1 /*
2 * Copyright (c) 1997, 2026, 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* compile_phase_name) const {
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", compile_phase_name);
593 // Print out all nodes in ascending order of index.
594 // It is important that we traverse both inputs and outputs of nodes,
595 // so that we reach all nodes that are connected to Root.
596 root()->dump_bfs(MaxNodeLimit, nullptr, "-+S$", &ss);
597 } else {
598 // Dump the node blockwise if we have a scheduling
599 _output->print_scheduling(&ss);
600 }
601
602 // Check that the lock is not broken by a safepoint.
603 NoSafepointVerifier nsv;
604 ttyLocker ttyl;
605 if (xtty != nullptr) {
606 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
607 compile_id(),
608 is_osr_compilation() ? " compile_kind='osr'" : "",
609 compile_phase_name);
610 }
611
612 tty->print("%s", ss.as_string());
613
614 if (xtty != nullptr) {
615 xtty->tail("ideal");
616 }
617 }
618 #endif
619
620 // ============================================================================
621 //------------------------------Compile standard-------------------------------
622
623 // Compile a method. entry_bci is -1 for normal compilations and indicates
624 // the continuation bci for on stack replacement.
625
626
627 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
628 Options options, DirectiveSet* directive)
629 : Phase(Compiler),
630 _compile_id(ci_env->compile_id()),
631 _options(options),
632 _method(target),
633 _entry_bci(osr_bci),
634 _ilt(nullptr),
635 _stub_function(nullptr),
636 _stub_name(nullptr),
637 _stub_id(StubId::NO_STUBID),
638 _stub_entry_point(nullptr),
639 _max_node_limit(MaxNodeLimit),
640 _post_loop_opts_phase(false),
641 _merge_stores_phase(false),
642 _allow_macro_nodes(true),
643 _inlining_progress(false),
644 _inlining_incrementally(false),
645 _do_cleanup(false),
646 _has_reserved_stack_access(target->has_reserved_stack_access()),
647 #ifndef PRODUCT
648 _igv_idx(0),
649 _trace_opto_output(directive->TraceOptoOutputOption),
650 #endif
651 _clinit_barrier_on_entry(false),
652 _stress_seed(0),
653 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
654 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
655 _env(ci_env),
656 _directive(directive),
657 _log(ci_env->log()),
658 _first_failure_details(nullptr),
659 _intrinsics(comp_arena(), 0, 0, nullptr),
660 _macro_nodes(comp_arena(), 8, 0, nullptr),
661 _parse_predicates(comp_arena(), 8, 0, nullptr),
662 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
663 _expensive_nodes(comp_arena(), 8, 0, nullptr),
664 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
665 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
666 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
667 _coarsened_locks(comp_arena(), 8, 0, nullptr),
668 _congraph(nullptr),
669 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
670 _unique(0),
671 _dead_node_count(0),
672 _dead_node_list(comp_arena()),
673 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
674 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
675 _node_arena(&_node_arena_one),
676 _mach_constant_base_node(nullptr),
677 _Compile_types(mtCompiler, Arena::Tag::tag_type),
678 _initial_gvn(nullptr),
679 _igvn_worklist(nullptr),
680 _types(nullptr),
681 _node_hash(nullptr),
682 _late_inlines(comp_arena(), 2, 0, nullptr),
683 _string_late_inlines(comp_arena(), 2, 0, nullptr),
684 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
685 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
686 _late_inlines_pos(0),
687 _has_mh_late_inlines(false),
688 _oom(false),
689 _replay_inline_data(nullptr),
690 _inline_printer(this),
691 _java_calls(0),
692 _inner_loops(0),
693 _FIRST_STACK_mask(comp_arena()),
694 _interpreter_frame_size(0),
695 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
696 _output(nullptr)
697 #ifndef PRODUCT
698 ,
699 _in_dump_cnt(0)
700 #endif
701 {
702 C = this;
703 CompileWrapper cw(this);
704
705 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
706 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
707
708 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
709 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
710 // We can always print a disassembly, either abstract (hex dump) or
711 // with the help of a suitable hsdis library. Thus, we should not
712 // couple print_assembly and print_opto_assembly controls.
713 // But: always print opto and regular assembly on compile command 'print'.
714 bool print_assembly = directive->PrintAssemblyOption;
715 set_print_assembly(print_opto_assembly || print_assembly);
716 #else
717 set_print_assembly(false); // must initialize.
718 #endif
719
720 #ifndef PRODUCT
721 set_parsed_irreducible_loop(false);
722 #endif
723
724 if (directive->ReplayInlineOption) {
725 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
726 }
727 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
728 set_print_intrinsics(directive->PrintIntrinsicsOption);
729 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
730
731 if (ProfileTraps) {
732 // Make sure the method being compiled gets its own MDO,
733 // so we can at least track the decompile_count().
734 method()->ensure_method_data();
735 }
736
737 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
738 StressIncrementalInlining || StressMacroExpansion ||
739 StressMacroElimination || StressUnstableIfTraps ||
740 StressBailout || StressLoopPeeling) {
741 initialize_stress_seed(directive);
742 }
743
744 Init(/*do_aliasing=*/ true);
745
746 print_compile_messages();
747
748 _ilt = InlineTree::build_inline_tree_root();
749
750 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
751 assert(num_alias_types() >= AliasIdxRaw, "");
752
753 #define MINIMUM_NODE_HASH 1023
754
755 // GVN that will be run immediately on new nodes
756 uint estimated_size = method()->code_size()*4+64;
757 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
758 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
759 _types = new (comp_arena()) Type_Array(comp_arena());
760 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
761 PhaseGVN gvn;
762 set_initial_gvn(&gvn);
763
764 { // Scope for timing the parser
765 TracePhase tp(_t_parser);
766
767 // Put top into the hash table ASAP.
768 initial_gvn()->transform(top());
769
770 // Set up tf(), start(), and find a CallGenerator.
771 CallGenerator* cg = nullptr;
772 if (is_osr_compilation()) {
773 const TypeTuple *domain = StartOSRNode::osr_domain();
774 const TypeTuple *range = TypeTuple::make_range(method()->signature());
775 init_tf(TypeFunc::make(domain, range));
776 StartNode* s = new StartOSRNode(root(), domain);
777 initial_gvn()->set_type_bottom(s);
778 verify_start(s);
779 cg = CallGenerator::for_osr(method(), entry_bci());
780 } else {
781 // Normal case.
782 init_tf(TypeFunc::make(method()));
783 StartNode* s = new StartNode(root(), tf()->domain());
784 initial_gvn()->set_type_bottom(s);
785 verify_start(s);
786 float past_uses = method()->interpreter_invocation_count();
787 float expected_uses = past_uses;
788 cg = CallGenerator::for_inline(method(), expected_uses);
789 }
790 if (failing()) return;
791 if (cg == nullptr) {
792 const char* reason = InlineTree::check_can_parse(method());
793 assert(reason != nullptr, "expect reason for parse failure");
794 stringStream ss;
795 ss.print("cannot parse method: %s", reason);
796 record_method_not_compilable(ss.as_string());
797 return;
798 }
799
800 gvn.set_type(root(), root()->bottom_type());
801
802 JVMState* jvms = build_start_state(start(), tf());
803 if ((jvms = cg->generate(jvms)) == nullptr) {
804 assert(failure_reason() != nullptr, "expect reason for parse failure");
805 stringStream ss;
806 ss.print("method parse failed: %s", failure_reason());
807 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
808 return;
809 }
810 GraphKit kit(jvms);
811
812 if (!kit.stopped()) {
813 // Accept return values, and transfer control we know not where.
814 // This is done by a special, unique ReturnNode bound to root.
815 return_values(kit.jvms());
816 }
817
818 if (kit.has_exceptions()) {
819 // Any exceptions that escape from this call must be rethrown
820 // to whatever caller is dynamically above us on the stack.
821 // This is done by a special, unique RethrowNode bound to root.
822 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
823 }
824
825 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
826
827 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
828 inline_string_calls(true);
829 }
830
831 if (failing()) return;
832
833 // Remove clutter produced by parsing.
834 if (!failing()) {
835 ResourceMark rm;
836 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
837 }
838 }
839
840 // Note: Large methods are capped off in do_one_bytecode().
841 if (failing()) return;
842
843 // After parsing, node notes are no longer automagic.
844 // They must be propagated by register_new_node_with_optimizer(),
845 // clone(), or the like.
846 set_default_node_notes(nullptr);
847
848 #ifndef PRODUCT
849 if (should_print_igv(1)) {
850 _igv_printer->print_inlining();
851 }
852 #endif
853
854 if (failing()) return;
855 NOT_PRODUCT( verify_graph_edges(); )
856
857 // Now optimize
858 Optimize();
859 if (failing()) return;
860 NOT_PRODUCT( verify_graph_edges(); )
861
862 #ifndef PRODUCT
863 if (should_print_ideal()) {
864 print_ideal_ir("PrintIdeal");
865 }
866 #endif
867
868 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
869
870 #ifdef ASSERT
871 bs->verify_gc_barriers(this, BarrierSetC2::AfterOptimize);
872 #endif
873
874 bs->final_refinement(this);
875
876 #ifdef ASSERT
877 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
878 #endif
879
880 // Dump compilation data to replay it.
881 if (directive->DumpReplayOption) {
882 env()->dump_replay_data(_compile_id);
883 }
884 if (directive->DumpInlineOption && (ilt() != nullptr)) {
885 env()->dump_inline_data(_compile_id);
886 }
887
888 // Now that we know the size of all the monitors we can add a fixed slot
889 // for the original deopt pc.
890 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
891 set_fixed_slots(next_slot);
892
893 // Compute when to use implicit null checks. Used by matching trap based
894 // nodes and NullCheck optimization.
895 set_allowed_deopt_reasons();
896
897 // Now generate code
898 Code_Gen();
899 }
900
901 //------------------------------Compile----------------------------------------
902 // Compile a runtime stub
903 Compile::Compile(ciEnv* ci_env,
904 TypeFunc_generator generator,
905 address stub_function,
906 const char* stub_name,
907 StubId stub_id,
908 int is_fancy_jump,
909 bool pass_tls,
910 bool return_pc,
911 DirectiveSet* directive)
912 : Phase(Compiler),
913 _compile_id(0),
914 _options(Options::for_runtime_stub()),
915 _method(nullptr),
916 _entry_bci(InvocationEntryBci),
917 _stub_function(stub_function),
918 _stub_name(stub_name),
919 _stub_id(stub_id),
920 _stub_entry_point(nullptr),
921 _max_node_limit(MaxNodeLimit),
922 _post_loop_opts_phase(false),
923 _merge_stores_phase(false),
924 _allow_macro_nodes(true),
925 _inlining_progress(false),
926 _inlining_incrementally(false),
927 _has_reserved_stack_access(false),
928 #ifndef PRODUCT
929 _igv_idx(0),
930 _trace_opto_output(directive->TraceOptoOutputOption),
931 #endif
932 _clinit_barrier_on_entry(false),
933 _stress_seed(0),
934 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
935 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
936 _env(ci_env),
937 _directive(directive),
938 _log(ci_env->log()),
939 _first_failure_details(nullptr),
940 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
941 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
942 _congraph(nullptr),
943 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
944 _unique(0),
945 _dead_node_count(0),
946 _dead_node_list(comp_arena()),
947 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
948 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
949 _node_arena(&_node_arena_one),
950 _mach_constant_base_node(nullptr),
951 _Compile_types(mtCompiler, Arena::Tag::tag_type),
952 _initial_gvn(nullptr),
953 _igvn_worklist(nullptr),
954 _types(nullptr),
955 _node_hash(nullptr),
956 _has_mh_late_inlines(false),
957 _oom(false),
958 _replay_inline_data(nullptr),
959 _inline_printer(this),
960 _java_calls(0),
961 _inner_loops(0),
962 _FIRST_STACK_mask(comp_arena()),
963 _interpreter_frame_size(0),
964 _regmask_arena(mtCompiler, Arena::Tag::tag_regmask),
965 _output(nullptr),
966 #ifndef PRODUCT
967 _in_dump_cnt(0),
968 #endif
969 _allowed_reasons(0) {
970 C = this;
971
972 // try to reuse an existing stub
973 {
974 BlobId blob_id = StubInfo::blob(_stub_id);
975 CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::C2Blob, blob_id);
976 if (blob != nullptr) {
977 RuntimeStub* rs = blob->as_runtime_stub();
978 _stub_entry_point = rs->entry_point();
979 return;
980 }
981 }
982
983 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
984 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
985
986 #ifndef PRODUCT
987 set_print_assembly(PrintFrameConverterAssembly);
988 set_parsed_irreducible_loop(false);
989 #else
990 set_print_assembly(false); // Must initialize.
991 #endif
992 set_has_irreducible_loop(false); // no loops
993
994 CompileWrapper cw(this);
995 Init(/*do_aliasing=*/ false);
996 init_tf((*generator)());
997
998 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
999 _types = new (comp_arena()) Type_Array(comp_arena());
1000 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1001
1002 if (StressLCM || StressGCM || StressBailout) {
1003 initialize_stress_seed(directive);
1004 }
1005
1006 {
1007 PhaseGVN gvn;
1008 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1009 gvn.transform(top());
1010
1011 GraphKit kit;
1012 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1013 }
1014
1015 NOT_PRODUCT( verify_graph_edges(); )
1016
1017 Code_Gen();
1018 }
1019
1020 Compile::~Compile() {
1021 delete _first_failure_details;
1022 };
1023
1024 //------------------------------Init-------------------------------------------
1025 // Prepare for a single compilation
1026 void Compile::Init(bool aliasing) {
1027 _do_aliasing = aliasing;
1028 _unique = 0;
1029 _regalloc = nullptr;
1030
1031 _tf = nullptr; // filled in later
1032 _top = nullptr; // cached later
1033 _matcher = nullptr; // filled in later
1034 _cfg = nullptr; // filled in later
1035
1036 _node_note_array = nullptr;
1037 _default_node_notes = nullptr;
1038 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1039
1040 _immutable_memory = nullptr; // filled in at first inquiry
1041
1042 #ifdef ASSERT
1043 _phase_optimize_finished = false;
1044 _phase_verify_ideal_loop = false;
1045 _exception_backedge = false;
1046 _type_verify = nullptr;
1047 #endif
1048
1049 // Globally visible Nodes
1050 // First set TOP to null to give safe behavior during creation of RootNode
1051 set_cached_top_node(nullptr);
1052 set_root(new RootNode());
1053 // Now that you have a Root to point to, create the real TOP
1054 set_cached_top_node( new ConNode(Type::TOP) );
1055 set_recent_alloc(nullptr, nullptr);
1056
1057 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1058 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1059 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1060 env()->set_dependencies(new Dependencies(env()));
1061
1062 _fixed_slots = 0;
1063 set_has_split_ifs(false);
1064 set_has_loops(false); // first approximation
1065 set_has_stringbuilder(false);
1066 set_has_boxed_value(false);
1067 _trap_can_recompile = false; // no traps emitted yet
1068 _major_progress = true; // start out assuming good things will happen
1069 set_has_unsafe_access(false);
1070 set_max_vector_size(0);
1071 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1072 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1073 set_decompile_count(0);
1074
1075 #ifndef PRODUCT
1076 _phase_counter = 0;
1077 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1078 #endif
1079
1080 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1081 _loop_opts_cnt = LoopOptsCount;
1082 set_do_inlining(Inline);
1083 set_max_inline_size(MaxInlineSize);
1084 set_freq_inline_size(FreqInlineSize);
1085 set_do_scheduling(OptoScheduling);
1086
1087 set_do_vector_loop(false);
1088 set_has_monitors(false);
1089 set_has_scoped_access(false);
1090
1091 if (AllowVectorizeOnDemand) {
1092 if (has_method() && _directive->VectorizeOption) {
1093 set_do_vector_loop(true);
1094 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());})
1095 } else if (has_method() && method()->name() != nullptr &&
1096 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1097 set_do_vector_loop(true);
1098 }
1099 }
1100 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1101 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());})
1102
1103 _max_node_limit = _directive->MaxNodeLimitOption;
1104
1105 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1106 set_clinit_barrier_on_entry(true);
1107 }
1108 if (debug_info()->recording_non_safepoints()) {
1109 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1110 (comp_arena(), 8, 0, nullptr));
1111 set_default_node_notes(Node_Notes::make(this));
1112 }
1113
1114 const int grow_ats = 16;
1115 _max_alias_types = grow_ats;
1116 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1117 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1118 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1119 {
1120 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1121 }
1122 // Initialize the first few types.
1123 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1124 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1125 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1126 _num_alias_types = AliasIdxRaw+1;
1127 // Zero out the alias type cache.
1128 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1129 // A null adr_type hits in the cache right away. Preload the right answer.
1130 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1131 }
1132
1133 #ifdef ASSERT
1134 // Verify that the current StartNode is valid.
1135 void Compile::verify_start(StartNode* s) const {
1136 assert(failing_internal() || s == start(), "should be StartNode");
1137 }
1138 #endif
1139
1140 /**
1141 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1142 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1143 * the ideal graph.
1144 */
1145 StartNode* Compile::start() const {
1146 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1147 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1148 Node* start = root()->fast_out(i);
1149 if (start->is_Start()) {
1150 return start->as_Start();
1151 }
1152 }
1153 fatal("Did not find Start node!");
1154 return nullptr;
1155 }
1156
1157 //-------------------------------immutable_memory-------------------------------------
1158 // Access immutable memory
1159 Node* Compile::immutable_memory() {
1160 if (_immutable_memory != nullptr) {
1161 return _immutable_memory;
1162 }
1163 StartNode* s = start();
1164 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1165 Node *p = s->fast_out(i);
1166 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1167 _immutable_memory = p;
1168 return _immutable_memory;
1169 }
1170 }
1171 ShouldNotReachHere();
1172 return nullptr;
1173 }
1174
1175 //----------------------set_cached_top_node------------------------------------
1176 // Install the cached top node, and make sure Node::is_top works correctly.
1177 void Compile::set_cached_top_node(Node* tn) {
1178 if (tn != nullptr) verify_top(tn);
1179 Node* old_top = _top;
1180 _top = tn;
1181 // Calling Node::setup_is_top allows the nodes the chance to adjust
1182 // their _out arrays.
1183 if (_top != nullptr) _top->setup_is_top();
1184 if (old_top != nullptr) old_top->setup_is_top();
1185 assert(_top == nullptr || top()->is_top(), "");
1186 }
1187
1188 #ifdef ASSERT
1189 uint Compile::count_live_nodes_by_graph_walk() {
1190 Unique_Node_List useful(comp_arena());
1191 // Get useful node list by walking the graph.
1192 identify_useful_nodes(useful);
1193 return useful.size();
1194 }
1195
1196 void Compile::print_missing_nodes() {
1197
1198 // Return if CompileLog is null and PrintIdealNodeCount is false.
1199 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1200 return;
1201 }
1202
1203 // This is an expensive function. It is executed only when the user
1204 // specifies VerifyIdealNodeCount option or otherwise knows the
1205 // additional work that needs to be done to identify reachable nodes
1206 // by walking the flow graph and find the missing ones using
1207 // _dead_node_list.
1208
1209 Unique_Node_List useful(comp_arena());
1210 // Get useful node list by walking the graph.
