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/inlinetypenode.hpp"
57 #include "opto/loopnode.hpp"
58 #include "opto/machnode.hpp"
59 #include "opto/macro.hpp"
60 #include "opto/matcher.hpp"
61 #include "opto/mathexactnode.hpp"
62 #include "opto/memnode.hpp"
63 #include "opto/mulnode.hpp"
64 #include "opto/narrowptrnode.hpp"
65 #include "opto/node.hpp"
66 #include "opto/opcodes.hpp"
67 #include "opto/output.hpp"
68 #include "opto/parse.hpp"
69 #include "opto/phaseX.hpp"
70 #include "opto/rootnode.hpp"
71 #include "opto/runtime.hpp"
72 #include "opto/stringopts.hpp"
73 #include "opto/type.hpp"
74 #include "opto/vector.hpp"
75 #include "opto/vectornode.hpp"
76 #include "runtime/globals_extension.hpp"
77 #include "runtime/sharedRuntime.hpp"
78 #include "runtime/signature.hpp"
79 #include "runtime/stubRoutines.hpp"
80 #include "runtime/timer.hpp"
81 #include "utilities/align.hpp"
82 #include "utilities/copy.hpp"
83 #include "utilities/macros.hpp"
84 #include "utilities/resourceHash.hpp"
85
86 // -------------------- Compile::mach_constant_base_node -----------------------
87 // Constant table base node singleton.
88 MachConstantBaseNode* Compile::mach_constant_base_node() {
89 if (_mach_constant_base_node == nullptr) {
90 _mach_constant_base_node = new MachConstantBaseNode();
91 _mach_constant_base_node->add_req(C->root());
92 }
93 return _mach_constant_base_node;
94 }
95
96
97 /// Support for intrinsics.
98
99 // Return the index at which m must be inserted (or already exists).
100 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
101 class IntrinsicDescPair {
102 private:
103 ciMethod* _m;
104 bool _is_virtual;
105 public:
106 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
107 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
108 ciMethod* m= elt->method();
109 ciMethod* key_m = key->_m;
110 if (key_m < m) return -1;
111 else if (key_m > m) return 1;
112 else {
113 bool is_virtual = elt->is_virtual();
114 bool key_virtual = key->_is_virtual;
115 if (key_virtual < is_virtual) return -1;
116 else if (key_virtual > is_virtual) return 1;
117 else return 0;
118 }
119 }
120 };
121 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
122 #ifdef ASSERT
123 for (int i = 1; i < _intrinsics.length(); i++) {
124 CallGenerator* cg1 = _intrinsics.at(i-1);
125 CallGenerator* cg2 = _intrinsics.at(i);
126 assert(cg1->method() != cg2->method()
127 ? cg1->method() < cg2->method()
128 : cg1->is_virtual() < cg2->is_virtual(),
129 "compiler intrinsics list must stay sorted");
130 }
131 #endif
132 IntrinsicDescPair pair(m, is_virtual);
133 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
134 }
135
136 void Compile::register_intrinsic(CallGenerator* cg) {
137 bool found = false;
138 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
139 assert(!found, "registering twice");
140 _intrinsics.insert_before(index, cg);
141 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
142 }
143
144 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
145 assert(m->is_loaded(), "don't try this on unloaded methods");
146 if (_intrinsics.length() > 0) {
147 bool found = false;
148 int index = intrinsic_insertion_index(m, is_virtual, found);
149 if (found) {
150 return _intrinsics.at(index);
151 }
152 }
153 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
154 if (m->intrinsic_id() != vmIntrinsics::_none &&
155 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
156 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
157 if (cg != nullptr) {
158 // Save it for next time:
159 register_intrinsic(cg);
160 return cg;
161 } else {
162 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
163 }
164 }
165 return nullptr;
166 }
167
168 // Compile::make_vm_intrinsic is defined in library_call.cpp.
169
170 #ifndef PRODUCT
171 // statistics gathering...
172
173 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
174 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
175
176 inline int as_int(vmIntrinsics::ID id) {
177 return vmIntrinsics::as_int(id);
178 }
179
180 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
181 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
182 int oflags = _intrinsic_hist_flags[as_int(id)];
183 assert(flags != 0, "what happened?");
184 if (is_virtual) {
185 flags |= _intrinsic_virtual;
186 }
187 bool changed = (flags != oflags);
188 if ((flags & _intrinsic_worked) != 0) {
189 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
190 if (count == 1) {
191 changed = true; // first time
192 }
193 // increment the overall count also:
194 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
195 }
196 if (changed) {
197 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
198 // Something changed about the intrinsic's virtuality.
199 if ((flags & _intrinsic_virtual) != 0) {
200 // This is the first use of this intrinsic as a virtual call.
201 if (oflags != 0) {
202 // We already saw it as a non-virtual, so note both cases.
203 flags |= _intrinsic_both;
204 }
205 } else if ((oflags & _intrinsic_both) == 0) {
206 // This is the first use of this intrinsic as a non-virtual
207 flags |= _intrinsic_both;
208 }
209 }
210 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
211 }
212 // update the overall flags also:
213 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
214 return changed;
215 }
216
217 static char* format_flags(int flags, char* buf) {
218 buf[0] = 0;
219 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
220 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
221 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
222 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
223 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
224 if (buf[0] == 0) strcat(buf, ",");
225 assert(buf[0] == ',', "must be");
226 return &buf[1];
227 }
228
229 void Compile::print_intrinsic_statistics() {
230 char flagsbuf[100];
231 ttyLocker ttyl;
232 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
233 tty->print_cr("Compiler intrinsic usage:");
234 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
235 if (total == 0) total = 1; // avoid div0 in case of no successes
236 #define PRINT_STAT_LINE(name, c, f) \
237 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
238 for (auto id : EnumRange<vmIntrinsicID>{}) {
239 int flags = _intrinsic_hist_flags[as_int(id)];
240 juint count = _intrinsic_hist_count[as_int(id)];
241 if ((flags | count) != 0) {
242 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
243 }
244 }
245 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
246 if (xtty != nullptr) xtty->tail("statistics");
247 }
248
249 void Compile::print_statistics() {
250 { ttyLocker ttyl;
251 if (xtty != nullptr) xtty->head("statistics type='opto'");
252 Parse::print_statistics();
253 PhaseStringOpts::print_statistics();
254 PhaseCCP::print_statistics();
255 PhaseRegAlloc::print_statistics();
256 PhaseOutput::print_statistics();
257 PhasePeephole::print_statistics();
258 PhaseIdealLoop::print_statistics();
259 ConnectionGraph::print_statistics();
260 PhaseMacroExpand::print_statistics();
261 if (xtty != nullptr) xtty->tail("statistics");
262 }
263 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
264 // put this under its own <statistics> element.
265 print_intrinsic_statistics();
266 }
267 }
268 #endif //PRODUCT
269
270 void Compile::gvn_replace_by(Node* n, Node* nn) {
271 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
272 Node* use = n->last_out(i);
273 bool is_in_table = initial_gvn()->hash_delete(use);
274 uint uses_found = 0;
275 for (uint j = 0; j < use->len(); j++) {
276 if (use->in(j) == n) {
277 if (j < use->req())
278 use->set_req(j, nn);
279 else
280 use->set_prec(j, nn);
281 uses_found++;
282 }
283 }
284 if (is_in_table) {
285 // reinsert into table
286 initial_gvn()->hash_find_insert(use);
287 }
288 record_for_igvn(use);
289 i -= uses_found; // we deleted 1 or more copies of this edge
290 }
291 }
292
293
294 // Identify all nodes that are reachable from below, useful.
295 // Use breadth-first pass that records state in a Unique_Node_List,
296 // recursive traversal is slower.
297 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
298 int estimated_worklist_size = live_nodes();
299 useful.map( estimated_worklist_size, nullptr ); // preallocate space
300
301 // Initialize worklist
302 if (root() != nullptr) { useful.push(root()); }
303 // If 'top' is cached, declare it useful to preserve cached node
304 if (cached_top_node()) { useful.push(cached_top_node()); }
305
306 // Push all useful nodes onto the list, breadthfirst
307 for( uint next = 0; next < useful.size(); ++next ) {
308 assert( next < unique(), "Unique useful nodes < total nodes");
309 Node *n = useful.at(next);
310 uint max = n->len();
311 for( uint i = 0; i < max; ++i ) {
312 Node *m = n->in(i);
313 if (not_a_node(m)) continue;
314 useful.push(m);
315 }
316 }
317 }
318
319 // Update dead_node_list with any missing dead nodes using useful
320 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
321 void Compile::update_dead_node_list(Unique_Node_List &useful) {
322 uint max_idx = unique();
323 VectorSet& useful_node_set = useful.member_set();
324
325 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
326 // If node with index node_idx is not in useful set,
327 // mark it as dead in dead node list.
328 if (!useful_node_set.test(node_idx)) {
329 record_dead_node(node_idx);
330 }
331 }
332 }
333
334 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
335 int shift = 0;
336 for (int i = 0; i < inlines->length(); i++) {
337 CallGenerator* cg = inlines->at(i);
338 if (useful.member(cg->call_node())) {
339 if (shift > 0) {
340 inlines->at_put(i - shift, cg);
341 }
342 } else {
343 shift++; // skip over the dead element
344 }
345 }
346 if (shift > 0) {
347 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
348 }
349 }
350
351 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
352 assert(dead != nullptr && dead->is_Call(), "sanity");
353 int found = 0;
354 for (int i = 0; i < inlines->length(); i++) {
355 if (inlines->at(i)->call_node() == dead) {
356 inlines->remove_at(i);
357 found++;
358 NOT_DEBUG( break; ) // elements are unique, so exit early
359 }
360 }
361 assert(found <= 1, "not unique");
362 }
363
364 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) {
365 for (int i = node_list.length() - 1; i >= 0; i--) {
366 Node* n = node_list.at(i);
367 if (!useful.member(n)) {
368 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
369 }
370 }
371 }
372
373 void Compile::remove_useless_node(Node* dead) {
374 remove_modified_node(dead);
375
376 // Constant node that has no out-edges and has only one in-edge from
377 // root is usually dead. However, sometimes reshaping walk makes
378 // it reachable by adding use edges. So, we will NOT count Con nodes
379 // as dead to be conservative about the dead node count at any
380 // given time.
381 if (!dead->is_Con()) {
382 record_dead_node(dead->_idx);
383 }
384 if (dead->is_macro()) {
385 remove_macro_node(dead);
386 }
387 if (dead->is_expensive()) {
388 remove_expensive_node(dead);
389 }
390 if (dead->Opcode() == Op_Opaque4) {
391 remove_template_assertion_predicate_opaq(dead);
392 }
393 if (dead->for_post_loop_opts_igvn()) {
394 remove_from_post_loop_opts_igvn(dead);
395 }
396 if (dead->is_InlineType()) {
397 remove_inline_type(dead);
398 }
399 if (dead->is_Call()) {
400 remove_useless_late_inlines( &_late_inlines, dead);
401 remove_useless_late_inlines( &_string_late_inlines, dead);
402 remove_useless_late_inlines( &_boxing_late_inlines, dead);
403 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
404
405 if (dead->is_CallStaticJava()) {
406 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
407 }
408 }
409 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
410 bs->unregister_potential_barrier_node(dead);
411 }
412
413 // Disconnect all useless nodes by disconnecting those at the boundary.
414 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
415 uint next = 0;
416 while (next < useful.size()) {
417 Node *n = useful.at(next++);
418 if (n->is_SafePoint()) {
419 // We're done with a parsing phase. Replaced nodes are not valid
420 // beyond that point.
421 n->as_SafePoint()->delete_replaced_nodes();
422 }
423 // Use raw traversal of out edges since this code removes out edges
424 int max = n->outcnt();
425 for (int j = 0; j < max; ++j) {
426 Node* child = n->raw_out(j);
427 if (!useful.member(child)) {
428 assert(!child->is_top() || child != top(),
429 "If top is cached in Compile object it is in useful list");
430 // Only need to remove this out-edge to the useless node
431 n->raw_del_out(j);
432 --j;
433 --max;
434 }
435 }
436 if (n->outcnt() == 1 && n->has_special_unique_user()) {
437 assert(useful.member(n->unique_out()), "do not push a useless node");
438 worklist.push(n->unique_out());
439 }
440 if (n->outcnt() == 0) {
441 worklist.push(n);
442 }
443 }
444
445 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
446 remove_useless_nodes(_parse_predicate_opaqs, useful); // remove useless Parse Predicate opaque nodes
447 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
448 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
449 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
450 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
451 #ifdef ASSERT
452 if (_modified_nodes != nullptr) {
453 _modified_nodes->remove_useless_nodes(useful.member_set());
454 }
455 #endif
456 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
457 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
458 #ifdef ASSERT
459 if (_modified_nodes != nullptr) {
460 _modified_nodes->remove_useless_nodes(useful.member_set());
461 }
462 #endif
463
464 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
465 bs->eliminate_useless_gc_barriers(useful, this);
466 // clean up the late inline lists
467 remove_useless_late_inlines( &_late_inlines, useful);
468 remove_useless_late_inlines( &_string_late_inlines, useful);
469 remove_useless_late_inlines( &_boxing_late_inlines, useful);
470 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
471 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
472 }
473
474 // ============================================================================
475 //------------------------------CompileWrapper---------------------------------
476 class CompileWrapper : public StackObj {
477 Compile *const _compile;
478 public:
479 CompileWrapper(Compile* compile);
480
481 ~CompileWrapper();
482 };
483
484 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
485 // the Compile* pointer is stored in the current ciEnv:
486 ciEnv* env = compile->env();
487 assert(env == ciEnv::current(), "must already be a ciEnv active");
488 assert(env->compiler_data() == nullptr, "compile already active?");
489 env->set_compiler_data(compile);
490 assert(compile == Compile::current(), "sanity");
491
492 compile->set_type_dict(nullptr);
493 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
494 compile->clone_map().set_clone_idx(0);
495 compile->set_type_last_size(0);
496 compile->set_last_tf(nullptr, nullptr);
497 compile->set_indexSet_arena(nullptr);
498 compile->set_indexSet_free_block_list(nullptr);
499 compile->init_type_arena();
500 Type::Initialize(compile);
501 _compile->begin_method();
502 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
503 }
504 CompileWrapper::~CompileWrapper() {
505 // simulate crash during compilation
506 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
507
508 _compile->end_method();
509 _compile->env()->set_compiler_data(nullptr);
510 }
511
512
513 //----------------------------print_compile_messages---------------------------
514 void Compile::print_compile_messages() {
515 #ifndef PRODUCT
516 // Check if recompiling
517 if (!subsume_loads() && PrintOpto) {
518 // Recompiling without allowing machine instructions to subsume loads
519 tty->print_cr("*********************************************************");
520 tty->print_cr("** Bailout: Recompile without subsuming loads **");
521 tty->print_cr("*********************************************************");
522 }
523 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
524 // Recompiling without escape analysis
525 tty->print_cr("*********************************************************");
526 tty->print_cr("** Bailout: Recompile without escape analysis **");
527 tty->print_cr("*********************************************************");
528 }
529 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
530 // Recompiling without iterative escape analysis
531 tty->print_cr("*********************************************************");
532 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
533 tty->print_cr("*********************************************************");
534 }
535 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
536 // Recompiling without reducing allocation merges
537 tty->print_cr("*********************************************************");
538 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
539 tty->print_cr("*********************************************************");
540 }
541 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
542 // Recompiling without boxing elimination
543 tty->print_cr("*********************************************************");
544 tty->print_cr("** Bailout: Recompile without boxing elimination **");
545 tty->print_cr("*********************************************************");
546 }
547 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
548 // Recompiling without locks coarsening
549 tty->print_cr("*********************************************************");
550 tty->print_cr("** Bailout: Recompile without locks coarsening **");
551 tty->print_cr("*********************************************************");
552 }
553 if (env()->break_at_compile()) {
554 // Open the debugger when compiling this method.
555 tty->print("### Breaking when compiling: ");
556 method()->print_short_name();
557 tty->cr();
558 BREAKPOINT;
559 }
560
561 if( PrintOpto ) {
562 if (is_osr_compilation()) {
563 tty->print("[OSR]%3d", _compile_id);
564 } else {
565 tty->print("%3d", _compile_id);
566 }
567 }
568 #endif
569 }
570
571 #ifndef PRODUCT
572 void Compile::print_ideal_ir(const char* phase_name) {
573 // keep the following output all in one block
574 // This output goes directly to the tty, not the compiler log.
575 // To enable tools to match it up with the compilation activity,
576 // be sure to tag this tty output with the compile ID.
577
578 // Node dumping can cause a safepoint, which can break the tty lock.
579 // Buffer all node dumps, so that all safepoints happen before we lock.
580 ResourceMark rm;
581 stringStream ss;
582
583 if (_output == nullptr) {
584 ss.print_cr("AFTER: %s", phase_name);
585 // Print out all nodes in ascending order of index.
586 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
587 } else {
588 // Dump the node blockwise if we have a scheduling
589 _output->print_scheduling(&ss);
590 }
591
592 // Check that the lock is not broken by a safepoint.
593 NoSafepointVerifier nsv;
594 ttyLocker ttyl;
595 if (xtty != nullptr) {
596 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
597 compile_id(),
598 is_osr_compilation() ? " compile_kind='osr'" : "",
599 phase_name);
600 xtty->print("%s", ss.as_string()); // print to tty would use xml escape encoding
601 xtty->tail("ideal");
602 } else {
603 tty->print("%s", ss.as_string());
604 }
605 }
606 #endif
607
608 // ============================================================================
609 //------------------------------Compile standard-------------------------------
610
611 // Compile a method. entry_bci is -1 for normal compilations and indicates
612 // the continuation bci for on stack replacement.
613
614
615 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
616 Options options, DirectiveSet* directive)
617 : Phase(Compiler),
618 _compile_id(ci_env->compile_id()),
619 _options(options),
620 _method(target),
621 _entry_bci(osr_bci),
622 _ilt(nullptr),
623 _stub_function(nullptr),
624 _stub_name(nullptr),
625 _stub_entry_point(nullptr),
626 _max_node_limit(MaxNodeLimit),
627 _post_loop_opts_phase(false),
628 _inlining_progress(false),
629 _inlining_incrementally(false),
630 _do_cleanup(false),
631 _has_reserved_stack_access(target->has_reserved_stack_access()),
632 _has_circular_inline_type(false),
633 #ifndef PRODUCT
634 _igv_idx(0),
635 _trace_opto_output(directive->TraceOptoOutputOption),
636 #endif
637 _has_method_handle_invokes(false),
638 _clinit_barrier_on_entry(false),
639 _stress_seed(0),
640 _comp_arena(mtCompiler),
641 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
642 _env(ci_env),
643 _directive(directive),
644 _log(ci_env->log()),
645 _failure_reason(nullptr),
646 _intrinsics (comp_arena(), 0, 0, nullptr),
647 _macro_nodes (comp_arena(), 8, 0, nullptr),
648 _parse_predicate_opaqs (comp_arena(), 8, 0, nullptr),
649 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
650 _expensive_nodes (comp_arena(), 8, 0, nullptr),
651 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
652 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
653 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
654 _coarsened_locks (comp_arena(), 8, 0, nullptr),
655 _congraph(nullptr),
656 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
657 _unique(0),
658 _dead_node_count(0),
659 _dead_node_list(comp_arena()),
660 _node_arena_one(mtCompiler),
661 _node_arena_two(mtCompiler),
662 _node_arena(&_node_arena_one),
663 _mach_constant_base_node(nullptr),
664 _Compile_types(mtCompiler),
665 _initial_gvn(nullptr),
666 _igvn_worklist(nullptr),
667 _types(nullptr),
668 _node_hash(nullptr),
669 _late_inlines(comp_arena(), 2, 0, nullptr),
670 _string_late_inlines(comp_arena(), 2, 0, nullptr),
671 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
672 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
673 _late_inlines_pos(0),
674 _number_of_mh_late_inlines(0),
675 _print_inlining_stream(new (mtCompiler) stringStream()),
676 _print_inlining_list(nullptr),
677 _print_inlining_idx(0),
678 _print_inlining_output(nullptr),
679 _replay_inline_data(nullptr),
680 _java_calls(0),
681 _inner_loops(0),
682 _interpreter_frame_size(0),
683 _output(nullptr)
684 #ifndef PRODUCT
685 , _in_dump_cnt(0)
686 #endif
687 {
688 C = this;
689 CompileWrapper cw(this);
690
691 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
692 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
693
694 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
695 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
696 // We can always print a disassembly, either abstract (hex dump) or
697 // with the help of a suitable hsdis library. Thus, we should not
698 // couple print_assembly and print_opto_assembly controls.
699 // But: always print opto and regular assembly on compile command 'print'.
700 bool print_assembly = directive->PrintAssemblyOption;
701 set_print_assembly(print_opto_assembly || print_assembly);
702 #else
703 set_print_assembly(false); // must initialize.
704 #endif
705
706 #ifndef PRODUCT
707 set_parsed_irreducible_loop(false);
708 #endif
709
710 if (directive->ReplayInlineOption) {
711 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
712 }
713 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
714 set_print_intrinsics(directive->PrintIntrinsicsOption);
715 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
716
717 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
718 // Make sure the method being compiled gets its own MDO,
719 // so we can at least track the decompile_count().
720 // Need MDO to record RTM code generation state.
721 method()->ensure_method_data();
722 }
723
724 Init(/*do_aliasing=*/ true);
725
726 print_compile_messages();
727
728 _ilt = InlineTree::build_inline_tree_root();
729
730 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
731 assert(num_alias_types() >= AliasIdxRaw, "");
732
733 #define MINIMUM_NODE_HASH 1023
734
735 // GVN that will be run immediately on new nodes
736 uint estimated_size = method()->code_size()*4+64;
737 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
738 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
739 _types = new (comp_arena()) Type_Array(comp_arena());
740 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
741 PhaseGVN gvn;
742 set_initial_gvn(&gvn);
743
744 print_inlining_init();
745 { // Scope for timing the parser
746 TracePhase tp("parse", &timers[_t_parser]);
747
748 // Put top into the hash table ASAP.
749 initial_gvn()->transform_no_reclaim(top());
750
751 // Set up tf(), start(), and find a CallGenerator.
752 CallGenerator* cg = nullptr;
753 if (is_osr_compilation()) {
754 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
755 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
756 initial_gvn()->set_type_bottom(s);
757 init_start(s);
758 cg = CallGenerator::for_osr(method(), entry_bci());
759 } else {
760 // Normal case.
761 init_tf(TypeFunc::make(method()));
762 StartNode* s = new StartNode(root(), tf()->domain_cc());
763 initial_gvn()->set_type_bottom(s);
764 init_start(s);
765 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
766 // With java.lang.ref.reference.get() we must go through the
767 // intrinsic - even when get() is the root
768 // method of the compile - so that, if necessary, the value in
769 // the referent field of the reference object gets recorded by
770 // the pre-barrier code.
