log->stamp();
1192 _log->end_head();
1193 }
1194 VectorSet& useful_member_set = useful.member_set();
1195 int last_idx = l_nodes_by_walk;
1196 for (int i = 0; i < last_idx; i++) {
1197 if (useful_member_set.test(i)) {
1198 if (_dead_node_list.test(i)) {
1199 if (_log != nullptr) {
1200 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1201 }
1202 if (PrintIdealNodeCount) {
1203 // Print the log message to tty
1204 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1205 useful.at(i)->dump();
1206 }
1207 }
1208 }
1209 else if (! _dead_node_list.test(i)) {
1210 if (_log != nullptr) {
1211 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1212 }
1213 if (PrintIdealNodeCount) {
1214 // Print the log message to tty
1215 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1216 }
1217 }
1218 }
1219 if (_log != nullptr) {
1220 _log->tail("mismatched_nodes");
1221 }
1222 }
1223 }
1224 void Compile::record_modified_node(Node* n) {
1225 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1226 _modified_nodes->push(n);
1227 }
1228 }
1229
1230 void Compile::remove_modified_node(Node* n) {
1231 if (_modified_nodes != nullptr) {
1232 _modified_nodes->remove(n);
1233 }
1234 }
1235 #endif
1236
1237 #ifndef PRODUCT
1238 void Compile::verify_top(Node* tn) const {
1239 if (tn != nullptr) {
1240 assert(tn->is_Con(), "top node must be a constant");
1241 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1242 assert(tn->in(0) != nullptr, "must have live top node");
1243 }
1244 }
1245 #endif
1246
1247
1248 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1249
1250 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1251 guarantee(arr != nullptr, "");
1252 int num_blocks = arr->length();
1253 if (grow_by < num_blocks) grow_by = num_blocks;
1254 int num_notes = grow_by * _node_notes_block_size;
1255 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1256 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1257 while (num_notes > 0) {
1258 arr->append(notes);
1259 notes += _node_notes_block_size;
1260 num_notes -= _node_notes_block_size;
1261 }
1262 assert(num_notes == 0, "exact multiple, please");
1263 }
1264
1265 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1266 if (source == nullptr || dest == nullptr) return false;
1267
1268 if (dest->is_Con())
1269 return false; // Do not push debug info onto constants.
1270
1271 #ifdef ASSERT
1272 // Leave a bread crumb trail pointing to the original node:
1273 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1274 dest->set_debug_orig(source);
1275 }
1276 #endif
1277
1278 if (node_note_array() == nullptr)
1279 return false; // Not collecting any notes now.
1280
1281 // This is a copy onto a pre-existing node, which may already have notes.
1282 // If both nodes have notes, do not overwrite any pre-existing notes.
1283 Node_Notes* source_notes = node_notes_at(source->_idx);
1284 if (source_notes == nullptr || source_notes->is_clear()) return false;
1285 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1286 if (dest_notes == nullptr || dest_notes->is_clear()) {
1287 return set_node_notes_at(dest->_idx, source_notes);
1288 }
1289
1290 Node_Notes merged_notes = (*source_notes);
1291 // The order of operations here ensures that dest notes will win...
1292 merged_notes.update_from(dest_notes);
1293 return set_node_notes_at(dest->_idx, &merged_notes);
1294 }
1295
1296
1297 //--------------------------allow_range_check_smearing-------------------------
1298 // Gating condition for coalescing similar range checks.
1299 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1300 // single covering check that is at least as strong as any of them.
1301 // If the optimization succeeds, the simplified (strengthened) range check
1302 // will always succeed. If it fails, we will deopt, and then give up
1303 // on the optimization.
1304 bool Compile::allow_range_check_smearing() const {
1305 // If this method has already thrown a range-check,
1306 // assume it was because we already tried range smearing
1307 // and it failed.
1308 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1309 return !already_trapped;
1310 }
1311
1312
1313 //------------------------------flatten_alias_type-----------------------------
1314 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1315 assert(do_aliasing(), "Aliasing should be enabled");
1316 int offset = tj->offset();
1317 TypePtr::PTR ptr = tj->ptr();
1318
1319 // Known instance (scalarizable allocation) alias only with itself.
1320 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1321 tj->is_oopptr()->is_known_instance();
1322
1323 // Process weird unsafe references.
1324 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1325 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1326 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1327 tj = TypeOopPtr::BOTTOM;
1328 ptr = tj->ptr();
1329 offset = tj->offset();
1330 }
1331
1332 // Array pointers need some flattening
1333 const TypeAryPtr* ta = tj->isa_aryptr();
1334 if (ta && ta->is_stable()) {
1335 // Erase stability property for alias analysis.
1336 tj = ta = ta->cast_to_stable(false);
1337 }
1338 if( ta && is_known_inst ) {
1339 if ( offset != Type::OffsetBot &&
1340 offset > arrayOopDesc::length_offset_in_bytes() ) {
1341 offset = Type::OffsetBot; // Flatten constant access into array body only
1342 tj = ta = ta->
1343 remove_speculative()->
1344 cast_to_ptr_type(ptr)->
1345 with_offset(offset);
1346 }
1347 } else if (ta) {
1348 // For arrays indexed by constant indices, we flatten the alias
1349 // space to include all of the array body. Only the header, klass
1350 // and array length can be accessed un-aliased.
1351 if( offset != Type::OffsetBot ) {
1352 if( ta->const_oop() ) { // MethodData* or Method*
1353 offset = Type::OffsetBot; // Flatten constant access into array body
1354 tj = ta = ta->
1355 remove_speculative()->
1356 cast_to_ptr_type(ptr)->
1357 cast_to_exactness(false)->
1358 with_offset(offset);
1359 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1360 // range is OK as-is.
1361 tj = ta = TypeAryPtr::RANGE;
1362 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1363 tj = TypeInstPtr::KLASS; // all klass loads look alike
1364 ta = TypeAryPtr::RANGE; // generic ignored junk
1365 ptr = TypePtr::BotPTR;
1366 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1367 tj = TypeInstPtr::MARK;
1368 ta = TypeAryPtr::RANGE; // generic ignored junk
1369 ptr = TypePtr::BotPTR;
1370 } else { // Random constant offset into array body
1371 offset = Type::OffsetBot; // Flatten constant access into array body
1372 tj = ta = ta->
1373 remove_speculative()->
1374 cast_to_ptr_type(ptr)->
1375 cast_to_exactness(false)->
1376 with_offset(offset);
1377 }
1378 }
1379 // Arrays of fixed size alias with arrays of unknown size.
1380 if (ta->size() != TypeInt::POS) {
1381 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1382 tj = ta = ta->
1383 remove_speculative()->
1384 cast_to_ptr_type(ptr)->
1385 with_ary(tary)->
1386 cast_to_exactness(false);
1387 }
1388 // Arrays of known objects become arrays of unknown objects.
1389 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1390 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1391 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1392 }
1393 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1394 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1395 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1396 }
1397 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1398 // cannot be distinguished by bytecode alone.
1399 if (ta->elem() == TypeInt::BOOL) {
1400 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1401 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1402 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1403 }
1404 // During the 2nd round of IterGVN, NotNull castings are removed.
1405 // Make sure the Bottom and NotNull variants alias the same.
1406 // Also, make sure exact and non-exact variants alias the same.
1407 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1408 tj = ta = ta->
1409 remove_speculative()->
1410 cast_to_ptr_type(TypePtr::BotPTR)->
1411 cast_to_exactness(false)->
1412 with_offset(offset);
1413 }
1414 }
1415
1416 // Oop pointers need some flattening
1417 const TypeInstPtr *to = tj->isa_instptr();
1418 if (to && to != TypeOopPtr::BOTTOM) {
1419 ciInstanceKlass* ik = to->instance_klass();
1420 if( ptr == TypePtr::Constant ) {
1421 if (ik != ciEnv::current()->Class_klass() ||
1422 offset < ik->layout_helper_size_in_bytes()) {
1423 // No constant oop pointers (such as Strings); they alias with
1424 // unknown strings.
1425 assert(!is_known_inst, "not scalarizable allocation");
1426 tj = to = to->
1427 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1428 remove_speculative()->
1429 cast_to_ptr_type(TypePtr::BotPTR)->
1430 cast_to_exactness(false);
1431 }
1432 } else if( is_known_inst ) {
1433 tj = to; // Keep NotNull and klass_is_exact for instance type
1434 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1435 // During the 2nd round of IterGVN, NotNull castings are removed.
1436 // Make sure the Bottom and NotNull variants alias the same.
1437 // Also, make sure exact and non-exact variants alias the same.
1438 tj = to = to->
1439 remove_speculative()->
1440 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1441 cast_to_ptr_type(TypePtr::BotPTR)->
1442 cast_to_exactness(false);
1443 }
1444 if (to->speculative() != nullptr) {
1445 tj = to = to->remove_speculative();
1446 }
1447 // Canonicalize the holder of this field
1448 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1449 // First handle header references such as a LoadKlassNode, even if the
1450 // object's klass is unloaded at compile time (4965979).
1451 if (!is_known_inst) { // Do it only for non-instance types
1452 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1453 }
1454 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1455 // Static fields are in the space above the normal instance
1456 // fields in the java.lang.Class instance.
1457 if (ik != ciEnv::current()->Class_klass()) {
1458 to = nullptr;
1459 tj = TypeOopPtr::BOTTOM;
1460 offset = tj->offset();
1461 }
1462 } else {
1463 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1464 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1465 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1466 if( is_known_inst ) {
1467 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id());
1468 } else {
1469 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset);
1470 }
1471 }
1472 }
1473 }
1474
1475 // Klass pointers to object array klasses need some flattening
1476 const TypeKlassPtr *tk = tj->isa_klassptr();
1477 if( tk ) {
1478 // If we are referencing a field within a Klass, we need
1479 // to assume the worst case of an Object. Both exact and
1480 // inexact types must flatten to the same alias class so
1481 // use NotNull as the PTR.
1482 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1483 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1484 env()->Object_klass(),
1485 offset);
1486 }
1487
1488 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1489 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1490 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1491 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1492 } else {
1493 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1494 }
1495 }
1496
1497 // Check for precise loads from the primary supertype array and force them
1498 // to the supertype cache alias index. Check for generic array loads from
1499 // the primary supertype array and also force them to the supertype cache
1500 // alias index. Since the same load can reach both, we need to merge
1501 // these 2 disparate memories into the same alias class. Since the
1502 // primary supertype array is read-only, there's no chance of confusion
1503 // where we bypass an array load and an array store.
1504 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1505 if (offset == Type::OffsetBot ||
1506 (offset >= primary_supers_offset &&
1507 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1508 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1509 offset = in_bytes(Klass::secondary_super_cache_offset());
1510 tj = tk = tk->with_offset(offset);
1511 }
1512 }
1513
1514 // Flatten all Raw pointers together.
1515 if (tj->base() == Type::RawPtr)
1516 tj = TypeRawPtr::BOTTOM;
1517
1518 if (tj->base() == Type::AnyPtr)
1519 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1520
1521 offset = tj->offset();
1522 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1523
1524 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1525 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1526 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1527 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1528 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1529 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1530 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1531 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1532 assert( tj->ptr() != TypePtr::TopPTR &&
1533 tj->ptr() != TypePtr::AnyNull &&
1534 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1535 // assert( tj->ptr() != TypePtr::Constant ||
1536 // tj->base() == Type::RawPtr ||
1537 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1538
1539 return tj;
1540 }
1541
1542 void Compile::AliasType::Init(int i, const TypePtr* at) {
1543 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1544 _index = i;
1545 _adr_type = at;
1546 _field = nullptr;
1547 _element = nullptr;
1548 _is_rewritable = true; // default
1549 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1550 if (atoop != nullptr && atoop->is_known_instance()) {
1551 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1552 _general_index = Compile::current()->get_alias_index(gt);
1553 } else {
1554 _general_index = 0;
1555 }
1556 }
1557
1558 BasicType Compile::AliasType::basic_type() const {
1559 if (element() != nullptr) {
1560 const Type* element = adr_type()->is_aryptr()->elem();
1561 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1562 } if (field() != nullptr) {
1563 return field()->layout_type();
1564 } else {
1565 return T_ILLEGAL; // unknown
1566 }
1567 }
1568
1569 //---------------------------------print_on------------------------------------
1570 #ifndef PRODUCT
1571 void Compile::AliasType::print_on(outputStream* st) {
1572 if (index() < 10)
1573 st->print("@ <%d> ", index());
1574 else st->print("@ <%d>", index());
1575 st->print(is_rewritable() ? " " : " RO");
1576 int offset = adr_type()->offset();
1577 if (offset == Type::OffsetBot)
1578 st->print(" +any");
1579 else st->print(" +%-3d", offset);
1580 st->print(" in ");
1581 adr_type()->dump_on(st);
1582 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1583 if (field() != nullptr && tjp) {
1584 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1585 tjp->offset() != field()->offset_in_bytes()) {
1586 st->print(" != ");
1587 field()->print();
1588 st->print(" ***");
1589 }
1590 }
1591 }
1592
1593 void print_alias_types() {
1594 Compile* C = Compile::current();
1595 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1596 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1597 C->alias_type(idx)->print_on(tty);
1598 tty->cr();
1599 }
1600 }
1601 #endif
1602
1603
1604 //----------------------------probe_alias_cache--------------------------------
1605 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1606 intptr_t key = (intptr_t) adr_type;
1607 key ^= key >> logAliasCacheSize;
1608 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1609 }
1610
1611
1612 //-----------------------------grow_alias_types--------------------------------
1613 void Compile::grow_alias_types() {
1614 const int old_ats = _max_alias_types; // how many before?
1615 const int new_ats = old_ats; // how many more?
1616 const int grow_ats = old_ats+new_ats; // how many now?
1617 _max_alias_types = grow_ats;
1618 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1619 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1620 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1621 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1622 }
1623
1624
1625 //--------------------------------find_alias_type------------------------------
1626 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1627 if (!do_aliasing()) {
1628 return alias_type(AliasIdxBot);
1629 }
1630
1631 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1632 if (ace->_adr_type == adr_type) {
1633 return alias_type(ace->_index);
1634 }
1635
1636 // Handle special cases.
1637 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1638 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1639
1640 // Do it the slow way.
1641 const TypePtr* flat = flatten_alias_type(adr_type);
1642
1643 #ifdef ASSERT
1644 {
1645 ResourceMark rm;
1646 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1647 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1648 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1649 Type::str(adr_type));
1650 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1651 const TypeOopPtr* foop = flat->is_oopptr();
1652 // Scalarizable allocations have exact klass always.
1653 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1654 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1655 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1656 Type::str(foop), Type::str(xoop));
1657 }
1658 }
1659 #endif
1660
1661 int idx = AliasIdxTop;
1662 for (int i = 0; i < num_alias_types(); i++) {
1663 if (alias_type(i)->adr_type() == flat) {
1664 idx = i;
1665 break;
1666 }
1667 }
1668
1669 if (idx == AliasIdxTop) {
1670 if (no_create) return nullptr;
1671 // Grow the array if necessary.
1672 if (_num_alias_types == _max_alias_types) grow_alias_types();
1673 // Add a new alias type.
1674 idx = _num_alias_types++;
1675 _alias_types[idx]->Init(idx, flat);
1676 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1677 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1678 if (flat->isa_instptr()) {
1679 if (flat->offset() == java_lang_Class::klass_offset()
1680 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1681 alias_type(idx)->set_rewritable(false);
1682 }
1683 if (flat->isa_aryptr()) {
1684 #ifdef ASSERT
1685 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1686 // (T_BYTE has the weakest alignment and size restrictions...)
1687 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1688 #endif
1689 if (flat->offset() == TypePtr::OffsetBot) {
1690 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1691 }
1692 }
1693 if (flat->isa_klassptr()) {
1694 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1695 alias_type(idx)->set_rewritable(false);
1696 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1697 alias_type(idx)->set_rewritable(false);
1698 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1699 alias_type(idx)->set_rewritable(false);
1700 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1701 alias_type(idx)->set_rewritable(false);
1702 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1703 alias_type(idx)->set_rewritable(false);
1704 }
1705 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1706 // but the base pointer type is not distinctive enough to identify
1707 // references into JavaThread.)
