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
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/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 "jvm_io.h"
40 #include "memory/allocation.hpp"
41 #include "memory/resourceArea.hpp"
42 #include "opto/addnode.hpp"
43 #include "opto/block.hpp"
44 #include "opto/c2compiler.hpp"
45 #include "opto/callGenerator.hpp"
46 #include "opto/callnode.hpp"
47 #include "opto/castnode.hpp"
48 #include "opto/cfgnode.hpp"
49 #include "opto/chaitin.hpp"
50 #include "opto/compile.hpp"
51 #include "opto/connode.hpp"
52 #include "opto/convertnode.hpp"
53 #include "opto/divnode.hpp"
54 #include "opto/escape.hpp"
55 #include "opto/idealGraphPrinter.hpp"
56 #include "opto/inlinetypenode.hpp"
57 #include "opto/loopnode.hpp"
58 #include "opto/machnode.hpp"
59 #include "opto/macro.hpp"
60 #include "opto/matcher.hpp"
61 #include "opto/mathexactnode.hpp"
62 #include "opto/memnode.hpp"
63 #include "opto/mulnode.hpp"
64 #include "opto/narrowptrnode.hpp"
65 #include "opto/node.hpp"
66 #include "opto/opcodes.hpp"
67 #include "opto/output.hpp"
68 #include "opto/parse.hpp"
69 #include "opto/phaseX.hpp"
70 #include "opto/rootnode.hpp"
71 #include "opto/runtime.hpp"
72 #include "opto/stringopts.hpp"
73 #include "opto/type.hpp"
74 #include "opto/vector.hpp"
75 #include "opto/vectornode.hpp"
76 #include "runtime/globals_extension.hpp"
77 #include "runtime/sharedRuntime.hpp"
78 #include "runtime/signature.hpp"
79 #include "runtime/stubRoutines.hpp"
80 #include "runtime/timer.hpp"
81 #include "utilities/align.hpp"
82 #include "utilities/copy.hpp"
83 #include "utilities/macros.hpp"
84 #include "utilities/resourceHash.hpp"
85
86 // -------------------- Compile::mach_constant_base_node -----------------------
87 // Constant table base node singleton.
88 MachConstantBaseNode* Compile::mach_constant_base_node() {
89 if (_mach_constant_base_node == NULL) {
90 _mach_constant_base_node = new MachConstantBaseNode();
91 _mach_constant_base_node->add_req(C->root());
92 }
93 return _mach_constant_base_node;
94 }
95
96
97 /// Support for intrinsics.
98
99 // Return the index at which m must be inserted (or already exists).
100 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
101 class IntrinsicDescPair {
102 private:
103 ciMethod* _m;
104 bool _is_virtual;
105 public:
106 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
107 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
108 ciMethod* m= elt->method();
109 ciMethod* key_m = key->_m;
110 if (key_m < m) return -1;
111 else if (key_m > m) return 1;
112 else {
113 bool is_virtual = elt->is_virtual();
114 bool key_virtual = key->_is_virtual;
115 if (key_virtual < is_virtual) return -1;
116 else if (key_virtual > is_virtual) return 1;
117 else return 0;
118 }
119 }
120 };
121 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
122 #ifdef ASSERT
123 for (int i = 1; i < _intrinsics.length(); i++) {
124 CallGenerator* cg1 = _intrinsics.at(i-1);
125 CallGenerator* cg2 = _intrinsics.at(i);
126 assert(cg1->method() != cg2->method()
127 ? cg1->method() < cg2->method()
128 : cg1->is_virtual() < cg2->is_virtual(),
129 "compiler intrinsics list must stay sorted");
130 }
131 #endif
132 IntrinsicDescPair pair(m, is_virtual);
133 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
134 }
135
136 void Compile::register_intrinsic(CallGenerator* cg) {
137 bool found = false;
138 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
139 assert(!found, "registering twice");
140 _intrinsics.insert_before(index, cg);
141 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
142 }
143
144 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
145 assert(m->is_loaded(), "don't try this on unloaded methods");
146 if (_intrinsics.length() > 0) {
147 bool found = false;
148 int index = intrinsic_insertion_index(m, is_virtual, found);
149 if (found) {
150 return _intrinsics.at(index);
151 }
152 }
153 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
154 if (m->intrinsic_id() != vmIntrinsics::_none &&
155 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
156 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
157 if (cg != NULL) {
158 // Save it for next time:
159 register_intrinsic(cg);
160 return cg;
161 } else {
162 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
163 }
164 }
165 return NULL;
166 }
167
168 // Compile::make_vm_intrinsic is defined in library_call.cpp.
169
170 #ifndef PRODUCT
171 // statistics gathering...
172
173 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
174 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
175
176 inline int as_int(vmIntrinsics::ID id) {
177 return vmIntrinsics::as_int(id);
178 }
179
180 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
181 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
182 int oflags = _intrinsic_hist_flags[as_int(id)];
183 assert(flags != 0, "what happened?");
184 if (is_virtual) {
185 flags |= _intrinsic_virtual;
186 }
187 bool changed = (flags != oflags);
188 if ((flags & _intrinsic_worked) != 0) {
189 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
190 if (count == 1) {
191 changed = true; // first time
192 }
193 // increment the overall count also:
194 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
195 }
196 if (changed) {
197 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
198 // Something changed about the intrinsic's virtuality.
199 if ((flags & _intrinsic_virtual) != 0) {
200 // This is the first use of this intrinsic as a virtual call.
201 if (oflags != 0) {
202 // We already saw it as a non-virtual, so note both cases.
203 flags |= _intrinsic_both;
204 }
205 } else if ((oflags & _intrinsic_both) == 0) {
206 // This is the first use of this intrinsic as a non-virtual
207 flags |= _intrinsic_both;
208 }
209 }
210 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
211 }
212 // update the overall flags also:
213 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
214 return changed;
215 }
216
217 static char* format_flags(int flags, char* buf) {
218 buf[0] = 0;
219 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
220 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
221 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
222 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
223 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
224 if (buf[0] == 0) strcat(buf, ",");
225 assert(buf[0] == ',', "must be");
226 return &buf[1];
227 }
228
229 void Compile::print_intrinsic_statistics() {
230 char flagsbuf[100];
231 ttyLocker ttyl;
232 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
233 tty->print_cr("Compiler intrinsic usage:");
234 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
235 if (total == 0) total = 1; // avoid div0 in case of no successes
236 #define PRINT_STAT_LINE(name, c, f) \
237 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
238 for (auto id : EnumRange<vmIntrinsicID>{}) {
239 int flags = _intrinsic_hist_flags[as_int(id)];
240 juint count = _intrinsic_hist_count[as_int(id)];
241 if ((flags | count) != 0) {
242 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
243 }
244 }
245 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
246 if (xtty != NULL) xtty->tail("statistics");
247 }
248
249 void Compile::print_statistics() {
250 { ttyLocker ttyl;
251 if (xtty != NULL) xtty->head("statistics type='opto'");
252 Parse::print_statistics();
253 PhaseStringOpts::print_statistics();
254 PhaseCCP::print_statistics();
255 PhaseRegAlloc::print_statistics();
256 PhaseOutput::print_statistics();
257 PhasePeephole::print_statistics();
258 PhaseIdealLoop::print_statistics();
259 ConnectionGraph::print_statistics();
260 PhaseMacroExpand::print_statistics();
261 if (xtty != NULL) xtty->tail("statistics");
262 }
263 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
264 // put this under its own <statistics> element.
265 print_intrinsic_statistics();
266 }
267 }
268 #endif //PRODUCT
269
270 void Compile::gvn_replace_by(Node* n, Node* nn) {
271 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
272 Node* use = n->last_out(i);
273 bool is_in_table = initial_gvn()->hash_delete(use);
274 uint uses_found = 0;
275 for (uint j = 0; j < use->len(); j++) {
276 if (use->in(j) == n) {
277 if (j < use->req())
278 use->set_req(j, nn);
279 else
280 use->set_prec(j, nn);
281 uses_found++;
282 }
283 }
284 if (is_in_table) {
285 // reinsert into table
286 initial_gvn()->hash_find_insert(use);
287 }
288 record_for_igvn(use);
289 i -= uses_found; // we deleted 1 or more copies of this edge
290 }
291 }
292
293
294 // Identify all nodes that are reachable from below, useful.
295 // Use breadth-first pass that records state in a Unique_Node_List,
296 // recursive traversal is slower.
297 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
298 int estimated_worklist_size = live_nodes();
299 useful.map( estimated_worklist_size, NULL ); // preallocate space
300
301 // Initialize worklist
302 if (root() != NULL) { useful.push(root()); }
303 // If 'top' is cached, declare it useful to preserve cached node
304 if( cached_top_node() ) { useful.push(cached_top_node()); }
305
306 // Push all useful nodes onto the list, breadthfirst
307 for( uint next = 0; next < useful.size(); ++next ) {
308 assert( next < unique(), "Unique useful nodes < total nodes");
309 Node *n = useful.at(next);
310 uint max = n->len();
311 for( uint i = 0; i < max; ++i ) {
312 Node *m = n->in(i);
313 if (not_a_node(m)) continue;
314 useful.push(m);
315 }
316 }
317 }
318
319 // Update dead_node_list with any missing dead nodes using useful
320 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
321 void Compile::update_dead_node_list(Unique_Node_List &useful) {
322 uint max_idx = unique();
323 VectorSet& useful_node_set = useful.member_set();
324
325 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
326 // If node with index node_idx is not in useful set,
327 // mark it as dead in dead node list.
328 if (!useful_node_set.test(node_idx)) {
329 record_dead_node(node_idx);
330 }
331 }
332 }
333
334 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
335 int shift = 0;
336 for (int i = 0; i < inlines->length(); i++) {
337 CallGenerator* cg = inlines->at(i);
338 if (useful.member(cg->call_node())) {
339 if (shift > 0) {
340 inlines->at_put(i - shift, cg);
341 }
342 } else {
343 shift++; // skip over the dead element
344 }
345 }
346 if (shift > 0) {
347 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
348 }
349 }
350
351 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
352 assert(dead != NULL && dead->is_Call(), "sanity");
353 int found = 0;
354 for (int i = 0; i < inlines->length(); i++) {
355 if (inlines->at(i)->call_node() == dead) {
356 inlines->remove_at(i);
357 found++;
358 NOT_DEBUG( break; ) // elements are unique, so exit early
359 }
360 }
361 assert(found <= 1, "not unique");
362 }
363
364 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) {
365 for (int i = node_list.length() - 1; i >= 0; i--) {
366 Node* n = node_list.at(i);
367 if (!useful.member(n)) {
368 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
369 }
370 }
371 }
372
373 void Compile::remove_useless_node(Node* dead) {
374 remove_modified_node(dead);
375
376 // Constant node that has no out-edges and has only one in-edge from
377 // root is usually dead. However, sometimes reshaping walk makes
378 // it reachable by adding use edges. So, we will NOT count Con nodes
379 // as dead to be conservative about the dead node count at any
380 // given time.
381 if (!dead->is_Con()) {
382 record_dead_node(dead->_idx);
383 }
384 if (dead->is_macro()) {
385 remove_macro_node(dead);
386 }
387 if (dead->is_expensive()) {
388 remove_expensive_node(dead);
389 }
390 if (dead->Opcode() == Op_Opaque4) {
391 remove_skeleton_predicate_opaq(dead);
392 }
393 if (dead->for_post_loop_opts_igvn()) {
394 remove_from_post_loop_opts_igvn(dead);
395 }
396 if (dead->is_InlineType()) {
397 remove_inline_type(dead);
398 }
399 if (dead->is_Call()) {
400 remove_useless_late_inlines( &_late_inlines, dead);
401 remove_useless_late_inlines( &_string_late_inlines, dead);
402 remove_useless_late_inlines( &_boxing_late_inlines, dead);
403 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
404
405 if (dead->is_CallStaticJava()) {
406 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
407 }
408 }
409 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
410 bs->unregister_potential_barrier_node(dead);
411 }
412
413 // Disconnect all useless nodes by disconnecting those at the boundary.
414 void Compile::disconnect_useless_nodes(Unique_Node_List &useful, Unique_Node_List* worklist) {
415 uint next = 0;
416 while (next < useful.size()) {
417 Node *n = useful.at(next++);
418 if (n->is_SafePoint()) {
419 // We're done with a parsing phase. Replaced nodes are not valid
420 // beyond that point.
421 n->as_SafePoint()->delete_replaced_nodes();
422 }
423 // Use raw traversal of out edges since this code removes out edges
424 int max = n->outcnt();
425 for (int j = 0; j < max; ++j) {
426 Node* child = n->raw_out(j);
427 if (!useful.member(child)) {
428 assert(!child->is_top() || child != top(),
429 "If top is cached in Compile object it is in useful list");
430 // Only need to remove this out-edge to the useless node
431 n->raw_del_out(j);
432 --j;
433 --max;
434 }
435 }
436 if (n->outcnt() == 1 && n->has_special_unique_user()) {
437 worklist->push(n->unique_out());
438 }
439 if (n->outcnt() == 0) {
440 worklist->push(n);
441 }
442 }
443
444 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
445 remove_useless_nodes(_predicate_opaqs, useful); // remove useless predicate opaque nodes
446 remove_useless_nodes(_skeleton_predicate_opaqs, useful);
447 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
448 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
449 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
450 #ifdef ASSERT
451 if (_modified_nodes != NULL) {
452 _modified_nodes->remove_useless_nodes(useful.member_set());
453 }
454 #endif
455 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
456 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
457 #ifdef ASSERT
458 if (_modified_nodes != NULL) {
459 _modified_nodes->remove_useless_nodes(useful.member_set());
460 }
461 #endif
462
463 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
464 bs->eliminate_useless_gc_barriers(useful, this);
465 // clean up the late inline lists
466 remove_useless_late_inlines( &_late_inlines, useful);
467 remove_useless_late_inlines( &_string_late_inlines, useful);
468 remove_useless_late_inlines( &_boxing_late_inlines, useful);
469 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
470 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
471 }
472
473 // ============================================================================
474 //------------------------------CompileWrapper---------------------------------
475 class CompileWrapper : public StackObj {
476 Compile *const _compile;
477 public:
478 CompileWrapper(Compile* compile);
479
480 ~CompileWrapper();
481 };
482
483 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
484 // the Compile* pointer is stored in the current ciEnv:
485 ciEnv* env = compile->env();
486 assert(env == ciEnv::current(), "must already be a ciEnv active");
487 assert(env->compiler_data() == NULL, "compile already active?");
488 env->set_compiler_data(compile);
489 assert(compile == Compile::current(), "sanity");
490
491 compile->set_type_dict(NULL);
492 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
493 compile->clone_map().set_clone_idx(0);
494 compile->set_type_last_size(0);
495 compile->set_last_tf(NULL, NULL);
496 compile->set_indexSet_arena(NULL);
497 compile->set_indexSet_free_block_list(NULL);
498 compile->init_type_arena();
499 Type::Initialize(compile);
500 _compile->begin_method();
501 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
502 }
503 CompileWrapper::~CompileWrapper() {
504 // simulate crash during compilation
505 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
506
507 _compile->end_method();
508 _compile->env()->set_compiler_data(NULL);
509 }
510
511
512 //----------------------------print_compile_messages---------------------------
513 void Compile::print_compile_messages() {
514 #ifndef PRODUCT
515 // Check if recompiling
516 if (!subsume_loads() && PrintOpto) {
517 // Recompiling without allowing machine instructions to subsume loads
518 tty->print_cr("*********************************************************");
519 tty->print_cr("** Bailout: Recompile without subsuming loads **");
520 tty->print_cr("*********************************************************");
521 }
522 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
523 // Recompiling without escape analysis
524 tty->print_cr("*********************************************************");
525 tty->print_cr("** Bailout: Recompile without escape analysis **");
526 tty->print_cr("*********************************************************");
527 }
528 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
529 // Recompiling without iterative escape analysis
530 tty->print_cr("*********************************************************");
531 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
532 tty->print_cr("*********************************************************");
533 }
534 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
535 // Recompiling without boxing elimination
536 tty->print_cr("*********************************************************");
537 tty->print_cr("** Bailout: Recompile without boxing elimination **");
538 tty->print_cr("*********************************************************");
539 }
540 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
541 // Recompiling without locks coarsening
542 tty->print_cr("*********************************************************");
543 tty->print_cr("** Bailout: Recompile without locks coarsening **");
544 tty->print_cr("*********************************************************");
545 }
546 if (env()->break_at_compile()) {
547 // Open the debugger when compiling this method.