771 cg = find_intrinsic(method(), false);
772 }
773 if (cg == nullptr) {
774 float past_uses = method()->interpreter_invocation_count();
775 float expected_uses = past_uses;
776 cg = CallGenerator::for_inline(method(), expected_uses);
777 }
778 }
779 if (failing()) return;
780 if (cg == nullptr) {
781 const char* reason = InlineTree::check_can_parse(method());
782 assert(reason != nullptr, "expect reason for parse failure");
783 stringStream ss;
784 ss.print("cannot parse method: %s", reason);
785 record_method_not_compilable(ss.as_string());
786 return;
787 }
788
789 gvn.set_type(root(), root()->bottom_type());
790
791 JVMState* jvms = build_start_state(start(), tf());
792 if ((jvms = cg->generate(jvms)) == nullptr) {
793 assert(failure_reason() != nullptr, "expect reason for parse failure");
794 stringStream ss;
795 ss.print("method parse failed: %s", failure_reason());
796 record_method_not_compilable(ss.as_string());
797 return;
798 }
799 GraphKit kit(jvms);
800
801 if (!kit.stopped()) {
802 // Accept return values, and transfer control we know not where.
803 // This is done by a special, unique ReturnNode bound to root.
804 return_values(kit.jvms());
805 }
806
807 if (kit.has_exceptions()) {
808 // Any exceptions that escape from this call must be rethrown
809 // to whatever caller is dynamically above us on the stack.
810 // This is done by a special, unique RethrowNode bound to root.
811 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
812 }
813
814 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
815
816 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
817 inline_string_calls(true);
818 }
819
820 if (failing()) return;
821
822 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
823
824 // Remove clutter produced by parsing.
825 if (!failing()) {
826 ResourceMark rm;
827 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
828 }
829 }
830
831 // Note: Large methods are capped off in do_one_bytecode().
832 if (failing()) return;
833
834 // After parsing, node notes are no longer automagic.
835 // They must be propagated by register_new_node_with_optimizer(),
836 // clone(), or the like.
837 set_default_node_notes(nullptr);
838
839 #ifndef PRODUCT
840 if (should_print_igv(1)) {
841 _igv_printer->print_inlining();
842 }
843 #endif
844
845 if (failing()) return;
846 NOT_PRODUCT( verify_graph_edges(); )
847
848 // If any phase is randomized for stress testing, seed random number
849 // generation and log the seed for repeatability.
850 if (StressLCM || StressGCM || StressIGVN || StressCCP) {
851 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
852 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
853 FLAG_SET_ERGO(StressSeed, _stress_seed);
854 } else {
855 _stress_seed = StressSeed;
856 }
857 if (_log != nullptr) {
858 _log->elem("stress_test seed='%u'", _stress_seed);
859 }
860 }
861
862 // Now optimize
863 Optimize();
864 if (failing()) return;
865 NOT_PRODUCT( verify_graph_edges(); )
866
867 #ifndef PRODUCT
868 if (should_print_ideal()) {
869 print_ideal_ir("print_ideal");
870 }
871 #endif
872
873 #ifdef ASSERT
874 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
875 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
876 #endif
877
878 // Dump compilation data to replay it.
879 if (directive->DumpReplayOption) {
880 env()->dump_replay_data(_compile_id);
881 }
882 if (directive->DumpInlineOption && (ilt() != nullptr)) {
883 env()->dump_inline_data(_compile_id);
884 }
885
886 // Now that we know the size of all the monitors we can add a fixed slot
887 // for the original deopt pc.
888 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
889 if (needs_stack_repair()) {
890 // One extra slot for the special stack increment value
891 next_slot += 2;
892 }
893 // TODO 8284443 Only reserve extra slot if needed
894 if (InlineTypeReturnedAsFields) {
895 // One extra slot to hold the IsInit information for a nullable
896 // inline type return if we run out of registers.
897 next_slot += 2;
898 }
899 set_fixed_slots(next_slot);
900
901 // Compute when to use implicit null checks. Used by matching trap based
902 // nodes and NullCheck optimization.
903 set_allowed_deopt_reasons();
904
905 // Now generate code
906 Code_Gen();
907 }
908
909 //------------------------------Compile----------------------------------------
910 // Compile a runtime stub
911 Compile::Compile( ciEnv* ci_env,
912 TypeFunc_generator generator,
913 address stub_function,
914 const char *stub_name,
915 int is_fancy_jump,
916 bool pass_tls,
917 bool return_pc,
918 DirectiveSet* directive)
919 : Phase(Compiler),
920 _compile_id(0),
921 _options(Options::for_runtime_stub()),
922 _method(nullptr),
923 _entry_bci(InvocationEntryBci),
924 _stub_function(stub_function),
925 _stub_name(stub_name),
926 _stub_entry_point(nullptr),
927 _max_node_limit(MaxNodeLimit),
928 _post_loop_opts_phase(false),
929 _inlining_progress(false),
930 _inlining_incrementally(false),
931 _has_reserved_stack_access(false),
932 _has_circular_inline_type(false),
933 #ifndef PRODUCT
934 _igv_idx(0),
935 _trace_opto_output(directive->TraceOptoOutputOption),
936 #endif
937 _has_method_handle_invokes(false),
938 _clinit_barrier_on_entry(false),
939 _stress_seed(0),
940 _comp_arena(mtCompiler),
941 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
942 _env(ci_env),
943 _directive(directive),
944 _log(ci_env->log()),
945 _failure_reason(nullptr),
946 _congraph(nullptr),
947 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
948 _unique(0),
949 _dead_node_count(0),
950 _dead_node_list(comp_arena()),
951 _node_arena_one(mtCompiler),
952 _node_arena_two(mtCompiler),
953 _node_arena(&_node_arena_one),
954 _mach_constant_base_node(nullptr),
955 _Compile_types(mtCompiler),
956 _initial_gvn(nullptr),
957 _igvn_worklist(nullptr),
958 _types(nullptr),
959 _node_hash(nullptr),
960 _number_of_mh_late_inlines(0),
961 _print_inlining_stream(new (mtCompiler) stringStream()),
962 _print_inlining_list(nullptr),
963 _print_inlining_idx(0),
964 _print_inlining_output(nullptr),
965 _replay_inline_data(nullptr),
966 _java_calls(0),
967 _inner_loops(0),
968 _interpreter_frame_size(0),
969 _output(nullptr),
970 #ifndef PRODUCT
971 _in_dump_cnt(0),
972 #endif
973 _allowed_reasons(0) {
974 C = this;
975
976 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
977 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
978
979 #ifndef PRODUCT
980 set_print_assembly(PrintFrameConverterAssembly);
981 set_parsed_irreducible_loop(false);
982 #else
983 set_print_assembly(false); // Must initialize.
984 #endif
985 set_has_irreducible_loop(false); // no loops
986
987 CompileWrapper cw(this);
988 Init(/*do_aliasing=*/ false);
989 init_tf((*generator)());
990
991 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
992 _types = new (comp_arena()) Type_Array(comp_arena());
993 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
994 {
995 PhaseGVN gvn;
996 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
997 gvn.transform_no_reclaim(top());
998
999 GraphKit kit;
1000 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1001 }
1002
1003 NOT_PRODUCT( verify_graph_edges(); )
1004
1005 Code_Gen();
1006 }
1007
1008 //------------------------------Init-------------------------------------------
1009 // Prepare for a single compilation
1010 void Compile::Init(bool aliasing) {
1011 _do_aliasing = aliasing;
1012 _unique = 0;
1013 _regalloc = nullptr;
1014
1015 _tf = nullptr; // filled in later
1016 _top = nullptr; // cached later
1017 _matcher = nullptr; // filled in later
1018 _cfg = nullptr; // filled in later
1019
1020 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1021
1022 _node_note_array = nullptr;
1023 _default_node_notes = nullptr;
1024 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1025
1026 _immutable_memory = nullptr; // filled in at first inquiry
1027
1028 #ifdef ASSERT
1029 _phase_optimize_finished = false;
1030 _exception_backedge = false;
1031 _type_verify = nullptr;
1032 #endif
1033
1034 // Globally visible Nodes
1035 // First set TOP to null to give safe behavior during creation of RootNode
1036 set_cached_top_node(nullptr);
1037 set_root(new RootNode());
1038 // Now that you have a Root to point to, create the real TOP
1039 set_cached_top_node( new ConNode(Type::TOP) );
1040 set_recent_alloc(nullptr, nullptr);
1041
1042 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1043 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1044 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1045 env()->set_dependencies(new Dependencies(env()));
1046
1047 _fixed_slots = 0;
1048 set_has_split_ifs(false);
1049 set_has_loops(false); // first approximation
1050 set_has_stringbuilder(false);
1051 set_has_boxed_value(false);
1052 _trap_can_recompile = false; // no traps emitted yet
1053 _major_progress = true; // start out assuming good things will happen
1054 set_has_unsafe_access(false);
1055 set_max_vector_size(0);
1056 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1057 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1058 set_decompile_count(0);
1059
1060 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1061 _loop_opts_cnt = LoopOptsCount;
1062 _has_flat_accesses = false;
1063 _flat_accesses_share_alias = true;
1064 _scalarize_in_safepoints = false;
1065
1066 set_do_inlining(Inline);
1067 set_max_inline_size(MaxInlineSize);
1068 set_freq_inline_size(FreqInlineSize);
1069 set_do_scheduling(OptoScheduling);
1070
1071 set_do_vector_loop(false);
1072 set_has_monitors(false);
1073
1074 if (AllowVectorizeOnDemand) {
1075 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1076 set_do_vector_loop(true);
1077 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());})
1078 } else if (has_method() && method()->name() != 0 &&
1079 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1080 set_do_vector_loop(true);
1081 }
1082 }
1083 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1084 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());})
1085
1086 set_rtm_state(NoRTM); // No RTM lock eliding by default
1087 _max_node_limit = _directive->MaxNodeLimitOption;
1088
1089 #if INCLUDE_RTM_OPT
1090 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) {
1091 int rtm_state = method()->method_data()->rtm_state();
1092 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1093 // Don't generate RTM lock eliding code.
1094 set_rtm_state(NoRTM);
1095 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1096 // Generate RTM lock eliding code without abort ratio calculation code.
1097 set_rtm_state(UseRTM);
1098 } else if (UseRTMDeopt) {
1099 // Generate RTM lock eliding code and include abort ratio calculation
1100 // code if UseRTMDeopt is on.
1101 set_rtm_state(ProfileRTM);
1102 }
1103 }
1104 #endif
1105 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1106 set_clinit_barrier_on_entry(true);
1107 }
1108 if (debug_info()->recording_non_safepoints()) {
1109 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1110 (comp_arena(), 8, 0, nullptr));
1111 set_default_node_notes(Node_Notes::make(this));
1112 }
1113
1114 const int grow_ats = 16;
1115 _max_alias_types = grow_ats;
1116 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1117 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1118 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1119 {
1120 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1121 }
1122 // Initialize the first few types.
1123 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1124 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1125 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1126 _num_alias_types = AliasIdxRaw+1;
1127 // Zero out the alias type cache.
1128 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1129 // A null adr_type hits in the cache right away. Preload the right answer.
1130 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1131 }
1132
1133 //---------------------------init_start----------------------------------------
1134 // Install the StartNode on this compile object.
1135 void Compile::init_start(StartNode* s) {
1136 if (failing())
1137 return; // already failing
1138 assert(s == start(), "");
1139 }
1140
1141 /**
1142 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1143 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1144 * the ideal graph.
1145 */
1146 StartNode* Compile::start() const {
1147 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1148 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1149 Node* start = root()->fast_out(i);
1150 if (start->is_Start()) {
1151 return start->as_Start();
1152 }
1153 }
1154 fatal("Did not find Start node!");
1155 return nullptr;
1156 }
1157
1158 //-------------------------------immutable_memory-------------------------------------
1159 // Access immutable memory
1160 Node* Compile::immutable_memory() {
1161 if (_immutable_memory != nullptr) {
1162 return _immutable_memory;
1163 }
1164 StartNode* s = start();
1165 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1166 Node *p = s->fast_out(i);
1167 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1168 _immutable_memory = p;
1169 return _immutable_memory;
1170 }
1171 }
1172 ShouldNotReachHere();
1173 return nullptr;
1174 }
1175
1176 //----------------------set_cached_top_node------------------------------------
1177 // Install the cached top node, and make sure Node::is_top works correctly.
1178 void Compile::set_cached_top_node(Node* tn) {
1179 if (tn != nullptr) verify_top(tn);
1180 Node* old_top = _top;
1181 _top = tn;
1182 // Calling Node::setup_is_top allows the nodes the chance to adjust
1183 // their _out arrays.
1184 if (_top != nullptr) _top->setup_is_top();
1185 if (old_top != nullptr) old_top->setup_is_top();
1186 assert(_top == nullptr || top()->is_top(), "");
1187 }
1188
1189 #ifdef ASSERT
1190 uint Compile::count_live_nodes_by_graph_walk() {
1191 Unique_Node_List useful(comp_arena());
1192 // Get useful node list by walking the graph.
1193 identify_useful_nodes(useful);
1194 return useful.size();
1195 }
1196
1197 void Compile::print_missing_nodes() {
1198
1199 // Return if CompileLog is null and PrintIdealNodeCount is false.
1200 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1201 return;
1202 }
1203
1204 // This is an expensive function. It is executed only when the user
1205 // specifies VerifyIdealNodeCount option or otherwise knows the
1206 // additional work that needs to be done to identify reachable nodes
1207 // by walking the flow graph and find the missing ones using
1208 // _dead_node_list.
1209
1210 Unique_Node_List useful(comp_arena());
1211 // Get useful node list by walking the graph.
1212 identify_useful_nodes(useful);
1213
1214 uint l_nodes = C->live_nodes();
1215 uint l_nodes_by_walk = useful.size();
1216
1217 if (l_nodes != l_nodes_by_walk) {
1218 if (_log != nullptr) {
1219 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1220 _log->stamp();
1221 _log->end_head();
1222 }
1223 VectorSet& useful_member_set = useful.member_set();
1224 int last_idx = l_nodes_by_walk;
1225 for (int i = 0; i < last_idx; i++) {
1226 if (useful_member_set.test(i)) {
1227 if (_dead_node_list.test(i)) {
1228 if (_log != nullptr) {
1229 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1230 }
1231 if (PrintIdealNodeCount) {
1232 // Print the log message to tty
1233 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1234 useful.at(i)->dump();
1235 }
1236 }
1237 }
1238 else if (! _dead_node_list.test(i)) {
1239 if (_log != nullptr) {
1240 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1241 }
1242 if (PrintIdealNodeCount) {
1243 // Print the log message to tty
1244 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1245 }
1246 }
1247 }
1248 if (_log != nullptr) {
1249 _log->tail("mismatched_nodes");
1250 }
1251 }
1252 }
1253 void Compile::record_modified_node(Node* n) {
1254 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1255 _modified_nodes->push(n);
1256 }
1257 }
1258
1259 void Compile::remove_modified_node(Node* n) {
1260 if (_modified_nodes != nullptr) {
1261 _modified_nodes->remove(n);
1262 }
1263 }
1264 #endif
1265
1266 #ifndef PRODUCT
1267 void Compile::verify_top(Node* tn) const {
1268 if (tn != nullptr) {
1269 assert(tn->is_Con(), "top node must be a constant");
1270 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1271 assert(tn->in(0) != nullptr, "must have live top node");
1272 }
1273 }
1274 #endif
1275
1276
1277 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1278
1279 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1280 guarantee(arr != nullptr, "");
1281 int num_blocks = arr->length();
1282 if (grow_by < num_blocks) grow_by = num_blocks;
1283 int num_notes = grow_by * _node_notes_block_size;
1284 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1285 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1286 while (num_notes > 0) {
1287 arr->append(notes);
1288 notes += _node_notes_block_size;
1289 num_notes -= _node_notes_block_size;
1290 }
1291 assert(num_notes == 0, "exact multiple, please");
1292 }
1293
1294 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1295 if (source == nullptr || dest == nullptr) return false;
1296
1297 if (dest->is_Con())
1298 return false; // Do not push debug info onto constants.
1299
1300 #ifdef ASSERT
1301 // Leave a bread crumb trail pointing to the original node:
1302 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1303 dest->set_debug_orig(source);
1304 }
1305 #endif
1306
1307 if (node_note_array() == nullptr)
1308 return false; // Not collecting any notes now.
1309
1310 // This is a copy onto a pre-existing node, which may already have notes.
1311 // If both nodes have notes, do not overwrite any pre-existing notes.
1312 Node_Notes* source_notes = node_notes_at(source->_idx);
1313 if (source_notes == nullptr || source_notes->is_clear()) return false;
1314 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1315 if (dest_notes == nullptr || dest_notes->is_clear()) {
1316 return set_node_notes_at(dest->_idx, source_notes);
1317 }
1318
1319 Node_Notes merged_notes = (*source_notes);
1320 // The order of operations here ensures that dest notes will win...
1321 merged_notes.update_from(dest_notes);
1322 return set_node_notes_at(dest->_idx, &merged_notes);
1323 }
1324
1325
1326 //--------------------------allow_range_check_smearing-------------------------
1327 // Gating condition for coalescing similar range checks.
1328 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1329 // single covering check that is at least as strong as any of them.
1330 // If the optimization succeeds, the simplified (strengthened) range check
1331 // will always succeed. If it fails, we will deopt, and then give up
1332 // on the optimization.
1333 bool Compile::allow_range_check_smearing() const {
1334 // If this method has already thrown a range-check,
1335 // assume it was because we already tried range smearing
1336 // and it failed.
1337 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1338 return !already_trapped;
1339 }
1340
1341
1342 //------------------------------flatten_alias_type-----------------------------
1343 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1344 assert(do_aliasing(), "Aliasing should be enabled");
1345 int offset = tj->offset();
1346 TypePtr::PTR ptr = tj->ptr();
1347
1348 // Known instance (scalarizable allocation) alias only with itself.
1349 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1350 tj->is_oopptr()->is_known_instance();
1351
1352 // Process weird unsafe references.
1353 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1354 bool default_value_load = EnableValhalla && tj->is_instptr()->instance_klass() == ciEnv::current()->Class_klass();
1355 assert(InlineUnsafeOps || StressReflectiveCode || default_value_load, "indeterminate pointers come only from unsafe ops");
1356 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1357 tj = TypeOopPtr::BOTTOM;
1358 ptr = tj->ptr();
1359 offset = tj->offset();
1360 }
1361
1362 // Array pointers need some flattening
1363 const TypeAryPtr* ta = tj->isa_aryptr();
1364 if (ta && ta->is_stable()) {
1365 // Erase stability property for alias analysis.
1366 tj = ta = ta->cast_to_stable(false);
1367 }
1368 if (ta && ta->is_not_flat()) {
1369 // Erase not flat property for alias analysis.
1370 tj = ta = ta->cast_to_not_flat(false);
1371 }
1372 if (ta && ta->is_not_null_free()) {
1373 // Erase not null free property for alias analysis.
1374 tj = ta = ta->cast_to_not_null_free(false);
1375 }
1376
1377 if( ta && is_known_inst ) {
1378 if ( offset != Type::OffsetBot &&
1379 offset > arrayOopDesc::length_offset_in_bytes() ) {
1380 offset = Type::OffsetBot; // Flatten constant access into array body only
1381 tj = ta = ta->
1382 remove_speculative()->
1383 cast_to_ptr_type(ptr)->
1384 with_offset(offset);
1385 }
1386 } else if (ta) {
1387 // For arrays indexed by constant indices, we flatten the alias
1388 // space to include all of the array body. Only the header, klass
1389 // and array length can be accessed un-aliased.
1390 // For flat inline type array, each field has its own slice so
1391 // we must include the field offset.
1392 if( offset != Type::OffsetBot ) {
1393 if( ta->const_oop() ) { // MethodData* or Method*
1394 offset = Type::OffsetBot; // Flatten constant access into array body
1395 tj = ta = ta->
1396 remove_speculative()->
1397 cast_to_ptr_type(ptr)->
1398 cast_to_exactness(false)->
1399 with_offset(offset);
1400 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1401 // range is OK as-is.
1402 tj = ta = TypeAryPtr::RANGE;
1403 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1404 tj = TypeInstPtr::KLASS; // all klass loads look alike
1405 ta = TypeAryPtr::RANGE; // generic ignored junk
1406 ptr = TypePtr::BotPTR;
1407 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1408 tj = TypeInstPtr::MARK;
1409 ta = TypeAryPtr::RANGE; // generic ignored junk
1410 ptr = TypePtr::BotPTR;
1411 } else { // Random constant offset into array body
1412 offset = Type::OffsetBot; // Flatten constant access into array body
1413 tj = ta = ta->
1414 remove_speculative()->
1415 cast_to_ptr_type(ptr)->
1416 cast_to_exactness(false)->
1417 with_offset(offset);
1418 }
1419 }
1420 // Arrays of fixed size alias with arrays of unknown size.
1421 if (ta->size() != TypeInt::POS) {
1422 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1423 tj = ta = ta->
1424 remove_speculative()->
1425 cast_to_ptr_type(ptr)->
1426 with_ary(tary)->
1427 cast_to_exactness(false);
1428 }
1429 // Arrays of known objects become arrays of unknown objects.
1430 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1431 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1432 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1433 }
1434 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1435 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1436 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1437 }
1438 // Initially all flattened array accesses share a single slice
1439 if (ta->is_flat() && ta->elem() != TypeInstPtr::BOTTOM && _flat_accesses_share_alias) {
1440 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size(), /* stable= */ false, /* flat= */ true);
1441 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1442 }
1443 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1444 // cannot be distinguished by bytecode alone.
1445 if (ta->elem() == TypeInt::BOOL) {
1446 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1447 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1448 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1449 }
1450 // During the 2nd round of IterGVN, NotNull castings are removed.
1451 // Make sure the Bottom and NotNull variants alias the same.
1452 // Also, make sure exact and non-exact variants alias the same.
1453 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1454 tj = ta = ta->
1455 remove_speculative()->
1456 cast_to_ptr_type(TypePtr::BotPTR)->
1457 cast_to_exactness(false)->
1458 with_offset(offset);
1459 }
1460 }
1461
1462 // Oop pointers need some flattening
1463 const TypeInstPtr *to = tj->isa_instptr();
1464 if (to && to != TypeOopPtr::BOTTOM) {
1465 ciInstanceKlass* ik = to->instance_klass();
1466 if( ptr == TypePtr::Constant ) {
1467 if (ik != ciEnv::current()->Class_klass() ||
1468 offset < ik->layout_helper_size_in_bytes()) {
1469 // No constant oop pointers (such as Strings); they alias with
1470 // unknown strings.
1471 assert(!is_known_inst, "not scalarizable allocation");
1472 tj = to = to->
1473 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1474 remove_speculative()->
1475 cast_to_ptr_type(TypePtr::BotPTR)->
1476 cast_to_exactness(false);
1477 }
1478 } else if( is_known_inst ) {
1479 tj = to; // Keep NotNull and klass_is_exact for instance type
1480 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1481 // During the 2nd round of IterGVN, NotNull castings are removed.