1708
1709 // Check for final fields.
1710 const TypeInstPtr* tinst = flat->isa_instptr();
1711 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1712 ciField* field;
1713 if (tinst->const_oop() != nullptr &&
1714 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1715 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1716 // static field
1717 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1718 field = k->get_field_by_offset(tinst->offset(), true);
1719 } else {
1720 ciInstanceKlass *k = tinst->instance_klass();
1721 field = k->get_field_by_offset(tinst->offset(), false);
1722 }
1723 assert(field == nullptr ||
1724 original_field == nullptr ||
1725 (field->holder() == original_field->holder() &&
1726 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1727 field->is_static() == original_field->is_static()), "wrong field?");
1728 // Set field() and is_rewritable() attributes.
1729 if (field != nullptr) alias_type(idx)->set_field(field);
1730 }
1731 }
1732
1733 // Fill the cache for next time.
1734 ace->_adr_type = adr_type;
1735 ace->_index = idx;
1736 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1737
1738 // Might as well try to fill the cache for the flattened version, too.
1739 AliasCacheEntry* face = probe_alias_cache(flat);
1740 if (face->_adr_type == nullptr) {
1741 face->_adr_type = flat;
1742 face->_index = idx;
1743 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1744 }
1745
1746 return alias_type(idx);
1747 }
1748
1749
1750 Compile::AliasType* Compile::alias_type(ciField* field) {
1751 const TypeOopPtr* t;
1752 if (field->is_static())
1753 t = TypeInstPtr::make(field->holder()->java_mirror());
1754 else
1755 t = TypeOopPtr::make_from_klass_raw(field->holder());
1756 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1757 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1758 return atp;
1759 }
1760
1761
1762 //------------------------------have_alias_type--------------------------------
1763 bool Compile::have_alias_type(const TypePtr* adr_type) {
1764 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1765 if (ace->_adr_type == adr_type) {
1766 return true;
1767 }
1768
1769 // Handle special cases.
1770 if (adr_type == nullptr) return true;
1771 if (adr_type == TypePtr::BOTTOM) return true;
1772
1773 return find_alias_type(adr_type, true, nullptr) != nullptr;
1774 }
1775
1776 //-----------------------------must_alias--------------------------------------
1777 // True if all values of the given address type are in the given alias category.
1778 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1779 if (alias_idx == AliasIdxBot) return true; // the universal category
1780 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1781 if (alias_idx == AliasIdxTop) return false; // the empty category
1782 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1783
1784 // the only remaining possible overlap is identity
1785 int adr_idx = get_alias_index(adr_type);
1786 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1787 assert(adr_idx == alias_idx ||
1788 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1789 && adr_type != TypeOopPtr::BOTTOM),
1790 "should not be testing for overlap with an unsafe pointer");
1791 return adr_idx == alias_idx;
1792 }
1793
1794 //------------------------------can_alias--------------------------------------
1795 // True if any values of the given address type are in the given alias category.
1796 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1797 if (alias_idx == AliasIdxTop) return false; // the empty category
1798 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1799 // Known instance doesn't alias with bottom memory
1800 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1801 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1802
1803 // the only remaining possible overlap is identity
1804 int adr_idx = get_alias_index(adr_type);
1805 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1806 return adr_idx == alias_idx;
1807 }
1808
1809 // Remove the opaque nodes that protect the Parse Predicates so that all unused
1810 // checks and uncommon_traps will be eliminated from the ideal graph.
1811 void Compile::cleanup_parse_predicates(PhaseIterGVN& igvn) const {
1812 if (parse_predicate_count() == 0) {
1813 return;
1814 }
1815 for (int i = parse_predicate_count(); i > 0; i--) {
1816 Node* n = parse_predicate_opaque1_node(i - 1);
1817 assert(n->Opcode() == Op_Opaque1, "must be");
1818 igvn.replace_node(n, n->in(1));
1819 }
1820 assert(parse_predicate_count() == 0, "should be clean!");
1821 }
1822
1823 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1824 if (!n->for_post_loop_opts_igvn()) {
1825 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1826 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1827 _for_post_loop_igvn.append(n);
1828 }
1829 }
1830
1831 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1832 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1833 _for_post_loop_igvn.remove(n);
1834 }
1835
1836 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1837 // Verify that all previous optimizations produced a valid graph
1838 // at least to this point, even if no loop optimizations were done.
1839 PhaseIdealLoop::verify(igvn);
1840
1841 C->set_post_loop_opts_phase(); // no more loop opts allowed
1842
1843 assert(!C->major_progress(), "not cleared");
1844
1845 if (_for_post_loop_igvn.length() > 0) {
1846 while (_for_post_loop_igvn.length() > 0) {
1847 Node* n = _for_post_loop_igvn.pop();
1848 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1849 igvn._worklist.push(n);
1850 }
1851 igvn.optimize();
1852 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1853
1854 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1855 if (C->major_progress()) {
1856 C->clear_major_progress(); // ensure that major progress is now clear
1857 }
1858 }
1859 }
1860
1861 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1862 if (OptimizeUnstableIf) {
1863 _unstable_if_traps.append(trap);
1864 }
1865 }
1866
1867 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1868 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1869 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1870 Node* n = trap->uncommon_trap();
1871 if (!useful.member(n)) {
1872 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1873 }
1874 }
1875 }
1876
1877 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1878 // or fold-compares case. Return true if succeed or not found.
1879 //
1880 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1881 // false and ask fold-compares to yield.
1882 //
1883 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1884 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1885 // when deoptimization does happen.
1886 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1887 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1888 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1889 if (trap->uncommon_trap() == unc) {
1890 if (yield && trap->modified()) {
1891 return false;
1892 }
1893 _unstable_if_traps.delete_at(i);
1894 break;
1895 }
1896 }
1897 return true;
1898 }
1899
1900 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1901 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1902 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1903 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1904 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1905 CallStaticJavaNode* unc = trap->uncommon_trap();
1906 int next_bci = trap->next_bci();
1907 bool modified = trap->modified();
1908
1909 if (next_bci != -1 && !modified) {
1910 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1911 JVMState* jvms = unc->jvms();
1912 ciMethod* method = jvms->method();
1913 ciBytecodeStream iter(method);
1914
1915 iter.force_bci(jvms->bci());
1916 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1917 Bytecodes::Code c = iter.cur_bc();
1918 Node* lhs = nullptr;
1919 Node* rhs = nullptr;
1920 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1921 lhs = unc->peek_operand(0);
1922 rhs = unc->peek_operand(1);
1923 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1924 lhs = unc->peek_operand(0);
1925 }
1926
1927 ResourceMark rm;
1928 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1929 assert(live_locals.is_valid(), "broken liveness info");
1930 int len = (int)live_locals.size();
1931
1932 for (int i = 0; i < len; i++) {
1933 Node* local = unc->local(jvms, i);
1934 // kill local using the liveness of next_bci.
1935 // give up when the local looks like an operand to secure reexecution.
1936 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1937 uint idx = jvms->locoff() + i;
1938 #ifdef ASSERT
1939 if (Verbose) {
1940 tty->print("[unstable_if] kill local#%d: ", idx);
1941 local->dump();
1942 tty->cr();
1943 }
1944 #endif
1945 igvn.replace_input_of(unc, idx, top());
1946 modified = true;
1947 }
1948 }
1949 }
1950
1951 // keep the mondified trap for late query
1952 if (modified) {
1953 trap->set_modified();
1954 } else {
1955 _unstable_if_traps.delete_at(i);
1956 }
1957 }
1958 igvn.optimize();
1959 }
1960
1961 // StringOpts and late inlining of string methods
1962 void Compile::inline_string_calls(bool parse_time) {
1963 {
1964 // remove useless nodes to make the usage analysis simpler
1965 ResourceMark rm;
1966 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1967 }
1968
1969 {
1970 ResourceMark rm;
1971 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1972 PhaseStringOpts pso(initial_gvn());
1973 print_method(PHASE_AFTER_STRINGOPTS, 3);
1974 }
1975
1976 // now inline anything that we skipped the first time around
1977 if (!parse_time) {
1978 _late_inlines_pos = _late_inlines.length();
1979 }
1980
1981 while (_string_late_inlines.length() > 0) {
1982 CallGenerator* cg = _string_late_inlines.pop();
1983 cg->do_late_inline();
1984 if (failing()) return;
1985 }
1986 _string_late_inlines.trunc_to(0);
1987 }
1988
1989 // Late inlining of boxing methods
1990 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1991 if (_boxing_late_inlines.length() > 0) {
1992 assert(has_boxed_value(), "inconsistent");
1993
1994 PhaseGVN* gvn = initial_gvn();
1995 set_inlining_incrementally(true);
1996
1997 igvn_worklist()->ensure_empty(); // should be done with igvn
1998
1999 _late_inlines_pos = _late_inlines.length();
2000
2001 while (_boxing_late_inlines.length() > 0) {
2002 CallGenerator* cg = _boxing_late_inlines.pop();
2003 cg->do_late_inline();
2004 if (failing()) return;
2005 }
2006 _boxing_late_inlines.trunc_to(0);
2007
2008 inline_incrementally_cleanup(igvn);
2009
2010 set_inlining_incrementally(false);
2011 }
2012 }
2013
2014 bool Compile::inline_incrementally_one() {
2015 assert(IncrementalInline, "incremental inlining should be on");
2016
2017 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2018
2019 set_inlining_progress(false);
2020 set_do_cleanup(false);
2021
2022 for (int i = 0; i < _late_inlines.length(); i++) {
2023 _late_inlines_pos = i+1;
2024 CallGenerator* cg = _late_inlines.at(i);
2025 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2026 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2027 cg->do_late_inline();
2028 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2029 if (failing()) {
2030 return false;
2031 } else if (inlining_progress()) {
2032 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2033 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2034 break; // process one call site at a time
2035 }
2036 } else {
2037 // Ignore late inline direct calls when inlining is not allowed.
2038 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2039 }
2040 }
2041 // Remove processed elements.
2042 _late_inlines.remove_till(_late_inlines_pos);
2043 _late_inlines_pos = 0;
2044
2045 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2046
2047 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2048
2049 set_inlining_progress(false);
2050 set_do_cleanup(false);
2051
2052 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2053 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2054 }
2055
2056 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2057 {
2058 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2059 ResourceMark rm;
2060 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2061 }
2062 {
2063 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2064 igvn.reset_from_gvn(initial_gvn());
2065 igvn.optimize();
2066 }
2067 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2068 }
2069
2070 // Perform incremental inlining until bound on number of live nodes is reached
2071 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2072 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2073
2074 set_inlining_incrementally(true);
2075 uint low_live_nodes = 0;
2076
2077 while (_late_inlines.length() > 0) {
2078 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2079 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2080 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2081 // PhaseIdealLoop is expensive so we only try it once we are
2082 // out of live nodes and we only try it again if the previous
2083 // helped got the number of nodes down significantly
2084 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2085 if (failing()) return;
2086 low_live_nodes = live_nodes();
2087 _major_progress = true;
2088 }
2089
2090 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2091 bool do_print_inlining = print_inlining() || print_intrinsics();
2092 if (do_print_inlining || log() != nullptr) {
2093 // Print inlining message for candidates that we couldn't inline for lack of space.
2094 for (int i = 0; i < _late_inlines.length(); i++) {
2095 CallGenerator* cg = _late_inlines.at(i);
2096 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2097 if (do_print_inlining) {
2098 cg->print_inlining_late(msg);
2099 }
2100 log_late_inline_failure(cg, msg);
2101 }
2102 }
2103 break; // finish
2104 }
2105 }
2106
2107 igvn_worklist()->ensure_empty(); // should be done with igvn
2108
2109 while (inline_incrementally_one()) {
2110 assert(!failing(), "inconsistent");
2111 }
2112 if (failing()) return;
2113
2114 inline_incrementally_cleanup(igvn);
2115
2116 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2117
2118 if (failing()) return;
2119
2120 if (_late_inlines.length() == 0) {
2121 break; // no more progress
2122 }
2123 }
2124
2125 igvn_worklist()->ensure_empty(); // should be done with igvn
2126
2127 if (_string_late_inlines.length() > 0) {
2128 assert(has_stringbuilder(), "inconsistent");
2129
2130 inline_string_calls(false);
2131
2132 if (failing()) return;
2133
2134 inline_incrementally_cleanup(igvn);
2135 }
2136
2137 set_inlining_incrementally(false);
2138 }
2139
2140 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2141 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2142 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2143 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2144 // as if "inlining_incrementally() == true" were set.
2145 assert(inlining_incrementally() == false, "not allowed");
2146 assert(_modified_nodes == nullptr, "not allowed");
2147 assert(_late_inlines.length() > 0, "sanity");
2148
2149 while (_late_inlines.length() > 0) {
2150 igvn_worklist()->ensure_empty(); // should be done with igvn
2151
2152 while (inline_incrementally_one()) {
2153 assert(!failing(), "inconsistent");
2154 }
2155 if (failing()) return;
2156
2157 inline_incrementally_cleanup(igvn);
2158 }
2159 }
2160
2161 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2162 if (_loop_opts_cnt > 0) {
2163 while (major_progress() && (_loop_opts_cnt > 0)) {
2164 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2165 PhaseIdealLoop::optimize(igvn, mode);
2166 _loop_opts_cnt--;
2167 if (failing()) return false;
2168 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2169 }
2170 }
2171 return true;
2172 }
2173
2174 // Remove edges from "root" to each SafePoint at a backward branch.
2175 // They were inserted during parsing (see add_safepoint()) to make
2176 // infinite loops without calls or exceptions visible to root, i.e.,
2177 // useful.
2178 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2179 Node *r = root();
2180 if (r != nullptr) {
2181 for (uint i = r->req(); i < r->len(); ++i) {
2182 Node *n = r->in(i);
2183 if (n != nullptr && n->is_SafePoint()) {
2184 r->rm_prec(i);
2185 if (n->outcnt() == 0) {
2186 igvn.remove_dead_node(n);
2187 }
2188 --i;
2189 }
2190 }
2191 // Parsing may have added top inputs to the root node (Path
2192 // leading to the Halt node proven dead). Make sure we get a
2193 // chance to clean them up.
2194 igvn._worklist.push(r);
2195 igvn.optimize();
2196 }
2197 }
2198
2199 //------------------------------Optimize---------------------------------------
2200 // Given a graph, optimize it.
2201 void Compile::Optimize() {
2202 TracePhase tp("optimizer", &timers[_t_optimizer]);
2203
2204 #ifndef PRODUCT
2205 if (env()->break_at_compile()) {
2206 BREAKPOINT;
2207 }
2208
2209 #endif
2210
2211 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2212 #ifdef ASSERT
2213 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2214 #endif
2215
2216 ResourceMark rm;
2217
2218 print_inlining_reinit();
2219
2220 NOT_PRODUCT( verify_graph_edges(); )
2221
2222 print_method(PHASE_AFTER_PARSING, 1);
2223
2224 {
2225 // Iterative Global Value Numbering, including ideal transforms
2226 // Initialize IterGVN with types and values from parse-time GVN
2227 PhaseIterGVN igvn(initial_gvn());
2228 #ifdef ASSERT
2229 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2230 #endif
2231 {
2232 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2233 igvn.optimize();
2234 }
2235
2236 if (failing()) return;
2237
2238 print_method(PHASE_ITER_GVN1, 2);
2239
2240 process_for_unstable_if_traps(igvn);
2241
2242 if (failing()) return;
2243
2244 inline_incrementally(igvn);
2245
2246 print_method(PHASE_INCREMENTAL_INLINE, 2);
2247
2248 if (failing()) return;
2249
2250 if (eliminate_boxing()) {
2251 // Inline valueOf() methods now.
2252 inline_boxing_calls(igvn);
2253
2254 if (failing()) return;
2255
2256 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2257 inline_incrementally(igvn);
2258 }
2259
2260 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2261
2262 if (failing()) return;
2263 }
2264
2265 // Remove the speculative part of types and clean up the graph from
2266 // the extra CastPP nodes whose only purpose is to carry them. Do
2267 // that early so that optimizations are not disrupted by the extra
2268 // CastPP nodes.