548 tty->print("### Breaking when compiling: ");
549 method()->print_short_name();
550 tty->cr();
551 BREAKPOINT;
552 }
553
554 if( PrintOpto ) {
555 if (is_osr_compilation()) {
556 tty->print("[OSR]%3d", _compile_id);
557 } else {
558 tty->print("%3d", _compile_id);
559 }
560 }
561 #endif
562 }
563
564 #ifndef PRODUCT
565 void Compile::print_ideal_ir(const char* phase_name) {
566 ttyLocker ttyl;
567 // keep the following output all in one block
568 // This output goes directly to the tty, not the compiler log.
569 // To enable tools to match it up with the compilation activity,
570 // be sure to tag this tty output with the compile ID.
571 if (xtty != NULL) {
572 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
573 compile_id(),
574 is_osr_compilation() ? " compile_kind='osr'" : "",
575 phase_name);
576 }
577 if (_output == nullptr) {
578 tty->print_cr("AFTER: %s", phase_name);
579 // Print out all nodes in ascending order of index.
580 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$");
581 } else {
582 // Dump the node blockwise if we have a scheduling
583 _output->print_scheduling();
584 }
585
586 if (xtty != NULL) {
587 xtty->tail("ideal");
588 }
589 }
590 #endif
591
592 // ============================================================================
593 //------------------------------Compile standard-------------------------------
594 debug_only( int Compile::_debug_idx = 100000; )
595
596 // Compile a method. entry_bci is -1 for normal compilations and indicates
597 // the continuation bci for on stack replacement.
598
599
600 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
601 Options options, DirectiveSet* directive)
602 : Phase(Compiler),
603 _compile_id(ci_env->compile_id()),
604 _options(options),
605 _method(target),
606 _entry_bci(osr_bci),
607 _ilt(NULL),
608 _stub_function(NULL),
609 _stub_name(NULL),
610 _stub_entry_point(NULL),
611 _max_node_limit(MaxNodeLimit),
612 _post_loop_opts_phase(false),
613 _inlining_progress(false),
614 _inlining_incrementally(false),
615 _do_cleanup(false),
616 _has_reserved_stack_access(target->has_reserved_stack_access()),
617 _has_circular_inline_type(false),
618 #ifndef PRODUCT
619 _igv_idx(0),
620 _trace_opto_output(directive->TraceOptoOutputOption),
621 #endif
622 _has_method_handle_invokes(false),
623 _clinit_barrier_on_entry(false),
624 _stress_seed(0),
625 _comp_arena(mtCompiler),
626 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
627 _env(ci_env),
628 _directive(directive),
629 _log(ci_env->log()),
630 _failure_reason(NULL),
631 _intrinsics (comp_arena(), 0, 0, NULL),
632 _macro_nodes (comp_arena(), 8, 0, NULL),
633 _predicate_opaqs (comp_arena(), 8, 0, NULL),
634 _skeleton_predicate_opaqs (comp_arena(), 8, 0, NULL),
635 _expensive_nodes (comp_arena(), 8, 0, NULL),
636 _for_post_loop_igvn(comp_arena(), 8, 0, NULL),
637 _inline_type_nodes (comp_arena(), 8, 0, NULL),
638 _unstable_if_traps (comp_arena(), 8, 0, NULL),
639 _coarsened_locks (comp_arena(), 8, 0, NULL),
640 _congraph(NULL),
641 NOT_PRODUCT(_igv_printer(NULL) COMMA)
642 _dead_node_list(comp_arena()),
643 _dead_node_count(0),
644 _node_arena(mtCompiler),
645 _old_arena(mtCompiler),
646 _mach_constant_base_node(NULL),
647 _Compile_types(mtCompiler),
648 _initial_gvn(NULL),
649 _for_igvn(NULL),
650 _late_inlines(comp_arena(), 2, 0, NULL),
651 _string_late_inlines(comp_arena(), 2, 0, NULL),
652 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
653 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL),
654 _late_inlines_pos(0),
655 _number_of_mh_late_inlines(0),
656 _print_inlining_stream(new (mtCompiler) stringStream()),
657 _print_inlining_list(NULL),
658 _print_inlining_idx(0),
659 _print_inlining_output(NULL),
660 _replay_inline_data(NULL),
661 _java_calls(0),
662 _inner_loops(0),
663 _interpreter_frame_size(0),
664 _output(NULL)
665 #ifndef PRODUCT
666 , _in_dump_cnt(0)
667 #endif
668 {
669 C = this;
670 CompileWrapper cw(this);
671
672 if (CITimeVerbose) {
673 tty->print(" ");
674 target->holder()->name()->print();
675 tty->print(".");
676 target->print_short_name();
677 tty->print(" ");
678 }
679 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
680 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
681
682 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
683 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
684 // We can always print a disassembly, either abstract (hex dump) or
685 // with the help of a suitable hsdis library. Thus, we should not
686 // couple print_assembly and print_opto_assembly controls.
687 // But: always print opto and regular assembly on compile command 'print'.
688 bool print_assembly = directive->PrintAssemblyOption;
689 set_print_assembly(print_opto_assembly || print_assembly);
690 #else
691 set_print_assembly(false); // must initialize.
692 #endif
693
694 #ifndef PRODUCT
695 set_parsed_irreducible_loop(false);
696
697 if (directive->ReplayInlineOption) {
698 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
699 }
700 #endif
701 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
702 set_print_intrinsics(directive->PrintIntrinsicsOption);
703 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
704
705 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
706 // Make sure the method being compiled gets its own MDO,
707 // so we can at least track the decompile_count().
708 // Need MDO to record RTM code generation state.
709 method()->ensure_method_data();
710 }
711
712 Init(/*do_aliasing=*/ true);
713
714 print_compile_messages();
715
716 _ilt = InlineTree::build_inline_tree_root();
717
718 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
719 assert(num_alias_types() >= AliasIdxRaw, "");
720
721 #define MINIMUM_NODE_HASH 1023
722 // Node list that Iterative GVN will start with
723 Unique_Node_List for_igvn(comp_arena());
724 set_for_igvn(&for_igvn);
725
726 // GVN that will be run immediately on new nodes
727 uint estimated_size = method()->code_size()*4+64;
728 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
729 PhaseGVN gvn(node_arena(), estimated_size);
730 set_initial_gvn(&gvn);
731
732 print_inlining_init();
733 { // Scope for timing the parser
734 TracePhase tp("parse", &timers[_t_parser]);
735
736 // Put top into the hash table ASAP.
737 initial_gvn()->transform_no_reclaim(top());
738
739 // Set up tf(), start(), and find a CallGenerator.
740 CallGenerator* cg = NULL;
741 if (is_osr_compilation()) {
742 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
743 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
744 initial_gvn()->set_type_bottom(s);
745 init_start(s);
746 cg = CallGenerator::for_osr(method(), entry_bci());
747 } else {
748 // Normal case.
749 init_tf(TypeFunc::make(method()));
750 StartNode* s = new StartNode(root(), tf()->domain_cc());
751 initial_gvn()->set_type_bottom(s);
752 init_start(s);
753 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
754 // With java.lang.ref.reference.get() we must go through the
755 // intrinsic - even when get() is the root
756 // method of the compile - so that, if necessary, the value in
757 // the referent field of the reference object gets recorded by
758 // the pre-barrier code.
759 cg = find_intrinsic(method(), false);
760 }
761 if (cg == NULL) {
762 float past_uses = method()->interpreter_invocation_count();
763 float expected_uses = past_uses;
764 cg = CallGenerator::for_inline(method(), expected_uses);
765 }
766 }
767 if (failing()) return;
768 if (cg == NULL) {
769 record_method_not_compilable("cannot parse method");
770 return;
771 }
772 JVMState* jvms = build_start_state(start(), tf());
773 if ((jvms = cg->generate(jvms)) == NULL) {
774 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
775 record_method_not_compilable("method parse failed");
776 }
777 return;
778 }
779 GraphKit kit(jvms);
780
781 if (!kit.stopped()) {
782 // Accept return values, and transfer control we know not where.
783 // This is done by a special, unique ReturnNode bound to root.
784 return_values(kit.jvms());
785 }
786
787 if (kit.has_exceptions()) {
788 // Any exceptions that escape from this call must be rethrown
789 // to whatever caller is dynamically above us on the stack.
790 // This is done by a special, unique RethrowNode bound to root.
791 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
792 }
793
794 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
795
796 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
797 inline_string_calls(true);
798 }
799
800 if (failing()) return;
801
802 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
803
804 // Remove clutter produced by parsing.
805 if (!failing()) {
806 ResourceMark rm;
807 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
808 }
809 }
810
811 // Note: Large methods are capped off in do_one_bytecode().
812 if (failing()) return;
813
814 // After parsing, node notes are no longer automagic.
815 // They must be propagated by register_new_node_with_optimizer(),
816 // clone(), or the like.
817 set_default_node_notes(NULL);
818
819 #ifndef PRODUCT
820 if (should_print_igv(1)) {
821 _igv_printer->print_inlining();
822 }
823 #endif
824
825 if (failing()) return;
826 NOT_PRODUCT( verify_graph_edges(); )
827
828 // If any phase is randomized for stress testing, seed random number
829 // generation and log the seed for repeatability.
830 if (StressLCM || StressGCM || StressIGVN || StressCCP) {
831 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) {
832 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
833 FLAG_SET_ERGO(StressSeed, _stress_seed);
834 } else {
835 _stress_seed = StressSeed;
836 }
837 if (_log != NULL) {
838 _log->elem("stress_test seed='%u'", _stress_seed);
839 }
840 }
841
842 // Now optimize
843 Optimize();
844 if (failing()) return;
845 NOT_PRODUCT( verify_graph_edges(); )
846
847 #ifndef PRODUCT
848 if (should_print_ideal()) {
849 print_ideal_ir("print_ideal");
850 }
851 #endif
852
853 #ifdef ASSERT
854 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
855 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
856 #endif
857
858 // Dump compilation data to replay it.
859 if (directive->DumpReplayOption) {
860 env()->dump_replay_data(_compile_id);
861 }
862 if (directive->DumpInlineOption && (ilt() != NULL)) {
863 env()->dump_inline_data(_compile_id);
864 }
865
866 // Now that we know the size of all the monitors we can add a fixed slot
867 // for the original deopt pc.
868 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
869 if (needs_stack_repair()) {
870 // One extra slot for the special stack increment value
871 next_slot += 2;
872 }
873 // TODO 8284443 Only reserve extra slot if needed
874 if (InlineTypeReturnedAsFields) {
875 // One extra slot to hold the IsInit information for a nullable
876 // inline type return if we run out of registers.
877 next_slot += 2;
878 }
879 set_fixed_slots(next_slot);
880
881 // Compute when to use implicit null checks. Used by matching trap based
882 // nodes and NullCheck optimization.
883 set_allowed_deopt_reasons();
884
885 // Now generate code
886 Code_Gen();
887 }
888
889 //------------------------------Compile----------------------------------------
890 // Compile a runtime stub
891 Compile::Compile( ciEnv* ci_env,
892 TypeFunc_generator generator,
893 address stub_function,
894 const char *stub_name,
895 int is_fancy_jump,
896 bool pass_tls,
897 bool return_pc,
898 DirectiveSet* directive)
899 : Phase(Compiler),
900 _compile_id(0),
901 _options(Options::for_runtime_stub()),
902 _method(NULL),
903 _entry_bci(InvocationEntryBci),
904 _stub_function(stub_function),
905 _stub_name(stub_name),
906 _stub_entry_point(NULL),
907 _max_node_limit(MaxNodeLimit),
908 _post_loop_opts_phase(false),
909 _inlining_progress(false),
910 _inlining_incrementally(false),
911 _has_reserved_stack_access(false),
912 _has_circular_inline_type(false),
913 #ifndef PRODUCT
914 _igv_idx(0),
915 _trace_opto_output(directive->TraceOptoOutputOption),
916 #endif
917 _has_method_handle_invokes(false),
918 _clinit_barrier_on_entry(false),
919 _stress_seed(0),
920 _comp_arena(mtCompiler),
921 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
922 _env(ci_env),
923 _directive(directive),
924 _log(ci_env->log()),
925 _failure_reason(NULL),
926 _congraph(NULL),
927 NOT_PRODUCT(_igv_printer(NULL) COMMA)
928 _dead_node_list(comp_arena()),
929 _dead_node_count(0),
930 _node_arena(mtCompiler),
931 _old_arena(mtCompiler),
932 _mach_constant_base_node(NULL),
933 _Compile_types(mtCompiler),
934 _initial_gvn(NULL),
935 _for_igvn(NULL),
936 _number_of_mh_late_inlines(0),
937 _print_inlining_stream(new (mtCompiler) stringStream()),
938 _print_inlining_list(NULL),
939 _print_inlining_idx(0),
940 _print_inlining_output(NULL),
941 _replay_inline_data(NULL),
942 _java_calls(0),
943 _inner_loops(0),
944 _interpreter_frame_size(0),
945 _output(NULL),
946 #ifndef PRODUCT
947 _in_dump_cnt(0),
948 #endif
949 _allowed_reasons(0) {
950 C = this;
951
952 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
953 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
954
955 #ifndef PRODUCT
956 set_print_assembly(PrintFrameConverterAssembly);
957 set_parsed_irreducible_loop(false);
958 #else
959 set_print_assembly(false); // Must initialize.
960 #endif
961 set_has_irreducible_loop(false); // no loops
962
963 CompileWrapper cw(this);
964 Init(/*do_aliasing=*/ false);
965 init_tf((*generator)());
966
967 {
968 // The following is a dummy for the sake of GraphKit::gen_stub
969 Unique_Node_List for_igvn(comp_arena());
970 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
971 PhaseGVN gvn(Thread::current()->resource_area(),255);
972 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
973 gvn.transform_no_reclaim(top());
974
975 GraphKit kit;
976 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
977 }
978
979 NOT_PRODUCT( verify_graph_edges(); )
980
981 Code_Gen();
982 }
983
984 //------------------------------Init-------------------------------------------
985 // Prepare for a single compilation
986 void Compile::Init(bool aliasing) {
987 _do_aliasing = aliasing;
988 _unique = 0;
989 _regalloc = NULL;
990
991 _tf = NULL; // filled in later
992 _top = NULL; // cached later
993 _matcher = NULL; // filled in later
994 _cfg = NULL; // filled in later
995
996 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
997
998 _node_note_array = NULL;
999 _default_node_notes = NULL;
1000 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1001
1002 _immutable_memory = NULL; // filled in at first inquiry
1003
1004 // Globally visible Nodes
1005 // First set TOP to NULL to give safe behavior during creation of RootNode
1006 set_cached_top_node(NULL);
1007 set_root(new RootNode());
1008 // Now that you have a Root to point to, create the real TOP
1009 set_cached_top_node( new ConNode(Type::TOP) );
1010 set_recent_alloc(NULL, NULL);
1011
1012 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1013 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1014 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1015 env()->set_dependencies(new Dependencies(env()));
1016
1017 _fixed_slots = 0;
1018 set_has_split_ifs(false);
1019 set_has_loops(false); // first approximation
1020 set_has_stringbuilder(false);
1021 set_has_boxed_value(false);
1022 _trap_can_recompile = false; // no traps emitted yet
1023 _major_progress = true; // start out assuming good things will happen
1024 set_has_unsafe_access(false);
1025 set_max_vector_size(0);
1026 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1027 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1028 set_decompile_count(0);
1029
1030 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1031 _loop_opts_cnt = LoopOptsCount;
1032 _has_flattened_accesses = false;
1033 _flattened_accesses_share_alias = true;
1034 _scalarize_in_safepoints = false;
1035
1036 set_do_inlining(Inline);
1037 set_max_inline_size(MaxInlineSize);
1038 set_freq_inline_size(FreqInlineSize);
1039 set_do_scheduling(OptoScheduling);
1040
1041 set_do_vector_loop(false);
1042 set_has_monitors(false);
1043
1044 if (AllowVectorizeOnDemand) {
1045 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1046 set_do_vector_loop(true);
1047 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());})
1048 } else if (has_method() && method()->name() != 0 &&
1049 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1050 set_do_vector_loop(true);
1051 }
1052 }
1053 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1054 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());})
1055
1056 set_rtm_state(NoRTM); // No RTM lock eliding by default
1057 _max_node_limit = _directive->MaxNodeLimitOption;
1058
1059 #if INCLUDE_RTM_OPT
1060 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1061 int rtm_state = method()->method_data()->rtm_state();
1062 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1063 // Don't generate RTM lock eliding code.
