1 /*
2 * Copyright (c) 1997, 2025, 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 "asm/macroAssembler.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "ci/ciReplay.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "code/exceptionHandlerTable.hpp"
30 #include "code/nmethod.hpp"
31 #include "compiler/compilationFailureInfo.hpp"
32 #include "compiler/compilationMemoryStatistic.hpp"
33 #include "compiler/compileBroker.hpp"
34 #include "compiler/compileLog.hpp"
35 #include "compiler/compilerOracle.hpp"
36 #include "compiler/compiler_globals.hpp"
37 #include "compiler/disassembler.hpp"
38 #include "compiler/oopMap.hpp"
39 #include "gc/shared/barrierSet.hpp"
40 #include "gc/shared/c2/barrierSetC2.hpp"
41 #include "jfr/jfrEvents.hpp"
42 #include "jvm_io.h"
43 #include "memory/allocation.hpp"
44 #include "memory/arena.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/movenode.hpp"
69 #include "opto/mulnode.hpp"
70 #include "opto/narrowptrnode.hpp"
71 #include "opto/node.hpp"
72 #include "opto/opcodes.hpp"
73 #include "opto/output.hpp"
74 #include "opto/parse.hpp"
75 #include "opto/phaseX.hpp"
76 #include "opto/rootnode.hpp"
77 #include "opto/runtime.hpp"
78 #include "opto/stringopts.hpp"
79 #include "opto/type.hpp"
80 #include "opto/vector.hpp"
81 #include "opto/vectornode.hpp"
82 #include "runtime/globals_extension.hpp"
83 #include "runtime/sharedRuntime.hpp"
84 #include "runtime/signature.hpp"
85 #include "runtime/stubRoutines.hpp"
86 #include "runtime/timer.hpp"
87 #include "utilities/align.hpp"
88 #include "utilities/copy.hpp"
89 #include "utilities/macros.hpp"
90 #include "utilities/resourceHash.hpp"
91
92 // -------------------- Compile::mach_constant_base_node -----------------------
93 // Constant table base node singleton.
94 MachConstantBaseNode* Compile::mach_constant_base_node() {
95 if (_mach_constant_base_node == nullptr) {
96 _mach_constant_base_node = new MachConstantBaseNode();
97 _mach_constant_base_node->add_req(C->root());
98 }
99 return _mach_constant_base_node;
100 }
101
102
103 /// Support for intrinsics.
104
105 // Return the index at which m must be inserted (or already exists).
106 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
107 class IntrinsicDescPair {
108 private:
109 ciMethod* _m;
110 bool _is_virtual;
111 public:
112 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
113 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
114 ciMethod* m= elt->method();
115 ciMethod* key_m = key->_m;
116 if (key_m < m) return -1;
117 else if (key_m > m) return 1;
118 else {
119 bool is_virtual = elt->is_virtual();
120 bool key_virtual = key->_is_virtual;
121 if (key_virtual < is_virtual) return -1;
122 else if (key_virtual > is_virtual) return 1;
123 else return 0;
124 }
125 }
126 };
127 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
128 #ifdef ASSERT
129 for (int i = 1; i < _intrinsics.length(); i++) {
130 CallGenerator* cg1 = _intrinsics.at(i-1);
131 CallGenerator* cg2 = _intrinsics.at(i);
132 assert(cg1->method() != cg2->method()
133 ? cg1->method() < cg2->method()
134 : cg1->is_virtual() < cg2->is_virtual(),
135 "compiler intrinsics list must stay sorted");
136 }
137 #endif
138 IntrinsicDescPair pair(m, is_virtual);
139 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
140 }
141
142 void Compile::register_intrinsic(CallGenerator* cg) {
143 bool found = false;
144 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
145 assert(!found, "registering twice");
146 _intrinsics.insert_before(index, cg);
147 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
148 }
149
150 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
151 assert(m->is_loaded(), "don't try this on unloaded methods");
152 if (_intrinsics.length() > 0) {
153 bool found = false;
154 int index = intrinsic_insertion_index(m, is_virtual, found);
155 if (found) {
156 return _intrinsics.at(index);
157 }
158 }
159 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
160 if (m->intrinsic_id() != vmIntrinsics::_none &&
161 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
162 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
163 if (cg != nullptr) {
164 // Save it for next time:
165 register_intrinsic(cg);
166 return cg;
167 } else {
168 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
169 }
170 }
171 return nullptr;
172 }
173
174 // Compile::make_vm_intrinsic is defined in library_call.cpp.
175
176 #ifndef PRODUCT
177 // statistics gathering...
178
179 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
180 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
181
182 inline int as_int(vmIntrinsics::ID id) {
183 return vmIntrinsics::as_int(id);
184 }
185
186 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
187 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
188 int oflags = _intrinsic_hist_flags[as_int(id)];
189 assert(flags != 0, "what happened?");
190 if (is_virtual) {
191 flags |= _intrinsic_virtual;
192 }
193 bool changed = (flags != oflags);
194 if ((flags & _intrinsic_worked) != 0) {
195 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
196 if (count == 1) {
197 changed = true; // first time
198 }
199 // increment the overall count also:
200 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
201 }
202 if (changed) {
203 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
204 // Something changed about the intrinsic's virtuality.
205 if ((flags & _intrinsic_virtual) != 0) {
206 // This is the first use of this intrinsic as a virtual call.
207 if (oflags != 0) {
208 // We already saw it as a non-virtual, so note both cases.
209 flags |= _intrinsic_both;
210 }
211 } else if ((oflags & _intrinsic_both) == 0) {
212 // This is the first use of this intrinsic as a non-virtual
213 flags |= _intrinsic_both;
214 }
215 }
216 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
217 }
218 // update the overall flags also:
219 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
220 return changed;
221 }
222
223 static char* format_flags(int flags, char* buf) {
224 buf[0] = 0;
225 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
226 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
227 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
228 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
229 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
230 if (buf[0] == 0) strcat(buf, ",");
231 assert(buf[0] == ',', "must be");
232 return &buf[1];
233 }
234
235 void Compile::print_intrinsic_statistics() {
236 char flagsbuf[100];
237 ttyLocker ttyl;
238 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
239 tty->print_cr("Compiler intrinsic usage:");
240 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
241 if (total == 0) total = 1; // avoid div0 in case of no successes
242 #define PRINT_STAT_LINE(name, c, f) \
243 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
244 for (auto id : EnumRange<vmIntrinsicID>{}) {
245 int flags = _intrinsic_hist_flags[as_int(id)];
246 juint count = _intrinsic_hist_count[as_int(id)];
247 if ((flags | count) != 0) {
248 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
249 }
250 }
251 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
252 if (xtty != nullptr) xtty->tail("statistics");
253 }
254
255 void Compile::print_statistics() {
256 { ttyLocker ttyl;
257 if (xtty != nullptr) xtty->head("statistics type='opto'");
258 Parse::print_statistics();
259 PhaseStringOpts::print_statistics();
260 PhaseCCP::print_statistics();
261 PhaseRegAlloc::print_statistics();
262 PhaseOutput::print_statistics();
263 PhasePeephole::print_statistics();
264 PhaseIdealLoop::print_statistics();
265 ConnectionGraph::print_statistics();
266 PhaseMacroExpand::print_statistics();
267 if (xtty != nullptr) xtty->tail("statistics");
268 }
269 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
270 // put this under its own <statistics> element.
271 print_intrinsic_statistics();
272 }
273 }
274 #endif //PRODUCT
275
276 void Compile::gvn_replace_by(Node* n, Node* nn) {
277 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
278 Node* use = n->last_out(i);
279 bool is_in_table = initial_gvn()->hash_delete(use);
280 uint uses_found = 0;
281 for (uint j = 0; j < use->len(); j++) {
282 if (use->in(j) == n) {
283 if (j < use->req())
284 use->set_req(j, nn);
285 else
286 use->set_prec(j, nn);
287 uses_found++;
288 }
289 }
290 if (is_in_table) {
291 // reinsert into table
292 initial_gvn()->hash_find_insert(use);
293 }
294 record_for_igvn(use);
295 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
296 i -= uses_found; // we deleted 1 or more copies of this edge
297 }
298 }
299
300
301 // Identify all nodes that are reachable from below, useful.
302 // Use breadth-first pass that records state in a Unique_Node_List,
303 // recursive traversal is slower.
304 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
305 int estimated_worklist_size = live_nodes();
306 useful.map( estimated_worklist_size, nullptr ); // preallocate space
307
308 // Initialize worklist
309 if (root() != nullptr) { useful.push(root()); }
310 // If 'top' is cached, declare it useful to preserve cached node
311 if (cached_top_node()) { useful.push(cached_top_node()); }
312
313 // Push all useful nodes onto the list, breadthfirst
314 for( uint next = 0; next < useful.size(); ++next ) {
315 assert( next < unique(), "Unique useful nodes < total nodes");
316 Node *n = useful.at(next);
317 uint max = n->len();
318 for( uint i = 0; i < max; ++i ) {
319 Node *m = n->in(i);
320 if (not_a_node(m)) continue;
321 useful.push(m);
322 }
323 }
324 }
325
326 // Update dead_node_list with any missing dead nodes using useful
327 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
328 void Compile::update_dead_node_list(Unique_Node_List &useful) {
329 uint max_idx = unique();
330 VectorSet& useful_node_set = useful.member_set();
331
332 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
333 // If node with index node_idx is not in useful set,
334 // mark it as dead in dead node list.
335 if (!useful_node_set.test(node_idx)) {
336 record_dead_node(node_idx);
337 }
338 }
339 }
340
341 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
342 int shift = 0;
343 for (int i = 0; i < inlines->length(); i++) {
344 CallGenerator* cg = inlines->at(i);
345 if (useful.member(cg->call_node())) {
346 if (shift > 0) {
347 inlines->at_put(i - shift, cg);
348 }
349 } else {
350 shift++; // skip over the dead element
351 }
352 }
353 if (shift > 0) {
354 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
355 }
356 }
357
358 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
359 assert(dead != nullptr && dead->is_Call(), "sanity");
360 int found = 0;
361 for (int i = 0; i < inlines->length(); i++) {
362 if (inlines->at(i)->call_node() == dead) {
363 inlines->remove_at(i);
364 found++;
365 NOT_DEBUG( break; ) // elements are unique, so exit early
366 }
367 }
368 assert(found <= 1, "not unique");
369 }
370
371 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
372 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
373 for (int i = node_list.length() - 1; i >= 0; i--) {
374 N* node = node_list.at(i);
375 if (!useful.member(node)) {
376 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
377 }
378 }
379 }
380
381 void Compile::remove_useless_node(Node* dead) {
382 remove_modified_node(dead);
383
384 // Constant node that has no out-edges and has only one in-edge from
385 // root is usually dead. However, sometimes reshaping walk makes
386 // it reachable by adding use edges. So, we will NOT count Con nodes
387 // as dead to be conservative about the dead node count at any
388 // given time.
389 if (!dead->is_Con()) {
390 record_dead_node(dead->_idx);
391 }
392 if (dead->is_macro()) {
393 remove_macro_node(dead);
394 }
395 if (dead->is_expensive()) {
396 remove_expensive_node(dead);
397 }
398 if (dead->is_OpaqueTemplateAssertionPredicate()) {
399 remove_template_assertion_predicate_opaq(dead);
400 }
401 if (dead->is_ParsePredicate()) {
402 remove_parse_predicate(dead->as_ParsePredicate());
403 }
404 if (dead->for_post_loop_opts_igvn()) {
405 remove_from_post_loop_opts_igvn(dead);
406 }
407 if (dead->is_InlineType()) {
408 remove_inline_type(dead);
409 }
410 if (dead->is_Call()) {
411 remove_useless_late_inlines( &_late_inlines, dead);
412 remove_useless_late_inlines( &_string_late_inlines, dead);
413 remove_useless_late_inlines( &_boxing_late_inlines, dead);
414 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
415
416 if (dead->is_CallStaticJava()) {
417 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
418 }
419 }
420 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
421 bs->unregister_potential_barrier_node(dead);
422 }
423
424 // Disconnect all useless nodes by disconnecting those at the boundary.
425 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) {
426 uint next = 0;
427 while (next < useful.size()) {
428 Node *n = useful.at(next++);
429 if (n->is_SafePoint()) {
430 // We're done with a parsing phase. Replaced nodes are not valid
431 // beyond that point.
432 n->as_SafePoint()->delete_replaced_nodes();
433 }
434 // Use raw traversal of out edges since this code removes out edges
435 int max = n->outcnt();
436 for (int j = 0; j < max; ++j) {
437 Node* child = n->raw_out(j);
438 if (!useful.member(child)) {
439 assert(!child->is_top() || child != top(),
440 "If top is cached in Compile object it is in useful list");
441 // Only need to remove this out-edge to the useless node
442 n->raw_del_out(j);
443 --j;
444 --max;
445 if (child->is_data_proj_of_pure_function(n)) {
446 worklist.push(n);
447 }
448 }
449 }
450 if (n->outcnt() == 1 && n->has_special_unique_user()) {
451 assert(useful.member(n->unique_out()), "do not push a useless node");
452 worklist.push(n->unique_out());
453 }
454 if (n->outcnt() == 0) {
455 worklist.push(n);
456 }
457 }
458
459 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
460 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
461 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes
462 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
463 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
464 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
465 #ifdef ASSERT
466 if (_modified_nodes != nullptr) {
467 _modified_nodes->remove_useless_nodes(useful.member_set());
468 }
469 #endif
470 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
471 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
472 #ifdef ASSERT
473 if (_modified_nodes != nullptr) {
474 _modified_nodes->remove_useless_nodes(useful.member_set());
475 }
476 #endif
477
478 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
479 bs->eliminate_useless_gc_barriers(useful, this);
480 // clean up the late inline lists
481 remove_useless_late_inlines( &_late_inlines, useful);
482 remove_useless_late_inlines( &_string_late_inlines, useful);
483 remove_useless_late_inlines( &_boxing_late_inlines, useful);
484 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
485 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
486 }
487
488 // ============================================================================
489 //------------------------------CompileWrapper---------------------------------
490 class CompileWrapper : public StackObj {
491 Compile *const _compile;
492 public:
493 CompileWrapper(Compile* compile);
494
495 ~CompileWrapper();
496 };
497
498 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
499 // the Compile* pointer is stored in the current ciEnv:
500 ciEnv* env = compile->env();
501 assert(env == ciEnv::current(), "must already be a ciEnv active");
502 assert(env->compiler_data() == nullptr, "compile already active?");
503 env->set_compiler_data(compile);
504 assert(compile == Compile::current(), "sanity");
505
506 compile->set_type_dict(nullptr);
507 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
508 compile->clone_map().set_clone_idx(0);
509 compile->set_type_last_size(0);
510 compile->set_last_tf(nullptr, nullptr);
511 compile->set_indexSet_arena(nullptr);
512 compile->set_indexSet_free_block_list(nullptr);
513 compile->init_type_arena();
514 Type::Initialize(compile);
515 _compile->begin_method();
516 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
517 }
518 CompileWrapper::~CompileWrapper() {
519 // simulate crash during compilation
520 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
521
522 _compile->end_method();
523 _compile->env()->set_compiler_data(nullptr);
524 }
525
526
527 //----------------------------print_compile_messages---------------------------
528 void Compile::print_compile_messages() {
529 #ifndef PRODUCT
530 // Check if recompiling
531 if (!subsume_loads() && PrintOpto) {
532 // Recompiling without allowing machine instructions to subsume loads
533 tty->print_cr("*********************************************************");
534 tty->print_cr("** Bailout: Recompile without subsuming loads **");
535 tty->print_cr("*********************************************************");
536 }
537 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
538 // Recompiling without escape analysis
539 tty->print_cr("*********************************************************");
540 tty->print_cr("** Bailout: Recompile without escape analysis **");
541 tty->print_cr("*********************************************************");
542 }
543 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
544 // Recompiling without iterative escape analysis
545 tty->print_cr("*********************************************************");
546 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
547 tty->print_cr("*********************************************************");
548 }
549 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
550 // Recompiling without reducing allocation merges
551 tty->print_cr("*********************************************************");
552 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
553 tty->print_cr("*********************************************************");
554 }
555 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
556 // Recompiling without boxing elimination
557 tty->print_cr("*********************************************************");
558 tty->print_cr("** Bailout: Recompile without boxing elimination **");
559 tty->print_cr("*********************************************************");
560 }
561 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
562 // Recompiling without locks coarsening
563 tty->print_cr("*********************************************************");
564 tty->print_cr("** Bailout: Recompile without locks coarsening **");
565 tty->print_cr("*********************************************************");
566 }
567 if (env()->break_at_compile()) {
568 // Open the debugger when compiling this method.
569 tty->print("### Breaking when compiling: ");
570 method()->print_short_name();
571 tty->cr();
572 BREAKPOINT;
573 }
574
575 if( PrintOpto ) {
576 if (is_osr_compilation()) {
577 tty->print("[OSR]%3d", _compile_id);
578 } else {
579 tty->print("%3d", _compile_id);
580 }
581 }
582 #endif
583 }
584
585 #ifndef PRODUCT
586 void Compile::print_ideal_ir(const char* phase_name) {
587 // keep the following output all in one block
588 // This output goes directly to the tty, not the compiler log.
589 // To enable tools to match it up with the compilation activity,
590 // be sure to tag this tty output with the compile ID.
591
592 // Node dumping can cause a safepoint, which can break the tty lock.
593 // Buffer all node dumps, so that all safepoints happen before we lock.
594 ResourceMark rm;
595 stringStream ss;
596
597 if (_output == nullptr) {
598 ss.print_cr("AFTER: %s", phase_name);
599 // Print out all nodes in ascending order of index.
600 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
601 } else {
602 // Dump the node blockwise if we have a scheduling
603 _output->print_scheduling(&ss);
604 }
605
606 // Check that the lock is not broken by a safepoint.
607 NoSafepointVerifier nsv;
608 ttyLocker ttyl;
609 if (xtty != nullptr) {
610 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
611 compile_id(),
612 is_osr_compilation() ? " compile_kind='osr'" : "",
613 phase_name);
614 }
615
616 tty->print("%s", ss.as_string());
617
618 if (xtty != nullptr) {
619 xtty->tail("ideal");
620 }
621 }
622 #endif
623
624 // ============================================================================
625 //------------------------------Compile standard-------------------------------
626
627 // Compile a method. entry_bci is -1 for normal compilations and indicates
628 // the continuation bci for on stack replacement.
629
630
631 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
632 Options options, DirectiveSet* directive)
633 : Phase(Compiler),
634 _compile_id(ci_env->compile_id()),
635 _options(options),
636 _method(target),
637 _entry_bci(osr_bci),
638 _ilt(nullptr),
639 _stub_function(nullptr),
640 _stub_name(nullptr),
641 _stub_entry_point(nullptr),
642 _max_node_limit(MaxNodeLimit),
643 _post_loop_opts_phase(false),
644 _allow_macro_nodes(true),
645 _inlining_progress(false),
646 _inlining_incrementally(false),
647 _do_cleanup(false),
648 _has_reserved_stack_access(target->has_reserved_stack_access()),
649 _has_circular_inline_type(false),
650 #ifndef PRODUCT
651 _igv_idx(0),
652 _trace_opto_output(directive->TraceOptoOutputOption),
653 #endif
654 _has_method_handle_invokes(false),
655 _clinit_barrier_on_entry(false),
656 _stress_seed(0),
657 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
658 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
659 _env(ci_env),
660 _directive(directive),
661 _log(ci_env->log()),
662 _first_failure_details(nullptr),
663 _intrinsics(comp_arena(), 0, 0, nullptr),
664 _macro_nodes(comp_arena(), 8, 0, nullptr),
665 _parse_predicates(comp_arena(), 8, 0, nullptr),
666 _template_assertion_predicate_opaqs(comp_arena(), 8, 0, nullptr),
667 _expensive_nodes(comp_arena(), 8, 0, nullptr),
668 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
669 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
670 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
671 _coarsened_locks(comp_arena(), 8, 0, nullptr),
672 _congraph(nullptr),
673 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
674 _unique(0),
675 _dead_node_count(0),
676 _dead_node_list(comp_arena()),
677 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
678 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
679 _node_arena(&_node_arena_one),
680 _mach_constant_base_node(nullptr),
681 _Compile_types(mtCompiler, Arena::Tag::tag_type),
682 _initial_gvn(nullptr),
683 _igvn_worklist(nullptr),
684 _types(nullptr),
685 _node_hash(nullptr),
686 _late_inlines(comp_arena(), 2, 0, nullptr),
687 _string_late_inlines(comp_arena(), 2, 0, nullptr),
688 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
689 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
690 _late_inlines_pos(0),
691 _number_of_mh_late_inlines(0),
692 _oom(false),
693 _replay_inline_data(nullptr),
694 _inline_printer(this),
695 _java_calls(0),
696 _inner_loops(0),
697 _interpreter_frame_size(0),
698 _output(nullptr)
699 #ifndef PRODUCT
700 ,
701 _in_dump_cnt(0)
702 #endif
703 {
704 C = this;
705 CompileWrapper cw(this);
706
707 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
708 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
709
710 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
711 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
712 // We can always print a disassembly, either abstract (hex dump) or
713 // with the help of a suitable hsdis library. Thus, we should not
714 // couple print_assembly and print_opto_assembly controls.
715 // But: always print opto and regular assembly on compile command 'print'.
716 bool print_assembly = directive->PrintAssemblyOption;
717 set_print_assembly(print_opto_assembly || print_assembly);
718 #else
719 set_print_assembly(false); // must initialize.
720 #endif
721
722 #ifndef PRODUCT
723 set_parsed_irreducible_loop(false);
724 #endif
725
726 if (directive->ReplayInlineOption) {
727 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
728 }
729 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
730 set_print_intrinsics(directive->PrintIntrinsicsOption);
731 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
732
733 if (ProfileTraps) {
734 // Make sure the method being compiled gets its own MDO,
735 // so we can at least track the decompile_count().
736 method()->ensure_method_data();
737 }
738
739 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
740 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout) {
741 initialize_stress_seed(directive);
742 }
743
744 Init(/*do_aliasing=*/ true);
745
746 print_compile_messages();
747
748 _ilt = InlineTree::build_inline_tree_root();
749
750 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
751 assert(num_alias_types() >= AliasIdxRaw, "");
752
753 #define MINIMUM_NODE_HASH 1023
754
755 // GVN that will be run immediately on new nodes
756 uint estimated_size = method()->code_size()*4+64;
757 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
758 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
759 _types = new (comp_arena()) Type_Array(comp_arena());
760 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
761 PhaseGVN gvn;
762 set_initial_gvn(&gvn);
763
764 { // Scope for timing the parser
765 TracePhase tp(_t_parser);
766
767 // Put top into the hash table ASAP.
