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 (UseCompactObjectHeaders) {
1695 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1696 alias_type(idx)->set_rewritable(false);
1697 }
1698 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1699 alias_type(idx)->set_rewritable(false);
1700 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1701 alias_type(idx)->set_rewritable(false);
1702 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1703 alias_type(idx)->set_rewritable(false);
1704 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1705 alias_type(idx)->set_rewritable(false);
1706 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1707 alias_type(idx)->set_rewritable(false);
1708 }
1709 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1710 // but the base pointer type is not distinctive enough to identify
1711 // references into JavaThread.)
1712
1713 // Check for final fields.
1714 const TypeInstPtr* tinst = flat->isa_instptr();
1715 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1716 ciField* field;
1717 if (tinst->const_oop() != nullptr &&
1718 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1719 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1720 // static field
1721 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1722 field = k->get_field_by_offset(tinst->offset(), true);
1723 } else {
1724 ciInstanceKlass *k = tinst->instance_klass();
1725 field = k->get_field_by_offset(tinst->offset(), false);
1726 }
1727 assert(field == nullptr ||
1728 original_field == nullptr ||
1729 (field->holder() == original_field->holder() &&
1730 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1731 field->is_static() == original_field->is_static()), "wrong field?");
1732 // Set field() and is_rewritable() attributes.
1733 if (field != nullptr) alias_type(idx)->set_field(field);
1734 }
1735 }
1736
1737 // Fill the cache for next time.
1738 ace->_adr_type = adr_type;
1739 ace->_index = idx;
1740 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1741
1742 // Might as well try to fill the cache for the flattened version, too.
1743 AliasCacheEntry* face = probe_alias_cache(flat);
1744 if (face->_adr_type == nullptr) {
1745 face->_adr_type = flat;
1746 face->_index = idx;
1747 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1748 }
1749
1750 return alias_type(idx);
1751 }
1752
1753
1754 Compile::AliasType* Compile::alias_type(ciField* field) {
1755 const TypeOopPtr* t;
1756 if (field->is_static())
1757 t = TypeInstPtr::make(field->holder()->java_mirror());
1758 else
1759 t = TypeOopPtr::make_from_klass_raw(field->holder());
1760 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1761 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1762 return atp;
1763 }
1764
1765
1766 //------------------------------have_alias_type--------------------------------
1767 bool Compile::have_alias_type(const TypePtr* adr_type) {
1768 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1769 if (ace->_adr_type == adr_type) {
1770 return true;
1771 }
1772
1773 // Handle special cases.
1774 if (adr_type == nullptr) return true;
1775 if (adr_type == TypePtr::BOTTOM) return true;
1776
1777 return find_alias_type(adr_type, true, nullptr) != nullptr;
1778 }
1779
1780 //-----------------------------must_alias--------------------------------------
1781 // True if all values of the given address type are in the given alias category.
1782 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1783 if (alias_idx == AliasIdxBot) return true; // the universal category
1784 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1785 if (alias_idx == AliasIdxTop) return false; // the empty category
1786 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1787
1788 // the only remaining possible overlap is identity
1789 int adr_idx = get_alias_index(adr_type);
1790 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1791 assert(adr_idx == alias_idx ||
1792 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1793 && adr_type != TypeOopPtr::BOTTOM),
1794 "should not be testing for overlap with an unsafe pointer");
1795 return adr_idx == alias_idx;
1796 }
1797
1798 //------------------------------can_alias--------------------------------------
1799 // True if any values of the given address type are in the given alias category.
1800 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1801 if (alias_idx == AliasIdxTop) return false; // the empty category
1802 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1803 // Known instance doesn't alias with bottom memory
1804 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1805 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1806
1807 // the only remaining possible overlap is identity
1808 int adr_idx = get_alias_index(adr_type);
1809 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1810 return adr_idx == alias_idx;
1811 }
1812
1813 // Remove the opaque nodes that protect the Parse Predicates so that all unused
1814 // checks and uncommon_traps will be eliminated from the ideal graph.
1815 void Compile::cleanup_parse_predicates(PhaseIterGVN& igvn) const {
1816 if (parse_predicate_count() == 0) {
1817 return;
1818 }
1819 for (int i = parse_predicate_count(); i > 0; i--) {
1820 Node* n = parse_predicate_opaque1_node(i - 1);
1821 assert(n->Opcode() == Op_Opaque1, "must be");
1822 igvn.replace_node(n, n->in(1));
1823 }
1824 assert(parse_predicate_count() == 0, "should be clean!");
1825 }
1826
1827 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1828 if (!n->for_post_loop_opts_igvn()) {
1829 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1830 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1831 _for_post_loop_igvn.append(n);
1832 }
1833 }
1834
1835 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1836 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1837 _for_post_loop_igvn.remove(n);
1838 }
1839
1840 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1841 // Verify that all previous optimizations produced a valid graph
1842 // at least to this point, even if no loop optimizations were done.
1843 PhaseIdealLoop::verify(igvn);
1844
1845 C->set_post_loop_opts_phase(); // no more loop opts allowed
1846
1847 assert(!C->major_progress(), "not cleared");
1848
1849 if (_for_post_loop_igvn.length() > 0) {
1850 while (_for_post_loop_igvn.length() > 0) {
1851 Node* n = _for_post_loop_igvn.pop();
1852 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1853 igvn._worklist.push(n);
1854 }
1855 igvn.optimize();
1856 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1857
1858 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1859 if (C->major_progress()) {
1860 C->clear_major_progress(); // ensure that major progress is now clear
1861 }
1862 }
1863 }
1864
1865 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1866 if (OptimizeUnstableIf) {
1867 _unstable_if_traps.append(trap);
1868 }
1869 }
1870
1871 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1872 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1873 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1874 Node* n = trap->uncommon_trap();
1875 if (!useful.member(n)) {
1876 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1877 }
1878 }
1879 }
1880
1881 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1882 // or fold-compares case. Return true if succeed or not found.
1883 //
1884 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1885 // false and ask fold-compares to yield.
1886 //
1887 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1888 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1889 // when deoptimization does happen.
1890 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1891 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1892 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1893 if (trap->uncommon_trap() == unc) {
1894 if (yield && trap->modified()) {
1895 return false;
1896 }
1897 _unstable_if_traps.delete_at(i);
1898 break;
1899 }
1900 }
1901 return true;
1902 }
1903
1904 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1905 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1906 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1907 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1908 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1909 CallStaticJavaNode* unc = trap->uncommon_trap();
1910 int next_bci = trap->next_bci();
1911 bool modified = trap->modified();
1912
1913 if (next_bci != -1 && !modified) {
1914 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1915 JVMState* jvms = unc->jvms();
1916 ciMethod* method = jvms->method();
1917 ciBytecodeStream iter(method);
1918
1919 iter.force_bci(jvms->bci());
1920 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1921 Bytecodes::Code c = iter.cur_bc();
1922 Node* lhs = nullptr;
1923 Node* rhs = nullptr;
1924 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1925 lhs = unc->peek_operand(0);
1926 rhs = unc->peek_operand(1);
1927 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1928 lhs = unc->peek_operand(0);
1929 }
1930
1931 ResourceMark rm;
1932 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1933 assert(live_locals.is_valid(), "broken liveness info");
1934 int len = (int)live_locals.size();
1935
1936 for (int i = 0; i < len; i++) {
1937 Node* local = unc->local(jvms, i);
1938 // kill local using the liveness of next_bci.
1939 // give up when the local looks like an operand to secure reexecution.
1940 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1941 uint idx = jvms->locoff() + i;
1942 #ifdef ASSERT
1943 if (Verbose) {
1944 tty->print("[unstable_if] kill local#%d: ", idx);
1945 local->dump();
1946 tty->cr();
1947 }
1948 #endif
1949 igvn.replace_input_of(unc, idx, top());
1950 modified = true;
1951 }
1952 }
1953 }
1954
1955 // keep the mondified trap for late query
1956 if (modified) {
1957 trap->set_modified();
1958 } else {
1959 _unstable_if_traps.delete_at(i);
1960 }
1961 }
1962 igvn.optimize();
1963 }
1964
1965 // StringOpts and late inlining of string methods
1966 void Compile::inline_string_calls(bool parse_time) {
1967 {
1968 // remove useless nodes to make the usage analysis simpler
1969 ResourceMark rm;
1970 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
1971 }
1972
1973 {
1974 ResourceMark rm;
1975 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1976 PhaseStringOpts pso(initial_gvn());
1977 print_method(PHASE_AFTER_STRINGOPTS, 3);
1978 }
1979
1980 // now inline anything that we skipped the first time around
1981 if (!parse_time) {
1982 _late_inlines_pos = _late_inlines.length();
1983 }
1984
1985 while (_string_late_inlines.length() > 0) {
1986 CallGenerator* cg = _string_late_inlines.pop();
1987 cg->do_late_inline();
1988 if (failing()) return;
1989 }
1990 _string_late_inlines.trunc_to(0);
1991 }
1992
1993 // Late inlining of boxing methods
1994 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1995 if (_boxing_late_inlines.length() > 0) {
1996 assert(has_boxed_value(), "inconsistent");
1997
1998 PhaseGVN* gvn = initial_gvn();
1999 set_inlining_incrementally(true);
2000
2001 igvn_worklist()->ensure_empty(); // should be done with igvn
2002
2003 _late_inlines_pos = _late_inlines.length();
2004
2005 while (_boxing_late_inlines.length() > 0) {
2006 CallGenerator* cg = _boxing_late_inlines.pop();
2007 cg->do_late_inline();
2008 if (failing()) return;
2009 }
2010 _boxing_late_inlines.trunc_to(0);
2011
2012 inline_incrementally_cleanup(igvn);
2013
2014 set_inlining_incrementally(false);
2015 }
2016 }
2017
2018 bool Compile::inline_incrementally_one() {
2019 assert(IncrementalInline, "incremental inlining should be on");
2020
2021 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2022
2023 set_inlining_progress(false);
2024 set_do_cleanup(false);
2025
2026 for (int i = 0; i < _late_inlines.length(); i++) {
2027 _late_inlines_pos = i+1;
2028 CallGenerator* cg = _late_inlines.at(i);
2029 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2030 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2031 cg->do_late_inline();
2032 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2033 if (failing()) {
2034 return false;
2035 } else if (inlining_progress()) {
2036 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2037 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2038 break; // process one call site at a time
2039 }
2040 } else {
2041 // Ignore late inline direct calls when inlining is not allowed.
2042 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2043 }
2044 }
2045 // Remove processed elements.
2046 _late_inlines.remove_till(_late_inlines_pos);
2047 _late_inlines_pos = 0;
2048
2049 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2050
2051 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2052
2053 set_inlining_progress(false);
2054 set_do_cleanup(false);
2055
2056 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2057 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2058 }
2059
2060 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2061 {
2062 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2063 ResourceMark rm;
2064 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2065 }
2066 {
2067 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2068 igvn.reset_from_gvn(initial_gvn());
2069 igvn.optimize();
2070 }
2071 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2072 }
2073
2074 // Perform incremental inlining until bound on number of live nodes is reached
2075 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2076 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2077
2078 set_inlining_incrementally(true);
2079 uint low_live_nodes = 0;
2080
2081 while (_late_inlines.length() > 0) {
2082 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2083 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2084 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2085 // PhaseIdealLoop is expensive so we only try it once we are
2086 // out of live nodes and we only try it again if the previous
2087 // helped got the number of nodes down significantly
2088 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2089 if (failing()) return;
2090 low_live_nodes = live_nodes();
2091 _major_progress = true;
2092 }
2093
2094 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2095 bool do_print_inlining = print_inlining() || print_intrinsics();
2096 if (do_print_inlining || log() != nullptr) {
2097 // Print inlining message for candidates that we couldn't inline for lack of space.
2098 for (int i = 0; i < _late_inlines.length(); i++) {
2099 CallGenerator* cg = _late_inlines.at(i);
2100 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2101 if (do_print_inlining) {
2102 cg->print_inlining_late(msg);
2103 }
2104 log_late_inline_failure(cg, msg);
2105 }
2106 }
2107 break; // finish
2108 }
2109 }
2110
2111 igvn_worklist()->ensure_empty(); // should be done with igvn
2112
2113 while (inline_incrementally_one()) {
2114 assert(!failing(), "inconsistent");
2115 }
2116 if (failing()) return;
2117
2118 inline_incrementally_cleanup(igvn);
2119
2120 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2121
2122 if (failing()) return;
2123
2124 if (_late_inlines.length() == 0) {
2125 break; // no more progress
2126 }
2127 }
2128
2129 igvn_worklist()->ensure_empty(); // should be done with igvn
2130
2131 if (_string_late_inlines.length() > 0) {
2132 assert(has_stringbuilder(), "inconsistent");
2133
2134 inline_string_calls(false);
2135
2136 if (failing()) return;
2137
2138 inline_incrementally_cleanup(igvn);
2139 }
2140
2141 set_inlining_incrementally(false);
2142 }
2143
2144 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2145 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2146 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2147 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2148 // as if "inlining_incrementally() == true" were set.
2149 assert(inlining_incrementally() == false, "not allowed");
2150 assert(_modified_nodes == nullptr, "not allowed");
2151 assert(_late_inlines.length() > 0, "sanity");
2152
2153 while (_late_inlines.length() > 0) {
2154 igvn_worklist()->ensure_empty(); // should be done with igvn
2155
2156 while (inline_incrementally_one()) {
2157 assert(!failing(), "inconsistent");
2158 }
2159 if (failing()) return;
2160
2161 inline_incrementally_cleanup(igvn);
2162 }
2163 }
2164
2165 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2166 if (_loop_opts_cnt > 0) {
2167 while (major_progress() && (_loop_opts_cnt > 0)) {
2168 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2169 PhaseIdealLoop::optimize(igvn, mode);
2170 _loop_opts_cnt--;
2171 if (failing()) return false;
2172 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2173 }
2174 }
2175 return true;
2176 }
2177
2178 // Remove edges from "root" to each SafePoint at a backward branch.
2179 // They were inserted during parsing (see add_safepoint()) to make
2180 // infinite loops without calls or exceptions visible to root, i.e.,
2181 // useful.