1064 set_rtm_state(NoRTM);
1065 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1066 // Generate RTM lock eliding code without abort ratio calculation code.
1067 set_rtm_state(UseRTM);
1068 } else if (UseRTMDeopt) {
1069 // Generate RTM lock eliding code and include abort ratio calculation
1070 // code if UseRTMDeopt is on.
1071 set_rtm_state(ProfileRTM);
1072 }
1073 }
1074 #endif
1075 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1076 set_clinit_barrier_on_entry(true);
1077 }
1078 if (debug_info()->recording_non_safepoints()) {
1079 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1080 (comp_arena(), 8, 0, NULL));
1081 set_default_node_notes(Node_Notes::make(this));
1082 }
1083
1084 const int grow_ats = 16;
1085 _max_alias_types = grow_ats;
1086 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1087 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1088 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1089 {
1090 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1091 }
1092 // Initialize the first few types.
1093 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1094 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1095 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1096 _num_alias_types = AliasIdxRaw+1;
1097 // Zero out the alias type cache.
1098 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1099 // A NULL adr_type hits in the cache right away. Preload the right answer.
1100 probe_alias_cache(NULL)->_index = AliasIdxTop;
1101
1102 #ifdef ASSERT
1103 _type_verify_symmetry = true;
1104 _phase_optimize_finished = false;
1105 _exception_backedge = false;
1106 #endif
1107 }
1108
1109 //---------------------------init_start----------------------------------------
1110 // Install the StartNode on this compile object.
1111 void Compile::init_start(StartNode* s) {
1112 if (failing())
1113 return; // already failing
1114 assert(s == start(), "");
1115 }
1116
1117 /**
1118 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1119 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1120 * the ideal graph.
1121 */
1122 StartNode* Compile::start() const {
1123 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1124 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1125 Node* start = root()->fast_out(i);
1126 if (start->is_Start()) {
1127 return start->as_Start();
1128 }
1129 }
1130 fatal("Did not find Start node!");
1131 return NULL;
1132 }
1133
1134 //-------------------------------immutable_memory-------------------------------------
1135 // Access immutable memory
1136 Node* Compile::immutable_memory() {
1137 if (_immutable_memory != NULL) {
1138 return _immutable_memory;
1139 }
1140 StartNode* s = start();
1141 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1142 Node *p = s->fast_out(i);
1143 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1144 _immutable_memory = p;
1145 return _immutable_memory;
1146 }
1147 }
1148 ShouldNotReachHere();
1149 return NULL;
1150 }
1151
1152 //----------------------set_cached_top_node------------------------------------
1153 // Install the cached top node, and make sure Node::is_top works correctly.
1154 void Compile::set_cached_top_node(Node* tn) {
1155 if (tn != NULL) verify_top(tn);
1156 Node* old_top = _top;
1157 _top = tn;
1158 // Calling Node::setup_is_top allows the nodes the chance to adjust
1159 // their _out arrays.
1160 if (_top != NULL) _top->setup_is_top();
1161 if (old_top != NULL) old_top->setup_is_top();
1162 assert(_top == NULL || top()->is_top(), "");
1163 }
1164
1165 #ifdef ASSERT
1166 uint Compile::count_live_nodes_by_graph_walk() {
1167 Unique_Node_List useful(comp_arena());
1168 // Get useful node list by walking the graph.
1169 identify_useful_nodes(useful);
1170 return useful.size();
1171 }
1172
1173 void Compile::print_missing_nodes() {
1174
1175 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1176 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1177 return;
1178 }
1179
1180 // This is an expensive function. It is executed only when the user
1181 // specifies VerifyIdealNodeCount option or otherwise knows the
1182 // additional work that needs to be done to identify reachable nodes
1183 // by walking the flow graph and find the missing ones using
1184 // _dead_node_list.
1185
1186 Unique_Node_List useful(comp_arena());
1187 // Get useful node list by walking the graph.
1188 identify_useful_nodes(useful);
1189
1190 uint l_nodes = C->live_nodes();
1191 uint l_nodes_by_walk = useful.size();
1192
1193 if (l_nodes != l_nodes_by_walk) {
1194 if (_log != NULL) {
1195 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1196 _log->stamp();
1197 _log->end_head();
1198 }
1199 VectorSet& useful_member_set = useful.member_set();
1200 int last_idx = l_nodes_by_walk;
1201 for (int i = 0; i < last_idx; i++) {
1202 if (useful_member_set.test(i)) {
1203 if (_dead_node_list.test(i)) {
1204 if (_log != NULL) {
1205 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1206 }
1207 if (PrintIdealNodeCount) {
1208 // Print the log message to tty
1209 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1210 useful.at(i)->dump();
1211 }
1212 }
1213 }
1214 else if (! _dead_node_list.test(i)) {
1215 if (_log != NULL) {
1216 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1217 }
1218 if (PrintIdealNodeCount) {
1219 // Print the log message to tty
1220 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1221 }
1222 }
1223 }
1224 if (_log != NULL) {
1225 _log->tail("mismatched_nodes");
1226 }
1227 }
1228 }
1229 void Compile::record_modified_node(Node* n) {
1230 if (_modified_nodes != NULL && !_inlining_incrementally && !n->is_Con()) {
1231 _modified_nodes->push(n);
1232 }
1233 }
1234
1235 void Compile::remove_modified_node(Node* n) {
1236 if (_modified_nodes != NULL) {
1237 _modified_nodes->remove(n);
1238 }
1239 }
1240 #endif
1241
1242 #ifndef PRODUCT
1243 void Compile::verify_top(Node* tn) const {
1244 if (tn != NULL) {
1245 assert(tn->is_Con(), "top node must be a constant");
1246 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1247 assert(tn->in(0) != NULL, "must have live top node");
1248 }
1249 }
1250 #endif
1251
1252
1253 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1254
1255 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1256 guarantee(arr != NULL, "");
1257 int num_blocks = arr->length();
1258 if (grow_by < num_blocks) grow_by = num_blocks;
1259 int num_notes = grow_by * _node_notes_block_size;
1260 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1261 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1262 while (num_notes > 0) {
1263 arr->append(notes);
1264 notes += _node_notes_block_size;
1265 num_notes -= _node_notes_block_size;
1266 }
1267 assert(num_notes == 0, "exact multiple, please");
1268 }
1269
1270 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1271 if (source == NULL || dest == NULL) return false;
1272
1273 if (dest->is_Con())
1274 return false; // Do not push debug info onto constants.
1275
1276 #ifdef ASSERT
1277 // Leave a bread crumb trail pointing to the original node:
1278 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1279 dest->set_debug_orig(source);
1280 }
1281 #endif
1282
1283 if (node_note_array() == NULL)
1284 return false; // Not collecting any notes now.
1285
1286 // This is a copy onto a pre-existing node, which may already have notes.
1287 // If both nodes have notes, do not overwrite any pre-existing notes.
1288 Node_Notes* source_notes = node_notes_at(source->_idx);
1289 if (source_notes == NULL || source_notes->is_clear()) return false;
1290 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1291 if (dest_notes == NULL || dest_notes->is_clear()) {
1292 return set_node_notes_at(dest->_idx, source_notes);
1293 }
1294
1295 Node_Notes merged_notes = (*source_notes);
1296 // The order of operations here ensures that dest notes will win...
1297 merged_notes.update_from(dest_notes);
1298 return set_node_notes_at(dest->_idx, &merged_notes);
1299 }
1300
1301
1302 //--------------------------allow_range_check_smearing-------------------------
1303 // Gating condition for coalescing similar range checks.
1304 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1305 // single covering check that is at least as strong as any of them.
1306 // If the optimization succeeds, the simplified (strengthened) range check
1307 // will always succeed. If it fails, we will deopt, and then give up
1308 // on the optimization.
1309 bool Compile::allow_range_check_smearing() const {
1310 // If this method has already thrown a range-check,
1311 // assume it was because we already tried range smearing
1312 // and it failed.
1313 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1314 return !already_trapped;
1315 }
1316
1317
1318 //------------------------------flatten_alias_type-----------------------------
1319 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1320 assert(do_aliasing(), "Aliasing should be enabled");
1321 int offset = tj->offset();
1322 TypePtr::PTR ptr = tj->ptr();
1323
1324 // Known instance (scalarizable allocation) alias only with itself.
1325 bool is_known_inst = tj->isa_oopptr() != NULL &&
1326 tj->is_oopptr()->is_known_instance();
1327
1328 // Process weird unsafe references.
1329 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1330 bool default_value_load = EnableValhalla && tj->is_instptr()->instance_klass() == ciEnv::current()->Class_klass();
1331 assert(InlineUnsafeOps || StressReflectiveCode || default_value_load, "indeterminate pointers come only from unsafe ops");
1332 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1333 tj = TypeOopPtr::BOTTOM;
1334 ptr = tj->ptr();
1335 offset = tj->offset();
1336 }
1337
1338 // Array pointers need some flattening
1339 const TypeAryPtr* ta = tj->isa_aryptr();
1340 if (ta && ta->is_stable()) {
1341 // Erase stability property for alias analysis.
1342 tj = ta = ta->cast_to_stable(false);
1343 }
1344 if (ta && ta->is_not_flat()) {
1345 // Erase not flat property for alias analysis.
1346 tj = ta = ta->cast_to_not_flat(false);
1347 }
1348 if (ta && ta->is_not_null_free()) {
1349 // Erase not null free property for alias analysis.
1350 tj = ta = ta->cast_to_not_null_free(false);
1351 }
1352
1353 if( ta && is_known_inst ) {
1354 if ( offset != Type::OffsetBot &&
1355 offset > arrayOopDesc::length_offset_in_bytes() ) {
1356 offset = Type::OffsetBot; // Flatten constant access into array body only
1357 tj = ta = ta->
1358 remove_speculative()->
1359 cast_to_ptr_type(ptr)->
1360 with_offset(offset);
1361 }
1362 } else if (ta) {
1363 // For arrays indexed by constant indices, we flatten the alias
1364 // space to include all of the array body. Only the header, klass
1365 // and array length can be accessed un-aliased.
1366 // For flattened inline type array, each field has its own slice so
1367 // we must include the field offset.
1368 if( offset != Type::OffsetBot ) {
1369 if( ta->const_oop() ) { // MethodData* or Method*
1370 offset = Type::OffsetBot; // Flatten constant access into array body
1371 tj = ta = ta->
1372 remove_speculative()->
1373 cast_to_ptr_type(ptr)->
1374 cast_to_exactness(false)->
1375 with_offset(offset);
1376 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1377 // range is OK as-is.
1378 tj = ta = TypeAryPtr::RANGE;
1379 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1380 tj = TypeInstPtr::KLASS; // all klass loads look alike
1381 ta = TypeAryPtr::RANGE; // generic ignored junk
1382 ptr = TypePtr::BotPTR;
1383 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1384 tj = TypeInstPtr::MARK;
1385 ta = TypeAryPtr::RANGE; // generic ignored junk
1386 ptr = TypePtr::BotPTR;
1387 } else { // Random constant offset into array body
1388 offset = Type::OffsetBot; // Flatten constant access into array body
1389 tj = ta = ta->
1390 remove_speculative()->
1391 cast_to_ptr_type(ptr)->
1392 cast_to_exactness(false)->
1393 with_offset(offset);
1394 }
1395 }
1396 // Arrays of fixed size alias with arrays of unknown size.
1397 if (ta->size() != TypeInt::POS) {
1398 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1399 tj = ta = ta->
1400 remove_speculative()->
1401 cast_to_ptr_type(ptr)->
1402 with_ary(tary)->
1403 cast_to_exactness(false);
1404 }
1405 // Arrays of known objects become arrays of unknown objects.
1406 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1407 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1408 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1409 }
1410 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1411 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1412 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1413 }
1414 // Initially all flattened array accesses share a single slice
1415 if (ta->is_flat() && ta->elem() != TypeInstPtr::BOTTOM && _flattened_accesses_share_alias) {
1416 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size(), /* stable= */ false, /* flat= */ true);
1417 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1418 }
1419 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1420 // cannot be distinguished by bytecode alone.
1421 if (ta->elem() == TypeInt::BOOL) {
1422 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1423 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1424 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1425 }
1426 // During the 2nd round of IterGVN, NotNull castings are removed.
1427 // Make sure the Bottom and NotNull variants alias the same.
1428 // Also, make sure exact and non-exact variants alias the same.
1429 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1430 tj = ta = ta->
1431 remove_speculative()->
1432 cast_to_ptr_type(TypePtr::BotPTR)->
1433 cast_to_exactness(false)->
1434 with_offset(offset);
1435 }
1436 }
1437
1438 // Oop pointers need some flattening
1439 const TypeInstPtr *to = tj->isa_instptr();
1440 if (to && to != TypeOopPtr::BOTTOM) {
1441 ciInstanceKlass* ik = to->instance_klass();
1442 if( ptr == TypePtr::Constant ) {
1443 if (ik != ciEnv::current()->Class_klass() ||
1444 offset < ik->layout_helper_size_in_bytes()) {
1445 // No constant oop pointers (such as Strings); they alias with
1446 // unknown strings.
1447 assert(!is_known_inst, "not scalarizable allocation");
1448 tj = to = to->
1449 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1450 remove_speculative()->
1451 cast_to_ptr_type(TypePtr::BotPTR)->
1452 cast_to_exactness(false);
1453 }
1454 } else if( is_known_inst ) {
1455 tj = to; // Keep NotNull and klass_is_exact for instance type
1456 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1457 // During the 2nd round of IterGVN, NotNull castings are removed.
1458 // Make sure the Bottom and NotNull variants alias the same.
1459 // Also, make sure exact and non-exact variants alias the same.
1460 tj = to = to->
1461 remove_speculative()->
1462 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1463 cast_to_ptr_type(TypePtr::BotPTR)->
1464 cast_to_exactness(false);
1465 }
1466 if (to->speculative() != NULL) {
1467 tj = to = to->remove_speculative();
1468 }
1469 // Canonicalize the holder of this field
1470 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1471 // First handle header references such as a LoadKlassNode, even if the
1472 // object's klass is unloaded at compile time (4965979).
1473 if (!is_known_inst) { // Do it only for non-instance types
1474 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset));
1475 }
1476 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1477 // Static fields are in the space above the normal instance
1478 // fields in the java.lang.Class instance.
1479 if (ik != ciEnv::current()->Class_klass()) {
1480 to = NULL;
1481 tj = TypeOopPtr::BOTTOM;
1482 offset = tj->offset();
1483 }
1484 } else {
1485 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1486 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1487 if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1488 if( is_known_inst ) {
1489 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id());
1490 } else {
1491 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset));
1492 }
1493 }
1494 }
1495 }
1496
1497 // Klass pointers to object array klasses need some flattening
1498 const TypeKlassPtr *tk = tj->isa_klassptr();
1499 if( tk ) {
1500 // If we are referencing a field within a Klass, we need
1501 // to assume the worst case of an Object. Both exact and
1502 // inexact types must flatten to the same alias class so
1503 // use NotNull as the PTR.
1504 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1505 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1506 env()->Object_klass(),
1507 Type::Offset(offset));
1508 }
1509
1510 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1511 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1512 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1513 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset));
1514 } else {
1515 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());
1516 }
1517 }
1518 // Check for precise loads from the primary supertype array and force them
1519 // to the supertype cache alias index. Check for generic array loads from
1520 // the primary supertype array and also force them to the supertype cache
1521 // alias index. Since the same load can reach both, we need to merge
1522 // these 2 disparate memories into the same alias class. Since the
1523 // primary supertype array is read-only, there's no chance of confusion
1524 // where we bypass an array load and an array store.
1525 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1526 if (offset == Type::OffsetBot ||
1527 (offset >= primary_supers_offset &&
1528 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1529 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1530 offset = in_bytes(Klass::secondary_super_cache_offset());
1531 tj = tk = tk->with_offset(offset);
1532 }
1533 }
1534
1535 // Flatten all Raw pointers together.