768 initial_gvn()->transform(top());
769
770 // Set up tf(), start(), and find a CallGenerator.
771 CallGenerator* cg = nullptr;
772 if (is_osr_compilation()) {
773 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
774 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
775 initial_gvn()->set_type_bottom(s);
776 verify_start(s);
777 cg = CallGenerator::for_osr(method(), entry_bci());
778 } else {
779 // Normal case.
780 init_tf(TypeFunc::make(method()));
781 StartNode* s = new StartNode(root(), tf()->domain_cc());
782 initial_gvn()->set_type_bottom(s);
783 verify_start(s);
784 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
785 // With java.lang.ref.reference.get() we must go through the
786 // intrinsic - even when get() is the root
787 // method of the compile - so that, if necessary, the value in
788 // the referent field of the reference object gets recorded by
789 // the pre-barrier code.
790 cg = find_intrinsic(method(), false);
791 }
792 if (cg == nullptr) {
793 float past_uses = method()->interpreter_invocation_count();
794 float expected_uses = past_uses;
795 cg = CallGenerator::for_inline(method(), expected_uses);
796 }
797 }
798 if (failing()) return;
799 if (cg == nullptr) {
800 const char* reason = InlineTree::check_can_parse(method());
801 assert(reason != nullptr, "expect reason for parse failure");
802 stringStream ss;
803 ss.print("cannot parse method: %s", reason);
804 record_method_not_compilable(ss.as_string());
805 return;
806 }
807
808 gvn.set_type(root(), root()->bottom_type());
809
810 JVMState* jvms = build_start_state(start(), tf());
811 if ((jvms = cg->generate(jvms)) == nullptr) {
812 assert(failure_reason() != nullptr, "expect reason for parse failure");
813 stringStream ss;
814 ss.print("method parse failed: %s", failure_reason());
815 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
816 return;
817 }
818 GraphKit kit(jvms);
819
820 if (!kit.stopped()) {
821 // Accept return values, and transfer control we know not where.
822 // This is done by a special, unique ReturnNode bound to root.
823 return_values(kit.jvms());
824 }
825
826 if (kit.has_exceptions()) {
827 // Any exceptions that escape from this call must be rethrown
828 // to whatever caller is dynamically above us on the stack.
829 // This is done by a special, unique RethrowNode bound to root.
830 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
831 }
832
833 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
834
835 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
836 inline_string_calls(true);
837 }
838
839 if (failing()) return;
840
841 // Remove clutter produced by parsing.
842 if (!failing()) {
843 ResourceMark rm;
844 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
845 }
846 }
847
848 // Note: Large methods are capped off in do_one_bytecode().
849 if (failing()) return;
850
851 // After parsing, node notes are no longer automagic.
852 // They must be propagated by register_new_node_with_optimizer(),
853 // clone(), or the like.
854 set_default_node_notes(nullptr);
855
856 #ifndef PRODUCT
857 if (should_print_igv(1)) {
858 _igv_printer->print_inlining();
859 }
860 #endif
861
862 if (failing()) return;
863 NOT_PRODUCT( verify_graph_edges(); )
864
865 // Now optimize
866 Optimize();
867 if (failing()) return;
868 NOT_PRODUCT( verify_graph_edges(); )
869
870 #ifndef PRODUCT
871 if (should_print_ideal()) {
872 print_ideal_ir("print_ideal");
873 }
874 #endif
875
876 #ifdef ASSERT
877 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
878 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
879 #endif
880
881 // Dump compilation data to replay it.
882 if (directive->DumpReplayOption) {
883 env()->dump_replay_data(_compile_id);
884 }
885 if (directive->DumpInlineOption && (ilt() != nullptr)) {
886 env()->dump_inline_data(_compile_id);
887 }
888
889 // Now that we know the size of all the monitors we can add a fixed slot
890 // for the original deopt pc.
891 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
892 if (needs_stack_repair()) {
893 // One extra slot for the special stack increment value
894 next_slot += 2;
895 }
896 // TODO 8284443 Only reserve extra slot if needed
897 if (InlineTypeReturnedAsFields) {
898 // One extra slot to hold the IsInit information for a nullable
899 // inline type return if we run out of registers.
900 next_slot += 2;
901 }
902 set_fixed_slots(next_slot);
903
904 // Compute when to use implicit null checks. Used by matching trap based
905 // nodes and NullCheck optimization.
906 set_allowed_deopt_reasons();
907
908 // Now generate code
909 Code_Gen();
910 }
911
912 //------------------------------Compile----------------------------------------
913 // Compile a runtime stub
914 Compile::Compile(ciEnv* ci_env,
915 TypeFunc_generator generator,
916 address stub_function,
917 const char *stub_name,
918 int is_fancy_jump,
919 bool pass_tls,
920 bool return_pc,
921 DirectiveSet* directive)
922 : Phase(Compiler),
923 _compile_id(0),
924 _options(Options::for_runtime_stub()),
925 _method(nullptr),
926 _entry_bci(InvocationEntryBci),
927 _stub_function(stub_function),
928 _stub_name(stub_name),
929 _stub_entry_point(nullptr),
930 _max_node_limit(MaxNodeLimit),
931 _post_loop_opts_phase(false),
932 _allow_macro_nodes(true),
933 _inlining_progress(false),
934 _inlining_incrementally(false),
935 _has_reserved_stack_access(false),
936 _has_circular_inline_type(false),
937 #ifndef PRODUCT
938 _igv_idx(0),
939 _trace_opto_output(directive->TraceOptoOutputOption),
940 #endif
941 _has_method_handle_invokes(false),
942 _clinit_barrier_on_entry(false),
943 _stress_seed(0),
944 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
945 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
946 _env(ci_env),
947 _directive(directive),
948 _log(ci_env->log()),
949 _first_failure_details(nullptr),
950 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
951 _congraph(nullptr),
952 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
953 _unique(0),
954 _dead_node_count(0),
955 _dead_node_list(comp_arena()),
956 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
957 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
958 _node_arena(&_node_arena_one),
959 _mach_constant_base_node(nullptr),
960 _Compile_types(mtCompiler, Arena::Tag::tag_type),
961 _initial_gvn(nullptr),
962 _igvn_worklist(nullptr),
963 _types(nullptr),
964 _node_hash(nullptr),
965 _number_of_mh_late_inlines(0),
966 _oom(false),
967 _replay_inline_data(nullptr),
968 _inline_printer(this),
969 _java_calls(0),
970 _inner_loops(0),
971 _interpreter_frame_size(0),
972 _output(nullptr),
973 #ifndef PRODUCT
974 _in_dump_cnt(0),
975 #endif
976 _allowed_reasons(0) {
977 C = this;
978
979 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
980 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
981
982 #ifndef PRODUCT
983 set_print_assembly(PrintFrameConverterAssembly);
984 set_parsed_irreducible_loop(false);
985 #else
986 set_print_assembly(false); // Must initialize.
987 #endif
988 set_has_irreducible_loop(false); // no loops
989
990 CompileWrapper cw(this);
991 Init(/*do_aliasing=*/ false);
992 init_tf((*generator)());
993
994 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
995 _types = new (comp_arena()) Type_Array(comp_arena());
996 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
997
998 if (StressLCM || StressGCM || StressBailout) {
999 initialize_stress_seed(directive);
1000 }
1001
1002 {
1003 PhaseGVN gvn;
1004 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1005 gvn.transform(top());
1006
1007 GraphKit kit;
1008 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1009 }
1010
1011 NOT_PRODUCT( verify_graph_edges(); )
1012
1013 Code_Gen();
1014 }
1015
1016 Compile::~Compile() {
1017 delete _first_failure_details;
1018 };
1019
1020 //------------------------------Init-------------------------------------------
1021 // Prepare for a single compilation
1022 void Compile::Init(bool aliasing) {
1023 _do_aliasing = aliasing;
1024 _unique = 0;
1025 _regalloc = nullptr;
1026
1027 _tf = nullptr; // filled in later
1028 _top = nullptr; // cached later
1029 _matcher = nullptr; // filled in later
1030 _cfg = nullptr; // filled in later
1031
1032 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1033
1034 _node_note_array = nullptr;
1035 _default_node_notes = nullptr;
1036 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1037
1038 _immutable_memory = nullptr; // filled in at first inquiry
1039
1040 #ifdef ASSERT
1041 _phase_optimize_finished = false;
1042 _phase_verify_ideal_loop = false;
1043 _exception_backedge = false;
1044 _type_verify = nullptr;
1045 #endif
1046
1047 // Globally visible Nodes
1048 // First set TOP to null to give safe behavior during creation of RootNode
1049 set_cached_top_node(nullptr);
1050 set_root(new RootNode());
1051 // Now that you have a Root to point to, create the real TOP
1052 set_cached_top_node( new ConNode(Type::TOP) );
1053 set_recent_alloc(nullptr, nullptr);
1054
1055 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1056 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1057 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1058 env()->set_dependencies(new Dependencies(env()));
1059
1060 _fixed_slots = 0;
1061 set_has_split_ifs(false);
1062 set_has_loops(false); // first approximation
1063 set_has_stringbuilder(false);
1064 set_has_boxed_value(false);
1065 _trap_can_recompile = false; // no traps emitted yet
1066 _major_progress = true; // start out assuming good things will happen
1067 set_has_unsafe_access(false);
1068 set_max_vector_size(0);
1069 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1070 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1071 set_decompile_count(0);
1072
1073 #ifndef PRODUCT
1074 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1075 #endif
1076
1077 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1078 _loop_opts_cnt = LoopOptsCount;
1079 _has_flat_accesses = false;
1080 _flat_accesses_share_alias = true;
1081 _scalarize_in_safepoints = false;
1082
1083 set_do_inlining(Inline);
1084 set_max_inline_size(MaxInlineSize);
1085 set_freq_inline_size(FreqInlineSize);
1086 set_do_scheduling(OptoScheduling);
1087
1088 set_do_vector_loop(false);
1089 set_has_monitors(false);
1090 set_has_scoped_access(false);
1091
1092 if (AllowVectorizeOnDemand) {
1093 if (has_method() && _directive->VectorizeOption) {
1094 set_do_vector_loop(true);
1095 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());})
1096 } else if (has_method() && method()->name() != nullptr &&
1097 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1098 set_do_vector_loop(true);
1099 }
1100 }
1101 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1102 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());})
1103
1104 _max_node_limit = _directive->MaxNodeLimitOption;
1105
1106 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1107 set_clinit_barrier_on_entry(true);
1108 }
1109 if (debug_info()->recording_non_safepoints()) {
1110 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1111 (comp_arena(), 8, 0, nullptr));
1112 set_default_node_notes(Node_Notes::make(this));
1113 }
1114
1115 const int grow_ats = 16;
1116 _max_alias_types = grow_ats;
1117 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1118 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1119 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1120 {
1121 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1122 }
1123 // Initialize the first few types.
1124 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1125 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1126 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1127 _num_alias_types = AliasIdxRaw+1;
1128 // Zero out the alias type cache.
1129 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1130 // A null adr_type hits in the cache right away. Preload the right answer.
1131 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1132 }
1133
1134 #ifdef ASSERT
1135 // Verify that the current StartNode is valid.
1136 void Compile::verify_start(StartNode* s) const {
1137 assert(failing_internal() || s == start(), "should be StartNode");
1138 }
1139 #endif
1140
1141 /**
1142 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1143 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1144 * the ideal graph.
1145 */
1146 StartNode* Compile::start() const {
1147 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1148 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1149 Node* start = root()->fast_out(i);
1150 if (start->is_Start()) {
1151 return start->as_Start();
1152 }
1153 }
1154 fatal("Did not find Start node!");
1155 return nullptr;
1156 }
1157
1158 //-------------------------------immutable_memory-------------------------------------
1159 // Access immutable memory
1160 Node* Compile::immutable_memory() {
1161 if (_immutable_memory != nullptr) {
1162 return _immutable_memory;
1163 }
1164 StartNode* s = start();
1165 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1166 Node *p = s->fast_out(i);
1167 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1168 _immutable_memory = p;
1169 return _immutable_memory;
1170 }
1171 }
1172 ShouldNotReachHere();
1173 return nullptr;
1174 }
1175
1176 //----------------------set_cached_top_node------------------------------------
1177 // Install the cached top node, and make sure Node::is_top works correctly.
1178 void Compile::set_cached_top_node(Node* tn) {
1179 if (tn != nullptr) verify_top(tn);
1180 Node* old_top = _top;
1181 _top = tn;
1182 // Calling Node::setup_is_top allows the nodes the chance to adjust
1183 // their _out arrays.
1184 if (_top != nullptr) _top->setup_is_top();
1185 if (old_top != nullptr) old_top->setup_is_top();
1186 assert(_top == nullptr || top()->is_top(), "");
1187 }
1188
1189 #ifdef ASSERT
1190 uint Compile::count_live_nodes_by_graph_walk() {
1191 Unique_Node_List useful(comp_arena());
1192 // Get useful node list by walking the graph.
1193 identify_useful_nodes(useful);
1194 return useful.size();
1195 }
1196
1197 void Compile::print_missing_nodes() {
1198
1199 // Return if CompileLog is null and PrintIdealNodeCount is false.
1200 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1201 return;
1202 }
1203
1204 // This is an expensive function. It is executed only when the user
1205 // specifies VerifyIdealNodeCount option or otherwise knows the
1206 // additional work that needs to be done to identify reachable nodes
1207 // by walking the flow graph and find the missing ones using
1208 // _dead_node_list.
1209
1210 Unique_Node_List useful(comp_arena());
1211 // Get useful node list by walking the graph.
1212 identify_useful_nodes(useful);
1213
1214 uint l_nodes = C->live_nodes();
1215 uint l_nodes_by_walk = useful.size();
1216
1217 if (l_nodes != l_nodes_by_walk) {
1218 if (_log != nullptr) {
1219 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1220 _log->stamp();
1221 _log->end_head();
1222 }
1223 VectorSet& useful_member_set = useful.member_set();
1224 int last_idx = l_nodes_by_walk;
1225 for (int i = 0; i < last_idx; i++) {
1226 if (useful_member_set.test(i)) {
1227 if (_dead_node_list.test(i)) {
1228 if (_log != nullptr) {
1229 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1230 }
1231 if (PrintIdealNodeCount) {
1232 // Print the log message to tty
1233 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1234 useful.at(i)->dump();
1235 }
1236 }
1237 }
1238 else if (! _dead_node_list.test(i)) {
1239 if (_log != nullptr) {
1240 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1241 }
1242 if (PrintIdealNodeCount) {
1243 // Print the log message to tty
1244 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1245 }
1246 }
1247 }
1248 if (_log != nullptr) {
1249 _log->tail("mismatched_nodes");
1250 }
1251 }
1252 }
1253 void Compile::record_modified_node(Node* n) {
1254 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1255 _modified_nodes->push(n);
1256 }
1257 }
1258
1259 void Compile::remove_modified_node(Node* n) {
1260 if (_modified_nodes != nullptr) {
1261 _modified_nodes->remove(n);
1262 }
1263 }
1264 #endif
1265
1266 #ifndef PRODUCT
1267 void Compile::verify_top(Node* tn) const {
1268 if (tn != nullptr) {
1269 assert(tn->is_Con(), "top node must be a constant");
1270 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1271 assert(tn->in(0) != nullptr, "must have live top node");
1272 }
1273 }
1274 #endif
1275
1276
1277 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1278
1279 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1280 guarantee(arr != nullptr, "");
1281 int num_blocks = arr->length();
1282 if (grow_by < num_blocks) grow_by = num_blocks;
1283 int num_notes = grow_by * _node_notes_block_size;
1284 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1285 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1286 while (num_notes > 0) {
1287 arr->append(notes);
1288 notes += _node_notes_block_size;
1289 num_notes -= _node_notes_block_size;
1290 }
1291 assert(num_notes == 0, "exact multiple, please");
1292 }
1293
1294 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1295 if (source == nullptr || dest == nullptr) return false;
1296
1297 if (dest->is_Con())
1298 return false; // Do not push debug info onto constants.
1299
1300 #ifdef ASSERT
1301 // Leave a bread crumb trail pointing to the original node:
1302 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1303 dest->set_debug_orig(source);
1304 }
1305 #endif
1306
1307 if (node_note_array() == nullptr)
1308 return false; // Not collecting any notes now.
1309
1310 // This is a copy onto a pre-existing node, which may already have notes.
1311 // If both nodes have notes, do not overwrite any pre-existing notes.
1312 Node_Notes* source_notes = node_notes_at(source->_idx);
1313 if (source_notes == nullptr || source_notes->is_clear()) return false;
1314 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1315 if (dest_notes == nullptr || dest_notes->is_clear()) {
1316 return set_node_notes_at(dest->_idx, source_notes);
1317 }
1318
1319 Node_Notes merged_notes = (*source_notes);
1320 // The order of operations here ensures that dest notes will win...
1321 merged_notes.update_from(dest_notes);
1322 return set_node_notes_at(dest->_idx, &merged_notes);
1323 }
1324
1325
1326 //--------------------------allow_range_check_smearing-------------------------
1327 // Gating condition for coalescing similar range checks.
1328 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1329 // single covering check that is at least as strong as any of them.
1330 // If the optimization succeeds, the simplified (strengthened) range check
1331 // will always succeed. If it fails, we will deopt, and then give up
1332 // on the optimization.
1333 bool Compile::allow_range_check_smearing() const {
1334 // If this method has already thrown a range-check,
1335 // assume it was because we already tried range smearing
1336 // and it failed.
1337 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1338 return !already_trapped;
1339 }
1340
1341
1342 //------------------------------flatten_alias_type-----------------------------
1343 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1344 assert(do_aliasing(), "Aliasing should be enabled");
1345 int offset = tj->offset();
1346 TypePtr::PTR ptr = tj->ptr();
1347
1348 // Known instance (scalarizable allocation) alias only with itself.
1349 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1350 tj->is_oopptr()->is_known_instance();
1351
1352 // Process weird unsafe references.
1353 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1354 bool default_value_load = EnableValhalla && tj->is_instptr()->instance_klass() == ciEnv::current()->Class_klass();
1355 assert(InlineUnsafeOps || StressReflectiveCode || default_value_load, "indeterminate pointers come only from unsafe ops");
1356 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1357 tj = TypeOopPtr::BOTTOM;
1358 ptr = tj->ptr();
1359 offset = tj->offset();
1360 }
1361
1362 // Array pointers need some flattening
1363 const TypeAryPtr* ta = tj->isa_aryptr();
1364 if (ta && ta->is_stable()) {
1365 // Erase stability property for alias analysis.
1366 tj = ta = ta->cast_to_stable(false);
1367 }
1368 if (ta && ta->is_not_flat()) {
1369 // Erase not flat property for alias analysis.
1370 tj = ta = ta->cast_to_not_flat(false);
1371 }
1372 if (ta && ta->is_not_null_free()) {
1373 // Erase not null free property for alias analysis.
1374 tj = ta = ta->cast_to_not_null_free(false);
1375 }
1376
1377 if( ta && is_known_inst ) {
1378 if ( offset != Type::OffsetBot &&
1379 offset > arrayOopDesc::length_offset_in_bytes() ) {
1380 offset = Type::OffsetBot; // Flatten constant access into array body only
1381 tj = ta = ta->
1382 remove_speculative()->
1383 cast_to_ptr_type(ptr)->
1384 with_offset(offset);
1385 }
1386 } else if (ta) {
1387 // For arrays indexed by constant indices, we flatten the alias
1388 // space to include all of the array body. Only the header, klass
1389 // and array length can be accessed un-aliased.
1390 // For flat inline type array, each field has its own slice so
1391 // we must include the field offset.
1392 if( offset != Type::OffsetBot ) {
1393 if( ta->const_oop() ) { // MethodData* or Method*
1394 offset = Type::OffsetBot; // Flatten constant access into array body
1395 tj = ta = ta->
1396 remove_speculative()->
1397 cast_to_ptr_type(ptr)->
1398 cast_to_exactness(false)->
1399 with_offset(offset);
1400 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1401 // range is OK as-is.
1402 tj = ta = TypeAryPtr::RANGE;
1403 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1404 tj = TypeInstPtr::KLASS; // all klass loads look alike
1405 ta = TypeAryPtr::RANGE; // generic ignored junk
1406 ptr = TypePtr::BotPTR;
1407 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1408 tj = TypeInstPtr::MARK;
1409 ta = TypeAryPtr::RANGE; // generic ignored junk
1410 ptr = TypePtr::BotPTR;
1411 } else { // Random constant offset into array body
1412 offset = Type::OffsetBot; // Flatten constant access into array body
1413 tj = ta = ta->
1414 remove_speculative()->
1415 cast_to_ptr_type(ptr)->
1416 cast_to_exactness(false)->
1417 with_offset(offset);
1418 }
1419 }
1420 // Arrays of fixed size alias with arrays of unknown size.
1421 if (ta->size() != TypeInt::POS) {
1422 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1423 tj = ta = ta->
1424 remove_speculative()->
1425 cast_to_ptr_type(ptr)->
1426 with_ary(tary)->
1427 cast_to_exactness(false);
1428 }
1429 // Arrays of known objects become arrays of unknown objects.
1430 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1431 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1432 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1433 }
1434 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1435 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1436 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1437 }
1438 // Initially all flattened array accesses share a single slice
1439 if (ta->is_flat() && ta->elem() != TypeInstPtr::BOTTOM && _flat_accesses_share_alias) {
1440 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size(), /* stable= */ false, /* flat= */ true);
1441 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1442 }
1443 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1444 // cannot be distinguished by bytecode alone.
1445 if (ta->elem() == TypeInt::BOOL) {
1446 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1447 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1448 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1449 }
1450 // During the 2nd round of IterGVN, NotNull castings are removed.
1451 // Make sure the Bottom and NotNull variants alias the same.
1452 // Also, make sure exact and non-exact variants alias the same.
1453 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1454 tj = ta = ta->
1455 remove_speculative()->
1456 cast_to_ptr_type(TypePtr::BotPTR)->
1457 cast_to_exactness(false)->
1458 with_offset(offset);
1459 }
1460 }
1461
1462 // Oop pointers need some flattening
1463 const TypeInstPtr *to = tj->isa_instptr();
1464 if (to && to != TypeOopPtr::BOTTOM) {
1465 ciInstanceKlass* ik = to->instance_klass();
1466 if( ptr == TypePtr::Constant ) {
1467 if (ik != ciEnv::current()->Class_klass() ||
1468 offset < ik->layout_helper_size_in_bytes()) {
1469 // No constant oop pointers (such as Strings); they alias with
1470 // unknown strings.
