1 /*
2 * Copyright (c) 1997, 2022, 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 #include "precompiled.hpp"
25 #include "jvm_io.h"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSet.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "jfr/jfrEvents.hpp"
39 #include "memory/resourceArea.hpp"
40 #include "opto/addnode.hpp"
41 #include "opto/block.hpp"
42 #include "opto/c2compiler.hpp"
43 #include "opto/callGenerator.hpp"
44 #include "opto/callnode.hpp"
45 #include "opto/castnode.hpp"
46 #include "opto/cfgnode.hpp"
47 #include "opto/chaitin.hpp"
48 #include "opto/compile.hpp"
49 #include "opto/connode.hpp"
50 #include "opto/convertnode.hpp"
51 #include "opto/divnode.hpp"
52 #include "opto/escape.hpp"
53 #include "opto/idealGraphPrinter.hpp"
54 #include "opto/inlinetypenode.hpp"
55 #include "opto/loopnode.hpp"
56 #include "opto/machnode.hpp"
57 #include "opto/macro.hpp"
58 #include "opto/matcher.hpp"
59 #include "opto/mathexactnode.hpp"
60 #include "opto/memnode.hpp"
61 #include "opto/mulnode.hpp"
62 #include "opto/narrowptrnode.hpp"
63 #include "opto/node.hpp"
64 #include "opto/opcodes.hpp"
65 #include "opto/output.hpp"
66 #include "opto/parse.hpp"
67 #include "opto/phaseX.hpp"
68 #include "opto/rootnode.hpp"
69 #include "opto/runtime.hpp"
70 #include "opto/stringopts.hpp"
71 #include "opto/type.hpp"
72 #include "opto/vector.hpp"
73 #include "opto/vectornode.hpp"
74 #include "runtime/globals_extension.hpp"
75 #include "runtime/sharedRuntime.hpp"
76 #include "runtime/signature.hpp"
77 #include "runtime/stubRoutines.hpp"
78 #include "runtime/timer.hpp"
79 #include "utilities/align.hpp"
80 #include "utilities/copy.hpp"
81 #include "utilities/macros.hpp"
82 #include "utilities/resourceHash.hpp"
83
84
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == NULL) {
89 _mach_constant_base_node = new MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
94
95
96 /// Support for intrinsics.
97
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 class IntrinsicDescPair {
101 private:
102 ciMethod* _m;
103 bool _is_virtual;
104 public:
105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
107 ciMethod* m= elt->method();
108 ciMethod* key_m = key->_m;
109 if (key_m < m) return -1;
110 else if (key_m > m) return 1;
111 else {
112 bool is_virtual = elt->is_virtual();
113 bool key_virtual = key->_is_virtual;
114 if (key_virtual < is_virtual) return -1;
115 else if (key_virtual > is_virtual) return 1;
116 else return 0;
117 }
118 }
119 };
120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
121 #ifdef ASSERT
122 for (int i = 1; i < _intrinsics.length(); i++) {
123 CallGenerator* cg1 = _intrinsics.at(i-1);
124 CallGenerator* cg2 = _intrinsics.at(i);
125 assert(cg1->method() != cg2->method()
126 ? cg1->method() < cg2->method()
127 : cg1->is_virtual() < cg2->is_virtual(),
128 "compiler intrinsics list must stay sorted");
129 }
130 #endif
131 IntrinsicDescPair pair(m, is_virtual);
132 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
133 }
134
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 bool found = false;
137 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
138 assert(!found, "registering twice");
139 _intrinsics.insert_before(index, cg);
140 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
141 }
142
143 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
144 assert(m->is_loaded(), "don't try this on unloaded methods");
145 if (_intrinsics.length() > 0) {
146 bool found = false;
147 int index = intrinsic_insertion_index(m, is_virtual, found);
148 if (found) {
149 return _intrinsics.at(index);
150 }
151 }
152 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
153 if (m->intrinsic_id() != vmIntrinsics::_none &&
154 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
155 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
156 if (cg != NULL) {
157 // Save it for next time:
158 register_intrinsic(cg);
159 return cg;
160 } else {
161 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
162 }
163 }
164 return NULL;
165 }
166
167 // Compile::make_vm_intrinsic is defined in library_call.cpp.
168
169 #ifndef PRODUCT
170 // statistics gathering...
171
172 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
174
175 inline int as_int(vmIntrinsics::ID id) {
176 return vmIntrinsics::as_int(id);
177 }
178
179 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
180 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
181 int oflags = _intrinsic_hist_flags[as_int(id)];
182 assert(flags != 0, "what happened?");
183 if (is_virtual) {
184 flags |= _intrinsic_virtual;
185 }
186 bool changed = (flags != oflags);
187 if ((flags & _intrinsic_worked) != 0) {
188 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
189 if (count == 1) {
190 changed = true; // first time
191 }
192 // increment the overall count also:
193 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
194 }
195 if (changed) {
196 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
197 // Something changed about the intrinsic's virtuality.
198 if ((flags & _intrinsic_virtual) != 0) {
199 // This is the first use of this intrinsic as a virtual call.
200 if (oflags != 0) {
201 // We already saw it as a non-virtual, so note both cases.
202 flags |= _intrinsic_both;
203 }
204 } else if ((oflags & _intrinsic_both) == 0) {
205 // This is the first use of this intrinsic as a non-virtual
206 flags |= _intrinsic_both;
207 }
208 }
209 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
210 }
211 // update the overall flags also:
212 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
213 return changed;
214 }
215
216 static char* format_flags(int flags, char* buf) {
217 buf[0] = 0;
218 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
219 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
220 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
221 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
222 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
223 if (buf[0] == 0) strcat(buf, ",");
224 assert(buf[0] == ',', "must be");
225 return &buf[1];
226 }
227
228 void Compile::print_intrinsic_statistics() {
229 char flagsbuf[100];
230 ttyLocker ttyl;
231 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
232 tty->print_cr("Compiler intrinsic usage:");
233 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
234 if (total == 0) total = 1; // avoid div0 in case of no successes
235 #define PRINT_STAT_LINE(name, c, f) \
236 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
237 for (auto id : EnumRange<vmIntrinsicID>{}) {
238 int flags = _intrinsic_hist_flags[as_int(id)];
239 juint count = _intrinsic_hist_count[as_int(id)];
240 if ((flags | count) != 0) {
241 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
242 }
243 }
244 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
245 if (xtty != NULL) xtty->tail("statistics");
246 }
247
248 void Compile::print_statistics() {
249 { ttyLocker ttyl;
250 if (xtty != NULL) xtty->head("statistics type='opto'");
251 Parse::print_statistics();
252 PhaseStringOpts::print_statistics();
253 PhaseCCP::print_statistics();
254 PhaseRegAlloc::print_statistics();
255 PhaseOutput::print_statistics();
256 PhasePeephole::print_statistics();
257 PhaseIdealLoop::print_statistics();
258 ConnectionGraph::print_statistics();
259 PhaseMacroExpand::print_statistics();
260 if (xtty != NULL) xtty->tail("statistics");
261 }
262 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
263 // put this under its own <statistics> element.
264 print_intrinsic_statistics();
265 }
266 }
267 #endif //PRODUCT
268
269 void Compile::gvn_replace_by(Node* n, Node* nn) {
270 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
271 Node* use = n->last_out(i);
272 bool is_in_table = initial_gvn()->hash_delete(use);
273 uint uses_found = 0;
274 for (uint j = 0; j < use->len(); j++) {
275 if (use->in(j) == n) {
276 if (j < use->req())
277 use->set_req(j, nn);
278 else
279 use->set_prec(j, nn);
280 uses_found++;
281 }
282 }
283 if (is_in_table) {
284 // reinsert into table
285 initial_gvn()->hash_find_insert(use);
286 }
287 record_for_igvn(use);
288 i -= uses_found; // we deleted 1 or more copies of this edge
289 }
290 }
291
292
293 // Identify all nodes that are reachable from below, useful.
294 // Use breadth-first pass that records state in a Unique_Node_List,
295 // recursive traversal is slower.
296 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
297 int estimated_worklist_size = live_nodes();
298 useful.map( estimated_worklist_size, NULL ); // preallocate space
299
300 // Initialize worklist
301 if (root() != NULL) { useful.push(root()); }
302 // If 'top' is cached, declare it useful to preserve cached node
303 if( cached_top_node() ) { useful.push(cached_top_node()); }
304
305 // Push all useful nodes onto the list, breadthfirst
306 for( uint next = 0; next < useful.size(); ++next ) {
307 assert( next < unique(), "Unique useful nodes < total nodes");
308 Node *n = useful.at(next);
309 uint max = n->len();
310 for( uint i = 0; i < max; ++i ) {
311 Node *m = n->in(i);
312 if (not_a_node(m)) continue;
313 useful.push(m);
314 }
315 }
316 }
317
318 // Update dead_node_list with any missing dead nodes using useful
319 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
320 void Compile::update_dead_node_list(Unique_Node_List &useful) {
321 uint max_idx = unique();
322 VectorSet& useful_node_set = useful.member_set();
323
324 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
325 // If node with index node_idx is not in useful set,
326 // mark it as dead in dead node list.
327 if (!useful_node_set.test(node_idx)) {
328 record_dead_node(node_idx);
329 }
330 }
331 }
332
333 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
334 int shift = 0;
335 for (int i = 0; i < inlines->length(); i++) {
336 CallGenerator* cg = inlines->at(i);
337 if (useful.member(cg->call_node())) {
338 if (shift > 0) {
339 inlines->at_put(i - shift, cg);
340 }
341 } else {
342 shift++; // skip over the dead element
343 }
344 }
345 if (shift > 0) {
346 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
347 }
348 }
349
350 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
351 assert(dead != NULL && dead->is_Call(), "sanity");
352 int found = 0;
353 for (int i = 0; i < inlines->length(); i++) {
354 if (inlines->at(i)->call_node() == dead) {
355 inlines->remove_at(i);
356 found++;
357 NOT_DEBUG( break; ) // elements are unique, so exit early
358 }
359 }
360 assert(found <= 1, "not unique");
361 }
362
363 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) {
364 for (int i = node_list.length() - 1; i >= 0; i--) {
365 Node* n = node_list.at(i);
366 if (!useful.member(n)) {
367 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
368 }
369 }
370 }
371
372 void Compile::remove_useless_node(Node* dead) {
373 remove_modified_node(dead);
374
375 // Constant node that has no out-edges and has only one in-edge from
376 // root is usually dead. However, sometimes reshaping walk makes
377 // it reachable by adding use edges. So, we will NOT count Con nodes
378 // as dead to be conservative about the dead node count at any
379 // given time.
380 if (!dead->is_Con()) {
381 record_dead_node(dead->_idx);
382 }
383 if (dead->is_macro()) {
384 remove_macro_node(dead);
385 }
386 if (dead->is_expensive()) {
387 remove_expensive_node(dead);
388 }
389 if (dead->Opcode() == Op_Opaque4) {
390 remove_skeleton_predicate_opaq(dead);
391 }
392 if (dead->for_post_loop_opts_igvn()) {
393 remove_from_post_loop_opts_igvn(dead);
394 }
395 if (dead->is_InlineTypeBase()) {
396 remove_inline_type(dead);
397 }
398 if (dead->is_Call()) {
399 remove_useless_late_inlines( &_late_inlines, dead);
400 remove_useless_late_inlines( &_string_late_inlines, dead);
401 remove_useless_late_inlines( &_boxing_late_inlines, dead);
402 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
403
404 if (dead->is_CallStaticJava()) {
405 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
406 }
407 }
408 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
409 bs->unregister_potential_barrier_node(dead);
410 }
411
412 // Disconnect all useless nodes by disconnecting those at the boundary.
413 void Compile::disconnect_useless_nodes(Unique_Node_List &useful, Unique_Node_List* worklist) {
414 uint next = 0;
415 while (next < useful.size()) {
416 Node *n = useful.at(next++);
417 if (n->is_SafePoint()) {
418 // We're done with a parsing phase. Replaced nodes are not valid
419 // beyond that point.
420 n->as_SafePoint()->delete_replaced_nodes();
421 }
422 // Use raw traversal of out edges since this code removes out edges
423 int max = n->outcnt();
424 for (int j = 0; j < max; ++j) {
425 Node* child = n->raw_out(j);
426 if (!useful.member(child)) {
427 assert(!child->is_top() || child != top(),
428 "If top is cached in Compile object it is in useful list");
429 // Only need to remove this out-edge to the useless node
430 n->raw_del_out(j);
431 --j;
432 --max;
433 }
434 }
435 if (n->outcnt() == 1 && n->has_special_unique_user()) {
436 worklist->push(n->unique_out());
437 }
438 if (n->outcnt() == 0) {
439 worklist->push(n);
440 }
441 }
442
443 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
444 remove_useless_nodes(_predicate_opaqs, useful); // remove useless predicate opaque nodes
445 remove_useless_nodes(_skeleton_predicate_opaqs, useful);
446 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
447 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
448 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
449 #ifdef ASSERT
450 if (_modified_nodes != NULL) {
451 _modified_nodes->remove_useless_nodes(useful.member_set());
452 }
453 #endif
454 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
455 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
456 #ifdef ASSERT
457 if (_modified_nodes != NULL) {
458 _modified_nodes->remove_useless_nodes(useful.member_set());
459 }
460 #endif
461
462 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
463 bs->eliminate_useless_gc_barriers(useful, this);
464 // clean up the late inline lists
465 remove_useless_late_inlines( &_late_inlines, useful);
466 remove_useless_late_inlines( &_string_late_inlines, useful);
467 remove_useless_late_inlines( &_boxing_late_inlines, useful);
468 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
469 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
470 }
471
472 // ============================================================================
473 //------------------------------CompileWrapper---------------------------------
474 class CompileWrapper : public StackObj {
475 Compile *const _compile;
476 public:
477 CompileWrapper(Compile* compile);
478
479 ~CompileWrapper();
480 };
481
482 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
483 // the Compile* pointer is stored in the current ciEnv:
484 ciEnv* env = compile->env();
485 assert(env == ciEnv::current(), "must already be a ciEnv active");
486 assert(env->compiler_data() == NULL, "compile already active?");
487 env->set_compiler_data(compile);
488 assert(compile == Compile::current(), "sanity");
489
490 compile->set_type_dict(NULL);
491 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
492 compile->clone_map().set_clone_idx(0);
493 compile->set_type_last_size(0);
494 compile->set_last_tf(NULL, NULL);
495 compile->set_indexSet_arena(NULL);
496 compile->set_indexSet_free_block_list(NULL);
497 compile->init_type_arena();
498 Type::Initialize(compile);
499 _compile->begin_method();
500 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
501 }
502 CompileWrapper::~CompileWrapper() {
503 // simulate crash during compilation
504 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
505
506 _compile->end_method();
507 _compile->env()->set_compiler_data(NULL);
508 }
509
510
511 //----------------------------print_compile_messages---------------------------
512 void Compile::print_compile_messages() {
513 #ifndef PRODUCT
514 // Check if recompiling
515 if (!subsume_loads() && PrintOpto) {
516 // Recompiling without allowing machine instructions to subsume loads
517 tty->print_cr("*********************************************************");
518 tty->print_cr("** Bailout: Recompile without subsuming loads **");
519 tty->print_cr("*********************************************************");
520 }
521 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
522 // Recompiling without escape analysis
523 tty->print_cr("*********************************************************");
524 tty->print_cr("** Bailout: Recompile without escape analysis **");
525 tty->print_cr("*********************************************************");
526 }
527 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
528 // Recompiling without iterative escape analysis
529 tty->print_cr("*********************************************************");
530 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
531 tty->print_cr("*********************************************************");
532 }
533 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
534 // Recompiling without boxing elimination
535 tty->print_cr("*********************************************************");
536 tty->print_cr("** Bailout: Recompile without boxing elimination **");
537 tty->print_cr("*********************************************************");
538 }
539 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
540 // Recompiling without locks coarsening
541 tty->print_cr("*********************************************************");
542 tty->print_cr("** Bailout: Recompile without locks coarsening **");
543 tty->print_cr("*********************************************************");
544 }
545 if (env()->break_at_compile()) {
546 // Open the debugger when compiling this method.
547 tty->print("### Breaking when compiling: ");
548 method()->print_short_name();
549 tty->cr();
550 BREAKPOINT;
551 }
552
553 if( PrintOpto ) {
554 if (is_osr_compilation()) {
555 tty->print("[OSR]%3d", _compile_id);
556 } else {
557 tty->print("%3d", _compile_id);
558 }
559 }
560 #endif
561 }
562
563 #ifndef PRODUCT
564 void Compile::print_ideal_ir(const char* phase_name) {
565 ttyLocker ttyl;
566 // keep the following output all in one block
567 // This output goes directly to the tty, not the compiler log.
568 // To enable tools to match it up with the compilation activity,
569 // be sure to tag this tty output with the compile ID.
570 if (xtty != NULL) {
571 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
572 compile_id(),
573 is_osr_compilation() ? " compile_kind='osr'" : "",
574 phase_name);
575 }
576 if (_output == nullptr) {
577 tty->print_cr("AFTER: %s", phase_name);
578 // Print out all nodes in ascending order of index.
579 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$");
580 } else {
581 // Dump the node blockwise if we have a scheduling
582 _output->print_scheduling();
583 }
584
585 if (xtty != NULL) {
586 xtty->tail("ideal");
587 }
588 }
589 #endif
590
591 // ============================================================================
592 //------------------------------Compile standard-------------------------------
593 debug_only( int Compile::_debug_idx = 100000; )
594
595 // Compile a method. entry_bci is -1 for normal compilations and indicates
596 // the continuation bci for on stack replacement.
597
598
599 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
600 Options options, DirectiveSet* directive)
601 : Phase(Compiler),
602 _compile_id(ci_env->compile_id()),
603 _options(options),
604 _method(target),
605 _entry_bci(osr_bci),
606 _ilt(NULL),
607 _stub_function(NULL),
608 _stub_name(NULL),
609 _stub_entry_point(NULL),
610 _max_node_limit(MaxNodeLimit),
611 _post_loop_opts_phase(false),
612 _inlining_progress(false),
613 _inlining_incrementally(false),
614 _do_cleanup(false),
615 _has_reserved_stack_access(target->has_reserved_stack_access()),
616 #ifndef PRODUCT
617 _igv_idx(0),
618 _trace_opto_output(directive->TraceOptoOutputOption),
619 #endif
620 _has_method_handle_invokes(false),
621 _clinit_barrier_on_entry(false),
622 _stress_seed(0),
623 _comp_arena(mtCompiler),
624 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
625 _env(ci_env),
626 _directive(directive),
627 _log(ci_env->log()),
628 _failure_reason(NULL),
629 _intrinsics (comp_arena(), 0, 0, NULL),
630 _macro_nodes (comp_arena(), 8, 0, NULL),
631 _predicate_opaqs (comp_arena(), 8, 0, NULL),
632 _skeleton_predicate_opaqs (comp_arena(), 8, 0, NULL),
633 _expensive_nodes (comp_arena(), 8, 0, NULL),
634 _for_post_loop_igvn(comp_arena(), 8, 0, NULL),
635 _inline_type_nodes (comp_arena(), 8, 0, NULL),
636 _unstable_if_traps (comp_arena(), 8, 0, NULL),
637 _coarsened_locks (comp_arena(), 8, 0, NULL),
638 _congraph(NULL),
639 NOT_PRODUCT(_igv_printer(NULL) COMMA)
640 _dead_node_list(comp_arena()),
641 _dead_node_count(0),
642 _node_arena(mtCompiler),
643 _old_arena(mtCompiler),
644 _mach_constant_base_node(NULL),
645 _Compile_types(mtCompiler),
646 _initial_gvn(NULL),
647 _for_igvn(NULL),
648 _late_inlines(comp_arena(), 2, 0, NULL),
649 _string_late_inlines(comp_arena(), 2, 0, NULL),
650 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
651 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL),
652 _late_inlines_pos(0),
653 _number_of_mh_late_inlines(0),
654 _print_inlining_stream(new stringStream()),
655 _print_inlining_list(NULL),
656 _print_inlining_idx(0),
657 _print_inlining_output(NULL),
658 _replay_inline_data(NULL),
659 _java_calls(0),
660 _inner_loops(0),
661 _interpreter_frame_size(0),
662 _output(NULL)
663 #ifndef PRODUCT
664 , _in_dump_cnt(0)
665 #endif
666 {
667 C = this;
668 CompileWrapper cw(this);
669
670 if (CITimeVerbose) {
671 tty->print(" ");
672 target->holder()->name()->print();
673 tty->print(".");
674 target->print_short_name();
675 tty->print(" ");
676 }
677 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
678 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
679
680 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
681 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
682 // We can always print a disassembly, either abstract (hex dump) or
683 // with the help of a suitable hsdis library. Thus, we should not
684 // couple print_assembly and print_opto_assembly controls.
685 // But: always print opto and regular assembly on compile command 'print'.
686 bool print_assembly = directive->PrintAssemblyOption;
687 set_print_assembly(print_opto_assembly || print_assembly);
688 #else
689 set_print_assembly(false); // must initialize.
690 #endif
691
692 #ifndef PRODUCT
693 set_parsed_irreducible_loop(false);
694
695 if (directive->ReplayInlineOption) {
696 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
697 }
698 #endif
699 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
700 set_print_intrinsics(directive->PrintIntrinsicsOption);
701 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
702
703 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
704 // Make sure the method being compiled gets its own MDO,
705 // so we can at least track the decompile_count().
706 // Need MDO to record RTM code generation state.
707 method()->ensure_method_data();
708 }
709
710 Init(::AliasLevel);
711
712
713 print_compile_messages();
714
715 _ilt = InlineTree::build_inline_tree_root();
716
717 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
718 assert(num_alias_types() >= AliasIdxRaw, "");
719
720 #define MINIMUM_NODE_HASH 1023
721 // Node list that Iterative GVN will start with
722 Unique_Node_List for_igvn(comp_arena());
723 set_for_igvn(&for_igvn);
724
725 // GVN that will be run immediately on new nodes
726 uint estimated_size = method()->code_size()*4+64;
727 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
728 PhaseGVN gvn(node_arena(), estimated_size);
729 set_initial_gvn(&gvn);
730
731 print_inlining_init();
732 { // Scope for timing the parser
733 TracePhase tp("parse", &timers[_t_parser]);
734
735 // Put top into the hash table ASAP.
736 initial_gvn()->transform_no_reclaim(top());
737
738 // Set up tf(), start(), and find a CallGenerator.
739 CallGenerator* cg = NULL;
740 if (is_osr_compilation()) {
741 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
742 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
743 initial_gvn()->set_type_bottom(s);
744 init_start(s);
745 cg = CallGenerator::for_osr(method(), entry_bci());
746 } else {
747 // Normal case.