1211 identify_useful_nodes(useful);
1212
1213 uint l_nodes = C->live_nodes();
1214 uint l_nodes_by_walk = useful.size();
1215
1216 if (l_nodes != l_nodes_by_walk) {
1217 if (_log != nullptr) {
1218 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1219 _log->stamp();
1220 _log->end_head();
1221 }
1222 VectorSet& useful_member_set = useful.member_set();
1223 int last_idx = l_nodes_by_walk;
1224 for (int i = 0; i < last_idx; i++) {
1225 if (useful_member_set.test(i)) {
1226 if (_dead_node_list.test(i)) {
1227 if (_log != nullptr) {
1228 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1229 }
1230 if (PrintIdealNodeCount) {
1231 // Print the log message to tty
1232 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1233 useful.at(i)->dump();
1234 }
1235 }
1236 }
1237 else if (! _dead_node_list.test(i)) {
1238 if (_log != nullptr) {
1239 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1240 }
1241 if (PrintIdealNodeCount) {
1242 // Print the log message to tty
1243 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1244 }
1245 }
1246 }
1247 if (_log != nullptr) {
1248 _log->tail("mismatched_nodes");
1249 }
1250 }
1251 }
1252 void Compile::record_modified_node(Node* n) {
1253 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1254 _modified_nodes->push(n);
1255 }
1256 }
1257
1258 void Compile::remove_modified_node(Node* n) {
1259 if (_modified_nodes != nullptr) {
1260 _modified_nodes->remove(n);
1261 }
1262 }
1263 #endif
1264
1265 #ifndef PRODUCT
1266 void Compile::verify_top(Node* tn) const {
1267 if (tn != nullptr) {
1268 assert(tn->is_Con(), "top node must be a constant");
1269 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1270 assert(tn->in(0) != nullptr, "must have live top node");
1271 }
1272 }
1273 #endif
1274
1275
1276 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1277
1278 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1279 guarantee(arr != nullptr, "");
1280 int num_blocks = arr->length();
1281 if (grow_by < num_blocks) grow_by = num_blocks;
1282 int num_notes = grow_by * _node_notes_block_size;
1283 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1284 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1285 while (num_notes > 0) {
1286 arr->append(notes);
1287 notes += _node_notes_block_size;
1288 num_notes -= _node_notes_block_size;
1289 }
1290 assert(num_notes == 0, "exact multiple, please");
1291 }
1292
1293 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1294 if (source == nullptr || dest == nullptr) return false;
1295
1296 if (dest->is_Con())
1297 return false; // Do not push debug info onto constants.
1298
1299 #ifdef ASSERT
1300 // Leave a bread crumb trail pointing to the original node:
1301 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1302 dest->set_debug_orig(source);
1303 }
1304 #endif
1305
1306 if (node_note_array() == nullptr)
1307 return false; // Not collecting any notes now.
1308
1309 // This is a copy onto a pre-existing node, which may already have notes.
1310 // If both nodes have notes, do not overwrite any pre-existing notes.
1311 Node_Notes* source_notes = node_notes_at(source->_idx);
1312 if (source_notes == nullptr || source_notes->is_clear()) return false;
1313 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1314 if (dest_notes == nullptr || dest_notes->is_clear()) {
1315 return set_node_notes_at(dest->_idx, source_notes);
1316 }
1317
1318 Node_Notes merged_notes = (*source_notes);
1319 // The order of operations here ensures that dest notes will win...
1320 merged_notes.update_from(dest_notes);
1321 return set_node_notes_at(dest->_idx, &merged_notes);
1322 }
1323
1324
1325 //--------------------------allow_range_check_smearing-------------------------
1326 // Gating condition for coalescing similar range checks.
1327 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1328 // single covering check that is at least as strong as any of them.
1329 // If the optimization succeeds, the simplified (strengthened) range check
1330 // will always succeed. If it fails, we will deopt, and then give up
1331 // on the optimization.
1332 bool Compile::allow_range_check_smearing() const {
1333 // If this method has already thrown a range-check,
1334 // assume it was because we already tried range smearing
1335 // and it failed.
1336 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1337 return !already_trapped;
1338 }
1339
1340
1341 //------------------------------flatten_alias_type-----------------------------
1342 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1343 assert(do_aliasing(), "Aliasing should be enabled");
1344 int offset = tj->offset();
1345 TypePtr::PTR ptr = tj->ptr();
1346
1347 // Known instance (scalarizable allocation) alias only with itself.
1348 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1349 tj->is_oopptr()->is_known_instance();
1350
1351 // Process weird unsafe references.
1352 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1353 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1354 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1355 tj = TypeOopPtr::BOTTOM;
1356 ptr = tj->ptr();
1357 offset = tj->offset();
1358 }
1359
1360 // Array pointers need some flattening
1361 const TypeAryPtr* ta = tj->isa_aryptr();
1362 if (ta && ta->is_stable()) {
1363 // Erase stability property for alias analysis.
1364 tj = ta = ta->cast_to_stable(false);
1365 }
1366 if( ta && is_known_inst ) {
1367 if ( offset != Type::OffsetBot &&
1368 offset > arrayOopDesc::length_offset_in_bytes() ) {
1369 offset = Type::OffsetBot; // Flatten constant access into array body only
1370 tj = ta = ta->
1371 remove_speculative()->
1372 cast_to_ptr_type(ptr)->
1373 with_offset(offset);
1374 }
1375 } else if (ta) {
1376 // For arrays indexed by constant indices, we flatten the alias
1377 // space to include all of the array body. Only the header, klass
1378 // and array length can be accessed un-aliased.
1379 if( offset != Type::OffsetBot ) {
1380 if( ta->const_oop() ) { // MethodData* or Method*
1381 offset = Type::OffsetBot; // Flatten constant access into array body
1382 tj = ta = ta->
1383 remove_speculative()->
1384 cast_to_ptr_type(ptr)->
1385 cast_to_exactness(false)->
1386 with_offset(offset);
1387 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1388 // range is OK as-is.
1389 tj = ta = TypeAryPtr::RANGE;
1390 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1391 tj = TypeInstPtr::KLASS; // all klass loads look alike
1392 ta = TypeAryPtr::RANGE; // generic ignored junk
1393 ptr = TypePtr::BotPTR;
1394 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1395 tj = TypeInstPtr::MARK;
1396 ta = TypeAryPtr::RANGE; // generic ignored junk
1397 ptr = TypePtr::BotPTR;
1398 } else { // Random constant offset into array body
1399 offset = Type::OffsetBot; // Flatten constant access into array body
1400 tj = ta = ta->
1401 remove_speculative()->
1402 cast_to_ptr_type(ptr)->
1403 cast_to_exactness(false)->
1404 with_offset(offset);
1405 }
1406 }
1407 // Arrays of fixed size alias with arrays of unknown size.
1408 if (ta->size() != TypeInt::POS) {
1409 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1410 tj = ta = ta->
1411 remove_speculative()->
1412 cast_to_ptr_type(ptr)->
1413 with_ary(tary)->
1414 cast_to_exactness(false);
1415 }
1416 // Arrays of known objects become arrays of unknown objects.
1417 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1418 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1419 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1420 }
1421 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1422 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1423 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset);
1424 }
1425 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1426 // cannot be distinguished by bytecode alone.
1427 if (ta->elem() == TypeInt::BOOL) {
1428 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1429 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1430 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1431 }
1432 // During the 2nd round of IterGVN, NotNull castings are removed.
1433 // Make sure the Bottom and NotNull variants alias the same.
1434 // Also, make sure exact and non-exact variants alias the same.
1435 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1436 tj = ta = ta->
1437 remove_speculative()->
1438 cast_to_ptr_type(TypePtr::BotPTR)->
1439 cast_to_exactness(false)->
1440 with_offset(offset);
1441 }
1442 }
1443
1444 // Oop pointers need some flattening
1445 const TypeInstPtr *to = tj->isa_instptr();
1446 if (to && to != TypeOopPtr::BOTTOM) {
1447 ciInstanceKlass* ik = to->instance_klass();
1448 if( ptr == TypePtr::Constant ) {
1449 if (ik != ciEnv::current()->Class_klass() ||
1450 offset < ik->layout_helper_size_in_bytes()) {
1451 // No constant oop pointers (such as Strings); they alias with
1452 // unknown strings.
1453 assert(!is_known_inst, "not scalarizable allocation");
1454 tj = to = to->
1455 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1456 remove_speculative()->
1457 cast_to_ptr_type(TypePtr::BotPTR)->
1458 cast_to_exactness(false);
1459 }
1460 } else if( is_known_inst ) {
1461 tj = to; // Keep NotNull and klass_is_exact for instance type
1462 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1463 // During the 2nd round of IterGVN, NotNull castings are removed.
1464 // Make sure the Bottom and NotNull variants alias the same.
1465 // Also, make sure exact and non-exact variants alias the same.
1466 tj = to = to->
1467 remove_speculative()->
1468 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1469 cast_to_ptr_type(TypePtr::BotPTR)->
1470 cast_to_exactness(false);
1471 }
1472 if (to->speculative() != nullptr) {
1473 tj = to = to->remove_speculative();
1474 }
1475 // Canonicalize the holder of this field
1476 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1477 // First handle header references such as a LoadKlassNode, even if the
1478 // object's klass is unloaded at compile time (4965979).
1479 if (!is_known_inst) { // Do it only for non-instance types
1480 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset);
1481 }
1482 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1483 // Static fields are in the space above the normal instance
1484 // fields in the java.lang.Class instance.
1485 if (ik != ciEnv::current()->Class_klass()) {
1486 to = nullptr;
1487 tj = TypeOopPtr::BOTTOM;
1488 offset = tj->offset();
1489 }
1490 } else {
1491 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1492 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1493 assert(tj->offset() == offset, "no change to offset expected");
1494 bool xk = to->klass_is_exact();
1495 int instance_id = to->instance_id();
1496
1497 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1498 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1499 // its interfaces are included.
1500 if (xk && ik->equals(canonical_holder)) {
1501 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id), "exact type should be canonical type");
1502 } else {
1503 assert(xk || !is_known_inst, "Known instance should be exact type");
1504 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, offset, instance_id);
1505 }
1506 }
1507 }
1508
1509 // Klass pointers to object array klasses need some flattening
1510 const TypeKlassPtr *tk = tj->isa_klassptr();
1511 if( tk ) {
1512 // If we are referencing a field within a Klass, we need
1513 // to assume the worst case of an Object. Both exact and
1514 // inexact types must flatten to the same alias class so
1515 // use NotNull as the PTR.
1516 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1517 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1518 env()->Object_klass(),
1519 offset);
1520 }
1521
1522 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1523 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1524 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1525 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1526 } else {
1527 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1528 }
1529 }
1530
1531 // Check for precise loads from the primary supertype array and force them
1532 // to the supertype cache alias index. Check for generic array loads from
1533 // the primary supertype array and also force them to the supertype cache
1534 // alias index. Since the same load can reach both, we need to merge
1535 // these 2 disparate memories into the same alias class. Since the
1536 // primary supertype array is read-only, there's no chance of confusion
1537 // where we bypass an array load and an array store.
1538 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1539 if (offset == Type::OffsetBot ||
1540 (offset >= primary_supers_offset &&
1541 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1542 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1543 offset = in_bytes(Klass::secondary_super_cache_offset());
1544 tj = tk = tk->with_offset(offset);
1545 }
1546 }
1547
1548 // Flatten all Raw pointers together.
1549 if (tj->base() == Type::RawPtr)
1550 tj = TypeRawPtr::BOTTOM;
1551
1552 if (tj->base() == Type::AnyPtr)
1553 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1554
1555 offset = tj->offset();
1556 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1557
1558 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1559 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1560 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1561 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1562 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1563 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1564 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1565 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1566 assert( tj->ptr() != TypePtr::TopPTR &&
1567 tj->ptr() != TypePtr::AnyNull &&
1568 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1569 // assert( tj->ptr() != TypePtr::Constant ||
1570 // tj->base() == Type::RawPtr ||
1571 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1572
1573 return tj;
1574 }
1575
1576 void Compile::AliasType::Init(int i, const TypePtr* at) {
1577 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1578 _index = i;
1579 _adr_type = at;
1580 _field = nullptr;
1581 _element = nullptr;
1582 _is_rewritable = true; // default
1583 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1584 if (atoop != nullptr && atoop->is_known_instance()) {
1585 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1586 _general_index = Compile::current()->get_alias_index(gt);
1587 } else {
1588 _general_index = 0;
1589 }
1590 }
1591
1592 BasicType Compile::AliasType::basic_type() const {
1593 if (element() != nullptr) {
1594 const Type* element = adr_type()->is_aryptr()->elem();
1595 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1596 } if (field() != nullptr) {
1597 return field()->layout_type();
1598 } else {
1599 return T_ILLEGAL; // unknown
1600 }
1601 }
1602
1603 //---------------------------------print_on------------------------------------
1604 #ifndef PRODUCT
1605 void Compile::AliasType::print_on(outputStream* st) {
1606 if (index() < 10)
1607 st->print("@ <%d> ", index());
1608 else st->print("@ <%d>", index());
1609 st->print(is_rewritable() ? " " : " RO");
1610 int offset = adr_type()->offset();
1611 if (offset == Type::OffsetBot)
1612 st->print(" +any");
1613 else st->print(" +%-3d", offset);
1614 st->print(" in ");
1615 adr_type()->dump_on(st);
1616 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1617 if (field() != nullptr && tjp) {
1618 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1619 tjp->offset() != field()->offset_in_bytes()) {
1620 st->print(" != ");
1621 field()->print();
1622 st->print(" ***");
1623 }
1624 }
1625 }
1626
1627 void print_alias_types() {
1628 Compile* C = Compile::current();
1629 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1630 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1631 C->alias_type(idx)->print_on(tty);
1632 tty->cr();
1633 }
1634 }
1635 #endif
1636
1637
1638 //----------------------------probe_alias_cache--------------------------------
1639 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1640 intptr_t key = (intptr_t) adr_type;
1641 key ^= key >> logAliasCacheSize;
1642 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1643 }
1644
1645
1646 //-----------------------------grow_alias_types--------------------------------
1647 void Compile::grow_alias_types() {
1648 const int old_ats = _max_alias_types; // how many before?
1649 const int new_ats = old_ats; // how many more?
1650 const int grow_ats = old_ats+new_ats; // how many now?
1651 _max_alias_types = grow_ats;
1652 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1653 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1654 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1655 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1656 }
1657
1658
1659 //--------------------------------find_alias_type------------------------------
1660 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1661 if (!do_aliasing()) {
1662 return alias_type(AliasIdxBot);
1663 }
1664
1665 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1666 if (ace->_adr_type == adr_type) {
1667 return alias_type(ace->_index);
1668 }
1669
1670 // Handle special cases.
1671 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1672 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1673
1674 // Do it the slow way.
1675 const TypePtr* flat = flatten_alias_type(adr_type);
1676
1677 #ifdef ASSERT
1678 {
1679 ResourceMark rm;
1680 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1681 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1682 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1683 Type::str(adr_type));
1684 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1685 const TypeOopPtr* foop = flat->is_oopptr();
1686 // Scalarizable allocations have exact klass always.
1687 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1688 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1689 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1690 Type::str(foop), Type::str(xoop));
1691 }
1692 }
1693 #endif
1694
1695 int idx = AliasIdxTop;
1696 for (int i = 0; i < num_alias_types(); i++) {
1697 if (alias_type(i)->adr_type() == flat) {
1698 idx = i;
1699 break;
1700 }
1701 }
1702
1703 if (idx == AliasIdxTop) {
1704 if (no_create) return nullptr;
1705 // Grow the array if necessary.
1706 if (_num_alias_types == _max_alias_types) grow_alias_types();
1707 // Add a new alias type.
1708 idx = _num_alias_types++;
1709 _alias_types[idx]->Init(idx, flat);
1710 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1711 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1712 if (flat->isa_instptr()) {
1713 if (flat->offset() == java_lang_Class::klass_offset()
1714 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1715 alias_type(idx)->set_rewritable(false);
1716 }
1717 if (flat->isa_aryptr()) {
1718 #ifdef ASSERT
1719 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1720 // (T_BYTE has the weakest alignment and size restrictions...)
1721 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1722 #endif
1723 if (flat->offset() == TypePtr::OffsetBot) {
1724 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1725 }
1726 }
1727 if (flat->isa_klassptr()) {
1728 if (UseCompactObjectHeaders) {
1729 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1730 alias_type(idx)->set_rewritable(false);
1731 }
1732 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1733 alias_type(idx)->set_rewritable(false);
1734 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1735 alias_type(idx)->set_rewritable(false);
1736 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1737 alias_type(idx)->set_rewritable(false);
1738 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1739 alias_type(idx)->set_rewritable(false);
1740 }
1741
1742 if (flat->isa_instklassptr()) {
1743 if (flat->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1744 alias_type(idx)->set_rewritable(false);
1745 }
1746 }
1747 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1748 // but the base pointer type is not distinctive enough to identify
1749 // references into JavaThread.)
1750
1751 // Check for final fields.
1752 const TypeInstPtr* tinst = flat->isa_instptr();
1753 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1754 ciField* field;
1755 if (tinst->const_oop() != nullptr &&
1756 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1757 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1758 // static field
1759 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1760 field = k->get_field_by_offset(tinst->offset(), true);
1761 } else {
1762 ciInstanceKlass *k = tinst->instance_klass();
1763 field = k->get_field_by_offset(tinst->offset(), false);
1764 }
1765 assert(field == nullptr ||
1766 original_field == nullptr ||
1767 (field->holder() == original_field->holder() &&
1768 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1769 field->is_static() == original_field->is_static()), "wrong field?");
1770 // Set field() and is_rewritable() attributes.
1771 if (field != nullptr) alias_type(idx)->set_field(field);
1772 }
1773 }
1774
1775 // Fill the cache for next time.
1776 ace->_adr_type = adr_type;
1777 ace->_index = idx;
1778 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1779
1780 // Might as well try to fill the cache for the flattened version, too.
1781 AliasCacheEntry* face = probe_alias_cache(flat);
1782 if (face->_adr_type == nullptr) {
1783 face->_adr_type = flat;
1784 face->_index = idx;
1785 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1786 }
1787
1788 return alias_type(idx);
1789 }
1790
1791
1792 Compile::AliasType* Compile::alias_type(ciField* field) {
1793 const TypeOopPtr* t;
1794 if (field->is_static())
1795 t = TypeInstPtr::make(field->holder()->java_mirror());
1796 else
1797 t = TypeOopPtr::make_from_klass_raw(field->holder());
1798 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1799 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1800 return atp;
1801 }
1802
1803
1804 //------------------------------have_alias_type--------------------------------
1805 bool Compile::have_alias_type(const TypePtr* adr_type) {
1806 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1807 if (ace->_adr_type == adr_type) {
1808 return true;
1809 }
1810
1811 // Handle special cases.
1812 if (adr_type == nullptr) return true;
1813 if (adr_type == TypePtr::BOTTOM) return true;
1814
1815 return find_alias_type(adr_type, true, nullptr) != nullptr;
1816 }
1817
1818 //-----------------------------must_alias--------------------------------------
1819 // True if all values of the given address type are in the given alias category.
1820 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1821 if (alias_idx == AliasIdxBot) return true; // the universal category
1822 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1823 if (alias_idx == AliasIdxTop) return false; // the empty category
1824 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1825
1826 // the only remaining possible overlap is identity
1827 int adr_idx = get_alias_index(adr_type);
1828 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1829 assert(adr_idx == alias_idx ||
1830 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1831 && adr_type != TypeOopPtr::BOTTOM),
1832 "should not be testing for overlap with an unsafe pointer");
1833 return adr_idx == alias_idx;
1834 }
1835
1836 //------------------------------can_alias--------------------------------------
1837 // True if any values of the given address type are in the given alias category.
1838 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1839 if (alias_idx == AliasIdxTop) return false; // the empty category
1840 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1841 // Known instance doesn't alias with bottom memory
1842 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1843 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1844
1845 // the only remaining possible overlap is identity
1846 int adr_idx = get_alias_index(adr_type);
1847 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1848 return adr_idx == alias_idx;
1849 }
1850
1851 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1852 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1853 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1854 if (parse_predicate_count() == 0) {
1855 return;
1856 }
1857 for (int i = 0; i < parse_predicate_count(); i++) {
1858 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1859 parse_predicate->mark_useless(igvn);
1860 }
1861 _parse_predicates.clear();
1862 }
1863
1864 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1865 if (!n->for_post_loop_opts_igvn()) {
1866 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1867 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1868 _for_post_loop_igvn.append(n);
1869 }
1870 }
1871
1872 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1873 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1874 _for_post_loop_igvn.remove(n);
1875 }
1876
1877 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1878 // Verify that all previous optimizations produced a valid graph
1879 // at least to this point, even if no loop optimizations were done.
