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/aotCodeCache.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compilationFailureInfo.hpp"
33 #include "compiler/compilationMemoryStatistic.hpp"
34 #include "compiler/compileBroker.hpp"
35 #include "compiler/compileLog.hpp"
36 #include "compiler/compiler_globals.hpp"
37 #include "compiler/compilerDefinitions.hpp"
38 #include "compiler/compilerOracle.hpp"
39 #include "compiler/disassembler.hpp"
40 #include "compiler/oopMap.hpp"
41 #include "gc/shared/barrierSet.hpp"
42 #include "gc/shared/c2/barrierSetC2.hpp"
43 #include "jfr/jfrEvents.hpp"
44 #include "jvm_io.h"
45 #include "memory/allocation.hpp"
46 #include "memory/arena.hpp"
47 #include "memory/resourceArea.hpp"
48 #include "opto/addnode.hpp"
49 #include "opto/block.hpp"
50 #include "opto/c2compiler.hpp"
51 #include "opto/callGenerator.hpp"
52 #include "opto/callnode.hpp"
53 #include "opto/castnode.hpp"
54 #include "opto/cfgnode.hpp"
55 #include "opto/chaitin.hpp"
56 #include "opto/compile.hpp"
57 #include "opto/connode.hpp"
58 #include "opto/convertnode.hpp"
59 #include "opto/divnode.hpp"
60 #include "opto/escape.hpp"
61 #include "opto/idealGraphPrinter.hpp"
62 #include "opto/locknode.hpp"
63 #include "opto/loopnode.hpp"
64 #include "opto/machnode.hpp"
65 #include "opto/macro.hpp"
66 #include "opto/matcher.hpp"
67 #include "opto/mathexactnode.hpp"
68 #include "opto/memnode.hpp"
69 #include "opto/mulnode.hpp"
70 #include "opto/narrowptrnode.hpp"
71 #include "opto/node.hpp"
72 #include "opto/opaquenode.hpp"
73 #include "opto/opcodes.hpp"
74 #include "opto/output.hpp"
75 #include "opto/parse.hpp"
76 #include "opto/phaseX.hpp"
77 #include "opto/rootnode.hpp"
78 #include "opto/runtime.hpp"
79 #include "opto/stringopts.hpp"
80 #include "opto/type.hpp"
81 #include "opto/vector.hpp"
82 #include "opto/vectornode.hpp"
83 #include "runtime/globals_extension.hpp"
84 #include "runtime/sharedRuntime.hpp"
85 #include "runtime/signature.hpp"
86 #include "runtime/stubRoutines.hpp"
87 #include "runtime/timer.hpp"
88 #include "utilities/align.hpp"
89 #include "utilities/copy.hpp"
90 #include "utilities/hashTable.hpp"
91 #include "utilities/macros.hpp"
92
93 // -------------------- Compile::mach_constant_base_node -----------------------
94 // Constant table base node singleton.
95 MachConstantBaseNode* Compile::mach_constant_base_node() {
96 if (_mach_constant_base_node == nullptr) {
97 _mach_constant_base_node = new MachConstantBaseNode();
98 _mach_constant_base_node->add_req(C->root());
99 }
100 return _mach_constant_base_node;
101 }
102
103
104 /// Support for intrinsics.
105
106 // Return the index at which m must be inserted (or already exists).
107 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
108 class IntrinsicDescPair {
109 private:
110 ciMethod* _m;
111 bool _is_virtual;
112 public:
113 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
114 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
115 ciMethod* m= elt->method();
116 ciMethod* key_m = key->_m;
117 if (key_m < m) return -1;
118 else if (key_m > m) return 1;
119 else {
120 bool is_virtual = elt->is_virtual();
121 bool key_virtual = key->_is_virtual;
122 if (key_virtual < is_virtual) return -1;
123 else if (key_virtual > is_virtual) return 1;
124 else return 0;
125 }
126 }
127 };
128 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
129 #ifdef ASSERT
130 for (int i = 1; i < _intrinsics.length(); i++) {
131 CallGenerator* cg1 = _intrinsics.at(i-1);
132 CallGenerator* cg2 = _intrinsics.at(i);
133 assert(cg1->method() != cg2->method()
134 ? cg1->method() < cg2->method()
135 : cg1->is_virtual() < cg2->is_virtual(),
136 "compiler intrinsics list must stay sorted");
137 }
138 #endif
139 IntrinsicDescPair pair(m, is_virtual);
140 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
141 }
142
143 void Compile::register_intrinsic(CallGenerator* cg) {
144 bool found = false;
145 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
146 assert(!found, "registering twice");
147 _intrinsics.insert_before(index, cg);
148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
149 }
150
151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
152 assert(m->is_loaded(), "don't try this on unloaded methods");
153 if (_intrinsics.length() > 0) {
154 bool found = false;
155 int index = intrinsic_insertion_index(m, is_virtual, found);
156 if (found) {
157 return _intrinsics.at(index);
158 }
159 }
160 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
161 if (m->intrinsic_id() != vmIntrinsics::_none &&
162 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
163 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
164 if (cg != nullptr) {
165 // Save it for next time:
166 register_intrinsic(cg);
167 return cg;
168 } else {
169 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
170 }
171 }
172 return nullptr;
173 }
174
175 // Compile::make_vm_intrinsic is defined in library_call.cpp.
176
177 #ifndef PRODUCT
178 // statistics gathering...
179
180 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
182
183 inline int as_int(vmIntrinsics::ID id) {
184 return vmIntrinsics::as_int(id);
185 }
186
187 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
188 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
189 int oflags = _intrinsic_hist_flags[as_int(id)];
190 assert(flags != 0, "what happened?");
191 if (is_virtual) {
192 flags |= _intrinsic_virtual;
193 }
194 bool changed = (flags != oflags);
195 if ((flags & _intrinsic_worked) != 0) {
196 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
197 if (count == 1) {
198 changed = true; // first time
199 }
200 // increment the overall count also:
201 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
202 }
203 if (changed) {
204 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
205 // Something changed about the intrinsic's virtuality.
206 if ((flags & _intrinsic_virtual) != 0) {
207 // This is the first use of this intrinsic as a virtual call.
208 if (oflags != 0) {
209 // We already saw it as a non-virtual, so note both cases.
210 flags |= _intrinsic_both;
211 }
212 } else if ((oflags & _intrinsic_both) == 0) {
213 // This is the first use of this intrinsic as a non-virtual
214 flags |= _intrinsic_both;
215 }
216 }
217 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
218 }
219 // update the overall flags also:
220 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
221 return changed;
222 }
223
224 static char* format_flags(int flags, char* buf) {
225 buf[0] = 0;
226 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
227 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
228 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
229 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
230 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
231 if (buf[0] == 0) strcat(buf, ",");
232 assert(buf[0] == ',', "must be");
233 return &buf[1];
234 }
235
236 void Compile::print_intrinsic_statistics() {
237 char flagsbuf[100];
238 ttyLocker ttyl;
239 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
240 tty->print_cr("Compiler intrinsic usage:");
241 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
242 if (total == 0) total = 1; // avoid div0 in case of no successes
243 #define PRINT_STAT_LINE(name, c, f) \
244 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
245 for (auto id : EnumRange<vmIntrinsicID>{}) {
246 int flags = _intrinsic_hist_flags[as_int(id)];
247 juint count = _intrinsic_hist_count[as_int(id)];
248 if ((flags | count) != 0) {
249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
250 }
251 }
252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
253 if (xtty != nullptr) xtty->tail("statistics");
254 }
255
256 void Compile::print_statistics() {
257 { ttyLocker ttyl;
258 if (xtty != nullptr) xtty->head("statistics type='opto'");
259 Parse::print_statistics();
260 PhaseStringOpts::print_statistics();
261 PhaseCCP::print_statistics();
262 PhaseRegAlloc::print_statistics();
263 PhaseOutput::print_statistics();
264 PhasePeephole::print_statistics();
265 PhaseIdealLoop::print_statistics();
266 ConnectionGraph::print_statistics();
267 PhaseMacroExpand::print_statistics();
268 if (xtty != nullptr) xtty->tail("statistics");
269 }
270 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
271 // put this under its own <statistics> element.
272 print_intrinsic_statistics();
273 }
274 }
275 #endif //PRODUCT
276
277 void Compile::gvn_replace_by(Node* n, Node* nn) {
278 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
279 Node* use = n->last_out(i);
280 bool is_in_table = initial_gvn()->hash_delete(use);
281 uint uses_found = 0;
282 for (uint j = 0; j < use->len(); j++) {
283 if (use->in(j) == n) {
284 if (j < use->req())
285 use->set_req(j, nn);
286 else
287 use->set_prec(j, nn);
288 uses_found++;
289 }
290 }
291 if (is_in_table) {
292 // reinsert into table
293 initial_gvn()->hash_find_insert(use);
294 }
295 record_for_igvn(use);
296 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
297 i -= uses_found; // we deleted 1 or more copies of this edge
298 }
299 }
300
301
302 // Identify all nodes that are reachable from below, useful.
303 // Use breadth-first pass that records state in a Unique_Node_List,
304 // recursive traversal is slower.
305 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
306 int estimated_worklist_size = live_nodes();
307 useful.map( estimated_worklist_size, nullptr ); // preallocate space
308
309 // Initialize worklist
310 if (root() != nullptr) { useful.push(root()); }
311 // If 'top' is cached, declare it useful to preserve cached node
312 if (cached_top_node()) { useful.push(cached_top_node()); }
313
314 // Push all useful nodes onto the list, breadthfirst
315 for( uint next = 0; next < useful.size(); ++next ) {
316 assert( next < unique(), "Unique useful nodes < total nodes");
317 Node *n = useful.at(next);
318 uint max = n->len();
319 for( uint i = 0; i < max; ++i ) {
320 Node *m = n->in(i);
321 if (not_a_node(m)) continue;
322 useful.push(m);
323 }
324 }
325 }
326
327 // Update dead_node_list with any missing dead nodes using useful
328 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
329 void Compile::update_dead_node_list(Unique_Node_List &useful) {
330 uint max_idx = unique();
331 VectorSet& useful_node_set = useful.member_set();
332
333 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
334 // If node with index node_idx is not in useful set,
335 // mark it as dead in dead node list.
336 if (!useful_node_set.test(node_idx)) {
337 record_dead_node(node_idx);
338 }
339 }
340 }
341
342 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
343 int shift = 0;
344 for (int i = 0; i < inlines->length(); i++) {
345 CallGenerator* cg = inlines->at(i);
346 if (useful.member(cg->call_node())) {
347 if (shift > 0) {
348 inlines->at_put(i - shift, cg);
349 }
350 } else {
351 shift++; // skip over the dead element
352 }
353 }
354 if (shift > 0) {
355 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
356 }
357 }
358
359 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
360 assert(dead != nullptr && dead->is_Call(), "sanity");
361 int found = 0;
362 for (int i = 0; i < inlines->length(); i++) {
363 if (inlines->at(i)->call_node() == dead) {
364 inlines->remove_at(i);
365 found++;
366 NOT_DEBUG( break; ) // elements are unique, so exit early
367 }
368 }
369 assert(found <= 1, "not unique");
370 }
371
372 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
373 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
374 for (int i = node_list.length() - 1; i >= 0; i--) {
375 N* node = node_list.at(i);
376 if (!useful.member(node)) {
377 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
378 }
379 }
380 }
381
382 void Compile::remove_useless_node(Node* dead) {
383 remove_modified_node(dead);
384
385 // Constant node that has no out-edges and has only one in-edge from
386 // root is usually dead. However, sometimes reshaping walk makes
387 // it reachable by adding use edges. So, we will NOT count Con nodes
388 // as dead to be conservative about the dead node count at any
389 // given time.
390 if (!dead->is_Con()) {
391 record_dead_node(dead->_idx);
392 }
393 if (dead->is_macro()) {
394 remove_macro_node(dead);
395 }
396 if (dead->is_expensive()) {
397 remove_expensive_node(dead);
398 }
399 if (dead->is_OpaqueTemplateAssertionPredicate()) {
400 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
401 }
402 if (dead->is_ParsePredicate()) {
403 remove_parse_predicate(dead->as_ParsePredicate());
404 }
405 if (dead->for_post_loop_opts_igvn()) {
406 remove_from_post_loop_opts_igvn(dead);
407 }
408 if (dead->for_merge_stores_igvn()) {
409 remove_from_merge_stores_igvn(dead);
410 }
411 if (dead->is_Call()) {
412 remove_useless_late_inlines( &_late_inlines, dead);
413 remove_useless_late_inlines( &_string_late_inlines, dead);
414 remove_useless_late_inlines( &_boxing_late_inlines, dead);
415 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
416
417 if (dead->is_CallStaticJava()) {
418 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
419 }
420 }
421 }
422
423 // Disconnect all useless nodes by disconnecting those at the boundary.
424 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
425 uint next = 0;
426 while (next < useful.size()) {
427 Node *n = useful.at(next++);
428 if (n->is_SafePoint()) {
429 // We're done with a parsing phase. Replaced nodes are not valid
430 // beyond that point.
431 n->as_SafePoint()->delete_replaced_nodes();
432 }
433 // Use raw traversal of out edges since this code removes out edges
434 int max = n->outcnt();
435 for (int j = 0; j < max; ++j) {
436 Node* child = n->raw_out(j);
437 if (!useful.member(child)) {
438 assert(!child->is_top() || child != top(),
439 "If top is cached in Compile object it is in useful list");
440 // Only need to remove this out-edge to the useless node
441 n->raw_del_out(j);
442 --j;
443 --max;
444 if (child->is_data_proj_of_pure_function(n)) {
445 worklist.push(n);
446 }
447 }
448 }
449 if (n->outcnt() == 1 && n->has_special_unique_user()) {
450 assert(useful.member(n->unique_out()), "do not push a useless node");
451 worklist.push(n->unique_out());
452 }
453 }
454
455 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
456 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
457 // Remove useless Template Assertion Predicate opaque nodes
458 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
459 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
460 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
461 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
462 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
463 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
464 #ifdef ASSERT
465 if (_modified_nodes != nullptr) {
466 _modified_nodes->remove_useless_nodes(useful.member_set());
467 }
468 #endif
469
470 // clean up the late inline lists
471 remove_useless_late_inlines( &_late_inlines, useful);
472 remove_useless_late_inlines( &_string_late_inlines, useful);
473 remove_useless_late_inlines( &_boxing_late_inlines, useful);
474 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
475 DEBUG_ONLY(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
476 }
477
478 // ============================================================================
479 //------------------------------CompileWrapper---------------------------------
480 class CompileWrapper : public StackObj {
481 Compile *const _compile;
482 public:
483 CompileWrapper(Compile* compile);
484
485 ~CompileWrapper();
486 };
487
488 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
489 // the Compile* pointer is stored in the current ciEnv:
490 ciEnv* env = compile->env();
491 assert(env == ciEnv::current(), "must already be a ciEnv active");
492 assert(env->compiler_data() == nullptr, "compile already active?");
493 env->set_compiler_data(compile);
494 assert(compile == Compile::current(), "sanity");
495
496 compile->set_type_dict(nullptr);
497 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
498 compile->clone_map().set_clone_idx(0);
499 compile->set_type_last_size(0);
500 compile->set_last_tf(nullptr, nullptr);
501 compile->set_indexSet_arena(nullptr);
502 compile->set_indexSet_free_block_list(nullptr);
503 compile->init_type_arena();
504 Type::Initialize(compile);
505 _compile->begin_method();
506 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
507 }
508 CompileWrapper::~CompileWrapper() {
509 // simulate crash during compilation
510 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
511
512 _compile->end_method();
513 _compile->env()->set_compiler_data(nullptr);
514 }
515
516
517 //----------------------------print_compile_messages---------------------------
518 void Compile::print_compile_messages() {
519 #ifndef PRODUCT
520 // Check if recompiling
521 if (!subsume_loads() && PrintOpto) {
522 // Recompiling without allowing machine instructions to subsume loads
523 tty->print_cr("*********************************************************");
524 tty->print_cr("** Bailout: Recompile without subsuming loads **");
525 tty->print_cr("*********************************************************");
526 }
527 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
528 // Recompiling without escape analysis
529 tty->print_cr("*********************************************************");
530 tty->print_cr("** Bailout: Recompile without escape analysis **");
531 tty->print_cr("*********************************************************");
532 }
533 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
534 // Recompiling without iterative escape analysis
535 tty->print_cr("*********************************************************");
536 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
537 tty->print_cr("*********************************************************");
538 }
539 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
540 // Recompiling without reducing allocation merges
541 tty->print_cr("*********************************************************");
542 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
543 tty->print_cr("*********************************************************");
544 }
545 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
546 // Recompiling without boxing elimination
547 tty->print_cr("*********************************************************");
548 tty->print_cr("** Bailout: Recompile without boxing elimination **");
549 tty->print_cr("*********************************************************");
550 }
551 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
552 // Recompiling without locks coarsening
553 tty->print_cr("*********************************************************");
554 tty->print_cr("** Bailout: Recompile without locks coarsening **");
555 tty->print_cr("*********************************************************");
556 }
557 if (env()->break_at_compile()) {
558 // Open the debugger when compiling this method.
559 tty->print("### Breaking when compiling: ");
560 method()->print_short_name();
561 tty->cr();
562 BREAKPOINT;
563 }
564
565 if( PrintOpto ) {
566 if (is_osr_compilation()) {
567 tty->print("[OSR]%3d", _compile_id);
568 } else {
569 tty->print("%3d", _compile_id);
570 }
571 }
572 #endif
573 }
574
575 #ifndef PRODUCT
576 void Compile::print_phase(const char* phase_name) {
577 tty->print_cr("%u.\t%s", ++_phase_counter, phase_name);
578 }
579
580 void Compile::print_ideal_ir(const char* phase_name) {
581 // keep the following output all in one block
582 // This output goes directly to the tty, not the compiler log.
583 // To enable tools to match it up with the compilation activity,
584 // be sure to tag this tty output with the compile ID.
585
586 // Node dumping can cause a safepoint, which can break the tty lock.
587 // Buffer all node dumps, so that all safepoints happen before we lock.
588 ResourceMark rm;
589 stringStream ss;
590
591 if (_output == nullptr) {
592 ss.print_cr("AFTER: %s", phase_name);
593 // Print out all nodes in ascending order of index.
594 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
595 } else {
596 // Dump the node blockwise if we have a scheduling
597 _output->print_scheduling(&ss);
598 }
599
600 // Check that the lock is not broken by a safepoint.
601 NoSafepointVerifier nsv;
602 ttyLocker ttyl;
603 if (xtty != nullptr) {
604 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
605 compile_id(),
606 is_osr_compilation() ? " compile_kind='osr'" : "",
607 phase_name);
608 }
609
610 tty->print("%s", ss.as_string());
611
612 if (xtty != nullptr) {
613 xtty->tail("ideal");
614 }
615 }
616 #endif
617
618 // ============================================================================
619 //------------------------------Compile standard-------------------------------
620
621 // Compile a method. entry_bci is -1 for normal compilations and indicates
622 // the continuation bci for on stack replacement.
623
624
625 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
626 Options options, DirectiveSet* directive)
627 : Phase(Compiler),
628 _compile_id(ci_env->compile_id()),
629 _options(options),
630 _method(target),
631 _entry_bci(osr_bci),
632 _ilt(nullptr),
633 _stub_function(nullptr),
634 _stub_name(nullptr),
635 _stub_id(StubId::NO_STUBID),
636 _stub_entry_point(nullptr),
637 _max_node_limit(MaxNodeLimit),
638 _post_loop_opts_phase(false),
639 _merge_stores_phase(false),
640 _allow_macro_nodes(true),
641 _inlining_progress(false),
642 _inlining_incrementally(false),
643 _do_cleanup(false),
644 _has_reserved_stack_access(target->has_reserved_stack_access()),
645 #ifndef PRODUCT
646 _igv_idx(0),
647 _trace_opto_output(directive->TraceOptoOutputOption),
648 #endif
649 _clinit_barrier_on_entry(false),
650 _stress_seed(0),
651 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
652 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
653 _env(ci_env),
654 _directive(directive),
655 _log(ci_env->log()),
656 _first_failure_details(nullptr),
657 _intrinsics(comp_arena(), 0, 0, nullptr),
658 _macro_nodes(comp_arena(), 8, 0, nullptr),
659 _parse_predicates(comp_arena(), 8, 0, nullptr),
660 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
661 _expensive_nodes(comp_arena(), 8, 0, nullptr),
662 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
663 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
664 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
665 _coarsened_locks(comp_arena(), 8, 0, nullptr),
666 _congraph(nullptr),
667 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
668 _unique(0),
669 _dead_node_count(0),
670 _dead_node_list(comp_arena()),
671 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
672 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
673 _node_arena(&_node_arena_one),
674 _mach_constant_base_node(nullptr),
675 _Compile_types(mtCompiler, Arena::Tag::tag_type),
676 _initial_gvn(nullptr),
677 _igvn_worklist(nullptr),
678 _types(nullptr),
679 _node_hash(nullptr),
680 _late_inlines(comp_arena(), 2, 0, nullptr),
681 _string_late_inlines(comp_arena(), 2, 0, nullptr),
682 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
683 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
684 _late_inlines_pos(0),
685 _number_of_mh_late_inlines(0),
686 _oom(false),
687 _replay_inline_data(nullptr),
688 _inline_printer(this),
689 _java_calls(0),
690 _inner_loops(0),
691 _FIRST_STACK_mask(comp_arena()),
692 _interpreter_frame_size(0),
693 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
694 _output(nullptr)
695 #ifndef PRODUCT
696 ,
697 _in_dump_cnt(0)
698 #endif
699 {
700 C = this;
701 CompileWrapper cw(this);
702
703 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
704 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
705
706 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
707 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
708 // We can always print a disassembly, either abstract (hex dump) or
709 // with the help of a suitable hsdis library. Thus, we should not
710 // couple print_assembly and print_opto_assembly controls.