748 init_tf(TypeFunc::make(method()));
749 StartNode* s = new StartNode(root(), tf()->domain_cc());
750 initial_gvn()->set_type_bottom(s);
751 init_start(s);
752 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
753 // With java.lang.ref.reference.get() we must go through the
754 // intrinsic - even when get() is the root
755 // method of the compile - so that, if necessary, the value in
756 // the referent field of the reference object gets recorded by
757 // the pre-barrier code.
758 cg = find_intrinsic(method(), false);
759 }
760 if (cg == NULL) {
761 float past_uses = method()->interpreter_invocation_count();
762 float expected_uses = past_uses;
763 cg = CallGenerator::for_inline(method(), expected_uses);
764 }
765 }
766 if (failing()) return;
767 if (cg == NULL) {
768 record_method_not_compilable("cannot parse method");
769 return;
770 }
771 JVMState* jvms = build_start_state(start(), tf());
772 if ((jvms = cg->generate(jvms)) == NULL) {
773 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
774 record_method_not_compilable("method parse failed");
775 }
776 return;
777 }
778 GraphKit kit(jvms);
779
780 if (!kit.stopped()) {
781 // Accept return values, and transfer control we know not where.
782 // This is done by a special, unique ReturnNode bound to root.
783 return_values(kit.jvms());
784 }
785
786 if (kit.has_exceptions()) {
787 // Any exceptions that escape from this call must be rethrown
788 // to whatever caller is dynamically above us on the stack.
789 // This is done by a special, unique RethrowNode bound to root.
790 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
791 }
792
793 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
794
795 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
796 inline_string_calls(true);
797 }
798
799 if (failing()) return;
800
801 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
802
803 // Remove clutter produced by parsing.
804 if (!failing()) {
805 ResourceMark rm;
806 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
807 }
808 }
809
810 // Note: Large methods are capped off in do_one_bytecode().
811 if (failing()) return;
812
813 // After parsing, node notes are no longer automagic.
814 // They must be propagated by register_new_node_with_optimizer(),
815 // clone(), or the like.
816 set_default_node_notes(NULL);
817
818 #ifndef PRODUCT
819 if (should_print_igv(1)) {
820 _igv_printer->print_inlining();
821 }
822 #endif
823
824 if (failing()) return;
825 NOT_PRODUCT( verify_graph_edges(); )
826
827 // If any phase is randomized for stress testing, seed random number
828 // generation and log the seed for repeatability.
829 if (StressLCM || StressGCM || StressIGVN || StressCCP) {
830 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) {
831 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
832 FLAG_SET_ERGO(StressSeed, _stress_seed);
833 } else {
834 _stress_seed = StressSeed;
835 }
836 if (_log != NULL) {
837 _log->elem("stress_test seed='%u'", _stress_seed);
838 }
839 }
840
841 // Now optimize
842 Optimize();
843 if (failing()) return;
844 NOT_PRODUCT( verify_graph_edges(); )
845
846 #ifndef PRODUCT
847 if (should_print_ideal()) {
848 print_ideal_ir("print_ideal");
849 }
850 #endif
851
852 #ifdef ASSERT
853 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
854 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
855 #endif
856
857 // Dump compilation data to replay it.
858 if (directive->DumpReplayOption) {
859 env()->dump_replay_data(_compile_id);
860 }
861 if (directive->DumpInlineOption && (ilt() != NULL)) {
862 env()->dump_inline_data(_compile_id);
863 }
864
865 // Now that we know the size of all the monitors we can add a fixed slot
866 // for the original deopt pc.
867 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
868 if (needs_stack_repair()) {
869 // One extra slot for the special stack increment value
870 next_slot += 2;
871 }
872 // TODO 8284443 Only reserve extra slot if needed
873 if (InlineTypeReturnedAsFields) {
874 // One extra slot to hold the IsInit information for a nullable
875 // inline type return if we run out of registers.
876 next_slot += 2;
877 }
878 set_fixed_slots(next_slot);
879
880 // Compute when to use implicit null checks. Used by matching trap based
881 // nodes and NullCheck optimization.
882 set_allowed_deopt_reasons();
883
884 // Now generate code
885 Code_Gen();
886 }
887
888 //------------------------------Compile----------------------------------------
889 // Compile a runtime stub
890 Compile::Compile( ciEnv* ci_env,
891 TypeFunc_generator generator,
892 address stub_function,
893 const char *stub_name,
894 int is_fancy_jump,
895 bool pass_tls,
896 bool return_pc,
897 DirectiveSet* directive)
898 : Phase(Compiler),
899 _compile_id(0),
900 _options(Options::for_runtime_stub()),
901 _method(NULL),
902 _entry_bci(InvocationEntryBci),
903 _stub_function(stub_function),
904 _stub_name(stub_name),
905 _stub_entry_point(NULL),
906 _max_node_limit(MaxNodeLimit),
907 _post_loop_opts_phase(false),
908 _inlining_progress(false),
909 _inlining_incrementally(false),
910 _has_reserved_stack_access(false),
911 #ifndef PRODUCT
912 _igv_idx(0),
913 _trace_opto_output(directive->TraceOptoOutputOption),
914 #endif
915 _has_method_handle_invokes(false),
916 _clinit_barrier_on_entry(false),
917 _stress_seed(0),
918 _comp_arena(mtCompiler),
919 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
920 _env(ci_env),
921 _directive(directive),
922 _log(ci_env->log()),
923 _failure_reason(NULL),
924 _congraph(NULL),
925 NOT_PRODUCT(_igv_printer(NULL) COMMA)
926 _dead_node_list(comp_arena()),
927 _dead_node_count(0),
928 _node_arena(mtCompiler),
929 _old_arena(mtCompiler),
930 _mach_constant_base_node(NULL),
931 _Compile_types(mtCompiler),
932 _initial_gvn(NULL),
933 _for_igvn(NULL),
934 _number_of_mh_late_inlines(0),
935 _print_inlining_stream(new stringStream()),
936 _print_inlining_list(NULL),
937 _print_inlining_idx(0),
938 _print_inlining_output(NULL),
939 _replay_inline_data(NULL),
940 _java_calls(0),
941 _inner_loops(0),
942 _interpreter_frame_size(0),
943 _output(NULL),
944 #ifndef PRODUCT
945 _in_dump_cnt(0),
946 #endif
947 _allowed_reasons(0) {
948 C = this;
949
950 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
951 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
952
953 #ifndef PRODUCT
954 set_print_assembly(PrintFrameConverterAssembly);
955 set_parsed_irreducible_loop(false);
956 #else
957 set_print_assembly(false); // Must initialize.
958 #endif
959 set_has_irreducible_loop(false); // no loops
960
961 CompileWrapper cw(this);
962 Init(/*AliasLevel=*/ 0);
963 init_tf((*generator)());
964
965 {
966 // The following is a dummy for the sake of GraphKit::gen_stub
967 Unique_Node_List for_igvn(comp_arena());
968 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
969 PhaseGVN gvn(Thread::current()->resource_area(),255);
970 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
971 gvn.transform_no_reclaim(top());
972
973 GraphKit kit;
974 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
975 }
976
977 NOT_PRODUCT( verify_graph_edges(); )
978
979 Code_Gen();
980 }
981
982 //------------------------------Init-------------------------------------------
983 // Prepare for a single compilation
984 void Compile::Init(int aliaslevel) {
985 _unique = 0;
986 _regalloc = NULL;
987
988 _tf = NULL; // filled in later
989 _top = NULL; // cached later
990 _matcher = NULL; // filled in later
991 _cfg = NULL; // filled in later
992
993 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
994
995 _node_note_array = NULL;
996 _default_node_notes = NULL;
997 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
998
999 _immutable_memory = NULL; // filled in at first inquiry
1000
1001 // Globally visible Nodes
1002 // First set TOP to NULL to give safe behavior during creation of RootNode
1003 set_cached_top_node(NULL);
1004 set_root(new RootNode());
1005 // Now that you have a Root to point to, create the real TOP
1006 set_cached_top_node( new ConNode(Type::TOP) );
1007 set_recent_alloc(NULL, NULL);
1008
1009 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1010 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1011 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1012 env()->set_dependencies(new Dependencies(env()));
1013
1014 _fixed_slots = 0;
1015 set_has_split_ifs(false);
1016 set_has_loops(false); // first approximation
1017 set_has_stringbuilder(false);
1018 set_has_boxed_value(false);
1019 _trap_can_recompile = false; // no traps emitted yet
1020 _major_progress = true; // start out assuming good things will happen
1021 set_has_unsafe_access(false);
1022 set_max_vector_size(0);
1023 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1024 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1025 set_decompile_count(0);
1026
1027 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1028 _loop_opts_cnt = LoopOptsCount;
1029 _has_flattened_accesses = false;
1030 _flattened_accesses_share_alias = true;
1031 _scalarize_in_safepoints = false;
1032
1033 set_do_inlining(Inline);
1034 set_max_inline_size(MaxInlineSize);
1035 set_freq_inline_size(FreqInlineSize);
1036 set_do_scheduling(OptoScheduling);
1037
1038 set_do_vector_loop(false);
1039 set_has_monitors(false);
1040
1041 if (AllowVectorizeOnDemand) {
1042 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1043 set_do_vector_loop(true);
1044 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());})
1045 } else if (has_method() && method()->name() != 0 &&
1046 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1047 set_do_vector_loop(true);
1048 }
1049 }
1050 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1051 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());})
1052
1053 set_rtm_state(NoRTM); // No RTM lock eliding by default
1054 _max_node_limit = _directive->MaxNodeLimitOption;
1055
1056 #if INCLUDE_RTM_OPT
1057 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1058 int rtm_state = method()->method_data()->rtm_state();
1059 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1060 // Don't generate RTM lock eliding code.
1061 set_rtm_state(NoRTM);
1062 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1063 // Generate RTM lock eliding code without abort ratio calculation code.
1064 set_rtm_state(UseRTM);
1065 } else if (UseRTMDeopt) {
1066 // Generate RTM lock eliding code and include abort ratio calculation
1067 // code if UseRTMDeopt is on.
1068 set_rtm_state(ProfileRTM);
1069 }
1070 }
1071 #endif
1072 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1073 set_clinit_barrier_on_entry(true);
1074 }
1075 if (debug_info()->recording_non_safepoints()) {
1076 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1077 (comp_arena(), 8, 0, NULL));
1078 set_default_node_notes(Node_Notes::make(this));
1079 }
1080
1081 // // -- Initialize types before each compile --
1082 // // Update cached type information
1083 // if( _method && _method->constants() )
1084 // Type::update_loaded_types(_method, _method->constants());
1085
1086 // Init alias_type map.
1087 if (!do_escape_analysis() && aliaslevel == 3) {
1088 aliaslevel = 2; // No unique types without escape analysis
1089 }
1090 _AliasLevel = aliaslevel;
1091 const int grow_ats = 16;
1092 _max_alias_types = grow_ats;
1093 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1094 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1095 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1096 {
1097 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1098 }
1099 // Initialize the first few types.
1100 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1101 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1102 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1103 _num_alias_types = AliasIdxRaw+1;
1104 // Zero out the alias type cache.
1105 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1106 // A NULL adr_type hits in the cache right away. Preload the right answer.
1107 probe_alias_cache(NULL)->_index = AliasIdxTop;
1108
1109 #ifdef ASSERT
1110 _type_verify_symmetry = true;
1111 _phase_optimize_finished = false;
1112 _exception_backedge = false;
1113 #endif
1114 }
1115
1116 //---------------------------init_start----------------------------------------
1117 // Install the StartNode on this compile object.
1118 void Compile::init_start(StartNode* s) {
1119 if (failing())
1120 return; // already failing
1121 assert(s == start(), "");
1122 }
1123
1124 /**
1125 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1126 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1127 * the ideal graph.
1128 */
1129 StartNode* Compile::start() const {
1130 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1131 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1132 Node* start = root()->fast_out(i);
1133 if (start->is_Start()) {
1134 return start->as_Start();
1135 }
1136 }
1137 fatal("Did not find Start node!");
1138 return NULL;
1139 }
1140
1141 //-------------------------------immutable_memory-------------------------------------
1142 // Access immutable memory
1143 Node* Compile::immutable_memory() {
1144 if (_immutable_memory != NULL) {
1145 return _immutable_memory;
1146 }
1147 StartNode* s = start();
1148 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1149 Node *p = s->fast_out(i);
1150 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1151 _immutable_memory = p;
1152 return _immutable_memory;
1153 }
1154 }
1155 ShouldNotReachHere();
1156 return NULL;
1157 }
1158
1159 //----------------------set_cached_top_node------------------------------------
1160 // Install the cached top node, and make sure Node::is_top works correctly.
1161 void Compile::set_cached_top_node(Node* tn) {
1162 if (tn != NULL) verify_top(tn);
1163 Node* old_top = _top;
1164 _top = tn;
1165 // Calling Node::setup_is_top allows the nodes the chance to adjust
1166 // their _out arrays.
1167 if (_top != NULL) _top->setup_is_top();
1168 if (old_top != NULL) old_top->setup_is_top();
1169 assert(_top == NULL || top()->is_top(), "");
1170 }
1171
1172 #ifdef ASSERT
1173 uint Compile::count_live_nodes_by_graph_walk() {
1174 Unique_Node_List useful(comp_arena());
1175 // Get useful node list by walking the graph.
1176 identify_useful_nodes(useful);
1177 return useful.size();
1178 }
1179
1180 void Compile::print_missing_nodes() {
1181
1182 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1183 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1184 return;
1185 }
1186
1187 // This is an expensive function. It is executed only when the user
1188 // specifies VerifyIdealNodeCount option or otherwise knows the
1189 // additional work that needs to be done to identify reachable nodes
1190 // by walking the flow graph and find the missing ones using
1191 // _dead_node_list.
1192
1193 Unique_Node_List useful(comp_arena());
1194 // Get useful node list by walking the graph.
1195 identify_useful_nodes(useful);
1196
1197 uint l_nodes = C->live_nodes();
1198 uint l_nodes_by_walk = useful.size();
1199
1200 if (l_nodes != l_nodes_by_walk) {
1201 if (_log != NULL) {
1202 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1203 _log->stamp();
1204 _log->end_head();
1205 }
1206 VectorSet& useful_member_set = useful.member_set();
1207 int last_idx = l_nodes_by_walk;
1208 for (int i = 0; i < last_idx; i++) {
1209 if (useful_member_set.test(i)) {
1210 if (_dead_node_list.test(i)) {
1211 if (_log != NULL) {
1212 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1213 }
1214 if (PrintIdealNodeCount) {
1215 // Print the log message to tty
1216 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1217 useful.at(i)->dump();
1218 }
1219 }
1220 }
1221 else if (! _dead_node_list.test(i)) {
1222 if (_log != NULL) {
1223 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1224 }
1225 if (PrintIdealNodeCount) {
1226 // Print the log message to tty
1227 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1228 }
1229 }
1230 }
1231 if (_log != NULL) {
1232 _log->tail("mismatched_nodes");
1233 }
1234 }
1235 }
1236 void Compile::record_modified_node(Node* n) {
1237 if (_modified_nodes != NULL && !_inlining_incrementally && !n->is_Con()) {
1238 _modified_nodes->push(n);
1239 }
1240 }
1241
1242 void Compile::remove_modified_node(Node* n) {
1243 if (_modified_nodes != NULL) {
1244 _modified_nodes->remove(n);
1245 }
1246 }
1247 #endif
1248
1249 #ifndef PRODUCT
1250 void Compile::verify_top(Node* tn) const {
1251 if (tn != NULL) {
1252 assert(tn->is_Con(), "top node must be a constant");
1253 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1254 assert(tn->in(0) != NULL, "must have live top node");
1255 }
1256 }
1257 #endif
1258
1259
1260 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1261
1262 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1263 guarantee(arr != NULL, "");
1264 int num_blocks = arr->length();
1265 if (grow_by < num_blocks) grow_by = num_blocks;
1266 int num_notes = grow_by * _node_notes_block_size;
1267 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1268 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1269 while (num_notes > 0) {
1270 arr->append(notes);
1271 notes += _node_notes_block_size;
1272 num_notes -= _node_notes_block_size;
1273 }
1274 assert(num_notes == 0, "exact multiple, please");
1275 }
1276
1277 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1278 if (source == NULL || dest == NULL) return false;
1279
1280 if (dest->is_Con())
1281 return false; // Do not push debug info onto constants.
1282
1283 #ifdef ASSERT
1284 // Leave a bread crumb trail pointing to the original node:
1285 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1286 dest->set_debug_orig(source);
1287 }
1288 #endif
1289
1290 if (node_note_array() == NULL)
1291 return false; // Not collecting any notes now.
1292
1293 // This is a copy onto a pre-existing node, which may already have notes.
1294 // If both nodes have notes, do not overwrite any pre-existing notes.
1295 Node_Notes* source_notes = node_notes_at(source->_idx);
1296 if (source_notes == NULL || source_notes->is_clear()) return false;
1297 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1298 if (dest_notes == NULL || dest_notes->is_clear()) {
1299 return set_node_notes_at(dest->_idx, source_notes);
1300 }
1301
1302 Node_Notes merged_notes = (*source_notes);
1303 // The order of operations here ensures that dest notes will win...
1304 merged_notes.update_from(dest_notes);
1305 return set_node_notes_at(dest->_idx, &merged_notes);
1306 }
1307
1308
1309 //--------------------------allow_range_check_smearing-------------------------
1310 // Gating condition for coalescing similar range checks.
1311 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1312 // single covering check that is at least as strong as any of them.
1313 // If the optimization succeeds, the simplified (strengthened) range check
1314 // will always succeed. If it fails, we will deopt, and then give up
1315 // on the optimization.
1316 bool Compile::allow_range_check_smearing() const {
1317 // If this method has already thrown a range-check,
1318 // assume it was because we already tried range smearing
1319 // and it failed.
1320 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1321 return !already_trapped;
1322 }
1323
1324
1325 //------------------------------flatten_alias_type-----------------------------
1326 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1327 int offset = tj->offset();
1328 TypePtr::PTR ptr = tj->ptr();
1329
1330 // Known instance (scalarizable allocation) alias only with itself.
1331 bool is_known_inst = tj->isa_oopptr() != NULL &&
1332 tj->is_oopptr()->is_known_instance();
1333
1334 // Process weird unsafe references.
1335 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1336 bool default_value_load = EnableValhalla && tj->is_instptr()->instance_klass() == ciEnv::current()->Class_klass();
1337 assert(InlineUnsafeOps || StressReflectiveCode || default_value_load, "indeterminate pointers come only from unsafe ops");
1338 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1339 tj = TypeOopPtr::BOTTOM;
1340 ptr = tj->ptr();
1341 offset = tj->offset();
1342 }
1343
1344 // Array pointers need some flattening
1345 const TypeAryPtr* ta = tj->isa_aryptr();
1346 if (ta && ta->is_stable()) {
1347 // Erase stability property for alias analysis.
1348 tj = ta = ta->cast_to_stable(false);
1349 }
1350 if (ta && ta->is_not_flat()) {
1351 // Erase not flat property for alias analysis.
1352 tj = ta = ta->cast_to_not_flat(false);
1353 }
1354 if (ta && ta->is_not_null_free()) {
1355 // Erase not null free property for alias analysis.
1356 tj = ta = ta->cast_to_not_null_free(false);
1357 }
1358
1359 if( ta && is_known_inst ) {
1360 if ( offset != Type::OffsetBot &&
1361 offset > arrayOopDesc::length_offset_in_bytes() ) {
1362 offset = Type::OffsetBot; // Flatten constant access into array body only
1363 tj = ta = ta->
1364 remove_speculative()->
1365 cast_to_ptr_type(ptr)->
1366 with_offset(offset);
1367 }
1368 } else if( ta && _AliasLevel >= 2 ) {
1369 // For arrays indexed by constant indices, we flatten the alias
1370 // space to include all of the array body. Only the header, klass
1371 // and array length can be accessed un-aliased.
1372 // For flattened inline type array, each field has its own slice so
1373 // we must include the field offset.
1374 if( offset != Type::OffsetBot ) {
1375 if( ta->const_oop() ) { // MethodData* or Method*
1376 offset = Type::OffsetBot; // Flatten constant access into array body
1377 tj = ta = ta->
1378 remove_speculative()->
1379 cast_to_ptr_type(ptr)->
1380 cast_to_exactness(false)->
1381 with_offset(offset);
1382 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1383 // range is OK as-is.
1384 tj = ta = TypeAryPtr::RANGE;
1385 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1386 tj = TypeInstPtr::KLASS; // all klass loads look alike
1387 ta = TypeAryPtr::RANGE; // generic ignored junk
1388 ptr = TypePtr::BotPTR;
1389 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1390 tj = TypeInstPtr::MARK;
1391 ta = TypeAryPtr::RANGE; // generic ignored junk
1392 ptr = TypePtr::BotPTR;
1393 } else { // Random constant offset into array body
1394 offset = Type::OffsetBot; // Flatten constant access into array body
1395 tj = ta = ta->
1396 remove_speculative()->
1397 cast_to_ptr_type(ptr)->
1398 cast_to_exactness(false)->
1399 with_offset(offset);
1400 }
1401 }
1402 // Arrays of fixed size alias with arrays of unknown size.
1403 if (ta->size() != TypeInt::POS) {
1404 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1405 tj = ta = ta->
1406 remove_speculative()->
1407 cast_to_ptr_type(ptr)->
1408 with_ary(tary)->
1409 cast_to_exactness(false);
1410 }
1411 // Arrays of known objects become arrays of unknown objects.
1412 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1413 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1414 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1415 }
1416 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1417 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1418 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1419 }
1420 // Initially all flattened array accesses share a single slice
1421 if (ta->is_flat() && ta->elem() != TypeInlineType::BOTTOM && _flattened_accesses_share_alias) {
1422 const TypeAry *tary = TypeAry::make(TypeInlineType::BOTTOM, ta->size());
1423 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
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,Type::Offset(offset), ta->field_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() != NULL) {
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 && _AliasLevel >= 2 && 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() != NULL) {
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, NULL, Type::Offset(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 = NULL;
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 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1494 if( is_known_inst ) {
1495 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), canonical_holder->flatten_array(), to->instance_id());
1496 } else {
1497 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset));
1498 }
1499 }
1500 }
1501 }
1502
1503 // Klass pointers to object array klasses need some flattening
1504 const TypeKlassPtr *tk = tj->isa_klassptr();
1505 if( tk ) {
1506 // If we are referencing a field within a Klass, we need
1507 // to assume the worst case of an Object. Both exact and
1508 // inexact types must flatten to the same alias class so
1509 // use NotNull as the PTR.