2269 remove_speculative_types(igvn);
2270
2271 if (failing()) return;
2272
2273 // No more new expensive nodes will be added to the list from here
2274 // so keep only the actual candidates for optimizations.
2275 cleanup_expensive_nodes(igvn);
2276
2277 if (failing()) return;
2278
2279 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2280 if (EnableVectorSupport && has_vbox_nodes()) {
2281 TracePhase tp("", &timers[_t_vector]);
2282 PhaseVector pv(igvn);
2283 pv.optimize_vector_boxes();
2284 if (failing()) return;
2285 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2286 }
2287 assert(!has_vbox_nodes(), "sanity");
2288
2289 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2290 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2291 igvn_worklist()->ensure_empty(); // should be done with igvn
2292 {
2293 ResourceMark rm;
2294 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2295 }
2296 igvn.reset_from_gvn(initial_gvn());
2297 igvn.optimize();
2298 }
2299
2300 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2301 // safepoints
2302 remove_root_to_sfpts_edges(igvn);
2303
2304 if (failing()) return;
2305
2306 // Perform escape analysis
2307 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2308 if (has_loops()) {
2309 // Cleanup graph (remove dead nodes).
2310 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2311 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2312 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2313 if (failing()) return;
2314 }
2315 bool progress;
2316 do {
2317 ConnectionGraph::do_analysis(this, &igvn);
2318
2319 if (failing()) return;
2320
2321 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2322
2323 // Optimize out fields loads from scalar replaceable allocations.
2324 igvn.optimize();
2325 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2326
2327 if (failing()) return;
2328
2329 if (congraph() != nullptr && macro_count() > 0) {
2330 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2331 PhaseMacroExpand mexp(igvn);
2332 mexp.eliminate_macro_nodes();
2333 igvn.set_delay_transform(false);
2334
2335 igvn.optimize();
2336 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2337
2338 if (failing()) return;
2339 }
2340 progress = do_iterative_escape_analysis() &&
2341 (macro_count() < mcount) &&
2342 ConnectionGraph::has_candidates(this);
2343 // Try again if candidates exist and made progress
2344 // by removing some allocations and/or locks.
2345 } while (progress);
2346 }
2347
2348 // Loop transforms on the ideal graph. Range Check Elimination,
2349 // peeling, unrolling, etc.
2350
2351 // Set loop opts counter
2352 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2353 {
2354 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2355 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2356 _loop_opts_cnt--;
2357 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2358 if (failing()) return;
2359 }
2360 // Loop opts pass if partial peeling occurred in previous pass
2361 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2362 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2363 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2364 _loop_opts_cnt--;
2365 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2366 if (failing()) return;
2367 }
2368 // Loop opts pass for loop-unrolling before CCP
2369 if(major_progress() && (_loop_opts_cnt > 0)) {
2370 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2371 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2372 _loop_opts_cnt--;
2373 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2374 }
2375 if (!failing()) {
2376 // Verify that last round of loop opts produced a valid graph
2377 PhaseIdealLoop::verify(igvn);
2378 }
2379 }
2380 if (failing()) return;
2381
2382 // Conditional Constant Propagation;
2383 PhaseCCP ccp( &igvn );
2384 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2385 {
2386 TracePhase tp("ccp", &timers[_t_ccp]);
2387 ccp.do_transform();
2388 }
2389 print_method(PHASE_CCP1, 2);
2390
2391 assert( true, "Break here to ccp.dump_old2new_map()");
2392
2393 // Iterative Global Value Numbering, including ideal transforms
2394 {
2395 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2396 igvn.reset_from_igvn(&ccp);
2397 igvn.optimize();
2398 }
2399 print_method(PHASE_ITER_GVN2, 2);
2400
2401 if (failing()) return;
2402
2403 // Loop transforms on the ideal graph. Range Check Elimination,
2404 // peeling, unrolling, etc.
2405 if (!optimize_loops(igvn, LoopOptsDefault)) {
2406 return;
2407 }
2408
2409 if (failing()) return;
2410
2411 C->clear_major_progress(); // ensure that major progress is now clear
2412
2413 process_for_post_loop_opts_igvn(igvn);
2414
2415 if (failing()) return;
2416
2417 #ifdef ASSERT
2418 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2419 #endif
2420
2421 {
2422 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2423 PhaseMacroExpand mex(igvn);
2424 if (mex.expand_macro_nodes()) {
2425 assert(failing(), "must bail out w/ explicit message");
2426 return;
2427 }
2428 print_method(PHASE_MACRO_EXPANSION, 2);
2429 }
2430
2431 {
2432 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2433 if (bs->expand_barriers(this, igvn)) {
2434 assert(failing(), "must bail out w/ explicit message");
2435 return;
2436 }
2437 print_method(PHASE_BARRIER_EXPANSION, 2);
2438 }
2439
2440 if (C->max_vector_size() > 0) {
2441 C->optimize_logic_cones(igvn);
2442 igvn.optimize();
2443 }
2444
2445 DEBUG_ONLY( _modified_nodes = nullptr; )
2446
2447 assert(igvn._worklist.size() == 0, "not empty");
2448
2449 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2450
2451 if (_late_inlines.length() > 0) {
2452 // More opportunities to optimize virtual and MH calls.
2453 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2454 process_late_inline_calls_no_inline(igvn);
2455 if (failing()) return;
2456 }
2457 } // (End scope of igvn; run destructor if necessary for asserts.)
2458
2459 check_no_dead_use();
2460
2461 process_print_inlining();
2462
2463 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2464 // to remove hashes to unlock nodes for modifications.
2465 C->node_hash()->clear();
2466
2467 // A method with only infinite loops has no edges entering loops from root
2468 {
2469 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2470 if (final_graph_reshaping()) {
2471 assert(failing(), "must bail out w/ explicit message");
2472 return;
2473 }
2474 }
2475
2476 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2477 DEBUG_ONLY(set_phase_optimize_finished();)
2478 }
2479
2480 #ifdef ASSERT
2481 void Compile::check_no_dead_use() const {
2482 ResourceMark rm;
2483 Unique_Node_List wq;
2484 wq.push(root());
2485 for (uint i = 0; i < wq.size(); ++i) {
2486 Node* n = wq.at(i);
2487 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2488 Node* u = n->fast_out(j);
2489 if (u->outcnt() == 0 && !u->is_Con()) {
2490 u->dump();
2491 fatal("no reachable node should have no use");
2492 }
2493 wq.push(u);
2494 }
2495 }
2496 }
2497 #endif
2498
2499 void Compile::inline_vector_reboxing_calls() {
2500 if (C->_vector_reboxing_late_inlines.length() > 0) {
2501 _late_inlines_pos = C->_late_inlines.length();
2502 while (_vector_reboxing_late_inlines.length() > 0) {
2503 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2504 cg->do_late_inline();
2505 if (failing()) return;
2506 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2507 }
2508 _vector_reboxing_late_inlines.trunc_to(0);
2509 }
2510 }
2511
2512 bool Compile::has_vbox_nodes() {
2513 if (C->_vector_reboxing_late_inlines.length() > 0) {
2514 return true;
2515 }
2516 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2517 Node * n = C->macro_node(macro_idx);
2518 assert(n->is_macro(), "only macro nodes expected here");
2519 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2520 return true;
2521 }
2522 }
2523 return false;
2524 }
2525
2526 //---------------------------- Bitwise operation packing optimization ---------------------------
2527
2528 static bool is_vector_unary_bitwise_op(Node* n) {
2529 return n->Opcode() == Op_XorV &&
2530 VectorNode::is_vector_bitwise_not_pattern(n);
2531 }
2532
2533 static bool is_vector_binary_bitwise_op(Node* n) {
2534 switch (n->Opcode()) {
2535 case Op_AndV:
2536 case Op_OrV:
2537 return true;
2538
2539 case Op_XorV:
2540 return !is_vector_unary_bitwise_op(n);
2541
2542 default:
2543 return false;
2544 }
2545 }
2546
2547 static bool is_vector_ternary_bitwise_op(Node* n) {
2548 return n->Opcode() == Op_MacroLogicV;
2549 }
2550
2551 static bool is_vector_bitwise_op(Node* n) {
2552 return is_vector_unary_bitwise_op(n) ||
2553 is_vector_binary_bitwise_op(n) ||
2554 is_vector_ternary_bitwise_op(n);
2555 }
2556
2557 static bool is_vector_bitwise_cone_root(Node* n) {
2558 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2559 return false;
2560 }
2561 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2562 if (is_vector_bitwise_op(n->fast_out(i))) {
2563 return false;
2564 }
2565 }
2566 return true;
2567 }
2568
2569 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2570 uint cnt = 0;
2571 if (is_vector_bitwise_op(n)) {
2572 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2573 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2574 for (uint i = 1; i < inp_cnt; i++) {
2575 Node* in = n->in(i);
2576 bool skip = VectorNode::is_all_ones_vector(in);
2577 if (!skip && !inputs.member(in)) {
2578 inputs.push(in);
2579 cnt++;
2580 }
2581 }
2582 assert(cnt <= 1, "not unary");
2583 } else {
2584 uint last_req = inp_cnt;
2585 if (is_vector_ternary_bitwise_op(n)) {
2586 last_req = inp_cnt - 1; // skip last input
2587 }
2588 for (uint i = 1; i < last_req; i++) {
2589 Node* def = n->in(i);
2590 if (!inputs.member(def)) {
2591 inputs.push(def);
2592 cnt++;
2593 }
2594 }
2595 }
2596 } else { // not a bitwise operations
2597 if (!inputs.member(n)) {
2598 inputs.push(n);
2599 cnt++;
2600 }
2601 }
2602 return cnt;
2603 }
2604
2605 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2606 Unique_Node_List useful_nodes;
2607 C->identify_useful_nodes(useful_nodes);
2608
2609 for (uint i = 0; i < useful_nodes.size(); i++) {
2610 Node* n = useful_nodes.at(i);
2611 if (is_vector_bitwise_cone_root(n)) {
2612 list.push(n);
2613 }
2614 }
2615 }
2616
2617 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2618 const TypeVect* vt,
2619 Unique_Node_List& partition,
2620 Unique_Node_List& inputs) {
2621 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2622 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2623 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2624
2625 Node* in1 = inputs.at(0);
2626 Node* in2 = inputs.at(1);
2627 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2628
2629 uint func = compute_truth_table(partition, inputs);
2630
2631 Node* pn = partition.at(partition.size() - 1);
2632 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2633 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2634 }
2635
2636 static uint extract_bit(uint func, uint pos) {
2637 return (func & (1 << pos)) >> pos;
2638 }
2639
2640 //
2641 // A macro logic node represents a truth table. It has 4 inputs,
2642 // First three inputs corresponds to 3 columns of a truth table
2643 // and fourth input captures the logic function.
2644 //
2645 // eg. fn = (in1 AND in2) OR in3;
2646 //
2647 // MacroNode(in1,in2,in3,fn)
2648 //
2649 // -----------------
2650 // in1 in2 in3 fn
2651 // -----------------
2652 // 0 0 0 0
2653 // 0 0 1 1
2654 // 0 1 0 0
2655 // 0 1 1 1
2656 // 1 0 0 0
2657 // 1 0 1 1
2658 // 1 1 0 1
2659 // 1 1 1 1
2660 //
2661
2662 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2663 int res = 0;
2664 for (int i = 0; i < 8; i++) {
2665 int bit1 = extract_bit(in1, i);
2666 int bit2 = extract_bit(in2, i);
2667 int bit3 = extract_bit(in3, i);
2668
2669 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2670 int func_bit = extract_bit(func, func_bit_pos);
2671
2672 res |= func_bit << i;
2673 }
2674 return res;
2675 }
2676
2677 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2678 assert(n != nullptr, "");
2679 assert(eval_map.contains(n), "absent");
2680 return *(eval_map.get(n));
2681 }
2682
2683 static void eval_operands(Node* n,
2684 uint& func1, uint& func2, uint& func3,
2685 ResourceHashtable<Node*,uint>& eval_map) {
2686 assert(is_vector_bitwise_op(n), "");
2687
2688 if (is_vector_unary_bitwise_op(n)) {
2689 Node* opnd = n->in(1);
2690 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2691 opnd = n->in(2);
2692 }
2693 func1 = eval_operand(opnd, eval_map);
2694 } else if (is_vector_binary_bitwise_op(n)) {
2695 func1 = eval_operand(n->in(1), eval_map);
2696 func2 = eval_operand(n->in(2), eval_map);
2697 } else {
2698 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2699 func1 = eval_operand(n->in(1), eval_map);
2700 func2 = eval_operand(n->in(2), eval_map);
2701 func3 = eval_operand(n->in(3), eval_map);
2702 }
2703 }
2704
2705 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2706 assert(inputs.size() <= 3, "sanity");
2707 ResourceMark rm;
2708 uint res = 0;
2709 ResourceHashtable<Node*,uint> eval_map;
2710
2711 // Populate precomputed functions for inputs.
2712 // Each input corresponds to one column of 3 input truth-table.
2713 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2714 0xCC, // (_, b, _) -> b
2715 0xF0 }; // (a, _, _) -> a
2716 for (uint i = 0; i < inputs.size(); i++) {
2717 eval_map.put(inputs.at(i), input_funcs[2-i]);
2718 }
2719
2720 for (uint i = 0; i < partition.size(); i++) {
2721 Node* n = partition.at(i);
2722
2723 uint func1 = 0, func2 = 0, func3 = 0;
2724 eval_operands(n, func1, func2, func3, eval_map);
2725
2726 switch (n->Opcode()) {
2727 case Op_OrV:
2728 assert(func3 == 0, "not binary");
2729 res = func1 | func2;
2730 break;
2731 case Op_AndV:
2732 assert(func3 == 0, "not binary");
2733 res = func1 & func2;
2734 break;
2735 case Op_XorV:
2736 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2737 assert(func2 == 0 && func3 == 0, "not unary");
2738 res = (~func1) & 0xFF;
2739 } else {
2740 assert(func3 == 0, "not binary");
2741 res = func1 ^ func2;
2742 }
2743 break;
2744 case Op_MacroLogicV:
2745 // Ordering of inputs may change during evaluation of sub-tree
2746 // containing MacroLogic node as a child node, thus a re-evaluation
2747 // makes sure that function is evaluated in context of current
2748 // inputs.
2749 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2750 break;
2751
2752 default: assert(false, "not supported: %s", n->Name());
2753 }
2754 assert(res <= 0xFF, "invalid");
2755 eval_map.put(n, res);
2756 }
2757 return res;
2758 }
2759
2760 // Criteria under which nodes gets packed into a macro logic node:-
2761 // 1) Parent and both child nodes are all unmasked or masked with
2762 // same predicates.
2763 // 2) Masked parent can be packed with left child if it is predicated
2764 // and both have same predicates.
2765 // 3) Masked parent can be packed with right child if its un-predicated
2766 // or has matching predication condition.
2767 // 4) An unmasked parent can be packed with an unmasked child.