1471 assert(!is_known_inst, "not scalarizable allocation");
1472 tj = to = to->
1473 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1474 remove_speculative()->
1475 cast_to_ptr_type(TypePtr::BotPTR)->
1476 cast_to_exactness(false);
1477 }
1478 } else if( is_known_inst ) {
1479 tj = to; // Keep NotNull and klass_is_exact for instance type
1480 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1481 // During the 2nd round of IterGVN, NotNull castings are removed.
1482 // Make sure the Bottom and NotNull variants alias the same.
1483 // Also, make sure exact and non-exact variants alias the same.
1484 tj = to = to->
1485 remove_speculative()->
1486 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1487 cast_to_ptr_type(TypePtr::BotPTR)->
1488 cast_to_exactness(false);
1489 }
1490 if (to->speculative() != nullptr) {
1491 tj = to = to->remove_speculative();
1492 }
1493 // Canonicalize the holder of this field
1494 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1495 // First handle header references such as a LoadKlassNode, even if the
1496 // object's klass is unloaded at compile time (4965979).
1497 if (!is_known_inst) { // Do it only for non-instance types
1498 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1499 }
1500 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1501 // Static fields are in the space above the normal instance
1502 // fields in the java.lang.Class instance.
1503 if (ik != ciEnv::current()->Class_klass()) {
1504 to = nullptr;
1505 tj = TypeOopPtr::BOTTOM;
1506 offset = tj->offset();
1507 }
1508 } else {
1509 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1510 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1511 assert(tj->offset() == offset, "no change to offset expected");
1512 bool xk = to->klass_is_exact();
1513 int instance_id = to->instance_id();
1514
1515 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1516 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1517 // its interfaces are included.
1518 if (xk && ik->equals(canonical_holder)) {
1519 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id), "exact type should be canonical type");
1520 } else {
1521 assert(xk || !is_known_inst, "Known instance should be exact type");
1522 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id);
1523 }
1524 }
1525 }
1526
1527 // Klass pointers to object array klasses need some flattening
1528 const TypeKlassPtr *tk = tj->isa_klassptr();
1529 if( tk ) {
1530 // If we are referencing a field within a Klass, we need
1531 // to assume the worst case of an Object. Both exact and
1532 // inexact types must flatten to the same alias class so
1533 // use NotNull as the PTR.
1534 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1535 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1536 env()->Object_klass(),
1537 Type::Offset(offset));
1538 }
1539
1540 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1541 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1542 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1543 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset));
1544 } else {
1545 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_flat(), tk->is_null_free());
1546 }
1547 }
1548 // Check for precise loads from the primary supertype array and force them
1549 // to the supertype cache alias index. Check for generic array loads from
1550 // the primary supertype array and also force them to the supertype cache
1551 // alias index. Since the same load can reach both, we need to merge
1552 // these 2 disparate memories into the same alias class. Since the
1553 // primary supertype array is read-only, there's no chance of confusion
1554 // where we bypass an array load and an array store.
1555 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1556 if (offset == Type::OffsetBot ||
1557 (offset >= primary_supers_offset &&
1558 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1559 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1560 offset = in_bytes(Klass::secondary_super_cache_offset());
1561 tj = tk = tk->with_offset(offset);
1562 }
1563 }
1564
1565 // Flatten all Raw pointers together.
1566 if (tj->base() == Type::RawPtr)
1567 tj = TypeRawPtr::BOTTOM;
1568
1569 if (tj->base() == Type::AnyPtr)
1570 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1571
1572 offset = tj->offset();
1573 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1574
1575 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1576 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1577 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1578 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1579 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1580 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1581 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1582 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1583 assert( tj->ptr() != TypePtr::TopPTR &&
1584 tj->ptr() != TypePtr::AnyNull &&
1585 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1586 // assert( tj->ptr() != TypePtr::Constant ||
1587 // tj->base() == Type::RawPtr ||
1588 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1589
1590 return tj;
1591 }
1592
1593 void Compile::AliasType::Init(int i, const TypePtr* at) {
1594 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1595 _index = i;
1596 _adr_type = at;
1597 _field = nullptr;
1598 _element = nullptr;
1599 _is_rewritable = true; // default
1600 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1601 if (atoop != nullptr && atoop->is_known_instance()) {
1602 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1603 _general_index = Compile::current()->get_alias_index(gt);
1604 } else {
1605 _general_index = 0;
1606 }
1607 }
1608
1609 BasicType Compile::AliasType::basic_type() const {
1610 if (element() != nullptr) {
1611 const Type* element = adr_type()->is_aryptr()->elem();
1612 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1613 } if (field() != nullptr) {
1614 return field()->layout_type();
1615 } else {
1616 return T_ILLEGAL; // unknown
1617 }
1618 }
1619
1620 //---------------------------------print_on------------------------------------
1621 #ifndef PRODUCT
1622 void Compile::AliasType::print_on(outputStream* st) {
1623 if (index() < 10)
1624 st->print("@ <%d> ", index());
1625 else st->print("@ <%d>", index());
1626 st->print(is_rewritable() ? " " : " RO");
1627 int offset = adr_type()->offset();
1628 if (offset == Type::OffsetBot)
1629 st->print(" +any");
1630 else st->print(" +%-3d", offset);
1631 st->print(" in ");
1632 adr_type()->dump_on(st);
1633 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1634 if (field() != nullptr && tjp) {
1635 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1636 tjp->offset() != field()->offset_in_bytes()) {
1637 st->print(" != ");
1638 field()->print();
1639 st->print(" ***");
1640 }
1641 }
1642 }
1643
1644 void print_alias_types() {
1645 Compile* C = Compile::current();
1646 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1647 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1648 C->alias_type(idx)->print_on(tty);
1649 tty->cr();
1650 }
1651 }
1652 #endif
1653
1654
1655 //----------------------------probe_alias_cache--------------------------------
1656 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1657 intptr_t key = (intptr_t) adr_type;
1658 key ^= key >> logAliasCacheSize;
1659 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1660 }
1661
1662
1663 //-----------------------------grow_alias_types--------------------------------
1664 void Compile::grow_alias_types() {
1665 const int old_ats = _max_alias_types; // how many before?
1666 const int new_ats = old_ats; // how many more?
1667 const int grow_ats = old_ats+new_ats; // how many now?
1668 _max_alias_types = grow_ats;
1669 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1670 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1671 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1672 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1673 }
1674
1675
1676 //--------------------------------find_alias_type------------------------------
1677 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1678 if (!do_aliasing()) {
1679 return alias_type(AliasIdxBot);
1680 }
1681
1682 AliasCacheEntry* ace = nullptr;
1683 if (!uncached) {
1684 ace = probe_alias_cache(adr_type);
1685 if (ace->_adr_type == adr_type) {
1686 return alias_type(ace->_index);
1687 }
1688 }
1689
1690 // Handle special cases.
1691 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1692 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1693
1694 // Do it the slow way.
1695 const TypePtr* flat = flatten_alias_type(adr_type);
1696
1697 #ifdef ASSERT
1698 {
1699 ResourceMark rm;
1700 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1701 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1702 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1703 Type::str(adr_type));
1704 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1705 const TypeOopPtr* foop = flat->is_oopptr();
1706 // Scalarizable allocations have exact klass always.
1707 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1708 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1709 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1710 Type::str(foop), Type::str(xoop));
1711 }
1712 }
1713 #endif
1714
1715 int idx = AliasIdxTop;
1716 for (int i = 0; i < num_alias_types(); i++) {
1717 if (alias_type(i)->adr_type() == flat) {
1718 idx = i;
1719 break;
1720 }
1721 }
1722
1723 if (idx == AliasIdxTop) {
1724 if (no_create) return nullptr;
1725 // Grow the array if necessary.
1726 if (_num_alias_types == _max_alias_types) grow_alias_types();
1727 // Add a new alias type.
1728 idx = _num_alias_types++;
1729 _alias_types[idx]->Init(idx, flat);
1730 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1731 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1732 if (flat->isa_instptr()) {
1733 if (flat->offset() == java_lang_Class::klass_offset()
1734 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1735 alias_type(idx)->set_rewritable(false);
1736 }
1737 ciField* field = nullptr;
1738 if (flat->isa_aryptr()) {
1739 #ifdef ASSERT
1740 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1741 // (T_BYTE has the weakest alignment and size restrictions...)
1742 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1743 #endif
1744 const Type* elemtype = flat->is_aryptr()->elem();
1745 if (flat->offset() == TypePtr::OffsetBot) {
1746 alias_type(idx)->set_element(elemtype);
1747 }
1748 int field_offset = flat->is_aryptr()->field_offset().get();
1749 if (flat->is_flat() &&
1750 field_offset != Type::OffsetBot) {
1751 ciInlineKlass* vk = elemtype->inline_klass();
1752 field_offset += vk->payload_offset();
1753 field = vk->get_field_by_offset(field_offset, false);
1754 }
1755 }
1756 if (flat->isa_klassptr()) {
1757 if (UseCompactObjectHeaders) {
1758 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1759 alias_type(idx)->set_rewritable(false);
1760 }
1761 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1762 alias_type(idx)->set_rewritable(false);
1763 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1764 alias_type(idx)->set_rewritable(false);
1765 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1766 alias_type(idx)->set_rewritable(false);
1767 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1768 alias_type(idx)->set_rewritable(false);
1769 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1770 alias_type(idx)->set_rewritable(false);
1771 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1772 alias_type(idx)->set_rewritable(false);
1773 }
1774 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1775 // but the base pointer type is not distinctive enough to identify
1776 // references into JavaThread.)
1777
1778 // Check for final fields.
1779 const TypeInstPtr* tinst = flat->isa_instptr();
1780 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1781 if (tinst->const_oop() != nullptr &&
1782 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1783 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1784 // static field
1785 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1786 field = k->get_field_by_offset(tinst->offset(), true);
1787 } else if (tinst->is_inlinetypeptr()) {
1788 // Inline type field
1789 ciInlineKlass* vk = tinst->inline_klass();
1790 field = vk->get_field_by_offset(tinst->offset(), false);
1791 } else {
1792 ciInstanceKlass *k = tinst->instance_klass();
1793 field = k->get_field_by_offset(tinst->offset(), false);
1794 }
1795 }
1796 assert(field == nullptr ||
1797 original_field == nullptr ||
1798 (field->holder() == original_field->holder() &&
1799 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1800 field->is_static() == original_field->is_static()), "wrong field?");
1801 // Set field() and is_rewritable() attributes.
1802 if (field != nullptr) {
1803 alias_type(idx)->set_field(field);
1804 if (flat->isa_aryptr()) {
1805 // Fields of flat arrays are rewritable although they are declared final
1806 assert(flat->is_flat(), "must be a flat array");
1807 alias_type(idx)->set_rewritable(true);
1808 }
1809 }
1810 }
1811
1812 // Fill the cache for next time.
1813 if (!uncached) {
1814 ace->_adr_type = adr_type;
1815 ace->_index = idx;
1816 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1817
1818 // Might as well try to fill the cache for the flattened version, too.
1819 AliasCacheEntry* face = probe_alias_cache(flat);
1820 if (face->_adr_type == nullptr) {
1821 face->_adr_type = flat;
1822 face->_index = idx;
1823 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1824 }
1825 }
1826
1827 return alias_type(idx);
1828 }
1829
1830
1831 Compile::AliasType* Compile::alias_type(ciField* field) {
1832 const TypeOopPtr* t;
1833 if (field->is_static())
1834 t = TypeInstPtr::make(field->holder()->java_mirror());
1835 else
1836 t = TypeOopPtr::make_from_klass_raw(field->holder());
1837 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1838 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1839 return atp;
1840 }
1841
1842
1843 //------------------------------have_alias_type--------------------------------
1844 bool Compile::have_alias_type(const TypePtr* adr_type) {
1845 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1846 if (ace->_adr_type == adr_type) {
1847 return true;
1848 }
1849
1850 // Handle special cases.
1851 if (adr_type == nullptr) return true;
1852 if (adr_type == TypePtr::BOTTOM) return true;
1853
1854 return find_alias_type(adr_type, true, nullptr) != nullptr;
1855 }
1856
1857 //-----------------------------must_alias--------------------------------------
1858 // True if all values of the given address type are in the given alias category.
1859 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1860 if (alias_idx == AliasIdxBot) return true; // the universal category
1861 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1862 if (alias_idx == AliasIdxTop) return false; // the empty category
1863 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1864
1865 // the only remaining possible overlap is identity
1866 int adr_idx = get_alias_index(adr_type);
1867 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1868 assert(adr_idx == alias_idx ||
1869 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1870 && adr_type != TypeOopPtr::BOTTOM),
1871 "should not be testing for overlap with an unsafe pointer");
1872 return adr_idx == alias_idx;
1873 }
1874
1875 //------------------------------can_alias--------------------------------------
1876 // True if any values of the given address type are in the given alias category.
1877 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1878 if (alias_idx == AliasIdxTop) return false; // the empty category
1879 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1880 // Known instance doesn't alias with bottom memory
1881 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1882 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1883
1884 // the only remaining possible overlap is identity
1885 int adr_idx = get_alias_index(adr_type);
1886 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1887 return adr_idx == alias_idx;
1888 }
1889
1890 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1891 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1892 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1893 if (parse_predicate_count() == 0) {
1894 return;
1895 }
1896 for (int i = 0; i < parse_predicate_count(); i++) {
1897 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1898 parse_predicate->mark_useless();
1899 igvn._worklist.push(parse_predicate);
1900 }
1901 _parse_predicates.clear();
1902 }
1903
1904 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1905 if (!n->for_post_loop_opts_igvn()) {
1906 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1907 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1908 _for_post_loop_igvn.append(n);
1909 }
1910 }
1911
1912 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1913 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1914 _for_post_loop_igvn.remove(n);
1915 }
1916
1917 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1918 // Verify that all previous optimizations produced a valid graph
1919 // at least to this point, even if no loop optimizations were done.
1920 PhaseIdealLoop::verify(igvn);
1921
1922 C->set_post_loop_opts_phase(); // no more loop opts allowed
1923
1924 assert(!C->major_progress(), "not cleared");
1925
1926 if (_for_post_loop_igvn.length() > 0) {
1927 while (_for_post_loop_igvn.length() > 0) {
1928 Node* n = _for_post_loop_igvn.pop();
1929 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1930 igvn._worklist.push(n);
1931 }
1932 igvn.optimize();
1933 if (failing()) return;
1934 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1935 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1936
1937 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1938 if (C->major_progress()) {
1939 C->clear_major_progress(); // ensure that major progress is now clear
1940 }
1941 }
1942 }
1943
1944 void Compile::add_inline_type(Node* n) {
1945 assert(n->is_InlineType(), "unexpected node");
1946 _inline_type_nodes.push(n);
1947 }
1948
1949 void Compile::remove_inline_type(Node* n) {
1950 assert(n->is_InlineType(), "unexpected node");
1951 if (_inline_type_nodes.contains(n)) {
1952 _inline_type_nodes.remove(n);
1953 }
1954 }
1955
1956 // Does the return value keep otherwise useless inline type allocations alive?
1957 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1958 ResourceMark rm;
1959 Unique_Node_List wq;
1960 wq.push(ret_val);
1961 bool some_allocations = false;
1962 for (uint i = 0; i < wq.size(); i++) {
1963 Node* n = wq.at(i);
1964 if (n->outcnt() > 1) {
1965 // Some other use for the allocation
1966 return false;
1967 } else if (n->is_InlineType()) {
1968 wq.push(n->in(1));
1969 } else if (n->is_Phi()) {
1970 for (uint j = 1; j < n->req(); j++) {
1971 wq.push(n->in(j));
1972 }
1973 } else if (n->is_CheckCastPP() &&
1974 n->in(1)->is_Proj() &&
1975 n->in(1)->in(0)->is_Allocate()) {
1976 some_allocations = true;
1977 } else if (n->is_CheckCastPP()) {
1978 wq.push(n->in(1));
1979 }
1980 }
1981 return some_allocations;
1982 }
1983
1984 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1985 // Make sure that the return value does not keep an otherwise unused allocation alive
1986 if (tf()->returns_inline_type_as_fields()) {
1987 Node* ret = nullptr;
1988 for (uint i = 1; i < root()->req(); i++) {
1989 Node* in = root()->in(i);
1990 if (in->Opcode() == Op_Return) {
1991 assert(ret == nullptr, "only one return");
1992 ret = in;
1993 }
1994 }
1995 if (ret != nullptr) {
1996 Node* ret_val = ret->in(TypeFunc::Parms);
1997 if (igvn.type(ret_val)->isa_oopptr() &&
1998 return_val_keeps_allocations_alive(ret_val)) {
1999 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
2000 assert(ret_val->outcnt() == 0, "should be dead now");
2001 igvn.remove_dead_node(ret_val);
2002 }
2003 }
2004 }
2005 if (_inline_type_nodes.length() == 0) {
2006 return;
2007 }
2008 // Scalarize inline types in safepoint debug info.
2009 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2010 set_scalarize_in_safepoints(true);
2011 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2012 InlineTypeNode* vt = _inline_type_nodes.at(i)->as_InlineType();
2013 vt->make_scalar_in_safepoints(&igvn);
2014 igvn.record_for_igvn(vt);
2015 }
2016 if (remove) {
2017 // Remove inline type nodes by replacing them with their oop input
2018 while (_inline_type_nodes.length() > 0) {
2019 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2020 if (vt->outcnt() == 0) {
2021 igvn.remove_dead_node(vt);
2022 continue;
2023 }
2024 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2025 DEBUG_ONLY(bool must_be_buffered = false);
2026 Node* u = vt->out(i);
2027 // Check if any users are blackholes. If so, rewrite them to use either the
2028 // allocated buffer, or individual components, instead of the inline type node
2029 // that goes away.
2030 if (u->is_Blackhole()) {
2031 BlackholeNode* bh = u->as_Blackhole();
2032
2033 // Unlink the old input
2034 int idx = bh->find_edge(vt);
2035 assert(idx != -1, "The edge should be there");
2036 bh->del_req(idx);
2037 --i;
2038
2039 if (vt->is_allocated(&igvn)) {
2040 // Already has the allocated instance, blackhole that
2041 bh->add_req(vt->get_oop());
2042 } else {
2043 // Not allocated yet, blackhole the components
2044 for (uint c = 0; c < vt->field_count(); c++) {
2045 bh->add_req(vt->field_value(c));
2046 }
2047 }
2048
2049 // Node modified, record for IGVN
2050 igvn.record_for_igvn(bh);
2051 }
2052 #ifdef ASSERT
2053 // Verify that inline type is buffered when replacing by oop
2054 else if (u->is_InlineType()) {
2055 // InlineType uses don't need buffering because they are about to be replaced as well
2056 } else if (u->is_Phi()) {
2057 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2058 } else {
2059 must_be_buffered = true;
2060 }
2061 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2062 vt->dump(0);
2063 u->dump(0);
2064 assert(false, "Should have been buffered");
2065 }
2066 #endif
2067 }
2068 igvn.replace_node(vt, vt->get_oop());
2069 }
2070 }
2071 igvn.optimize();
2072 }
2073
2074 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2075 if (!_has_flat_accesses) {
2076 return;
2077 }
2078 // Initially, all flat array accesses share the same slice to
2079 // keep dependencies with Object[] array accesses (that could be
2080 // to a flat array) correct. We're done with parsing so we
2081 // now know all flat array accesses in this compile
2082 // unit. Let's move flat array accesses to their own slice,
2083 // one per element field. This should help memory access
2084 // optimizations.
2085 ResourceMark rm;
2086 Unique_Node_List wq;
2087 wq.push(root());
2088
2089 Node_List mergememnodes;
2090 Node_List memnodes;
2091
2092 // Alias index currently shared by all flat memory accesses
2093 int index = get_alias_index(TypeAryPtr::INLINES);
2094
2095 // Find MergeMem nodes and flat array accesses
2096 for (uint i = 0; i < wq.size(); i++) {
2097 Node* n = wq.at(i);
2098 if (n->is_Mem()) {
2099 const TypePtr* adr_type = nullptr;
2100 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2101 if (adr_type == TypeAryPtr::INLINES) {
2102 memnodes.push(n);
2103 }
2104 } else if (n->is_MergeMem()) {
2105 MergeMemNode* mm = n->as_MergeMem();
2106 if (mm->memory_at(index) != mm->base_memory()) {
2107 mergememnodes.push(n);
2108 }
2109 }
2110 for (uint j = 0; j < n->req(); j++) {
2111 Node* m = n->in(j);
2112 if (m != nullptr) {
2113 wq.push(m);
2114 }
2115 }
2116 }
2117
2118 if (memnodes.size() > 0) {
2119 _flat_accesses_share_alias = false;
2120
2121 // We are going to change the slice for the flat array
2122 // accesses so we need to clear the cache entries that refer to
2123 // them.
2124 for (uint i = 0; i < AliasCacheSize; i++) {
2125 AliasCacheEntry* ace = &_alias_cache[i];
2126 if (ace->_adr_type != nullptr &&
2127 ace->_adr_type->is_flat()) {
2128 ace->_adr_type = nullptr;
2129 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2130 }
2131 }
2132
2133 // Find what aliases we are going to add
2134 int start_alias = num_alias_types()-1;
2135 int stop_alias = 0;
2136
2137 for (uint i = 0; i < memnodes.size(); i++) {
2138 Node* m = memnodes.at(i);
2139 const TypePtr* adr_type = nullptr;
2140 adr_type = m->adr_type();
2141 #ifdef ASSERT
2142 m->as_Mem()->set_adr_type(adr_type);
2143 #endif
2144 int idx = get_alias_index(adr_type);
2145 start_alias = MIN2(start_alias, idx);
2146 stop_alias = MAX2(stop_alias, idx);
2147 }
2148
2149 assert(stop_alias >= start_alias, "should have expanded aliases");
2150
2151 Node_Stack stack(0);
2152 #ifdef ASSERT
2153 VectorSet seen(Thread::current()->resource_area());
2154 #endif
2155 // Now let's fix the memory graph so each flat array access
2156 // is moved to the right slice. Start from the MergeMem nodes.
2157 uint last = unique();
2158 for (uint i = 0; i < mergememnodes.size(); i++) {
2159 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2160 Node* n = current->memory_at(index);
2161 MergeMemNode* mm = nullptr;
2162 do {
2163 // Follow memory edges through memory accesses, phis and
2164 // narrow membars and push nodes on the stack. Once we hit
2165 // bottom memory, we pop element off the stack one at a
2166 // time, in reverse order, and move them to the right slice
2167 // by changing their memory edges.
2168 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2169 assert(!seen.test_set(n->_idx), "");
2170 // Uses (a load for instance) will need to be moved to the
2171 // right slice as well and will get a new memory state
2172 // that we don't know yet. The use could also be the
2173 // backedge of a loop. We put a place holder node between
2174 // the memory node and its uses. We replace that place
2175 // holder with the correct memory state once we know it,
2176 // i.e. when nodes are popped off the stack. Using the
2177 // place holder make the logic work in the presence of
2178 // loops.