1536 if (tj->base() == Type::RawPtr)
1537 tj = TypeRawPtr::BOTTOM;
1538
1539 if (tj->base() == Type::AnyPtr)
1540 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1541
1542 offset = tj->offset();
1543 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1544
1545 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1546 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1547 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1548 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1549 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1550 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1551 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1552 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1553 assert( tj->ptr() != TypePtr::TopPTR &&
1554 tj->ptr() != TypePtr::AnyNull &&
1555 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1556 // assert( tj->ptr() != TypePtr::Constant ||
1557 // tj->base() == Type::RawPtr ||
1558 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1559
1560 return tj;
1561 }
1562
1563 void Compile::AliasType::Init(int i, const TypePtr* at) {
1564 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1565 _index = i;
1566 _adr_type = at;
1567 _field = NULL;
1568 _element = NULL;
1569 _is_rewritable = true; // default
1570 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1571 if (atoop != NULL && atoop->is_known_instance()) {
1572 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1573 _general_index = Compile::current()->get_alias_index(gt);
1574 } else {
1575 _general_index = 0;
1576 }
1577 }
1578
1579 BasicType Compile::AliasType::basic_type() const {
1580 if (element() != NULL) {
1581 const Type* element = adr_type()->is_aryptr()->elem();
1582 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1583 } if (field() != NULL) {
1584 return field()->layout_type();
1585 } else {
1586 return T_ILLEGAL; // unknown
1587 }
1588 }
1589
1590 //---------------------------------print_on------------------------------------
1591 #ifndef PRODUCT
1592 void Compile::AliasType::print_on(outputStream* st) {
1593 if (index() < 10)
1594 st->print("@ <%d> ", index());
1595 else st->print("@ <%d>", index());
1596 st->print(is_rewritable() ? " " : " RO");
1597 int offset = adr_type()->offset();
1598 if (offset == Type::OffsetBot)
1599 st->print(" +any");
1600 else st->print(" +%-3d", offset);
1601 st->print(" in ");
1602 adr_type()->dump_on(st);
1603 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1604 if (field() != NULL && tjp) {
1605 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1606 tjp->offset() != field()->offset_in_bytes()) {
1607 st->print(" != ");
1608 field()->print();
1609 st->print(" ***");
1610 }
1611 }
1612 }
1613
1614 void print_alias_types() {
1615 Compile* C = Compile::current();
1616 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1617 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1618 C->alias_type(idx)->print_on(tty);
1619 tty->cr();
1620 }
1621 }
1622 #endif
1623
1624
1625 //----------------------------probe_alias_cache--------------------------------
1626 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1627 intptr_t key = (intptr_t) adr_type;
1628 key ^= key >> logAliasCacheSize;
1629 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1630 }
1631
1632
1633 //-----------------------------grow_alias_types--------------------------------
1634 void Compile::grow_alias_types() {
1635 const int old_ats = _max_alias_types; // how many before?
1636 const int new_ats = old_ats; // how many more?
1637 const int grow_ats = old_ats+new_ats; // how many now?
1638 _max_alias_types = grow_ats;
1639 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1640 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1641 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1642 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1643 }
1644
1645
1646 //--------------------------------find_alias_type------------------------------
1647 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1648 if (!do_aliasing()) {
1649 return alias_type(AliasIdxBot);
1650 }
1651
1652 AliasCacheEntry* ace = NULL;
1653 if (!uncached) {
1654 ace = probe_alias_cache(adr_type);
1655 if (ace->_adr_type == adr_type) {
1656 return alias_type(ace->_index);
1657 }
1658 }
1659
1660 // Handle special cases.
1661 if (adr_type == NULL) return alias_type(AliasIdxTop);
1662 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1663
1664 // Do it the slow way.
1665 const TypePtr* flat = flatten_alias_type(adr_type);
1666
1667 #ifdef ASSERT
1668 {
1669 ResourceMark rm;
1670 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1671 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1672 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1673 Type::str(adr_type));
1674 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1675 const TypeOopPtr* foop = flat->is_oopptr();
1676 // Scalarizable allocations have exact klass always.
1677 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1678 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1679 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1680 Type::str(foop), Type::str(xoop));
1681 }
1682 }
1683 #endif
1684
1685 int idx = AliasIdxTop;
1686 for (int i = 0; i < num_alias_types(); i++) {
1687 if (alias_type(i)->adr_type() == flat) {
1688 idx = i;
1689 break;
1690 }
1691 }
1692
1693 if (idx == AliasIdxTop) {
1694 if (no_create) return NULL;
1695 // Grow the array if necessary.
1696 if (_num_alias_types == _max_alias_types) grow_alias_types();
1697 // Add a new alias type.
1698 idx = _num_alias_types++;
1699 _alias_types[idx]->Init(idx, flat);
1700 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1701 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1702 if (flat->isa_instptr()) {
1703 if (flat->offset() == java_lang_Class::klass_offset()
1704 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1705 alias_type(idx)->set_rewritable(false);
1706 }
1707 ciField* field = NULL;
1708 if (flat->isa_aryptr()) {
1709 #ifdef ASSERT
1710 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1711 // (T_BYTE has the weakest alignment and size restrictions...)
1712 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1713 #endif
1714 const Type* elemtype = flat->is_aryptr()->elem();
1715 if (flat->offset() == TypePtr::OffsetBot) {
1716 alias_type(idx)->set_element(elemtype);
1717 }
1718 int field_offset = flat->is_aryptr()->field_offset().get();
1719 if (flat->is_flat() &&
1720 field_offset != Type::OffsetBot) {
1721 ciInlineKlass* vk = elemtype->inline_klass();
1722 field_offset += vk->first_field_offset();
1723 field = vk->get_field_by_offset(field_offset, false);
1724 }
1725 }
1726 if (flat->isa_klassptr()) {
1727 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1728 alias_type(idx)->set_rewritable(false);
1729 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1730 alias_type(idx)->set_rewritable(false);
1731 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1732 alias_type(idx)->set_rewritable(false);
1733 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1734 alias_type(idx)->set_rewritable(false);
1735 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1736 alias_type(idx)->set_rewritable(false);
1737 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1738 alias_type(idx)->set_rewritable(false);
1739 }
1740 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1741 // but the base pointer type is not distinctive enough to identify
1742 // references into JavaThread.)
1743
1744 // Check for final fields.
1745 const TypeInstPtr* tinst = flat->isa_instptr();
1746 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1747 if (tinst->const_oop() != NULL &&
1748 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1749 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1750 // static field
1751 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1752 field = k->get_field_by_offset(tinst->offset(), true);
1753 } else if (tinst->is_inlinetypeptr()) {
1754 // Inline type field
1755 ciInlineKlass* vk = tinst->inline_klass();
1756 field = vk->get_field_by_offset(tinst->offset(), false);
1757 } else {
1758 ciInstanceKlass *k = tinst->instance_klass();
1759 field = k->get_field_by_offset(tinst->offset(), false);
1760 }
1761 }
1762 assert(field == NULL ||
1763 original_field == NULL ||
1764 (field->holder() == original_field->holder() &&
1765 field->offset() == original_field->offset() &&
1766 field->is_static() == original_field->is_static()), "wrong field?");
1767 // Set field() and is_rewritable() attributes.
1768 if (field != NULL) {
1769 alias_type(idx)->set_field(field);
1770 if (flat->isa_aryptr()) {
1771 // Fields of flat arrays are rewritable although they are declared final
1772 assert(flat->is_flat(), "must be a flat array");
1773 alias_type(idx)->set_rewritable(true);
1774 }
1775 }
1776 }
1777
1778 // Fill the cache for next time.
1779 if (!uncached) {
1780 ace->_adr_type = adr_type;
1781 ace->_index = idx;
1782 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1783
1784 // Might as well try to fill the cache for the flattened version, too.
1785 AliasCacheEntry* face = probe_alias_cache(flat);
1786 if (face->_adr_type == NULL) {
1787 face->_adr_type = flat;
1788 face->_index = idx;
1789 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1790 }
1791 }
1792
1793 return alias_type(idx);
1794 }
1795
1796
1797 Compile::AliasType* Compile::alias_type(ciField* field) {
1798 const TypeOopPtr* t;
1799 if (field->is_static())
1800 t = TypeInstPtr::make(field->holder()->java_mirror());
1801 else
1802 t = TypeOopPtr::make_from_klass_raw(field->holder());
1803 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1804 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1805 return atp;
1806 }
1807
1808
1809 //------------------------------have_alias_type--------------------------------
1810 bool Compile::have_alias_type(const TypePtr* adr_type) {
1811 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1812 if (ace->_adr_type == adr_type) {
1813 return true;
1814 }
1815
1816 // Handle special cases.
1817 if (adr_type == NULL) return true;
1818 if (adr_type == TypePtr::BOTTOM) return true;
1819
1820 return find_alias_type(adr_type, true, NULL) != NULL;
1821 }
1822
1823 //-----------------------------must_alias--------------------------------------
1824 // True if all values of the given address type are in the given alias category.
1825 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1826 if (alias_idx == AliasIdxBot) return true; // the universal category
1827 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1828 if (alias_idx == AliasIdxTop) return false; // the empty category
1829 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1830
1831 // the only remaining possible overlap is identity
1832 int adr_idx = get_alias_index(adr_type);
1833 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1834 assert(adr_idx == alias_idx ||
1835 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1836 && adr_type != TypeOopPtr::BOTTOM),
1837 "should not be testing for overlap with an unsafe pointer");
1838 return adr_idx == alias_idx;
1839 }
1840
1841 //------------------------------can_alias--------------------------------------
1842 // True if any values of the given address type are in the given alias category.
1843 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1844 if (alias_idx == AliasIdxTop) return false; // the empty category
1845 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1846 // Known instance doesn't alias with bottom memory
1847 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1848 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1849
1850 // the only remaining possible overlap is identity
1851 int adr_idx = get_alias_index(adr_type);
1852 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1853 return adr_idx == alias_idx;
1854 }
1855
1856 //---------------------cleanup_loop_predicates-----------------------
1857 // Remove the opaque nodes that protect the predicates so that all unused
1858 // checks and uncommon_traps will be eliminated from the ideal graph
1859 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1860 if (predicate_count()==0) return;
1861 for (int i = predicate_count(); i > 0; i--) {
1862 Node * n = predicate_opaque1_node(i-1);
1863 assert(n->Opcode() == Op_Opaque1, "must be");
1864 igvn.replace_node(n, n->in(1));
1865 }
1866 assert(predicate_count()==0, "should be clean!");
1867 }
1868
1869 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1870 if (!n->for_post_loop_opts_igvn()) {
1871 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1872 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1873 _for_post_loop_igvn.append(n);
1874 }
1875 }
1876
1877 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1878 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1879 _for_post_loop_igvn.remove(n);
1880 }
1881
1882 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1883 // Verify that all previous optimizations produced a valid graph
1884 // at least to this point, even if no loop optimizations were done.
1885 PhaseIdealLoop::verify(igvn);
1886
1887 C->set_post_loop_opts_phase(); // no more loop opts allowed
1888
1889 assert(!C->major_progress(), "not cleared");
1890
1891 if (_for_post_loop_igvn.length() > 0) {
1892 while (_for_post_loop_igvn.length() > 0) {
1893 Node* n = _for_post_loop_igvn.pop();
1894 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1895 igvn._worklist.push(n);
1896 }
1897 igvn.optimize();
1898 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1899
1900 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1901 if (C->major_progress()) {
1902 C->clear_major_progress(); // ensure that major progress is now clear
1903 }
1904 }
1905 }
1906
1907 void Compile::add_inline_type(Node* n) {
1908 assert(n->is_InlineType(), "unexpected node");
1909 _inline_type_nodes.push(n);
1910 }
1911
1912 void Compile::remove_inline_type(Node* n) {
1913 assert(n->is_InlineType(), "unexpected node");
1914 if (_inline_type_nodes.contains(n)) {
1915 _inline_type_nodes.remove(n);
1916 }
1917 }
1918
1919 // Does the return value keep otherwise useless inline type allocations alive?
1920 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1921 ResourceMark rm;
1922 Unique_Node_List wq;
1923 wq.push(ret_val);
1924 bool some_allocations = false;
1925 for (uint i = 0; i < wq.size(); i++) {
1926 Node* n = wq.at(i);
1927 assert(n == ret_val || !n->is_InlineType(), "chain of inline type nodes");
1928 if (n->outcnt() > 1) {
1929 // Some other use for the allocation
1930 return false;
1931 } else if (n->is_InlineType()) {
1932 wq.push(n->in(1));
1933 } else if (n->is_Phi()) {
1934 for (uint j = 1; j < n->req(); j++) {
1935 wq.push(n->in(j));
1936 }
1937 } else if (n->is_CheckCastPP() &&
1938 n->in(1)->is_Proj() &&
1939 n->in(1)->in(0)->is_Allocate()) {
1940 some_allocations = true;
1941 } else if (n->is_CheckCastPP()) {
1942 wq.push(n->in(1));
1943 }
1944 }
1945 return some_allocations;
1946 }
1947
1948 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1949 // Make sure that the return value does not keep an otherwise unused allocation alive
1950 if (tf()->returns_inline_type_as_fields()) {
1951 Node* ret = NULL;
1952 for (uint i = 1; i < root()->req(); i++) {
1953 Node* in = root()->in(i);
1954 if (in->Opcode() == Op_Return) {
1955 assert(ret == NULL, "only one return");
1956 ret = in;
1957 }
1958 }
1959 if (ret != NULL) {
1960 Node* ret_val = ret->in(TypeFunc::Parms);
1961 if (igvn.type(ret_val)->isa_oopptr() &&
1962 return_val_keeps_allocations_alive(ret_val)) {
1963 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
1964 assert(ret_val->outcnt() == 0, "should be dead now");
1965 igvn.remove_dead_node(ret_val);
1966 }
1967 }
1968 }
1969 if (_inline_type_nodes.length() == 0) {
1970 return;
1971 }
1972 // Scalarize inline types in safepoint debug info.
1973 // Delay this until all inlining is over to avoid getting inconsistent debug info.
1974 set_scalarize_in_safepoints(true);
1975 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
1976 _inline_type_nodes.at(i)->as_InlineType()->make_scalar_in_safepoints(&igvn);
1977 }
1978 if (remove) {
1979 // Remove inline type nodes by replacing them with their oop input
1980 while (_inline_type_nodes.length() > 0) {
1981 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
1982 if (vt->outcnt() == 0) {
1983 igvn.remove_dead_node(vt);
1984 continue;
1985 }
1986 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
1987 DEBUG_ONLY(bool must_be_buffered = false);
1988 Node* u = vt->out(i);
1989 // Check if any users are blackholes. If so, rewrite them to use either the
1990 // allocated buffer, or individual components, instead of the inline type node
1991 // that goes away.
1992 if (u->is_Blackhole()) {
1993 BlackholeNode* bh = u->as_Blackhole();
1994
1995 // Unlink the old input
1996 int idx = bh->find_edge(vt);
1997 assert(idx != -1, "The edge should be there");
1998 bh->del_req(idx);
1999 --i;
2000
2001 if (vt->is_allocated(&igvn)) {
2002 // Already has the allocated instance, blackhole that
2003 bh->add_req(vt->get_oop());
2004 } else {
2005 // Not allocated yet, blackhole the components
2006 for (uint c = 0; c < vt->field_count(); c++) {
2007 bh->add_req(vt->field_value(c));
2008 }
2009 }
2010
2011 // Node modified, record for IGVN
2012 igvn.record_for_igvn(bh);
2013 }
2014 #ifdef ASSERT
2015 // Verify that inline type is buffered when replacing by oop
2016 else if (u->is_InlineType()) {
2017 InlineTypeNode* vt2 = u->as_InlineType();
2018 for (uint i = 0; i < vt2->field_count(); ++i) {
2019 if (vt2->field_value(i) == vt && !vt2->field_is_flattened(i)) {
2020 // Use in non-flat field
2021 must_be_buffered = true;
2022 }
2023 }
2024 } else if (u->Opcode() != Op_Return || !tf()->returns_inline_type_as_fields()) {
2025 must_be_buffered = true;
2026 }
2027 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2028 vt->dump(-3);
2029 assert(false, "Should have been buffered");
2030 }
2031 #endif
2032 }
2033 igvn.replace_node(vt, vt->get_oop());
2034 }
2035 }
2036 igvn.optimize();
2037 }
2038
2039 void Compile::adjust_flattened_array_access_aliases(PhaseIterGVN& igvn) {
2040 if (!_has_flattened_accesses) {
2041 return;
2042 }
2043 // Initially, all flattened array accesses share the same slice to
2044 // keep dependencies with Object[] array accesses (that could be
2045 // to a flattened array) correct. We're done with parsing so we
2046 // now know all flattened array accesses in this compile
2047 // unit. Let's move flattened array accesses to their own slice,
2048 // one per element field. This should help memory access
2049 // optimizations.