1880 PhaseIdealLoop::verify(igvn);
1881
1882 if (_print_phase_loop_opts) {
1883 print_method(PHASE_AFTER_LOOP_OPTS, 2);
1884 }
1885 C->set_post_loop_opts_phase(); // no more loop opts allowed
1886
1887 assert(!C->major_progress(), "not cleared");
1888
1889 if (_for_post_loop_igvn.length() > 0) {
1890 while (_for_post_loop_igvn.length() > 0) {
1891 Node* n = _for_post_loop_igvn.pop();
1892 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1893 igvn._worklist.push(n);
1894 }
1895 igvn.optimize();
1896 if (failing()) return;
1897 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1898 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1899
1900 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1901 if (C->major_progress()) {
1902 C->clear_major_progress(); // ensure that major progress is now clear
1903 }
1904 }
1905 }
1906
1907 void Compile::record_for_merge_stores_igvn(Node* n) {
1908 if (!n->for_merge_stores_igvn()) {
1909 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
1910 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1911 _for_merge_stores_igvn.append(n);
1912 }
1913 }
1914
1915 void Compile::remove_from_merge_stores_igvn(Node* n) {
1916 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1917 _for_merge_stores_igvn.remove(n);
1918 }
1919
1920 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
1921 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
1922 // the stores, and we merge the wrong sequence of stores.
1923 // Example:
1924 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
1925 // Apply MergeStores:
1926 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
1927 // Remove more RangeChecks:
1928 // StoreI [ StoreL ] StoreI
1929 // But now it would have been better to do this instead:
1930 // [ StoreL ] [ StoreL ]
1931 //
1932 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
1933 // since we never unset _merge_stores_phase.
1934 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
1935 C->set_merge_stores_phase();
1936
1937 if (_for_merge_stores_igvn.length() > 0) {
1938 while (_for_merge_stores_igvn.length() > 0) {
1939 Node* n = _for_merge_stores_igvn.pop();
1940 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
1941 igvn._worklist.push(n);
1942 }
1943 igvn.optimize();
1944 if (failing()) return;
1945 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
1946 print_method(PHASE_AFTER_MERGE_STORES, 3);
1947 }
1948 }
1949
1950 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1951 if (OptimizeUnstableIf) {
1952 _unstable_if_traps.append(trap);
1953 }
1954 }
1955
1956 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1957 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1958 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1959 Node* n = trap->uncommon_trap();
1960 if (!useful.member(n)) {
1961 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1962 }
1963 }
1964 }
1965
1966 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1967 // or fold-compares case. Return true if succeed or not found.
1968 //
1969 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1970 // false and ask fold-compares to yield.
1971 //
1972 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1973 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1974 // when deoptimization does happen.
1975 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1976 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1977 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1978 if (trap->uncommon_trap() == unc) {
1979 if (yield && trap->modified()) {
1980 return false;
1981 }
1982 _unstable_if_traps.delete_at(i);
1983 break;
1984 }
1985 }
1986 return true;
1987 }
1988
1989 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1990 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1991 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1992 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1993 UnstableIfTrap* trap = _unstable_if_traps.at(i);
1994 CallStaticJavaNode* unc = trap->uncommon_trap();
1995 int next_bci = trap->next_bci();
1996 bool modified = trap->modified();
1997
1998 if (next_bci != -1 && !modified) {
1999 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2000 JVMState* jvms = unc->jvms();
2001 ciMethod* method = jvms->method();
2002 ciBytecodeStream iter(method);
2003
2004 iter.force_bci(jvms->bci());
2005 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2006 Bytecodes::Code c = iter.cur_bc();
2007 Node* lhs = nullptr;
2008 Node* rhs = nullptr;
2009 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2010 lhs = unc->peek_operand(0);
2011 rhs = unc->peek_operand(1);
2012 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2013 lhs = unc->peek_operand(0);
2014 }
2015
2016 ResourceMark rm;
2017 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2018 assert(live_locals.is_valid(), "broken liveness info");
2019 int len = (int)live_locals.size();
2020
2021 for (int i = 0; i < len; i++) {
2022 Node* local = unc->local(jvms, i);
2023 // kill local using the liveness of next_bci.
2024 // give up when the local looks like an operand to secure reexecution.
2025 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2026 uint idx = jvms->locoff() + i;
2027 #ifdef ASSERT
2028 if (PrintOpto && Verbose) {
2029 tty->print("[unstable_if] kill local#%d: ", idx);
2030 local->dump();
2031 tty->cr();
2032 }
2033 #endif
2034 igvn.replace_input_of(unc, idx, top());
2035 modified = true;
2036 }
2037 }
2038 }
2039
2040 // keep the mondified trap for late query
2041 if (modified) {
2042 trap->set_modified();
2043 } else {
2044 _unstable_if_traps.delete_at(i);
2045 }
2046 }
2047 igvn.optimize();
2048 }
2049
2050 // StringOpts and late inlining of string methods
2051 void Compile::inline_string_calls(bool parse_time) {
2052 {
2053 // remove useless nodes to make the usage analysis simpler
2054 ResourceMark rm;
2055 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2056 }
2057
2058 {
2059 ResourceMark rm;
2060 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2061 PhaseStringOpts pso(initial_gvn());
2062 print_method(PHASE_AFTER_STRINGOPTS, 3);
2063 }
2064
2065 // now inline anything that we skipped the first time around
2066 if (!parse_time) {
2067 _late_inlines_pos = _late_inlines.length();
2068 }
2069
2070 while (_string_late_inlines.length() > 0) {
2071 CallGenerator* cg = _string_late_inlines.pop();
2072 cg->do_late_inline();
2073 if (failing()) return;
2074 }
2075 _string_late_inlines.trunc_to(0);
2076 }
2077
2078 // Late inlining of boxing methods
2079 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2080 if (_boxing_late_inlines.length() > 0) {
2081 assert(has_boxed_value(), "inconsistent");
2082
2083 set_inlining_incrementally(true);
2084
2085 igvn_worklist()->ensure_empty(); // should be done with igvn
2086
2087 _late_inlines_pos = _late_inlines.length();
2088
2089 while (_boxing_late_inlines.length() > 0) {
2090 CallGenerator* cg = _boxing_late_inlines.pop();
2091 cg->do_late_inline();
2092 if (failing()) return;
2093 }
2094 _boxing_late_inlines.trunc_to(0);
2095
2096 inline_incrementally_cleanup(igvn);
2097
2098 set_inlining_incrementally(false);
2099 }
2100 }
2101
2102 bool Compile::inline_incrementally_one() {
2103 assert(IncrementalInline, "incremental inlining should be on");
2104 assert(_late_inlines.length() > 0, "should have been checked by caller");
2105
2106 TracePhase tp(_t_incrInline_inline);
2107
2108 set_inlining_progress(false);
2109 set_do_cleanup(false);
2110
2111 for (int i = 0; i < _late_inlines.length(); i++) {
2112 _late_inlines_pos = i+1;
2113 CallGenerator* cg = _late_inlines.at(i);
2114 bool is_scheduled_for_igvn_before = C->igvn_worklist()->member(cg->call_node());
2115 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2116 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2117 if (should_stress_inlining()) {
2118 // randomly add repeated inline attempt if stress-inlining
2119 cg->call_node()->set_generator(cg);
2120 C->igvn_worklist()->push(cg->call_node());
2121 continue;
2122 }
2123 cg->do_late_inline();
2124 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2125 if (failing()) {
2126 return false;
2127 } else if (inlining_progress()) {
2128 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2129 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2130 break; // process one call site at a time
2131 } else {
2132 bool is_scheduled_for_igvn_after = C->igvn_worklist()->member(cg->call_node());
2133 if (!is_scheduled_for_igvn_before && is_scheduled_for_igvn_after) {
2134 // Avoid potential infinite loop if node already in the IGVN list
2135 assert(false, "scheduled for IGVN during inlining attempt");
2136 } else {
2137 // Ensure call node has not disappeared from IGVN worklist during a failed inlining attempt
2138 assert(!is_scheduled_for_igvn_before || is_scheduled_for_igvn_after, "call node removed from IGVN list during inlining pass");
2139 cg->call_node()->set_generator(cg);
2140 }
2141 }
2142 } else {
2143 // Ignore late inline direct calls when inlining is not allowed.
2144 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2145 }
2146 }
2147 // Remove processed elements.
2148 _late_inlines.remove_till(_late_inlines_pos);
2149 _late_inlines_pos = 0;
2150
2151 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2152
2153 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2154
2155 set_inlining_progress(false);
2156 set_do_cleanup(false);
2157
2158 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2159 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2160 }
2161
2162 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2163 {
2164 TracePhase tp(_t_incrInline_pru);
2165 ResourceMark rm;
2166 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2167 }
2168 {
2169 TracePhase tp(_t_incrInline_igvn);
2170 igvn.reset();
2171 igvn.optimize();
2172 if (failing()) return;
2173 }
2174 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2175 }
2176
2177 template<typename E>
2178 static void shuffle_array(Compile& C, GrowableArray<E>& array) {
2179 if (array.length() < 2) {
2180 return;
2181 }
2182 for (uint i = array.length() - 1; i >= 1; i--) {
2183 uint j = C.random() % (i + 1);
2184 swap(array.at(i), array.at(j));
2185 }
2186 }
2187
2188 void Compile::shuffle_late_inlines() {
2189 shuffle_array(*C, _late_inlines);
2190 }
2191
2192 // Perform incremental inlining until bound on number of live nodes is reached
2193 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2194 TracePhase tp(_t_incrInline);
2195
2196 set_inlining_incrementally(true);
2197 uint low_live_nodes = 0;
2198
2199 if (StressIncrementalInlining) {
2200 shuffle_late_inlines();
2201 }
2202
2203 while (_late_inlines.length() > 0) {
2204 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2205 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2206 TracePhase tp(_t_incrInline_ideal);
2207 // PhaseIdealLoop is expensive so we only try it once we are
2208 // out of live nodes and we only try it again if the previous
2209 // helped got the number of nodes down significantly
2210 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2211 if (failing()) return;
2212 low_live_nodes = live_nodes();
2213 _major_progress = true;
2214 }
2215
2216 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2217 bool do_print_inlining = print_inlining() || print_intrinsics();
2218 if (do_print_inlining || log() != nullptr) {
2219 // Print inlining message for candidates that we couldn't inline for lack of space.
2220 for (int i = 0; i < _late_inlines.length(); i++) {
2221 CallGenerator* cg = _late_inlines.at(i);
2222 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2223 if (do_print_inlining) {
2224 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2225 }
2226 log_late_inline_failure(cg, msg);
2227 }
2228 }
2229 break; // finish
2230 }
2231 }
2232
2233 igvn_worklist()->ensure_empty(); // should be done with igvn
2234
2235 if (_late_inlines.length() == 0) {
2236 break; // no more progress
2237 }
2238
2239 while (inline_incrementally_one()) {
2240 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2241 }
2242 if (failing()) return;
2243
2244 inline_incrementally_cleanup(igvn);
2245
2246 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2247
2248 if (failing()) return;
2249 }
2250
2251 igvn_worklist()->ensure_empty(); // should be done with igvn
2252
2253 if (_string_late_inlines.length() > 0) {
2254 assert(has_stringbuilder(), "inconsistent");
2255
2256 inline_string_calls(false);
2257
2258 if (failing()) return;
2259
2260 inline_incrementally_cleanup(igvn);
2261 }
2262
2263 set_inlining_incrementally(false);
2264 }
2265
2266 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2267 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2268 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2269 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2270 // as if "inlining_incrementally() == true" were set.
2271 assert(inlining_incrementally() == false, "not allowed");
2272 assert(_modified_nodes == nullptr, "not allowed");
2273 assert(_late_inlines.length() > 0, "sanity");
2274
2275 if (StressIncrementalInlining) {
2276 shuffle_late_inlines();
2277 }
2278
2279 while (_late_inlines.length() > 0) {
2280 igvn_worklist()->ensure_empty(); // should be done with igvn
2281
2282 while (inline_incrementally_one()) {
2283 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2284 }
2285 if (failing()) return;
2286
2287 inline_incrementally_cleanup(igvn);
2288 }
2289 }
2290
2291 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2292 if (_loop_opts_cnt > 0) {
2293 while (major_progress() && (_loop_opts_cnt > 0)) {
2294 TracePhase tp(_t_idealLoop);
2295 PhaseIdealLoop::optimize(igvn, mode);
2296 _loop_opts_cnt--;
2297 if (failing()) return false;
2298 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2299 }
2300 }
2301 return true;
2302 }
2303
2304 // Remove edges from "root" to each SafePoint at a backward branch.
2305 // They were inserted during parsing (see add_safepoint()) to make
2306 // infinite loops without calls or exceptions visible to root, i.e.,
2307 // useful.
2308 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2309 Node *r = root();
2310 if (r != nullptr) {
2311 for (uint i = r->req(); i < r->len(); ++i) {
2312 Node *n = r->in(i);
2313 if (n != nullptr && n->is_SafePoint()) {
2314 r->rm_prec(i);
2315 if (n->outcnt() == 0) {
2316 igvn.remove_dead_node(n);
2317 }
2318 --i;
2319 }
2320 }
2321 // Parsing may have added top inputs to the root node (Path
2322 // leading to the Halt node proven dead). Make sure we get a
2323 // chance to clean them up.
2324 igvn._worklist.push(r);
2325 igvn.optimize();
2326 }
2327 }
2328
2329 //------------------------------Optimize---------------------------------------
2330 // Given a graph, optimize it.
2331 void Compile::Optimize() {
2332 TracePhase tp(_t_optimizer);
2333
2334 #ifndef PRODUCT
2335 if (env()->break_at_compile()) {
2336 BREAKPOINT;
2337 }
2338
2339 #endif
2340
2341 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2342 #ifdef ASSERT
2343 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2344 #endif
2345
2346 ResourceMark rm;
2347
2348 NOT_PRODUCT( verify_graph_edges(); )
2349
2350 print_method(PHASE_AFTER_PARSING, 1);
2351
2352 {
2353 // Iterative Global Value Numbering, including ideal transforms
2354 PhaseIterGVN igvn;
2355 #ifdef ASSERT
2356 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2357 #endif
2358 {
2359 TracePhase tp(_t_iterGVN);
2360 igvn.optimize();
2361 }
2362
2363 if (failing()) return;
2364
2365 print_method(PHASE_ITER_GVN1, 2);
2366
2367 process_for_unstable_if_traps(igvn);
2368
2369 if (failing()) return;
2370
2371 inline_incrementally(igvn);
2372
2373 print_method(PHASE_INCREMENTAL_INLINE, 2);
2374
2375 if (failing()) return;
2376
2377 if (eliminate_boxing()) {
2378 // Inline valueOf() methods now.
2379 inline_boxing_calls(igvn);
2380
2381 if (failing()) return;
2382
2383 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2384 inline_incrementally(igvn);
2385 }
2386
2387 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2388
2389 if (failing()) return;
2390 }
2391
2392 // Remove the speculative part of types and clean up the graph from
2393 // the extra CastPP nodes whose only purpose is to carry them. Do
2394 // that early so that optimizations are not disrupted by the extra
2395 // CastPP nodes.
2396 remove_speculative_types(igvn);
2397
2398 if (failing()) return;
2399
2400 // No more new expensive nodes will be added to the list from here
2401 // so keep only the actual candidates for optimizations.
2402 cleanup_expensive_nodes(igvn);
2403
2404 if (failing()) return;
2405
2406 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2407 if (EnableVectorSupport && has_vbox_nodes()) {
2408 TracePhase tp(_t_vector);
2409 PhaseVector pv(igvn);
2410 pv.optimize_vector_boxes();
2411 if (failing()) return;
2412 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2413 }
2414 assert(!has_vbox_nodes(), "sanity");
2415
2416 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2417 Compile::TracePhase tp(_t_renumberLive);
2418 igvn_worklist()->ensure_empty(); // should be done with igvn
2419 {
2420 ResourceMark rm;
2421 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2422 }
2423 igvn.reset();
2424 igvn.optimize();
2425 if (failing()) return;
2426 }
2427
2428 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2429 // safepoints
2430 remove_root_to_sfpts_edges(igvn);
2431
2432 if (failing()) return;
2433
2434 _print_phase_loop_opts = has_loops();
2435 if (_print_phase_loop_opts) {
2436 print_method(PHASE_BEFORE_LOOP_OPTS, 2);
2437 }
2438
2439 // Perform escape analysis
2440 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2441 if (has_loops()) {
2442 // Cleanup graph (remove dead nodes).
2443 TracePhase tp(_t_idealLoop);
2444 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2445 if (failing()) return;
2446 }
2447 bool progress;
2448 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2449 do {
2450 ConnectionGraph::do_analysis(this, &igvn);
2451
2452 if (failing()) return;
2453
2454 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2455
2456 // Optimize out fields loads from scalar replaceable allocations.
2457 igvn.optimize();
2458 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2459
2460 if (failing()) return;
2461
2462 if (congraph() != nullptr && macro_count() > 0) {
2463 TracePhase tp(_t_macroEliminate);
2464 PhaseMacroExpand mexp(igvn);
2465 mexp.eliminate_macro_nodes();
2466 if (failing()) return;
2467 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2468
2469 igvn.set_delay_transform(false);
2470 igvn.optimize();
2471 if (failing()) return;
2472
2473 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2474 }
2475
2476 ConnectionGraph::verify_ram_nodes(this, root());
2477 if (failing()) return;
2478
2479 progress = do_iterative_escape_analysis() &&
2480 (macro_count() < mcount) &&
2481 ConnectionGraph::has_candidates(this);
2482 // Try again if candidates exist and made progress
2483 // by removing some allocations and/or locks.
2484 } while (progress);
2485 }
2486
2487 // Loop transforms on the ideal graph. Range Check Elimination,
2488 // peeling, unrolling, etc.
2489
2490 // Set loop opts counter
2491 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2492 {
2493 TracePhase tp(_t_idealLoop);
2494 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2495 _loop_opts_cnt--;
2496 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2497 if (failing()) return;
2498 }
2499 // Loop opts pass if partial peeling occurred in previous pass
2500 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2501 TracePhase tp(_t_idealLoop);
2502 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2503 _loop_opts_cnt--;
2504 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2505 if (failing()) return;
2506 }
2507 // Loop opts pass for loop-unrolling before CCP
2508 if(major_progress() && (_loop_opts_cnt > 0)) {
2509 TracePhase tp(_t_idealLoop);
2510 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2511 _loop_opts_cnt--;
2512 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2513 }
2514 if (!failing()) {
2515 // Verify that last round of loop opts produced a valid graph
2516 PhaseIdealLoop::verify(igvn);
2517 }
2518 }
2519 if (failing()) return;
2520
2521 // Conditional Constant Propagation;
2522 print_method(PHASE_BEFORE_CCP1, 2);
2523 PhaseCCP ccp( &igvn );
2524 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2525 {
2526 TracePhase tp(_t_ccp);
2527 ccp.do_transform();
2528 }
2529 print_method(PHASE_CCP1, 2);
2530
2531 assert( true, "Break here to ccp.dump_old2new_map()");
2532
2533 // Iterative Global Value Numbering, including ideal transforms
2534 {
2535 TracePhase tp(_t_iterGVN2);
2536 igvn.reset_from_igvn(&ccp);
2537 igvn.optimize();
2538 }
2539 print_method(PHASE_ITER_GVN2, 2);
2540
2541 if (failing()) return;
2542
2543 // Loop transforms on the ideal graph. Range Check Elimination,
2544 // peeling, unrolling, etc.