2182 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2183 Node *r = root();
2184 if (r != nullptr) {
2185 for (uint i = r->req(); i < r->len(); ++i) {
2186 Node *n = r->in(i);
2187 if (n != nullptr && n->is_SafePoint()) {
2188 r->rm_prec(i);
2189 if (n->outcnt() == 0) {
2190 igvn.remove_dead_node(n);
2191 }
2192 --i;
2193 }
2194 }
2195 // Parsing may have added top inputs to the root node (Path
2196 // leading to the Halt node proven dead). Make sure we get a
2197 // chance to clean them up.
2198 igvn._worklist.push(r);
2199 igvn.optimize();
2200 }
2201 }
2202
2203 //------------------------------Optimize---------------------------------------
2204 // Given a graph, optimize it.
2205 void Compile::Optimize() {
2206 TracePhase tp("optimizer", &timers[_t_optimizer]);
2207
2208 #ifndef PRODUCT
2209 if (env()->break_at_compile()) {
2210 BREAKPOINT;
2211 }
2212
2213 #endif
2214
2215 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2216 #ifdef ASSERT
2217 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2218 #endif
2219
2220 ResourceMark rm;
2221
2222 print_inlining_reinit();
2223
2224 NOT_PRODUCT( verify_graph_edges(); )
2225
2226 print_method(PHASE_AFTER_PARSING, 1);
2227
2228 {
2229 // Iterative Global Value Numbering, including ideal transforms
2230 // Initialize IterGVN with types and values from parse-time GVN
2231 PhaseIterGVN igvn(initial_gvn());
2232 #ifdef ASSERT
2233 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2234 #endif
2235 {
2236 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2237 igvn.optimize();
2238 }
2239
2240 if (failing()) return;
2241
2242 print_method(PHASE_ITER_GVN1, 2);
2243
2244 process_for_unstable_if_traps(igvn);
2245
2246 if (failing()) return;
2247
2248 inline_incrementally(igvn);
2249
2250 print_method(PHASE_INCREMENTAL_INLINE, 2);
2251
2252 if (failing()) return;
2253
2254 if (eliminate_boxing()) {
2255 // Inline valueOf() methods now.
2256 inline_boxing_calls(igvn);
2257
2258 if (failing()) return;
2259
2260 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2261 inline_incrementally(igvn);
2262 }
2263
2264 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2265
2266 if (failing()) return;
2267 }
2268
2269 // Remove the speculative part of types and clean up the graph from
2270 // the extra CastPP nodes whose only purpose is to carry them. Do
2271 // that early so that optimizations are not disrupted by the extra
2272 // CastPP nodes.
2273 remove_speculative_types(igvn);
2274
2275 if (failing()) return;
2276
2277 // No more new expensive nodes will be added to the list from here
2278 // so keep only the actual candidates for optimizations.
2279 cleanup_expensive_nodes(igvn);
2280
2281 if (failing()) return;
2282
2283 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2284 if (EnableVectorSupport && has_vbox_nodes()) {
2285 TracePhase tp("", &timers[_t_vector]);
2286 PhaseVector pv(igvn);
2287 pv.optimize_vector_boxes();
2288 if (failing()) return;
2289 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2290 }
2291 assert(!has_vbox_nodes(), "sanity");
2292
2293 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2294 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2295 igvn_worklist()->ensure_empty(); // should be done with igvn
2296 {
2297 ResourceMark rm;
2298 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2299 }
2300 igvn.reset_from_gvn(initial_gvn());
2301 igvn.optimize();
2302 }
2303
2304 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2305 // safepoints
2306 remove_root_to_sfpts_edges(igvn);
2307
2308 if (failing()) return;
2309
2310 // Perform escape analysis
2311 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2312 if (has_loops()) {
2313 // Cleanup graph (remove dead nodes).
2314 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2315 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2316 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2317 if (failing()) return;
2318 }
2319 bool progress;
2320 do {
2321 ConnectionGraph::do_analysis(this, &igvn);
2322
2323 if (failing()) return;
2324
2325 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2326
2327 // Optimize out fields loads from scalar replaceable allocations.
2328 igvn.optimize();
2329 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2330
2331 if (failing()) return;
2332
2333 if (congraph() != nullptr && macro_count() > 0) {
2334 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2335 PhaseMacroExpand mexp(igvn);
2336 mexp.eliminate_macro_nodes();
2337 igvn.set_delay_transform(false);
2338
2339 igvn.optimize();
2340 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2341
2342 if (failing()) return;
2343 }
2344 progress = do_iterative_escape_analysis() &&
2345 (macro_count() < mcount) &&
2346 ConnectionGraph::has_candidates(this);
2347 // Try again if candidates exist and made progress
2348 // by removing some allocations and/or locks.
2349 } while (progress);
2350 }
2351
2352 // Loop transforms on the ideal graph. Range Check Elimination,
2353 // peeling, unrolling, etc.
2354
2355 // Set loop opts counter
2356 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2357 {
2358 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2359 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2360 _loop_opts_cnt--;
2361 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2362 if (failing()) return;
2363 }
2364 // Loop opts pass if partial peeling occurred in previous pass
2365 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2366 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2367 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2368 _loop_opts_cnt--;
2369 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2370 if (failing()) return;
2371 }
2372 // Loop opts pass for loop-unrolling before CCP
2373 if(major_progress() && (_loop_opts_cnt > 0)) {
2374 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2375 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2376 _loop_opts_cnt--;
2377 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2378 }
2379 if (!failing()) {
2380 // Verify that last round of loop opts produced a valid graph
2381 PhaseIdealLoop::verify(igvn);
2382 }
2383 }
2384 if (failing()) return;
2385
2386 // Conditional Constant Propagation;
2387 PhaseCCP ccp( &igvn );
2388 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2389 {
2390 TracePhase tp("ccp", &timers[_t_ccp]);
2391 ccp.do_transform();
2392 }
2393 print_method(PHASE_CCP1, 2);
2394
2395 assert( true, "Break here to ccp.dump_old2new_map()");
2396
2397 // Iterative Global Value Numbering, including ideal transforms
2398 {
2399 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2400 igvn.reset_from_igvn(&ccp);
2401 igvn.optimize();
2402 }
2403 print_method(PHASE_ITER_GVN2, 2);
2404
2405 if (failing()) return;
2406
2407 // Loop transforms on the ideal graph. Range Check Elimination,
2408 // peeling, unrolling, etc.
2409 if (!optimize_loops(igvn, LoopOptsDefault)) {
2410 return;
2411 }
2412
2413 if (failing()) return;
2414
2415 C->clear_major_progress(); // ensure that major progress is now clear
2416
2417 process_for_post_loop_opts_igvn(igvn);
2418
2419 if (failing()) return;
2420
2421 #ifdef ASSERT
2422 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2423 #endif
2424
2425 {
2426 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2427 PhaseMacroExpand mex(igvn);
2428 if (mex.expand_macro_nodes()) {
2429 assert(failing(), "must bail out w/ explicit message");
2430 return;
2431 }
2432 print_method(PHASE_MACRO_EXPANSION, 2);
2433 }
2434
2435 {
2436 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2437 if (bs->expand_barriers(this, igvn)) {
2438 assert(failing(), "must bail out w/ explicit message");
2439 return;
2440 }
2441 print_method(PHASE_BARRIER_EXPANSION, 2);
2442 }
2443
2444 if (C->max_vector_size() > 0) {
2445 C->optimize_logic_cones(igvn);
2446 igvn.optimize();
2447 }
2448
2449 DEBUG_ONLY( _modified_nodes = nullptr; )
2450
2451 assert(igvn._worklist.size() == 0, "not empty");
2452
2453 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2454
2455 if (_late_inlines.length() > 0) {
2456 // More opportunities to optimize virtual and MH calls.
2457 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2458 process_late_inline_calls_no_inline(igvn);
2459 if (failing()) return;
2460 }
2461 } // (End scope of igvn; run destructor if necessary for asserts.)
2462
2463 check_no_dead_use();
2464
2465 process_print_inlining();
2466
2467 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2468 // to remove hashes to unlock nodes for modifications.
2469 C->node_hash()->clear();
2470
2471 // A method with only infinite loops has no edges entering loops from root
2472 {
2473 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2474 if (final_graph_reshaping()) {
2475 assert(failing(), "must bail out w/ explicit message");
2476 return;
2477 }
2478 }
2479
2480 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2481 DEBUG_ONLY(set_phase_optimize_finished();)
2482 }
2483
2484 #ifdef ASSERT
2485 void Compile::check_no_dead_use() const {
2486 ResourceMark rm;
2487 Unique_Node_List wq;
2488 wq.push(root());
2489 for (uint i = 0; i < wq.size(); ++i) {
2490 Node* n = wq.at(i);
2491 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2492 Node* u = n->fast_out(j);
2493 if (u->outcnt() == 0 && !u->is_Con()) {
2494 u->dump();
2495 fatal("no reachable node should have no use");
2496 }
2497 wq.push(u);
2498 }
2499 }
2500 }
2501 #endif
2502
2503 void Compile::inline_vector_reboxing_calls() {
2504 if (C->_vector_reboxing_late_inlines.length() > 0) {
2505 _late_inlines_pos = C->_late_inlines.length();
2506 while (_vector_reboxing_late_inlines.length() > 0) {
2507 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2508 cg->do_late_inline();
2509 if (failing()) return;
2510 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2511 }
2512 _vector_reboxing_late_inlines.trunc_to(0);
2513 }
2514 }
2515
2516 bool Compile::has_vbox_nodes() {
2517 if (C->_vector_reboxing_late_inlines.length() > 0) {
2518 return true;
2519 }
2520 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2521 Node * n = C->macro_node(macro_idx);
2522 assert(n->is_macro(), "only macro nodes expected here");
2523 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2524 return true;
2525 }
2526 }
2527 return false;
2528 }
2529
2530 //---------------------------- Bitwise operation packing optimization ---------------------------
2531
2532 static bool is_vector_unary_bitwise_op(Node* n) {
2533 return n->Opcode() == Op_XorV &&
2534 VectorNode::is_vector_bitwise_not_pattern(n);
2535 }
2536
2537 static bool is_vector_binary_bitwise_op(Node* n) {
2538 switch (n->Opcode()) {
2539 case Op_AndV:
2540 case Op_OrV:
2541 return true;
2542
2543 case Op_XorV:
2544 return !is_vector_unary_bitwise_op(n);
2545
2546 default:
2547 return false;
2548 }
2549 }
2550
2551 static bool is_vector_ternary_bitwise_op(Node* n) {
2552 return n->Opcode() == Op_MacroLogicV;
2553 }
2554
2555 static bool is_vector_bitwise_op(Node* n) {
2556 return is_vector_unary_bitwise_op(n) ||
2557 is_vector_binary_bitwise_op(n) ||
2558 is_vector_ternary_bitwise_op(n);
2559 }
2560
2561 static bool is_vector_bitwise_cone_root(Node* n) {
2562 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2563 return false;
2564 }
2565 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2566 if (is_vector_bitwise_op(n->fast_out(i))) {
2567 return false;
2568 }
2569 }
2570 return true;
2571 }
2572
2573 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2574 uint cnt = 0;
2575 if (is_vector_bitwise_op(n)) {
2576 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2577 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2578 for (uint i = 1; i < inp_cnt; i++) {
2579 Node* in = n->in(i);
2580 bool skip = VectorNode::is_all_ones_vector(in);
2581 if (!skip && !inputs.member(in)) {
2582 inputs.push(in);
2583 cnt++;
2584 }
2585 }
2586 assert(cnt <= 1, "not unary");
2587 } else {
2588 uint last_req = inp_cnt;
2589 if (is_vector_ternary_bitwise_op(n)) {
2590 last_req = inp_cnt - 1; // skip last input
2591 }
2592 for (uint i = 1; i < last_req; i++) {
2593 Node* def = n->in(i);
2594 if (!inputs.member(def)) {
2595 inputs.push(def);
2596 cnt++;
2597 }
2598 }
2599 }
2600 } else { // not a bitwise operations
2601 if (!inputs.member(n)) {
2602 inputs.push(n);
2603 cnt++;
2604 }
2605 }
2606 return cnt;
2607 }
2608
2609 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2610 Unique_Node_List useful_nodes;
2611 C->identify_useful_nodes(useful_nodes);
2612
2613 for (uint i = 0; i < useful_nodes.size(); i++) {
2614 Node* n = useful_nodes.at(i);
2615 if (is_vector_bitwise_cone_root(n)) {
2616 list.push(n);
2617 }
2618 }
2619 }
2620
2621 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2622 const TypeVect* vt,
2623 Unique_Node_List& partition,
2624 Unique_Node_List& inputs) {
2625 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2626 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2627 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2628
2629 Node* in1 = inputs.at(0);
2630 Node* in2 = inputs.at(1);
2631 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2632
2633 uint func = compute_truth_table(partition, inputs);
2634
2635 Node* pn = partition.at(partition.size() - 1);
2636 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2637 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2638 }
2639
2640 static uint extract_bit(uint func, uint pos) {
2641 return (func & (1 << pos)) >> pos;
2642 }
2643
2644 //
2645 // A macro logic node represents a truth table. It has 4 inputs,
2646 // First three inputs corresponds to 3 columns of a truth table
2647 // and fourth input captures the logic function.