2768 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2769 assert(partition.size() == 0, "not empty");
2770 assert(inputs.size() == 0, "not empty");
2771 if (is_vector_ternary_bitwise_op(n)) {
2772 return false;
2773 }
2774
2775 bool is_unary_op = is_vector_unary_bitwise_op(n);
2776 if (is_unary_op) {
2777 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2778 return false; // too few inputs
2779 }
2780
2781 bool pack_left_child = true;
2782 bool pack_right_child = true;
2783
2784 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2785 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2786
2787 int left_child_input_cnt = 0;
2788 int right_child_input_cnt = 0;
2789
2790 bool parent_is_predicated = n->is_predicated_vector();
2791 bool left_child_predicated = n->in(1)->is_predicated_vector();
2792 bool right_child_predicated = n->in(2)->is_predicated_vector();
2793
2794 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2795 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2796 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2797
2798 do {
2799 if (pack_left_child && left_child_LOP &&
2800 ((!parent_is_predicated && !left_child_predicated) ||
2801 ((parent_is_predicated && left_child_predicated &&
2802 parent_pred == left_child_pred)))) {
2803 partition.push(n->in(1));
2804 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2805 } else {
2806 inputs.push(n->in(1));
2807 left_child_input_cnt = 1;
2808 }
2809
2810 if (pack_right_child && right_child_LOP &&
2811 (!right_child_predicated ||
2812 (right_child_predicated && parent_is_predicated &&
2813 parent_pred == right_child_pred))) {
2814 partition.push(n->in(2));
2815 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2816 } else {
2817 inputs.push(n->in(2));
2818 right_child_input_cnt = 1;
2819 }
2820
2821 if (inputs.size() > 3) {
2822 assert(partition.size() > 0, "");
2823 inputs.clear();
2824 partition.clear();
2825 if (left_child_input_cnt > right_child_input_cnt) {
2826 pack_left_child = false;
2827 } else {
2828 pack_right_child = false;
2829 }
2830 } else {
2831 break;
2832 }
2833 } while(true);
2834
2835 if(partition.size()) {
2836 partition.push(n);
2837 }
2838
2839 return (partition.size() == 2 || partition.size() == 3) &&
2840 (inputs.size() == 2 || inputs.size() == 3);
2841 }
2842
2843 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2844 assert(is_vector_bitwise_op(n), "not a root");
2845
2846 visited.set(n->_idx);
2847
2848 // 1) Do a DFS walk over the logic cone.
2849 for (uint i = 1; i < n->req(); i++) {
2850 Node* in = n->in(i);
2851 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2852 process_logic_cone_root(igvn, in, visited);
2853 }
2854 }
2855
2856 // 2) Bottom up traversal: Merge node[s] with
2857 // the parent to form macro logic node.
2858 Unique_Node_List partition;
2859 Unique_Node_List inputs;
2860 if (compute_logic_cone(n, partition, inputs)) {
2861 const TypeVect* vt = n->bottom_type()->is_vect();
2862 Node* pn = partition.at(partition.size() - 1);
2863 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2864 if (mask == nullptr ||
2865 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2866 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2867 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2868 igvn.replace_node(n, macro_logic);
2869 }
2870 }
2871 }
2872
2873 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2874 ResourceMark rm;
2875 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2876 Unique_Node_List list;
2877 collect_logic_cone_roots(list);
2878
2879 while (list.size() > 0) {
2880 Node* n = list.pop();
2881 const TypeVect* vt = n->bottom_type()->is_vect();
2882 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2883 if (supported) {
2884 VectorSet visited(comp_arena());
2885 process_logic_cone_root(igvn, n, visited);
2886 }
2887 }
2888 }
2889 }
2890
2891 //------------------------------Code_Gen---------------------------------------
2892 // Given a graph, generate code for it
2893 void Compile::Code_Gen() {
2894 if (failing()) {
2895 return;
2896 }
2897
2898 // Perform instruction selection. You might think we could reclaim Matcher
2899 // memory PDQ, but actually the Matcher is used in generating spill code.
2900 // Internals of the Matcher (including some VectorSets) must remain live
2901 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2902 // set a bit in reclaimed memory.
2903
2904 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2905 // nodes. Mapping is only valid at the root of each matched subtree.
2906 NOT_PRODUCT( verify_graph_edges(); )
2907
2908 Matcher matcher;
2909 _matcher = &matcher;
2910 {
2911 TracePhase tp("matcher", &timers[_t_matcher]);
2912 matcher.match();
2913 if (failing()) {
2914 return;
2915 }
2916 }
2917 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2918 // nodes. Mapping is only valid at the root of each matched subtree.
2919 NOT_PRODUCT( verify_graph_edges(); )
2920
2921 // If you have too many nodes, or if matching has failed, bail out
2922 check_node_count(0, "out of nodes matching instructions");
2923 if (failing()) {
2924 return;
2925 }
2926
2927 print_method(PHASE_MATCHING, 2);
2928
2929 // Build a proper-looking CFG
2930 PhaseCFG cfg(node_arena(), root(), matcher);
2931 _cfg = &cfg;
2932 {
2933 TracePhase tp("scheduler", &timers[_t_scheduler]);
2934 bool success = cfg.do_global_code_motion();
2935 if (!success) {
2936 return;
2937 }
2938
2939 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2940 NOT_PRODUCT( verify_graph_edges(); )
2941 cfg.verify();
2942 }
2943
2944 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2945 _regalloc = ®alloc;
2946 {
2947 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2948 // Perform register allocation. After Chaitin, use-def chains are
2949 // no longer accurate (at spill code) and so must be ignored.
2950 // Node->LRG->reg mappings are still accurate.
2951 _regalloc->Register_Allocate();
2952
2953 // Bail out if the allocator builds too many nodes
2954 if (failing()) {
2955 return;
2956 }
2957 }
2958
2959 // Prior to register allocation we kept empty basic blocks in case the
2960 // the allocator needed a place to spill. After register allocation we
2961 // are not adding any new instructions. If any basic block is empty, we
2962 // can now safely remove it.
2963 {
2964 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2965 cfg.remove_empty_blocks();
2966 if (do_freq_based_layout()) {
2967 PhaseBlockLayout layout(cfg);
2968 } else {
2969 cfg.set_loop_alignment();
2970 }
2971 cfg.fixup_flow();
2972 cfg.remove_unreachable_blocks();
2973 cfg.verify_dominator_tree();
2974 }
2975
2976 // Apply peephole optimizations
2977 if( OptoPeephole ) {
2978 TracePhase tp("peephole", &timers[_t_peephole]);
2979 PhasePeephole peep( _regalloc, cfg);
2980 peep.do_transform();
2981 }
2982
2983 // Do late expand if CPU requires this.
2984 if (Matcher::require_postalloc_expand) {
2985 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2986 cfg.postalloc_expand(_regalloc);
2987 }
2988
2989 // Convert Nodes to instruction bits in a buffer
2990 {
2991 TracePhase tp("output", &timers[_t_output]);
2992 PhaseOutput output;
2993 output.Output();
2994 if (failing()) return;
2995 output.install();
2996 }
2997
2998 print_method(PHASE_FINAL_CODE, 1);
2999
3000 // He's dead, Jim.
3001 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3002 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3003 }
3004
3005 //------------------------------Final_Reshape_Counts---------------------------
3006 // This class defines counters to help identify when a method
3007 // may/must be executed using hardware with only 24-bit precision.
3008 struct Final_Reshape_Counts : public StackObj {
3009 int _call_count; // count non-inlined 'common' calls
3010 int _float_count; // count float ops requiring 24-bit precision
3011 int _double_count; // count double ops requiring more precision
3012 int _java_call_count; // count non-inlined 'java' calls
3013 int _inner_loop_count; // count loops which need alignment
3014 VectorSet _visited; // Visitation flags
3015 Node_List _tests; // Set of IfNodes & PCTableNodes
3016
3017 Final_Reshape_Counts() :
3018 _call_count(0), _float_count(0), _double_count(0),
3019 _java_call_count(0), _inner_loop_count(0) { }
3020
3021 void inc_call_count () { _call_count ++; }
3022 void inc_float_count () { _float_count ++; }
3023 void inc_double_count() { _double_count++; }
3024 void inc_java_call_count() { _java_call_count++; }
3025 void inc_inner_loop_count() { _inner_loop_count++; }
3026
3027 int get_call_count () const { return _call_count ; }
3028 int get_float_count () const { return _float_count ; }
3029 int get_double_count() const { return _double_count; }
3030 int get_java_call_count() const { return _java_call_count; }
3031 int get_inner_loop_count() const { return _inner_loop_count; }
3032 };
3033
3034 // Eliminate trivially redundant StoreCMs and accumulate their
3035 // precedence edges.
3036 void Compile::eliminate_redundant_card_marks(Node* n) {
3037 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3038 if (n->in(MemNode::Address)->outcnt() > 1) {
3039 // There are multiple users of the same address so it might be
3040 // possible to eliminate some of the StoreCMs
3041 Node* mem = n->in(MemNode::Memory);
3042 Node* adr = n->in(MemNode::Address);
3043 Node* val = n->in(MemNode::ValueIn);
3044 Node* prev = n;
3045 bool done = false;
3046 // Walk the chain of StoreCMs eliminating ones that match. As
3047 // long as it's a chain of single users then the optimization is
3048 // safe. Eliminating partially redundant StoreCMs would require
3049 // cloning copies down the other paths.
3050 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3051 if (adr == mem->in(MemNode::Address) &&
3052 val == mem->in(MemNode::ValueIn)) {
3053 // redundant StoreCM
3054 if (mem->req() > MemNode::OopStore) {
3055 // Hasn't been processed by this code yet.
3056 n->add_prec(mem->in(MemNode::OopStore));
3057 } else {
3058 // Already converted to precedence edge
3059 for (uint i = mem->req(); i < mem->len(); i++) {
3060 // Accumulate any precedence edges
3061 if (mem->in(i) != nullptr) {
3062 n->add_prec(mem->in(i));
3063 }
3064 }
3065 // Everything above this point has been processed.
3066 done = true;
3067 }
3068 // Eliminate the previous StoreCM
3069 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3070 assert(mem->outcnt() == 0, "should be dead");
3071 mem->disconnect_inputs(this);
3072 } else {
3073 prev = mem;
3074 }
3075 mem = prev->in(MemNode::Memory);
3076 }
3077 }
3078 }
3079
3080 //------------------------------final_graph_reshaping_impl----------------------
3081 // Implement items 1-5 from final_graph_reshaping below.
3082 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3083
3084 if ( n->outcnt() == 0 ) return; // dead node
3085 uint nop = n->Opcode();
3086
3087 // Check for 2-input instruction with "last use" on right input.
3088 // Swap to left input. Implements item (2).
3089 if( n->req() == 3 && // two-input instruction
3090 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3091 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3092 n->in(2)->outcnt() == 1 &&// right use IS a last use
3093 !n->in(2)->is_Con() ) { // right use is not a constant
3094 // Check for commutative opcode
3095 switch( nop ) {
3096 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3097 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3098 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3099 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3100 case Op_AndL: case Op_XorL: case Op_OrL:
3101 case Op_AndI: case Op_XorI: case Op_OrI: {
3102 // Move "last use" input to left by swapping inputs
3103 n->swap_edges(1, 2);
3104 break;
3105 }
3106 default:
3107 break;
3108 }
3109 }
3110
3111 #ifdef ASSERT
3112 if( n->is_Mem() ) {
3113 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3114 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3115 // oop will be recorded in oop map if load crosses safepoint
3116 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3117 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3118 "raw memory operations should have control edge");
3119 }
3120 if (n->is_MemBar()) {
3121 MemBarNode* mb = n->as_MemBar();
3122 if (mb->trailing_store() || mb->trailing_load_store()) {
3123 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3124 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3125 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3126 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3127 } else if (mb->leading()) {
3128 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3129 }
3130 }
3131 #endif
3132 // Count FPU ops and common calls, implements item (3)
3133 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3134 if (!gc_handled) {
3135 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3136 }
3137
3138 // Collect CFG split points
3139 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3140 frc._tests.push(n);
3141 }
3142 }
3143
3144 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3145 switch( nop ) {
3146 // Count all float operations that may use FPU
3147 case Op_AddF:
3148 case Op_SubF:
3149 case Op_MulF:
3150 case Op_DivF:
3151 case Op_NegF:
3152 case Op_ModF:
3153 case Op_ConvI2F:
3154 case Op_ConF:
3155 case Op_CmpF:
3156 case Op_CmpF3:
3157 case Op_StoreF:
3158 case Op_LoadF:
3159 // case Op_ConvL2F: // longs are split into 32-bit halves
3160 frc.inc_float_count();
3161 break;
3162
3163 case Op_ConvF2D:
3164 case Op_ConvD2F:
3165 frc.inc_float_count();
3166 frc.inc_double_count();
3167 break;
3168
3169 // Count all double operations that may use FPU
3170 case Op_AddD:
3171 case Op_SubD:
3172 case Op_MulD:
3173 case Op_DivD:
3174 case Op_NegD:
3175 case Op_ModD:
3176 case Op_ConvI2D:
3177 case Op_ConvD2I:
3178 // case Op_ConvL2D: // handled by leaf call
3179 // case Op_ConvD2L: // handled by leaf call
3180 case Op_ConD:
3181 case Op_CmpD:
3182 case Op_CmpD3:
3183 case Op_StoreD:
3184 case Op_LoadD:
3185 case Op_LoadD_unaligned:
3186 frc.inc_double_count();
3187 break;
3188 case Op_Opaque1: // Remove Opaque Nodes before matching
3189 case Op_Opaque3:
3190 n->subsume_by(n->in(1), this);
3191 break;
3192 case Op_CallStaticJava:
3193 case Op_CallJava:
3194 case Op_CallDynamicJava:
3195 frc.inc_java_call_count(); // Count java call site;
3196 case Op_CallRuntime:
3197 case Op_CallLeaf:
3198 case Op_CallLeafVector:
3199 case Op_CallLeafNoFP: {
3200 assert (n->is_Call(), "");
3201 CallNode *call = n->as_Call();
3202 // Count call sites where the FP mode bit would have to be flipped.
3203 // Do not count uncommon runtime calls:
3204 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3205 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3206 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3207 frc.inc_call_count(); // Count the call site
3208 } else { // See if uncommon argument is shared
3209 Node *n = call->in(TypeFunc::Parms);
3210 int nop = n->Opcode();
3211 // Clone shared simple arguments to uncommon calls, item (1).
3212 if (n->outcnt() > 1 &&
3213 !n->is_Proj() &&
3214 nop != Op_CreateEx &&
3215 nop != Op_CheckCastPP &&
3216 nop != Op_DecodeN &&
3217 nop != Op_DecodeNKlass &&
3218 !n->is_Mem() &&
3219 !n->is_Phi()) {
3220 Node *x = n->clone();
3221 call->set_req(TypeFunc::Parms, x);
3222 }
3223 }
3224 break;
3225 }
3226
3227 case Op_StoreCM:
3228 {
3229 // Convert OopStore dependence into precedence edge
3230 Node* prec = n->in(MemNode::OopStore);
3231 n->del_req(MemNode::OopStore);
3232 n->add_prec(prec);
3233 eliminate_redundant_card_marks(n);
3234 }
3235
3236 // fall through
3237
3238 case Op_StoreB:
3239 case Op_StoreC:
3240 case Op_StoreI:
3241 case Op_StoreL:
3242 case Op_CompareAndSwapB:
3243 case Op_CompareAndSwapS:
3244 case Op_CompareAndSwapI:
3245 case Op_CompareAndSwapL:
3246 case Op_CompareAndSwapP:
3247 case Op_CompareAndSwapN:
3248 case Op_WeakCompareAndSwapB:
3249 case Op_WeakCompareAndSwapS:
3250 case Op_WeakCompareAndSwapI:
3251 case Op_WeakCompareAndSwapL:
3252 case Op_WeakCompareAndSwapP:
3253 case Op_WeakCompareAndSwapN:
3254 case Op_CompareAndExchangeB:
3255 case Op_CompareAndExchangeS:
3256 case Op_CompareAndExchangeI:
3257 case Op_CompareAndExchangeL:
3258 case Op_CompareAndExchangeP:
3259 case Op_CompareAndExchangeN:
3260 case Op_GetAndAddS:
3261 case Op_GetAndAddB:
3262 case Op_GetAndAddI:
3263 case Op_GetAndAddL:
3264 case Op_GetAndSetS:
3265 case Op_GetAndSetB:
3266 case Op_GetAndSetI:
3267 case Op_GetAndSetL:
3268 case Op_GetAndSetP:
3269 case Op_GetAndSetN:
3270 case Op_StoreP:
3271 case Op_StoreN:
3272 case Op_StoreNKlass:
3273 case Op_LoadB:
3274 case Op_LoadUB:
3275 case Op_LoadUS:
3276 case Op_LoadI:
3277 case Op_LoadKlass:
3278 case Op_LoadNKlass:
3279 case Op_LoadL:
3280 case Op_LoadL_unaligned:
3281 case Op_LoadP:
3282 case Op_LoadN:
3283 case Op_LoadRange:
3284 case Op_LoadS:
3285 break;
3286
3287 case Op_AddP: { // Assert sane base pointers
3288 Node *addp = n->in(AddPNode::Address);
3289 assert( !addp->is_AddP() ||
3290 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3291 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3292 "Base pointers must match (addp %u)", addp->_idx );
3293 #ifdef _LP64
3294 if ((UseCompressedOops || UseCompressedClassPointers) &&
3295 addp->Opcode() == Op_ConP &&
3296 addp == n->in(AddPNode::Base) &&
3297 n->in(AddPNode::Offset)->is_Con()) {
3298 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3299 // on the platform and on the compressed oops mode.