2545 if (!optimize_loops(igvn, LoopOptsDefault)) {
2546 return;
2547 }
2548
2549 if (failing()) return;
2550
2551 C->clear_major_progress(); // ensure that major progress is now clear
2552
2553 process_for_post_loop_opts_igvn(igvn);
2554
2555 process_for_merge_stores_igvn(igvn);
2556
2557 if (failing()) return;
2558
2559 #ifdef ASSERT
2560 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2561 #endif
2562
2563 {
2564 TracePhase tp(_t_macroExpand);
2565 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2566 PhaseMacroExpand mex(igvn);
2567 // Do not allow new macro nodes once we start to eliminate and expand
2568 C->reset_allow_macro_nodes();
2569 // Last attempt to eliminate macro nodes before expand
2570 mex.eliminate_macro_nodes();
2571 if (failing()) {
2572 return;
2573 }
2574 mex.eliminate_opaque_looplimit_macro_nodes();
2575 if (failing()) {
2576 return;
2577 }
2578 print_method(PHASE_AFTER_MACRO_ELIMINATION, 2);
2579 if (mex.expand_macro_nodes()) {
2580 assert(failing(), "must bail out w/ explicit message");
2581 return;
2582 }
2583 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2584 }
2585
2586 {
2587 TracePhase tp(_t_barrierExpand);
2588 if (bs->expand_barriers(this, igvn)) {
2589 assert(failing(), "must bail out w/ explicit message");
2590 return;
2591 }
2592 print_method(PHASE_BARRIER_EXPANSION, 2);
2593 }
2594
2595 if (C->max_vector_size() > 0) {
2596 C->optimize_logic_cones(igvn);
2597 igvn.optimize();
2598 if (failing()) return;
2599 }
2600
2601 DEBUG_ONLY( _modified_nodes = nullptr; )
2602
2603 assert(igvn._worklist.size() == 0, "not empty");
2604
2605 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2606
2607 if (_late_inlines.length() > 0) {
2608 // More opportunities to optimize virtual and MH calls.
2609 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2610 process_late_inline_calls_no_inline(igvn);
2611 if (failing()) return;
2612 }
2613 } // (End scope of igvn; run destructor if necessary for asserts.)
2614
2615 check_no_dead_use();
2616
2617 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
2618 // to remove hashes to unlock nodes for modifications.
2619 C->node_hash()->clear();
2620
2621 // A method with only infinite loops has no edges entering loops from root
2622 {
2623 TracePhase tp(_t_graphReshaping);
2624 if (final_graph_reshaping()) {
2625 assert(failing(), "must bail out w/ explicit message");
2626 return;
2627 }
2628 }
2629
2630 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2631 DEBUG_ONLY(set_phase_optimize_finished();)
2632 }
2633
2634 #ifdef ASSERT
2635 void Compile::check_no_dead_use() const {
2636 ResourceMark rm;
2637 Unique_Node_List wq;
2638 wq.push(root());
2639 for (uint i = 0; i < wq.size(); ++i) {
2640 Node* n = wq.at(i);
2641 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2642 Node* u = n->fast_out(j);
2643 if (u->outcnt() == 0 && !u->is_Con()) {
2644 u->dump();
2645 fatal("no reachable node should have no use");
2646 }
2647 wq.push(u);
2648 }
2649 }
2650 }
2651 #endif
2652
2653 void Compile::inline_vector_reboxing_calls() {
2654 if (C->_vector_reboxing_late_inlines.length() > 0) {
2655 _late_inlines_pos = C->_late_inlines.length();
2656 while (_vector_reboxing_late_inlines.length() > 0) {
2657 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2658 cg->do_late_inline();
2659 if (failing()) return;
2660 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2661 }
2662 _vector_reboxing_late_inlines.trunc_to(0);
2663 }
2664 }
2665
2666 bool Compile::has_vbox_nodes() {
2667 if (C->_vector_reboxing_late_inlines.length() > 0) {
2668 return true;
2669 }
2670 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2671 Node * n = C->macro_node(macro_idx);
2672 assert(n->is_macro(), "only macro nodes expected here");
2673 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2674 return true;
2675 }
2676 }
2677 return false;
2678 }
2679
2680 //---------------------------- Bitwise operation packing optimization ---------------------------
2681
2682 static bool is_vector_unary_bitwise_op(Node* n) {
2683 return n->Opcode() == Op_XorV &&
2684 VectorNode::is_vector_bitwise_not_pattern(n);
2685 }
2686
2687 static bool is_vector_binary_bitwise_op(Node* n) {
2688 switch (n->Opcode()) {
2689 case Op_AndV:
2690 case Op_OrV:
2691 return true;
2692
2693 case Op_XorV:
2694 return !is_vector_unary_bitwise_op(n);
2695
2696 default:
2697 return false;
2698 }
2699 }
2700
2701 static bool is_vector_ternary_bitwise_op(Node* n) {
2702 return n->Opcode() == Op_MacroLogicV;
2703 }
2704
2705 static bool is_vector_bitwise_op(Node* n) {
2706 return is_vector_unary_bitwise_op(n) ||
2707 is_vector_binary_bitwise_op(n) ||
2708 is_vector_ternary_bitwise_op(n);
2709 }
2710
2711 static bool is_vector_bitwise_cone_root(Node* n) {
2712 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2713 return false;
2714 }
2715 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2716 if (is_vector_bitwise_op(n->fast_out(i))) {
2717 return false;
2718 }
2719 }
2720 return true;
2721 }
2722
2723 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2724 uint cnt = 0;
2725 if (is_vector_bitwise_op(n)) {
2726 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2727 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2728 for (uint i = 1; i < inp_cnt; i++) {
2729 Node* in = n->in(i);
2730 bool skip = VectorNode::is_all_ones_vector(in);
2731 if (!skip && !inputs.member(in)) {
2732 inputs.push(in);
2733 cnt++;
2734 }
2735 }
2736 assert(cnt <= 1, "not unary");
2737 } else {
2738 uint last_req = inp_cnt;
2739 if (is_vector_ternary_bitwise_op(n)) {
2740 last_req = inp_cnt - 1; // skip last input
2741 }
2742 for (uint i = 1; i < last_req; i++) {
2743 Node* def = n->in(i);
2744 if (!inputs.member(def)) {
2745 inputs.push(def);
2746 cnt++;
2747 }
2748 }
2749 }
2750 } else { // not a bitwise operations
2751 if (!inputs.member(n)) {
2752 inputs.push(n);
2753 cnt++;
2754 }
2755 }
2756 return cnt;
2757 }
2758
2759 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2760 Unique_Node_List useful_nodes;
2761 C->identify_useful_nodes(useful_nodes);
2762
2763 for (uint i = 0; i < useful_nodes.size(); i++) {
2764 Node* n = useful_nodes.at(i);
2765 if (is_vector_bitwise_cone_root(n)) {
2766 list.push(n);
2767 }
2768 }
2769 }
2770
2771 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2772 const TypeVect* vt,
2773 Unique_Node_List& partition,
2774 Unique_Node_List& inputs) {
2775 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2776 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2777 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2778
2779 Node* in1 = inputs.at(0);
2780 Node* in2 = inputs.at(1);
2781 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2782
2783 uint func = compute_truth_table(partition, inputs);
2784
2785 Node* pn = partition.at(partition.size() - 1);
2786 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
2787 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2788 }
2789
2790 static uint extract_bit(uint func, uint pos) {
2791 return (func & (1 << pos)) >> pos;
2792 }
2793
2794 //
2795 // A macro logic node represents a truth table. It has 4 inputs,
2796 // First three inputs corresponds to 3 columns of a truth table
2797 // and fourth input captures the logic function.
2798 //
2799 // eg. fn = (in1 AND in2) OR in3;
2800 //
2801 // MacroNode(in1,in2,in3,fn)
2802 //
2803 // -----------------
2804 // in1 in2 in3 fn
2805 // -----------------
2806 // 0 0 0 0
2807 // 0 0 1 1
2808 // 0 1 0 0
2809 // 0 1 1 1
2810 // 1 0 0 0
2811 // 1 0 1 1
2812 // 1 1 0 1
2813 // 1 1 1 1
2814 //
2815
2816 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2817 int res = 0;
2818 for (int i = 0; i < 8; i++) {
2819 int bit1 = extract_bit(in1, i);
2820 int bit2 = extract_bit(in2, i);
2821 int bit3 = extract_bit(in3, i);
2822
2823 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2824 int func_bit = extract_bit(func, func_bit_pos);
2825
2826 res |= func_bit << i;
2827 }
2828 return res;
2829 }
2830
2831 static uint eval_operand(Node* n, HashTable<Node*,uint>& eval_map) {
2832 assert(n != nullptr, "");
2833 assert(eval_map.contains(n), "absent");
2834 return *(eval_map.get(n));
2835 }
2836
2837 static void eval_operands(Node* n,
2838 uint& func1, uint& func2, uint& func3,
2839 HashTable<Node*,uint>& eval_map) {
2840 assert(is_vector_bitwise_op(n), "");
2841
2842 if (is_vector_unary_bitwise_op(n)) {
2843 Node* opnd = n->in(1);
2844 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2845 opnd = n->in(2);
2846 }
2847 func1 = eval_operand(opnd, eval_map);
2848 } else if (is_vector_binary_bitwise_op(n)) {
2849 func1 = eval_operand(n->in(1), eval_map);
2850 func2 = eval_operand(n->in(2), eval_map);
2851 } else {
2852 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2853 func1 = eval_operand(n->in(1), eval_map);
2854 func2 = eval_operand(n->in(2), eval_map);
2855 func3 = eval_operand(n->in(3), eval_map);
2856 }
2857 }
2858
2859 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2860 assert(inputs.size() <= 3, "sanity");
2861 ResourceMark rm;
2862 uint res = 0;
2863 HashTable<Node*,uint> eval_map;
2864
2865 // Populate precomputed functions for inputs.
2866 // Each input corresponds to one column of 3 input truth-table.
2867 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2868 0xCC, // (_, b, _) -> b
2869 0xF0 }; // (a, _, _) -> a
2870 for (uint i = 0; i < inputs.size(); i++) {
2871 eval_map.put(inputs.at(i), input_funcs[2-i]);
2872 }
2873
2874 for (uint i = 0; i < partition.size(); i++) {
2875 Node* n = partition.at(i);
2876
2877 uint func1 = 0, func2 = 0, func3 = 0;
2878 eval_operands(n, func1, func2, func3, eval_map);
2879
2880 switch (n->Opcode()) {
2881 case Op_OrV:
2882 assert(func3 == 0, "not binary");
2883 res = func1 | func2;
2884 break;
2885 case Op_AndV:
2886 assert(func3 == 0, "not binary");
2887 res = func1 & func2;
2888 break;
2889 case Op_XorV:
2890 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2891 assert(func2 == 0 && func3 == 0, "not unary");
2892 res = (~func1) & 0xFF;
2893 } else {
2894 assert(func3 == 0, "not binary");
2895 res = func1 ^ func2;
2896 }
2897 break;
2898 case Op_MacroLogicV:
2899 // Ordering of inputs may change during evaluation of sub-tree
2900 // containing MacroLogic node as a child node, thus a re-evaluation
2901 // makes sure that function is evaluated in context of current
2902 // inputs.
2903 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2904 break;
2905
2906 default: assert(false, "not supported: %s", n->Name());
2907 }
2908 assert(res <= 0xFF, "invalid");
2909 eval_map.put(n, res);
2910 }
2911 return res;
2912 }
2913
2914 // Criteria under which nodes gets packed into a macro logic node:-
2915 // 1) Parent and both child nodes are all unmasked or masked with
2916 // same predicates.
2917 // 2) Masked parent can be packed with left child if it is predicated
2918 // and both have same predicates.
2919 // 3) Masked parent can be packed with right child if its un-predicated
2920 // or has matching predication condition.
2921 // 4) An unmasked parent can be packed with an unmasked child.
2922 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2923 assert(partition.size() == 0, "not empty");
2924 assert(inputs.size() == 0, "not empty");
2925 if (is_vector_ternary_bitwise_op(n)) {
2926 return false;
2927 }
2928
2929 bool is_unary_op = is_vector_unary_bitwise_op(n);
2930 if (is_unary_op) {
2931 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2932 return false; // too few inputs
2933 }
2934
2935 bool pack_left_child = true;
2936 bool pack_right_child = true;
2937
2938 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2939 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2940
2941 int left_child_input_cnt = 0;
2942 int right_child_input_cnt = 0;
2943
2944 bool parent_is_predicated = n->is_predicated_vector();
2945 bool left_child_predicated = n->in(1)->is_predicated_vector();
2946 bool right_child_predicated = n->in(2)->is_predicated_vector();
2947
2948 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
2949 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2950 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
2951
2952 do {
2953 if (pack_left_child && left_child_LOP &&
2954 ((!parent_is_predicated && !left_child_predicated) ||
2955 ((parent_is_predicated && left_child_predicated &&
2956 parent_pred == left_child_pred)))) {
2957 partition.push(n->in(1));
2958 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2959 } else {
2960 inputs.push(n->in(1));
2961 left_child_input_cnt = 1;
2962 }
2963
2964 if (pack_right_child && right_child_LOP &&
2965 (!right_child_predicated ||
2966 (right_child_predicated && parent_is_predicated &&
2967 parent_pred == right_child_pred))) {
2968 partition.push(n->in(2));
2969 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2970 } else {
2971 inputs.push(n->in(2));
2972 right_child_input_cnt = 1;
2973 }
2974
2975 if (inputs.size() > 3) {
2976 assert(partition.size() > 0, "");
2977 inputs.clear();
2978 partition.clear();
2979 if (left_child_input_cnt > right_child_input_cnt) {
2980 pack_left_child = false;
2981 } else {
2982 pack_right_child = false;
2983 }
2984 } else {
2985 break;
2986 }
2987 } while(true);
2988
2989 if(partition.size()) {
2990 partition.push(n);
2991 }
2992
2993 return (partition.size() == 2 || partition.size() == 3) &&
2994 (inputs.size() == 2 || inputs.size() == 3);
2995 }
2996
2997 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2998 assert(is_vector_bitwise_op(n), "not a root");
2999
3000 visited.set(n->_idx);
3001
3002 // 1) Do a DFS walk over the logic cone.
3003 for (uint i = 1; i < n->req(); i++) {
3004 Node* in = n->in(i);
3005 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3006 process_logic_cone_root(igvn, in, visited);
3007 }
3008 }
3009
3010 // 2) Bottom up traversal: Merge node[s] with
3011 // the parent to form macro logic node.
3012 Unique_Node_List partition;
3013 Unique_Node_List inputs;
3014 if (compute_logic_cone(n, partition, inputs)) {
3015 const TypeVect* vt = n->bottom_type()->is_vect();
3016 Node* pn = partition.at(partition.size() - 1);
3017 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3018 if (mask == nullptr ||
3019 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3020 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3021 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3022 igvn.replace_node(n, macro_logic);
3023 }
3024 }
3025 }
3026
3027 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3028 ResourceMark rm;
3029 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3030 Unique_Node_List list;
3031 collect_logic_cone_roots(list);
3032
3033 while (list.size() > 0) {
3034 Node* n = list.pop();
3035 const TypeVect* vt = n->bottom_type()->is_vect();
3036 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3037 if (supported) {
3038 VectorSet visited(comp_arena());
3039 process_logic_cone_root(igvn, n, visited);
3040 }
3041 }
3042 }
3043 }
3044
3045 //------------------------------Code_Gen---------------------------------------
3046 // Given a graph, generate code for it
3047 void Compile::Code_Gen() {
3048 if (failing()) {
3049 return;
3050 }
3051
3052 // Perform instruction selection. You might think we could reclaim Matcher
3053 // memory PDQ, but actually the Matcher is used in generating spill code.
3054 // Internals of the Matcher (including some VectorSets) must remain live
3055 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3056 // set a bit in reclaimed memory.
3057
3058 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3059 // nodes. Mapping is only valid at the root of each matched subtree.
3060 NOT_PRODUCT( verify_graph_edges(); )
3061
3062 Matcher matcher;
3063 _matcher = &matcher;
3064 {
3065 TracePhase tp(_t_matcher);
3066 matcher.match();
3067 if (failing()) {
3068 return;
3069 }
3070 }
3071 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3072 // nodes. Mapping is only valid at the root of each matched subtree.
3073 NOT_PRODUCT( verify_graph_edges(); )
3074
3075 // If you have too many nodes, or if matching has failed, bail out
3076 check_node_count(0, "out of nodes matching instructions");
3077 if (failing()) {
3078 return;
3079 }
3080
3081 print_method(PHASE_MATCHING, 2);
3082
3083 // Build a proper-looking CFG
3084 PhaseCFG cfg(node_arena(), root(), matcher);
3085 if (failing()) {
3086 return;
3087 }
3088 _cfg = &cfg;
3089 {
3090 TracePhase tp(_t_scheduler);
3091 bool success = cfg.do_global_code_motion();
3092 if (!success) {
3093 return;
3094 }
3095
3096 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3097 NOT_PRODUCT( verify_graph_edges(); )
3098 cfg.verify();
3099 if (failing()) {
3100 return;
3101 }
3102 }
3103
3104 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3105 _regalloc = ®alloc;
3106 {
3107 TracePhase tp(_t_registerAllocation);
3108 // Perform register allocation. After Chaitin, use-def chains are
3109 // no longer accurate (at spill code) and so must be ignored.
3110 // Node->LRG->reg mappings are still accurate.
3111 _regalloc->Register_Allocate();
3112
3113 // Bail out if the allocator builds too many nodes
3114 if (failing()) {
3115 return;
3116 }
3117
3118 print_method(PHASE_REGISTER_ALLOCATION, 2);
3119 }
3120
3121 // Prior to register allocation we kept empty basic blocks in case the
3122 // the allocator needed a place to spill. After register allocation we
3123 // are not adding any new instructions. If any basic block is empty, we
3124 // can now safely remove it.
3125 {
3126 TracePhase tp(_t_blockOrdering);
3127 cfg.remove_empty_blocks();
3128 if (do_freq_based_layout()) {
3129 PhaseBlockLayout layout(cfg);
3130 } else {
3131 cfg.set_loop_alignment();
3132 }
3133 cfg.fixup_flow();
3134 cfg.remove_unreachable_blocks();
3135 cfg.verify_dominator_tree();
3136 print_method(PHASE_BLOCK_ORDERING, 3);
3137 }
3138
3139 // Apply peephole optimizations
3140 if( OptoPeephole ) {
3141 TracePhase tp(_t_peephole);
3142 PhasePeephole peep( _regalloc, cfg);
3143 peep.do_transform();
3144 print_method(PHASE_PEEPHOLE, 3);
3145 }
3146
3147 // Do late expand if CPU requires this.
3148 if (Matcher::require_postalloc_expand) {
3149 TracePhase tp(_t_postalloc_expand);
3150 cfg.postalloc_expand(_regalloc);
3151 print_method(PHASE_POSTALLOC_EXPAND, 3);
3152 }
3153
3154 #ifdef ASSERT
3155 {
3156 CompilationMemoryStatistic::do_test_allocations();
3157 if (failing()) return;
3158 }
3159 #endif
3160
3161 // Convert Nodes to instruction bits in a buffer
3162 {
3163 TracePhase tp(_t_output);
3164 PhaseOutput output;
3165 output.Output();
3166 if (failing()) return;
3167 output.install();
3168 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3169 }
3170
3171 // He's dead, Jim.
3172 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3173 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3174 }
3175
3176 //------------------------------Final_Reshape_Counts---------------------------
3177 // This class defines counters to help identify when a method
3178 // may/must be executed using hardware with only 24-bit precision.
3179 struct Final_Reshape_Counts : public StackObj {
3180 int _call_count; // count non-inlined 'common' calls
3181 int _float_count; // count float ops requiring 24-bit precision
3182 int _double_count; // count double ops requiring more precision
3183 int _java_call_count; // count non-inlined 'java' calls
3184 int _inner_loop_count; // count loops which need alignment
3185 VectorSet _visited; // Visitation flags
3186 Node_List _tests; // Set of IfNodes & PCTableNodes
3187
3188 Final_Reshape_Counts() :
3189 _call_count(0), _float_count(0), _double_count(0),
3190 _java_call_count(0), _inner_loop_count(0) { }
3191
3192 void inc_call_count () { _call_count ++; }
3193 void inc_float_count () { _float_count ++; }
3194 void inc_double_count() { _double_count++; }
3195 void inc_java_call_count() { _java_call_count++; }
3196 void inc_inner_loop_count() { _inner_loop_count++; }
3197
3198 int get_call_count () const { return _call_count ; }
3199 int get_float_count () const { return _float_count ; }
3200 int get_double_count() const { return _double_count; }
3201 int get_java_call_count() const { return _java_call_count; }
3202 int get_inner_loop_count() const { return _inner_loop_count; }
3203 };
3204
3205 //------------------------------final_graph_reshaping_impl----------------------
3206 // Implement items 1-5 from final_graph_reshaping below.
