1 /*
2 * Copyright (c) 1997, 2024, 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/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compilerOracle.hpp"
37 #include "compiler/compiler_globals.hpp"
38 #include "compiler/disassembler.hpp"
39 #include "compiler/oopMap.hpp"
40 #include "gc/shared/barrierSet.hpp"
41 #include "gc/shared/c2/barrierSetC2.hpp"
42 #include "jfr/jfrEvents.hpp"
43 #include "jvm_io.h"
44 #include "memory/allocation.hpp"
45 #include "memory/resourceArea.hpp"
46 #include "opto/addnode.hpp"
47 #include "opto/block.hpp"
48 #include "opto/c2compiler.hpp"
49 #include "opto/callGenerator.hpp"
50 #include "opto/callnode.hpp"
51 #include "opto/castnode.hpp"
52 #include "opto/cfgnode.hpp"
53 #include "opto/chaitin.hpp"
54 #include "opto/compile.hpp"
55 #include "opto/connode.hpp"
56 #include "opto/convertnode.hpp"
57 #include "opto/divnode.hpp"
58 #include "opto/escape.hpp"
59 #include "opto/idealGraphPrinter.hpp"
60 #include "opto/inlinetypenode.hpp"
61 #include "opto/locknode.hpp"
62 #include "opto/loopnode.hpp"
63 #include "opto/machnode.hpp"
64 #include "opto/macro.hpp"
65 #include "opto/matcher.hpp"
66 #include "opto/mathexactnode.hpp"
67 #include "opto/memnode.hpp"
68 #include "opto/mulnode.hpp"
69 #include "opto/narrowptrnode.hpp"
70 #include "opto/node.hpp"
71 #include "opto/opcodes.hpp"
72 #include "opto/output.hpp"
73 #include "opto/parse.hpp"
74 #include "opto/phaseX.hpp"
75 #include "opto/rootnode.hpp"
76 #include "opto/runtime.hpp"
77 #include "opto/stringopts.hpp"
78 #include "opto/type.hpp"
79 #include "opto/vector.hpp"
80 #include "opto/vectornode.hpp"
81 #include "runtime/globals_extension.hpp"
82 #include "runtime/sharedRuntime.hpp"
83 #include "runtime/signature.hpp"
84 #include "runtime/stubRoutines.hpp"
85 #include "runtime/timer.hpp"
86 #include "utilities/align.hpp"
87 #include "utilities/copy.hpp"
88 #include "utilities/macros.hpp"
89 #include "utilities/resourceHash.hpp"
90
91 // -------------------- Compile::mach_constant_base_node -----------------------
92 // Constant table base node singleton.
93 MachConstantBaseNode* Compile::mach_constant_base_node() {
94 if (_mach_constant_base_node == nullptr) {
95 _mach_constant_base_node = new MachConstantBaseNode();
96 _mach_constant_base_node->add_req(C->root());
97 }
98 return _mach_constant_base_node;
99 }
100
101
102 /// Support for intrinsics.
103
104 // Return the index at which m must be inserted (or already exists).
105 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
106 class IntrinsicDescPair {
107 private:
108 ciMethod* _m;
109 bool _is_virtual;
110 public:
111 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
112 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
113 ciMethod* m= elt->method();
114 ciMethod* key_m = key->_m;
115 if (key_m < m) return -1;
116 else if (key_m > m) return 1;
117 else {
118 bool is_virtual = elt->is_virtual();
119 bool key_virtual = key->_is_virtual;
120 if (key_virtual < is_virtual) return -1;
121 else if (key_virtual > is_virtual) return 1;
122 else return 0;
123 }
124 }
125 };
126 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
127 #ifdef ASSERT
128 for (int i = 1; i < _intrinsics.length(); i++) {
129 CallGenerator* cg1 = _intrinsics.at(i-1);
130 CallGenerator* cg2 = _intrinsics.at(i);
131 assert(cg1->method() != cg2->method()
132 ? cg1->method() < cg2->method()
133 : cg1->is_virtual() < cg2->is_virtual(),
134 "compiler intrinsics list must stay sorted");
135 }
136 #endif
137 IntrinsicDescPair pair(m, is_virtual);
138 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
139 }
140
141 void Compile::register_intrinsic(CallGenerator* cg) {
142 bool found = false;
143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
144 assert(!found, "registering twice");
145 _intrinsics.insert_before(index, cg);
146 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
147 }
148
149 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
150 assert(m->is_loaded(), "don't try this on unloaded methods");
151 if (_intrinsics.length() > 0) {
152 bool found = false;
153 int index = intrinsic_insertion_index(m, is_virtual, found);
154 if (found) {
155 return _intrinsics.at(index);
156 }
157 }
158 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
159 if (m->intrinsic_id() != vmIntrinsics::_none &&
160 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
161 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
162 if (cg != nullptr) {
163 // Save it for next time:
164 register_intrinsic(cg);
165 return cg;
166 } else {
167 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
168 }
169 }
170 return nullptr;
171 }
172
173 // Compile::make_vm_intrinsic is defined in library_call.cpp.
174
175 #ifndef PRODUCT
176 // statistics gathering...
177
178 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
179 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
180
181 inline int as_int(vmIntrinsics::ID id) {
182 return vmIntrinsics::as_int(id);
183 }
184
185 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
186 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
187 int oflags = _intrinsic_hist_flags[as_int(id)];
188 assert(flags != 0, "what happened?");
189 if (is_virtual) {
190 flags |= _intrinsic_virtual;
191 }
192 bool changed = (flags != oflags);
193 if ((flags & _intrinsic_worked) != 0) {
194 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
195 if (count == 1) {
196 changed = true; // first time
197 }
198 // increment the overall count also:
199 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
200 }
201 if (changed) {
202 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
203 // Something changed about the intrinsic's virtuality.
204 if ((flags & _intrinsic_virtual) != 0) {
205 // This is the first use of this intrinsic as a virtual call.
206 if (oflags != 0) {
207 // We already saw it as a non-virtual, so note both cases.
208 flags |= _intrinsic_both;
209 }
210 } else if ((oflags & _intrinsic_both) == 0) {
211 // This is the first use of this intrinsic as a non-virtual
212 flags |= _intrinsic_both;
213 }
214 }
215 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
216 }
217 // update the overall flags also:
218 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
219 return changed;
220 }
221
222 static char* format_flags(int flags, char* buf) {
223 buf[0] = 0;
224 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
225 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
226 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
227 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
228 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
229 if (buf[0] == 0) strcat(buf, ",");
230 assert(buf[0] == ',', "must be");
231 return &buf[1];
232 }
233
234 void Compile::print_intrinsic_statistics() {
235 char flagsbuf[100];
236 ttyLocker ttyl;
237 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
238 tty->print_cr("Compiler intrinsic usage:");
239 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
240 if (total == 0) total = 1; // avoid div0 in case of no successes
241 #define PRINT_STAT_LINE(name, c, f) \
242 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
243 for (auto id : EnumRange<vmIntrinsicID>{}) {
244 int flags = _intrinsic_hist_flags[as_int(id)];
245 juint count = _intrinsic_hist_count[as_int(id)];
246 if ((flags | count) != 0) {
247 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
248 }
249 }
250 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
251 if (xtty != nullptr) xtty->tail("statistics");
252 }
253
254 void Compile::print_statistics() {
255 { ttyLocker ttyl;
256 if (xtty != nullptr) xtty->head("statistics type='opto'");
257 Parse::print_statistics();
258 PhaseStringOpts::print_statistics();
259 PhaseCCP::print_statistics();
260 PhaseRegAlloc::print_statistics();
261 PhaseOutput::print_statistics();
262 PhasePeephole::print_statistics();
263 PhaseIdealLoop::print_statistics();
264 ConnectionGraph::print_statistics();
265 PhaseMacroExpand::print_statistics();
266 if (xtty != nullptr) xtty->tail("statistics");
267 }
268 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
269 // put this under its own <statistics> element.
270 print_intrinsic_statistics();
271 }
272 }
273 #endif //PRODUCT
274
275 void Compile::gvn_replace_by(Node* n, Node* nn) {
276 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
277 Node* use = n->last_out(i);
278 bool is_in_table = initial_gvn()->hash_delete(use);
279 uint uses_found = 0;
280 for (uint j = 0; j < use->len(); j++) {
281 if (use->in(j) == n) {
282 if (j < use->req())
283 use->set_req(j, nn);
284 else
285 use->set_prec(j, nn);
286 uses_found++;
287 }
288 }
289 if (is_in_table) {
290 // reinsert into table
291 initial_gvn()->hash_find_insert(use);
292 }
293 record_for_igvn(use);
294 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
295 i -= uses_found; // we deleted 1 or more copies of this edge
296 }
297 }
298
299
300 // Identify all nodes that are reachable from below, useful.
301 // Use breadth-first pass that records state in a Unique_Node_List,
302 // recursive traversal is slower.
303 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
304 int estimated_worklist_size = live_nodes();
305 useful.map( estimated_worklist_size, nullptr ); // preallocate space
306
307 // Initialize worklist
308 if (root() != nullptr) { useful.push(root()); }
309 // If 'top' is cached, declare it useful to preserve cached node
310 if (cached_top_node()) { useful.push(cached_top_node()); }
311
312 // Push all useful nodes onto the list, breadthfirst
313 for( uint next = 0; next < useful.size(); ++next ) {
314 assert( next < unique(), "Unique useful nodes < total nodes");
315 Node *n = useful.at(next);
316 uint max = n->len();
317 for( uint i = 0; i < max; ++i ) {
318 Node *m = n->in(i);
319 if (not_a_node(m)) continue;
320 useful.push(m);
321 }
322 }
323 }
324
325 // Update dead_node_list with any missing dead nodes using useful
326 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
327 void Compile::update_dead_node_list(Unique_Node_List &useful) {
328 uint max_idx = unique();
329 VectorSet& useful_node_set = useful.member_set();
330
331 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
332 // If node with index node_idx is not in useful set,
333 // mark it as dead in dead node list.
334 if (!useful_node_set.test(node_idx)) {
335 record_dead_node(node_idx);
336 }
337 }
338 }
339
340 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
341 int shift = 0;
342 for (int i = 0; i < inlines->length(); i++) {
343 CallGenerator* cg = inlines->at(i);
344 if (useful.member(cg->call_node())) {
345 if (shift > 0) {
346 inlines->at_put(i - shift, cg);
347 }
348 } else {
349 shift++; // skip over the dead element
350 }
351 }
352 if (shift > 0) {
353 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
354 }
355 }
356
357 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
358 assert(dead != nullptr && dead->is_Call(), "sanity");
359 int found = 0;
360 for (int i = 0; i < inlines->length(); i++) {
361 if (inlines->at(i)->call_node() == dead) {
362 inlines->remove_at(i);
363 found++;
364 NOT_DEBUG( break; ) // elements are unique, so exit early
365 }
366 }
367 assert(found <= 1, "not unique");
368 }
369
370 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
371 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
372 for (int i = node_list.length() - 1; i >= 0; i--) {
373 N* node = node_list.at(i);
374 if (!useful.member(node)) {
375 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
376 }
377 }
378 }
379
380 void Compile::remove_useless_node(Node* dead) {
381 remove_modified_node(dead);
382
383 // Constant node that has no out-edges and has only one in-edge from
384 // root is usually dead. However, sometimes reshaping walk makes
385 // it reachable by adding use edges. So, we will NOT count Con nodes
386 // as dead to be conservative about the dead node count at any
387 // given time.
388 if (!dead->is_Con()) {
389 record_dead_node(dead->_idx);
390 }
391 if (dead->is_macro()) {
392 remove_macro_node(dead);
393 }
394 if (dead->is_expensive()) {
395 remove_expensive_node(dead);
396 }
397 if (dead->Opcode() == Op_Opaque4) {
398 remove_template_assertion_predicate_opaq(dead);
399 }
400 if (dead->is_ParsePredicate()) {
401 remove_parse_predicate(dead->as_ParsePredicate());
402 }
403 if (dead->for_post_loop_opts_igvn()) {
404 remove_from_post_loop_opts_igvn(dead);
405 }
406 if (dead->is_InlineType()) {
407 remove_inline_type(dead);
408 }
409 if (dead->is_Call()) {
410 remove_useless_late_inlines( &_late_inlines, dead);
411 remove_useless_late_inlines( &_string_late_inlines, dead);
412 remove_useless_late_inlines( &_boxing_late_inlines, dead);
413 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
414
415 if (dead->is_CallStaticJava()) {
416 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
417 }
418 }
419 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
420 bs->unregister_potential_barrier_node(dead);
421 }
422
423 // Disconnect all useless nodes by disconnecting those at the boundary.
424 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
425 uint next = 0;
426 while (next < useful.size()) {
427 Node *n = useful.at(next++);
428 if (n->is_SafePoint()) {
429 // We're done with a parsing phase. Replaced nodes are not valid
430 // beyond that point.
431 n->as_SafePoint()->delete_replaced_nodes();
432 }
433 // Use raw traversal of out edges since this code removes out edges
434 int max = n->outcnt();
435 for (int j = 0; j < max; ++j) {
436 Node* child = n->raw_out(j);
437 if (!useful.member(child)) {
438 assert(!child->is_top() || child != top(),
439 "If top is cached in Compile object it is in useful list");
440 // Only need to remove this out-edge to the useless node
441 n->raw_del_out(j);
442 --j;
443 --max;
444 }
445 }
446 if (n->outcnt() == 1 && n->has_special_unique_user()) {
447 assert(useful.member(n->unique_out()), "do not push a useless node");
448 worklist.push(n->unique_out());
449 }
450 if (n->outcnt() == 0) {
451 worklist.push(n);
452 }
453 }
454
455 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
456 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
457 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
458 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
459 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
460 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
461 #ifdef ASSERT
462 if (_modified_nodes != nullptr) {
463 _modified_nodes->remove_useless_nodes(useful.member_set());
464 }
465 #endif
466 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
467 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
468 #ifdef ASSERT
469 if (_modified_nodes != nullptr) {
470 _modified_nodes->remove_useless_nodes(useful.member_set());
471 }
472 #endif
473
474 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
475 bs->eliminate_useless_gc_barriers(useful, this);
476 // clean up the late inline lists
477 remove_useless_late_inlines( &_late_inlines, useful);
478 remove_useless_late_inlines( &_string_late_inlines, useful);
479 remove_useless_late_inlines( &_boxing_late_inlines, useful);
480 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
481 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
482 }
483
484 // ============================================================================
485 //------------------------------CompileWrapper---------------------------------
486 class CompileWrapper : public StackObj {
487 Compile *const _compile;
488 public:
489 CompileWrapper(Compile* compile);
490
491 ~CompileWrapper();
492 };
493
494 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
495 // the Compile* pointer is stored in the current ciEnv:
496 ciEnv* env = compile->env();
497 assert(env == ciEnv::current(), "must already be a ciEnv active");
498 assert(env->compiler_data() == nullptr, "compile already active?");
499 env->set_compiler_data(compile);
500 assert(compile == Compile::current(), "sanity");
501
502 compile->set_type_dict(nullptr);
503 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
504 compile->clone_map().set_clone_idx(0);
505 compile->set_type_last_size(0);
506 compile->set_last_tf(nullptr, nullptr);
507 compile->set_indexSet_arena(nullptr);
508 compile->set_indexSet_free_block_list(nullptr);
509 compile->init_type_arena();
510 Type::Initialize(compile);
511 _compile->begin_method();
512 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
513 }
514 CompileWrapper::~CompileWrapper() {
515 // simulate crash during compilation
516 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
517
518 _compile->end_method();
519 _compile->env()->set_compiler_data(nullptr);
520 }
521
522
523 //----------------------------print_compile_messages---------------------------
524 void Compile::print_compile_messages() {
525 #ifndef PRODUCT
526 // Check if recompiling
527 if (!subsume_loads() && PrintOpto) {
528 // Recompiling without allowing machine instructions to subsume loads
529 tty->print_cr("*********************************************************");
530 tty->print_cr("** Bailout: Recompile without subsuming loads **");
531 tty->print_cr("*********************************************************");
532 }
533 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
534 // Recompiling without escape analysis
535 tty->print_cr("*********************************************************");
536 tty->print_cr("** Bailout: Recompile without escape analysis **");
537 tty->print_cr("*********************************************************");
538 }
539 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
540 // Recompiling without iterative escape analysis
541 tty->print_cr("*********************************************************");
542 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
543 tty->print_cr("*********************************************************");
544 }
545 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
546 // Recompiling without reducing allocation merges
547 tty->print_cr("*********************************************************");
548 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
549 tty->print_cr("*********************************************************");
550 }
551 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
552 // Recompiling without boxing elimination
553 tty->print_cr("*********************************************************");
554 tty->print_cr("** Bailout: Recompile without boxing elimination **");
555 tty->print_cr("*********************************************************");
556 }
557 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
558 // Recompiling without locks coarsening
559 tty->print_cr("*********************************************************");
560 tty->print_cr("** Bailout: Recompile without locks coarsening **");
561 tty->print_cr("*********************************************************");
562 }
563 if (env()->break_at_compile()) {
564 // Open the debugger when compiling this method.
565 tty->print("### Breaking when compiling: ");
566 method()->print_short_name();
567 tty->cr();
568 BREAKPOINT;
569 }
570
571 if( PrintOpto ) {
572 if (is_osr_compilation()) {
573 tty->print("[OSR]%3d", _compile_id);
574 } else {
575 tty->print("%3d", _compile_id);
576 }
577 }
578 #endif
579 }
580
581 #ifndef PRODUCT
582 void Compile::print_ideal_ir(const char* phase_name) {
583 // keep the following output all in one block
584 // This output goes directly to the tty, not the compiler log.
585 // To enable tools to match it up with the compilation activity,
586 // be sure to tag this tty output with the compile ID.
587
588 // Node dumping can cause a safepoint, which can break the tty lock.
589 // Buffer all node dumps, so that all safepoints happen before we lock.
590 ResourceMark rm;
591 stringStream ss;
592
593 if (_output == nullptr) {
594 ss.print_cr("AFTER: %s", phase_name);
595 // Print out all nodes in ascending order of index.
596 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
597 } else {
598 // Dump the node blockwise if we have a scheduling
599 _output->print_scheduling(&ss);
600 }
601
602 // Check that the lock is not broken by a safepoint.
603 NoSafepointVerifier nsv;
604 ttyLocker ttyl;
605 if (xtty != nullptr) {
606 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
607 compile_id(),
608 is_osr_compilation() ? " compile_kind='osr'" : "",
609 phase_name);
610 }
611
612 tty->print("%s", ss.as_string());
613
614 if (xtty != nullptr) {
615 xtty->tail("ideal");
616 }
617 }
618 #endif
619
620 // ============================================================================
621 //------------------------------Compile standard-------------------------------
622
623 // Compile a method. entry_bci is -1 for normal compilations and indicates
624 // the continuation bci for on stack replacement.
625
626
627 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
628 Options options, DirectiveSet* directive)
629 : Phase(Compiler),
630 _compile_id(ci_env->compile_id()),
631 _options(options),
632 _method(target),
633 _entry_bci(osr_bci),
634 _ilt(nullptr),
635 _stub_function(nullptr),
636 _stub_name(nullptr),
637 _stub_entry_point(nullptr),
638 _max_node_limit(MaxNodeLimit),
639 _post_loop_opts_phase(false),
640 _allow_macro_nodes(true),
641 _inlining_progress(false),
642 _inlining_incrementally(false),
643 _do_cleanup(false),
644 _has_reserved_stack_access(target->has_reserved_stack_access()),
645 _has_circular_inline_type(false),
646 #ifndef PRODUCT
647 _igv_idx(0),
648 _trace_opto_output(directive->TraceOptoOutputOption),
649 #endif
650 _has_method_handle_invokes(false),
651 _clinit_barrier_on_entry(false),
652 _stress_seed(0),
653 _comp_arena(mtCompiler),
654 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
655 _env(ci_env),
656 _directive(directive),
657 _log(ci_env->log()),
658 _first_failure_details(nullptr),
659 _intrinsics (comp_arena(), 0, 0, nullptr),
660 _macro_nodes (comp_arena(), 8, 0, nullptr),
661 _parse_predicates (comp_arena(), 8, 0, nullptr),
662 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr),
663 _expensive_nodes (comp_arena(), 8, 0, nullptr),
664 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
665 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
666 _unstable_if_traps (comp_arena(), 8, 0, nullptr),
667 _coarsened_locks (comp_arena(), 8, 0, nullptr),
668 _congraph(nullptr),
669 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
670 _unique(0),
671 _dead_node_count(0),
672 _dead_node_list(comp_arena()),
673 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
674 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
675 _node_arena(&_node_arena_one),
676 _mach_constant_base_node(nullptr),
677 _Compile_types(mtCompiler),
678 _initial_gvn(nullptr),
679 _igvn_worklist(nullptr),
680 _types(nullptr),
681 _node_hash(nullptr),
682 _late_inlines(comp_arena(), 2, 0, nullptr),
683 _string_late_inlines(comp_arena(), 2, 0, nullptr),
684 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
685 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
686 _late_inlines_pos(0),
687 _number_of_mh_late_inlines(0),
688 _oom(false),
689 _print_inlining_stream(new (mtCompiler) stringStream()),
690 _print_inlining_list(nullptr),
691 _print_inlining_idx(0),
692 _print_inlining_output(nullptr),
693 _replay_inline_data(nullptr),
694 _java_calls(0),
695 _inner_loops(0),
696 _interpreter_frame_size(0),
697 _output(nullptr)
698 #ifndef PRODUCT
699 , _in_dump_cnt(0)
700 #endif
701 {
702 C = this;
703 CompileWrapper cw(this);
704
705 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
706 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
707
708 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
709 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
710 // We can always print a disassembly, either abstract (hex dump) or
711 // with the help of a suitable hsdis library. Thus, we should not
712 // couple print_assembly and print_opto_assembly controls.
713 // But: always print opto and regular assembly on compile command 'print'.
714 bool print_assembly = directive->PrintAssemblyOption;
715 set_print_assembly(print_opto_assembly || print_assembly);
716 #else
717 set_print_assembly(false); // must initialize.
718 #endif
719
720 #ifndef PRODUCT
721 set_parsed_irreducible_loop(false);
722 #endif
723
724 if (directive->ReplayInlineOption) {
725 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
726 }
727 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
728 set_print_intrinsics(directive->PrintIntrinsicsOption);
729 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
730
731 if (ProfileTraps) {
732 // Make sure the method being compiled gets its own MDO,
733 // so we can at least track the decompile_count().
734 method()->ensure_method_data();
735 }
736
737 Init(/*do_aliasing=*/ true);
738
739 print_compile_messages();
740
741 _ilt = InlineTree::build_inline_tree_root();
742
743 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
744 assert(num_alias_types() >= AliasIdxRaw, "");
745
746 #define MINIMUM_NODE_HASH 1023
747
748 // GVN that will be run immediately on new nodes
749 uint estimated_size = method()->code_size()*4+64;
750 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
751 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
752 _types = new (comp_arena()) Type_Array(comp_arena());
753 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
754 PhaseGVN gvn;
755 set_initial_gvn(&gvn);
756
757 print_inlining_init();
758 { // Scope for timing the parser
759 TracePhase tp("parse", &timers[_t_parser]);
760
761 // Put top into the hash table ASAP.
762 initial_gvn()->transform(top());
763
764 // Set up tf(), start(), and find a CallGenerator.
765 CallGenerator* cg = nullptr;
766 if (is_osr_compilation()) {
767 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
768 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
769 initial_gvn()->set_type_bottom(s);
770 init_start(s);
771 cg = CallGenerator::for_osr(method(), entry_bci());
772 } else {
773 // Normal case.