711 // But: always print opto and regular assembly on compile command 'print'.
712 bool print_assembly = directive->PrintAssemblyOption;
713 set_print_assembly(print_opto_assembly || print_assembly);
714 #else
715 set_print_assembly(false); // must initialize.
716 #endif
717
718 #ifndef PRODUCT
719 set_parsed_irreducible_loop(false);
720 #endif
721
722 if (directive->ReplayInlineOption) {
723 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
724 }
725 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
726 set_print_intrinsics(directive->PrintIntrinsicsOption);
727 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
728
729 if (ProfileTraps) {
730 // Make sure the method being compiled gets its own MDO,
731 // so we can at least track the decompile_count().
732 method()->ensure_method_data();
733 }
734
735 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
736 StressIncrementalInlining || StressMacroExpansion ||
737 StressMacroElimination || StressUnstableIfTraps ||
738 StressBailout || StressLoopPeeling) {
739 initialize_stress_seed(directive);
740 }
741
742 Init(/*do_aliasing=*/ true);
743
744 print_compile_messages();
745
746 _ilt = InlineTree::build_inline_tree_root();
747
748 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
749 assert(num_alias_types() >= AliasIdxRaw, "");
750
751 #define MINIMUM_NODE_HASH 1023
752
753 // GVN that will be run immediately on new nodes
754 uint estimated_size = method()->code_size()*4+64;
755 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
756 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
757 _types = new (comp_arena()) Type_Array(comp_arena());
758 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
759 PhaseGVN gvn;
760 set_initial_gvn(&gvn);
761
762 { // Scope for timing the parser
763 TracePhase tp(_t_parser);
764
765 // Put top into the hash table ASAP.
766 initial_gvn()->transform(top());
767
768 // Set up tf(), start(), and find a CallGenerator.
769 CallGenerator* cg = nullptr;
770 if (is_osr_compilation()) {
771 const TypeTuple *domain = StartOSRNode::osr_domain();
772 const TypeTuple *range = TypeTuple::make_range(method()->signature());
773 init_tf(TypeFunc::make(domain, range));
774 StartNode* s = new StartOSRNode(root(), domain);
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());
782 initial_gvn()->set_type_bottom(s);
783 verify_start(s);
784 float past_uses = method()->interpreter_invocation_count();
785 float expected_uses = past_uses;
786 cg = CallGenerator::for_inline(method(), expected_uses);
787 }
788 if (failing()) return;
789 if (cg == nullptr) {
790 const char* reason = InlineTree::check_can_parse(method());
791 assert(reason != nullptr, "expect reason for parse failure");
792 stringStream ss;
793 ss.print("cannot parse method: %s", reason);
794 record_method_not_compilable(ss.as_string());
795 return;
796 }
797
798 gvn.set_type(root(), root()->bottom_type());
799
800 JVMState* jvms = build_start_state(start(), tf());
801 if ((jvms = cg->generate(jvms)) == nullptr) {
802 assert(failure_reason() != nullptr, "expect reason for parse failure");
803 stringStream ss;
804 ss.print("method parse failed: %s", failure_reason());
805 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
806 return;
807 }
808 GraphKit kit(jvms);
809
810 if (!kit.stopped()) {
811 // Accept return values, and transfer control we know not where.
812 // This is done by a special, unique ReturnNode bound to root.
813 return_values(kit.jvms());
814 }
815
816 if (kit.has_exceptions()) {
817 // Any exceptions that escape from this call must be rethrown
818 // to whatever caller is dynamically above us on the stack.
819 // This is done by a special, unique RethrowNode bound to root.
820 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
821 }
822
823 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
824
825 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
826 inline_string_calls(true);
827 }
828
829 if (failing()) return;
830
831 // Remove clutter produced by parsing.
832 if (!failing()) {
833 ResourceMark rm;
834 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
835 }
836 }
837
838 // Note: Large methods are capped off in do_one_bytecode().
839 if (failing()) return;
840
841 // After parsing, node notes are no longer automagic.
842 // They must be propagated by register_new_node_with_optimizer(),
843 // clone(), or the like.
844 set_default_node_notes(nullptr);
845
846 #ifndef PRODUCT
847 if (should_print_igv(1)) {
848 _igv_printer->print_inlining();
849 }
850 #endif
851
852 if (failing()) return;
853 NOT_PRODUCT( verify_graph_edges(); )
854
855 // Now optimize
856 Optimize();
857 if (failing()) return;
858 NOT_PRODUCT( verify_graph_edges(); )
859
860 #ifndef PRODUCT
861 if (should_print_ideal()) {
862 print_ideal_ir("print_ideal");
863 }
864 #endif
865
866 #ifdef ASSERT
867 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
868 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
869 #endif
870
871 // Dump compilation data to replay it.
872 if (directive->DumpReplayOption) {
873 env()->dump_replay_data(_compile_id);
874 }
875 if (directive->DumpInlineOption && (ilt() != nullptr)) {
876 env()->dump_inline_data(_compile_id);
877 }
878
879 // Now that we know the size of all the monitors we can add a fixed slot
880 // for the original deopt pc.
881 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
882 set_fixed_slots(next_slot);
883
884 // Compute when to use implicit null checks. Used by matching trap based
885 // nodes and NullCheck optimization.
886 set_allowed_deopt_reasons();
887
888 // Now generate code
889 Code_Gen();
890 }
891
892 //------------------------------Compile----------------------------------------
893 // Compile a runtime stub
894 Compile::Compile(ciEnv* ci_env,
895 TypeFunc_generator generator,
896 address stub_function,
897 const char* stub_name,
898 StubId stub_id,
899 int is_fancy_jump,
900 bool pass_tls,
901 bool return_pc,
902 DirectiveSet* directive)
903 : Phase(Compiler),
904 _compile_id(0),
905 _options(Options::for_runtime_stub()),
906 _method(nullptr),
907 _entry_bci(InvocationEntryBci),
908 _stub_function(stub_function),
909 _stub_name(stub_name),
910 _stub_id(stub_id),
911 _stub_entry_point(nullptr),
912 _max_node_limit(MaxNodeLimit),
913 _post_loop_opts_phase(false),
914 _merge_stores_phase(false),
915 _allow_macro_nodes(true),
916 _inlining_progress(false),
917 _inlining_incrementally(false),
918 _has_reserved_stack_access(false),
919 #ifndef PRODUCT
920 _igv_idx(0),
921 _trace_opto_output(directive->TraceOptoOutputOption),
922 #endif
923 _clinit_barrier_on_entry(false),
924 _stress_seed(0),
925 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
926 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
927 _env(ci_env),
928 _directive(directive),
929 _log(ci_env->log()),
930 _first_failure_details(nullptr),
931 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
932 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
933 _congraph(nullptr),
934 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
935 _unique(0),
936 _dead_node_count(0),
937 _dead_node_list(comp_arena()),
938 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
939 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
940 _node_arena(&_node_arena_one),
941 _mach_constant_base_node(nullptr),
942 _Compile_types(mtCompiler, Arena::Tag::tag_type),
943 _initial_gvn(nullptr),
944 _igvn_worklist(nullptr),
945 _types(nullptr),
946 _node_hash(nullptr),
947 _number_of_mh_late_inlines(0),
948 _oom(false),
949 _replay_inline_data(nullptr),
950 _inline_printer(this),
951 _java_calls(0),
952 _inner_loops(0),
953 _FIRST_STACK_mask(comp_arena()),
954 _interpreter_frame_size(0),
955 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
956 _output(nullptr),
957 #ifndef PRODUCT
958 _in_dump_cnt(0),
959 #endif
960 _allowed_reasons(0) {
961 C = this;
962
963 // try to reuse an existing stub
964 {
965 BlobId blob_id = StubInfo::blob(_stub_id);
966 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
967 if (blob != nullptr) {
968 RuntimeStub* rs = blob->as_runtime_stub();
969 _stub_entry_point = rs->entry_point();
970 return;
971 }
972 }
973
974 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
975 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
976
977 #ifndef PRODUCT
978 set_print_assembly(PrintFrameConverterAssembly);
979 set_parsed_irreducible_loop(false);
980 #else
981 set_print_assembly(false); // Must initialize.
982 #endif
983 set_has_irreducible_loop(false); // no loops
984
985 CompileWrapper cw(this);
986 Init(/*do_aliasing=*/ false);
987 init_tf((*generator)());
988
989 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
990 _types = new (comp_arena()) Type_Array(comp_arena());
991 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
992
993 if (StressLCM || StressGCM || StressBailout) {
994 initialize_stress_seed(directive);
995 }
996
997 {
998 PhaseGVN gvn;
999 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1000 gvn.transform(top());
1001
1002 GraphKit kit;
1003 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1004 }
1005
1006 NOT_PRODUCT( verify_graph_edges(); )
1007
1008 Code_Gen();
1009 }
1010
1011 Compile::~Compile() {
1012 delete _first_failure_details;
1013 };
1014
1015 //------------------------------Init-------------------------------------------
1016 // Prepare for a single compilation
1017 void Compile::Init(bool aliasing) {
1018 _do_aliasing = aliasing;
1019 _unique = 0;
1020 _regalloc = nullptr;
1021
1022 _tf = nullptr; // filled in later
1023 _top = nullptr; // cached later
1024 _matcher = nullptr; // filled in later
1025 _cfg = nullptr; // filled in later
1026
1027 _node_note_array = nullptr;
1028 _default_node_notes = nullptr;
1029 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1030
1031 _immutable_memory = nullptr; // filled in at first inquiry
1032
1033 #ifdef ASSERT
1034 _phase_optimize_finished = false;
1035 _phase_verify_ideal_loop = false;
1036 _exception_backedge = false;
1037 _type_verify = nullptr;
1038 #endif
1039
1040 // Globally visible Nodes
1041 // First set TOP to null to give safe behavior during creation of RootNode
1042 set_cached_top_node(nullptr);
1043 set_root(new RootNode());
1044 // Now that you have a Root to point to, create the real TOP
1045 set_cached_top_node( new ConNode(Type::TOP) );
1046 set_recent_alloc(nullptr, nullptr);
1047
1048 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1049 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1050 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1051 env()->set_dependencies(new Dependencies(env()));
1052
1053 _fixed_slots = 0;
1054 set_has_split_ifs(false);
1055 set_has_loops(false); // first approximation
1056 set_has_stringbuilder(false);
1057 set_has_boxed_value(false);
1058 _trap_can_recompile = false; // no traps emitted yet
1059 _major_progress = true; // start out assuming good things will happen
1060 set_has_unsafe_access(false);
1061 set_max_vector_size(0);
1062 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1063 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1064 set_decompile_count(0);
1065
1066 #ifndef PRODUCT
1067 _phase_counter = 0;
1068 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1069 #endif
1070
1071 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1072 _loop_opts_cnt = LoopOptsCount;
1073 set_do_inlining(Inline);
1074 set_max_inline_size(MaxInlineSize);
1075 set_freq_inline_size(FreqInlineSize);
1076 set_do_scheduling(OptoScheduling);
1077
1078 set_do_vector_loop(false);
1079 set_has_monitors(false);
1080 set_has_scoped_access(false);
1081
1082 if (AllowVectorizeOnDemand) {
1083 if (has_method() && _directive->VectorizeOption) {
1084 set_do_vector_loop(true);
1085 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());})
1086 } else if (has_method() && method()->name() != nullptr &&
1087 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1088 set_do_vector_loop(true);
1089 }
1090 }
1091 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1092 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());})
1093
1094 _max_node_limit = _directive->MaxNodeLimitOption;
1095
1096 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1097 set_clinit_barrier_on_entry(true);
1098 }
1099 if (debug_info()->recording_non_safepoints()) {
1100 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1101 (comp_arena(), 8, 0, nullptr));
1102 set_default_node_notes(Node_Notes::make(this));
1103 }
1104
1105 const int grow_ats = 16;
1106 _max_alias_types = grow_ats;
1107 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1108 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1109 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1110 {
1111 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1112 }
1113 // Initialize the first few types.
1114 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1115 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1116 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1117 _num_alias_types = AliasIdxRaw+1;
1118 // Zero out the alias type cache.
1119 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1120 // A null adr_type hits in the cache right away. Preload the right answer.
1121 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1122 }
1123
1124 #ifdef ASSERT
1125 // Verify that the current StartNode is valid.
1126 void Compile::verify_start(StartNode* s) const {
1127 assert(failing_internal() || s == start(), "should be StartNode");
1128 }
1129 #endif
1130
1131 /**
1132 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1133 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1134 * the ideal graph.
1135 */
1136 StartNode* Compile::start() const {
1137 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1138 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1139 Node* start = root()->fast_out(i);
1140 if (start->is_Start()) {
1141 return start->as_Start();
1142 }
1143 }
1144 fatal("Did not find Start node!");
1145 return nullptr;
1146 }
1147
1148 //-------------------------------immutable_memory-------------------------------------
1149 // Access immutable memory
1150 Node* Compile::immutable_memory() {
1151 if (_immutable_memory != nullptr) {
1152 return _immutable_memory;
1153 }
1154 StartNode* s = start();
1155 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1156 Node *p = s->fast_out(i);
1157 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1158 _immutable_memory = p;
1159 return _immutable_memory;
1160 }
1161 }
1162 ShouldNotReachHere();
1163 return nullptr;
1164 }
1165
1166 //----------------------set_cached_top_node------------------------------------
1167 // Install the cached top node, and make sure Node::is_top works correctly.
1168 void Compile::set_cached_top_node(Node* tn) {
1169 if (tn != nullptr) verify_top(tn);
1170 Node* old_top = _top;
1171 _top = tn;
1172 // Calling Node::setup_is_top allows the nodes the chance to adjust
1173 // their _out arrays.
1174 if (_top != nullptr) _top->setup_is_top();
1175 if (old_top != nullptr) old_top->setup_is_top();
1176 assert(_top == nullptr || top()->is_top(), "");
1177 }
1178
1179 #ifdef ASSERT
1180 uint Compile::count_live_nodes_by_graph_walk() {
1181 Unique_Node_List useful(comp_arena());
1182 // Get useful node list by walking the graph.
1183 identify_useful_nodes(useful);
1184 return useful.size();
1185 }
1186
1187 void Compile::print_missing_nodes() {
1188
1189 // Return if CompileLog is null and PrintIdealNodeCount is false.
1190 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1191 return;
1192 }
1193
1194 // This is an expensive function. It is executed only when the user
1195 // specifies VerifyIdealNodeCount option or otherwise knows the
1196 // additional work that needs to be done to identify reachable nodes
1197 // by walking the flow graph and find the missing ones using
1198 // _dead_node_list.
1199
1200 Unique_Node_List useful(comp_arena());
1201 // Get useful node list by walking the graph.
1202 identify_useful_nodes(useful);
1203
1204 uint l_nodes = C->live_nodes();
1205 uint l_nodes_by_walk = useful.size();
1206
1207 if (l_nodes != l_nodes_by_walk) {
1208 if (_log != nullptr) {
1209 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1210 _log->stamp();
1211 _log->end_head();
1212 }
1213 VectorSet& useful_member_set = useful.member_set();
1214 int last_idx = l_nodes_by_walk;
1215 for (int i = 0; i < last_idx; i++) {
1216 if (useful_member_set.test(i)) {
1217 if (_dead_node_list.test(i)) {
1218 if (_log != nullptr) {
1219 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1220 }
1221 if (PrintIdealNodeCount) {
1222 // Print the log message to tty
1223 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1224 useful.at(i)->dump();
1225 }
1226 }
1227 }
1228 else if (! _dead_node_list.test(i)) {
1229 if (_log != nullptr) {
1230 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1231 }
1232 if (PrintIdealNodeCount) {
1233 // Print the log message to tty
1234 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1235 }
1236 }
1237 }
1238 if (_log != nullptr) {
1239 _log->tail("mismatched_nodes");
1240 }
1241 }
1242 }
1243 void Compile::record_modified_node(Node* n) {
1244 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1245 _modified_nodes->push(n);
1246 }
1247 }
1248
1249 void Compile::remove_modified_node(Node* n) {
1250 if (_modified_nodes != nullptr) {
1251 _modified_nodes->remove(n);
1252 }
1253 }
1254 #endif
1255
1256 #ifndef PRODUCT
1257 void Compile::verify_top(Node* tn) const {
1258 if (tn != nullptr) {
1259 assert(tn->is_Con(), "top node must be a constant");
1260 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1261 assert(tn->in(0) != nullptr, "must have live top node");
1262 }
1263 }
1264 #endif
1265
1266
1267 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1268
1269 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1270 guarantee(arr != nullptr, "");
1271 int num_blocks = arr->length();
1272 if (grow_by < num_blocks) grow_by = num_blocks;
1273 int num_notes = grow_by * _node_notes_block_size;
1274 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1275 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1276 while (num_notes > 0) {
1277 arr->append(notes);
1278 notes += _node_notes_block_size;
1279 num_notes -= _node_notes_block_size;
1280 }
1281 assert(num_notes == 0, "exact multiple, please");
1282 }
1283
1284 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1285 if (source == nullptr || dest == nullptr) return false;
1286
1287 if (dest->is_Con())
1288 return false; // Do not push debug info onto constants.
1289
1290 #ifdef ASSERT
1291 // Leave a bread crumb trail pointing to the original node:
1292 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1293 dest->set_debug_orig(source);
1294 }
1295 #endif
1296
1297 if (node_note_array() == nullptr)
1298 return false; // Not collecting any notes now.
1299
1300 // This is a copy onto a pre-existing node, which may already have notes.
1301 // If both nodes have notes, do not overwrite any pre-existing notes.
1302 Node_Notes* source_notes = node_notes_at(source->_idx);
1303 if (source_notes == nullptr || source_notes->is_clear()) return false;
1304 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1305 if (dest_notes == nullptr || dest_notes->is_clear()) {
1306 return set_node_notes_at(dest->_idx, source_notes);
1307 }
1308
1309 Node_Notes merged_notes = (*source_notes);
1310 // The order of operations here ensures that dest notes will win...
1311 merged_notes.update_from(dest_notes);
1312 return set_node_notes_at(dest->_idx, &merged_notes);
1313 }
1314
1315
1316 //--------------------------allow_range_check_smearing-------------------------
1317 // Gating condition for coalescing similar range checks.
1318 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1319 // single covering check that is at least as strong as any of them.
1320 // If the optimization succeeds, the simplified (strengthened) range check
1321 // will always succeed. If it fails, we will deopt, and then give up
1322 // on the optimization.
1323 bool Compile::allow_range_check_smearing() const {
1324 // If this method has already thrown a range-check,
1325 // assume it was because we already tried range smearing
1326 // and it failed.
1327 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1328 return !already_trapped;
1329 }
1330
1331
1332 //------------------------------flatten_alias_type-----------------------------
1333 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1334 assert(do_aliasing(), "Aliasing should be enabled");
1335 int offset = tj->offset();
1336 TypePtr::PTR ptr = tj->ptr();
1337
1338 // Known instance (scalarizable allocation) alias only with itself.
1339 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1340 tj->is_oopptr()->is_known_instance();
1341
1342 // Process weird unsafe references.
1343 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1344 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1345 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1346 tj = TypeOopPtr::BOTTOM;
1347 ptr = tj->ptr();
1348 offset = tj->offset();
1349 }
1350
1351 // Array pointers need some flattening
1352 const TypeAryPtr* ta = tj->isa_aryptr();
1353 if (ta && ta->is_stable()) {
1354 // Erase stability property for alias analysis.
1355 tj = ta = ta->cast_to_stable(false);
1356 }
1357 if( ta && is_known_inst ) {
1358 if ( offset != Type::OffsetBot &&
1359 offset > arrayOopDesc::length_offset_in_bytes() ) {
1360 offset = Type::OffsetBot; // Flatten constant access into array body only
1361 tj = ta = ta->
1362 remove_speculative()->
1363 cast_to_ptr_type(ptr)->
1364 with_offset(offset);
1365 }
1366 } else if (ta) {
1367 // For arrays indexed by constant indices, we flatten the alias
1368 // space to include all of the array body. Only the header, klass
1369 // and array length can be accessed un-aliased.
1370 if( offset != Type::OffsetBot ) {
1371 if( ta->const_oop() ) { // MethodData* or Method*
1372 offset = Type::OffsetBot; // Flatten constant access into array body
1373 tj = ta = ta->
1374 remove_speculative()->
1375 cast_to_ptr_type(ptr)->
1376 cast_to_exactness(false)->
1377 with_offset(offset);
1378 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1379 // range is OK as-is.
1380 tj = ta = TypeAryPtr::RANGE;
1381 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1382 tj = TypeInstPtr::KLASS; // all klass loads look alike
1383 ta = TypeAryPtr::RANGE; // generic ignored junk
1384 ptr = TypePtr::BotPTR;
1385 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1386 tj = TypeInstPtr::MARK;
1387 ta = TypeAryPtr::RANGE; // generic ignored junk
1388 ptr = TypePtr::BotPTR;
1389 } else { // Random constant offset into array body
1390 offset = Type::OffsetBot; // Flatten constant access into array body
1391 tj = ta = ta->
1392 remove_speculative()->
1393 cast_to_ptr_type(ptr)->
1394 cast_to_exactness(false)->
1395 with_offset(offset);
1396 }
1397 }
1398 // Arrays of fixed size alias with arrays of unknown size.