1482 // Make sure the Bottom and NotNull variants alias the same.
1483 // Also, make sure exact and non-exact variants alias the same.
1484 tj = to = to->
1485 remove_speculative()->
1486 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1487 cast_to_ptr_type(TypePtr::BotPTR)->
1488 cast_to_exactness(false);
1489 }
1490 if (to->speculative() != nullptr) {
1491 tj = to = to->remove_speculative();
1492 }
1493 // Canonicalize the holder of this field
1494 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1495 // First handle header references such as a LoadKlassNode, even if the
1496 // object's klass is unloaded at compile time (4965979).
1497 if (!is_known_inst) { // Do it only for non-instance types
1498 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1499 }
1500 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1501 // Static fields are in the space above the normal instance
1502 // fields in the java.lang.Class instance.
1503 if (ik != ciEnv::current()->Class_klass()) {
1504 to = nullptr;
1505 tj = TypeOopPtr::BOTTOM;
1506 offset = tj->offset();
1507 }
1508 } else {
1509 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1510 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1511 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1512 if( is_known_inst ) {
1513 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, Type::Offset(offset), to->instance_id());
1514 } else {
1515 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, Type::Offset(offset));
1516 }
1517 }
1518 }
1519 }
1520
1521 // Klass pointers to object array klasses need some flattening
1522 const TypeKlassPtr *tk = tj->isa_klassptr();
1523 if( tk ) {
1524 // If we are referencing a field within a Klass, we need
1525 // to assume the worst case of an Object. Both exact and
1526 // inexact types must flatten to the same alias class so
1527 // use NotNull as the PTR.
1528 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1529 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1530 env()->Object_klass(),
1531 Type::Offset(offset));
1532 }
1533
1534 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1535 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1536 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1537 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset));
1538 } else {
1539 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, Type::Offset(offset), tk->is_not_flat(), tk->is_not_null_free(), tk->is_null_free());
1540 }
1541 }
1542 // Check for precise loads from the primary supertype array and force them
1543 // to the supertype cache alias index. Check for generic array loads from
1544 // the primary supertype array and also force them to the supertype cache
1545 // alias index. Since the same load can reach both, we need to merge
1546 // these 2 disparate memories into the same alias class. Since the
1547 // primary supertype array is read-only, there's no chance of confusion
1548 // where we bypass an array load and an array store.
1549 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1550 if (offset == Type::OffsetBot ||
1551 (offset >= primary_supers_offset &&
1552 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1553 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1554 offset = in_bytes(Klass::secondary_super_cache_offset());
1555 tj = tk = tk->with_offset(offset);
1556 }
1557 }
1558
1559 // Flatten all Raw pointers together.
1560 if (tj->base() == Type::RawPtr)
1561 tj = TypeRawPtr::BOTTOM;
1562
1563 if (tj->base() == Type::AnyPtr)
1564 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1565
1566 offset = tj->offset();
1567 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1568
1569 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1570 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1571 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1572 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1573 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1574 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1575 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1576 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1577 assert( tj->ptr() != TypePtr::TopPTR &&
1578 tj->ptr() != TypePtr::AnyNull &&
1579 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1580 // assert( tj->ptr() != TypePtr::Constant ||
1581 // tj->base() == Type::RawPtr ||
1582 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1583
1584 return tj;
1585 }
1586
1587 void Compile::AliasType::Init(int i, const TypePtr* at) {
1588 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1589 _index = i;
1590 _adr_type = at;
1591 _field = nullptr;
1592 _element = nullptr;
1593 _is_rewritable = true; // default
1594 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1595 if (atoop != nullptr && atoop->is_known_instance()) {
1596 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1597 _general_index = Compile::current()->get_alias_index(gt);
1598 } else {
1599 _general_index = 0;
1600 }
1601 }
1602
1603 BasicType Compile::AliasType::basic_type() const {
1604 if (element() != nullptr) {
1605 const Type* element = adr_type()->is_aryptr()->elem();
1606 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1607 } if (field() != nullptr) {
1608 return field()->layout_type();
1609 } else {
1610 return T_ILLEGAL; // unknown
1611 }
1612 }
1613
1614 //---------------------------------print_on------------------------------------
1615 #ifndef PRODUCT
1616 void Compile::AliasType::print_on(outputStream* st) {
1617 if (index() < 10)
1618 st->print("@ <%d> ", index());
1619 else st->print("@ <%d>", index());
1620 st->print(is_rewritable() ? " " : " RO");
1621 int offset = adr_type()->offset();
1622 if (offset == Type::OffsetBot)
1623 st->print(" +any");
1624 else st->print(" +%-3d", offset);
1625 st->print(" in ");
1626 adr_type()->dump_on(st);
1627 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1628 if (field() != nullptr && tjp) {
1629 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1630 tjp->offset() != field()->offset_in_bytes()) {
1631 st->print(" != ");
1632 field()->print();
1633 st->print(" ***");
1634 }
1635 }
1636 }
1637
1638 void print_alias_types() {
1639 Compile* C = Compile::current();
1640 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1641 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1642 C->alias_type(idx)->print_on(tty);
1643 tty->cr();
1644 }
1645 }
1646 #endif
1647
1648
1649 //----------------------------probe_alias_cache--------------------------------
1650 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1651 intptr_t key = (intptr_t) adr_type;
1652 key ^= key >> logAliasCacheSize;
1653 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1654 }
1655
1656
1657 //-----------------------------grow_alias_types--------------------------------
1658 void Compile::grow_alias_types() {
1659 const int old_ats = _max_alias_types; // how many before?
1660 const int new_ats = old_ats; // how many more?
1661 const int grow_ats = old_ats+new_ats; // how many now?
1662 _max_alias_types = grow_ats;
1663 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1664 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1665 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1666 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1667 }
1668
1669
1670 //--------------------------------find_alias_type------------------------------
1671 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1672 if (!do_aliasing()) {
1673 return alias_type(AliasIdxBot);
1674 }
1675
1676 AliasCacheEntry* ace = nullptr;
1677 if (!uncached) {
1678 ace = probe_alias_cache(adr_type);
1679 if (ace->_adr_type == adr_type) {
1680 return alias_type(ace->_index);
1681 }
1682 }
1683
1684 // Handle special cases.
1685 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1686 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1687
1688 // Do it the slow way.
1689 const TypePtr* flat = flatten_alias_type(adr_type);
1690
1691 #ifdef ASSERT
1692 {
1693 ResourceMark rm;
1694 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1695 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1696 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1697 Type::str(adr_type));
1698 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1699 const TypeOopPtr* foop = flat->is_oopptr();
1700 // Scalarizable allocations have exact klass always.
1701 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1702 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1703 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1704 Type::str(foop), Type::str(xoop));
1705 }
1706 }
1707 #endif
1708
1709 int idx = AliasIdxTop;
1710 for (int i = 0; i < num_alias_types(); i++) {
1711 if (alias_type(i)->adr_type() == flat) {
1712 idx = i;
1713 break;
1714 }
1715 }
1716
1717 if (idx == AliasIdxTop) {
1718 if (no_create) return nullptr;
1719 // Grow the array if necessary.
1720 if (_num_alias_types == _max_alias_types) grow_alias_types();
1721 // Add a new alias type.
1722 idx = _num_alias_types++;
1723 _alias_types[idx]->Init(idx, flat);
1724 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1725 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1726 if (flat->isa_instptr()) {
1727 if (flat->offset() == java_lang_Class::klass_offset()
1728 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1729 alias_type(idx)->set_rewritable(false);
1730 }
1731 ciField* field = nullptr;
1732 if (flat->isa_aryptr()) {
1733 #ifdef ASSERT
1734 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1735 // (T_BYTE has the weakest alignment and size restrictions...)
1736 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1737 #endif
1738 const Type* elemtype = flat->is_aryptr()->elem();
1739 if (flat->offset() == TypePtr::OffsetBot) {
1740 alias_type(idx)->set_element(elemtype);
1741 }
1742 int field_offset = flat->is_aryptr()->field_offset().get();
1743 if (flat->is_flat() &&
1744 field_offset != Type::OffsetBot) {
1745 ciInlineKlass* vk = elemtype->inline_klass();
1746 field_offset += vk->first_field_offset();
1747 field = vk->get_field_by_offset(field_offset, false);
1748 }
1749 }
1750 if (flat->isa_klassptr()) {
1751 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1752 alias_type(idx)->set_rewritable(false);
1753 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1754 alias_type(idx)->set_rewritable(false);
1755 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1756 alias_type(idx)->set_rewritable(false);
1757 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1758 alias_type(idx)->set_rewritable(false);
1759 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1760 alias_type(idx)->set_rewritable(false);
1761 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1762 alias_type(idx)->set_rewritable(false);
1763 }
1764 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1765 // but the base pointer type is not distinctive enough to identify
1766 // references into JavaThread.)
1767
1768 // Check for final fields.
1769 const TypeInstPtr* tinst = flat->isa_instptr();
1770 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1771 if (tinst->const_oop() != nullptr &&
1772 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1773 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1774 // static field
1775 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1776 field = k->get_field_by_offset(tinst->offset(), true);
1777 } else if (tinst->is_inlinetypeptr()) {
1778 // Inline type field
1779 ciInlineKlass* vk = tinst->inline_klass();
1780 field = vk->get_field_by_offset(tinst->offset(), false);
1781 } else {
1782 ciInstanceKlass *k = tinst->instance_klass();
1783 field = k->get_field_by_offset(tinst->offset(), false);
1784 }
1785 }
1786 assert(field == nullptr ||
1787 original_field == nullptr ||
1788 (field->holder() == original_field->holder() &&
1789 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1790 field->is_static() == original_field->is_static()), "wrong field?");
1791 // Set field() and is_rewritable() attributes.
1792 if (field != nullptr) {
1793 alias_type(idx)->set_field(field);
1794 if (flat->isa_aryptr()) {
1795 // Fields of flat arrays are rewritable although they are declared final
1796 assert(flat->is_flat(), "must be a flat array");
1797 alias_type(idx)->set_rewritable(true);
1798 }
1799 }
1800 }
1801
1802 // Fill the cache for next time.
1803 if (!uncached) {
1804 ace->_adr_type = adr_type;
1805 ace->_index = idx;
1806 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1807
1808 // Might as well try to fill the cache for the flattened version, too.
1809 AliasCacheEntry* face = probe_alias_cache(flat);
1810 if (face->_adr_type == nullptr) {
1811 face->_adr_type = flat;
1812 face->_index = idx;
1813 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1814 }
1815 }
1816
1817 return alias_type(idx);
1818 }
1819
1820
1821 Compile::AliasType* Compile::alias_type(ciField* field) {
1822 const TypeOopPtr* t;
1823 if (field->is_static())
1824 t = TypeInstPtr::make(field->holder()->java_mirror());
1825 else
1826 t = TypeOopPtr::make_from_klass_raw(field->holder());
1827 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1828 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1829 return atp;
1830 }
1831
1832
1833 //------------------------------have_alias_type--------------------------------
1834 bool Compile::have_alias_type(const TypePtr* adr_type) {
1835 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1836 if (ace->_adr_type == adr_type) {
1837 return true;
1838 }
1839
1840 // Handle special cases.
1841 if (adr_type == nullptr) return true;
1842 if (adr_type == TypePtr::BOTTOM) return true;
1843
1844 return find_alias_type(adr_type, true, nullptr) != nullptr;
1845 }
1846
1847 //-----------------------------must_alias--------------------------------------
1848 // True if all values of the given address type are in the given alias category.
1849 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1850 if (alias_idx == AliasIdxBot) return true; // the universal category
1851 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1852 if (alias_idx == AliasIdxTop) return false; // the empty category
1853 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1854
1855 // the only remaining possible overlap is identity
1856 int adr_idx = get_alias_index(adr_type);
1857 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1858 assert(adr_idx == alias_idx ||
1859 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1860 && adr_type != TypeOopPtr::BOTTOM),
1861 "should not be testing for overlap with an unsafe pointer");
1862 return adr_idx == alias_idx;
1863 }
1864
1865 //------------------------------can_alias--------------------------------------
1866 // True if any values of the given address type are in the given alias category.
1867 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1868 if (alias_idx == AliasIdxTop) return false; // the empty category
1869 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1870 // Known instance doesn't alias with bottom memory
1871 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1872 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1873
1874 // the only remaining possible overlap is identity
1875 int adr_idx = get_alias_index(adr_type);
1876 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1877 return adr_idx == alias_idx;
1878 }
1879
1880 // Remove the opaque nodes that protect the Parse Predicates so that all unused
1881 // checks and uncommon_traps will be eliminated from the ideal graph.
1882 void Compile::cleanup_parse_predicates(PhaseIterGVN& igvn) const {
1883 if (parse_predicate_count() == 0) {
1884 return;
1885 }
1886 for (int i = parse_predicate_count(); i > 0; i--) {
1887 Node* n = parse_predicate_opaque1_node(i - 1);
1888 assert(n->Opcode() == Op_Opaque1, "must be");
1889 igvn.replace_node(n, n->in(1));
1890 }
1891 assert(parse_predicate_count() == 0, "should be clean!");
1892 }
1893
1894 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1895 if (!n->for_post_loop_opts_igvn()) {
1896 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1897 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1898 _for_post_loop_igvn.append(n);
1899 }
1900 }
1901
1902 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1903 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1904 _for_post_loop_igvn.remove(n);
1905 }
1906
1907 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1908 // Verify that all previous optimizations produced a valid graph
1909 // at least to this point, even if no loop optimizations were done.
1910 PhaseIdealLoop::verify(igvn);
1911
1912 C->set_post_loop_opts_phase(); // no more loop opts allowed
1913
1914 assert(!C->major_progress(), "not cleared");
1915
1916 if (_for_post_loop_igvn.length() > 0) {
1917 while (_for_post_loop_igvn.length() > 0) {
1918 Node* n = _for_post_loop_igvn.pop();
1919 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1920 igvn._worklist.push(n);
1921 }
1922 igvn.optimize();
1923 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1924
1925 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1926 if (C->major_progress()) {
1927 C->clear_major_progress(); // ensure that major progress is now clear
1928 }
1929 }
1930 }
1931
1932 void Compile::add_inline_type(Node* n) {
1933 assert(n->is_InlineType(), "unexpected node");
1934 _inline_type_nodes.push(n);
1935 }
1936
1937 void Compile::remove_inline_type(Node* n) {
1938 assert(n->is_InlineType(), "unexpected node");
1939 if (_inline_type_nodes.contains(n)) {
1940 _inline_type_nodes.remove(n);
1941 }
1942 }
1943
1944 // Does the return value keep otherwise useless inline type allocations alive?
1945 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1946 ResourceMark rm;
1947 Unique_Node_List wq;
1948 wq.push(ret_val);
1949 bool some_allocations = false;
1950 for (uint i = 0; i < wq.size(); i++) {
1951 Node* n = wq.at(i);
1952 if (n->outcnt() > 1) {
1953 // Some other use for the allocation
1954 return false;
1955 } else if (n->is_InlineType()) {
1956 wq.push(n->in(1));
1957 } else if (n->is_Phi()) {
1958 for (uint j = 1; j < n->req(); j++) {
1959 wq.push(n->in(j));
1960 }
1961 } else if (n->is_CheckCastPP() &&
1962 n->in(1)->is_Proj() &&
1963 n->in(1)->in(0)->is_Allocate()) {
1964 some_allocations = true;
1965 } else if (n->is_CheckCastPP()) {
1966 wq.push(n->in(1));
1967 }
1968 }
1969 return some_allocations;
1970 }
1971
1972 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1973 // Make sure that the return value does not keep an otherwise unused allocation alive
1974 if (tf()->returns_inline_type_as_fields()) {
1975 Node* ret = nullptr;
1976 for (uint i = 1; i < root()->req(); i++) {
1977 Node* in = root()->in(i);
1978 if (in->Opcode() == Op_Return) {
1979 assert(ret == nullptr, "only one return");
1980 ret = in;
1981 }
1982 }
1983 if (ret != nullptr) {
1984 Node* ret_val = ret->in(TypeFunc::Parms);
1985 if (igvn.type(ret_val)->isa_oopptr() &&
1986 return_val_keeps_allocations_alive(ret_val)) {
1987 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
1988 assert(ret_val->outcnt() == 0, "should be dead now");
1989 igvn.remove_dead_node(ret_val);
1990 }
1991 }
1992 }
1993 if (_inline_type_nodes.length() == 0) {
1994 return;
1995 }
1996 // Scalarize inline types in safepoint debug info.
1997 // Delay this until all inlining is over to avoid getting inconsistent debug info.
1998 set_scalarize_in_safepoints(true);
1999 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2000 _inline_type_nodes.at(i)->as_InlineType()->make_scalar_in_safepoints(&igvn);
2001 }
2002 if (remove) {
2003 // Remove inline type nodes by replacing them with their oop input
2004 while (_inline_type_nodes.length() > 0) {
2005 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2006 if (vt->outcnt() == 0) {
2007 igvn.remove_dead_node(vt);
2008 continue;
2009 }
2010 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2011 DEBUG_ONLY(bool must_be_buffered = false);
2012 Node* u = vt->out(i);
2013 // Check if any users are blackholes. If so, rewrite them to use either the
2014 // allocated buffer, or individual components, instead of the inline type node
2015 // that goes away.
2016 if (u->is_Blackhole()) {
2017 BlackholeNode* bh = u->as_Blackhole();
2018
2019 // Unlink the old input
2020 int idx = bh->find_edge(vt);
2021 assert(idx != -1, "The edge should be there");
2022 bh->del_req(idx);
2023 --i;
2024
2025 if (vt->is_allocated(&igvn)) {
2026 // Already has the allocated instance, blackhole that
2027 bh->add_req(vt->get_oop());
2028 } else {
2029 // Not allocated yet, blackhole the components
2030 for (uint c = 0; c < vt->field_count(); c++) {
2031 bh->add_req(vt->field_value(c));
2032 }
2033 }
2034
2035 // Node modified, record for IGVN
2036 igvn.record_for_igvn(bh);
2037 }
2038 #ifdef ASSERT
2039 // Verify that inline type is buffered when replacing by oop
2040 else if (u->is_InlineType()) {
2041 // InlineType uses don't need buffering because they are about to be replaced as well
2042 } else if (u->is_Phi()) {
2043 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2044 } else {
2045 must_be_buffered = true;
2046 }
2047 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2048 vt->dump(0);
2049 u->dump(0);
2050 assert(false, "Should have been buffered");
2051 }
2052 #endif
2053 }
2054 igvn.replace_node(vt, vt->get_oop());
2055 }
2056 }
2057 igvn.optimize();
2058 }
2059
2060 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2061 if (!_has_flat_accesses) {
2062 return;
2063 }
2064 // Initially, all flat array accesses share the same slice to
2065 // keep dependencies with Object[] array accesses (that could be
2066 // to a flat array) correct. We're done with parsing so we
2067 // now know all flat array accesses in this compile
2068 // unit. Let's move flat array accesses to their own slice,
2069 // one per element field. This should help memory access
2070 // optimizations.
2071 ResourceMark rm;
2072 Unique_Node_List wq;
2073 wq.push(root());
2074
2075 Node_List mergememnodes;
2076 Node_List memnodes;
2077
2078 // Alias index currently shared by all flat memory accesses
2079 int index = get_alias_index(TypeAryPtr::INLINES);
2080
2081 // Find MergeMem nodes and flat array accesses
2082 for (uint i = 0; i < wq.size(); i++) {
2083 Node* n = wq.at(i);
2084 if (n->is_Mem()) {
2085 const TypePtr* adr_type = nullptr;
2086 if (n->Opcode() == Op_StoreCM) {
2087 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type()));
2088 } else {
2089 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2090 }
2091 if (adr_type == TypeAryPtr::INLINES) {
2092 memnodes.push(n);
2093 }
2094 } else if (n->is_MergeMem()) {
2095 MergeMemNode* mm = n->as_MergeMem();
2096 if (mm->memory_at(index) != mm->base_memory()) {
2097 mergememnodes.push(n);
2098 }
2099 }
2100 for (uint j = 0; j < n->req(); j++) {
2101 Node* m = n->in(j);
2102 if (m != nullptr) {
2103 wq.push(m);
2104 }
2105 }
2106 }
2107
2108 if (memnodes.size() > 0) {
2109 _flat_accesses_share_alias = false;
2110
2111 // We are going to change the slice for the flat array
2112 // accesses so we need to clear the cache entries that refer to
2113 // them.
2114 for (uint i = 0; i < AliasCacheSize; i++) {
2115 AliasCacheEntry* ace = &_alias_cache[i];
2116 if (ace->_adr_type != nullptr &&
2117 ace->_adr_type->is_flat()) {
2118 ace->_adr_type = nullptr;
2119 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2120 }
2121 }
2122
2123 // Find what aliases we are going to add
2124 int start_alias = num_alias_types()-1;
2125 int stop_alias = 0;
2126
2127 for (uint i = 0; i < memnodes.size(); i++) {
2128 Node* m = memnodes.at(i);
2129 const TypePtr* adr_type = nullptr;
2130 if (m->Opcode() == Op_StoreCM) {
2131 adr_type = m->in(MemNode::OopStore)->adr_type();
2132 if (adr_type != TypeAryPtr::INLINES) {
2133 // store was optimized out and we lost track of the adr_type
2134 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address),
2135 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore),
2136 get_alias_index(adr_type));
2137 igvn.register_new_node_with_optimizer(clone);
2138 igvn.replace_node(m, clone);
2139 }
2140 } else {
2141 adr_type = m->adr_type();
2142 #ifdef ASSERT
2143 m->as_Mem()->set_adr_type(adr_type);
2144 #endif
2145 }
2146 int idx = get_alias_index(adr_type);
2147 start_alias = MIN2(start_alias, idx);
2148 stop_alias = MAX2(stop_alias, idx);
2149 }
2150
2151 assert(stop_alias >= start_alias, "should have expanded aliases");
2152
2153 Node_Stack stack(0);
2154 #ifdef ASSERT
2155 VectorSet seen(Thread::current()->resource_area());
2156 #endif
2157 // Now let's fix the memory graph so each flat array access
2158 // is moved to the right slice. Start from the MergeMem nodes.
2159 uint last = unique();
2160 for (uint i = 0; i < mergememnodes.size(); i++) {
2161 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2162 Node* n = current->memory_at(index);
2163 MergeMemNode* mm = nullptr;
2164 do {
2165 // Follow memory edges through memory accesses, phis and
2166 // narrow membars and push nodes on the stack. Once we hit
2167 // bottom memory, we pop element off the stack one at a
2168 // time, in reverse order, and move them to the right slice
2169 // by changing their memory edges.
2170 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2171 assert(!seen.test_set(n->_idx), "");
2172 // Uses (a load for instance) will need to be moved to the
2173 // right slice as well and will get a new memory state
2174 // that we don't know yet. The use could also be the
2175 // backedge of a loop. We put a place holder node between
2176 // the memory node and its uses. We replace that place
2177 // holder with the correct memory state once we know it,
2178 // i.e. when nodes are popped off the stack. Using the
2179 // place holder make the logic work in the presence of
2180 // loops.