2648 //
2649 // eg. fn = (in1 AND in2) OR in3;
2650 //
2651 // MacroNode(in1,in2,in3,fn)
2652 //
2653 // -----------------
2654 // in1 in2 in3 fn
2655 // -----------------
2656 // 0 0 0 0
2657 // 0 0 1 1
2658 // 0 1 0 0
2659 // 0 1 1 1
2660 // 1 0 0 0
2661 // 1 0 1 1
2662 // 1 1 0 1
2663 // 1 1 1 1
2664 //
2665
2666 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2667 int res = 0;
2668 for (int i = 0; i < 8; i++) {
2669 int bit1 = extract_bit(in1, i);
2670 int bit2 = extract_bit(in2, i);
2671 int bit3 = extract_bit(in3, i);
2672
2673 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2674 int func_bit = extract_bit(func, func_bit_pos);
2675
2676 res |= func_bit << i;
2677 }
2678 return res;
2679 }
2680
2681 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2682 assert(n != nullptr, "");
2683 assert(eval_map.contains(n), "absent");
2684 return *(eval_map.get(n));
2685 }
2686
2687 static void eval_operands(Node* n,
2688 uint& func1, uint& func2, uint& func3,
2689 ResourceHashtable<Node*,uint>& eval_map) {
2690 assert(is_vector_bitwise_op(n), "");
2691
2692 if (is_vector_unary_bitwise_op(n)) {
2693 Node* opnd = n->in(1);
2694 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2695 opnd = n->in(2);
2696 }
2697 func1 = eval_operand(opnd, eval_map);
2698 } else if (is_vector_binary_bitwise_op(n)) {
2699 func1 = eval_operand(n->in(1), eval_map);
2700 func2 = eval_operand(n->in(2), eval_map);
2701 } else {
2702 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2703 func1 = eval_operand(n->in(1), eval_map);
2704 func2 = eval_operand(n->in(2), eval_map);
2705 func3 = eval_operand(n->in(3), eval_map);
2706 }
2707 }
2708
2709 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2710 assert(inputs.size() <= 3, "sanity");
2711 ResourceMark rm;
2712 uint res = 0;
2713 ResourceHashtable<Node*,uint> eval_map;
2714
2715 // Populate precomputed functions for inputs.
2716 // Each input corresponds to one column of 3 input truth-table.
2717 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2718 0xCC, // (_, b, _) -> b
2719 0xF0 }; // (a, _, _) -> a
2720 for (uint i = 0; i < inputs.size(); i++) {
2721 eval_map.put(inputs.at(i), input_funcs[2-i]);
2722 }
2723
2724 for (uint i = 0; i < partition.size(); i++) {
2725 Node* n = partition.at(i);
2726
2727 uint func1 = 0, func2 = 0, func3 = 0;
2728 eval_operands(n, func1, func2, func3, eval_map);
2729
2730 switch (n->Opcode()) {
2731 case Op_OrV:
2732 assert(func3 == 0, "not binary");
2733 res = func1 | func2;
2734 break;
2735 case Op_AndV:
2736 assert(func3 == 0, "not binary");
2737 res = func1 & func2;
2738 break;
2739 case Op_XorV:
2740 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2741 assert(func2 == 0 && func3 == 0, "not unary");
2742 res = (~func1) & 0xFF;
2743 } else {
2744 assert(func3 == 0, "not binary");
2745 res = func1 ^ func2;
2746 }
2747 break;
2748 case Op_MacroLogicV:
2749 // Ordering of inputs may change during evaluation of sub-tree
2750 // containing MacroLogic node as a child node, thus a re-evaluation
2751 // makes sure that function is evaluated in context of current
2752 // inputs.
2753 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2754 break;
2755
2756 default: assert(false, "not supported: %s", n->Name());
2757 }
2758 assert(res <= 0xFF, "invalid");
2759 eval_map.put(n, res);
2760 }
2761 return res;
2762 }
2763
2764 // Criteria under which nodes gets packed into a macro logic node:-
2765 // 1) Parent and both child nodes are all unmasked or masked with
2766 // same predicates.
2767 // 2) Masked parent can be packed with left child if it is predicated
2768 // and both have same predicates.
2769 // 3) Masked parent can be packed with right child if its un-predicated
2770 // or has matching predication condition.
2771 // 4) An unmasked parent can be packed with an unmasked child.
2772 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2773 assert(partition.size() == 0, "not empty");
2774 assert(inputs.size() == 0, "not empty");
2775 if (is_vector_ternary_bitwise_op(n)) {
2776 return false;
2777 }
2778
2779 bool is_unary_op = is_vector_unary_bitwise_op(n);
2780 if (is_unary_op) {
2781 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2782 return false; // too few inputs
2783 }
2784
2785 bool pack_left_child = true;
2786 bool pack_right_child = true;
2787
2788 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2789 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2790
2791 int left_child_input_cnt = 0;
2792 int right_child_input_cnt = 0;
2793
2794 bool parent_is_predicated = n->is_predicated_vector();
2795 bool left_child_predicated = n->in(1)->is_predicated_vector();
2796 bool right_child_predicated = n->in(2)->is_predicated_vector();
2797
2798 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2799 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2800 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2801
2802 do {
2803 if (pack_left_child && left_child_LOP &&
2804 ((!parent_is_predicated && !left_child_predicated) ||
2805 ((parent_is_predicated && left_child_predicated &&
2806 parent_pred == left_child_pred)))) {
2807 partition.push(n->in(1));
2808 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2809 } else {
2810 inputs.push(n->in(1));
2811 left_child_input_cnt = 1;
2812 }
2813
2814 if (pack_right_child && right_child_LOP &&
2815 (!right_child_predicated ||
2816 (right_child_predicated && parent_is_predicated &&
2817 parent_pred == right_child_pred))) {
2818 partition.push(n->in(2));
2819 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2820 } else {
2821 inputs.push(n->in(2));
2822 right_child_input_cnt = 1;
2823 }
2824
2825 if (inputs.size() > 3) {
2826 assert(partition.size() > 0, "");
2827 inputs.clear();
2828 partition.clear();
2829 if (left_child_input_cnt > right_child_input_cnt) {
2830 pack_left_child = false;
2831 } else {
2832 pack_right_child = false;
2833 }
2834 } else {
2835 break;
2836 }
2837 } while(true);
2838
2839 if(partition.size()) {
2840 partition.push(n);
2841 }
2842
2843 return (partition.size() == 2 || partition.size() == 3) &&
2844 (inputs.size() == 2 || inputs.size() == 3);
2845 }
2846
2847 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2848 assert(is_vector_bitwise_op(n), "not a root");
2849
2850 visited.set(n->_idx);
2851
2852 // 1) Do a DFS walk over the logic cone.
2853 for (uint i = 1; i < n->req(); i++) {
2854 Node* in = n->in(i);
2855 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2856 process_logic_cone_root(igvn, in, visited);
2857 }
2858 }
2859
2860 // 2) Bottom up traversal: Merge node[s] with
2861 // the parent to form macro logic node.
2862 Unique_Node_List partition;
2863 Unique_Node_List inputs;
2864 if (compute_logic_cone(n, partition, inputs)) {
2865 const TypeVect* vt = n->bottom_type()->is_vect();
2866 Node* pn = partition.at(partition.size() - 1);
2867 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2868 if (mask == nullptr ||
2869 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2870 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2871 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2872 igvn.replace_node(n, macro_logic);
2873 }
2874 }
2875 }
2876
2877 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2878 ResourceMark rm;
2879 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2880 Unique_Node_List list;
2881 collect_logic_cone_roots(list);
2882
2883 while (list.size() > 0) {
2884 Node* n = list.pop();
2885 const TypeVect* vt = n->bottom_type()->is_vect();
2886 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2887 if (supported) {
2888 VectorSet visited(comp_arena());
2889 process_logic_cone_root(igvn, n, visited);
2890 }
2891 }
2892 }
2893 }
2894
2895 //------------------------------Code_Gen---------------------------------------
2896 // Given a graph, generate code for it
2897 void Compile::Code_Gen() {
2898 if (failing()) {
2899 return;
2900 }
2901
2902 // Perform instruction selection. You might think we could reclaim Matcher
2903 // memory PDQ, but actually the Matcher is used in generating spill code.
2904 // Internals of the Matcher (including some VectorSets) must remain live
2905 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2906 // set a bit in reclaimed memory.
2907
2908 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2909 // nodes. Mapping is only valid at the root of each matched subtree.
2910 NOT_PRODUCT( verify_graph_edges(); )
2911
2912 Matcher matcher;
2913 _matcher = &matcher;
2914 {
2915 TracePhase tp("matcher", &timers[_t_matcher]);
2916 matcher.match();
2917 if (failing()) {
2918 return;
2919 }
2920 }
2921 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2922 // nodes. Mapping is only valid at the root of each matched subtree.
2923 NOT_PRODUCT( verify_graph_edges(); )
2924
2925 // If you have too many nodes, or if matching has failed, bail out
2926 check_node_count(0, "out of nodes matching instructions");
2927 if (failing()) {
2928 return;
2929 }
2930
2931 print_method(PHASE_MATCHING, 2);
2932
2933 // Build a proper-looking CFG
2934 PhaseCFG cfg(node_arena(), root(), matcher);
2935 _cfg = &cfg;
2936 {
2937 TracePhase tp("scheduler", &timers[_t_scheduler]);
2938 bool success = cfg.do_global_code_motion();
2939 if (!success) {
2940 return;
2941 }
2942
2943 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2944 NOT_PRODUCT( verify_graph_edges(); )
2945 cfg.verify();
2946 }
2947
2948 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2949 _regalloc = ®alloc;
2950 {
2951 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2952 // Perform register allocation. After Chaitin, use-def chains are
2953 // no longer accurate (at spill code) and so must be ignored.
2954 // Node->LRG->reg mappings are still accurate.
2955 _regalloc->Register_Allocate();
2956
2957 // Bail out if the allocator builds too many nodes
2958 if (failing()) {
2959 return;
2960 }
2961 }
2962
2963 // Prior to register allocation we kept empty basic blocks in case the
2964 // the allocator needed a place to spill. After register allocation we
2965 // are not adding any new instructions. If any basic block is empty, we
2966 // can now safely remove it.
2967 {
2968 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2969 cfg.remove_empty_blocks();
2970 if (do_freq_based_layout()) {
2971 PhaseBlockLayout layout(cfg);
2972 } else {
2973 cfg.set_loop_alignment();
2974 }
2975 cfg.fixup_flow();
2976 cfg.remove_unreachable_blocks();
2977 cfg.verify_dominator_tree();
2978 }
2979
2980 // Apply peephole optimizations
2981 if( OptoPeephole ) {
2982 TracePhase tp("peephole", &timers[_t_peephole]);
2983 PhasePeephole peep( _regalloc, cfg);
2984 peep.do_transform();
2985 }
2986
2987 // Do late expand if CPU requires this.
2988 if (Matcher::require_postalloc_expand) {
2989 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2990 cfg.postalloc_expand(_regalloc);
2991 }
2992
2993 // Convert Nodes to instruction bits in a buffer
2994 {
2995 TracePhase tp("output", &timers[_t_output]);
2996 PhaseOutput output;
2997 output.Output();
2998 if (failing()) return;
2999 output.install();
3000 }
3001
3002 print_method(PHASE_FINAL_CODE, 1);
3003
3004 // He's dead, Jim.
3005 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3006 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3007 }
3008
3009 //------------------------------Final_Reshape_Counts---------------------------
3010 // This class defines counters to help identify when a method
3011 // may/must be executed using hardware with only 24-bit precision.
3012 struct Final_Reshape_Counts : public StackObj {
3013 int _call_count; // count non-inlined 'common' calls
3014 int _float_count; // count float ops requiring 24-bit precision
3015 int _double_count; // count double ops requiring more precision
3016 int _java_call_count; // count non-inlined 'java' calls
3017 int _inner_loop_count; // count loops which need alignment
3018 VectorSet _visited; // Visitation flags
3019 Node_List _tests; // Set of IfNodes & PCTableNodes
3020
3021 Final_Reshape_Counts() :
3022 _call_count(0), _float_count(0), _double_count(0),
3023 _java_call_count(0), _inner_loop_count(0) { }
3024
3025 void inc_call_count () { _call_count ++; }
3026 void inc_float_count () { _float_count ++; }
3027 void inc_double_count() { _double_count++; }
3028 void inc_java_call_count() { _java_call_count++; }
3029 void inc_inner_loop_count() { _inner_loop_count++; }
3030
3031 int get_call_count () const { return _call_count ; }
3032 int get_float_count () const { return _float_count ; }
3033 int get_double_count() const { return _double_count; }
3034 int get_java_call_count() const { return _java_call_count; }
3035 int get_inner_loop_count() const { return _inner_loop_count; }
3036 };
3037
3038 // Eliminate trivially redundant StoreCMs and accumulate their
3039 // precedence edges.
3040 void Compile::eliminate_redundant_card_marks(Node* n) {
3041 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3042 if (n->in(MemNode::Address)->outcnt() > 1) {
3043 // There are multiple users of the same address so it might be
3044 // possible to eliminate some of the StoreCMs
3045 Node* mem = n->in(MemNode::Memory);
3046 Node* adr = n->in(MemNode::Address);
3047 Node* val = n->in(MemNode::ValueIn);
3048 Node* prev = n;
3049 bool done = false;
3050 // Walk the chain of StoreCMs eliminating ones that match. As
3051 // long as it's a chain of single users then the optimization is
3052 // safe. Eliminating partially redundant StoreCMs would require
3053 // cloning copies down the other paths.
3054 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3055 if (adr == mem->in(MemNode::Address) &&
3056 val == mem->in(MemNode::ValueIn)) {
3057 // redundant StoreCM
3058 if (mem->req() > MemNode::OopStore) {
3059 // Hasn't been processed by this code yet.
3060 n->add_prec(mem->in(MemNode::OopStore));
3061 } else {
3062 // Already converted to precedence edge
3063 for (uint i = mem->req(); i < mem->len(); i++) {
3064 // Accumulate any precedence edges
3065 if (mem->in(i) != nullptr) {
3066 n->add_prec(mem->in(i));
3067 }
3068 }
3069 // Everything above this point has been processed.
3070 done = true;
3071 }
3072 // Eliminate the previous StoreCM
3073 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3074 assert(mem->outcnt() == 0, "should be dead");
3075 mem->disconnect_inputs(this);
3076 } else {
3077 prev = mem;
3078 }
3079 mem = prev->in(MemNode::Memory);
3080 }
3081 }
3082 }
3083
3084 //------------------------------final_graph_reshaping_impl----------------------
3085 // Implement items 1-5 from final_graph_reshaping below.
3086 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3087
3088 if ( n->outcnt() == 0 ) return; // dead node
3089 uint nop = n->Opcode();
3090
3091 // Check for 2-input instruction with "last use" on right input.
3092 // Swap to left input. Implements item (2).