1399 if (ta->size() != TypeInt::POS) {
1400 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1401 tj = ta = ta->
1402 remove_speculative()->
1403 cast_to_ptr_type(ptr)->
1404 with_ary(tary)->
1405 cast_to_exactness(false);
1406 }
1407 // Arrays of known objects become arrays of unknown objects.
1408 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1409 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1410 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1411 }
1412 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1413 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1414 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1415 }
1416 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1417 // cannot be distinguished by bytecode alone.
1418 if (ta->elem() == TypeInt::BOOL) {
1419 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1420 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1421 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1422 }
1423 // During the 2nd round of IterGVN, NotNull castings are removed.
1424 // Make sure the Bottom and NotNull variants alias the same.
1425 // Also, make sure exact and non-exact variants alias the same.
1426 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1427 tj = ta = ta->
1428 remove_speculative()->
1429 cast_to_ptr_type(TypePtr::BotPTR)->
1430 cast_to_exactness(false)->
1431 with_offset(offset);
1432 }
1433 }
1434
1435 // Oop pointers need some flattening
1436 const TypeInstPtr *to = tj->isa_instptr();
1437 if (to && to != TypeOopPtr::BOTTOM) {
1438 ciInstanceKlass* ik = to->instance_klass();
1439 if( ptr == TypePtr::Constant ) {
1440 if (ik != ciEnv::current()->Class_klass() ||
1441 offset < ik->layout_helper_size_in_bytes()) {
1442 // No constant oop pointers (such as Strings); they alias with
1443 // unknown strings.
1444 assert(!is_known_inst, "not scalarizable allocation");
1445 tj = to = to->
1446 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1447 remove_speculative()->
1448 cast_to_ptr_type(TypePtr::BotPTR)->
1449 cast_to_exactness(false);
1450 }
1451 } else if( is_known_inst ) {
1452 tj = to; // Keep NotNull and klass_is_exact for instance type
1453 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1454 // During the 2nd round of IterGVN, NotNull castings are removed.
1455 // Make sure the Bottom and NotNull variants alias the same.
1456 // Also, make sure exact and non-exact variants alias the same.
1457 tj = to = to->
1458 remove_speculative()->
1459 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1460 cast_to_ptr_type(TypePtr::BotPTR)->
1461 cast_to_exactness(false);
1462 }
1463 if (to->speculative() != nullptr) {
1464 tj = to = to->remove_speculative();
1465 }
1466 // Canonicalize the holder of this field
1467 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1468 // First handle header references such as a LoadKlassNode, even if the
1469 // object's klass is unloaded at compile time (4965979).
1470 if (!is_known_inst) { // Do it only for non-instance types
1471 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1472 }
1473 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1474 // Static fields are in the space above the normal instance
1475 // fields in the java.lang.Class instance.
1476 if (ik != ciEnv::current()->Class_klass()) {
1477 to = nullptr;
1478 tj = TypeOopPtr::BOTTOM;
1479 offset = tj->offset();
1480 }
1481 } else {
1482 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1483 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1484 assert(tj->offset() == offset, "no change to offset expected");
1485 bool xk = to->klass_is_exact();
1486 int instance_id = to->instance_id();
1487
1488 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1489 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1490 // its interfaces are included.
1491 if (xk && ik->equals(canonical_holder)) {
1492 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1493 } else {
1494 assert(xk || !is_known_inst, "Known instance should be exact type");
1495 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1496 }
1497 }
1498 }
1499
1500 // Klass pointers to object array klasses need some flattening
1501 const TypeKlassPtr *tk = tj->isa_klassptr();
1502 if( tk ) {
1503 // If we are referencing a field within a Klass, we need
1504 // to assume the worst case of an Object. Both exact and
1505 // inexact types must flatten to the same alias class so
1506 // use NotNull as the PTR.
1507 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1508 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1509 env()->Object_klass(),
1510 offset);
1511 }
1512
1513 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1514 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1515 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1516 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1517 } else {
1518 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1519 }
1520 }
1521
1522 // Check for precise loads from the primary supertype array and force them
1523 // to the supertype cache alias index. Check for generic array loads from
1524 // the primary supertype array and also force them to the supertype cache
1525 // alias index. Since the same load can reach both, we need to merge
1526 // these 2 disparate memories into the same alias class. Since the
1527 // primary supertype array is read-only, there's no chance of confusion
1528 // where we bypass an array load and an array store.
1529 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1530 if (offset == Type::OffsetBot ||
1531 (offset >= primary_supers_offset &&
1532 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1533 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1534 offset = in_bytes(Klass::secondary_super_cache_offset());
1535 tj = tk = tk->with_offset(offset);
1536 }
1537 }
1538
1539 // Flatten all Raw pointers together.
1540 if (tj->base() == Type::RawPtr)
1541 tj = TypeRawPtr::BOTTOM;
1542
1543 if (tj->base() == Type::AnyPtr)
1544 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1545
1546 offset = tj->offset();
1547 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1548
1549 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1550 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1551 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1552 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1553 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1554 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1555 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1556 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1557 assert( tj->ptr() != TypePtr::TopPTR &&
1558 tj->ptr() != TypePtr::AnyNull &&
1559 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1560 // assert( tj->ptr() != TypePtr::Constant ||
1561 // tj->base() == Type::RawPtr ||
1562 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1563
1564 return tj;
1565 }
1566
1567 void Compile::AliasType::Init(int i, const TypePtr* at) {
1568 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1569 _index = i;
1570 _adr_type = at;
1571 _field = nullptr;
1572 _element = nullptr;
1573 _is_rewritable = true; // default
1574 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1575 if (atoop != nullptr && atoop->is_known_instance()) {
1576 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1577 _general_index = Compile::current()->get_alias_index(gt);
1578 } else {
1579 _general_index = 0;
1580 }
1581 }
1582
1583 BasicType Compile::AliasType::basic_type() const {
1584 if (element() != nullptr) {
1585 const Type* element = adr_type()->is_aryptr()->elem();
1586 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1587 } if (field() != nullptr) {
1588 return field()->layout_type();
1589 } else {
1590 return T_ILLEGAL; // unknown
1591 }
1592 }
1593
1594 //---------------------------------print_on------------------------------------
1595 #ifndef PRODUCT
1596 void Compile::AliasType::print_on(outputStream* st) {
1597 if (index() < 10)
1598 st->print("@ <%d> ", index());
1599 else st->print("@ <%d>", index());
1600 st->print(is_rewritable() ? " " : " RO");
1601 int offset = adr_type()->offset();
1602 if (offset == Type::OffsetBot)
1603 st->print(" +any");
1604 else st->print(" +%-3d", offset);
1605 st->print(" in ");
1606 adr_type()->dump_on(st);
1607 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1608 if (field() != nullptr && tjp) {
1609 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1610 tjp->offset() != field()->offset_in_bytes()) {
1611 st->print(" != ");
1612 field()->print();
1613 st->print(" ***");
1614 }
1615 }
1616 }
1617
1618 void print_alias_types() {
1619 Compile* C = Compile::current();
1620 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1621 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1622 C->alias_type(idx)->print_on(tty);
1623 tty->cr();
1624 }
1625 }
1626 #endif
1627
1628
1629 //----------------------------probe_alias_cache--------------------------------
1630 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1631 intptr_t key = (intptr_t) adr_type;
1632 key ^= key >> logAliasCacheSize;
1633 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1634 }
1635
1636
1637 //-----------------------------grow_alias_types--------------------------------
1638 void Compile::grow_alias_types() {
1639 const int old_ats = _max_alias_types; // how many before?
1640 const int new_ats = old_ats; // how many more?
1641 const int grow_ats = old_ats+new_ats; // how many now?
1642 _max_alias_types = grow_ats;
1643 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1644 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1645 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1646 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1647 }
1648
1649
1650 //--------------------------------find_alias_type------------------------------
1651 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1652 if (!do_aliasing()) {
1653 return alias_type(AliasIdxBot);
1654 }
1655
1656 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1657 if (ace->_adr_type == adr_type) {
1658 return alias_type(ace->_index);
1659 }
1660
1661 // Handle special cases.
1662 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1663 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1664
1665 // Do it the slow way.
1666 const TypePtr* flat = flatten_alias_type(adr_type);
1667
1668 #ifdef ASSERT
1669 {
1670 ResourceMark rm;
1671 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1672 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1673 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1674 Type::str(adr_type));
1675 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1676 const TypeOopPtr* foop = flat->is_oopptr();
1677 // Scalarizable allocations have exact klass always.
1678 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1679 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1680 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1681 Type::str(foop), Type::str(xoop));
1682 }
1683 }
1684 #endif
1685
1686 int idx = AliasIdxTop;
1687 for (int i = 0; i < num_alias_types(); i++) {
1688 if (alias_type(i)->adr_type() == flat) {
1689 idx = i;
1690 break;
1691 }
1692 }
1693
1694 if (idx == AliasIdxTop) {
1695 if (no_create) return nullptr;
1696 // Grow the array if necessary.
1697 if (_num_alias_types == _max_alias_types) grow_alias_types();
1698 // Add a new alias type.
1699 idx = _num_alias_types++;
1700 _alias_types[idx]->Init(idx, flat);
1701 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1702 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1703 if (flat->isa_instptr()) {
1704 if (flat->offset() == java_lang_Class::klass_offset()
1705 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1706 alias_type(idx)->set_rewritable(false);
1707 }
1708 if (flat->isa_aryptr()) {
1709 #ifdef ASSERT
1710 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1711 // (T_BYTE has the weakest alignment and size restrictions...)
1712 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1713 #endif
1714 if (flat->offset() == TypePtr::OffsetBot) {
1715 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1716 }
1717 }
1718 if (flat->isa_klassptr()) {
1719 if (UseCompactObjectHeaders) {
1720 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1721 alias_type(idx)->set_rewritable(false);
1722 }
1723 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1724 alias_type(idx)->set_rewritable(false);
1725 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1726 alias_type(idx)->set_rewritable(false);
1727 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1728 alias_type(idx)->set_rewritable(false);
1729 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1730 alias_type(idx)->set_rewritable(false);
1731 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1732 alias_type(idx)->set_rewritable(false);
1733 }
1734 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1735 // but the base pointer type is not distinctive enough to identify
1736 // references into JavaThread.)
1737
1738 // Check for final fields.
1739 const TypeInstPtr* tinst = flat->isa_instptr();
1740 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1741 ciField* field;
1742 if (tinst->const_oop() != nullptr &&
1743 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1744 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1745 // static field
1746 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1747 field = k->get_field_by_offset(tinst->offset(), true);
1748 } else {
1749 ciInstanceKlass *k = tinst->instance_klass();
1750 field = k->get_field_by_offset(tinst->offset(), false);
1751 }
1752 assert(field == nullptr ||
1753 original_field == nullptr ||
1754 (field->holder() == original_field->holder() &&
1755 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1756 field->is_static() == original_field->is_static()), "wrong field?");
1757 // Set field() and is_rewritable() attributes.
1758 if (field != nullptr) alias_type(idx)->set_field(field);
1759 }
1760 }
1761
1762 // Fill the cache for next time.
1763 ace->_adr_type = adr_type;
1764 ace->_index = idx;
1765 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1766
1767 // Might as well try to fill the cache for the flattened version, too.
1768 AliasCacheEntry* face = probe_alias_cache(flat);
1769 if (face->_adr_type == nullptr) {
1770 face->_adr_type = flat;
1771 face->_index = idx;
1772 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1773 }
1774
1775 return alias_type(idx);
1776 }
1777
1778
1779 Compile::AliasType* Compile::alias_type(ciField* field) {
1780 const TypeOopPtr* t;
1781 if (field->is_static())
1782 t = TypeInstPtr::make(field->holder()->java_mirror());
1783 else
1784 t = TypeOopPtr::make_from_klass_raw(field->holder());
1785 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1786 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1787 return atp;
1788 }
1789
1790
1791 //------------------------------have_alias_type--------------------------------
1792 bool Compile::have_alias_type(const TypePtr* adr_type) {
1793 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1794 if (ace->_adr_type == adr_type) {
1795 return true;
1796 }
1797
1798 // Handle special cases.
1799 if (adr_type == nullptr) return true;
1800 if (adr_type == TypePtr::BOTTOM) return true;
1801
1802 return find_alias_type(adr_type, true, nullptr) != nullptr;
1803 }
1804
1805 //-----------------------------must_alias--------------------------------------
1806 // True if all values of the given address type are in the given alias category.
1807 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1808 if (alias_idx == AliasIdxBot) return true; // the universal category
1809 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1810 if (alias_idx == AliasIdxTop) return false; // the empty category
1811 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1812
1813 // the only remaining possible overlap is identity
1814 int adr_idx = get_alias_index(adr_type);
1815 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1816 assert(adr_idx == alias_idx ||
1817 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1818 && adr_type != TypeOopPtr::BOTTOM),
1819 "should not be testing for overlap with an unsafe pointer");
1820 return adr_idx == alias_idx;
1821 }
1822
1823 //------------------------------can_alias--------------------------------------
1824 // True if any values of the given address type are in the given alias category.
1825 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1826 if (alias_idx == AliasIdxTop) return false; // the empty category
1827 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1828 // Known instance doesn't alias with bottom memory
1829 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1830 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1831
1832 // the only remaining possible overlap is identity
1833 int adr_idx = get_alias_index(adr_type);
1834 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1835 return adr_idx == alias_idx;
1836 }
1837
1838 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1839 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1840 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1841 if (parse_predicate_count() == 0) {
1842 return;
1843 }
1844 for (int i = 0; i < parse_predicate_count(); i++) {
1845 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1846 parse_predicate->mark_useless(igvn);
1847 }
1848 _parse_predicates.clear();
1849 }
1850
1851 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1852 if (!n->for_post_loop_opts_igvn()) {
1853 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1854 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1855 _for_post_loop_igvn.append(n);
1856 }
1857 }
1858
1859 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1860 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1861 _for_post_loop_igvn.remove(n);
1862 }
1863
1864 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1865 // Verify that all previous optimizations produced a valid graph
1866 // at least to this point, even if no loop optimizations were done.
1867 PhaseIdealLoop::verify(igvn);
1868
1869 if (has_loops() || _loop_opts_cnt > 0) {
1870 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1871 }
1872 C->set_post_loop_opts_phase(); // no more loop opts allowed
1873
1874 assert(!C->major_progress(), "not cleared");
1875
1876 if (_for_post_loop_igvn.length() > 0) {
1877 while (_for_post_loop_igvn.length() > 0) {
1878 Node* n = _for_post_loop_igvn.pop();
1879 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1880 igvn._worklist.push(n);
1881 }
1882 igvn.optimize();
1883 if (failing()) return;
1884 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1885 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1886
1887 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1888 if (C->major_progress()) {
1889 C->clear_major_progress(); // ensure that major progress is now clear
1890 }
1891 }
1892 }
1893
1894 void Compile::record_for_merge_stores_igvn(Node* n) {
1895 if (!n->for_merge_stores_igvn()) {
1896 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1897 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1898 _for_merge_stores_igvn.append(n);
1899 }
1900 }
1901
1902 void Compile::remove_from_merge_stores_igvn(Node* n) {
1903 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1904 _for_merge_stores_igvn.remove(n);
1905 }
1906
1907 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1908 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1909 // the stores, and we merge the wrong sequence of stores.
1910 // Example:
1911 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1912 // Apply MergeStores:
1913 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1914 // Remove more RangeChecks:
1915 // StoreI [ StoreL ] StoreI
1916 // But now it would have been better to do this instead:
1917 // [ StoreL ] [ StoreL ]
1918 //
1919 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1920 // since we never unset _merge_stores_phase.
1921 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1922 C->set_merge_stores_phase();
1923
1924 if (_for_merge_stores_igvn.length() > 0) {
1925 while (_for_merge_stores_igvn.length() > 0) {
1926 Node* n = _for_merge_stores_igvn.pop();
1927 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1928 igvn._worklist.push(n);
1929 }
1930 igvn.optimize();
1931 if (failing()) return;
1932 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1933 print_method(PHASE_AFTER_MERGE_STORES, 3);
1934 }
1935 }
1936
1937 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1938 if (OptimizeUnstableIf) {
1939 _unstable_if_traps.append(trap);
1940 }
1941 }
1942
1943 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1944 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1945 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1946 Node* n = trap->uncommon_trap();
1947 if (!useful.member(n)) {
1948 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1949 }
1950 }
1951 }
1952
1953 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1954 // or fold-compares case. Return true if succeed or not found.
1955 //
1956 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1957 // false and ask fold-compares to yield.
1958 //
1959 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1960 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1961 // when deoptimization does happen.
1962 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1963 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1964 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1965 if (trap->uncommon_trap() == unc) {
1966 if (yield && trap->modified()) {
1967 return false;
1968 }
1969 _unstable_if_traps.delete_at(i);
1970 break;
1971 }
1972 }
1973 return true;
1974 }
1975
1976 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1977 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1978 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1979 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1980 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1981 CallStaticJavaNode* unc = trap->uncommon_trap();
1982 int next_bci = trap->next_bci();
1983 bool modified = trap->modified();
1984
1985 if (next_bci != -1 && !modified) {
1986 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1987 JVMState* jvms = unc->jvms();
1988 ciMethod* method = jvms->method();
1989 ciBytecodeStream iter(method);
1990
1991 iter.force_bci(jvms->bci());
1992 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1993 Bytecodes::Code c = iter.cur_bc();
1994 Node* lhs = nullptr;
1995 Node* rhs = nullptr;
1996 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1997 lhs = unc->peek_operand(0);
1998 rhs = unc->peek_operand(1);
1999 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2000 lhs = unc->peek_operand(0);
2001 }
2002
2003 ResourceMark rm;
2004 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2005 assert(live_locals.is_valid(), "broken liveness info");
2006 int len = (int)live_locals.size();
2007
2008 for (int i = 0; i < len; i++) {
2009 Node* local = unc->local(jvms, i);
2010 // kill local using the liveness of next_bci.
2011 // give up when the local looks like an operand to secure reexecution.
2012 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2013 uint idx = jvms->locoff() + i;
2014 #ifdef ASSERT
2015 if (PrintOpto && Verbose) {
2016 tty->print("[unstable_if] kill local#%d: ", idx);
2017 local->dump();
2018 tty->cr();
2019 }
2020 #endif
2021 igvn.replace_input_of(unc, idx, top());
2022 modified = true;
2023 }
2024 }
2025 }
2026
2027 // keep the mondified trap for late query
2028 if (modified) {
2029 trap->set_modified();
2030 } else {
2031 _unstable_if_traps.delete_at(i);
2032 }
2033 }
2034 igvn.optimize();
2035 }
2036
2037 // StringOpts and late inlining of string methods
2038 void Compile::inline_string_calls(bool parse_time) {
2039 {
2040 // remove useless nodes to make the usage analysis simpler
2041 ResourceMark rm;
2042 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2043 }
2044
2045 {
2046 ResourceMark rm;
2047 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2048 PhaseStringOpts pso(initial_gvn());
2049 print_method(PHASE_AFTER_STRINGOPTS, 3);
2050 }
2051
2052 // now inline anything that we skipped the first time around
2053 if (!parse_time) {
2054 _late_inlines_pos = _late_inlines.length();
2055 }
2056
2057 while (_string_late_inlines.length() > 0) {
2058 CallGenerator* cg = _string_late_inlines.pop();
2059 cg->do_late_inline();
2060 if (failing()) return;
2061 }
2062 _string_late_inlines.trunc_to(0);
2063 }
2064
2065 // Late inlining of boxing methods
2066 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2067 if (_boxing_late_inlines.length() > 0) {
2068 assert(has_boxed_value(), "inconsistent");
2069
2070 set_inlining_incrementally(true);
2071
2072 igvn_worklist()->ensure_empty(); // should be done with igvn
2073
2074 _late_inlines_pos = _late_inlines.length();
2075
2076 while (_boxing_late_inlines.length() > 0) {
2077 CallGenerator* cg = _boxing_late_inlines.pop();
2078 cg->do_late_inline();
2079 if (failing()) return;
2080 }
2081 _boxing_late_inlines.trunc_to(0);
2082
2083 inline_incrementally_cleanup(igvn);
2084
2085 set_inlining_incrementally(false);
2086 }
2087 }
2088
2089 bool Compile::inline_incrementally_one() {
2090 assert(IncrementalInline, "incremental inlining should be on");
2091
2092 TracePhase tp(_t_incrInline_inline);
2093
2094 set_inlining_progress(false);
2095 set_do_cleanup(false);
2096
2097 for (int i = 0; i < _late_inlines.length(); i++) {
2098 _late_inlines_pos = i+1;
2099 CallGenerator* cg = _late_inlines.at(i);
2100 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2101 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2102 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2103 if (should_stress_inlining()) {
2104 // randomly add repeated inline attempt if stress-inlining
2105 cg->call_node()->set_generator(cg);
2106 C->igvn_worklist()->push(cg->call_node());
2107 continue;
2108 }
2109 cg->do_late_inline();
2110 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2111 if (failing()) {
2112 return false;
2113 } else if (inlining_progress()) {
2114 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2115 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2116 break; // process one call site at a time
2117 } else {
2118 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2119 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2120 // Avoid potential infinite loop if node already in the IGVN list
2121 assert(false, "scheduled for IGVN during inlining attempt");
2122 } else {
2123 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2124 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2125 cg->call_node()->set_generator(cg);
2126 }
2127 }
2128 } else {
2129 // Ignore late inline direct calls when inlining is not allowed.