3300 // Use addressing with narrow klass to load with offset on x86.
3301 // Some platforms can use the constant pool to load ConP.
3302 // Do this transformation here since IGVN will convert ConN back to ConP.
3303 const Type* t = addp->bottom_type();
3304 bool is_oop = t->isa_oopptr() != nullptr;
3305 bool is_klass = t->isa_klassptr() != nullptr;
3306
3307 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3308 (is_klass && Matcher::const_klass_prefer_decode())) {
3309 Node* nn = nullptr;
3310
3311 int op = is_oop ? Op_ConN : Op_ConNKlass;
3312
3313 // Look for existing ConN node of the same exact type.
3314 Node* r = root();
3315 uint cnt = r->outcnt();
3316 for (uint i = 0; i < cnt; i++) {
3317 Node* m = r->raw_out(i);
3318 if (m!= nullptr && m->Opcode() == op &&
3319 m->bottom_type()->make_ptr() == t) {
3320 nn = m;
3321 break;
3322 }
3323 }
3324 if (nn != nullptr) {
3325 // Decode a narrow oop to match address
3326 // [R12 + narrow_oop_reg<<3 + offset]
3327 if (is_oop) {
3328 nn = new DecodeNNode(nn, t);
3329 } else {
3330 nn = new DecodeNKlassNode(nn, t);
3331 }
3332 // Check for succeeding AddP which uses the same Base.
3333 // Otherwise we will run into the assertion above when visiting that guy.
3334 for (uint i = 0; i < n->outcnt(); ++i) {
3335 Node *out_i = n->raw_out(i);
3336 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3337 out_i->set_req(AddPNode::Base, nn);
3338 #ifdef ASSERT
3339 for (uint j = 0; j < out_i->outcnt(); ++j) {
3340 Node *out_j = out_i->raw_out(j);
3341 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3342 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3343 }
3344 #endif
3345 }
3346 }
3347 n->set_req(AddPNode::Base, nn);
3348 n->set_req(AddPNode::Address, nn);
3349 if (addp->outcnt() == 0) {
3350 addp->disconnect_inputs(this);
3351 }
3352 }
3353 }
3354 }
3355 #endif
3356 break;
3357 }
3358
3359 case Op_CastPP: {
3360 // Remove CastPP nodes to gain more freedom during scheduling but
3361 // keep the dependency they encode as control or precedence edges
3362 // (if control is set already) on memory operations. Some CastPP
3363 // nodes don't have a control (don't carry a dependency): skip
3364 // those.
3365 if (n->in(0) != nullptr) {
3366 ResourceMark rm;
3367 Unique_Node_List wq;
3368 wq.push(n);
3369 for (uint next = 0; next < wq.size(); ++next) {
3370 Node *m = wq.at(next);
3371 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3372 Node* use = m->fast_out(i);
3373 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3374 use->ensure_control_or_add_prec(n->in(0));
3375 } else {
3376 switch(use->Opcode()) {
3377 case Op_AddP:
3378 case Op_DecodeN:
3379 case Op_DecodeNKlass:
3380 case Op_CheckCastPP:
3381 case Op_CastPP:
3382 wq.push(use);
3383 break;
3384 }
3385 }
3386 }
3387 }
3388 }
3389 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3390 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3391 Node* in1 = n->in(1);
3392 const Type* t = n->bottom_type();
3393 Node* new_in1 = in1->clone();
3394 new_in1->as_DecodeN()->set_type(t);
3395
3396 if (!Matcher::narrow_oop_use_complex_address()) {
3397 //
3398 // x86, ARM and friends can handle 2 adds in addressing mode
3399 // and Matcher can fold a DecodeN node into address by using
3400 // a narrow oop directly and do implicit null check in address:
3401 //
3402 // [R12 + narrow_oop_reg<<3 + offset]
3403 // NullCheck narrow_oop_reg
3404 //
3405 // On other platforms (Sparc) we have to keep new DecodeN node and
3406 // use it to do implicit null check in address:
3407 //
3408 // decode_not_null narrow_oop_reg, base_reg
3409 // [base_reg + offset]
3410 // NullCheck base_reg
3411 //
3412 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3413 // to keep the information to which null check the new DecodeN node
3414 // corresponds to use it as value in implicit_null_check().
3415 //
3416 new_in1->set_req(0, n->in(0));
3417 }
3418
3419 n->subsume_by(new_in1, this);
3420 if (in1->outcnt() == 0) {
3421 in1->disconnect_inputs(this);
3422 }
3423 } else {
3424 n->subsume_by(n->in(1), this);
3425 if (n->outcnt() == 0) {
3426 n->disconnect_inputs(this);
3427 }
3428 }
3429 break;
3430 }
3431 #ifdef _LP64
3432 case Op_CmpP:
3433 // Do this transformation here to preserve CmpPNode::sub() and
3434 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3435 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3436 Node* in1 = n->in(1);
3437 Node* in2 = n->in(2);
3438 if (!in1->is_DecodeNarrowPtr()) {
3439 in2 = in1;
3440 in1 = n->in(2);
3441 }
3442 assert(in1->is_DecodeNarrowPtr(), "sanity");
3443
3444 Node* new_in2 = nullptr;
3445 if (in2->is_DecodeNarrowPtr()) {
3446 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3447 new_in2 = in2->in(1);
3448 } else if (in2->Opcode() == Op_ConP) {
3449 const Type* t = in2->bottom_type();
3450 if (t == TypePtr::NULL_PTR) {
3451 assert(in1->is_DecodeN(), "compare klass to null?");
3452 // Don't convert CmpP null check into CmpN if compressed
3453 // oops implicit null check is not generated.
3454 // This will allow to generate normal oop implicit null check.
3455 if (Matcher::gen_narrow_oop_implicit_null_checks())
3456 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3457 //
3458 // This transformation together with CastPP transformation above
3459 // will generated code for implicit null checks for compressed oops.
3460 //
3461 // The original code after Optimize()
3462 //
3463 // LoadN memory, narrow_oop_reg
3464 // decode narrow_oop_reg, base_reg
3465 // CmpP base_reg, nullptr
3466 // CastPP base_reg // NotNull
3467 // Load [base_reg + offset], val_reg
3468 //
3469 // after these transformations will be
3470 //
3471 // LoadN memory, narrow_oop_reg
3472 // CmpN narrow_oop_reg, nullptr
3473 // decode_not_null narrow_oop_reg, base_reg
3474 // Load [base_reg + offset], val_reg
3475 //
3476 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3477 // since narrow oops can be used in debug info now (see the code in
3478 // final_graph_reshaping_walk()).
3479 //
3480 // At the end the code will be matched to
3481 // on x86:
3482 //
3483 // Load_narrow_oop memory, narrow_oop_reg
3484 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3485 // NullCheck narrow_oop_reg
3486 //
3487 // and on sparc:
3488 //
3489 // Load_narrow_oop memory, narrow_oop_reg
3490 // decode_not_null narrow_oop_reg, base_reg
3491 // Load [base_reg + offset], val_reg
3492 // NullCheck base_reg
3493 //
3494 } else if (t->isa_oopptr()) {
3495 new_in2 = ConNode::make(t->make_narrowoop());
3496 } else if (t->isa_klassptr()) {
3497 new_in2 = ConNode::make(t->make_narrowklass());
3498 }
3499 }
3500 if (new_in2 != nullptr) {
3501 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3502 n->subsume_by(cmpN, this);
3503 if (in1->outcnt() == 0) {
3504 in1->disconnect_inputs(this);
3505 }
3506 if (in2->outcnt() == 0) {
3507 in2->disconnect_inputs(this);
3508 }
3509 }
3510 }
3511 break;
3512
3513 case Op_DecodeN:
3514 case Op_DecodeNKlass:
3515 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3516 // DecodeN could be pinned when it can't be fold into
3517 // an address expression, see the code for Op_CastPP above.
3518 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3519 break;
3520
3521 case Op_EncodeP:
3522 case Op_EncodePKlass: {
3523 Node* in1 = n->in(1);
3524 if (in1->is_DecodeNarrowPtr()) {
3525 n->subsume_by(in1->in(1), this);
3526 } else if (in1->Opcode() == Op_ConP) {
3527 const Type* t = in1->bottom_type();
3528 if (t == TypePtr::NULL_PTR) {
3529 assert(t->isa_oopptr(), "null klass?");
3530 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3531 } else if (t->isa_oopptr()) {
3532 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3533 } else if (t->isa_klassptr()) {
3534 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3535 }
3536 }
3537 if (in1->outcnt() == 0) {
3538 in1->disconnect_inputs(this);
3539 }
3540 break;
3541 }
3542
3543 case Op_Proj: {
3544 if (OptimizeStringConcat || IncrementalInline) {
3545 ProjNode* proj = n->as_Proj();
3546 if (proj->_is_io_use) {
3547 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3548 // Separate projections were used for the exception path which
3549 // are normally removed by a late inline. If it wasn't inlined
3550 // then they will hang around and should just be replaced with
3551 // the original one. Merge them.
3552 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3553 if (non_io_proj != nullptr) {
3554 proj->subsume_by(non_io_proj , this);
3555 }
3556 }
3557 }
3558 break;
3559 }
3560
3561 case Op_Phi:
3562 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3563 // The EncodeP optimization may create Phi with the same edges
3564 // for all paths. It is not handled well by Register Allocator.
3565 Node* unique_in = n->in(1);
3566 assert(unique_in != nullptr, "");
3567 uint cnt = n->req();
3568 for (uint i = 2; i < cnt; i++) {
3569 Node* m = n->in(i);
3570 assert(m != nullptr, "");
3571 if (unique_in != m)
3572 unique_in = nullptr;
3573 }
3574 if (unique_in != nullptr) {
3575 n->subsume_by(unique_in, this);
3576 }
3577 }
3578 break;
3579
3580 #endif
3581
3582 #ifdef ASSERT
3583 case Op_CastII:
3584 // Verify that all range check dependent CastII nodes were removed.
3585 if (n->isa_CastII()->has_range_check()) {
3586 n->dump(3);
3587 assert(false, "Range check dependent CastII node was not removed");
3588 }
3589 break;
3590 #endif
3591
3592 case Op_ModI:
3593 if (UseDivMod) {
3594 // Check if a%b and a/b both exist
3595 Node* d = n->find_similar(Op_DivI);
3596 if (d) {
3597 // Replace them with a fused divmod if supported
3598 if (Matcher::has_match_rule(Op_DivModI)) {
3599 DivModINode* divmod = DivModINode::make(n);
3600 d->subsume_by(divmod->div_proj(), this);
3601 n->subsume_by(divmod->mod_proj(), this);
3602 } else {
3603 // replace a%b with a-((a/b)*b)
3604 Node* mult = new MulINode(d, d->in(2));
3605 Node* sub = new SubINode(d->in(1), mult);
3606 n->subsume_by(sub, this);
3607 }
3608 }
3609 }
3610 break;
3611
3612 case Op_ModL:
3613 if (UseDivMod) {
3614 // Check if a%b and a/b both exist
3615 Node* d = n->find_similar(Op_DivL);
3616 if (d) {
3617 // Replace them with a fused divmod if supported
3618 if (Matcher::has_match_rule(Op_DivModL)) {
3619 DivModLNode* divmod = DivModLNode::make(n);
3620 d->subsume_by(divmod->div_proj(), this);
3621 n->subsume_by(divmod->mod_proj(), this);
3622 } else {
3623 // replace a%b with a-((a/b)*b)
3624 Node* mult = new MulLNode(d, d->in(2));
3625 Node* sub = new SubLNode(d->in(1), mult);
3626 n->subsume_by(sub, this);
3627 }
3628 }
3629 }
3630 break;
3631
3632 case Op_UModI:
3633 if (UseDivMod) {
3634 // Check if a%b and a/b both exist
3635 Node* d = n->find_similar(Op_UDivI);
3636 if (d) {
3637 // Replace them with a fused unsigned divmod if supported
3638 if (Matcher::has_match_rule(Op_UDivModI)) {
3639 UDivModINode* divmod = UDivModINode::make(n);
3640 d->subsume_by(divmod->div_proj(), this);
3641 n->subsume_by(divmod->mod_proj(), this);
3642 } else {
3643 // replace a%b with a-((a/b)*b)
3644 Node* mult = new MulINode(d, d->in(2));
3645 Node* sub = new SubINode(d->in(1), mult);
3646 n->subsume_by(sub, this);
3647 }
3648 }
3649 }
3650 break;
3651
3652 case Op_UModL:
3653 if (UseDivMod) {
3654 // Check if a%b and a/b both exist
3655 Node* d = n->find_similar(Op_UDivL);
3656 if (d) {
3657 // Replace them with a fused unsigned divmod if supported
3658 if (Matcher::has_match_rule(Op_UDivModL)) {
3659 UDivModLNode* divmod = UDivModLNode::make(n);
3660 d->subsume_by(divmod->div_proj(), this);
3661 n->subsume_by(divmod->mod_proj(), this);
3662 } else {
3663 // replace a%b with a-((a/b)*b)
3664 Node* mult = new MulLNode(d, d->in(2));
3665 Node* sub = new SubLNode(d->in(1), mult);
3666 n->subsume_by(sub, this);
3667 }
3668 }
3669 }
3670 break;
3671
3672 case Op_LoadVector:
3673 case Op_StoreVector:
3674 case Op_LoadVectorGather:
3675 case Op_StoreVectorScatter:
3676 case Op_LoadVectorGatherMasked:
3677 case Op_StoreVectorScatterMasked:
3678 case Op_VectorCmpMasked:
3679 case Op_VectorMaskGen:
3680 case Op_LoadVectorMasked:
3681 case Op_StoreVectorMasked:
3682 break;
3683
3684 case Op_AddReductionVI:
3685 case Op_AddReductionVL:
3686 case Op_AddReductionVF:
3687 case Op_AddReductionVD:
3688 case Op_MulReductionVI:
3689 case Op_MulReductionVL:
3690 case Op_MulReductionVF:
3691 case Op_MulReductionVD:
3692 case Op_MinReductionV:
3693 case Op_MaxReductionV:
3694 case Op_AndReductionV:
3695 case Op_OrReductionV:
3696 case Op_XorReductionV:
3697 break;
3698
3699 case Op_PackB:
3700 case Op_PackS:
3701 case Op_PackI:
3702 case Op_PackF:
3703 case Op_PackL:
3704 case Op_PackD:
3705 if (n->req()-1 > 2) {
3706 // Replace many operand PackNodes with a binary tree for matching
3707 PackNode* p = (PackNode*) n;
3708 Node* btp = p->binary_tree_pack(1, n->req());
3709 n->subsume_by(btp, this);
3710 }
3711 break;
3712 case Op_Loop:
3713 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3714 case Op_CountedLoop:
3715 case Op_LongCountedLoop:
3716 case Op_OuterStripMinedLoop:
3717 if (n->as_Loop()->is_inner_loop()) {
3718 frc.inc_inner_loop_count();
3719 }
3720 n->as_Loop()->verify_strip_mined(0);
3721 break;
3722 case Op_LShiftI:
3723 case Op_RShiftI:
3724 case Op_URShiftI:
3725 case Op_LShiftL:
3726 case Op_RShiftL:
3727 case Op_URShiftL:
3728 if (Matcher::need_masked_shift_count) {
3729 // The cpu's shift instructions don't restrict the count to the
3730 // lower 5/6 bits. We need to do the masking ourselves.