2179 if (n->outcnt() > 1) {
2180 Node* place_holder = nullptr;
2181 assert(!n->has_out_with(Op_Node), "");
2182 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2183 Node* u = n->out(k);
2184 if (u != current && u->_idx < last) {
2185 bool success = false;
2186 for (uint l = 0; l < u->req(); l++) {
2187 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2188 continue;
2189 }
2190 Node* in = u->in(l);
2191 if (in == n) {
2192 if (place_holder == nullptr) {
2193 place_holder = new Node(1);
2194 place_holder->init_req(0, n);
2195 }
2196 igvn.replace_input_of(u, l, place_holder);
2197 success = true;
2198 }
2199 }
2200 if (success) {
2201 --k;
2202 }
2203 }
2204 }
2205 }
2206 if (n->is_Phi()) {
2207 stack.push(n, 1);
2208 n = n->in(1);
2209 } else if (n->is_Mem()) {
2210 stack.push(n, n->req());
2211 n = n->in(MemNode::Memory);
2212 } else {
2213 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2214 stack.push(n, n->req());
2215 n = n->in(0)->in(TypeFunc::Memory);
2216 }
2217 } else {
2218 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2219 // Build a new MergeMem node to carry the new memory state
2220 // as we build it. IGVN should fold extraneous MergeMem
2221 // nodes.
2222 mm = MergeMemNode::make(n);
2223 igvn.register_new_node_with_optimizer(mm);
2224 while (stack.size() > 0) {
2225 Node* m = stack.node();
2226 uint idx = stack.index();
2227 if (m->is_Mem()) {
2228 // Move memory node to its new slice
2229 const TypePtr* adr_type = m->adr_type();
2230 int alias = get_alias_index(adr_type);
2231 Node* prev = mm->memory_at(alias);
2232 igvn.replace_input_of(m, MemNode::Memory, prev);
2233 mm->set_memory_at(alias, m);
2234 } else if (m->is_Phi()) {
2235 // We need as many new phis as there are new aliases
2236 igvn.replace_input_of(m, idx, mm);
2237 if (idx == m->req()-1) {
2238 Node* r = m->in(0);
2239 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2240 const TypePtr* adr_type = get_adr_type(j);
2241 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2242 continue;
2243 }
2244 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2245 igvn.register_new_node_with_optimizer(phi);
2246 for (uint k = 1; k < m->req(); k++) {
2247 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2248 }
2249 mm->set_memory_at(j, phi);
2250 }
2251 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2252 igvn.register_new_node_with_optimizer(base_phi);
2253 for (uint k = 1; k < m->req(); k++) {
2254 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2255 }
2256 mm->set_base_memory(base_phi);
2257 }
2258 } else {
2259 // This is a MemBarCPUOrder node from
2260 // Parse::array_load()/Parse::array_store(), in the
2261 // branch that handles flat arrays hidden under
2262 // an Object[] array. We also need one new membar per
2263 // new alias to keep the unknown access that the
2264 // membars protect properly ordered with accesses to
2265 // known flat array.
2266 assert(m->is_Proj(), "projection expected");
2267 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2268 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2269 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2270 const TypePtr* adr_type = get_adr_type(j);
2271 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2272 continue;
2273 }
2274 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2275 igvn.register_new_node_with_optimizer(mb);
2276 Node* mem = mm->memory_at(j);
2277 mb->init_req(TypeFunc::Control, ctrl);
2278 mb->init_req(TypeFunc::Memory, mem);
2279 ctrl = new ProjNode(mb, TypeFunc::Control);
2280 igvn.register_new_node_with_optimizer(ctrl);
2281 mem = new ProjNode(mb, TypeFunc::Memory);
2282 igvn.register_new_node_with_optimizer(mem);
2283 mm->set_memory_at(j, mem);
2284 }
2285 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2286 }
2287 if (idx < m->req()-1) {
2288 idx += 1;
2289 stack.set_index(idx);
2290 n = m->in(idx);
2291 break;
2292 }
2293 // Take care of place holder nodes
2294 if (m->has_out_with(Op_Node)) {
2295 Node* place_holder = m->find_out_with(Op_Node);
2296 if (place_holder != nullptr) {
2297 Node* mm_clone = mm->clone();
2298 igvn.register_new_node_with_optimizer(mm_clone);
2299 Node* hook = new Node(1);
2300 hook->init_req(0, mm);
2301 igvn.replace_node(place_holder, mm_clone);
2302 hook->destruct(&igvn);
2303 }
2304 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2305 }
2306 stack.pop();
2307 }
2308 }
2309 } while(stack.size() > 0);
2310 // Fix the memory state at the MergeMem we started from
2311 igvn.rehash_node_delayed(current);
2312 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2313 const TypePtr* adr_type = get_adr_type(j);
2314 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2315 continue;
2316 }
2317 current->set_memory_at(j, mm);
2318 }
2319 current->set_memory_at(index, current->base_memory());
2320 }
2321 igvn.optimize();
2322 }
2323 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2324 #ifdef ASSERT
2325 if (!_flat_accesses_share_alias) {
2326 wq.clear();
2327 wq.push(root());
2328 for (uint i = 0; i < wq.size(); i++) {
2329 Node* n = wq.at(i);
2330 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2331 for (uint j = 0; j < n->req(); j++) {
2332 Node* m = n->in(j);
2333 if (m != nullptr) {
2334 wq.push(m);
2335 }
2336 }
2337 }
2338 }
2339 #endif
2340 }
2341
2342 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2343 if (OptimizeUnstableIf) {
2344 _unstable_if_traps.append(trap);
2345 }
2346 }
2347
2348 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2349 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2350 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2351 Node* n = trap->uncommon_trap();
2352 if (!useful.member(n)) {
2353 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2354 }
2355 }
2356 }
2357
2358 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2359 // or fold-compares case. Return true if succeed or not found.
2360 //
2361 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2362 // false and ask fold-compares to yield.
2363 //
2364 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2365 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2366 // when deoptimization does happen.
2367 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2368 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2369 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2370 if (trap->uncommon_trap() == unc) {
2371 if (yield && trap->modified()) {
2372 return false;
2373 }
2374 _unstable_if_traps.delete_at(i);
2375 break;
2376 }
2377 }
2378 return true;
2379 }
2380
2381 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2382 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2383 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2384 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2385 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2386 CallStaticJavaNode* unc = trap->uncommon_trap();
2387 int next_bci = trap->next_bci();
2388 bool modified = trap->modified();
2389
2390 if (next_bci != -1 && !modified) {
2391 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2392 JVMState* jvms = unc->jvms();
2393 ciMethod* method = jvms->method();
2394 ciBytecodeStream iter(method);
2395
2396 iter.force_bci(jvms->bci());
2397 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2398 Bytecodes::Code c = iter.cur_bc();
2399 Node* lhs = nullptr;
2400 Node* rhs = nullptr;
2401 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2402 lhs = unc->peek_operand(0);
2403 rhs = unc->peek_operand(1);
2404 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2405 lhs = unc->peek_operand(0);
2406 }
2407
2408 ResourceMark rm;
2409 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2410 assert(live_locals.is_valid(), "broken liveness info");
2411 int len = (int)live_locals.size();
2412
2413 for (int i = 0; i < len; i++) {
2414 Node* local = unc->local(jvms, i);
2415 // kill local using the liveness of next_bci.
2416 // give up when the local looks like an operand to secure reexecution.
2417 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2418 uint idx = jvms->locoff() + i;
2419 #ifdef ASSERT
2420 if (PrintOpto && Verbose) {
2421 tty->print("[unstable_if] kill local#%d: ", idx);
2422 local->dump();
2423 tty->cr();
2424 }
2425 #endif
2426 igvn.replace_input_of(unc, idx, top());
2427 modified = true;
2428 }
2429 }
2430 }
2431
2432 // keep the modified trap for late query
2433 if (modified) {
2434 trap->set_modified();
2435 } else {
2436 _unstable_if_traps.delete_at(i);
2437 }
2438 }
2439 igvn.optimize();
2440 }
2441
2442 // StringOpts and late inlining of string methods
2443 void Compile::inline_string_calls(bool parse_time) {
2444 {
2445 // remove useless nodes to make the usage analysis simpler
2446 ResourceMark rm;
2447 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2448 }
2449
2450 {
2451 ResourceMark rm;
2452 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2453 PhaseStringOpts pso(initial_gvn());
2454 print_method(PHASE_AFTER_STRINGOPTS, 3);
2455 }
2456
2457 // now inline anything that we skipped the first time around
2458 if (!parse_time) {
2459 _late_inlines_pos = _late_inlines.length();
2460 }
2461
2462 while (_string_late_inlines.length() > 0) {
2463 CallGenerator* cg = _string_late_inlines.pop();
2464 cg->do_late_inline();
2465 if (failing()) return;
2466 }
2467 _string_late_inlines.trunc_to(0);
2468 }
2469
2470 // Late inlining of boxing methods
2471 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2472 if (_boxing_late_inlines.length() > 0) {
2473 assert(has_boxed_value(), "inconsistent");
2474
2475 set_inlining_incrementally(true);
2476
2477 igvn_worklist()->ensure_empty(); // should be done with igvn
2478
2479 _late_inlines_pos = _late_inlines.length();
2480
2481 while (_boxing_late_inlines.length() > 0) {
2482 CallGenerator* cg = _boxing_late_inlines.pop();
2483 cg->do_late_inline();
2484 if (failing()) return;
2485 }
2486 _boxing_late_inlines.trunc_to(0);
2487
2488 inline_incrementally_cleanup(igvn);
2489
2490 set_inlining_incrementally(false);
2491 }
2492 }
2493
2494 bool Compile::inline_incrementally_one() {
2495 assert(IncrementalInline, "incremental inlining should be on");
2496
2497 TracePhase tp(_t_incrInline_inline);
2498
2499 set_inlining_progress(false);
2500 set_do_cleanup(false);
2501
2502 for (int i = 0; i < _late_inlines.length(); i++) {
2503 _late_inlines_pos = i+1;
2504 CallGenerator* cg = _late_inlines.at(i);
2505 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2506 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2507 cg->do_late_inline();
2508 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2509 if (failing()) {
2510 return false;
2511 } else if (inlining_progress()) {
2512 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2513 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2514 break; // process one call site at a time
2515 }
2516 } else {
2517 // Ignore late inline direct calls when inlining is not allowed.
2518 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2519 }
2520 }
2521 // Remove processed elements.
2522 _late_inlines.remove_till(_late_inlines_pos);
2523 _late_inlines_pos = 0;
2524
2525 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2526
2527 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2528
2529 set_inlining_progress(false);
2530 set_do_cleanup(false);
2531
2532 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2533 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2534 }
2535
2536 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2537 {
2538 TracePhase tp(_t_incrInline_pru);
2539 ResourceMark rm;
2540 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2541 }
2542 {
2543 TracePhase tp(_t_incrInline_igvn);
2544 igvn.reset_from_gvn(initial_gvn());
2545 igvn.optimize();
2546 if (failing()) return;
2547 }
2548 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2549 }
2550
2551 // Perform incremental inlining until bound on number of live nodes is reached
2552 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2553 TracePhase tp(_t_incrInline);
2554
2555 set_inlining_incrementally(true);
2556 uint low_live_nodes = 0;
2557
2558 while (_late_inlines.length() > 0) {
2559 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2560 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2561 TracePhase tp(_t_incrInline_ideal);
2562 // PhaseIdealLoop is expensive so we only try it once we are
2563 // out of live nodes and we only try it again if the previous
2564 // helped got the number of nodes down significantly
2565 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2566 if (failing()) return;
2567 low_live_nodes = live_nodes();
2568 _major_progress = true;
2569 }
2570
2571 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2572 bool do_print_inlining = print_inlining() || print_intrinsics();
2573 if (do_print_inlining || log() != nullptr) {
2574 // Print inlining message for candidates that we couldn't inline for lack of space.
2575 for (int i = 0; i < _late_inlines.length(); i++) {
2576 CallGenerator* cg = _late_inlines.at(i);
2577 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2578 if (do_print_inlining) {
2579 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2580 }
2581 log_late_inline_failure(cg, msg);
2582 }
2583 }
2584 break; // finish
2585 }
2586 }
2587
2588 igvn_worklist()->ensure_empty(); // should be done with igvn
2589
2590 while (inline_incrementally_one()) {
2591 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2592 }
2593 if (failing()) return;
2594
2595 inline_incrementally_cleanup(igvn);
2596
2597 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2598
2599 if (failing()) return;
2600
2601 if (_late_inlines.length() == 0) {
2602 break; // no more progress
2603 }
2604 }
2605
2606 igvn_worklist()->ensure_empty(); // should be done with igvn
2607
2608 if (_string_late_inlines.length() > 0) {
2609 assert(has_stringbuilder(), "inconsistent");
2610
2611 inline_string_calls(false);
2612
2613 if (failing()) return;
2614
2615 inline_incrementally_cleanup(igvn);
2616 }
2617
2618 set_inlining_incrementally(false);
2619 }
2620
2621 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2622 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2623 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2624 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2625 // as if "inlining_incrementally() == true" were set.
2626 assert(inlining_incrementally() == false, "not allowed");
2627 #ifdef ASSERT
2628 Unique_Node_List* modified_nodes = _modified_nodes;
2629 _modified_nodes = nullptr;
2630 #endif
2631 assert(_late_inlines.length() > 0, "sanity");
2632
2633 while (_late_inlines.length() > 0) {
2634 igvn_worklist()->ensure_empty(); // should be done with igvn
2635
2636 while (inline_incrementally_one()) {
2637 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2638 }
2639 if (failing()) return;
2640
2641 inline_incrementally_cleanup(igvn);
2642 }
2643 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2644 }
2645
2646 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2647 if (_loop_opts_cnt > 0) {
2648 while (major_progress() && (_loop_opts_cnt > 0)) {
2649 TracePhase tp(_t_idealLoop);
2650 PhaseIdealLoop::optimize(igvn, mode);
2651 _loop_opts_cnt--;
2652 if (failing()) return false;
2653 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2654 }
2655 }
2656 return true;
2657 }
2658
2659 // Remove edges from "root" to each SafePoint at a backward branch.
2660 // They were inserted during parsing (see add_safepoint()) to make
2661 // infinite loops without calls or exceptions visible to root, i.e.,
2662 // useful.
2663 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2664 Node *r = root();
2665 if (r != nullptr) {
2666 for (uint i = r->req(); i < r->len(); ++i) {
2667 Node *n = r->in(i);
2668 if (n != nullptr && n->is_SafePoint()) {
2669 r->rm_prec(i);
2670 if (n->outcnt() == 0) {
2671 igvn.remove_dead_node(n);
2672 }
2673 --i;
2674 }
2675 }
2676 // Parsing may have added top inputs to the root node (Path
2677 // leading to the Halt node proven dead). Make sure we get a
2678 // chance to clean them up.
2679 igvn._worklist.push(r);
2680 igvn.optimize();
2681 }
2682 }
2683
2684 //------------------------------Optimize---------------------------------------
2685 // Given a graph, optimize it.
2686 void Compile::Optimize() {
2687 TracePhase tp(_t_optimizer);
2688
2689 #ifndef PRODUCT
2690 if (env()->break_at_compile()) {
2691 BREAKPOINT;
2692 }
2693
2694 #endif
2695
2696 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2697 #ifdef ASSERT
2698 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2699 #endif
2700
2701 ResourceMark rm;
2702
2703 NOT_PRODUCT( verify_graph_edges(); )
2704
2705 print_method(PHASE_AFTER_PARSING, 1);
2706
2707 {
2708 // Iterative Global Value Numbering, including ideal transforms
2709 // Initialize IterGVN with types and values from parse-time GVN
2710 PhaseIterGVN igvn(initial_gvn());
2711 #ifdef ASSERT
2712 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2713 #endif
2714 {
2715 TracePhase tp(_t_iterGVN);
2716 igvn.optimize();
2717 }
2718
2719 if (failing()) return;
2720
2721 print_method(PHASE_ITER_GVN1, 2);
2722
2723 process_for_unstable_if_traps(igvn);
2724
2725 if (failing()) return;
2726
2727 inline_incrementally(igvn);
2728
2729 print_method(PHASE_INCREMENTAL_INLINE, 2);
2730
2731 if (failing()) return;
2732
2733 if (eliminate_boxing()) {
2734 // Inline valueOf() methods now.
2735 inline_boxing_calls(igvn);
2736
2737 if (failing()) return;
2738
2739 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2740 inline_incrementally(igvn);
2741 }
2742
2743 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2744
2745 if (failing()) return;
2746 }
2747
2748 // Remove the speculative part of types and clean up the graph from
2749 // the extra CastPP nodes whose only purpose is to carry them. Do
2750 // that early so that optimizations are not disrupted by the extra
2751 // CastPP nodes.
2752 remove_speculative_types(igvn);
2753
2754 if (failing()) return;
2755
2756 // No more new expensive nodes will be added to the list from here
2757 // so keep only the actual candidates for optimizations.
2758 cleanup_expensive_nodes(igvn);
2759
2760 if (failing()) return;
2761
2762 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2763 if (EnableVectorSupport && has_vbox_nodes()) {
2764 TracePhase tp(_t_vector);
2765 PhaseVector pv(igvn);
2766 pv.optimize_vector_boxes();
2767 if (failing()) return;
2768 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2769 }
2770 assert(!has_vbox_nodes(), "sanity");
2771
2772 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2773 Compile::TracePhase tp(_t_renumberLive);
2774 igvn_worklist()->ensure_empty(); // should be done with igvn
2775 {
2776 ResourceMark rm;
2777 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2778 }
2779 igvn.reset_from_gvn(initial_gvn());
2780 igvn.optimize();
2781 if (failing()) return;
2782 }
2783
2784 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2785 // safepoints
2786 remove_root_to_sfpts_edges(igvn);
2787
2788 // Process inline type nodes now that all inlining is over
2789 process_inline_types(igvn);
2790
2791 adjust_flat_array_access_aliases(igvn);
2792
2793 if (failing()) return;
2794
2795 // Perform escape analysis
2796 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2797 if (has_loops()) {
2798 // Cleanup graph (remove dead nodes).
2799 TracePhase tp(_t_idealLoop);
2800 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2801 if (failing()) return;
2802 }
2803 bool progress;
2804 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2805 do {
2806 ConnectionGraph::do_analysis(this, &igvn);
2807
2808 if (failing()) return;
2809
2810 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2811
2812 // Optimize out fields loads from scalar replaceable allocations.
2813 igvn.optimize();
2814 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2815
2816 if (failing()) return;
2817
2818 if (congraph() != nullptr && macro_count() > 0) {
2819 TracePhase tp(_t_macroEliminate);
2820 PhaseMacroExpand mexp(igvn);
2821 mexp.eliminate_macro_nodes();
2822 if (failing()) return;
2823
2824 igvn.set_delay_transform(false);
2825 igvn.optimize();
2826 if (failing()) return;
2827
2828 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2829 }
2830
2831 ConnectionGraph::verify_ram_nodes(this, root());
2832 if (failing()) return;
2833
2834 progress = do_iterative_escape_analysis() &&
2835 (macro_count() < mcount) &&
2836 ConnectionGraph::has_candidates(this);
2837 // Try again if candidates exist and made progress
2838 // by removing some allocations and/or locks.
2839 } while (progress);
2840 }
2841
2842 // Loop transforms on the ideal graph. Range Check Elimination,
2843 // peeling, unrolling, etc.
2844
2845 // Set loop opts counter
2846 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2847 {
2848 TracePhase tp(_t_idealLoop);
2849 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2850 _loop_opts_cnt--;
2851 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2852 if (failing()) return;
2853 }
2854 // Loop opts pass if partial peeling occurred in previous pass
2855 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2856 TracePhase tp(_t_idealLoop);
2857 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2858 _loop_opts_cnt--;
2859 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2860 if (failing()) return;
2861 }
2862 // Loop opts pass for loop-unrolling before CCP
2863 if(major_progress() && (_loop_opts_cnt > 0)) {
2864 TracePhase tp(_t_idealLoop);
2865 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2866 _loop_opts_cnt--;
2867 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2868 }
2869 if (!failing()) {
2870 // Verify that last round of loop opts produced a valid graph
2871 PhaseIdealLoop::verify(igvn);
2872 }
2873 }
2874 if (failing()) return;
2875
2876 // Conditional Constant Propagation;
2877 print_method(PHASE_BEFORE_CCP1, 2);
2878 PhaseCCP ccp( &igvn );
2879 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2880 {
2881 TracePhase tp(_t_ccp);
2882 ccp.do_transform();
2883 }
2884 print_method(PHASE_CCP1, 2);
2885
2886 assert( true, "Break here to ccp.dump_old2new_map()");
2887
2888 // Iterative Global Value Numbering, including ideal transforms
2889 {
2890 TracePhase tp(_t_iterGVN2);
2891 igvn.reset_from_igvn(&ccp);
2892 igvn.optimize();
2893 }
2894 print_method(PHASE_ITER_GVN2, 2);
2895
2896 if (failing()) return;
2897
2898 // Loop transforms on the ideal graph. Range Check Elimination,
2899 // peeling, unrolling, etc.
2900 if (!optimize_loops(igvn, LoopOptsDefault)) {
2901 return;
2902 }
2903
2904 if (failing()) return;
2905
2906 C->clear_major_progress(); // ensure that major progress is now clear
2907
2908 process_for_post_loop_opts_igvn(igvn);
2909
2910 if (failing()) return;
2911
2912 #ifdef ASSERT
2913 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2914 #endif
2915
2916 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2917
2918 if (_late_inlines.length() > 0) {
2919 // More opportunities to optimize virtual and MH calls.
2920 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2921 process_late_inline_calls_no_inline(igvn);
2922 }
2923
2924 {
2925 TracePhase tp(_t_macroExpand);
2926 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2927 PhaseMacroExpand mex(igvn);
2928 if (mex.expand_macro_nodes()) {
2929 assert(failing(), "must bail out w/ explicit message");
2930 return;
2931 }
2932 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2933 }
2934
2935 // Process inline type nodes again and remove them. From here
2936 // on we don't need to keep track of field values anymore.
2937 process_inline_types(igvn, /* remove= */ true);
2938
2939 {
2940 TracePhase tp(_t_barrierExpand);
2941 if (bs->expand_barriers(this, igvn)) {
2942 assert(failing(), "must bail out w/ explicit message");
2943 return;
2944 }
2945 print_method(PHASE_BARRIER_EXPANSION, 2);
2946 }
2947
2948 if (C->max_vector_size() > 0) {
2949 C->optimize_logic_cones(igvn);
2950 igvn.optimize();
2951 if (failing()) return;
2952 }
2953
2954 DEBUG_ONLY( _modified_nodes = nullptr; )
2955 DEBUG_ONLY( _late_inlines.clear(); )
2956
2957 assert(igvn._worklist.size() == 0, "not empty");
2958 } // (End scope of igvn; run destructor if necessary for asserts.)