3207 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3208
3209 if ( n->outcnt() == 0 ) return; // dead node
3210 uint nop = n->Opcode();
3211
3212 // Check for 2-input instruction with "last use" on right input.
3213 // Swap to left input. Implements item (2).
3214 if( n->req() == 3 && // two-input instruction
3215 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3216 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3217 n->in(2)->outcnt() == 1 &&// right use IS a last use
3218 !n->in(2)->is_Con() ) { // right use is not a constant
3219 // Check for commutative opcode
3220 switch( nop ) {
3221 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3222 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3223 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3224 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3225 case Op_AndL: case Op_XorL: case Op_OrL:
3226 case Op_AndI: case Op_XorI: case Op_OrI: {
3227 // Move "last use" input to left by swapping inputs
3228 n->swap_edges(1, 2);
3229 break;
3230 }
3231 default:
3232 break;
3233 }
3234 }
3235
3236 #ifdef ASSERT
3237 if( n->is_Mem() ) {
3238 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3239 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3240 // oop will be recorded in oop map if load crosses safepoint
3241 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3242 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3243 "raw memory operations should have control edge");
3244 }
3245 if (n->is_MemBar()) {
3246 MemBarNode* mb = n->as_MemBar();
3247 if (mb->trailing_store() || mb->trailing_load_store()) {
3248 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3249 Node* mem = mb->in(MemBarNode::Precedent);
3250 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3251 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3252 } else if (mb->leading()) {
3253 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3254 }
3255 }
3256 #endif
3257 // Count FPU ops and common calls, implements item (3)
3258 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3259
3260 // Collect CFG split points
3261 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3262 frc._tests.push(n);
3263 }
3264 }
3265
3266 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3267 if (!UseDivMod) {
3268 return;
3269 }
3270
3271 // Check if "a % b" and "a / b" both exist
3272 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3273 if (d == nullptr) {
3274 return;
3275 }
3276
3277 // Replace them with a fused divmod if supported
3278 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3279 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3280 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3281 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3282 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3283 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3284 // DivMod node so the dependency is not lost.
3285 divmod->add_prec_from(n);
3286 divmod->add_prec_from(d);
3287 d->subsume_by(divmod->div_proj(), this);
3288 n->subsume_by(divmod->mod_proj(), this);
3289 } else {
3290 // Replace "a % b" with "a - ((a / b) * b)"
3291 Node* mult = MulNode::make(d, d->in(2), bt);
3292 Node* sub = SubNode::make(d->in(1), mult, bt);
3293 n->subsume_by(sub, this);
3294 }
3295 }
3296
3297 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3298 switch( nop ) {
3299 // Count all float operations that may use FPU
3300 case Op_AddF:
3301 case Op_SubF:
3302 case Op_MulF:
3303 case Op_DivF:
3304 case Op_NegF:
3305 case Op_ModF:
3306 case Op_ConvI2F:
3307 case Op_ConF:
3308 case Op_CmpF:
3309 case Op_CmpF3:
3310 case Op_StoreF:
3311 case Op_LoadF:
3312 // case Op_ConvL2F: // longs are split into 32-bit halves
3313 frc.inc_float_count();
3314 break;
3315
3316 case Op_ConvF2D:
3317 case Op_ConvD2F:
3318 frc.inc_float_count();
3319 frc.inc_double_count();
3320 break;
3321
3322 // Count all double operations that may use FPU
3323 case Op_AddD:
3324 case Op_SubD:
3325 case Op_MulD:
3326 case Op_DivD:
3327 case Op_NegD:
3328 case Op_ModD:
3329 case Op_ConvI2D:
3330 case Op_ConvD2I:
3331 // case Op_ConvL2D: // handled by leaf call
3332 // case Op_ConvD2L: // handled by leaf call
3333 case Op_ConD:
3334 case Op_CmpD:
3335 case Op_CmpD3:
3336 case Op_StoreD:
3337 case Op_LoadD:
3338 case Op_LoadD_unaligned:
3339 frc.inc_double_count();
3340 break;
3341 case Op_Opaque1: // Remove Opaque Nodes before matching
3342 n->subsume_by(n->in(1), this);
3343 break;
3344 case Op_CallLeafPure: {
3345 // If the pure call is not supported, then lower to a CallLeaf.
3346 if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
3347 CallNode* call = n->as_Call();
3348 CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
3349 call->_name, TypeRawPtr::BOTTOM);
3350 new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
3351 new_call->init_req(TypeFunc::I_O, C->top());
3352 new_call->init_req(TypeFunc::Memory, C->top());
3353 new_call->init_req(TypeFunc::ReturnAdr, C->top());
3354 new_call->init_req(TypeFunc::FramePtr, C->top());
3355 for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
3356 new_call->init_req(i, call->in(i));
3357 }
3358 n->subsume_by(new_call, this);
3359 }
3360 frc.inc_call_count();
3361 break;
3362 }
3363 case Op_CallStaticJava:
3364 case Op_CallJava:
3365 case Op_CallDynamicJava:
3366 frc.inc_java_call_count(); // Count java call site;
3367 case Op_CallRuntime:
3368 case Op_CallLeaf:
3369 case Op_CallLeafVector:
3370 case Op_CallLeafNoFP: {
3371 assert (n->is_Call(), "");
3372 CallNode *call = n->as_Call();
3373 // Count call sites where the FP mode bit would have to be flipped.
3374 // Do not count uncommon runtime calls:
3375 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3376 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3377 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3378 frc.inc_call_count(); // Count the call site
3379 } else { // See if uncommon argument is shared
3380 Node *n = call->in(TypeFunc::Parms);
3381 int nop = n->Opcode();
3382 // Clone shared simple arguments to uncommon calls, item (1).
3383 if (n->outcnt() > 1 &&
3384 !n->is_Proj() &&
3385 nop != Op_CreateEx &&
3386 nop != Op_CheckCastPP &&
3387 nop != Op_DecodeN &&
3388 nop != Op_DecodeNKlass &&
3389 !n->is_Mem() &&
3390 !n->is_Phi()) {
3391 Node *x = n->clone();
3392 call->set_req(TypeFunc::Parms, x);
3393 }
3394 }
3395 break;
3396 }
3397 case Op_StoreB:
3398 case Op_StoreC:
3399 case Op_StoreI:
3400 case Op_StoreL:
3401 case Op_CompareAndSwapB:
3402 case Op_CompareAndSwapS:
3403 case Op_CompareAndSwapI:
3404 case Op_CompareAndSwapL:
3405 case Op_CompareAndSwapP:
3406 case Op_CompareAndSwapN:
3407 case Op_WeakCompareAndSwapB:
3408 case Op_WeakCompareAndSwapS:
3409 case Op_WeakCompareAndSwapI:
3410 case Op_WeakCompareAndSwapL:
3411 case Op_WeakCompareAndSwapP:
3412 case Op_WeakCompareAndSwapN:
3413 case Op_CompareAndExchangeB:
3414 case Op_CompareAndExchangeS:
3415 case Op_CompareAndExchangeI:
3416 case Op_CompareAndExchangeL:
3417 case Op_CompareAndExchangeP:
3418 case Op_CompareAndExchangeN:
3419 case Op_GetAndAddS:
3420 case Op_GetAndAddB:
3421 case Op_GetAndAddI:
3422 case Op_GetAndAddL:
3423 case Op_GetAndSetS:
3424 case Op_GetAndSetB:
3425 case Op_GetAndSetI:
3426 case Op_GetAndSetL:
3427 case Op_GetAndSetP:
3428 case Op_GetAndSetN:
3429 case Op_StoreP:
3430 case Op_StoreN:
3431 case Op_StoreNKlass:
3432 case Op_LoadB:
3433 case Op_LoadUB:
3434 case Op_LoadUS:
3435 case Op_LoadI:
3436 case Op_LoadKlass:
3437 case Op_LoadNKlass:
3438 case Op_LoadL:
3439 case Op_LoadL_unaligned:
3440 case Op_LoadP:
3441 case Op_LoadN:
3442 case Op_LoadRange:
3443 case Op_LoadS:
3444 break;
3445
3446 case Op_AddP: { // Assert sane base pointers
3447 Node *addp = n->in(AddPNode::Address);
3448 assert(n->as_AddP()->address_input_has_same_base(), "Base pointers must match (addp %u)", addp->_idx );
3449 #ifdef _LP64
3450 if ((UseCompressedOops || UseCompressedClassPointers) &&
3451 addp->Opcode() == Op_ConP &&
3452 addp == n->in(AddPNode::Base) &&
3453 n->in(AddPNode::Offset)->is_Con()) {
3454 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3455 // on the platform and on the compressed oops mode.
3456 // Use addressing with narrow klass to load with offset on x86.
3457 // Some platforms can use the constant pool to load ConP.
3458 // Do this transformation here since IGVN will convert ConN back to ConP.
3459 const Type* t = addp->bottom_type();
3460 bool is_oop = t->isa_oopptr() != nullptr;
3461 bool is_klass = t->isa_klassptr() != nullptr;
3462
3463 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3464 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3465 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3466 Node* nn = nullptr;
3467
3468 int op = is_oop ? Op_ConN : Op_ConNKlass;
3469
3470 // Look for existing ConN node of the same exact type.
3471 Node* r = root();
3472 uint cnt = r->outcnt();
3473 for (uint i = 0; i < cnt; i++) {
3474 Node* m = r->raw_out(i);
3475 if (m!= nullptr && m->Opcode() == op &&
3476 m->bottom_type()->make_ptr() == t) {
3477 nn = m;
3478 break;
3479 }
3480 }
3481 if (nn != nullptr) {
3482 // Decode a narrow oop to match address
3483 // [R12 + narrow_oop_reg<<3 + offset]
3484 if (is_oop) {
3485 nn = new DecodeNNode(nn, t);
3486 } else {
3487 nn = new DecodeNKlassNode(nn, t);
3488 }
3489 // Check for succeeding AddP which uses the same Base.
3490 // Otherwise we will run into the assertion above when visiting that guy.
3491 for (uint i = 0; i < n->outcnt(); ++i) {
3492 Node *out_i = n->raw_out(i);
3493 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3494 out_i->set_req(AddPNode::Base, nn);
3495 #ifdef ASSERT
3496 for (uint j = 0; j < out_i->outcnt(); ++j) {
3497 Node *out_j = out_i->raw_out(j);
3498 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3499 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3500 }
3501 #endif
3502 }
3503 }
3504 n->set_req(AddPNode::Base, nn);
3505 n->set_req(AddPNode::Address, nn);
3506 if (addp->outcnt() == 0) {
3507 addp->disconnect_inputs(this);
3508 }
3509 }
3510 }
3511 }
3512 #endif
3513 break;
3514 }
3515
3516 case Op_CastPP: {
3517 // Remove CastPP nodes to gain more freedom during scheduling but
3518 // keep the dependency they encode as control or precedence edges
3519 // (if control is set already) on memory operations. Some CastPP
3520 // nodes don't have a control (don't carry a dependency): skip
3521 // those.
3522 if (n->in(0) != nullptr) {
3523 ResourceMark rm;
3524 Unique_Node_List wq;
3525 wq.push(n);
3526 for (uint next = 0; next < wq.size(); ++next) {
3527 Node *m = wq.at(next);
3528 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3529 Node* use = m->fast_out(i);
3530 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3531 use->ensure_control_or_add_prec(n->in(0));
3532 } else {
3533 switch(use->Opcode()) {
3534 case Op_AddP:
3535 case Op_DecodeN:
3536 case Op_DecodeNKlass:
3537 case Op_CheckCastPP:
3538 case Op_CastPP:
3539 wq.push(use);
3540 break;
3541 }
3542 }
3543 }
3544 }
3545 }
3546 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3547 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3548 Node* in1 = n->in(1);
3549 const Type* t = n->bottom_type();
3550 Node* new_in1 = in1->clone();
3551 new_in1->as_DecodeN()->set_type(t);
3552
3553 if (!Matcher::narrow_oop_use_complex_address()) {
3554 //
3555 // x86, ARM and friends can handle 2 adds in addressing mode
3556 // and Matcher can fold a DecodeN node into address by using
3557 // a narrow oop directly and do implicit null check in address:
3558 //
3559 // [R12 + narrow_oop_reg<<3 + offset]
3560 // NullCheck narrow_oop_reg
3561 //
3562 // On other platforms (Sparc) we have to keep new DecodeN node and
3563 // use it to do implicit null check in address:
3564 //
3565 // decode_not_null narrow_oop_reg, base_reg
3566 // [base_reg + offset]
3567 // NullCheck base_reg
3568 //
3569 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3570 // to keep the information to which null check the new DecodeN node
3571 // corresponds to use it as value in implicit_null_check().
3572 //
3573 new_in1->set_req(0, n->in(0));
3574 }
3575
3576 n->subsume_by(new_in1, this);
3577 if (in1->outcnt() == 0) {
3578 in1->disconnect_inputs(this);
3579 }
3580 } else {
3581 n->subsume_by(n->in(1), this);
3582 if (n->outcnt() == 0) {
3583 n->disconnect_inputs(this);
3584 }
3585 }
3586 break;
3587 }
3588 case Op_CastII: {
3589 n->as_CastII()->remove_range_check_cast(this);
3590 break;
3591 }
3592 #ifdef _LP64
3593 case Op_CmpP:
3594 // Do this transformation here to preserve CmpPNode::sub() and
3595 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3596 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3597 Node* in1 = n->in(1);
3598 Node* in2 = n->in(2);
3599 if (!in1->is_DecodeNarrowPtr()) {
3600 in2 = in1;
3601 in1 = n->in(2);
3602 }
3603 assert(in1->is_DecodeNarrowPtr(), "sanity");
3604
3605 Node* new_in2 = nullptr;
3606 if (in2->is_DecodeNarrowPtr()) {
3607 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3608 new_in2 = in2->in(1);
3609 } else if (in2->Opcode() == Op_ConP) {
3610 const Type* t = in2->bottom_type();
3611 if (t == TypePtr::NULL_PTR) {
3612 assert(in1->is_DecodeN(), "compare klass to null?");
3613 // Don't convert CmpP null check into CmpN if compressed
3614 // oops implicit null check is not generated.
3615 // This will allow to generate normal oop implicit null check.
3616 if (Matcher::gen_narrow_oop_implicit_null_checks())
3617 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3618 //
3619 // This transformation together with CastPP transformation above
3620 // will generated code for implicit null checks for compressed oops.
3621 //
3622 // The original code after Optimize()
3623 //
3624 // LoadN memory, narrow_oop_reg
3625 // decode narrow_oop_reg, base_reg
3626 // CmpP base_reg, nullptr
3627 // CastPP base_reg // NotNull
3628 // Load [base_reg + offset], val_reg
3629 //
3630 // after these transformations will be
3631 //
3632 // LoadN memory, narrow_oop_reg
3633 // CmpN narrow_oop_reg, nullptr
3634 // decode_not_null narrow_oop_reg, base_reg
3635 // Load [base_reg + offset], val_reg
3636 //
3637 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
3638 // since narrow oops can be used in debug info now (see the code in
3639 // final_graph_reshaping_walk()).
3640 //
3641 // At the end the code will be matched to
3642 // on x86:
3643 //
3644 // Load_narrow_oop memory, narrow_oop_reg
3645 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3646 // NullCheck narrow_oop_reg
3647 //
3648 // and on sparc:
3649 //
3650 // Load_narrow_oop memory, narrow_oop_reg
3651 // decode_not_null narrow_oop_reg, base_reg
3652 // Load [base_reg + offset], val_reg
3653 // NullCheck base_reg
3654 //
3655 } else if (t->isa_oopptr()) {
3656 new_in2 = ConNode::make(t->make_narrowoop());
3657 } else if (t->isa_klassptr()) {
3658 ciKlass* klass = t->is_klassptr()->exact_klass();
3659 if (klass->is_in_encoding_range()) {
3660 new_in2 = ConNode::make(t->make_narrowklass());
3661 }
3662 }
3663 }
3664 if (new_in2 != nullptr) {
3665 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3666 n->subsume_by(cmpN, this);
3667 if (in1->outcnt() == 0) {
3668 in1->disconnect_inputs(this);
3669 }
3670 if (in2->outcnt() == 0) {
3671 in2->disconnect_inputs(this);
3672 }
3673 }
3674 }
3675 break;
3676
3677 case Op_DecodeN:
3678 case Op_DecodeNKlass:
3679 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3680 // DecodeN could be pinned when it can't be fold into
3681 // an address expression, see the code for Op_CastPP above.
3682 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3683 break;
3684
3685 case Op_EncodeP:
3686 case Op_EncodePKlass: {
3687 Node* in1 = n->in(1);
3688 if (in1->is_DecodeNarrowPtr()) {
3689 n->subsume_by(in1->in(1), this);
3690 } else if (in1->Opcode() == Op_ConP) {
3691 const Type* t = in1->bottom_type();
3692 if (t == TypePtr::NULL_PTR) {
3693 assert(t->isa_oopptr(), "null klass?");
3694 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3695 } else if (t->isa_oopptr()) {
3696 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3697 } else if (t->isa_klassptr()) {
3698 ciKlass* klass = t->is_klassptr()->exact_klass();
3699 if (klass->is_in_encoding_range()) {
3700 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3701 } else {
3702 assert(false, "unencodable klass in ConP -> EncodeP");
3703 C->record_failure("unencodable klass in ConP -> EncodeP");
3704 }
3705 }
3706 }
3707 if (in1->outcnt() == 0) {
3708 in1->disconnect_inputs(this);
3709 }
3710 break;
3711 }
3712
3713 case Op_Proj: {
3714 if (OptimizeStringConcat || IncrementalInline) {
3715 ProjNode* proj = n->as_Proj();
3716 if (proj->_is_io_use) {
3717 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3718 // Separate projections were used for the exception path which
3719 // are normally removed by a late inline. If it wasn't inlined
3720 // then they will hang around and should just be replaced with
3721 // the original one. Merge them.
3722 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3723 if (non_io_proj != nullptr) {
3724 proj->subsume_by(non_io_proj , this);
3725 }
3726 }
3727 }
3728 break;
3729 }
3730
3731 case Op_Phi:
3732 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3733 // The EncodeP optimization may create Phi with the same edges
3734 // for all paths. It is not handled well by Register Allocator.
3735 Node* unique_in = n->in(1);
3736 assert(unique_in != nullptr, "");
3737 uint cnt = n->req();
3738 for (uint i = 2; i < cnt; i++) {
3739 Node* m = n->in(i);
3740 assert(m != nullptr, "");
3741 if (unique_in != m)
3742 unique_in = nullptr;
3743 }
3744 if (unique_in != nullptr) {
3745 n->subsume_by(unique_in, this);
3746 }
3747 }
3748 break;
3749
3750 #endif
3751
3752 case Op_ModI:
3753 handle_div_mod_op(n, T_INT, false);
3754 break;
3755
3756 case Op_ModL:
3757 handle_div_mod_op(n, T_LONG, false);
3758 break;
3759
3760 case Op_UModI:
3761 handle_div_mod_op(n, T_INT, true);
3762 break;
3763
3764 case Op_UModL:
3765 handle_div_mod_op(n, T_LONG, true);
3766 break;
3767
3768 case Op_LoadVector:
3769 case Op_StoreVector:
3770 #ifdef ASSERT
3771 // Add VerifyVectorAlignment node between adr and load / store.
3772 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
3773 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
3774 n->as_StoreVector()->must_verify_alignment();
3775 if (must_verify_alignment) {
3776 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
3777 n->as_StoreVector()->memory_size();
3778 // The memory access should be aligned to the vector width in bytes.