3731 Node* in2 = n->in(2);
3732 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3733 const TypeInt* t = in2->find_int_type();
3734 if (t != nullptr && t->is_con()) {
3735 juint shift = t->get_con();
3736 if (shift > mask) { // Unsigned cmp
3737 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3738 }
3739 } else {
3740 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3741 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3742 n->set_req(2, shift);
3743 }
3744 }
3745 if (in2->outcnt() == 0) { // Remove dead node
3746 in2->disconnect_inputs(this);
3747 }
3748 }
3749 break;
3750 case Op_MemBarStoreStore:
3751 case Op_MemBarRelease:
3752 // Break the link with AllocateNode: it is no longer useful and
3753 // confuses register allocation.
3754 if (n->req() > MemBarNode::Precedent) {
3755 n->set_req(MemBarNode::Precedent, top());
3756 }
3757 break;
3758 case Op_MemBarAcquire: {
3759 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3760 // At parse time, the trailing MemBarAcquire for a volatile load
3761 // is created with an edge to the load. After optimizations,
3762 // that input may be a chain of Phis. If those phis have no
3763 // other use, then the MemBarAcquire keeps them alive and
3764 // register allocation can be confused.
3765 dead_nodes.push(n->in(MemBarNode::Precedent));
3766 n->set_req(MemBarNode::Precedent, top());
3767 }
3768 break;
3769 }
3770 case Op_Blackhole:
3771 break;
3772 case Op_RangeCheck: {
3773 RangeCheckNode* rc = n->as_RangeCheck();
3774 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3775 n->subsume_by(iff, this);
3776 frc._tests.push(iff);
3777 break;
3778 }
3779 case Op_ConvI2L: {
3780 if (!Matcher::convi2l_type_required) {
3781 // Code generation on some platforms doesn't need accurate
3782 // ConvI2L types. Widening the type can help remove redundant
3783 // address computations.
3784 n->as_Type()->set_type(TypeLong::INT);
3785 ResourceMark rm;
3786 Unique_Node_List wq;
3787 wq.push(n);
3788 for (uint next = 0; next < wq.size(); next++) {
3789 Node *m = wq.at(next);
3790
3791 for(;;) {
3792 // Loop over all nodes with identical inputs edges as m
3793 Node* k = m->find_similar(m->Opcode());
3794 if (k == nullptr) {
3795 break;
3796 }
3797 // Push their uses so we get a chance to remove node made
3798 // redundant
3799 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3800 Node* u = k->fast_out(i);
3801 if (u->Opcode() == Op_LShiftL ||
3802 u->Opcode() == Op_AddL ||
3803 u->Opcode() == Op_SubL ||
3804 u->Opcode() == Op_AddP) {
3805 wq.push(u);
3806 }
3807 }
3808 // Replace all nodes with identical edges as m with m
3809 k->subsume_by(m, this);
3810 }
3811 }
3812 }
3813 break;
3814 }
3815 case Op_CmpUL: {
3816 if (!Matcher::has_match_rule(Op_CmpUL)) {
3817 // No support for unsigned long comparisons
3818 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3819 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3820 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3821 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3822 Node* andl = new AndLNode(orl, remove_sign_mask);
3823 Node* cmp = new CmpLNode(andl, n->in(2));
3824 n->subsume_by(cmp, this);
3825 }
3826 break;
3827 }
3828 default:
3829 assert(!n->is_Call(), "");
3830 assert(!n->is_Mem(), "");
3831 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3832 break;
3833 }
3834 }
3835
3836 //------------------------------final_graph_reshaping_walk---------------------
3837 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3838 // requires that the walk visits a node's inputs before visiting the node.
3839 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3840 Unique_Node_List sfpt;
3841
3842 frc._visited.set(root->_idx); // first, mark node as visited
3843 uint cnt = root->req();
3844 Node *n = root;
3845 uint i = 0;
3846 while (true) {
3847 if (i < cnt) {
3848 // Place all non-visited non-null inputs onto stack
3849 Node* m = n->in(i);
3850 ++i;
3851 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3852 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3853 // compute worst case interpreter size in case of a deoptimization
3854 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3855
3856 sfpt.push(m);
3857 }
3858 cnt = m->req();
3859 nstack.push(n, i); // put on stack parent and next input's index
3860 n = m;
3861 i = 0;
3862 }
3863 } else {
3864 // Now do post-visit work
3865 final_graph_reshaping_impl(n, frc, dead_nodes);
3866 if (nstack.is_empty())
3867 break; // finished
3868 n = nstack.node(); // Get node from stack
3869 cnt = n->req();
3870 i = nstack.index();
3871 nstack.pop(); // Shift to the next node on stack
3872 }
3873 }
3874
3875 // Skip next transformation if compressed oops are not used.
3876 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3877 (!UseCompressedOops && !UseCompressedClassPointers))
3878 return;
3879
3880 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3881 // It could be done for an uncommon traps or any safepoints/calls
3882 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3883 while (sfpt.size() > 0) {
3884 n = sfpt.pop();
3885 JVMState *jvms = n->as_SafePoint()->jvms();
3886 assert(jvms != nullptr, "sanity");
3887 int start = jvms->debug_start();
3888 int end = n->req();
3889 bool is_uncommon = (n->is_CallStaticJava() &&
3890 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3891 for (int j = start; j < end; j++) {
3892 Node* in = n->in(j);
3893 if (in->is_DecodeNarrowPtr()) {
3894 bool safe_to_skip = true;
3895 if (!is_uncommon ) {
3896 // Is it safe to skip?
3897 for (uint i = 0; i < in->outcnt(); i++) {
3898 Node* u = in->raw_out(i);
3899 if (!u->is_SafePoint() ||
3900 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3901 safe_to_skip = false;
3902 }
3903 }
3904 }
3905 if (safe_to_skip) {
3906 n->set_req(j, in->in(1));
3907 }
3908 if (in->outcnt() == 0) {
3909 in->disconnect_inputs(this);
3910 }
3911 }
3912 }
3913 }
3914 }
3915
3916 //------------------------------final_graph_reshaping--------------------------
3917 // Final Graph Reshaping.
3918 //
3919 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3920 // and not commoned up and forced early. Must come after regular
3921 // optimizations to avoid GVN undoing the cloning. Clone constant
3922 // inputs to Loop Phis; these will be split by the allocator anyways.
3923 // Remove Opaque nodes.
3924 // (2) Move last-uses by commutative operations to the left input to encourage
3925 // Intel update-in-place two-address operations and better register usage
3926 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3927 // calls canonicalizing them back.
3928 // (3) Count the number of double-precision FP ops, single-precision FP ops
3929 // and call sites. On Intel, we can get correct rounding either by
3930 // forcing singles to memory (requires extra stores and loads after each
3931 // FP bytecode) or we can set a rounding mode bit (requires setting and
3932 // clearing the mode bit around call sites). The mode bit is only used
3933 // if the relative frequency of single FP ops to calls is low enough.
3934 // This is a key transform for SPEC mpeg_audio.
3935 // (4) Detect infinite loops; blobs of code reachable from above but not
3936 // below. Several of the Code_Gen algorithms fail on such code shapes,
3937 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3938 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3939 // Detection is by looking for IfNodes where only 1 projection is
3940 // reachable from below or CatchNodes missing some targets.
3941 // (5) Assert for insane oop offsets in debug mode.
3942
3943 bool Compile::final_graph_reshaping() {
3944 // an infinite loop may have been eliminated by the optimizer,
3945 // in which case the graph will be empty.
3946 if (root()->req() == 1) {
3947 // Do not compile method that is only a trivial infinite loop,
3948 // since the content of the loop may have been eliminated.
3949 record_method_not_compilable("trivial infinite loop");
3950 return true;
3951 }
3952
3953 // Expensive nodes have their control input set to prevent the GVN
3954 // from freely commoning them. There's no GVN beyond this point so
3955 // no need to keep the control input. We want the expensive nodes to
3956 // be freely moved to the least frequent code path by gcm.
3957 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3958 for (int i = 0; i < expensive_count(); i++) {
3959 _expensive_nodes.at(i)->set_req(0, nullptr);
3960 }
3961
3962 Final_Reshape_Counts frc;
3963
3964 // Visit everybody reachable!
3965 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3966 Node_Stack nstack(live_nodes() >> 1);
3967 Unique_Node_List dead_nodes;
3968 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
3969
3970 // Check for unreachable (from below) code (i.e., infinite loops).
3971 for( uint i = 0; i < frc._tests.size(); i++ ) {
3972 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3973 // Get number of CFG targets.
3974 // Note that PCTables include exception targets after calls.
3975 uint required_outcnt = n->required_outcnt();
3976 if (n->outcnt() != required_outcnt) {
3977 // Check for a few special cases. Rethrow Nodes never take the
3978 // 'fall-thru' path, so expected kids is 1 less.
3979 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3980 if (n->in(0)->in(0)->is_Call()) {
3981 CallNode* call = n->in(0)->in(0)->as_Call();
3982 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3983 required_outcnt--; // Rethrow always has 1 less kid
3984 } else if (call->req() > TypeFunc::Parms &&
3985 call->is_CallDynamicJava()) {
3986 // Check for null receiver. In such case, the optimizer has
3987 // detected that the virtual call will always result in a null
3988 // pointer exception. The fall-through projection of this CatchNode
3989 // will not be populated.
3990 Node* arg0 = call->in(TypeFunc::Parms);
3991 if (arg0->is_Type() &&
3992 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3993 required_outcnt--;
3994 }
3995 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
3996 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3997 // Check for illegal array length. In such case, the optimizer has
3998 // detected that the allocation attempt will always result in an
3999 // exception. There is no fall-through projection of this CatchNode .
4000 assert(call->is_CallStaticJava(), "static call expected");
4001 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4002 uint valid_length_test_input = call->req() - 1;
4003 Node* valid_length_test = call->in(valid_length_test_input);
4004 call->del_req(valid_length_test_input);
4005 if (valid_length_test->find_int_con(1) == 0) {
4006 required_outcnt--;
4007 }
4008 dead_nodes.push(valid_length_test);
4009 assert(n->outcnt() == required_outcnt, "malformed control flow");
4010 continue;
4011 }
4012 }
4013 }
4014
4015 // Recheck with a better notion of 'required_outcnt'
4016 if (n->outcnt() != required_outcnt) {
4017 record_method_not_compilable("malformed control flow");
4018 return true; // Not all targets reachable!
4019 }
4020 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4021 CallNode* call = n->in(0)->in(0)->as_Call();
4022 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4023 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4024 assert(call->is_CallStaticJava(), "static call expected");
4025 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4026 uint valid_length_test_input = call->req() - 1;
4027 dead_nodes.push(call->in(valid_length_test_input));
4028 call->del_req(valid_length_test_input); // valid length test useless now
4029 }
4030 }
4031 // Check that I actually visited all kids. Unreached kids
4032 // must be infinite loops.
4033 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4034 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4035 record_method_not_compilable("infinite loop");
4036 return true; // Found unvisited kid; must be unreach
4037 }
4038
4039 // Here so verification code in final_graph_reshaping_walk()
4040 // always see an OuterStripMinedLoopEnd
4041 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4042 IfNode* init_iff = n->as_If();
4043 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4044 n->subsume_by(iff, this);
4045 }
4046 }
4047
4048 while (dead_nodes.size() > 0) {
4049 Node* m = dead_nodes.pop();
4050 if (m->outcnt() == 0 && m != top()) {
4051 for (uint j = 0; j < m->req(); j++) {
4052 Node* in = m->in(j);
4053 if (in != nullptr) {
4054 dead_nodes.push(in);
4055 }
4056 }
4057 m->disconnect_inputs(this);
4058 }
4059 }
4060
4061 #ifdef IA32
4062 // If original bytecodes contained a mixture of floats and doubles
4063 // check if the optimizer has made it homogeneous, item (3).
4064 if (UseSSE == 0 &&
4065 frc.get_float_count() > 32 &&
4066 frc.get_double_count() == 0 &&
4067 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4068 set_24_bit_selection_and_mode(false, true);
4069 }
4070 #endif // IA32
4071
4072 set_java_calls(frc.get_java_call_count());
4073 set_inner_loops(frc.get_inner_loop_count());
4074
4075 // No infinite loops, no reason to bail out.
4076 return false;
4077 }
4078
4079 //-----------------------------too_many_traps----------------------------------
4080 // Report if there are too many traps at the current method and bci.
4081 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4082 bool Compile::too_many_traps(ciMethod* method,
4083 int bci,
4084 Deoptimization::DeoptReason reason) {
4085 ciMethodData* md = method->method_data();
4086 if (md->is_empty()) {
4087 // Assume the trap has not occurred, or that it occurred only
4088 // because of a transient condition during start-up in the interpreter.
4089 return false;
4090 }
4091 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4092 if (md->has_trap_at(bci, m, reason) != 0) {
4093 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4094 // Also, if there are multiple reasons, or if there is no per-BCI record,
4095 // assume the worst.
4096 if (log())
4097 log()->elem("observe trap='%s' count='%d'",
4098 Deoptimization::trap_reason_name(reason),
4099 md->trap_count(reason));
4100 return true;
4101 } else {
4102 // Ignore method/bci and see if there have been too many globally.
4103 return too_many_traps(reason, md);
4104 }
4105 }
4106
4107 // Less-accurate variant which does not require a method and bci.
4108 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4109 ciMethodData* logmd) {
4110 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4111 // Too many traps globally.
4112 // Note that we use cumulative trap_count, not just md->trap_count.
4113 if (log()) {
4114 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4115 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4116 Deoptimization::trap_reason_name(reason),
4117 mcount, trap_count(reason));
4118 }
4119 return true;
4120 } else {
4121 // The coast is clear.
4122 return false;
4123 }
4124 }
4125
4126 //--------------------------too_many_recompiles--------------------------------
4127 // Report if there are too many recompiles at the current method and bci.
4128 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4129 // Is not eager to return true, since this will cause the compiler to use
4130 // Action_none for a trap point, to avoid too many recompilations.
4131 bool Compile::too_many_recompiles(ciMethod* method,
4132 int bci,
4133 Deoptimization::DeoptReason reason) {
4134 ciMethodData* md = method->method_data();
4135 if (md->is_empty()) {
4136 // Assume the trap has not occurred, or that it occurred only
4137 // because of a transient condition during start-up in the interpreter.
4138 return false;
4139 }
4140 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4141 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4142 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4143 Deoptimization::DeoptReason per_bc_reason
4144 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4145 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4146 if ((per_bc_reason == Deoptimization::Reason_none
4147 || md->has_trap_at(bci, m, reason) != 0)
4148 // The trap frequency measure we care about is the recompile count:
4149 && md->trap_recompiled_at(bci, m)
4150 && md->overflow_recompile_count() >= bc_cutoff) {
4151 // Do not emit a trap here if it has already caused recompilations.
4152 // Also, if there are multiple reasons, or if there is no per-BCI record,
4153 // assume the worst.
4154 if (log())
4155 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4156 Deoptimization::trap_reason_name(reason),
4157 md->trap_count(reason),
4158 md->overflow_recompile_count());
4159 return true;
4160 } else if (trap_count(reason) != 0
4161 && decompile_count() >= m_cutoff) {
4162 // Too many recompiles globally, and we have seen this sort of trap.
4163 // Use cumulative decompile_count, not just md->decompile_count.
4164 if (log())
4165 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4166 Deoptimization::trap_reason_name(reason),
4167 md->trap_count(reason), trap_count(reason),
4168 md->decompile_count(), decompile_count());
4169 return true;
4170 } else {
4171 // The coast is clear.
4172 return false;
4173 }
4174 }
4175
4176 // Compute when not to trap. Used by matching trap based nodes and
4177 // NullCheck optimization.
4178 void Compile::set_allowed_deopt_reasons() {
4179 _allowed_reasons = 0;
4180 if (is_method_compilation()) {
4181 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4182 assert(rs < BitsPerInt, "recode bit map");
4183 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4184 _allowed_reasons |= nth_bit(rs);
4185 }
4186 }
4187 }
4188 }
4189
4190 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4191 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4192 }
4193
4194 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4195 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4196 }
4197
4198 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4199 if (holder->is_initialized()) {
4200 return false;
4201 }
4202 if (holder->is_being_initialized()) {
4203 if (accessing_method->holder() == holder) {
4204 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4205 // <init>, or a static method. In all those cases, there was an initialization
4206 // barrier on the holder klass passed.