2130 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2131 }
2132 }
2133 // Remove processed elements.
2134 _late_inlines.remove_till(_late_inlines_pos);
2135 _late_inlines_pos = 0;
2136
2137 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2138
2139 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2140
2141 set_inlining_progress(false);
2142 set_do_cleanup(false);
2143
2144 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2145 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2146 }
2147
2148 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2149 {
2150 TracePhase tp(_t_incrInline_pru);
2151 ResourceMark rm;
2152 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2153 }
2154 {
2155 TracePhase tp(_t_incrInline_igvn);
2156 igvn.reset();
2157 igvn.optimize();
2158 if (failing()) return;
2159 }
2160 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2161 }
2162
2163 // Perform incremental inlining until bound on number of live nodes is reached
2164 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2165 TracePhase tp(_t_incrInline);
2166
2167 set_inlining_incrementally(true);
2168 uint low_live_nodes = 0;
2169
2170 while (_late_inlines.length() > 0) {
2171 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2172 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2173 TracePhase tp(_t_incrInline_ideal);
2174 // PhaseIdealLoop is expensive so we only try it once we are
2175 // out of live nodes and we only try it again if the previous
2176 // helped got the number of nodes down significantly
2177 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2178 if (failing()) return;
2179 low_live_nodes = live_nodes();
2180 _major_progress = true;
2181 }
2182
2183 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2184 bool do_print_inlining = print_inlining() || print_intrinsics();
2185 if (do_print_inlining || log() != nullptr) {
2186 // Print inlining message for candidates that we couldn't inline for lack of space.
2187 for (int i = 0; i < _late_inlines.length(); i++) {
2188 CallGenerator* cg = _late_inlines.at(i);
2189 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2190 if (do_print_inlining) {
2191 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2192 }
2193 log_late_inline_failure(cg, msg);
2194 }
2195 }
2196 break; // finish
2197 }
2198 }
2199
2200 igvn_worklist()->ensure_empty(); // should be done with igvn
2201
2202 while (inline_incrementally_one()) {
2203 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2204 }
2205 if (failing()) return;
2206
2207 inline_incrementally_cleanup(igvn);
2208
2209 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2210
2211 if (failing()) return;
2212
2213 if (_late_inlines.length() == 0) {
2214 break; // no more progress
2215 }
2216 }
2217
2218 igvn_worklist()->ensure_empty(); // should be done with igvn
2219
2220 if (_string_late_inlines.length() > 0) {
2221 assert(has_stringbuilder(), "inconsistent");
2222
2223 inline_string_calls(false);
2224
2225 if (failing()) return;
2226
2227 inline_incrementally_cleanup(igvn);
2228 }
2229
2230 set_inlining_incrementally(false);
2231 }
2232
2233 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2234 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2235 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2236 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2237 // as if "inlining_incrementally() == true" were set.
2238 assert(inlining_incrementally() == false, "not allowed");
2239 assert(_modified_nodes == nullptr, "not allowed");
2240 assert(_late_inlines.length() > 0, "sanity");
2241
2242 while (_late_inlines.length() > 0) {
2243 igvn_worklist()->ensure_empty(); // should be done with igvn
2244
2245 while (inline_incrementally_one()) {
2246 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2247 }
2248 if (failing()) return;
2249
2250 inline_incrementally_cleanup(igvn);
2251 }
2252 }
2253
2254 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2255 if (_loop_opts_cnt > 0) {
2256 while (major_progress() && (_loop_opts_cnt > 0)) {
2257 TracePhase tp(_t_idealLoop);
2258 PhaseIdealLoop::optimize(igvn, mode);
2259 _loop_opts_cnt--;
2260 if (failing()) return false;
2261 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2262 }
2263 }
2264 return true;
2265 }
2266
2267 // Remove edges from "root" to each SafePoint at a backward branch.
2268 // They were inserted during parsing (see add_safepoint()) to make
2269 // infinite loops without calls or exceptions visible to root, i.e.,
2270 // useful.
2271 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2272 Node *r = root();
2273 if (r != nullptr) {
2274 for (uint i = r->req(); i < r->len(); ++i) {
2275 Node *n = r->in(i);
2276 if (n != nullptr && n->is_SafePoint()) {
2277 r->rm_prec(i);
2278 if (n->outcnt() == 0) {
2279 igvn.remove_dead_node(n);
2280 }
2281 --i;
2282 }
2283 }
2284 // Parsing may have added top inputs to the root node (Path
2285 // leading to the Halt node proven dead). Make sure we get a
2286 // chance to clean them up.
2287 igvn._worklist.push(r);
2288 igvn.optimize();
2289 }
2290 }
2291
2292 //------------------------------Optimize---------------------------------------
2293 // Given a graph, optimize it.
2294 void Compile::Optimize() {
2295 TracePhase tp(_t_optimizer);
2296
2297 #ifndef PRODUCT
2298 if (env()->break_at_compile()) {
2299 BREAKPOINT;
2300 }
2301
2302 #endif
2303
2304 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2305 #ifdef ASSERT
2306 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2307 #endif
2308
2309 ResourceMark rm;
2310
2311 NOT_PRODUCT( verify_graph_edges(); )
2312
2313 print_method(PHASE_AFTER_PARSING, 1);
2314
2315 {
2316 // Iterative Global Value Numbering, including ideal transforms
2317 PhaseIterGVN igvn;
2318 #ifdef ASSERT
2319 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2320 #endif
2321 {
2322 TracePhase tp(_t_iterGVN);
2323 igvn.optimize();
2324 }
2325
2326 if (failing()) return;
2327
2328 print_method(PHASE_ITER_GVN1, 2);
2329
2330 process_for_unstable_if_traps(igvn);
2331
2332 if (failing()) return;
2333
2334 inline_incrementally(igvn);
2335
2336 print_method(PHASE_INCREMENTAL_INLINE, 2);
2337
2338 if (failing()) return;
2339
2340 if (eliminate_boxing()) {
2341 // Inline valueOf() methods now.
2342 inline_boxing_calls(igvn);
2343
2344 if (failing()) return;
2345
2346 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2347 inline_incrementally(igvn);
2348 }
2349
2350 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2351
2352 if (failing()) return;
2353 }
2354
2355 // Remove the speculative part of types and clean up the graph from
2356 // the extra CastPP nodes whose only purpose is to carry them. Do
2357 // that early so that optimizations are not disrupted by the extra
2358 // CastPP nodes.
2359 remove_speculative_types(igvn);
2360
2361 if (failing()) return;
2362
2363 // No more new expensive nodes will be added to the list from here
2364 // so keep only the actual candidates for optimizations.
2365 cleanup_expensive_nodes(igvn);
2366
2367 if (failing()) return;
2368
2369 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2370 if (EnableVectorSupport && has_vbox_nodes()) {
2371 TracePhase tp(_t_vector);
2372 PhaseVector pv(igvn);
2373 pv.optimize_vector_boxes();
2374 if (failing()) return;
2375 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2376 }
2377 assert(!has_vbox_nodes(), "sanity");
2378
2379 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2380 Compile::TracePhase tp(_t_renumberLive);
2381 igvn_worklist()->ensure_empty(); // should be done with igvn
2382 {
2383 ResourceMark rm;
2384 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2385 }
2386 igvn.reset();
2387 igvn.optimize();
2388 if (failing()) return;
2389 }
2390
2391 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2392 // safepoints
2393 remove_root_to_sfpts_edges(igvn);
2394
2395 if (failing()) return;
2396
2397 if (has_loops()) {
2398 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2399 }
2400
2401 // Perform escape analysis
2402 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2403 if (has_loops()) {
2404 // Cleanup graph (remove dead nodes).
2405 TracePhase tp(_t_idealLoop);
2406 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2407 if (failing()) return;
2408 }
2409 bool progress;
2410 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2411 do {
2412 ConnectionGraph::do_analysis(this, &igvn);
2413
2414 if (failing()) return;
2415
2416 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2417
2418 // Optimize out fields loads from scalar replaceable allocations.
2419 igvn.optimize();
2420 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2421
2422 if (failing()) return;
2423
2424 if (congraph() != nullptr && macro_count() > 0) {
2425 TracePhase tp(_t_macroEliminate);
2426 PhaseMacroExpand mexp(igvn);
2427 mexp.eliminate_macro_nodes();
2428 if (failing()) return;
2429 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2430
2431 igvn.set_delay_transform(false);
2432 igvn.optimize();
2433 if (failing()) return;
2434
2435 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2436 }
2437
2438 ConnectionGraph::verify_ram_nodes(this, root());
2439 if (failing()) return;
2440
2441 progress = do_iterative_escape_analysis() &&
2442 (macro_count() < mcount) &&
2443 ConnectionGraph::has_candidates(this);
2444 // Try again if candidates exist and made progress
2445 // by removing some allocations and/or locks.
2446 } while (progress);
2447 }
2448
2449 // Loop transforms on the ideal graph. Range Check Elimination,
2450 // peeling, unrolling, etc.
2451
2452 // Set loop opts counter
2453 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2454 {
2455 TracePhase tp(_t_idealLoop);
2456 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2457 _loop_opts_cnt--;
2458 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2459 if (failing()) return;
2460 }
2461 // Loop opts pass if partial peeling occurred in previous pass
2462 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2463 TracePhase tp(_t_idealLoop);
2464 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2465 _loop_opts_cnt--;
2466 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2467 if (failing()) return;
2468 }
2469 // Loop opts pass for loop-unrolling before CCP
2470 if(major_progress() && (_loop_opts_cnt > 0)) {
2471 TracePhase tp(_t_idealLoop);
2472 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2473 _loop_opts_cnt--;
2474 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2475 }
2476 if (!failing()) {
2477 // Verify that last round of loop opts produced a valid graph
2478 PhaseIdealLoop::verify(igvn);
2479 }
2480 }
2481 if (failing()) return;
2482
2483 // Conditional Constant Propagation;
2484 print_method(PHASE_BEFORE_CCP1, 2);
2485 PhaseCCP ccp( &igvn );
2486 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2487 {
2488 TracePhase tp(_t_ccp);
2489 ccp.do_transform();
2490 }
2491 print_method(PHASE_CCP1, 2);
2492
2493 assert( true, "Break here to ccp.dump_old2new_map()");
2494
2495 // Iterative Global Value Numbering, including ideal transforms
2496 {
2497 TracePhase tp(_t_iterGVN2);
2498 igvn.reset_from_igvn(&ccp);
2499 igvn.optimize();
2500 }
2501 print_method(PHASE_ITER_GVN2, 2);
2502
2503 if (failing()) return;
2504
2505 // Loop transforms on the ideal graph. Range Check Elimination,
2506 // peeling, unrolling, etc.
2507 if (!optimize_loops(igvn, LoopOptsDefault)) {
2508 return;
2509 }
2510
2511 if (failing()) return;
2512
2513 C->clear_major_progress(); // ensure that major progress is now clear
2514
2515 process_for_post_loop_opts_igvn(igvn);
2516
2517 process_for_merge_stores_igvn(igvn);
2518
2519 if (failing()) return;
2520
2521 #ifdef ASSERT
2522 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2523 #endif
2524
2525 {
2526 TracePhase tp(_t_macroExpand);
2527 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2528 PhaseMacroExpand mex(igvn);
2529 // Do not allow new macro nodes once we start to eliminate and expand
2530 C->reset_allow_macro_nodes();
2531 // Last attempt to eliminate macro nodes before expand
2532 mex.eliminate_macro_nodes();
2533 if (failing()) {
2534 return;
2535 }
2536 mex.eliminate_opaque_looplimit_macro_nodes();
2537 if (failing()) {
2538 return;
2539 }
2540 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2541 if (mex.expand_macro_nodes()) {
2542 assert(failing(), "must bail out w/ explicit message");
2543 return;
2544 }
2545 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2546 }
2547
2548 {
2549 TracePhase tp(_t_barrierExpand);
2550 if (bs->expand_barriers(this, igvn)) {
2551 assert(failing(), "must bail out w/ explicit message");
2552 return;
2553 }
2554 print_method(PHASE_BARRIER_EXPANSION, 2);
2555 }
2556
2557 if (C->max_vector_size() > 0) {
2558 C->optimize_logic_cones(igvn);
2559 igvn.optimize();
2560 if (failing()) return;
2561 }
2562
2563 DEBUG_ONLY( _modified_nodes = nullptr; )
2564
2565 assert(igvn._worklist.size() == 0, "not empty");
2566
2567 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2568
2569 if (_late_inlines.length() > 0) {
2570 // More opportunities to optimize virtual and MH calls.
2571 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2572 process_late_inline_calls_no_inline(igvn);
2573 if (failing()) return;
2574 }
2575 } // (End scope of igvn; run destructor if necessary for asserts.)
2576
2577 check_no_dead_use();
2578
2579 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2580 // to remove hashes to unlock nodes for modifications.
2581 C->node_hash()->clear();
2582
2583 // A method with only infinite loops has no edges entering loops from root
2584 {
2585 TracePhase tp(_t_graphReshaping);
2586 if (final_graph_reshaping()) {
2587 assert(failing(), "must bail out w/ explicit message");
2588 return;
2589 }
2590 }
2591
2592 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2593 DEBUG_ONLY(set_phase_optimize_finished();)
2594 }
2595
2596 #ifdef ASSERT
2597 void Compile::check_no_dead_use() const {
2598 ResourceMark rm;
2599 Unique_Node_List wq;
2600 wq.push(root());
2601 for (uint i = 0; i < wq.size(); ++i) {
2602 Node* n = wq.at(i);
2603 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2604 Node* u = n->fast_out(j);
2605 if (u->outcnt() == 0 && !u->is_Con()) {
2606 u->dump();
2607 fatal("no reachable node should have no use");
2608 }
2609 wq.push(u);
2610 }
2611 }
2612 }
2613 #endif
2614
2615 void Compile::inline_vector_reboxing_calls() {
2616 if (C->_vector_reboxing_late_inlines.length() > 0) {
2617 _late_inlines_pos = C->_late_inlines.length();
2618 while (_vector_reboxing_late_inlines.length() > 0) {
2619 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2620 cg->do_late_inline();
2621 if (failing()) return;
2622 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2623 }
2624 _vector_reboxing_late_inlines.trunc_to(0);
2625 }
2626 }
2627
2628 bool Compile::has_vbox_nodes() {
2629 if (C->_vector_reboxing_late_inlines.length() > 0) {
2630 return true;
2631 }
2632 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2633 Node * n = C->macro_node(macro_idx);
2634 assert(n->is_macro(), "only macro nodes expected here");
2635 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2636 return true;
2637 }
2638 }
2639 return false;
2640 }
2641
2642 //---------------------------- Bitwise operation packing optimization ---------------------------
2643
2644 static bool is_vector_unary_bitwise_op(Node* n) {
2645 return n->Opcode() == Op_XorV &&
2646 VectorNode::is_vector_bitwise_not_pattern(n);
2647 }
2648
2649 static bool is_vector_binary_bitwise_op(Node* n) {
2650 switch (n->Opcode()) {
2651 case Op_AndV:
2652 case Op_OrV:
2653 return true;
2654
2655 case Op_XorV:
2656 return !is_vector_unary_bitwise_op(n);
2657
2658 default:
2659 return false;
2660 }
2661 }
2662
2663 static bool is_vector_ternary_bitwise_op(Node* n) {
2664 return n->Opcode() == Op_MacroLogicV;
2665 }
2666
2667 static bool is_vector_bitwise_op(Node* n) {
2668 return is_vector_unary_bitwise_op(n) ||
2669 is_vector_binary_bitwise_op(n) ||
2670 is_vector_ternary_bitwise_op(n);
2671 }
2672
2673 static bool is_vector_bitwise_cone_root(Node* n) {
2674 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2675 return false;
2676 }
2677 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2678 if (is_vector_bitwise_op(n->fast_out(i))) {
2679 return false;
2680 }
2681 }
2682 return true;
2683 }
2684
2685 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2686 uint cnt = 0;
2687 if (is_vector_bitwise_op(n)) {
2688 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2689 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2690 for (uint i = 1; i < inp_cnt; i++) {
2691 Node* in = n->in(i);
2692 bool skip = VectorNode::is_all_ones_vector(in);
2693 if (!skip && !inputs.member(in)) {
2694 inputs.push(in);
2695 cnt++;
2696 }
2697 }
2698 assert(cnt <= 1, "not unary");
2699 } else {
2700 uint last_req = inp_cnt;
2701 if (is_vector_ternary_bitwise_op(n)) {
2702 last_req = inp_cnt - 1; // skip last input
2703 }
2704 for (uint i = 1; i < last_req; i++) {
2705 Node* def = n->in(i);
2706 if (!inputs.member(def)) {
2707 inputs.push(def);
2708 cnt++;
2709 }
2710 }
2711 }
2712 } else { // not a bitwise operations
2713 if (!inputs.member(n)) {
2714 inputs.push(n);
2715 cnt++;
2716 }
2717 }
2718 return cnt;
2719 }
2720
2721 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2722 Unique_Node_List useful_nodes;
2723 C->identify_useful_nodes(useful_nodes);
2724
2725 for (uint i = 0; i < useful_nodes.size(); i++) {
2726 Node* n = useful_nodes.at(i);
2727 if (is_vector_bitwise_cone_root(n)) {
2728 list.push(n);
2729 }
2730 }
2731 }
2732
2733 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2734 const TypeVect* vt,
2735 Unique_Node_List& partition,
2736 Unique_Node_List& inputs) {
2737 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2738 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2739 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2740
2741 Node* in1 = inputs.at(0);
2742 Node* in2 = inputs.at(1);
2743 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2744
2745 uint func = compute_truth_table(partition, inputs);
2746
2747 Node* pn = partition.at(partition.size() - 1);
2748 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2749 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2750 }
2751
2752 static uint extract_bit(uint func, uint pos) {
2753 return (func & (1 << pos)) >> pos;
2754 }
2755
2756 //
2757 // A macro logic node represents a truth table. It has 4 inputs,
2758 // First three inputs corresponds to 3 columns of a truth table
2759 // and fourth input captures the logic function.
2760 //
2761 // eg. fn = (in1 AND in2) OR in3;
2762 //
2763 // MacroNode(in1,in2,in3,fn)
2764 //
2765 // -----------------
2766 // in1 in2 in3 fn
2767 // -----------------
2768 // 0 0 0 0
2769 // 0 0 1 1
2770 // 0 1 0 0
2771 // 0 1 1 1
2772 // 1 0 0 0
2773 // 1 0 1 1
2774 // 1 1 0 1
2775 // 1 1 1 1
2776 //
2777
2778 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2779 int res = 0;
2780 for (int i = 0; i < 8; i++) {
2781 int bit1 = extract_bit(in1, i);
2782 int bit2 = extract_bit(in2, i);
2783 int bit3 = extract_bit(in3, i);
2784
2785 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2786 int func_bit = extract_bit(func, func_bit_pos);
2787
2788 res |= func_bit << i;
2789 }
2790 return res;
2791 }
2792
2793 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
2794 assert(n != nullptr, "");
2795 assert(eval_map.contains(n), "absent");
2796 return *(eval_map.get(n));
2797 }
2798
2799 static void eval_operands(Node* n,
2800 uint& func1, uint& func2, uint& func3,
2801 HashTable<Node*,uint>& eval_map) {
2802 assert(is_vector_bitwise_op(n), "");
2803
2804 if (is_vector_unary_bitwise_op(n)) {
2805 Node* opnd = n->in(1);
2806 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2807 opnd = n->in(2);
2808 }
2809 func1 = eval_operand(opnd, eval_map);
2810 } else if (is_vector_binary_bitwise_op(n)) {
2811 func1 = eval_operand(n->in(1), eval_map);
2812 func2 = eval_operand(n->in(2), eval_map);
2813 } else {
2814 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2815 func1 = eval_operand(n->in(1), eval_map);
2816 func2 = eval_operand(n->in(2), eval_map);
2817 func3 = eval_operand(n->in(3), eval_map);
2818 }
2819 }
2820
2821 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2822 assert(inputs.size() <= 3, "sanity");
2823 ResourceMark rm;
2824 uint res = 0;
2825 HashTable<Node*,uint> eval_map;
2826
2827 // Populate precomputed functions for inputs.
2828 // Each input corresponds to one column of 3 input truth-table.
2829 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2830 0xCC, // (_, b, _) -> b
2831 0xF0 }; // (a, _, _) -> a
2832 for (uint i = 0; i < inputs.size(); i++) {
2833 eval_map.put(inputs.at(i), input_funcs[2-i]);
2834 }
2835
2836 for (uint i = 0; i < partition.size(); i++) {
2837 Node* n = partition.at(i);
2838
2839 uint func1 = 0, func2 = 0, func3 = 0;
2840 eval_operands(n, func1, func2, func3, eval_map);
2841
2842 switch (n->Opcode()) {
2843 case Op_OrV:
2844 assert(func3 == 0, "not binary");
2845 res = func1 | func2;
2846 break;
2847 case Op_AndV:
2848 assert(func3 == 0, "not binary");
2849 res = func1 & func2;
2850 break;
2851 case Op_XorV:
2852 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2853 assert(func2 == 0 && func3 == 0, "not unary");
2854 res = (~func1) & 0xFF;
2855 } else {
2856 assert(func3 == 0, "not binary");
2857 res = func1 ^ func2;
2858 }
2859 break;
2860 case Op_MacroLogicV:
2861 // Ordering of inputs may change during evaluation of sub-tree
2862 // containing MacroLogic node as a child node, thus a re-evaluation
2863 // makes sure that function is evaluated in context of current
2864 // inputs.