3779 // However, the underlying array is possibly less well aligned, but at least
3780 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
3781 // a loop we can expect at least the following alignment:
3782 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
3783 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
3784 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
3785 // Create mask from alignment. e.g. 0b1000 -> 0b0111
3786 jlong mask = guaranteed_alignment - 1;
3787 Node* mask_con = ConLNode::make(mask);
3788 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
3789 n->set_req(MemNode::Address, va);
3790 }
3791 }
3792 #endif
3793 break;
3794
3795 case Op_LoadVectorGather:
3796 case Op_StoreVectorScatter:
3797 case Op_LoadVectorGatherMasked:
3798 case Op_StoreVectorScatterMasked:
3799 case Op_VectorCmpMasked:
3800 case Op_VectorMaskGen:
3801 case Op_LoadVectorMasked:
3802 case Op_StoreVectorMasked:
3803 break;
3804
3805 case Op_AddReductionVI:
3806 case Op_AddReductionVL:
3807 case Op_AddReductionVF:
3808 case Op_AddReductionVD:
3809 case Op_MulReductionVI:
3810 case Op_MulReductionVL:
3811 case Op_MulReductionVF:
3812 case Op_MulReductionVD:
3813 case Op_MinReductionV:
3814 case Op_MaxReductionV:
3815 case Op_UMinReductionV:
3816 case Op_UMaxReductionV:
3817 case Op_AndReductionV:
3818 case Op_OrReductionV:
3819 case Op_XorReductionV:
3820 break;
3821
3822 case Op_PackB:
3823 case Op_PackS:
3824 case Op_PackI:
3825 case Op_PackF:
3826 case Op_PackL:
3827 case Op_PackD:
3828 if (n->req()-1 > 2) {
3829 // Replace many operand PackNodes with a binary tree for matching
3830 PackNode* p = (PackNode*) n;
3831 Node* btp = p->binary_tree_pack(1, n->req());
3832 n->subsume_by(btp, this);
3833 }
3834 break;
3835 case Op_Loop:
3836 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3837 case Op_CountedLoop:
3838 case Op_LongCountedLoop:
3839 case Op_OuterStripMinedLoop:
3840 if (n->as_Loop()->is_inner_loop()) {
3841 frc.inc_inner_loop_count();
3842 }
3843 n->as_Loop()->verify_strip_mined(0);
3844 break;
3845 case Op_LShiftI:
3846 case Op_RShiftI:
3847 case Op_URShiftI:
3848 case Op_LShiftL:
3849 case Op_RShiftL:
3850 case Op_URShiftL:
3851 if (Matcher::need_masked_shift_count) {
3852 // The cpu's shift instructions don't restrict the count to the
3853 // lower 5/6 bits. We need to do the masking ourselves.
3854 Node* in2 = n->in(2);
3855 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3856 const TypeInt* t = in2->find_int_type();
3857 if (t != nullptr && t->is_con()) {
3858 juint shift = t->get_con();
3859 if (shift > mask) { // Unsigned cmp
3860 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3861 }
3862 } else {
3863 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
3864 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3865 n->set_req(2, shift);
3866 }
3867 }
3868 if (in2->outcnt() == 0) { // Remove dead node
3869 in2->disconnect_inputs(this);
3870 }
3871 }
3872 break;
3873 case Op_MemBarStoreStore:
3874 case Op_MemBarRelease:
3875 // Break the link with AllocateNode: it is no longer useful and
3876 // confuses register allocation.
3877 if (n->req() > MemBarNode::Precedent) {
3878 n->set_req(MemBarNode::Precedent, top());
3879 }
3880 break;
3881 case Op_MemBarAcquire: {
3882 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3883 // At parse time, the trailing MemBarAcquire for a volatile load
3884 // is created with an edge to the load. After optimizations,
3885 // that input may be a chain of Phis. If those phis have no
3886 // other use, then the MemBarAcquire keeps them alive and
3887 // register allocation can be confused.
3888 dead_nodes.push(n->in(MemBarNode::Precedent));
3889 n->set_req(MemBarNode::Precedent, top());
3890 }
3891 break;
3892 }
3893 case Op_Blackhole:
3894 break;
3895 case Op_RangeCheck: {
3896 RangeCheckNode* rc = n->as_RangeCheck();
3897 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3898 n->subsume_by(iff, this);
3899 frc._tests.push(iff);
3900 break;
3901 }
3902 case Op_ConvI2L: {
3903 if (!Matcher::convi2l_type_required) {
3904 // Code generation on some platforms doesn't need accurate
3905 // ConvI2L types. Widening the type can help remove redundant
3906 // address computations.
3907 n->as_Type()->set_type(TypeLong::INT);
3908 ResourceMark rm;
3909 Unique_Node_List wq;
3910 wq.push(n);
3911 for (uint next = 0; next < wq.size(); next++) {
3912 Node *m = wq.at(next);
3913
3914 for(;;) {
3915 // Loop over all nodes with identical inputs edges as m
3916 Node* k = m->find_similar(m->Opcode());
3917 if (k == nullptr) {
3918 break;
3919 }
3920 // Push their uses so we get a chance to remove node made
3921 // redundant
3922 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3923 Node* u = k->fast_out(i);
3924 if (u->Opcode() == Op_LShiftL ||
3925 u->Opcode() == Op_AddL ||
3926 u->Opcode() == Op_SubL ||
3927 u->Opcode() == Op_AddP) {
3928 wq.push(u);
3929 }
3930 }
3931 // Replace all nodes with identical edges as m with m
3932 k->subsume_by(m, this);
3933 }
3934 }
3935 }
3936 break;
3937 }
3938 case Op_CmpUL: {
3939 if (!Matcher::has_match_rule(Op_CmpUL)) {
3940 // No support for unsigned long comparisons
3941 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3942 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3943 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3944 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3945 Node* andl = new AndLNode(orl, remove_sign_mask);
3946 Node* cmp = new CmpLNode(andl, n->in(2));
3947 n->subsume_by(cmp, this);
3948 }
3949 break;
3950 }
3951 #ifdef ASSERT
3952 case Op_ConNKlass: {
3953 const TypePtr* tp = n->as_Type()->type()->make_ptr();
3954 ciKlass* klass = tp->is_klassptr()->exact_klass();
3955 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
3956 break;
3957 }
3958 #endif
3959 default:
3960 assert(!n->is_Call(), "");
3961 assert(!n->is_Mem(), "");
3962 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3963 break;
3964 }
3965 }
3966
3967 //------------------------------final_graph_reshaping_walk---------------------
3968 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3969 // requires that the walk visits a node's inputs before visiting the node.
3970 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3971 Unique_Node_List sfpt;
3972
3973 frc._visited.set(root->_idx); // first, mark node as visited
3974 uint cnt = root->req();
3975 Node *n = root;
3976 uint i = 0;
3977 while (true) {
3978 if (i < cnt) {
3979 // Place all non-visited non-null inputs onto stack
3980 Node* m = n->in(i);
3981 ++i;
3982 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
3983 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
3984 // compute worst case interpreter size in case of a deoptimization
3985 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3986
3987 sfpt.push(m);
3988 }
3989 cnt = m->req();
3990 nstack.push(n, i); // put on stack parent and next input's index
3991 n = m;
3992 i = 0;
3993 }
3994 } else {
3995 // Now do post-visit work
3996 final_graph_reshaping_impl(n, frc, dead_nodes);
3997 if (nstack.is_empty())
3998 break; // finished
3999 n = nstack.node(); // Get node from stack
4000 cnt = n->req();
4001 i = nstack.index();
4002 nstack.pop(); // Shift to the next node on stack
4003 }
4004 }
4005
4006 // Skip next transformation if compressed oops are not used.
4007 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4008 (!UseCompressedOops && !UseCompressedClassPointers))
4009 return;
4010
4011 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4012 // It could be done for an uncommon traps or any safepoints/calls
4013 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4014 while (sfpt.size() > 0) {
4015 n = sfpt.pop();
4016 JVMState *jvms = n->as_SafePoint()->jvms();
4017 assert(jvms != nullptr, "sanity");
4018 int start = jvms->debug_start();
4019 int end = n->req();
4020 bool is_uncommon = (n->is_CallStaticJava() &&
4021 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4022 for (int j = start; j < end; j++) {
4023 Node* in = n->in(j);
4024 if (in->is_DecodeNarrowPtr()) {
4025 bool safe_to_skip = true;
4026 if (!is_uncommon ) {
4027 // Is it safe to skip?
4028 for (uint i = 0; i < in->outcnt(); i++) {
4029 Node* u = in->raw_out(i);
4030 if (!u->is_SafePoint() ||
4031 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4032 safe_to_skip = false;
4033 }
4034 }
4035 }
4036 if (safe_to_skip) {
4037 n->set_req(j, in->in(1));
4038 }
4039 if (in->outcnt() == 0) {
4040 in->disconnect_inputs(this);
4041 }
4042 }
4043 }
4044 }
4045 }
4046
4047 //------------------------------final_graph_reshaping--------------------------
4048 // Final Graph Reshaping.
4049 //
4050 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4051 // and not commoned up and forced early. Must come after regular
4052 // optimizations to avoid GVN undoing the cloning. Clone constant
4053 // inputs to Loop Phis; these will be split by the allocator anyways.
4054 // Remove Opaque nodes.
4055 // (2) Move last-uses by commutative operations to the left input to encourage
4056 // Intel update-in-place two-address operations and better register usage
4057 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4058 // calls canonicalizing them back.
4059 // (3) Count the number of double-precision FP ops, single-precision FP ops
4060 // and call sites. On Intel, we can get correct rounding either by
4061 // forcing singles to memory (requires extra stores and loads after each
4062 // FP bytecode) or we can set a rounding mode bit (requires setting and
4063 // clearing the mode bit around call sites). The mode bit is only used
4064 // if the relative frequency of single FP ops to calls is low enough.
4065 // This is a key transform for SPEC mpeg_audio.
4066 // (4) Detect infinite loops; blobs of code reachable from above but not
4067 // below. Several of the Code_Gen algorithms fail on such code shapes,
4068 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4069 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4070 // Detection is by looking for IfNodes where only 1 projection is
4071 // reachable from below or CatchNodes missing some targets.
4072 // (5) Assert for insane oop offsets in debug mode.
4073
4074 bool Compile::final_graph_reshaping() {
4075 // an infinite loop may have been eliminated by the optimizer,
4076 // in which case the graph will be empty.
4077 if (root()->req() == 1) {
4078 // Do not compile method that is only a trivial infinite loop,
4079 // since the content of the loop may have been eliminated.
4080 record_method_not_compilable("trivial infinite loop");
4081 return true;
4082 }
4083
4084 // Expensive nodes have their control input set to prevent the GVN
4085 // from freely commoning them. There's no GVN beyond this point so
4086 // no need to keep the control input. We want the expensive nodes to
4087 // be freely moved to the least frequent code path by gcm.
4088 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4089 for (int i = 0; i < expensive_count(); i++) {
4090 _expensive_nodes.at(i)->set_req(0, nullptr);
4091 }
4092
4093 Final_Reshape_Counts frc;
4094
4095 // Visit everybody reachable!
4096 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4097 Node_Stack nstack(live_nodes() >> 1);
4098 Unique_Node_List dead_nodes;
4099 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4100
4101 // Check for unreachable (from below) code (i.e., infinite loops).
4102 for( uint i = 0; i < frc._tests.size(); i++ ) {
4103 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4104 // Get number of CFG targets.
4105 // Note that PCTables include exception targets after calls.
4106 uint required_outcnt = n->required_outcnt();
4107 if (n->outcnt() != required_outcnt) {
4108 // Check for a few special cases. Rethrow Nodes never take the
4109 // 'fall-thru' path, so expected kids is 1 less.
4110 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4111 if (n->in(0)->in(0)->is_Call()) {
4112 CallNode* call = n->in(0)->in(0)->as_Call();
4113 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4114 required_outcnt--; // Rethrow always has 1 less kid
4115 } else if (call->req() > TypeFunc::Parms &&
4116 call->is_CallDynamicJava()) {
4117 // Check for null receiver. In such case, the optimizer has
4118 // detected that the virtual call will always result in a null
4119 // pointer exception. The fall-through projection of this CatchNode
4120 // will not be populated.
4121 Node* arg0 = call->in(TypeFunc::Parms);
4122 if (arg0->is_Type() &&
4123 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4124 required_outcnt--;
4125 }
4126 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4127 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4128 // Check for illegal array length. In such case, the optimizer has
4129 // detected that the allocation attempt will always result in an
4130 // exception. There is no fall-through projection of this CatchNode .
4131 assert(call->is_CallStaticJava(), "static call expected");
4132 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4133 uint valid_length_test_input = call->req() - 1;
4134 Node* valid_length_test = call->in(valid_length_test_input);
4135 call->del_req(valid_length_test_input);
4136 if (valid_length_test->find_int_con(1) == 0) {
4137 required_outcnt--;
4138 }
4139 dead_nodes.push(valid_length_test);
4140 assert(n->outcnt() == required_outcnt, "malformed control flow");
4141 continue;
4142 }
4143 }
4144 }
4145
4146 // Recheck with a better notion of 'required_outcnt'
4147 if (n->outcnt() != required_outcnt) {
4148 record_method_not_compilable("malformed control flow");
4149 return true; // Not all targets reachable!
4150 }
4151 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4152 CallNode* call = n->in(0)->in(0)->as_Call();
4153 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4154 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4155 assert(call->is_CallStaticJava(), "static call expected");
4156 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4157 uint valid_length_test_input = call->req() - 1;
4158 dead_nodes.push(call->in(valid_length_test_input));
4159 call->del_req(valid_length_test_input); // valid length test useless now
4160 }
4161 }
4162 // Check that I actually visited all kids. Unreached kids
4163 // must be infinite loops.
4164 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4165 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4166 record_method_not_compilable("infinite loop");
4167 return true; // Found unvisited kid; must be unreach
4168 }
4169
4170 // Here so verification code in final_graph_reshaping_walk()
4171 // always see an OuterStripMinedLoopEnd
4172 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4173 IfNode* init_iff = n->as_If();
4174 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4175 n->subsume_by(iff, this);
4176 }
4177 }
4178
4179 while (dead_nodes.size() > 0) {
4180 Node* m = dead_nodes.pop();
4181 if (m->outcnt() == 0 && m != top()) {
4182 for (uint j = 0; j < m->req(); j++) {
4183 Node* in = m->in(j);
4184 if (in != nullptr) {
4185 dead_nodes.push(in);
4186 }
4187 }
4188 m->disconnect_inputs(this);
4189 }
4190 }
4191
4192 set_java_calls(frc.get_java_call_count());
4193 set_inner_loops(frc.get_inner_loop_count());
4194
4195 // No infinite loops, no reason to bail out.
4196 return false;
4197 }
4198
4199 //-----------------------------too_many_traps----------------------------------
4200 // Report if there are too many traps at the current method and bci.
4201 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4202 bool Compile::too_many_traps(ciMethod* method,
4203 int bci,
4204 Deoptimization::DeoptReason reason) {
4205 ciMethodData* md = method->method_data();
4206 if (md->is_empty()) {
4207 // Assume the trap has not occurred, or that it occurred only
4208 // because of a transient condition during start-up in the interpreter.
4209 return false;
4210 }
4211 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4212 if (md->has_trap_at(bci, m, reason) != 0) {
4213 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4214 // Also, if there are multiple reasons, or if there is no per-BCI record,
4215 // assume the worst.
4216 if (log())
4217 log()->elem("observe trap='%s' count='%d'",
4218 Deoptimization::trap_reason_name(reason),
4219 md->trap_count(reason));
4220 return true;
4221 } else {
4222 // Ignore method/bci and see if there have been too many globally.
4223 return too_many_traps(reason, md);
4224 }
4225 }
4226
4227 // Less-accurate variant which does not require a method and bci.
4228 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4229 ciMethodData* logmd) {
4230 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4231 // Too many traps globally.
4232 // Note that we use cumulative trap_count, not just md->trap_count.
4233 if (log()) {
4234 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4235 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4236 Deoptimization::trap_reason_name(reason),
4237 mcount, trap_count(reason));
4238 }
4239 return true;
4240 } else {
4241 // The coast is clear.
4242 return false;
4243 }
4244 }
4245
4246 //--------------------------too_many_recompiles--------------------------------
4247 // Report if there are too many recompiles at the current method and bci.
4248 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4249 // Is not eager to return true, since this will cause the compiler to use
4250 // Action_none for a trap point, to avoid too many recompilations.
4251 bool Compile::too_many_recompiles(ciMethod* method,
4252 int bci,
4253 Deoptimization::DeoptReason reason) {
4254 ciMethodData* md = method->method_data();
4255 if (md->is_empty()) {
4256 // Assume the trap has not occurred, or that it occurred only
4257 // because of a transient condition during start-up in the interpreter.
4258 return false;
4259 }
4260 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4261 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4262 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4263 Deoptimization::DeoptReason per_bc_reason
4264 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4265 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4266 if ((per_bc_reason == Deoptimization::Reason_none
4267 || md->has_trap_at(bci, m, reason) != 0)
4268 // The trap frequency measure we care about is the recompile count:
4269 && md->trap_recompiled_at(bci, m)
4270 && md->overflow_recompile_count() >= bc_cutoff) {
4271 // Do not emit a trap here if it has already caused recompilations.
4272 // Also, if there are multiple reasons, or if there is no per-BCI record,
4273 // assume the worst.
4274 if (log())
4275 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4276 Deoptimization::trap_reason_name(reason),
4277 md->trap_count(reason),
4278 md->overflow_recompile_count());
4279 return true;
4280 } else if (trap_count(reason) != 0
4281 && decompile_count() >= m_cutoff) {
4282 // Too many recompiles globally, and we have seen this sort of trap.
4283 // Use cumulative decompile_count, not just md->decompile_count.
4284 if (log())
4285 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4286 Deoptimization::trap_reason_name(reason),
4287 md->trap_count(reason), trap_count(reason),
4288 md->decompile_count(), decompile_count());
4289 return true;
4290 } else {
4291 // The coast is clear.
4292 return false;
4293 }
4294 }
4295
4296 // Compute when not to trap. Used by matching trap based nodes and
4297 // NullCheck optimization.
4298 void Compile::set_allowed_deopt_reasons() {
4299 _allowed_reasons = 0;
4300 if (is_method_compilation()) {
4301 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4302 assert(rs < BitsPerInt, "recode bit map");
4303 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4304 _allowed_reasons |= nth_bit(rs);
4305 }
4306 }
4307 }
4308 }
4309
4310 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4311 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4312 }
4313
4314 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4315 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4316 }
4317
4318 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4319 if (holder->is_initialized()) {
4320 return false;
4321 }
4322 if (holder->is_being_initialized()) {
4323 if (accessing_method->holder() == holder) {
4324 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4325 // <init>, or a static method. In all those cases, there was an initialization
4326 // barrier on the holder klass passed.
4327 if (accessing_method->is_static_initializer() ||
4328 accessing_method->is_object_initializer() ||
4329 accessing_method->is_static()) {
4330 return false;
4331 }
4332 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4333 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4334 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4335 // child class can become fully initialized while its parent class is still being initialized.
4336 if (accessing_method->is_static_initializer()) {
4337 return false;
4338 }
4339 }
4340 ciMethod* root = method(); // the root method of compilation
4341 if (root != accessing_method) {
4342 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4343 }
4344 }
4345 return true;
4346 }
4347
4348 #ifndef PRODUCT
4349 //------------------------------verify_bidirectional_edges---------------------
4350 // For each input edge to a node (ie - for each Use-Def edge), verify that
4351 // there is a corresponding Def-Use edge.
4352 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4353 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4354 uint stack_size = live_nodes() >> 4;
4355 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4356 if (root_and_safepoints != nullptr) {
4357 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4358 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4359 Node* root_or_safepoint = root_and_safepoints->at(i);
4360 // If the node is a safepoint, let's check if it still has a control input
4361 // Lack of control input signifies that this node was killed by CCP or
4362 // recursively by remove_globally_dead_node and it shouldn't be a starting
4363 // point.