2959
2960 check_no_dead_use();
2961
2962 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2963 // to remove hashes to unlock nodes for modifications.
2964 C->node_hash()->clear();
2965
2966 // A method with only infinite loops has no edges entering loops from root
2967 {
2968 TracePhase tp(_t_graphReshaping);
2969 if (final_graph_reshaping()) {
2970 assert(failing(), "must bail out w/ explicit message");
2971 return;
2972 }
2973 }
2974
2975 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2976 DEBUG_ONLY(set_phase_optimize_finished();)
2977 }
2978
2979 #ifdef ASSERT
2980 void Compile::check_no_dead_use() const {
2981 ResourceMark rm;
2982 Unique_Node_List wq;
2983 wq.push(root());
2984 for (uint i = 0; i < wq.size(); ++i) {
2985 Node* n = wq.at(i);
2986 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2987 Node* u = n->fast_out(j);
2988 if (u->outcnt() == 0 && !u->is_Con()) {
2989 u->dump();
2990 fatal("no reachable node should have no use");
2991 }
2992 wq.push(u);
2993 }
2994 }
2995 }
2996 #endif
2997
2998 void Compile::inline_vector_reboxing_calls() {
2999 if (C->_vector_reboxing_late_inlines.length() > 0) {
3000 _late_inlines_pos = C->_late_inlines.length();
3001 while (_vector_reboxing_late_inlines.length() > 0) {
3002 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
3003 cg->do_late_inline();
3004 if (failing()) return;
3005 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
3006 }
3007 _vector_reboxing_late_inlines.trunc_to(0);
3008 }
3009 }
3010
3011 bool Compile::has_vbox_nodes() {
3012 if (C->_vector_reboxing_late_inlines.length() > 0) {
3013 return true;
3014 }
3015 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3016 Node * n = C->macro_node(macro_idx);
3017 assert(n->is_macro(), "only macro nodes expected here");
3018 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3019 return true;
3020 }
3021 }
3022 return false;
3023 }
3024
3025 //---------------------------- Bitwise operation packing optimization ---------------------------
3026
3027 static bool is_vector_unary_bitwise_op(Node* n) {
3028 return n->Opcode() == Op_XorV &&
3029 VectorNode::is_vector_bitwise_not_pattern(n);
3030 }
3031
3032 static bool is_vector_binary_bitwise_op(Node* n) {
3033 switch (n->Opcode()) {
3034 case Op_AndV:
3035 case Op_OrV:
3036 return true;
3037
3038 case Op_XorV:
3039 return !is_vector_unary_bitwise_op(n);
3040
3041 default:
3042 return false;
3043 }
3044 }
3045
3046 static bool is_vector_ternary_bitwise_op(Node* n) {
3047 return n->Opcode() == Op_MacroLogicV;
3048 }
3049
3050 static bool is_vector_bitwise_op(Node* n) {
3051 return is_vector_unary_bitwise_op(n) ||
3052 is_vector_binary_bitwise_op(n) ||
3053 is_vector_ternary_bitwise_op(n);
3054 }
3055
3056 static bool is_vector_bitwise_cone_root(Node* n) {
3057 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
3058 return false;
3059 }
3060 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3061 if (is_vector_bitwise_op(n->fast_out(i))) {
3062 return false;
3063 }
3064 }
3065 return true;
3066 }
3067
3068 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3069 uint cnt = 0;
3070 if (is_vector_bitwise_op(n)) {
3071 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3072 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3073 for (uint i = 1; i < inp_cnt; i++) {
3074 Node* in = n->in(i);
3075 bool skip = VectorNode::is_all_ones_vector(in);
3076 if (!skip && !inputs.member(in)) {
3077 inputs.push(in);
3078 cnt++;
3079 }
3080 }
3081 assert(cnt <= 1, "not unary");
3082 } else {
3083 uint last_req = inp_cnt;
3084 if (is_vector_ternary_bitwise_op(n)) {
3085 last_req = inp_cnt - 1; // skip last input
3086 }
3087 for (uint i = 1; i < last_req; i++) {
3088 Node* def = n->in(i);
3089 if (!inputs.member(def)) {
3090 inputs.push(def);
3091 cnt++;
3092 }
3093 }
3094 }
3095 } else { // not a bitwise operations
3096 if (!inputs.member(n)) {
3097 inputs.push(n);
3098 cnt++;
3099 }
3100 }
3101 return cnt;
3102 }
3103
3104 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3105 Unique_Node_List useful_nodes;
3106 C->identify_useful_nodes(useful_nodes);
3107
3108 for (uint i = 0; i < useful_nodes.size(); i++) {
3109 Node* n = useful_nodes.at(i);
3110 if (is_vector_bitwise_cone_root(n)) {
3111 list.push(n);
3112 }
3113 }
3114 }
3115
3116 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3117 const TypeVect* vt,
3118 Unique_Node_List& partition,
3119 Unique_Node_List& inputs) {
3120 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3121 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3122 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3123
3124 Node* in1 = inputs.at(0);
3125 Node* in2 = inputs.at(1);
3126 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3127
3128 uint func = compute_truth_table(partition, inputs);
3129
3130 Node* pn = partition.at(partition.size() - 1);
3131 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3132 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3133 }
3134
3135 static uint extract_bit(uint func, uint pos) {
3136 return (func & (1 << pos)) >> pos;
3137 }
3138
3139 //
3140 // A macro logic node represents a truth table. It has 4 inputs,
3141 // First three inputs corresponds to 3 columns of a truth table
3142 // and fourth input captures the logic function.
3143 //
3144 // eg. fn = (in1 AND in2) OR in3;
3145 //
3146 // MacroNode(in1,in2,in3,fn)
3147 //
3148 // -----------------
3149 // in1 in2 in3 fn
3150 // -----------------
3151 // 0 0 0 0
3152 // 0 0 1 1
3153 // 0 1 0 0
3154 // 0 1 1 1
3155 // 1 0 0 0
3156 // 1 0 1 1
3157 // 1 1 0 1
3158 // 1 1 1 1
3159 //
3160
3161 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3162 int res = 0;
3163 for (int i = 0; i < 8; i++) {
3164 int bit1 = extract_bit(in1, i);
3165 int bit2 = extract_bit(in2, i);
3166 int bit3 = extract_bit(in3, i);
3167
3168 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3169 int func_bit = extract_bit(func, func_bit_pos);
3170
3171 res |= func_bit << i;
3172 }
3173 return res;
3174 }
3175
3176 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
3177 assert(n != nullptr, "");
3178 assert(eval_map.contains(n), "absent");
3179 return *(eval_map.get(n));
3180 }
3181
3182 static void eval_operands(Node* n,
3183 uint& func1, uint& func2, uint& func3,
3184 ResourceHashtable<Node*,uint>& eval_map) {
3185 assert(is_vector_bitwise_op(n), "");
3186
3187 if (is_vector_unary_bitwise_op(n)) {
3188 Node* opnd = n->in(1);
3189 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3190 opnd = n->in(2);
3191 }
3192 func1 = eval_operand(opnd, eval_map);
3193 } else if (is_vector_binary_bitwise_op(n)) {
3194 func1 = eval_operand(n->in(1), eval_map);
3195 func2 = eval_operand(n->in(2), eval_map);
3196 } else {
3197 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3198 func1 = eval_operand(n->in(1), eval_map);
3199 func2 = eval_operand(n->in(2), eval_map);
3200 func3 = eval_operand(n->in(3), eval_map);
3201 }
3202 }
3203
3204 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3205 assert(inputs.size() <= 3, "sanity");
3206 ResourceMark rm;
3207 uint res = 0;
3208 ResourceHashtable<Node*,uint> eval_map;
3209
3210 // Populate precomputed functions for inputs.
3211 // Each input corresponds to one column of 3 input truth-table.
3212 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3213 0xCC, // (_, b, _) -> b
3214 0xF0 }; // (a, _, _) -> a
3215 for (uint i = 0; i < inputs.size(); i++) {
3216 eval_map.put(inputs.at(i), input_funcs[2-i]);
3217 }
3218
3219 for (uint i = 0; i < partition.size(); i++) {
3220 Node* n = partition.at(i);
3221
3222 uint func1 = 0, func2 = 0, func3 = 0;
3223 eval_operands(n, func1, func2, func3, eval_map);
3224
3225 switch (n->Opcode()) {
3226 case Op_OrV:
3227 assert(func3 == 0, "not binary");
3228 res = func1 | func2;
3229 break;
3230 case Op_AndV:
3231 assert(func3 == 0, "not binary");
3232 res = func1 & func2;
3233 break;
3234 case Op_XorV:
3235 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3236 assert(func2 == 0 && func3 == 0, "not unary");
3237 res = (~func1) & 0xFF;
3238 } else {
3239 assert(func3 == 0, "not binary");
3240 res = func1 ^ func2;
3241 }
3242 break;
3243 case Op_MacroLogicV:
3244 // Ordering of inputs may change during evaluation of sub-tree
3245 // containing MacroLogic node as a child node, thus a re-evaluation
3246 // makes sure that function is evaluated in context of current
3247 // inputs.
3248 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3249 break;
3250
3251 default: assert(false, "not supported: %s", n->Name());
3252 }
3253 assert(res <= 0xFF, "invalid");
3254 eval_map.put(n, res);
3255 }
3256 return res;
3257 }
3258
3259 // Criteria under which nodes gets packed into a macro logic node:-
3260 // 1) Parent and both child nodes are all unmasked or masked with
3261 // same predicates.
3262 // 2) Masked parent can be packed with left child if it is predicated
3263 // and both have same predicates.
3264 // 3) Masked parent can be packed with right child if its un-predicated
3265 // or has matching predication condition.
3266 // 4) An unmasked parent can be packed with an unmasked child.
3267 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3268 assert(partition.size() == 0, "not empty");
3269 assert(inputs.size() == 0, "not empty");
3270 if (is_vector_ternary_bitwise_op(n)) {
3271 return false;
3272 }
3273
3274 bool is_unary_op = is_vector_unary_bitwise_op(n);
3275 if (is_unary_op) {
3276 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3277 return false; // too few inputs
3278 }
3279
3280 bool pack_left_child = true;
3281 bool pack_right_child = true;
3282
3283 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3284 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3285
3286 int left_child_input_cnt = 0;
3287 int right_child_input_cnt = 0;
3288
3289 bool parent_is_predicated = n->is_predicated_vector();
3290 bool left_child_predicated = n->in(1)->is_predicated_vector();
3291 bool right_child_predicated = n->in(2)->is_predicated_vector();
3292
3293 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
3294 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3295 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3296
3297 do {
3298 if (pack_left_child && left_child_LOP &&
3299 ((!parent_is_predicated && !left_child_predicated) ||
3300 ((parent_is_predicated && left_child_predicated &&
3301 parent_pred == left_child_pred)))) {
3302 partition.push(n->in(1));
3303 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3304 } else {
3305 inputs.push(n->in(1));
3306 left_child_input_cnt = 1;
3307 }
3308
3309 if (pack_right_child && right_child_LOP &&
3310 (!right_child_predicated ||
3311 (right_child_predicated && parent_is_predicated &&
3312 parent_pred == right_child_pred))) {
3313 partition.push(n->in(2));
3314 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3315 } else {
3316 inputs.push(n->in(2));
3317 right_child_input_cnt = 1;
3318 }
3319
3320 if (inputs.size() > 3) {
3321 assert(partition.size() > 0, "");
3322 inputs.clear();
3323 partition.clear();
3324 if (left_child_input_cnt > right_child_input_cnt) {
3325 pack_left_child = false;
3326 } else {
3327 pack_right_child = false;
3328 }
3329 } else {
3330 break;
3331 }
3332 } while(true);
3333
3334 if(partition.size()) {
3335 partition.push(n);
3336 }
3337
3338 return (partition.size() == 2 || partition.size() == 3) &&
3339 (inputs.size() == 2 || inputs.size() == 3);
3340 }
3341
3342 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3343 assert(is_vector_bitwise_op(n), "not a root");
3344
3345 visited.set(n->_idx);
3346
3347 // 1) Do a DFS walk over the logic cone.
3348 for (uint i = 1; i < n->req(); i++) {
3349 Node* in = n->in(i);
3350 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3351 process_logic_cone_root(igvn, in, visited);
3352 }
3353 }
3354
3355 // 2) Bottom up traversal: Merge node[s] with
3356 // the parent to form macro logic node.
3357 Unique_Node_List partition;
3358 Unique_Node_List inputs;
3359 if (compute_logic_cone(n, partition, inputs)) {
3360 const TypeVect* vt = n->bottom_type()->is_vect();
3361 Node* pn = partition.at(partition.size() - 1);
3362 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3363 if (mask == nullptr ||
3364 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3365 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3366 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3367 igvn.replace_node(n, macro_logic);
3368 }
3369 }
3370 }
3371
3372 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3373 ResourceMark rm;
3374 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3375 Unique_Node_List list;
3376 collect_logic_cone_roots(list);
3377
3378 while (list.size() > 0) {
3379 Node* n = list.pop();
3380 const TypeVect* vt = n->bottom_type()->is_vect();
3381 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3382 if (supported) {
3383 VectorSet visited(comp_arena());
3384 process_logic_cone_root(igvn, n, visited);
3385 }
3386 }
3387 }
3388 }
3389
3390 //------------------------------Code_Gen---------------------------------------
3391 // Given a graph, generate code for it
3392 void Compile::Code_Gen() {
3393 if (failing()) {
3394 return;
3395 }
3396
3397 // Perform instruction selection. You might think we could reclaim Matcher
3398 // memory PDQ, but actually the Matcher is used in generating spill code.
3399 // Internals of the Matcher (including some VectorSets) must remain live
3400 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3401 // set a bit in reclaimed memory.
3402
3403 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3404 // nodes. Mapping is only valid at the root of each matched subtree.
3405 NOT_PRODUCT( verify_graph_edges(); )
3406
3407 Matcher matcher;
3408 _matcher = &matcher;
3409 {
3410 TracePhase tp(_t_matcher);
3411 matcher.match();
3412 if (failing()) {
3413 return;
3414 }
3415 }
3416 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3417 // nodes. Mapping is only valid at the root of each matched subtree.
3418 NOT_PRODUCT( verify_graph_edges(); )
3419
3420 // If you have too many nodes, or if matching has failed, bail out
3421 check_node_count(0, "out of nodes matching instructions");
3422 if (failing()) {
3423 return;
3424 }
3425
3426 print_method(PHASE_MATCHING, 2);
3427
3428 // Build a proper-looking CFG
3429 PhaseCFG cfg(node_arena(), root(), matcher);
3430 if (failing()) {
3431 return;
3432 }
3433 _cfg = &cfg;
3434 {
3435 TracePhase tp(_t_scheduler);
3436 bool success = cfg.do_global_code_motion();
3437 if (!success) {
3438 return;
3439 }
3440
3441 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3442 NOT_PRODUCT( verify_graph_edges(); )
3443 cfg.verify();
3444 if (failing()) {
3445 return;
3446 }
3447 }
3448
3449 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3450 _regalloc = ®alloc;
3451 {
3452 TracePhase tp(_t_registerAllocation);
3453 // Perform register allocation. After Chaitin, use-def chains are
3454 // no longer accurate (at spill code) and so must be ignored.
3455 // Node->LRG->reg mappings are still accurate.
3456 _regalloc->Register_Allocate();
3457
3458 // Bail out if the allocator builds too many nodes
3459 if (failing()) {
3460 return;
3461 }
3462
3463 print_method(PHASE_REGISTER_ALLOCATION, 2);
3464 }
3465
3466 // Prior to register allocation we kept empty basic blocks in case the
3467 // the allocator needed a place to spill. After register allocation we
3468 // are not adding any new instructions. If any basic block is empty, we
3469 // can now safely remove it.
3470 {
3471 TracePhase tp(_t_blockOrdering);
3472 cfg.remove_empty_blocks();
3473 if (do_freq_based_layout()) {
3474 PhaseBlockLayout layout(cfg);
3475 } else {
3476 cfg.set_loop_alignment();
3477 }
3478 cfg.fixup_flow();
3479 cfg.remove_unreachable_blocks();
3480 cfg.verify_dominator_tree();
3481 print_method(PHASE_BLOCK_ORDERING, 3);
3482 }
3483
3484 // Apply peephole optimizations
3485 if( OptoPeephole ) {
3486 TracePhase tp(_t_peephole);
3487 PhasePeephole peep( _regalloc, cfg);
3488 peep.do_transform();
3489 print_method(PHASE_PEEPHOLE, 3);
3490 }
3491
3492 // Do late expand if CPU requires this.
3493 if (Matcher::require_postalloc_expand) {
3494 TracePhase tp(_t_postalloc_expand);
3495 cfg.postalloc_expand(_regalloc);
3496 print_method(PHASE_POSTALLOC_EXPAND, 3);
3497 }
3498
3499 #ifdef ASSERT
3500 {
3501 CompilationMemoryStatistic::do_test_allocations();
3502 if (failing()) return;
3503 }
3504 #endif
3505
3506 // Convert Nodes to instruction bits in a buffer
3507 {
3508 TracePhase tp(_t_output);
3509 PhaseOutput output;
3510 output.Output();
3511 if (failing()) return;
3512 output.install();
3513 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3514 }
3515
3516 // He's dead, Jim.
3517 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3518 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3519 }
3520
3521 //------------------------------Final_Reshape_Counts---------------------------
3522 // This class defines counters to help identify when a method
3523 // may/must be executed using hardware with only 24-bit precision.
3524 struct Final_Reshape_Counts : public StackObj {
3525 int _call_count; // count non-inlined 'common' calls
3526 int _float_count; // count float ops requiring 24-bit precision
3527 int _double_count; // count double ops requiring more precision
3528 int _java_call_count; // count non-inlined 'java' calls
3529 int _inner_loop_count; // count loops which need alignment
3530 VectorSet _visited; // Visitation flags
3531 Node_List _tests; // Set of IfNodes & PCTableNodes
3532
3533 Final_Reshape_Counts() :
3534 _call_count(0), _float_count(0), _double_count(0),
3535 _java_call_count(0), _inner_loop_count(0) { }
3536
3537 void inc_call_count () { _call_count ++; }
3538 void inc_float_count () { _float_count ++; }
3539 void inc_double_count() { _double_count++; }
3540 void inc_java_call_count() { _java_call_count++; }
3541 void inc_inner_loop_count() { _inner_loop_count++; }
3542
3543 int get_call_count () const { return _call_count ; }
3544 int get_float_count () const { return _float_count ; }
3545 int get_double_count() const { return _double_count; }
3546 int get_java_call_count() const { return _java_call_count; }
3547 int get_inner_loop_count() const { return _inner_loop_count; }
3548 };
3549
3550 //------------------------------final_graph_reshaping_impl----------------------
3551 // Implement items 1-5 from final_graph_reshaping below.
3552 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3553
3554 if ( n->outcnt() == 0 ) return; // dead node
3555 uint nop = n->Opcode();
3556
3557 // Check for 2-input instruction with "last use" on right input.
3558 // Swap to left input. Implements item (2).
3559 if( n->req() == 3 && // two-input instruction
3560 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3561 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3562 n->in(2)->outcnt() == 1 &&// right use IS a last use
3563 !n->in(2)->is_Con() ) { // right use is not a constant
3564 // Check for commutative opcode
3565 switch( nop ) {
3566 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3567 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3568 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3569 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3570 case Op_AndL: case Op_XorL: case Op_OrL:
3571 case Op_AndI: case Op_XorI: case Op_OrI: {
3572 // Move "last use" input to left by swapping inputs
3573 n->swap_edges(1, 2);
3574 break;
3575 }
3576 default:
3577 break;
3578 }
3579 }
3580
3581 #ifdef ASSERT
3582 if( n->is_Mem() ) {
3583 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3584 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3585 // oop will be recorded in oop map if load crosses safepoint
3586 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3587 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3588 "raw memory operations should have control edge");
3589 }
3590 if (n->is_MemBar()) {
3591 MemBarNode* mb = n->as_MemBar();
3592 if (mb->trailing_store() || mb->trailing_load_store()) {
3593 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3594 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3595 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3596 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3597 } else if (mb->leading()) {
3598 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3599 }
3600 }
3601 #endif
3602 // Count FPU ops and common calls, implements item (3)
3603 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3604 if (!gc_handled) {
3605 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3606 }
3607
3608 // Collect CFG split points
3609 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3610 frc._tests.push(n);
3611 }
3612 }
3613
3614 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3615 if (!UseDivMod) {
3616 return;
3617 }
3618
3619 // Check if "a % b" and "a / b" both exist
3620 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3621 if (d == nullptr) {
3622 return;
3623 }
3624
3625 // Replace them with a fused divmod if supported
3626 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3627 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3628 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3629 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3630 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3631 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3632 // DivMod node so the dependency is not lost.
3633 divmod->add_prec_from(n);
3634 divmod->add_prec_from(d);
3635 d->subsume_by(divmod->div_proj(), this);
3636 n->subsume_by(divmod->mod_proj(), this);
3637 } else {
3638 // Replace "a % b" with "a - ((a / b) * b)"
3639 Node* mult = MulNode::make(d, d->in(2), bt);
3640 Node* sub = SubNode::make(d->in(1), mult, bt);
3641 n->subsume_by(sub, this);
3642 }
3643 }
3644
3645 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3646 switch( nop ) {
3647 // Count all float operations that may use FPU
3648 case Op_AddF:
3649 case Op_SubF:
3650 case Op_MulF:
3651 case Op_DivF:
3652 case Op_NegF:
3653 case Op_ModF:
3654 case Op_ConvI2F:
3655 case Op_ConF:
3656 case Op_CmpF:
3657 case Op_CmpF3:
3658 case Op_StoreF:
3659 case Op_LoadF:
3660 // case Op_ConvL2F: // longs are split into 32-bit halves
3661 frc.inc_float_count();
3662 break;
3663
3664 case Op_ConvF2D:
3665 case Op_ConvD2F:
3666 frc.inc_float_count();
3667 frc.inc_double_count();
3668 break;
3669
3670 // Count all double operations that may use FPU
3671 case Op_AddD:
3672 case Op_SubD:
3673 case Op_MulD:
3674 case Op_DivD:
3675 case Op_NegD:
3676 case Op_ModD:
3677 case Op_ConvI2D:
3678 case Op_ConvD2I:
3679 // case Op_ConvL2D: // handled by leaf call
3680 // case Op_ConvD2L: // handled by leaf call
3681 case Op_ConD:
3682 case Op_CmpD:
3683 case Op_CmpD3:
3684 case Op_StoreD:
3685 case Op_LoadD:
3686 case Op_LoadD_unaligned:
3687 frc.inc_double_count();
3688 break;
3689 case Op_Opaque1: // Remove Opaque Nodes before matching
3690 n->subsume_by(n->in(1), this);
3691 break;
3692 case Op_CallStaticJava:
3693 case Op_CallJava:
3694 case Op_CallDynamicJava:
3695 frc.inc_java_call_count(); // Count java call site;
3696 case Op_CallRuntime:
3697 case Op_CallLeaf:
3698 case Op_CallLeafVector:
3699 case Op_CallLeafNoFP: {
3700 assert (n->is_Call(), "");
3701 CallNode *call = n->as_Call();
3702 // Count call sites where the FP mode bit would have to be flipped.