1510 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1511 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1512 env()->Object_klass(),
1513 Type::Offset(offset));
1514 }
1515
1516 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1517 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1518 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1519 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset));
1520 } else {
1521 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, Type::Offset(offset), tk->is_not_flat(), tk->is_not_null_free(), tk->is_null_free());
1522 }
1523 }
1524
1525 // Check for precise loads from the primary supertype array and force them
1526 // to the supertype cache alias index. Check for generic array loads from
1527 // the primary supertype array and also force them to the supertype cache
1528 // alias index. Since the same load can reach both, we need to merge
1529 // these 2 disparate memories into the same alias class. Since the
1530 // primary supertype array is read-only, there's no chance of confusion
1531 // where we bypass an array load and an array store.
1532 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1533 if (offset == Type::OffsetBot ||
1534 (offset >= primary_supers_offset &&
1535 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1536 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1537 offset = in_bytes(Klass::secondary_super_cache_offset());
1538 tj = tk = tk->with_offset(offset);
1539 }
1540 }
1541
1542 // Flatten all Raw pointers together.
1543 if (tj->base() == Type::RawPtr)
1544 tj = TypeRawPtr::BOTTOM;
1545
1546 if (tj->base() == Type::AnyPtr)
1547 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1548
1549 // Flatten all to bottom for now
1550 switch( _AliasLevel ) {
1551 case 0:
1552 tj = TypePtr::BOTTOM;
1553 break;
1554 case 1: // Flatten to: oop, static, field or array
1555 switch (tj->base()) {
1556 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1557 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1558 case Type::AryPtr: // do not distinguish arrays at all
1559 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1560 case Type::KlassPtr:
1561 case Type::AryKlassPtr:
1562 case Type::InstKlassPtr: tj = TypeInstKlassPtr::OBJECT; break;
1563 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1564 default: ShouldNotReachHere();
1565 }
1566 break;
1567 case 2: // No collapsing at level 2; keep all splits
1568 case 3: // No collapsing at level 3; keep all splits
1569 break;
1570 default:
1571 Unimplemented();
1572 }
1573
1574 offset = tj->offset();
1575 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1576
1577 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1578 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1579 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1580 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1581 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1582 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1583 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1584 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1585 assert( tj->ptr() != TypePtr::TopPTR &&
1586 tj->ptr() != TypePtr::AnyNull &&
1587 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1588 // assert( tj->ptr() != TypePtr::Constant ||
1589 // tj->base() == Type::RawPtr ||
1590 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1591
1592 return tj;
1593 }
1594
1595 void Compile::AliasType::Init(int i, const TypePtr* at) {
1596 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1597 _index = i;
1598 _adr_type = at;
1599 _field = NULL;
1600 _element = NULL;
1601 _is_rewritable = true; // default
1602 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1603 if (atoop != NULL && atoop->is_known_instance()) {
1604 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1605 _general_index = Compile::current()->get_alias_index(gt);
1606 } else {
1607 _general_index = 0;
1608 }
1609 }
1610
1611 BasicType Compile::AliasType::basic_type() const {
1612 if (element() != NULL) {
1613 const Type* element = adr_type()->is_aryptr()->elem();
1614 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1615 } if (field() != NULL) {
1616 return field()->layout_type();
1617 } else {
1618 return T_ILLEGAL; // unknown
1619 }
1620 }
1621
1622 //---------------------------------print_on------------------------------------
1623 #ifndef PRODUCT
1624 void Compile::AliasType::print_on(outputStream* st) {
1625 if (index() < 10)
1626 st->print("@ <%d> ", index());
1627 else st->print("@ <%d>", index());
1628 st->print(is_rewritable() ? " " : " RO");
1629 int offset = adr_type()->offset();
1630 if (offset == Type::OffsetBot)
1631 st->print(" +any");
1632 else st->print(" +%-3d", offset);
1633 st->print(" in ");
1634 adr_type()->dump_on(st);
1635 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1636 if (field() != NULL && tjp) {
1637 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1638 tjp->offset() != field()->offset_in_bytes()) {
1639 st->print(" != ");
1640 field()->print();
1641 st->print(" ***");
1642 }
1643 }
1644 }
1645
1646 void print_alias_types() {
1647 Compile* C = Compile::current();
1648 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1649 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1650 C->alias_type(idx)->print_on(tty);
1651 tty->cr();
1652 }
1653 }
1654 #endif
1655
1656
1657 //----------------------------probe_alias_cache--------------------------------
1658 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1659 intptr_t key = (intptr_t) adr_type;
1660 key ^= key >> logAliasCacheSize;
1661 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1662 }
1663
1664
1665 //-----------------------------grow_alias_types--------------------------------
1666 void Compile::grow_alias_types() {
1667 const int old_ats = _max_alias_types; // how many before?
1668 const int new_ats = old_ats; // how many more?
1669 const int grow_ats = old_ats+new_ats; // how many now?
1670 _max_alias_types = grow_ats;
1671 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1672 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1673 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1674 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1675 }
1676
1677
1678 //--------------------------------find_alias_type------------------------------
1679 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1680 if (_AliasLevel == 0)
1681 return alias_type(AliasIdxBot);
1682
1683 AliasCacheEntry* ace = NULL;
1684 if (!uncached) {
1685 ace = probe_alias_cache(adr_type);
1686 if (ace->_adr_type == adr_type) {
1687 return alias_type(ace->_index);
1688 }
1689 }
1690
1691 // Handle special cases.
1692 if (adr_type == NULL) return alias_type(AliasIdxTop);
1693 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1694
1695 // Do it the slow way.
1696 const TypePtr* flat = flatten_alias_type(adr_type);
1697
1698 #ifdef ASSERT
1699 {
1700 ResourceMark rm;
1701 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1702 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1703 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1704 Type::str(adr_type));
1705 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1706 const TypeOopPtr* foop = flat->is_oopptr();
1707 // Scalarizable allocations have exact klass always.
1708 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1709 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1710 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1711 Type::str(foop), Type::str(xoop));
1712 }
1713 }
1714 #endif
1715
1716 int idx = AliasIdxTop;
1717 for (int i = 0; i < num_alias_types(); i++) {
1718 if (alias_type(i)->adr_type() == flat) {
1719 idx = i;
1720 break;
1721 }
1722 }
1723
1724 if (idx == AliasIdxTop) {
1725 if (no_create) return NULL;
1726 // Grow the array if necessary.
1727 if (_num_alias_types == _max_alias_types) grow_alias_types();
1728 // Add a new alias type.
1729 idx = _num_alias_types++;
1730 _alias_types[idx]->Init(idx, flat);
1731 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1732 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1733 if (flat->isa_instptr()) {
1734 if (flat->offset() == java_lang_Class::klass_offset()
1735 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1736 alias_type(idx)->set_rewritable(false);
1737 }
1738 ciField* field = NULL;
1739 if (flat->isa_aryptr()) {
1740 #ifdef ASSERT
1741 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1742 // (T_BYTE has the weakest alignment and size restrictions...)
1743 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1744 #endif
1745 const Type* elemtype = flat->is_aryptr()->elem();
1746 if (flat->offset() == TypePtr::OffsetBot) {
1747 alias_type(idx)->set_element(elemtype);
1748 }
1749 int field_offset = flat->is_aryptr()->field_offset().get();
1750 if (elemtype->isa_inlinetype() &&
1751 field_offset != Type::OffsetBot) {
1752 ciInlineKlass* vk = elemtype->inline_klass();
1753 field_offset += vk->first_field_offset();
1754 field = vk->get_field_by_offset(field_offset, false);
1755 }
1756 }
1757 if (flat->isa_klassptr()) {
1758 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1759 alias_type(idx)->set_rewritable(false);
1760 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1761 alias_type(idx)->set_rewritable(false);
1762 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1763 alias_type(idx)->set_rewritable(false);
1764 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1765 alias_type(idx)->set_rewritable(false);
1766 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1767 alias_type(idx)->set_rewritable(false);
1768 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1769 alias_type(idx)->set_rewritable(false);
1770 }
1771 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1772 // but the base pointer type is not distinctive enough to identify
1773 // references into JavaThread.)
1774
1775 // Check for final fields.
1776 const TypeInstPtr* tinst = flat->isa_instptr();
1777 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1778 if (tinst->const_oop() != NULL &&
1779 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1780 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1781 // static field
1782 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1783 field = k->get_field_by_offset(tinst->offset(), true);
1784 } else if (tinst->is_inlinetypeptr()) {
1785 // Inline type field
1786 ciInlineKlass* vk = tinst->inline_klass();
1787 field = vk->get_field_by_offset(tinst->offset(), false);
1788 } else {
1789 ciInstanceKlass *k = tinst->instance_klass();
1790 field = k->get_field_by_offset(tinst->offset(), false);
1791 }
1792 }
1793 assert(field == NULL ||
1794 original_field == NULL ||
1795 (field->holder() == original_field->holder() &&
1796 field->offset() == original_field->offset() &&
1797 field->is_static() == original_field->is_static()), "wrong field?");
1798 // Set field() and is_rewritable() attributes.
1799 if (field != NULL) {
1800 alias_type(idx)->set_field(field);
1801 if (flat->isa_aryptr()) {
1802 // Fields of flat arrays are rewritable although they are declared final
1803 assert(flat->is_aryptr()->is_flat(), "must be a flat array");
1804 alias_type(idx)->set_rewritable(true);
1805 }
1806 }
1807 }
1808
1809 // Fill the cache for next time.
1810 if (!uncached) {
1811 ace->_adr_type = adr_type;
1812 ace->_index = idx;
1813 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1814
1815 // Might as well try to fill the cache for the flattened version, too.
1816 AliasCacheEntry* face = probe_alias_cache(flat);
1817 if (face->_adr_type == NULL) {
1818 face->_adr_type = flat;
1819 face->_index = idx;
1820 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1821 }
1822 }
1823
1824 return alias_type(idx);
1825 }
1826
1827
1828 Compile::AliasType* Compile::alias_type(ciField* field) {
1829 const TypeOopPtr* t;
1830 if (field->is_static())
1831 t = TypeInstPtr::make(field->holder()->java_mirror());
1832 else
1833 t = TypeOopPtr::make_from_klass_raw(field->holder());
1834 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1835 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1836 return atp;
1837 }
1838
1839
1840 //------------------------------have_alias_type--------------------------------
1841 bool Compile::have_alias_type(const TypePtr* adr_type) {
1842 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1843 if (ace->_adr_type == adr_type) {
1844 return true;
1845 }
1846
1847 // Handle special cases.
1848 if (adr_type == NULL) return true;
1849 if (adr_type == TypePtr::BOTTOM) return true;
1850
1851 return find_alias_type(adr_type, true, NULL) != NULL;
1852 }
1853
1854 //-----------------------------must_alias--------------------------------------
1855 // True if all values of the given address type are in the given alias category.
1856 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1857 if (alias_idx == AliasIdxBot) return true; // the universal category
1858 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1859 if (alias_idx == AliasIdxTop) return false; // the empty category
1860 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1861
1862 // the only remaining possible overlap is identity
1863 int adr_idx = get_alias_index(adr_type);
1864 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1865 assert(adr_idx == alias_idx ||
1866 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1867 && adr_type != TypeOopPtr::BOTTOM),
1868 "should not be testing for overlap with an unsafe pointer");
1869 return adr_idx == alias_idx;
1870 }
1871
1872 //------------------------------can_alias--------------------------------------
1873 // True if any values of the given address type are in the given alias category.
1874 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1875 if (alias_idx == AliasIdxTop) return false; // the empty category
1876 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1877 // Known instance doesn't alias with bottom memory
1878 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1879 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1880
1881 // the only remaining possible overlap is identity
1882 int adr_idx = get_alias_index(adr_type);
1883 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1884 return adr_idx == alias_idx;
1885 }
1886
1887 //---------------------cleanup_loop_predicates-----------------------
1888 // Remove the opaque nodes that protect the predicates so that all unused
1889 // checks and uncommon_traps will be eliminated from the ideal graph
1890 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1891 if (predicate_count()==0) return;
1892 for (int i = predicate_count(); i > 0; i--) {
1893 Node * n = predicate_opaque1_node(i-1);
1894 assert(n->Opcode() == Op_Opaque1, "must be");
1895 igvn.replace_node(n, n->in(1));
1896 }
1897 assert(predicate_count()==0, "should be clean!");
1898 }
1899
1900 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1901 if (!n->for_post_loop_opts_igvn()) {
1902 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1903 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1904 _for_post_loop_igvn.append(n);
1905 }
1906 }
1907
1908 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1909 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1910 _for_post_loop_igvn.remove(n);
1911 }
1912
1913 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1914 // Verify that all previous optimizations produced a valid graph
1915 // at least to this point, even if no loop optimizations were done.
1916 PhaseIdealLoop::verify(igvn);
1917
1918 C->set_post_loop_opts_phase(); // no more loop opts allowed
1919
1920 assert(!C->major_progress(), "not cleared");
1921
1922 if (_for_post_loop_igvn.length() > 0) {
1923 while (_for_post_loop_igvn.length() > 0) {
1924 Node* n = _for_post_loop_igvn.pop();
1925 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1926 igvn._worklist.push(n);
1927 }
1928 igvn.optimize();
1929 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1930
1931 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1932 if (C->major_progress()) {
1933 C->clear_major_progress(); // ensure that major progress is now clear
1934 }
1935 }
1936 }
1937
1938 void Compile::add_inline_type(Node* n) {
1939 assert(n->is_InlineTypeBase(), "unexpected node");
1940 _inline_type_nodes.push(n);
1941 }
1942
1943 void Compile::remove_inline_type(Node* n) {
1944 assert(n->is_InlineTypeBase(), "unexpected node");
1945 if (_inline_type_nodes.contains(n)) {
1946 _inline_type_nodes.remove(n);
1947 }
1948 }
1949
1950 // Does the return value keep otherwise useless inline type allocations alive?
1951 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1952 ResourceMark rm;
1953 Unique_Node_List wq;
1954 wq.push(ret_val);
1955 bool some_allocations = false;
1956 for (uint i = 0; i < wq.size(); i++) {
1957 Node* n = wq.at(i);
1958 assert(!n->is_InlineType(), "chain of inline type nodes");
1959 if (n->outcnt() > 1) {
1960 // Some other use for the allocation
1961 return false;
1962 } else if (n->is_InlineTypePtr()) {
1963 wq.push(n->in(1));
1964 } else if (n->is_Phi()) {
1965 for (uint j = 1; j < n->req(); j++) {
1966 wq.push(n->in(j));
1967 }
1968 } else if (n->is_CheckCastPP() &&
1969 n->in(1)->is_Proj() &&
1970 n->in(1)->in(0)->is_Allocate()) {
1971 some_allocations = true;
1972 } else if (n->is_CheckCastPP()) {
1973 wq.push(n->in(1));
1974 }
1975 }
1976 return some_allocations;
1977 }
1978
1979 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1980 // Make sure that the return value does not keep an otherwise unused allocation alive
1981 if (tf()->returns_inline_type_as_fields()) {
1982 Node* ret = NULL;
1983 for (uint i = 1; i < root()->req(); i++) {
1984 Node* in = root()->in(i);
1985 if (in->Opcode() == Op_Return) {
1986 assert(ret == NULL, "only one return");
1987 ret = in;
1988 }
1989 }
1990 if (ret != NULL) {
1991 Node* ret_val = ret->in(TypeFunc::Parms);
1992 if (igvn.type(ret_val)->isa_oopptr() &&
1993 return_val_keeps_allocations_alive(ret_val)) {
1994 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
1995 assert(ret_val->outcnt() == 0, "should be dead now");
1996 igvn.remove_dead_node(ret_val);
1997 }
1998 }
1999 }
2000 if (_inline_type_nodes.length() == 0) {
2001 return;
2002 }
2003 // Scalarize inline types in safepoint debug info.
2004 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2005 set_scalarize_in_safepoints(true);
2006 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2007 _inline_type_nodes.at(i)->as_InlineTypeBase()->make_scalar_in_safepoints(&igvn);
2008 }
2009 if (remove) {
2010 // Remove inline type nodes
2011 while (_inline_type_nodes.length() > 0) {
2012 InlineTypeBaseNode* vt = _inline_type_nodes.pop()->as_InlineTypeBase();
2013 if (vt->outcnt() == 0) {
2014 igvn.remove_dead_node(vt);
2015 } else if (vt->is_InlineTypePtr()) {
2016 igvn.replace_node(vt, vt->get_oop());
2017 } else {
2018 // Check if any users are blackholes. If so, rewrite them to use either the
2019 // allocated buffer, or individual components, instead of the inline type node
2020 // that goes away.
2021 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2022 if (vt->out(i)->is_Blackhole()) {
2023 BlackholeNode* bh = vt->out(i)->as_Blackhole();
2024
2025 // Unlink the old input
2026 int idx = bh->find_edge(vt);
2027 assert(idx != -1, "The edge should be there");
2028 bh->del_req(idx);
2029 --i;
2030
2031 if (vt->is_allocated(&igvn)) {
2032 // Already has the allocated instance, blackhole that
2033 bh->add_req(vt->get_oop());
2034 } else {
2035 // Not allocated yet, blackhole the components
2036 for (uint c = 0; c < vt->field_count(); c++) {
2037 bh->add_req(vt->field_value(c));
2038 }
2039 }
2040
2041 // Node modified, record for IGVN
2042 igvn.record_for_igvn(bh);
2043 }
2044 }
2045
2046 #ifdef ASSERT
2047 for (DUIterator_Fast imax, i = vt->fast_outs(imax); i < imax; i++) {
2048 assert(vt->fast_out(i)->is_InlineTypeBase(), "Unexpected inline type user");
2049 }
2050 #endif
2051 igvn.replace_node(vt, igvn.C->top());
2052 }
2053 }
2054 }
2055 igvn.optimize();
2056 }
2057
2058 void Compile::adjust_flattened_array_access_aliases(PhaseIterGVN& igvn) {
2059 if (!_has_flattened_accesses) {
2060 return;
2061 }
2062 // Initially, all flattened array accesses share the same slice to
2063 // keep dependencies with Object[] array accesses (that could be
2064 // to a flattened array) correct. We're done with parsing so we
2065 // now know all flattened array accesses in this compile
2066 // unit. Let's move flattened array accesses to their own slice,
2067 // one per element field. This should help memory access
2068 // optimizations.
2069 ResourceMark rm;
2070 Unique_Node_List wq;
2071 wq.push(root());
2072
2073 Node_List mergememnodes;
2074 Node_List memnodes;
2075
2076 // Alias index currently shared by all flattened memory accesses
2077 int index = get_alias_index(TypeAryPtr::INLINES);
2078
2079 // Find MergeMem nodes and flattened array accesses
2080 for (uint i = 0; i < wq.size(); i++) {
2081 Node* n = wq.at(i);
2082 if (n->is_Mem()) {
2083 const TypePtr* adr_type = NULL;
2084 if (n->Opcode() == Op_StoreCM) {
2085 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type()));
2086 } else {
2087 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2088 }
2089 if (adr_type == TypeAryPtr::INLINES) {
2090 memnodes.push(n);
2091 }
2092 } else if (n->is_MergeMem()) {
2093 MergeMemNode* mm = n->as_MergeMem();
2094 if (mm->memory_at(index) != mm->base_memory()) {
2095 mergememnodes.push(n);
2096 }
2097 }
2098 for (uint j = 0; j < n->req(); j++) {
2099 Node* m = n->in(j);
2100 if (m != NULL) {
2101 wq.push(m);
2102 }
2103 }
2104 }
2105
2106 if (memnodes.size() > 0) {
2107 _flattened_accesses_share_alias = false;
2108
2109 // We are going to change the slice for the flattened array
2110 // accesses so we need to clear the cache entries that refer to
2111 // them.
2112 for (uint i = 0; i < AliasCacheSize; i++) {
2113 AliasCacheEntry* ace = &_alias_cache[i];
2114 if (ace->_adr_type != NULL &&
2115 ace->_adr_type->isa_aryptr() &&
2116 ace->_adr_type->is_aryptr()->is_flat()) {
2117 ace->_adr_type = NULL;
2118 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the NULL adr_type resolves to AliasIdxTop
2119 }
2120 }
2121
2122 // Find what aliases we are going to add
2123 int start_alias = num_alias_types()-1;
2124 int stop_alias = 0;
2125
2126 for (uint i = 0; i < memnodes.size(); i++) {
2127 Node* m = memnodes.at(i);
2128 const TypePtr* adr_type = NULL;
2129 if (m->Opcode() == Op_StoreCM) {
2130 adr_type = m->in(MemNode::OopStore)->adr_type();
2131 if (adr_type != TypeAryPtr::INLINES) {
2132 // store was optimized out and we lost track of the adr_type
2133 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address),
2134 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore),
2135 get_alias_index(adr_type));
2136 igvn.register_new_node_with_optimizer(clone);
2137 igvn.replace_node(m, clone);
2138 }
2139 } else {
2140 adr_type = m->adr_type();
2141 #ifdef ASSERT
2142 m->as_Mem()->set_adr_type(adr_type);
2143 #endif
2144 }
2145 int idx = get_alias_index(adr_type);
2146 start_alias = MIN2(start_alias, idx);
2147 stop_alias = MAX2(stop_alias, idx);
2148 }
2149
2150 assert(stop_alias >= start_alias, "should have expanded aliases");
2151
2152 Node_Stack stack(0);
2153 #ifdef ASSERT
2154 VectorSet seen(Thread::current()->resource_area());
2155 #endif
2156 // Now let's fix the memory graph so each flattened array access
2157 // is moved to the right slice. Start from the MergeMem nodes.
2158 uint last = unique();
2159 for (uint i = 0; i < mergememnodes.size(); i++) {
2160 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2161 Node* n = current->memory_at(index);
2162 MergeMemNode* mm = NULL;
2163 do {
2164 // Follow memory edges through memory accesses, phis and
2165 // narrow membars and push nodes on the stack. Once we hit
2166 // bottom memory, we pop element off the stack one at a
2167 // time, in reverse order, and move them to the right slice
2168 // by changing their memory edges.
2169 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2170 assert(!seen.test_set(n->_idx), "");
2171 // Uses (a load for instance) will need to be moved to the
2172 // right slice as well and will get a new memory state
2173 // that we don't know yet. The use could also be the
2174 // backedge of a loop. We put a place holder node between
2175 // the memory node and its uses. We replace that place
2176 // holder with the correct memory state once we know it,
2177 // i.e. when nodes are popped off the stack. Using the
2178 // place holder make the logic work in the presence of
2179 // loops.