2181 if (n->outcnt() > 1) {
2182 Node* place_holder = nullptr;
2183 assert(!n->has_out_with(Op_Node), "");
2184 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2185 Node* u = n->out(k);
2186 if (u != current && u->_idx < last) {
2187 bool success = false;
2188 for (uint l = 0; l < u->req(); l++) {
2189 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2190 continue;
2191 }
2192 Node* in = u->in(l);
2193 if (in == n) {
2194 if (place_holder == nullptr) {
2195 place_holder = new Node(1);
2196 place_holder->init_req(0, n);
2197 }
2198 igvn.replace_input_of(u, l, place_holder);
2199 success = true;
2200 }
2201 }
2202 if (success) {
2203 --k;
2204 }
2205 }
2206 }
2207 }
2208 if (n->is_Phi()) {
2209 stack.push(n, 1);
2210 n = n->in(1);
2211 } else if (n->is_Mem()) {
2212 stack.push(n, n->req());
2213 n = n->in(MemNode::Memory);
2214 } else {
2215 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2216 stack.push(n, n->req());
2217 n = n->in(0)->in(TypeFunc::Memory);
2218 }
2219 } else {
2220 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2221 // Build a new MergeMem node to carry the new memory state
2222 // as we build it. IGVN should fold extraneous MergeMem
2223 // nodes.
2224 mm = MergeMemNode::make(n);
2225 igvn.register_new_node_with_optimizer(mm);
2226 while (stack.size() > 0) {
2227 Node* m = stack.node();
2228 uint idx = stack.index();
2229 if (m->is_Mem()) {
2230 // Move memory node to its new slice
2231 const TypePtr* adr_type = m->adr_type();
2232 int alias = get_alias_index(adr_type);
2233 Node* prev = mm->memory_at(alias);
2234 igvn.replace_input_of(m, MemNode::Memory, prev);
2235 mm->set_memory_at(alias, m);
2236 } else if (m->is_Phi()) {
2237 // We need as many new phis as there are new aliases
2238 igvn.replace_input_of(m, idx, mm);
2239 if (idx == m->req()-1) {
2240 Node* r = m->in(0);
2241 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2242 const TypePtr* adr_type = get_adr_type(j);
2243 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2244 continue;
2245 }
2246 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2247 igvn.register_new_node_with_optimizer(phi);
2248 for (uint k = 1; k < m->req(); k++) {
2249 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2250 }
2251 mm->set_memory_at(j, phi);
2252 }
2253 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2254 igvn.register_new_node_with_optimizer(base_phi);
2255 for (uint k = 1; k < m->req(); k++) {
2256 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2257 }
2258 mm->set_base_memory(base_phi);
2259 }
2260 } else {
2261 // This is a MemBarCPUOrder node from
2262 // Parse::array_load()/Parse::array_store(), in the
2263 // branch that handles flat arrays hidden under
2264 // an Object[] array. We also need one new membar per
2265 // new alias to keep the unknown access that the
2266 // membars protect properly ordered with accesses to
2267 // known flat array.
2268 assert(m->is_Proj(), "projection expected");
2269 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2270 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2271 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2272 const TypePtr* adr_type = get_adr_type(j);
2273 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2274 continue;
2275 }
2276 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2277 igvn.register_new_node_with_optimizer(mb);
2278 Node* mem = mm->memory_at(j);
2279 mb->init_req(TypeFunc::Control, ctrl);
2280 mb->init_req(TypeFunc::Memory, mem);
2281 ctrl = new ProjNode(mb, TypeFunc::Control);
2282 igvn.register_new_node_with_optimizer(ctrl);
2283 mem = new ProjNode(mb, TypeFunc::Memory);
2284 igvn.register_new_node_with_optimizer(mem);
2285 mm->set_memory_at(j, mem);
2286 }
2287 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2288 }
2289 if (idx < m->req()-1) {
2290 idx += 1;
2291 stack.set_index(idx);
2292 n = m->in(idx);
2293 break;
2294 }
2295 // Take care of place holder nodes
2296 if (m->has_out_with(Op_Node)) {
2297 Node* place_holder = m->find_out_with(Op_Node);
2298 if (place_holder != nullptr) {
2299 Node* mm_clone = mm->clone();
2300 igvn.register_new_node_with_optimizer(mm_clone);
2301 Node* hook = new Node(1);
2302 hook->init_req(0, mm);
2303 igvn.replace_node(place_holder, mm_clone);
2304 hook->destruct(&igvn);
2305 }
2306 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2307 }
2308 stack.pop();
2309 }
2310 }
2311 } while(stack.size() > 0);
2312 // Fix the memory state at the MergeMem we started from
2313 igvn.rehash_node_delayed(current);
2314 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2315 const TypePtr* adr_type = get_adr_type(j);
2316 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2317 continue;
2318 }
2319 current->set_memory_at(j, mm);
2320 }
2321 current->set_memory_at(index, current->base_memory());
2322 }
2323 igvn.optimize();
2324 }
2325 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2326 #ifdef ASSERT
2327 if (!_flat_accesses_share_alias) {
2328 wq.clear();
2329 wq.push(root());
2330 for (uint i = 0; i < wq.size(); i++) {
2331 Node* n = wq.at(i);
2332 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2333 for (uint j = 0; j < n->req(); j++) {
2334 Node* m = n->in(j);
2335 if (m != nullptr) {
2336 wq.push(m);
2337 }
2338 }
2339 }
2340 }
2341 #endif
2342 }
2343
2344 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2345 if (OptimizeUnstableIf) {
2346 _unstable_if_traps.append(trap);
2347 }
2348 }
2349
2350 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2351 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2352 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2353 Node* n = trap->uncommon_trap();
2354 if (!useful.member(n)) {
2355 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2356 }
2357 }
2358 }
2359
2360 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2361 // or fold-compares case. Return true if succeed or not found.
2362 //
2363 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2364 // false and ask fold-compares to yield.
2365 //
2366 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2367 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2368 // when deoptimization does happen.
2369 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2370 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2371 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2372 if (trap->uncommon_trap() == unc) {
2373 if (yield && trap->modified()) {
2374 return false;
2375 }
2376 _unstable_if_traps.delete_at(i);
2377 break;
2378 }
2379 }
2380 return true;
2381 }
2382
2383 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2384 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2385 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2386 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2387 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2388 CallStaticJavaNode* unc = trap->uncommon_trap();
2389 int next_bci = trap->next_bci();
2390 bool modified = trap->modified();
2391
2392 if (next_bci != -1 && !modified) {
2393 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2394 JVMState* jvms = unc->jvms();
2395 ciMethod* method = jvms->method();
2396 ciBytecodeStream iter(method);
2397
2398 iter.force_bci(jvms->bci());
2399 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2400 Bytecodes::Code c = iter.cur_bc();
2401 Node* lhs = nullptr;
2402 Node* rhs = nullptr;
2403 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2404 lhs = unc->peek_operand(0);
2405 rhs = unc->peek_operand(1);
2406 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2407 lhs = unc->peek_operand(0);
2408 }
2409
2410 ResourceMark rm;
2411 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2412 assert(live_locals.is_valid(), "broken liveness info");
2413 int len = (int)live_locals.size();
2414
2415 for (int i = 0; i < len; i++) {
2416 Node* local = unc->local(jvms, i);
2417 // kill local using the liveness of next_bci.
2418 // give up when the local looks like an operand to secure reexecution.
2419 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2420 uint idx = jvms->locoff() + i;
2421 #ifdef ASSERT
2422 if (PrintOpto && Verbose) {
2423 tty->print("[unstable_if] kill local#%d: ", idx);
2424 local->dump();
2425 tty->cr();
2426 }
2427 #endif
2428 igvn.replace_input_of(unc, idx, top());
2429 modified = true;
2430 }
2431 }
2432 }
2433
2434 // keep the mondified trap for late query
2435 if (modified) {
2436 trap->set_modified();
2437 } else {
2438 _unstable_if_traps.delete_at(i);
2439 }
2440 }
2441 igvn.optimize();
2442 }
2443
2444 // StringOpts and late inlining of string methods
2445 void Compile::inline_string_calls(bool parse_time) {
2446 {
2447 // remove useless nodes to make the usage analysis simpler
2448 ResourceMark rm;
2449 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2450 }
2451
2452 {
2453 ResourceMark rm;
2454 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2455 PhaseStringOpts pso(initial_gvn());
2456 print_method(PHASE_AFTER_STRINGOPTS, 3);
2457 }
2458
2459 // now inline anything that we skipped the first time around
2460 if (!parse_time) {
2461 _late_inlines_pos = _late_inlines.length();
2462 }
2463
2464 while (_string_late_inlines.length() > 0) {
2465 CallGenerator* cg = _string_late_inlines.pop();
2466 cg->do_late_inline();
2467 if (failing()) return;
2468 }
2469 _string_late_inlines.trunc_to(0);
2470 }
2471
2472 // Late inlining of boxing methods
2473 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2474 if (_boxing_late_inlines.length() > 0) {
2475 assert(has_boxed_value(), "inconsistent");
2476
2477 PhaseGVN* gvn = initial_gvn();
2478 set_inlining_incrementally(true);
2479
2480 igvn_worklist()->ensure_empty(); // should be done with igvn
2481
2482 _late_inlines_pos = _late_inlines.length();
2483
2484 while (_boxing_late_inlines.length() > 0) {
2485 CallGenerator* cg = _boxing_late_inlines.pop();
2486 cg->do_late_inline();
2487 if (failing()) return;
2488 }
2489 _boxing_late_inlines.trunc_to(0);
2490
2491 inline_incrementally_cleanup(igvn);
2492
2493 set_inlining_incrementally(false);
2494 }
2495 }
2496
2497 bool Compile::inline_incrementally_one() {
2498 assert(IncrementalInline, "incremental inlining should be on");
2499
2500 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2501
2502 set_inlining_progress(false);
2503 set_do_cleanup(false);
2504
2505 for (int i = 0; i < _late_inlines.length(); i++) {
2506 _late_inlines_pos = i+1;
2507 CallGenerator* cg = _late_inlines.at(i);
2508 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2509 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2510 cg->do_late_inline();
2511 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2512 if (failing()) {
2513 return false;
2514 } else if (inlining_progress()) {
2515 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2516 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2517 break; // process one call site at a time
2518 }
2519 } else {
2520 // Ignore late inline direct calls when inlining is not allowed.
2521 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2522 }
2523 }
2524 // Remove processed elements.
2525 _late_inlines.remove_till(_late_inlines_pos);
2526 _late_inlines_pos = 0;
2527
2528 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2529
2530 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2531
2532 set_inlining_progress(false);
2533 set_do_cleanup(false);
2534
2535 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2536 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2537 }
2538
2539 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2540 {
2541 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2542 ResourceMark rm;
2543 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2544 }
2545 {
2546 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2547 igvn.reset_from_gvn(initial_gvn());
2548 igvn.optimize();
2549 }
2550 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2551 }
2552
2553 // Perform incremental inlining until bound on number of live nodes is reached
2554 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2555 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2556
2557 set_inlining_incrementally(true);
2558 uint low_live_nodes = 0;
2559
2560 while (_late_inlines.length() > 0) {
2561 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2562 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2563 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2564 // PhaseIdealLoop is expensive so we only try it once we are
2565 // out of live nodes and we only try it again if the previous
2566 // helped got the number of nodes down significantly
2567 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2568 if (failing()) return;
2569 low_live_nodes = live_nodes();
2570 _major_progress = true;
2571 }
2572
2573 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2574 bool do_print_inlining = print_inlining() || print_intrinsics();
2575 if (do_print_inlining || log() != nullptr) {
2576 // Print inlining message for candidates that we couldn't inline for lack of space.
2577 for (int i = 0; i < _late_inlines.length(); i++) {
2578 CallGenerator* cg = _late_inlines.at(i);
2579 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2580 if (do_print_inlining) {
2581 cg->print_inlining_late(msg);
2582 }
2583 log_late_inline_failure(cg, msg);
2584 }
2585 }
2586 break; // finish
2587 }
2588 }
2589
2590 igvn_worklist()->ensure_empty(); // should be done with igvn
2591
2592 while (inline_incrementally_one()) {
2593 assert(!failing(), "inconsistent");
2594 }
2595 if (failing()) return;
2596
2597 inline_incrementally_cleanup(igvn);
2598
2599 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2600
2601 if (failing()) return;
2602
2603 if (_late_inlines.length() == 0) {
2604 break; // no more progress
2605 }
2606 }
2607
2608 igvn_worklist()->ensure_empty(); // should be done with igvn
2609
2610 if (_string_late_inlines.length() > 0) {
2611 assert(has_stringbuilder(), "inconsistent");
2612
2613 inline_string_calls(false);
2614
2615 if (failing()) return;
2616
2617 inline_incrementally_cleanup(igvn);
2618 }
2619
2620 set_inlining_incrementally(false);
2621 }
2622
2623 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2624 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2625 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2626 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2627 // as if "inlining_incrementally() == true" were set.
2628 assert(inlining_incrementally() == false, "not allowed");
2629 #ifdef ASSERT
2630 Unique_Node_List* modified_nodes = _modified_nodes;
2631 _modified_nodes = nullptr;
2632 #endif
2633 assert(_late_inlines.length() > 0, "sanity");
2634
2635 while (_late_inlines.length() > 0) {
2636 igvn_worklist()->ensure_empty(); // should be done with igvn
2637
2638 while (inline_incrementally_one()) {
2639 assert(!failing(), "inconsistent");
2640 }
2641 if (failing()) return;
2642
2643 inline_incrementally_cleanup(igvn);
2644 }
2645 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2646 }
2647
2648 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2649 if (_loop_opts_cnt > 0) {
2650 while (major_progress() && (_loop_opts_cnt > 0)) {
2651 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2652 PhaseIdealLoop::optimize(igvn, mode);
2653 _loop_opts_cnt--;
2654 if (failing()) return false;
2655 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2656 }
2657 }
2658 return true;
2659 }
2660
2661 // Remove edges from "root" to each SafePoint at a backward branch.
2662 // They were inserted during parsing (see add_safepoint()) to make
2663 // infinite loops without calls or exceptions visible to root, i.e.,
2664 // useful.
2665 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2666 Node *r = root();
2667 if (r != nullptr) {
2668 for (uint i = r->req(); i < r->len(); ++i) {
2669 Node *n = r->in(i);
2670 if (n != nullptr && n->is_SafePoint()) {
2671 r->rm_prec(i);
2672 if (n->outcnt() == 0) {
2673 igvn.remove_dead_node(n);
2674 }
2675 --i;
2676 }
2677 }
2678 // Parsing may have added top inputs to the root node (Path
2679 // leading to the Halt node proven dead). Make sure we get a
2680 // chance to clean them up.
2681 igvn._worklist.push(r);
2682 igvn.optimize();
2683 }
2684 }
2685
2686 //------------------------------Optimize---------------------------------------
2687 // Given a graph, optimize it.
2688 void Compile::Optimize() {
2689 TracePhase tp("optimizer", &timers[_t_optimizer]);
2690
2691 #ifndef PRODUCT
2692 if (env()->break_at_compile()) {
2693 BREAKPOINT;
2694 }
2695
2696 #endif
2697
2698 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2699 #ifdef ASSERT
2700 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2701 #endif
2702
2703 ResourceMark rm;
2704
2705 print_inlining_reinit();
2706
2707 NOT_PRODUCT( verify_graph_edges(); )
2708
2709 print_method(PHASE_AFTER_PARSING, 1);
2710
2711 {
2712 // Iterative Global Value Numbering, including ideal transforms
2713 // Initialize IterGVN with types and values from parse-time GVN
2714 PhaseIterGVN igvn(initial_gvn());
2715 #ifdef ASSERT
2716 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2717 #endif
2718 {
2719 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2720 igvn.optimize();
2721 }
2722
2723 if (failing()) return;
2724
2725 print_method(PHASE_ITER_GVN1, 2);
2726
2727 process_for_unstable_if_traps(igvn);
2728
2729 inline_incrementally(igvn);
2730
2731 print_method(PHASE_INCREMENTAL_INLINE, 2);
2732
2733 if (failing()) return;
2734
2735 if (eliminate_boxing()) {
2736 // Inline valueOf() methods now.
2737 inline_boxing_calls(igvn);
2738
2739 if (AlwaysIncrementalInline) {
2740 inline_incrementally(igvn);
2741 }
2742
2743 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2744
2745 if (failing()) return;
2746 }
2747
2748 // Remove the speculative part of types and clean up the graph from
2749 // the extra CastPP nodes whose only purpose is to carry them. Do
2750 // that early so that optimizations are not disrupted by the extra
2751 // CastPP nodes.
2752 remove_speculative_types(igvn);
2753
2754 // No more new expensive nodes will be added to the list from here
2755 // so keep only the actual candidates for optimizations.
2756 cleanup_expensive_nodes(igvn);
2757
2758 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2759 if (EnableVectorSupport && has_vbox_nodes()) {
2760 TracePhase tp("", &timers[_t_vector]);
2761 PhaseVector pv(igvn);
2762 pv.optimize_vector_boxes();
2763
2764 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2765 }
2766 assert(!has_vbox_nodes(), "sanity");
2767
2768 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2769 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2770 igvn_worklist()->ensure_empty(); // should be done with igvn
2771 {
2772 ResourceMark rm;
2773 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2774 }
2775 igvn.reset_from_gvn(initial_gvn());
2776 igvn.optimize();
2777 }
2778
2779 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2780 // safepoints
2781 remove_root_to_sfpts_edges(igvn);
2782
2783 // Process inline type nodes now that all inlining is over
2784 process_inline_types(igvn);
2785
2786 adjust_flat_array_access_aliases(igvn);
2787
2788 // Perform escape analysis
2789 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2790 if (has_loops()) {
2791 // Cleanup graph (remove dead nodes).
2792 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2793 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2794 if (failing()) return;
2795 }
2796 bool progress;
2797 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2798 do {
2799 ConnectionGraph::do_analysis(this, &igvn);
2800
2801 if (failing()) return;
2802
2803 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2804
2805 // Optimize out fields loads from scalar replaceable allocations.
2806 igvn.optimize();
2807 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2808
2809 if (failing()) return;
2810
2811 if (congraph() != nullptr && macro_count() > 0) {
2812 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2813 PhaseMacroExpand mexp(igvn);
2814 mexp.eliminate_macro_nodes();
2815 igvn.set_delay_transform(false);
2816
2817 igvn.optimize();
2818 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2819 }
2820
2821 ConnectionGraph::verify_ram_nodes(this, root());
2822 if (failing()) return;
2823
2824 progress = do_iterative_escape_analysis() &&
2825 (macro_count() < mcount) &&
2826 ConnectionGraph::has_candidates(this);
2827 // Try again if candidates exist and made progress
2828 // by removing some allocations and/or locks.
2829 } while (progress);
2830 }
2831
2832 // Loop transforms on the ideal graph. Range Check Elimination,
2833 // peeling, unrolling, etc.
2834
2835 // Set loop opts counter
2836 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2837 {
2838 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2839 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2840 _loop_opts_cnt--;
2841 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2842 if (failing()) return;
2843 }
2844 // Loop opts pass if partial peeling occurred in previous pass
2845 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2846 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2847 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2848 _loop_opts_cnt--;
2849 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2850 if (failing()) return;
2851 }
2852 // Loop opts pass for loop-unrolling before CCP
2853 if(major_progress() && (_loop_opts_cnt > 0)) {
2854 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2855 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2856 _loop_opts_cnt--;
2857 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2858 }
2859 if (!failing()) {
2860 // Verify that last round of loop opts produced a valid graph
2861 PhaseIdealLoop::verify(igvn);
2862 }
2863 }
2864 if (failing()) return;
2865
2866 // Conditional Constant Propagation;
2867 PhaseCCP ccp( &igvn );
2868 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2869 {
2870 TracePhase tp("ccp", &timers[_t_ccp]);
2871 ccp.do_transform();
2872 }
2873 print_method(PHASE_CCP1, 2);
2874
2875 assert( true, "Break here to ccp.dump_old2new_map()");
2876
2877 // Iterative Global Value Numbering, including ideal transforms
2878 {
2879 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2880 igvn.reset_from_igvn(&ccp);
2881 igvn.optimize();
2882 }
2883 print_method(PHASE_ITER_GVN2, 2);
2884
2885 if (failing()) return;
2886
2887 // Loop transforms on the ideal graph. Range Check Elimination,
2888 // peeling, unrolling, etc.
2889 if (!optimize_loops(igvn, LoopOptsDefault)) {
2890 return;
2891 }
2892
2893 if (failing()) return;
2894
2895 C->clear_major_progress(); // ensure that major progress is now clear
2896
2897 process_for_post_loop_opts_igvn(igvn);
2898
2899 #ifdef ASSERT
2900 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2901 #endif
2902
2903 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2904
2905 if (_late_inlines.length() > 0) {
2906 // More opportunities to optimize virtual and MH calls.
2907 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2908 process_late_inline_calls_no_inline(igvn);
2909 }
2910
2911 {
2912 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2913 PhaseMacroExpand mex(igvn);
2914 if (mex.expand_macro_nodes()) {
2915 assert(failing(), "must bail out w/ explicit message");
2916 return;
2917 }
2918 print_method(PHASE_MACRO_EXPANSION, 2);
2919 }
2920
2921 // Process inline type nodes again and remove them. From here
2922 // on we don't need to keep track of field values anymore.
2923 process_inline_types(igvn, /* remove= */ true);
2924
2925 {
2926 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2927 if (bs->expand_barriers(this, igvn)) {
2928 assert(failing(), "must bail out w/ explicit message");
2929 return;
2930 }
2931 print_method(PHASE_BARRIER_EXPANSION, 2);
2932 }
2933
2934 if (C->max_vector_size() > 0) {
2935 C->optimize_logic_cones(igvn);
2936 igvn.optimize();
2937 }
2938
2939 DEBUG_ONLY( _modified_nodes = nullptr; )
2940 DEBUG_ONLY( _late_inlines.clear(); )
2941
2942 assert(igvn._worklist.size() == 0, "not empty");
2943 } // (End scope of igvn; run destructor if necessary for asserts.)
2944
2945 check_no_dead_use();
2946
2947 process_print_inlining();
2948
2949 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2950 // to remove hashes to unlock nodes for modifications.