2865 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2866 break;
2867
2868 default: assert(false, "not supported: %s", n->Name());
2869 }
2870 assert(res <= 0xFF, "invalid");
2871 eval_map.put(n, res);
2872 }
2873 return res;
2874 }
2875
2876 // Criteria under which nodes gets packed into a macro logic node:-
2877 // 1) Parent and both child nodes are all unmasked or masked with
2878 // same predicates.
2879 // 2) Masked parent can be packed with left child if it is predicated
2880 // and both have same predicates.
2881 // 3) Masked parent can be packed with right child if its un-predicated
2882 // or has matching predication condition.
2883 // 4) An unmasked parent can be packed with an unmasked child.
2884 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2885 assert(partition.size() == 0, "not empty");
2886 assert(inputs.size() == 0, "not empty");
2887 if (is_vector_ternary_bitwise_op(n)) {
2888 return false;
2889 }
2890
2891 bool is_unary_op = is_vector_unary_bitwise_op(n);
2892 if (is_unary_op) {
2893 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2894 return false; // too few inputs
2895 }
2896
2897 bool pack_left_child = true;
2898 bool pack_right_child = true;
2899
2900 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2901 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2902
2903 int left_child_input_cnt = 0;
2904 int right_child_input_cnt = 0;
2905
2906 bool parent_is_predicated = n->is_predicated_vector();
2907 bool left_child_predicated = n->in(1)->is_predicated_vector();
2908 bool right_child_predicated = n->in(2)->is_predicated_vector();
2909
2910 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2911 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2912 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2913
2914 do {
2915 if (pack_left_child && left_child_LOP &&
2916 ((!parent_is_predicated && !left_child_predicated) ||
2917 ((parent_is_predicated && left_child_predicated &&
2918 parent_pred == left_child_pred)))) {
2919 partition.push(n->in(1));
2920 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2921 } else {
2922 inputs.push(n->in(1));
2923 left_child_input_cnt = 1;
2924 }
2925
2926 if (pack_right_child && right_child_LOP &&
2927 (!right_child_predicated ||
2928 (right_child_predicated && parent_is_predicated &&
2929 parent_pred == right_child_pred))) {
2930 partition.push(n->in(2));
2931 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2932 } else {
2933 inputs.push(n->in(2));
2934 right_child_input_cnt = 1;
2935 }
2936
2937 if (inputs.size() > 3) {
2938 assert(partition.size() > 0, "");
2939 inputs.clear();
2940 partition.clear();
2941 if (left_child_input_cnt > right_child_input_cnt) {
2942 pack_left_child = false;
2943 } else {
2944 pack_right_child = false;
2945 }
2946 } else {
2947 break;
2948 }
2949 } while(true);
2950
2951 if(partition.size()) {
2952 partition.push(n);
2953 }
2954
2955 return (partition.size() == 2 || partition.size() == 3) &&
2956 (inputs.size() == 2 || inputs.size() == 3);
2957 }
2958
2959 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2960 assert(is_vector_bitwise_op(n), "not a root");
2961
2962 visited.set(n->_idx);
2963
2964 // 1) Do a DFS walk over the logic cone.
2965 for (uint i = 1; i < n->req(); i++) {
2966 Node* in = n->in(i);
2967 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2968 process_logic_cone_root(igvn, in, visited);
2969 }
2970 }
2971
2972 // 2) Bottom up traversal: Merge node[s] with
2973 // the parent to form macro logic node.
2974 Unique_Node_List partition;
2975 Unique_Node_List inputs;
2976 if (compute_logic_cone(n, partition, inputs)) {
2977 const TypeVect* vt = n->bottom_type()->is_vect();
2978 Node* pn = partition.at(partition.size() - 1);
2979 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2980 if (mask == nullptr ||
2981 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2982 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2983 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
2984 igvn.replace_node(n, macro_logic);
2985 }
2986 }
2987 }
2988
2989 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2990 ResourceMark rm;
2991 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2992 Unique_Node_List list;
2993 collect_logic_cone_roots(list);
2994
2995 while (list.size() > 0) {
2996 Node* n = list.pop();
2997 const TypeVect* vt = n->bottom_type()->is_vect();
2998 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2999 if (supported) {
3000 VectorSet visited(comp_arena());
3001 process_logic_cone_root(igvn, n, visited);
3002 }
3003 }
3004 }
3005 }
3006
3007 //------------------------------Code_Gen---------------------------------------
3008 // Given a graph, generate code for it
3009 void Compile::Code_Gen() {
3010 if (failing()) {
3011 return;
3012 }
3013
3014 // Perform instruction selection. You might think we could reclaim Matcher
3015 // memory PDQ, but actually the Matcher is used in generating spill code.
3016 // Internals of the Matcher (including some VectorSets) must remain live
3017 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3018 // set a bit in reclaimed memory.
3019
3020 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3021 // nodes. Mapping is only valid at the root of each matched subtree.
3022 NOT_PRODUCT( verify_graph_edges(); )
3023
3024 Matcher matcher;
3025 _matcher = &matcher;
3026 {
3027 TracePhase tp(_t_matcher);
3028 matcher.match();
3029 if (failing()) {
3030 return;
3031 }
3032 }
3033 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3034 // nodes. Mapping is only valid at the root of each matched subtree.
3035 NOT_PRODUCT( verify_graph_edges(); )
3036
3037 // If you have too many nodes, or if matching has failed, bail out
3038 check_node_count(0, "out of nodes matching instructions");
3039 if (failing()) {
3040 return;
3041 }
3042
3043 print_method(PHASE_MATCHING, 2);
3044
3045 // Build a proper-looking CFG
3046 PhaseCFG cfg(node_arena(), root(), matcher);
3047 if (failing()) {
3048 return;
3049 }
3050 _cfg = &cfg;
3051 {
3052 TracePhase tp(_t_scheduler);
3053 bool success = cfg.do_global_code_motion();
3054 if (!success) {
3055 return;
3056 }
3057
3058 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3059 NOT_PRODUCT( verify_graph_edges(); )
3060 cfg.verify();
3061 if (failing()) {
3062 return;
3063 }
3064 }
3065
3066 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3067 _regalloc = ®alloc;
3068 {
3069 TracePhase tp(_t_registerAllocation);
3070 // Perform register allocation. After Chaitin, use-def chains are
3071 // no longer accurate (at spill code) and so must be ignored.
3072 // Node->LRG->reg mappings are still accurate.
3073 _regalloc->Register_Allocate();
3074
3075 // Bail out if the allocator builds too many nodes
3076 if (failing()) {
3077 return;
3078 }
3079
3080 print_method(PHASE_REGISTER_ALLOCATION, 2);
3081 }
3082
3083 // Prior to register allocation we kept empty basic blocks in case the
3084 // the allocator needed a place to spill. After register allocation we
3085 // are not adding any new instructions. If any basic block is empty, we
3086 // can now safely remove it.
3087 {
3088 TracePhase tp(_t_blockOrdering);
3089 cfg.remove_empty_blocks();
3090 if (do_freq_based_layout()) {
3091 PhaseBlockLayout layout(cfg);
3092 } else {
3093 cfg.set_loop_alignment();
3094 }
3095 cfg.fixup_flow();
3096 cfg.remove_unreachable_blocks();
3097 cfg.verify_dominator_tree();
3098 print_method(PHASE_BLOCK_ORDERING, 3);
3099 }
3100
3101 // Apply peephole optimizations
3102 if( OptoPeephole ) {
3103 TracePhase tp(_t_peephole);
3104 PhasePeephole peep( _regalloc, cfg);
3105 peep.do_transform();
3106 print_method(PHASE_PEEPHOLE, 3);
3107 }
3108
3109 // Do late expand if CPU requires this.
3110 if (Matcher::require_postalloc_expand) {
3111 TracePhase tp(_t_postalloc_expand);
3112 cfg.postalloc_expand(_regalloc);
3113 print_method(PHASE_POSTALLOC_EXPAND, 3);
3114 }
3115
3116 #ifdef ASSERT
3117 {
3118 CompilationMemoryStatistic::do_test_allocations();
3119 if (failing()) return;
3120 }
3121 #endif
3122
3123 // Convert Nodes to instruction bits in a buffer
3124 {
3125 TracePhase tp(_t_output);
3126 PhaseOutput output;
3127 output.Output();
3128 if (failing()) return;
3129 output.install();
3130 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3131 }
3132
3133 // He's dead, Jim.
3134 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3135 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3136 }
3137
3138 //------------------------------Final_Reshape_Counts---------------------------
3139 // This class defines counters to help identify when a method
3140 // may/must be executed using hardware with only 24-bit precision.
3141 struct Final_Reshape_Counts : public StackObj {
3142 int _call_count; // count non-inlined 'common' calls
3143 int _float_count; // count float ops requiring 24-bit precision
3144 int _double_count; // count double ops requiring more precision
3145 int _java_call_count; // count non-inlined 'java' calls
3146 int _inner_loop_count; // count loops which need alignment
3147 VectorSet _visited; // Visitation flags
3148 Node_List _tests; // Set of IfNodes & PCTableNodes
3149
3150 Final_Reshape_Counts() :
3151 _call_count(0), _float_count(0), _double_count(0),
3152 _java_call_count(0), _inner_loop_count(0) { }
3153
3154 void inc_call_count () { _call_count ++; }
3155 void inc_float_count () { _float_count ++; }
3156 void inc_double_count() { _double_count++; }
3157 void inc_java_call_count() { _java_call_count++; }
3158 void inc_inner_loop_count() { _inner_loop_count++; }
3159
3160 int get_call_count () const { return _call_count ; }
3161 int get_float_count () const { return _float_count ; }
3162 int get_double_count() const { return _double_count; }
3163 int get_java_call_count() const { return _java_call_count; }
3164 int get_inner_loop_count() const { return _inner_loop_count; }
3165 };
3166
3167 //------------------------------final_graph_reshaping_impl----------------------
3168 // Implement items 1-5 from final_graph_reshaping below.
3169 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3170
3171 if ( n->outcnt() == 0 ) return; // dead node
3172 uint nop = n->Opcode();
3173
3174 // Check for 2-input instruction with "last use" on right input.
3175 // Swap to left input. Implements item (2).
3176 if( n->req() == 3 && // two-input instruction
3177 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3178 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3179 n->in(2)->outcnt() == 1 &&// right use IS a last use
3180 !n->in(2)->is_Con() ) { // right use is not a constant
3181 // Check for commutative opcode
3182 switch( nop ) {
3183 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3184 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3185 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3186 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3187 case Op_AndL: case Op_XorL: case Op_OrL:
3188 case Op_AndI: case Op_XorI: case Op_OrI: {
3189 // Move "last use" input to left by swapping inputs
3190 n->swap_edges(1, 2);
3191 break;
3192 }
3193 default:
3194 break;
3195 }
3196 }
3197
3198 #ifdef ASSERT
3199 if( n->is_Mem() ) {
3200 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3201 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3202 // oop will be recorded in oop map if load crosses safepoint
3203 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3204 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3205 "raw memory operations should have control edge");
3206 }
3207 if (n->is_MemBar()) {
3208 MemBarNode* mb = n->as_MemBar();
3209 if (mb->trailing_store() || mb->trailing_load_store()) {
3210 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3211 Node* mem = mb->in(MemBarNode::Precedent);
3212 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3213 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3214 } else if (mb->leading()) {
3215 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3216 }
3217 }
3218 #endif
3219 // Count FPU ops and common calls, implements item (3)
3220 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3221
3222 // Collect CFG split points
3223 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3224 frc._tests.push(n);
3225 }
3226 }
3227
3228 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3229 if (!UseDivMod) {
3230 return;
3231 }
3232
3233 // Check if "a % b" and "a / b" both exist
3234 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3235 if (d == nullptr) {
3236 return;
3237 }
3238
3239 // Replace them with a fused divmod if supported
3240 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3241 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3242 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3243 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3244 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3245 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3246 // DivMod node so the dependency is not lost.
3247 divmod->add_prec_from(n);
3248 divmod->add_prec_from(d);
3249 d->subsume_by(divmod->div_proj(), this);
3250 n->subsume_by(divmod->mod_proj(), this);
3251 } else {
3252 // Replace "a % b" with "a - ((a / b) * b)"
3253 Node* mult = MulNode::make(d, d->in(2), bt);
3254 Node* sub = SubNode::make(d->in(1), mult, bt);
3255 n->subsume_by(sub, this);
3256 }
3257 }
3258
3259 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3260 switch( nop ) {
3261 // Count all float operations that may use FPU
3262 case Op_AddF:
3263 case Op_SubF:
3264 case Op_MulF:
3265 case Op_DivF:
3266 case Op_NegF:
3267 case Op_ModF:
3268 case Op_ConvI2F:
3269 case Op_ConF:
3270 case Op_CmpF:
3271 case Op_CmpF3:
3272 case Op_StoreF:
3273 case Op_LoadF:
3274 // case Op_ConvL2F: // longs are split into 32-bit halves
3275 frc.inc_float_count();
3276 break;
3277
3278 case Op_ConvF2D:
3279 case Op_ConvD2F:
3280 frc.inc_float_count();
3281 frc.inc_double_count();
3282 break;
3283
3284 // Count all double operations that may use FPU
3285 case Op_AddD:
3286 case Op_SubD:
3287 case Op_MulD:
3288 case Op_DivD:
3289 case Op_NegD:
3290 case Op_ModD:
3291 case Op_ConvI2D:
3292 case Op_ConvD2I:
3293 // case Op_ConvL2D: // handled by leaf call
3294 // case Op_ConvD2L: // handled by leaf call
3295 case Op_ConD:
3296 case Op_CmpD:
3297 case Op_CmpD3:
3298 case Op_StoreD:
3299 case Op_LoadD:
3300 case Op_LoadD_unaligned:
3301 frc.inc_double_count();
3302 break;
3303 case Op_Opaque1: // Remove Opaque Nodes before matching
3304 n->subsume_by(n->in(1), this);
3305 break;
3306 case Op_CallLeafPure: {
3307 // If the pure call is not supported, then lower to a CallLeaf.
3308 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3309 CallNode* call = n->as_Call();
3310 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3311 call->_name, TypeRawPtr::BOTTOM);
3312 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3313 new_call->init_req(TypeFunc::I_O, C->top());
3314 new_call->init_req(TypeFunc::Memory, C->top());
3315 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3316 new_call->init_req(TypeFunc::FramePtr, C->top());
3317 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
3318 new_call->init_req(i, call->in(i));
3319 }
3320 n->subsume_by(new_call, this);
3321 }
3322 frc.inc_call_count();
3323 break;
3324 }
3325 case Op_CallStaticJava:
3326 case Op_CallJava:
3327 case Op_CallDynamicJava:
3328 frc.inc_java_call_count(); // Count java call site;
3329 case Op_CallRuntime:
3330 case Op_CallLeaf:
3331 case Op_CallLeafVector:
3332 case Op_CallLeafNoFP: {
3333 assert (n->is_Call(), "");
3334 CallNode *call = n->as_Call();
3335 // Count call sites where the FP mode bit would have to be flipped.
3336 // Do not count uncommon runtime calls:
3337 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3338 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3339 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3340 frc.inc_call_count(); // Count the call site
3341 } else { // See if uncommon argument is shared
3342 Node *n = call->in(TypeFunc::Parms);
3343 int nop = n->Opcode();
3344 // Clone shared simple arguments to uncommon calls, item (1).
3345 if (n->outcnt() > 1 &&
3346 !n->is_Proj() &&
3347 nop != Op_CreateEx &&
3348 nop != Op_CheckCastPP &&
3349 nop != Op_DecodeN &&
3350 nop != Op_DecodeNKlass &&
3351 !n->is_Mem() &&
3352 !n->is_Phi()) {
3353 Node *x = n->clone();
3354 call->set_req(TypeFunc::Parms, x);
3355 }
3356 }
3357 break;
3358 }
3359 case Op_StoreB:
3360 case Op_StoreC:
3361 case Op_StoreI:
3362 case Op_StoreL:
3363 case Op_CompareAndSwapB:
3364 case Op_CompareAndSwapS:
3365 case Op_CompareAndSwapI:
3366 case Op_CompareAndSwapL:
3367 case Op_CompareAndSwapP:
3368 case Op_CompareAndSwapN:
3369 case Op_WeakCompareAndSwapB:
3370 case Op_WeakCompareAndSwapS:
3371 case Op_WeakCompareAndSwapI:
3372 case Op_WeakCompareAndSwapL:
3373 case Op_WeakCompareAndSwapP:
3374 case Op_WeakCompareAndSwapN:
3375 case Op_CompareAndExchangeB:
3376 case Op_CompareAndExchangeS:
3377 case Op_CompareAndExchangeI:
3378 case Op_CompareAndExchangeL:
3379 case Op_CompareAndExchangeP:
3380 case Op_CompareAndExchangeN:
3381 case Op_GetAndAddS:
3382 case Op_GetAndAddB:
3383 case Op_GetAndAddI:
3384 case Op_GetAndAddL:
3385 case Op_GetAndSetS:
3386 case Op_GetAndSetB:
3387 case Op_GetAndSetI:
3388 case Op_GetAndSetL:
3389 case Op_GetAndSetP:
3390 case Op_GetAndSetN:
3391 case Op_StoreP:
3392 case Op_StoreN:
3393 case Op_StoreNKlass:
3394 case Op_LoadB:
3395 case Op_LoadUB:
3396 case Op_LoadUS:
3397 case Op_LoadI:
3398 case Op_LoadKlass:
3399 case Op_LoadNKlass:
3400 case Op_LoadL:
3401 case Op_LoadL_unaligned:
3402 case Op_LoadP:
3403 case Op_LoadN:
3404 case Op_LoadRange:
3405 case Op_LoadS:
3406 break;
3407
3408 case Op_AddP: { // Assert sane base pointers
3409 Node *addp = n->in(AddPNode::Address);
3410 assert( !addp->is_AddP() ||
3411 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3412 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3413 "Base pointers must match (addp %u)", addp->_idx );
3414 #ifdef _LP64
3415 if ((UseCompressedOops || UseCompressedClassPointers) &&
3416 addp->Opcode() == Op_ConP &&
3417 addp == n->in(AddPNode::Base) &&
3418 n->in(AddPNode::Offset)->is_Con()) {
3419 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3420 // on the platform and on the compressed oops mode.
3421 // Use addressing with narrow klass to load with offset on x86.
3422 // Some platforms can use the constant pool to load ConP.
3423 // Do this transformation here since IGVN will convert ConN back to ConP.
3424 const Type* t = addp->bottom_type();
3425 bool is_oop = t->isa_oopptr() != nullptr;
3426 bool is_klass = t->isa_klassptr() != nullptr;
3427
3428 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3429 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3430 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3431 Node* nn = nullptr;
3432
3433 int op = is_oop ? Op_ConN : Op_ConNKlass;
3434
3435 // Look for existing ConN node of the same exact type.
3436 Node* r = root();
3437 uint cnt = r->outcnt();
3438 for (uint i = 0; i < cnt; i++) {
3439 Node* m = r->raw_out(i);
3440 if (m!= nullptr && m->Opcode() == op &&
3441 m->bottom_type()->make_ptr() == t) {
3442 nn = m;
3443 break;
3444 }
3445 }
3446 if (nn != nullptr) {
3447 // Decode a narrow oop to match address
3448 // [R12 + narrow_oop_reg<<3 + offset]
3449 if (is_oop) {
3450 nn = new DecodeNNode(nn, t);
3451 } else {
3452 nn = new DecodeNKlassNode(nn, t);
3453 }
3454 // Check for succeeding AddP which uses the same Base.
3455 // Otherwise we will run into the assertion above when visiting that guy.
3456 for (uint i = 0; i < n->outcnt(); ++i) {
3457 Node *out_i = n->raw_out(i);
3458 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3459 out_i->set_req(AddPNode::Base, nn);
3460 #ifdef ASSERT
3461 for (uint j = 0; j < out_i->outcnt(); ++j) {
3462 Node *out_j = out_i->raw_out(j);
3463 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3464 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3465 }
3466 #endif
3467 }
3468 }
3469 n->set_req(AddPNode::Base, nn);
3470 n->set_req(AddPNode::Address, nn);
3471 if (addp->outcnt() == 0) {
3472 addp->disconnect_inputs(this);
3473 }
3474 }
3475 }
3476 }
3477 #endif
3478 break;
3479 }
3480
3481 case Op_CastPP: {
3482 // Remove CastPP nodes to gain more freedom during scheduling but
3483 // keep the dependency they encode as control or precedence edges
3484 // (if control is set already) on memory operations. Some CastPP
3485 // nodes don't have a control (don't carry a dependency): skip
3486 // those.
3487 if (n->in(0) != nullptr) {
3488 ResourceMark rm;
3489 Unique_Node_List wq;
3490 wq.push(n);
3491 for (uint next = 0; next < wq.size(); ++next) {
3492 Node *m = wq.at(next);
3493 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3494 Node* use = m->fast_out(i);
3495 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3496 use->ensure_control_or_add_prec(n->in(0));
3497 } else {
3498 switch(use->Opcode()) {
3499 case Op_AddP:
3500 case Op_DecodeN:
3501 case Op_DecodeNKlass:
3502 case Op_CheckCastPP:
3503 case Op_CastPP:
3504 wq.push(use);
3505 break;
3506 }
3507 }
3508 }
3509 }
3510 }
3511 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3512 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3513 Node* in1 = n->in(1);
3514 const Type* t = n->bottom_type();
3515 Node* new_in1 = in1->clone();
3516 new_in1->as_DecodeN()->set_type(t);
3517
3518 if (!Matcher::narrow_oop_use_complex_address()) {
3519 //
3520 // x86, ARM and friends can handle 2 adds in addressing mode
3521 // and Matcher can fold a DecodeN node into address by using
3522 // a narrow oop directly and do implicit null check in address:
3523 //
3524 // [R12 + narrow_oop_reg<<3 + offset]
3525 // NullCheck narrow_oop_reg
3526 //
3527 // On other platforms (Sparc) we have to keep new DecodeN node and
3528 // use it to do implicit null check in address:
3529 //
3530 // decode_not_null narrow_oop_reg, base_reg
3531 // [base_reg + offset]
3532 // NullCheck base_reg
3533 //
3534 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3535 // to keep the information to which null check the new DecodeN node
3536 // corresponds to use it as value in implicit_null_check().