774 init_tf(TypeFunc::make(method()));
775 StartNode* s = new StartNode(root(), tf()->domain_cc());
776 initial_gvn()->set_type_bottom(s);
777 init_start(s);
778 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
779 // With java.lang.ref.reference.get() we must go through the
780 // intrinsic - even when get() is the root
781 // method of the compile - so that, if necessary, the value in
782 // the referent field of the reference object gets recorded by
783 // the pre-barrier code.
784 cg = find_intrinsic(method(), false);
785 }
786 if (cg == nullptr) {
787 float past_uses = method()->interpreter_invocation_count();
788 float expected_uses = past_uses;
789 cg = CallGenerator::for_inline(method(), expected_uses);
790 }
791 }
792 if (failing()) return;
793 if (cg == nullptr) {
794 const char* reason = InlineTree::check_can_parse(method());
795 assert(reason != nullptr, "expect reason for parse failure");
796 stringStream ss;
797 ss.print("cannot parse method: %s", reason);
798 record_method_not_compilable(ss.as_string());
799 return;
800 }
801
802 gvn.set_type(root(), root()->bottom_type());
803
804 JVMState* jvms = build_start_state(start(), tf());
805 if ((jvms = cg->generate(jvms)) == nullptr) {
806 assert(failure_reason() != nullptr, "expect reason for parse failure");
807 stringStream ss;
808 ss.print("method parse failed: %s", failure_reason());
809 record_method_not_compilable(ss.as_string());
810 return;
811 }
812 GraphKit kit(jvms);
813
814 if (!kit.stopped()) {
815 // Accept return values, and transfer control we know not where.
816 // This is done by a special, unique ReturnNode bound to root.
817 return_values(kit.jvms());
818 }
819
820 if (kit.has_exceptions()) {
821 // Any exceptions that escape from this call must be rethrown
822 // to whatever caller is dynamically above us on the stack.
823 // This is done by a special, unique RethrowNode bound to root.
824 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
825 }
826
827 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
828
829 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
830 inline_string_calls(true);
831 }
832
833 if (failing()) return;
834
835 // Remove clutter produced by parsing.
836 if (!failing()) {
837 ResourceMark rm;
838 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
839 }
840 }
841
842 // Note: Large methods are capped off in do_one_bytecode().
843 if (failing()) return;
844
845 // After parsing, node notes are no longer automagic.
846 // They must be propagated by register_new_node_with_optimizer(),
847 // clone(), or the like.
848 set_default_node_notes(nullptr);
849
850 #ifndef PRODUCT
851 if (should_print_igv(1)) {
852 _igv_printer->print_inlining();
853 }
854 #endif
855
856 if (failing()) return;
857 NOT_PRODUCT( verify_graph_edges(); )
858
859 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
860 StressIncrementalInlining || StressMacroExpansion) {
861 initialize_stress_seed(directive);
862 }
863
864 // Now optimize
865 Optimize();
866 if (failing()) return;
867 NOT_PRODUCT( verify_graph_edges(); )
868
869 #ifndef PRODUCT
870 if (should_print_ideal()) {
871 print_ideal_ir("print_ideal");
872 }
873 #endif
874
875 #ifdef ASSERT
876 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
877 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
878 #endif
879
880 // Dump compilation data to replay it.
881 if (directive->DumpReplayOption) {
882 env()->dump_replay_data(_compile_id);
883 }
884 if (directive->DumpInlineOption && (ilt() != nullptr)) {
885 env()->dump_inline_data(_compile_id);
886 }
887
888 // Now that we know the size of all the monitors we can add a fixed slot
889 // for the original deopt pc.
890 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
891 if (needs_stack_repair()) {
892 // One extra slot for the special stack increment value
893 next_slot += 2;
894 }
895 // TODO 8284443 Only reserve extra slot if needed
896 if (InlineTypeReturnedAsFields) {
897 // One extra slot to hold the IsInit information for a nullable
898 // inline type return if we run out of registers.
899 next_slot += 2;
900 }
901 set_fixed_slots(next_slot);
902
903 // Compute when to use implicit null checks. Used by matching trap based
904 // nodes and NullCheck optimization.
905 set_allowed_deopt_reasons();
906
907 // Now generate code
908 Code_Gen();
909 }
910
911 //------------------------------Compile----------------------------------------
912 // Compile a runtime stub
913 Compile::Compile( ciEnv* ci_env,
914 TypeFunc_generator generator,
915 address stub_function,
916 const char *stub_name,
917 int is_fancy_jump,
918 bool pass_tls,
919 bool return_pc,
920 DirectiveSet* directive)
921 : Phase(Compiler),
922 _compile_id(0),
923 _options(Options::for_runtime_stub()),
924 _method(nullptr),
925 _entry_bci(InvocationEntryBci),
926 _stub_function(stub_function),
927 _stub_name(stub_name),
928 _stub_entry_point(nullptr),
929 _max_node_limit(MaxNodeLimit),
930 _post_loop_opts_phase(false),
931 _allow_macro_nodes(true),
932 _inlining_progress(false),
933 _inlining_incrementally(false),
934 _has_reserved_stack_access(false),
935 _has_circular_inline_type(false),
936 #ifndef PRODUCT
937 _igv_idx(0),
938 _trace_opto_output(directive->TraceOptoOutputOption),
939 #endif
940 _has_method_handle_invokes(false),
941 _clinit_barrier_on_entry(false),
942 _stress_seed(0),
943 _comp_arena(mtCompiler),
944 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
945 _env(ci_env),
946 _directive(directive),
947 _log(ci_env->log()),
948 _first_failure_details(nullptr),
949 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
950 _congraph(nullptr),
951 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
952 _unique(0),
953 _dead_node_count(0),
954 _dead_node_list(comp_arena()),
955 _node_arena_one(mtCompiler),
956 _node_arena_two(mtCompiler),
957 _node_arena(&_node_arena_one),
958 _mach_constant_base_node(nullptr),
959 _Compile_types(mtCompiler),
960 _initial_gvn(nullptr),
961 _igvn_worklist(nullptr),
962 _types(nullptr),
963 _node_hash(nullptr),
964 _number_of_mh_late_inlines(0),
965 _oom(false),
966 _print_inlining_stream(new (mtCompiler) stringStream()),
967 _print_inlining_list(nullptr),
968 _print_inlining_idx(0),
969 _print_inlining_output(nullptr),
970 _replay_inline_data(nullptr),
971 _java_calls(0),
972 _inner_loops(0),
973 _interpreter_frame_size(0),
974 _output(nullptr),
975 #ifndef PRODUCT
976 _in_dump_cnt(0),
977 #endif
978 _allowed_reasons(0) {
979 C = this;
980
981 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
982 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
983
984 #ifndef PRODUCT
985 set_print_assembly(PrintFrameConverterAssembly);
986 set_parsed_irreducible_loop(false);
987 #else
988 set_print_assembly(false); // Must initialize.
989 #endif
990 set_has_irreducible_loop(false); // no loops
991
992 CompileWrapper cw(this);
993 Init(/*do_aliasing=*/ false);
994 init_tf((*generator)());
995
996 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
997 _types = new (comp_arena()) Type_Array(comp_arena());
998 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
999
1000 if (StressLCM || StressGCM) {
1001 initialize_stress_seed(directive);
1002 }
1003
1004 {
1005 PhaseGVN gvn;
1006 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1007 gvn.transform(top());
1008
1009 GraphKit kit;
1010 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1011 }
1012
1013 NOT_PRODUCT( verify_graph_edges(); )
1014
1015 Code_Gen();
1016 }
1017
1018 Compile::~Compile() {
1019 delete _print_inlining_stream;
1020 delete _first_failure_details;
1021 };
1022
1023 //------------------------------Init-------------------------------------------
1024 // Prepare for a single compilation
1025 void Compile::Init(bool aliasing) {
1026 _do_aliasing = aliasing;
1027 _unique = 0;
1028 _regalloc = nullptr;
1029
1030 _tf = nullptr; // filled in later
1031 _top = nullptr; // cached later
1032 _matcher = nullptr; // filled in later
1033 _cfg = nullptr; // filled in later
1034
1035 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1036
1037 _node_note_array = nullptr;
1038 _default_node_notes = nullptr;
1039 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1040
1041 _immutable_memory = nullptr; // filled in at first inquiry
1042
1043 #ifdef ASSERT
1044 _phase_optimize_finished = false;
1045 _exception_backedge = false;
1046 _type_verify = nullptr;
1047 #endif
1048
1049 // Globally visible Nodes
1050 // First set TOP to null to give safe behavior during creation of RootNode
1051 set_cached_top_node(nullptr);
1052 set_root(new RootNode());
1053 // Now that you have a Root to point to, create the real TOP
1054 set_cached_top_node( new ConNode(Type::TOP) );
1055 set_recent_alloc(nullptr, nullptr);
1056
1057 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1058 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1059 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1060 env()->set_dependencies(new Dependencies(env()));
1061
1062 _fixed_slots = 0;
1063 set_has_split_ifs(false);
1064 set_has_loops(false); // first approximation
1065 set_has_stringbuilder(false);
1066 set_has_boxed_value(false);
1067 _trap_can_recompile = false; // no traps emitted yet
1068 _major_progress = true; // start out assuming good things will happen
1069 set_has_unsafe_access(false);
1070 set_max_vector_size(0);
1071 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1072 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1073 set_decompile_count(0);
1074
1075 #ifndef PRODUCT
1076 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1077 #endif
1078
1079 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1080 _loop_opts_cnt = LoopOptsCount;
1081 _has_flat_accesses = false;
1082 _flat_accesses_share_alias = true;
1083 _scalarize_in_safepoints = false;
1084
1085 set_do_inlining(Inline);
1086 set_max_inline_size(MaxInlineSize);
1087 set_freq_inline_size(FreqInlineSize);
1088 set_do_scheduling(OptoScheduling);
1089
1090 set_do_vector_loop(false);
1091 set_has_monitors(false);
1092 set_has_scoped_access(false);
1093
1094 if (AllowVectorizeOnDemand) {
1095 if (has_method() && _directive->VectorizeOption) {
1096 set_do_vector_loop(true);
1097 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());})
1098 } else if (has_method() && method()->name() != nullptr &&
1099 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1100 set_do_vector_loop(true);
1101 }
1102 }
1103 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1104 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());})
1105
1106 _max_node_limit = _directive->MaxNodeLimitOption;
1107
1108 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1109 set_clinit_barrier_on_entry(true);
1110 }
1111 if (debug_info()->recording_non_safepoints()) {
1112 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1113 (comp_arena(), 8, 0, nullptr));
1114 set_default_node_notes(Node_Notes::make(this));
1115 }
1116
1117 const int grow_ats = 16;
1118 _max_alias_types = grow_ats;
1119 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1120 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1121 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1122 {
1123 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1124 }
1125 // Initialize the first few types.
1126 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1127 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1128 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1129 _num_alias_types = AliasIdxRaw+1;
1130 // Zero out the alias type cache.
1131 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1132 // A null adr_type hits in the cache right away. Preload the right answer.
1133 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1134 }
1135
1136 //---------------------------init_start----------------------------------------
1137 // Install the StartNode on this compile object.
1138 void Compile::init_start(StartNode* s) {
1139 if (failing())
1140 return; // already failing
1141 assert(s == start(), "");
1142 }
1143
1144 /**
1145 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1146 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1147 * the ideal graph.
1148 */
1149 StartNode* Compile::start() const {
1150 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1151 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1152 Node* start = root()->fast_out(i);
1153 if (start->is_Start()) {
1154 return start->as_Start();
1155 }
1156 }
1157 fatal("Did not find Start node!");
1158 return nullptr;
1159 }
1160
1161 //-------------------------------immutable_memory-------------------------------------
1162 // Access immutable memory
1163 Node* Compile::immutable_memory() {
1164 if (_immutable_memory != nullptr) {
1165 return _immutable_memory;
1166 }
1167 StartNode* s = start();
1168 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1169 Node *p = s->fast_out(i);
1170 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1171 _immutable_memory = p;
1172 return _immutable_memory;
1173 }
1174 }
1175 ShouldNotReachHere();
1176 return nullptr;
1177 }
1178
1179 //----------------------set_cached_top_node------------------------------------
1180 // Install the cached top node, and make sure Node::is_top works correctly.
1181 void Compile::set_cached_top_node(Node* tn) {
1182 if (tn != nullptr) verify_top(tn);
1183 Node* old_top = _top;
1184 _top = tn;
1185 // Calling Node::setup_is_top allows the nodes the chance to adjust
1186 // their _out arrays.
1187 if (_top != nullptr) _top->setup_is_top();
1188 if (old_top != nullptr) old_top->setup_is_top();
1189 assert(_top == nullptr || top()->is_top(), "");
1190 }
1191
1192 #ifdef ASSERT
1193 uint Compile::count_live_nodes_by_graph_walk() {
1194 Unique_Node_List useful(comp_arena());
1195 // Get useful node list by walking the graph.
1196 identify_useful_nodes(useful);
1197 return useful.size();
1198 }
1199
1200 void Compile::print_missing_nodes() {
1201
1202 // Return if CompileLog is null and PrintIdealNodeCount is false.
1203 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1204 return;
1205 }
1206
1207 // This is an expensive function. It is executed only when the user
1208 // specifies VerifyIdealNodeCount option or otherwise knows the
1209 // additional work that needs to be done to identify reachable nodes
1210 // by walking the flow graph and find the missing ones using
1211 // _dead_node_list.
1212
1213 Unique_Node_List useful(comp_arena());
1214 // Get useful node list by walking the graph.
1215 identify_useful_nodes(useful);
1216
1217 uint l_nodes = C->live_nodes();
1218 uint l_nodes_by_walk = useful.size();
1219
1220 if (l_nodes != l_nodes_by_walk) {
1221 if (_log != nullptr) {
1222 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1223 _log->stamp();
1224 _log->end_head();
1225 }
1226 VectorSet& useful_member_set = useful.member_set();
1227 int last_idx = l_nodes_by_walk;
1228 for (int i = 0; i < last_idx; i++) {
1229 if (useful_member_set.test(i)) {
1230 if (_dead_node_list.test(i)) {
1231 if (_log != nullptr) {
1232 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1233 }
1234 if (PrintIdealNodeCount) {
1235 // Print the log message to tty
1236 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1237 useful.at(i)->dump();
1238 }
1239 }
1240 }
1241 else if (! _dead_node_list.test(i)) {
1242 if (_log != nullptr) {
1243 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1244 }
1245 if (PrintIdealNodeCount) {
1246 // Print the log message to tty
1247 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1248 }
1249 }
1250 }
1251 if (_log != nullptr) {
1252 _log->tail("mismatched_nodes");
1253 }
1254 }
1255 }
1256 void Compile::record_modified_node(Node* n) {
1257 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1258 _modified_nodes->push(n);
1259 }
1260 }
1261
1262 void Compile::remove_modified_node(Node* n) {
1263 if (_modified_nodes != nullptr) {
1264 _modified_nodes->remove(n);
1265 }
1266 }
1267 #endif
1268
1269 #ifndef PRODUCT
1270 void Compile::verify_top(Node* tn) const {
1271 if (tn != nullptr) {
1272 assert(tn->is_Con(), "top node must be a constant");
1273 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1274 assert(tn->in(0) != nullptr, "must have live top node");
1275 }
1276 }
1277 #endif
1278
1279
1280 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1281
1282 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1283 guarantee(arr != nullptr, "");
1284 int num_blocks = arr->length();
1285 if (grow_by < num_blocks) grow_by = num_blocks;
1286 int num_notes = grow_by * _node_notes_block_size;
1287 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1288 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1289 while (num_notes > 0) {
1290 arr->append(notes);
1291 notes += _node_notes_block_size;
1292 num_notes -= _node_notes_block_size;
1293 }
1294 assert(num_notes == 0, "exact multiple, please");
1295 }
1296
1297 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1298 if (source == nullptr || dest == nullptr) return false;
1299
1300 if (dest->is_Con())
1301 return false; // Do not push debug info onto constants.
1302
1303 #ifdef ASSERT
1304 // Leave a bread crumb trail pointing to the original node:
1305 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1306 dest->set_debug_orig(source);
1307 }
1308 #endif
1309
1310 if (node_note_array() == nullptr)
1311 return false; // Not collecting any notes now.
1312
1313 // This is a copy onto a pre-existing node, which may already have notes.
1314 // If both nodes have notes, do not overwrite any pre-existing notes.
1315 Node_Notes* source_notes = node_notes_at(source->_idx);
1316 if (source_notes == nullptr || source_notes->is_clear()) return false;
1317 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1318 if (dest_notes == nullptr || dest_notes->is_clear()) {
1319 return set_node_notes_at(dest->_idx, source_notes);
1320 }
1321
1322 Node_Notes merged_notes = (*source_notes);
1323 // The order of operations here ensures that dest notes will win...
1324 merged_notes.update_from(dest_notes);
1325 return set_node_notes_at(dest->_idx, &merged_notes);
1326 }
1327
1328
1329 //--------------------------allow_range_check_smearing-------------------------
1330 // Gating condition for coalescing similar range checks.
1331 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1332 // single covering check that is at least as strong as any of them.
1333 // If the optimization succeeds, the simplified (strengthened) range check
1334 // will always succeed. If it fails, we will deopt, and then give up
1335 // on the optimization.
1336 bool Compile::allow_range_check_smearing() const {
1337 // If this method has already thrown a range-check,
1338 // assume it was because we already tried range smearing
1339 // and it failed.
1340 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1341 return !already_trapped;
1342 }
1343
1344
1345 //------------------------------flatten_alias_type-----------------------------
1346 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1347 assert(do_aliasing(), "Aliasing should be enabled");
1348 int offset = tj->offset();
1349 TypePtr::PTR ptr = tj->ptr();
1350
1351 // Known instance (scalarizable allocation) alias only with itself.
1352 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1353 tj->is_oopptr()->is_known_instance();
1354
1355 // Process weird unsafe references.
1356 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1357 bool default_value_load = EnableValhalla && tj->is_instptr()->instance_klass() == ciEnv::current()->Class_klass();
1358 assert(InlineUnsafeOps || StressReflectiveCode || default_value_load, "indeterminate pointers come only from unsafe ops");
1359 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1360 tj = TypeOopPtr::BOTTOM;
1361 ptr = tj->ptr();
1362 offset = tj->offset();
1363 }
1364
1365 // Array pointers need some flattening
1366 const TypeAryPtr* ta = tj->isa_aryptr();
1367 if (ta && ta->is_stable()) {
1368 // Erase stability property for alias analysis.
1369 tj = ta = ta->cast_to_stable(false);
1370 }
1371 if (ta && ta->is_not_flat()) {
1372 // Erase not flat property for alias analysis.
1373 tj = ta = ta->cast_to_not_flat(false);
1374 }
1375 if (ta && ta->is_not_null_free()) {
1376 // Erase not null free property for alias analysis.
1377 tj = ta = ta->cast_to_not_null_free(false);
1378 }
1379
1380 if( ta && is_known_inst ) {
1381 if ( offset != Type::OffsetBot &&
1382 offset > arrayOopDesc::length_offset_in_bytes() ) {
1383 offset = Type::OffsetBot; // Flatten constant access into array body only
1384 tj = ta = ta->
1385 remove_speculative()->
1386 cast_to_ptr_type(ptr)->
1387 with_offset(offset);
1388 }
1389 } else if (ta) {
1390 // For arrays indexed by constant indices, we flatten the alias
1391 // space to include all of the array body. Only the header, klass
1392 // and array length can be accessed un-aliased.
1393 // For flat inline type array, each field has its own slice so
1394 // we must include the field offset.
1395 if( offset != Type::OffsetBot ) {
1396 if( ta->const_oop() ) { // MethodData* or Method*
1397 offset = Type::OffsetBot; // Flatten constant access into array body
1398 tj = ta = ta->
1399 remove_speculative()->
1400 cast_to_ptr_type(ptr)->
1401 cast_to_exactness(false)->
1402 with_offset(offset);
1403 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1404 // range is OK as-is.
1405 tj = ta = TypeAryPtr::RANGE;
1406 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1407 tj = TypeInstPtr::KLASS; // all klass loads look alike
1408 ta = TypeAryPtr::RANGE; // generic ignored junk
1409 ptr = TypePtr::BotPTR;
1410 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1411 tj = TypeInstPtr::MARK;
1412 ta = TypeAryPtr::RANGE; // generic ignored junk
1413 ptr = TypePtr::BotPTR;
1414 } else { // Random constant offset into array body
1415 offset = Type::OffsetBot; // Flatten constant access into array body
1416 tj = ta = ta->
1417 remove_speculative()->
1418 cast_to_ptr_type(ptr)->
1419 cast_to_exactness(false)->
1420 with_offset(offset);
1421 }
1422 }
1423 // Arrays of fixed size alias with arrays of unknown size.
1424 if (ta->size() != TypeInt::POS) {
1425 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1426 tj = ta = ta->
1427 remove_speculative()->
1428 cast_to_ptr_type(ptr)->
1429 with_ary(tary)->
1430 cast_to_exactness(false);
1431 }
1432 // Arrays of known objects become arrays of unknown objects.
1433 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1434 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1435 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1436 }
1437 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1438 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1439 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1440 }
1441 // Initially all flattened array accesses share a single slice
1442 if (ta->is_flat() && ta->elem() != TypeInstPtr::BOTTOM && _flat_accesses_share_alias) {
1443 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size(), /* stable= */ false, /* flat= */ true);
1444 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1445 }
1446 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1447 // cannot be distinguished by bytecode alone.
1448 if (ta->elem() == TypeInt::BOOL) {
1449 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1450 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1451 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1452 }
1453 // During the 2nd round of IterGVN, NotNull castings are removed.
1454 // Make sure the Bottom and NotNull variants alias the same.
1455 // Also, make sure exact and non-exact variants alias the same.
1456 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1457 tj = ta = ta->
1458 remove_speculative()->
1459 cast_to_ptr_type(TypePtr::BotPTR)->
1460 cast_to_exactness(false)->
1461 with_offset(offset);
1462 }
1463 }
1464
1465 // Oop pointers need some flattening
1466 const TypeInstPtr *to = tj->isa_instptr();
1467 if (to && to != TypeOopPtr::BOTTOM) {
1468 ciInstanceKlass* ik = to->instance_klass();
1469 if( ptr == TypePtr::Constant ) {
1470 if (ik != ciEnv::current()->Class_klass() ||
1471 offset < ik->layout_helper_size_in_bytes()) {
1472 // No constant oop pointers (such as Strings); they alias with
1473 // unknown strings.
1474 assert(!is_known_inst, "not scalarizable allocation");
1475 tj = to = to->
1476 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1477 remove_speculative()->
1478 cast_to_ptr_type(TypePtr::BotPTR)->
1479 cast_to_exactness(false);
1480 }
1481 } else if( is_known_inst ) {
1482 tj = to; // Keep NotNull and klass_is_exact for instance type
1483 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1484 // During the 2nd round of IterGVN, NotNull castings are removed.
1485 // Make sure the Bottom and NotNull variants alias the same.
1486 // Also, make sure exact and non-exact variants alias the same.
1487 tj = to = to->
1488 remove_speculative()->
1489 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1490 cast_to_ptr_type(TypePtr::BotPTR)->
1491 cast_to_exactness(false);
1492 }
1493 if (to->speculative() != nullptr) {
1494 tj = to = to->remove_speculative();
1495 }
1496 // Canonicalize the holder of this field
1497 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1498 // First handle header references such as a LoadKlassNode, even if the
1499 // object's klass is unloaded at compile time (4965979).