3703 // Do not count uncommon runtime calls:
3704 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3705 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3706 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3707 frc.inc_call_count(); // Count the call site
3708 } else { // See if uncommon argument is shared
3709 Node *n = call->in(TypeFunc::Parms);
3710 int nop = n->Opcode();
3711 // Clone shared simple arguments to uncommon calls, item (1).
3712 if (n->outcnt() > 1 &&
3713 !n->is_Proj() &&
3714 nop != Op_CreateEx &&
3715 nop != Op_CheckCastPP &&
3716 nop != Op_DecodeN &&
3717 nop != Op_DecodeNKlass &&
3718 !n->is_Mem() &&
3719 !n->is_Phi()) {
3720 Node *x = n->clone();
3721 call->set_req(TypeFunc::Parms, x);
3722 }
3723 }
3724 break;
3725 }
3726 case Op_StoreB:
3727 case Op_StoreC:
3728 case Op_StoreI:
3729 case Op_StoreL:
3730 case Op_StoreLSpecial:
3731 case Op_CompareAndSwapB:
3732 case Op_CompareAndSwapS:
3733 case Op_CompareAndSwapI:
3734 case Op_CompareAndSwapL:
3735 case Op_CompareAndSwapP:
3736 case Op_CompareAndSwapN:
3737 case Op_WeakCompareAndSwapB:
3738 case Op_WeakCompareAndSwapS:
3739 case Op_WeakCompareAndSwapI:
3740 case Op_WeakCompareAndSwapL:
3741 case Op_WeakCompareAndSwapP:
3742 case Op_WeakCompareAndSwapN:
3743 case Op_CompareAndExchangeB:
3744 case Op_CompareAndExchangeS:
3745 case Op_CompareAndExchangeI:
3746 case Op_CompareAndExchangeL:
3747 case Op_CompareAndExchangeP:
3748 case Op_CompareAndExchangeN:
3749 case Op_GetAndAddS:
3750 case Op_GetAndAddB:
3751 case Op_GetAndAddI:
3752 case Op_GetAndAddL:
3753 case Op_GetAndSetS:
3754 case Op_GetAndSetB:
3755 case Op_GetAndSetI:
3756 case Op_GetAndSetL:
3757 case Op_GetAndSetP:
3758 case Op_GetAndSetN:
3759 case Op_StoreP:
3760 case Op_StoreN:
3761 case Op_StoreNKlass:
3762 case Op_LoadB:
3763 case Op_LoadUB:
3764 case Op_LoadUS:
3765 case Op_LoadI:
3766 case Op_LoadKlass:
3767 case Op_LoadNKlass:
3768 case Op_LoadL:
3769 case Op_LoadL_unaligned:
3770 case Op_LoadP:
3771 case Op_LoadN:
3772 case Op_LoadRange:
3773 case Op_LoadS:
3774 break;
3775
3776 case Op_AddP: { // Assert sane base pointers
3777 Node *addp = n->in(AddPNode::Address);
3778 assert( !addp->is_AddP() ||
3779 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3780 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3781 "Base pointers must match (addp %u)", addp->_idx );
3782 #ifdef _LP64
3783 if ((UseCompressedOops || UseCompressedClassPointers) &&
3784 addp->Opcode() == Op_ConP &&
3785 addp == n->in(AddPNode::Base) &&
3786 n->in(AddPNode::Offset)->is_Con()) {
3787 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3788 // on the platform and on the compressed oops mode.
3789 // Use addressing with narrow klass to load with offset on x86.
3790 // Some platforms can use the constant pool to load ConP.
3791 // Do this transformation here since IGVN will convert ConN back to ConP.
3792 const Type* t = addp->bottom_type();
3793 bool is_oop = t->isa_oopptr() != nullptr;
3794 bool is_klass = t->isa_klassptr() != nullptr;
3795
3796 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3797 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3798 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3799 Node* nn = nullptr;
3800
3801 int op = is_oop ? Op_ConN : Op_ConNKlass;
3802
3803 // Look for existing ConN node of the same exact type.
3804 Node* r = root();
3805 uint cnt = r->outcnt();
3806 for (uint i = 0; i < cnt; i++) {
3807 Node* m = r->raw_out(i);
3808 if (m!= nullptr && m->Opcode() == op &&
3809 m->bottom_type()->make_ptr() == t) {
3810 nn = m;
3811 break;
3812 }
3813 }
3814 if (nn != nullptr) {
3815 // Decode a narrow oop to match address
3816 // [R12 + narrow_oop_reg<<3 + offset]
3817 if (is_oop) {
3818 nn = new DecodeNNode(nn, t);
3819 } else {
3820 nn = new DecodeNKlassNode(nn, t);
3821 }
3822 // Check for succeeding AddP which uses the same Base.
3823 // Otherwise we will run into the assertion above when visiting that guy.
3824 for (uint i = 0; i < n->outcnt(); ++i) {
3825 Node *out_i = n->raw_out(i);
3826 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3827 out_i->set_req(AddPNode::Base, nn);
3828 #ifdef ASSERT
3829 for (uint j = 0; j < out_i->outcnt(); ++j) {
3830 Node *out_j = out_i->raw_out(j);
3831 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3832 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3833 }
3834 #endif
3835 }
3836 }
3837 n->set_req(AddPNode::Base, nn);
3838 n->set_req(AddPNode::Address, nn);
3839 if (addp->outcnt() == 0) {
3840 addp->disconnect_inputs(this);
3841 }
3842 }
3843 }
3844 }
3845 #endif
3846 break;
3847 }
3848
3849 case Op_CastPP: {
3850 // Remove CastPP nodes to gain more freedom during scheduling but
3851 // keep the dependency they encode as control or precedence edges
3852 // (if control is set already) on memory operations. Some CastPP
3853 // nodes don't have a control (don't carry a dependency): skip
3854 // those.
3855 if (n->in(0) != nullptr) {
3856 ResourceMark rm;
3857 Unique_Node_List wq;
3858 wq.push(n);
3859 for (uint next = 0; next < wq.size(); ++next) {
3860 Node *m = wq.at(next);
3861 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3862 Node* use = m->fast_out(i);
3863 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3864 use->ensure_control_or_add_prec(n->in(0));
3865 } else {
3866 switch(use->Opcode()) {
3867 case Op_AddP:
3868 case Op_DecodeN:
3869 case Op_DecodeNKlass:
3870 case Op_CheckCastPP:
3871 case Op_CastPP:
3872 wq.push(use);
3873 break;
3874 }
3875 }
3876 }
3877 }
3878 }
3879 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3880 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3881 Node* in1 = n->in(1);
3882 const Type* t = n->bottom_type();
3883 Node* new_in1 = in1->clone();
3884 new_in1->as_DecodeN()->set_type(t);
3885
3886 if (!Matcher::narrow_oop_use_complex_address()) {
3887 //
3888 // x86, ARM and friends can handle 2 adds in addressing mode
3889 // and Matcher can fold a DecodeN node into address by using
3890 // a narrow oop directly and do implicit null check in address:
3891 //
3892 // [R12 + narrow_oop_reg<<3 + offset]
3893 // NullCheck narrow_oop_reg
3894 //
3895 // On other platforms (Sparc) we have to keep new DecodeN node and
3896 // use it to do implicit null check in address:
3897 //
3898 // decode_not_null narrow_oop_reg, base_reg
3899 // [base_reg + offset]
3900 // NullCheck base_reg
3901 //
3902 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3903 // to keep the information to which null check the new DecodeN node
3904 // corresponds to use it as value in implicit_null_check().
3905 //
3906 new_in1->set_req(0, n->in(0));
3907 }
3908
3909 n->subsume_by(new_in1, this);
3910 if (in1->outcnt() == 0) {
3911 in1->disconnect_inputs(this);
3912 }
3913 } else {
3914 n->subsume_by(n->in(1), this);
3915 if (n->outcnt() == 0) {
3916 n->disconnect_inputs(this);
3917 }
3918 }
3919 break;
3920 }
3921 case Op_CastII: {
3922 n->as_CastII()->remove_range_check_cast(this);
3923 break;
3924 }
3925 #ifdef _LP64
3926 case Op_CmpP:
3927 // Do this transformation here to preserve CmpPNode::sub() and
3928 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3929 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3930 Node* in1 = n->in(1);
3931 Node* in2 = n->in(2);
3932 if (!in1->is_DecodeNarrowPtr()) {
3933 in2 = in1;
3934 in1 = n->in(2);
3935 }
3936 assert(in1->is_DecodeNarrowPtr(), "sanity");
3937
3938 Node* new_in2 = nullptr;
3939 if (in2->is_DecodeNarrowPtr()) {
3940 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3941 new_in2 = in2->in(1);
3942 } else if (in2->Opcode() == Op_ConP) {
3943 const Type* t = in2->bottom_type();
3944 if (t == TypePtr::NULL_PTR) {
3945 assert(in1->is_DecodeN(), "compare klass to null?");
3946 // Don't convert CmpP null check into CmpN if compressed
3947 // oops implicit null check is not generated.
3948 // This will allow to generate normal oop implicit null check.
3949 if (Matcher::gen_narrow_oop_implicit_null_checks())
3950 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3951 //
3952 // This transformation together with CastPP transformation above
3953 // will generated code for implicit null checks for compressed oops.
3954 //
3955 // The original code after Optimize()
3956 //
3957 // LoadN memory, narrow_oop_reg
3958 // decode narrow_oop_reg, base_reg
3959 // CmpP base_reg, nullptr
3960 // CastPP base_reg // NotNull
3961 // Load [base_reg + offset], val_reg
3962 //
3963 // after these transformations will be
3964 //
3965 // LoadN memory, narrow_oop_reg
3966 // CmpN narrow_oop_reg, nullptr
3967 // decode_not_null narrow_oop_reg, base_reg
3968 // Load [base_reg + offset], val_reg
3969 //
3970 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3971 // since narrow oops can be used in debug info now (see the code in
3972 // final_graph_reshaping_walk()).
3973 //
3974 // At the end the code will be matched to
3975 // on x86:
3976 //
3977 // Load_narrow_oop memory, narrow_oop_reg
3978 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3979 // NullCheck narrow_oop_reg
3980 //
3981 // and on sparc:
3982 //
3983 // Load_narrow_oop memory, narrow_oop_reg
3984 // decode_not_null narrow_oop_reg, base_reg
3985 // Load [base_reg + offset], val_reg
3986 // NullCheck base_reg
3987 //
3988 } else if (t->isa_oopptr()) {
3989 new_in2 = ConNode::make(t->make_narrowoop());
3990 } else if (t->isa_klassptr()) {
3991 new_in2 = ConNode::make(t->make_narrowklass());
3992 }
3993 }
3994 if (new_in2 != nullptr) {
3995 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3996 n->subsume_by(cmpN, this);
3997 if (in1->outcnt() == 0) {
3998 in1->disconnect_inputs(this);
3999 }
4000 if (in2->outcnt() == 0) {
4001 in2->disconnect_inputs(this);
4002 }
4003 }
4004 }
4005 break;
4006
4007 case Op_DecodeN:
4008 case Op_DecodeNKlass:
4009 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4010 // DecodeN could be pinned when it can't be fold into
4011 // an address expression, see the code for Op_CastPP above.
4012 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4013 break;
4014
4015 case Op_EncodeP:
4016 case Op_EncodePKlass: {
4017 Node* in1 = n->in(1);
4018 if (in1->is_DecodeNarrowPtr()) {
4019 n->subsume_by(in1->in(1), this);
4020 } else if (in1->Opcode() == Op_ConP) {
4021 const Type* t = in1->bottom_type();
4022 if (t == TypePtr::NULL_PTR) {
4023 assert(t->isa_oopptr(), "null klass?");
4024 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4025 } else if (t->isa_oopptr()) {
4026 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4027 } else if (t->isa_klassptr()) {
4028 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4029 }
4030 }
4031 if (in1->outcnt() == 0) {
4032 in1->disconnect_inputs(this);
4033 }
4034 break;
4035 }
4036
4037 case Op_Proj: {
4038 if (OptimizeStringConcat || IncrementalInline) {
4039 ProjNode* proj = n->as_Proj();
4040 if (proj->_is_io_use) {
4041 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4042 // Separate projections were used for the exception path which
4043 // are normally removed by a late inline. If it wasn't inlined
4044 // then they will hang around and should just be replaced with
4045 // the original one. Merge them.
4046 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4047 if (non_io_proj != nullptr) {
4048 proj->subsume_by(non_io_proj , this);
4049 }
4050 }
4051 }
4052 break;
4053 }
4054
4055 case Op_Phi:
4056 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4057 // The EncodeP optimization may create Phi with the same edges
4058 // for all paths. It is not handled well by Register Allocator.
4059 Node* unique_in = n->in(1);
4060 assert(unique_in != nullptr, "");
4061 uint cnt = n->req();
4062 for (uint i = 2; i < cnt; i++) {
4063 Node* m = n->in(i);
4064 assert(m != nullptr, "");
4065 if (unique_in != m)
4066 unique_in = nullptr;
4067 }
4068 if (unique_in != nullptr) {
4069 n->subsume_by(unique_in, this);
4070 }
4071 }
4072 break;
4073
4074 #endif
4075
4076 case Op_ModI:
4077 handle_div_mod_op(n, T_INT, false);
4078 break;
4079
4080 case Op_ModL:
4081 handle_div_mod_op(n, T_LONG, false);
4082 break;
4083
4084 case Op_UModI:
4085 handle_div_mod_op(n, T_INT, true);
4086 break;
4087
4088 case Op_UModL:
4089 handle_div_mod_op(n, T_LONG, true);
4090 break;
4091
4092 case Op_LoadVector:
4093 case Op_StoreVector:
4094 #ifdef ASSERT
4095 // Add VerifyVectorAlignment node between adr and load / store.
4096 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
4097 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
4098 n->as_StoreVector()->must_verify_alignment();
4099 if (must_verify_alignment) {
4100 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
4101 n->as_StoreVector()->memory_size();
4102 // The memory access should be aligned to the vector width in bytes.
4103 // However, the underlying array is possibly less well aligned, but at least
4104 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
4105 // a loop we can expect at least the following alignment:
4106 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
4107 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
4108 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
4109 // Create mask from alignment. e.g. 0b1000 -> 0b0111
4110 jlong mask = guaranteed_alignment - 1;
4111 Node* mask_con = ConLNode::make(mask);
4112 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
4113 n->set_req(MemNode::Address, va);
4114 }
4115 }
4116 #endif
4117 break;
4118
4119 case Op_LoadVectorGather:
4120 case Op_StoreVectorScatter:
4121 case Op_LoadVectorGatherMasked:
4122 case Op_StoreVectorScatterMasked:
4123 case Op_VectorCmpMasked:
4124 case Op_VectorMaskGen:
4125 case Op_LoadVectorMasked:
4126 case Op_StoreVectorMasked:
4127 break;
4128
4129 case Op_AddReductionVI:
4130 case Op_AddReductionVL:
4131 case Op_AddReductionVF:
4132 case Op_AddReductionVD:
4133 case Op_MulReductionVI:
4134 case Op_MulReductionVL:
4135 case Op_MulReductionVF:
4136 case Op_MulReductionVD:
4137 case Op_MinReductionV:
4138 case Op_MaxReductionV:
4139 case Op_AndReductionV:
4140 case Op_OrReductionV:
4141 case Op_XorReductionV:
4142 break;
4143
4144 case Op_PackB:
4145 case Op_PackS:
4146 case Op_PackI:
4147 case Op_PackF:
4148 case Op_PackL:
4149 case Op_PackD:
4150 if (n->req()-1 > 2) {
4151 // Replace many operand PackNodes with a binary tree for matching
4152 PackNode* p = (PackNode*) n;
4153 Node* btp = p->binary_tree_pack(1, n->req());
4154 n->subsume_by(btp, this);
4155 }
4156 break;
4157 case Op_Loop:
4158 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4159 case Op_CountedLoop:
4160 case Op_LongCountedLoop:
4161 case Op_OuterStripMinedLoop:
4162 if (n->as_Loop()->is_inner_loop()) {
4163 frc.inc_inner_loop_count();
4164 }
4165 n->as_Loop()->verify_strip_mined(0);
4166 break;
4167 case Op_LShiftI:
4168 case Op_RShiftI:
4169 case Op_URShiftI:
4170 case Op_LShiftL:
4171 case Op_RShiftL:
4172 case Op_URShiftL:
4173 if (Matcher::need_masked_shift_count) {
4174 // The cpu's shift instructions don't restrict the count to the
4175 // lower 5/6 bits. We need to do the masking ourselves.
4176 Node* in2 = n->in(2);
4177 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4178 const TypeInt* t = in2->find_int_type();
4179 if (t != nullptr && t->is_con()) {
4180 juint shift = t->get_con();
4181 if (shift > mask) { // Unsigned cmp
4182 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4183 }
4184 } else {
4185 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
4186 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4187 n->set_req(2, shift);
4188 }
4189 }
4190 if (in2->outcnt() == 0) { // Remove dead node
4191 in2->disconnect_inputs(this);
4192 }
4193 }
4194 break;
4195 case Op_MemBarStoreStore:
4196 case Op_MemBarRelease:
4197 // Break the link with AllocateNode: it is no longer useful and
4198 // confuses register allocation.
4199 if (n->req() > MemBarNode::Precedent) {
4200 n->set_req(MemBarNode::Precedent, top());
4201 }
4202 break;
4203 case Op_MemBarAcquire: {
4204 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4205 // At parse time, the trailing MemBarAcquire for a volatile load
4206 // is created with an edge to the load. After optimizations,
4207 // that input may be a chain of Phis. If those phis have no
4208 // other use, then the MemBarAcquire keeps them alive and
4209 // register allocation can be confused.
4210 dead_nodes.push(n->in(MemBarNode::Precedent));
4211 n->set_req(MemBarNode::Precedent, top());
4212 }
4213 break;
4214 }
4215 case Op_Blackhole:
4216 break;
4217 case Op_RangeCheck: {
4218 RangeCheckNode* rc = n->as_RangeCheck();
4219 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4220 n->subsume_by(iff, this);
4221 frc._tests.push(iff);
4222 break;
4223 }
4224 case Op_ConvI2L: {
4225 if (!Matcher::convi2l_type_required) {
4226 // Code generation on some platforms doesn't need accurate
4227 // ConvI2L types. Widening the type can help remove redundant
4228 // address computations.
4229 n->as_Type()->set_type(TypeLong::INT);
4230 ResourceMark rm;
4231 Unique_Node_List wq;
4232 wq.push(n);
4233 for (uint next = 0; next < wq.size(); next++) {
4234 Node *m = wq.at(next);
4235
4236 for(;;) {
4237 // Loop over all nodes with identical inputs edges as m
4238 Node* k = m->find_similar(m->Opcode());
4239 if (k == nullptr) {
4240 break;
4241 }
4242 // Push their uses so we get a chance to remove node made
4243 // redundant
4244 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4245 Node* u = k->fast_out(i);
4246 if (u->Opcode() == Op_LShiftL ||
4247 u->Opcode() == Op_AddL ||
4248 u->Opcode() == Op_SubL ||
4249 u->Opcode() == Op_AddP) {
4250 wq.push(u);
4251 }
4252 }
4253 // Replace all nodes with identical edges as m with m
4254 k->subsume_by(m, this);
4255 }
4256 }
4257 }
4258 break;
4259 }
4260 case Op_CmpUL: {
4261 if (!Matcher::has_match_rule(Op_CmpUL)) {
4262 // No support for unsigned long comparisons
4263 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4264 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4265 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4266 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4267 Node* andl = new AndLNode(orl, remove_sign_mask);
4268 Node* cmp = new CmpLNode(andl, n->in(2));
4269 n->subsume_by(cmp, this);
4270 }
4271 break;
4272 }
4273 #ifdef ASSERT
4274 case Op_InlineType: {
4275 n->dump(-1);
4276 assert(false, "inline type node was not removed");
4277 break;
4278 }
4279 case Op_ConNKlass: {
4280 const TypePtr* tp = n->as_Type()->type()->make_ptr();
4281 ciKlass* klass = tp->is_klassptr()->exact_klass();
4282 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
4283 break;
4284 }
4285 #endif
4286 default:
4287 assert(!n->is_Call(), "");
4288 assert(!n->is_Mem(), "");
4289 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4290 break;
4291 }
4292 }
4293
4294 //------------------------------final_graph_reshaping_walk---------------------
4295 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4296 // requires that the walk visits a node's inputs before visiting the node.
4297 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4298 Unique_Node_List sfpt;
4299
4300 frc._visited.set(root->_idx); // first, mark node as visited
4301 uint cnt = root->req();
4302 Node *n = root;
4303 uint i = 0;
4304 while (true) {
4305 if (i < cnt) {
4306 // Place all non-visited non-null inputs onto stack
4307 Node* m = n->in(i);
4308 ++i;
4309 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
4310 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
4311 // compute worst case interpreter size in case of a deoptimization
4312 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4313
4314 sfpt.push(m);
4315 }
4316 cnt = m->req();
4317 nstack.push(n, i); // put on stack parent and next input's index
4318 n = m;
4319 i = 0;
4320 }
4321 } else {
4322 // Now do post-visit work
4323 final_graph_reshaping_impl(n, frc, dead_nodes);
4324 if (nstack.is_empty())
4325 break; // finished
4326 n = nstack.node(); // Get node from stack
4327 cnt = n->req();
4328 i = nstack.index();
4329 nstack.pop(); // Shift to the next node on stack
4330 }
4331 }
4332
4333 // Skip next transformation if compressed oops are not used.