2951 C->node_hash()->clear();
2952
2953 // A method with only infinite loops has no edges entering loops from root
2954 {
2955 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2956 if (final_graph_reshaping()) {
2957 assert(failing(), "must bail out w/ explicit message");
2958 return;
2959 }
2960 }
2961
2962 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2963 DEBUG_ONLY(set_phase_optimize_finished();)
2964 }
2965
2966 #ifdef ASSERT
2967 void Compile::check_no_dead_use() const {
2968 ResourceMark rm;
2969 Unique_Node_List wq;
2970 wq.push(root());
2971 for (uint i = 0; i < wq.size(); ++i) {
2972 Node* n = wq.at(i);
2973 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2974 Node* u = n->fast_out(j);
2975 if (u->outcnt() == 0 && !u->is_Con()) {
2976 u->dump();
2977 fatal("no reachable node should have no use");
2978 }
2979 wq.push(u);
2980 }
2981 }
2982 }
2983 #endif
2984
2985 void Compile::inline_vector_reboxing_calls() {
2986 if (C->_vector_reboxing_late_inlines.length() > 0) {
2987 _late_inlines_pos = C->_late_inlines.length();
2988 while (_vector_reboxing_late_inlines.length() > 0) {
2989 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2990 cg->do_late_inline();
2991 if (failing()) return;
2992 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2993 }
2994 _vector_reboxing_late_inlines.trunc_to(0);
2995 }
2996 }
2997
2998 bool Compile::has_vbox_nodes() {
2999 if (C->_vector_reboxing_late_inlines.length() > 0) {
3000 return true;
3001 }
3002 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3003 Node * n = C->macro_node(macro_idx);
3004 assert(n->is_macro(), "only macro nodes expected here");
3005 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3006 return true;
3007 }
3008 }
3009 return false;
3010 }
3011
3012 //---------------------------- Bitwise operation packing optimization ---------------------------
3013
3014 static bool is_vector_unary_bitwise_op(Node* n) {
3015 return n->Opcode() == Op_XorV &&
3016 VectorNode::is_vector_bitwise_not_pattern(n);
3017 }
3018
3019 static bool is_vector_binary_bitwise_op(Node* n) {
3020 switch (n->Opcode()) {
3021 case Op_AndV:
3022 case Op_OrV:
3023 return true;
3024
3025 case Op_XorV:
3026 return !is_vector_unary_bitwise_op(n);
3027
3028 default:
3029 return false;
3030 }
3031 }
3032
3033 static bool is_vector_ternary_bitwise_op(Node* n) {
3034 return n->Opcode() == Op_MacroLogicV;
3035 }
3036
3037 static bool is_vector_bitwise_op(Node* n) {
3038 return is_vector_unary_bitwise_op(n) ||
3039 is_vector_binary_bitwise_op(n) ||
3040 is_vector_ternary_bitwise_op(n);
3041 }
3042
3043 static bool is_vector_bitwise_cone_root(Node* n) {
3044 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
3045 return false;
3046 }
3047 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3048 if (is_vector_bitwise_op(n->fast_out(i))) {
3049 return false;
3050 }
3051 }
3052 return true;
3053 }
3054
3055 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3056 uint cnt = 0;
3057 if (is_vector_bitwise_op(n)) {
3058 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3059 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3060 for (uint i = 1; i < inp_cnt; i++) {
3061 Node* in = n->in(i);
3062 bool skip = VectorNode::is_all_ones_vector(in);
3063 if (!skip && !inputs.member(in)) {
3064 inputs.push(in);
3065 cnt++;
3066 }
3067 }
3068 assert(cnt <= 1, "not unary");
3069 } else {
3070 uint last_req = inp_cnt;
3071 if (is_vector_ternary_bitwise_op(n)) {
3072 last_req = inp_cnt - 1; // skip last input
3073 }
3074 for (uint i = 1; i < last_req; i++) {
3075 Node* def = n->in(i);
3076 if (!inputs.member(def)) {
3077 inputs.push(def);
3078 cnt++;
3079 }
3080 }
3081 }
3082 } else { // not a bitwise operations
3083 if (!inputs.member(n)) {
3084 inputs.push(n);
3085 cnt++;
3086 }
3087 }
3088 return cnt;
3089 }
3090
3091 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3092 Unique_Node_List useful_nodes;
3093 C->identify_useful_nodes(useful_nodes);
3094
3095 for (uint i = 0; i < useful_nodes.size(); i++) {
3096 Node* n = useful_nodes.at(i);
3097 if (is_vector_bitwise_cone_root(n)) {
3098 list.push(n);
3099 }
3100 }
3101 }
3102
3103 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3104 const TypeVect* vt,
3105 Unique_Node_List& partition,
3106 Unique_Node_List& inputs) {
3107 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3108 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3109 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3110
3111 Node* in1 = inputs.at(0);
3112 Node* in2 = inputs.at(1);
3113 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3114
3115 uint func = compute_truth_table(partition, inputs);
3116
3117 Node* pn = partition.at(partition.size() - 1);
3118 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3119 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3120 }
3121
3122 static uint extract_bit(uint func, uint pos) {
3123 return (func & (1 << pos)) >> pos;
3124 }
3125
3126 //
3127 // A macro logic node represents a truth table. It has 4 inputs,
3128 // First three inputs corresponds to 3 columns of a truth table
3129 // and fourth input captures the logic function.
3130 //
3131 // eg. fn = (in1 AND in2) OR in3;
3132 //
3133 // MacroNode(in1,in2,in3,fn)
3134 //
3135 // -----------------
3136 // in1 in2 in3 fn
3137 // -----------------
3138 // 0 0 0 0
3139 // 0 0 1 1
3140 // 0 1 0 0
3141 // 0 1 1 1
3142 // 1 0 0 0
3143 // 1 0 1 1
3144 // 1 1 0 1
3145 // 1 1 1 1
3146 //
3147
3148 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3149 int res = 0;
3150 for (int i = 0; i < 8; i++) {
3151 int bit1 = extract_bit(in1, i);
3152 int bit2 = extract_bit(in2, i);
3153 int bit3 = extract_bit(in3, i);
3154
3155 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3156 int func_bit = extract_bit(func, func_bit_pos);
3157
3158 res |= func_bit << i;
3159 }
3160 return res;
3161 }
3162
3163 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
3164 assert(n != nullptr, "");
3165 assert(eval_map.contains(n), "absent");
3166 return *(eval_map.get(n));
3167 }
3168
3169 static void eval_operands(Node* n,
3170 uint& func1, uint& func2, uint& func3,
3171 ResourceHashtable<Node*,uint>& eval_map) {
3172 assert(is_vector_bitwise_op(n), "");
3173
3174 if (is_vector_unary_bitwise_op(n)) {
3175 Node* opnd = n->in(1);
3176 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3177 opnd = n->in(2);
3178 }
3179 func1 = eval_operand(opnd, eval_map);
3180 } else if (is_vector_binary_bitwise_op(n)) {
3181 func1 = eval_operand(n->in(1), eval_map);
3182 func2 = eval_operand(n->in(2), eval_map);
3183 } else {
3184 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3185 func1 = eval_operand(n->in(1), eval_map);
3186 func2 = eval_operand(n->in(2), eval_map);
3187 func3 = eval_operand(n->in(3), eval_map);
3188 }
3189 }
3190
3191 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3192 assert(inputs.size() <= 3, "sanity");
3193 ResourceMark rm;
3194 uint res = 0;
3195 ResourceHashtable<Node*,uint> eval_map;
3196
3197 // Populate precomputed functions for inputs.
3198 // Each input corresponds to one column of 3 input truth-table.
3199 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3200 0xCC, // (_, b, _) -> b
3201 0xF0 }; // (a, _, _) -> a
3202 for (uint i = 0; i < inputs.size(); i++) {
3203 eval_map.put(inputs.at(i), input_funcs[2-i]);
3204 }
3205
3206 for (uint i = 0; i < partition.size(); i++) {
3207 Node* n = partition.at(i);
3208
3209 uint func1 = 0, func2 = 0, func3 = 0;
3210 eval_operands(n, func1, func2, func3, eval_map);
3211
3212 switch (n->Opcode()) {
3213 case Op_OrV:
3214 assert(func3 == 0, "not binary");
3215 res = func1 | func2;
3216 break;
3217 case Op_AndV:
3218 assert(func3 == 0, "not binary");
3219 res = func1 & func2;
3220 break;
3221 case Op_XorV:
3222 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3223 assert(func2 == 0 && func3 == 0, "not unary");
3224 res = (~func1) & 0xFF;
3225 } else {
3226 assert(func3 == 0, "not binary");
3227 res = func1 ^ func2;
3228 }
3229 break;
3230 case Op_MacroLogicV:
3231 // Ordering of inputs may change during evaluation of sub-tree
3232 // containing MacroLogic node as a child node, thus a re-evaluation
3233 // makes sure that function is evaluated in context of current
3234 // inputs.
3235 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3236 break;
3237
3238 default: assert(false, "not supported: %s", n->Name());
3239 }
3240 assert(res <= 0xFF, "invalid");
3241 eval_map.put(n, res);
3242 }
3243 return res;
3244 }
3245
3246 // Criteria under which nodes gets packed into a macro logic node:-
3247 // 1) Parent and both child nodes are all unmasked or masked with
3248 // same predicates.
3249 // 2) Masked parent can be packed with left child if it is predicated
3250 // and both have same predicates.
3251 // 3) Masked parent can be packed with right child if its un-predicated
3252 // or has matching predication condition.
3253 // 4) An unmasked parent can be packed with an unmasked child.
3254 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3255 assert(partition.size() == 0, "not empty");
3256 assert(inputs.size() == 0, "not empty");
3257 if (is_vector_ternary_bitwise_op(n)) {
3258 return false;
3259 }
3260
3261 bool is_unary_op = is_vector_unary_bitwise_op(n);
3262 if (is_unary_op) {
3263 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3264 return false; // too few inputs
3265 }
3266
3267 bool pack_left_child = true;
3268 bool pack_right_child = true;
3269
3270 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3271 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3272
3273 int left_child_input_cnt = 0;
3274 int right_child_input_cnt = 0;
3275
3276 bool parent_is_predicated = n->is_predicated_vector();
3277 bool left_child_predicated = n->in(1)->is_predicated_vector();
3278 bool right_child_predicated = n->in(2)->is_predicated_vector();
3279
3280 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
3281 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3282 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3283
3284 do {
3285 if (pack_left_child && left_child_LOP &&
3286 ((!parent_is_predicated && !left_child_predicated) ||
3287 ((parent_is_predicated && left_child_predicated &&
3288 parent_pred == left_child_pred)))) {
3289 partition.push(n->in(1));
3290 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3291 } else {
3292 inputs.push(n->in(1));
3293 left_child_input_cnt = 1;
3294 }
3295
3296 if (pack_right_child && right_child_LOP &&
3297 (!right_child_predicated ||
3298 (right_child_predicated && parent_is_predicated &&
3299 parent_pred == right_child_pred))) {
3300 partition.push(n->in(2));
3301 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3302 } else {
3303 inputs.push(n->in(2));
3304 right_child_input_cnt = 1;
3305 }
3306
3307 if (inputs.size() > 3) {
3308 assert(partition.size() > 0, "");
3309 inputs.clear();
3310 partition.clear();
3311 if (left_child_input_cnt > right_child_input_cnt) {
3312 pack_left_child = false;
3313 } else {
3314 pack_right_child = false;
3315 }
3316 } else {
3317 break;
3318 }
3319 } while(true);
3320
3321 if(partition.size()) {
3322 partition.push(n);
3323 }
3324
3325 return (partition.size() == 2 || partition.size() == 3) &&
3326 (inputs.size() == 2 || inputs.size() == 3);
3327 }
3328
3329 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3330 assert(is_vector_bitwise_op(n), "not a root");
3331
3332 visited.set(n->_idx);
3333
3334 // 1) Do a DFS walk over the logic cone.
3335 for (uint i = 1; i < n->req(); i++) {
3336 Node* in = n->in(i);
3337 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3338 process_logic_cone_root(igvn, in, visited);
3339 }
3340 }
3341
3342 // 2) Bottom up traversal: Merge node[s] with
3343 // the parent to form macro logic node.
3344 Unique_Node_List partition;
3345 Unique_Node_List inputs;
3346 if (compute_logic_cone(n, partition, inputs)) {
3347 const TypeVect* vt = n->bottom_type()->is_vect();
3348 Node* pn = partition.at(partition.size() - 1);
3349 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3350 if (mask == nullptr ||
3351 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3352 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3353 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3354 igvn.replace_node(n, macro_logic);
3355 }
3356 }
3357 }
3358
3359 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3360 ResourceMark rm;
3361 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3362 Unique_Node_List list;
3363 collect_logic_cone_roots(list);
3364
3365 while (list.size() > 0) {
3366 Node* n = list.pop();
3367 const TypeVect* vt = n->bottom_type()->is_vect();
3368 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3369 if (supported) {
3370 VectorSet visited(comp_arena());
3371 process_logic_cone_root(igvn, n, visited);
3372 }
3373 }
3374 }
3375 }
3376
3377 //------------------------------Code_Gen---------------------------------------
3378 // Given a graph, generate code for it
3379 void Compile::Code_Gen() {
3380 if (failing()) {
3381 return;
3382 }
3383
3384 // Perform instruction selection. You might think we could reclaim Matcher
3385 // memory PDQ, but actually the Matcher is used in generating spill code.
3386 // Internals of the Matcher (including some VectorSets) must remain live
3387 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3388 // set a bit in reclaimed memory.
3389
3390 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3391 // nodes. Mapping is only valid at the root of each matched subtree.
3392 NOT_PRODUCT( verify_graph_edges(); )
3393
3394 Matcher matcher;
3395 _matcher = &matcher;
3396 {
3397 TracePhase tp("matcher", &timers[_t_matcher]);
3398 matcher.match();
3399 if (failing()) {
3400 return;
3401 }
3402 }
3403 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3404 // nodes. Mapping is only valid at the root of each matched subtree.
3405 NOT_PRODUCT( verify_graph_edges(); )
3406
3407 // If you have too many nodes, or if matching has failed, bail out
3408 check_node_count(0, "out of nodes matching instructions");
3409 if (failing()) {
3410 return;
3411 }
3412
3413 print_method(PHASE_MATCHING, 2);
3414
3415 // Build a proper-looking CFG
3416 PhaseCFG cfg(node_arena(), root(), matcher);
3417 _cfg = &cfg;
3418 {
3419 TracePhase tp("scheduler", &timers[_t_scheduler]);
3420 bool success = cfg.do_global_code_motion();
3421 if (!success) {
3422 return;
3423 }
3424
3425 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3426 NOT_PRODUCT( verify_graph_edges(); )
3427 cfg.verify();
3428 }
3429
3430 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3431 _regalloc = ®alloc;
3432 {
3433 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
3434 // Perform register allocation. After Chaitin, use-def chains are
3435 // no longer accurate (at spill code) and so must be ignored.
3436 // Node->LRG->reg mappings are still accurate.
3437 _regalloc->Register_Allocate();
3438
3439 // Bail out if the allocator builds too many nodes
3440 if (failing()) {
3441 return;
3442 }
3443 }
3444
3445 // Prior to register allocation we kept empty basic blocks in case the
3446 // the allocator needed a place to spill. After register allocation we
3447 // are not adding any new instructions. If any basic block is empty, we
3448 // can now safely remove it.
3449 {
3450 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
3451 cfg.remove_empty_blocks();
3452 if (do_freq_based_layout()) {
3453 PhaseBlockLayout layout(cfg);
3454 } else {
3455 cfg.set_loop_alignment();
3456 }
3457 cfg.fixup_flow();
3458 cfg.remove_unreachable_blocks();
3459 cfg.verify_dominator_tree();
3460 }
3461
3462 // Apply peephole optimizations
3463 if( OptoPeephole ) {
3464 TracePhase tp("peephole", &timers[_t_peephole]);
3465 PhasePeephole peep( _regalloc, cfg);
3466 peep.do_transform();
3467 }
3468
3469 // Do late expand if CPU requires this.
3470 if (Matcher::require_postalloc_expand) {
3471 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3472 cfg.postalloc_expand(_regalloc);
3473 }
3474
3475 // Convert Nodes to instruction bits in a buffer
3476 {
3477 TracePhase tp("output", &timers[_t_output]);
3478 PhaseOutput output;
3479 output.Output();
3480 if (failing()) return;
3481 output.install();
3482 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3483 }
3484
3485 // He's dead, Jim.
3486 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3487 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3488 }
3489
3490 //------------------------------Final_Reshape_Counts---------------------------
3491 // This class defines counters to help identify when a method
3492 // may/must be executed using hardware with only 24-bit precision.
3493 struct Final_Reshape_Counts : public StackObj {
3494 int _call_count; // count non-inlined 'common' calls
3495 int _float_count; // count float ops requiring 24-bit precision
3496 int _double_count; // count double ops requiring more precision
3497 int _java_call_count; // count non-inlined 'java' calls
3498 int _inner_loop_count; // count loops which need alignment
3499 VectorSet _visited; // Visitation flags
3500 Node_List _tests; // Set of IfNodes & PCTableNodes
3501
3502 Final_Reshape_Counts() :
3503 _call_count(0), _float_count(0), _double_count(0),
3504 _java_call_count(0), _inner_loop_count(0) { }
3505
3506 void inc_call_count () { _call_count ++; }
3507 void inc_float_count () { _float_count ++; }
3508 void inc_double_count() { _double_count++; }
3509 void inc_java_call_count() { _java_call_count++; }
3510 void inc_inner_loop_count() { _inner_loop_count++; }
3511
3512 int get_call_count () const { return _call_count ; }
3513 int get_float_count () const { return _float_count ; }
3514 int get_double_count() const { return _double_count; }
3515 int get_java_call_count() const { return _java_call_count; }
3516 int get_inner_loop_count() const { return _inner_loop_count; }
3517 };
3518
3519 // Eliminate trivially redundant StoreCMs and accumulate their
3520 // precedence edges.
3521 void Compile::eliminate_redundant_card_marks(Node* n) {
3522 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3523 if (n->in(MemNode::Address)->outcnt() > 1) {
3524 // There are multiple users of the same address so it might be
3525 // possible to eliminate some of the StoreCMs
3526 Node* mem = n->in(MemNode::Memory);
3527 Node* adr = n->in(MemNode::Address);
3528 Node* val = n->in(MemNode::ValueIn);
3529 Node* prev = n;
3530 bool done = false;
3531 // Walk the chain of StoreCMs eliminating ones that match. As
3532 // long as it's a chain of single users then the optimization is
3533 // safe. Eliminating partially redundant StoreCMs would require
3534 // cloning copies down the other paths.
3535 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3536 if (adr == mem->in(MemNode::Address) &&
3537 val == mem->in(MemNode::ValueIn)) {
3538 // redundant StoreCM
3539 if (mem->req() > MemNode::OopStore) {
3540 // Hasn't been processed by this code yet.
3541 n->add_prec(mem->in(MemNode::OopStore));
3542 } else {
3543 // Already converted to precedence edge
3544 for (uint i = mem->req(); i < mem->len(); i++) {
3545 // Accumulate any precedence edges
3546 if (mem->in(i) != nullptr) {
3547 n->add_prec(mem->in(i));
3548 }
3549 }
3550 // Everything above this point has been processed.
3551 done = true;
3552 }
3553 // Eliminate the previous StoreCM
3554 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3555 assert(mem->outcnt() == 0, "should be dead");
3556 mem->disconnect_inputs(this);
3557 } else {
3558 prev = mem;
3559 }
3560 mem = prev->in(MemNode::Memory);
3561 }
3562 }
3563 }
3564
3565
3566 //------------------------------final_graph_reshaping_impl----------------------
3567 // Implement items 1-5 from final_graph_reshaping below.
3568 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3569
3570 if ( n->outcnt() == 0 ) return; // dead node
3571 uint nop = n->Opcode();
3572
3573 // Check for 2-input instruction with "last use" on right input.
3574 // Swap to left input. Implements item (2).
3575 if( n->req() == 3 && // two-input instruction
3576 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3577 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3578 n->in(2)->outcnt() == 1 &&// right use IS a last use
3579 !n->in(2)->is_Con() ) { // right use is not a constant
3580 // Check for commutative opcode
3581 switch( nop ) {
3582 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3583 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3584 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3585 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3586 case Op_AndL: case Op_XorL: case Op_OrL:
3587 case Op_AndI: case Op_XorI: case Op_OrI: {
3588 // Move "last use" input to left by swapping inputs
3589 n->swap_edges(1, 2);
3590 break;
3591 }
3592 default:
3593 break;
3594 }
3595 }
3596
3597 #ifdef ASSERT
3598 if( n->is_Mem() ) {
3599 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3600 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3601 // oop will be recorded in oop map if load crosses safepoint
3602 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3603 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3604 "raw memory operations should have control edge");
3605 }
3606 if (n->is_MemBar()) {
3607 MemBarNode* mb = n->as_MemBar();
3608 if (mb->trailing_store() || mb->trailing_load_store()) {
3609 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3610 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3611 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3612 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3613 } else if (mb->leading()) {
3614 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3615 }
3616 }
3617 #endif
3618 // Count FPU ops and common calls, implements item (3)
3619 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3620 if (!gc_handled) {
3621 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3622 }
3623
3624 // Collect CFG split points
3625 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3626 frc._tests.push(n);
3627 }
3628 }
3629
3630 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3631 switch( nop ) {
3632 // Count all float operations that may use FPU
3633 case Op_AddF:
3634 case Op_SubF:
3635 case Op_MulF:
3636 case Op_DivF:
3637 case Op_NegF:
3638 case Op_ModF:
3639 case Op_ConvI2F:
3640 case Op_ConF:
3641 case Op_CmpF:
3642 case Op_CmpF3:
3643 case Op_StoreF:
3644 case Op_LoadF:
3645 // case Op_ConvL2F: // longs are split into 32-bit halves
3646 frc.inc_float_count();
3647 break;
3648
3649 case Op_ConvF2D:
3650 case Op_ConvD2F:
3651 frc.inc_float_count();
3652 frc.inc_double_count();
3653 break;
3654
3655 // Count all double operations that may use FPU
3656 case Op_AddD:
3657 case Op_SubD:
3658 case Op_MulD:
3659 case Op_DivD:
3660 case Op_NegD:
3661 case Op_ModD:
3662 case Op_ConvI2D:
3663 case Op_ConvD2I:
3664 // case Op_ConvL2D: // handled by leaf call
3665 // case Op_ConvD2L: // handled by leaf call
3666 case Op_ConD:
3667 case Op_CmpD:
3668 case Op_CmpD3:
3669 case Op_StoreD:
3670 case Op_LoadD:
3671 case Op_LoadD_unaligned:
3672 frc.inc_double_count();
3673 break;
3674 case Op_Opaque1: // Remove Opaque Nodes before matching
3675 case Op_Opaque3:
3676 n->subsume_by(n->in(1), this);
3677 break;
3678 case Op_CallStaticJava:
3679 case Op_CallJava:
3680 case Op_CallDynamicJava:
3681 frc.inc_java_call_count(); // Count java call site;
3682 case Op_CallRuntime:
3683 case Op_CallLeaf:
3684 case Op_CallLeafVector:
3685 case Op_CallLeafNoFP: {
3686 assert (n->is_Call(), "");
3687 CallNode *call = n->as_Call();
3688 // Count call sites where the FP mode bit would have to be flipped.
3689 // Do not count uncommon runtime calls:
3690 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3691 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3692 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3693 frc.inc_call_count(); // Count the call site
3694 } else { // See if uncommon argument is shared
3695 Node *n = call->in(TypeFunc::Parms);
3696 int nop = n->Opcode();
3697 // Clone shared simple arguments to uncommon calls, item (1).