1500 if (!is_known_inst) { // Do it only for non-instance types
1501 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1502 }
1503 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1504 // Static fields are in the space above the normal instance
1505 // fields in the java.lang.Class instance.
1506 if (ik != ciEnv::current()->Class_klass()) {
1507 to = nullptr;
1508 tj = TypeOopPtr::BOTTOM;
1509 offset = tj->offset();
1510 }
1511 } else {
1512 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1513 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1514 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1515 if( is_known_inst ) {
1516 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, Type::Offset(offset), to->instance_id());
1517 } else {
1518 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, Type::Offset(offset));
1519 }
1520 }
1521 }
1522 }
1523
1524 // Klass pointers to object array klasses need some flattening
1525 const TypeKlassPtr *tk = tj->isa_klassptr();
1526 if( tk ) {
1527 // If we are referencing a field within a Klass, we need
1528 // to assume the worst case of an Object. Both exact and
1529 // inexact types must flatten to the same alias class so
1530 // use NotNull as the PTR.
1531 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1532 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1533 env()->Object_klass(),
1534 Type::Offset(offset));
1535 }
1536
1537 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1538 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1539 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1540 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset));
1541 } else {
1542 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());
1543 }
1544 }
1545 // Check for precise loads from the primary supertype array and force them
1546 // to the supertype cache alias index. Check for generic array loads from
1547 // the primary supertype array and also force them to the supertype cache
1548 // alias index. Since the same load can reach both, we need to merge
1549 // these 2 disparate memories into the same alias class. Since the
1550 // primary supertype array is read-only, there's no chance of confusion
1551 // where we bypass an array load and an array store.
1552 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1553 if (offset == Type::OffsetBot ||
1554 (offset >= primary_supers_offset &&
1555 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1556 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1557 offset = in_bytes(Klass::secondary_super_cache_offset());
1558 tj = tk = tk->with_offset(offset);
1559 }
1560 }
1561
1562 // Flatten all Raw pointers together.
1563 if (tj->base() == Type::RawPtr)
1564 tj = TypeRawPtr::BOTTOM;
1565
1566 if (tj->base() == Type::AnyPtr)
1567 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1568
1569 offset = tj->offset();
1570 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1571
1572 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1573 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1574 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1575 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1576 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1577 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1578 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1579 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1580 assert( tj->ptr() != TypePtr::TopPTR &&
1581 tj->ptr() != TypePtr::AnyNull &&
1582 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1583 // assert( tj->ptr() != TypePtr::Constant ||
1584 // tj->base() == Type::RawPtr ||
1585 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1586
1587 return tj;
1588 }
1589
1590 void Compile::AliasType::Init(int i, const TypePtr* at) {
1591 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1592 _index = i;
1593 _adr_type = at;
1594 _field = nullptr;
1595 _element = nullptr;
1596 _is_rewritable = true; // default
1597 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1598 if (atoop != nullptr && atoop->is_known_instance()) {
1599 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1600 _general_index = Compile::current()->get_alias_index(gt);
1601 } else {
1602 _general_index = 0;
1603 }
1604 }
1605
1606 BasicType Compile::AliasType::basic_type() const {
1607 if (element() != nullptr) {
1608 const Type* element = adr_type()->is_aryptr()->elem();
1609 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1610 } if (field() != nullptr) {
1611 return field()->layout_type();
1612 } else {
1613 return T_ILLEGAL; // unknown
1614 }
1615 }
1616
1617 //---------------------------------print_on------------------------------------
1618 #ifndef PRODUCT
1619 void Compile::AliasType::print_on(outputStream* st) {
1620 if (index() < 10)
1621 st->print("@ <%d> ", index());
1622 else st->print("@ <%d>", index());
1623 st->print(is_rewritable() ? " " : " RO");
1624 int offset = adr_type()->offset();
1625 if (offset == Type::OffsetBot)
1626 st->print(" +any");
1627 else st->print(" +%-3d", offset);
1628 st->print(" in ");
1629 adr_type()->dump_on(st);
1630 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1631 if (field() != nullptr && tjp) {
1632 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1633 tjp->offset() != field()->offset_in_bytes()) {
1634 st->print(" != ");
1635 field()->print();
1636 st->print(" ***");
1637 }
1638 }
1639 }
1640
1641 void print_alias_types() {
1642 Compile* C = Compile::current();
1643 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1644 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1645 C->alias_type(idx)->print_on(tty);
1646 tty->cr();
1647 }
1648 }
1649 #endif
1650
1651
1652 //----------------------------probe_alias_cache--------------------------------
1653 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1654 intptr_t key = (intptr_t) adr_type;
1655 key ^= key >> logAliasCacheSize;
1656 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1657 }
1658
1659
1660 //-----------------------------grow_alias_types--------------------------------
1661 void Compile::grow_alias_types() {
1662 const int old_ats = _max_alias_types; // how many before?
1663 const int new_ats = old_ats; // how many more?
1664 const int grow_ats = old_ats+new_ats; // how many now?
1665 _max_alias_types = grow_ats;
1666 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1667 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1668 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1669 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1670 }
1671
1672
1673 //--------------------------------find_alias_type------------------------------
1674 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1675 if (!do_aliasing()) {
1676 return alias_type(AliasIdxBot);
1677 }
1678
1679 AliasCacheEntry* ace = nullptr;
1680 if (!uncached) {
1681 ace = probe_alias_cache(adr_type);
1682 if (ace->_adr_type == adr_type) {
1683 return alias_type(ace->_index);
1684 }
1685 }
1686
1687 // Handle special cases.
1688 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1689 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1690
1691 // Do it the slow way.
1692 const TypePtr* flat = flatten_alias_type(adr_type);
1693
1694 #ifdef ASSERT
1695 {
1696 ResourceMark rm;
1697 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1698 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1699 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1700 Type::str(adr_type));
1701 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1702 const TypeOopPtr* foop = flat->is_oopptr();
1703 // Scalarizable allocations have exact klass always.
1704 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1705 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1706 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1707 Type::str(foop), Type::str(xoop));
1708 }
1709 }
1710 #endif
1711
1712 int idx = AliasIdxTop;
1713 for (int i = 0; i < num_alias_types(); i++) {
1714 if (alias_type(i)->adr_type() == flat) {
1715 idx = i;
1716 break;
1717 }
1718 }
1719
1720 if (idx == AliasIdxTop) {
1721 if (no_create) return nullptr;
1722 // Grow the array if necessary.
1723 if (_num_alias_types == _max_alias_types) grow_alias_types();
1724 // Add a new alias type.
1725 idx = _num_alias_types++;
1726 _alias_types[idx]->Init(idx, flat);
1727 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1728 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1729 if (flat->isa_instptr()) {
1730 if (flat->offset() == java_lang_Class::klass_offset()
1731 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1732 alias_type(idx)->set_rewritable(false);
1733 }
1734 ciField* field = nullptr;
1735 if (flat->isa_aryptr()) {
1736 #ifdef ASSERT
1737 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1738 // (T_BYTE has the weakest alignment and size restrictions...)
1739 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1740 #endif
1741 const Type* elemtype = flat->is_aryptr()->elem();
1742 if (flat->offset() == TypePtr::OffsetBot) {
1743 alias_type(idx)->set_element(elemtype);
1744 }
1745 int field_offset = flat->is_aryptr()->field_offset().get();
1746 if (flat->is_flat() &&
1747 field_offset != Type::OffsetBot) {
1748 ciInlineKlass* vk = elemtype->inline_klass();
1749 field_offset += vk->first_field_offset();
1750 field = vk->get_field_by_offset(field_offset, false);
1751 }
1752 }
1753 if (flat->isa_klassptr()) {
1754 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1755 alias_type(idx)->set_rewritable(false);
1756 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1757 alias_type(idx)->set_rewritable(false);
1758 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1759 alias_type(idx)->set_rewritable(false);
1760 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1761 alias_type(idx)->set_rewritable(false);
1762 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1763 alias_type(idx)->set_rewritable(false);
1764 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1765 alias_type(idx)->set_rewritable(false);
1766 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1767 alias_type(idx)->set_rewritable(false);
1768 }
1769 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1770 // but the base pointer type is not distinctive enough to identify
1771 // references into JavaThread.)
1772
1773 // Check for final fields.
1774 const TypeInstPtr* tinst = flat->isa_instptr();
1775 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1776 if (tinst->const_oop() != nullptr &&
1777 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1778 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1779 // static field
1780 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1781 field = k->get_field_by_offset(tinst->offset(), true);
1782 } else if (tinst->is_inlinetypeptr()) {
1783 // Inline type field
1784 ciInlineKlass* vk = tinst->inline_klass();
1785 field = vk->get_field_by_offset(tinst->offset(), false);
1786 } else {
1787 ciInstanceKlass *k = tinst->instance_klass();
1788 field = k->get_field_by_offset(tinst->offset(), false);
1789 }
1790 }
1791 assert(field == nullptr ||
1792 original_field == nullptr ||
1793 (field->holder() == original_field->holder() &&
1794 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1795 field->is_static() == original_field->is_static()), "wrong field?");
1796 // Set field() and is_rewritable() attributes.
1797 if (field != nullptr) {
1798 alias_type(idx)->set_field(field);
1799 if (flat->isa_aryptr()) {
1800 // Fields of flat arrays are rewritable although they are declared final
1801 assert(flat->is_flat(), "must be a flat array");
1802 alias_type(idx)->set_rewritable(true);
1803 }
1804 }
1805 }
1806
1807 // Fill the cache for next time.
1808 if (!uncached) {
1809 ace->_adr_type = adr_type;
1810 ace->_index = idx;
1811 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1812
1813 // Might as well try to fill the cache for the flattened version, too.
1814 AliasCacheEntry* face = probe_alias_cache(flat);
1815 if (face->_adr_type == nullptr) {
1816 face->_adr_type = flat;
1817 face->_index = idx;
1818 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1819 }
1820 }
1821
1822 return alias_type(idx);
1823 }
1824
1825
1826 Compile::AliasType* Compile::alias_type(ciField* field) {
1827 const TypeOopPtr* t;
1828 if (field->is_static())
1829 t = TypeInstPtr::make(field->holder()->java_mirror());
1830 else
1831 t = TypeOopPtr::make_from_klass_raw(field->holder());
1832 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1833 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1834 return atp;
1835 }
1836
1837
1838 //------------------------------have_alias_type--------------------------------
1839 bool Compile::have_alias_type(const TypePtr* adr_type) {
1840 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1841 if (ace->_adr_type == adr_type) {
1842 return true;
1843 }
1844
1845 // Handle special cases.
1846 if (adr_type == nullptr) return true;
1847 if (adr_type == TypePtr::BOTTOM) return true;
1848
1849 return find_alias_type(adr_type, true, nullptr) != nullptr;
1850 }
1851
1852 //-----------------------------must_alias--------------------------------------
1853 // True if all values of the given address type are in the given alias category.
1854 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1855 if (alias_idx == AliasIdxBot) return true; // the universal category
1856 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1857 if (alias_idx == AliasIdxTop) return false; // the empty category
1858 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1859
1860 // the only remaining possible overlap is identity
1861 int adr_idx = get_alias_index(adr_type);
1862 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1863 assert(adr_idx == alias_idx ||
1864 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1865 && adr_type != TypeOopPtr::BOTTOM),
1866 "should not be testing for overlap with an unsafe pointer");
1867 return adr_idx == alias_idx;
1868 }
1869
1870 //------------------------------can_alias--------------------------------------
1871 // True if any values of the given address type are in the given alias category.
1872 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1873 if (alias_idx == AliasIdxTop) return false; // the empty category
1874 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1875 // Known instance doesn't alias with bottom memory
1876 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1877 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1878
1879 // the only remaining possible overlap is identity
1880 int adr_idx = get_alias_index(adr_type);
1881 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1882 return adr_idx == alias_idx;
1883 }
1884
1885 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1886 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1887 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1888 if (parse_predicate_count() == 0) {
1889 return;
1890 }
1891 for (int i = 0; i < parse_predicate_count(); i++) {
1892 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1893 parse_predicate->mark_useless();
1894 igvn._worklist.push(parse_predicate);
1895 }
1896 _parse_predicates.clear();
1897 }
1898
1899 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1900 if (!n->for_post_loop_opts_igvn()) {
1901 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1902 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1903 _for_post_loop_igvn.append(n);
1904 }
1905 }
1906
1907 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1908 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1909 _for_post_loop_igvn.remove(n);
1910 }
1911
1912 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1913 // Verify that all previous optimizations produced a valid graph
1914 // at least to this point, even if no loop optimizations were done.
1915 PhaseIdealLoop::verify(igvn);
1916
1917 C->set_post_loop_opts_phase(); // no more loop opts allowed
1918
1919 assert(!C->major_progress(), "not cleared");
1920
1921 if (_for_post_loop_igvn.length() > 0) {
1922 while (_for_post_loop_igvn.length() > 0) {
1923 Node* n = _for_post_loop_igvn.pop();
1924 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1925 igvn._worklist.push(n);
1926 }
1927 igvn.optimize();
1928 if (failing()) return;
1929 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1930 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1931
1932 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1933 if (C->major_progress()) {
1934 C->clear_major_progress(); // ensure that major progress is now clear
1935 }
1936 }
1937 }
1938
1939 void Compile::add_inline_type(Node* n) {
1940 assert(n->is_InlineType(), "unexpected node");
1941 _inline_type_nodes.push(n);
1942 }
1943
1944 void Compile::remove_inline_type(Node* n) {
1945 assert(n->is_InlineType(), "unexpected node");
1946 if (_inline_type_nodes.contains(n)) {
1947 _inline_type_nodes.remove(n);
1948 }
1949 }
1950
1951 // Does the return value keep otherwise useless inline type allocations alive?
1952 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1953 ResourceMark rm;
1954 Unique_Node_List wq;
1955 wq.push(ret_val);
1956 bool some_allocations = false;
1957 for (uint i = 0; i < wq.size(); i++) {
1958 Node* n = wq.at(i);
1959 if (n->outcnt() > 1) {
1960 // Some other use for the allocation
1961 return false;
1962 } else if (n->is_InlineType()) {
1963 wq.push(n->in(1));
1964 } else if (n->is_Phi()) {
1965 for (uint j = 1; j < n->req(); j++) {
1966 wq.push(n->in(j));
1967 }
1968 } else if (n->is_CheckCastPP() &&
1969 n->in(1)->is_Proj() &&
1970 n->in(1)->in(0)->is_Allocate()) {
1971 some_allocations = true;
1972 } else if (n->is_CheckCastPP()) {
1973 wq.push(n->in(1));
1974 }
1975 }
1976 return some_allocations;
1977 }
1978
1979 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1980 // Make sure that the return value does not keep an otherwise unused allocation alive
1981 if (tf()->returns_inline_type_as_fields()) {
1982 Node* ret = nullptr;
1983 for (uint i = 1; i < root()->req(); i++) {
1984 Node* in = root()->in(i);
1985 if (in->Opcode() == Op_Return) {
1986 assert(ret == nullptr, "only one return");
1987 ret = in;
1988 }
1989 }
1990 if (ret != nullptr) {
1991 Node* ret_val = ret->in(TypeFunc::Parms);
1992 if (igvn.type(ret_val)->isa_oopptr() &&
1993 return_val_keeps_allocations_alive(ret_val)) {
1994 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
1995 assert(ret_val->outcnt() == 0, "should be dead now");
1996 igvn.remove_dead_node(ret_val);
1997 }
1998 }
1999 }
2000 if (_inline_type_nodes.length() == 0) {
2001 return;
2002 }
2003 // Scalarize inline types in safepoint debug info.
2004 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2005 set_scalarize_in_safepoints(true);
2006 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2007 InlineTypeNode* vt = _inline_type_nodes.at(i)->as_InlineType();
2008 vt->make_scalar_in_safepoints(&igvn);
2009 igvn.record_for_igvn(vt);
2010 }
2011 if (remove) {
2012 // Remove inline type nodes by replacing them with their oop input
2013 while (_inline_type_nodes.length() > 0) {
2014 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2015 if (vt->outcnt() == 0) {
2016 igvn.remove_dead_node(vt);
2017 continue;
2018 }
2019 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2020 DEBUG_ONLY(bool must_be_buffered = false);
2021 Node* u = vt->out(i);
2022 // Check if any users are blackholes. If so, rewrite them to use either the
2023 // allocated buffer, or individual components, instead of the inline type node
2024 // that goes away.
2025 if (u->is_Blackhole()) {
2026 BlackholeNode* bh = u->as_Blackhole();
2027
2028 // Unlink the old input
2029 int idx = bh->find_edge(vt);
2030 assert(idx != -1, "The edge should be there");
2031 bh->del_req(idx);
2032 --i;
2033
2034 if (vt->is_allocated(&igvn)) {
2035 // Already has the allocated instance, blackhole that
2036 bh->add_req(vt->get_oop());
2037 } else {
2038 // Not allocated yet, blackhole the components
2039 for (uint c = 0; c < vt->field_count(); c++) {
2040 bh->add_req(vt->field_value(c));
2041 }
2042 }
2043
2044 // Node modified, record for IGVN
2045 igvn.record_for_igvn(bh);
2046 }
2047 #ifdef ASSERT
2048 // Verify that inline type is buffered when replacing by oop
2049 else if (u->is_InlineType()) {
2050 // InlineType uses don't need buffering because they are about to be replaced as well
2051 } else if (u->is_Phi()) {
2052 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2053 } else {
2054 must_be_buffered = true;
2055 }
2056 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2057 vt->dump(0);
2058 u->dump(0);
2059 assert(false, "Should have been buffered");
2060 }
2061 #endif
2062 }
2063 igvn.replace_node(vt, vt->get_oop());
2064 }
2065 }
2066 igvn.optimize();
2067 }
2068
2069 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2070 if (!_has_flat_accesses) {
2071 return;
2072 }
2073 // Initially, all flat array accesses share the same slice to
2074 // keep dependencies with Object[] array accesses (that could be
2075 // to a flat array) correct. We're done with parsing so we
2076 // now know all flat array accesses in this compile
2077 // unit. Let's move flat array accesses to their own slice,
2078 // one per element field. This should help memory access
2079 // optimizations.
2080 ResourceMark rm;
2081 Unique_Node_List wq;
2082 wq.push(root());
2083
2084 Node_List mergememnodes;
2085 Node_List memnodes;
2086
2087 // Alias index currently shared by all flat memory accesses
2088 int index = get_alias_index(TypeAryPtr::INLINES);
2089
2090 // Find MergeMem nodes and flat array accesses
2091 for (uint i = 0; i < wq.size(); i++) {
2092 Node* n = wq.at(i);
2093 if (n->is_Mem()) {
2094 const TypePtr* adr_type = nullptr;
2095 if (n->Opcode() == Op_StoreCM) {
2096 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type()));
2097 } else {
2098 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2099 }
2100 if (adr_type == TypeAryPtr::INLINES) {
2101 memnodes.push(n);
2102 }
2103 } else if (n->is_MergeMem()) {
2104 MergeMemNode* mm = n->as_MergeMem();
2105 if (mm->memory_at(index) != mm->base_memory()) {
2106 mergememnodes.push(n);
2107 }
2108 }
2109 for (uint j = 0; j < n->req(); j++) {
2110 Node* m = n->in(j);
2111 if (m != nullptr) {
2112 wq.push(m);
2113 }
2114 }
2115 }
2116
2117 if (memnodes.size() > 0) {
2118 _flat_accesses_share_alias = false;
2119
2120 // We are going to change the slice for the flat array
2121 // accesses so we need to clear the cache entries that refer to
2122 // them.
2123 for (uint i = 0; i < AliasCacheSize; i++) {
2124 AliasCacheEntry* ace = &_alias_cache[i];
2125 if (ace->_adr_type != nullptr &&
2126 ace->_adr_type->is_flat()) {
2127 ace->_adr_type = nullptr;
2128 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2129 }
2130 }
2131
2132 // Find what aliases we are going to add
2133 int start_alias = num_alias_types()-1;
2134 int stop_alias = 0;
2135
2136 for (uint i = 0; i < memnodes.size(); i++) {
2137 Node* m = memnodes.at(i);
2138 const TypePtr* adr_type = nullptr;
2139 if (m->Opcode() == Op_StoreCM) {
2140 adr_type = m->in(MemNode::OopStore)->adr_type();
2141 if (adr_type != TypeAryPtr::INLINES) {
2142 // store was optimized out and we lost track of the adr_type
2143 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address),
2144 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore),
2145 get_alias_index(adr_type));
2146 igvn.register_new_node_with_optimizer(clone);
2147 igvn.replace_node(m, clone);
2148 }
2149 } else {
2150 adr_type = m->adr_type();
2151 #ifdef ASSERT
2152 m->as_Mem()->set_adr_type(adr_type);
2153 #endif
2154 }
2155 int idx = get_alias_index(adr_type);
2156 start_alias = MIN2(start_alias, idx);
2157 stop_alias = MAX2(stop_alias, idx);
2158 }
2159
2160 assert(stop_alias >= start_alias, "should have expanded aliases");
2161
2162 Node_Stack stack(0);
2163 #ifdef ASSERT
2164 VectorSet seen(Thread::current()->resource_area());
2165 #endif
2166 // Now let's fix the memory graph so each flat array access
2167 // is moved to the right slice. Start from the MergeMem nodes.
2168 uint last = unique();
2169 for (uint i = 0; i < mergememnodes.size(); i++) {
2170 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2171 Node* n = current->memory_at(index);
2172 MergeMemNode* mm = nullptr;
2173 do {
2174 // Follow memory edges through memory accesses, phis and
2175 // narrow membars and push nodes on the stack. Once we hit
2176 // bottom memory, we pop element off the stack one at a
2177 // time, in reverse order, and move them to the right slice
2178 // by changing their memory edges.
2179 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2180 assert(!seen.test_set(n->_idx), "");
2181 // Uses (a load for instance) will need to be moved to the
2182 // right slice as well and will get a new memory state
2183 // that we don't know yet. The use could also be the
2184 // backedge of a loop. We put a place holder node between
2185 // the memory node and its uses. We replace that place
2186 // holder with the correct memory state once we know it,
2187 // i.e. when nodes are popped off the stack. Using the
2188 // place holder make the logic work in the presence of
2189 // loops.
2190 if (n->outcnt() > 1) {
2191 Node* place_holder = nullptr;
2192 assert(!n->has_out_with(Op_Node), "");
2193 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2194 Node* u = n->out(k);
2195 if (u != current && u->_idx < last) {
2196 bool success = false;
2197 for (uint l = 0; l < u->req(); l++) {
2198 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2199 continue;
2200 }
2201 Node* in = u->in(l);
2202 if (in == n) {
2203 if (place_holder == nullptr) {
2204 place_holder = new Node(1);
2205 place_holder->init_req(0, n);
2206 }
2207 igvn.replace_input_of(u, l, place_holder);
2208 success = true;
2209 }
2210 }
2211 if (success) {
2212 --k;
2213 }
2214 }
2215 }
2216 }
2217 if (n->is_Phi()) {
2218 stack.push(n, 1);
2219 n = n->in(1);
2220 } else if (n->is_Mem()) {
2221 stack.push(n, n->req());
2222 n = n->in(MemNode::Memory);
2223 } else {
2224 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2225 stack.push(n, n->req());
2226 n = n->in(0)->in(TypeFunc::Memory);
2227 }
2228 } else {
2229 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2230 // Build a new MergeMem node to carry the new memory state
2231 // as we build it. IGVN should fold extraneous MergeMem
2232 // nodes.