2180 if (n->outcnt() > 1) {
2181 Node* place_holder = NULL;
2182 assert(!n->has_out_with(Op_Node), "");
2183 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2184 Node* u = n->out(k);
2185 if (u != current && u->_idx < last) {
2186 bool success = false;
2187 for (uint l = 0; l < u->req(); l++) {
2188 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2189 continue;
2190 }
2191 Node* in = u->in(l);
2192 if (in == n) {
2193 if (place_holder == NULL) {
2194 place_holder = new Node(1);
2195 place_holder->init_req(0, n);
2196 }
2197 igvn.replace_input_of(u, l, place_holder);
2198 success = true;
2199 }
2200 }
2201 if (success) {
2202 --k;
2203 }
2204 }
2205 }
2206 }
2207 if (n->is_Phi()) {
2208 stack.push(n, 1);
2209 n = n->in(1);
2210 } else if (n->is_Mem()) {
2211 stack.push(n, n->req());
2212 n = n->in(MemNode::Memory);
2213 } else {
2214 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2215 stack.push(n, n->req());
2216 n = n->in(0)->in(TypeFunc::Memory);
2217 }
2218 } else {
2219 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2220 // Build a new MergeMem node to carry the new memory state
2221 // as we build it. IGVN should fold extraneous MergeMem
2222 // nodes.
2223 mm = MergeMemNode::make(n);
2224 igvn.register_new_node_with_optimizer(mm);
2225 while (stack.size() > 0) {
2226 Node* m = stack.node();
2227 uint idx = stack.index();
2228 if (m->is_Mem()) {
2229 // Move memory node to its new slice
2230 const TypePtr* adr_type = m->adr_type();
2231 int alias = get_alias_index(adr_type);
2232 Node* prev = mm->memory_at(alias);
2233 igvn.replace_input_of(m, MemNode::Memory, prev);
2234 mm->set_memory_at(alias, m);
2235 } else if (m->is_Phi()) {
2236 // We need as many new phis as there are new aliases
2237 igvn.replace_input_of(m, idx, mm);
2238 if (idx == m->req()-1) {
2239 Node* r = m->in(0);
2240 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2241 const Type* adr_type = get_adr_type(j);
2242 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->is_flat() || j == (uint)index) {
2243 continue;
2244 }
2245 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2246 igvn.register_new_node_with_optimizer(phi);
2247 for (uint k = 1; k < m->req(); k++) {
2248 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2249 }
2250 mm->set_memory_at(j, phi);
2251 }
2252 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2253 igvn.register_new_node_with_optimizer(base_phi);
2254 for (uint k = 1; k < m->req(); k++) {
2255 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2256 }
2257 mm->set_base_memory(base_phi);
2258 }
2259 } else {
2260 // This is a MemBarCPUOrder node from
2261 // Parse::array_load()/Parse::array_store(), in the
2262 // branch that handles flattened arrays hidden under
2263 // an Object[] array. We also need one new membar per
2264 // new alias to keep the unknown access that the
2265 // membars protect properly ordered with accesses to
2266 // known flattened array.
2267 assert(m->is_Proj(), "projection expected");
2268 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2269 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2270 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2271 const Type* adr_type = get_adr_type(j);
2272 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->is_flat() || j == (uint)index) {
2273 continue;
2274 }
2275 MemBarNode* mb = new MemBarCPUOrderNode(this, j, NULL);
2276 igvn.register_new_node_with_optimizer(mb);
2277 Node* mem = mm->memory_at(j);
2278 mb->init_req(TypeFunc::Control, ctrl);
2279 mb->init_req(TypeFunc::Memory, mem);
2280 ctrl = new ProjNode(mb, TypeFunc::Control);
2281 igvn.register_new_node_with_optimizer(ctrl);
2282 mem = new ProjNode(mb, TypeFunc::Memory);
2283 igvn.register_new_node_with_optimizer(mem);
2284 mm->set_memory_at(j, mem);
2285 }
2286 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2287 }
2288 if (idx < m->req()-1) {
2289 idx += 1;
2290 stack.set_index(idx);
2291 n = m->in(idx);
2292 break;
2293 }
2294 // Take care of place holder nodes
2295 if (m->has_out_with(Op_Node)) {
2296 Node* place_holder = m->find_out_with(Op_Node);
2297 if (place_holder != NULL) {
2298 Node* mm_clone = mm->clone();
2299 igvn.register_new_node_with_optimizer(mm_clone);
2300 Node* hook = new Node(1);
2301 hook->init_req(0, mm);
2302 igvn.replace_node(place_holder, mm_clone);
2303 hook->destruct(&igvn);
2304 }
2305 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2306 }
2307 stack.pop();
2308 }
2309 }
2310 } while(stack.size() > 0);
2311 // Fix the memory state at the MergeMem we started from
2312 igvn.rehash_node_delayed(current);
2313 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2314 const Type* adr_type = get_adr_type(j);
2315 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->is_flat()) {
2316 continue;
2317 }
2318 current->set_memory_at(j, mm);
2319 }
2320 current->set_memory_at(index, current->base_memory());
2321 }
2322 igvn.optimize();
2323 }
2324 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2325 #ifdef ASSERT
2326 if (!_flattened_accesses_share_alias) {
2327 wq.clear();
2328 wq.push(root());
2329 for (uint i = 0; i < wq.size(); i++) {
2330 Node* n = wq.at(i);
2331 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2332 for (uint j = 0; j < n->req(); j++) {
2333 Node* m = n->in(j);
2334 if (m != NULL) {
2335 wq.push(m);
2336 }
2337 }
2338 }
2339 }
2340 #endif
2341 }
2342
2343 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2344 if (OptimizeUnstableIf) {
2345 _unstable_if_traps.append(trap);
2346 }
2347 }
2348
2349 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2350 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2351 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2352 Node* n = trap->uncommon_trap();
2353 if (!useful.member(n)) {
2354 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2355 }
2356 }
2357 }
2358
2359 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2360 // or fold-compares case. Return true if succeed or not found.
2361 //
2362 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2363 // false and ask fold-compares to yield.
2364 //
2365 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2366 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2367 // when deoptimization does happen.
2368 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2369 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2370 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2371 if (trap->uncommon_trap() == unc) {
2372 if (yield && trap->modified()) {
2373 return false;
2374 }
2375 _unstable_if_traps.delete_at(i);
2376 break;
2377 }
2378 }
2379 return true;
2380 }
2381
2382 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2383 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2384 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2385 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2386 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2387 CallStaticJavaNode* unc = trap->uncommon_trap();
2388 int next_bci = trap->next_bci();
2389 bool modified = trap->modified();
2390
2391 if (next_bci != -1 && !modified) {
2392 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2393 JVMState* jvms = unc->jvms();
2394 ciMethod* method = jvms->method();
2395 ciBytecodeStream iter(method);
2396
2397 iter.force_bci(jvms->bci());
2398 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2399 Bytecodes::Code c = iter.cur_bc();
2400 Node* lhs = nullptr;
2401 Node* rhs = nullptr;
2402 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2403 lhs = unc->peek_operand(0);
2404 rhs = unc->peek_operand(1);
2405 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2406 lhs = unc->peek_operand(0);
2407 }
2408
2409 ResourceMark rm;
2410 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2411 assert(live_locals.is_valid(), "broken liveness info");
2412 int len = (int)live_locals.size();
2413
2414 for (int i = 0; i < len; i++) {
2415 Node* local = unc->local(jvms, i);
2416 // kill local using the liveness of next_bci.
2417 // give up when the local looks like an operand to secure reexecution.
2418 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2419 uint idx = jvms->locoff() + i;
2420 #ifdef ASSERT
2421 if (Verbose) {
2422 tty->print("[unstable_if] kill local#%d: ", idx);
2423 local->dump();
2424 tty->cr();
2425 }
2426 #endif
2427 igvn.replace_input_of(unc, idx, top());
2428 modified = true;
2429 }
2430 }
2431 }
2432
2433 // keep the mondified trap for late query
2434 if (modified) {
2435 trap->set_modified();
2436 } else {
2437 _unstable_if_traps.delete_at(i);
2438 }
2439 }
2440 igvn.optimize();
2441 }
2442
2443 // StringOpts and late inlining of string methods
2444 void Compile::inline_string_calls(bool parse_time) {
2445 {
2446 // remove useless nodes to make the usage analysis simpler
2447 ResourceMark rm;
2448 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2449 }
2450
2451 {
2452 ResourceMark rm;
2453 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2454 PhaseStringOpts pso(initial_gvn());
2455 print_method(PHASE_AFTER_STRINGOPTS, 3);
2456 }
2457
2458 // now inline anything that we skipped the first time around
2459 if (!parse_time) {
2460 _late_inlines_pos = _late_inlines.length();
2461 }
2462
2463 while (_string_late_inlines.length() > 0) {
2464 CallGenerator* cg = _string_late_inlines.pop();
2465 cg->do_late_inline();
2466 if (failing()) return;
2467 }
2468 _string_late_inlines.trunc_to(0);
2469 }
2470
2471 // Late inlining of boxing methods
2472 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2473 if (_boxing_late_inlines.length() > 0) {
2474 assert(has_boxed_value(), "inconsistent");
2475
2476 PhaseGVN* gvn = initial_gvn();
2477 set_inlining_incrementally(true);
2478
2479 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2480 for_igvn()->clear();
2481 gvn->replace_with(&igvn);
2482
2483 _late_inlines_pos = _late_inlines.length();
2484
2485 while (_boxing_late_inlines.length() > 0) {
2486 CallGenerator* cg = _boxing_late_inlines.pop();
2487 cg->do_late_inline();
2488 if (failing()) return;
2489 }
2490 _boxing_late_inlines.trunc_to(0);
2491
2492 inline_incrementally_cleanup(igvn);
2493
2494 set_inlining_incrementally(false);
2495 }
2496 }
2497
2498 bool Compile::inline_incrementally_one() {
2499 assert(IncrementalInline, "incremental inlining should be on");
2500
2501 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2502
2503 set_inlining_progress(false);
2504 set_do_cleanup(false);
2505
2506 for (int i = 0; i < _late_inlines.length(); i++) {
2507 _late_inlines_pos = i+1;
2508 CallGenerator* cg = _late_inlines.at(i);
2509 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2510 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2511 cg->do_late_inline();
2512 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2513 if (failing()) {
2514 return false;
2515 } else if (inlining_progress()) {
2516 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2517 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2518 break; // process one call site at a time
2519 }
2520 } else {
2521 // Ignore late inline direct calls when inlining is not allowed.
2522 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2523 }
2524 }
2525 // Remove processed elements.
2526 _late_inlines.remove_till(_late_inlines_pos);
2527 _late_inlines_pos = 0;
2528
2529 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2530
2531 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2532
2533 set_inlining_progress(false);
2534 set_do_cleanup(false);
2535
2536 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2537 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2538 }
2539
2540 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2541 {
2542 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2543 ResourceMark rm;
2544 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2545 }
2546 {
2547 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2548 igvn = PhaseIterGVN(initial_gvn());
2549 igvn.optimize();
2550 }
2551 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2552 }
2553
2554 // Perform incremental inlining until bound on number of live nodes is reached
2555 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2556 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2557
2558 set_inlining_incrementally(true);
2559 uint low_live_nodes = 0;
2560
2561 while (_late_inlines.length() > 0) {
2562 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2563 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2564 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2565 // PhaseIdealLoop is expensive so we only try it once we are
2566 // out of live nodes and we only try it again if the previous
2567 // helped got the number of nodes down significantly
2568 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2569 if (failing()) return;
2570 low_live_nodes = live_nodes();
2571 _major_progress = true;
2572 }
2573
2574 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2575 bool do_print_inlining = print_inlining() || print_intrinsics();
2576 if (do_print_inlining || log() != NULL) {
2577 // Print inlining message for candidates that we couldn't inline for lack of space.
2578 for (int i = 0; i < _late_inlines.length(); i++) {
2579 CallGenerator* cg = _late_inlines.at(i);
2580 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2581 if (do_print_inlining) {
2582 cg->print_inlining_late(msg);
2583 }
2584 log_late_inline_failure(cg, msg);
2585 }
2586 }
2587 break; // finish
2588 }
2589 }
2590
2591 for_igvn()->clear();
2592 initial_gvn()->replace_with(&igvn);
2593
2594 while (inline_incrementally_one()) {
2595 assert(!failing(), "inconsistent");
2596 }
2597 if (failing()) return;
2598
2599 inline_incrementally_cleanup(igvn);
2600
2601 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2602
2603 if (failing()) return;
2604
2605 if (_late_inlines.length() == 0) {
2606 break; // no more progress
2607 }
2608 }
2609 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2610
2611 if (_string_late_inlines.length() > 0) {
2612 assert(has_stringbuilder(), "inconsistent");
2613 for_igvn()->clear();
2614 initial_gvn()->replace_with(&igvn);
2615
2616 inline_string_calls(false);
2617
2618 if (failing()) return;
2619
2620 inline_incrementally_cleanup(igvn);
2621 }
2622
2623 set_inlining_incrementally(false);
2624 }
2625
2626 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2627 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2628 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2629 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == NULL"
2630 // as if "inlining_incrementally() == true" were set.
2631 assert(inlining_incrementally() == false, "not allowed");
2632 #ifdef ASSERT
2633 Unique_Node_List* modified_nodes = _modified_nodes;
2634 _modified_nodes = NULL;
2635 #endif
2636 assert(_late_inlines.length() > 0, "sanity");
2637
2638 while (_late_inlines.length() > 0) {
2639 for_igvn()->clear();
2640 initial_gvn()->replace_with(&igvn);
2641
2642 while (inline_incrementally_one()) {
2643 assert(!failing(), "inconsistent");
2644 }
2645 if (failing()) return;
2646
2647 inline_incrementally_cleanup(igvn);
2648 }
2649 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2650 }
2651
2652 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2653 if (_loop_opts_cnt > 0) {
2654 while (major_progress() && (_loop_opts_cnt > 0)) {
2655 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2656 PhaseIdealLoop::optimize(igvn, mode);
2657 _loop_opts_cnt--;
2658 if (failing()) return false;
2659 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2660 }
2661 }
2662 return true;
2663 }
2664
2665 // Remove edges from "root" to each SafePoint at a backward branch.
2666 // They were inserted during parsing (see add_safepoint()) to make
2667 // infinite loops without calls or exceptions visible to root, i.e.,
2668 // useful.
2669 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2670 Node *r = root();
2671 if (r != NULL) {
2672 for (uint i = r->req(); i < r->len(); ++i) {
2673 Node *n = r->in(i);
2674 if (n != NULL && n->is_SafePoint()) {
2675 r->rm_prec(i);
2676 if (n->outcnt() == 0) {
2677 igvn.remove_dead_node(n);
2678 }
2679 --i;
2680 }
2681 }
2682 // Parsing may have added top inputs to the root node (Path
2683 // leading to the Halt node proven dead). Make sure we get a
2684 // chance to clean them up.
2685 igvn._worklist.push(r);
2686 igvn.optimize();
2687 }
2688 }
2689
2690 //------------------------------Optimize---------------------------------------
2691 // Given a graph, optimize it.
2692 void Compile::Optimize() {
2693 TracePhase tp("optimizer", &timers[_t_optimizer]);
2694
2695 #ifndef PRODUCT
2696 if (env()->break_at_compile()) {
2697 BREAKPOINT;
2698 }
2699
2700 #endif
2701
2702 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2703 #ifdef ASSERT
2704 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2705 #endif
2706
2707 ResourceMark rm;
2708
2709 print_inlining_reinit();
2710
2711 NOT_PRODUCT( verify_graph_edges(); )
2712
2713 print_method(PHASE_AFTER_PARSING, 1);
2714
2715 {
2716 // Iterative Global Value Numbering, including ideal transforms
2717 // Initialize IterGVN with types and values from parse-time GVN
2718 PhaseIterGVN igvn(initial_gvn());
2719 #ifdef ASSERT
2720 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2721 #endif
2722 {
2723 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2724 igvn.optimize();
2725 }
2726
2727 if (failing()) return;
2728
2729 print_method(PHASE_ITER_GVN1, 2);
2730
2731 process_for_unstable_if_traps(igvn);
2732
2733 inline_incrementally(igvn);
2734
2735 print_method(PHASE_INCREMENTAL_INLINE, 2);
2736
2737 if (failing()) return;
2738
2739 if (eliminate_boxing()) {
2740 // Inline valueOf() methods now.
2741 inline_boxing_calls(igvn);
2742
2743 if (AlwaysIncrementalInline) {
2744 inline_incrementally(igvn);
2745 }
2746
2747 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2748
2749 if (failing()) return;
2750 }
2751
2752 // Remove the speculative part of types and clean up the graph from
2753 // the extra CastPP nodes whose only purpose is to carry them. Do
2754 // that early so that optimizations are not disrupted by the extra
2755 // CastPP nodes.
2756 remove_speculative_types(igvn);
2757
2758 // No more new expensive nodes will be added to the list from here
2759 // so keep only the actual candidates for optimizations.
2760 cleanup_expensive_nodes(igvn);
2761
2762 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2763 if (EnableVectorSupport && has_vbox_nodes()) {
2764 TracePhase tp("", &timers[_t_vector]);
2765 PhaseVector pv(igvn);
2766 pv.optimize_vector_boxes();
2767
2768 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2769 }
2770 assert(!has_vbox_nodes(), "sanity");
2771
2772 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2773 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2774 initial_gvn()->replace_with(&igvn);
2775 Unique_Node_List* old_worklist = for_igvn();
2776 old_worklist->clear();
2777 Unique_Node_List new_worklist(C->comp_arena());
2778 {
2779 ResourceMark rm;
2780 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2781 }
2782 Unique_Node_List* save_for_igvn = for_igvn();
2783 set_for_igvn(&new_worklist);
2784 igvn = PhaseIterGVN(initial_gvn());
2785 igvn.optimize();
2786 set_for_igvn(old_worklist); // new_worklist is dead beyond this point
2787 }
2788
2789 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2790 // safepoints
2791 remove_root_to_sfpts_edges(igvn);
2792
2793 // Process inline type nodes now that all inlining is over
2794 process_inline_types(igvn);
2795
2796 adjust_flattened_array_access_aliases(igvn);
2797
2798 // Perform escape analysis
2799 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2800 if (has_loops()) {
2801 // Cleanup graph (remove dead nodes).
2802 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2803 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2804 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2805 if (failing()) return;
2806 }
2807 bool progress;
2808 do {
2809 ConnectionGraph::do_analysis(this, &igvn);
2810
2811 if (failing()) return;
2812
2813 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2814
2815 // Optimize out fields loads from scalar replaceable allocations.
2816 igvn.optimize();
2817 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2818
2819 if (failing()) return;
2820
2821 if (congraph() != NULL && macro_count() > 0) {
2822 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2823 PhaseMacroExpand mexp(igvn);
2824 mexp.eliminate_macro_nodes();
2825 igvn.set_delay_transform(false);
2826
2827 igvn.optimize();
2828 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2829
2830 if (failing()) return;
2831 }
2832 progress = do_iterative_escape_analysis() &&
2833 (macro_count() < mcount) &&
2834 ConnectionGraph::has_candidates(this);
2835 // Try again if candidates exist and made progress
2836 // by removing some allocations and/or locks.
2837 } while (progress);
2838 }
2839
2840 // Loop transforms on the ideal graph. Range Check Elimination,
2841 // peeling, unrolling, etc.
2842
2843 // Set loop opts counter
2844 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2845 {
2846 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2847 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2848 _loop_opts_cnt--;
2849 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2850 if (failing()) return;
2851 }
2852 // Loop opts pass if partial peeling occurred in previous pass
2853 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2854 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2855 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2856 _loop_opts_cnt--;
2857 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2858 if (failing()) return;
2859 }
2860 // Loop opts pass for loop-unrolling before CCP
2861 if(major_progress() && (_loop_opts_cnt > 0)) {
2862 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2863 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2864 _loop_opts_cnt--;
2865 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2866 }
2867 if (!failing()) {
2868 // Verify that last round of loop opts produced a valid graph
2869 PhaseIdealLoop::verify(igvn);
2870 }
2871 }
2872 if (failing()) return;
2873
2874 // Conditional Constant Propagation;
2875 PhaseCCP ccp( &igvn );
2876 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2877 {
2878 TracePhase tp("ccp", &timers[_t_ccp]);
2879 ccp.do_transform();
2880 }
2881 print_method(PHASE_CCP1, 2);
2882
2883 assert( true, "Break here to ccp.dump_old2new_map()");
2884
2885 // Iterative Global Value Numbering, including ideal transforms
2886 {
2887 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2888 igvn = ccp;
2889 igvn.optimize();
2890 }
2891 print_method(PHASE_ITER_GVN2, 2);
2892
2893 if (failing()) return;
2894
2895 // Loop transforms on the ideal graph. Range Check Elimination,
2896 // peeling, unrolling, etc.
2897 if (!optimize_loops(igvn, LoopOptsDefault)) {
2898 return;
2899 }
2900
2901 if (failing()) return;
2902
2903 C->clear_major_progress(); // ensure that major progress is now clear
2904
2905 process_for_post_loop_opts_igvn(igvn);
2906
2907 #ifdef ASSERT
2908 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2909 #endif
2910
2911 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2912
2913 if (_late_inlines.length() > 0) {
2914 // More opportunities to optimize virtual and MH calls.
2915 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2916 process_late_inline_calls_no_inline(igvn);
2917 }
2918
2919 {
2920 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2921 PhaseMacroExpand mex(igvn);
2922 if (mex.expand_macro_nodes()) {
2923 assert(failing(), "must bail out w/ explicit message");
2924 return;
2925 }
2926 print_method(PHASE_MACRO_EXPANSION, 2);
2927 }
2928
2929 // Process inline type nodes again and remove them. From here
2930 // on we don't need to keep track of field values anymore.
2931 process_inline_types(igvn, /* remove= */ true);
2932
2933 {
2934 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2935 if (bs->expand_barriers(this, igvn)) {
2936 assert(failing(), "must bail out w/ explicit message");
2937 return;
2938 }
2939 print_method(PHASE_BARRIER_EXPANSION, 2);
2940 }
2941
2942 if (C->max_vector_size() > 0) {
2943 C->optimize_logic_cones(igvn);
2944 igvn.optimize();
2945 }
2946
2947 DEBUG_ONLY( _modified_nodes = NULL; )
2948 DEBUG_ONLY( _late_inlines.clear(); )
2949
2950 assert(igvn._worklist.size() == 0, "not empty");
2951 } // (End scope of igvn; run destructor if necessary for asserts.)