3537 //
3538 new_in1->set_req(0, n->in(0));
3539 }
3540
3541 n->subsume_by(new_in1, this);
3542 if (in1->outcnt() == 0) {
3543 in1->disconnect_inputs(this);
3544 }
3545 } else {
3546 n->subsume_by(n->in(1), this);
3547 if (n->outcnt() == 0) {
3548 n->disconnect_inputs(this);
3549 }
3550 }
3551 break;
3552 }
3553 case Op_CastII: {
3554 n->as_CastII()->remove_range_check_cast(this);
3555 break;
3556 }
3557 #ifdef _LP64
3558 case Op_CmpP:
3559 // Do this transformation here to preserve CmpPNode::sub() and
3560 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3561 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3562 Node* in1 = n->in(1);
3563 Node* in2 = n->in(2);
3564 if (!in1->is_DecodeNarrowPtr()) {
3565 in2 = in1;
3566 in1 = n->in(2);
3567 }
3568 assert(in1->is_DecodeNarrowPtr(), "sanity");
3569
3570 Node* new_in2 = nullptr;
3571 if (in2->is_DecodeNarrowPtr()) {
3572 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3573 new_in2 = in2->in(1);
3574 } else if (in2->Opcode() == Op_ConP) {
3575 const Type* t = in2->bottom_type();
3576 if (t == TypePtr::NULL_PTR) {
3577 assert(in1->is_DecodeN(), "compare klass to null?");
3578 // Don't convert CmpP null check into CmpN if compressed
3579 // oops implicit null check is not generated.
3580 // This will allow to generate normal oop implicit null check.
3581 if (Matcher::gen_narrow_oop_implicit_null_checks())
3582 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3583 //
3584 // This transformation together with CastPP transformation above
3585 // will generated code for implicit null checks for compressed oops.
3586 //
3587 // The original code after Optimize()
3588 //
3589 // LoadN memory, narrow_oop_reg
3590 // decode narrow_oop_reg, base_reg
3591 // CmpP base_reg, nullptr
3592 // CastPP base_reg // NotNull
3593 // Load [base_reg + offset], val_reg
3594 //
3595 // after these transformations will be
3596 //
3597 // LoadN memory, narrow_oop_reg
3598 // CmpN narrow_oop_reg, nullptr
3599 // decode_not_null narrow_oop_reg, base_reg
3600 // Load [base_reg + offset], val_reg
3601 //
3602 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3603 // since narrow oops can be used in debug info now (see the code in
3604 // final_graph_reshaping_walk()).
3605 //
3606 // At the end the code will be matched to
3607 // on x86:
3608 //
3609 // Load_narrow_oop memory, narrow_oop_reg
3610 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3611 // NullCheck narrow_oop_reg
3612 //
3613 // and on sparc:
3614 //
3615 // Load_narrow_oop memory, narrow_oop_reg
3616 // decode_not_null narrow_oop_reg, base_reg
3617 // Load [base_reg + offset], val_reg
3618 // NullCheck base_reg
3619 //
3620 } else if (t->isa_oopptr()) {
3621 new_in2 = ConNode::make(t->make_narrowoop());
3622 } else if (t->isa_klassptr()) {
3623 ciKlass* klass = t->is_klassptr()->exact_klass();
3624 if (klass->is_in_encoding_range()) {
3625 new_in2 = ConNode::make(t->make_narrowklass());
3626 }
3627 }
3628 }
3629 if (new_in2 != nullptr) {
3630 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3631 n->subsume_by(cmpN, this);
3632 if (in1->outcnt() == 0) {
3633 in1->disconnect_inputs(this);
3634 }
3635 if (in2->outcnt() == 0) {
3636 in2->disconnect_inputs(this);
3637 }
3638 }
3639 }
3640 break;
3641
3642 case Op_DecodeN:
3643 case Op_DecodeNKlass:
3644 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3645 // DecodeN could be pinned when it can't be fold into
3646 // an address expression, see the code for Op_CastPP above.
3647 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3648 break;
3649
3650 case Op_EncodeP:
3651 case Op_EncodePKlass: {
3652 Node* in1 = n->in(1);
3653 if (in1->is_DecodeNarrowPtr()) {
3654 n->subsume_by(in1->in(1), this);
3655 } else if (in1->Opcode() == Op_ConP) {
3656 const Type* t = in1->bottom_type();
3657 if (t == TypePtr::NULL_PTR) {
3658 assert(t->isa_oopptr(), "null klass?");
3659 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3660 } else if (t->isa_oopptr()) {
3661 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3662 } else if (t->isa_klassptr()) {
3663 ciKlass* klass = t->is_klassptr()->exact_klass();
3664 if (klass->is_in_encoding_range()) {
3665 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3666 } else {
3667 assert(false, "unencodable klass in ConP -> EncodeP");
3668 C->record_failure("unencodable klass in ConP -> EncodeP");
3669 }
3670 }
3671 }
3672 if (in1->outcnt() == 0) {
3673 in1->disconnect_inputs(this);
3674 }
3675 break;
3676 }
3677
3678 case Op_Proj: {
3679 if (OptimizeStringConcat || IncrementalInline) {
3680 ProjNode* proj = n->as_Proj();
3681 if (proj->_is_io_use) {
3682 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3683 // Separate projections were used for the exception path which
3684 // are normally removed by a late inline. If it wasn't inlined
3685 // then they will hang around and should just be replaced with
3686 // the original one. Merge them.
3687 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3688 if (non_io_proj != nullptr) {
3689 proj->subsume_by(non_io_proj , this);
3690 }
3691 }
3692 }
3693 break;
3694 }
3695
3696 case Op_Phi:
3697 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3698 // The EncodeP optimization may create Phi with the same edges
3699 // for all paths. It is not handled well by Register Allocator.
3700 Node* unique_in = n->in(1);
3701 assert(unique_in != nullptr, "");
3702 uint cnt = n->req();
3703 for (uint i = 2; i < cnt; i++) {
3704 Node* m = n->in(i);
3705 assert(m != nullptr, "");
3706 if (unique_in != m)
3707 unique_in = nullptr;
3708 }
3709 if (unique_in != nullptr) {
3710 n->subsume_by(unique_in, this);
3711 }
3712 }
3713 break;
3714
3715 #endif
3716
3717 case Op_ModI:
3718 handle_div_mod_op(n, T_INT, false);
3719 break;
3720
3721 case Op_ModL:
3722 handle_div_mod_op(n, T_LONG, false);
3723 break;
3724
3725 case Op_UModI:
3726 handle_div_mod_op(n, T_INT, true);
3727 break;
3728
3729 case Op_UModL:
3730 handle_div_mod_op(n, T_LONG, true);
3731 break;
3732
3733 case Op_LoadVector:
3734 case Op_StoreVector:
3735 #ifdef ASSERT
3736 // Add VerifyVectorAlignment node between adr and load / store.
3737 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3738 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3739 n->as_StoreVector()->must_verify_alignment();
3740 if (must_verify_alignment) {
3741 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3742 n->as_StoreVector()->memory_size();
3743 // The memory access should be aligned to the vector width in bytes.
3744 // However, the underlying array is possibly less well aligned, but at least
3745 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3746 // a loop we can expect at least the following alignment:
3747 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3748 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3749 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3750 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3751 jlong mask = guaranteed_alignment - 1;
3752 Node* mask_con = ConLNode::make(mask);
3753 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3754 n->set_req(MemNode::Address, va);
3755 }
3756 }
3757 #endif
3758 break;
3759
3760 case Op_LoadVectorGather:
3761 case Op_StoreVectorScatter:
3762 case Op_LoadVectorGatherMasked:
3763 case Op_StoreVectorScatterMasked:
3764 case Op_VectorCmpMasked:
3765 case Op_VectorMaskGen:
3766 case Op_LoadVectorMasked:
3767 case Op_StoreVectorMasked:
3768 break;
3769
3770 case Op_AddReductionVI:
3771 case Op_AddReductionVL:
3772 case Op_AddReductionVF:
3773 case Op_AddReductionVD:
3774 case Op_MulReductionVI:
3775 case Op_MulReductionVL:
3776 case Op_MulReductionVF:
3777 case Op_MulReductionVD:
3778 case Op_MinReductionV:
3779 case Op_MaxReductionV:
3780 case Op_AndReductionV:
3781 case Op_OrReductionV:
3782 case Op_XorReductionV:
3783 break;
3784
3785 case Op_PackB:
3786 case Op_PackS:
3787 case Op_PackI:
3788 case Op_PackF:
3789 case Op_PackL:
3790 case Op_PackD:
3791 if (n->req()-1 > 2) {
3792 // Replace many operand PackNodes with a binary tree for matching
3793 PackNode* p = (PackNode*) n;
3794 Node* btp = p->binary_tree_pack(1, n->req());
3795 n->subsume_by(btp, this);
3796 }
3797 break;
3798 case Op_Loop:
3799 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3800 case Op_CountedLoop:
3801 case Op_LongCountedLoop:
3802 case Op_OuterStripMinedLoop:
3803 if (n->as_Loop()->is_inner_loop()) {
3804 frc.inc_inner_loop_count();
3805 }
3806 n->as_Loop()->verify_strip_mined(0);
3807 break;
3808 case Op_LShiftI:
3809 case Op_RShiftI:
3810 case Op_URShiftI:
3811 case Op_LShiftL:
3812 case Op_RShiftL:
3813 case Op_URShiftL:
3814 if (Matcher::need_masked_shift_count) {
3815 // The cpu's shift instructions don't restrict the count to the
3816 // lower 5/6 bits. We need to do the masking ourselves.
3817 Node* in2 = n->in(2);
3818 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3819 const TypeInt* t = in2->find_int_type();
3820 if (t != nullptr && t->is_con()) {
3821 juint shift = t->get_con();
3822 if (shift > mask) { // Unsigned cmp
3823 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3824 }
3825 } else {
3826 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3827 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3828 n->set_req(2, shift);
3829 }
3830 }
3831 if (in2->outcnt() == 0) { // Remove dead node
3832 in2->disconnect_inputs(this);
3833 }
3834 }
3835 break;
3836 case Op_MemBarStoreStore:
3837 case Op_MemBarRelease:
3838 // Break the link with AllocateNode: it is no longer useful and
3839 // confuses register allocation.
3840 if (n->req() > MemBarNode::Precedent) {
3841 n->set_req(MemBarNode::Precedent, top());
3842 }
3843 break;
3844 case Op_MemBarAcquire: {
3845 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3846 // At parse time, the trailing MemBarAcquire for a volatile load
3847 // is created with an edge to the load. After optimizations,
3848 // that input may be a chain of Phis. If those phis have no
3849 // other use, then the MemBarAcquire keeps them alive and
3850 // register allocation can be confused.
3851 dead_nodes.push(n->in(MemBarNode::Precedent));
3852 n->set_req(MemBarNode::Precedent, top());
3853 }
3854 break;
3855 }
3856 case Op_Blackhole:
3857 break;
3858 case Op_RangeCheck: {
3859 RangeCheckNode* rc = n->as_RangeCheck();
3860 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3861 n->subsume_by(iff, this);
3862 frc._tests.push(iff);
3863 break;
3864 }
3865 case Op_ConvI2L: {
3866 if (!Matcher::convi2l_type_required) {
3867 // Code generation on some platforms doesn't need accurate
3868 // ConvI2L types. Widening the type can help remove redundant
3869 // address computations.
3870 n->as_Type()->set_type(TypeLong::INT);
3871 ResourceMark rm;
3872 Unique_Node_List wq;
3873 wq.push(n);
3874 for (uint next = 0; next < wq.size(); next++) {
3875 Node *m = wq.at(next);
3876
3877 for(;;) {
3878 // Loop over all nodes with identical inputs edges as m
3879 Node* k = m->find_similar(m->Opcode());
3880 if (k == nullptr) {
3881 break;
3882 }
3883 // Push their uses so we get a chance to remove node made
3884 // redundant
3885 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3886 Node* u = k->fast_out(i);
3887 if (u->Opcode() == Op_LShiftL ||
3888 u->Opcode() == Op_AddL ||
3889 u->Opcode() == Op_SubL ||
3890 u->Opcode() == Op_AddP) {
3891 wq.push(u);
3892 }
3893 }
3894 // Replace all nodes with identical edges as m with m
3895 k->subsume_by(m, this);
3896 }
3897 }
3898 }
3899 break;
3900 }
3901 case Op_CmpUL: {
3902 if (!Matcher::has_match_rule(Op_CmpUL)) {
3903 // No support for unsigned long comparisons
3904 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3905 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3906 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3907 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3908 Node* andl = new AndLNode(orl, remove_sign_mask);
3909 Node* cmp = new CmpLNode(andl, n->in(2));
3910 n->subsume_by(cmp, this);
3911 }
3912 break;
3913 }
3914 #ifdef ASSERT
3915 case Op_ConNKlass: {
3916 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3917 ciKlass* klass = tp->is_klassptr()->exact_klass();
3918 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3919 break;
3920 }
3921 #endif
3922 default:
3923 assert(!n->is_Call(), "");
3924 assert(!n->is_Mem(), "");
3925 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3926 break;
3927 }
3928 }
3929
3930 //------------------------------final_graph_reshaping_walk---------------------
3931 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3932 // requires that the walk visits a node's inputs before visiting the node.
3933 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3934 Unique_Node_List sfpt;
3935
3936 frc._visited.set(root->_idx); // first, mark node as visited
3937 uint cnt = root->req();
3938 Node *n = root;
3939 uint i = 0;
3940 while (true) {
3941 if (i < cnt) {
3942 // Place all non-visited non-null inputs onto stack
3943 Node* m = n->in(i);
3944 ++i;
3945 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3946 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3947 // compute worst case interpreter size in case of a deoptimization
3948 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3949
3950 sfpt.push(m);
3951 }
3952 cnt = m->req();
3953 nstack.push(n, i); // put on stack parent and next input's index
3954 n = m;
3955 i = 0;
3956 }
3957 } else {
3958 // Now do post-visit work
3959 final_graph_reshaping_impl(n, frc, dead_nodes);
3960 if (nstack.is_empty())
3961 break; // finished
3962 n = nstack.node(); // Get node from stack
3963 cnt = n->req();
3964 i = nstack.index();
3965 nstack.pop(); // Shift to the next node on stack
3966 }
3967 }
3968
3969 // Skip next transformation if compressed oops are not used.
3970 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3971 (!UseCompressedOops && !UseCompressedClassPointers))
3972 return;
3973
3974 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3975 // It could be done for an uncommon traps or any safepoints/calls
3976 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3977 while (sfpt.size() > 0) {
3978 n = sfpt.pop();
3979 JVMState *jvms = n->as_SafePoint()->jvms();
3980 assert(jvms != nullptr, "sanity");
3981 int start = jvms->debug_start();
3982 int end = n->req();
3983 bool is_uncommon = (n->is_CallStaticJava() &&
3984 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3985 for (int j = start; j < end; j++) {
3986 Node* in = n->in(j);
3987 if (in->is_DecodeNarrowPtr()) {
3988 bool safe_to_skip = true;
3989 if (!is_uncommon ) {
3990 // Is it safe to skip?
3991 for (uint i = 0; i < in->outcnt(); i++) {
3992 Node* u = in->raw_out(i);
3993 if (!u->is_SafePoint() ||
3994 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3995 safe_to_skip = false;
3996 }
3997 }
3998 }
3999 if (safe_to_skip) {
4000 n->set_req(j, in->in(1));
4001 }
4002 if (in->outcnt() == 0) {
4003 in->disconnect_inputs(this);
4004 }
4005 }
4006 }
4007 }
4008 }
4009
4010 //------------------------------final_graph_reshaping--------------------------
4011 // Final Graph Reshaping.
4012 //
4013 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4014 // and not commoned up and forced early. Must come after regular
4015 // optimizations to avoid GVN undoing the cloning. Clone constant
4016 // inputs to Loop Phis; these will be split by the allocator anyways.
4017 // Remove Opaque nodes.
4018 // (2) Move last-uses by commutative operations to the left input to encourage
4019 // Intel update-in-place two-address operations and better register usage
4020 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4021 // calls canonicalizing them back.
4022 // (3) Count the number of double-precision FP ops, single-precision FP ops
4023 // and call sites. On Intel, we can get correct rounding either by
4024 // forcing singles to memory (requires extra stores and loads after each
4025 // FP bytecode) or we can set a rounding mode bit (requires setting and
4026 // clearing the mode bit around call sites). The mode bit is only used
4027 // if the relative frequency of single FP ops to calls is low enough.
4028 // This is a key transform for SPEC mpeg_audio.
4029 // (4) Detect infinite loops; blobs of code reachable from above but not
4030 // below. Several of the Code_Gen algorithms fail on such code shapes,
4031 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4032 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4033 // Detection is by looking for IfNodes where only 1 projection is
4034 // reachable from below or CatchNodes missing some targets.
4035 // (5) Assert for insane oop offsets in debug mode.
4036
4037 bool Compile::final_graph_reshaping() {
4038 // an infinite loop may have been eliminated by the optimizer,
4039 // in which case the graph will be empty.
4040 if (root()->req() == 1) {
4041 // Do not compile method that is only a trivial infinite loop,
4042 // since the content of the loop may have been eliminated.
4043 record_method_not_compilable("trivial infinite loop");
4044 return true;
4045 }
4046
4047 // Expensive nodes have their control input set to prevent the GVN
4048 // from freely commoning them. There's no GVN beyond this point so
4049 // no need to keep the control input. We want the expensive nodes to
4050 // be freely moved to the least frequent code path by gcm.
4051 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4052 for (int i = 0; i < expensive_count(); i++) {
4053 _expensive_nodes.at(i)->set_req(0, nullptr);
4054 }
4055
4056 Final_Reshape_Counts frc;
4057
4058 // Visit everybody reachable!
4059 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4060 Node_Stack nstack(live_nodes() >> 1);
4061 Unique_Node_List dead_nodes;
4062 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4063
4064 // Check for unreachable (from below) code (i.e., infinite loops).
4065 for( uint i = 0; i < frc._tests.size(); i++ ) {
4066 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4067 // Get number of CFG targets.
4068 // Note that PCTables include exception targets after calls.
4069 uint required_outcnt = n->required_outcnt();
4070 if (n->outcnt() != required_outcnt) {
4071 // Check for a few special cases. Rethrow Nodes never take the
4072 // 'fall-thru' path, so expected kids is 1 less.
4073 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4074 if (n->in(0)->in(0)->is_Call()) {
4075 CallNode* call = n->in(0)->in(0)->as_Call();
4076 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4077 required_outcnt--; // Rethrow always has 1 less kid
4078 } else if (call->req() > TypeFunc::Parms &&
4079 call->is_CallDynamicJava()) {
4080 // Check for null receiver. In such case, the optimizer has
4081 // detected that the virtual call will always result in a null
4082 // pointer exception. The fall-through projection of this CatchNode
4083 // will not be populated.
4084 Node* arg0 = call->in(TypeFunc::Parms);
4085 if (arg0->is_Type() &&
4086 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4087 required_outcnt--;
4088 }
4089 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4090 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4091 // Check for illegal array length. In such case, the optimizer has
4092 // detected that the allocation attempt will always result in an
4093 // exception. There is no fall-through projection of this CatchNode .
4094 assert(call->is_CallStaticJava(), "static call expected");
4095 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4096 uint valid_length_test_input = call->req() - 1;
4097 Node* valid_length_test = call->in(valid_length_test_input);
4098 call->del_req(valid_length_test_input);
4099 if (valid_length_test->find_int_con(1) == 0) {
4100 required_outcnt--;
4101 }
4102 dead_nodes.push(valid_length_test);
4103 assert(n->outcnt() == required_outcnt, "malformed control flow");
4104 continue;
4105 }
4106 }
4107 }
4108
4109 // Recheck with a better notion of 'required_outcnt'
4110 if (n->outcnt() != required_outcnt) {
4111 record_method_not_compilable("malformed control flow");
4112 return true; // Not all targets reachable!
4113 }
4114 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4115 CallNode* call = n->in(0)->in(0)->as_Call();
4116 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4117 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4118 assert(call->is_CallStaticJava(), "static call expected");
4119 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4120 uint valid_length_test_input = call->req() - 1;
4121 dead_nodes.push(call->in(valid_length_test_input));
4122 call->del_req(valid_length_test_input); // valid length test useless now
4123 }
4124 }
4125 // Check that I actually visited all kids. Unreached kids
4126 // must be infinite loops.