2050 ResourceMark rm;
2051 Unique_Node_List wq;
2052 wq.push(root());
2053
2054 Node_List mergememnodes;
2055 Node_List memnodes;
2056
2057 // Alias index currently shared by all flattened memory accesses
2058 int index = get_alias_index(TypeAryPtr::INLINES);
2059
2060 // Find MergeMem nodes and flattened array accesses
2061 for (uint i = 0; i < wq.size(); i++) {
2062 Node* n = wq.at(i);
2063 if (n->is_Mem()) {
2064 const TypePtr* adr_type = NULL;
2065 if (n->Opcode() == Op_StoreCM) {
2066 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type()));
2067 } else {
2068 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2069 }
2070 if (adr_type == TypeAryPtr::INLINES) {
2071 memnodes.push(n);
2072 }
2073 } else if (n->is_MergeMem()) {
2074 MergeMemNode* mm = n->as_MergeMem();
2075 if (mm->memory_at(index) != mm->base_memory()) {
2076 mergememnodes.push(n);
2077 }
2078 }
2079 for (uint j = 0; j < n->req(); j++) {
2080 Node* m = n->in(j);
2081 if (m != NULL) {
2082 wq.push(m);
2083 }
2084 }
2085 }
2086
2087 if (memnodes.size() > 0) {
2088 _flattened_accesses_share_alias = false;
2089
2090 // We are going to change the slice for the flattened array
2091 // accesses so we need to clear the cache entries that refer to
2092 // them.
2093 for (uint i = 0; i < AliasCacheSize; i++) {
2094 AliasCacheEntry* ace = &_alias_cache[i];
2095 if (ace->_adr_type != NULL &&
2096 ace->_adr_type->is_flat()) {
2097 ace->_adr_type = NULL;
2098 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the NULL adr_type resolves to AliasIdxTop
2099 }
2100 }
2101
2102 // Find what aliases we are going to add
2103 int start_alias = num_alias_types()-1;
2104 int stop_alias = 0;
2105
2106 for (uint i = 0; i < memnodes.size(); i++) {
2107 Node* m = memnodes.at(i);
2108 const TypePtr* adr_type = NULL;
2109 if (m->Opcode() == Op_StoreCM) {
2110 adr_type = m->in(MemNode::OopStore)->adr_type();
2111 if (adr_type != TypeAryPtr::INLINES) {
2112 // store was optimized out and we lost track of the adr_type
2113 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address),
2114 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore),
2115 get_alias_index(adr_type));
2116 igvn.register_new_node_with_optimizer(clone);
2117 igvn.replace_node(m, clone);
2118 }
2119 } else {
2120 adr_type = m->adr_type();
2121 #ifdef ASSERT
2122 m->as_Mem()->set_adr_type(adr_type);
2123 #endif
2124 }
2125 int idx = get_alias_index(adr_type);
2126 start_alias = MIN2(start_alias, idx);
2127 stop_alias = MAX2(stop_alias, idx);
2128 }
2129
2130 assert(stop_alias >= start_alias, "should have expanded aliases");
2131
2132 Node_Stack stack(0);
2133 #ifdef ASSERT
2134 VectorSet seen(Thread::current()->resource_area());
2135 #endif
2136 // Now let's fix the memory graph so each flattened array access
2137 // is moved to the right slice. Start from the MergeMem nodes.
2138 uint last = unique();
2139 for (uint i = 0; i < mergememnodes.size(); i++) {
2140 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2141 Node* n = current->memory_at(index);
2142 MergeMemNode* mm = NULL;
2143 do {
2144 // Follow memory edges through memory accesses, phis and
2145 // narrow membars and push nodes on the stack. Once we hit
2146 // bottom memory, we pop element off the stack one at a
2147 // time, in reverse order, and move them to the right slice
2148 // by changing their memory edges.
2149 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2150 assert(!seen.test_set(n->_idx), "");
2151 // Uses (a load for instance) will need to be moved to the
2152 // right slice as well and will get a new memory state
2153 // that we don't know yet. The use could also be the
2154 // backedge of a loop. We put a place holder node between
2155 // the memory node and its uses. We replace that place
2156 // holder with the correct memory state once we know it,
2157 // i.e. when nodes are popped off the stack. Using the
2158 // place holder make the logic work in the presence of
2159 // loops.
2160 if (n->outcnt() > 1) {
2161 Node* place_holder = NULL;
2162 assert(!n->has_out_with(Op_Node), "");
2163 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2164 Node* u = n->out(k);
2165 if (u != current && u->_idx < last) {
2166 bool success = false;
2167 for (uint l = 0; l < u->req(); l++) {
2168 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2169 continue;
2170 }
2171 Node* in = u->in(l);
2172 if (in == n) {
2173 if (place_holder == NULL) {
2174 place_holder = new Node(1);
2175 place_holder->init_req(0, n);
2176 }
2177 igvn.replace_input_of(u, l, place_holder);
2178 success = true;
2179 }
2180 }
2181 if (success) {
2182 --k;
2183 }
2184 }
2185 }
2186 }
2187 if (n->is_Phi()) {
2188 stack.push(n, 1);
2189 n = n->in(1);
2190 } else if (n->is_Mem()) {
2191 stack.push(n, n->req());
2192 n = n->in(MemNode::Memory);
2193 } else {
2194 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2195 stack.push(n, n->req());
2196 n = n->in(0)->in(TypeFunc::Memory);
2197 }
2198 } else {
2199 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2200 // Build a new MergeMem node to carry the new memory state
2201 // as we build it. IGVN should fold extraneous MergeMem
2202 // nodes.
2203 mm = MergeMemNode::make(n);
2204 igvn.register_new_node_with_optimizer(mm);
2205 while (stack.size() > 0) {
2206 Node* m = stack.node();
2207 uint idx = stack.index();
2208 if (m->is_Mem()) {
2209 // Move memory node to its new slice
2210 const TypePtr* adr_type = m->adr_type();
2211 int alias = get_alias_index(adr_type);
2212 Node* prev = mm->memory_at(alias);
2213 igvn.replace_input_of(m, MemNode::Memory, prev);
2214 mm->set_memory_at(alias, m);
2215 } else if (m->is_Phi()) {
2216 // We need as many new phis as there are new aliases
2217 igvn.replace_input_of(m, idx, mm);
2218 if (idx == m->req()-1) {
2219 Node* r = m->in(0);
2220 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2221 const TypePtr* adr_type = get_adr_type(j);
2222 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2223 continue;
2224 }
2225 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2226 igvn.register_new_node_with_optimizer(phi);
2227 for (uint k = 1; k < m->req(); k++) {
2228 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2229 }
2230 mm->set_memory_at(j, phi);
2231 }
2232 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2233 igvn.register_new_node_with_optimizer(base_phi);
2234 for (uint k = 1; k < m->req(); k++) {
2235 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2236 }
2237 mm->set_base_memory(base_phi);
2238 }
2239 } else {
2240 // This is a MemBarCPUOrder node from
2241 // Parse::array_load()/Parse::array_store(), in the
2242 // branch that handles flattened arrays hidden under
2243 // an Object[] array. We also need one new membar per
2244 // new alias to keep the unknown access that the
2245 // membars protect properly ordered with accesses to
2246 // known flattened array.
2247 assert(m->is_Proj(), "projection expected");
2248 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2249 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2250 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2251 const TypePtr* adr_type = get_adr_type(j);
2252 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2253 continue;
2254 }
2255 MemBarNode* mb = new MemBarCPUOrderNode(this, j, NULL);
2256 igvn.register_new_node_with_optimizer(mb);
2257 Node* mem = mm->memory_at(j);
2258 mb->init_req(TypeFunc::Control, ctrl);
2259 mb->init_req(TypeFunc::Memory, mem);
2260 ctrl = new ProjNode(mb, TypeFunc::Control);
2261 igvn.register_new_node_with_optimizer(ctrl);
2262 mem = new ProjNode(mb, TypeFunc::Memory);
2263 igvn.register_new_node_with_optimizer(mem);
2264 mm->set_memory_at(j, mem);
2265 }
2266 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2267 }
2268 if (idx < m->req()-1) {
2269 idx += 1;
2270 stack.set_index(idx);
2271 n = m->in(idx);
2272 break;
2273 }
2274 // Take care of place holder nodes
2275 if (m->has_out_with(Op_Node)) {
2276 Node* place_holder = m->find_out_with(Op_Node);
2277 if (place_holder != NULL) {
2278 Node* mm_clone = mm->clone();
2279 igvn.register_new_node_with_optimizer(mm_clone);
2280 Node* hook = new Node(1);
2281 hook->init_req(0, mm);
2282 igvn.replace_node(place_holder, mm_clone);
2283 hook->destruct(&igvn);
2284 }
2285 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2286 }
2287 stack.pop();
2288 }
2289 }
2290 } while(stack.size() > 0);
2291 // Fix the memory state at the MergeMem we started from
2292 igvn.rehash_node_delayed(current);
2293 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2294 const TypePtr* adr_type = get_adr_type(j);
2295 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2296 continue;
2297 }
2298 current->set_memory_at(j, mm);
2299 }
2300 current->set_memory_at(index, current->base_memory());
2301 }
2302 igvn.optimize();
2303 }
2304 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2305 #ifdef ASSERT
2306 if (!_flattened_accesses_share_alias) {
2307 wq.clear();
2308 wq.push(root());
2309 for (uint i = 0; i < wq.size(); i++) {
2310 Node* n = wq.at(i);
2311 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2312 for (uint j = 0; j < n->req(); j++) {
2313 Node* m = n->in(j);
2314 if (m != NULL) {
2315 wq.push(m);
2316 }
2317 }
2318 }
2319 }
2320 #endif
2321 }
2322
2323 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2324 if (OptimizeUnstableIf) {
2325 _unstable_if_traps.append(trap);
2326 }
2327 }
2328
2329 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2330 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2331 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2332 Node* n = trap->uncommon_trap();
2333 if (!useful.member(n)) {
2334 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2335 }
2336 }
2337 }
2338
2339 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2340 // or fold-compares case. Return true if succeed or not found.
2341 //
2342 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2343 // false and ask fold-compares to yield.
2344 //
2345 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2346 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2347 // when deoptimization does happen.
2348 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2349 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2350 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2351 if (trap->uncommon_trap() == unc) {
2352 if (yield && trap->modified()) {
2353 return false;
2354 }
2355 _unstable_if_traps.delete_at(i);
2356 break;
2357 }
2358 }
2359 return true;
2360 }
2361
2362 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2363 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2364 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2365 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2366 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2367 CallStaticJavaNode* unc = trap->uncommon_trap();
2368 int next_bci = trap->next_bci();
2369 bool modified = trap->modified();
2370
2371 if (next_bci != -1 && !modified) {
2372 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2373 JVMState* jvms = unc->jvms();
2374 ciMethod* method = jvms->method();
2375 ciBytecodeStream iter(method);
2376
2377 iter.force_bci(jvms->bci());
2378 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2379 Bytecodes::Code c = iter.cur_bc();
2380 Node* lhs = nullptr;
2381 Node* rhs = nullptr;
2382 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2383 lhs = unc->peek_operand(0);
2384 rhs = unc->peek_operand(1);
2385 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2386 lhs = unc->peek_operand(0);
2387 }
2388
2389 ResourceMark rm;
2390 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2391 assert(live_locals.is_valid(), "broken liveness info");
2392 int len = (int)live_locals.size();
2393
2394 for (int i = 0; i < len; i++) {
2395 Node* local = unc->local(jvms, i);
2396 // kill local using the liveness of next_bci.
2397 // give up when the local looks like an operand to secure reexecution.
2398 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
2399 uint idx = jvms->locoff() + i;
2400 #ifdef ASSERT
2401 if (Verbose) {
2402 tty->print("[unstable_if] kill local#%d: ", idx);
2403 local->dump();
2404 tty->cr();
2405 }
2406 #endif
2407 igvn.replace_input_of(unc, idx, top());
2408 modified = true;
2409 }
2410 }
2411 }
2412
2413 // keep the mondified trap for late query
2414 if (modified) {
2415 trap->set_modified();
2416 } else {
2417 _unstable_if_traps.delete_at(i);
2418 }
2419 }
2420 igvn.optimize();
2421 }
2422
2423 // StringOpts and late inlining of string methods
2424 void Compile::inline_string_calls(bool parse_time) {
2425 {
2426 // remove useless nodes to make the usage analysis simpler
2427 ResourceMark rm;
2428 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2429 }
2430
2431 {
2432 ResourceMark rm;
2433 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2434 PhaseStringOpts pso(initial_gvn());
2435 print_method(PHASE_AFTER_STRINGOPTS, 3);
2436 }
2437
2438 // now inline anything that we skipped the first time around
2439 if (!parse_time) {
2440 _late_inlines_pos = _late_inlines.length();
2441 }
2442
2443 while (_string_late_inlines.length() > 0) {
2444 CallGenerator* cg = _string_late_inlines.pop();
2445 cg->do_late_inline();
2446 if (failing()) return;
2447 }
2448 _string_late_inlines.trunc_to(0);
2449 }
2450
2451 // Late inlining of boxing methods
2452 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2453 if (_boxing_late_inlines.length() > 0) {
2454 assert(has_boxed_value(), "inconsistent");
2455
2456 PhaseGVN* gvn = initial_gvn();
2457 set_inlining_incrementally(true);
2458
2459 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2460 for_igvn()->clear();
2461 gvn->replace_with(&igvn);
2462
2463 _late_inlines_pos = _late_inlines.length();
2464
2465 while (_boxing_late_inlines.length() > 0) {
2466 CallGenerator* cg = _boxing_late_inlines.pop();
2467 cg->do_late_inline();
2468 if (failing()) return;
2469 }
2470 _boxing_late_inlines.trunc_to(0);
2471
2472 inline_incrementally_cleanup(igvn);
2473
2474 set_inlining_incrementally(false);
2475 }
2476 }
2477
2478 bool Compile::inline_incrementally_one() {
2479 assert(IncrementalInline, "incremental inlining should be on");
2480
2481 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2482
2483 set_inlining_progress(false);
2484 set_do_cleanup(false);
2485
2486 for (int i = 0; i < _late_inlines.length(); i++) {
2487 _late_inlines_pos = i+1;
2488 CallGenerator* cg = _late_inlines.at(i);
2489 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2490 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2491 cg->do_late_inline();
2492 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2493 if (failing()) {
2494 return false;
2495 } else if (inlining_progress()) {
2496 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2497 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2498 break; // process one call site at a time
2499 }
2500 } else {
2501 // Ignore late inline direct calls when inlining is not allowed.
2502 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2503 }
2504 }
2505 // Remove processed elements.
2506 _late_inlines.remove_till(_late_inlines_pos);
2507 _late_inlines_pos = 0;
2508
2509 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2510
2511 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2512
2513 set_inlining_progress(false);
2514 set_do_cleanup(false);
2515
2516 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2517 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2518 }
2519
2520 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2521 {
2522 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2523 ResourceMark rm;
2524 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2525 }
2526 {
2527 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2528 igvn = PhaseIterGVN(initial_gvn());
2529 igvn.optimize();
2530 }
2531 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2532 }
2533
2534 // Perform incremental inlining until bound on number of live nodes is reached
2535 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2536 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2537
2538 set_inlining_incrementally(true);
2539 uint low_live_nodes = 0;
2540
2541 while (_late_inlines.length() > 0) {
2542 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2543 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2544 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2545 // PhaseIdealLoop is expensive so we only try it once we are
2546 // out of live nodes and we only try it again if the previous
2547 // helped got the number of nodes down significantly
2548 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2549 if (failing()) return;
2550 low_live_nodes = live_nodes();
2551 _major_progress = true;
2552 }
2553
2554 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2555 bool do_print_inlining = print_inlining() || print_intrinsics();
2556 if (do_print_inlining || log() != NULL) {
2557 // Print inlining message for candidates that we couldn't inline for lack of space.