2233 mm = MergeMemNode::make(n);
2234 igvn.register_new_node_with_optimizer(mm);
2235 while (stack.size() > 0) {
2236 Node* m = stack.node();
2237 uint idx = stack.index();
2238 if (m->is_Mem()) {
2239 // Move memory node to its new slice
2240 const TypePtr* adr_type = m->adr_type();
2241 int alias = get_alias_index(adr_type);
2242 Node* prev = mm->memory_at(alias);
2243 igvn.replace_input_of(m, MemNode::Memory, prev);
2244 mm->set_memory_at(alias, m);
2245 } else if (m->is_Phi()) {
2246 // We need as many new phis as there are new aliases
2247 igvn.replace_input_of(m, idx, mm);
2248 if (idx == m->req()-1) {
2249 Node* r = m->in(0);
2250 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2251 const TypePtr* adr_type = get_adr_type(j);
2252 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2253 continue;
2254 }
2255 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2256 igvn.register_new_node_with_optimizer(phi);
2257 for (uint k = 1; k < m->req(); k++) {
2258 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2259 }
2260 mm->set_memory_at(j, phi);
2261 }
2262 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2263 igvn.register_new_node_with_optimizer(base_phi);
2264 for (uint k = 1; k < m->req(); k++) {
2265 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2266 }
2267 mm->set_base_memory(base_phi);
2268 }
2269 } else {
2270 // This is a MemBarCPUOrder node from
2271 // Parse::array_load()/Parse::array_store(), in the
2272 // branch that handles flat arrays hidden under
2273 // an Object[] array. We also need one new membar per
2274 // new alias to keep the unknown access that the
2275 // membars protect properly ordered with accesses to
2276 // known flat array.
2277 assert(m->is_Proj(), "projection expected");
2278 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2279 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2280 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2281 const TypePtr* adr_type = get_adr_type(j);
2282 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2283 continue;
2284 }
2285 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2286 igvn.register_new_node_with_optimizer(mb);
2287 Node* mem = mm->memory_at(j);
2288 mb->init_req(TypeFunc::Control, ctrl);
2289 mb->init_req(TypeFunc::Memory, mem);
2290 ctrl = new ProjNode(mb, TypeFunc::Control);
2291 igvn.register_new_node_with_optimizer(ctrl);
2292 mem = new ProjNode(mb, TypeFunc::Memory);
2293 igvn.register_new_node_with_optimizer(mem);
2294 mm->set_memory_at(j, mem);
2295 }
2296 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2297 }
2298 if (idx < m->req()-1) {
2299 idx += 1;
2300 stack.set_index(idx);
2301 n = m->in(idx);
2302 break;
2303 }
2304 // Take care of place holder nodes
2305 if (m->has_out_with(Op_Node)) {
2306 Node* place_holder = m->find_out_with(Op_Node);
2307 if (place_holder != nullptr) {
2308 Node* mm_clone = mm->clone();
2309 igvn.register_new_node_with_optimizer(mm_clone);
2310 Node* hook = new Node(1);
2311 hook->init_req(0, mm);
2312 igvn.replace_node(place_holder, mm_clone);
2313 hook->destruct(&igvn);
2314 }
2315 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2316 }
2317 stack.pop();
2318 }
2319 }
2320 } while(stack.size() > 0);
2321 // Fix the memory state at the MergeMem we started from
2322 igvn.rehash_node_delayed(current);
2323 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2324 const TypePtr* adr_type = get_adr_type(j);
2325 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2326 continue;
2327 }
2328 current->set_memory_at(j, mm);
2329 }
2330 current->set_memory_at(index, current->base_memory());
2331 }
2332 igvn.optimize();
2333 }
2334 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2335 #ifdef ASSERT
2336 if (!_flat_accesses_share_alias) {
2337 wq.clear();
2338 wq.push(root());
2339 for (uint i = 0; i < wq.size(); i++) {
2340 Node* n = wq.at(i);
2341 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2342 for (uint j = 0; j < n->req(); j++) {
2343 Node* m = n->in(j);
2344 if (m != nullptr) {
2345 wq.push(m);
2346 }
2347 }
2348 }
2349 }
2350 #endif
2351 }
2352
2353 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2354 if (OptimizeUnstableIf) {
2355 _unstable_if_traps.append(trap);
2356 }
2357 }
2358
2359 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2360 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2361 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2362 Node* n = trap->uncommon_trap();
2363 if (!useful.member(n)) {
2364 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2365 }
2366 }
2367 }
2368
2369 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2370 // or fold-compares case. Return true if succeed or not found.
2371 //
2372 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2373 // false and ask fold-compares to yield.
2374 //
2375 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2376 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2377 // when deoptimization does happen.
2378 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2379 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2380 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2381 if (trap->uncommon_trap() == unc) {
2382 if (yield && trap->modified()) {
2383 return false;
2384 }
2385 _unstable_if_traps.delete_at(i);
2386 break;
2387 }
2388 }
2389 return true;
2390 }
2391
2392 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2393 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2394 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2395 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2396 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2397 CallStaticJavaNode* unc = trap->uncommon_trap();
2398 int next_bci = trap->next_bci();
2399 bool modified = trap->modified();
2400
2401 if (next_bci != -1 && !modified) {
2402 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2403 JVMState* jvms = unc->jvms();
2404 ciMethod* method = jvms->method();
2405 ciBytecodeStream iter(method);
2406
2407 iter.force_bci(jvms->bci());
2408 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2409 Bytecodes::Code c = iter.cur_bc();
2410 Node* lhs = nullptr;
2411 Node* rhs = nullptr;
2412 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2413 lhs = unc->peek_operand(0);
2414 rhs = unc->peek_operand(1);
2415 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2416 lhs = unc->peek_operand(0);
2417 }
2418
2419 ResourceMark rm;
2420 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2421 assert(live_locals.is_valid(), "broken liveness info");
2422 int len = (int)live_locals.size();
2423
2424 for (int i = 0; i < len; i++) {
2425 Node* local = unc->local(jvms, i);
2426 // kill local using the liveness of next_bci.
2427 // give up when the local looks like an operand to secure reexecution.
2428 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2429 uint idx = jvms->locoff() + i;
2430 #ifdef ASSERT
2431 if (PrintOpto && Verbose) {
2432 tty->print("[unstable_if] kill local#%d: ", idx);
2433 local->dump();
2434 tty->cr();
2435 }
2436 #endif
2437 igvn.replace_input_of(unc, idx, top());
2438 modified = true;
2439 }
2440 }
2441 }
2442
2443 // keep the modified trap for late query
2444 if (modified) {
2445 trap->set_modified();
2446 } else {
2447 _unstable_if_traps.delete_at(i);
2448 }
2449 }
2450 igvn.optimize();
2451 }
2452
2453 // StringOpts and late inlining of string methods
2454 void Compile::inline_string_calls(bool parse_time) {
2455 {
2456 // remove useless nodes to make the usage analysis simpler
2457 ResourceMark rm;
2458 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2459 }
2460
2461 {
2462 ResourceMark rm;
2463 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2464 PhaseStringOpts pso(initial_gvn());
2465 print_method(PHASE_AFTER_STRINGOPTS, 3);
2466 }
2467
2468 // now inline anything that we skipped the first time around
2469 if (!parse_time) {
2470 _late_inlines_pos = _late_inlines.length();
2471 }
2472
2473 while (_string_late_inlines.length() > 0) {
2474 CallGenerator* cg = _string_late_inlines.pop();
2475 cg->do_late_inline();
2476 if (failing()) return;
2477 }
2478 _string_late_inlines.trunc_to(0);
2479 }
2480
2481 // Late inlining of boxing methods
2482 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2483 if (_boxing_late_inlines.length() > 0) {
2484 assert(has_boxed_value(), "inconsistent");
2485
2486 set_inlining_incrementally(true);
2487
2488 igvn_worklist()->ensure_empty(); // should be done with igvn
2489
2490 _late_inlines_pos = _late_inlines.length();
2491
2492 while (_boxing_late_inlines.length() > 0) {
2493 CallGenerator* cg = _boxing_late_inlines.pop();
2494 cg->do_late_inline();
2495 if (failing()) return;
2496 }
2497 _boxing_late_inlines.trunc_to(0);
2498
2499 inline_incrementally_cleanup(igvn);
2500
2501 set_inlining_incrementally(false);
2502 }
2503 }
2504
2505 bool Compile::inline_incrementally_one() {
2506 assert(IncrementalInline, "incremental inlining should be on");
2507
2508 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2509
2510 set_inlining_progress(false);
2511 set_do_cleanup(false);
2512
2513 for (int i = 0; i < _late_inlines.length(); i++) {
2514 _late_inlines_pos = i+1;
2515 CallGenerator* cg = _late_inlines.at(i);
2516 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2517 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2518 cg->do_late_inline();
2519 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2520 if (failing()) {
2521 return false;
2522 } else if (inlining_progress()) {
2523 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2524 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2525 break; // process one call site at a time
2526 }
2527 } else {
2528 // Ignore late inline direct calls when inlining is not allowed.
2529 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2530 }
2531 }
2532 // Remove processed elements.
2533 _late_inlines.remove_till(_late_inlines_pos);
2534 _late_inlines_pos = 0;
2535
2536 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2537
2538 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2539
2540 set_inlining_progress(false);
2541 set_do_cleanup(false);
2542
2543 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2544 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2545 }
2546
2547 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2548 {
2549 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2550 ResourceMark rm;
2551 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2552 }
2553 {
2554 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2555 igvn.reset_from_gvn(initial_gvn());
2556 igvn.optimize();
2557 if (failing()) return;
2558 }
2559 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2560 }
2561
2562 // Perform incremental inlining until bound on number of live nodes is reached
2563 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2564 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2565
2566 set_inlining_incrementally(true);
2567 uint low_live_nodes = 0;
2568
2569 while (_late_inlines.length() > 0) {
2570 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2571 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2572 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2573 // PhaseIdealLoop is expensive so we only try it once we are
2574 // out of live nodes and we only try it again if the previous
2575 // helped got the number of nodes down significantly
2576 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2577 if (failing()) return;
2578 low_live_nodes = live_nodes();
2579 _major_progress = true;
2580 }
2581
2582 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2583 bool do_print_inlining = print_inlining() || print_intrinsics();
2584 if (do_print_inlining || log() != nullptr) {
2585 // Print inlining message for candidates that we couldn't inline for lack of space.
2586 for (int i = 0; i < _late_inlines.length(); i++) {
2587 CallGenerator* cg = _late_inlines.at(i);
2588 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2589 if (do_print_inlining) {
2590 cg->print_inlining_late(InliningResult::FAILURE, msg);
2591 }
2592 log_late_inline_failure(cg, msg);
2593 }
2594 }
2595 break; // finish
2596 }
2597 }
2598
2599 igvn_worklist()->ensure_empty(); // should be done with igvn
2600
2601 while (inline_incrementally_one()) {
2602 assert(!failing(), "inconsistent");
2603 }
2604 if (failing()) return;
2605
2606 inline_incrementally_cleanup(igvn);
2607
2608 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2609
2610 if (failing()) return;
2611
2612 if (_late_inlines.length() == 0) {
2613 break; // no more progress
2614 }
2615 }
2616
2617 igvn_worklist()->ensure_empty(); // should be done with igvn
2618
2619 if (_string_late_inlines.length() > 0) {
2620 assert(has_stringbuilder(), "inconsistent");
2621
2622 inline_string_calls(false);
2623
2624 if (failing()) return;
2625
2626 inline_incrementally_cleanup(igvn);
2627 }
2628
2629 set_inlining_incrementally(false);
2630 }
2631
2632 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2633 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2634 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2635 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2636 // as if "inlining_incrementally() == true" were set.
2637 assert(inlining_incrementally() == false, "not allowed");
2638 #ifdef ASSERT
2639 Unique_Node_List* modified_nodes = _modified_nodes;
2640 _modified_nodes = nullptr;
2641 #endif
2642 assert(_late_inlines.length() > 0, "sanity");
2643
2644 while (_late_inlines.length() > 0) {
2645 igvn_worklist()->ensure_empty(); // should be done with igvn
2646
2647 while (inline_incrementally_one()) {
2648 assert(!failing(), "inconsistent");
2649 }
2650 if (failing()) return;
2651
2652 inline_incrementally_cleanup(igvn);
2653 }
2654 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2655 }
2656
2657 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2658 if (_loop_opts_cnt > 0) {
2659 while (major_progress() && (_loop_opts_cnt > 0)) {
2660 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2661 PhaseIdealLoop::optimize(igvn, mode);
2662 _loop_opts_cnt--;
2663 if (failing()) return false;
2664 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2665 }
2666 }
2667 return true;
2668 }
2669
2670 // Remove edges from "root" to each SafePoint at a backward branch.
2671 // They were inserted during parsing (see add_safepoint()) to make
2672 // infinite loops without calls or exceptions visible to root, i.e.,
2673 // useful.
2674 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2675 Node *r = root();
2676 if (r != nullptr) {
2677 for (uint i = r->req(); i < r->len(); ++i) {
2678 Node *n = r->in(i);
2679 if (n != nullptr && n->is_SafePoint()) {
2680 r->rm_prec(i);
2681 if (n->outcnt() == 0) {
2682 igvn.remove_dead_node(n);
2683 }
2684 --i;
2685 }
2686 }
2687 // Parsing may have added top inputs to the root node (Path
2688 // leading to the Halt node proven dead). Make sure we get a
2689 // chance to clean them up.
2690 igvn._worklist.push(r);
2691 igvn.optimize();
2692 }
2693 }
2694
2695 //------------------------------Optimize---------------------------------------
2696 // Given a graph, optimize it.
2697 void Compile::Optimize() {
2698 TracePhase tp("optimizer", &timers[_t_optimizer]);
2699
2700 #ifndef PRODUCT
2701 if (env()->break_at_compile()) {
2702 BREAKPOINT;
2703 }
2704
2705 #endif
2706
2707 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2708 #ifdef ASSERT
2709 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2710 #endif
2711
2712 ResourceMark rm;
2713
2714 print_inlining_reinit();
2715
2716 NOT_PRODUCT( verify_graph_edges(); )
2717
2718 print_method(PHASE_AFTER_PARSING, 1);
2719
2720 {
2721 // Iterative Global Value Numbering, including ideal transforms
2722 // Initialize IterGVN with types and values from parse-time GVN
2723 PhaseIterGVN igvn(initial_gvn());
2724 #ifdef ASSERT
2725 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2726 #endif
2727 {
2728 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2729 igvn.optimize();
2730 }
2731
2732 if (failing()) return;
2733
2734 print_method(PHASE_ITER_GVN1, 2);
2735
2736 process_for_unstable_if_traps(igvn);
2737
2738 if (failing()) return;
2739
2740 inline_incrementally(igvn);
2741
2742 print_method(PHASE_INCREMENTAL_INLINE, 2);
2743
2744 if (failing()) return;
2745
2746 if (eliminate_boxing()) {
2747 // Inline valueOf() methods now.
2748 inline_boxing_calls(igvn);
2749
2750 if (failing()) return;
2751
2752 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2753 inline_incrementally(igvn);
2754 }
2755
2756 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2757
2758 if (failing()) return;
2759 }
2760
2761 // Remove the speculative part of types and clean up the graph from
2762 // the extra CastPP nodes whose only purpose is to carry them. Do
2763 // that early so that optimizations are not disrupted by the extra
2764 // CastPP nodes.
2765 remove_speculative_types(igvn);
2766
2767 if (failing()) return;
2768
2769 // No more new expensive nodes will be added to the list from here
2770 // so keep only the actual candidates for optimizations.
2771 cleanup_expensive_nodes(igvn);
2772
2773 if (failing()) return;
2774
2775 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2776 if (EnableVectorSupport && has_vbox_nodes()) {
2777 TracePhase tp("", &timers[_t_vector]);
2778 PhaseVector pv(igvn);
2779 pv.optimize_vector_boxes();
2780 if (failing()) return;
2781 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2782 }
2783 assert(!has_vbox_nodes(), "sanity");
2784
2785 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2786 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2787 igvn_worklist()->ensure_empty(); // should be done with igvn
2788 {
2789 ResourceMark rm;
2790 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2791 }
2792 igvn.reset_from_gvn(initial_gvn());
2793 igvn.optimize();
2794 if (failing()) return;
2795 }
2796
2797 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2798 // safepoints
2799 remove_root_to_sfpts_edges(igvn);
2800
2801 // Process inline type nodes now that all inlining is over
2802 process_inline_types(igvn);
2803
2804 adjust_flat_array_access_aliases(igvn);
2805
2806 if (failing()) return;
2807
2808 // Perform escape analysis
2809 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2810 if (has_loops()) {
2811 // Cleanup graph (remove dead nodes).
2812 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2813 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2814 if (failing()) return;
2815 }
2816 bool progress;
2817 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2818 do {
2819 ConnectionGraph::do_analysis(this, &igvn);
2820
2821 if (failing()) return;
2822
2823 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2824
2825 // Optimize out fields loads from scalar replaceable allocations.
2826 igvn.optimize();
2827 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2828
2829 if (failing()) return;
2830
2831 if (congraph() != nullptr && macro_count() > 0) {
2832 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2833 PhaseMacroExpand mexp(igvn);
2834 mexp.eliminate_macro_nodes();
2835 if (failing()) return;
2836
2837 igvn.set_delay_transform(false);
2838 igvn.optimize();
2839 if (failing()) return;
2840
2841 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2842 }
2843
2844 ConnectionGraph::verify_ram_nodes(this, root());
2845 if (failing()) return;
2846
2847 progress = do_iterative_escape_analysis() &&
2848 (macro_count() < mcount) &&
2849 ConnectionGraph::has_candidates(this);
2850 // Try again if candidates exist and made progress
2851 // by removing some allocations and/or locks.
2852 } while (progress);
2853 }
2854
2855 // Loop transforms on the ideal graph. Range Check Elimination,
2856 // peeling, unrolling, etc.
2857
2858 // Set loop opts counter
2859 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2860 {
2861 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2862 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2863 _loop_opts_cnt--;
2864 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2865 if (failing()) return;
2866 }
2867 // Loop opts pass if partial peeling occurred in previous pass
2868 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2869 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2870 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2871 _loop_opts_cnt--;
2872 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2873 if (failing()) return;
2874 }
2875 // Loop opts pass for loop-unrolling before CCP
2876 if(major_progress() && (_loop_opts_cnt > 0)) {
2877 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2878 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2879 _loop_opts_cnt--;
2880 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2881 }
2882 if (!failing()) {
2883 // Verify that last round of loop opts produced a valid graph
2884 PhaseIdealLoop::verify(igvn);
2885 }
2886 }
2887 if (failing()) return;
2888
2889 // Conditional Constant Propagation;
2890 print_method(PHASE_BEFORE_CCP1, 2);
2891 PhaseCCP ccp( &igvn );
2892 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2893 {
2894 TracePhase tp("ccp", &timers[_t_ccp]);
2895 ccp.do_transform();
2896 }
2897 print_method(PHASE_CCP1, 2);
2898
2899 assert( true, "Break here to ccp.dump_old2new_map()");
2900
2901 // Iterative Global Value Numbering, including ideal transforms
2902 {
2903 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2904 igvn.reset_from_igvn(&ccp);
2905 igvn.optimize();
2906 }
2907 print_method(PHASE_ITER_GVN2, 2);
2908
2909 if (failing()) return;
2910
2911 // Loop transforms on the ideal graph. Range Check Elimination,
2912 // peeling, unrolling, etc.
2913 if (!optimize_loops(igvn, LoopOptsDefault)) {
2914 return;
2915 }
2916
2917 if (failing()) return;
2918
2919 C->clear_major_progress(); // ensure that major progress is now clear
2920
2921 process_for_post_loop_opts_igvn(igvn);
2922
2923 if (failing()) return;
2924
2925 #ifdef ASSERT
2926 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2927 #endif
2928
2929 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2930
2931 if (_late_inlines.length() > 0) {
2932 // More opportunities to optimize virtual and MH calls.
2933 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2934 process_late_inline_calls_no_inline(igvn);
2935 }
2936
2937 {
2938 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2939 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2940 PhaseMacroExpand mex(igvn);
2941 if (mex.expand_macro_nodes()) {
2942 assert(failing(), "must bail out w/ explicit message");
2943 return;
2944 }
2945 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2946 }
2947
2948 // Process inline type nodes again and remove them. From here
2949 // on we don't need to keep track of field values anymore.
2950 process_inline_types(igvn, /* remove= */ true);
2951
2952 {
2953 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2954 if (bs->expand_barriers(this, igvn)) {
2955 assert(failing(), "must bail out w/ explicit message");
2956 return;
2957 }
2958 print_method(PHASE_BARRIER_EXPANSION, 2);
2959 }
2960
2961 if (C->max_vector_size() > 0) {
2962 C->optimize_logic_cones(igvn);
2963 igvn.optimize();
2964 if (failing()) return;
2965 }
2966
2967 DEBUG_ONLY( _modified_nodes = nullptr; )
2968 DEBUG_ONLY( _late_inlines.clear(); )
2969
2970 assert(igvn._worklist.size() == 0, "not empty");
2971 } // (End scope of igvn; run destructor if necessary for asserts.)
2972
2973 check_no_dead_use();
2974
2975 process_print_inlining();
2976
2977 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2978 // to remove hashes to unlock nodes for modifications.