3093 if( n->req() == 3 && // two-input instruction
3094 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3095 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3096 n->in(2)->outcnt() == 1 &&// right use IS a last use
3097 !n->in(2)->is_Con() ) { // right use is not a constant
3098 // Check for commutative opcode
3099 switch( nop ) {
3100 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3101 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3102 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3103 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3104 case Op_AndL: case Op_XorL: case Op_OrL:
3105 case Op_AndI: case Op_XorI: case Op_OrI: {
3106 // Move "last use" input to left by swapping inputs
3107 n->swap_edges(1, 2);
3108 break;
3109 }
3110 default:
3111 break;
3112 }
3113 }
3114
3115 #ifdef ASSERT
3116 if( n->is_Mem() ) {
3117 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3118 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3119 // oop will be recorded in oop map if load crosses safepoint
3120 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3121 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3122 "raw memory operations should have control edge");
3123 }
3124 if (n->is_MemBar()) {
3125 MemBarNode* mb = n->as_MemBar();
3126 if (mb->trailing_store() || mb->trailing_load_store()) {
3127 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3128 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3129 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3130 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3131 } else if (mb->leading()) {
3132 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3133 }
3134 }
3135 #endif
3136 // Count FPU ops and common calls, implements item (3)
3137 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3138 if (!gc_handled) {
3139 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3140 }
3141
3142 // Collect CFG split points
3143 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3144 frc._tests.push(n);
3145 }
3146 }
3147
3148 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3149 switch( nop ) {
3150 // Count all float operations that may use FPU
3151 case Op_AddF:
3152 case Op_SubF:
3153 case Op_MulF:
3154 case Op_DivF:
3155 case Op_NegF:
3156 case Op_ModF:
3157 case Op_ConvI2F:
3158 case Op_ConF:
3159 case Op_CmpF:
3160 case Op_CmpF3:
3161 case Op_StoreF:
3162 case Op_LoadF:
3163 // case Op_ConvL2F: // longs are split into 32-bit halves
3164 frc.inc_float_count();
3165 break;
3166
3167 case Op_ConvF2D:
3168 case Op_ConvD2F:
3169 frc.inc_float_count();
3170 frc.inc_double_count();
3171 break;
3172
3173 // Count all double operations that may use FPU
3174 case Op_AddD:
3175 case Op_SubD:
3176 case Op_MulD:
3177 case Op_DivD:
3178 case Op_NegD:
3179 case Op_ModD:
3180 case Op_ConvI2D:
3181 case Op_ConvD2I:
3182 // case Op_ConvL2D: // handled by leaf call
3183 // case Op_ConvD2L: // handled by leaf call
3184 case Op_ConD:
3185 case Op_CmpD:
3186 case Op_CmpD3:
3187 case Op_StoreD:
3188 case Op_LoadD:
3189 case Op_LoadD_unaligned:
3190 frc.inc_double_count();
3191 break;
3192 case Op_Opaque1: // Remove Opaque Nodes before matching
3193 case Op_Opaque3:
3194 n->subsume_by(n->in(1), this);
3195 break;
3196 case Op_CallStaticJava:
3197 case Op_CallJava:
3198 case Op_CallDynamicJava:
3199 frc.inc_java_call_count(); // Count java call site;
3200 case Op_CallRuntime:
3201 case Op_CallLeaf:
3202 case Op_CallLeafVector:
3203 case Op_CallLeafNoFP: {
3204 assert (n->is_Call(), "");
3205 CallNode *call = n->as_Call();
3206 // Count call sites where the FP mode bit would have to be flipped.
3207 // Do not count uncommon runtime calls:
3208 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3209 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3210 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3211 frc.inc_call_count(); // Count the call site
3212 } else { // See if uncommon argument is shared
3213 Node *n = call->in(TypeFunc::Parms);
3214 int nop = n->Opcode();
3215 // Clone shared simple arguments to uncommon calls, item (1).
3216 if (n->outcnt() > 1 &&
3217 !n->is_Proj() &&
3218 nop != Op_CreateEx &&
3219 nop != Op_CheckCastPP &&
3220 nop != Op_DecodeN &&
3221 nop != Op_DecodeNKlass &&
3222 !n->is_Mem() &&
3223 !n->is_Phi()) {
3224 Node *x = n->clone();
3225 call->set_req(TypeFunc::Parms, x);
3226 }
3227 }
3228 break;
3229 }
3230
3231 case Op_StoreCM:
3232 {
3233 // Convert OopStore dependence into precedence edge
3234 Node* prec = n->in(MemNode::OopStore);
3235 n->del_req(MemNode::OopStore);
3236 n->add_prec(prec);
3237 eliminate_redundant_card_marks(n);
3238 }
3239
3240 // fall through
3241
3242 case Op_StoreB:
3243 case Op_StoreC:
3244 case Op_StoreI:
3245 case Op_StoreL:
3246 case Op_CompareAndSwapB:
3247 case Op_CompareAndSwapS:
3248 case Op_CompareAndSwapI:
3249 case Op_CompareAndSwapL:
3250 case Op_CompareAndSwapP:
3251 case Op_CompareAndSwapN:
3252 case Op_WeakCompareAndSwapB:
3253 case Op_WeakCompareAndSwapS:
3254 case Op_WeakCompareAndSwapI:
3255 case Op_WeakCompareAndSwapL:
3256 case Op_WeakCompareAndSwapP:
3257 case Op_WeakCompareAndSwapN:
3258 case Op_CompareAndExchangeB:
3259 case Op_CompareAndExchangeS:
3260 case Op_CompareAndExchangeI:
3261 case Op_CompareAndExchangeL:
3262 case Op_CompareAndExchangeP:
3263 case Op_CompareAndExchangeN:
3264 case Op_GetAndAddS:
3265 case Op_GetAndAddB:
3266 case Op_GetAndAddI:
3267 case Op_GetAndAddL:
3268 case Op_GetAndSetS:
3269 case Op_GetAndSetB:
3270 case Op_GetAndSetI:
3271 case Op_GetAndSetL:
3272 case Op_GetAndSetP:
3273 case Op_GetAndSetN:
3274 case Op_StoreP:
3275 case Op_StoreN:
3276 case Op_StoreNKlass:
3277 case Op_LoadB:
3278 case Op_LoadUB:
3279 case Op_LoadUS:
3280 case Op_LoadI:
3281 case Op_LoadKlass:
3282 case Op_LoadNKlass:
3283 case Op_LoadL:
3284 case Op_LoadL_unaligned:
3285 case Op_LoadP:
3286 case Op_LoadN:
3287 case Op_LoadRange:
3288 case Op_LoadS:
3289 break;
3290
3291 case Op_AddP: { // Assert sane base pointers
3292 Node *addp = n->in(AddPNode::Address);
3293 assert( !addp->is_AddP() ||
3294 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3295 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3296 "Base pointers must match (addp %u)", addp->_idx );
3297 #ifdef _LP64
3298 if ((UseCompressedOops || UseCompressedClassPointers) &&
3299 addp->Opcode() == Op_ConP &&
3300 addp == n->in(AddPNode::Base) &&
3301 n->in(AddPNode::Offset)->is_Con()) {
3302 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3303 // on the platform and on the compressed oops mode.
3304 // Use addressing with narrow klass to load with offset on x86.
3305 // Some platforms can use the constant pool to load ConP.
3306 // Do this transformation here since IGVN will convert ConN back to ConP.
3307 const Type* t = addp->bottom_type();
3308 bool is_oop = t->isa_oopptr() != nullptr;
3309 bool is_klass = t->isa_klassptr() != nullptr;
3310
3311 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3312 (is_klass && Matcher::const_klass_prefer_decode())) {
3313 Node* nn = nullptr;
3314
3315 int op = is_oop ? Op_ConN : Op_ConNKlass;
3316
3317 // Look for existing ConN node of the same exact type.
3318 Node* r = root();
3319 uint cnt = r->outcnt();
3320 for (uint i = 0; i < cnt; i++) {
3321 Node* m = r->raw_out(i);
3322 if (m!= nullptr && m->Opcode() == op &&
3323 m->bottom_type()->make_ptr() == t) {
3324 nn = m;
3325 break;
3326 }
3327 }
3328 if (nn != nullptr) {
3329 // Decode a narrow oop to match address
3330 // [R12 + narrow_oop_reg<<3 + offset]
3331 if (is_oop) {
3332 nn = new DecodeNNode(nn, t);
3333 } else {
3334 nn = new DecodeNKlassNode(nn, t);
3335 }
3336 // Check for succeeding AddP which uses the same Base.
3337 // Otherwise we will run into the assertion above when visiting that guy.
3338 for (uint i = 0; i < n->outcnt(); ++i) {
3339 Node *out_i = n->raw_out(i);
3340 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3341 out_i->set_req(AddPNode::Base, nn);
3342 #ifdef ASSERT
3343 for (uint j = 0; j < out_i->outcnt(); ++j) {
3344 Node *out_j = out_i->raw_out(j);
3345 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3346 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3347 }
3348 #endif
3349 }
3350 }
3351 n->set_req(AddPNode::Base, nn);
3352 n->set_req(AddPNode::Address, nn);
3353 if (addp->outcnt() == 0) {
3354 addp->disconnect_inputs(this);
3355 }
3356 }
3357 }
3358 }
3359 #endif
3360 break;
3361 }
3362
3363 case Op_CastPP: {
3364 // Remove CastPP nodes to gain more freedom during scheduling but
3365 // keep the dependency they encode as control or precedence edges
3366 // (if control is set already) on memory operations. Some CastPP
3367 // nodes don't have a control (don't carry a dependency): skip
3368 // those.
3369 if (n->in(0) != nullptr) {
3370 ResourceMark rm;
3371 Unique_Node_List wq;
3372 wq.push(n);
3373 for (uint next = 0; next < wq.size(); ++next) {
3374 Node *m = wq.at(next);
3375 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3376 Node* use = m->fast_out(i);
3377 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3378 use->ensure_control_or_add_prec(n->in(0));
3379 } else {
3380 switch(use->Opcode()) {
3381 case Op_AddP:
3382 case Op_DecodeN:
3383 case Op_DecodeNKlass:
3384 case Op_CheckCastPP:
3385 case Op_CastPP:
3386 wq.push(use);
3387 break;
3388 }
3389 }
3390 }
3391 }
3392 }
3393 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3394 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3395 Node* in1 = n->in(1);
3396 const Type* t = n->bottom_type();
3397 Node* new_in1 = in1->clone();
3398 new_in1->as_DecodeN()->set_type(t);
3399
3400 if (!Matcher::narrow_oop_use_complex_address()) {
3401 //
3402 // x86, ARM and friends can handle 2 adds in addressing mode
3403 // and Matcher can fold a DecodeN node into address by using
3404 // a narrow oop directly and do implicit null check in address:
3405 //
3406 // [R12 + narrow_oop_reg<<3 + offset]
3407 // NullCheck narrow_oop_reg
3408 //
3409 // On other platforms (Sparc) we have to keep new DecodeN node and
3410 // use it to do implicit null check in address:
3411 //
3412 // decode_not_null narrow_oop_reg, base_reg
3413 // [base_reg + offset]
3414 // NullCheck base_reg
3415 //
3416 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3417 // to keep the information to which null check the new DecodeN node
3418 // corresponds to use it as value in implicit_null_check().
3419 //
3420 new_in1->set_req(0, n->in(0));
3421 }
3422
3423 n->subsume_by(new_in1, this);
3424 if (in1->outcnt() == 0) {
3425 in1->disconnect_inputs(this);
3426 }
3427 } else {
3428 n->subsume_by(n->in(1), this);
3429 if (n->outcnt() == 0) {
3430 n->disconnect_inputs(this);
3431 }
3432 }
3433 break;
3434 }
3435 #ifdef _LP64
3436 case Op_CmpP:
3437 // Do this transformation here to preserve CmpPNode::sub() and
3438 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3439 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3440 Node* in1 = n->in(1);
3441 Node* in2 = n->in(2);
3442 if (!in1->is_DecodeNarrowPtr()) {
3443 in2 = in1;
3444 in1 = n->in(2);
3445 }
3446 assert(in1->is_DecodeNarrowPtr(), "sanity");
3447
3448 Node* new_in2 = nullptr;
3449 if (in2->is_DecodeNarrowPtr()) {
3450 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3451 new_in2 = in2->in(1);
3452 } else if (in2->Opcode() == Op_ConP) {
3453 const Type* t = in2->bottom_type();
3454 if (t == TypePtr::NULL_PTR) {
3455 assert(in1->is_DecodeN(), "compare klass to null?");
3456 // Don't convert CmpP null check into CmpN if compressed
3457 // oops implicit null check is not generated.
3458 // This will allow to generate normal oop implicit null check.
3459 if (Matcher::gen_narrow_oop_implicit_null_checks())
3460 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3461 //
3462 // This transformation together with CastPP transformation above
3463 // will generated code for implicit null checks for compressed oops.
3464 //
3465 // The original code after Optimize()
3466 //
3467 // LoadN memory, narrow_oop_reg
3468 // decode narrow_oop_reg, base_reg
3469 // CmpP base_reg, nullptr
3470 // CastPP base_reg // NotNull
3471 // Load [base_reg + offset], val_reg
3472 //
3473 // after these transformations will be
3474 //
3475 // LoadN memory, narrow_oop_reg
3476 // CmpN narrow_oop_reg, nullptr
3477 // decode_not_null narrow_oop_reg, base_reg
3478 // Load [base_reg + offset], val_reg
3479 //
3480 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3481 // since narrow oops can be used in debug info now (see the code in
3482 // final_graph_reshaping_walk()).
3483 //
3484 // At the end the code will be matched to
3485 // on x86:
3486 //
3487 // Load_narrow_oop memory, narrow_oop_reg
3488 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3489 // NullCheck narrow_oop_reg
3490 //
3491 // and on sparc:
3492 //
3493 // Load_narrow_oop memory, narrow_oop_reg
3494 // decode_not_null narrow_oop_reg, base_reg
3495 // Load [base_reg + offset], val_reg
3496 // NullCheck base_reg
3497 //
3498 } else if (t->isa_oopptr()) {
3499 new_in2 = ConNode::make(t->make_narrowoop());
3500 } else if (t->isa_klassptr()) {
3501 new_in2 = ConNode::make(t->make_narrowklass());
3502 }
3503 }
3504 if (new_in2 != nullptr) {
3505 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3506 n->subsume_by(cmpN, this);
3507 if (in1->outcnt() == 0) {
3508 in1->disconnect_inputs(this);
3509 }
3510 if (in2->outcnt() == 0) {
3511 in2->disconnect_inputs(this);
3512 }
3513 }
3514 }
3515 break;
3516
3517 case Op_DecodeN:
3518 case Op_DecodeNKlass:
3519 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3520 // DecodeN could be pinned when it can't be fold into
3521 // an address expression, see the code for Op_CastPP above.