2558 for (int i = 0; i < _late_inlines.length(); i++) {
2559 CallGenerator* cg = _late_inlines.at(i);
2560 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2561 if (do_print_inlining) {
2562 cg->print_inlining_late(msg);
2563 }
2564 log_late_inline_failure(cg, msg);
2565 }
2566 }
2567 break; // finish
2568 }
2569 }
2570
2571 for_igvn()->clear();
2572 initial_gvn()->replace_with(&igvn);
2573
2574 while (inline_incrementally_one()) {
2575 assert(!failing(), "inconsistent");
2576 }
2577 if (failing()) return;
2578
2579 inline_incrementally_cleanup(igvn);
2580
2581 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2582
2583 if (failing()) return;
2584
2585 if (_late_inlines.length() == 0) {
2586 break; // no more progress
2587 }
2588 }
2589 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2590
2591 if (_string_late_inlines.length() > 0) {
2592 assert(has_stringbuilder(), "inconsistent");
2593 for_igvn()->clear();
2594 initial_gvn()->replace_with(&igvn);
2595
2596 inline_string_calls(false);
2597
2598 if (failing()) return;
2599
2600 inline_incrementally_cleanup(igvn);
2601 }
2602
2603 set_inlining_incrementally(false);
2604 }
2605
2606 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2607 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2608 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2609 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == NULL"
2610 // as if "inlining_incrementally() == true" were set.
2611 assert(inlining_incrementally() == false, "not allowed");
2612 #ifdef ASSERT
2613 Unique_Node_List* modified_nodes = _modified_nodes;
2614 _modified_nodes = NULL;
2615 #endif
2616 assert(_late_inlines.length() > 0, "sanity");
2617
2618 while (_late_inlines.length() > 0) {
2619 for_igvn()->clear();
2620 initial_gvn()->replace_with(&igvn);
2621
2622 while (inline_incrementally_one()) {
2623 assert(!failing(), "inconsistent");
2624 }
2625 if (failing()) return;
2626
2627 inline_incrementally_cleanup(igvn);
2628 }
2629 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2630 }
2631
2632 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2633 if (_loop_opts_cnt > 0) {
2634 while (major_progress() && (_loop_opts_cnt > 0)) {
2635 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2636 PhaseIdealLoop::optimize(igvn, mode);
2637 _loop_opts_cnt--;
2638 if (failing()) return false;
2639 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2640 }
2641 }
2642 return true;
2643 }
2644
2645 // Remove edges from "root" to each SafePoint at a backward branch.
2646 // They were inserted during parsing (see add_safepoint()) to make
2647 // infinite loops without calls or exceptions visible to root, i.e.,
2648 // useful.
2649 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2650 Node *r = root();
2651 if (r != NULL) {
2652 for (uint i = r->req(); i < r->len(); ++i) {
2653 Node *n = r->in(i);
2654 if (n != NULL && n->is_SafePoint()) {
2655 r->rm_prec(i);
2656 if (n->outcnt() == 0) {
2657 igvn.remove_dead_node(n);
2658 }
2659 --i;
2660 }
2661 }
2662 // Parsing may have added top inputs to the root node (Path
2663 // leading to the Halt node proven dead). Make sure we get a
2664 // chance to clean them up.
2665 igvn._worklist.push(r);
2666 igvn.optimize();
2667 }
2668 }
2669
2670 //------------------------------Optimize---------------------------------------
2671 // Given a graph, optimize it.
2672 void Compile::Optimize() {
2673 TracePhase tp("optimizer", &timers[_t_optimizer]);
2674
2675 #ifndef PRODUCT
2676 if (env()->break_at_compile()) {
2677 BREAKPOINT;
2678 }
2679
2680 #endif
2681
2682 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2683 #ifdef ASSERT
2684 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2685 #endif
2686
2687 ResourceMark rm;
2688
2689 print_inlining_reinit();
2690
2691 NOT_PRODUCT( verify_graph_edges(); )
2692
2693 print_method(PHASE_AFTER_PARSING, 1);
2694
2695 {
2696 // Iterative Global Value Numbering, including ideal transforms
2697 // Initialize IterGVN with types and values from parse-time GVN
2698 PhaseIterGVN igvn(initial_gvn());
2699 #ifdef ASSERT
2700 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2701 #endif
2702 {
2703 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2704 igvn.optimize();
2705 }
2706
2707 if (failing()) return;
2708
2709 print_method(PHASE_ITER_GVN1, 2);
2710
2711 process_for_unstable_if_traps(igvn);
2712
2713 inline_incrementally(igvn);
2714
2715 print_method(PHASE_INCREMENTAL_INLINE, 2);
2716
2717 if (failing()) return;
2718
2719 if (eliminate_boxing()) {
2720 // Inline valueOf() methods now.
2721 inline_boxing_calls(igvn);
2722
2723 if (AlwaysIncrementalInline) {
2724 inline_incrementally(igvn);
2725 }
2726
2727 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2728
2729 if (failing()) return;
2730 }
2731
2732 // Remove the speculative part of types and clean up the graph from
2733 // the extra CastPP nodes whose only purpose is to carry them. Do
2734 // that early so that optimizations are not disrupted by the extra
2735 // CastPP nodes.
2736 remove_speculative_types(igvn);
2737
2738 // No more new expensive nodes will be added to the list from here
2739 // so keep only the actual candidates for optimizations.
2740 cleanup_expensive_nodes(igvn);
2741
2742 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2743 if (EnableVectorSupport && has_vbox_nodes()) {
2744 TracePhase tp("", &timers[_t_vector]);
2745 PhaseVector pv(igvn);
2746 pv.optimize_vector_boxes();
2747
2748 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2749 }
2750 assert(!has_vbox_nodes(), "sanity");
2751
2752 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2753 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2754 initial_gvn()->replace_with(&igvn);
2755 Unique_Node_List* old_worklist = for_igvn();
2756 old_worklist->clear();
2757 Unique_Node_List new_worklist(C->comp_arena());
2758 {
2759 ResourceMark rm;
2760 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2761 }
2762 Unique_Node_List* save_for_igvn = for_igvn();
2763 set_for_igvn(&new_worklist);
2764 igvn = PhaseIterGVN(initial_gvn());
2765 igvn.optimize();
2766 set_for_igvn(old_worklist); // new_worklist is dead beyond this point
2767 }
2768
2769 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2770 // safepoints
2771 remove_root_to_sfpts_edges(igvn);
2772
2773 // Process inline type nodes now that all inlining is over
2774 process_inline_types(igvn);
2775
2776 adjust_flattened_array_access_aliases(igvn);
2777
2778 // Perform escape analysis
2779 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2780 if (has_loops()) {
2781 // Cleanup graph (remove dead nodes).
2782 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2783 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2784 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2785 if (failing()) return;
2786 }
2787 bool progress;
2788 do {
2789 ConnectionGraph::do_analysis(this, &igvn);
2790
2791 if (failing()) return;
2792
2793 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2794
2795 // Optimize out fields loads from scalar replaceable allocations.
2796 igvn.optimize();
2797 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2798
2799 if (failing()) return;
2800
2801 if (congraph() != NULL && macro_count() > 0) {
2802 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2803 PhaseMacroExpand mexp(igvn);
2804 mexp.eliminate_macro_nodes();
2805 igvn.set_delay_transform(false);
2806
2807 igvn.optimize();
2808 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2809
2810 if (failing()) return;
2811 }
2812 progress = do_iterative_escape_analysis() &&
2813 (macro_count() < mcount) &&
2814 ConnectionGraph::has_candidates(this);
2815 // Try again if candidates exist and made progress
2816 // by removing some allocations and/or locks.
2817 } while (progress);
2818 }
2819
2820 // Loop transforms on the ideal graph. Range Check Elimination,
2821 // peeling, unrolling, etc.
2822
2823 // Set loop opts counter
2824 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2825 {
2826 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2827 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2828 _loop_opts_cnt--;
2829 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2830 if (failing()) return;
2831 }
2832 // Loop opts pass if partial peeling occurred in previous pass
2833 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2834 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2835 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2836 _loop_opts_cnt--;
2837 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2838 if (failing()) return;
2839 }
2840 // Loop opts pass for loop-unrolling before CCP
2841 if(major_progress() && (_loop_opts_cnt > 0)) {
2842 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2843 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2844 _loop_opts_cnt--;
2845 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2846 }
2847 if (!failing()) {
2848 // Verify that last round of loop opts produced a valid graph
2849 PhaseIdealLoop::verify(igvn);
2850 }
2851 }
2852 if (failing()) return;
2853
2854 // Conditional Constant Propagation;
2855 PhaseCCP ccp( &igvn );
2856 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2857 {
2858 TracePhase tp("ccp", &timers[_t_ccp]);
2859 ccp.do_transform();
2860 }
2861 print_method(PHASE_CCP1, 2);
2862
2863 assert( true, "Break here to ccp.dump_old2new_map()");
2864
2865 // Iterative Global Value Numbering, including ideal transforms
2866 {
2867 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2868 igvn = ccp;
2869 igvn.optimize();
2870 }
2871 print_method(PHASE_ITER_GVN2, 2);
2872
2873 if (failing()) return;
2874
2875 // Loop transforms on the ideal graph. Range Check Elimination,
2876 // peeling, unrolling, etc.
2877 if (!optimize_loops(igvn, LoopOptsDefault)) {
2878 return;
2879 }
2880
2881 if (failing()) return;
2882
2883 C->clear_major_progress(); // ensure that major progress is now clear
2884
2885 process_for_post_loop_opts_igvn(igvn);
2886
2887 #ifdef ASSERT
2888 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2889 #endif
2890
2891 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2892
2893 if (_late_inlines.length() > 0) {
2894 // More opportunities to optimize virtual and MH calls.
2895 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2896 process_late_inline_calls_no_inline(igvn);
2897 }
2898
2899 {
2900 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2901 PhaseMacroExpand mex(igvn);
2902 if (mex.expand_macro_nodes()) {
2903 assert(failing(), "must bail out w/ explicit message");
2904 return;
2905 }
2906 print_method(PHASE_MACRO_EXPANSION, 2);
2907 }
2908
2909 // Process inline type nodes again and remove them. From here
2910 // on we don't need to keep track of field values anymore.
2911 process_inline_types(igvn, /* remove= */ true);
2912
2913 {
2914 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2915 if (bs->expand_barriers(this, igvn)) {
2916 assert(failing(), "must bail out w/ explicit message");
2917 return;
2918 }
2919 print_method(PHASE_BARRIER_EXPANSION, 2);
2920 }
2921
2922 if (C->max_vector_size() > 0) {
2923 C->optimize_logic_cones(igvn);
2924 igvn.optimize();
2925 }
2926
2927 DEBUG_ONLY( _modified_nodes = NULL; )
2928 DEBUG_ONLY( _late_inlines.clear(); )
2929
2930 assert(igvn._worklist.size() == 0, "not empty");
2931 } // (End scope of igvn; run destructor if necessary for asserts.)
2932
2933 check_no_dead_use();
2934
2935 process_print_inlining();
2936
2937 // A method with only infinite loops has no edges entering loops from root
2938 {
2939 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2940 if (final_graph_reshaping()) {
2941 assert(failing(), "must bail out w/ explicit message");
2942 return;
2943 }
2944 }
2945
2946 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2947 DEBUG_ONLY(set_phase_optimize_finished();)
2948 }
2949
2950 #ifdef ASSERT
2951 void Compile::check_no_dead_use() const {
2952 ResourceMark rm;
2953 Unique_Node_List wq;
2954 wq.push(root());
2955 for (uint i = 0; i < wq.size(); ++i) {
2956 Node* n = wq.at(i);
2957 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2958 Node* u = n->fast_out(j);
2959 if (u->outcnt() == 0 && !u->is_Con()) {
2960 u->dump();
2961 fatal("no reachable node should have no use");
2962 }
2963 wq.push(u);
2964 }
2965 }
2966 }
2967 #endif
2968
2969 void Compile::inline_vector_reboxing_calls() {
2970 if (C->_vector_reboxing_late_inlines.length() > 0) {
2971 _late_inlines_pos = C->_late_inlines.length();
2972 while (_vector_reboxing_late_inlines.length() > 0) {
2973 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2974 cg->do_late_inline();
2975 if (failing()) return;
2976 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2977 }
2978 _vector_reboxing_late_inlines.trunc_to(0);
2979 }
2980 }
2981
2982 bool Compile::has_vbox_nodes() {
2983 if (C->_vector_reboxing_late_inlines.length() > 0) {
2984 return true;
2985 }
2986 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2987 Node * n = C->macro_node(macro_idx);
2988 assert(n->is_macro(), "only macro nodes expected here");
2989 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2990 return true;
2991 }
2992 }
2993 return false;
2994 }
2995
2996 //---------------------------- Bitwise operation packing optimization ---------------------------
2997
2998 static bool is_vector_unary_bitwise_op(Node* n) {
2999 return n->Opcode() == Op_XorV &&
3000 VectorNode::is_vector_bitwise_not_pattern(n);
3001 }
3002
3003 static bool is_vector_binary_bitwise_op(Node* n) {
3004 switch (n->Opcode()) {
3005 case Op_AndV:
3006 case Op_OrV:
3007 return true;
3008
3009 case Op_XorV:
3010 return !is_vector_unary_bitwise_op(n);
3011
3012 default:
3013 return false;
3014 }
3015 }
3016
3017 static bool is_vector_ternary_bitwise_op(Node* n) {
3018 return n->Opcode() == Op_MacroLogicV;
3019 }
3020
3021 static bool is_vector_bitwise_op(Node* n) {
3022 return is_vector_unary_bitwise_op(n) ||
3023 is_vector_binary_bitwise_op(n) ||
3024 is_vector_ternary_bitwise_op(n);
3025 }
3026
3027 static bool is_vector_bitwise_cone_root(Node* n) {
3028 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
3029 return false;
3030 }
3031 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3032 if (is_vector_bitwise_op(n->fast_out(i))) {
3033 return false;
3034 }
3035 }
3036 return true;
3037 }
3038
3039 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3040 uint cnt = 0;
3041 if (is_vector_bitwise_op(n)) {
3042 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3043 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3044 for (uint i = 1; i < inp_cnt; i++) {
3045 Node* in = n->in(i);
3046 bool skip = VectorNode::is_all_ones_vector(in);
3047 if (!skip && !inputs.member(in)) {
3048 inputs.push(in);
3049 cnt++;
3050 }
3051 }
3052 assert(cnt <= 1, "not unary");
3053 } else {
3054 uint last_req = inp_cnt;
3055 if (is_vector_ternary_bitwise_op(n)) {
3056 last_req = inp_cnt - 1; // skip last input
3057 }
3058 for (uint i = 1; i < last_req; i++) {
3059 Node* def = n->in(i);
3060 if (!inputs.member(def)) {
3061 inputs.push(def);
3062 cnt++;
3063 }
3064 }
3065 }
3066 } else { // not a bitwise operations
3067 if (!inputs.member(n)) {
3068 inputs.push(n);
3069 cnt++;
3070 }
3071 }
3072 return cnt;
3073 }
3074
3075 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3076 Unique_Node_List useful_nodes;
3077 C->identify_useful_nodes(useful_nodes);
3078
3079 for (uint i = 0; i < useful_nodes.size(); i++) {
3080 Node* n = useful_nodes.at(i);
3081 if (is_vector_bitwise_cone_root(n)) {
3082 list.push(n);
3083 }
3084 }
3085 }
3086
3087 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3088 const TypeVect* vt,
3089 Unique_Node_List& partition,
3090 Unique_Node_List& inputs) {
3091 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3092 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3093 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3094
3095 Node* in1 = inputs.at(0);
3096 Node* in2 = inputs.at(1);
3097 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3098
3099 uint func = compute_truth_table(partition, inputs);
3100
3101 Node* pn = partition.at(partition.size() - 1);
3102 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : NULL;
3103 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3104 }
3105
3106 static uint extract_bit(uint func, uint pos) {
3107 return (func & (1 << pos)) >> pos;
3108 }
3109
3110 //
3111 // A macro logic node represents a truth table. It has 4 inputs,
3112 // First three inputs corresponds to 3 columns of a truth table
3113 // and fourth input captures the logic function.