2952
2953 check_no_dead_use();
2954
2955 process_print_inlining();
2956
2957 // A method with only infinite loops has no edges entering loops from root
2958 {
2959 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2960 if (final_graph_reshaping()) {
2961 assert(failing(), "must bail out w/ explicit message");
2962 return;
2963 }
2964 }
2965
2966 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2967 DEBUG_ONLY(set_phase_optimize_finished();)
2968 }
2969
2970 #ifdef ASSERT
2971 void Compile::check_no_dead_use() const {
2972 ResourceMark rm;
2973 Unique_Node_List wq;
2974 wq.push(root());
2975 for (uint i = 0; i < wq.size(); ++i) {
2976 Node* n = wq.at(i);
2977 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2978 Node* u = n->fast_out(j);
2979 if (u->outcnt() == 0 && !u->is_Con()) {
2980 u->dump();
2981 fatal("no reachable node should have no use");
2982 }
2983 wq.push(u);
2984 }
2985 }
2986 }
2987 #endif
2988
2989 void Compile::inline_vector_reboxing_calls() {
2990 if (C->_vector_reboxing_late_inlines.length() > 0) {
2991 _late_inlines_pos = C->_late_inlines.length();
2992 while (_vector_reboxing_late_inlines.length() > 0) {
2993 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2994 cg->do_late_inline();
2995 if (failing()) return;
2996 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2997 }
2998 _vector_reboxing_late_inlines.trunc_to(0);
2999 }
3000 }
3001
3002 bool Compile::has_vbox_nodes() {
3003 if (C->_vector_reboxing_late_inlines.length() > 0) {
3004 return true;
3005 }
3006 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3007 Node * n = C->macro_node(macro_idx);
3008 assert(n->is_macro(), "only macro nodes expected here");
3009 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3010 return true;
3011 }
3012 }
3013 return false;
3014 }
3015
3016 //---------------------------- Bitwise operation packing optimization ---------------------------
3017
3018 static bool is_vector_unary_bitwise_op(Node* n) {
3019 return n->Opcode() == Op_XorV &&
3020 VectorNode::is_vector_bitwise_not_pattern(n);
3021 }
3022
3023 static bool is_vector_binary_bitwise_op(Node* n) {
3024 switch (n->Opcode()) {
3025 case Op_AndV:
3026 case Op_OrV:
3027 return true;
3028
3029 case Op_XorV:
3030 return !is_vector_unary_bitwise_op(n);
3031
3032 default:
3033 return false;
3034 }
3035 }
3036
3037 static bool is_vector_ternary_bitwise_op(Node* n) {
3038 return n->Opcode() == Op_MacroLogicV;
3039 }
3040
3041 static bool is_vector_bitwise_op(Node* n) {
3042 return is_vector_unary_bitwise_op(n) ||
3043 is_vector_binary_bitwise_op(n) ||
3044 is_vector_ternary_bitwise_op(n);
3045 }
3046
3047 static bool is_vector_bitwise_cone_root(Node* n) {
3048 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
3049 return false;
3050 }
3051 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3052 if (is_vector_bitwise_op(n->fast_out(i))) {
3053 return false;
3054 }
3055 }
3056 return true;
3057 }
3058
3059 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3060 uint cnt = 0;
3061 if (is_vector_bitwise_op(n)) {
3062 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3063 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3064 for (uint i = 1; i < inp_cnt; i++) {
3065 Node* in = n->in(i);
3066 bool skip = VectorNode::is_all_ones_vector(in);
3067 if (!skip && !inputs.member(in)) {
3068 inputs.push(in);
3069 cnt++;
3070 }
3071 }
3072 assert(cnt <= 1, "not unary");
3073 } else {
3074 uint last_req = inp_cnt;
3075 if (is_vector_ternary_bitwise_op(n)) {
3076 last_req = inp_cnt - 1; // skip last input
3077 }
3078 for (uint i = 1; i < last_req; i++) {
3079 Node* def = n->in(i);
3080 if (!inputs.member(def)) {
3081 inputs.push(def);
3082 cnt++;
3083 }
3084 }
3085 }
3086 } else { // not a bitwise operations
3087 if (!inputs.member(n)) {
3088 inputs.push(n);
3089 cnt++;
3090 }
3091 }
3092 return cnt;
3093 }
3094
3095 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3096 Unique_Node_List useful_nodes;
3097 C->identify_useful_nodes(useful_nodes);
3098
3099 for (uint i = 0; i < useful_nodes.size(); i++) {
3100 Node* n = useful_nodes.at(i);
3101 if (is_vector_bitwise_cone_root(n)) {
3102 list.push(n);
3103 }
3104 }
3105 }
3106
3107 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3108 const TypeVect* vt,
3109 Unique_Node_List& partition,
3110 Unique_Node_List& inputs) {
3111 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3112 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3113 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3114
3115 Node* in1 = inputs.at(0);
3116 Node* in2 = inputs.at(1);
3117 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3118
3119 uint func = compute_truth_table(partition, inputs);
3120
3121 Node* pn = partition.at(partition.size() - 1);
3122 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : NULL;
3123 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3124 }
3125
3126 static uint extract_bit(uint func, uint pos) {
3127 return (func & (1 << pos)) >> pos;
3128 }
3129
3130 //
3131 // A macro logic node represents a truth table. It has 4 inputs,
3132 // First three inputs corresponds to 3 columns of a truth table
3133 // and fourth input captures the logic function.
3134 //
3135 // eg. fn = (in1 AND in2) OR in3;
3136 //
3137 // MacroNode(in1,in2,in3,fn)
3138 //
3139 // -----------------
3140 // in1 in2 in3 fn
3141 // -----------------
3142 // 0 0 0 0
3143 // 0 0 1 1
3144 // 0 1 0 0
3145 // 0 1 1 1
3146 // 1 0 0 0
3147 // 1 0 1 1
3148 // 1 1 0 1
3149 // 1 1 1 1
3150 //
3151
3152 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3153 int res = 0;
3154 for (int i = 0; i < 8; i++) {
3155 int bit1 = extract_bit(in1, i);
3156 int bit2 = extract_bit(in2, i);
3157 int bit3 = extract_bit(in3, i);
3158
3159 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3160 int func_bit = extract_bit(func, func_bit_pos);
3161
3162 res |= func_bit << i;
3163 }
3164 return res;
3165 }
3166
3167 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
3168 assert(n != NULL, "");
3169 assert(eval_map.contains(n), "absent");
3170 return *(eval_map.get(n));
3171 }
3172
3173 static void eval_operands(Node* n,
3174 uint& func1, uint& func2, uint& func3,
3175 ResourceHashtable<Node*,uint>& eval_map) {
3176 assert(is_vector_bitwise_op(n), "");
3177
3178 if (is_vector_unary_bitwise_op(n)) {
3179 Node* opnd = n->in(1);
3180 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3181 opnd = n->in(2);
3182 }
3183 func1 = eval_operand(opnd, eval_map);
3184 } else if (is_vector_binary_bitwise_op(n)) {
3185 func1 = eval_operand(n->in(1), eval_map);
3186 func2 = eval_operand(n->in(2), eval_map);
3187 } else {
3188 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3189 func1 = eval_operand(n->in(1), eval_map);
3190 func2 = eval_operand(n->in(2), eval_map);
3191 func3 = eval_operand(n->in(3), eval_map);
3192 }
3193 }
3194
3195 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3196 assert(inputs.size() <= 3, "sanity");
3197 ResourceMark rm;
3198 uint res = 0;
3199 ResourceHashtable<Node*,uint> eval_map;
3200
3201 // Populate precomputed functions for inputs.
3202 // Each input corresponds to one column of 3 input truth-table.
3203 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3204 0xCC, // (_, b, _) -> b
3205 0xF0 }; // (a, _, _) -> a
3206 for (uint i = 0; i < inputs.size(); i++) {
3207 eval_map.put(inputs.at(i), input_funcs[2-i]);
3208 }
3209
3210 for (uint i = 0; i < partition.size(); i++) {
3211 Node* n = partition.at(i);
3212
3213 uint func1 = 0, func2 = 0, func3 = 0;
3214 eval_operands(n, func1, func2, func3, eval_map);
3215
3216 switch (n->Opcode()) {
3217 case Op_OrV:
3218 assert(func3 == 0, "not binary");
3219 res = func1 | func2;
3220 break;
3221 case Op_AndV:
3222 assert(func3 == 0, "not binary");
3223 res = func1 & func2;
3224 break;
3225 case Op_XorV:
3226 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3227 assert(func2 == 0 && func3 == 0, "not unary");
3228 res = (~func1) & 0xFF;
3229 } else {
3230 assert(func3 == 0, "not binary");
3231 res = func1 ^ func2;
3232 }
3233 break;
3234 case Op_MacroLogicV:
3235 // Ordering of inputs may change during evaluation of sub-tree
3236 // containing MacroLogic node as a child node, thus a re-evaluation
3237 // makes sure that function is evaluated in context of current
3238 // inputs.
3239 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3240 break;
3241
3242 default: assert(false, "not supported: %s", n->Name());
3243 }
3244 assert(res <= 0xFF, "invalid");
3245 eval_map.put(n, res);
3246 }
3247 return res;
3248 }
3249
3250 // Criteria under which nodes gets packed into a macro logic node:-
3251 // 1) Parent and both child nodes are all unmasked or masked with
3252 // same predicates.
3253 // 2) Masked parent can be packed with left child if it is predicated
3254 // and both have same predicates.
3255 // 3) Masked parent can be packed with right child if its un-predicated
3256 // or has matching predication condition.
3257 // 4) An unmasked parent can be packed with an unmasked child.
3258 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3259 assert(partition.size() == 0, "not empty");
3260 assert(inputs.size() == 0, "not empty");
3261 if (is_vector_ternary_bitwise_op(n)) {
3262 return false;
3263 }
3264
3265 bool is_unary_op = is_vector_unary_bitwise_op(n);
3266 if (is_unary_op) {
3267 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3268 return false; // too few inputs
3269 }
3270
3271 bool pack_left_child = true;
3272 bool pack_right_child = true;
3273
3274 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3275 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3276
3277 int left_child_input_cnt = 0;
3278 int right_child_input_cnt = 0;
3279
3280 bool parent_is_predicated = n->is_predicated_vector();
3281 bool left_child_predicated = n->in(1)->is_predicated_vector();
3282 bool right_child_predicated = n->in(2)->is_predicated_vector();
3283
3284 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : NULL;
3285 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : NULL;
3286 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : NULL;
3287
3288 do {
3289 if (pack_left_child && left_child_LOP &&
3290 ((!parent_is_predicated && !left_child_predicated) ||
3291 ((parent_is_predicated && left_child_predicated &&
3292 parent_pred == left_child_pred)))) {
3293 partition.push(n->in(1));
3294 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3295 } else {
3296 inputs.push(n->in(1));
3297 left_child_input_cnt = 1;
3298 }
3299
3300 if (pack_right_child && right_child_LOP &&
3301 (!right_child_predicated ||
3302 (right_child_predicated && parent_is_predicated &&
3303 parent_pred == right_child_pred))) {
3304 partition.push(n->in(2));
3305 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3306 } else {
3307 inputs.push(n->in(2));
3308 right_child_input_cnt = 1;
3309 }
3310
3311 if (inputs.size() > 3) {
3312 assert(partition.size() > 0, "");
3313 inputs.clear();
3314 partition.clear();
3315 if (left_child_input_cnt > right_child_input_cnt) {
3316 pack_left_child = false;
3317 } else {
3318 pack_right_child = false;
3319 }
3320 } else {
3321 break;
3322 }
3323 } while(true);
3324
3325 if(partition.size()) {
3326 partition.push(n);
3327 }
3328
3329 return (partition.size() == 2 || partition.size() == 3) &&
3330 (inputs.size() == 2 || inputs.size() == 3);
3331 }
3332
3333 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3334 assert(is_vector_bitwise_op(n), "not a root");
3335
3336 visited.set(n->_idx);
3337
3338 // 1) Do a DFS walk over the logic cone.
3339 for (uint i = 1; i < n->req(); i++) {
3340 Node* in = n->in(i);
3341 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3342 process_logic_cone_root(igvn, in, visited);
3343 }
3344 }
3345
3346 // 2) Bottom up traversal: Merge node[s] with
3347 // the parent to form macro logic node.
3348 Unique_Node_List partition;
3349 Unique_Node_List inputs;
3350 if (compute_logic_cone(n, partition, inputs)) {
3351 const TypeVect* vt = n->bottom_type()->is_vect();
3352 Node* pn = partition.at(partition.size() - 1);
3353 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : NULL;
3354 if (mask == NULL ||
3355 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3356 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3357 #ifdef ASSERT
3358 if (TraceNewVectors) {
3359 tty->print("new Vector node: ");
3360 macro_logic->dump();
3361 }
3362 #endif
3363 igvn.replace_node(n, macro_logic);
3364 }
3365 }
3366 }
3367
3368 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3369 ResourceMark rm;
3370 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3371 Unique_Node_List list;
3372 collect_logic_cone_roots(list);
3373
3374 while (list.size() > 0) {
3375 Node* n = list.pop();
3376 const TypeVect* vt = n->bottom_type()->is_vect();
3377 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3378 if (supported) {
3379 VectorSet visited(comp_arena());
3380 process_logic_cone_root(igvn, n, visited);
3381 }
3382 }
3383 }
3384 }
3385
3386 //------------------------------Code_Gen---------------------------------------
3387 // Given a graph, generate code for it
3388 void Compile::Code_Gen() {
3389 if (failing()) {
3390 return;
3391 }
3392
3393 // Perform instruction selection. You might think we could reclaim Matcher
3394 // memory PDQ, but actually the Matcher is used in generating spill code.
3395 // Internals of the Matcher (including some VectorSets) must remain live
3396 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3397 // set a bit in reclaimed memory.
3398
3399 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3400 // nodes. Mapping is only valid at the root of each matched subtree.
3401 NOT_PRODUCT( verify_graph_edges(); )
3402
3403 Matcher matcher;
3404 _matcher = &matcher;
3405 {
3406 TracePhase tp("matcher", &timers[_t_matcher]);
3407 matcher.match();
3408 if (failing()) {
3409 return;
3410 }
3411 }
3412 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3413 // nodes. Mapping is only valid at the root of each matched subtree.
3414 NOT_PRODUCT( verify_graph_edges(); )
3415
3416 // If you have too many nodes, or if matching has failed, bail out
3417 check_node_count(0, "out of nodes matching instructions");
3418 if (failing()) {
3419 return;
3420 }
3421
3422 print_method(PHASE_MATCHING, 2);
3423
3424 // Build a proper-looking CFG
3425 PhaseCFG cfg(node_arena(), root(), matcher);
3426 _cfg = &cfg;
3427 {
3428 TracePhase tp("scheduler", &timers[_t_scheduler]);
3429 bool success = cfg.do_global_code_motion();
3430 if (!success) {
3431 return;
3432 }
3433
3434 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3435 NOT_PRODUCT( verify_graph_edges(); )
3436 cfg.verify();
3437 }
3438
3439 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3440 _regalloc = ®alloc;
3441 {
3442 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
3443 // Perform register allocation. After Chaitin, use-def chains are
3444 // no longer accurate (at spill code) and so must be ignored.
3445 // Node->LRG->reg mappings are still accurate.
3446 _regalloc->Register_Allocate();
3447
3448 // Bail out if the allocator builds too many nodes
3449 if (failing()) {
3450 return;
3451 }
3452 }
3453
3454 // Prior to register allocation we kept empty basic blocks in case the
3455 // the allocator needed a place to spill. After register allocation we
3456 // are not adding any new instructions. If any basic block is empty, we
3457 // can now safely remove it.
3458 {
3459 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
3460 cfg.remove_empty_blocks();
3461 if (do_freq_based_layout()) {
3462 PhaseBlockLayout layout(cfg);
3463 } else {
3464 cfg.set_loop_alignment();
3465 }
3466 cfg.fixup_flow();
3467 cfg.remove_unreachable_blocks();
3468 cfg.verify_dominator_tree();
3469 }
3470
3471 // Apply peephole optimizations
3472 if( OptoPeephole ) {
3473 TracePhase tp("peephole", &timers[_t_peephole]);
3474 PhasePeephole peep( _regalloc, cfg);
3475 peep.do_transform();
3476 }
3477
3478 // Do late expand if CPU requires this.
3479 if (Matcher::require_postalloc_expand) {
3480 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3481 cfg.postalloc_expand(_regalloc);
3482 }
3483
3484 // Convert Nodes to instruction bits in a buffer
3485 {
3486 TracePhase tp("output", &timers[_t_output]);
3487 PhaseOutput output;
3488 output.Output();
3489 if (failing()) return;
3490 output.install();
3491 }
3492
3493 print_method(PHASE_FINAL_CODE, 1);
3494
3495 // He's dead, Jim.
3496 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3497 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3498 }
3499
3500 //------------------------------Final_Reshape_Counts---------------------------
3501 // This class defines counters to help identify when a method
3502 // may/must be executed using hardware with only 24-bit precision.
3503 struct Final_Reshape_Counts : public StackObj {
3504 int _call_count; // count non-inlined 'common' calls
3505 int _float_count; // count float ops requiring 24-bit precision
3506 int _double_count; // count double ops requiring more precision
3507 int _java_call_count; // count non-inlined 'java' calls
3508 int _inner_loop_count; // count loops which need alignment
3509 VectorSet _visited; // Visitation flags
3510 Node_List _tests; // Set of IfNodes & PCTableNodes
3511
3512 Final_Reshape_Counts() :
3513 _call_count(0), _float_count(0), _double_count(0),
3514 _java_call_count(0), _inner_loop_count(0) { }
3515
3516 void inc_call_count () { _call_count ++; }
3517 void inc_float_count () { _float_count ++; }
3518 void inc_double_count() { _double_count++; }
3519 void inc_java_call_count() { _java_call_count++; }
3520 void inc_inner_loop_count() { _inner_loop_count++; }
3521
3522 int get_call_count () const { return _call_count ; }
3523 int get_float_count () const { return _float_count ; }
3524 int get_double_count() const { return _double_count; }
3525 int get_java_call_count() const { return _java_call_count; }
3526 int get_inner_loop_count() const { return _inner_loop_count; }
3527 };
3528
3529 // Eliminate trivially redundant StoreCMs and accumulate their
3530 // precedence edges.
3531 void Compile::eliminate_redundant_card_marks(Node* n) {
3532 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3533 if (n->in(MemNode::Address)->outcnt() > 1) {
3534 // There are multiple users of the same address so it might be
3535 // possible to eliminate some of the StoreCMs
3536 Node* mem = n->in(MemNode::Memory);
3537 Node* adr = n->in(MemNode::Address);
3538 Node* val = n->in(MemNode::ValueIn);
3539 Node* prev = n;
3540 bool done = false;
3541 // Walk the chain of StoreCMs eliminating ones that match. As
3542 // long as it's a chain of single users then the optimization is
3543 // safe. Eliminating partially redundant StoreCMs would require
3544 // cloning copies down the other paths.
3545 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3546 if (adr == mem->in(MemNode::Address) &&
3547 val == mem->in(MemNode::ValueIn)) {
3548 // redundant StoreCM
3549 if (mem->req() > MemNode::OopStore) {
3550 // Hasn't been processed by this code yet.
3551 n->add_prec(mem->in(MemNode::OopStore));
3552 } else {
3553 // Already converted to precedence edge
3554 for (uint i = mem->req(); i < mem->len(); i++) {
3555 // Accumulate any precedence edges
3556 if (mem->in(i) != NULL) {
3557 n->add_prec(mem->in(i));
3558 }
3559 }
3560 // Everything above this point has been processed.
3561 done = true;
3562 }
3563 // Eliminate the previous StoreCM
3564 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3565 assert(mem->outcnt() == 0, "should be dead");
3566 mem->disconnect_inputs(this);
3567 } else {
3568 prev = mem;
3569 }
3570 mem = prev->in(MemNode::Memory);
3571 }
3572 }
3573 }
3574
3575
3576 //------------------------------final_graph_reshaping_impl----------------------
3577 // Implement items 1-5 from final_graph_reshaping below.
3578 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
3579
3580 if ( n->outcnt() == 0 ) return; // dead node
3581 uint nop = n->Opcode();
3582
3583 // Check for 2-input instruction with "last use" on right input.
3584 // Swap to left input. Implements item (2).
3585 if( n->req() == 3 && // two-input instruction
3586 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3587 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3588 n->in(2)->outcnt() == 1 &&// right use IS a last use
3589 !n->in(2)->is_Con() ) { // right use is not a constant
3590 // Check for commutative opcode
3591 switch( nop ) {
3592 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3593 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3594 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3595 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3596 case Op_AndL: case Op_XorL: case Op_OrL:
3597 case Op_AndI: case Op_XorI: case Op_OrI: {
3598 // Move "last use" input to left by swapping inputs
3599 n->swap_edges(1, 2);
3600 break;
3601 }
3602 default:
3603 break;
3604 }
3605 }
3606
3607 #ifdef ASSERT
3608 if( n->is_Mem() ) {
3609 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3610 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
3611 // oop will be recorded in oop map if load crosses safepoint
3612 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3613 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3614 "raw memory operations should have control edge");
3615 }
3616 if (n->is_MemBar()) {
3617 MemBarNode* mb = n->as_MemBar();
3618 if (mb->trailing_store() || mb->trailing_load_store()) {
3619 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3620 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3621 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3622 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3623 } else if (mb->leading()) {
3624 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3625 }
3626 }
3627 #endif
3628 // Count FPU ops and common calls, implements item (3)
3629 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
3630 if (!gc_handled) {
3631 final_graph_reshaping_main_switch(n, frc, nop);
3632 }
3633
3634 // Collect CFG split points
3635 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3636 frc._tests.push(n);
3637 }
3638 }
3639
3640 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
3641 switch( nop ) {
3642 // Count all float operations that may use FPU
3643 case Op_AddF:
3644 case Op_SubF:
3645 case Op_MulF:
3646 case Op_DivF:
3647 case Op_NegF:
3648 case Op_ModF:
3649 case Op_ConvI2F:
3650 case Op_ConF:
3651 case Op_CmpF:
3652 case Op_CmpF3:
3653 case Op_StoreF:
3654 case Op_LoadF:
3655 // case Op_ConvL2F: // longs are split into 32-bit halves
3656 frc.inc_float_count();
3657 break;
3658
3659 case Op_ConvF2D:
3660 case Op_ConvD2F:
3661 frc.inc_float_count();
3662 frc.inc_double_count();
3663 break;
3664
3665 // Count all double operations that may use FPU
3666 case Op_AddD:
3667 case Op_SubD:
3668 case Op_MulD:
3669 case Op_DivD:
3670 case Op_NegD:
3671 case Op_ModD:
3672 case Op_ConvI2D:
3673 case Op_ConvD2I:
3674 // case Op_ConvL2D: // handled by leaf call
3675 // case Op_ConvD2L: // handled by leaf call
3676 case Op_ConD:
3677 case Op_CmpD:
3678 case Op_CmpD3:
3679 case Op_StoreD:
3680 case Op_LoadD:
3681 case Op_LoadD_unaligned:
3682 frc.inc_double_count();
3683 break;
3684 case Op_Opaque1: // Remove Opaque Nodes before matching
3685 case Op_Opaque2: // Remove Opaque Nodes before matching
3686 case Op_Opaque3:
3687 n->subsume_by(n->in(1), this);
3688 break;
3689 case Op_CallStaticJava:
3690 case Op_CallJava:
3691 case Op_CallDynamicJava:
3692 frc.inc_java_call_count(); // Count java call site;
3693 case Op_CallRuntime:
3694 case Op_CallLeaf:
3695 case Op_CallLeafVector:
3696 case Op_CallLeafNoFP: {
3697 assert (n->is_Call(), "");
3698 CallNode *call = n->as_Call();
3699 // Count call sites where the FP mode bit would have to be flipped.