3522 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3523 break;
3524
3525 case Op_EncodeP:
3526 case Op_EncodePKlass: {
3527 Node* in1 = n->in(1);
3528 if (in1->is_DecodeNarrowPtr()) {
3529 n->subsume_by(in1->in(1), this);
3530 } else if (in1->Opcode() == Op_ConP) {
3531 const Type* t = in1->bottom_type();
3532 if (t == TypePtr::NULL_PTR) {
3533 assert(t->isa_oopptr(), "null klass?");
3534 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3535 } else if (t->isa_oopptr()) {
3536 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3537 } else if (t->isa_klassptr()) {
3538 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3539 }
3540 }
3541 if (in1->outcnt() == 0) {
3542 in1->disconnect_inputs(this);
3543 }
3544 break;
3545 }
3546
3547 case Op_Proj: {
3548 if (OptimizeStringConcat || IncrementalInline) {
3549 ProjNode* proj = n->as_Proj();
3550 if (proj->_is_io_use) {
3551 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3552 // Separate projections were used for the exception path which
3553 // are normally removed by a late inline. If it wasn't inlined
3554 // then they will hang around and should just be replaced with
3555 // the original one. Merge them.
3556 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3557 if (non_io_proj != nullptr) {
3558 proj->subsume_by(non_io_proj , this);
3559 }
3560 }
3561 }
3562 break;
3563 }
3564
3565 case Op_Phi:
3566 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3567 // The EncodeP optimization may create Phi with the same edges
3568 // for all paths. It is not handled well by Register Allocator.
3569 Node* unique_in = n->in(1);
3570 assert(unique_in != nullptr, "");
3571 uint cnt = n->req();
3572 for (uint i = 2; i < cnt; i++) {
3573 Node* m = n->in(i);
3574 assert(m != nullptr, "");
3575 if (unique_in != m)
3576 unique_in = nullptr;
3577 }
3578 if (unique_in != nullptr) {
3579 n->subsume_by(unique_in, this);
3580 }
3581 }
3582 break;
3583
3584 #endif
3585
3586 #ifdef ASSERT
3587 case Op_CastII:
3588 // Verify that all range check dependent CastII nodes were removed.
3589 if (n->isa_CastII()->has_range_check()) {
3590 n->dump(3);
3591 assert(false, "Range check dependent CastII node was not removed");
3592 }
3593 break;
3594 #endif
3595
3596 case Op_ModI:
3597 if (UseDivMod) {
3598 // Check if a%b and a/b both exist
3599 Node* d = n->find_similar(Op_DivI);
3600 if (d) {
3601 // Replace them with a fused divmod if supported
3602 if (Matcher::has_match_rule(Op_DivModI)) {
3603 DivModINode* divmod = DivModINode::make(n);
3604 d->subsume_by(divmod->div_proj(), this);
3605 n->subsume_by(divmod->mod_proj(), this);
3606 } else {
3607 // replace a%b with a-((a/b)*b)
3608 Node* mult = new MulINode(d, d->in(2));
3609 Node* sub = new SubINode(d->in(1), mult);
3610 n->subsume_by(sub, this);
3611 }
3612 }
3613 }
3614 break;
3615
3616 case Op_ModL:
3617 if (UseDivMod) {
3618 // Check if a%b and a/b both exist
3619 Node* d = n->find_similar(Op_DivL);
3620 if (d) {
3621 // Replace them with a fused divmod if supported
3622 if (Matcher::has_match_rule(Op_DivModL)) {
3623 DivModLNode* divmod = DivModLNode::make(n);
3624 d->subsume_by(divmod->div_proj(), this);
3625 n->subsume_by(divmod->mod_proj(), this);
3626 } else {
3627 // replace a%b with a-((a/b)*b)
3628 Node* mult = new MulLNode(d, d->in(2));
3629 Node* sub = new SubLNode(d->in(1), mult);
3630 n->subsume_by(sub, this);
3631 }
3632 }
3633 }
3634 break;
3635
3636 case Op_UModI:
3637 if (UseDivMod) {
3638 // Check if a%b and a/b both exist
3639 Node* d = n->find_similar(Op_UDivI);
3640 if (d) {
3641 // Replace them with a fused unsigned divmod if supported
3642 if (Matcher::has_match_rule(Op_UDivModI)) {
3643 UDivModINode* divmod = UDivModINode::make(n);
3644 d->subsume_by(divmod->div_proj(), this);
3645 n->subsume_by(divmod->mod_proj(), this);
3646 } else {
3647 // replace a%b with a-((a/b)*b)
3648 Node* mult = new MulINode(d, d->in(2));
3649 Node* sub = new SubINode(d->in(1), mult);
3650 n->subsume_by(sub, this);
3651 }
3652 }
3653 }
3654 break;
3655
3656 case Op_UModL:
3657 if (UseDivMod) {
3658 // Check if a%b and a/b both exist
3659 Node* d = n->find_similar(Op_UDivL);
3660 if (d) {
3661 // Replace them with a fused unsigned divmod if supported
3662 if (Matcher::has_match_rule(Op_UDivModL)) {
3663 UDivModLNode* divmod = UDivModLNode::make(n);
3664 d->subsume_by(divmod->div_proj(), this);
3665 n->subsume_by(divmod->mod_proj(), this);
3666 } else {
3667 // replace a%b with a-((a/b)*b)
3668 Node* mult = new MulLNode(d, d->in(2));
3669 Node* sub = new SubLNode(d->in(1), mult);
3670 n->subsume_by(sub, this);
3671 }
3672 }
3673 }
3674 break;
3675
3676 case Op_LoadVector:
3677 case Op_StoreVector:
3678 case Op_LoadVectorGather:
3679 case Op_StoreVectorScatter:
3680 case Op_LoadVectorGatherMasked:
3681 case Op_StoreVectorScatterMasked:
3682 case Op_VectorCmpMasked:
3683 case Op_VectorMaskGen:
3684 case Op_LoadVectorMasked:
3685 case Op_StoreVectorMasked:
3686 break;
3687
3688 case Op_AddReductionVI:
3689 case Op_AddReductionVL:
3690 case Op_AddReductionVF:
3691 case Op_AddReductionVD:
3692 case Op_MulReductionVI:
3693 case Op_MulReductionVL:
3694 case Op_MulReductionVF:
3695 case Op_MulReductionVD:
3696 case Op_MinReductionV:
3697 case Op_MaxReductionV:
3698 case Op_AndReductionV:
3699 case Op_OrReductionV:
3700 case Op_XorReductionV:
3701 break;
3702
3703 case Op_PackB:
3704 case Op_PackS:
3705 case Op_PackI:
3706 case Op_PackF:
3707 case Op_PackL:
3708 case Op_PackD:
3709 if (n->req()-1 > 2) {
3710 // Replace many operand PackNodes with a binary tree for matching
3711 PackNode* p = (PackNode*) n;
3712 Node* btp = p->binary_tree_pack(1, n->req());
3713 n->subsume_by(btp, this);
3714 }
3715 break;
3716 case Op_Loop:
3717 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3718 case Op_CountedLoop:
3719 case Op_LongCountedLoop:
3720 case Op_OuterStripMinedLoop:
3721 if (n->as_Loop()->is_inner_loop()) {
3722 frc.inc_inner_loop_count();
3723 }
3724 n->as_Loop()->verify_strip_mined(0);
3725 break;
3726 case Op_LShiftI:
3727 case Op_RShiftI:
3728 case Op_URShiftI:
3729 case Op_LShiftL:
3730 case Op_RShiftL:
3731 case Op_URShiftL:
3732 if (Matcher::need_masked_shift_count) {
3733 // The cpu's shift instructions don't restrict the count to the
3734 // lower 5/6 bits. We need to do the masking ourselves.
3735 Node* in2 = n->in(2);
3736 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3737 const TypeInt* t = in2->find_int_type();
3738 if (t != nullptr && t->is_con()) {
3739 juint shift = t->get_con();
3740 if (shift > mask) { // Unsigned cmp
3741 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3742 }
3743 } else {
3744 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3745 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3746 n->set_req(2, shift);
3747 }
3748 }
3749 if (in2->outcnt() == 0) { // Remove dead node
3750 in2->disconnect_inputs(this);
3751 }
3752 }
3753 break;
3754 case Op_MemBarStoreStore:
3755 case Op_MemBarRelease:
3756 // Break the link with AllocateNode: it is no longer useful and
3757 // confuses register allocation.
3758 if (n->req() > MemBarNode::Precedent) {
3759 n->set_req(MemBarNode::Precedent, top());
3760 }
3761 break;
3762 case Op_MemBarAcquire: {
3763 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3764 // At parse time, the trailing MemBarAcquire for a volatile load
3765 // is created with an edge to the load. After optimizations,
3766 // that input may be a chain of Phis. If those phis have no
3767 // other use, then the MemBarAcquire keeps them alive and
3768 // register allocation can be confused.
3769 dead_nodes.push(n->in(MemBarNode::Precedent));
3770 n->set_req(MemBarNode::Precedent, top());
3771 }
3772 break;
3773 }
3774 case Op_Blackhole:
3775 break;
3776 case Op_RangeCheck: {
3777 RangeCheckNode* rc = n->as_RangeCheck();
3778 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3779 n->subsume_by(iff, this);
3780 frc._tests.push(iff);
3781 break;
3782 }
3783 case Op_ConvI2L: {
3784 if (!Matcher::convi2l_type_required) {
3785 // Code generation on some platforms doesn't need accurate
3786 // ConvI2L types. Widening the type can help remove redundant
3787 // address computations.
3788 n->as_Type()->set_type(TypeLong::INT);
3789 ResourceMark rm;
3790 Unique_Node_List wq;
3791 wq.push(n);
3792 for (uint next = 0; next < wq.size(); next++) {
3793 Node *m = wq.at(next);
3794
3795 for(;;) {
3796 // Loop over all nodes with identical inputs edges as m
3797 Node* k = m->find_similar(m->Opcode());
3798 if (k == nullptr) {
3799 break;
3800 }
3801 // Push their uses so we get a chance to remove node made
3802 // redundant
3803 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3804 Node* u = k->fast_out(i);
3805 if (u->Opcode() == Op_LShiftL ||
3806 u->Opcode() == Op_AddL ||
3807 u->Opcode() == Op_SubL ||
3808 u->Opcode() == Op_AddP) {
3809 wq.push(u);
3810 }
3811 }
3812 // Replace all nodes with identical edges as m with m
3813 k->subsume_by(m, this);
3814 }
3815 }
3816 }
3817 break;
3818 }
3819 case Op_CmpUL: {
3820 if (!Matcher::has_match_rule(Op_CmpUL)) {
3821 // No support for unsigned long comparisons
3822 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3823 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3824 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3825 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3826 Node* andl = new AndLNode(orl, remove_sign_mask);
3827 Node* cmp = new CmpLNode(andl, n->in(2));
3828 n->subsume_by(cmp, this);
3829 }
3830 break;
3831 }
3832 default:
3833 assert(!n->is_Call(), "");
3834 assert(!n->is_Mem(), "");
3835 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3836 break;
3837 }
3838 }
3839
3840 //------------------------------final_graph_reshaping_walk---------------------
3841 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3842 // requires that the walk visits a node's inputs before visiting the node.
3843 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3844 Unique_Node_List sfpt;
3845
3846 frc._visited.set(root->_idx); // first, mark node as visited
3847 uint cnt = root->req();
3848 Node *n = root;
3849 uint i = 0;
3850 while (true) {
3851 if (i < cnt) {
3852 // Place all non-visited non-null inputs onto stack
3853 Node* m = n->in(i);
3854 ++i;
3855 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3856 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3857 // compute worst case interpreter size in case of a deoptimization
3858 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3859
3860 sfpt.push(m);
3861 }
3862 cnt = m->req();
3863 nstack.push(n, i); // put on stack parent and next input's index
3864 n = m;
3865 i = 0;
3866 }
3867 } else {
3868 // Now do post-visit work
3869 final_graph_reshaping_impl(n, frc, dead_nodes);
3870 if (nstack.is_empty())
3871 break; // finished
3872 n = nstack.node(); // Get node from stack
3873 cnt = n->req();
3874 i = nstack.index();
3875 nstack.pop(); // Shift to the next node on stack
3876 }
3877 }
3878
3879 // Skip next transformation if compressed oops are not used.
3880 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3881 (!UseCompressedOops && !UseCompressedClassPointers))
3882 return;
3883
3884 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3885 // It could be done for an uncommon traps or any safepoints/calls
3886 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3887 while (sfpt.size() > 0) {
3888 n = sfpt.pop();
3889 JVMState *jvms = n->as_SafePoint()->jvms();
3890 assert(jvms != nullptr, "sanity");
3891 int start = jvms->debug_start();
3892 int end = n->req();
3893 bool is_uncommon = (n->is_CallStaticJava() &&
3894 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3895 for (int j = start; j < end; j++) {
3896 Node* in = n->in(j);
3897 if (in->is_DecodeNarrowPtr()) {
3898 bool safe_to_skip = true;
3899 if (!is_uncommon ) {
3900 // Is it safe to skip?
3901 for (uint i = 0; i < in->outcnt(); i++) {
3902 Node* u = in->raw_out(i);
3903 if (!u->is_SafePoint() ||
3904 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3905 safe_to_skip = false;
3906 }
3907 }
3908 }
3909 if (safe_to_skip) {
3910 n->set_req(j, in->in(1));
3911 }
3912 if (in->outcnt() == 0) {
3913 in->disconnect_inputs(this);
3914 }
3915 }
3916 }
3917 }
3918 }
3919
3920 //------------------------------final_graph_reshaping--------------------------
3921 // Final Graph Reshaping.
3922 //
3923 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3924 // and not commoned up and forced early. Must come after regular
3925 // optimizations to avoid GVN undoing the cloning. Clone constant
3926 // inputs to Loop Phis; these will be split by the allocator anyways.