4207 if (accessing_method->is_static_initializer() ||
4208 accessing_method->is_object_initializer() ||
4209 accessing_method->is_static()) {
4210 return false;
4211 }
4212 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4213 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4214 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4215 // child class can become fully initialized while its parent class is still being initialized.
4216 if (accessing_method->is_static_initializer()) {
4217 return false;
4218 }
4219 }
4220 ciMethod* root = method(); // the root method of compilation
4221 if (root != accessing_method) {
4222 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4223 }
4224 }
4225 return true;
4226 }
4227
4228 #ifndef PRODUCT
4229 //------------------------------verify_bidirectional_edges---------------------
4230 // For each input edge to a node (ie - for each Use-Def edge), verify that
4231 // there is a corresponding Def-Use edge.
4232 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4233 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4234 uint stack_size = live_nodes() >> 4;
4235 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4236 nstack.push(_root);
4237
4238 while (nstack.size() > 0) {
4239 Node* n = nstack.pop();
4240 if (visited.member(n)) {
4241 continue;
4242 }
4243 visited.push(n);
4244
4245 // Walk over all input edges, checking for correspondence
4246 uint length = n->len();
4247 for (uint i = 0; i < length; i++) {
4248 Node* in = n->in(i);
4249 if (in != nullptr && !visited.member(in)) {
4250 nstack.push(in); // Put it on stack
4251 }
4252 if (in != nullptr && !in->is_top()) {
4253 // Count instances of `next`
4254 int cnt = 0;
4255 for (uint idx = 0; idx < in->_outcnt; idx++) {
4256 if (in->_out[idx] == n) {
4257 cnt++;
4258 }
4259 }
4260 assert(cnt > 0, "Failed to find Def-Use edge.");
4261 // Check for duplicate edges
4262 // walk the input array downcounting the input edges to n
4263 for (uint j = 0; j < length; j++) {
4264 if (n->in(j) == in) {
4265 cnt--;
4266 }
4267 }
4268 assert(cnt == 0, "Mismatched edge count.");
4269 } else if (in == nullptr) {
4270 assert(i == 0 || i >= n->req() ||
4271 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4272 (n->is_Unlock() && i == (n->req() - 1)) ||
4273 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4274 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4275 } else {
4276 assert(in->is_top(), "sanity");
4277 // Nothing to check.
4278 }
4279 }
4280 }
4281 }
4282
4283 //------------------------------verify_graph_edges---------------------------
4284 // Walk the Graph and verify that there is a one-to-one correspondence
4285 // between Use-Def edges and Def-Use edges in the graph.
4286 void Compile::verify_graph_edges(bool no_dead_code) {
4287 if (VerifyGraphEdges) {
4288 Unique_Node_List visited;
4289
4290 // Call graph walk to check edges
4291 verify_bidirectional_edges(visited);
4292 if (no_dead_code) {
4293 // Now make sure that no visited node is used by an unvisited node.
4294 bool dead_nodes = false;
4295 Unique_Node_List checked;
4296 while (visited.size() > 0) {
4297 Node* n = visited.pop();
4298 checked.push(n);
4299 for (uint i = 0; i < n->outcnt(); i++) {
4300 Node* use = n->raw_out(i);
4301 if (checked.member(use)) continue; // already checked
4302 if (visited.member(use)) continue; // already in the graph
4303 if (use->is_Con()) continue; // a dead ConNode is OK
4304 // At this point, we have found a dead node which is DU-reachable.
4305 if (!dead_nodes) {
4306 tty->print_cr("*** Dead nodes reachable via DU edges:");
4307 dead_nodes = true;
4308 }
4309 use->dump(2);
4310 tty->print_cr("---");
4311 checked.push(use); // No repeats; pretend it is now checked.
4312 }
4313 }
4314 assert(!dead_nodes, "using nodes must be reachable from root");
4315 }
4316 }
4317 }
4318 #endif
4319
4320 // The Compile object keeps track of failure reasons separately from the ciEnv.
4321 // This is required because there is not quite a 1-1 relation between the
4322 // ciEnv and its compilation task and the Compile object. Note that one
4323 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4324 // to backtrack and retry without subsuming loads. Other than this backtracking
4325 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4326 // by the logic in C2Compiler.
4327 void Compile::record_failure(const char* reason) {
4328 if (log() != nullptr) {
4329 log()->elem("failure reason='%s' phase='compile'", reason);
4330 }
4331 if (_failure_reason == nullptr) {
4332 // Record the first failure reason.
4333 _failure_reason = reason;
4334 }
4335
4336 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4337 C->print_method(PHASE_FAILURE, 1);
4338 }
4339 _root = nullptr; // flush the graph, too
4340 }
4341
4342 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4343 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4344 _phase_name(name), _dolog(CITimeVerbose)
4345 {
4346 if (_dolog) {
4347 C = Compile::current();
4348 _log = C->log();
4349 } else {
4350 C = nullptr;
4351 _log = nullptr;
4352 }
4353 if (_log != nullptr) {
4354 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4355 _log->stamp();
4356 _log->end_head();
4357 }
4358 }
4359
4360 Compile::TracePhase::~TracePhase() {
4361
4362 C = Compile::current();
4363 if (_dolog) {
4364 _log = C->log();
4365 } else {
4366 _log = nullptr;
4367 }
4368
4369 #ifdef ASSERT
4370 if (PrintIdealNodeCount) {
4371 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4372 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4373 }
4374
4375 if (VerifyIdealNodeCount) {
4376 Compile::current()->print_missing_nodes();
4377 }
4378 #endif
4379
4380 if (_log != nullptr) {
4381 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4382 }
4383 }
4384
4385 //----------------------------static_subtype_check-----------------------------
4386 // Shortcut important common cases when superklass is exact:
4387 // (0) superklass is java.lang.Object (can occur in reflective code)
4388 // (1) subklass is already limited to a subtype of superklass => always ok
4389 // (2) subklass does not overlap with superklass => always fail
4390 // (3) superklass has NO subtypes and we can check with a simple compare.
4391 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4392 if (skip) {
4393 return SSC_full_test; // Let caller generate the general case.
4394 }
4395
4396 if (subk->is_java_subtype_of(superk)) {
4397 return SSC_always_true; // (0) and (1) this test cannot fail
4398 }
4399
4400 if (!subk->maybe_java_subtype_of(superk)) {
4401 return SSC_always_false; // (2) true path dead; no dynamic test needed
4402 }
4403
4404 const Type* superelem = superk;
4405 if (superk->isa_aryklassptr()) {
4406 int ignored;
4407 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4408 }
4409
4410 if (superelem->isa_instklassptr()) {
4411 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4412 if (!ik->has_subklass()) {
4413 if (!ik->is_final()) {
4414 // Add a dependency if there is a chance of a later subclass.
4415 dependencies()->assert_leaf_type(ik);
4416 }
4417 if (!superk->maybe_java_subtype_of(subk)) {
4418 return SSC_always_false;
4419 }
4420 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4421 }
4422 } else {
4423 // A primitive array type has no subtypes.
4424 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4425 }
4426
4427 return SSC_full_test;
4428 }
4429
4430 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4431 #ifdef _LP64
4432 // The scaled index operand to AddP must be a clean 64-bit value.
4433 // Java allows a 32-bit int to be incremented to a negative
4434 // value, which appears in a 64-bit register as a large
4435 // positive number. Using that large positive number as an
4436 // operand in pointer arithmetic has bad consequences.
4437 // On the other hand, 32-bit overflow is rare, and the possibility
4438 // can often be excluded, if we annotate the ConvI2L node with
4439 // a type assertion that its value is known to be a small positive
4440 // number. (The prior range check has ensured this.)
4441 // This assertion is used by ConvI2LNode::Ideal.
4442 int index_max = max_jint - 1; // array size is max_jint, index is one less
4443 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4444 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4445 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4446 #endif
4447 return idx;
4448 }
4449
4450 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4451 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4452 if (ctrl != nullptr) {
4453 // Express control dependency by a CastII node with a narrow type.
4454 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4455 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4456 // node from floating above the range check during loop optimizations. Otherwise, the
4457 // ConvI2L node may be eliminated independently of the range check, causing the data path
4458 // to become TOP while the control path is still there (although it's unreachable).
4459 value->set_req(0, ctrl);
4460 value = phase->transform(value);
4461 }
4462 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4463 return phase->transform(new ConvI2LNode(value, ltype));
4464 }
4465
4466 // The message about the current inlining is accumulated in
4467 // _print_inlining_stream and transferred into the _print_inlining_list
4468 // once we know whether inlining succeeds or not. For regular
4469 // inlining, messages are appended to the buffer pointed by
4470 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4471 // a new buffer is added after _print_inlining_idx in the list. This
4472 // way we can update the inlining message for late inlining call site
4473 // when the inlining is attempted again.
4474 void Compile::print_inlining_init() {
4475 if (print_inlining() || print_intrinsics()) {
4476 // print_inlining_init is actually called several times.
4477 print_inlining_reset();
4478 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4479 }
4480 }
4481
4482 void Compile::print_inlining_reinit() {
4483 if (print_inlining() || print_intrinsics()) {
4484 print_inlining_reset();
4485 }
4486 }
4487
4488 void Compile::print_inlining_reset() {
4489 _print_inlining_stream->reset();
4490 }
4491
4492 void Compile::print_inlining_commit() {
4493 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4494 // Transfer the message from _print_inlining_stream to the current
4495 // _print_inlining_list buffer and clear _print_inlining_stream.
4496 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4497 print_inlining_reset();
4498 }
4499
4500 void Compile::print_inlining_push() {
4501 // Add new buffer to the _print_inlining_list at current position
4502 _print_inlining_idx++;
4503 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4504 }
4505
4506 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4507 return _print_inlining_list->at(_print_inlining_idx);
4508 }
4509
4510 void Compile::print_inlining_update(CallGenerator* cg) {
4511 if (print_inlining() || print_intrinsics()) {
4512 if (cg->is_late_inline()) {
4513 if (print_inlining_current()->cg() != cg &&
4514 (print_inlining_current()->cg() != nullptr ||
4515 print_inlining_current()->ss()->size() != 0)) {
4516 print_inlining_push();
4517 }
4518 print_inlining_commit();
4519 print_inlining_current()->set_cg(cg);
4520 } else {
4521 if (print_inlining_current()->cg() != nullptr) {
4522 print_inlining_push();
4523 }
4524 print_inlining_commit();
4525 }
4526 }
4527 }
4528
4529 void Compile::print_inlining_move_to(CallGenerator* cg) {
4530 // We resume inlining at a late inlining call site. Locate the
4531 // corresponding inlining buffer so that we can update it.
4532 if (print_inlining() || print_intrinsics()) {
4533 for (int i = 0; i < _print_inlining_list->length(); i++) {
4534 if (_print_inlining_list->at(i)->cg() == cg) {
4535 _print_inlining_idx = i;
4536 return;
4537 }
4538 }
4539 ShouldNotReachHere();
4540 }
4541 }
4542
4543 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4544 if (print_inlining() || print_intrinsics()) {
4545 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4546 assert(print_inlining_current()->cg() == cg, "wrong entry");
4547 // replace message with new message
4548 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4549 print_inlining_commit();
4550 print_inlining_current()->set_cg(cg);
4551 }
4552 }
4553
4554 void Compile::print_inlining_assert_ready() {
4555 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4556 }
4557
4558 void Compile::process_print_inlining() {
4559 assert(_late_inlines.length() == 0, "not drained yet");
4560 if (print_inlining() || print_intrinsics()) {
4561 ResourceMark rm;
4562 stringStream ss;
4563 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4564 for (int i = 0; i < _print_inlining_list->length(); i++) {
4565 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4566 ss.print("%s", pib->ss()->freeze());
4567 delete pib;
4568 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4569 }
4570 // Reset _print_inlining_list, it only contains destructed objects.
4571 // It is on the arena, so it will be freed when the arena is reset.
4572 _print_inlining_list = nullptr;
4573 // _print_inlining_stream won't be used anymore, either.
4574 print_inlining_reset();
4575 size_t end = ss.size();
4576 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4577 strncpy(_print_inlining_output, ss.freeze(), end+1);
4578 _print_inlining_output[end] = 0;
4579 }
4580 }
4581
4582 void Compile::dump_print_inlining() {
4583 if (_print_inlining_output != nullptr) {
4584 tty->print_raw(_print_inlining_output);
4585 }
4586 }
4587
4588 void Compile::log_late_inline(CallGenerator* cg) {
4589 if (log() != nullptr) {
4590 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4591 cg->unique_id());
4592 JVMState* p = cg->call_node()->jvms();
4593 while (p != nullptr) {
4594 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4595 p = p->caller();
4596 }
4597 log()->tail("late_inline");
4598 }
4599 }
4600
4601 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4602 log_late_inline(cg);
4603 if (log() != nullptr) {
4604 log()->inline_fail(msg);
4605 }
4606 }
4607
4608 void Compile::log_inline_id(CallGenerator* cg) {
4609 if (log() != nullptr) {
4610 // The LogCompilation tool needs a unique way to identify late
4611 // inline call sites. This id must be unique for this call site in
4612 // this compilation. Try to have it unique across compilations as
4613 // well because it can be convenient when grepping through the log
4614 // file.
4615 // Distinguish OSR compilations from others in case CICountOSR is
4616 // on.
4617 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4618 cg->set_unique_id(id);
4619 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4620 }
4621 }
4622
4623 void Compile::log_inline_failure(const char* msg) {
4624 if (C->log() != nullptr) {
4625 C->log()->inline_fail(msg);
4626 }
4627 }
4628
4629
4630 // Dump inlining replay data to the stream.
4631 // Don't change thread state and acquire any locks.
4632 void Compile::dump_inline_data(outputStream* out) {
4633 InlineTree* inl_tree = ilt();
4634 if (inl_tree != nullptr) {
4635 out->print(" inline %d", inl_tree->count());
4636 inl_tree->dump_replay_data(out);
4637 }
4638 }
4639
4640 void Compile::dump_inline_data_reduced(outputStream* out) {
4641 assert(ReplayReduce, "");
4642
4643 InlineTree* inl_tree = ilt();
4644 if (inl_tree == nullptr) {
4645 return;
4646 }
4647 // Enable iterative replay file reduction
4648 // Output "compile" lines for depth 1 subtrees,
4649 // simulating that those trees were compiled
4650 // instead of inlined.
4651 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4652 InlineTree* sub = inl_tree->subtrees().at(i);
4653 if (sub->inline_level() != 1) {
4654 continue;
4655 }
4656
4657 ciMethod* method = sub->method();
4658 int entry_bci = -1;
4659 int comp_level = env()->task()->comp_level();
4660 out->print("compile ");
4661 method->dump_name_as_ascii(out);
4662 out->print(" %d %d", entry_bci, comp_level);
4663 out->print(" inline %d", sub->count());
4664 sub->dump_replay_data(out, -1);
4665 out->cr();
4666 }
4667 }
4668
4669 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4670 if (n1->Opcode() < n2->Opcode()) return -1;
4671 else if (n1->Opcode() > n2->Opcode()) return 1;
4672
4673 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4674 for (uint i = 1; i < n1->req(); i++) {
4675 if (n1->in(i) < n2->in(i)) return -1;
4676 else if (n1->in(i) > n2->in(i)) return 1;
4677 }
4678
4679 return 0;
4680 }
4681
4682 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4683 Node* n1 = *n1p;
4684 Node* n2 = *n2p;
4685
4686 return cmp_expensive_nodes(n1, n2);
4687 }
4688
4689 void Compile::sort_expensive_nodes() {
4690 if (!expensive_nodes_sorted()) {
4691 _expensive_nodes.sort(cmp_expensive_nodes);
4692 }
4693 }
4694
4695 bool Compile::expensive_nodes_sorted() const {
4696 for (int i = 1; i < _expensive_nodes.length(); i++) {
4697 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4698 return false;
4699 }
4700 }
4701 return true;
4702 }
4703
4704 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4705 if (_expensive_nodes.length() == 0) {
4706 return false;
4707 }
4708
4709 assert(OptimizeExpensiveOps, "optimization off?");
4710
4711 // Take this opportunity to remove dead nodes from the list
4712 int j = 0;
4713 for (int i = 0; i < _expensive_nodes.length(); i++) {
4714 Node* n = _expensive_nodes.at(i);
4715 if (!n->is_unreachable(igvn)) {
4716 assert(n->is_expensive(), "should be expensive");
4717 _expensive_nodes.at_put(j, n);
4718 j++;
4719 }
4720 }
4721 _expensive_nodes.trunc_to(j);
4722
4723 // Then sort the list so that similar nodes are next to each other
4724 // and check for at least two nodes of identical kind with same data
4725 // inputs.