3700 // Do not count uncommon runtime calls:
3701 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3702 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3703 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3704 frc.inc_call_count(); // Count the call site
3705 } else { // See if uncommon argument is shared
3706 Node *n = call->in(TypeFunc::Parms);
3707 int nop = n->Opcode();
3708 // Clone shared simple arguments to uncommon calls, item (1).
3709 if (n->outcnt() > 1 &&
3710 !n->is_Proj() &&
3711 nop != Op_CreateEx &&
3712 nop != Op_CheckCastPP &&
3713 nop != Op_DecodeN &&
3714 nop != Op_DecodeNKlass &&
3715 !n->is_Mem() &&
3716 !n->is_Phi()) {
3717 Node *x = n->clone();
3718 call->set_req(TypeFunc::Parms, x);
3719 }
3720 }
3721 break;
3722 }
3723
3724 case Op_StoreCM:
3725 {
3726 // Convert OopStore dependence into precedence edge
3727 Node* prec = n->in(MemNode::OopStore);
3728 n->del_req(MemNode::OopStore);
3729 if (prec->is_MergeMem()) {
3730 MergeMemNode* mm = prec->as_MergeMem();
3731 Node* base = mm->base_memory();
3732 for (int i = AliasIdxRaw + 1; i < num_alias_types(); i++) {
3733 const Type* adr_type = get_adr_type(i);
3734 if (adr_type->isa_aryptr() && adr_type->is_aryptr()->is_flat()) {
3735 Node* m = mm->memory_at(i);
3736 n->add_prec(m);
3737 }
3738 }
3739 if (mm->outcnt() == 0) {
3740 mm->disconnect_inputs(this);
3741 }
3742 } else {
3743 n->add_prec(prec);
3744 }
3745 eliminate_redundant_card_marks(n);
3746 }
3747
3748 // fall through
3749
3750 case Op_StoreB:
3751 case Op_StoreC:
3752 case Op_StoreI:
3753 case Op_StoreL:
3754 case Op_CompareAndSwapB:
3755 case Op_CompareAndSwapS:
3756 case Op_CompareAndSwapI:
3757 case Op_CompareAndSwapL:
3758 case Op_CompareAndSwapP:
3759 case Op_CompareAndSwapN:
3760 case Op_WeakCompareAndSwapB:
3761 case Op_WeakCompareAndSwapS:
3762 case Op_WeakCompareAndSwapI:
3763 case Op_WeakCompareAndSwapL:
3764 case Op_WeakCompareAndSwapP:
3765 case Op_WeakCompareAndSwapN:
3766 case Op_CompareAndExchangeB:
3767 case Op_CompareAndExchangeS:
3768 case Op_CompareAndExchangeI:
3769 case Op_CompareAndExchangeL:
3770 case Op_CompareAndExchangeP:
3771 case Op_CompareAndExchangeN:
3772 case Op_GetAndAddS:
3773 case Op_GetAndAddB:
3774 case Op_GetAndAddI:
3775 case Op_GetAndAddL:
3776 case Op_GetAndSetS:
3777 case Op_GetAndSetB:
3778 case Op_GetAndSetI:
3779 case Op_GetAndSetL:
3780 case Op_GetAndSetP:
3781 case Op_GetAndSetN:
3782 case Op_StoreP:
3783 case Op_StoreN:
3784 case Op_StoreNKlass:
3785 case Op_LoadB:
3786 case Op_LoadUB:
3787 case Op_LoadUS:
3788 case Op_LoadI:
3789 case Op_LoadKlass:
3790 case Op_LoadNKlass:
3791 case Op_LoadL:
3792 case Op_LoadL_unaligned:
3793 case Op_LoadP:
3794 case Op_LoadN:
3795 case Op_LoadRange:
3796 case Op_LoadS:
3797 break;
3798
3799 case Op_AddP: { // Assert sane base pointers
3800 Node *addp = n->in(AddPNode::Address);
3801 assert( !addp->is_AddP() ||
3802 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3803 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3804 "Base pointers must match (addp %u)", addp->_idx );
3805 #ifdef _LP64
3806 if ((UseCompressedOops || UseCompressedClassPointers) &&
3807 addp->Opcode() == Op_ConP &&
3808 addp == n->in(AddPNode::Base) &&
3809 n->in(AddPNode::Offset)->is_Con()) {
3810 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3811 // on the platform and on the compressed oops mode.
3812 // Use addressing with narrow klass to load with offset on x86.
3813 // Some platforms can use the constant pool to load ConP.
3814 // Do this transformation here since IGVN will convert ConN back to ConP.
3815 const Type* t = addp->bottom_type();
3816 bool is_oop = t->isa_oopptr() != NULL;
3817 bool is_klass = t->isa_klassptr() != NULL;
3818
3819 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3820 (is_klass && Matcher::const_klass_prefer_decode())) {
3821 Node* nn = NULL;
3822
3823 int op = is_oop ? Op_ConN : Op_ConNKlass;
3824
3825 // Look for existing ConN node of the same exact type.
3826 Node* r = root();
3827 uint cnt = r->outcnt();
3828 for (uint i = 0; i < cnt; i++) {
3829 Node* m = r->raw_out(i);
3830 if (m!= NULL && m->Opcode() == op &&
3831 m->bottom_type()->make_ptr() == t) {
3832 nn = m;
3833 break;
3834 }
3835 }
3836 if (nn != NULL) {
3837 // Decode a narrow oop to match address
3838 // [R12 + narrow_oop_reg<<3 + offset]
3839 if (is_oop) {
3840 nn = new DecodeNNode(nn, t);
3841 } else {
3842 nn = new DecodeNKlassNode(nn, t);
3843 }
3844 // Check for succeeding AddP which uses the same Base.
3845 // Otherwise we will run into the assertion above when visiting that guy.
3846 for (uint i = 0; i < n->outcnt(); ++i) {
3847 Node *out_i = n->raw_out(i);
3848 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3849 out_i->set_req(AddPNode::Base, nn);
3850 #ifdef ASSERT
3851 for (uint j = 0; j < out_i->outcnt(); ++j) {
3852 Node *out_j = out_i->raw_out(j);
3853 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3854 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3855 }
3856 #endif
3857 }
3858 }
3859 n->set_req(AddPNode::Base, nn);
3860 n->set_req(AddPNode::Address, nn);
3861 if (addp->outcnt() == 0) {
3862 addp->disconnect_inputs(this);
3863 }
3864 }
3865 }
3866 }
3867 #endif
3868 break;
3869 }
3870
3871 case Op_CastPP: {
3872 // Remove CastPP nodes to gain more freedom during scheduling but
3873 // keep the dependency they encode as control or precedence edges
3874 // (if control is set already) on memory operations. Some CastPP
3875 // nodes don't have a control (don't carry a dependency): skip
3876 // those.
3877 if (n->in(0) != NULL) {
3878 ResourceMark rm;
3879 Unique_Node_List wq;
3880 wq.push(n);
3881 for (uint next = 0; next < wq.size(); ++next) {
3882 Node *m = wq.at(next);
3883 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3884 Node* use = m->fast_out(i);
3885 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3886 use->ensure_control_or_add_prec(n->in(0));
3887 } else {
3888 switch(use->Opcode()) {
3889 case Op_AddP:
3890 case Op_DecodeN:
3891 case Op_DecodeNKlass:
3892 case Op_CheckCastPP:
3893 case Op_CastPP:
3894 wq.push(use);
3895 break;
3896 }
3897 }
3898 }
3899 }
3900 }
3901 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3902 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3903 Node* in1 = n->in(1);
3904 const Type* t = n->bottom_type();
3905 Node* new_in1 = in1->clone();
3906 new_in1->as_DecodeN()->set_type(t);
3907
3908 if (!Matcher::narrow_oop_use_complex_address()) {
3909 //
3910 // x86, ARM and friends can handle 2 adds in addressing mode
3911 // and Matcher can fold a DecodeN node into address by using
3912 // a narrow oop directly and do implicit NULL check in address:
3913 //
3914 // [R12 + narrow_oop_reg<<3 + offset]
3915 // NullCheck narrow_oop_reg
3916 //
3917 // On other platforms (Sparc) we have to keep new DecodeN node and
3918 // use it to do implicit NULL check in address:
3919 //
3920 // decode_not_null narrow_oop_reg, base_reg
3921 // [base_reg + offset]
3922 // NullCheck base_reg
3923 //
3924 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3925 // to keep the information to which NULL check the new DecodeN node
3926 // corresponds to use it as value in implicit_null_check().
3927 //
3928 new_in1->set_req(0, n->in(0));
3929 }
3930
3931 n->subsume_by(new_in1, this);
3932 if (in1->outcnt() == 0) {
3933 in1->disconnect_inputs(this);
3934 }
3935 } else {
3936 n->subsume_by(n->in(1), this);
3937 if (n->outcnt() == 0) {
3938 n->disconnect_inputs(this);
3939 }
3940 }
3941 break;
3942 }
3943 #ifdef _LP64
3944 case Op_CmpP:
3945 // Do this transformation here to preserve CmpPNode::sub() and
3946 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3947 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3948 Node* in1 = n->in(1);
3949 Node* in2 = n->in(2);
3950 if (!in1->is_DecodeNarrowPtr()) {
3951 in2 = in1;
3952 in1 = n->in(2);
3953 }
3954 assert(in1->is_DecodeNarrowPtr(), "sanity");
3955
3956 Node* new_in2 = NULL;
3957 if (in2->is_DecodeNarrowPtr()) {
3958 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3959 new_in2 = in2->in(1);
3960 } else if (in2->Opcode() == Op_ConP) {
3961 const Type* t = in2->bottom_type();
3962 if (t == TypePtr::NULL_PTR) {
3963 assert(in1->is_DecodeN(), "compare klass to null?");
3964 // Don't convert CmpP null check into CmpN if compressed
3965 // oops implicit null check is not generated.
3966 // This will allow to generate normal oop implicit null check.
3967 if (Matcher::gen_narrow_oop_implicit_null_checks())
3968 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3969 //
3970 // This transformation together with CastPP transformation above
3971 // will generated code for implicit NULL checks for compressed oops.
3972 //
3973 // The original code after Optimize()
3974 //
3975 // LoadN memory, narrow_oop_reg
3976 // decode narrow_oop_reg, base_reg
3977 // CmpP base_reg, NULL
3978 // CastPP base_reg // NotNull
3979 // Load [base_reg + offset], val_reg
3980 //
3981 // after these transformations will be
3982 //
3983 // LoadN memory, narrow_oop_reg
3984 // CmpN narrow_oop_reg, NULL
3985 // decode_not_null narrow_oop_reg, base_reg
3986 // Load [base_reg + offset], val_reg
3987 //
3988 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3989 // since narrow oops can be used in debug info now (see the code in
3990 // final_graph_reshaping_walk()).
3991 //
3992 // At the end the code will be matched to
3993 // on x86:
3994 //
3995 // Load_narrow_oop memory, narrow_oop_reg
3996 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3997 // NullCheck narrow_oop_reg
3998 //
3999 // and on sparc:
4000 //
4001 // Load_narrow_oop memory, narrow_oop_reg
4002 // decode_not_null narrow_oop_reg, base_reg
4003 // Load [base_reg + offset], val_reg
4004 // NullCheck base_reg
4005 //
4006 } else if (t->isa_oopptr()) {
4007 new_in2 = ConNode::make(t->make_narrowoop());
4008 } else if (t->isa_klassptr()) {
4009 new_in2 = ConNode::make(t->make_narrowklass());
4010 }
4011 }
4012 if (new_in2 != NULL) {
4013 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
4014 n->subsume_by(cmpN, this);
4015 if (in1->outcnt() == 0) {
4016 in1->disconnect_inputs(this);
4017 }
4018 if (in2->outcnt() == 0) {
4019 in2->disconnect_inputs(this);
4020 }
4021 }
4022 }
4023 break;
4024
4025 case Op_DecodeN:
4026 case Op_DecodeNKlass:
4027 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4028 // DecodeN could be pinned when it can't be fold into
4029 // an address expression, see the code for Op_CastPP above.
4030 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4031 break;
4032
4033 case Op_EncodeP:
4034 case Op_EncodePKlass: {
4035 Node* in1 = n->in(1);
4036 if (in1->is_DecodeNarrowPtr()) {
4037 n->subsume_by(in1->in(1), this);
4038 } else if (in1->Opcode() == Op_ConP) {
4039 const Type* t = in1->bottom_type();
4040 if (t == TypePtr::NULL_PTR) {
4041 assert(t->isa_oopptr(), "null klass?");
4042 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4043 } else if (t->isa_oopptr()) {
4044 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4045 } else if (t->isa_klassptr()) {
4046 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4047 }
4048 }
4049 if (in1->outcnt() == 0) {
4050 in1->disconnect_inputs(this);
4051 }
4052 break;
4053 }
4054
4055 case Op_Proj: {
4056 if (OptimizeStringConcat || IncrementalInline) {
4057 ProjNode* proj = n->as_Proj();
4058 if (proj->_is_io_use) {
4059 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4060 // Separate projections were used for the exception path which
4061 // are normally removed by a late inline. If it wasn't inlined
4062 // then they will hang around and should just be replaced with
4063 // the original one. Merge them.
4064 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4065 if (non_io_proj != NULL) {
4066 proj->subsume_by(non_io_proj , this);
4067 }
4068 }
4069 }
4070 break;
4071 }
4072
4073 case Op_Phi:
4074 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4075 // The EncodeP optimization may create Phi with the same edges
4076 // for all paths. It is not handled well by Register Allocator.
4077 Node* unique_in = n->in(1);
4078 assert(unique_in != NULL, "");
4079 uint cnt = n->req();
4080 for (uint i = 2; i < cnt; i++) {
4081 Node* m = n->in(i);
4082 assert(m != NULL, "");
4083 if (unique_in != m)
4084 unique_in = NULL;
4085 }
4086 if (unique_in != NULL) {
4087 n->subsume_by(unique_in, this);
4088 }
4089 }
4090 break;
4091
4092 #endif
4093
4094 #ifdef ASSERT
4095 case Op_CastII:
4096 // Verify that all range check dependent CastII nodes were removed.
4097 if (n->isa_CastII()->has_range_check()) {
4098 n->dump(3);
4099 assert(false, "Range check dependent CastII node was not removed");
4100 }
4101 break;
4102 #endif
4103
4104 case Op_ModI:
4105 if (UseDivMod) {
4106 // Check if a%b and a/b both exist
4107 Node* d = n->find_similar(Op_DivI);
4108 if (d) {
4109 // Replace them with a fused divmod if supported
4110 if (Matcher::has_match_rule(Op_DivModI)) {
4111 DivModINode* divmod = DivModINode::make(n);
4112 d->subsume_by(divmod->div_proj(), this);
4113 n->subsume_by(divmod->mod_proj(), this);
4114 } else {
4115 // replace a%b with a-((a/b)*b)
4116 Node* mult = new MulINode(d, d->in(2));
4117 Node* sub = new SubINode(d->in(1), mult);
4118 n->subsume_by(sub, this);
4119 }
4120 }
4121 }
4122 break;
4123
4124 case Op_ModL:
4125 if (UseDivMod) {
4126 // Check if a%b and a/b both exist
4127 Node* d = n->find_similar(Op_DivL);
4128 if (d) {
4129 // Replace them with a fused divmod if supported
4130 if (Matcher::has_match_rule(Op_DivModL)) {
4131 DivModLNode* divmod = DivModLNode::make(n);
4132 d->subsume_by(divmod->div_proj(), this);
4133 n->subsume_by(divmod->mod_proj(), this);
4134 } else {
4135 // replace a%b with a-((a/b)*b)
4136 Node* mult = new MulLNode(d, d->in(2));
4137 Node* sub = new SubLNode(d->in(1), mult);
4138 n->subsume_by(sub, this);
4139 }
4140 }
4141 }
4142 break;
4143
4144 case Op_UModI:
4145 if (UseDivMod) {
4146 // Check if a%b and a/b both exist
4147 Node* d = n->find_similar(Op_UDivI);
4148 if (d) {
4149 // Replace them with a fused unsigned divmod if supported
4150 if (Matcher::has_match_rule(Op_UDivModI)) {
4151 UDivModINode* divmod = UDivModINode::make(n);
4152 d->subsume_by(divmod->div_proj(), this);
4153 n->subsume_by(divmod->mod_proj(), this);
4154 } else {
4155 // replace a%b with a-((a/b)*b)
4156 Node* mult = new MulINode(d, d->in(2));
4157 Node* sub = new SubINode(d->in(1), mult);
4158 n->subsume_by(sub, this);
4159 }
4160 }
4161 }
4162 break;
4163
4164 case Op_UModL:
4165 if (UseDivMod) {
4166 // Check if a%b and a/b both exist
4167 Node* d = n->find_similar(Op_UDivL);
4168 if (d) {
4169 // Replace them with a fused unsigned divmod if supported
4170 if (Matcher::has_match_rule(Op_UDivModL)) {
4171 UDivModLNode* divmod = UDivModLNode::make(n);
4172 d->subsume_by(divmod->div_proj(), this);
4173 n->subsume_by(divmod->mod_proj(), this);
4174 } else {
4175 // replace a%b with a-((a/b)*b)
4176 Node* mult = new MulLNode(d, d->in(2));
4177 Node* sub = new SubLNode(d->in(1), mult);
4178 n->subsume_by(sub, this);
4179 }
4180 }
4181 }
4182 break;
4183
4184 case Op_LoadVector:
4185 case Op_StoreVector:
4186 case Op_LoadVectorGather:
4187 case Op_StoreVectorScatter:
4188 case Op_LoadVectorGatherMasked:
4189 case Op_StoreVectorScatterMasked:
4190 case Op_VectorCmpMasked:
4191 case Op_VectorMaskGen:
4192 case Op_LoadVectorMasked:
4193 case Op_StoreVectorMasked:
4194 break;
4195
4196 case Op_AddReductionVI:
4197 case Op_AddReductionVL:
4198 case Op_AddReductionVF:
4199 case Op_AddReductionVD:
4200 case Op_MulReductionVI:
4201 case Op_MulReductionVL:
4202 case Op_MulReductionVF:
4203 case Op_MulReductionVD:
4204 case Op_MinReductionV:
4205 case Op_MaxReductionV:
4206 case Op_AndReductionV:
4207 case Op_OrReductionV:
4208 case Op_XorReductionV:
4209 break;
4210
4211 case Op_PackB:
4212 case Op_PackS:
4213 case Op_PackI:
4214 case Op_PackF:
4215 case Op_PackL:
4216 case Op_PackD:
4217 if (n->req()-1 > 2) {
4218 // Replace many operand PackNodes with a binary tree for matching
4219 PackNode* p = (PackNode*) n;
4220 Node* btp = p->binary_tree_pack(1, n->req());
4221 n->subsume_by(btp, this);
4222 }
4223 break;
4224 case Op_Loop:
4225 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4226 case Op_CountedLoop:
4227 case Op_LongCountedLoop:
4228 case Op_OuterStripMinedLoop:
4229 if (n->as_Loop()->is_inner_loop()) {
4230 frc.inc_inner_loop_count();
4231 }
4232 n->as_Loop()->verify_strip_mined(0);
4233 break;
4234 case Op_LShiftI:
4235 case Op_RShiftI:
4236 case Op_URShiftI:
4237 case Op_LShiftL:
4238 case Op_RShiftL:
4239 case Op_URShiftL:
4240 if (Matcher::need_masked_shift_count) {
4241 // The cpu's shift instructions don't restrict the count to the
4242 // lower 5/6 bits. We need to do the masking ourselves.
4243 Node* in2 = n->in(2);
4244 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4245 const TypeInt* t = in2->find_int_type();
4246 if (t != NULL && t->is_con()) {
4247 juint shift = t->get_con();
4248 if (shift > mask) { // Unsigned cmp
4249 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4250 }
4251 } else {
4252 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
4253 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4254 n->set_req(2, shift);
4255 }
4256 }
4257 if (in2->outcnt() == 0) { // Remove dead node
4258 in2->disconnect_inputs(this);
4259 }
4260 }
4261 break;
4262 case Op_MemBarStoreStore:
4263 case Op_MemBarRelease:
4264 // Break the link with AllocateNode: it is no longer useful and
4265 // confuses register allocation.
4266 if (n->req() > MemBarNode::Precedent) {
4267 n->set_req(MemBarNode::Precedent, top());
4268 }
4269 break;
4270 case Op_MemBarAcquire: {
4271 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4272 // At parse time, the trailing MemBarAcquire for a volatile load
4273 // is created with an edge to the load. After optimizations,
4274 // that input may be a chain of Phis. If those phis have no
4275 // other use, then the MemBarAcquire keeps them alive and
4276 // register allocation can be confused.
4277 ResourceMark rm;
4278 Unique_Node_List wq;
4279 wq.push(n->in(MemBarNode::Precedent));
4280 n->set_req(MemBarNode::Precedent, top());
4281 while (wq.size() > 0) {
4282 Node* m = wq.pop();
4283 if (m->outcnt() == 0 && m != top()) {
4284 for (uint j = 0; j < m->req(); j++) {
4285 Node* in = m->in(j);
4286 if (in != NULL) {
4287 wq.push(in);
4288 }
4289 }
4290 m->disconnect_inputs(this);
4291 }
4292 }
4293 }
4294 break;
4295 }
4296 case Op_Blackhole:
4297 break;
4298 case Op_RangeCheck: {
4299 RangeCheckNode* rc = n->as_RangeCheck();
4300 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4301 n->subsume_by(iff, this);
4302 frc._tests.push(iff);
4303 break;
4304 }
4305 case Op_ConvI2L: {
4306 if (!Matcher::convi2l_type_required) {
4307 // Code generation on some platforms doesn't need accurate
4308 // ConvI2L types. Widening the type can help remove redundant
4309 // address computations.