3698 if (n->outcnt() > 1 &&
3699 !n->is_Proj() &&
3700 nop != Op_CreateEx &&
3701 nop != Op_CheckCastPP &&
3702 nop != Op_DecodeN &&
3703 nop != Op_DecodeNKlass &&
3704 !n->is_Mem() &&
3705 !n->is_Phi()) {
3706 Node *x = n->clone();
3707 call->set_req(TypeFunc::Parms, x);
3708 }
3709 }
3710 break;
3711 }
3712
3713 case Op_StoreCM:
3714 {
3715 // Convert OopStore dependence into precedence edge
3716 Node* prec = n->in(MemNode::OopStore);
3717 n->del_req(MemNode::OopStore);
3718 if (prec->is_MergeMem()) {
3719 MergeMemNode* mm = prec->as_MergeMem();
3720 Node* base = mm->base_memory();
3721 for (int i = AliasIdxRaw + 1; i < num_alias_types(); i++) {
3722 const TypePtr* adr_type = get_adr_type(i);
3723 if (adr_type->is_flat()) {
3724 Node* m = mm->memory_at(i);
3725 n->add_prec(m);
3726 }
3727 }
3728 if (mm->outcnt() == 0) {
3729 mm->disconnect_inputs(this);
3730 }
3731 } else {
3732 n->add_prec(prec);
3733 }
3734 eliminate_redundant_card_marks(n);
3735 }
3736
3737 // fall through
3738
3739 case Op_StoreB:
3740 case Op_StoreC:
3741 case Op_StoreI:
3742 case Op_StoreL:
3743 case Op_CompareAndSwapB:
3744 case Op_CompareAndSwapS:
3745 case Op_CompareAndSwapI:
3746 case Op_CompareAndSwapL:
3747 case Op_CompareAndSwapP:
3748 case Op_CompareAndSwapN:
3749 case Op_WeakCompareAndSwapB:
3750 case Op_WeakCompareAndSwapS:
3751 case Op_WeakCompareAndSwapI:
3752 case Op_WeakCompareAndSwapL:
3753 case Op_WeakCompareAndSwapP:
3754 case Op_WeakCompareAndSwapN:
3755 case Op_CompareAndExchangeB:
3756 case Op_CompareAndExchangeS:
3757 case Op_CompareAndExchangeI:
3758 case Op_CompareAndExchangeL:
3759 case Op_CompareAndExchangeP:
3760 case Op_CompareAndExchangeN:
3761 case Op_GetAndAddS:
3762 case Op_GetAndAddB:
3763 case Op_GetAndAddI:
3764 case Op_GetAndAddL:
3765 case Op_GetAndSetS:
3766 case Op_GetAndSetB:
3767 case Op_GetAndSetI:
3768 case Op_GetAndSetL:
3769 case Op_GetAndSetP:
3770 case Op_GetAndSetN:
3771 case Op_StoreP:
3772 case Op_StoreN:
3773 case Op_StoreNKlass:
3774 case Op_LoadB:
3775 case Op_LoadUB:
3776 case Op_LoadUS:
3777 case Op_LoadI:
3778 case Op_LoadKlass:
3779 case Op_LoadNKlass:
3780 case Op_LoadL:
3781 case Op_LoadL_unaligned:
3782 case Op_LoadP:
3783 case Op_LoadN:
3784 case Op_LoadRange:
3785 case Op_LoadS:
3786 break;
3787
3788 case Op_AddP: { // Assert sane base pointers
3789 Node *addp = n->in(AddPNode::Address);
3790 assert( !addp->is_AddP() ||
3791 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3792 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3793 "Base pointers must match (addp %u)", addp->_idx );
3794 #ifdef _LP64
3795 if ((UseCompressedOops || UseCompressedClassPointers) &&
3796 addp->Opcode() == Op_ConP &&
3797 addp == n->in(AddPNode::Base) &&
3798 n->in(AddPNode::Offset)->is_Con()) {
3799 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3800 // on the platform and on the compressed oops mode.
3801 // Use addressing with narrow klass to load with offset on x86.
3802 // Some platforms can use the constant pool to load ConP.
3803 // Do this transformation here since IGVN will convert ConN back to ConP.
3804 const Type* t = addp->bottom_type();
3805 bool is_oop = t->isa_oopptr() != nullptr;
3806 bool is_klass = t->isa_klassptr() != nullptr;
3807
3808 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3809 (is_klass && Matcher::const_klass_prefer_decode())) {
3810 Node* nn = nullptr;
3811
3812 int op = is_oop ? Op_ConN : Op_ConNKlass;
3813
3814 // Look for existing ConN node of the same exact type.
3815 Node* r = root();
3816 uint cnt = r->outcnt();
3817 for (uint i = 0; i < cnt; i++) {
3818 Node* m = r->raw_out(i);
3819 if (m!= nullptr && m->Opcode() == op &&
3820 m->bottom_type()->make_ptr() == t) {
3821 nn = m;
3822 break;
3823 }
3824 }
3825 if (nn != nullptr) {
3826 // Decode a narrow oop to match address
3827 // [R12 + narrow_oop_reg<<3 + offset]
3828 if (is_oop) {
3829 nn = new DecodeNNode(nn, t);
3830 } else {
3831 nn = new DecodeNKlassNode(nn, t);
3832 }
3833 // Check for succeeding AddP which uses the same Base.
3834 // Otherwise we will run into the assertion above when visiting that guy.
3835 for (uint i = 0; i < n->outcnt(); ++i) {
3836 Node *out_i = n->raw_out(i);
3837 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3838 out_i->set_req(AddPNode::Base, nn);
3839 #ifdef ASSERT
3840 for (uint j = 0; j < out_i->outcnt(); ++j) {
3841 Node *out_j = out_i->raw_out(j);
3842 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3843 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3844 }
3845 #endif
3846 }
3847 }
3848 n->set_req(AddPNode::Base, nn);
3849 n->set_req(AddPNode::Address, nn);
3850 if (addp->outcnt() == 0) {
3851 addp->disconnect_inputs(this);
3852 }
3853 }
3854 }
3855 }
3856 #endif
3857 break;
3858 }
3859
3860 case Op_CastPP: {
3861 // Remove CastPP nodes to gain more freedom during scheduling but
3862 // keep the dependency they encode as control or precedence edges
3863 // (if control is set already) on memory operations. Some CastPP
3864 // nodes don't have a control (don't carry a dependency): skip
3865 // those.
3866 if (n->in(0) != nullptr) {
3867 ResourceMark rm;
3868 Unique_Node_List wq;
3869 wq.push(n);
3870 for (uint next = 0; next < wq.size(); ++next) {
3871 Node *m = wq.at(next);
3872 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3873 Node* use = m->fast_out(i);
3874 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3875 use->ensure_control_or_add_prec(n->in(0));
3876 } else {
3877 switch(use->Opcode()) {
3878 case Op_AddP:
3879 case Op_DecodeN:
3880 case Op_DecodeNKlass:
3881 case Op_CheckCastPP:
3882 case Op_CastPP:
3883 wq.push(use);
3884 break;
3885 }
3886 }
3887 }
3888 }
3889 }
3890 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3891 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3892 Node* in1 = n->in(1);
3893 const Type* t = n->bottom_type();
3894 Node* new_in1 = in1->clone();
3895 new_in1->as_DecodeN()->set_type(t);
3896
3897 if (!Matcher::narrow_oop_use_complex_address()) {
3898 //
3899 // x86, ARM and friends can handle 2 adds in addressing mode
3900 // and Matcher can fold a DecodeN node into address by using
3901 // a narrow oop directly and do implicit null check in address:
3902 //
3903 // [R12 + narrow_oop_reg<<3 + offset]
3904 // NullCheck narrow_oop_reg
3905 //
3906 // On other platforms (Sparc) we have to keep new DecodeN node and
3907 // use it to do implicit null check in address:
3908 //
3909 // decode_not_null narrow_oop_reg, base_reg
3910 // [base_reg + offset]
3911 // NullCheck base_reg
3912 //
3913 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3914 // to keep the information to which null check the new DecodeN node
3915 // corresponds to use it as value in implicit_null_check().
3916 //
3917 new_in1->set_req(0, n->in(0));
3918 }
3919
3920 n->subsume_by(new_in1, this);
3921 if (in1->outcnt() == 0) {
3922 in1->disconnect_inputs(this);
3923 }
3924 } else {
3925 n->subsume_by(n->in(1), this);
3926 if (n->outcnt() == 0) {
3927 n->disconnect_inputs(this);
3928 }
3929 }
3930 break;
3931 }
3932 #ifdef _LP64
3933 case Op_CmpP:
3934 // Do this transformation here to preserve CmpPNode::sub() and
3935 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3936 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3937 Node* in1 = n->in(1);
3938 Node* in2 = n->in(2);
3939 if (!in1->is_DecodeNarrowPtr()) {
3940 in2 = in1;
3941 in1 = n->in(2);
3942 }
3943 assert(in1->is_DecodeNarrowPtr(), "sanity");
3944
3945 Node* new_in2 = nullptr;
3946 if (in2->is_DecodeNarrowPtr()) {
3947 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3948 new_in2 = in2->in(1);
3949 } else if (in2->Opcode() == Op_ConP) {
3950 const Type* t = in2->bottom_type();
3951 if (t == TypePtr::NULL_PTR) {
3952 assert(in1->is_DecodeN(), "compare klass to null?");
3953 // Don't convert CmpP null check into CmpN if compressed
3954 // oops implicit null check is not generated.
3955 // This will allow to generate normal oop implicit null check.
3956 if (Matcher::gen_narrow_oop_implicit_null_checks())
3957 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3958 //
3959 // This transformation together with CastPP transformation above
3960 // will generated code for implicit null checks for compressed oops.
3961 //
3962 // The original code after Optimize()
3963 //
3964 // LoadN memory, narrow_oop_reg
3965 // decode narrow_oop_reg, base_reg
3966 // CmpP base_reg, nullptr
3967 // CastPP base_reg // NotNull
3968 // Load [base_reg + offset], val_reg
3969 //
3970 // after these transformations will be
3971 //
3972 // LoadN memory, narrow_oop_reg
3973 // CmpN narrow_oop_reg, nullptr
3974 // decode_not_null narrow_oop_reg, base_reg
3975 // Load [base_reg + offset], val_reg
3976 //
3977 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3978 // since narrow oops can be used in debug info now (see the code in
3979 // final_graph_reshaping_walk()).
3980 //
3981 // At the end the code will be matched to
3982 // on x86:
3983 //
3984 // Load_narrow_oop memory, narrow_oop_reg
3985 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3986 // NullCheck narrow_oop_reg
3987 //
3988 // and on sparc:
3989 //
3990 // Load_narrow_oop memory, narrow_oop_reg
3991 // decode_not_null narrow_oop_reg, base_reg
3992 // Load [base_reg + offset], val_reg
3993 // NullCheck base_reg
3994 //
3995 } else if (t->isa_oopptr()) {
3996 new_in2 = ConNode::make(t->make_narrowoop());
3997 } else if (t->isa_klassptr()) {
3998 new_in2 = ConNode::make(t->make_narrowklass());
3999 }
4000 }
4001 if (new_in2 != nullptr) {
4002 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
4003 n->subsume_by(cmpN, this);
4004 if (in1->outcnt() == 0) {
4005 in1->disconnect_inputs(this);
4006 }
4007 if (in2->outcnt() == 0) {
4008 in2->disconnect_inputs(this);
4009 }
4010 }
4011 }
4012 break;
4013
4014 case Op_DecodeN:
4015 case Op_DecodeNKlass:
4016 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4017 // DecodeN could be pinned when it can't be fold into
4018 // an address expression, see the code for Op_CastPP above.
4019 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4020 break;
4021
4022 case Op_EncodeP:
4023 case Op_EncodePKlass: {
4024 Node* in1 = n->in(1);
4025 if (in1->is_DecodeNarrowPtr()) {
4026 n->subsume_by(in1->in(1), this);
4027 } else if (in1->Opcode() == Op_ConP) {
4028 const Type* t = in1->bottom_type();
4029 if (t == TypePtr::NULL_PTR) {
4030 assert(t->isa_oopptr(), "null klass?");
4031 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4032 } else if (t->isa_oopptr()) {
4033 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4034 } else if (t->isa_klassptr()) {
4035 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4036 }
4037 }
4038 if (in1->outcnt() == 0) {
4039 in1->disconnect_inputs(this);
4040 }
4041 break;
4042 }
4043
4044 case Op_Proj: {
4045 if (OptimizeStringConcat || IncrementalInline) {
4046 ProjNode* proj = n->as_Proj();
4047 if (proj->_is_io_use) {
4048 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4049 // Separate projections were used for the exception path which
4050 // are normally removed by a late inline. If it wasn't inlined
4051 // then they will hang around and should just be replaced with
4052 // the original one. Merge them.
4053 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4054 if (non_io_proj != nullptr) {
4055 proj->subsume_by(non_io_proj , this);
4056 }
4057 }
4058 }
4059 break;
4060 }
4061
4062 case Op_Phi:
4063 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4064 // The EncodeP optimization may create Phi with the same edges
4065 // for all paths. It is not handled well by Register Allocator.
4066 Node* unique_in = n->in(1);
4067 assert(unique_in != nullptr, "");
4068 uint cnt = n->req();
4069 for (uint i = 2; i < cnt; i++) {
4070 Node* m = n->in(i);
4071 assert(m != nullptr, "");
4072 if (unique_in != m)
4073 unique_in = nullptr;
4074 }
4075 if (unique_in != nullptr) {
4076 n->subsume_by(unique_in, this);
4077 }
4078 }
4079 break;
4080
4081 #endif
4082
4083 #ifdef ASSERT
4084 case Op_CastII:
4085 // Verify that all range check dependent CastII nodes were removed.
4086 if (n->isa_CastII()->has_range_check()) {
4087 n->dump(3);
4088 assert(false, "Range check dependent CastII node was not removed");
4089 }
4090 break;
4091 #endif
4092
4093 case Op_ModI:
4094 if (UseDivMod) {
4095 // Check if a%b and a/b both exist
4096 Node* d = n->find_similar(Op_DivI);
4097 if (d) {
4098 // Replace them with a fused divmod if supported
4099 if (Matcher::has_match_rule(Op_DivModI)) {
4100 DivModINode* divmod = DivModINode::make(n);
4101 d->subsume_by(divmod->div_proj(), this);
4102 n->subsume_by(divmod->mod_proj(), this);
4103 } else {
4104 // replace a%b with a-((a/b)*b)
4105 Node* mult = new MulINode(d, d->in(2));
4106 Node* sub = new SubINode(d->in(1), mult);
4107 n->subsume_by(sub, this);
4108 }
4109 }
4110 }
4111 break;
4112
4113 case Op_ModL:
4114 if (UseDivMod) {
4115 // Check if a%b and a/b both exist
4116 Node* d = n->find_similar(Op_DivL);
4117 if (d) {
4118 // Replace them with a fused divmod if supported
4119 if (Matcher::has_match_rule(Op_DivModL)) {
4120 DivModLNode* divmod = DivModLNode::make(n);
4121 d->subsume_by(divmod->div_proj(), this);
4122 n->subsume_by(divmod->mod_proj(), this);
4123 } else {
4124 // replace a%b with a-((a/b)*b)
4125 Node* mult = new MulLNode(d, d->in(2));
4126 Node* sub = new SubLNode(d->in(1), mult);
4127 n->subsume_by(sub, this);
4128 }
4129 }
4130 }
4131 break;
4132
4133 case Op_UModI:
4134 if (UseDivMod) {
4135 // Check if a%b and a/b both exist
4136 Node* d = n->find_similar(Op_UDivI);
4137 if (d) {
4138 // Replace them with a fused unsigned divmod if supported
4139 if (Matcher::has_match_rule(Op_UDivModI)) {
4140 UDivModINode* divmod = UDivModINode::make(n);
4141 d->subsume_by(divmod->div_proj(), this);
4142 n->subsume_by(divmod->mod_proj(), this);
4143 } else {
4144 // replace a%b with a-((a/b)*b)
4145 Node* mult = new MulINode(d, d->in(2));
4146 Node* sub = new SubINode(d->in(1), mult);
4147 n->subsume_by(sub, this);
4148 }
4149 }
4150 }
4151 break;
4152
4153 case Op_UModL:
4154 if (UseDivMod) {
4155 // Check if a%b and a/b both exist
4156 Node* d = n->find_similar(Op_UDivL);
4157 if (d) {
4158 // Replace them with a fused unsigned divmod if supported
4159 if (Matcher::has_match_rule(Op_UDivModL)) {
4160 UDivModLNode* divmod = UDivModLNode::make(n);
4161 d->subsume_by(divmod->div_proj(), this);
4162 n->subsume_by(divmod->mod_proj(), this);
4163 } else {
4164 // replace a%b with a-((a/b)*b)
4165 Node* mult = new MulLNode(d, d->in(2));
4166 Node* sub = new SubLNode(d->in(1), mult);
4167 n->subsume_by(sub, this);
4168 }
4169 }
4170 }
4171 break;
4172
4173 case Op_LoadVector:
4174 case Op_StoreVector:
4175 case Op_LoadVectorGather:
4176 case Op_StoreVectorScatter:
4177 case Op_LoadVectorGatherMasked:
4178 case Op_StoreVectorScatterMasked:
4179 case Op_VectorCmpMasked:
4180 case Op_VectorMaskGen:
4181 case Op_LoadVectorMasked:
4182 case Op_StoreVectorMasked:
4183 break;
4184
4185 case Op_AddReductionVI:
4186 case Op_AddReductionVL:
4187 case Op_AddReductionVF:
4188 case Op_AddReductionVD:
4189 case Op_MulReductionVI:
4190 case Op_MulReductionVL:
4191 case Op_MulReductionVF:
4192 case Op_MulReductionVD:
4193 case Op_MinReductionV:
4194 case Op_MaxReductionV:
4195 case Op_AndReductionV:
4196 case Op_OrReductionV:
4197 case Op_XorReductionV:
4198 break;
4199
4200 case Op_PackB:
4201 case Op_PackS:
4202 case Op_PackI:
4203 case Op_PackF:
4204 case Op_PackL:
4205 case Op_PackD:
4206 if (n->req()-1 > 2) {
4207 // Replace many operand PackNodes with a binary tree for matching
4208 PackNode* p = (PackNode*) n;
4209 Node* btp = p->binary_tree_pack(1, n->req());
4210 n->subsume_by(btp, this);
4211 }
4212 break;
4213 case Op_Loop:
4214 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4215 case Op_CountedLoop:
4216 case Op_LongCountedLoop:
4217 case Op_OuterStripMinedLoop:
4218 if (n->as_Loop()->is_inner_loop()) {
4219 frc.inc_inner_loop_count();
4220 }
4221 n->as_Loop()->verify_strip_mined(0);
4222 break;
4223 case Op_LShiftI:
4224 case Op_RShiftI:
4225 case Op_URShiftI:
4226 case Op_LShiftL:
4227 case Op_RShiftL:
4228 case Op_URShiftL:
4229 if (Matcher::need_masked_shift_count) {
4230 // The cpu's shift instructions don't restrict the count to the
4231 // lower 5/6 bits. We need to do the masking ourselves.
4232 Node* in2 = n->in(2);
4233 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4234 const TypeInt* t = in2->find_int_type();
4235 if (t != nullptr && t->is_con()) {
4236 juint shift = t->get_con();
4237 if (shift > mask) { // Unsigned cmp
4238 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4239 }
4240 } else {
4241 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
4242 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4243 n->set_req(2, shift);
4244 }
4245 }
4246 if (in2->outcnt() == 0) { // Remove dead node
4247 in2->disconnect_inputs(this);
4248 }
4249 }
4250 break;
4251 case Op_MemBarStoreStore:
4252 case Op_MemBarRelease:
4253 // Break the link with AllocateNode: it is no longer useful and
4254 // confuses register allocation.
4255 if (n->req() > MemBarNode::Precedent) {
4256 n->set_req(MemBarNode::Precedent, top());
4257 }
4258 break;
4259 case Op_MemBarAcquire: {
4260 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4261 // At parse time, the trailing MemBarAcquire for a volatile load
4262 // is created with an edge to the load. After optimizations,
4263 // that input may be a chain of Phis. If those phis have no
4264 // other use, then the MemBarAcquire keeps them alive and
4265 // register allocation can be confused.
4266 dead_nodes.push(n->in(MemBarNode::Precedent));
4267 n->set_req(MemBarNode::Precedent, top());
4268 }
4269 break;
4270 }
4271 case Op_Blackhole:
4272 break;
4273 case Op_RangeCheck: {
4274 RangeCheckNode* rc = n->as_RangeCheck();
4275 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4276 n->subsume_by(iff, this);
4277 frc._tests.push(iff);
4278 break;
4279 }
4280 case Op_ConvI2L: {
4281 if (!Matcher::convi2l_type_required) {
4282 // Code generation on some platforms doesn't need accurate
4283 // ConvI2L types. Widening the type can help remove redundant
4284 // address computations.
4285 n->as_Type()->set_type(TypeLong::INT);
4286 ResourceMark rm;
4287 Unique_Node_List wq;
4288 wq.push(n);
4289 for (uint next = 0; next < wq.size(); next++) {
4290 Node *m = wq.at(next);
4291
4292 for(;;) {
4293 // Loop over all nodes with identical inputs edges as m
4294 Node* k = m->find_similar(m->Opcode());
4295 if (k == nullptr) {
4296 break;
4297 }
4298 // Push their uses so we get a chance to remove node made
4299 // redundant
4300 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4301 Node* u = k->fast_out(i);
4302 if (u->Opcode() == Op_LShiftL ||
4303 u->Opcode() == Op_AddL ||
4304 u->Opcode() == Op_SubL ||
4305 u->Opcode() == Op_AddP) {
4306 wq.push(u);
4307 }
4308 }
4309 // Replace all nodes with identical edges as m with m
4310 k->subsume_by(m, this);
4311 }
4312 }
4313 }
4314 break;
4315 }
4316 case Op_CmpUL: {
4317 if (!Matcher::has_match_rule(Op_CmpUL)) {
4318 // No support for unsigned long comparisons
4319 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4320 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4321 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4322 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4323 Node* andl = new AndLNode(orl, remove_sign_mask);
4324 Node* cmp = new CmpLNode(andl, n->in(2));
4325 n->subsume_by(cmp, this);
4326 }
4327 break;
4328 }
4329 #ifdef ASSERT
4330 case Op_InlineType: {
4331 n->dump(-1);
4332 assert(false, "inline type node was not removed");
4333 break;
4334 }
4335 #endif
4336 default:
4337 assert(!n->is_Call(), "");
4338 assert(!n->is_Mem(), "");
4339 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4340 break;
4341 }
4342 }
4343
4344 //------------------------------final_graph_reshaping_walk---------------------
4345 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4346 // requires that the walk visits a node's inputs before visiting the node.