2979 C->node_hash()->clear();
2980
2981 // A method with only infinite loops has no edges entering loops from root
2982 {
2983 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2984 if (final_graph_reshaping()) {
2985 assert(failing(), "must bail out w/ explicit message");
2986 return;
2987 }
2988 }
2989
2990 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2991 DEBUG_ONLY(set_phase_optimize_finished();)
2992 }
2993
2994 #ifdef ASSERT
2995 void Compile::check_no_dead_use() const {
2996 ResourceMark rm;
2997 Unique_Node_List wq;
2998 wq.push(root());
2999 for (uint i = 0; i < wq.size(); ++i) {
3000 Node* n = wq.at(i);
3001 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3002 Node* u = n->fast_out(j);
3003 if (u->outcnt() == 0 && !u->is_Con()) {
3004 u->dump();
3005 fatal("no reachable node should have no use");
3006 }
3007 wq.push(u);
3008 }
3009 }
3010 }
3011 #endif
3012
3013 void Compile::inline_vector_reboxing_calls() {
3014 if (C->_vector_reboxing_late_inlines.length() > 0) {
3015 _late_inlines_pos = C->_late_inlines.length();
3016 while (_vector_reboxing_late_inlines.length() > 0) {
3017 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
3018 cg->do_late_inline();
3019 if (failing()) return;
3020 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
3021 }
3022 _vector_reboxing_late_inlines.trunc_to(0);
3023 }
3024 }
3025
3026 bool Compile::has_vbox_nodes() {
3027 if (C->_vector_reboxing_late_inlines.length() > 0) {
3028 return true;
3029 }
3030 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3031 Node * n = C->macro_node(macro_idx);
3032 assert(n->is_macro(), "only macro nodes expected here");
3033 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3034 return true;
3035 }
3036 }
3037 return false;
3038 }
3039
3040 //---------------------------- Bitwise operation packing optimization ---------------------------
3041
3042 static bool is_vector_unary_bitwise_op(Node* n) {
3043 return n->Opcode() == Op_XorV &&
3044 VectorNode::is_vector_bitwise_not_pattern(n);
3045 }
3046
3047 static bool is_vector_binary_bitwise_op(Node* n) {
3048 switch (n->Opcode()) {
3049 case Op_AndV:
3050 case Op_OrV:
3051 return true;
3052
3053 case Op_XorV:
3054 return !is_vector_unary_bitwise_op(n);
3055
3056 default:
3057 return false;
3058 }
3059 }
3060
3061 static bool is_vector_ternary_bitwise_op(Node* n) {
3062 return n->Opcode() == Op_MacroLogicV;
3063 }
3064
3065 static bool is_vector_bitwise_op(Node* n) {
3066 return is_vector_unary_bitwise_op(n) ||
3067 is_vector_binary_bitwise_op(n) ||
3068 is_vector_ternary_bitwise_op(n);
3069 }
3070
3071 static bool is_vector_bitwise_cone_root(Node* n) {
3072 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
3073 return false;
3074 }
3075 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3076 if (is_vector_bitwise_op(n->fast_out(i))) {
3077 return false;
3078 }
3079 }
3080 return true;
3081 }
3082
3083 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3084 uint cnt = 0;
3085 if (is_vector_bitwise_op(n)) {
3086 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3087 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3088 for (uint i = 1; i < inp_cnt; i++) {
3089 Node* in = n->in(i);
3090 bool skip = VectorNode::is_all_ones_vector(in);
3091 if (!skip && !inputs.member(in)) {
3092 inputs.push(in);
3093 cnt++;
3094 }
3095 }
3096 assert(cnt <= 1, "not unary");
3097 } else {
3098 uint last_req = inp_cnt;
3099 if (is_vector_ternary_bitwise_op(n)) {
3100 last_req = inp_cnt - 1; // skip last input
3101 }
3102 for (uint i = 1; i < last_req; i++) {
3103 Node* def = n->in(i);
3104 if (!inputs.member(def)) {
3105 inputs.push(def);
3106 cnt++;
3107 }
3108 }
3109 }
3110 } else { // not a bitwise operations
3111 if (!inputs.member(n)) {
3112 inputs.push(n);
3113 cnt++;
3114 }
3115 }
3116 return cnt;
3117 }
3118
3119 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3120 Unique_Node_List useful_nodes;
3121 C->identify_useful_nodes(useful_nodes);
3122
3123 for (uint i = 0; i < useful_nodes.size(); i++) {
3124 Node* n = useful_nodes.at(i);
3125 if (is_vector_bitwise_cone_root(n)) {
3126 list.push(n);
3127 }
3128 }
3129 }
3130
3131 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3132 const TypeVect* vt,
3133 Unique_Node_List& partition,
3134 Unique_Node_List& inputs) {
3135 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3136 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3137 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3138
3139 Node* in1 = inputs.at(0);
3140 Node* in2 = inputs.at(1);
3141 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3142
3143 uint func = compute_truth_table(partition, inputs);
3144
3145 Node* pn = partition.at(partition.size() - 1);
3146 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3147 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3148 }
3149
3150 static uint extract_bit(uint func, uint pos) {
3151 return (func & (1 << pos)) >> pos;
3152 }
3153
3154 //
3155 // A macro logic node represents a truth table. It has 4 inputs,
3156 // First three inputs corresponds to 3 columns of a truth table
3157 // and fourth input captures the logic function.
3158 //
3159 // eg. fn = (in1 AND in2) OR in3;
3160 //
3161 // MacroNode(in1,in2,in3,fn)
3162 //
3163 // -----------------
3164 // in1 in2 in3 fn
3165 // -----------------
3166 // 0 0 0 0
3167 // 0 0 1 1
3168 // 0 1 0 0
3169 // 0 1 1 1
3170 // 1 0 0 0
3171 // 1 0 1 1
3172 // 1 1 0 1
3173 // 1 1 1 1
3174 //
3175
3176 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3177 int res = 0;
3178 for (int i = 0; i < 8; i++) {
3179 int bit1 = extract_bit(in1, i);
3180 int bit2 = extract_bit(in2, i);
3181 int bit3 = extract_bit(in3, i);
3182
3183 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3184 int func_bit = extract_bit(func, func_bit_pos);
3185
3186 res |= func_bit << i;
3187 }
3188 return res;
3189 }
3190
3191 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
3192 assert(n != nullptr, "");
3193 assert(eval_map.contains(n), "absent");
3194 return *(eval_map.get(n));
3195 }
3196
3197 static void eval_operands(Node* n,
3198 uint& func1, uint& func2, uint& func3,
3199 ResourceHashtable<Node*,uint>& eval_map) {
3200 assert(is_vector_bitwise_op(n), "");
3201
3202 if (is_vector_unary_bitwise_op(n)) {
3203 Node* opnd = n->in(1);
3204 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3205 opnd = n->in(2);
3206 }
3207 func1 = eval_operand(opnd, eval_map);
3208 } else if (is_vector_binary_bitwise_op(n)) {
3209 func1 = eval_operand(n->in(1), eval_map);
3210 func2 = eval_operand(n->in(2), eval_map);
3211 } else {
3212 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3213 func1 = eval_operand(n->in(1), eval_map);
3214 func2 = eval_operand(n->in(2), eval_map);
3215 func3 = eval_operand(n->in(3), eval_map);
3216 }
3217 }
3218
3219 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3220 assert(inputs.size() <= 3, "sanity");
3221 ResourceMark rm;
3222 uint res = 0;
3223 ResourceHashtable<Node*,uint> eval_map;
3224
3225 // Populate precomputed functions for inputs.
3226 // Each input corresponds to one column of 3 input truth-table.
3227 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3228 0xCC, // (_, b, _) -> b
3229 0xF0 }; // (a, _, _) -> a
3230 for (uint i = 0; i < inputs.size(); i++) {
3231 eval_map.put(inputs.at(i), input_funcs[2-i]);
3232 }
3233
3234 for (uint i = 0; i < partition.size(); i++) {
3235 Node* n = partition.at(i);
3236
3237 uint func1 = 0, func2 = 0, func3 = 0;
3238 eval_operands(n, func1, func2, func3, eval_map);
3239
3240 switch (n->Opcode()) {
3241 case Op_OrV:
3242 assert(func3 == 0, "not binary");
3243 res = func1 | func2;
3244 break;
3245 case Op_AndV:
3246 assert(func3 == 0, "not binary");
3247 res = func1 & func2;
3248 break;
3249 case Op_XorV:
3250 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3251 assert(func2 == 0 && func3 == 0, "not unary");
3252 res = (~func1) & 0xFF;
3253 } else {
3254 assert(func3 == 0, "not binary");
3255 res = func1 ^ func2;
3256 }
3257 break;
3258 case Op_MacroLogicV:
3259 // Ordering of inputs may change during evaluation of sub-tree
3260 // containing MacroLogic node as a child node, thus a re-evaluation
3261 // makes sure that function is evaluated in context of current
3262 // inputs.
3263 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3264 break;
3265
3266 default: assert(false, "not supported: %s", n->Name());
3267 }
3268 assert(res <= 0xFF, "invalid");
3269 eval_map.put(n, res);
3270 }
3271 return res;
3272 }
3273
3274 // Criteria under which nodes gets packed into a macro logic node:-
3275 // 1) Parent and both child nodes are all unmasked or masked with
3276 // same predicates.
3277 // 2) Masked parent can be packed with left child if it is predicated
3278 // and both have same predicates.
3279 // 3) Masked parent can be packed with right child if its un-predicated
3280 // or has matching predication condition.
3281 // 4) An unmasked parent can be packed with an unmasked child.
3282 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3283 assert(partition.size() == 0, "not empty");
3284 assert(inputs.size() == 0, "not empty");
3285 if (is_vector_ternary_bitwise_op(n)) {
3286 return false;
3287 }
3288
3289 bool is_unary_op = is_vector_unary_bitwise_op(n);
3290 if (is_unary_op) {
3291 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3292 return false; // too few inputs
3293 }
3294
3295 bool pack_left_child = true;
3296 bool pack_right_child = true;
3297
3298 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3299 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3300
3301 int left_child_input_cnt = 0;
3302 int right_child_input_cnt = 0;
3303
3304 bool parent_is_predicated = n->is_predicated_vector();
3305 bool left_child_predicated = n->in(1)->is_predicated_vector();
3306 bool right_child_predicated = n->in(2)->is_predicated_vector();
3307
3308 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
3309 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3310 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3311
3312 do {
3313 if (pack_left_child && left_child_LOP &&
3314 ((!parent_is_predicated && !left_child_predicated) ||
3315 ((parent_is_predicated && left_child_predicated &&
3316 parent_pred == left_child_pred)))) {
3317 partition.push(n->in(1));
3318 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3319 } else {
3320 inputs.push(n->in(1));
3321 left_child_input_cnt = 1;
3322 }
3323
3324 if (pack_right_child && right_child_LOP &&
3325 (!right_child_predicated ||
3326 (right_child_predicated && parent_is_predicated &&
3327 parent_pred == right_child_pred))) {
3328 partition.push(n->in(2));
3329 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3330 } else {
3331 inputs.push(n->in(2));
3332 right_child_input_cnt = 1;
3333 }
3334
3335 if (inputs.size() > 3) {
3336 assert(partition.size() > 0, "");
3337 inputs.clear();
3338 partition.clear();
3339 if (left_child_input_cnt > right_child_input_cnt) {
3340 pack_left_child = false;
3341 } else {
3342 pack_right_child = false;
3343 }
3344 } else {
3345 break;
3346 }
3347 } while(true);
3348
3349 if(partition.size()) {
3350 partition.push(n);
3351 }
3352
3353 return (partition.size() == 2 || partition.size() == 3) &&
3354 (inputs.size() == 2 || inputs.size() == 3);
3355 }
3356
3357 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3358 assert(is_vector_bitwise_op(n), "not a root");
3359
3360 visited.set(n->_idx);
3361
3362 // 1) Do a DFS walk over the logic cone.
3363 for (uint i = 1; i < n->req(); i++) {
3364 Node* in = n->in(i);
3365 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3366 process_logic_cone_root(igvn, in, visited);
3367 }
3368 }
3369
3370 // 2) Bottom up traversal: Merge node[s] with
3371 // the parent to form macro logic node.
3372 Unique_Node_List partition;
3373 Unique_Node_List inputs;
3374 if (compute_logic_cone(n, partition, inputs)) {
3375 const TypeVect* vt = n->bottom_type()->is_vect();
3376 Node* pn = partition.at(partition.size() - 1);
3377 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3378 if (mask == nullptr ||
3379 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3380 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3381 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3382 igvn.replace_node(n, macro_logic);
3383 }
3384 }
3385 }
3386
3387 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3388 ResourceMark rm;
3389 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3390 Unique_Node_List list;
3391 collect_logic_cone_roots(list);
3392
3393 while (list.size() > 0) {
3394 Node* n = list.pop();
3395 const TypeVect* vt = n->bottom_type()->is_vect();
3396 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3397 if (supported) {
3398 VectorSet visited(comp_arena());
3399 process_logic_cone_root(igvn, n, visited);
3400 }
3401 }
3402 }
3403 }
3404
3405 //------------------------------Code_Gen---------------------------------------
3406 // Given a graph, generate code for it
3407 void Compile::Code_Gen() {
3408 if (failing()) {
3409 return;
3410 }
3411
3412 // Perform instruction selection. You might think we could reclaim Matcher
3413 // memory PDQ, but actually the Matcher is used in generating spill code.
3414 // Internals of the Matcher (including some VectorSets) must remain live
3415 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3416 // set a bit in reclaimed memory.
3417
3418 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3419 // nodes. Mapping is only valid at the root of each matched subtree.
3420 NOT_PRODUCT( verify_graph_edges(); )
3421
3422 Matcher matcher;
3423 _matcher = &matcher;
3424 {
3425 TracePhase tp("matcher", &timers[_t_matcher]);
3426 matcher.match();
3427 if (failing()) {
3428 return;
3429 }
3430 }
3431 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3432 // nodes. Mapping is only valid at the root of each matched subtree.
3433 NOT_PRODUCT( verify_graph_edges(); )
3434
3435 // If you have too many nodes, or if matching has failed, bail out
3436 check_node_count(0, "out of nodes matching instructions");
3437 if (failing()) {
3438 return;
3439 }
3440
3441 print_method(PHASE_MATCHING, 2);
3442
3443 // Build a proper-looking CFG
3444 PhaseCFG cfg(node_arena(), root(), matcher);
3445 _cfg = &cfg;
3446 {
3447 TracePhase tp("scheduler", &timers[_t_scheduler]);
3448 bool success = cfg.do_global_code_motion();
3449 if (!success) {
3450 return;
3451 }
3452
3453 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3454 NOT_PRODUCT( verify_graph_edges(); )
3455 cfg.verify();
3456 }
3457
3458 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3459 _regalloc = ®alloc;
3460 {
3461 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
3462 // Perform register allocation. After Chaitin, use-def chains are
3463 // no longer accurate (at spill code) and so must be ignored.
3464 // Node->LRG->reg mappings are still accurate.
3465 _regalloc->Register_Allocate();
3466
3467 // Bail out if the allocator builds too many nodes
3468 if (failing()) {
3469 return;
3470 }
3471
3472 print_method(PHASE_REGISTER_ALLOCATION, 2);
3473 }
3474
3475 // Prior to register allocation we kept empty basic blocks in case the
3476 // the allocator needed a place to spill. After register allocation we
3477 // are not adding any new instructions. If any basic block is empty, we
3478 // can now safely remove it.
3479 {
3480 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
3481 cfg.remove_empty_blocks();
3482 if (do_freq_based_layout()) {
3483 PhaseBlockLayout layout(cfg);
3484 } else {
3485 cfg.set_loop_alignment();
3486 }
3487 cfg.fixup_flow();
3488 cfg.remove_unreachable_blocks();
3489 cfg.verify_dominator_tree();
3490 print_method(PHASE_BLOCK_ORDERING, 3);
3491 }
3492
3493 // Apply peephole optimizations
3494 if( OptoPeephole ) {
3495 TracePhase tp("peephole", &timers[_t_peephole]);
3496 PhasePeephole peep( _regalloc, cfg);
3497 peep.do_transform();
3498 print_method(PHASE_PEEPHOLE, 3);
3499 }
3500
3501 // Do late expand if CPU requires this.
3502 if (Matcher::require_postalloc_expand) {
3503 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3504 cfg.postalloc_expand(_regalloc);
3505 print_method(PHASE_POSTALLOC_EXPAND, 3);
3506 }
3507
3508 // Convert Nodes to instruction bits in a buffer
3509 {
3510 TracePhase tp("output", &timers[_t_output]);
3511 PhaseOutput output;
3512 output.Output();
3513 if (failing()) return;
3514 output.install();
3515 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3516 }
3517
3518 // He's dead, Jim.
3519 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3520 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3521 }
3522
3523 //------------------------------Final_Reshape_Counts---------------------------
3524 // This class defines counters to help identify when a method
3525 // may/must be executed using hardware with only 24-bit precision.
3526 struct Final_Reshape_Counts : public StackObj {
3527 int _call_count; // count non-inlined 'common' calls
3528 int _float_count; // count float ops requiring 24-bit precision
3529 int _double_count; // count double ops requiring more precision
3530 int _java_call_count; // count non-inlined 'java' calls
3531 int _inner_loop_count; // count loops which need alignment
3532 VectorSet _visited; // Visitation flags
3533 Node_List _tests; // Set of IfNodes & PCTableNodes
3534
3535 Final_Reshape_Counts() :
3536 _call_count(0), _float_count(0), _double_count(0),
3537 _java_call_count(0), _inner_loop_count(0) { }
3538
3539 void inc_call_count () { _call_count ++; }
3540 void inc_float_count () { _float_count ++; }
3541 void inc_double_count() { _double_count++; }
3542 void inc_java_call_count() { _java_call_count++; }
3543 void inc_inner_loop_count() { _inner_loop_count++; }
3544
3545 int get_call_count () const { return _call_count ; }
3546 int get_float_count () const { return _float_count ; }
3547 int get_double_count() const { return _double_count; }
3548 int get_java_call_count() const { return _java_call_count; }
3549 int get_inner_loop_count() const { return _inner_loop_count; }
3550 };
3551
3552 // Eliminate trivially redundant StoreCMs and accumulate their
3553 // precedence edges.
3554 void Compile::eliminate_redundant_card_marks(Node* n) {
3555 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3556 if (n->in(MemNode::Address)->outcnt() > 1) {
3557 // There are multiple users of the same address so it might be
3558 // possible to eliminate some of the StoreCMs
3559 Node* mem = n->in(MemNode::Memory);
3560 Node* adr = n->in(MemNode::Address);
3561 Node* val = n->in(MemNode::ValueIn);
3562 Node* prev = n;
3563 bool done = false;
3564 // Walk the chain of StoreCMs eliminating ones that match. As
3565 // long as it's a chain of single users then the optimization is
3566 // safe. Eliminating partially redundant StoreCMs would require
3567 // cloning copies down the other paths.
3568 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3569 if (adr == mem->in(MemNode::Address) &&
3570 val == mem->in(MemNode::ValueIn)) {
3571 // redundant StoreCM
3572 if (mem->req() > MemNode::OopStore) {
3573 // Hasn't been processed by this code yet.
3574 n->add_prec(mem->in(MemNode::OopStore));
3575 } else {
3576 // Already converted to precedence edge
3577 for (uint i = mem->req(); i < mem->len(); i++) {
3578 // Accumulate any precedence edges
3579 if (mem->in(i) != nullptr) {
3580 n->add_prec(mem->in(i));
3581 }
3582 }
3583 // Everything above this point has been processed.
3584 done = true;
3585 }
3586 // Eliminate the previous StoreCM
3587 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3588 assert(mem->outcnt() == 0, "should be dead");
3589 mem->disconnect_inputs(this);
3590 } else {
3591 prev = mem;
3592 }
3593 mem = prev->in(MemNode::Memory);
3594 }
3595 }
3596 }
3597
3598
3599 //------------------------------final_graph_reshaping_impl----------------------
3600 // Implement items 1-5 from final_graph_reshaping below.
3601 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3602
3603 if ( n->outcnt() == 0 ) return; // dead node
3604 uint nop = n->Opcode();
3605
3606 // Check for 2-input instruction with "last use" on right input.
3607 // Swap to left input. Implements item (2).
3608 if( n->req() == 3 && // two-input instruction
3609 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3610 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3611 n->in(2)->outcnt() == 1 &&// right use IS a last use
3612 !n->in(2)->is_Con() ) { // right use is not a constant
3613 // Check for commutative opcode
3614 switch( nop ) {
3615 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3616 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3617 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3618 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3619 case Op_AndL: case Op_XorL: case Op_OrL:
3620 case Op_AndI: case Op_XorI: case Op_OrI: {
3621 // Move "last use" input to left by swapping inputs
3622 n->swap_edges(1, 2);
3623 break;
3624 }
3625 default:
3626 break;
3627 }
3628 }
3629
3630 #ifdef ASSERT
3631 if( n->is_Mem() ) {
3632 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3633 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3634 // oop will be recorded in oop map if load crosses safepoint
3635 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3636 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3637 "raw memory operations should have control edge");
3638 }
3639 if (n->is_MemBar()) {
3640 MemBarNode* mb = n->as_MemBar();
3641 if (mb->trailing_store() || mb->trailing_load_store()) {
3642 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3643 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3644 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3645 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3646 } else if (mb->leading()) {
3647 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3648 }
3649 }
3650 #endif
3651 // Count FPU ops and common calls, implements item (3)
3652 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3653 if (!gc_handled) {
3654 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3655 }
3656
3657 // Collect CFG split points
3658 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3659 frc._tests.push(n);
3660 }
3661 }
3662
3663 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3664 switch( nop ) {
3665 // Count all float operations that may use FPU
3666 case Op_AddF:
3667 case Op_SubF:
3668 case Op_MulF:
3669 case Op_DivF:
3670 case Op_NegF:
3671 case Op_ModF:
3672 case Op_ConvI2F:
3673 case Op_ConF:
3674 case Op_CmpF:
3675 case Op_CmpF3:
3676 case Op_StoreF:
3677 case Op_LoadF:
3678 // case Op_ConvL2F: // longs are split into 32-bit halves
3679 frc.inc_float_count();
3680 break;
3681
3682 case Op_ConvF2D:
3683 case Op_ConvD2F:
3684 frc.inc_float_count();
3685 frc.inc_double_count();
3686 break;
3687
3688 // Count all double operations that may use FPU
3689 case Op_AddD:
3690 case Op_SubD:
3691 case Op_MulD:
3692 case Op_DivD:
3693 case Op_NegD:
3694 case Op_ModD:
3695 case Op_ConvI2D:
3696 case Op_ConvD2I:
3697 // case Op_ConvL2D: // handled by leaf call
3698 // case Op_ConvD2L: // handled by leaf call
3699 case Op_ConD:
3700 case Op_CmpD:
3701 case Op_CmpD3:
3702 case Op_StoreD:
3703 case Op_LoadD:
3704 case Op_LoadD_unaligned:
3705 frc.inc_double_count();
3706 break;
3707 case Op_Opaque1: // Remove Opaque Nodes before matching
3708 n->subsume_by(n->in(1), this);
3709 break;
3710 case Op_CallStaticJava:
3711 case Op_CallJava:
3712 case Op_CallDynamicJava:
3713 frc.inc_java_call_count(); // Count java call site;
3714 case Op_CallRuntime:
3715 case Op_CallLeaf:
3716 case Op_CallLeafVector:
3717 case Op_CallLeafNoFP: {
3718 assert (n->is_Call(), "");
3719 CallNode *call = n->as_Call();
3720 // Count call sites where the FP mode bit would have to be flipped.
3721 // Do not count uncommon runtime calls:
3722 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3723 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3724 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3725 frc.inc_call_count(); // Count the call site
3726 } else { // See if uncommon argument is shared
3727 Node *n = call->in(TypeFunc::Parms);
3728 int nop = n->Opcode();
3729 // Clone shared simple arguments to uncommon calls, item (1).