3114 //
3115 // eg. fn = (in1 AND in2) OR in3;
3116 //
3117 // MacroNode(in1,in2,in3,fn)
3118 //
3119 // -----------------
3120 // in1 in2 in3 fn
3121 // -----------------
3122 // 0 0 0 0
3123 // 0 0 1 1
3124 // 0 1 0 0
3125 // 0 1 1 1
3126 // 1 0 0 0
3127 // 1 0 1 1
3128 // 1 1 0 1
3129 // 1 1 1 1
3130 //
3131
3132 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3133 int res = 0;
3134 for (int i = 0; i < 8; i++) {
3135 int bit1 = extract_bit(in1, i);
3136 int bit2 = extract_bit(in2, i);
3137 int bit3 = extract_bit(in3, i);
3138
3139 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3140 int func_bit = extract_bit(func, func_bit_pos);
3141
3142 res |= func_bit << i;
3143 }
3144 return res;
3145 }
3146
3147 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
3148 assert(n != NULL, "");
3149 assert(eval_map.contains(n), "absent");
3150 return *(eval_map.get(n));
3151 }
3152
3153 static void eval_operands(Node* n,
3154 uint& func1, uint& func2, uint& func3,
3155 ResourceHashtable<Node*,uint>& eval_map) {
3156 assert(is_vector_bitwise_op(n), "");
3157
3158 if (is_vector_unary_bitwise_op(n)) {
3159 Node* opnd = n->in(1);
3160 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3161 opnd = n->in(2);
3162 }
3163 func1 = eval_operand(opnd, eval_map);
3164 } else if (is_vector_binary_bitwise_op(n)) {
3165 func1 = eval_operand(n->in(1), eval_map);
3166 func2 = eval_operand(n->in(2), eval_map);
3167 } else {
3168 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3169 func1 = eval_operand(n->in(1), eval_map);
3170 func2 = eval_operand(n->in(2), eval_map);
3171 func3 = eval_operand(n->in(3), eval_map);
3172 }
3173 }
3174
3175 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3176 assert(inputs.size() <= 3, "sanity");
3177 ResourceMark rm;
3178 uint res = 0;
3179 ResourceHashtable<Node*,uint> eval_map;
3180
3181 // Populate precomputed functions for inputs.
3182 // Each input corresponds to one column of 3 input truth-table.
3183 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3184 0xCC, // (_, b, _) -> b
3185 0xF0 }; // (a, _, _) -> a
3186 for (uint i = 0; i < inputs.size(); i++) {
3187 eval_map.put(inputs.at(i), input_funcs[2-i]);
3188 }
3189
3190 for (uint i = 0; i < partition.size(); i++) {
3191 Node* n = partition.at(i);
3192
3193 uint func1 = 0, func2 = 0, func3 = 0;
3194 eval_operands(n, func1, func2, func3, eval_map);
3195
3196 switch (n->Opcode()) {
3197 case Op_OrV:
3198 assert(func3 == 0, "not binary");
3199 res = func1 | func2;
3200 break;
3201 case Op_AndV:
3202 assert(func3 == 0, "not binary");
3203 res = func1 & func2;
3204 break;
3205 case Op_XorV:
3206 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3207 assert(func2 == 0 && func3 == 0, "not unary");
3208 res = (~func1) & 0xFF;
3209 } else {
3210 assert(func3 == 0, "not binary");
3211 res = func1 ^ func2;
3212 }
3213 break;
3214 case Op_MacroLogicV:
3215 // Ordering of inputs may change during evaluation of sub-tree
3216 // containing MacroLogic node as a child node, thus a re-evaluation
3217 // makes sure that function is evaluated in context of current
3218 // inputs.
3219 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3220 break;
3221
3222 default: assert(false, "not supported: %s", n->Name());
3223 }
3224 assert(res <= 0xFF, "invalid");
3225 eval_map.put(n, res);
3226 }
3227 return res;
3228 }
3229
3230 // Criteria under which nodes gets packed into a macro logic node:-
3231 // 1) Parent and both child nodes are all unmasked or masked with
3232 // same predicates.
3233 // 2) Masked parent can be packed with left child if it is predicated
3234 // and both have same predicates.
3235 // 3) Masked parent can be packed with right child if its un-predicated
3236 // or has matching predication condition.
3237 // 4) An unmasked parent can be packed with an unmasked child.
3238 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3239 assert(partition.size() == 0, "not empty");
3240 assert(inputs.size() == 0, "not empty");
3241 if (is_vector_ternary_bitwise_op(n)) {
3242 return false;
3243 }
3244
3245 bool is_unary_op = is_vector_unary_bitwise_op(n);
3246 if (is_unary_op) {
3247 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3248 return false; // too few inputs
3249 }
3250
3251 bool pack_left_child = true;
3252 bool pack_right_child = true;
3253
3254 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3255 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3256
3257 int left_child_input_cnt = 0;
3258 int right_child_input_cnt = 0;
3259
3260 bool parent_is_predicated = n->is_predicated_vector();
3261 bool left_child_predicated = n->in(1)->is_predicated_vector();
3262 bool right_child_predicated = n->in(2)->is_predicated_vector();
3263
3264 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : NULL;
3265 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : NULL;
3266 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : NULL;
3267
3268 do {
3269 if (pack_left_child && left_child_LOP &&
3270 ((!parent_is_predicated && !left_child_predicated) ||
3271 ((parent_is_predicated && left_child_predicated &&
3272 parent_pred == left_child_pred)))) {
3273 partition.push(n->in(1));
3274 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3275 } else {
3276 inputs.push(n->in(1));
3277 left_child_input_cnt = 1;
3278 }
3279
3280 if (pack_right_child && right_child_LOP &&
3281 (!right_child_predicated ||
3282 (right_child_predicated && parent_is_predicated &&
3283 parent_pred == right_child_pred))) {
3284 partition.push(n->in(2));
3285 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3286 } else {
3287 inputs.push(n->in(2));
3288 right_child_input_cnt = 1;
3289 }
3290
3291 if (inputs.size() > 3) {
3292 assert(partition.size() > 0, "");
3293 inputs.clear();
3294 partition.clear();
3295 if (left_child_input_cnt > right_child_input_cnt) {
3296 pack_left_child = false;
3297 } else {
3298 pack_right_child = false;
3299 }
3300 } else {
3301 break;
3302 }
3303 } while(true);
3304
3305 if(partition.size()) {
3306 partition.push(n);
3307 }
3308
3309 return (partition.size() == 2 || partition.size() == 3) &&
3310 (inputs.size() == 2 || inputs.size() == 3);
3311 }
3312
3313 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3314 assert(is_vector_bitwise_op(n), "not a root");
3315
3316 visited.set(n->_idx);
3317
3318 // 1) Do a DFS walk over the logic cone.
3319 for (uint i = 1; i < n->req(); i++) {
3320 Node* in = n->in(i);
3321 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3322 process_logic_cone_root(igvn, in, visited);
3323 }
3324 }
3325
3326 // 2) Bottom up traversal: Merge node[s] with
3327 // the parent to form macro logic node.
3328 Unique_Node_List partition;
3329 Unique_Node_List inputs;
3330 if (compute_logic_cone(n, partition, inputs)) {
3331 const TypeVect* vt = n->bottom_type()->is_vect();
3332 Node* pn = partition.at(partition.size() - 1);
3333 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : NULL;
3334 if (mask == NULL ||
3335 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3336 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3337 #ifdef ASSERT
3338 if (TraceNewVectors) {
3339 tty->print("new Vector node: ");
3340 macro_logic->dump();
3341 }
3342 #endif
3343 igvn.replace_node(n, macro_logic);
3344 }
3345 }
3346 }
3347
3348 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3349 ResourceMark rm;
3350 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3351 Unique_Node_List list;
3352 collect_logic_cone_roots(list);
3353
3354 while (list.size() > 0) {
3355 Node* n = list.pop();
3356 const TypeVect* vt = n->bottom_type()->is_vect();
3357 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3358 if (supported) {
3359 VectorSet visited(comp_arena());
3360 process_logic_cone_root(igvn, n, visited);
3361 }
3362 }
3363 }
3364 }
3365
3366 //------------------------------Code_Gen---------------------------------------
3367 // Given a graph, generate code for it
3368 void Compile::Code_Gen() {
3369 if (failing()) {
3370 return;
3371 }
3372
3373 // Perform instruction selection. You might think we could reclaim Matcher
3374 // memory PDQ, but actually the Matcher is used in generating spill code.
3375 // Internals of the Matcher (including some VectorSets) must remain live
3376 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3377 // set a bit in reclaimed memory.
3378
3379 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3380 // nodes. Mapping is only valid at the root of each matched subtree.
3381 NOT_PRODUCT( verify_graph_edges(); )
3382
3383 Matcher matcher;
3384 _matcher = &matcher;
3385 {
3386 TracePhase tp("matcher", &timers[_t_matcher]);
3387 matcher.match();
3388 if (failing()) {
3389 return;
3390 }
3391 }
3392 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3393 // nodes. Mapping is only valid at the root of each matched subtree.
3394 NOT_PRODUCT( verify_graph_edges(); )
3395
3396 // If you have too many nodes, or if matching has failed, bail out
3397 check_node_count(0, "out of nodes matching instructions");
3398 if (failing()) {
3399 return;
3400 }
3401
3402 print_method(PHASE_MATCHING, 2);
3403
3404 // Build a proper-looking CFG
3405 PhaseCFG cfg(node_arena(), root(), matcher);
3406 _cfg = &cfg;
3407 {
3408 TracePhase tp("scheduler", &timers[_t_scheduler]);
3409 bool success = cfg.do_global_code_motion();
3410 if (!success) {
3411 return;
3412 }
3413
3414 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3415 NOT_PRODUCT( verify_graph_edges(); )
3416 cfg.verify();
3417 }
3418
3419 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3420 _regalloc = ®alloc;
3421 {
3422 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
3423 // Perform register allocation. After Chaitin, use-def chains are
3424 // no longer accurate (at spill code) and so must be ignored.
3425 // Node->LRG->reg mappings are still accurate.
3426 _regalloc->Register_Allocate();
3427
3428 // Bail out if the allocator builds too many nodes
3429 if (failing()) {
3430 return;
3431 }
3432 }
3433
3434 // Prior to register allocation we kept empty basic blocks in case the
3435 // the allocator needed a place to spill. After register allocation we
3436 // are not adding any new instructions. If any basic block is empty, we
3437 // can now safely remove it.
3438 {
3439 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
3440 cfg.remove_empty_blocks();
3441 if (do_freq_based_layout()) {
3442 PhaseBlockLayout layout(cfg);
3443 } else {
3444 cfg.set_loop_alignment();
3445 }
3446 cfg.fixup_flow();
3447 cfg.remove_unreachable_blocks();
3448 cfg.verify_dominator_tree();
3449 }
3450
3451 // Apply peephole optimizations
3452 if( OptoPeephole ) {
3453 TracePhase tp("peephole", &timers[_t_peephole]);
3454 PhasePeephole peep( _regalloc, cfg);
3455 peep.do_transform();
3456 }
3457
3458 // Do late expand if CPU requires this.
3459 if (Matcher::require_postalloc_expand) {
3460 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3461 cfg.postalloc_expand(_regalloc);
3462 }
3463
3464 // Convert Nodes to instruction bits in a buffer
3465 {
3466 TracePhase tp("output", &timers[_t_output]);
3467 PhaseOutput output;
3468 output.Output();
3469 if (failing()) return;
3470 output.install();
3471 }
3472
3473 print_method(PHASE_FINAL_CODE, 1);
3474
3475 // He's dead, Jim.
3476 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3477 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3478 }
3479
3480 //------------------------------Final_Reshape_Counts---------------------------
3481 // This class defines counters to help identify when a method
3482 // may/must be executed using hardware with only 24-bit precision.
3483 struct Final_Reshape_Counts : public StackObj {
3484 int _call_count; // count non-inlined 'common' calls
3485 int _float_count; // count float ops requiring 24-bit precision
3486 int _double_count; // count double ops requiring more precision
3487 int _java_call_count; // count non-inlined 'java' calls
3488 int _inner_loop_count; // count loops which need alignment
3489 VectorSet _visited; // Visitation flags
3490 Node_List _tests; // Set of IfNodes & PCTableNodes
3491
3492 Final_Reshape_Counts() :
3493 _call_count(0), _float_count(0), _double_count(0),
3494 _java_call_count(0), _inner_loop_count(0) { }
3495
3496 void inc_call_count () { _call_count ++; }
3497 void inc_float_count () { _float_count ++; }
3498 void inc_double_count() { _double_count++; }
3499 void inc_java_call_count() { _java_call_count++; }
3500 void inc_inner_loop_count() { _inner_loop_count++; }
3501
3502 int get_call_count () const { return _call_count ; }
3503 int get_float_count () const { return _float_count ; }
3504 int get_double_count() const { return _double_count; }
3505 int get_java_call_count() const { return _java_call_count; }
3506 int get_inner_loop_count() const { return _inner_loop_count; }
3507 };
3508
3509 // Eliminate trivially redundant StoreCMs and accumulate their
3510 // precedence edges.
3511 void Compile::eliminate_redundant_card_marks(Node* n) {
3512 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3513 if (n->in(MemNode::Address)->outcnt() > 1) {
3514 // There are multiple users of the same address so it might be
3515 // possible to eliminate some of the StoreCMs
3516 Node* mem = n->in(MemNode::Memory);
3517 Node* adr = n->in(MemNode::Address);
3518 Node* val = n->in(MemNode::ValueIn);
3519 Node* prev = n;
3520 bool done = false;
3521 // Walk the chain of StoreCMs eliminating ones that match. As
3522 // long as it's a chain of single users then the optimization is
3523 // safe. Eliminating partially redundant StoreCMs would require
3524 // cloning copies down the other paths.
3525 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3526 if (adr == mem->in(MemNode::Address) &&
3527 val == mem->in(MemNode::ValueIn)) {
3528 // redundant StoreCM
3529 if (mem->req() > MemNode::OopStore) {
3530 // Hasn't been processed by this code yet.
3531 n->add_prec(mem->in(MemNode::OopStore));
3532 } else {
3533 // Already converted to precedence edge
3534 for (uint i = mem->req(); i < mem->len(); i++) {
3535 // Accumulate any precedence edges
3536 if (mem->in(i) != NULL) {
3537 n->add_prec(mem->in(i));
3538 }
3539 }
3540 // Everything above this point has been processed.
3541 done = true;
3542 }
3543 // Eliminate the previous StoreCM
3544 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3545 assert(mem->outcnt() == 0, "should be dead");
3546 mem->disconnect_inputs(this);
3547 } else {
3548 prev = mem;
3549 }
3550 mem = prev->in(MemNode::Memory);
3551 }
3552 }
3553 }
3554
3555
3556 //------------------------------final_graph_reshaping_impl----------------------
3557 // Implement items 1-5 from final_graph_reshaping below.
3558 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3559
3560 if ( n->outcnt() == 0 ) return; // dead node
3561 uint nop = n->Opcode();
3562
3563 // Check for 2-input instruction with "last use" on right input.
3564 // Swap to left input. Implements item (2).