4347 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4348 Unique_Node_List sfpt;
4349
4350 frc._visited.set(root->_idx); // first, mark node as visited
4351 uint cnt = root->req();
4352 Node *n = root;
4353 uint i = 0;
4354 while (true) {
4355 if (i < cnt) {
4356 // Place all non-visited non-null inputs onto stack
4357 Node* m = n->in(i);
4358 ++i;
4359 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
4360 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
4361 // compute worst case interpreter size in case of a deoptimization
4362 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4363
4364 sfpt.push(m);
4365 }
4366 cnt = m->req();
4367 nstack.push(n, i); // put on stack parent and next input's index
4368 n = m;
4369 i = 0;
4370 }
4371 } else {
4372 // Now do post-visit work
4373 final_graph_reshaping_impl(n, frc, dead_nodes);
4374 if (nstack.is_empty())
4375 break; // finished
4376 n = nstack.node(); // Get node from stack
4377 cnt = n->req();
4378 i = nstack.index();
4379 nstack.pop(); // Shift to the next node on stack
4380 }
4381 }
4382
4383 // Skip next transformation if compressed oops are not used.
4384 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4385 (!UseCompressedOops && !UseCompressedClassPointers))
4386 return;
4387
4388 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4389 // It could be done for an uncommon traps or any safepoints/calls
4390 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4391 while (sfpt.size() > 0) {
4392 n = sfpt.pop();
4393 JVMState *jvms = n->as_SafePoint()->jvms();
4394 assert(jvms != nullptr, "sanity");
4395 int start = jvms->debug_start();
4396 int end = n->req();
4397 bool is_uncommon = (n->is_CallStaticJava() &&
4398 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4399 for (int j = start; j < end; j++) {
4400 Node* in = n->in(j);
4401 if (in->is_DecodeNarrowPtr()) {
4402 bool safe_to_skip = true;
4403 if (!is_uncommon ) {
4404 // Is it safe to skip?
4405 for (uint i = 0; i < in->outcnt(); i++) {
4406 Node* u = in->raw_out(i);
4407 if (!u->is_SafePoint() ||
4408 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4409 safe_to_skip = false;
4410 }
4411 }
4412 }
4413 if (safe_to_skip) {
4414 n->set_req(j, in->in(1));
4415 }
4416 if (in->outcnt() == 0) {
4417 in->disconnect_inputs(this);
4418 }
4419 }
4420 }
4421 }
4422 }
4423
4424 //------------------------------final_graph_reshaping--------------------------
4425 // Final Graph Reshaping.
4426 //
4427 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4428 // and not commoned up and forced early. Must come after regular
4429 // optimizations to avoid GVN undoing the cloning. Clone constant
4430 // inputs to Loop Phis; these will be split by the allocator anyways.
4431 // Remove Opaque nodes.
4432 // (2) Move last-uses by commutative operations to the left input to encourage
4433 // Intel update-in-place two-address operations and better register usage
4434 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4435 // calls canonicalizing them back.
4436 // (3) Count the number of double-precision FP ops, single-precision FP ops
4437 // and call sites. On Intel, we can get correct rounding either by
4438 // forcing singles to memory (requires extra stores and loads after each
4439 // FP bytecode) or we can set a rounding mode bit (requires setting and
4440 // clearing the mode bit around call sites). The mode bit is only used
4441 // if the relative frequency of single FP ops to calls is low enough.
4442 // This is a key transform for SPEC mpeg_audio.
4443 // (4) Detect infinite loops; blobs of code reachable from above but not
4444 // below. Several of the Code_Gen algorithms fail on such code shapes,
4445 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4446 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4447 // Detection is by looking for IfNodes where only 1 projection is
4448 // reachable from below or CatchNodes missing some targets.
4449 // (5) Assert for insane oop offsets in debug mode.
4450
4451 bool Compile::final_graph_reshaping() {
4452 // an infinite loop may have been eliminated by the optimizer,
4453 // in which case the graph will be empty.
4454 if (root()->req() == 1) {
4455 // Do not compile method that is only a trivial infinite loop,
4456 // since the content of the loop may have been eliminated.
4457 record_method_not_compilable("trivial infinite loop");
4458 return true;
4459 }
4460
4461 // Expensive nodes have their control input set to prevent the GVN
4462 // from freely commoning them. There's no GVN beyond this point so
4463 // no need to keep the control input. We want the expensive nodes to
4464 // be freely moved to the least frequent code path by gcm.
4465 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4466 for (int i = 0; i < expensive_count(); i++) {
4467 _expensive_nodes.at(i)->set_req(0, nullptr);
4468 }
4469
4470 Final_Reshape_Counts frc;
4471
4472 // Visit everybody reachable!
4473 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4474 Node_Stack nstack(live_nodes() >> 1);
4475 Unique_Node_List dead_nodes;
4476 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4477
4478 // Check for unreachable (from below) code (i.e., infinite loops).
4479 for( uint i = 0; i < frc._tests.size(); i++ ) {
4480 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4481 // Get number of CFG targets.
4482 // Note that PCTables include exception targets after calls.
4483 uint required_outcnt = n->required_outcnt();
4484 if (n->outcnt() != required_outcnt) {
4485 // Check for a few special cases. Rethrow Nodes never take the
4486 // 'fall-thru' path, so expected kids is 1 less.
4487 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4488 if (n->in(0)->in(0)->is_Call()) {
4489 CallNode* call = n->in(0)->in(0)->as_Call();
4490 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4491 required_outcnt--; // Rethrow always has 1 less kid
4492 } else if (call->req() > TypeFunc::Parms &&
4493 call->is_CallDynamicJava()) {
4494 // Check for null receiver. In such case, the optimizer has
4495 // detected that the virtual call will always result in a null
4496 // pointer exception. The fall-through projection of this CatchNode
4497 // will not be populated.
4498 Node* arg0 = call->in(TypeFunc::Parms);
4499 if (arg0->is_Type() &&
4500 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4501 required_outcnt--;
4502 }
4503 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4504 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4505 // Check for illegal array length. In such case, the optimizer has
4506 // detected that the allocation attempt will always result in an
4507 // exception. There is no fall-through projection of this CatchNode .
4508 assert(call->is_CallStaticJava(), "static call expected");
4509 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4510 uint valid_length_test_input = call->req() - 1;
4511 Node* valid_length_test = call->in(valid_length_test_input);
4512 call->del_req(valid_length_test_input);
4513 if (valid_length_test->find_int_con(1) == 0) {
4514 required_outcnt--;
4515 }
4516 dead_nodes.push(valid_length_test);
4517 assert(n->outcnt() == required_outcnt, "malformed control flow");
4518 continue;
4519 }
4520 }
4521 }
4522
4523 // Recheck with a better notion of 'required_outcnt'
4524 if (n->outcnt() != required_outcnt) {
4525 DEBUG_ONLY( n->dump_bfs(1, 0, "-"); );
4526 assert(false, "malformed control flow");
4527 record_method_not_compilable("malformed control flow");
4528 return true; // Not all targets reachable!
4529 }
4530 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4531 CallNode* call = n->in(0)->in(0)->as_Call();
4532 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4533 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4534 assert(call->is_CallStaticJava(), "static call expected");
4535 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4536 uint valid_length_test_input = call->req() - 1;
4537 dead_nodes.push(call->in(valid_length_test_input));
4538 call->del_req(valid_length_test_input); // valid length test useless now
4539 }
4540 }
4541 // Check that I actually visited all kids. Unreached kids
4542 // must be infinite loops.
4543 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4544 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4545 record_method_not_compilable("infinite loop");
4546 return true; // Found unvisited kid; must be unreach
4547 }
4548
4549 // Here so verification code in final_graph_reshaping_walk()
4550 // always see an OuterStripMinedLoopEnd
4551 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4552 IfNode* init_iff = n->as_If();
4553 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4554 n->subsume_by(iff, this);
4555 }
4556 }
4557
4558 while (dead_nodes.size() > 0) {
4559 Node* m = dead_nodes.pop();
4560 if (m->outcnt() == 0 && m != top()) {
4561 for (uint j = 0; j < m->req(); j++) {
4562 Node* in = m->in(j);
4563 if (in != nullptr) {
4564 dead_nodes.push(in);
4565 }
4566 }
4567 m->disconnect_inputs(this);
4568 }
4569 }
4570
4571 #ifdef IA32
4572 // If original bytecodes contained a mixture of floats and doubles
4573 // check if the optimizer has made it homogeneous, item (3).
4574 if (UseSSE == 0 &&
4575 frc.get_float_count() > 32 &&
4576 frc.get_double_count() == 0 &&
4577 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4578 set_24_bit_selection_and_mode(false, true);
4579 }
4580 #endif // IA32
4581
4582 set_java_calls(frc.get_java_call_count());
4583 set_inner_loops(frc.get_inner_loop_count());
4584
4585 // No infinite loops, no reason to bail out.
4586 return false;
4587 }
4588
4589 //-----------------------------too_many_traps----------------------------------
4590 // Report if there are too many traps at the current method and bci.
4591 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4592 bool Compile::too_many_traps(ciMethod* method,
4593 int bci,
4594 Deoptimization::DeoptReason reason) {
4595 ciMethodData* md = method->method_data();
4596 if (md->is_empty()) {
4597 // Assume the trap has not occurred, or that it occurred only
4598 // because of a transient condition during start-up in the interpreter.
4599 return false;
4600 }
4601 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4602 if (md->has_trap_at(bci, m, reason) != 0) {
4603 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4604 // Also, if there are multiple reasons, or if there is no per-BCI record,
4605 // assume the worst.
4606 if (log())
4607 log()->elem("observe trap='%s' count='%d'",
4608 Deoptimization::trap_reason_name(reason),
4609 md->trap_count(reason));
4610 return true;
4611 } else {
4612 // Ignore method/bci and see if there have been too many globally.
4613 return too_many_traps(reason, md);
4614 }
4615 }
4616
4617 // Less-accurate variant which does not require a method and bci.
4618 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4619 ciMethodData* logmd) {
4620 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4621 // Too many traps globally.
4622 // Note that we use cumulative trap_count, not just md->trap_count.
4623 if (log()) {
4624 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4625 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4626 Deoptimization::trap_reason_name(reason),
4627 mcount, trap_count(reason));
4628 }
4629 return true;
4630 } else {
4631 // The coast is clear.
4632 return false;
4633 }
4634 }
4635
4636 //--------------------------too_many_recompiles--------------------------------
4637 // Report if there are too many recompiles at the current method and bci.
4638 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4639 // Is not eager to return true, since this will cause the compiler to use
4640 // Action_none for a trap point, to avoid too many recompilations.
4641 bool Compile::too_many_recompiles(ciMethod* method,
4642 int bci,
4643 Deoptimization::DeoptReason reason) {
4644 ciMethodData* md = method->method_data();
4645 if (md->is_empty()) {
4646 // Assume the trap has not occurred, or that it occurred only
4647 // because of a transient condition during start-up in the interpreter.
4648 return false;
4649 }
4650 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4651 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4652 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4653 Deoptimization::DeoptReason per_bc_reason
4654 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4655 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4656 if ((per_bc_reason == Deoptimization::Reason_none
4657 || md->has_trap_at(bci, m, reason) != 0)
4658 // The trap frequency measure we care about is the recompile count:
4659 && md->trap_recompiled_at(bci, m)
4660 && md->overflow_recompile_count() >= bc_cutoff) {
4661 // Do not emit a trap here if it has already caused recompilations.
4662 // Also, if there are multiple reasons, or if there is no per-BCI record,
4663 // assume the worst.
4664 if (log())
4665 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4666 Deoptimization::trap_reason_name(reason),
4667 md->trap_count(reason),
4668 md->overflow_recompile_count());
4669 return true;
4670 } else if (trap_count(reason) != 0
4671 && decompile_count() >= m_cutoff) {
4672 // Too many recompiles globally, and we have seen this sort of trap.
4673 // Use cumulative decompile_count, not just md->decompile_count.
4674 if (log())
4675 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4676 Deoptimization::trap_reason_name(reason),
4677 md->trap_count(reason), trap_count(reason),
4678 md->decompile_count(), decompile_count());
4679 return true;
4680 } else {
4681 // The coast is clear.
4682 return false;
4683 }
4684 }
4685
4686 // Compute when not to trap. Used by matching trap based nodes and
4687 // NullCheck optimization.
4688 void Compile::set_allowed_deopt_reasons() {
4689 _allowed_reasons = 0;
4690 if (is_method_compilation()) {
4691 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4692 assert(rs < BitsPerInt, "recode bit map");
4693 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4694 _allowed_reasons |= nth_bit(rs);
4695 }
4696 }
4697 }
4698 }
4699
4700 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4701 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4702 }
4703
4704 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4705 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4706 }
4707
4708 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4709 if (holder->is_initialized()) {
4710 return false;
4711 }
4712 if (holder->is_being_initialized()) {
4713 if (accessing_method->holder() == holder) {
4714 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4715 // <init>, or a static method. In all those cases, there was an initialization
4716 // barrier on the holder klass passed.
4717 if (accessing_method->is_class_initializer() ||
4718 accessing_method->is_object_constructor() ||
4719 accessing_method->is_static()) {
4720 return false;
4721 }
4722 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4723 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4724 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4725 // child class can become fully initialized while its parent class is still being initialized.
4726 if (accessing_method->is_class_initializer()) {
4727 return false;
4728 }
4729 }
4730 ciMethod* root = method(); // the root method of compilation
4731 if (root != accessing_method) {
4732 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4733 }
4734 }
4735 return true;
4736 }
4737
4738 #ifndef PRODUCT
4739 //------------------------------verify_bidirectional_edges---------------------
4740 // For each input edge to a node (ie - for each Use-Def edge), verify that
4741 // there is a corresponding Def-Use edge.
4742 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4743 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4744 uint stack_size = live_nodes() >> 4;
4745 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4746 nstack.push(_root);
4747
4748 while (nstack.size() > 0) {
4749 Node* n = nstack.pop();
4750 if (visited.member(n)) {
4751 continue;
4752 }
4753 visited.push(n);
4754
4755 // Walk over all input edges, checking for correspondence
4756 uint length = n->len();
4757 for (uint i = 0; i < length; i++) {
4758 Node* in = n->in(i);
4759 if (in != nullptr && !visited.member(in)) {
4760 nstack.push(in); // Put it on stack
4761 }
4762 if (in != nullptr && !in->is_top()) {
4763 // Count instances of `next`
4764 int cnt = 0;
4765 for (uint idx = 0; idx < in->_outcnt; idx++) {
4766 if (in->_out[idx] == n) {
4767 cnt++;
4768 }
4769 }
4770 assert(cnt > 0, "Failed to find Def-Use edge.");
4771 // Check for duplicate edges
4772 // walk the input array downcounting the input edges to n
4773 for (uint j = 0; j < length; j++) {
4774 if (n->in(j) == in) {
4775 cnt--;
4776 }
4777 }
4778 assert(cnt == 0, "Mismatched edge count.");
4779 } else if (in == nullptr) {
4780 assert(i == 0 || i >= n->req() ||
4781 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4782 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
4783 (n->is_Unlock() && i == (n->req() - 1)) ||
4784 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4785 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
4786 } else {
4787 assert(in->is_top(), "sanity");
4788 // Nothing to check.
4789 }
4790 }
4791 }
4792 }
4793
4794 //------------------------------verify_graph_edges---------------------------
4795 // Walk the Graph and verify that there is a one-to-one correspondence
4796 // between Use-Def edges and Def-Use edges in the graph.
4797 void Compile::verify_graph_edges(bool no_dead_code) {
4798 if (VerifyGraphEdges) {
4799 Unique_Node_List visited;
4800
4801 // Call graph walk to check edges
4802 verify_bidirectional_edges(visited);
4803 if (no_dead_code) {
4804 // Now make sure that no visited node is used by an unvisited node.
4805 bool dead_nodes = false;
4806 Unique_Node_List checked;
4807 while (visited.size() > 0) {
4808 Node* n = visited.pop();
4809 checked.push(n);
4810 for (uint i = 0; i < n->outcnt(); i++) {
4811 Node* use = n->raw_out(i);
4812 if (checked.member(use)) continue; // already checked
4813 if (visited.member(use)) continue; // already in the graph
4814 if (use->is_Con()) continue; // a dead ConNode is OK
4815 // At this point, we have found a dead node which is DU-reachable.
4816 if (!dead_nodes) {
4817 tty->print_cr("*** Dead nodes reachable via DU edges:");
4818 dead_nodes = true;
4819 }
4820 use->dump(2);
4821 tty->print_cr("---");
4822 checked.push(use); // No repeats; pretend it is now checked.
4823 }
4824 }
4825 assert(!dead_nodes, "using nodes must be reachable from root");
4826 }
4827 }
4828 }
4829 #endif
4830
4831 // The Compile object keeps track of failure reasons separately from the ciEnv.
4832 // This is required because there is not quite a 1-1 relation between the
4833 // ciEnv and its compilation task and the Compile object. Note that one
4834 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4835 // to backtrack and retry without subsuming loads. Other than this backtracking
4836 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4837 // by the logic in C2Compiler.
4838 void Compile::record_failure(const char* reason) {
4839 if (log() != nullptr) {
4840 log()->elem("failure reason='%s' phase='compile'", reason);
4841 }
4842 if (_failure_reason == nullptr) {
4843 // Record the first failure reason.
4844 _failure_reason = reason;
4845 }
4846
4847 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4848 C->print_method(PHASE_FAILURE, 1);
4849 }
4850 _root = nullptr; // flush the graph, too
4851 }
4852
4853 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4854 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4855 _compile(nullptr), _log(nullptr), _phase_name(name), _dolog(CITimeVerbose)
4856 {
4857 if (_dolog) {
4858 _compile = Compile::current();
4859 _log = _compile->log();
4860 }
4861 if (_log != nullptr) {
4862 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4863 _log->stamp();
4864 _log->end_head();
4865 }
4866 }
4867
4868 Compile::TracePhase::~TracePhase() {
4869 #ifdef ASSERT
4870 if (PrintIdealNodeCount) {
4871 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4872 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4873 }
4874
4875 if (VerifyIdealNodeCount) {
4876 Compile::current()->print_missing_nodes();
4877 }
4878 #endif
4879
4880 if (_log != nullptr) {
4881 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4882 }
4883 }
4884
4885 //----------------------------static_subtype_check-----------------------------
4886 // Shortcut important common cases when superklass is exact:
4887 // (0) superklass is java.lang.Object (can occur in reflective code)
4888 // (1) subklass is already limited to a subtype of superklass => always ok
4889 // (2) subklass does not overlap with superklass => always fail
4890 // (3) superklass has NO subtypes and we can check with a simple compare.
4891 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4892 if (skip) {
4893 return SSC_full_test; // Let caller generate the general case.
4894 }
4895
4896 if (subk->is_java_subtype_of(superk)) {
4897 return SSC_always_true; // (0) and (1) this test cannot fail
4898 }
4899
4900 if (!subk->maybe_java_subtype_of(superk)) {
4901 return SSC_always_false; // (2) true path dead; no dynamic test needed
4902 }
4903
4904 const Type* superelem = superk;
4905 if (superk->isa_aryklassptr()) {
4906 int ignored;
4907 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4908
4909 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays.
4910 // [QMyValue is a subtype of [LMyValue but the klass for [QMyValue is not equal to
4911 // the klass for [LMyValue. Perform a full test.
4912 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
4913 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
4914 return SSC_full_test;
4915 }
4916 }
4917
4918 if (superelem->isa_instklassptr()) {
4919 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4920 if (!ik->has_subklass()) {
4921 if (!ik->is_final()) {
4922 // Add a dependency if there is a chance of a later subclass.
4923 dependencies()->assert_leaf_type(ik);
4924 }
4925 if (!superk->maybe_java_subtype_of(subk)) {
4926 return SSC_always_false;
4927 }
4928 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4929 }
4930 } else {
4931 // A primitive array type has no subtypes.
4932 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4933 }
4934
4935 return SSC_full_test;
4936 }
4937
4938 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4939 #ifdef _LP64
4940 // The scaled index operand to AddP must be a clean 64-bit value.
4941 // Java allows a 32-bit int to be incremented to a negative
4942 // value, which appears in a 64-bit register as a large
4943 // positive number. Using that large positive number as an
4944 // operand in pointer arithmetic has bad consequences.
4945 // On the other hand, 32-bit overflow is rare, and the possibility
4946 // can often be excluded, if we annotate the ConvI2L node with
4947 // a type assertion that its value is known to be a small positive
4948 // number. (The prior range check has ensured this.)
4949 // This assertion is used by ConvI2LNode::Ideal.
4950 int index_max = max_jint - 1; // array size is max_jint, index is one less
4951 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4952 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4953 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4954 #endif
4955 return idx;
4956 }
4957
4958 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4959 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4960 if (ctrl != nullptr) {
4961 // Express control dependency by a CastII node with a narrow type.
4962 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4963 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4964 // node from floating above the range check during loop optimizations. Otherwise, the
4965 // ConvI2L node may be eliminated independently of the range check, causing the data path
4966 // to become TOP while the control path is still there (although it's unreachable).
4967 value->set_req(0, ctrl);
4968 value = phase->transform(value);
4969 }
4970 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4971 return phase->transform(new ConvI2LNode(value, ltype));
4972 }
4973
4974 // The message about the current inlining is accumulated in
4975 // _print_inlining_stream and transferred into the _print_inlining_list
4976 // once we know whether inlining succeeds or not. For regular
4977 // inlining, messages are appended to the buffer pointed by
4978 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4979 // a new buffer is added after _print_inlining_idx in the list. This
4980 // way we can update the inlining message for late inlining call site
4981 // when the inlining is attempted again.
4982 void Compile::print_inlining_init() {
4983 if (print_inlining() || print_intrinsics()) {
4984 // print_inlining_init is actually called several times.
4985 print_inlining_reset();
4986 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4987 }
4988 }
4989
4990 void Compile::print_inlining_reinit() {
4991 if (print_inlining() || print_intrinsics()) {
4992 print_inlining_reset();
4993 }
4994 }
4995
4996 void Compile::print_inlining_reset() {
4997 _print_inlining_stream->reset();
4998 }
4999
5000 void Compile::print_inlining_commit() {
5001 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
5002 // Transfer the message from _print_inlining_stream to the current
5003 // _print_inlining_list buffer and clear _print_inlining_stream.
5004 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
5005 print_inlining_reset();
5006 }
5007
5008 void Compile::print_inlining_push() {
5009 // Add new buffer to the _print_inlining_list at current position
5010 _print_inlining_idx++;
5011 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
5012 }
5013
5014 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
5015 return _print_inlining_list->at(_print_inlining_idx);
5016 }
5017
5018 void Compile::print_inlining_update(CallGenerator* cg) {
5019 if (print_inlining() || print_intrinsics()) {
5020 if (cg->is_late_inline()) {
5021 if (print_inlining_current()->cg() != cg &&
5022 (print_inlining_current()->cg() != nullptr ||
5023 print_inlining_current()->ss()->size() != 0)) {
5024 print_inlining_push();
5025 }
5026 print_inlining_commit();
5027 print_inlining_current()->set_cg(cg);
5028 } else {
5029 if (print_inlining_current()->cg() != nullptr) {
5030 print_inlining_push();
5031 }
5032 print_inlining_commit();
5033 }
5034 }
5035 }
5036
5037 void Compile::print_inlining_move_to(CallGenerator* cg) {
5038 // We resume inlining at a late inlining call site. Locate the
5039 // corresponding inlining buffer so that we can update it.