3927 // Remove Opaque nodes.
3928 // (2) Move last-uses by commutative operations to the left input to encourage
3929 // Intel update-in-place two-address operations and better register usage
3930 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3931 // calls canonicalizing them back.
3932 // (3) Count the number of double-precision FP ops, single-precision FP ops
3933 // and call sites. On Intel, we can get correct rounding either by
3934 // forcing singles to memory (requires extra stores and loads after each
3935 // FP bytecode) or we can set a rounding mode bit (requires setting and
3936 // clearing the mode bit around call sites). The mode bit is only used
3937 // if the relative frequency of single FP ops to calls is low enough.
3938 // This is a key transform for SPEC mpeg_audio.
3939 // (4) Detect infinite loops; blobs of code reachable from above but not
3940 // below. Several of the Code_Gen algorithms fail on such code shapes,
3941 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3942 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3943 // Detection is by looking for IfNodes where only 1 projection is
3944 // reachable from below or CatchNodes missing some targets.
3945 // (5) Assert for insane oop offsets in debug mode.
3946
3947 bool Compile::final_graph_reshaping() {
3948 // an infinite loop may have been eliminated by the optimizer,
3949 // in which case the graph will be empty.
3950 if (root()->req() == 1) {
3951 // Do not compile method that is only a trivial infinite loop,
3952 // since the content of the loop may have been eliminated.
3953 record_method_not_compilable("trivial infinite loop");
3954 return true;
3955 }
3956
3957 // Expensive nodes have their control input set to prevent the GVN
3958 // from freely commoning them. There's no GVN beyond this point so
3959 // no need to keep the control input. We want the expensive nodes to
3960 // be freely moved to the least frequent code path by gcm.
3961 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3962 for (int i = 0; i < expensive_count(); i++) {
3963 _expensive_nodes.at(i)->set_req(0, nullptr);
3964 }
3965
3966 Final_Reshape_Counts frc;
3967
3968 // Visit everybody reachable!
3969 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3970 Node_Stack nstack(live_nodes() >> 1);
3971 Unique_Node_List dead_nodes;
3972 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
3973
3974 // Check for unreachable (from below) code (i.e., infinite loops).
3975 for( uint i = 0; i < frc._tests.size(); i++ ) {
3976 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3977 // Get number of CFG targets.
3978 // Note that PCTables include exception targets after calls.
3979 uint required_outcnt = n->required_outcnt();
3980 if (n->outcnt() != required_outcnt) {
3981 // Check for a few special cases. Rethrow Nodes never take the
3982 // 'fall-thru' path, so expected kids is 1 less.
3983 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3984 if (n->in(0)->in(0)->is_Call()) {
3985 CallNode* call = n->in(0)->in(0)->as_Call();
3986 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3987 required_outcnt--; // Rethrow always has 1 less kid
3988 } else if (call->req() > TypeFunc::Parms &&
3989 call->is_CallDynamicJava()) {
3990 // Check for null receiver. In such case, the optimizer has
3991 // detected that the virtual call will always result in a null
3992 // pointer exception. The fall-through projection of this CatchNode
3993 // will not be populated.
3994 Node* arg0 = call->in(TypeFunc::Parms);
3995 if (arg0->is_Type() &&
3996 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3997 required_outcnt--;
3998 }
3999 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4000 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4001 // Check for illegal array length. In such case, the optimizer has
4002 // detected that the allocation attempt will always result in an
4003 // exception. There is no fall-through projection of this CatchNode .
4004 assert(call->is_CallStaticJava(), "static call expected");
4005 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4006 uint valid_length_test_input = call->req() - 1;
4007 Node* valid_length_test = call->in(valid_length_test_input);
4008 call->del_req(valid_length_test_input);
4009 if (valid_length_test->find_int_con(1) == 0) {
4010 required_outcnt--;
4011 }
4012 dead_nodes.push(valid_length_test);
4013 assert(n->outcnt() == required_outcnt, "malformed control flow");
4014 continue;
4015 }
4016 }
4017 }
4018
4019 // Recheck with a better notion of 'required_outcnt'
4020 if (n->outcnt() != required_outcnt) {
4021 record_method_not_compilable("malformed control flow");
4022 return true; // Not all targets reachable!
4023 }
4024 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4025 CallNode* call = n->in(0)->in(0)->as_Call();
4026 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4027 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4028 assert(call->is_CallStaticJava(), "static call expected");
4029 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4030 uint valid_length_test_input = call->req() - 1;
4031 dead_nodes.push(call->in(valid_length_test_input));
4032 call->del_req(valid_length_test_input); // valid length test useless now
4033 }
4034 }
4035 // Check that I actually visited all kids. Unreached kids
4036 // must be infinite loops.
4037 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4038 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4039 record_method_not_compilable("infinite loop");
4040 return true; // Found unvisited kid; must be unreach
4041 }
4042
4043 // Here so verification code in final_graph_reshaping_walk()
4044 // always see an OuterStripMinedLoopEnd
4045 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4046 IfNode* init_iff = n->as_If();
4047 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4048 n->subsume_by(iff, this);
4049 }
4050 }
4051
4052 while (dead_nodes.size() > 0) {
4053 Node* m = dead_nodes.pop();
4054 if (m->outcnt() == 0 && m != top()) {
4055 for (uint j = 0; j < m->req(); j++) {
4056 Node* in = m->in(j);
4057 if (in != nullptr) {
4058 dead_nodes.push(in);
4059 }
4060 }
4061 m->disconnect_inputs(this);
4062 }
4063 }
4064
4065 #ifdef IA32
4066 // If original bytecodes contained a mixture of floats and doubles
4067 // check if the optimizer has made it homogeneous, item (3).
4068 if (UseSSE == 0 &&
4069 frc.get_float_count() > 32 &&
4070 frc.get_double_count() == 0 &&
4071 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4072 set_24_bit_selection_and_mode(false, true);
4073 }
4074 #endif // IA32
4075
4076 set_java_calls(frc.get_java_call_count());
4077 set_inner_loops(frc.get_inner_loop_count());
4078
4079 // No infinite loops, no reason to bail out.
4080 return false;
4081 }
4082
4083 //-----------------------------too_many_traps----------------------------------
4084 // Report if there are too many traps at the current method and bci.
4085 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4086 bool Compile::too_many_traps(ciMethod* method,
4087 int bci,
4088 Deoptimization::DeoptReason reason) {
4089 ciMethodData* md = method->method_data();
4090 if (md->is_empty()) {
4091 // Assume the trap has not occurred, or that it occurred only
4092 // because of a transient condition during start-up in the interpreter.
4093 return false;
4094 }
4095 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4096 if (md->has_trap_at(bci, m, reason) != 0) {
4097 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4098 // Also, if there are multiple reasons, or if there is no per-BCI record,
4099 // assume the worst.
4100 if (log())
4101 log()->elem("observe trap='%s' count='%d'",
4102 Deoptimization::trap_reason_name(reason),
4103 md->trap_count(reason));
4104 return true;
4105 } else {
4106 // Ignore method/bci and see if there have been too many globally.
4107 return too_many_traps(reason, md);
4108 }
4109 }
4110
4111 // Less-accurate variant which does not require a method and bci.
4112 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4113 ciMethodData* logmd) {
4114 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4115 // Too many traps globally.
4116 // Note that we use cumulative trap_count, not just md->trap_count.
4117 if (log()) {
4118 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4119 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4120 Deoptimization::trap_reason_name(reason),
4121 mcount, trap_count(reason));
4122 }
4123 return true;
4124 } else {
4125 // The coast is clear.
4126 return false;
4127 }
4128 }
4129
4130 //--------------------------too_many_recompiles--------------------------------
4131 // Report if there are too many recompiles at the current method and bci.
4132 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4133 // Is not eager to return true, since this will cause the compiler to use
4134 // Action_none for a trap point, to avoid too many recompilations.
4135 bool Compile::too_many_recompiles(ciMethod* method,
4136 int bci,
4137 Deoptimization::DeoptReason reason) {
4138 ciMethodData* md = method->method_data();
4139 if (md->is_empty()) {
4140 // Assume the trap has not occurred, or that it occurred only
4141 // because of a transient condition during start-up in the interpreter.
4142 return false;
4143 }
4144 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4145 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4146 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4147 Deoptimization::DeoptReason per_bc_reason
4148 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4149 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4150 if ((per_bc_reason == Deoptimization::Reason_none
4151 || md->has_trap_at(bci, m, reason) != 0)
4152 // The trap frequency measure we care about is the recompile count:
4153 && md->trap_recompiled_at(bci, m)
4154 && md->overflow_recompile_count() >= bc_cutoff) {
4155 // Do not emit a trap here if it has already caused recompilations.
4156 // Also, if there are multiple reasons, or if there is no per-BCI record,
4157 // assume the worst.
4158 if (log())
4159 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4160 Deoptimization::trap_reason_name(reason),
4161 md->trap_count(reason),
4162 md->overflow_recompile_count());
4163 return true;
4164 } else if (trap_count(reason) != 0
4165 && decompile_count() >= m_cutoff) {
4166 // Too many recompiles globally, and we have seen this sort of trap.
4167 // Use cumulative decompile_count, not just md->decompile_count.
4168 if (log())
4169 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4170 Deoptimization::trap_reason_name(reason),
4171 md->trap_count(reason), trap_count(reason),
4172 md->decompile_count(), decompile_count());
4173 return true;
4174 } else {
4175 // The coast is clear.
4176 return false;
4177 }
4178 }
4179
4180 // Compute when not to trap. Used by matching trap based nodes and
4181 // NullCheck optimization.
4182 void Compile::set_allowed_deopt_reasons() {
4183 _allowed_reasons = 0;
4184 if (is_method_compilation()) {
4185 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4186 assert(rs < BitsPerInt, "recode bit map");
4187 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4188 _allowed_reasons |= nth_bit(rs);
4189 }
4190 }
4191 }
4192 }
4193
4194 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4195 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4196 }
4197
4198 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4199 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4200 }
4201
4202 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4203 if (holder->is_initialized()) {
4204 return false;
4205 }
4206 if (holder->is_being_initialized()) {
4207 if (accessing_method->holder() == holder) {
4208 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4209 // <init>, or a static method. In all those cases, there was an initialization
4210 // barrier on the holder klass passed.
4211 if (accessing_method->is_static_initializer() ||
4212 accessing_method->is_object_initializer() ||
4213 accessing_method->is_static()) {
4214 return false;
4215 }
4216 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4217 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4218 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4219 // child class can become fully initialized while its parent class is still being initialized.
4220 if (accessing_method->is_static_initializer()) {
4221 return false;
4222 }
4223 }
4224 ciMethod* root = method(); // the root method of compilation
4225 if (root != accessing_method) {
4226 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4227 }
4228 }
4229 return true;
4230 }
4231
4232 #ifndef PRODUCT
4233 //------------------------------verify_bidirectional_edges---------------------
4234 // For each input edge to a node (ie - for each Use-Def edge), verify that
4235 // there is a corresponding Def-Use edge.
4236 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4237 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4238 uint stack_size = live_nodes() >> 4;
4239 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4240 nstack.push(_root);
4241
4242 while (nstack.size() > 0) {
4243 Node* n = nstack.pop();
4244 if (visited.member(n)) {
4245 continue;
4246 }
4247 visited.push(n);
4248
4249 // Walk over all input edges, checking for correspondence
4250 uint length = n->len();
4251 for (uint i = 0; i < length; i++) {
4252 Node* in = n->in(i);
4253 if (in != nullptr && !visited.member(in)) {
4254 nstack.push(in); // Put it on stack
4255 }
4256 if (in != nullptr && !in->is_top()) {
4257 // Count instances of `next`
4258 int cnt = 0;
4259 for (uint idx = 0; idx < in->_outcnt; idx++) {
4260 if (in->_out[idx] == n) {
4261 cnt++;
4262 }
4263 }
4264 assert(cnt > 0, "Failed to find Def-Use edge.");
4265 // Check for duplicate edges
4266 // walk the input array downcounting the input edges to n
4267 for (uint j = 0; j < length; j++) {
4268 if (n->in(j) == in) {
4269 cnt--;
4270 }
4271 }
4272 assert(cnt == 0, "Mismatched edge count.");
4273 } else if (in == nullptr) {
4274 assert(i == 0 || i >= n->req() ||
4275 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4276 (n->is_Unlock() && i == (n->req() - 1)) ||
4277 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4278 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4279 } else {
4280 assert(in->is_top(), "sanity");
4281 // Nothing to check.
4282 }
4283 }
4284 }
4285 }
4286
4287 //------------------------------verify_graph_edges---------------------------
4288 // Walk the Graph and verify that there is a one-to-one correspondence
4289 // between Use-Def edges and Def-Use edges in the graph.
4290 void Compile::verify_graph_edges(bool no_dead_code) {
4291 if (VerifyGraphEdges) {
4292 Unique_Node_List visited;
4293
4294 // Call graph walk to check edges
4295 verify_bidirectional_edges(visited);
4296 if (no_dead_code) {
4297 // Now make sure that no visited node is used by an unvisited node.
4298 bool dead_nodes = false;
4299 Unique_Node_List checked;
4300 while (visited.size() > 0) {
4301 Node* n = visited.pop();
4302 checked.push(n);
4303 for (uint i = 0; i < n->outcnt(); i++) {
4304 Node* use = n->raw_out(i);
4305 if (checked.member(use)) continue; // already checked
4306 if (visited.member(use)) continue; // already in the graph
4307 if (use->is_Con()) continue; // a dead ConNode is OK
4308 // At this point, we have found a dead node which is DU-reachable.
4309 if (!dead_nodes) {
4310 tty->print_cr("*** Dead nodes reachable via DU edges:");
4311 dead_nodes = true;
4312 }
4313 use->dump(2);
4314 tty->print_cr("---");
4315 checked.push(use); // No repeats; pretend it is now checked.
4316 }
4317 }
4318 assert(!dead_nodes, "using nodes must be reachable from root");
4319 }
4320 }
4321 }
4322 #endif
4323
4324 // The Compile object keeps track of failure reasons separately from the ciEnv.
4325 // This is required because there is not quite a 1-1 relation between the
4326 // ciEnv and its compilation task and the Compile object. Note that one
4327 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4328 // to backtrack and retry without subsuming loads. Other than this backtracking
4329 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4330 // by the logic in C2Compiler.