4334 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4335 (!UseCompressedOops && !UseCompressedClassPointers))
4336 return;
4337
4338 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4339 // It could be done for an uncommon traps or any safepoints/calls
4340 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4341 while (sfpt.size() > 0) {
4342 n = sfpt.pop();
4343 JVMState *jvms = n->as_SafePoint()->jvms();
4344 assert(jvms != nullptr, "sanity");
4345 int start = jvms->debug_start();
4346 int end = n->req();
4347 bool is_uncommon = (n->is_CallStaticJava() &&
4348 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4349 for (int j = start; j < end; j++) {
4350 Node* in = n->in(j);
4351 if (in->is_DecodeNarrowPtr()) {
4352 bool safe_to_skip = true;
4353 if (!is_uncommon ) {
4354 // Is it safe to skip?
4355 for (uint i = 0; i < in->outcnt(); i++) {
4356 Node* u = in->raw_out(i);
4357 if (!u->is_SafePoint() ||
4358 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4359 safe_to_skip = false;
4360 }
4361 }
4362 }
4363 if (safe_to_skip) {
4364 n->set_req(j, in->in(1));
4365 }
4366 if (in->outcnt() == 0) {
4367 in->disconnect_inputs(this);
4368 }
4369 }
4370 }
4371 }
4372 }
4373
4374 //------------------------------final_graph_reshaping--------------------------
4375 // Final Graph Reshaping.
4376 //
4377 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4378 // and not commoned up and forced early. Must come after regular
4379 // optimizations to avoid GVN undoing the cloning. Clone constant
4380 // inputs to Loop Phis; these will be split by the allocator anyways.
4381 // Remove Opaque nodes.
4382 // (2) Move last-uses by commutative operations to the left input to encourage
4383 // Intel update-in-place two-address operations and better register usage
4384 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4385 // calls canonicalizing them back.
4386 // (3) Count the number of double-precision FP ops, single-precision FP ops
4387 // and call sites. On Intel, we can get correct rounding either by
4388 // forcing singles to memory (requires extra stores and loads after each
4389 // FP bytecode) or we can set a rounding mode bit (requires setting and
4390 // clearing the mode bit around call sites). The mode bit is only used
4391 // if the relative frequency of single FP ops to calls is low enough.
4392 // This is a key transform for SPEC mpeg_audio.
4393 // (4) Detect infinite loops; blobs of code reachable from above but not
4394 // below. Several of the Code_Gen algorithms fail on such code shapes,
4395 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4396 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4397 // Detection is by looking for IfNodes where only 1 projection is
4398 // reachable from below or CatchNodes missing some targets.
4399 // (5) Assert for insane oop offsets in debug mode.
4400
4401 bool Compile::final_graph_reshaping() {
4402 // an infinite loop may have been eliminated by the optimizer,
4403 // in which case the graph will be empty.
4404 if (root()->req() == 1) {
4405 // Do not compile method that is only a trivial infinite loop,
4406 // since the content of the loop may have been eliminated.
4407 record_method_not_compilable("trivial infinite loop");
4408 return true;
4409 }
4410
4411 // Expensive nodes have their control input set to prevent the GVN
4412 // from freely commoning them. There's no GVN beyond this point so
4413 // no need to keep the control input. We want the expensive nodes to
4414 // be freely moved to the least frequent code path by gcm.
4415 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4416 for (int i = 0; i < expensive_count(); i++) {
4417 _expensive_nodes.at(i)->set_req(0, nullptr);
4418 }
4419
4420 Final_Reshape_Counts frc;
4421
4422 // Visit everybody reachable!
4423 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4424 Node_Stack nstack(live_nodes() >> 1);
4425 Unique_Node_List dead_nodes;
4426 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4427
4428 // Check for unreachable (from below) code (i.e., infinite loops).
4429 for( uint i = 0; i < frc._tests.size(); i++ ) {
4430 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4431 // Get number of CFG targets.
4432 // Note that PCTables include exception targets after calls.
4433 uint required_outcnt = n->required_outcnt();
4434 if (n->outcnt() != required_outcnt) {
4435 // Check for a few special cases. Rethrow Nodes never take the
4436 // 'fall-thru' path, so expected kids is 1 less.
4437 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4438 if (n->in(0)->in(0)->is_Call()) {
4439 CallNode* call = n->in(0)->in(0)->as_Call();
4440 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4441 required_outcnt--; // Rethrow always has 1 less kid
4442 } else if (call->req() > TypeFunc::Parms &&
4443 call->is_CallDynamicJava()) {
4444 // Check for null receiver. In such case, the optimizer has
4445 // detected that the virtual call will always result in a null
4446 // pointer exception. The fall-through projection of this CatchNode
4447 // will not be populated.
4448 Node* arg0 = call->in(TypeFunc::Parms);
4449 if (arg0->is_Type() &&
4450 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4451 required_outcnt--;
4452 }
4453 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4454 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4455 // Check for illegal array length. In such case, the optimizer has
4456 // detected that the allocation attempt will always result in an
4457 // exception. There is no fall-through projection of this CatchNode .
4458 assert(call->is_CallStaticJava(), "static call expected");
4459 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4460 uint valid_length_test_input = call->req() - 1;
4461 Node* valid_length_test = call->in(valid_length_test_input);
4462 call->del_req(valid_length_test_input);
4463 if (valid_length_test->find_int_con(1) == 0) {
4464 required_outcnt--;
4465 }
4466 dead_nodes.push(valid_length_test);
4467 assert(n->outcnt() == required_outcnt, "malformed control flow");
4468 continue;
4469 }
4470 }
4471 }
4472
4473 // Recheck with a better notion of 'required_outcnt'
4474 if (n->outcnt() != required_outcnt) {
4475 record_method_not_compilable("malformed control flow");
4476 return true; // Not all targets reachable!
4477 }
4478 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4479 CallNode* call = n->in(0)->in(0)->as_Call();
4480 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4481 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4482 assert(call->is_CallStaticJava(), "static call expected");
4483 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4484 uint valid_length_test_input = call->req() - 1;
4485 dead_nodes.push(call->in(valid_length_test_input));
4486 call->del_req(valid_length_test_input); // valid length test useless now
4487 }
4488 }
4489 // Check that I actually visited all kids. Unreached kids
4490 // must be infinite loops.
4491 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4492 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4493 record_method_not_compilable("infinite loop");
4494 return true; // Found unvisited kid; must be unreach
4495 }
4496
4497 // Here so verification code in final_graph_reshaping_walk()
4498 // always see an OuterStripMinedLoopEnd
4499 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4500 IfNode* init_iff = n->as_If();
4501 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4502 n->subsume_by(iff, this);
4503 }
4504 }
4505
4506 while (dead_nodes.size() > 0) {
4507 Node* m = dead_nodes.pop();
4508 if (m->outcnt() == 0 && m != top()) {
4509 for (uint j = 0; j < m->req(); j++) {
4510 Node* in = m->in(j);
4511 if (in != nullptr) {
4512 dead_nodes.push(in);
4513 }
4514 }
4515 m->disconnect_inputs(this);
4516 }
4517 }
4518
4519 #ifdef IA32
4520 // If original bytecodes contained a mixture of floats and doubles
4521 // check if the optimizer has made it homogeneous, item (3).
4522 if (UseSSE == 0 &&
4523 frc.get_float_count() > 32 &&
4524 frc.get_double_count() == 0 &&
4525 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4526 set_24_bit_selection_and_mode(false, true);
4527 }
4528 #endif // IA32
4529
4530 set_java_calls(frc.get_java_call_count());
4531 set_inner_loops(frc.get_inner_loop_count());
4532
4533 // No infinite loops, no reason to bail out.
4534 return false;
4535 }
4536
4537 //-----------------------------too_many_traps----------------------------------
4538 // Report if there are too many traps at the current method and bci.
4539 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4540 bool Compile::too_many_traps(ciMethod* method,
4541 int bci,
4542 Deoptimization::DeoptReason reason) {
4543 ciMethodData* md = method->method_data();
4544 if (md->is_empty()) {
4545 // Assume the trap has not occurred, or that it occurred only
4546 // because of a transient condition during start-up in the interpreter.
4547 return false;
4548 }
4549 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4550 if (md->has_trap_at(bci, m, reason) != 0) {
4551 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4552 // Also, if there are multiple reasons, or if there is no per-BCI record,
4553 // assume the worst.
4554 if (log())
4555 log()->elem("observe trap='%s' count='%d'",
4556 Deoptimization::trap_reason_name(reason),
4557 md->trap_count(reason));
4558 return true;
4559 } else {
4560 // Ignore method/bci and see if there have been too many globally.
4561 return too_many_traps(reason, md);
4562 }
4563 }
4564
4565 // Less-accurate variant which does not require a method and bci.
4566 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4567 ciMethodData* logmd) {
4568 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4569 // Too many traps globally.
4570 // Note that we use cumulative trap_count, not just md->trap_count.
4571 if (log()) {
4572 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4573 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4574 Deoptimization::trap_reason_name(reason),
4575 mcount, trap_count(reason));
4576 }
4577 return true;
4578 } else {
4579 // The coast is clear.
4580 return false;
4581 }
4582 }
4583
4584 //--------------------------too_many_recompiles--------------------------------
4585 // Report if there are too many recompiles at the current method and bci.
4586 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4587 // Is not eager to return true, since this will cause the compiler to use
4588 // Action_none for a trap point, to avoid too many recompilations.
4589 bool Compile::too_many_recompiles(ciMethod* method,
4590 int bci,
4591 Deoptimization::DeoptReason reason) {
4592 ciMethodData* md = method->method_data();
4593 if (md->is_empty()) {
4594 // Assume the trap has not occurred, or that it occurred only
4595 // because of a transient condition during start-up in the interpreter.
4596 return false;
4597 }
4598 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4599 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4600 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4601 Deoptimization::DeoptReason per_bc_reason
4602 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4603 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4604 if ((per_bc_reason == Deoptimization::Reason_none
4605 || md->has_trap_at(bci, m, reason) != 0)
4606 // The trap frequency measure we care about is the recompile count:
4607 && md->trap_recompiled_at(bci, m)
4608 && md->overflow_recompile_count() >= bc_cutoff) {
4609 // Do not emit a trap here if it has already caused recompilations.
4610 // Also, if there are multiple reasons, or if there is no per-BCI record,
4611 // assume the worst.
4612 if (log())
4613 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4614 Deoptimization::trap_reason_name(reason),
4615 md->trap_count(reason),
4616 md->overflow_recompile_count());
4617 return true;
4618 } else if (trap_count(reason) != 0
4619 && decompile_count() >= m_cutoff) {
4620 // Too many recompiles globally, and we have seen this sort of trap.
4621 // Use cumulative decompile_count, not just md->decompile_count.
4622 if (log())
4623 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4624 Deoptimization::trap_reason_name(reason),
4625 md->trap_count(reason), trap_count(reason),
4626 md->decompile_count(), decompile_count());
4627 return true;
4628 } else {
4629 // The coast is clear.
4630 return false;
4631 }
4632 }
4633
4634 // Compute when not to trap. Used by matching trap based nodes and
4635 // NullCheck optimization.
4636 void Compile::set_allowed_deopt_reasons() {
4637 _allowed_reasons = 0;
4638 if (is_method_compilation()) {
4639 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4640 assert(rs < BitsPerInt, "recode bit map");
4641 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4642 _allowed_reasons |= nth_bit(rs);
4643 }
4644 }
4645 }
4646 }
4647
4648 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4649 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4650 }
4651
4652 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4653 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4654 }
4655
4656 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4657 if (holder->is_initialized()) {
4658 return false;
4659 }
4660 if (holder->is_being_initialized()) {
4661 if (accessing_method->holder() == holder) {
4662 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4663 // <init>, or a static method. In all those cases, there was an initialization
4664 // barrier on the holder klass passed.
4665 if (accessing_method->is_class_initializer() ||
4666 accessing_method->is_object_constructor() ||
4667 accessing_method->is_static()) {
4668 return false;
4669 }
4670 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4671 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4672 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4673 // child class can become fully initialized while its parent class is still being initialized.
4674 if (accessing_method->is_class_initializer()) {
4675 return false;
4676 }
4677 }
4678 ciMethod* root = method(); // the root method of compilation
4679 if (root != accessing_method) {
4680 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4681 }
4682 }
4683 return true;
4684 }
4685
4686 #ifndef PRODUCT
4687 //------------------------------verify_bidirectional_edges---------------------
4688 // For each input edge to a node (ie - for each Use-Def edge), verify that
4689 // there is a corresponding Def-Use edge.
4690 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4691 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4692 uint stack_size = live_nodes() >> 4;
4693 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4694 nstack.push(_root);
4695
4696 while (nstack.size() > 0) {
4697 Node* n = nstack.pop();
4698 if (visited.member(n)) {
4699 continue;
4700 }
4701 visited.push(n);
4702
4703 // Walk over all input edges, checking for correspondence
4704 uint length = n->len();
4705 for (uint i = 0; i < length; i++) {
4706 Node* in = n->in(i);
4707 if (in != nullptr && !visited.member(in)) {
4708 nstack.push(in); // Put it on stack
4709 }
4710 if (in != nullptr && !in->is_top()) {
4711 // Count instances of `next`
4712 int cnt = 0;
4713 for (uint idx = 0; idx < in->_outcnt; idx++) {
4714 if (in->_out[idx] == n) {
4715 cnt++;
4716 }
4717 }
4718 assert(cnt > 0, "Failed to find Def-Use edge.");
4719 // Check for duplicate edges
4720 // walk the input array downcounting the input edges to n
4721 for (uint j = 0; j < length; j++) {
4722 if (n->in(j) == in) {
4723 cnt--;
4724 }
4725 }
4726 assert(cnt == 0, "Mismatched edge count.");
4727 } else if (in == nullptr) {
4728 assert(i == 0 || i >= n->req() ||
4729 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4730 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
4731 (n->is_Unlock() && i == (n->req() - 1)) ||
4732 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4733 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
4734 } else {
4735 assert(in->is_top(), "sanity");
4736 // Nothing to check.
4737 }
4738 }
4739 }
4740 }
4741
4742 //------------------------------verify_graph_edges---------------------------
4743 // Walk the Graph and verify that there is a one-to-one correspondence
4744 // between Use-Def edges and Def-Use edges in the graph.
4745 void Compile::verify_graph_edges(bool no_dead_code) {
4746 if (VerifyGraphEdges) {
4747 Unique_Node_List visited;
4748
4749 // Call graph walk to check edges
4750 verify_bidirectional_edges(visited);
4751 if (no_dead_code) {
4752 // Now make sure that no visited node is used by an unvisited node.
4753 bool dead_nodes = false;
4754 Unique_Node_List checked;
4755 while (visited.size() > 0) {
4756 Node* n = visited.pop();
4757 checked.push(n);
4758 for (uint i = 0; i < n->outcnt(); i++) {
4759 Node* use = n->raw_out(i);
4760 if (checked.member(use)) continue; // already checked
4761 if (visited.member(use)) continue; // already in the graph
4762 if (use->is_Con()) continue; // a dead ConNode is OK
4763 // At this point, we have found a dead node which is DU-reachable.
4764 if (!dead_nodes) {
4765 tty->print_cr("*** Dead nodes reachable via DU edges:");
4766 dead_nodes = true;
4767 }
4768 use->dump(2);
4769 tty->print_cr("---");
4770 checked.push(use); // No repeats; pretend it is now checked.
4771 }
4772 }
4773 assert(!dead_nodes, "using nodes must be reachable from root");
4774 }
4775 }
4776 }
4777 #endif
4778
4779 // The Compile object keeps track of failure reasons separately from the ciEnv.
4780 // This is required because there is not quite a 1-1 relation between the
4781 // ciEnv and its compilation task and the Compile object. Note that one
4782 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4783 // to backtrack and retry without subsuming loads. Other than this backtracking
4784 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4785 // by the logic in C2Compiler.
4786 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4787 if (log() != nullptr) {
4788 log()->elem("failure reason='%s' phase='compile'", reason);
4789 }
4790 if (_failure_reason.get() == nullptr) {
4791 // Record the first failure reason.
4792 _failure_reason.set(reason);
4793 if (CaptureBailoutInformation) {
4794 _first_failure_details = new CompilationFailureInfo(reason);
4795 }
4796 } else {
4797 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4798 }
4799
4800 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4801 C->print_method(PHASE_FAILURE, 1);
4802 }
4803 _root = nullptr; // flush the graph, too
4804 }
4805
4806 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4807 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4808 _compile(Compile::current()),
4809 _log(nullptr),
4810 _dolog(CITimeVerbose)
4811 {
4812 assert(_compile != nullptr, "sanity check");
4813 assert(id != PhaseTraceId::_t_none, "Don't use none");
4814 if (_dolog) {
4815 _log = _compile->log();
4816 }
4817 if (_log != nullptr) {
4818 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4819 _log->stamp();
4820 _log->end_head();
4821 }
4822
4823 // Inform memory statistic, if enabled
4824 if (CompilationMemoryStatistic::enabled()) {
4825 CompilationMemoryStatistic::on_phase_start((int)id, name);
4826 }
4827 }
4828
4829 Compile::TracePhase::TracePhase(PhaseTraceId id)
4830 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4831
4832 Compile::TracePhase::~TracePhase() {
4833
4834 // Inform memory statistic, if enabled
4835 if (CompilationMemoryStatistic::enabled()) {
4836 CompilationMemoryStatistic::on_phase_end();
4837 }
4838
4839 if (_compile->failing_internal()) {
4840 if (_log != nullptr) {
4841 _log->done("phase");
4842 }
4843 return; // timing code, not stressing bailouts.
4844 }
4845 #ifdef ASSERT
4846 if (PrintIdealNodeCount) {
4847 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4848 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4849 }
4850
4851 if (VerifyIdealNodeCount) {
4852 _compile->print_missing_nodes();
4853 }
4854 #endif
4855
4856 if (_log != nullptr) {
4857 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4858 }
4859 }
4860
4861 //----------------------------static_subtype_check-----------------------------
4862 // Shortcut important common cases when superklass is exact:
4863 // (0) superklass is java.lang.Object (can occur in reflective code)
4864 // (1) subklass is already limited to a subtype of superklass => always ok
4865 // (2) subklass does not overlap with superklass => always fail
4866 // (3) superklass has NO subtypes and we can check with a simple compare.
4867 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4868 if (skip) {
4869 return SSC_full_test; // Let caller generate the general case.
4870 }
4871
4872 if (subk->is_java_subtype_of(superk)) {
4873 return SSC_always_true; // (0) and (1) this test cannot fail
4874 }
4875
4876 if (!subk->maybe_java_subtype_of(superk)) {
4877 return SSC_always_false; // (2) true path dead; no dynamic test needed
4878 }
4879
4880 const Type* superelem = superk;
4881 if (superk->isa_aryklassptr()) {
4882 int ignored;
4883 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4884
4885 // Do not fold the subtype check to an array klass pointer comparison for null-able inline type arrays
4886 // because null-free [LMyValue <: null-able [LMyValue but the klasses are different. Perform a full test.
4887 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
4888 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
4889 return SSC_full_test;
4890 }
4891 }
4892
4893 if (superelem->isa_instklassptr()) {
4894 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4895 if (!ik->has_subklass()) {
4896 if (!ik->is_final()) {
4897 // Add a dependency if there is a chance of a later subclass.
4898 dependencies()->assert_leaf_type(ik);
4899 }
4900 if (!superk->maybe_java_subtype_of(subk)) {
4901 return SSC_always_false;
4902 }
4903 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4904 }
4905 } else {
4906 // A primitive array type has no subtypes.
4907 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4908 }
4909
4910 return SSC_full_test;
4911 }
4912
4913 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4914 #ifdef _LP64
4915 // The scaled index operand to AddP must be a clean 64-bit value.
4916 // Java allows a 32-bit int to be incremented to a negative
4917 // value, which appears in a 64-bit register as a large
4918 // positive number. Using that large positive number as an
4919 // operand in pointer arithmetic has bad consequences.
4920 // On the other hand, 32-bit overflow is rare, and the possibility
4921 // can often be excluded, if we annotate the ConvI2L node with
4922 // a type assertion that its value is known to be a small positive
4923 // number. (The prior range check has ensured this.)
4924 // This assertion is used by ConvI2LNode::Ideal.
4925 int index_max = max_jint - 1; // array size is max_jint, index is one less
4926 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4927 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4928 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4929 #endif
4930 return idx;
4931 }
4932
4933 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4934 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4935 if (ctrl != nullptr) {
4936 // Express control dependency by a CastII node with a narrow type.
4937 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4938 // node from floating above the range check during loop optimizations. Otherwise, the
4939 // ConvI2L node may be eliminated independently of the range check, causing the data path
4940 // to become TOP while the control path is still there (although it's unreachable).
4941 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4942 value = phase->transform(value);
4943 }
4944 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4945 return phase->transform(new ConvI2LNode(value, ltype));
4946 }
4947
4948 void Compile::dump_print_inlining() {
4949 inline_printer()->print_on(tty);
4950 }
4951
4952 void Compile::log_late_inline(CallGenerator* cg) {
4953 if (log() != nullptr) {
4954 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4955 cg->unique_id());
4956 JVMState* p = cg->call_node()->jvms();
4957 while (p != nullptr) {
4958 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4959 p = p->caller();
4960 }
4961 log()->tail("late_inline");
4962 }
4963 }
4964
4965 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4966 log_late_inline(cg);
4967 if (log() != nullptr) {
4968 log()->inline_fail(msg);
4969 }
4970 }
4971
4972 void Compile::log_inline_id(CallGenerator* cg) {
4973 if (log() != nullptr) {
4974 // The LogCompilation tool needs a unique way to identify late
4975 // inline call sites. This id must be unique for this call site in
4976 // this compilation. Try to have it unique across compilations as
4977 // well because it can be convenient when grepping through the log
4978 // file.
4979 // Distinguish OSR compilations from others in case CICountOSR is
4980 // on.
4981 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4982 cg->set_unique_id(id);
4983 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4984 }
4985 }
4986
4987 void Compile::log_inline_failure(const char* msg) {
4988 if (C->log() != nullptr) {
4989 C->log()->inline_fail(msg);
4990 }
4991 }
4992
4993
4994 // Dump inlining replay data to the stream.
4995 // Don't change thread state and acquire any locks.
4996 void Compile::dump_inline_data(outputStream* out) {
4997 InlineTree* inl_tree = ilt();
4998 if (inl_tree != nullptr) {
4999 out->print(" inline %d", inl_tree->count());
5000 inl_tree->dump_replay_data(out);
5001 }
5002 }
5003
5004 void Compile::dump_inline_data_reduced(outputStream* out) {
5005 assert(ReplayReduce, "");
5006
5007 InlineTree* inl_tree = ilt();
5008 if (inl_tree == nullptr) {
5009 return;
5010 }
5011 // Enable iterative replay file reduction
5012 // Output "compile" lines for depth 1 subtrees,
5013 // simulating that those trees were compiled
5014 // instead of inlined.