3730 if (n->outcnt() > 1 &&
3731 !n->is_Proj() &&
3732 nop != Op_CreateEx &&
3733 nop != Op_CheckCastPP &&
3734 nop != Op_DecodeN &&
3735 nop != Op_DecodeNKlass &&
3736 !n->is_Mem() &&
3737 !n->is_Phi()) {
3738 Node *x = n->clone();
3739 call->set_req(TypeFunc::Parms, x);
3740 }
3741 }
3742 break;
3743 }
3744
3745 case Op_StoreCM:
3746 {
3747 // Convert OopStore dependence into precedence edge
3748 Node* prec = n->in(MemNode::OopStore);
3749 n->del_req(MemNode::OopStore);
3750 if (prec->is_MergeMem()) {
3751 MergeMemNode* mm = prec->as_MergeMem();
3752 Node* base = mm->base_memory();
3753 for (int i = AliasIdxRaw + 1; i < num_alias_types(); i++) {
3754 const TypePtr* adr_type = get_adr_type(i);
3755 if (adr_type->is_flat()) {
3756 Node* m = mm->memory_at(i);
3757 n->add_prec(m);
3758 }
3759 }
3760 if (mm->outcnt() == 0) {
3761 mm->disconnect_inputs(this);
3762 }
3763 } else {
3764 n->add_prec(prec);
3765 }
3766 eliminate_redundant_card_marks(n);
3767 }
3768
3769 // fall through
3770
3771 case Op_StoreB:
3772 case Op_StoreC:
3773 case Op_StoreI:
3774 case Op_StoreL:
3775 case Op_CompareAndSwapB:
3776 case Op_CompareAndSwapS:
3777 case Op_CompareAndSwapI:
3778 case Op_CompareAndSwapL:
3779 case Op_CompareAndSwapP:
3780 case Op_CompareAndSwapN:
3781 case Op_WeakCompareAndSwapB:
3782 case Op_WeakCompareAndSwapS:
3783 case Op_WeakCompareAndSwapI:
3784 case Op_WeakCompareAndSwapL:
3785 case Op_WeakCompareAndSwapP:
3786 case Op_WeakCompareAndSwapN:
3787 case Op_CompareAndExchangeB:
3788 case Op_CompareAndExchangeS:
3789 case Op_CompareAndExchangeI:
3790 case Op_CompareAndExchangeL:
3791 case Op_CompareAndExchangeP:
3792 case Op_CompareAndExchangeN:
3793 case Op_GetAndAddS:
3794 case Op_GetAndAddB:
3795 case Op_GetAndAddI:
3796 case Op_GetAndAddL:
3797 case Op_GetAndSetS:
3798 case Op_GetAndSetB:
3799 case Op_GetAndSetI:
3800 case Op_GetAndSetL:
3801 case Op_GetAndSetP:
3802 case Op_GetAndSetN:
3803 case Op_StoreP:
3804 case Op_StoreN:
3805 case Op_StoreNKlass:
3806 case Op_LoadB:
3807 case Op_LoadUB:
3808 case Op_LoadUS:
3809 case Op_LoadI:
3810 case Op_LoadKlass:
3811 case Op_LoadNKlass:
3812 case Op_LoadL:
3813 case Op_LoadL_unaligned:
3814 case Op_LoadP:
3815 case Op_LoadN:
3816 case Op_LoadRange:
3817 case Op_LoadS:
3818 break;
3819
3820 case Op_AddP: { // Assert sane base pointers
3821 Node *addp = n->in(AddPNode::Address);
3822 assert( !addp->is_AddP() ||
3823 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3824 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3825 "Base pointers must match (addp %u)", addp->_idx );
3826 #ifdef _LP64
3827 if ((UseCompressedOops || UseCompressedClassPointers) &&
3828 addp->Opcode() == Op_ConP &&
3829 addp == n->in(AddPNode::Base) &&
3830 n->in(AddPNode::Offset)->is_Con()) {
3831 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3832 // on the platform and on the compressed oops mode.
3833 // Use addressing with narrow klass to load with offset on x86.
3834 // Some platforms can use the constant pool to load ConP.
3835 // Do this transformation here since IGVN will convert ConN back to ConP.
3836 const Type* t = addp->bottom_type();
3837 bool is_oop = t->isa_oopptr() != nullptr;
3838 bool is_klass = t->isa_klassptr() != nullptr;
3839
3840 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3841 (is_klass && Matcher::const_klass_prefer_decode())) {
3842 Node* nn = nullptr;
3843
3844 int op = is_oop ? Op_ConN : Op_ConNKlass;
3845
3846 // Look for existing ConN node of the same exact type.
3847 Node* r = root();
3848 uint cnt = r->outcnt();
3849 for (uint i = 0; i < cnt; i++) {
3850 Node* m = r->raw_out(i);
3851 if (m!= nullptr && m->Opcode() == op &&
3852 m->bottom_type()->make_ptr() == t) {
3853 nn = m;
3854 break;
3855 }
3856 }
3857 if (nn != nullptr) {
3858 // Decode a narrow oop to match address
3859 // [R12 + narrow_oop_reg<<3 + offset]
3860 if (is_oop) {
3861 nn = new DecodeNNode(nn, t);
3862 } else {
3863 nn = new DecodeNKlassNode(nn, t);
3864 }
3865 // Check for succeeding AddP which uses the same Base.
3866 // Otherwise we will run into the assertion above when visiting that guy.
3867 for (uint i = 0; i < n->outcnt(); ++i) {
3868 Node *out_i = n->raw_out(i);
3869 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3870 out_i->set_req(AddPNode::Base, nn);
3871 #ifdef ASSERT
3872 for (uint j = 0; j < out_i->outcnt(); ++j) {
3873 Node *out_j = out_i->raw_out(j);
3874 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3875 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3876 }
3877 #endif
3878 }
3879 }
3880 n->set_req(AddPNode::Base, nn);
3881 n->set_req(AddPNode::Address, nn);
3882 if (addp->outcnt() == 0) {
3883 addp->disconnect_inputs(this);
3884 }
3885 }
3886 }
3887 }
3888 #endif
3889 break;
3890 }
3891
3892 case Op_CastPP: {
3893 // Remove CastPP nodes to gain more freedom during scheduling but
3894 // keep the dependency they encode as control or precedence edges
3895 // (if control is set already) on memory operations. Some CastPP
3896 // nodes don't have a control (don't carry a dependency): skip
3897 // those.
3898 if (n->in(0) != nullptr) {
3899 ResourceMark rm;
3900 Unique_Node_List wq;
3901 wq.push(n);
3902 for (uint next = 0; next < wq.size(); ++next) {
3903 Node *m = wq.at(next);
3904 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3905 Node* use = m->fast_out(i);
3906 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3907 use->ensure_control_or_add_prec(n->in(0));
3908 } else {
3909 switch(use->Opcode()) {
3910 case Op_AddP:
3911 case Op_DecodeN:
3912 case Op_DecodeNKlass:
3913 case Op_CheckCastPP:
3914 case Op_CastPP:
3915 wq.push(use);
3916 break;
3917 }
3918 }
3919 }
3920 }
3921 }
3922 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3923 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3924 Node* in1 = n->in(1);
3925 const Type* t = n->bottom_type();
3926 Node* new_in1 = in1->clone();
3927 new_in1->as_DecodeN()->set_type(t);
3928
3929 if (!Matcher::narrow_oop_use_complex_address()) {
3930 //
3931 // x86, ARM and friends can handle 2 adds in addressing mode
3932 // and Matcher can fold a DecodeN node into address by using
3933 // a narrow oop directly and do implicit null check in address:
3934 //
3935 // [R12 + narrow_oop_reg<<3 + offset]
3936 // NullCheck narrow_oop_reg
3937 //
3938 // On other platforms (Sparc) we have to keep new DecodeN node and
3939 // use it to do implicit null check in address:
3940 //
3941 // decode_not_null narrow_oop_reg, base_reg
3942 // [base_reg + offset]
3943 // NullCheck base_reg
3944 //
3945 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3946 // to keep the information to which null check the new DecodeN node
3947 // corresponds to use it as value in implicit_null_check().
3948 //
3949 new_in1->set_req(0, n->in(0));
3950 }
3951
3952 n->subsume_by(new_in1, this);
3953 if (in1->outcnt() == 0) {
3954 in1->disconnect_inputs(this);
3955 }
3956 } else {
3957 n->subsume_by(n->in(1), this);
3958 if (n->outcnt() == 0) {
3959 n->disconnect_inputs(this);
3960 }
3961 }
3962 break;
3963 }
3964 #ifdef _LP64
3965 case Op_CmpP:
3966 // Do this transformation here to preserve CmpPNode::sub() and
3967 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3968 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3969 Node* in1 = n->in(1);
3970 Node* in2 = n->in(2);
3971 if (!in1->is_DecodeNarrowPtr()) {
3972 in2 = in1;
3973 in1 = n->in(2);
3974 }
3975 assert(in1->is_DecodeNarrowPtr(), "sanity");
3976
3977 Node* new_in2 = nullptr;
3978 if (in2->is_DecodeNarrowPtr()) {
3979 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3980 new_in2 = in2->in(1);
3981 } else if (in2->Opcode() == Op_ConP) {
3982 const Type* t = in2->bottom_type();
3983 if (t == TypePtr::NULL_PTR) {
3984 assert(in1->is_DecodeN(), "compare klass to null?");
3985 // Don't convert CmpP null check into CmpN if compressed
3986 // oops implicit null check is not generated.
3987 // This will allow to generate normal oop implicit null check.
3988 if (Matcher::gen_narrow_oop_implicit_null_checks())
3989 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3990 //
3991 // This transformation together with CastPP transformation above
3992 // will generated code for implicit null checks for compressed oops.
3993 //
3994 // The original code after Optimize()
3995 //
3996 // LoadN memory, narrow_oop_reg
3997 // decode narrow_oop_reg, base_reg
3998 // CmpP base_reg, nullptr
3999 // CastPP base_reg // NotNull
4000 // Load [base_reg + offset], val_reg
4001 //
4002 // after these transformations will be
4003 //
4004 // LoadN memory, narrow_oop_reg
4005 // CmpN narrow_oop_reg, nullptr
4006 // decode_not_null narrow_oop_reg, base_reg
4007 // Load [base_reg + offset], val_reg
4008 //
4009 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
4010 // since narrow oops can be used in debug info now (see the code in
4011 // final_graph_reshaping_walk()).
4012 //
4013 // At the end the code will be matched to
4014 // on x86:
4015 //
4016 // Load_narrow_oop memory, narrow_oop_reg
4017 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
4018 // NullCheck narrow_oop_reg
4019 //
4020 // and on sparc:
4021 //
4022 // Load_narrow_oop memory, narrow_oop_reg
4023 // decode_not_null narrow_oop_reg, base_reg
4024 // Load [base_reg + offset], val_reg
4025 // NullCheck base_reg
4026 //
4027 } else if (t->isa_oopptr()) {
4028 new_in2 = ConNode::make(t->make_narrowoop());
4029 } else if (t->isa_klassptr()) {
4030 new_in2 = ConNode::make(t->make_narrowklass());
4031 }
4032 }
4033 if (new_in2 != nullptr) {
4034 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
4035 n->subsume_by(cmpN, this);
4036 if (in1->outcnt() == 0) {
4037 in1->disconnect_inputs(this);
4038 }
4039 if (in2->outcnt() == 0) {
4040 in2->disconnect_inputs(this);
4041 }
4042 }
4043 }
4044 break;
4045
4046 case Op_DecodeN:
4047 case Op_DecodeNKlass:
4048 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4049 // DecodeN could be pinned when it can't be fold into
4050 // an address expression, see the code for Op_CastPP above.
4051 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4052 break;
4053
4054 case Op_EncodeP:
4055 case Op_EncodePKlass: {
4056 Node* in1 = n->in(1);
4057 if (in1->is_DecodeNarrowPtr()) {
4058 n->subsume_by(in1->in(1), this);
4059 } else if (in1->Opcode() == Op_ConP) {
4060 const Type* t = in1->bottom_type();
4061 if (t == TypePtr::NULL_PTR) {
4062 assert(t->isa_oopptr(), "null klass?");
4063 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4064 } else if (t->isa_oopptr()) {
4065 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4066 } else if (t->isa_klassptr()) {
4067 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4068 }
4069 }
4070 if (in1->outcnt() == 0) {
4071 in1->disconnect_inputs(this);
4072 }
4073 break;
4074 }
4075
4076 case Op_Proj: {
4077 if (OptimizeStringConcat || IncrementalInline) {
4078 ProjNode* proj = n->as_Proj();
4079 if (proj->_is_io_use) {
4080 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4081 // Separate projections were used for the exception path which
4082 // are normally removed by a late inline. If it wasn't inlined
4083 // then they will hang around and should just be replaced with
4084 // the original one. Merge them.
4085 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4086 if (non_io_proj != nullptr) {
4087 proj->subsume_by(non_io_proj , this);
4088 }
4089 }
4090 }
4091 break;
4092 }
4093
4094 case Op_Phi:
4095 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4096 // The EncodeP optimization may create Phi with the same edges
4097 // for all paths. It is not handled well by Register Allocator.
4098 Node* unique_in = n->in(1);
4099 assert(unique_in != nullptr, "");
4100 uint cnt = n->req();
4101 for (uint i = 2; i < cnt; i++) {
4102 Node* m = n->in(i);
4103 assert(m != nullptr, "");
4104 if (unique_in != m)
4105 unique_in = nullptr;
4106 }
4107 if (unique_in != nullptr) {
4108 n->subsume_by(unique_in, this);
4109 }
4110 }
4111 break;
4112
4113 #endif
4114
4115 #ifdef ASSERT
4116 case Op_CastII:
4117 // Verify that all range check dependent CastII nodes were removed.
4118 if (n->isa_CastII()->has_range_check()) {
4119 n->dump(3);
4120 assert(false, "Range check dependent CastII node was not removed");
4121 }
4122 break;
4123 #endif
4124
4125 case Op_ModI:
4126 if (UseDivMod) {
4127 // Check if a%b and a/b both exist
4128 Node* d = n->find_similar(Op_DivI);
4129 if (d) {
4130 // Replace them with a fused divmod if supported
4131 if (Matcher::has_match_rule(Op_DivModI)) {
4132 DivModINode* divmod = DivModINode::make(n);
4133 d->subsume_by(divmod->div_proj(), this);
4134 n->subsume_by(divmod->mod_proj(), this);
4135 } else {
4136 // replace a%b with a-((a/b)*b)
4137 Node* mult = new MulINode(d, d->in(2));
4138 Node* sub = new SubINode(d->in(1), mult);
4139 n->subsume_by(sub, this);
4140 }
4141 }
4142 }
4143 break;
4144
4145 case Op_ModL:
4146 if (UseDivMod) {
4147 // Check if a%b and a/b both exist
4148 Node* d = n->find_similar(Op_DivL);
4149 if (d) {
4150 // Replace them with a fused divmod if supported
4151 if (Matcher::has_match_rule(Op_DivModL)) {
4152 DivModLNode* divmod = DivModLNode::make(n);
4153 d->subsume_by(divmod->div_proj(), this);
4154 n->subsume_by(divmod->mod_proj(), this);
4155 } else {
4156 // replace a%b with a-((a/b)*b)
4157 Node* mult = new MulLNode(d, d->in(2));
4158 Node* sub = new SubLNode(d->in(1), mult);
4159 n->subsume_by(sub, this);
4160 }
4161 }
4162 }
4163 break;
4164
4165 case Op_UModI:
4166 if (UseDivMod) {
4167 // Check if a%b and a/b both exist
4168 Node* d = n->find_similar(Op_UDivI);
4169 if (d) {
4170 // Replace them with a fused unsigned divmod if supported
4171 if (Matcher::has_match_rule(Op_UDivModI)) {
4172 UDivModINode* divmod = UDivModINode::make(n);
4173 d->subsume_by(divmod->div_proj(), this);
4174 n->subsume_by(divmod->mod_proj(), this);
4175 } else {
4176 // replace a%b with a-((a/b)*b)
4177 Node* mult = new MulINode(d, d->in(2));
4178 Node* sub = new SubINode(d->in(1), mult);
4179 n->subsume_by(sub, this);
4180 }
4181 }
4182 }
4183 break;
4184
4185 case Op_UModL:
4186 if (UseDivMod) {
4187 // Check if a%b and a/b both exist
4188 Node* d = n->find_similar(Op_UDivL);
4189 if (d) {
4190 // Replace them with a fused unsigned divmod if supported
4191 if (Matcher::has_match_rule(Op_UDivModL)) {
4192 UDivModLNode* divmod = UDivModLNode::make(n);
4193 d->subsume_by(divmod->div_proj(), this);
4194 n->subsume_by(divmod->mod_proj(), this);
4195 } else {
4196 // replace a%b with a-((a/b)*b)
4197 Node* mult = new MulLNode(d, d->in(2));
4198 Node* sub = new SubLNode(d->in(1), mult);
4199 n->subsume_by(sub, this);
4200 }
4201 }
4202 }
4203 break;
4204
4205 case Op_LoadVector:
4206 case Op_StoreVector:
4207 #ifdef ASSERT
4208 // Add VerifyVectorAlignment node between adr and load / store.
4209 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
4210 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
4211 n->as_StoreVector()->must_verify_alignment();
4212 if (must_verify_alignment) {
4213 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
4214 n->as_StoreVector()->memory_size();
4215 // The memory access should be aligned to the vector width in bytes.
4216 // However, the underlying array is possibly less well aligned, but at least
4217 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
4218 // a loop we can expect at least the following alignment:
4219 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
4220 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
4221 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
4222 // Create mask from alignment. e.g. 0b1000 -> 0b0111
4223 jlong mask = guaranteed_alignment - 1;
4224 Node* mask_con = ConLNode::make(mask);
4225 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
4226 n->set_req(MemNode::Address, va);
4227 }
4228 }
4229 #endif
4230 break;
4231
4232 case Op_LoadVectorGather:
4233 case Op_StoreVectorScatter:
4234 case Op_LoadVectorGatherMasked:
4235 case Op_StoreVectorScatterMasked:
4236 case Op_VectorCmpMasked:
4237 case Op_VectorMaskGen:
4238 case Op_LoadVectorMasked:
4239 case Op_StoreVectorMasked:
4240 break;
4241
4242 case Op_AddReductionVI:
4243 case Op_AddReductionVL:
4244 case Op_AddReductionVF:
4245 case Op_AddReductionVD:
4246 case Op_MulReductionVI:
4247 case Op_MulReductionVL:
4248 case Op_MulReductionVF:
4249 case Op_MulReductionVD:
4250 case Op_MinReductionV:
4251 case Op_MaxReductionV:
4252 case Op_AndReductionV:
4253 case Op_OrReductionV:
4254 case Op_XorReductionV:
4255 break;
4256
4257 case Op_PackB:
4258 case Op_PackS:
4259 case Op_PackI:
4260 case Op_PackF:
4261 case Op_PackL:
4262 case Op_PackD:
4263 if (n->req()-1 > 2) {
4264 // Replace many operand PackNodes with a binary tree for matching
4265 PackNode* p = (PackNode*) n;
4266 Node* btp = p->binary_tree_pack(1, n->req());
4267 n->subsume_by(btp, this);
4268 }
4269 break;
4270 case Op_Loop:
4271 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4272 case Op_CountedLoop:
4273 case Op_LongCountedLoop:
4274 case Op_OuterStripMinedLoop:
4275 if (n->as_Loop()->is_inner_loop()) {
4276 frc.inc_inner_loop_count();
4277 }
4278 n->as_Loop()->verify_strip_mined(0);
4279 break;
4280 case Op_LShiftI:
4281 case Op_RShiftI:
4282 case Op_URShiftI:
4283 case Op_LShiftL:
4284 case Op_RShiftL:
4285 case Op_URShiftL:
4286 if (Matcher::need_masked_shift_count) {
4287 // The cpu's shift instructions don't restrict the count to the
4288 // lower 5/6 bits. We need to do the masking ourselves.
4289 Node* in2 = n->in(2);
4290 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4291 const TypeInt* t = in2->find_int_type();
4292 if (t != nullptr && t->is_con()) {
4293 juint shift = t->get_con();
4294 if (shift > mask) { // Unsigned cmp
4295 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4296 }
4297 } else {
4298 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
4299 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4300 n->set_req(2, shift);
4301 }
4302 }
4303 if (in2->outcnt() == 0) { // Remove dead node
4304 in2->disconnect_inputs(this);
4305 }
4306 }
4307 break;
4308 case Op_MemBarStoreStore:
4309 case Op_MemBarRelease:
4310 // Break the link with AllocateNode: it is no longer useful and
4311 // confuses register allocation.
4312 if (n->req() > MemBarNode::Precedent) {
4313 n->set_req(MemBarNode::Precedent, top());
4314 }
4315 break;
4316 case Op_MemBarAcquire: {
4317 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4318 // At parse time, the trailing MemBarAcquire for a volatile load
4319 // is created with an edge to the load. After optimizations,
4320 // that input may be a chain of Phis. If those phis have no
4321 // other use, then the MemBarAcquire keeps them alive and
4322 // register allocation can be confused.
4323 dead_nodes.push(n->in(MemBarNode::Precedent));
4324 n->set_req(MemBarNode::Precedent, top());
4325 }
4326 break;
4327 }
4328 case Op_Blackhole:
4329 break;
4330 case Op_RangeCheck: {
4331 RangeCheckNode* rc = n->as_RangeCheck();
4332 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4333 n->subsume_by(iff, this);
4334 frc._tests.push(iff);
4335 break;
4336 }
4337 case Op_ConvI2L: {
4338 if (!Matcher::convi2l_type_required) {
4339 // Code generation on some platforms doesn't need accurate
4340 // ConvI2L types. Widening the type can help remove redundant
4341 // address computations.
4342 n->as_Type()->set_type(TypeLong::INT);
4343 ResourceMark rm;
4344 Unique_Node_List wq;
4345 wq.push(n);
4346 for (uint next = 0; next < wq.size(); next++) {
4347 Node *m = wq.at(next);
4348
4349 for(;;) {
4350 // Loop over all nodes with identical inputs edges as m
4351 Node* k = m->find_similar(m->Opcode());
4352 if (k == nullptr) {
4353 break;
4354 }
4355 // Push their uses so we get a chance to remove node made
4356 // redundant
4357 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4358 Node* u = k->fast_out(i);
4359 if (u->Opcode() == Op_LShiftL ||
4360 u->Opcode() == Op_AddL ||
4361 u->Opcode() == Op_SubL ||
4362 u->Opcode() == Op_AddP) {
4363 wq.push(u);
4364 }
4365 }
4366 // Replace all nodes with identical edges as m with m
4367 k->subsume_by(m, this);
4368 }
4369 }
4370 }
4371 break;
4372 }
4373 case Op_CmpUL: {
4374 if (!Matcher::has_match_rule(Op_CmpUL)) {
4375 // No support for unsigned long comparisons
4376 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4377 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4378 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4379 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4380 Node* andl = new AndLNode(orl, remove_sign_mask);
4381 Node* cmp = new CmpLNode(andl, n->in(2));
4382 n->subsume_by(cmp, this);
4383 }
4384 break;
4385 }
4386 #ifdef ASSERT
4387 case Op_InlineType: {
4388 n->dump(-1);
4389 assert(false, "inline type node was not removed");
4390 break;
4391 }
4392 #endif
4393 default:
4394 assert(!n->is_Call(), "");
4395 assert(!n->is_Mem(), "");
4396 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4397 break;
4398 }
4399 }
4400
4401 //------------------------------final_graph_reshaping_walk---------------------
4402 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4403 // requires that the walk visits a node's inputs before visiting the node.
4404 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4405 Unique_Node_List sfpt;
4406
4407 frc._visited.set(root->_idx); // first, mark node as visited
4408 uint cnt = root->req();
4409 Node *n = root;
4410 uint i = 0;
4411 while (true) {
4412 if (i < cnt) {
4413 // Place all non-visited non-null inputs onto stack
4414 Node* m = n->in(i);
4415 ++i;
4416 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
4417 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
4418 // compute worst case interpreter size in case of a deoptimization
4419 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4420
4421 sfpt.push(m);
4422 }
4423 cnt = m->req();
4424 nstack.push(n, i); // put on stack parent and next input's index
4425 n = m;
4426 i = 0;
4427 }
4428 } else {
4429 // Now do post-visit work
4430 final_graph_reshaping_impl(n, frc, dead_nodes);
4431 if (nstack.is_empty())
4432 break; // finished
4433 n = nstack.node(); // Get node from stack
4434 cnt = n->req();
4435 i = nstack.index();
4436 nstack.pop(); // Shift to the next node on stack
4437 }
4438 }
4439
4440 // Skip next transformation if compressed oops are not used.
4441 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4442 (!UseCompressedOops && !UseCompressedClassPointers))
4443 return;
4444
4445 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4446 // It could be done for an uncommon traps or any safepoints/calls
4447 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4448 while (sfpt.size() > 0) {
4449 n = sfpt.pop();
4450 JVMState *jvms = n->as_SafePoint()->jvms();
4451 assert(jvms != nullptr, "sanity");
4452 int start = jvms->debug_start();
4453 int end = n->req();
4454 bool is_uncommon = (n->is_CallStaticJava() &&
4455 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4456 for (int j = start; j < end; j++) {
4457 Node* in = n->in(j);
4458 if (in->is_DecodeNarrowPtr()) {
4459 bool safe_to_skip = true;
4460 if (!is_uncommon ) {
4461 // Is it safe to skip?