5040 if (print_inlining() || print_intrinsics()) {
5041 for (int i = 0; i < _print_inlining_list->length(); i++) {
5042 if (_print_inlining_list->at(i)->cg() == cg) {
5043 _print_inlining_idx = i;
5044 return;
5045 }
5046 }
5047 ShouldNotReachHere();
5048 }
5049 }
5050
5051 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
5052 if (print_inlining() || print_intrinsics()) {
5053 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
5054 assert(print_inlining_current()->cg() == cg, "wrong entry");
5055 // replace message with new message
5056 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
5057 print_inlining_commit();
5058 print_inlining_current()->set_cg(cg);
5059 }
5060 }
5061
5062 void Compile::print_inlining_assert_ready() {
5063 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
5064 }
5065
5066 void Compile::process_print_inlining() {
5067 assert(_late_inlines.length() == 0, "not drained yet");
5068 if (print_inlining() || print_intrinsics()) {
5069 ResourceMark rm;
5070 stringStream ss;
5071 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
5072 for (int i = 0; i < _print_inlining_list->length(); i++) {
5073 PrintInliningBuffer* pib = _print_inlining_list->at(i);
5074 ss.print("%s", pib->ss()->freeze());
5075 delete pib;
5076 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
5077 }
5078 // Reset _print_inlining_list, it only contains destructed objects.
5079 // It is on the arena, so it will be freed when the arena is reset.
5080 _print_inlining_list = nullptr;
5081 // _print_inlining_stream won't be used anymore, either.
5082 print_inlining_reset();
5083 size_t end = ss.size();
5084 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
5085 strncpy(_print_inlining_output, ss.freeze(), end+1);
5086 _print_inlining_output[end] = 0;
5087 }
5088 }
5089
5090 void Compile::dump_print_inlining() {
5091 if (_print_inlining_output != nullptr) {
5092 tty->print_raw(_print_inlining_output);
5093 }
5094 }
5095
5096 void Compile::log_late_inline(CallGenerator* cg) {
5097 if (log() != nullptr) {
5098 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5099 cg->unique_id());
5100 JVMState* p = cg->call_node()->jvms();
5101 while (p != nullptr) {
5102 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5103 p = p->caller();
5104 }
5105 log()->tail("late_inline");
5106 }
5107 }
5108
5109 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5110 log_late_inline(cg);
5111 if (log() != nullptr) {
5112 log()->inline_fail(msg);
5113 }
5114 }
5115
5116 void Compile::log_inline_id(CallGenerator* cg) {
5117 if (log() != nullptr) {
5118 // The LogCompilation tool needs a unique way to identify late
5119 // inline call sites. This id must be unique for this call site in
5120 // this compilation. Try to have it unique across compilations as
5121 // well because it can be convenient when grepping through the log
5122 // file.
5123 // Distinguish OSR compilations from others in case CICountOSR is
5124 // on.
5125 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5126 cg->set_unique_id(id);
5127 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5128 }
5129 }
5130
5131 void Compile::log_inline_failure(const char* msg) {
5132 if (C->log() != nullptr) {
5133 C->log()->inline_fail(msg);
5134 }
5135 }
5136
5137
5138 // Dump inlining replay data to the stream.
5139 // Don't change thread state and acquire any locks.
5140 void Compile::dump_inline_data(outputStream* out) {
5141 InlineTree* inl_tree = ilt();
5142 if (inl_tree != nullptr) {
5143 out->print(" inline %d", inl_tree->count());
5144 inl_tree->dump_replay_data(out);
5145 }
5146 }
5147
5148 void Compile::dump_inline_data_reduced(outputStream* out) {
5149 assert(ReplayReduce, "");
5150
5151 InlineTree* inl_tree = ilt();
5152 if (inl_tree == nullptr) {
5153 return;
5154 }
5155 // Enable iterative replay file reduction
5156 // Output "compile" lines for depth 1 subtrees,
5157 // simulating that those trees were compiled
5158 // instead of inlined.
5159 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5160 InlineTree* sub = inl_tree->subtrees().at(i);
5161 if (sub->inline_level() != 1) {
5162 continue;
5163 }
5164
5165 ciMethod* method = sub->method();
5166 int entry_bci = -1;
5167 int comp_level = env()->task()->comp_level();
5168 out->print("compile ");
5169 method->dump_name_as_ascii(out);
5170 out->print(" %d %d", entry_bci, comp_level);
5171 out->print(" inline %d", sub->count());
5172 sub->dump_replay_data(out, -1);
5173 out->cr();
5174 }
5175 }
5176
5177 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5178 if (n1->Opcode() < n2->Opcode()) return -1;
5179 else if (n1->Opcode() > n2->Opcode()) return 1;
5180
5181 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5182 for (uint i = 1; i < n1->req(); i++) {
5183 if (n1->in(i) < n2->in(i)) return -1;
5184 else if (n1->in(i) > n2->in(i)) return 1;
5185 }
5186
5187 return 0;
5188 }
5189
5190 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5191 Node* n1 = *n1p;
5192 Node* n2 = *n2p;
5193
5194 return cmp_expensive_nodes(n1, n2);
5195 }
5196
5197 void Compile::sort_expensive_nodes() {
5198 if (!expensive_nodes_sorted()) {
5199 _expensive_nodes.sort(cmp_expensive_nodes);
5200 }
5201 }
5202
5203 bool Compile::expensive_nodes_sorted() const {
5204 for (int i = 1; i < _expensive_nodes.length(); i++) {
5205 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5206 return false;
5207 }
5208 }
5209 return true;
5210 }
5211
5212 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5213 if (_expensive_nodes.length() == 0) {
5214 return false;
5215 }
5216
5217 assert(OptimizeExpensiveOps, "optimization off?");
5218
5219 // Take this opportunity to remove dead nodes from the list
5220 int j = 0;
5221 for (int i = 0; i < _expensive_nodes.length(); i++) {
5222 Node* n = _expensive_nodes.at(i);
5223 if (!n->is_unreachable(igvn)) {
5224 assert(n->is_expensive(), "should be expensive");
5225 _expensive_nodes.at_put(j, n);
5226 j++;
5227 }
5228 }
5229 _expensive_nodes.trunc_to(j);
5230
5231 // Then sort the list so that similar nodes are next to each other
5232 // and check for at least two nodes of identical kind with same data
5233 // inputs.
5234 sort_expensive_nodes();
5235
5236 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5237 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5238 return true;
5239 }
5240 }
5241
5242 return false;
5243 }
5244
5245 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5246 if (_expensive_nodes.length() == 0) {
5247 return;
5248 }
5249
5250 assert(OptimizeExpensiveOps, "optimization off?");
5251
5252 // Sort to bring similar nodes next to each other and clear the
5253 // control input of nodes for which there's only a single copy.
5254 sort_expensive_nodes();
5255
5256 int j = 0;
5257 int identical = 0;
5258 int i = 0;
5259 bool modified = false;
5260 for (; i < _expensive_nodes.length()-1; i++) {
5261 assert(j <= i, "can't write beyond current index");
5262 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5263 identical++;
5264 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5265 continue;
5266 }
5267 if (identical > 0) {
5268 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5269 identical = 0;
5270 } else {
5271 Node* n = _expensive_nodes.at(i);
5272 igvn.replace_input_of(n, 0, nullptr);
5273 igvn.hash_insert(n);
5274 modified = true;
5275 }
5276 }
5277 if (identical > 0) {
5278 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5279 } else if (_expensive_nodes.length() >= 1) {
5280 Node* n = _expensive_nodes.at(i);
5281 igvn.replace_input_of(n, 0, nullptr);
5282 igvn.hash_insert(n);
5283 modified = true;
5284 }
5285 _expensive_nodes.trunc_to(j);
5286 if (modified) {
5287 igvn.optimize();
5288 }
5289 }
5290
5291 void Compile::add_expensive_node(Node * n) {
5292 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5293 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5294 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5295 if (OptimizeExpensiveOps) {
5296 _expensive_nodes.append(n);
5297 } else {
5298 // Clear control input and let IGVN optimize expensive nodes if
5299 // OptimizeExpensiveOps is off.
5300 n->set_req(0, nullptr);
5301 }
5302 }
5303
5304 /**
5305 * Track coarsened Lock and Unlock nodes.
5306 */
5307
5308 class Lock_List : public Node_List {
5309 uint _origin_cnt;
5310 public:
5311 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5312 uint origin_cnt() const { return _origin_cnt; }
5313 };
5314
5315 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5316 int length = locks.length();
5317 if (length > 0) {
5318 // Have to keep this list until locks elimination during Macro nodes elimination.
5319 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5320 for (int i = 0; i < length; i++) {
5321 AbstractLockNode* lock = locks.at(i);
5322 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5323 locks_list->push(lock);
5324 }
5325 _coarsened_locks.append(locks_list);
5326 }
5327 }
5328
5329 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5330 int count = coarsened_count();
5331 for (int i = 0; i < count; i++) {
5332 Node_List* locks_list = _coarsened_locks.at(i);
5333 for (uint j = 0; j < locks_list->size(); j++) {
5334 Node* lock = locks_list->at(j);
5335 assert(lock->is_AbstractLock(), "sanity");
5336 if (!useful.member(lock)) {
5337 locks_list->yank(lock);
5338 }
5339 }
5340 }
5341 }
5342
5343 void Compile::remove_coarsened_lock(Node* n) {
5344 if (n->is_AbstractLock()) {
5345 int count = coarsened_count();
5346 for (int i = 0; i < count; i++) {
5347 Node_List* locks_list = _coarsened_locks.at(i);
5348 locks_list->yank(n);
5349 }
5350 }
5351 }
5352
5353 bool Compile::coarsened_locks_consistent() {
5354 int count = coarsened_count();
5355 for (int i = 0; i < count; i++) {
5356 bool unbalanced = false;
5357 bool modified = false; // track locks kind modifications
5358 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5359 uint size = locks_list->size();
5360 if (size == 0) {
5361 unbalanced = false; // All locks were eliminated - good
5362 } else if (size != locks_list->origin_cnt()) {
5363 unbalanced = true; // Some locks were removed from list
5364 } else {
5365 for (uint j = 0; j < size; j++) {
5366 Node* lock = locks_list->at(j);
5367 // All nodes in group should have the same state (modified or not)
5368 if (!lock->as_AbstractLock()->is_coarsened()) {
5369 if (j == 0) {
5370 // first on list was modified, the rest should be too for consistency
5371 modified = true;
5372 } else if (!modified) {
5373 // this lock was modified but previous locks on the list were not
5374 unbalanced = true;
5375 break;
5376 }
5377 } else if (modified) {
5378 // previous locks on list were modified but not this lock
5379 unbalanced = true;
5380 break;
5381 }
5382 }
5383 }
5384 if (unbalanced) {
5385 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5386 #ifdef ASSERT
5387 if (PrintEliminateLocks) {
5388 tty->print_cr("=== unbalanced coarsened locks ===");
5389 for (uint l = 0; l < size; l++) {
5390 locks_list->at(l)->dump();
5391 }
5392 }
5393 #endif
5394 record_failure(C2Compiler::retry_no_locks_coarsening());
5395 return false;
5396 }
5397 }
5398 return true;
5399 }
5400
5401 /**
5402 * Remove the speculative part of types and clean up the graph
5403 */
5404 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5405 if (UseTypeSpeculation) {
5406 Unique_Node_List worklist;
5407 worklist.push(root());
5408 int modified = 0;
5409 // Go over all type nodes that carry a speculative type, drop the
5410 // speculative part of the type and enqueue the node for an igvn
5411 // which may optimize it out.
5412 for (uint next = 0; next < worklist.size(); ++next) {
5413 Node *n = worklist.at(next);
5414 if (n->is_Type()) {
5415 TypeNode* tn = n->as_Type();
5416 const Type* t = tn->type();
5417 const Type* t_no_spec = t->remove_speculative();
5418 if (t_no_spec != t) {
5419 bool in_hash = igvn.hash_delete(n);
5420 assert(in_hash, "node should be in igvn hash table");
5421 tn->set_type(t_no_spec);
5422 igvn.hash_insert(n);
5423 igvn._worklist.push(n); // give it a chance to go away
5424 modified++;
5425 }
5426 }
5427 // Iterate over outs - endless loops is unreachable from below
5428 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5429 Node *m = n->fast_out(i);
5430 if (not_a_node(m)) {
5431 continue;
5432 }
5433 worklist.push(m);
5434 }
5435 }
5436 // Drop the speculative part of all types in the igvn's type table
5437 igvn.remove_speculative_types();
5438 if (modified > 0) {
5439 igvn.optimize();
5440 }
5441 #ifdef ASSERT
5442 // Verify that after the IGVN is over no speculative type has resurfaced
5443 worklist.clear();
5444 worklist.push(root());
5445 for (uint next = 0; next < worklist.size(); ++next) {
5446 Node *n = worklist.at(next);
5447 const Type* t = igvn.type_or_null(n);
5448 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5449 if (n->is_Type()) {
5450 t = n->as_Type()->type();
5451 assert(t == t->remove_speculative(), "no more speculative types");
5452 }
5453 // Iterate over outs - endless loops is unreachable from below
5454 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5455 Node *m = n->fast_out(i);
5456 if (not_a_node(m)) {
5457 continue;
5458 }
5459 worklist.push(m);
5460 }
5461 }
5462 igvn.check_no_speculative_types();
5463 #endif
5464 }
5465 }
5466
5467 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5468 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5469 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5470 if (!EnableValhalla || ta == nullptr || tb == nullptr ||
5471 ta->is_zero_type() || tb->is_zero_type() ||
5472 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5473 // Use old acmp if one operand is null or not an inline type
5474 return new CmpPNode(a, b);
5475 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5476 // We know that one operand is an inline type. Therefore,
5477 // new acmp will only return true if both operands are nullptr.
5478 // Check if both operands are null by or'ing the oops.
5479 a = phase->transform(new CastP2XNode(nullptr, a));
5480 b = phase->transform(new CastP2XNode(nullptr, b));
5481 a = phase->transform(new OrXNode(a, b));
5482 return new CmpXNode(a, phase->MakeConX(0));
5483 }
5484 // Use new acmp
5485 return nullptr;
5486 }
5487
5488 // Auxiliary methods to support randomized stressing/fuzzing.
5489
5490 int Compile::random() {
5491 _stress_seed = os::next_random(_stress_seed);
5492 return static_cast<int>(_stress_seed);
5493 }
5494
5495 // This method can be called the arbitrary number of times, with current count
5496 // as the argument. The logic allows selecting a single candidate from the
5497 // running list of candidates as follows:
5498 // int count = 0;
5499 // Cand* selected = null;
5500 // while(cand = cand->next()) {
5501 // if (randomized_select(++count)) {
5502 // selected = cand;
5503 // }
5504 // }
5505 //
5506 // Including count equalizes the chances any candidate is "selected".
5507 // This is useful when we don't have the complete list of candidates to choose
5508 // from uniformly. In this case, we need to adjust the randomicity of the
5509 // selection, or else we will end up biasing the selection towards the latter
5510 // candidates.
5511 //
5512 // Quick back-envelope calculation shows that for the list of n candidates
5513 // the equal probability for the candidate to persist as "best" can be
5514 // achieved by replacing it with "next" k-th candidate with the probability
5515 // of 1/k. It can be easily shown that by the end of the run, the
5516 // probability for any candidate is converged to 1/n, thus giving the
5517 // uniform distribution among all the candidates.
5518 //
5519 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5520 #define RANDOMIZED_DOMAIN_POW 29
5521 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5522 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5523 bool Compile::randomized_select(int count) {
5524 assert(count > 0, "only positive");
5525 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5526 }
5527
5528 CloneMap& Compile::clone_map() { return _clone_map; }
5529 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5530
5531 void NodeCloneInfo::dump_on(outputStream* st) const {
5532 st->print(" {%d:%d} ", idx(), gen());
5533 }
5534
5535 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5536 uint64_t val = value(old->_idx);
5537 NodeCloneInfo cio(val);
5538 assert(val != 0, "old node should be in the map");
5539 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5540 insert(nnn->_idx, cin.get());
5541 #ifndef PRODUCT
5542 if (is_debug()) {
5543 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5544 }
5545 #endif
5546 }
5547
5548 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5549 NodeCloneInfo cio(value(old->_idx));
5550 if (cio.get() == 0) {
5551 cio.set(old->_idx, 0);
5552 insert(old->_idx, cio.get());
5553 #ifndef PRODUCT
5554 if (is_debug()) {
5555 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5556 }
5557 #endif
5558 }
5559 clone(old, nnn, gen);
5560 }
5561
5562 int CloneMap::max_gen() const {
5563 int g = 0;
5564 DictI di(_dict);
5565 for(; di.test(); ++di) {
5566 int t = gen(di._key);
5567 if (g < t) {
5568 g = t;
5569 #ifndef PRODUCT
5570 if (is_debug()) {
5571 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5572 }
5573 #endif
5574 }
5575 }
5576 return g;
5577 }
5578
5579 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5580 uint64_t val = value(key);
5581 if (val != 0) {
5582 NodeCloneInfo ni(val);
5583 ni.dump_on(st);
5584 }
5585 }
5586
5587 // Move Allocate nodes to the start of the list
5588 void Compile::sort_macro_nodes() {
5589 int count = macro_count();
5590 int allocates = 0;
5591 for (int i = 0; i < count; i++) {
5592 Node* n = macro_node(i);
5593 if (n->is_Allocate()) {
5594 if (i != allocates) {
5595 Node* tmp = macro_node(allocates);
5596 _macro_nodes.at_put(allocates, n);
5597 _macro_nodes.at_put(i, tmp);
5598 }
5599 allocates++;
5600 }
5601 }
5602 }
5603
5604 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5605 EventCompilerPhase event;
5606 if (event.should_commit()) {
5607 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5608 }
5609 #ifndef PRODUCT
5610 ResourceMark rm;
5611 stringStream ss;
5612 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5613 if (n != nullptr) {
5614 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5615 }
5616
5617 const char* name = ss.as_string();
5618 if (should_print_igv(level)) {
5619 _igv_printer->print_method(name, level);
5620 }
5621 if (should_print_phase(cpt)) {
5622 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5623 }
5624 #endif
5625 C->_latest_stage_start_counter.stamp();
5626 }
5627
5628 // Only used from CompileWrapper
5629 void Compile::begin_method() {
5630 #ifndef PRODUCT
5631 if (_method != nullptr && should_print_igv(1)) {
5632 _igv_printer->begin_method();
5633 }
5634 #endif
5635 C->_latest_stage_start_counter.stamp();
5636 }
5637
5638 // Only used from CompileWrapper
5639 void Compile::end_method() {
5640 EventCompilerPhase event;
5641 if (event.should_commit()) {
5642 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5643 }
5644
5645 #ifndef PRODUCT
5646 if (_method != nullptr && should_print_igv(1)) {
5647 _igv_printer->end_method();
5648 }
5649 #endif
5650 }
5651
5652 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5653 #ifndef PRODUCT
5654 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) {
5655 return true;
5656 }
5657 #endif
5658 return false;
5659 }
5660
5661 bool Compile::should_print_igv(int level) {
5662 #ifndef PRODUCT
5663 if (PrintIdealGraphLevel < 0) { // disabled by the user
5664 return false;
5665 }
5666
5667 bool need = directive()->IGVPrintLevelOption >= level;
5668 if (need && !_igv_printer) {
5669 _igv_printer = IdealGraphPrinter::printer();
5670 _igv_printer->set_compile(this);
5671 }
5672 return need;
5673 #else
5674 return false;
5675 #endif
5676 }
5677
5678 #ifndef PRODUCT
5679 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5680 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5681
5682 // Called from debugger. Prints method to the default file with the default phase name.
5683 // This works regardless of any Ideal Graph Visualizer flags set or not.
5684 void igv_print() {
5685 Compile::current()->igv_print_method_to_file();
5686 }
5687
5688 // Same as igv_print() above but with a specified phase name.
5689 void igv_print(const char* phase_name) {
5690 Compile::current()->igv_print_method_to_file(phase_name);
5691 }
5692
5693 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5694 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5695 // This works regardless of any Ideal Graph Visualizer flags set or not.
5696 void igv_print(bool network) {
5697 if (network) {
5698 Compile::current()->igv_print_method_to_network();
5699 } else {
5700 Compile::current()->igv_print_method_to_file();
5701 }
5702 }
5703
5704 // Same as igv_print(bool network) above but with a specified phase name.
5705 void igv_print(bool network, const char* phase_name) {
5706 if (network) {
5707 Compile::current()->igv_print_method_to_network(phase_name);
5708 } else {
5709 Compile::current()->igv_print_method_to_file(phase_name);
5710 }
5711 }
5712
5713 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5714 void igv_print_default() {
5715 Compile::current()->print_method(PHASE_DEBUG, 0);
5716 }
5717
5718 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5719 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5720 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5721 void igv_append() {
5722 Compile::current()->igv_print_method_to_file("Debug", true);
5723 }
5724
5725 // Same as igv_append() above but with a specified phase name.
5726 void igv_append(const char* phase_name) {
5727 Compile::current()->igv_print_method_to_file(phase_name, true);
5728 }
5729
5730 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5731 const char* file_name = "custom_debug.xml";
5732 if (_debug_file_printer == nullptr) {
5733 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5734 } else {
5735 _debug_file_printer->update_compiled_method(C->method());
5736 }
5737 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5738 _debug_file_printer->print(phase_name, (Node*)C->root());
5739 }
5740
5741 void Compile::igv_print_method_to_network(const char* phase_name) {
5742 if (_debug_network_printer == nullptr) {
5743 _debug_network_printer = new IdealGraphPrinter(C);
5744 } else {
5745 _debug_network_printer->update_compiled_method(C->method());
5746 }
5747 tty->print_cr("Method printed over network stream to IGV");
5748 _debug_network_printer->print(phase_name, (Node*)C->root());
5749 }
5750 #endif
5751
5752 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5753 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5754 return value;
5755 }
5756 Node* result = nullptr;
5757 if (bt == T_BYTE) {
5758 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5759 result = new RShiftINode(result, phase->intcon(24));
5760 } else if (bt == T_BOOLEAN) {
5761 result = new AndINode(value, phase->intcon(0xFF));
5762 } else if (bt == T_CHAR) {
5763 result = new AndINode(value,phase->intcon(0xFFFF));
5764 } else {
5765 assert(bt == T_SHORT, "unexpected narrow type");
5766 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5767 result = new RShiftINode(result, phase->intcon(16));
5768 }
5769 if (transform_res) {
5770 result = phase->transform(result);
5771 }
5772 return result;
5773 }
5774