5015 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5016 InlineTree* sub = inl_tree->subtrees().at(i);
5017 if (sub->inline_level() != 1) {
5018 continue;
5019 }
5020
5021 ciMethod* method = sub->method();
5022 int entry_bci = -1;
5023 int comp_level = env()->task()->comp_level();
5024 out->print("compile ");
5025 method->dump_name_as_ascii(out);
5026 out->print(" %d %d", entry_bci, comp_level);
5027 out->print(" inline %d", sub->count());
5028 sub->dump_replay_data(out, -1);
5029 out->cr();
5030 }
5031 }
5032
5033 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5034 if (n1->Opcode() < n2->Opcode()) return -1;
5035 else if (n1->Opcode() > n2->Opcode()) return 1;
5036
5037 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5038 for (uint i = 1; i < n1->req(); i++) {
5039 if (n1->in(i) < n2->in(i)) return -1;
5040 else if (n1->in(i) > n2->in(i)) return 1;
5041 }
5042
5043 return 0;
5044 }
5045
5046 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5047 Node* n1 = *n1p;
5048 Node* n2 = *n2p;
5049
5050 return cmp_expensive_nodes(n1, n2);
5051 }
5052
5053 void Compile::sort_expensive_nodes() {
5054 if (!expensive_nodes_sorted()) {
5055 _expensive_nodes.sort(cmp_expensive_nodes);
5056 }
5057 }
5058
5059 bool Compile::expensive_nodes_sorted() const {
5060 for (int i = 1; i < _expensive_nodes.length(); i++) {
5061 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5062 return false;
5063 }
5064 }
5065 return true;
5066 }
5067
5068 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5069 if (_expensive_nodes.length() == 0) {
5070 return false;
5071 }
5072
5073 assert(OptimizeExpensiveOps, "optimization off?");
5074
5075 // Take this opportunity to remove dead nodes from the list
5076 int j = 0;
5077 for (int i = 0; i < _expensive_nodes.length(); i++) {
5078 Node* n = _expensive_nodes.at(i);
5079 if (!n->is_unreachable(igvn)) {
5080 assert(n->is_expensive(), "should be expensive");
5081 _expensive_nodes.at_put(j, n);
5082 j++;
5083 }
5084 }
5085 _expensive_nodes.trunc_to(j);
5086
5087 // Then sort the list so that similar nodes are next to each other
5088 // and check for at least two nodes of identical kind with same data
5089 // inputs.
5090 sort_expensive_nodes();
5091
5092 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5093 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5094 return true;
5095 }
5096 }
5097
5098 return false;
5099 }
5100
5101 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5102 if (_expensive_nodes.length() == 0) {
5103 return;
5104 }
5105
5106 assert(OptimizeExpensiveOps, "optimization off?");
5107
5108 // Sort to bring similar nodes next to each other and clear the
5109 // control input of nodes for which there's only a single copy.
5110 sort_expensive_nodes();
5111
5112 int j = 0;
5113 int identical = 0;
5114 int i = 0;
5115 bool modified = false;
5116 for (; i < _expensive_nodes.length()-1; i++) {
5117 assert(j <= i, "can't write beyond current index");
5118 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5119 identical++;
5120 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5121 continue;
5122 }
5123 if (identical > 0) {
5124 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5125 identical = 0;
5126 } else {
5127 Node* n = _expensive_nodes.at(i);
5128 igvn.replace_input_of(n, 0, nullptr);
5129 igvn.hash_insert(n);
5130 modified = true;
5131 }
5132 }
5133 if (identical > 0) {
5134 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5135 } else if (_expensive_nodes.length() >= 1) {
5136 Node* n = _expensive_nodes.at(i);
5137 igvn.replace_input_of(n, 0, nullptr);
5138 igvn.hash_insert(n);
5139 modified = true;
5140 }
5141 _expensive_nodes.trunc_to(j);
5142 if (modified) {
5143 igvn.optimize();
5144 }
5145 }
5146
5147 void Compile::add_expensive_node(Node * n) {
5148 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5149 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5150 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5151 if (OptimizeExpensiveOps) {
5152 _expensive_nodes.append(n);
5153 } else {
5154 // Clear control input and let IGVN optimize expensive nodes if
5155 // OptimizeExpensiveOps is off.
5156 n->set_req(0, nullptr);
5157 }
5158 }
5159
5160 /**
5161 * Track coarsened Lock and Unlock nodes.
5162 */
5163
5164 class Lock_List : public Node_List {
5165 uint _origin_cnt;
5166 public:
5167 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5168 uint origin_cnt() const { return _origin_cnt; }
5169 };
5170
5171 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5172 int length = locks.length();
5173 if (length > 0) {
5174 // Have to keep this list until locks elimination during Macro nodes elimination.
5175 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5176 AbstractLockNode* alock = locks.at(0);
5177 BoxLockNode* box = alock->box_node()->as_BoxLock();
5178 for (int i = 0; i < length; i++) {
5179 AbstractLockNode* lock = locks.at(i);
5180 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5181 locks_list->push(lock);
5182 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
5183 if (this_box != box) {
5184 // Locking regions (BoxLock) could be Unbalanced here:
5185 // - its coarsened locks were eliminated in earlier
5186 // macro nodes elimination followed by loop unroll
5187 // - it is OSR locking region (no Lock node)
5188 // Preserve Unbalanced status in such cases.
5189 if (!this_box->is_unbalanced()) {
5190 this_box->set_coarsened();
5191 }
5192 if (!box->is_unbalanced()) {
5193 box->set_coarsened();
5194 }
5195 }
5196 }
5197 _coarsened_locks.append(locks_list);
5198 }
5199 }
5200
5201 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5202 int count = coarsened_count();
5203 for (int i = 0; i < count; i++) {
5204 Node_List* locks_list = _coarsened_locks.at(i);
5205 for (uint j = 0; j < locks_list->size(); j++) {
5206 Node* lock = locks_list->at(j);
5207 assert(lock->is_AbstractLock(), "sanity");
5208 if (!useful.member(lock)) {
5209 locks_list->yank(lock);
5210 }
5211 }
5212 }
5213 }
5214
5215 void Compile::remove_coarsened_lock(Node* n) {
5216 if (n->is_AbstractLock()) {
5217 int count = coarsened_count();
5218 for (int i = 0; i < count; i++) {
5219 Node_List* locks_list = _coarsened_locks.at(i);
5220 locks_list->yank(n);
5221 }
5222 }
5223 }
5224
5225 bool Compile::coarsened_locks_consistent() {
5226 int count = coarsened_count();
5227 for (int i = 0; i < count; i++) {
5228 bool unbalanced = false;
5229 bool modified = false; // track locks kind modifications
5230 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5231 uint size = locks_list->size();
5232 if (size == 0) {
5233 unbalanced = false; // All locks were eliminated - good
5234 } else if (size != locks_list->origin_cnt()) {
5235 unbalanced = true; // Some locks were removed from list
5236 } else {
5237 for (uint j = 0; j < size; j++) {
5238 Node* lock = locks_list->at(j);
5239 // All nodes in group should have the same state (modified or not)
5240 if (!lock->as_AbstractLock()->is_coarsened()) {
5241 if (j == 0) {
5242 // first on list was modified, the rest should be too for consistency
5243 modified = true;
5244 } else if (!modified) {
5245 // this lock was modified but previous locks on the list were not
5246 unbalanced = true;
5247 break;
5248 }
5249 } else if (modified) {
5250 // previous locks on list were modified but not this lock
5251 unbalanced = true;
5252 break;
5253 }
5254 }
5255 }
5256 if (unbalanced) {
5257 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5258 #ifdef ASSERT
5259 if (PrintEliminateLocks) {
5260 tty->print_cr("=== unbalanced coarsened locks ===");
5261 for (uint l = 0; l < size; l++) {
5262 locks_list->at(l)->dump();
5263 }
5264 }
5265 #endif
5266 record_failure(C2Compiler::retry_no_locks_coarsening());
5267 return false;
5268 }
5269 }
5270 return true;
5271 }
5272
5273 // Mark locking regions (identified by BoxLockNode) as unbalanced if
5274 // locks coarsening optimization removed Lock/Unlock nodes from them.
5275 // Such regions become unbalanced because coarsening only removes part
5276 // of Lock/Unlock nodes in region. As result we can't execute other
5277 // locks elimination optimizations which assume all code paths have
5278 // corresponding pair of Lock/Unlock nodes - they are balanced.
5279 void Compile::mark_unbalanced_boxes() const {
5280 int count = coarsened_count();
5281 for (int i = 0; i < count; i++) {
5282 Node_List* locks_list = _coarsened_locks.at(i);
5283 uint size = locks_list->size();
5284 if (size > 0) {
5285 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
5286 BoxLockNode* box = alock->box_node()->as_BoxLock();
5287 if (alock->is_coarsened()) {
5288 // coarsened_locks_consistent(), which is called before this method, verifies
5289 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
5290 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5291 for (uint j = 1; j < size; j++) {
5292 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
5293 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
5294 if (box != this_box) {
5295 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5296 box->set_unbalanced();
5297 this_box->set_unbalanced();
5298 }
5299 }
5300 }
5301 }
5302 }
5303 }
5304
5305 /**
5306 * Remove the speculative part of types and clean up the graph
5307 */
5308 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5309 if (UseTypeSpeculation) {
5310 Unique_Node_List worklist;
5311 worklist.push(root());
5312 int modified = 0;
5313 // Go over all type nodes that carry a speculative type, drop the
5314 // speculative part of the type and enqueue the node for an igvn
5315 // which may optimize it out.
5316 for (uint next = 0; next < worklist.size(); ++next) {
5317 Node *n = worklist.at(next);
5318 if (n->is_Type()) {
5319 TypeNode* tn = n->as_Type();
5320 const Type* t = tn->type();
5321 const Type* t_no_spec = t->remove_speculative();
5322 if (t_no_spec != t) {
5323 bool in_hash = igvn.hash_delete(n);
5324 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5325 tn->set_type(t_no_spec);
5326 igvn.hash_insert(n);
5327 igvn._worklist.push(n); // give it a chance to go away
5328 modified++;
5329 }
5330 }
5331 // Iterate over outs - endless loops is unreachable from below
5332 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5333 Node *m = n->fast_out(i);
5334 if (not_a_node(m)) {
5335 continue;
5336 }
5337 worklist.push(m);
5338 }
5339 }
5340 // Drop the speculative part of all types in the igvn's type table
5341 igvn.remove_speculative_types();
5342 if (modified > 0) {
5343 igvn.optimize();
5344 if (failing()) return;
5345 }
5346 #ifdef ASSERT
5347 // Verify that after the IGVN is over no speculative type has resurfaced
5348 worklist.clear();
5349 worklist.push(root());
5350 for (uint next = 0; next < worklist.size(); ++next) {
5351 Node *n = worklist.at(next);
5352 const Type* t = igvn.type_or_null(n);
5353 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5354 if (n->is_Type()) {
5355 t = n->as_Type()->type();
5356 assert(t == t->remove_speculative(), "no more speculative types");
5357 }
5358 // Iterate over outs - endless loops is unreachable from below
5359 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5360 Node *m = n->fast_out(i);
5361 if (not_a_node(m)) {
5362 continue;
5363 }
5364 worklist.push(m);
5365 }
5366 }
5367 igvn.check_no_speculative_types();
5368 #endif
5369 }
5370 }
5371
5372 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5373 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5374 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5375 if (!EnableValhalla || ta == nullptr || tb == nullptr ||
5376 ta->is_zero_type() || tb->is_zero_type() ||
5377 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5378 // Use old acmp if one operand is null or not an inline type
5379 return new CmpPNode(a, b);
5380 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5381 // We know that one operand is an inline type. Therefore,
5382 // new acmp will only return true if both operands are nullptr.
5383 // Check if both operands are null by or'ing the oops.
5384 a = phase->transform(new CastP2XNode(nullptr, a));
5385 b = phase->transform(new CastP2XNode(nullptr, b));
5386 a = phase->transform(new OrXNode(a, b));
5387 return new CmpXNode(a, phase->MakeConX(0));
5388 }
5389 // Use new acmp
5390 return nullptr;
5391 }
5392
5393 // Auxiliary methods to support randomized stressing/fuzzing.
5394
5395 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5396 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5397 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5398 FLAG_SET_ERGO(StressSeed, _stress_seed);
5399 } else {
5400 _stress_seed = StressSeed;
5401 }
5402 if (_log != nullptr) {
5403 _log->elem("stress_test seed='%u'", _stress_seed);
5404 }
5405 }
5406
5407 int Compile::random() {
5408 _stress_seed = os::next_random(_stress_seed);
5409 return static_cast<int>(_stress_seed);
5410 }
5411
5412 // This method can be called the arbitrary number of times, with current count
5413 // as the argument. The logic allows selecting a single candidate from the
5414 // running list of candidates as follows:
5415 // int count = 0;
5416 // Cand* selected = null;
5417 // while(cand = cand->next()) {
5418 // if (randomized_select(++count)) {
5419 // selected = cand;
5420 // }
5421 // }
5422 //
5423 // Including count equalizes the chances any candidate is "selected".
5424 // This is useful when we don't have the complete list of candidates to choose
5425 // from uniformly. In this case, we need to adjust the randomicity of the
5426 // selection, or else we will end up biasing the selection towards the latter
5427 // candidates.
5428 //
5429 // Quick back-envelope calculation shows that for the list of n candidates
5430 // the equal probability for the candidate to persist as "best" can be
5431 // achieved by replacing it with "next" k-th candidate with the probability
5432 // of 1/k. It can be easily shown that by the end of the run, the
5433 // probability for any candidate is converged to 1/n, thus giving the
5434 // uniform distribution among all the candidates.
5435 //
5436 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5437 #define RANDOMIZED_DOMAIN_POW 29
5438 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5439 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5440 bool Compile::randomized_select(int count) {
5441 assert(count > 0, "only positive");
5442 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5443 }
5444
5445 #ifdef ASSERT
5446 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5447 bool Compile::fail_randomly() {
5448 if ((random() % StressBailoutMean) != 0) {
5449 return false;
5450 }
5451 record_failure("StressBailout");
5452 return true;
5453 }
5454
5455 bool Compile::failure_is_artificial() {
5456 return C->failure_reason_is("StressBailout");
5457 }
5458 #endif
5459
5460 CloneMap& Compile::clone_map() { return _clone_map; }
5461 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5462
5463 void NodeCloneInfo::dump_on(outputStream* st) const {
5464 st->print(" {%d:%d} ", idx(), gen());
5465 }
5466
5467 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5468 uint64_t val = value(old->_idx);
5469 NodeCloneInfo cio(val);
5470 assert(val != 0, "old node should be in the map");
5471 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5472 insert(nnn->_idx, cin.get());
5473 #ifndef PRODUCT
5474 if (is_debug()) {
5475 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5476 }
5477 #endif
5478 }
5479
5480 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5481 NodeCloneInfo cio(value(old->_idx));
5482 if (cio.get() == 0) {
5483 cio.set(old->_idx, 0);
5484 insert(old->_idx, cio.get());
5485 #ifndef PRODUCT
5486 if (is_debug()) {
5487 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5488 }
5489 #endif
5490 }
5491 clone(old, nnn, gen);
5492 }
5493
5494 int CloneMap::max_gen() const {
5495 int g = 0;
5496 DictI di(_dict);
5497 for(; di.test(); ++di) {
5498 int t = gen(di._key);
5499 if (g < t) {
5500 g = t;
5501 #ifndef PRODUCT
5502 if (is_debug()) {
5503 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5504 }
5505 #endif
5506 }
5507 }
5508 return g;
5509 }
5510
5511 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5512 uint64_t val = value(key);
5513 if (val != 0) {
5514 NodeCloneInfo ni(val);
5515 ni.dump_on(st);
5516 }
5517 }
5518
5519 void Compile::shuffle_macro_nodes() {
5520 if (_macro_nodes.length() < 2) {
5521 return;
5522 }
5523 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5524 uint j = C->random() % (i + 1);
5525 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5526 }
5527 }
5528
5529 // Move Allocate nodes to the start of the list
5530 void Compile::sort_macro_nodes() {
5531 int count = macro_count();
5532 int allocates = 0;
5533 for (int i = 0; i < count; i++) {
5534 Node* n = macro_node(i);
5535 if (n->is_Allocate()) {
5536 if (i != allocates) {
5537 Node* tmp = macro_node(allocates);
5538 _macro_nodes.at_put(allocates, n);
5539 _macro_nodes.at_put(i, tmp);
5540 }
5541 allocates++;
5542 }
5543 }
5544 }
5545
5546 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5547 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5548 EventCompilerPhase event(UNTIMED);
5549 if (event.should_commit()) {
5550 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5551 }
5552 #ifndef PRODUCT
5553 ResourceMark rm;
5554 stringStream ss;
5555 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5556 int iter = ++_igv_phase_iter[cpt];
5557 if (iter > 1) {
5558 ss.print(" %d", iter);
5559 }
5560 if (n != nullptr) {
5561 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5562 if (n->is_Call()) {
5563 CallNode* call = n->as_Call();
5564 if (call->_name != nullptr) {
5565 // E.g. uncommon traps etc.
5566 ss.print(" - %s", call->_name);
5567 } else if (call->is_CallJava()) {
5568 CallJavaNode* call_java = call->as_CallJava();
5569 if (call_java->method() != nullptr) {
5570 ss.print(" -");
5571 call_java->method()->print_short_name(&ss);
5572 }
5573 }
5574 }
5575 }
5576
5577 const char* name = ss.as_string();
5578 if (should_print_igv(level)) {
5579 _igv_printer->print_graph(name);
5580 }
5581 if (should_print_phase(cpt)) {
5582 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5583 }
5584 #endif
5585 C->_latest_stage_start_counter.stamp();
5586 }
5587
5588 // Only used from CompileWrapper
5589 void Compile::begin_method() {
5590 #ifndef PRODUCT
5591 if (_method != nullptr && should_print_igv(1)) {
5592 _igv_printer->begin_method();
5593 }
5594 #endif
5595 C->_latest_stage_start_counter.stamp();
5596 }
5597
5598 // Only used from CompileWrapper
5599 void Compile::end_method() {
5600 EventCompilerPhase event(UNTIMED);
5601 if (event.should_commit()) {
5602 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5603 }
5604
5605 #ifndef PRODUCT
5606 if (_method != nullptr && should_print_igv(1)) {
5607 _igv_printer->end_method();
5608 }
5609 #endif
5610 }
5611
5612 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5613 #ifndef PRODUCT
5614 if (_directive->should_print_phase(cpt)) {
5615 return true;
5616 }
5617 #endif
5618 return false;
5619 }
5620
5621 #ifndef PRODUCT
5622 void Compile::init_igv() {
5623 if (_igv_printer == nullptr) {
5624 _igv_printer = IdealGraphPrinter::printer();
5625 _igv_printer->set_compile(this);
5626 }
5627 }
5628 #endif
5629
5630 bool Compile::should_print_igv(const int level) {
5631 #ifndef PRODUCT
5632 if (PrintIdealGraphLevel < 0) { // disabled by the user
5633 return false;
5634 }
5635
5636 bool need = directive()->IGVPrintLevelOption >= level;
5637 if (need) {
5638 Compile::init_igv();
5639 }
5640 return need;
5641 #else
5642 return false;
5643 #endif
5644 }
5645
5646 #ifndef PRODUCT
5647 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5648 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5649
5650 // Called from debugger. Prints method to the default file with the default phase name.
5651 // This works regardless of any Ideal Graph Visualizer flags set or not.
5652 void igv_print() {
5653 Compile::current()->igv_print_method_to_file();
5654 }
5655
5656 // Same as igv_print() above but with a specified phase name.
5657 void igv_print(const char* phase_name) {
5658 Compile::current()->igv_print_method_to_file(phase_name);
5659 }
5660
5661 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5662 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5663 // This works regardless of any Ideal Graph Visualizer flags set or not.
5664 void igv_print(bool network) {
5665 if (network) {
5666 Compile::current()->igv_print_method_to_network();
5667 } else {
5668 Compile::current()->igv_print_method_to_file();
5669 }
5670 }
5671
5672 // Same as igv_print(bool network) above but with a specified phase name.
5673 void igv_print(bool network, const char* phase_name) {
5674 if (network) {
5675 Compile::current()->igv_print_method_to_network(phase_name);
5676 } else {
5677 Compile::current()->igv_print_method_to_file(phase_name);
5678 }
5679 }
5680
5681 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5682 void igv_print_default() {
5683 Compile::current()->print_method(PHASE_DEBUG, 0);
5684 }
5685
5686 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5687 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5688 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5689 void igv_append() {
5690 Compile::current()->igv_print_method_to_file("Debug", true);
5691 }
5692
5693 // Same as igv_append() above but with a specified phase name.
5694 void igv_append(const char* phase_name) {
5695 Compile::current()->igv_print_method_to_file(phase_name, true);
5696 }
5697
5698 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5699 const char* file_name = "custom_debug.xml";
5700 if (_debug_file_printer == nullptr) {
5701 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5702 } else {
5703 _debug_file_printer->update_compiled_method(C->method());
5704 }
5705 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5706 _debug_file_printer->print_graph(phase_name);
5707 }
5708
5709 void Compile::igv_print_method_to_network(const char* phase_name) {
5710 ResourceMark rm;
5711 GrowableArray<const Node*> empty_list;
5712 igv_print_graph_to_network(phase_name, (Node*) C->root(), empty_list);
5713 }
5714
5715 void Compile::igv_print_graph_to_network(const char* name, Node* node, GrowableArray<const Node*>& visible_nodes) {
5716 if (_debug_network_printer == nullptr) {
5717 _debug_network_printer = new IdealGraphPrinter(C);
5718 } else {
5719 _debug_network_printer->update_compiled_method(C->method());
5720 }
5721 tty->print_cr("Method printed over network stream to IGV");
5722 _debug_network_printer->print(name, C->root(), visible_nodes);
5723 }
5724 #endif
5725
5726 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5727 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5728 return value;
5729 }
5730 Node* result = nullptr;
5731 if (bt == T_BYTE) {
5732 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5733 result = new RShiftINode(result, phase->intcon(24));
5734 } else if (bt == T_BOOLEAN) {
5735 result = new AndINode(value, phase->intcon(0xFF));
5736 } else if (bt == T_CHAR) {
5737 result = new AndINode(value,phase->intcon(0xFFFF));
5738 } else if (bt == T_FLOAT) {
5739 result = new MoveI2FNode(value);
5740 } else {
5741 assert(bt == T_SHORT, "unexpected narrow type");
5742 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5743 result = new RShiftINode(result, phase->intcon(16));
5744 }
5745 if (transform_res) {
5746 result = phase->transform(result);
5747 }
5748 return result;
5749 }
5750
5751 void Compile::record_method_not_compilable_oom() {
5752 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5753 }