4331 void Compile::record_failure(const char* reason) {
4332 if (log() != nullptr) {
4333 log()->elem("failure reason='%s' phase='compile'", reason);
4334 }
4335 if (_failure_reason == nullptr) {
4336 // Record the first failure reason.
4337 _failure_reason = reason;
4338 }
4339
4340 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4341 C->print_method(PHASE_FAILURE, 1);
4342 }
4343 _root = nullptr; // flush the graph, too
4344 }
4345
4346 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4347 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4348 _phase_name(name), _dolog(CITimeVerbose)
4349 {
4350 if (_dolog) {
4351 C = Compile::current();
4352 _log = C->log();
4353 } else {
4354 C = nullptr;
4355 _log = nullptr;
4356 }
4357 if (_log != nullptr) {
4358 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4359 _log->stamp();
4360 _log->end_head();
4361 }
4362 }
4363
4364 Compile::TracePhase::~TracePhase() {
4365
4366 C = Compile::current();
4367 if (_dolog) {
4368 _log = C->log();
4369 } else {
4370 _log = nullptr;
4371 }
4372
4373 #ifdef ASSERT
4374 if (PrintIdealNodeCount) {
4375 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4376 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4377 }
4378
4379 if (VerifyIdealNodeCount) {
4380 Compile::current()->print_missing_nodes();
4381 }
4382 #endif
4383
4384 if (_log != nullptr) {
4385 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4386 }
4387 }
4388
4389 //----------------------------static_subtype_check-----------------------------
4390 // Shortcut important common cases when superklass is exact:
4391 // (0) superklass is java.lang.Object (can occur in reflective code)
4392 // (1) subklass is already limited to a subtype of superklass => always ok
4393 // (2) subklass does not overlap with superklass => always fail
4394 // (3) superklass has NO subtypes and we can check with a simple compare.
4395 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4396 if (skip) {
4397 return SSC_full_test; // Let caller generate the general case.
4398 }
4399
4400 if (subk->is_java_subtype_of(superk)) {
4401 return SSC_always_true; // (0) and (1) this test cannot fail
4402 }
4403
4404 if (!subk->maybe_java_subtype_of(superk)) {
4405 return SSC_always_false; // (2) true path dead; no dynamic test needed
4406 }
4407
4408 const Type* superelem = superk;
4409 if (superk->isa_aryklassptr()) {
4410 int ignored;
4411 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4412 }
4413
4414 if (superelem->isa_instklassptr()) {
4415 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4416 if (!ik->has_subklass()) {
4417 if (!ik->is_final()) {
4418 // Add a dependency if there is a chance of a later subclass.
4419 dependencies()->assert_leaf_type(ik);
4420 }
4421 if (!superk->maybe_java_subtype_of(subk)) {
4422 return SSC_always_false;
4423 }
4424 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4425 }
4426 } else {
4427 // A primitive array type has no subtypes.
4428 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4429 }
4430
4431 return SSC_full_test;
4432 }
4433
4434 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4435 #ifdef _LP64
4436 // The scaled index operand to AddP must be a clean 64-bit value.
4437 // Java allows a 32-bit int to be incremented to a negative
4438 // value, which appears in a 64-bit register as a large
4439 // positive number. Using that large positive number as an
4440 // operand in pointer arithmetic has bad consequences.
4441 // On the other hand, 32-bit overflow is rare, and the possibility
4442 // can often be excluded, if we annotate the ConvI2L node with
4443 // a type assertion that its value is known to be a small positive
4444 // number. (The prior range check has ensured this.)
4445 // This assertion is used by ConvI2LNode::Ideal.
4446 int index_max = max_jint - 1; // array size is max_jint, index is one less
4447 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4448 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4449 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4450 #endif
4451 return idx;
4452 }
4453
4454 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4455 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4456 if (ctrl != nullptr) {
4457 // Express control dependency by a CastII node with a narrow type.
4458 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4459 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4460 // node from floating above the range check during loop optimizations. Otherwise, the
4461 // ConvI2L node may be eliminated independently of the range check, causing the data path
4462 // to become TOP while the control path is still there (although it's unreachable).
4463 value->set_req(0, ctrl);
4464 value = phase->transform(value);
4465 }
4466 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4467 return phase->transform(new ConvI2LNode(value, ltype));
4468 }
4469
4470 // The message about the current inlining is accumulated in
4471 // _print_inlining_stream and transferred into the _print_inlining_list
4472 // once we know whether inlining succeeds or not. For regular
4473 // inlining, messages are appended to the buffer pointed by
4474 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4475 // a new buffer is added after _print_inlining_idx in the list. This
4476 // way we can update the inlining message for late inlining call site
4477 // when the inlining is attempted again.
4478 void Compile::print_inlining_init() {
4479 if (print_inlining() || print_intrinsics()) {
4480 // print_inlining_init is actually called several times.
4481 print_inlining_reset();
4482 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4483 }
4484 }
4485
4486 void Compile::print_inlining_reinit() {
4487 if (print_inlining() || print_intrinsics()) {
4488 print_inlining_reset();
4489 }
4490 }
4491
4492 void Compile::print_inlining_reset() {
4493 _print_inlining_stream->reset();
4494 }
4495
4496 void Compile::print_inlining_commit() {
4497 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4498 // Transfer the message from _print_inlining_stream to the current
4499 // _print_inlining_list buffer and clear _print_inlining_stream.
4500 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4501 print_inlining_reset();
4502 }
4503
4504 void Compile::print_inlining_push() {
4505 // Add new buffer to the _print_inlining_list at current position
4506 _print_inlining_idx++;
4507 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4508 }
4509
4510 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4511 return _print_inlining_list->at(_print_inlining_idx);
4512 }
4513
4514 void Compile::print_inlining_update(CallGenerator* cg) {
4515 if (print_inlining() || print_intrinsics()) {
4516 if (cg->is_late_inline()) {
4517 if (print_inlining_current()->cg() != cg &&
4518 (print_inlining_current()->cg() != nullptr ||
4519 print_inlining_current()->ss()->size() != 0)) {
4520 print_inlining_push();
4521 }
4522 print_inlining_commit();
4523 print_inlining_current()->set_cg(cg);
4524 } else {
4525 if (print_inlining_current()->cg() != nullptr) {
4526 print_inlining_push();
4527 }
4528 print_inlining_commit();
4529 }
4530 }
4531 }
4532
4533 void Compile::print_inlining_move_to(CallGenerator* cg) {
4534 // We resume inlining at a late inlining call site. Locate the
4535 // corresponding inlining buffer so that we can update it.
4536 if (print_inlining() || print_intrinsics()) {
4537 for (int i = 0; i < _print_inlining_list->length(); i++) {
4538 if (_print_inlining_list->at(i)->cg() == cg) {
4539 _print_inlining_idx = i;
4540 return;
4541 }
4542 }
4543 ShouldNotReachHere();
4544 }
4545 }
4546
4547 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4548 if (print_inlining() || print_intrinsics()) {
4549 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4550 assert(print_inlining_current()->cg() == cg, "wrong entry");
4551 // replace message with new message
4552 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4553 print_inlining_commit();
4554 print_inlining_current()->set_cg(cg);
4555 }
4556 }
4557
4558 void Compile::print_inlining_assert_ready() {
4559 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4560 }
4561
4562 void Compile::process_print_inlining() {
4563 assert(_late_inlines.length() == 0, "not drained yet");
4564 if (print_inlining() || print_intrinsics()) {
4565 ResourceMark rm;
4566 stringStream ss;
4567 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
4568 for (int i = 0; i < _print_inlining_list->length(); i++) {
4569 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4570 ss.print("%s", pib->ss()->freeze());
4571 delete pib;
4572 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
4573 }
4574 // Reset _print_inlining_list, it only contains destructed objects.
4575 // It is on the arena, so it will be freed when the arena is reset.
4576 _print_inlining_list = nullptr;
4577 // _print_inlining_stream won't be used anymore, either.
4578 print_inlining_reset();
4579 size_t end = ss.size();
4580 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4581 strncpy(_print_inlining_output, ss.freeze(), end+1);
4582 _print_inlining_output[end] = 0;
4583 }
4584 }
4585
4586 void Compile::dump_print_inlining() {
4587 if (_print_inlining_output != nullptr) {
4588 tty->print_raw(_print_inlining_output);
4589 }
4590 }
4591
4592 void Compile::log_late_inline(CallGenerator* cg) {
4593 if (log() != nullptr) {
4594 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4595 cg->unique_id());
4596 JVMState* p = cg->call_node()->jvms();
4597 while (p != nullptr) {
4598 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4599 p = p->caller();
4600 }
4601 log()->tail("late_inline");
4602 }
4603 }
4604
4605 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4606 log_late_inline(cg);
4607 if (log() != nullptr) {
4608 log()->inline_fail(msg);
4609 }
4610 }
4611
4612 void Compile::log_inline_id(CallGenerator* cg) {
4613 if (log() != nullptr) {
4614 // The LogCompilation tool needs a unique way to identify late
4615 // inline call sites. This id must be unique for this call site in
4616 // this compilation. Try to have it unique across compilations as
4617 // well because it can be convenient when grepping through the log
4618 // file.
4619 // Distinguish OSR compilations from others in case CICountOSR is
4620 // on.
4621 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4622 cg->set_unique_id(id);
4623 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4624 }
4625 }
4626
4627 void Compile::log_inline_failure(const char* msg) {
4628 if (C->log() != nullptr) {
4629 C->log()->inline_fail(msg);
4630 }
4631 }
4632
4633
4634 // Dump inlining replay data to the stream.
4635 // Don't change thread state and acquire any locks.
4636 void Compile::dump_inline_data(outputStream* out) {
4637 InlineTree* inl_tree = ilt();
4638 if (inl_tree != nullptr) {
4639 out->print(" inline %d", inl_tree->count());
4640 inl_tree->dump_replay_data(out);
4641 }
4642 }
4643
4644 void Compile::dump_inline_data_reduced(outputStream* out) {
4645 assert(ReplayReduce, "");
4646
4647 InlineTree* inl_tree = ilt();
4648 if (inl_tree == nullptr) {
4649 return;
4650 }
4651 // Enable iterative replay file reduction
4652 // Output "compile" lines for depth 1 subtrees,
4653 // simulating that those trees were compiled
4654 // instead of inlined.
4655 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4656 InlineTree* sub = inl_tree->subtrees().at(i);
4657 if (sub->inline_level() != 1) {
4658 continue;
4659 }
4660
4661 ciMethod* method = sub->method();
4662 int entry_bci = -1;
4663 int comp_level = env()->task()->comp_level();
4664 out->print("compile ");
4665 method->dump_name_as_ascii(out);
4666 out->print(" %d %d", entry_bci, comp_level);
4667 out->print(" inline %d", sub->count());
4668 sub->dump_replay_data(out, -1);
4669 out->cr();
4670 }
4671 }
4672
4673 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4674 if (n1->Opcode() < n2->Opcode()) return -1;
4675 else if (n1->Opcode() > n2->Opcode()) return 1;
4676
4677 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4678 for (uint i = 1; i < n1->req(); i++) {
4679 if (n1->in(i) < n2->in(i)) return -1;
4680 else if (n1->in(i) > n2->in(i)) return 1;
4681 }
4682
4683 return 0;
4684 }
4685
4686 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4687 Node* n1 = *n1p;
4688 Node* n2 = *n2p;
4689
4690 return cmp_expensive_nodes(n1, n2);
4691 }
4692
4693 void Compile::sort_expensive_nodes() {
4694 if (!expensive_nodes_sorted()) {
4695 _expensive_nodes.sort(cmp_expensive_nodes);
4696 }
4697 }
4698
4699 bool Compile::expensive_nodes_sorted() const {
4700 for (int i = 1; i < _expensive_nodes.length(); i++) {
4701 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4702 return false;
4703 }
4704 }
4705 return true;
4706 }
4707
4708 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4709 if (_expensive_nodes.length() == 0) {
4710 return false;
4711 }
4712
4713 assert(OptimizeExpensiveOps, "optimization off?");
4714
4715 // Take this opportunity to remove dead nodes from the list
4716 int j = 0;
4717 for (int i = 0; i < _expensive_nodes.length(); i++) {
4718 Node* n = _expensive_nodes.at(i);
4719 if (!n->is_unreachable(igvn)) {
4720 assert(n->is_expensive(), "should be expensive");
4721 _expensive_nodes.at_put(j, n);
4722 j++;
4723 }
4724 }
4725 _expensive_nodes.trunc_to(j);
4726
4727 // Then sort the list so that similar nodes are next to each other
4728 // and check for at least two nodes of identical kind with same data
4729 // inputs.
4730 sort_expensive_nodes();
4731
4732 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4733 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4734 return true;
4735 }
4736 }
4737
4738 return false;
4739 }
4740
4741 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4742 if (_expensive_nodes.length() == 0) {
4743 return;
4744 }
4745
4746 assert(OptimizeExpensiveOps, "optimization off?");
4747
4748 // Sort to bring similar nodes next to each other and clear the
4749 // control input of nodes for which there's only a single copy.
4750 sort_expensive_nodes();
4751
4752 int j = 0;
4753 int identical = 0;
4754 int i = 0;
4755 bool modified = false;
4756 for (; i < _expensive_nodes.length()-1; i++) {
4757 assert(j <= i, "can't write beyond current index");
4758 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4759 identical++;
4760 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4761 continue;
4762 }
4763 if (identical > 0) {
4764 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4765 identical = 0;
4766 } else {
4767 Node* n = _expensive_nodes.at(i);
4768 igvn.replace_input_of(n, 0, nullptr);
4769 igvn.hash_insert(n);
4770 modified = true;
4771 }
4772 }
4773 if (identical > 0) {
4774 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4775 } else if (_expensive_nodes.length() >= 1) {
4776 Node* n = _expensive_nodes.at(i);
4777 igvn.replace_input_of(n, 0, nullptr);
4778 igvn.hash_insert(n);
4779 modified = true;
4780 }
4781 _expensive_nodes.trunc_to(j);
4782 if (modified) {
4783 igvn.optimize();
4784 }
4785 }
4786
4787 void Compile::add_expensive_node(Node * n) {
4788 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4789 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4790 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4791 if (OptimizeExpensiveOps) {
4792 _expensive_nodes.append(n);
4793 } else {
4794 // Clear control input and let IGVN optimize expensive nodes if
4795 // OptimizeExpensiveOps is off.