4462 for (uint i = 0; i < in->outcnt(); i++) {
4463 Node* u = in->raw_out(i);
4464 if (!u->is_SafePoint() ||
4465 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4466 safe_to_skip = false;
4467 }
4468 }
4469 }
4470 if (safe_to_skip) {
4471 n->set_req(j, in->in(1));
4472 }
4473 if (in->outcnt() == 0) {
4474 in->disconnect_inputs(this);
4475 }
4476 }
4477 }
4478 }
4479 }
4480
4481 //------------------------------final_graph_reshaping--------------------------
4482 // Final Graph Reshaping.
4483 //
4484 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4485 // and not commoned up and forced early. Must come after regular
4486 // optimizations to avoid GVN undoing the cloning. Clone constant
4487 // inputs to Loop Phis; these will be split by the allocator anyways.
4488 // Remove Opaque nodes.
4489 // (2) Move last-uses by commutative operations to the left input to encourage
4490 // Intel update-in-place two-address operations and better register usage
4491 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4492 // calls canonicalizing them back.
4493 // (3) Count the number of double-precision FP ops, single-precision FP ops
4494 // and call sites. On Intel, we can get correct rounding either by
4495 // forcing singles to memory (requires extra stores and loads after each
4496 // FP bytecode) or we can set a rounding mode bit (requires setting and
4497 // clearing the mode bit around call sites). The mode bit is only used
4498 // if the relative frequency of single FP ops to calls is low enough.
4499 // This is a key transform for SPEC mpeg_audio.
4500 // (4) Detect infinite loops; blobs of code reachable from above but not
4501 // below. Several of the Code_Gen algorithms fail on such code shapes,
4502 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4503 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4504 // Detection is by looking for IfNodes where only 1 projection is
4505 // reachable from below or CatchNodes missing some targets.
4506 // (5) Assert for insane oop offsets in debug mode.
4507
4508 bool Compile::final_graph_reshaping() {
4509 // an infinite loop may have been eliminated by the optimizer,
4510 // in which case the graph will be empty.
4511 if (root()->req() == 1) {
4512 // Do not compile method that is only a trivial infinite loop,
4513 // since the content of the loop may have been eliminated.
4514 record_method_not_compilable("trivial infinite loop");
4515 return true;
4516 }
4517
4518 // Expensive nodes have their control input set to prevent the GVN
4519 // from freely commoning them. There's no GVN beyond this point so
4520 // no need to keep the control input. We want the expensive nodes to
4521 // be freely moved to the least frequent code path by gcm.
4522 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4523 for (int i = 0; i < expensive_count(); i++) {
4524 _expensive_nodes.at(i)->set_req(0, nullptr);
4525 }
4526
4527 Final_Reshape_Counts frc;
4528
4529 // Visit everybody reachable!
4530 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4531 Node_Stack nstack(live_nodes() >> 1);
4532 Unique_Node_List dead_nodes;
4533 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4534
4535 // Check for unreachable (from below) code (i.e., infinite loops).
4536 for( uint i = 0; i < frc._tests.size(); i++ ) {
4537 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4538 // Get number of CFG targets.
4539 // Note that PCTables include exception targets after calls.
4540 uint required_outcnt = n->required_outcnt();
4541 if (n->outcnt() != required_outcnt) {
4542 // Check for a few special cases. Rethrow Nodes never take the
4543 // 'fall-thru' path, so expected kids is 1 less.
4544 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4545 if (n->in(0)->in(0)->is_Call()) {
4546 CallNode* call = n->in(0)->in(0)->as_Call();
4547 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4548 required_outcnt--; // Rethrow always has 1 less kid
4549 } else if (call->req() > TypeFunc::Parms &&
4550 call->is_CallDynamicJava()) {
4551 // Check for null receiver. In such case, the optimizer has
4552 // detected that the virtual call will always result in a null
4553 // pointer exception. The fall-through projection of this CatchNode
4554 // will not be populated.
4555 Node* arg0 = call->in(TypeFunc::Parms);
4556 if (arg0->is_Type() &&
4557 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4558 required_outcnt--;
4559 }
4560 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4561 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4562 // Check for illegal array length. In such case, the optimizer has
4563 // detected that the allocation attempt will always result in an
4564 // exception. There is no fall-through projection of this CatchNode .
4565 assert(call->is_CallStaticJava(), "static call expected");
4566 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4567 uint valid_length_test_input = call->req() - 1;
4568 Node* valid_length_test = call->in(valid_length_test_input);
4569 call->del_req(valid_length_test_input);
4570 if (valid_length_test->find_int_con(1) == 0) {
4571 required_outcnt--;
4572 }
4573 dead_nodes.push(valid_length_test);
4574 assert(n->outcnt() == required_outcnt, "malformed control flow");
4575 continue;
4576 }
4577 }
4578 }
4579
4580 // Recheck with a better notion of 'required_outcnt'
4581 if (n->outcnt() != required_outcnt) {
4582 record_method_not_compilable("malformed control flow");
4583 return true; // Not all targets reachable!
4584 }
4585 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4586 CallNode* call = n->in(0)->in(0)->as_Call();
4587 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4588 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4589 assert(call->is_CallStaticJava(), "static call expected");
4590 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4591 uint valid_length_test_input = call->req() - 1;
4592 dead_nodes.push(call->in(valid_length_test_input));
4593 call->del_req(valid_length_test_input); // valid length test useless now
4594 }
4595 }
4596 // Check that I actually visited all kids. Unreached kids
4597 // must be infinite loops.
4598 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4599 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4600 record_method_not_compilable("infinite loop");
4601 return true; // Found unvisited kid; must be unreach
4602 }
4603
4604 // Here so verification code in final_graph_reshaping_walk()
4605 // always see an OuterStripMinedLoopEnd
4606 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4607 IfNode* init_iff = n->as_If();
4608 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4609 n->subsume_by(iff, this);
4610 }
4611 }
4612
4613 while (dead_nodes.size() > 0) {
4614 Node* m = dead_nodes.pop();
4615 if (m->outcnt() == 0 && m != top()) {
4616 for (uint j = 0; j < m->req(); j++) {
4617 Node* in = m->in(j);
4618 if (in != nullptr) {
4619 dead_nodes.push(in);
4620 }
4621 }
4622 m->disconnect_inputs(this);
4623 }
4624 }
4625
4626 #ifdef IA32
4627 // If original bytecodes contained a mixture of floats and doubles
4628 // check if the optimizer has made it homogeneous, item (3).
4629 if (UseSSE == 0 &&
4630 frc.get_float_count() > 32 &&
4631 frc.get_double_count() == 0 &&
4632 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4633 set_24_bit_selection_and_mode(false, true);
4634 }
4635 #endif // IA32
4636
4637 set_java_calls(frc.get_java_call_count());
4638 set_inner_loops(frc.get_inner_loop_count());
4639
4640 // No infinite loops, no reason to bail out.
4641 return false;
4642 }
4643
4644 //-----------------------------too_many_traps----------------------------------
4645 // Report if there are too many traps at the current method and bci.
4646 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4647 bool Compile::too_many_traps(ciMethod* method,
4648 int bci,
4649 Deoptimization::DeoptReason reason) {
4650 ciMethodData* md = method->method_data();
4651 if (md->is_empty()) {
4652 // Assume the trap has not occurred, or that it occurred only
4653 // because of a transient condition during start-up in the interpreter.
4654 return false;
4655 }
4656 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4657 if (md->has_trap_at(bci, m, reason) != 0) {
4658 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4659 // Also, if there are multiple reasons, or if there is no per-BCI record,
4660 // assume the worst.
4661 if (log())
4662 log()->elem("observe trap='%s' count='%d'",
4663 Deoptimization::trap_reason_name(reason),
4664 md->trap_count(reason));
4665 return true;
4666 } else {
4667 // Ignore method/bci and see if there have been too many globally.
4668 return too_many_traps(reason, md);
4669 }
4670 }
4671
4672 // Less-accurate variant which does not require a method and bci.
4673 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4674 ciMethodData* logmd) {
4675 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4676 // Too many traps globally.
4677 // Note that we use cumulative trap_count, not just md->trap_count.
4678 if (log()) {
4679 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4680 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4681 Deoptimization::trap_reason_name(reason),
4682 mcount, trap_count(reason));
4683 }
4684 return true;
4685 } else {
4686 // The coast is clear.
4687 return false;
4688 }
4689 }
4690
4691 //--------------------------too_many_recompiles--------------------------------
4692 // Report if there are too many recompiles at the current method and bci.
4693 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4694 // Is not eager to return true, since this will cause the compiler to use
4695 // Action_none for a trap point, to avoid too many recompilations.
4696 bool Compile::too_many_recompiles(ciMethod* method,
4697 int bci,
4698 Deoptimization::DeoptReason reason) {
4699 ciMethodData* md = method->method_data();
4700 if (md->is_empty()) {
4701 // Assume the trap has not occurred, or that it occurred only
4702 // because of a transient condition during start-up in the interpreter.
4703 return false;
4704 }
4705 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4706 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4707 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4708 Deoptimization::DeoptReason per_bc_reason
4709 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4710 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4711 if ((per_bc_reason == Deoptimization::Reason_none
4712 || md->has_trap_at(bci, m, reason) != 0)
4713 // The trap frequency measure we care about is the recompile count:
4714 && md->trap_recompiled_at(bci, m)
4715 && md->overflow_recompile_count() >= bc_cutoff) {
4716 // Do not emit a trap here if it has already caused recompilations.
4717 // Also, if there are multiple reasons, or if there is no per-BCI record,
4718 // assume the worst.
4719 if (log())
4720 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4721 Deoptimization::trap_reason_name(reason),
4722 md->trap_count(reason),
4723 md->overflow_recompile_count());
4724 return true;
4725 } else if (trap_count(reason) != 0
4726 && decompile_count() >= m_cutoff) {
4727 // Too many recompiles globally, and we have seen this sort of trap.
4728 // Use cumulative decompile_count, not just md->decompile_count.
4729 if (log())
4730 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4731 Deoptimization::trap_reason_name(reason),
4732 md->trap_count(reason), trap_count(reason),
4733 md->decompile_count(), decompile_count());
4734 return true;
4735 } else {
4736 // The coast is clear.
4737 return false;
4738 }
4739 }
4740
4741 // Compute when not to trap. Used by matching trap based nodes and
4742 // NullCheck optimization.
4743 void Compile::set_allowed_deopt_reasons() {
4744 _allowed_reasons = 0;
4745 if (is_method_compilation()) {
4746 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4747 assert(rs < BitsPerInt, "recode bit map");
4748 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4749 _allowed_reasons |= nth_bit(rs);
4750 }
4751 }
4752 }
4753 }
4754
4755 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4756 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4757 }
4758
4759 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4760 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4761 }
4762
4763 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4764 if (holder->is_initialized()) {
4765 return false;
4766 }
4767 if (holder->is_being_initialized()) {
4768 if (accessing_method->holder() == holder) {
4769 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4770 // <init>, or a static method. In all those cases, there was an initialization
4771 // barrier on the holder klass passed.
4772 if (accessing_method->is_class_initializer() ||
4773 accessing_method->is_object_constructor() ||
4774 accessing_method->is_static()) {
4775 return false;
4776 }
4777 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4778 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4779 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4780 // child class can become fully initialized while its parent class is still being initialized.
4781 if (accessing_method->is_class_initializer()) {
4782 return false;
4783 }
4784 }
4785 ciMethod* root = method(); // the root method of compilation
4786 if (root != accessing_method) {
4787 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4788 }
4789 }
4790 return true;
4791 }
4792
4793 #ifndef PRODUCT
4794 //------------------------------verify_bidirectional_edges---------------------
4795 // For each input edge to a node (ie - for each Use-Def edge), verify that
4796 // there is a corresponding Def-Use edge.
4797 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4798 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4799 uint stack_size = live_nodes() >> 4;
4800 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4801 nstack.push(_root);
4802
4803 while (nstack.size() > 0) {
4804 Node* n = nstack.pop();
4805 if (visited.member(n)) {
4806 continue;
4807 }
4808 visited.push(n);
4809
4810 // Walk over all input edges, checking for correspondence
4811 uint length = n->len();
4812 for (uint i = 0; i < length; i++) {
4813 Node* in = n->in(i);
4814 if (in != nullptr && !visited.member(in)) {
4815 nstack.push(in); // Put it on stack
4816 }
4817 if (in != nullptr && !in->is_top()) {
4818 // Count instances of `next`
4819 int cnt = 0;
4820 for (uint idx = 0; idx < in->_outcnt; idx++) {
4821 if (in->_out[idx] == n) {
4822 cnt++;
4823 }
4824 }
4825 assert(cnt > 0, "Failed to find Def-Use edge.");
4826 // Check for duplicate edges
4827 // walk the input array downcounting the input edges to n
4828 for (uint j = 0; j < length; j++) {
4829 if (n->in(j) == in) {
4830 cnt--;
4831 }
4832 }
4833 assert(cnt == 0, "Mismatched edge count.");
4834 } else if (in == nullptr) {
4835 assert(i == 0 || i >= n->req() ||
4836 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4837 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
4838 (n->is_Unlock() && i == (n->req() - 1)) ||
4839 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4840 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
4841 } else {
4842 assert(in->is_top(), "sanity");
4843 // Nothing to check.
4844 }
4845 }
4846 }
4847 }
4848
4849 //------------------------------verify_graph_edges---------------------------
4850 // Walk the Graph and verify that there is a one-to-one correspondence
4851 // between Use-Def edges and Def-Use edges in the graph.
4852 void Compile::verify_graph_edges(bool no_dead_code) {
4853 if (VerifyGraphEdges) {
4854 Unique_Node_List visited;
4855
4856 // Call graph walk to check edges
4857 verify_bidirectional_edges(visited);
4858 if (no_dead_code) {
4859 // Now make sure that no visited node is used by an unvisited node.
4860 bool dead_nodes = false;
4861 Unique_Node_List checked;
4862 while (visited.size() > 0) {
4863 Node* n = visited.pop();
4864 checked.push(n);
4865 for (uint i = 0; i < n->outcnt(); i++) {
4866 Node* use = n->raw_out(i);
4867 if (checked.member(use)) continue; // already checked
4868 if (visited.member(use)) continue; // already in the graph
4869 if (use->is_Con()) continue; // a dead ConNode is OK
4870 // At this point, we have found a dead node which is DU-reachable.
4871 if (!dead_nodes) {
4872 tty->print_cr("*** Dead nodes reachable via DU edges:");
4873 dead_nodes = true;
4874 }
4875 use->dump(2);
4876 tty->print_cr("---");
4877 checked.push(use); // No repeats; pretend it is now checked.
4878 }
4879 }
4880 assert(!dead_nodes, "using nodes must be reachable from root");
4881 }
4882 }
4883 }
4884 #endif
4885
4886 // The Compile object keeps track of failure reasons separately from the ciEnv.
4887 // This is required because there is not quite a 1-1 relation between the
4888 // ciEnv and its compilation task and the Compile object. Note that one
4889 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4890 // to backtrack and retry without subsuming loads. Other than this backtracking
4891 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4892 // by the logic in C2Compiler.
4893 void Compile::record_failure(const char* reason) {
4894 if (log() != nullptr) {
4895 log()->elem("failure reason='%s' phase='compile'", reason);
4896 }
4897 if (_failure_reason.get() == nullptr) {
4898 // Record the first failure reason.
4899 _failure_reason.set(reason);
4900 if (CaptureBailoutInformation) {
4901 _first_failure_details = new CompilationFailureInfo(reason);
4902 }
4903 }
4904
4905 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4906 C->print_method(PHASE_FAILURE, 1);
4907 }
4908 _root = nullptr; // flush the graph, too
4909 }
4910
4911 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4912 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4913 _compile(Compile::current()),
4914 _log(nullptr),
4915 _phase_name(name),
4916 _dolog(CITimeVerbose)
4917 {
4918 assert(_compile != nullptr, "sanity check");
4919 if (_dolog) {
4920 _log = _compile->log();
4921 }
4922 if (_log != nullptr) {
4923 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4924 _log->stamp();
4925 _log->end_head();
4926 }
4927 }
4928
4929 Compile::TracePhase::~TracePhase() {
4930 if (_compile->failing()) return;
4931 #ifdef ASSERT
4932 if (PrintIdealNodeCount) {
4933 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4934 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4935 }
4936
4937 if (VerifyIdealNodeCount) {
4938 _compile->print_missing_nodes();
4939 }
4940 #endif
4941
4942 if (_log != nullptr) {
4943 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes());
4944 }
4945 }
4946
4947 //----------------------------static_subtype_check-----------------------------
4948 // Shortcut important common cases when superklass is exact:
4949 // (0) superklass is java.lang.Object (can occur in reflective code)
4950 // (1) subklass is already limited to a subtype of superklass => always ok
4951 // (2) subklass does not overlap with superklass => always fail
4952 // (3) superklass has NO subtypes and we can check with a simple compare.
4953 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4954 if (skip) {
4955 return SSC_full_test; // Let caller generate the general case.
4956 }
4957
4958 if (subk->is_java_subtype_of(superk)) {
4959 return SSC_always_true; // (0) and (1) this test cannot fail
4960 }
4961
4962 if (!subk->maybe_java_subtype_of(superk)) {
4963 return SSC_always_false; // (2) true path dead; no dynamic test needed
4964 }
4965
4966 const Type* superelem = superk;
4967 if (superk->isa_aryklassptr()) {
4968 int ignored;
4969 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4970
4971 // Do not fold the subtype check to an array klass pointer comparison for null-able inline type arrays
4972 // because null-free [LMyValue <: null-able [LMyValue but the klasses are different. Perform a full test.
4973 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
4974 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
4975 return SSC_full_test;
4976 }
4977 }
4978
4979 if (superelem->isa_instklassptr()) {
4980 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4981 if (!ik->has_subklass()) {
4982 if (!ik->is_final()) {
4983 // Add a dependency if there is a chance of a later subclass.
4984 dependencies()->assert_leaf_type(ik);
4985 }
4986 if (!superk->maybe_java_subtype_of(subk)) {
4987 return SSC_always_false;
4988 }
4989 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4990 }
4991 } else {
4992 // A primitive array type has no subtypes.
4993 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4994 }
4995
4996 return SSC_full_test;
4997 }
4998
4999 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
5000 #ifdef _LP64
5001 // The scaled index operand to AddP must be a clean 64-bit value.
5002 // Java allows a 32-bit int to be incremented to a negative
5003 // value, which appears in a 64-bit register as a large
5004 // positive number. Using that large positive number as an
5005 // operand in pointer arithmetic has bad consequences.
5006 // On the other hand, 32-bit overflow is rare, and the possibility
5007 // can often be excluded, if we annotate the ConvI2L node with
5008 // a type assertion that its value is known to be a small positive
5009 // number. (The prior range check has ensured this.)
5010 // This assertion is used by ConvI2LNode::Ideal.
5011 int index_max = max_jint - 1; // array size is max_jint, index is one less
5012 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
5013 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
5014 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
5015 #endif
5016 return idx;
5017 }
5018
5019 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
5020 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
5021 if (ctrl != nullptr) {
5022 // Express control dependency by a CastII node with a narrow type.
5023 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
5024 // node from floating above the range check during loop optimizations. Otherwise, the
5025 // ConvI2L node may be eliminated independently of the range check, causing the data path
5026 // to become TOP while the control path is still there (although it's unreachable).
5027 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
5028 value = phase->transform(value);
5029 }
5030 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
5031 return phase->transform(new ConvI2LNode(value, ltype));
5032 }
5033
5034 // The message about the current inlining is accumulated in
5035 // _print_inlining_stream and transferred into the _print_inlining_list
5036 // once we know whether inlining succeeds or not. For regular
5037 // inlining, messages are appended to the buffer pointed by
5038 // _print_inlining_idx in the _print_inlining_list. For late inlining,
5039 // a new buffer is added after _print_inlining_idx in the list. This
5040 // way we can update the inlining message for late inlining call site
5041 // when the inlining is attempted again.
5042 void Compile::print_inlining_init() {
5043 if (print_inlining() || print_intrinsics()) {
5044 // print_inlining_init is actually called several times.
5045 print_inlining_reset();
5046 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
5047 }
5048 }
5049
5050 void Compile::print_inlining_reinit() {
5051 if (print_inlining() || print_intrinsics()) {
5052 print_inlining_reset();
5053 }
5054 }
5055
5056 void Compile::print_inlining_reset() {
5057 _print_inlining_stream->reset();
5058 }
5059
5060 void Compile::print_inlining_commit() {
5061 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
5062 // Transfer the message from _print_inlining_stream to the current
5063 // _print_inlining_list buffer and clear _print_inlining_stream.
5064 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
5065 print_inlining_reset();
5066 }
5067
5068 void Compile::print_inlining_push() {
5069 // Add new buffer to the _print_inlining_list at current position
5070 _print_inlining_idx++;
5071 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
5072 }
5073
5074 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
5075 return _print_inlining_list->at(_print_inlining_idx);
5076 }
5077
5078 void Compile::print_inlining_update(CallGenerator* cg) {
5079 if (print_inlining() || print_intrinsics()) {
5080 if (cg->is_late_inline()) {
5081 if (print_inlining_current()->cg() != cg &&
5082 (print_inlining_current()->cg() != nullptr ||
5083 print_inlining_current()->ss()->size() != 0)) {
5084 print_inlining_push();
5085 }
5086 print_inlining_commit();
5087 print_inlining_current()->set_cg(cg);
5088 } else {
5089 if (print_inlining_current()->cg() != nullptr) {
5090 print_inlining_push();
5091 }
5092 print_inlining_commit();
5093 }
5094 }
5095 }
5096
5097 void Compile::print_inlining_move_to(CallGenerator* cg) {
5098 // We resume inlining at a late inlining call site. Locate the
5099 // corresponding inlining buffer so that we can update it.
5100 if (print_inlining() || print_intrinsics()) {
5101 for (int i = 0; i < _print_inlining_list->length(); i++) {
5102 if (_print_inlining_list->at(i)->cg() == cg) {
5103 _print_inlining_idx = i;
5104 return;
5105 }
5106 }
5107 ShouldNotReachHere();
5108 }
5109 }
5110
5111 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
5112 if (print_inlining() || print_intrinsics()) {
5113 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
5114 assert(print_inlining_current()->cg() == cg, "wrong entry");
5115 // replace message with new message
5116 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
5117 print_inlining_commit();
5118 print_inlining_current()->set_cg(cg);
5119 }
5120 }
5121
5122 void Compile::print_inlining_assert_ready() {
5123 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
5124 }
5125
5126 void Compile::process_print_inlining() {
5127 assert(_late_inlines.length() == 0, "not drained yet");
5128 if (print_inlining() || print_intrinsics()) {
5129 ResourceMark rm;
5130 stringStream ss;
5131 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once.");
5132 for (int i = 0; i < _print_inlining_list->length(); i++) {
5133 PrintInliningBuffer* pib = _print_inlining_list->at(i);
5134 ss.print("%s", pib->ss()->freeze());
5135 delete pib;
5136 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr));
5137 }
5138 // Reset _print_inlining_list, it only contains destructed objects.
5139 // It is on the arena, so it will be freed when the arena is reset.
5140 _print_inlining_list = nullptr;
5141 // _print_inlining_stream won't be used anymore, either.