4127 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4128 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4129 record_method_not_compilable("infinite loop");
4130 return true; // Found unvisited kid; must be unreach
4131 }
4132
4133 // Here so verification code in final_graph_reshaping_walk()
4134 // always see an OuterStripMinedLoopEnd
4135 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4136 IfNode* init_iff = n->as_If();
4137 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4138 n->subsume_by(iff, this);
4139 }
4140 }
4141
4142 while (dead_nodes.size() > 0) {
4143 Node* m = dead_nodes.pop();
4144 if (m->outcnt() == 0 && m != top()) {
4145 for (uint j = 0; j < m->req(); j++) {
4146 Node* in = m->in(j);
4147 if (in != nullptr) {
4148 dead_nodes.push(in);
4149 }
4150 }
4151 m->disconnect_inputs(this);
4152 }
4153 }
4154
4155 set_java_calls(frc.get_java_call_count());
4156 set_inner_loops(frc.get_inner_loop_count());
4157
4158 // No infinite loops, no reason to bail out.
4159 return false;
4160 }
4161
4162 //-----------------------------too_many_traps----------------------------------
4163 // Report if there are too many traps at the current method and bci.
4164 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4165 bool Compile::too_many_traps(ciMethod* method,
4166 int bci,
4167 Deoptimization::DeoptReason reason) {
4168 ciMethodData* md = method->method_data();
4169 if (md->is_empty()) {
4170 // Assume the trap has not occurred, or that it occurred only
4171 // because of a transient condition during start-up in the interpreter.
4172 return false;
4173 }
4174 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4175 if (md->has_trap_at(bci, m, reason) != 0) {
4176 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4177 // Also, if there are multiple reasons, or if there is no per-BCI record,
4178 // assume the worst.
4179 if (log())
4180 log()->elem("observe trap='%s' count='%d'",
4181 Deoptimization::trap_reason_name(reason),
4182 md->trap_count(reason));
4183 return true;
4184 } else {
4185 // Ignore method/bci and see if there have been too many globally.
4186 return too_many_traps(reason, md);
4187 }
4188 }
4189
4190 // Less-accurate variant which does not require a method and bci.
4191 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4192 ciMethodData* logmd) {
4193 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4194 // Too many traps globally.
4195 // Note that we use cumulative trap_count, not just md->trap_count.
4196 if (log()) {
4197 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4198 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4199 Deoptimization::trap_reason_name(reason),
4200 mcount, trap_count(reason));
4201 }
4202 return true;
4203 } else {
4204 // The coast is clear.
4205 return false;
4206 }
4207 }
4208
4209 //--------------------------too_many_recompiles--------------------------------
4210 // Report if there are too many recompiles at the current method and bci.
4211 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4212 // Is not eager to return true, since this will cause the compiler to use
4213 // Action_none for a trap point, to avoid too many recompilations.
4214 bool Compile::too_many_recompiles(ciMethod* method,
4215 int bci,
4216 Deoptimization::DeoptReason reason) {
4217 ciMethodData* md = method->method_data();
4218 if (md->is_empty()) {
4219 // Assume the trap has not occurred, or that it occurred only
4220 // because of a transient condition during start-up in the interpreter.
4221 return false;
4222 }
4223 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4224 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4225 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4226 Deoptimization::DeoptReason per_bc_reason
4227 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4228 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4229 if ((per_bc_reason == Deoptimization::Reason_none
4230 || md->has_trap_at(bci, m, reason) != 0)
4231 // The trap frequency measure we care about is the recompile count:
4232 && md->trap_recompiled_at(bci, m)
4233 && md->overflow_recompile_count() >= bc_cutoff) {
4234 // Do not emit a trap here if it has already caused recompilations.
4235 // Also, if there are multiple reasons, or if there is no per-BCI record,
4236 // assume the worst.
4237 if (log())
4238 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4239 Deoptimization::trap_reason_name(reason),
4240 md->trap_count(reason),
4241 md->overflow_recompile_count());
4242 return true;
4243 } else if (trap_count(reason) != 0
4244 && decompile_count() >= m_cutoff) {
4245 // Too many recompiles globally, and we have seen this sort of trap.
4246 // Use cumulative decompile_count, not just md->decompile_count.
4247 if (log())
4248 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4249 Deoptimization::trap_reason_name(reason),
4250 md->trap_count(reason), trap_count(reason),
4251 md->decompile_count(), decompile_count());
4252 return true;
4253 } else {
4254 // The coast is clear.
4255 return false;
4256 }
4257 }
4258
4259 // Compute when not to trap. Used by matching trap based nodes and
4260 // NullCheck optimization.
4261 void Compile::set_allowed_deopt_reasons() {
4262 _allowed_reasons = 0;
4263 if (is_method_compilation()) {
4264 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4265 assert(rs < BitsPerInt, "recode bit map");
4266 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4267 _allowed_reasons |= nth_bit(rs);
4268 }
4269 }
4270 }
4271 }
4272
4273 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4274 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4275 }
4276
4277 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4278 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4279 }
4280
4281 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4282 if (holder->is_initialized()) {
4283 return false;
4284 }
4285 if (holder->is_being_initialized()) {
4286 if (accessing_method->holder() == holder) {
4287 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4288 // <init>, or a static method. In all those cases, there was an initialization
4289 // barrier on the holder klass passed.
4290 if (accessing_method->is_static_initializer() ||
4291 accessing_method->is_object_initializer() ||
4292 accessing_method->is_static()) {
4293 return false;
4294 }
4295 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4296 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4297 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4298 // child class can become fully initialized while its parent class is still being initialized.
4299 if (accessing_method->is_static_initializer()) {
4300 return false;
4301 }
4302 }
4303 ciMethod* root = method(); // the root method of compilation
4304 if (root != accessing_method) {
4305 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4306 }
4307 }
4308 return true;
4309 }
4310
4311 #ifndef PRODUCT
4312 //------------------------------verify_bidirectional_edges---------------------
4313 // For each input edge to a node (ie - for each Use-Def edge), verify that
4314 // there is a corresponding Def-Use edge.
4315 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4316 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4317 uint stack_size = live_nodes() >> 4;
4318 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4319 if (root_and_safepoints != nullptr) {
4320 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4321 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4322 Node* root_or_safepoint = root_and_safepoints->at(i);
4323 // If the node is a safepoint, let's check if it still has a control input
4324 // Lack of control input signifies that this node was killed by CCP or
4325 // recursively by remove_globally_dead_node and it shouldn't be a starting
4326 // point.
4327 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4328 nstack.push(root_or_safepoint);
4329 }
4330 }
4331 } else {
4332 nstack.push(_root);
4333 }
4334
4335 while (nstack.size() > 0) {
4336 Node* n = nstack.pop();
4337 if (visited.member(n)) {
4338 continue;
4339 }
4340 visited.push(n);
4341
4342 // Walk over all input edges, checking for correspondence
4343 uint length = n->len();
4344 for (uint i = 0; i < length; i++) {
4345 Node* in = n->in(i);
4346 if (in != nullptr && !visited.member(in)) {
4347 nstack.push(in); // Put it on stack
4348 }
4349 if (in != nullptr && !in->is_top()) {
4350 // Count instances of `next`
4351 int cnt = 0;
4352 for (uint idx = 0; idx < in->_outcnt; idx++) {
4353 if (in->_out[idx] == n) {
4354 cnt++;
4355 }
4356 }
4357 assert(cnt > 0, "Failed to find Def-Use edge.");
4358 // Check for duplicate edges
4359 // walk the input array downcounting the input edges to n
4360 for (uint j = 0; j < length; j++) {
4361 if (n->in(j) == in) {
4362 cnt--;
4363 }
4364 }
4365 assert(cnt == 0, "Mismatched edge count.");
4366 } else if (in == nullptr) {
4367 assert(i == 0 || i >= n->req() ||
4368 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4369 (n->is_Unlock() && i == (n->req() - 1)) ||
4370 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4371 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4372 } else {
4373 assert(in->is_top(), "sanity");
4374 // Nothing to check.
4375 }
4376 }
4377 }
4378 }
4379
4380 //------------------------------verify_graph_edges---------------------------
4381 // Walk the Graph and verify that there is a one-to-one correspondence
4382 // between Use-Def edges and Def-Use edges in the graph.
4383 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4384 if (VerifyGraphEdges) {
4385 Unique_Node_List visited;
4386
4387 // Call graph walk to check edges
4388 verify_bidirectional_edges(visited, root_and_safepoints);
4389 if (no_dead_code) {
4390 // Now make sure that no visited node is used by an unvisited node.
4391 bool dead_nodes = false;
4392 Unique_Node_List checked;
4393 while (visited.size() > 0) {
4394 Node* n = visited.pop();
4395 checked.push(n);
4396 for (uint i = 0; i < n->outcnt(); i++) {
4397 Node* use = n->raw_out(i);
4398 if (checked.member(use)) continue; // already checked
4399 if (visited.member(use)) continue; // already in the graph
4400 if (use->is_Con()) continue; // a dead ConNode is OK
4401 // At this point, we have found a dead node which is DU-reachable.
4402 if (!dead_nodes) {
4403 tty->print_cr("*** Dead nodes reachable via DU edges:");
4404 dead_nodes = true;
4405 }
4406 use->dump(2);
4407 tty->print_cr("---");
4408 checked.push(use); // No repeats; pretend it is now checked.
4409 }
4410 }
4411 assert(!dead_nodes, "using nodes must be reachable from root");
4412 }
4413 }
4414 }
4415 #endif
4416
4417 // The Compile object keeps track of failure reasons separately from the ciEnv.
4418 // This is required because there is not quite a 1-1 relation between the
4419 // ciEnv and its compilation task and the Compile object. Note that one
4420 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4421 // to backtrack and retry without subsuming loads. Other than this backtracking
4422 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4423 // by the logic in C2Compiler.
4424 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4425 if (log() != nullptr) {
4426 log()->elem("failure reason='%s' phase='compile'", reason);
4427 }
4428 if (_failure_reason.get() == nullptr) {
4429 // Record the first failure reason.
4430 _failure_reason.set(reason);
4431 if (CaptureBailoutInformation) {
4432 _first_failure_details = new CompilationFailureInfo(reason);
4433 }
4434 } else {
4435 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4436 }
4437
4438 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4439 C->print_method(PHASE_FAILURE, 1);
4440 }
4441 _root = nullptr; // flush the graph, too
4442 }
4443
4444 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4445 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4446 _compile(Compile::current()),
4447 _log(nullptr),
4448 _dolog(CITimeVerbose)
4449 {
4450 assert(_compile != nullptr, "sanity check");
4451 assert(id != PhaseTraceId::_t_none, "Don't use none");
4452 if (_dolog) {
4453 _log = _compile->log();
4454 }
4455 if (_log != nullptr) {
4456 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4457 _log->stamp();
4458 _log->end_head();
4459 }
4460
4461 // Inform memory statistic, if enabled
4462 if (CompilationMemoryStatistic::enabled()) {
4463 CompilationMemoryStatistic::on_phase_start((int)id, name);
4464 }
4465 }
4466
4467 Compile::TracePhase::TracePhase(PhaseTraceId id)
4468 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4469
4470 Compile::TracePhase::~TracePhase() {
4471
4472 // Inform memory statistic, if enabled
4473 if (CompilationMemoryStatistic::enabled()) {
4474 CompilationMemoryStatistic::on_phase_end();
4475 }
4476
4477 if (_compile->failing_internal()) {
4478 if (_log != nullptr) {
4479 _log->done("phase");
4480 }
4481 return; // timing code, not stressing bailouts.
4482 }
4483 #ifdef ASSERT
4484 if (PrintIdealNodeCount) {
4485 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4486 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4487 }
4488
4489 if (VerifyIdealNodeCount) {
4490 _compile->print_missing_nodes();
4491 }
4492 #endif
4493
4494 if (_log != nullptr) {
4495 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4496 }
4497 }
4498
4499 //----------------------------static_subtype_check-----------------------------
4500 // Shortcut important common cases when superklass is exact:
4501 // (0) superklass is java.lang.Object (can occur in reflective code)
4502 // (1) subklass is already limited to a subtype of superklass => always ok
4503 // (2) subklass does not overlap with superklass => always fail
4504 // (3) superklass has NO subtypes and we can check with a simple compare.
4505 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4506 if (skip) {
4507 return SSC_full_test; // Let caller generate the general case.
4508 }
4509
4510 if (subk->is_java_subtype_of(superk)) {
4511 return SSC_always_true; // (0) and (1) this test cannot fail
4512 }
4513
4514 if (!subk->maybe_java_subtype_of(superk)) {
4515 return SSC_always_false; // (2) true path dead; no dynamic test needed
4516 }
4517
4518 const Type* superelem = superk;
4519 if (superk->isa_aryklassptr()) {
4520 int ignored;
4521 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4522 }
4523
4524 if (superelem->isa_instklassptr()) {
4525 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4526 if (!ik->has_subklass()) {
4527 if (!ik->is_final()) {
4528 // Add a dependency if there is a chance of a later subclass.
4529 dependencies()->assert_leaf_type(ik);
4530 }
4531 if (!superk->maybe_java_subtype_of(subk)) {
4532 return SSC_always_false;
4533 }
4534 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4535 }
4536 } else {
4537 // A primitive array type has no subtypes.
4538 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4539 }
4540
4541 return SSC_full_test;
4542 }
4543
4544 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4545 #ifdef _LP64
4546 // The scaled index operand to AddP must be a clean 64-bit value.
4547 // Java allows a 32-bit int to be incremented to a negative
4548 // value, which appears in a 64-bit register as a large
4549 // positive number. Using that large positive number as an
4550 // operand in pointer arithmetic has bad consequences.
4551 // On the other hand, 32-bit overflow is rare, and the possibility
4552 // can often be excluded, if we annotate the ConvI2L node with
4553 // a type assertion that its value is known to be a small positive
4554 // number. (The prior range check has ensured this.)
4555 // This assertion is used by ConvI2LNode::Ideal.
4556 int index_max = max_jint - 1; // array size is max_jint, index is one less
4557 if (sizetype != nullptr && sizetype->_hi > 0) {
4558 index_max = sizetype->_hi - 1;
4559 }
4560 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4561 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4562 #endif
4563 return idx;
4564 }
4565
4566 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4567 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4568 if (ctrl != nullptr) {
4569 // Express control dependency by a CastII node with a narrow type.
4570 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4571 // node from floating above the range check during loop optimizations. Otherwise, the
4572 // ConvI2L node may be eliminated independently of the range check, causing the data path
4573 // to become TOP while the control path is still there (although it's unreachable).
4574 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4575 value = phase->transform(value);
4576 }
4577 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4578 return phase->transform(new ConvI2LNode(value, ltype));
4579 }
4580
4581 void Compile::dump_print_inlining() {
4582 inline_printer()->print_on(tty);
4583 }
4584
4585 void Compile::log_late_inline(CallGenerator* cg) {
4586 if (log() != nullptr) {
4587 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4588 cg->unique_id());
4589 JVMState* p = cg->call_node()->jvms();
4590 while (p != nullptr) {
4591 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4592 p = p->caller();
4593 }
4594 log()->tail("late_inline");
4595 }
4596 }
4597
4598 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4599 log_late_inline(cg);
4600 if (log() != nullptr) {
4601 log()->inline_fail(msg);
4602 }
4603 }
4604
4605 void Compile::log_inline_id(CallGenerator* cg) {
4606 if (log() != nullptr) {
4607 // The LogCompilation tool needs a unique way to identify late
4608 // inline call sites. This id must be unique for this call site in
4609 // this compilation. Try to have it unique across compilations as
4610 // well because it can be convenient when grepping through the log
4611 // file.
4612 // Distinguish OSR compilations from others in case CICountOSR is
4613 // on.
4614 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4615 cg->set_unique_id(id);
4616 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4617 }
4618 }
4619
4620 void Compile::log_inline_failure(const char* msg) {
4621 if (C->log() != nullptr) {
4622 C->log()->inline_fail(msg);
4623 }
4624 }
4625
4626
4627 // Dump inlining replay data to the stream.
4628 // Don't change thread state and acquire any locks.
4629 void Compile::dump_inline_data(outputStream* out) {
4630 InlineTree* inl_tree = ilt();
4631 if (inl_tree != nullptr) {
4632 out->print(" inline %d", inl_tree->count());
4633 inl_tree->dump_replay_data(out);
4634 }
4635 }
4636
4637 void Compile::dump_inline_data_reduced(outputStream* out) {
4638 assert(ReplayReduce, "");
4639
4640 InlineTree* inl_tree = ilt();
4641 if (inl_tree == nullptr) {
4642 return;
4643 }
4644 // Enable iterative replay file reduction
4645 // Output "compile" lines for depth 1 subtrees,
4646 // simulating that those trees were compiled
4647 // instead of inlined.
4648 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4649 InlineTree* sub = inl_tree->subtrees().at(i);
4650 if (sub->inline_level() != 1) {
4651 continue;
4652 }
4653
4654 ciMethod* method = sub->method();
4655 int entry_bci = -1;
4656 int comp_level = env()->task()->comp_level();
4657 out->print("compile ");
4658 method->dump_name_as_ascii(out);
4659 out->print(" %d %d", entry_bci, comp_level);
4660 out->print(" inline %d", sub->count());
4661 sub->dump_replay_data(out, -1);
4662 out->cr();
4663 }
4664 }
4665
4666 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4667 if (n1->Opcode() < n2->Opcode()) return -1;
4668 else if (n1->Opcode() > n2->Opcode()) return 1;
4669
4670 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4671 for (uint i = 1; i < n1->req(); i++) {
4672 if (n1->in(i) < n2->in(i)) return -1;
4673 else if (n1->in(i) > n2->in(i)) return 1;
4674 }
4675
4676 return 0;
4677 }
4678
4679 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4680 Node* n1 = *n1p;
4681 Node* n2 = *n2p;
4682
4683 return cmp_expensive_nodes(n1, n2);
4684 }
4685
4686 void Compile::sort_expensive_nodes() {
4687 if (!expensive_nodes_sorted()) {
4688 _expensive_nodes.sort(cmp_expensive_nodes);
4689 }
4690 }
4691
4692 bool Compile::expensive_nodes_sorted() const {
4693 for (int i = 1; i < _expensive_nodes.length(); i++) {
4694 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4695 return false;
4696 }
4697 }
4698 return true;
4699 }
4700
4701 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4702 if (_expensive_nodes.length() == 0) {
4703 return false;
4704 }
4705
4706 assert(OptimizeExpensiveOps, "optimization off?");
4707
4708 // Take this opportunity to remove dead nodes from the list
4709 int j = 0;
4710 for (int i = 0; i < _expensive_nodes.length(); i++) {
4711 Node* n = _expensive_nodes.at(i);
4712 if (!n->is_unreachable(igvn)) {
4713 assert(n->is_expensive(), "should be expensive");
4714 _expensive_nodes.at_put(j, n);
4715 j++;
4716 }
4717 }
4718 _expensive_nodes.trunc_to(j);
4719
4720 // Then sort the list so that similar nodes are next to each other
4721 // and check for at least two nodes of identical kind with same data
4722 // inputs.
4723 sort_expensive_nodes();
4724
4725 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4726 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4727 return true;
4728 }
4729 }
4730
4731 return false;
4732 }
4733
4734 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4735 if (_expensive_nodes.length() == 0) {
4736 return;
4737 }
4738
4739 assert(OptimizeExpensiveOps, "optimization off?");
4740
4741 // Sort to bring similar nodes next to each other and clear the
4742 // control input of nodes for which there's only a single copy.
4743 sort_expensive_nodes();
4744
4745 int j = 0;
4746 int identical = 0;
4747 int i = 0;
4748 bool modified = false;
4749 for (; i < _expensive_nodes.length()-1; i++) {
4750 assert(j <= i, "can't write beyond current index");
4751 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4752 identical++;
4753 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4754 continue;
4755 }
4756 if (identical > 0) {
4757 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4758 identical = 0;
4759 } else {
4760 Node* n = _expensive_nodes.at(i);
4761 igvn.replace_input_of(n, 0, nullptr);
4762 igvn.hash_insert(n);
4763 modified = true;
4764 }
4765 }
4766 if (identical > 0) {
4767 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4768 } else if (_expensive_nodes.length() >= 1) {
4769 Node* n = _expensive_nodes.at(i);
4770 igvn.replace_input_of(n, 0, nullptr);
4771 igvn.hash_insert(n);
4772 modified = true;
4773 }
4774 _expensive_nodes.trunc_to(j);
4775 if (modified) {
4776 igvn.optimize();
4777 }
4778 }
4779
4780 void Compile::add_expensive_node(Node * n) {
4781 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4782 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4783 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4784 if (OptimizeExpensiveOps) {
4785 _expensive_nodes.append(n);
4786 } else {
4787 // Clear control input and let IGVN optimize expensive nodes if
4788 // OptimizeExpensiveOps is off.
4789 n->set_req(0, nullptr);
4790 }
4791 }
4792
4793 /**
4794 * Track coarsened Lock and Unlock nodes.
4795 */
4796
4797 class Lock_List : public Node_List {
4798 uint _origin_cnt;
4799 public:
4800 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4801 uint origin_cnt() const { return _origin_cnt; }
4802 };
4803
4804 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4805 int length = locks.length();
4806 if (length > 0) {
4807 // Have to keep this list until locks elimination during Macro nodes elimination.
4808 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4809 AbstractLockNode* alock = locks.at(0);
4810 BoxLockNode* box = alock->box_node()->as_BoxLock();
4811 for (int i = 0; i < length; i++) {
4812 AbstractLockNode* lock = locks.at(i);
4813 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4814 locks_list->push(lock);
4815 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4816 if (this_box != box) {
4817 // Locking regions (BoxLock) could be Unbalanced here:
4818 // - its coarsened locks were eliminated in earlier
4819 // macro nodes elimination followed by loop unroll
4820 // - it is OSR locking region (no Lock node)
4821 // Preserve Unbalanced status in such cases.