4796 n->set_req(0, nullptr);
4797 }
4798 }
4799
4800 /**
4801 * Track coarsened Lock and Unlock nodes.
4802 */
4803
4804 class Lock_List : public Node_List {
4805 uint _origin_cnt;
4806 public:
4807 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4808 uint origin_cnt() const { return _origin_cnt; }
4809 };
4810
4811 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4812 int length = locks.length();
4813 if (length > 0) {
4814 // Have to keep this list until locks elimination during Macro nodes elimination.
4815 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4816 for (int i = 0; i < length; i++) {
4817 AbstractLockNode* lock = locks.at(i);
4818 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4819 locks_list->push(lock);
4820 }
4821 _coarsened_locks.append(locks_list);
4822 }
4823 }
4824
4825 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4826 int count = coarsened_count();
4827 for (int i = 0; i < count; i++) {
4828 Node_List* locks_list = _coarsened_locks.at(i);
4829 for (uint j = 0; j < locks_list->size(); j++) {
4830 Node* lock = locks_list->at(j);
4831 assert(lock->is_AbstractLock(), "sanity");
4832 if (!useful.member(lock)) {
4833 locks_list->yank(lock);
4834 }
4835 }
4836 }
4837 }
4838
4839 void Compile::remove_coarsened_lock(Node* n) {
4840 if (n->is_AbstractLock()) {
4841 int count = coarsened_count();
4842 for (int i = 0; i < count; i++) {
4843 Node_List* locks_list = _coarsened_locks.at(i);
4844 locks_list->yank(n);
4845 }
4846 }
4847 }
4848
4849 bool Compile::coarsened_locks_consistent() {
4850 int count = coarsened_count();
4851 for (int i = 0; i < count; i++) {
4852 bool unbalanced = false;
4853 bool modified = false; // track locks kind modifications
4854 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4855 uint size = locks_list->size();
4856 if (size == 0) {
4857 unbalanced = false; // All locks were eliminated - good
4858 } else if (size != locks_list->origin_cnt()) {
4859 unbalanced = true; // Some locks were removed from list
4860 } else {
4861 for (uint j = 0; j < size; j++) {
4862 Node* lock = locks_list->at(j);
4863 // All nodes in group should have the same state (modified or not)
4864 if (!lock->as_AbstractLock()->is_coarsened()) {
4865 if (j == 0) {
4866 // first on list was modified, the rest should be too for consistency
4867 modified = true;
4868 } else if (!modified) {
4869 // this lock was modified but previous locks on the list were not
4870 unbalanced = true;
4871 break;
4872 }
4873 } else if (modified) {
4874 // previous locks on list were modified but not this lock
4875 unbalanced = true;
4876 break;
4877 }
4878 }
4879 }
4880 if (unbalanced) {
4881 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4882 #ifdef ASSERT
4883 if (PrintEliminateLocks) {
4884 tty->print_cr("=== unbalanced coarsened locks ===");
4885 for (uint l = 0; l < size; l++) {
4886 locks_list->at(l)->dump();
4887 }
4888 }
4889 #endif
4890 record_failure(C2Compiler::retry_no_locks_coarsening());
4891 return false;
4892 }
4893 }
4894 return true;
4895 }
4896
4897 /**
4898 * Remove the speculative part of types and clean up the graph
4899 */
4900 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4901 if (UseTypeSpeculation) {
4902 Unique_Node_List worklist;
4903 worklist.push(root());
4904 int modified = 0;
4905 // Go over all type nodes that carry a speculative type, drop the
4906 // speculative part of the type and enqueue the node for an igvn
4907 // which may optimize it out.
4908 for (uint next = 0; next < worklist.size(); ++next) {
4909 Node *n = worklist.at(next);
4910 if (n->is_Type()) {
4911 TypeNode* tn = n->as_Type();
4912 const Type* t = tn->type();
4913 const Type* t_no_spec = t->remove_speculative();
4914 if (t_no_spec != t) {
4915 bool in_hash = igvn.hash_delete(n);
4916 assert(in_hash, "node should be in igvn hash table");
4917 tn->set_type(t_no_spec);
4918 igvn.hash_insert(n);
4919 igvn._worklist.push(n); // give it a chance to go away
4920 modified++;
4921 }
4922 }
4923 // Iterate over outs - endless loops is unreachable from below
4924 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4925 Node *m = n->fast_out(i);
4926 if (not_a_node(m)) {
4927 continue;
4928 }
4929 worklist.push(m);
4930 }
4931 }
4932 // Drop the speculative part of all types in the igvn's type table
4933 igvn.remove_speculative_types();
4934 if (modified > 0) {
4935 igvn.optimize();
4936 if (failing()) return;
4937 }
4938 #ifdef ASSERT
4939 // Verify that after the IGVN is over no speculative type has resurfaced
4940 worklist.clear();
4941 worklist.push(root());
4942 for (uint next = 0; next < worklist.size(); ++next) {
4943 Node *n = worklist.at(next);
4944 const Type* t = igvn.type_or_null(n);
4945 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4946 if (n->is_Type()) {
4947 t = n->as_Type()->type();
4948 assert(t == t->remove_speculative(), "no more speculative types");
4949 }
4950 // Iterate over outs - endless loops is unreachable from below
4951 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4952 Node *m = n->fast_out(i);
4953 if (not_a_node(m)) {
4954 continue;
4955 }
4956 worklist.push(m);
4957 }
4958 }
4959 igvn.check_no_speculative_types();
4960 #endif
4961 }
4962 }
4963
4964 // Auxiliary methods to support randomized stressing/fuzzing.
4965
4966 int Compile::random() {
4967 _stress_seed = os::next_random(_stress_seed);
4968 return static_cast<int>(_stress_seed);
4969 }
4970
4971 // This method can be called the arbitrary number of times, with current count
4972 // as the argument. The logic allows selecting a single candidate from the
4973 // running list of candidates as follows:
4974 // int count = 0;
4975 // Cand* selected = null;
4976 // while(cand = cand->next()) {
4977 // if (randomized_select(++count)) {
4978 // selected = cand;
4979 // }
4980 // }
4981 //
4982 // Including count equalizes the chances any candidate is "selected".
4983 // This is useful when we don't have the complete list of candidates to choose
4984 // from uniformly. In this case, we need to adjust the randomicity of the
4985 // selection, or else we will end up biasing the selection towards the latter
4986 // candidates.
4987 //
4988 // Quick back-envelope calculation shows that for the list of n candidates
4989 // the equal probability for the candidate to persist as "best" can be
4990 // achieved by replacing it with "next" k-th candidate with the probability
4991 // of 1/k. It can be easily shown that by the end of the run, the
4992 // probability for any candidate is converged to 1/n, thus giving the
4993 // uniform distribution among all the candidates.
4994 //
4995 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4996 #define RANDOMIZED_DOMAIN_POW 29
4997 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4998 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4999 bool Compile::randomized_select(int count) {
5000 assert(count > 0, "only positive");
5001 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5002 }
5003
5004 CloneMap& Compile::clone_map() { return _clone_map; }
5005 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5006
5007 void NodeCloneInfo::dump_on(outputStream* st) const {
5008 st->print(" {%d:%d} ", idx(), gen());
5009 }
5010
5011 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5012 uint64_t val = value(old->_idx);
5013 NodeCloneInfo cio(val);
5014 assert(val != 0, "old node should be in the map");
5015 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5016 insert(nnn->_idx, cin.get());
5017 #ifndef PRODUCT
5018 if (is_debug()) {
5019 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5020 }
5021 #endif
5022 }
5023
5024 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5025 NodeCloneInfo cio(value(old->_idx));
5026 if (cio.get() == 0) {
5027 cio.set(old->_idx, 0);
5028 insert(old->_idx, cio.get());
5029 #ifndef PRODUCT
5030 if (is_debug()) {
5031 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5032 }
5033 #endif
5034 }
5035 clone(old, nnn, gen);
5036 }
5037
5038 int CloneMap::max_gen() const {
5039 int g = 0;
5040 DictI di(_dict);
5041 for(; di.test(); ++di) {
5042 int t = gen(di._key);
5043 if (g < t) {
5044 g = t;
5045 #ifndef PRODUCT
5046 if (is_debug()) {
5047 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5048 }
5049 #endif
5050 }
5051 }
5052 return g;
5053 }
5054
5055 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5056 uint64_t val = value(key);
5057 if (val != 0) {
5058 NodeCloneInfo ni(val);
5059 ni.dump_on(st);
5060 }
5061 }
5062
5063 // Move Allocate nodes to the start of the list
5064 void Compile::sort_macro_nodes() {
5065 int count = macro_count();
5066 int allocates = 0;
5067 for (int i = 0; i < count; i++) {
5068 Node* n = macro_node(i);
5069 if (n->is_Allocate()) {
5070 if (i != allocates) {
5071 Node* tmp = macro_node(allocates);
5072 _macro_nodes.at_put(allocates, n);
5073 _macro_nodes.at_put(i, tmp);
5074 }
5075 allocates++;
5076 }
5077 }
5078 }
5079
5080 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5081 EventCompilerPhase event;
5082 if (event.should_commit()) {
5083 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5084 }
5085 #ifndef PRODUCT
5086 ResourceMark rm;
5087 stringStream ss;
5088 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5089 if (n != nullptr) {
5090 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5091 }
5092
5093 const char* name = ss.as_string();
5094 if (should_print_igv(level)) {
5095 _igv_printer->print_method(name, level);
5096 }
5097 if (should_print_phase(cpt)) {
5098 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5099 }
5100 #endif
5101 C->_latest_stage_start_counter.stamp();
5102 }
5103
5104 // Only used from CompileWrapper
5105 void Compile::begin_method() {
5106 #ifndef PRODUCT
5107 if (_method != nullptr && should_print_igv(1)) {
5108 _igv_printer->begin_method();
5109 }
5110 #endif
5111 C->_latest_stage_start_counter.stamp();
5112 }
5113
5114 // Only used from CompileWrapper
5115 void Compile::end_method() {
5116 EventCompilerPhase event;
5117 if (event.should_commit()) {
5118 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5119 }
5120
5121 #ifndef PRODUCT
5122 if (_method != nullptr && should_print_igv(1)) {
5123 _igv_printer->end_method();
5124 }
5125 #endif
5126 }
5127
5128 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5129 #ifndef PRODUCT
5130 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) {
5131 return true;
5132 }
5133 #endif
5134 return false;
5135 }
5136
5137 bool Compile::should_print_igv(int level) {
5138 #ifndef PRODUCT
5139 if (PrintIdealGraphLevel < 0) { // disabled by the user
5140 return false;
5141 }
5142
5143 bool need = directive()->IGVPrintLevelOption >= level;
5144 if (need && !_igv_printer) {
5145 _igv_printer = IdealGraphPrinter::printer();
5146 _igv_printer->set_compile(this);
5147 }
5148 return need;
5149 #else
5150 return false;
5151 #endif
5152 }
5153
5154 #ifndef PRODUCT
5155 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5156 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5157
5158 // Called from debugger. Prints method to the default file with the default phase name.
5159 // This works regardless of any Ideal Graph Visualizer flags set or not.
5160 void igv_print() {
5161 Compile::current()->igv_print_method_to_file();
5162 }
5163
5164 // Same as igv_print() above but with a specified phase name.
5165 void igv_print(const char* phase_name) {
5166 Compile::current()->igv_print_method_to_file(phase_name);
5167 }
5168
5169 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5170 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5171 // This works regardless of any Ideal Graph Visualizer flags set or not.
5172 void igv_print(bool network) {
5173 if (network) {
5174 Compile::current()->igv_print_method_to_network();
5175 } else {
5176 Compile::current()->igv_print_method_to_file();
5177 }
5178 }
5179
5180 // Same as igv_print(bool network) above but with a specified phase name.
5181 void igv_print(bool network, const char* phase_name) {
5182 if (network) {
5183 Compile::current()->igv_print_method_to_network(phase_name);
5184 } else {
5185 Compile::current()->igv_print_method_to_file(phase_name);
5186 }
5187 }
5188
5189 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5190 void igv_print_default() {
5191 Compile::current()->print_method(PHASE_DEBUG, 0);
5192 }
5193
5194 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5195 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5196 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5197 void igv_append() {
5198 Compile::current()->igv_print_method_to_file("Debug", true);
5199 }
5200
5201 // Same as igv_append() above but with a specified phase name.
5202 void igv_append(const char* phase_name) {
5203 Compile::current()->igv_print_method_to_file(phase_name, true);
5204 }
5205
5206 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5207 const char* file_name = "custom_debug.xml";
5208 if (_debug_file_printer == nullptr) {
5209 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5210 } else {
5211 _debug_file_printer->update_compiled_method(C->method());
5212 }
5213 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5214 _debug_file_printer->print(phase_name, (Node*)C->root());
5215 }
5216
5217 void Compile::igv_print_method_to_network(const char* phase_name) {
5218 if (_debug_network_printer == nullptr) {
5219 _debug_network_printer = new IdealGraphPrinter(C);
5220 } else {
5221 _debug_network_printer->update_compiled_method(C->method());
5222 }
5223 tty->print_cr("Method printed over network stream to IGV");
5224 _debug_network_printer->print(phase_name, (Node*)C->root());
5225 }
5226 #endif
5227
5228 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5229 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5230 return value;
5231 }
5232 Node* result = nullptr;
5233 if (bt == T_BYTE) {
5234 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5235 result = new RShiftINode(result, phase->intcon(24));
5236 } else if (bt == T_BOOLEAN) {
5237 result = new AndINode(value, phase->intcon(0xFF));
5238 } else if (bt == T_CHAR) {
5239 result = new AndINode(value,phase->intcon(0xFFFF));
5240 } else {
5241 assert(bt == T_SHORT, "unexpected narrow type");
5242 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5243 result = new RShiftINode(result, phase->intcon(16));
5244 }
5245 if (transform_res) {
5246 result = phase->transform(result);
5247 }
5248 return result;
5249 }
5250