4726 sort_expensive_nodes();
4727
4728 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4729 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4730 return true;
4731 }
4732 }
4733
4734 return false;
4735 }
4736
4737 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4738 if (_expensive_nodes.length() == 0) {
4739 return;
4740 }
4741
4742 assert(OptimizeExpensiveOps, "optimization off?");
4743
4744 // Sort to bring similar nodes next to each other and clear the
4745 // control input of nodes for which there's only a single copy.
4746 sort_expensive_nodes();
4747
4748 int j = 0;
4749 int identical = 0;
4750 int i = 0;
4751 bool modified = false;
4752 for (; i < _expensive_nodes.length()-1; i++) {
4753 assert(j <= i, "can't write beyond current index");
4754 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4755 identical++;
4756 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4757 continue;
4758 }
4759 if (identical > 0) {
4760 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4761 identical = 0;
4762 } else {
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 }
4769 if (identical > 0) {
4770 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4771 } else if (_expensive_nodes.length() >= 1) {
4772 Node* n = _expensive_nodes.at(i);
4773 igvn.replace_input_of(n, 0, nullptr);
4774 igvn.hash_insert(n);
4775 modified = true;
4776 }
4777 _expensive_nodes.trunc_to(j);
4778 if (modified) {
4779 igvn.optimize();
4780 }
4781 }
4782
4783 void Compile::add_expensive_node(Node * n) {
4784 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4785 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4786 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4787 if (OptimizeExpensiveOps) {
4788 _expensive_nodes.append(n);
4789 } else {
4790 // Clear control input and let IGVN optimize expensive nodes if
4791 // OptimizeExpensiveOps is off.
4792 n->set_req(0, nullptr);
4793 }
4794 }
4795
4796 /**
4797 * Track coarsened Lock and Unlock nodes.
4798 */
4799
4800 class Lock_List : public Node_List {
4801 uint _origin_cnt;
4802 public:
4803 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4804 uint origin_cnt() const { return _origin_cnt; }
4805 };
4806
4807 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4808 int length = locks.length();
4809 if (length > 0) {
4810 // Have to keep this list until locks elimination during Macro nodes elimination.
4811 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4812 for (int i = 0; i < length; i++) {
4813 AbstractLockNode* lock = locks.at(i);
4814 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4815 locks_list->push(lock);
4816 }
4817 _coarsened_locks.append(locks_list);
4818 }
4819 }
4820
4821 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4822 int count = coarsened_count();
4823 for (int i = 0; i < count; i++) {
4824 Node_List* locks_list = _coarsened_locks.at(i);
4825 for (uint j = 0; j < locks_list->size(); j++) {
4826 Node* lock = locks_list->at(j);
4827 assert(lock->is_AbstractLock(), "sanity");
4828 if (!useful.member(lock)) {
4829 locks_list->yank(lock);
4830 }
4831 }
4832 }
4833 }
4834
4835 void Compile::remove_coarsened_lock(Node* n) {
4836 if (n->is_AbstractLock()) {
4837 int count = coarsened_count();
4838 for (int i = 0; i < count; i++) {
4839 Node_List* locks_list = _coarsened_locks.at(i);
4840 locks_list->yank(n);
4841 }
4842 }
4843 }
4844
4845 bool Compile::coarsened_locks_consistent() {
4846 int count = coarsened_count();
4847 for (int i = 0; i < count; i++) {
4848 bool unbalanced = false;
4849 bool modified = false; // track locks kind modifications
4850 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4851 uint size = locks_list->size();
4852 if (size == 0) {
4853 unbalanced = false; // All locks were eliminated - good
4854 } else if (size != locks_list->origin_cnt()) {
4855 unbalanced = true; // Some locks were removed from list
4856 } else {
4857 for (uint j = 0; j < size; j++) {
4858 Node* lock = locks_list->at(j);
4859 // All nodes in group should have the same state (modified or not)
4860 if (!lock->as_AbstractLock()->is_coarsened()) {
4861 if (j == 0) {
4862 // first on list was modified, the rest should be too for consistency
4863 modified = true;
4864 } else if (!modified) {
4865 // this lock was modified but previous locks on the list were not
4866 unbalanced = true;
4867 break;
4868 }
4869 } else if (modified) {
4870 // previous locks on list were modified but not this lock
4871 unbalanced = true;
4872 break;
4873 }
4874 }
4875 }
4876 if (unbalanced) {
4877 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4878 #ifdef ASSERT
4879 if (PrintEliminateLocks) {
4880 tty->print_cr("=== unbalanced coarsened locks ===");
4881 for (uint l = 0; l < size; l++) {
4882 locks_list->at(l)->dump();
4883 }
4884 }
4885 #endif
4886 record_failure(C2Compiler::retry_no_locks_coarsening());
4887 return false;
4888 }
4889 }
4890 return true;
4891 }
4892
4893 /**
4894 * Remove the speculative part of types and clean up the graph
4895 */
4896 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4897 if (UseTypeSpeculation) {
4898 Unique_Node_List worklist;
4899 worklist.push(root());
4900 int modified = 0;
4901 // Go over all type nodes that carry a speculative type, drop the
4902 // speculative part of the type and enqueue the node for an igvn
4903 // which may optimize it out.
4904 for (uint next = 0; next < worklist.size(); ++next) {
4905 Node *n = worklist.at(next);
4906 if (n->is_Type()) {
4907 TypeNode* tn = n->as_Type();
4908 const Type* t = tn->type();
4909 const Type* t_no_spec = t->remove_speculative();
4910 if (t_no_spec != t) {
4911 bool in_hash = igvn.hash_delete(n);
4912 assert(in_hash, "node should be in igvn hash table");
4913 tn->set_type(t_no_spec);
4914 igvn.hash_insert(n);
4915 igvn._worklist.push(n); // give it a chance to go away
4916 modified++;
4917 }
4918 }
4919 // Iterate over outs - endless loops is unreachable from below
4920 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4921 Node *m = n->fast_out(i);
4922 if (not_a_node(m)) {
4923 continue;
4924 }
4925 worklist.push(m);
4926 }
4927 }
4928 // Drop the speculative part of all types in the igvn's type table
4929 igvn.remove_speculative_types();
4930 if (modified > 0) {
4931 igvn.optimize();
4932 if (failing()) return;
4933 }
4934 #ifdef ASSERT
4935 // Verify that after the IGVN is over no speculative type has resurfaced
4936 worklist.clear();
4937 worklist.push(root());
4938 for (uint next = 0; next < worklist.size(); ++next) {
4939 Node *n = worklist.at(next);
4940 const Type* t = igvn.type_or_null(n);
4941 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4942 if (n->is_Type()) {
4943 t = n->as_Type()->type();
4944 assert(t == t->remove_speculative(), "no more speculative types");
4945 }
4946 // Iterate over outs - endless loops is unreachable from below
4947 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4948 Node *m = n->fast_out(i);
4949 if (not_a_node(m)) {
4950 continue;
4951 }
4952 worklist.push(m);
4953 }
4954 }
4955 igvn.check_no_speculative_types();
4956 #endif
4957 }
4958 }
4959
4960 // Auxiliary methods to support randomized stressing/fuzzing.
4961
4962 int Compile::random() {
4963 _stress_seed = os::next_random(_stress_seed);
4964 return static_cast<int>(_stress_seed);
4965 }
4966
4967 // This method can be called the arbitrary number of times, with current count
4968 // as the argument. The logic allows selecting a single candidate from the
4969 // running list of candidates as follows:
4970 // int count = 0;
4971 // Cand* selected = null;
4972 // while(cand = cand->next()) {
4973 // if (randomized_select(++count)) {
4974 // selected = cand;
4975 // }
4976 // }
4977 //
4978 // Including count equalizes the chances any candidate is "selected".
4979 // This is useful when we don't have the complete list of candidates to choose
4980 // from uniformly. In this case, we need to adjust the randomicity of the
4981 // selection, or else we will end up biasing the selection towards the latter
4982 // candidates.
4983 //
4984 // Quick back-envelope calculation shows that for the list of n candidates
4985 // the equal probability for the candidate to persist as "best" can be
4986 // achieved by replacing it with "next" k-th candidate with the probability
4987 // of 1/k. It can be easily shown that by the end of the run, the
4988 // probability for any candidate is converged to 1/n, thus giving the
4989 // uniform distribution among all the candidates.
4990 //
4991 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4992 #define RANDOMIZED_DOMAIN_POW 29
4993 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4994 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4995 bool Compile::randomized_select(int count) {
4996 assert(count > 0, "only positive");
4997 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4998 }
4999
5000 CloneMap& Compile::clone_map() { return _clone_map; }
5001 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5002
5003 void NodeCloneInfo::dump_on(outputStream* st) const {
5004 st->print(" {%d:%d} ", idx(), gen());
5005 }
5006
5007 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5008 uint64_t val = value(old->_idx);
5009 NodeCloneInfo cio(val);
5010 assert(val != 0, "old node should be in the map");
5011 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5012 insert(nnn->_idx, cin.get());
5013 #ifndef PRODUCT
5014 if (is_debug()) {
5015 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5016 }
5017 #endif
5018 }
5019
5020 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5021 NodeCloneInfo cio(value(old->_idx));
5022 if (cio.get() == 0) {
5023 cio.set(old->_idx, 0);
5024 insert(old->_idx, cio.get());
5025 #ifndef PRODUCT
5026 if (is_debug()) {
5027 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5028 }
5029 #endif
5030 }
5031 clone(old, nnn, gen);
5032 }
5033
5034 int CloneMap::max_gen() const {
5035 int g = 0;
5036 DictI di(_dict);
5037 for(; di.test(); ++di) {
5038 int t = gen(di._key);
5039 if (g < t) {
5040 g = t;
5041 #ifndef PRODUCT
5042 if (is_debug()) {
5043 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5044 }
5045 #endif
5046 }
5047 }
5048 return g;
5049 }
5050
5051 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5052 uint64_t val = value(key);
5053 if (val != 0) {
5054 NodeCloneInfo ni(val);
5055 ni.dump_on(st);
5056 }
5057 }
5058
5059 // Move Allocate nodes to the start of the list
5060 void Compile::sort_macro_nodes() {
5061 int count = macro_count();
5062 int allocates = 0;
5063 for (int i = 0; i < count; i++) {
5064 Node* n = macro_node(i);
5065 if (n->is_Allocate()) {
5066 if (i != allocates) {
5067 Node* tmp = macro_node(allocates);
5068 _macro_nodes.at_put(allocates, n);
5069 _macro_nodes.at_put(i, tmp);
5070 }
5071 allocates++;
5072 }
5073 }
5074 }
5075
5076 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5077 EventCompilerPhase event;
5078 if (event.should_commit()) {
5079 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5080 }
5081 #ifndef PRODUCT
5082 ResourceMark rm;
5083 stringStream ss;
5084 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5085 if (n != nullptr) {
5086 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5087 }
5088
5089 const char* name = ss.as_string();
5090 if (should_print_igv(level)) {
5091 _igv_printer->print_method(name, level);
5092 }
5093 if (should_print_phase(cpt)) {
5094 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5095 }
5096 #endif
5097 C->_latest_stage_start_counter.stamp();
5098 }
5099
5100 // Only used from CompileWrapper
5101 void Compile::begin_method() {
5102 #ifndef PRODUCT
5103 if (_method != nullptr && should_print_igv(1)) {
5104 _igv_printer->begin_method();
5105 }
5106 #endif
5107 C->_latest_stage_start_counter.stamp();
5108 }
5109
5110 // Only used from CompileWrapper
5111 void Compile::end_method() {
5112 EventCompilerPhase event;
5113 if (event.should_commit()) {
5114 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5115 }
5116
5117 #ifndef PRODUCT
5118 if (_method != nullptr && should_print_igv(1)) {
5119 _igv_printer->end_method();
5120 }
5121 #endif
5122 }
5123
5124 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5125 #ifndef PRODUCT
5126 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) {
5127 return true;
5128 }
5129 #endif
5130 return false;
5131 }
5132
5133 bool Compile::should_print_igv(int level) {
5134 #ifndef PRODUCT
5135 if (PrintIdealGraphLevel < 0) { // disabled by the user
5136 return false;
5137 }
5138
5139 bool need = directive()->IGVPrintLevelOption >= level;
5140 if (need && !_igv_printer) {
5141 _igv_printer = IdealGraphPrinter::printer();
5142 _igv_printer->set_compile(this);
5143 }
5144 return need;
5145 #else
5146 return false;
5147 #endif
5148 }
5149
5150 #ifndef PRODUCT
5151 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5152 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5153
5154 // Called from debugger. Prints method to the default file with the default phase name.
5155 // This works regardless of any Ideal Graph Visualizer flags set or not.
5156 void igv_print() {
5157 Compile::current()->igv_print_method_to_file();
5158 }
5159
5160 // Same as igv_print() above but with a specified phase name.
5161 void igv_print(const char* phase_name) {
5162 Compile::current()->igv_print_method_to_file(phase_name);
5163 }
5164
5165 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5166 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5167 // This works regardless of any Ideal Graph Visualizer flags set or not.
5168 void igv_print(bool network) {
5169 if (network) {
5170 Compile::current()->igv_print_method_to_network();
5171 } else {
5172 Compile::current()->igv_print_method_to_file();
5173 }
5174 }
5175
5176 // Same as igv_print(bool network) above but with a specified phase name.
5177 void igv_print(bool network, const char* phase_name) {
5178 if (network) {
5179 Compile::current()->igv_print_method_to_network(phase_name);
5180 } else {
5181 Compile::current()->igv_print_method_to_file(phase_name);
5182 }
5183 }
5184
5185 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5186 void igv_print_default() {
5187 Compile::current()->print_method(PHASE_DEBUG, 0);
5188 }
5189
5190 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5191 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5192 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5193 void igv_append() {
5194 Compile::current()->igv_print_method_to_file("Debug", true);
5195 }
5196
5197 // Same as igv_append() above but with a specified phase name.
5198 void igv_append(const char* phase_name) {
5199 Compile::current()->igv_print_method_to_file(phase_name, true);
5200 }
5201
5202 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5203 const char* file_name = "custom_debug.xml";
5204 if (_debug_file_printer == nullptr) {
5205 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5206 } else {
5207 _debug_file_printer->update_compiled_method(C->method());
5208 }
5209 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5210 _debug_file_printer->print(phase_name, (Node*)C->root());
5211 }
5212
5213 void Compile::igv_print_method_to_network(const char* phase_name) {
5214 if (_debug_network_printer == nullptr) {
5215 _debug_network_printer = new IdealGraphPrinter(C);
5216 } else {
5217 _debug_network_printer->update_compiled_method(C->method());
5218 }
5219 tty->print_cr("Method printed over network stream to IGV");
5220 _debug_network_printer->print(phase_name, (Node*)C->root());
5221 }
5222 #endif
5223
5224 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5225 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5226 return value;
5227 }
5228 Node* result = nullptr;
5229 if (bt == T_BYTE) {
5230 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5231 result = new RShiftINode(result, phase->intcon(24));
5232 } else if (bt == T_BOOLEAN) {
5233 result = new AndINode(value, phase->intcon(0xFF));
5234 } else if (bt == T_CHAR) {
5235 result = new AndINode(value,phase->intcon(0xFFFF));
5236 } else {
5237 assert(bt == T_SHORT, "unexpected narrow type");
5238 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5239 result = new RShiftINode(result, phase->intcon(16));
5240 }
5241 if (transform_res) {
5242 result = phase->transform(result);
5243 }
5244 return result;
5245 }
5246