4364 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4365 nstack.push(root_or_safepoint);
4366 }
4367 }
4368 } else {
4369 nstack.push(_root);
4370 }
4371
4372 while (nstack.size() > 0) {
4373 Node* n = nstack.pop();
4374 if (visited.member(n)) {
4375 continue;
4376 }
4377 visited.push(n);
4378
4379 // Walk over all input edges, checking for correspondence
4380 uint length = n->len();
4381 for (uint i = 0; i < length; i++) {
4382 Node* in = n->in(i);
4383 if (in != nullptr && !visited.member(in)) {
4384 nstack.push(in); // Put it on stack
4385 }
4386 if (in != nullptr && !in->is_top()) {
4387 // Count instances of `next`
4388 int cnt = 0;
4389 for (uint idx = 0; idx < in->_outcnt; idx++) {
4390 if (in->_out[idx] == n) {
4391 cnt++;
4392 }
4393 }
4394 assert(cnt > 0, "Failed to find Def-Use edge.");
4395 // Check for duplicate edges
4396 // walk the input array downcounting the input edges to n
4397 for (uint j = 0; j < length; j++) {
4398 if (n->in(j) == in) {
4399 cnt--;
4400 }
4401 }
4402 assert(cnt == 0, "Mismatched edge count.");
4403 } else if (in == nullptr) {
4404 assert(i == 0 || i >= n->req() ||
4405 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4406 (n->is_Unlock() && i == (n->req() - 1)) ||
4407 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4408 "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4409 } else {
4410 assert(in->is_top(), "sanity");
4411 // Nothing to check.
4412 }
4413 }
4414 }
4415 }
4416
4417 //------------------------------verify_graph_edges---------------------------
4418 // Walk the Graph and verify that there is a one-to-one correspondence
4419 // between Use-Def edges and Def-Use edges in the graph.
4420 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4421 if (VerifyGraphEdges) {
4422 Unique_Node_List visited;
4423
4424 // Call graph walk to check edges
4425 verify_bidirectional_edges(visited, root_and_safepoints);
4426 if (no_dead_code) {
4427 // Now make sure that no visited node is used by an unvisited node.
4428 bool dead_nodes = false;
4429 Unique_Node_List checked;
4430 while (visited.size() > 0) {
4431 Node* n = visited.pop();
4432 checked.push(n);
4433 for (uint i = 0; i < n->outcnt(); i++) {
4434 Node* use = n->raw_out(i);
4435 if (checked.member(use)) continue; // already checked
4436 if (visited.member(use)) continue; // already in the graph
4437 if (use->is_Con()) continue; // a dead ConNode is OK
4438 // At this point, we have found a dead node which is DU-reachable.
4439 if (!dead_nodes) {
4440 tty->print_cr("*** Dead nodes reachable via DU edges:");
4441 dead_nodes = true;
4442 }
4443 use->dump(2);
4444 tty->print_cr("---");
4445 checked.push(use); // No repeats; pretend it is now checked.
4446 }
4447 }
4448 assert(!dead_nodes, "using nodes must be reachable from root");
4449 }
4450 }
4451 }
4452 #endif
4453
4454 // The Compile object keeps track of failure reasons separately from the ciEnv.
4455 // This is required because there is not quite a 1-1 relation between the
4456 // ciEnv and its compilation task and the Compile object. Note that one
4457 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4458 // to backtrack and retry without subsuming loads. Other than this backtracking
4459 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4460 // by the logic in C2Compiler.
4461 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4462 if (log() != nullptr) {
4463 log()->elem("failure reason='%s' phase='compile'", reason);
4464 }
4465 if (_failure_reason.get() == nullptr) {
4466 // Record the first failure reason.
4467 _failure_reason.set(reason);
4468 if (CaptureBailoutInformation) {
4469 _first_failure_details = new CompilationFailureInfo(reason);
4470 }
4471 } else {
4472 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4473 }
4474
4475 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4476 C->print_method(PHASE_FAILURE, 1);
4477 }
4478 _root = nullptr; // flush the graph, too
4479 }
4480
4481 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4482 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4483 _compile(Compile::current()),
4484 _log(nullptr),
4485 _dolog(CITimeVerbose)
4486 {
4487 assert(_compile != nullptr, "sanity check");
4488 assert(id != PhaseTraceId::_t_none, "Don't use none");
4489 if (_dolog) {
4490 _log = _compile->log();
4491 }
4492 if (_log != nullptr) {
4493 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4494 _log->stamp();
4495 _log->end_head();
4496 }
4497
4498 // Inform memory statistic, if enabled
4499 if (CompilationMemoryStatistic::enabled()) {
4500 CompilationMemoryStatistic::on_phase_start((int)id, name);
4501 }
4502 }
4503
4504 Compile::TracePhase::TracePhase(PhaseTraceId id)
4505 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4506
4507 Compile::TracePhase::~TracePhase() {
4508
4509 // Inform memory statistic, if enabled
4510 if (CompilationMemoryStatistic::enabled()) {
4511 CompilationMemoryStatistic::on_phase_end();
4512 }
4513
4514 if (_compile->failing_internal()) {
4515 if (_log != nullptr) {
4516 _log->done("phase");
4517 }
4518 return; // timing code, not stressing bailouts.
4519 }
4520 #ifdef ASSERT
4521 if (PrintIdealNodeCount) {
4522 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4523 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4524 }
4525
4526 if (VerifyIdealNodeCount) {
4527 _compile->print_missing_nodes();
4528 }
4529 #endif
4530
4531 if (_log != nullptr) {
4532 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4533 }
4534 }
4535
4536 //----------------------------static_subtype_check-----------------------------
4537 // Shortcut important common cases when superklass is exact:
4538 // (0) superklass is java.lang.Object (can occur in reflective code)
4539 // (1) subklass is already limited to a subtype of superklass => always ok
4540 // (2) subklass does not overlap with superklass => always fail
4541 // (3) superklass has NO subtypes and we can check with a simple compare.
4542 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4543 if (skip) {
4544 return SSC_full_test; // Let caller generate the general case.
4545 }
4546
4547 if (subk->is_java_subtype_of(superk)) {
4548 return SSC_always_true; // (0) and (1) this test cannot fail
4549 }
4550
4551 if (!subk->maybe_java_subtype_of(superk)) {
4552 return SSC_always_false; // (2) true path dead; no dynamic test needed
4553 }
4554
4555 const Type* superelem = superk;
4556 if (superk->isa_aryklassptr()) {
4557 int ignored;
4558 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4559 }
4560
4561 if (superelem->isa_instklassptr()) {
4562 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4563 if (!ik->has_subklass()) {
4564 if (!ik->is_final()) {
4565 // Add a dependency if there is a chance of a later subclass.
4566 dependencies()->assert_leaf_type(ik);
4567 }
4568 if (!superk->maybe_java_subtype_of(subk)) {
4569 return SSC_always_false;
4570 }
4571 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4572 }
4573 } else {
4574 // A primitive array type has no subtypes.
4575 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4576 }
4577
4578 return SSC_full_test;
4579 }
4580
4581 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4582 #ifdef _LP64
4583 // The scaled index operand to AddP must be a clean 64-bit value.
4584 // Java allows a 32-bit int to be incremented to a negative
4585 // value, which appears in a 64-bit register as a large
4586 // positive number. Using that large positive number as an
4587 // operand in pointer arithmetic has bad consequences.
4588 // On the other hand, 32-bit overflow is rare, and the possibility
4589 // can often be excluded, if we annotate the ConvI2L node with
4590 // a type assertion that its value is known to be a small positive
4591 // number. (The prior range check has ensured this.)
4592 // This assertion is used by ConvI2LNode::Ideal.
4593 int index_max = max_jint - 1; // array size is max_jint, index is one less
4594 if (sizetype != nullptr && sizetype->_hi > 0) {
4595 index_max = sizetype->_hi - 1;
4596 }
4597 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4598 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4599 #endif
4600 return idx;
4601 }
4602
4603 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4604 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4605 if (ctrl != nullptr) {
4606 // Express control dependency by a CastII node with a narrow type.
4607 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4608 // node from floating above the range check during loop optimizations. Otherwise, the
4609 // ConvI2L node may be eliminated independently of the range check, causing the data path
4610 // to become TOP while the control path is still there (although it's unreachable).
4611 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::DependencyType::NonFloatingNarrowing : ConstraintCastNode::DependencyType::FloatingNarrowing, true /* range check dependency */);
4612 value = phase->transform(value);
4613 }
4614 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4615 return phase->transform(new ConvI2LNode(value, ltype));
4616 }
4617
4618 void Compile::dump_print_inlining() {
4619 inline_printer()->print_on(tty);
4620 }
4621
4622 void Compile::log_late_inline(CallGenerator* cg) {
4623 if (log() != nullptr) {
4624 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4625 cg->unique_id());
4626 JVMState* p = cg->call_node()->jvms();
4627 while (p != nullptr) {
4628 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4629 p = p->caller();
4630 }
4631 log()->tail("late_inline");
4632 }
4633 }
4634
4635 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4636 log_late_inline(cg);
4637 if (log() != nullptr) {
4638 log()->inline_fail(msg);
4639 }
4640 }
4641
4642 void Compile::log_inline_id(CallGenerator* cg) {
4643 if (log() != nullptr) {
4644 // The LogCompilation tool needs a unique way to identify late
4645 // inline call sites. This id must be unique for this call site in
4646 // this compilation. Try to have it unique across compilations as
4647 // well because it can be convenient when grepping through the log
4648 // file.
4649 // Distinguish OSR compilations from others in case CICountOSR is
4650 // on.
4651 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4652 cg->set_unique_id(id);
4653 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4654 }
4655 }
4656
4657 void Compile::log_inline_failure(const char* msg) {
4658 if (C->log() != nullptr) {
4659 C->log()->inline_fail(msg);
4660 }
4661 }
4662
4663
4664 // Dump inlining replay data to the stream.
4665 // Don't change thread state and acquire any locks.
4666 void Compile::dump_inline_data(outputStream* out) {
4667 InlineTree* inl_tree = ilt();
4668 if (inl_tree != nullptr) {
4669 out->print(" inline %d", inl_tree->count());
4670 inl_tree->dump_replay_data(out);
4671 }
4672 }
4673
4674 void Compile::dump_inline_data_reduced(outputStream* out) {
4675 assert(ReplayReduce, "");
4676
4677 InlineTree* inl_tree = ilt();
4678 if (inl_tree == nullptr) {
4679 return;
4680 }
4681 // Enable iterative replay file reduction
4682 // Output "compile" lines for depth 1 subtrees,
4683 // simulating that those trees were compiled
4684 // instead of inlined.
4685 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4686 InlineTree* sub = inl_tree->subtrees().at(i);
4687 if (sub->inline_level() != 1) {
4688 continue;
4689 }
4690
4691 ciMethod* method = sub->method();
4692 int entry_bci = -1;
4693 int comp_level = env()->task()->comp_level();
4694 out->print("compile ");
4695 method->dump_name_as_ascii(out);
4696 out->print(" %d %d", entry_bci, comp_level);
4697 out->print(" inline %d", sub->count());
4698 sub->dump_replay_data(out, -1);
4699 out->cr();
4700 }
4701 }
4702
4703 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4704 if (n1->Opcode() < n2->Opcode()) return -1;
4705 else if (n1->Opcode() > n2->Opcode()) return 1;
4706
4707 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4708 for (uint i = 1; i < n1->req(); i++) {
4709 if (n1->in(i) < n2->in(i)) return -1;
4710 else if (n1->in(i) > n2->in(i)) return 1;
4711 }
4712
4713 return 0;
4714 }
4715
4716 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4717 Node* n1 = *n1p;
4718 Node* n2 = *n2p;
4719
4720 return cmp_expensive_nodes(n1, n2);
4721 }
4722
4723 void Compile::sort_expensive_nodes() {
4724 if (!expensive_nodes_sorted()) {
4725 _expensive_nodes.sort(cmp_expensive_nodes);
4726 }
4727 }
4728
4729 bool Compile::expensive_nodes_sorted() const {
4730 for (int i = 1; i < _expensive_nodes.length(); i++) {
4731 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4732 return false;
4733 }
4734 }
4735 return true;
4736 }
4737
4738 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4739 if (_expensive_nodes.length() == 0) {
4740 return false;
4741 }
4742
4743 assert(OptimizeExpensiveOps, "optimization off?");
4744
4745 // Take this opportunity to remove dead nodes from the list
4746 int j = 0;
4747 for (int i = 0; i < _expensive_nodes.length(); i++) {
4748 Node* n = _expensive_nodes.at(i);
4749 if (!n->is_unreachable(igvn)) {
4750 assert(n->is_expensive(), "should be expensive");
4751 _expensive_nodes.at_put(j, n);
4752 j++;
4753 }
4754 }
4755 _expensive_nodes.trunc_to(j);
4756
4757 // Then sort the list so that similar nodes are next to each other
4758 // and check for at least two nodes of identical kind with same data
4759 // inputs.
4760 sort_expensive_nodes();
4761
4762 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4763 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4764 return true;
4765 }
4766 }
4767
4768 return false;
4769 }
4770
4771 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4772 if (_expensive_nodes.length() == 0) {
4773 return;
4774 }
4775
4776 assert(OptimizeExpensiveOps, "optimization off?");
4777
4778 // Sort to bring similar nodes next to each other and clear the
4779 // control input of nodes for which there's only a single copy.
4780 sort_expensive_nodes();
4781
4782 int j = 0;
4783 int identical = 0;
4784 int i = 0;
4785 bool modified = false;
4786 for (; i < _expensive_nodes.length()-1; i++) {
4787 assert(j <= i, "can't write beyond current index");
4788 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4789 identical++;
4790 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4791 continue;
4792 }
4793 if (identical > 0) {
4794 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4795 identical = 0;
4796 } else {
4797 Node* n = _expensive_nodes.at(i);
4798 igvn.replace_input_of(n, 0, nullptr);
4799 igvn.hash_insert(n);
4800 modified = true;
4801 }
4802 }
4803 if (identical > 0) {
4804 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4805 } else if (_expensive_nodes.length() >= 1) {
4806 Node* n = _expensive_nodes.at(i);
4807 igvn.replace_input_of(n, 0, nullptr);
4808 igvn.hash_insert(n);
4809 modified = true;
4810 }
4811 _expensive_nodes.trunc_to(j);
4812 if (modified) {
4813 igvn.optimize();
4814 }
4815 }
4816
4817 void Compile::add_expensive_node(Node * n) {
4818 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4819 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4820 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4821 if (OptimizeExpensiveOps) {
4822 _expensive_nodes.append(n);
4823 } else {
4824 // Clear control input and let IGVN optimize expensive nodes if
4825 // OptimizeExpensiveOps is off.
4826 n->set_req(0, nullptr);
4827 }
4828 }
4829
4830 /**
4831 * Track coarsened Lock and Unlock nodes.
4832 */
4833
4834 class Lock_List : public Node_List {
4835 uint _origin_cnt;
4836 public:
4837 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4838 uint origin_cnt() const { return _origin_cnt; }
4839 };
4840
4841 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4842 int length = locks.length();
4843 if (length > 0) {
4844 // Have to keep this list until locks elimination during Macro nodes elimination.
4845 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4846 AbstractLockNode* alock = locks.at(0);
4847 BoxLockNode* box = alock->box_node()->as_BoxLock();
4848 for (int i = 0; i < length; i++) {
4849 AbstractLockNode* lock = locks.at(i);
4850 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4851 locks_list->push(lock);
4852 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
4853 if (this_box != box) {
4854 // Locking regions (BoxLock) could be Unbalanced here:
4855 // - its coarsened locks were eliminated in earlier
4856 // macro nodes elimination followed by loop unroll
4857 // - it is OSR locking region (no Lock node)
4858 // Preserve Unbalanced status in such cases.
4859 if (!this_box->is_unbalanced()) {
4860 this_box->set_coarsened();
4861 }
4862 if (!box->is_unbalanced()) {
4863 box->set_coarsened();
4864 }
4865 }
4866 }
4867 _coarsened_locks.append(locks_list);
4868 }
4869 }
4870
4871 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4872 int count = coarsened_count();
4873 for (int i = 0; i < count; i++) {
4874 Node_List* locks_list = _coarsened_locks.at(i);
4875 for (uint j = 0; j < locks_list->size(); j++) {
4876 Node* lock = locks_list->at(j);
4877 assert(lock->is_AbstractLock(), "sanity");
4878 if (!useful.member(lock)) {
4879 locks_list->yank(lock);
4880 }
4881 }
4882 }
4883 }
4884
4885 void Compile::remove_coarsened_lock(Node* n) {
4886 if (n->is_AbstractLock()) {
4887 int count = coarsened_count();
4888 for (int i = 0; i < count; i++) {
4889 Node_List* locks_list = _coarsened_locks.at(i);
4890 locks_list->yank(n);
4891 }
4892 }
4893 }
4894
4895 bool Compile::coarsened_locks_consistent() {
4896 int count = coarsened_count();
4897 for (int i = 0; i < count; i++) {
4898 bool unbalanced = false;
4899 bool modified = false; // track locks kind modifications
4900 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4901 uint size = locks_list->size();
4902 if (size == 0) {
4903 unbalanced = false; // All locks were eliminated - good
4904 } else if (size != locks_list->origin_cnt()) {
4905 unbalanced = true; // Some locks were removed from list
4906 } else {
4907 for (uint j = 0; j < size; j++) {
4908 Node* lock = locks_list->at(j);
4909 // All nodes in group should have the same state (modified or not)
4910 if (!lock->as_AbstractLock()->is_coarsened()) {
4911 if (j == 0) {
4912 // first on list was modified, the rest should be too for consistency
4913 modified = true;
4914 } else if (!modified) {
4915 // this lock was modified but previous locks on the list were not
4916 unbalanced = true;
4917 break;
4918 }
4919 } else if (modified) {
4920 // previous locks on list were modified but not this lock
4921 unbalanced = true;
4922 break;
4923 }
4924 }
4925 }
4926 if (unbalanced) {
4927 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4928 #ifdef ASSERT
4929 if (PrintEliminateLocks) {
4930 tty->print_cr("=== unbalanced coarsened locks ===");
4931 for (uint l = 0; l < size; l++) {
4932 locks_list->at(l)->dump();
4933 }
4934 }
4935 #endif
4936 record_failure(C2Compiler::retry_no_locks_coarsening());
4937 return false;
4938 }
4939 }
4940 return true;
4941 }
4942
4943 // Mark locking regions (identified by BoxLockNode) as unbalanced if
4944 // locks coarsening optimization removed Lock/Unlock nodes from them.
4945 // Such regions become unbalanced because coarsening only removes part
4946 // of Lock/Unlock nodes in region. As result we can't execute other
4947 // locks elimination optimizations which assume all code paths have
4948 // corresponding pair of Lock/Unlock nodes - they are balanced.
4949 void Compile::mark_unbalanced_boxes() const {
4950 int count = coarsened_count();
4951 for (int i = 0; i < count; i++) {
4952 Node_List* locks_list = _coarsened_locks.at(i);
4953 uint size = locks_list->size();
4954 if (size > 0) {
4955 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
4956 BoxLockNode* box = alock->box_node()->as_BoxLock();
4957 if (alock->is_coarsened()) {
4958 // coarsened_locks_consistent(), which is called before this method, verifies
4959 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
4960 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4961 for (uint j = 1; j < size; j++) {
4962 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
4963 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
4964 if (box != this_box) {
4965 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
4966 box->set_unbalanced();
4967 this_box->set_unbalanced();
4968 }
4969 }
4970 }
4971 }
4972 }
4973 }
4974
4975 /**
4976 * Remove the speculative part of types and clean up the graph
4977 */
4978 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4979 if (UseTypeSpeculation) {
4980 Unique_Node_List worklist;
4981 worklist.push(root());
4982 int modified = 0;
4983 // Go over all type nodes that carry a speculative type, drop the
4984 // speculative part of the type and enqueue the node for an igvn
4985 // which may optimize it out.