4822 if (!this_box->is_unbalanced()) {
4823 this_box->set_coarsened();
4824 }
4825 if (!box->is_unbalanced()) {
4826 box->set_coarsened();
4827 }
4828 }
4829 }
4830 _coarsened_locks.append(locks_list);
4831 }
4832 }
4833
4834 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4835 int count = coarsened_count();
4836 for (int i = 0; i < count; i++) {
4837 Node_List* locks_list = _coarsened_locks.at(i);
4838 for (uint j = 0; j < locks_list->size(); j++) {
4839 Node* lock = locks_list->at(j);
4840 assert(lock->is_AbstractLock(), "sanity");
4841 if (!useful.member(lock)) {
4842 locks_list->yank(lock);
4843 }
4844 }
4845 }
4846 }
4847
4848 void Compile::remove_coarsened_lock(Node* n) {
4849 if (n->is_AbstractLock()) {
4850 int count = coarsened_count();
4851 for (int i = 0; i < count; i++) {
4852 Node_List* locks_list = _coarsened_locks.at(i);
4853 locks_list->yank(n);
4854 }
4855 }
4856 }
4857
4858 bool Compile::coarsened_locks_consistent() {
4859 int count = coarsened_count();
4860 for (int i = 0; i < count; i++) {
4861 bool unbalanced = false;
4862 bool modified = false; // track locks kind modifications
4863 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4864 uint size = locks_list->size();
4865 if (size == 0) {
4866 unbalanced = false; // All locks were eliminated - good
4867 } else if (size != locks_list->origin_cnt()) {
4868 unbalanced = true; // Some locks were removed from list
4869 } else {
4870 for (uint j = 0; j < size; j++) {
4871 Node* lock = locks_list->at(j);
4872 // All nodes in group should have the same state (modified or not)
4873 if (!lock->as_AbstractLock()->is_coarsened()) {
4874 if (j == 0) {
4875 // first on list was modified, the rest should be too for consistency
4876 modified = true;
4877 } else if (!modified) {
4878 // this lock was modified but previous locks on the list were not
4879 unbalanced = true;
4880 break;
4881 }
4882 } else if (modified) {
4883 // previous locks on list were modified but not this lock
4884 unbalanced = true;
4885 break;
4886 }
4887 }
4888 }
4889 if (unbalanced) {
4890 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4891 #ifdef ASSERT
4892 if (PrintEliminateLocks) {
4893 tty->print_cr("=== unbalanced coarsened locks ===");
4894 for (uint l = 0; l < size; l++) {
4895 locks_list->at(l)->dump();
4896 }
4897 }
4898 #endif
4899 record_failure(C2Compiler::retry_no_locks_coarsening());
4900 return false;
4901 }
4902 }
4903 return true;
4904 }
4905
4906 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4907 // locks coarsening optimization removed Lock/Unlock nodes from them.
4908 // Such regions become unbalanced because coarsening only removes part
4909 // of Lock/Unlock nodes in region. As result we can't execute other
4910 // locks elimination optimizations which assume all code paths have
4911 // corresponding pair of Lock/Unlock nodes - they are balanced.
4912 void Compile::mark_unbalanced_boxes() const {
4913 int count = coarsened_count();
4914 for (int i = 0; i < count; i++) {
4915 Node_List* locks_list = _coarsened_locks.at(i);
4916 uint size = locks_list->size();
4917 if (size > 0) {
4918 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4919 BoxLockNode* box = alock->box_node()->as_BoxLock();
4920 if (alock->is_coarsened()) {
4921 // coarsened_locks_consistent(), which is called before this method, verifies
4922 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4923 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4924 for (uint j = 1; j < size; j++) {
4925 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4926 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4927 if (box != this_box) {
4928 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4929 box->set_unbalanced();
4930 this_box->set_unbalanced();
4931 }
4932 }
4933 }
4934 }
4935 }
4936 }
4937
4938 /**
4939 * Remove the speculative part of types and clean up the graph
4940 */
4941 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4942 if (UseTypeSpeculation) {
4943 Unique_Node_List worklist;
4944 worklist.push(root());
4945 int modified = 0;
4946 // Go over all type nodes that carry a speculative type, drop the
4947 // speculative part of the type and enqueue the node for an igvn
4948 // which may optimize it out.
4949 for (uint next = 0; next < worklist.size(); ++next) {
4950 Node *n = worklist.at(next);
4951 if (n->is_Type()) {
4952 TypeNode* tn = n->as_Type();
4953 const Type* t = tn->type();
4954 const Type* t_no_spec = t->remove_speculative();
4955 if (t_no_spec != t) {
4956 bool in_hash = igvn.hash_delete(n);
4957 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4958 tn->set_type(t_no_spec);
4959 igvn.hash_insert(n);
4960 igvn._worklist.push(n); // give it a chance to go away
4961 modified++;
4962 }
4963 }
4964 // Iterate over outs - endless loops is unreachable from below
4965 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4966 Node *m = n->fast_out(i);
4967 if (not_a_node(m)) {
4968 continue;
4969 }
4970 worklist.push(m);
4971 }
4972 }
4973 // Drop the speculative part of all types in the igvn's type table
4974 igvn.remove_speculative_types();
4975 if (modified > 0) {
4976 igvn.optimize();
4977 if (failing()) return;
4978 }
4979 #ifdef ASSERT
4980 // Verify that after the IGVN is over no speculative type has resurfaced
4981 worklist.clear();
4982 worklist.push(root());
4983 for (uint next = 0; next < worklist.size(); ++next) {
4984 Node *n = worklist.at(next);
4985 const Type* t = igvn.type_or_null(n);
4986 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
4987 if (n->is_Type()) {
4988 t = n->as_Type()->type();
4989 assert(t == t->remove_speculative(), "no more speculative types");
4990 }
4991 // Iterate over outs - endless loops is unreachable from below
4992 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4993 Node *m = n->fast_out(i);
4994 if (not_a_node(m)) {
4995 continue;
4996 }
4997 worklist.push(m);
4998 }
4999 }
5000 igvn.check_no_speculative_types();
5001 #endif
5002 }
5003 }
5004
5005 // Auxiliary methods to support randomized stressing/fuzzing.
5006
5007 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5008 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5009 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5010 FLAG_SET_ERGO(StressSeed, _stress_seed);
5011 } else {
5012 _stress_seed = StressSeed;
5013 }
5014 if (_log != nullptr) {
5015 _log->elem("stress_test seed='%u'", _stress_seed);
5016 }
5017 }
5018
5019 int Compile::random() {
5020 _stress_seed = os::next_random(_stress_seed);
5021 return static_cast<int>(_stress_seed);
5022 }
5023
5024 // This method can be called the arbitrary number of times, with current count
5025 // as the argument. The logic allows selecting a single candidate from the
5026 // running list of candidates as follows:
5027 // int count = 0;
5028 // Cand* selected = null;
5029 // while(cand = cand->next()) {
5030 // if (randomized_select(++count)) {
5031 // selected = cand;
5032 // }
5033 // }
5034 //
5035 // Including count equalizes the chances any candidate is "selected".
5036 // This is useful when we don't have the complete list of candidates to choose
5037 // from uniformly. In this case, we need to adjust the randomicity of the
5038 // selection, or else we will end up biasing the selection towards the latter
5039 // candidates.
5040 //
5041 // Quick back-envelope calculation shows that for the list of n candidates
5042 // the equal probability for the candidate to persist as "best" can be
5043 // achieved by replacing it with "next" k-th candidate with the probability
5044 // of 1/k. It can be easily shown that by the end of the run, the
5045 // probability for any candidate is converged to 1/n, thus giving the
5046 // uniform distribution among all the candidates.
5047 //
5048 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5049 #define RANDOMIZED_DOMAIN_POW 29
5050 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5051 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5052 bool Compile::randomized_select(int count) {
5053 assert(count > 0, "only positive");
5054 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5055 }
5056
5057 #ifdef ASSERT
5058 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5059 bool Compile::fail_randomly() {
5060 if ((random() % StressBailoutMean) != 0) {
5061 return false;
5062 }
5063 record_failure("StressBailout");
5064 return true;
5065 }
5066
5067 bool Compile::failure_is_artificial() {
5068 return C->failure_reason_is("StressBailout");
5069 }
5070 #endif
5071
5072 CloneMap& Compile::clone_map() { return _clone_map; }
5073 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5074
5075 void NodeCloneInfo::dump_on(outputStream* st) const {
5076 st->print(" {%d:%d} ", idx(), gen());
5077 }
5078
5079 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5080 uint64_t val = value(old->_idx);
5081 NodeCloneInfo cio(val);
5082 assert(val != 0, "old node should be in the map");
5083 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5084 insert(nnn->_idx, cin.get());
5085 #ifndef PRODUCT
5086 if (is_debug()) {
5087 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5088 }
5089 #endif
5090 }
5091
5092 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5093 NodeCloneInfo cio(value(old->_idx));
5094 if (cio.get() == 0) {
5095 cio.set(old->_idx, 0);
5096 insert(old->_idx, cio.get());
5097 #ifndef PRODUCT
5098 if (is_debug()) {
5099 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5100 }
5101 #endif
5102 }
5103 clone(old, nnn, gen);
5104 }
5105
5106 int CloneMap::max_gen() const {
5107 int g = 0;
5108 DictI di(_dict);
5109 for(; di.test(); ++di) {
5110 int t = gen(di._key);
5111 if (g < t) {
5112 g = t;
5113 #ifndef PRODUCT
5114 if (is_debug()) {
5115 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5116 }
5117 #endif
5118 }
5119 }
5120 return g;
5121 }
5122
5123 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5124 uint64_t val = value(key);
5125 if (val != 0) {
5126 NodeCloneInfo ni(val);
5127 ni.dump_on(st);
5128 }
5129 }
5130
5131 void Compile::shuffle_macro_nodes() {
5132 if (_macro_nodes.length() < 2) {
5133 return;
5134 }
5135 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5136 uint j = C->random() % (i + 1);
5137 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5138 }
5139 }
5140
5141 // Move Allocate nodes to the start of the list
5142 void Compile::sort_macro_nodes() {
5143 int count = macro_count();
5144 int allocates = 0;
5145 for (int i = 0; i < count; i++) {
5146 Node* n = macro_node(i);
5147 if (n->is_Allocate()) {
5148 if (i != allocates) {
5149 Node* tmp = macro_node(allocates);
5150 _macro_nodes.at_put(allocates, n);
5151 _macro_nodes.at_put(i, tmp);
5152 }
5153 allocates++;
5154 }
5155 }
5156 }
5157
5158 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5159 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5160 EventCompilerPhase event(UNTIMED);
5161 if (event.should_commit()) {
5162 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5163 }
5164 #ifndef PRODUCT
5165 ResourceMark rm;
5166 stringStream ss;
5167 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5168 int iter = ++_igv_phase_iter[cpt];
5169 if (iter > 1) {
5170 ss.print(" %d", iter);
5171 }
5172 if (n != nullptr) {
5173 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5174 if (n->is_Call()) {
5175 CallNode* call = n->as_Call();
5176 if (call->_name != nullptr) {
5177 // E.g. uncommon traps etc.
5178 ss.print(" - %s", call->_name);
5179 } else if (call->is_CallJava()) {
5180 CallJavaNode* call_java = call->as_CallJava();
5181 if (call_java->method() != nullptr) {
5182 ss.print(" -");
5183 call_java->method()->print_short_name(&ss);
5184 }
5185 }
5186 }
5187 }
5188
5189 const char* name = ss.as_string();
5190 if (should_print_igv(level)) {
5191 _igv_printer->print_graph(name);
5192 }
5193 if (should_print_phase(level)) {
5194 print_phase(name);
5195 }
5196 if (should_print_ideal_phase(cpt)) {
5197 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5198 }
5199 #endif
5200 C->_latest_stage_start_counter.stamp();
5201 }
5202
5203 // Only used from CompileWrapper
5204 void Compile::begin_method() {
5205 #ifndef PRODUCT
5206 if (_method != nullptr && should_print_igv(1)) {
5207 _igv_printer->begin_method();
5208 }
5209 #endif
5210 C->_latest_stage_start_counter.stamp();
5211 }
5212
5213 // Only used from CompileWrapper
5214 void Compile::end_method() {
5215 EventCompilerPhase event(UNTIMED);
5216 if (event.should_commit()) {
5217 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5218 }
5219
5220 #ifndef PRODUCT
5221 if (_method != nullptr && should_print_igv(1)) {
5222 _igv_printer->end_method();
5223 }
5224 #endif
5225 }
5226
5227 #ifndef PRODUCT
5228 bool Compile::should_print_phase(const int level) const {
5229 return PrintPhaseLevel > 0 && directive()->PhasePrintLevelOption >= level &&
5230 _method != nullptr; // Do not print phases for stubs.
5231 }
5232
5233 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5234 return _directive->should_print_ideal_phase(cpt);
5235 }
5236
5237 void Compile::init_igv() {
5238 if (_igv_printer == nullptr) {
5239 _igv_printer = IdealGraphPrinter::printer();
5240 _igv_printer->set_compile(this);
5241 }
5242 }
5243
5244 bool Compile::should_print_igv(const int level) {
5245 PRODUCT_RETURN_(return false;);
5246
5247 if (PrintIdealGraphLevel < 0) { // disabled by the user
5248 return false;
5249 }
5250
5251 bool need = directive()->IGVPrintLevelOption >= level;
5252 if (need) {
5253 Compile::init_igv();
5254 }
5255 return need;
5256 }
5257
5258 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5259 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5260
5261 // Called from debugger. Prints method to the default file with the default phase name.
5262 // This works regardless of any Ideal Graph Visualizer flags set or not.
5263 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5264 void igv_print(void* sp, void* fp, void* pc) {
5265 frame fr(sp, fp, pc);
5266 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5267 }
5268
5269 // Same as igv_print() above but with a specified phase name.
5270 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5271 frame fr(sp, fp, pc);
5272 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5273 }
5274
5275 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5276 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5277 // This works regardless of any Ideal Graph Visualizer flags set or not.
5278 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5279 void igv_print(bool network, void* sp, void* fp, void* pc) {
5280 frame fr(sp, fp, pc);
5281 if (network) {
5282 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5283 } else {
5284 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5285 }
5286 }
5287
5288 // Same as igv_print(bool network, ...) above but with a specified phase name.
5289 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5290 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5291 frame fr(sp, fp, pc);
5292 if (network) {
5293 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5294 } else {
5295 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5296 }
5297 }
5298
5299 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5300 void igv_print_default() {
5301 Compile::current()->print_method(PHASE_DEBUG, 0);
5302 }
5303
5304 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5305 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5306 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5307 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5308 void igv_append(void* sp, void* fp, void* pc) {
5309 frame fr(sp, fp, pc);
5310 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5311 }
5312
5313 // Same as igv_append(...) above but with a specified phase name.
5314 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5315 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5316 frame fr(sp, fp, pc);
5317 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5318 }
5319
5320 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5321 const char* file_name = "custom_debug.xml";
5322 if (_debug_file_printer == nullptr) {
5323 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5324 } else {
5325 _debug_file_printer->update_compiled_method(C->method());
5326 }
5327 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5328 _debug_file_printer->print_graph(phase_name, fr);
5329 }
5330
5331 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5332 ResourceMark rm;
5333 GrowableArray<const Node*> empty_list;
5334 igv_print_graph_to_network(phase_name, empty_list, fr);
5335 }
5336
5337 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5338 if (_debug_network_printer == nullptr) {
5339 _debug_network_printer = new IdealGraphPrinter(C);
5340 } else {
5341 _debug_network_printer->update_compiled_method(C->method());
5342 }
5343 tty->print_cr("Method printed over network stream to IGV");
5344 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5345 }
5346 #endif // !PRODUCT
5347
5348 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5349 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5350 return value;
5351 }
5352 Node* result = nullptr;
5353 if (bt == T_BYTE) {
5354 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5355 result = new RShiftINode(result, phase->intcon(24));
5356 } else if (bt == T_BOOLEAN) {
5357 result = new AndINode(value, phase->intcon(0xFF));
5358 } else if (bt == T_CHAR) {
5359 result = new AndINode(value,phase->intcon(0xFFFF));
5360 } else {
5361 assert(bt == T_SHORT, "unexpected narrow type");
5362 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5363 result = new RShiftINode(result, phase->intcon(16));
5364 }
5365 if (transform_res) {
5366 result = phase->transform(result);
5367 }
5368 return result;
5369 }
5370
5371 void Compile::record_method_not_compilable_oom() {
5372 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5373 }
5374
5375 #ifndef PRODUCT
5376 // Collects all the control inputs from nodes on the worklist and from their data dependencies
5377 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
5378 // Follow non-control edges until we reach CFG nodes
5379 for (uint i = 0; i < worklist.size(); i++) {
5380 const Node* n = worklist.at(i);
5381 for (uint j = 0; j < n->req(); j++) {
5382 Node* in = n->in(j);
5383 if (in == nullptr || in->is_Root()) {
5384 continue;
5385 }
5386 if (in->is_CFG()) {
5387 if (in->is_Call()) {
5388 // The return value of a call is only available if the call did not result in an exception
5389 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
5390 if (control_proj_use->is_Catch()) {
5391 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
5392 candidates.push(fall_through);
5393 continue;
5394 }
5395 }
5396
5397 if (in->is_Multi()) {
5398 // We got here by following data inputs so we should only have one control use
5399 // (no IfNode, etc)
5400 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
5401 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
5402 } else {
5403 candidates.push(in);
5404 }
5405 } else {
5406 worklist.push(in);
5407 }
5408 }
5409 }
5410 }
5411
5412 // Returns the candidate node that is a descendant to all the other candidates
5413 static Node* pick_control(Unique_Node_List& candidates) {
5414 Unique_Node_List worklist;
5415 worklist.copy(candidates);
5416
5417 // Traverse backwards through the CFG
5418 for (uint i = 0; i < worklist.size(); i++) {
5419 const Node* n = worklist.at(i);
5420 if (n->is_Root()) {
5421 continue;
5422 }
5423 for (uint j = 0; j < n->req(); j++) {
5424 // Skip backedge of loops to avoid cycles
5425 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
5426 continue;
5427 }
5428
5429 Node* pred = n->in(j);
5430 if (pred != nullptr && pred != n && pred->is_CFG()) {
5431 worklist.push(pred);
5432 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
5433 candidates.remove(pred);
5434 }
5435 }
5436 }
5437
5438 assert(candidates.size() == 1, "unexpected control flow");
5439 return candidates.at(0);
5440 }
5441
5442 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
5443 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
5444 call->init_req((*pos)++, parm);
5445 const BasicType bt = parm->bottom_type()->basic_type();
5446 if (bt == T_LONG || bt == T_DOUBLE) {
5447 call->init_req((*pos)++, half);
5448 }
5449 }
5450
5451 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
5452 Node* parm0, Node* parm1,
5453 Node* parm2, Node* parm3,
5454 Node* parm4, Node* parm5,
5455 Node* parm6) const {
5456 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
5457 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
5458 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
5459
5460 // find the most suitable control input
5461 Unique_Node_List worklist, candidates;
5462 if (parm0 != nullptr) { worklist.push(parm0);
5463 if (parm1 != nullptr) { worklist.push(parm1);
5464 if (parm2 != nullptr) { worklist.push(parm2);
5465 if (parm3 != nullptr) { worklist.push(parm3);
5466 if (parm4 != nullptr) { worklist.push(parm4);
5467 if (parm5 != nullptr) { worklist.push(parm5);
5468 if (parm6 != nullptr) { worklist.push(parm6);
5469 /* close each nested if ===> */ } } } } } } }
5470 find_candidate_control_inputs(worklist, candidates);
5471 Node* control = nullptr;
5472 if (candidates.size() == 0) {
5473 control = C->start()->proj_out(TypeFunc::Control);
5474 } else {
5475 control = pick_control(candidates);
5476 }
5477
5478 // find all the previous users of the control we picked
5479 GrowableArray<Node*> users_of_control;
5480 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
5481 Node* use = control->fast_out(i);
5482 if (use->is_CFG() && use != control) {
5483 users_of_control.push(use);
5484 }
5485 }
5486
5487 // we do not actually care about IO and memory as it uses neither
5488 call->init_req(TypeFunc::Control, control);
5489 call->init_req(TypeFunc::I_O, top());
5490 call->init_req(TypeFunc::Memory, top());
5491 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
5492 call->init_req(TypeFunc::ReturnAdr, top());
5493
5494 int pos = TypeFunc::Parms;
5495 call->init_req(pos++, str_node);
5496 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
5497 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
5498 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
5499 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
5500 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
5501 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
5502 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
5503 /* close each nested if ===> */ } } } } } } }
5504 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
5505
5506 call = gvn->transform(call);
5507 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
5508
5509 // rewire previous users to have the new call as control instead
5510 PhaseIterGVN* igvn = gvn->is_IterGVN();
5511 for (int i = 0; i < users_of_control.length(); i++) {
5512 Node* use = users_of_control.at(i);
5513 for (uint j = 0; j < use->req(); j++) {
5514 if (use->in(j) == control) {
5515 if (igvn != nullptr) {
5516 igvn->replace_input_of(use, j, call_control_proj);
5517 } else {
5518 gvn->hash_delete(use);
5519 use->set_req(j, call_control_proj);
5520 gvn->hash_insert(use);
5521 }
5522 }
5523 }
5524 }
5525
5526 return call;
5527 }
5528 #endif // !PRODUCT