5142 print_inlining_reset();
5143 size_t end = ss.size();
5144 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
5145 strncpy(_print_inlining_output, ss.freeze(), end+1);
5146 _print_inlining_output[end] = 0;
5147 }
5148 }
5149
5150 void Compile::dump_print_inlining() {
5151 if (_print_inlining_output != nullptr) {
5152 tty->print_raw(_print_inlining_output);
5153 }
5154 }
5155
5156 void Compile::log_late_inline(CallGenerator* cg) {
5157 if (log() != nullptr) {
5158 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5159 cg->unique_id());
5160 JVMState* p = cg->call_node()->jvms();
5161 while (p != nullptr) {
5162 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5163 p = p->caller();
5164 }
5165 log()->tail("late_inline");
5166 }
5167 }
5168
5169 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5170 log_late_inline(cg);
5171 if (log() != nullptr) {
5172 log()->inline_fail(msg);
5173 }
5174 }
5175
5176 void Compile::log_inline_id(CallGenerator* cg) {
5177 if (log() != nullptr) {
5178 // The LogCompilation tool needs a unique way to identify late
5179 // inline call sites. This id must be unique for this call site in
5180 // this compilation. Try to have it unique across compilations as
5181 // well because it can be convenient when grepping through the log
5182 // file.
5183 // Distinguish OSR compilations from others in case CICountOSR is
5184 // on.
5185 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5186 cg->set_unique_id(id);
5187 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5188 }
5189 }
5190
5191 void Compile::log_inline_failure(const char* msg) {
5192 if (C->log() != nullptr) {
5193 C->log()->inline_fail(msg);
5194 }
5195 }
5196
5197
5198 // Dump inlining replay data to the stream.
5199 // Don't change thread state and acquire any locks.
5200 void Compile::dump_inline_data(outputStream* out) {
5201 InlineTree* inl_tree = ilt();
5202 if (inl_tree != nullptr) {
5203 out->print(" inline %d", inl_tree->count());
5204 inl_tree->dump_replay_data(out);
5205 }
5206 }
5207
5208 void Compile::dump_inline_data_reduced(outputStream* out) {
5209 assert(ReplayReduce, "");
5210
5211 InlineTree* inl_tree = ilt();
5212 if (inl_tree == nullptr) {
5213 return;
5214 }
5215 // Enable iterative replay file reduction
5216 // Output "compile" lines for depth 1 subtrees,
5217 // simulating that those trees were compiled
5218 // instead of inlined.
5219 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5220 InlineTree* sub = inl_tree->subtrees().at(i);
5221 if (sub->inline_level() != 1) {
5222 continue;
5223 }
5224
5225 ciMethod* method = sub->method();
5226 int entry_bci = -1;
5227 int comp_level = env()->task()->comp_level();
5228 out->print("compile ");
5229 method->dump_name_as_ascii(out);
5230 out->print(" %d %d", entry_bci, comp_level);
5231 out->print(" inline %d", sub->count());
5232 sub->dump_replay_data(out, -1);
5233 out->cr();
5234 }
5235 }
5236
5237 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5238 if (n1->Opcode() < n2->Opcode()) return -1;
5239 else if (n1->Opcode() > n2->Opcode()) return 1;
5240
5241 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5242 for (uint i = 1; i < n1->req(); i++) {
5243 if (n1->in(i) < n2->in(i)) return -1;
5244 else if (n1->in(i) > n2->in(i)) return 1;
5245 }
5246
5247 return 0;
5248 }
5249
5250 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5251 Node* n1 = *n1p;
5252 Node* n2 = *n2p;
5253
5254 return cmp_expensive_nodes(n1, n2);
5255 }
5256
5257 void Compile::sort_expensive_nodes() {
5258 if (!expensive_nodes_sorted()) {
5259 _expensive_nodes.sort(cmp_expensive_nodes);
5260 }
5261 }
5262
5263 bool Compile::expensive_nodes_sorted() const {
5264 for (int i = 1; i < _expensive_nodes.length(); i++) {
5265 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5266 return false;
5267 }
5268 }
5269 return true;
5270 }
5271
5272 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5273 if (_expensive_nodes.length() == 0) {
5274 return false;
5275 }
5276
5277 assert(OptimizeExpensiveOps, "optimization off?");
5278
5279 // Take this opportunity to remove dead nodes from the list
5280 int j = 0;
5281 for (int i = 0; i < _expensive_nodes.length(); i++) {
5282 Node* n = _expensive_nodes.at(i);
5283 if (!n->is_unreachable(igvn)) {
5284 assert(n->is_expensive(), "should be expensive");
5285 _expensive_nodes.at_put(j, n);
5286 j++;
5287 }
5288 }
5289 _expensive_nodes.trunc_to(j);
5290
5291 // Then sort the list so that similar nodes are next to each other
5292 // and check for at least two nodes of identical kind with same data
5293 // inputs.
5294 sort_expensive_nodes();
5295
5296 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5297 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5298 return true;
5299 }
5300 }
5301
5302 return false;
5303 }
5304
5305 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5306 if (_expensive_nodes.length() == 0) {
5307 return;
5308 }
5309
5310 assert(OptimizeExpensiveOps, "optimization off?");
5311
5312 // Sort to bring similar nodes next to each other and clear the
5313 // control input of nodes for which there's only a single copy.
5314 sort_expensive_nodes();
5315
5316 int j = 0;
5317 int identical = 0;
5318 int i = 0;
5319 bool modified = false;
5320 for (; i < _expensive_nodes.length()-1; i++) {
5321 assert(j <= i, "can't write beyond current index");
5322 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5323 identical++;
5324 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5325 continue;
5326 }
5327 if (identical > 0) {
5328 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5329 identical = 0;
5330 } else {
5331 Node* n = _expensive_nodes.at(i);
5332 igvn.replace_input_of(n, 0, nullptr);
5333 igvn.hash_insert(n);
5334 modified = true;
5335 }
5336 }
5337 if (identical > 0) {
5338 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5339 } else if (_expensive_nodes.length() >= 1) {
5340 Node* n = _expensive_nodes.at(i);
5341 igvn.replace_input_of(n, 0, nullptr);
5342 igvn.hash_insert(n);
5343 modified = true;
5344 }
5345 _expensive_nodes.trunc_to(j);
5346 if (modified) {
5347 igvn.optimize();
5348 }
5349 }
5350
5351 void Compile::add_expensive_node(Node * n) {
5352 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5353 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5354 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5355 if (OptimizeExpensiveOps) {
5356 _expensive_nodes.append(n);
5357 } else {
5358 // Clear control input and let IGVN optimize expensive nodes if
5359 // OptimizeExpensiveOps is off.
5360 n->set_req(0, nullptr);
5361 }
5362 }
5363
5364 /**
5365 * Track coarsened Lock and Unlock nodes.
5366 */
5367
5368 class Lock_List : public Node_List {
5369 uint _origin_cnt;
5370 public:
5371 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5372 uint origin_cnt() const { return _origin_cnt; }
5373 };
5374
5375 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5376 int length = locks.length();
5377 if (length > 0) {
5378 // Have to keep this list until locks elimination during Macro nodes elimination.
5379 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5380 AbstractLockNode* alock = locks.at(0);
5381 BoxLockNode* box = alock->box_node()->as_BoxLock();
5382 for (int i = 0; i < length; i++) {
5383 AbstractLockNode* lock = locks.at(i);
5384 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5385 locks_list->push(lock);
5386 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
5387 if (this_box != box) {
5388 // Locking regions (BoxLock) could be Unbalanced here:
5389 // - its coarsened locks were eliminated in earlier
5390 // macro nodes elimination followed by loop unroll
5391 // - it is OSR locking region (no Lock node)
5392 // Preserve Unbalanced status in such cases.
5393 if (!this_box->is_unbalanced()) {
5394 this_box->set_coarsened();
5395 }
5396 if (!box->is_unbalanced()) {
5397 box->set_coarsened();
5398 }
5399 }
5400 }
5401 _coarsened_locks.append(locks_list);
5402 }
5403 }
5404
5405 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5406 int count = coarsened_count();
5407 for (int i = 0; i < count; i++) {
5408 Node_List* locks_list = _coarsened_locks.at(i);
5409 for (uint j = 0; j < locks_list->size(); j++) {
5410 Node* lock = locks_list->at(j);
5411 assert(lock->is_AbstractLock(), "sanity");
5412 if (!useful.member(lock)) {
5413 locks_list->yank(lock);
5414 }
5415 }
5416 }
5417 }
5418
5419 void Compile::remove_coarsened_lock(Node* n) {
5420 if (n->is_AbstractLock()) {
5421 int count = coarsened_count();
5422 for (int i = 0; i < count; i++) {
5423 Node_List* locks_list = _coarsened_locks.at(i);
5424 locks_list->yank(n);
5425 }
5426 }
5427 }
5428
5429 bool Compile::coarsened_locks_consistent() {
5430 int count = coarsened_count();
5431 for (int i = 0; i < count; i++) {
5432 bool unbalanced = false;
5433 bool modified = false; // track locks kind modifications
5434 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5435 uint size = locks_list->size();
5436 if (size == 0) {
5437 unbalanced = false; // All locks were eliminated - good
5438 } else if (size != locks_list->origin_cnt()) {
5439 unbalanced = true; // Some locks were removed from list
5440 } else {
5441 for (uint j = 0; j < size; j++) {
5442 Node* lock = locks_list->at(j);
5443 // All nodes in group should have the same state (modified or not)
5444 if (!lock->as_AbstractLock()->is_coarsened()) {
5445 if (j == 0) {
5446 // first on list was modified, the rest should be too for consistency
5447 modified = true;
5448 } else if (!modified) {
5449 // this lock was modified but previous locks on the list were not
5450 unbalanced = true;
5451 break;
5452 }
5453 } else if (modified) {
5454 // previous locks on list were modified but not this lock
5455 unbalanced = true;
5456 break;
5457 }
5458 }
5459 }
5460 if (unbalanced) {
5461 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5462 #ifdef ASSERT
5463 if (PrintEliminateLocks) {
5464 tty->print_cr("=== unbalanced coarsened locks ===");
5465 for (uint l = 0; l < size; l++) {
5466 locks_list->at(l)->dump();
5467 }
5468 }
5469 #endif
5470 record_failure(C2Compiler::retry_no_locks_coarsening());
5471 return false;
5472 }
5473 }
5474 return true;
5475 }
5476
5477 // Mark locking regions (identified by BoxLockNode) as unbalanced if
5478 // locks coarsening optimization removed Lock/Unlock nodes from them.
5479 // Such regions become unbalanced because coarsening only removes part
5480 // of Lock/Unlock nodes in region. As result we can't execute other
5481 // locks elimination optimizations which assume all code paths have
5482 // corresponding pair of Lock/Unlock nodes - they are balanced.
5483 void Compile::mark_unbalanced_boxes() const {
5484 int count = coarsened_count();
5485 for (int i = 0; i < count; i++) {
5486 Node_List* locks_list = _coarsened_locks.at(i);
5487 uint size = locks_list->size();
5488 if (size > 0) {
5489 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
5490 BoxLockNode* box = alock->box_node()->as_BoxLock();
5491 if (alock->is_coarsened()) {
5492 // coarsened_locks_consistent(), which is called before this method, verifies
5493 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
5494 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5495 for (uint j = 1; j < size; j++) {
5496 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
5497 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
5498 if (box != this_box) {
5499 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5500 box->set_unbalanced();
5501 this_box->set_unbalanced();
5502 }
5503 }
5504 }
5505 }
5506 }
5507 }
5508
5509 /**
5510 * Remove the speculative part of types and clean up the graph
5511 */
5512 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5513 if (UseTypeSpeculation) {
5514 Unique_Node_List worklist;
5515 worklist.push(root());
5516 int modified = 0;
5517 // Go over all type nodes that carry a speculative type, drop the
5518 // speculative part of the type and enqueue the node for an igvn
5519 // which may optimize it out.
5520 for (uint next = 0; next < worklist.size(); ++next) {
5521 Node *n = worklist.at(next);
5522 if (n->is_Type()) {
5523 TypeNode* tn = n->as_Type();
5524 const Type* t = tn->type();
5525 const Type* t_no_spec = t->remove_speculative();
5526 if (t_no_spec != t) {
5527 bool in_hash = igvn.hash_delete(n);
5528 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5529 tn->set_type(t_no_spec);
5530 igvn.hash_insert(n);
5531 igvn._worklist.push(n); // give it a chance to go away
5532 modified++;
5533 }
5534 }
5535 // Iterate over outs - endless loops is unreachable from below
5536 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5537 Node *m = n->fast_out(i);
5538 if (not_a_node(m)) {
5539 continue;
5540 }
5541 worklist.push(m);
5542 }
5543 }
5544 // Drop the speculative part of all types in the igvn's type table
5545 igvn.remove_speculative_types();
5546 if (modified > 0) {
5547 igvn.optimize();
5548 if (failing()) return;
5549 }
5550 #ifdef ASSERT
5551 // Verify that after the IGVN is over no speculative type has resurfaced
5552 worklist.clear();
5553 worklist.push(root());
5554 for (uint next = 0; next < worklist.size(); ++next) {
5555 Node *n = worklist.at(next);
5556 const Type* t = igvn.type_or_null(n);
5557 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5558 if (n->is_Type()) {
5559 t = n->as_Type()->type();
5560 assert(t == t->remove_speculative(), "no more speculative types");
5561 }
5562 // Iterate over outs - endless loops is unreachable from below
5563 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5564 Node *m = n->fast_out(i);
5565 if (not_a_node(m)) {
5566 continue;
5567 }
5568 worklist.push(m);
5569 }
5570 }
5571 igvn.check_no_speculative_types();
5572 #endif
5573 }
5574 }
5575
5576 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5577 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5578 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5579 if (!EnableValhalla || ta == nullptr || tb == nullptr ||
5580 ta->is_zero_type() || tb->is_zero_type() ||
5581 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5582 // Use old acmp if one operand is null or not an inline type
5583 return new CmpPNode(a, b);
5584 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5585 // We know that one operand is an inline type. Therefore,
5586 // new acmp will only return true if both operands are nullptr.
5587 // Check if both operands are null by or'ing the oops.
5588 a = phase->transform(new CastP2XNode(nullptr, a));
5589 b = phase->transform(new CastP2XNode(nullptr, b));
5590 a = phase->transform(new OrXNode(a, b));
5591 return new CmpXNode(a, phase->MakeConX(0));
5592 }
5593 // Use new acmp
5594 return nullptr;
5595 }
5596
5597 // Auxiliary methods to support randomized stressing/fuzzing.
5598
5599 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5600 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5601 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5602 FLAG_SET_ERGO(StressSeed, _stress_seed);
5603 } else {
5604 _stress_seed = StressSeed;
5605 }
5606 if (_log != nullptr) {
5607 _log->elem("stress_test seed='%u'", _stress_seed);
5608 }
5609 }
5610
5611 int Compile::random() {
5612 _stress_seed = os::next_random(_stress_seed);
5613 return static_cast<int>(_stress_seed);
5614 }
5615
5616 // This method can be called the arbitrary number of times, with current count
5617 // as the argument. The logic allows selecting a single candidate from the
5618 // running list of candidates as follows:
5619 // int count = 0;
5620 // Cand* selected = null;
5621 // while(cand = cand->next()) {
5622 // if (randomized_select(++count)) {
5623 // selected = cand;
5624 // }
5625 // }
5626 //
5627 // Including count equalizes the chances any candidate is "selected".
5628 // This is useful when we don't have the complete list of candidates to choose
5629 // from uniformly. In this case, we need to adjust the randomicity of the
5630 // selection, or else we will end up biasing the selection towards the latter
5631 // candidates.
5632 //
5633 // Quick back-envelope calculation shows that for the list of n candidates
5634 // the equal probability for the candidate to persist as "best" can be
5635 // achieved by replacing it with "next" k-th candidate with the probability
5636 // of 1/k. It can be easily shown that by the end of the run, the
5637 // probability for any candidate is converged to 1/n, thus giving the
5638 // uniform distribution among all the candidates.
5639 //
5640 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5641 #define RANDOMIZED_DOMAIN_POW 29
5642 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5643 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5644 bool Compile::randomized_select(int count) {
5645 assert(count > 0, "only positive");
5646 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5647 }
5648
5649 CloneMap& Compile::clone_map() { return _clone_map; }
5650 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5651
5652 void NodeCloneInfo::dump_on(outputStream* st) const {
5653 st->print(" {%d:%d} ", idx(), gen());
5654 }
5655
5656 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5657 uint64_t val = value(old->_idx);
5658 NodeCloneInfo cio(val);
5659 assert(val != 0, "old node should be in the map");
5660 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5661 insert(nnn->_idx, cin.get());
5662 #ifndef PRODUCT
5663 if (is_debug()) {
5664 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5665 }
5666 #endif
5667 }
5668
5669 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5670 NodeCloneInfo cio(value(old->_idx));
5671 if (cio.get() == 0) {
5672 cio.set(old->_idx, 0);
5673 insert(old->_idx, cio.get());
5674 #ifndef PRODUCT
5675 if (is_debug()) {
5676 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5677 }
5678 #endif
5679 }
5680 clone(old, nnn, gen);
5681 }
5682
5683 int CloneMap::max_gen() const {
5684 int g = 0;
5685 DictI di(_dict);
5686 for(; di.test(); ++di) {
5687 int t = gen(di._key);
5688 if (g < t) {
5689 g = t;
5690 #ifndef PRODUCT
5691 if (is_debug()) {
5692 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5693 }
5694 #endif
5695 }
5696 }
5697 return g;
5698 }
5699
5700 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5701 uint64_t val = value(key);
5702 if (val != 0) {
5703 NodeCloneInfo ni(val);
5704 ni.dump_on(st);
5705 }
5706 }
5707
5708 void Compile::shuffle_macro_nodes() {
5709 if (_macro_nodes.length() < 2) {
5710 return;
5711 }
5712 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5713 uint j = C->random() % (i + 1);
5714 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5715 }
5716 }
5717
5718 // Move Allocate nodes to the start of the list
5719 void Compile::sort_macro_nodes() {
5720 int count = macro_count();
5721 int allocates = 0;
5722 for (int i = 0; i < count; i++) {
5723 Node* n = macro_node(i);
5724 if (n->is_Allocate()) {
5725 if (i != allocates) {
5726 Node* tmp = macro_node(allocates);
5727 _macro_nodes.at_put(allocates, n);
5728 _macro_nodes.at_put(i, tmp);
5729 }
5730 allocates++;
5731 }
5732 }
5733 }
5734
5735 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5736 if (failing()) { return; }
5737 EventCompilerPhase event(UNTIMED);
5738 if (event.should_commit()) {
5739 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5740 }
5741 #ifndef PRODUCT
5742 ResourceMark rm;
5743 stringStream ss;
5744 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5745 int iter = ++_igv_phase_iter[cpt];
5746 if (iter > 1) {
5747 ss.print(" %d", iter);
5748 }
5749 if (n != nullptr) {
5750 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5751 }
5752
5753 const char* name = ss.as_string();
5754 if (should_print_igv(level)) {
5755 _igv_printer->print_method(name, level);
5756 }
5757 if (should_print_phase(cpt)) {
5758 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5759 }
5760 #endif
5761 C->_latest_stage_start_counter.stamp();
5762 }
5763
5764 // Only used from CompileWrapper
5765 void Compile::begin_method() {
5766 #ifndef PRODUCT
5767 if (_method != nullptr && should_print_igv(1)) {
5768 _igv_printer->begin_method();
5769 }
5770 #endif
5771 C->_latest_stage_start_counter.stamp();
5772 }
5773
5774 // Only used from CompileWrapper
5775 void Compile::end_method() {
5776 EventCompilerPhase event(UNTIMED);
5777 if (event.should_commit()) {
5778 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5779 }
5780
5781 #ifndef PRODUCT
5782 if (_method != nullptr && should_print_igv(1)) {
5783 _igv_printer->end_method();
5784 }
5785 #endif
5786 }
5787
5788 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5789 #ifndef PRODUCT
5790 if (_directive->should_print_phase(cpt)) {
5791 return true;
5792 }
5793 #endif
5794 return false;
5795 }
5796
5797 bool Compile::should_print_igv(const int level) {
5798 #ifndef PRODUCT
5799 if (PrintIdealGraphLevel < 0) { // disabled by the user
5800 return false;
5801 }
5802
5803 bool need = directive()->IGVPrintLevelOption >= level;
5804 if (need && !_igv_printer) {
5805 _igv_printer = IdealGraphPrinter::printer();
5806 _igv_printer->set_compile(this);
5807 }
5808 return need;
5809 #else
5810 return false;
5811 #endif
5812 }
5813
5814 #ifndef PRODUCT
5815 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5816 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5817
5818 // Called from debugger. Prints method to the default file with the default phase name.
5819 // This works regardless of any Ideal Graph Visualizer flags set or not.
5820 void igv_print() {
5821 Compile::current()->igv_print_method_to_file();
5822 }
5823
5824 // Same as igv_print() above but with a specified phase name.
5825 void igv_print(const char* phase_name) {
5826 Compile::current()->igv_print_method_to_file(phase_name);
5827 }
5828
5829 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5830 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5831 // This works regardless of any Ideal Graph Visualizer flags set or not.
5832 void igv_print(bool network) {
5833 if (network) {
5834 Compile::current()->igv_print_method_to_network();
5835 } else {
5836 Compile::current()->igv_print_method_to_file();
5837 }
5838 }
5839
5840 // Same as igv_print(bool network) above but with a specified phase name.
5841 void igv_print(bool network, const char* phase_name) {
5842 if (network) {
5843 Compile::current()->igv_print_method_to_network(phase_name);
5844 } else {
5845 Compile::current()->igv_print_method_to_file(phase_name);
5846 }
5847 }
5848
5849 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5850 void igv_print_default() {
5851 Compile::current()->print_method(PHASE_DEBUG, 0);
5852 }
5853
5854 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5855 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5856 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5857 void igv_append() {
5858 Compile::current()->igv_print_method_to_file("Debug", true);
5859 }
5860
5861 // Same as igv_append() above but with a specified phase name.
5862 void igv_append(const char* phase_name) {
5863 Compile::current()->igv_print_method_to_file(phase_name, true);
5864 }
5865
5866 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5867 const char* file_name = "custom_debug.xml";
5868 if (_debug_file_printer == nullptr) {
5869 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5870 } else {
5871 _debug_file_printer->update_compiled_method(C->method());
5872 }
5873 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5874 _debug_file_printer->print(phase_name, (Node*)C->root());
5875 }
5876
5877 void Compile::igv_print_method_to_network(const char* phase_name) {
5878 if (_debug_network_printer == nullptr) {
5879 _debug_network_printer = new IdealGraphPrinter(C);
5880 } else {
5881 _debug_network_printer->update_compiled_method(C->method());
5882 }
5883 tty->print_cr("Method printed over network stream to IGV");
5884 _debug_network_printer->print(phase_name, (Node*)C->root());
5885 }
5886 #endif
5887
5888 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5889 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5890 return value;
5891 }
5892 Node* result = nullptr;
5893 if (bt == T_BYTE) {
5894 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5895 result = new RShiftINode(result, phase->intcon(24));
5896 } else if (bt == T_BOOLEAN) {
5897 result = new AndINode(value, phase->intcon(0xFF));
5898 } else if (bt == T_CHAR) {
5899 result = new AndINode(value,phase->intcon(0xFFFF));
5900 } else {
5901 assert(bt == T_SHORT, "unexpected narrow type");
5902 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5903 result = new RShiftINode(result, phase->intcon(16));
5904 }
5905 if (transform_res) {
5906 result = phase->transform(result);
5907 }
5908 return result;
5909 }
5910
5911 void Compile::record_method_not_compilable_oom() {
5912 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5913 }