4986 for (uint next = 0; next < worklist.size(); ++next) {
4987 Node *n = worklist.at(next);
4988 if (n->is_Type()) {
4989 TypeNode* tn = n->as_Type();
4990 const Type* t = tn->type();
4991 const Type* t_no_spec = t->remove_speculative();
4992 if (t_no_spec != t) {
4993 bool in_hash = igvn.hash_delete(n);
4994 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4995 tn->set_type(t_no_spec);
4996 igvn.hash_insert(n);
4997 igvn._worklist.push(n); // give it a chance to go away
4998 modified++;
4999 }
5000 }
5001 // Iterate over outs - endless loops is unreachable from below
5002 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5003 Node *m = n->fast_out(i);
5004 if (not_a_node(m)) {
5005 continue;
5006 }
5007 worklist.push(m);
5008 }
5009 }
5010 // Drop the speculative part of all types in the igvn's type table
5011 igvn.remove_speculative_types();
5012 if (modified > 0) {
5013 igvn.optimize();
5014 if (failing()) return;
5015 }
5016 #ifdef ASSERT
5017 // Verify that after the IGVN is over no speculative type has resurfaced
5018 worklist.clear();
5019 worklist.push(root());
5020 for (uint next = 0; next < worklist.size(); ++next) {
5021 Node *n = worklist.at(next);
5022 const Type* t = igvn.type_or_null(n);
5023 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5024 if (n->is_Type()) {
5025 t = n->as_Type()->type();
5026 assert(t == t->remove_speculative(), "no more speculative types");
5027 }
5028 // Iterate over outs - endless loops is unreachable from below
5029 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5030 Node *m = n->fast_out(i);
5031 if (not_a_node(m)) {
5032 continue;
5033 }
5034 worklist.push(m);
5035 }
5036 }
5037 igvn.check_no_speculative_types();
5038 #endif
5039 }
5040 }
5041
5042 // Auxiliary methods to support randomized stressing/fuzzing.
5043
5044 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5045 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5046 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5047 FLAG_SET_ERGO(StressSeed, _stress_seed);
5048 } else {
5049 _stress_seed = StressSeed;
5050 }
5051 if (_log != nullptr) {
5052 _log->elem("stress_test seed='%u'", _stress_seed);
5053 }
5054 }
5055
5056 int Compile::random() {
5057 _stress_seed = os::next_random(_stress_seed);
5058 return static_cast<int>(_stress_seed);
5059 }
5060
5061 // This method can be called the arbitrary number of times, with current count
5062 // as the argument. The logic allows selecting a single candidate from the
5063 // running list of candidates as follows:
5064 // int count = 0;
5065 // Cand* selected = null;
5066 // while(cand = cand->next()) {
5067 // if (randomized_select(++count)) {
5068 // selected = cand;
5069 // }
5070 // }
5071 //
5072 // Including count equalizes the chances any candidate is "selected".
5073 // This is useful when we don't have the complete list of candidates to choose
5074 // from uniformly. In this case, we need to adjust the randomicity of the
5075 // selection, or else we will end up biasing the selection towards the latter
5076 // candidates.
5077 //
5078 // Quick back-envelope calculation shows that for the list of n candidates
5079 // the equal probability for the candidate to persist as "best" can be
5080 // achieved by replacing it with "next" k-th candidate with the probability
5081 // of 1/k. It can be easily shown that by the end of the run, the
5082 // probability for any candidate is converged to 1/n, thus giving the
5083 // uniform distribution among all the candidates.
5084 //
5085 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5086 #define RANDOMIZED_DOMAIN_POW 29
5087 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5088 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5089 bool Compile::randomized_select(int count) {
5090 assert(count > 0, "only positive");
5091 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5092 }
5093
5094 #ifdef ASSERT
5095 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5096 bool Compile::fail_randomly() {
5097 if ((random() % StressBailoutMean) != 0) {
5098 return false;
5099 }
5100 record_failure("StressBailout");
5101 return true;
5102 }
5103
5104 bool Compile::failure_is_artificial() {
5105 return C->failure_reason_is("StressBailout");
5106 }
5107 #endif
5108
5109 CloneMap& Compile::clone_map() { return _clone_map; }
5110 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5111
5112 void NodeCloneInfo::dump_on(outputStream* st) const {
5113 st->print(" {%d:%d} ", idx(), gen());
5114 }
5115
5116 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5117 uint64_t val = value(old->_idx);
5118 NodeCloneInfo cio(val);
5119 assert(val != 0, "old node should be in the map");
5120 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5121 insert(nnn->_idx, cin.get());
5122 #ifndef PRODUCT
5123 if (is_debug()) {
5124 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5125 }
5126 #endif
5127 }
5128
5129 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5130 NodeCloneInfo cio(value(old->_idx));
5131 if (cio.get() == 0) {
5132 cio.set(old->_idx, 0);
5133 insert(old->_idx, cio.get());
5134 #ifndef PRODUCT
5135 if (is_debug()) {
5136 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5137 }
5138 #endif
5139 }
5140 clone(old, nnn, gen);
5141 }
5142
5143 int CloneMap::max_gen() const {
5144 int g = 0;
5145 DictI di(_dict);
5146 for(; di.test(); ++di) {
5147 int t = gen(di._key);
5148 if (g < t) {
5149 g = t;
5150 #ifndef PRODUCT
5151 if (is_debug()) {
5152 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5153 }
5154 #endif
5155 }
5156 }
5157 return g;
5158 }
5159
5160 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5161 uint64_t val = value(key);
5162 if (val != 0) {
5163 NodeCloneInfo ni(val);
5164 ni.dump_on(st);
5165 }
5166 }
5167
5168 void Compile::shuffle_macro_nodes() {
5169 shuffle_array(*C, _macro_nodes);
5170 }
5171
5172 // Move Allocate nodes to the start of the list
5173 void Compile::sort_macro_nodes() {
5174 int count = macro_count();
5175 int allocates = 0;
5176 for (int i = 0; i < count; i++) {
5177 Node* n = macro_node(i);
5178 if (n->is_Allocate()) {
5179 if (i != allocates) {
5180 Node* tmp = macro_node(allocates);
5181 _macro_nodes.at_put(allocates, n);
5182 _macro_nodes.at_put(i, tmp);
5183 }
5184 allocates++;
5185 }
5186 }
5187 }
5188
5189 void Compile::print_method(CompilerPhaseType compile_phase, int level, Node* n) {
5190 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5191 EventCompilerPhase event(UNTIMED);
5192 if (event.should_commit()) {
5193 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, compile_phase, C->_compile_id, level);
5194 }
5195 #ifndef PRODUCT
5196 ResourceMark rm;
5197 stringStream ss;
5198 ss.print_raw(CompilerPhaseTypeHelper::to_description(compile_phase));
5199 int iter = ++_igv_phase_iter[compile_phase];
5200 if (iter > 1) {
5201 ss.print(" %d", iter);
5202 }
5203 if (n != nullptr) {
5204 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5205 if (n->is_Call()) {
5206 CallNode* call = n->as_Call();
5207 if (call->_name != nullptr) {
5208 // E.g. uncommon traps etc.
5209 ss.print(" - %s", call->_name);
5210 } else if (call->is_CallJava()) {
5211 CallJavaNode* call_java = call->as_CallJava();
5212 if (call_java->method() != nullptr) {
5213 ss.print(" -");
5214 call_java->method()->print_short_name(&ss);
5215 }
5216 }
5217 }
5218 }
5219
5220 const char* name = ss.as_string();
5221 if (should_print_igv(level)) {
5222 _igv_printer->print_graph(name);
5223 }
5224 if (should_print_phase(level)) {
5225 print_phase(name);
5226 }
5227 if (should_print_ideal_phase(compile_phase)) {
5228 print_ideal_ir(CompilerPhaseTypeHelper::to_name(compile_phase));
5229 }
5230 #endif
5231 C->_latest_stage_start_counter.stamp();
5232 }
5233
5234 // Only used from CompileWrapper
5235 void Compile::begin_method() {
5236 #ifndef PRODUCT
5237 if (_method != nullptr && should_print_igv(1)) {
5238 _igv_printer->begin_method();
5239 }
5240 #endif
5241 C->_latest_stage_start_counter.stamp();
5242 }
5243
5244 // Only used from CompileWrapper
5245 void Compile::end_method() {
5246 EventCompilerPhase event(UNTIMED);
5247 if (event.should_commit()) {
5248 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5249 }
5250
5251 #ifndef PRODUCT
5252 if (_method != nullptr && should_print_igv(1)) {
5253 _igv_printer->end_method();
5254 }
5255 #endif
5256 }
5257
5258 #ifndef PRODUCT
5259 bool Compile::should_print_phase(const int level) const {
5260 return PrintPhaseLevel >= 0 && directive()->PhasePrintLevelOption >= level &&
5261 _method != nullptr; // Do not print phases for stubs.
5262 }
5263
5264 bool Compile::should_print_ideal_phase(CompilerPhaseType cpt) const {
5265 return _directive->should_print_ideal_phase(cpt);
5266 }
5267
5268 void Compile::init_igv() {
5269 if (_igv_printer == nullptr) {
5270 _igv_printer = IdealGraphPrinter::printer();
5271 _igv_printer->set_compile(this);
5272 }
5273 }
5274
5275 bool Compile::should_print_igv(const int level) {
5276 PRODUCT_RETURN_(return false;);
5277
5278 if (PrintIdealGraphLevel < 0) { // disabled by the user
5279 return false;
5280 }
5281
5282 bool need = directive()->IGVPrintLevelOption >= level;
5283 if (need) {
5284 Compile::init_igv();
5285 }
5286 return need;
5287 }
5288
5289 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5290 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5291
5292 // Called from debugger. Prints method to the default file with the default phase name.
5293 // This works regardless of any Ideal Graph Visualizer flags set or not.
5294 // Use in debugger (gdb/rr): p igv_print($sp, $fp, $pc).
5295 void igv_print(void* sp, void* fp, void* pc) {
5296 frame fr(sp, fp, pc);
5297 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5298 }
5299
5300 // Same as igv_print() above but with a specified phase name.
5301 void igv_print(const char* phase_name, void* sp, void* fp, void* pc) {
5302 frame fr(sp, fp, pc);
5303 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5304 }
5305
5306 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5307 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5308 // This works regardless of any Ideal Graph Visualizer flags set or not.
5309 // Use in debugger (gdb/rr): p igv_print(true, $sp, $fp, $pc).
5310 void igv_print(bool network, void* sp, void* fp, void* pc) {
5311 frame fr(sp, fp, pc);
5312 if (network) {
5313 Compile::current()->igv_print_method_to_network(nullptr, &fr);
5314 } else {
5315 Compile::current()->igv_print_method_to_file(nullptr, false, &fr);
5316 }
5317 }
5318
5319 // Same as igv_print(bool network, ...) above but with a specified phase name.
5320 // Use in debugger (gdb/rr): p igv_print(true, "MyPhase", $sp, $fp, $pc).
5321 void igv_print(bool network, const char* phase_name, void* sp, void* fp, void* pc) {
5322 frame fr(sp, fp, pc);
5323 if (network) {
5324 Compile::current()->igv_print_method_to_network(phase_name, &fr);
5325 } else {
5326 Compile::current()->igv_print_method_to_file(phase_name, false, &fr);
5327 }
5328 }
5329
5330 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5331 void igv_print_default() {
5332 Compile::current()->print_method(PHASE_DEBUG, 0);
5333 }
5334
5335 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5336 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5337 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5338 // Use in debugger (gdb/rr): p igv_append($sp, $fp, $pc).
5339 void igv_append(void* sp, void* fp, void* pc) {
5340 frame fr(sp, fp, pc);
5341 Compile::current()->igv_print_method_to_file(nullptr, true, &fr);
5342 }
5343
5344 // Same as igv_append(...) above but with a specified phase name.
5345 // Use in debugger (gdb/rr): p igv_append("MyPhase", $sp, $fp, $pc).
5346 void igv_append(const char* phase_name, void* sp, void* fp, void* pc) {
5347 frame fr(sp, fp, pc);
5348 Compile::current()->igv_print_method_to_file(phase_name, true, &fr);
5349 }
5350
5351 void Compile::igv_print_method_to_file(const char* phase_name, bool append, const frame* fr) {
5352 const char* file_name = "custom_debug.xml";
5353 if (_debug_file_printer == nullptr) {
5354 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5355 } else {
5356 _debug_file_printer->update_compiled_method(C->method());
5357 }
5358 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5359 _debug_file_printer->print_graph(phase_name, fr);
5360 }
5361
5362 void Compile::igv_print_method_to_network(const char* phase_name, const frame* fr) {
5363 ResourceMark rm;
5364 GrowableArray<const Node*> empty_list;
5365 igv_print_graph_to_network(phase_name, empty_list, fr);
5366 }
5367
5368 void Compile::igv_print_graph_to_network(const char* name, GrowableArray<const Node*>& visible_nodes, const frame* fr) {
5369 if (_debug_network_printer == nullptr) {
5370 _debug_network_printer = new IdealGraphPrinter(C);
5371 } else {
5372 _debug_network_printer->update_compiled_method(C->method());
5373 }
5374 tty->print_cr("Method printed over network stream to IGV");
5375 _debug_network_printer->print(name, C->root(), visible_nodes, fr);
5376 }
5377 #endif // !PRODUCT
5378
5379 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5380 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5381 return value;
5382 }
5383 Node* result = nullptr;
5384 if (bt == T_BYTE) {
5385 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5386 result = new RShiftINode(result, phase->intcon(24));
5387 } else if (bt == T_BOOLEAN) {
5388 result = new AndINode(value, phase->intcon(0xFF));
5389 } else if (bt == T_CHAR) {
5390 result = new AndINode(value,phase->intcon(0xFFFF));
5391 } else {
5392 assert(bt == T_SHORT, "unexpected narrow type");
5393 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5394 result = new RShiftINode(result, phase->intcon(16));
5395 }
5396 if (transform_res) {
5397 result = phase->transform(result);
5398 }
5399 return result;
5400 }
5401
5402 void Compile::record_method_not_compilable_oom() {
5403 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5404 }
5405
5406 #ifndef PRODUCT
5407 // Collects all the control inputs from nodes on the worklist and from their data dependencies
5408 static void find_candidate_control_inputs(Unique_Node_List& worklist, Unique_Node_List& candidates) {
5409 // Follow non-control edges until we reach CFG nodes
5410 for (uint i = 0; i < worklist.size(); i++) {
5411 const Node* n = worklist.at(i);
5412 for (uint j = 0; j < n->req(); j++) {
5413 Node* in = n->in(j);
5414 if (in == nullptr || in->is_Root()) {
5415 continue;
5416 }
5417 if (in->is_CFG()) {
5418 if (in->is_Call()) {
5419 // The return value of a call is only available if the call did not result in an exception
5420 Node* control_proj_use = in->as_Call()->proj_out(TypeFunc::Control)->unique_out();
5421 if (control_proj_use->is_Catch()) {
5422 Node* fall_through = control_proj_use->as_Catch()->proj_out(CatchProjNode::fall_through_index);
5423 candidates.push(fall_through);
5424 continue;
5425 }
5426 }
5427
5428 if (in->is_Multi()) {
5429 // We got here by following data inputs so we should only have one control use
5430 // (no IfNode, etc)
5431 assert(!n->is_MultiBranch(), "unexpected node type: %s", n->Name());
5432 candidates.push(in->as_Multi()->proj_out(TypeFunc::Control));
5433 } else {
5434 candidates.push(in);
5435 }
5436 } else {
5437 worklist.push(in);
5438 }
5439 }
5440 }
5441 }
5442
5443 // Returns the candidate node that is a descendant to all the other candidates
5444 static Node* pick_control(Unique_Node_List& candidates) {
5445 Unique_Node_List worklist;
5446 worklist.copy(candidates);
5447
5448 // Traverse backwards through the CFG
5449 for (uint i = 0; i < worklist.size(); i++) {
5450 const Node* n = worklist.at(i);
5451 if (n->is_Root()) {
5452 continue;
5453 }
5454 for (uint j = 0; j < n->req(); j++) {
5455 // Skip backedge of loops to avoid cycles
5456 if (n->is_Loop() && j == LoopNode::LoopBackControl) {
5457 continue;
5458 }
5459
5460 Node* pred = n->in(j);
5461 if (pred != nullptr && pred != n && pred->is_CFG()) {
5462 worklist.push(pred);
5463 // if pred is an ancestor of n, then pred is an ancestor to at least one candidate
5464 candidates.remove(pred);
5465 }
5466 }
5467 }
5468
5469 assert(candidates.size() == 1, "unexpected control flow");
5470 return candidates.at(0);
5471 }
5472
5473 // Initialize a parameter input for a debug print call, using a placeholder for jlong and jdouble
5474 static void debug_print_init_parm(Node* call, Node* parm, Node* half, int* pos) {
5475 call->init_req((*pos)++, parm);
5476 const BasicType bt = parm->bottom_type()->basic_type();
5477 if (bt == T_LONG || bt == T_DOUBLE) {
5478 call->init_req((*pos)++, half);
5479 }
5480 }
5481
5482 Node* Compile::make_debug_print_call(const char* str, address call_addr, PhaseGVN* gvn,
5483 Node* parm0, Node* parm1,
5484 Node* parm2, Node* parm3,
5485 Node* parm4, Node* parm5,
5486 Node* parm6) const {
5487 Node* str_node = gvn->transform(new ConPNode(TypeRawPtr::make(((address) str))));
5488 const TypeFunc* type = OptoRuntime::debug_print_Type(parm0, parm1, parm2, parm3, parm4, parm5, parm6);
5489 Node* call = new CallLeafNode(type, call_addr, "debug_print", TypeRawPtr::BOTTOM);
5490
5491 // find the most suitable control input
5492 Unique_Node_List worklist, candidates;
5493 if (parm0 != nullptr) { worklist.push(parm0);
5494 if (parm1 != nullptr) { worklist.push(parm1);
5495 if (parm2 != nullptr) { worklist.push(parm2);
5496 if (parm3 != nullptr) { worklist.push(parm3);
5497 if (parm4 != nullptr) { worklist.push(parm4);
5498 if (parm5 != nullptr) { worklist.push(parm5);
5499 if (parm6 != nullptr) { worklist.push(parm6);
5500 /* close each nested if ===> */ } } } } } } }
5501 find_candidate_control_inputs(worklist, candidates);
5502 Node* control = nullptr;
5503 if (candidates.size() == 0) {
5504 control = C->start()->proj_out(TypeFunc::Control);
5505 } else {
5506 control = pick_control(candidates);
5507 }
5508
5509 // find all the previous users of the control we picked
5510 GrowableArray<Node*> users_of_control;
5511 for (DUIterator_Fast kmax, i = control->fast_outs(kmax); i < kmax; i++) {
5512 Node* use = control->fast_out(i);
5513 if (use->is_CFG() && use != control) {
5514 users_of_control.push(use);
5515 }
5516 }
5517
5518 // we do not actually care about IO and memory as it uses neither
5519 call->init_req(TypeFunc::Control, control);
5520 call->init_req(TypeFunc::I_O, top());
5521 call->init_req(TypeFunc::Memory, top());
5522 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr));
5523 call->init_req(TypeFunc::ReturnAdr, top());
5524
5525 int pos = TypeFunc::Parms;
5526 call->init_req(pos++, str_node);
5527 if (parm0 != nullptr) { debug_print_init_parm(call, parm0, top(), &pos);
5528 if (parm1 != nullptr) { debug_print_init_parm(call, parm1, top(), &pos);
5529 if (parm2 != nullptr) { debug_print_init_parm(call, parm2, top(), &pos);
5530 if (parm3 != nullptr) { debug_print_init_parm(call, parm3, top(), &pos);
5531 if (parm4 != nullptr) { debug_print_init_parm(call, parm4, top(), &pos);
5532 if (parm5 != nullptr) { debug_print_init_parm(call, parm5, top(), &pos);
5533 if (parm6 != nullptr) { debug_print_init_parm(call, parm6, top(), &pos);
5534 /* close each nested if ===> */ } } } } } } }
5535 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
5536
5537 call = gvn->transform(call);
5538 Node* call_control_proj = gvn->transform(new ProjNode(call, TypeFunc::Control));
5539
5540 // rewire previous users to have the new call as control instead
5541 PhaseIterGVN* igvn = gvn->is_IterGVN();
5542 for (int i = 0; i < users_of_control.length(); i++) {
5543 Node* use = users_of_control.at(i);
5544 for (uint j = 0; j < use->req(); j++) {
5545 if (use->in(j) == control) {
5546 if (igvn != nullptr) {
5547 igvn->replace_input_of(use, j, call_control_proj);
5548 } else {
5549 gvn->hash_delete(use);
5550 use->set_req(j, call_control_proj);
5551 gvn->hash_insert(use);
5552 }
5553 }
5554 }
5555 }
5556
5557 return call;
5558 }
5559 #endif // !PRODUCT