4310 n->as_Type()->set_type(TypeLong::INT);
4311 ResourceMark rm;
4312 Unique_Node_List wq;
4313 wq.push(n);
4314 for (uint next = 0; next < wq.size(); next++) {
4315 Node *m = wq.at(next);
4316
4317 for(;;) {
4318 // Loop over all nodes with identical inputs edges as m
4319 Node* k = m->find_similar(m->Opcode());
4320 if (k == NULL) {
4321 break;
4322 }
4323 // Push their uses so we get a chance to remove node made
4324 // redundant
4325 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4326 Node* u = k->fast_out(i);
4327 if (u->Opcode() == Op_LShiftL ||
4328 u->Opcode() == Op_AddL ||
4329 u->Opcode() == Op_SubL ||
4330 u->Opcode() == Op_AddP) {
4331 wq.push(u);
4332 }
4333 }
4334 // Replace all nodes with identical edges as m with m
4335 k->subsume_by(m, this);
4336 }
4337 }
4338 }
4339 break;
4340 }
4341 case Op_CmpUL: {
4342 if (!Matcher::has_match_rule(Op_CmpUL)) {
4343 // No support for unsigned long comparisons
4344 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4345 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4346 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4347 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4348 Node* andl = new AndLNode(orl, remove_sign_mask);
4349 Node* cmp = new CmpLNode(andl, n->in(2));
4350 n->subsume_by(cmp, this);
4351 }
4352 break;
4353 }
4354 #ifdef ASSERT
4355 case Op_InlineTypePtr:
4356 case Op_InlineType: {
4357 n->dump(-1);
4358 assert(false, "inline type node was not removed");
4359 break;
4360 }
4361 #endif
4362 default:
4363 assert(!n->is_Call(), "");
4364 assert(!n->is_Mem(), "");
4365 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4366 break;
4367 }
4368 }
4369
4370 //------------------------------final_graph_reshaping_walk---------------------
4371 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4372 // requires that the walk visits a node's inputs before visiting the node.
4373 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
4374 Unique_Node_List sfpt;
4375
4376 frc._visited.set(root->_idx); // first, mark node as visited
4377 uint cnt = root->req();
4378 Node *n = root;
4379 uint i = 0;
4380 while (true) {
4381 if (i < cnt) {
4382 // Place all non-visited non-null inputs onto stack
4383 Node* m = n->in(i);
4384 ++i;
4385 if (m != NULL && !frc._visited.test_set(m->_idx)) {
4386 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
4387 // compute worst case interpreter size in case of a deoptimization
4388 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4389
4390 sfpt.push(m);
4391 }
4392 cnt = m->req();
4393 nstack.push(n, i); // put on stack parent and next input's index
4394 n = m;
4395 i = 0;
4396 }
4397 } else {
4398 // Now do post-visit work
4399 final_graph_reshaping_impl( n, frc );
4400 if (nstack.is_empty())
4401 break; // finished
4402 n = nstack.node(); // Get node from stack
4403 cnt = n->req();
4404 i = nstack.index();
4405 nstack.pop(); // Shift to the next node on stack
4406 }
4407 }
4408
4409 // Skip next transformation if compressed oops are not used.
4410 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4411 (!UseCompressedOops && !UseCompressedClassPointers))
4412 return;
4413
4414 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4415 // It could be done for an uncommon traps or any safepoints/calls
4416 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4417 while (sfpt.size() > 0) {
4418 n = sfpt.pop();
4419 JVMState *jvms = n->as_SafePoint()->jvms();
4420 assert(jvms != NULL, "sanity");
4421 int start = jvms->debug_start();
4422 int end = n->req();
4423 bool is_uncommon = (n->is_CallStaticJava() &&
4424 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4425 for (int j = start; j < end; j++) {
4426 Node* in = n->in(j);
4427 if (in->is_DecodeNarrowPtr()) {
4428 bool safe_to_skip = true;
4429 if (!is_uncommon ) {
4430 // Is it safe to skip?
4431 for (uint i = 0; i < in->outcnt(); i++) {
4432 Node* u = in->raw_out(i);
4433 if (!u->is_SafePoint() ||
4434 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4435 safe_to_skip = false;
4436 }
4437 }
4438 }
4439 if (safe_to_skip) {
4440 n->set_req(j, in->in(1));
4441 }
4442 if (in->outcnt() == 0) {
4443 in->disconnect_inputs(this);
4444 }
4445 }
4446 }
4447 }
4448 }
4449
4450 //------------------------------final_graph_reshaping--------------------------
4451 // Final Graph Reshaping.
4452 //
4453 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4454 // and not commoned up and forced early. Must come after regular
4455 // optimizations to avoid GVN undoing the cloning. Clone constant
4456 // inputs to Loop Phis; these will be split by the allocator anyways.
4457 // Remove Opaque nodes.
4458 // (2) Move last-uses by commutative operations to the left input to encourage
4459 // Intel update-in-place two-address operations and better register usage
4460 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4461 // calls canonicalizing them back.
4462 // (3) Count the number of double-precision FP ops, single-precision FP ops
4463 // and call sites. On Intel, we can get correct rounding either by
4464 // forcing singles to memory (requires extra stores and loads after each
4465 // FP bytecode) or we can set a rounding mode bit (requires setting and
4466 // clearing the mode bit around call sites). The mode bit is only used
4467 // if the relative frequency of single FP ops to calls is low enough.
4468 // This is a key transform for SPEC mpeg_audio.
4469 // (4) Detect infinite loops; blobs of code reachable from above but not
4470 // below. Several of the Code_Gen algorithms fail on such code shapes,
4471 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4472 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4473 // Detection is by looking for IfNodes where only 1 projection is
4474 // reachable from below or CatchNodes missing some targets.
4475 // (5) Assert for insane oop offsets in debug mode.
4476
4477 bool Compile::final_graph_reshaping() {
4478 // an infinite loop may have been eliminated by the optimizer,
4479 // in which case the graph will be empty.
4480 if (root()->req() == 1) {
4481 record_method_not_compilable("trivial infinite loop");
4482 return true;
4483 }
4484
4485 // Expensive nodes have their control input set to prevent the GVN
4486 // from freely commoning them. There's no GVN beyond this point so
4487 // no need to keep the control input. We want the expensive nodes to
4488 // be freely moved to the least frequent code path by gcm.
4489 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4490 for (int i = 0; i < expensive_count(); i++) {
4491 _expensive_nodes.at(i)->set_req(0, NULL);
4492 }
4493
4494 Final_Reshape_Counts frc;
4495
4496 // Visit everybody reachable!
4497 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4498 Node_Stack nstack(live_nodes() >> 1);
4499 final_graph_reshaping_walk(nstack, root(), frc);
4500
4501 // Check for unreachable (from below) code (i.e., infinite loops).
4502 for( uint i = 0; i < frc._tests.size(); i++ ) {
4503 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4504 // Get number of CFG targets.
4505 // Note that PCTables include exception targets after calls.
4506 uint required_outcnt = n->required_outcnt();
4507 if (n->outcnt() != required_outcnt) {
4508 // Check for a few special cases. Rethrow Nodes never take the
4509 // 'fall-thru' path, so expected kids is 1 less.
4510 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4511 if (n->in(0)->in(0)->is_Call()) {
4512 CallNode *call = n->in(0)->in(0)->as_Call();
4513 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4514 required_outcnt--; // Rethrow always has 1 less kid
4515 } else if (call->req() > TypeFunc::Parms &&
4516 call->is_CallDynamicJava()) {
4517 // Check for null receiver. In such case, the optimizer has
4518 // detected that the virtual call will always result in a null
4519 // pointer exception. The fall-through projection of this CatchNode
4520 // will not be populated.
4521 Node *arg0 = call->in(TypeFunc::Parms);
4522 if (arg0->is_Type() &&
4523 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4524 required_outcnt--;
4525 }
4526 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
4527 call->req() > TypeFunc::Parms+1 &&
4528 call->is_CallStaticJava()) {
4529 // Check for negative array length. In such case, the optimizer has
4530 // detected that the allocation attempt will always result in an
4531 // exception. There is no fall-through projection of this CatchNode .
4532 Node *arg1 = call->in(TypeFunc::Parms+1);
4533 if (arg1->is_Type() &&
4534 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
4535 required_outcnt--;
4536 }
4537 }
4538 }
4539 }
4540 // Recheck with a better notion of 'required_outcnt'
4541 if (n->outcnt() != required_outcnt) {
4542 record_method_not_compilable("malformed control flow");
4543 return true; // Not all targets reachable!
4544 }
4545 }
4546 // Check that I actually visited all kids. Unreached kids
4547 // must be infinite loops.
4548 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4549 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4550 record_method_not_compilable("infinite loop");
4551 return true; // Found unvisited kid; must be unreach
4552 }
4553
4554 // Here so verification code in final_graph_reshaping_walk()
4555 // always see an OuterStripMinedLoopEnd
4556 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4557 IfNode* init_iff = n->as_If();
4558 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4559 n->subsume_by(iff, this);
4560 }
4561 }
4562
4563 #ifdef IA32
4564 // If original bytecodes contained a mixture of floats and doubles
4565 // check if the optimizer has made it homogeneous, item (3).
4566 if (UseSSE == 0 &&
4567 frc.get_float_count() > 32 &&
4568 frc.get_double_count() == 0 &&
4569 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4570 set_24_bit_selection_and_mode(false, true);
4571 }
4572 #endif // IA32
4573
4574 set_java_calls(frc.get_java_call_count());
4575 set_inner_loops(frc.get_inner_loop_count());
4576
4577 // No infinite loops, no reason to bail out.
4578 return false;
4579 }
4580
4581 //-----------------------------too_many_traps----------------------------------
4582 // Report if there are too many traps at the current method and bci.
4583 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4584 bool Compile::too_many_traps(ciMethod* method,
4585 int bci,
4586 Deoptimization::DeoptReason reason) {
4587 ciMethodData* md = method->method_data();
4588 if (md->is_empty()) {
4589 // Assume the trap has not occurred, or that it occurred only
4590 // because of a transient condition during start-up in the interpreter.
4591 return false;
4592 }
4593 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4594 if (md->has_trap_at(bci, m, reason) != 0) {
4595 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4596 // Also, if there are multiple reasons, or if there is no per-BCI record,
4597 // assume the worst.
4598 if (log())
4599 log()->elem("observe trap='%s' count='%d'",
4600 Deoptimization::trap_reason_name(reason),
4601 md->trap_count(reason));
4602 return true;
4603 } else {
4604 // Ignore method/bci and see if there have been too many globally.
4605 return too_many_traps(reason, md);
4606 }
4607 }
4608
4609 // Less-accurate variant which does not require a method and bci.
4610 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4611 ciMethodData* logmd) {
4612 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4613 // Too many traps globally.
4614 // Note that we use cumulative trap_count, not just md->trap_count.
4615 if (log()) {
4616 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4617 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4618 Deoptimization::trap_reason_name(reason),
4619 mcount, trap_count(reason));
4620 }
4621 return true;
4622 } else {
4623 // The coast is clear.
4624 return false;
4625 }
4626 }
4627
4628 //--------------------------too_many_recompiles--------------------------------
4629 // Report if there are too many recompiles at the current method and bci.
4630 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4631 // Is not eager to return true, since this will cause the compiler to use
4632 // Action_none for a trap point, to avoid too many recompilations.
4633 bool Compile::too_many_recompiles(ciMethod* method,
4634 int bci,
4635 Deoptimization::DeoptReason reason) {
4636 ciMethodData* md = method->method_data();
4637 if (md->is_empty()) {
4638 // Assume the trap has not occurred, or that it occurred only
4639 // because of a transient condition during start-up in the interpreter.
4640 return false;
4641 }
4642 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4643 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4644 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4645 Deoptimization::DeoptReason per_bc_reason
4646 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4647 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4648 if ((per_bc_reason == Deoptimization::Reason_none
4649 || md->has_trap_at(bci, m, reason) != 0)
4650 // The trap frequency measure we care about is the recompile count:
4651 && md->trap_recompiled_at(bci, m)
4652 && md->overflow_recompile_count() >= bc_cutoff) {
4653 // Do not emit a trap here if it has already caused recompilations.
4654 // Also, if there are multiple reasons, or if there is no per-BCI record,
4655 // assume the worst.
4656 if (log())
4657 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4658 Deoptimization::trap_reason_name(reason),
4659 md->trap_count(reason),
4660 md->overflow_recompile_count());
4661 return true;
4662 } else if (trap_count(reason) != 0
4663 && decompile_count() >= m_cutoff) {
4664 // Too many recompiles globally, and we have seen this sort of trap.
4665 // Use cumulative decompile_count, not just md->decompile_count.
4666 if (log())
4667 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4668 Deoptimization::trap_reason_name(reason),
4669 md->trap_count(reason), trap_count(reason),
4670 md->decompile_count(), decompile_count());
4671 return true;
4672 } else {
4673 // The coast is clear.
4674 return false;
4675 }
4676 }
4677
4678 // Compute when not to trap. Used by matching trap based nodes and
4679 // NullCheck optimization.
4680 void Compile::set_allowed_deopt_reasons() {
4681 _allowed_reasons = 0;
4682 if (is_method_compilation()) {
4683 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4684 assert(rs < BitsPerInt, "recode bit map");
4685 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4686 _allowed_reasons |= nth_bit(rs);
4687 }
4688 }
4689 }
4690 }
4691
4692 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4693 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4694 }
4695
4696 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4697 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4698 }
4699
4700 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4701 if (holder->is_initialized()) {
4702 return false;
4703 }
4704 if (holder->is_being_initialized()) {
4705 if (accessing_method->holder() == holder) {
4706 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4707 // <init>, or a static method. In all those cases, there was an initialization
4708 // barrier on the holder klass passed.
4709 if (accessing_method->is_class_initializer() ||
4710 accessing_method->is_object_constructor() ||
4711 accessing_method->is_static()) {
4712 return false;
4713 }
4714 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4715 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4716 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4717 // child class can become fully initialized while its parent class is still being initialized.
4718 if (accessing_method->is_class_initializer()) {
4719 return false;
4720 }
4721 }
4722 ciMethod* root = method(); // the root method of compilation
4723 if (root != accessing_method) {
4724 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4725 }
4726 }
4727 return true;
4728 }
4729
4730 #ifndef PRODUCT
4731 //------------------------------verify_graph_edges---------------------------
4732 // Walk the Graph and verify that there is a one-to-one correspondence
4733 // between Use-Def edges and Def-Use edges in the graph.
4734 void Compile::verify_graph_edges(bool no_dead_code) {
4735 if (VerifyGraphEdges) {
4736 Unique_Node_List visited;
4737 // Call recursive graph walk to check edges
4738 _root->verify_edges(visited);
4739 if (no_dead_code) {
4740 // Now make sure that no visited node is used by an unvisited node.
4741 bool dead_nodes = false;
4742 Unique_Node_List checked;
4743 while (visited.size() > 0) {
4744 Node* n = visited.pop();
4745 checked.push(n);
4746 for (uint i = 0; i < n->outcnt(); i++) {
4747 Node* use = n->raw_out(i);
4748 if (checked.member(use)) continue; // already checked
4749 if (visited.member(use)) continue; // already in the graph
4750 if (use->is_Con()) continue; // a dead ConNode is OK
4751 // At this point, we have found a dead node which is DU-reachable.
4752 if (!dead_nodes) {
4753 tty->print_cr("*** Dead nodes reachable via DU edges:");
4754 dead_nodes = true;
4755 }
4756 use->dump(2);
4757 tty->print_cr("---");
4758 checked.push(use); // No repeats; pretend it is now checked.
4759 }
4760 }
4761 assert(!dead_nodes, "using nodes must be reachable from root");
4762 }
4763 }
4764 }
4765 #endif
4766
4767 // The Compile object keeps track of failure reasons separately from the ciEnv.
4768 // This is required because there is not quite a 1-1 relation between the
4769 // ciEnv and its compilation task and the Compile object. Note that one
4770 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4771 // to backtrack and retry without subsuming loads. Other than this backtracking
4772 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4773 // by the logic in C2Compiler.
4774 void Compile::record_failure(const char* reason) {
4775 if (log() != NULL) {
4776 log()->elem("failure reason='%s' phase='compile'", reason);
4777 }
4778 if (_failure_reason == NULL) {
4779 // Record the first failure reason.
4780 _failure_reason = reason;
4781 }
4782
4783 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4784 C->print_method(PHASE_FAILURE, 1);
4785 }
4786 _root = NULL; // flush the graph, too
4787 }
4788
4789 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4790 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4791 _phase_name(name), _dolog(CITimeVerbose)
4792 {
4793 if (_dolog) {
4794 C = Compile::current();
4795 _log = C->log();
4796 } else {
4797 C = NULL;
4798 _log = NULL;
4799 }
4800 if (_log != NULL) {
4801 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4802 _log->stamp();
4803 _log->end_head();
4804 }
4805 }
4806
4807 Compile::TracePhase::~TracePhase() {
4808
4809 C = Compile::current();
4810 if (_dolog) {
4811 _log = C->log();
4812 } else {
4813 _log = NULL;
4814 }
4815
4816 #ifdef ASSERT
4817 if (PrintIdealNodeCount) {
4818 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4819 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4820 }
4821
4822 if (VerifyIdealNodeCount) {
4823 Compile::current()->print_missing_nodes();
4824 }
4825 #endif
4826
4827 if (_log != NULL) {
4828 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4829 }
4830 }
4831
4832 //----------------------------static_subtype_check-----------------------------
4833 // Shortcut important common cases when superklass is exact:
4834 // (0) superklass is java.lang.Object (can occur in reflective code)
4835 // (1) subklass is already limited to a subtype of superklass => always ok
4836 // (2) subklass does not overlap with superklass => always fail
4837 // (3) superklass has NO subtypes and we can check with a simple compare.
4838 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk) {
4839 if (StressReflectiveCode) {
4840 return SSC_full_test; // Let caller generate the general case.
4841 }
4842
4843 if (subk->is_java_subtype_of(superk)) {
4844 return SSC_always_true; // (0) and (1) this test cannot fail
4845 }
4846
4847 if (!subk->maybe_java_subtype_of(superk)) {
4848 return SSC_always_false; // (2) true path dead; no dynamic test needed
4849 }
4850
4851 const Type* superelem = superk;
4852 if (superk->isa_aryklassptr()) {
4853 int ignored;
4854 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4855
4856 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays.
4857 // [QMyValue is a subtype of [LMyValue but the klass for [QMyValue is not equal to
4858 // the klass for [LMyValue. Perform a full test.
4859 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
4860 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
4861 return SSC_full_test;
4862 }
4863 }
4864
4865 if (superelem->isa_instklassptr()) {
4866 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4867 if (!ik->has_subklass()) {
4868 if (!ik->is_final()) {
4869 // Add a dependency if there is a chance of a later subclass.
4870 dependencies()->assert_leaf_type(ik);
4871 }
4872 if (!superk->maybe_java_subtype_of(subk)) {
4873 return SSC_always_false;
4874 }
4875 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4876 }
4877 } else {
4878 // A primitive array type has no subtypes.
4879 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4880 }
4881
4882 return SSC_full_test;
4883 }
4884
4885 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4886 #ifdef _LP64
4887 // The scaled index operand to AddP must be a clean 64-bit value.
4888 // Java allows a 32-bit int to be incremented to a negative
4889 // value, which appears in a 64-bit register as a large
4890 // positive number. Using that large positive number as an
4891 // operand in pointer arithmetic has bad consequences.
4892 // On the other hand, 32-bit overflow is rare, and the possibility
4893 // can often be excluded, if we annotate the ConvI2L node with
4894 // a type assertion that its value is known to be a small positive
4895 // number. (The prior range check has ensured this.)
4896 // This assertion is used by ConvI2LNode::Ideal.
4897 int index_max = max_jint - 1; // array size is max_jint, index is one less
4898 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4899 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4900 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4901 #endif
4902 return idx;
4903 }
4904
4905 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4906 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4907 if (ctrl != NULL) {
4908 // Express control dependency by a CastII node with a narrow type.
4909 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4910 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4911 // node from floating above the range check during loop optimizations. Otherwise, the
4912 // ConvI2L node may be eliminated independently of the range check, causing the data path
4913 // to become TOP while the control path is still there (although it's unreachable).
4914 value->set_req(0, ctrl);
4915 value = phase->transform(value);
4916 }
4917 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4918 return phase->transform(new ConvI2LNode(value, ltype));
4919 }
4920
4921 // The message about the current inlining is accumulated in
4922 // _print_inlining_stream and transferred into the _print_inlining_list
4923 // once we know whether inlining succeeds or not. For regular
4924 // inlining, messages are appended to the buffer pointed by
4925 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4926 // a new buffer is added after _print_inlining_idx in the list. This
4927 // way we can update the inlining message for late inlining call site
4928 // when the inlining is attempted again.
4929 void Compile::print_inlining_init() {
4930 if (print_inlining() || print_intrinsics()) {
4931 // print_inlining_init is actually called several times.
4932 print_inlining_reset();
4933 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4934 }
4935 }
4936
4937 void Compile::print_inlining_reinit() {
4938 if (print_inlining() || print_intrinsics()) {
4939 print_inlining_reset();
4940 }
4941 }
4942
4943 void Compile::print_inlining_reset() {
4944 _print_inlining_stream->reset();
4945 }
4946
4947 void Compile::print_inlining_commit() {
4948 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4949 // Transfer the message from _print_inlining_stream to the current
4950 // _print_inlining_list buffer and clear _print_inlining_stream.
4951 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4952 print_inlining_reset();
4953 }
4954
4955 void Compile::print_inlining_push() {
4956 // Add new buffer to the _print_inlining_list at current position
4957 _print_inlining_idx++;
4958 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4959 }
4960
4961 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4962 return _print_inlining_list->at(_print_inlining_idx);
4963 }
4964
4965 void Compile::print_inlining_update(CallGenerator* cg) {
4966 if (print_inlining() || print_intrinsics()) {
4967 if (cg->is_late_inline()) {
4968 if (print_inlining_current()->cg() != cg &&
4969 (print_inlining_current()->cg() != NULL ||
4970 print_inlining_current()->ss()->size() != 0)) {
4971 print_inlining_push();
4972 }
4973 print_inlining_commit();
4974 print_inlining_current()->set_cg(cg);
4975 } else {
4976 if (print_inlining_current()->cg() != NULL) {
4977 print_inlining_push();
4978 }
4979 print_inlining_commit();
4980 }
4981 }
4982 }
4983
4984 void Compile::print_inlining_move_to(CallGenerator* cg) {
4985 // We resume inlining at a late inlining call site. Locate the
4986 // corresponding inlining buffer so that we can update it.
4987 if (print_inlining() || print_intrinsics()) {
4988 for (int i = 0; i < _print_inlining_list->length(); i++) {
4989 if (_print_inlining_list->at(i)->cg() == cg) {
4990 _print_inlining_idx = i;
4991 return;
4992 }
4993 }
4994 ShouldNotReachHere();
4995 }
4996 }
4997
4998 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4999 if (print_inlining() || print_intrinsics()) {
5000 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
5001 assert(print_inlining_current()->cg() == cg, "wrong entry");
5002 // replace message with new message
5003 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
5004 print_inlining_commit();
5005 print_inlining_current()->set_cg(cg);
5006 }
5007 }
5008
5009 void Compile::print_inlining_assert_ready() {
5010 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
5011 }
5012
5013 void Compile::process_print_inlining() {
5014 assert(_late_inlines.length() == 0, "not drained yet");
5015 if (print_inlining() || print_intrinsics()) {
5016 ResourceMark rm;
5017 stringStream ss;
5018 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
5019 for (int i = 0; i < _print_inlining_list->length(); i++) {
5020 PrintInliningBuffer* pib = _print_inlining_list->at(i);
5021 ss.print("%s", pib->ss()->freeze());
5022 delete pib;
5023 DEBUG_ONLY(_print_inlining_list->at_put(i, NULL));
5024 }
5025 // Reset _print_inlining_list, it only contains destructed objects.
5026 // It is on the arena, so it will be freed when the arena is reset.
5027 _print_inlining_list = NULL;
5028 // _print_inlining_stream won't be used anymore, either.
5029 print_inlining_reset();
5030 size_t end = ss.size();
5031 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
5032 strncpy(_print_inlining_output, ss.freeze(), end+1);
5033 _print_inlining_output[end] = 0;
5034 }
5035 }
5036
5037 void Compile::dump_print_inlining() {
5038 if (_print_inlining_output != NULL) {
5039 tty->print_raw(_print_inlining_output);
5040 }
5041 }
5042
5043 void Compile::log_late_inline(CallGenerator* cg) {
5044 if (log() != NULL) {
5045 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5046 cg->unique_id());
5047 JVMState* p = cg->call_node()->jvms();
5048 while (p != NULL) {
5049 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5050 p = p->caller();
5051 }
5052 log()->tail("late_inline");
5053 }
5054 }
5055
5056 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5057 log_late_inline(cg);
5058 if (log() != NULL) {
5059 log()->inline_fail(msg);
5060 }
5061 }
5062
5063 void Compile::log_inline_id(CallGenerator* cg) {
5064 if (log() != NULL) {
5065 // The LogCompilation tool needs a unique way to identify late
5066 // inline call sites. This id must be unique for this call site in
5067 // this compilation. Try to have it unique across compilations as
5068 // well because it can be convenient when grepping through the log
5069 // file.
5070 // Distinguish OSR compilations from others in case CICountOSR is
5071 // on.
5072 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5073 cg->set_unique_id(id);
5074 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5075 }
5076 }
5077
5078 void Compile::log_inline_failure(const char* msg) {
5079 if (C->log() != NULL) {
5080 C->log()->inline_fail(msg);
5081 }
5082 }
5083
5084
5085 // Dump inlining replay data to the stream.
5086 // Don't change thread state and acquire any locks.
5087 void Compile::dump_inline_data(outputStream* out) {
5088 InlineTree* inl_tree = ilt();
5089 if (inl_tree != NULL) {
5090 out->print(" inline %d", inl_tree->count());
5091 inl_tree->dump_replay_data(out);
5092 }
5093 }
5094
5095 void Compile::dump_inline_data_reduced(outputStream* out) {
5096 assert(ReplayReduce, "");
5097
5098 InlineTree* inl_tree = ilt();
5099 if (inl_tree == NULL) {
5100 return;
5101 }
5102 // Enable iterative replay file reduction
5103 // Output "compile" lines for depth 1 subtrees,
5104 // simulating that those trees were compiled
5105 // instead of inlined.
5106 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5107 InlineTree* sub = inl_tree->subtrees().at(i);
5108 if (sub->inline_level() != 1) {
5109 continue;
5110 }
5111
5112 ciMethod* method = sub->method();
5113 int entry_bci = -1;
5114 int comp_level = env()->task()->comp_level();
5115 out->print("compile ");
5116 method->dump_name_as_ascii(out);
5117 out->print(" %d %d", entry_bci, comp_level);
5118 out->print(" inline %d", sub->count());
5119 sub->dump_replay_data(out, -1);
5120 out->cr();
5121 }
5122 }
5123
5124 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5125 if (n1->Opcode() < n2->Opcode()) return -1;
5126 else if (n1->Opcode() > n2->Opcode()) return 1;
5127
5128 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5129 for (uint i = 1; i < n1->req(); i++) {
5130 if (n1->in(i) < n2->in(i)) return -1;
5131 else if (n1->in(i) > n2->in(i)) return 1;
5132 }
5133
5134 return 0;
5135 }
5136
5137 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5138 Node* n1 = *n1p;
5139 Node* n2 = *n2p;
5140
5141 return cmp_expensive_nodes(n1, n2);
5142 }
5143
5144 void Compile::sort_expensive_nodes() {
5145 if (!expensive_nodes_sorted()) {
5146 _expensive_nodes.sort(cmp_expensive_nodes);
5147 }
5148 }
5149
5150 bool Compile::expensive_nodes_sorted() const {
5151 for (int i = 1; i < _expensive_nodes.length(); i++) {
5152 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5153 return false;
5154 }
5155 }
5156 return true;
5157 }
5158
5159 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5160 if (_expensive_nodes.length() == 0) {
5161 return false;
5162 }
5163
5164 assert(OptimizeExpensiveOps, "optimization off?");
5165
5166 // Take this opportunity to remove dead nodes from the list
5167 int j = 0;
5168 for (int i = 0; i < _expensive_nodes.length(); i++) {
5169 Node* n = _expensive_nodes.at(i);
5170 if (!n->is_unreachable(igvn)) {
5171 assert(n->is_expensive(), "should be expensive");
5172 _expensive_nodes.at_put(j, n);
5173 j++;
5174 }
5175 }
5176 _expensive_nodes.trunc_to(j);
5177
5178 // Then sort the list so that similar nodes are next to each other
5179 // and check for at least two nodes of identical kind with same data
5180 // inputs.
5181 sort_expensive_nodes();
5182
5183 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5184 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5185 return true;
5186 }
5187 }
5188
5189 return false;
5190 }
5191
5192 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5193 if (_expensive_nodes.length() == 0) {
5194 return;
5195 }
5196
5197 assert(OptimizeExpensiveOps, "optimization off?");
5198
5199 // Sort to bring similar nodes next to each other and clear the
5200 // control input of nodes for which there's only a single copy.
5201 sort_expensive_nodes();
5202
5203 int j = 0;
5204 int identical = 0;
5205 int i = 0;
5206 bool modified = false;
5207 for (; i < _expensive_nodes.length()-1; i++) {
5208 assert(j <= i, "can't write beyond current index");
5209 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5210 identical++;
5211 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5212 continue;
5213 }
5214 if (identical > 0) {
5215 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5216 identical = 0;
5217 } else {
5218 Node* n = _expensive_nodes.at(i);
5219 igvn.replace_input_of(n, 0, NULL);
5220 igvn.hash_insert(n);
5221 modified = true;
5222 }
5223 }
5224 if (identical > 0) {
5225 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5226 } else if (_expensive_nodes.length() >= 1) {
5227 Node* n = _expensive_nodes.at(i);
5228 igvn.replace_input_of(n, 0, NULL);
5229 igvn.hash_insert(n);
5230 modified = true;
5231 }
5232 _expensive_nodes.trunc_to(j);
5233 if (modified) {
5234 igvn.optimize();
5235 }
5236 }
5237
5238 void Compile::add_expensive_node(Node * n) {
5239 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5240 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5241 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5242 if (OptimizeExpensiveOps) {
5243 _expensive_nodes.append(n);
5244 } else {
5245 // Clear control input and let IGVN optimize expensive nodes if
5246 // OptimizeExpensiveOps is off.
5247 n->set_req(0, NULL);
5248 }
5249 }
5250
5251 /**
5252 * Track coarsened Lock and Unlock nodes.
5253 */
5254
5255 class Lock_List : public Node_List {
5256 uint _origin_cnt;
5257 public:
5258 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5259 uint origin_cnt() const { return _origin_cnt; }
5260 };
5261
5262 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5263 int length = locks.length();
5264 if (length > 0) {
5265 // Have to keep this list until locks elimination during Macro nodes elimination.
5266 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5267 for (int i = 0; i < length; i++) {
5268 AbstractLockNode* lock = locks.at(i);
5269 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5270 locks_list->push(lock);
5271 }
5272 _coarsened_locks.append(locks_list);
5273 }
5274 }
5275
5276 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5277 int count = coarsened_count();
5278 for (int i = 0; i < count; i++) {
5279 Node_List* locks_list = _coarsened_locks.at(i);
5280 for (uint j = 0; j < locks_list->size(); j++) {
5281 Node* lock = locks_list->at(j);
5282 assert(lock->is_AbstractLock(), "sanity");
5283 if (!useful.member(lock)) {
5284 locks_list->yank(lock);
5285 }
5286 }
5287 }
5288 }
5289
5290 void Compile::remove_coarsened_lock(Node* n) {
5291 if (n->is_AbstractLock()) {
5292 int count = coarsened_count();
5293 for (int i = 0; i < count; i++) {
5294 Node_List* locks_list = _coarsened_locks.at(i);
5295 locks_list->yank(n);
5296 }
5297 }
5298 }
5299
5300 bool Compile::coarsened_locks_consistent() {
5301 int count = coarsened_count();
5302 for (int i = 0; i < count; i++) {
5303 bool unbalanced = false;
5304 bool modified = false; // track locks kind modifications
5305 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5306 uint size = locks_list->size();
5307 if (size == 0) {
5308 unbalanced = false; // All locks were eliminated - good
5309 } else if (size != locks_list->origin_cnt()) {
5310 unbalanced = true; // Some locks were removed from list
5311 } else {
5312 for (uint j = 0; j < size; j++) {
5313 Node* lock = locks_list->at(j);
5314 // All nodes in group should have the same state (modified or not)
5315 if (!lock->as_AbstractLock()->is_coarsened()) {
5316 if (j == 0) {
5317 // first on list was modified, the rest should be too for consistency
5318 modified = true;
5319 } else if (!modified) {
5320 // this lock was modified but previous locks on the list were not
5321 unbalanced = true;
5322 break;
5323 }
5324 } else if (modified) {
5325 // previous locks on list were modified but not this lock
5326 unbalanced = true;
5327 break;
5328 }
5329 }
5330 }
5331 if (unbalanced) {
5332 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5333 #ifdef ASSERT
5334 if (PrintEliminateLocks) {
5335 tty->print_cr("=== unbalanced coarsened locks ===");
5336 for (uint l = 0; l < size; l++) {
5337 locks_list->at(l)->dump();
5338 }
5339 }
5340 #endif
5341 record_failure(C2Compiler::retry_no_locks_coarsening());
5342 return false;
5343 }
5344 }
5345 return true;
5346 }
5347
5348 /**
5349 * Remove the speculative part of types and clean up the graph
5350 */
5351 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5352 if (UseTypeSpeculation) {
5353 Unique_Node_List worklist;
5354 worklist.push(root());
5355 int modified = 0;
5356 // Go over all type nodes that carry a speculative type, drop the
5357 // speculative part of the type and enqueue the node for an igvn
5358 // which may optimize it out.
5359 for (uint next = 0; next < worklist.size(); ++next) {
5360 Node *n = worklist.at(next);
5361 if (n->is_Type()) {
5362 TypeNode* tn = n->as_Type();
5363 const Type* t = tn->type();
5364 const Type* t_no_spec = t->remove_speculative();
5365 if (t_no_spec != t) {
5366 bool in_hash = igvn.hash_delete(n);
5367 assert(in_hash, "node should be in igvn hash table");
5368 tn->set_type(t_no_spec);
5369 igvn.hash_insert(n);
5370 igvn._worklist.push(n); // give it a chance to go away
5371 modified++;
5372 }
5373 }
5374 // Iterate over outs - endless loops is unreachable from below
5375 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5376 Node *m = n->fast_out(i);
5377 if (not_a_node(m)) {
5378 continue;
5379 }
5380 worklist.push(m);
5381 }
5382 }
5383 // Drop the speculative part of all types in the igvn's type table
5384 igvn.remove_speculative_types();
5385 if (modified > 0) {
5386 igvn.optimize();
5387 }
5388 #ifdef ASSERT
5389 // Verify that after the IGVN is over no speculative type has resurfaced
5390 worklist.clear();
5391 worklist.push(root());
5392 for (uint next = 0; next < worklist.size(); ++next) {
5393 Node *n = worklist.at(next);
5394 const Type* t = igvn.type_or_null(n);
5395 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
5396 if (n->is_Type()) {
5397 t = n->as_Type()->type();
5398 assert(t == t->remove_speculative(), "no more speculative types");
5399 }
5400 // Iterate over outs - endless loops is unreachable from below
5401 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5402 Node *m = n->fast_out(i);
5403 if (not_a_node(m)) {
5404 continue;
5405 }
5406 worklist.push(m);
5407 }
5408 }
5409 igvn.check_no_speculative_types();
5410 #endif
5411 }
5412 }
5413
5414 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5415 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5416 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5417 if (!EnableValhalla || ta == NULL || tb == NULL ||
5418 ta->is_zero_type() || tb->is_zero_type() ||
5419 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5420 // Use old acmp if one operand is null or not an inline type
5421 return new CmpPNode(a, b);
5422 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5423 // We know that one operand is an inline type. Therefore,
5424 // new acmp will only return true if both operands are NULL.
5425 // Check if both operands are null by or'ing the oops.
5426 a = phase->transform(new CastP2XNode(NULL, a));
5427 b = phase->transform(new CastP2XNode(NULL, b));
5428 a = phase->transform(new OrXNode(a, b));
5429 return new CmpXNode(a, phase->MakeConX(0));
5430 }
5431 // Use new acmp
5432 return NULL;
5433 }
5434
5435 // Auxiliary methods to support randomized stressing/fuzzing.
5436
5437 int Compile::random() {
5438 _stress_seed = os::next_random(_stress_seed);
5439 return static_cast<int>(_stress_seed);
5440 }
5441
5442 // This method can be called the arbitrary number of times, with current count
5443 // as the argument. The logic allows selecting a single candidate from the
5444 // running list of candidates as follows:
5445 // int count = 0;
5446 // Cand* selected = null;
5447 // while(cand = cand->next()) {
5448 // if (randomized_select(++count)) {
5449 // selected = cand;
5450 // }
5451 // }
5452 //
5453 // Including count equalizes the chances any candidate is "selected".
5454 // This is useful when we don't have the complete list of candidates to choose
5455 // from uniformly. In this case, we need to adjust the randomicity of the
5456 // selection, or else we will end up biasing the selection towards the latter
5457 // candidates.
5458 //
5459 // Quick back-envelope calculation shows that for the list of n candidates
5460 // the equal probability for the candidate to persist as "best" can be
5461 // achieved by replacing it with "next" k-th candidate with the probability
5462 // of 1/k. It can be easily shown that by the end of the run, the
5463 // probability for any candidate is converged to 1/n, thus giving the
5464 // uniform distribution among all the candidates.
5465 //
5466 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5467 #define RANDOMIZED_DOMAIN_POW 29
5468 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5469 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5470 bool Compile::randomized_select(int count) {
5471 assert(count > 0, "only positive");
5472 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5473 }
5474
5475 CloneMap& Compile::clone_map() { return _clone_map; }
5476 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5477
5478 void NodeCloneInfo::dump() const {
5479 tty->print(" {%d:%d} ", idx(), gen());
5480 }
5481
5482 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5483 uint64_t val = value(old->_idx);
5484 NodeCloneInfo cio(val);
5485 assert(val != 0, "old node should be in the map");
5486 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5487 insert(nnn->_idx, cin.get());
5488 #ifndef PRODUCT
5489 if (is_debug()) {
5490 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5491 }
5492 #endif
5493 }
5494
5495 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5496 NodeCloneInfo cio(value(old->_idx));
5497 if (cio.get() == 0) {
5498 cio.set(old->_idx, 0);
5499 insert(old->_idx, cio.get());
5500 #ifndef PRODUCT
5501 if (is_debug()) {
5502 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5503 }
5504 #endif
5505 }
5506 clone(old, nnn, gen);
5507 }
5508
5509 int CloneMap::max_gen() const {
5510 int g = 0;
5511 DictI di(_dict);
5512 for(; di.test(); ++di) {
5513 int t = gen(di._key);
5514 if (g < t) {
5515 g = t;
5516 #ifndef PRODUCT
5517 if (is_debug()) {
5518 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5519 }
5520 #endif
5521 }
5522 }
5523 return g;
5524 }
5525
5526 void CloneMap::dump(node_idx_t key) const {
5527 uint64_t val = value(key);
5528 if (val != 0) {
5529 NodeCloneInfo ni(val);
5530 ni.dump();
5531 }
5532 }
5533
5534 // Move Allocate nodes to the start of the list
5535 void Compile::sort_macro_nodes() {
5536 int count = macro_count();
5537 int allocates = 0;
5538 for (int i = 0; i < count; i++) {
5539 Node* n = macro_node(i);
5540 if (n->is_Allocate()) {
5541 if (i != allocates) {
5542 Node* tmp = macro_node(allocates);
5543 _macro_nodes.at_put(allocates, n);
5544 _macro_nodes.at_put(i, tmp);
5545 }
5546 allocates++;
5547 }
5548 }
5549 }
5550
5551 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5552 EventCompilerPhase event;
5553 if (event.should_commit()) {
5554 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5555 }
5556 #ifndef PRODUCT
5557 ResourceMark rm;
5558 stringStream ss;
5559 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5560 if (n != nullptr) {
5561 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5562 }
5563
5564 const char* name = ss.as_string();
5565 if (should_print_igv(level)) {
5566 _igv_printer->print_method(name, level);
5567 }
5568 if (should_print_phase(cpt)) {
5569 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5570 }
5571 #endif
5572 C->_latest_stage_start_counter.stamp();
5573 }
5574
5575 // Only used from CompileWrapper
5576 void Compile::begin_method() {
5577 #ifndef PRODUCT
5578 if (_method != NULL && should_print_igv(1)) {
5579 _igv_printer->begin_method();
5580 }
5581 #endif
5582 C->_latest_stage_start_counter.stamp();
5583 }
5584
5585 // Only used from CompileWrapper
5586 void Compile::end_method() {
5587 EventCompilerPhase event;
5588 if (event.should_commit()) {
5589 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5590 }
5591
5592 #ifndef PRODUCT
5593 if (_method != NULL && should_print_igv(1)) {
5594 _igv_printer->end_method();
5595 }
5596 #endif
5597 }
5598
5599 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5600 #ifndef PRODUCT
5601 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) {
5602 return true;
5603 }
5604 #endif
5605 return false;
5606 }
5607
5608 bool Compile::should_print_igv(int level) {
5609 #ifndef PRODUCT
5610 if (PrintIdealGraphLevel < 0) { // disabled by the user
5611 return false;
5612 }
5613
5614 bool need = directive()->IGVPrintLevelOption >= level;
5615 if (need && !_igv_printer) {
5616 _igv_printer = IdealGraphPrinter::printer();
5617 _igv_printer->set_compile(this);
5618 }
5619 return need;
5620 #else
5621 return false;
5622 #endif
5623 }
5624
5625 #ifndef PRODUCT
5626 IdealGraphPrinter* Compile::_debug_file_printer = NULL;
5627 IdealGraphPrinter* Compile::_debug_network_printer = NULL;
5628
5629 // Called from debugger. Prints method to the default file with the default phase name.
5630 // This works regardless of any Ideal Graph Visualizer flags set or not.
5631 void igv_print() {
5632 Compile::current()->igv_print_method_to_file();
5633 }
5634
5635 // Same as igv_print() above but with a specified phase name.
5636 void igv_print(const char* phase_name) {
5637 Compile::current()->igv_print_method_to_file(phase_name);
5638 }
5639
5640 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5641 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5642 // This works regardless of any Ideal Graph Visualizer flags set or not.
5643 void igv_print(bool network) {
5644 if (network) {
5645 Compile::current()->igv_print_method_to_network();
5646 } else {
5647 Compile::current()->igv_print_method_to_file();
5648 }
5649 }
5650
5651 // Same as igv_print(bool network) above but with a specified phase name.
5652 void igv_print(bool network, const char* phase_name) {
5653 if (network) {
5654 Compile::current()->igv_print_method_to_network(phase_name);
5655 } else {
5656 Compile::current()->igv_print_method_to_file(phase_name);
5657 }
5658 }
5659
5660 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5661 void igv_print_default() {
5662 Compile::current()->print_method(PHASE_DEBUG, 0);
5663 }
5664
5665 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5666 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5667 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5668 void igv_append() {
5669 Compile::current()->igv_print_method_to_file("Debug", true);
5670 }
5671
5672 // Same as igv_append() above but with a specified phase name.
5673 void igv_append(const char* phase_name) {
5674 Compile::current()->igv_print_method_to_file(phase_name, true);
5675 }
5676
5677 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5678 const char* file_name = "custom_debug.xml";
5679 if (_debug_file_printer == NULL) {
5680 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5681 } else {
5682 _debug_file_printer->update_compiled_method(C->method());
5683 }
5684 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5685 _debug_file_printer->print(phase_name, (Node*)C->root());
5686 }
5687
5688 void Compile::igv_print_method_to_network(const char* phase_name) {
5689 if (_debug_network_printer == NULL) {
5690 _debug_network_printer = new IdealGraphPrinter(C);
5691 } else {
5692 _debug_network_printer->update_compiled_method(C->method());
5693 }
5694 tty->print_cr("Method printed over network stream to IGV");
5695 _debug_network_printer->print(phase_name, (Node*)C->root());
5696 }
5697 #endif
5698
5699 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5700 if (type != NULL && phase->type(value)->higher_equal(type)) {
5701 return value;
5702 }
5703 Node* result = NULL;
5704 if (bt == T_BYTE) {
5705 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5706 result = new RShiftINode(result, phase->intcon(24));
5707 } else if (bt == T_BOOLEAN) {
5708 result = new AndINode(value, phase->intcon(0xFF));
5709 } else if (bt == T_CHAR) {
5710 result = new AndINode(value,phase->intcon(0xFFFF));
5711 } else {
5712 assert(bt == T_SHORT, "unexpected narrow type");
5713 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5714 result = new RShiftINode(result, phase->intcon(16));
5715 }
5716 if (transform_res) {
5717 result = phase->transform(result);
5718 }
5719 return result;
5720 }
5721