3565 if( n->req() == 3 && // two-input instruction
3566 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3567 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3568 n->in(2)->outcnt() == 1 &&// right use IS a last use
3569 !n->in(2)->is_Con() ) { // right use is not a constant
3570 // Check for commutative opcode
3571 switch( nop ) {
3572 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3573 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3574 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3575 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3576 case Op_AndL: case Op_XorL: case Op_OrL:
3577 case Op_AndI: case Op_XorI: case Op_OrI: {
3578 // Move "last use" input to left by swapping inputs
3579 n->swap_edges(1, 2);
3580 break;
3581 }
3582 default:
3583 break;
3584 }
3585 }
3586
3587 #ifdef ASSERT
3588 if( n->is_Mem() ) {
3589 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3590 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
3591 // oop will be recorded in oop map if load crosses safepoint
3592 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3593 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3594 "raw memory operations should have control edge");
3595 }
3596 if (n->is_MemBar()) {
3597 MemBarNode* mb = n->as_MemBar();
3598 if (mb->trailing_store() || mb->trailing_load_store()) {
3599 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3600 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3601 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3602 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3603 } else if (mb->leading()) {
3604 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3605 }
3606 }
3607 #endif
3608 // Count FPU ops and common calls, implements item (3)
3609 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3610 if (!gc_handled) {
3611 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3612 }
3613
3614 // Collect CFG split points
3615 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3616 frc._tests.push(n);
3617 }
3618 }
3619
3620 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3621 switch( nop ) {
3622 // Count all float operations that may use FPU
3623 case Op_AddF:
3624 case Op_SubF:
3625 case Op_MulF:
3626 case Op_DivF:
3627 case Op_NegF:
3628 case Op_ModF:
3629 case Op_ConvI2F:
3630 case Op_ConF:
3631 case Op_CmpF:
3632 case Op_CmpF3:
3633 case Op_StoreF:
3634 case Op_LoadF:
3635 // case Op_ConvL2F: // longs are split into 32-bit halves
3636 frc.inc_float_count();
3637 break;
3638
3639 case Op_ConvF2D:
3640 case Op_ConvD2F:
3641 frc.inc_float_count();
3642 frc.inc_double_count();
3643 break;
3644
3645 // Count all double operations that may use FPU
3646 case Op_AddD:
3647 case Op_SubD:
3648 case Op_MulD:
3649 case Op_DivD:
3650 case Op_NegD:
3651 case Op_ModD:
3652 case Op_ConvI2D:
3653 case Op_ConvD2I:
3654 // case Op_ConvL2D: // handled by leaf call
3655 // case Op_ConvD2L: // handled by leaf call
3656 case Op_ConD:
3657 case Op_CmpD:
3658 case Op_CmpD3:
3659 case Op_StoreD:
3660 case Op_LoadD:
3661 case Op_LoadD_unaligned:
3662 frc.inc_double_count();
3663 break;
3664 case Op_Opaque1: // Remove Opaque Nodes before matching
3665 case Op_Opaque3:
3666 n->subsume_by(n->in(1), this);
3667 break;
3668 case Op_CallStaticJava:
3669 case Op_CallJava:
3670 case Op_CallDynamicJava:
3671 frc.inc_java_call_count(); // Count java call site;
3672 case Op_CallRuntime:
3673 case Op_CallLeaf:
3674 case Op_CallLeafVector:
3675 case Op_CallLeafNoFP: {
3676 assert (n->is_Call(), "");
3677 CallNode *call = n->as_Call();
3678 // Count call sites where the FP mode bit would have to be flipped.
3679 // Do not count uncommon runtime calls:
3680 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3681 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3682 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3683 frc.inc_call_count(); // Count the call site
3684 } else { // See if uncommon argument is shared
3685 Node *n = call->in(TypeFunc::Parms);
3686 int nop = n->Opcode();
3687 // Clone shared simple arguments to uncommon calls, item (1).
3688 if (n->outcnt() > 1 &&
3689 !n->is_Proj() &&
3690 nop != Op_CreateEx &&
3691 nop != Op_CheckCastPP &&
3692 nop != Op_DecodeN &&
3693 nop != Op_DecodeNKlass &&
3694 !n->is_Mem() &&
3695 !n->is_Phi()) {
3696 Node *x = n->clone();
3697 call->set_req(TypeFunc::Parms, x);
3698 }
3699 }
3700 break;
3701 }
3702
3703 case Op_StoreCM:
3704 {
3705 // Convert OopStore dependence into precedence edge
3706 Node* prec = n->in(MemNode::OopStore);
3707 n->del_req(MemNode::OopStore);
3708 if (prec->is_MergeMem()) {
3709 MergeMemNode* mm = prec->as_MergeMem();
3710 Node* base = mm->base_memory();
3711 for (int i = AliasIdxRaw + 1; i < num_alias_types(); i++) {
3712 const TypePtr* adr_type = get_adr_type(i);
3713 if (adr_type->is_flat()) {
3714 Node* m = mm->memory_at(i);
3715 n->add_prec(m);
3716 }
3717 }
3718 if (mm->outcnt() == 0) {
3719 mm->disconnect_inputs(this);
3720 }
3721 } else {
3722 n->add_prec(prec);
3723 }
3724 eliminate_redundant_card_marks(n);
3725 }
3726
3727 // fall through
3728
3729 case Op_StoreB:
3730 case Op_StoreC:
3731 case Op_StoreI:
3732 case Op_StoreL:
3733 case Op_CompareAndSwapB:
3734 case Op_CompareAndSwapS:
3735 case Op_CompareAndSwapI:
3736 case Op_CompareAndSwapL:
3737 case Op_CompareAndSwapP:
3738 case Op_CompareAndSwapN:
3739 case Op_WeakCompareAndSwapB:
3740 case Op_WeakCompareAndSwapS:
3741 case Op_WeakCompareAndSwapI:
3742 case Op_WeakCompareAndSwapL:
3743 case Op_WeakCompareAndSwapP:
3744 case Op_WeakCompareAndSwapN:
3745 case Op_CompareAndExchangeB:
3746 case Op_CompareAndExchangeS:
3747 case Op_CompareAndExchangeI:
3748 case Op_CompareAndExchangeL:
3749 case Op_CompareAndExchangeP:
3750 case Op_CompareAndExchangeN:
3751 case Op_GetAndAddS:
3752 case Op_GetAndAddB:
3753 case Op_GetAndAddI:
3754 case Op_GetAndAddL:
3755 case Op_GetAndSetS:
3756 case Op_GetAndSetB:
3757 case Op_GetAndSetI:
3758 case Op_GetAndSetL:
3759 case Op_GetAndSetP:
3760 case Op_GetAndSetN:
3761 case Op_StoreP:
3762 case Op_StoreN:
3763 case Op_StoreNKlass:
3764 case Op_LoadB:
3765 case Op_LoadUB:
3766 case Op_LoadUS:
3767 case Op_LoadI:
3768 case Op_LoadKlass:
3769 case Op_LoadNKlass:
3770 case Op_LoadL:
3771 case Op_LoadL_unaligned:
3772 case Op_LoadP:
3773 case Op_LoadN:
3774 case Op_LoadRange:
3775 case Op_LoadS:
3776 break;
3777
3778 case Op_AddP: { // Assert sane base pointers
3779 Node *addp = n->in(AddPNode::Address);
3780 assert( !addp->is_AddP() ||
3781 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3782 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3783 "Base pointers must match (addp %u)", addp->_idx );
3784 #ifdef _LP64
3785 if ((UseCompressedOops || UseCompressedClassPointers) &&
3786 addp->Opcode() == Op_ConP &&
3787 addp == n->in(AddPNode::Base) &&
3788 n->in(AddPNode::Offset)->is_Con()) {
3789 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3790 // on the platform and on the compressed oops mode.
3791 // Use addressing with narrow klass to load with offset on x86.
3792 // Some platforms can use the constant pool to load ConP.
3793 // Do this transformation here since IGVN will convert ConN back to ConP.
3794 const Type* t = addp->bottom_type();
3795 bool is_oop = t->isa_oopptr() != NULL;
3796 bool is_klass = t->isa_klassptr() != NULL;
3797
3798 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3799 (is_klass && Matcher::const_klass_prefer_decode())) {
3800 Node* nn = NULL;
3801
3802 int op = is_oop ? Op_ConN : Op_ConNKlass;
3803
3804 // Look for existing ConN node of the same exact type.
3805 Node* r = root();
3806 uint cnt = r->outcnt();
3807 for (uint i = 0; i < cnt; i++) {
3808 Node* m = r->raw_out(i);
3809 if (m!= NULL && m->Opcode() == op &&
3810 m->bottom_type()->make_ptr() == t) {
3811 nn = m;
3812 break;
3813 }
3814 }
3815 if (nn != NULL) {
3816 // Decode a narrow oop to match address
3817 // [R12 + narrow_oop_reg<<3 + offset]
3818 if (is_oop) {
3819 nn = new DecodeNNode(nn, t);
3820 } else {
3821 nn = new DecodeNKlassNode(nn, t);
3822 }
3823 // Check for succeeding AddP which uses the same Base.
3824 // Otherwise we will run into the assertion above when visiting that guy.
3825 for (uint i = 0; i < n->outcnt(); ++i) {
3826 Node *out_i = n->raw_out(i);
3827 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3828 out_i->set_req(AddPNode::Base, nn);
3829 #ifdef ASSERT
3830 for (uint j = 0; j < out_i->outcnt(); ++j) {
3831 Node *out_j = out_i->raw_out(j);
3832 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3833 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3834 }
3835 #endif
3836 }
3837 }
3838 n->set_req(AddPNode::Base, nn);
3839 n->set_req(AddPNode::Address, nn);
3840 if (addp->outcnt() == 0) {
3841 addp->disconnect_inputs(this);
3842 }
3843 }
3844 }
3845 }
3846 #endif
3847 break;
3848 }
3849
3850 case Op_CastPP: {
3851 // Remove CastPP nodes to gain more freedom during scheduling but
3852 // keep the dependency they encode as control or precedence edges
3853 // (if control is set already) on memory operations. Some CastPP
3854 // nodes don't have a control (don't carry a dependency): skip
3855 // those.
3856 if (n->in(0) != NULL) {
3857 ResourceMark rm;
3858 Unique_Node_List wq;
3859 wq.push(n);
3860 for (uint next = 0; next < wq.size(); ++next) {
3861 Node *m = wq.at(next);
3862 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3863 Node* use = m->fast_out(i);
3864 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3865 use->ensure_control_or_add_prec(n->in(0));
3866 } else {
3867 switch(use->Opcode()) {
3868 case Op_AddP:
3869 case Op_DecodeN:
3870 case Op_DecodeNKlass:
3871 case Op_CheckCastPP:
3872 case Op_CastPP:
3873 wq.push(use);
3874 break;
3875 }
3876 }
3877 }
3878 }
3879 }
3880 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3881 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3882 Node* in1 = n->in(1);
3883 const Type* t = n->bottom_type();
3884 Node* new_in1 = in1->clone();
3885 new_in1->as_DecodeN()->set_type(t);
3886
3887 if (!Matcher::narrow_oop_use_complex_address()) {
3888 //
3889 // x86, ARM and friends can handle 2 adds in addressing mode
3890 // and Matcher can fold a DecodeN node into address by using
3891 // a narrow oop directly and do implicit NULL check in address:
3892 //
3893 // [R12 + narrow_oop_reg<<3 + offset]
3894 // NullCheck narrow_oop_reg
3895 //
3896 // On other platforms (Sparc) we have to keep new DecodeN node and
3897 // use it to do implicit NULL check in address:
3898 //
3899 // decode_not_null narrow_oop_reg, base_reg
3900 // [base_reg + offset]
3901 // NullCheck base_reg
3902 //
3903 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3904 // to keep the information to which NULL check the new DecodeN node
3905 // corresponds to use it as value in implicit_null_check().
3906 //
3907 new_in1->set_req(0, n->in(0));
3908 }
3909
3910 n->subsume_by(new_in1, this);
3911 if (in1->outcnt() == 0) {
3912 in1->disconnect_inputs(this);
3913 }
3914 } else {
3915 n->subsume_by(n->in(1), this);
3916 if (n->outcnt() == 0) {
3917 n->disconnect_inputs(this);
3918 }
3919 }
3920 break;
3921 }
3922 #ifdef _LP64
3923 case Op_CmpP:
3924 // Do this transformation here to preserve CmpPNode::sub() and
3925 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3926 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3927 Node* in1 = n->in(1);
3928 Node* in2 = n->in(2);
3929 if (!in1->is_DecodeNarrowPtr()) {
3930 in2 = in1;
3931 in1 = n->in(2);
3932 }
3933 assert(in1->is_DecodeNarrowPtr(), "sanity");
3934
3935 Node* new_in2 = NULL;
3936 if (in2->is_DecodeNarrowPtr()) {
3937 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3938 new_in2 = in2->in(1);
3939 } else if (in2->Opcode() == Op_ConP) {
3940 const Type* t = in2->bottom_type();
3941 if (t == TypePtr::NULL_PTR) {
3942 assert(in1->is_DecodeN(), "compare klass to null?");
3943 // Don't convert CmpP null check into CmpN if compressed
3944 // oops implicit null check is not generated.
3945 // This will allow to generate normal oop implicit null check.
3946 if (Matcher::gen_narrow_oop_implicit_null_checks())
3947 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3948 //
3949 // This transformation together with CastPP transformation above
3950 // will generated code for implicit NULL checks for compressed oops.
3951 //
3952 // The original code after Optimize()
3953 //
3954 // LoadN memory, narrow_oop_reg
3955 // decode narrow_oop_reg, base_reg
3956 // CmpP base_reg, NULL
3957 // CastPP base_reg // NotNull
3958 // Load [base_reg + offset], val_reg
3959 //
3960 // after these transformations will be
3961 //
3962 // LoadN memory, narrow_oop_reg
3963 // CmpN narrow_oop_reg, NULL
3964 // decode_not_null narrow_oop_reg, base_reg
3965 // Load [base_reg + offset], val_reg
3966 //
3967 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3968 // since narrow oops can be used in debug info now (see the code in
3969 // final_graph_reshaping_walk()).
3970 //
3971 // At the end the code will be matched to
3972 // on x86:
3973 //
3974 // Load_narrow_oop memory, narrow_oop_reg
3975 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3976 // NullCheck narrow_oop_reg
3977 //
3978 // and on sparc:
3979 //
3980 // Load_narrow_oop memory, narrow_oop_reg
3981 // decode_not_null narrow_oop_reg, base_reg
3982 // Load [base_reg + offset], val_reg
3983 // NullCheck base_reg
3984 //
3985 } else if (t->isa_oopptr()) {
3986 new_in2 = ConNode::make(t->make_narrowoop());
3987 } else if (t->isa_klassptr()) {
3988 new_in2 = ConNode::make(t->make_narrowklass());
3989 }
3990 }
3991 if (new_in2 != NULL) {
3992 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3993 n->subsume_by(cmpN, this);
3994 if (in1->outcnt() == 0) {
3995 in1->disconnect_inputs(this);
3996 }
3997 if (in2->outcnt() == 0) {
3998 in2->disconnect_inputs(this);
3999 }
4000 }
4001 }
4002 break;
4003
4004 case Op_DecodeN:
4005 case Op_DecodeNKlass:
4006 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4007 // DecodeN could be pinned when it can't be fold into
4008 // an address expression, see the code for Op_CastPP above.
4009 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4010 break;
4011
4012 case Op_EncodeP:
4013 case Op_EncodePKlass: {
4014 Node* in1 = n->in(1);
4015 if (in1->is_DecodeNarrowPtr()) {
4016 n->subsume_by(in1->in(1), this);
4017 } else if (in1->Opcode() == Op_ConP) {
4018 const Type* t = in1->bottom_type();
4019 if (t == TypePtr::NULL_PTR) {
4020 assert(t->isa_oopptr(), "null klass?");
4021 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4022 } else if (t->isa_oopptr()) {
4023 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4024 } else if (t->isa_klassptr()) {
4025 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4026 }
4027 }
4028 if (in1->outcnt() == 0) {
4029 in1->disconnect_inputs(this);
4030 }
4031 break;
4032 }
4033
4034 case Op_Proj: {
4035 if (OptimizeStringConcat || IncrementalInline) {
4036 ProjNode* proj = n->as_Proj();
4037 if (proj->_is_io_use) {
4038 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4039 // Separate projections were used for the exception path which
4040 // are normally removed by a late inline. If it wasn't inlined
4041 // then they will hang around and should just be replaced with
4042 // the original one. Merge them.
4043 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4044 if (non_io_proj != NULL) {
4045 proj->subsume_by(non_io_proj , this);
4046 }
4047 }
4048 }
4049 break;
4050 }
4051
4052 case Op_Phi:
4053 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4054 // The EncodeP optimization may create Phi with the same edges
4055 // for all paths. It is not handled well by Register Allocator.
4056 Node* unique_in = n->in(1);
4057 assert(unique_in != NULL, "");
4058 uint cnt = n->req();
4059 for (uint i = 2; i < cnt; i++) {
4060 Node* m = n->in(i);
4061 assert(m != NULL, "");
4062 if (unique_in != m)
4063 unique_in = NULL;
4064 }
4065 if (unique_in != NULL) {
4066 n->subsume_by(unique_in, this);
4067 }
4068 }
4069 break;
4070
4071 #endif
4072
4073 #ifdef ASSERT
4074 case Op_CastII:
4075 // Verify that all range check dependent CastII nodes were removed.
4076 if (n->isa_CastII()->has_range_check()) {
4077 n->dump(3);
4078 assert(false, "Range check dependent CastII node was not removed");
4079 }
4080 break;
4081 #endif
4082
4083 case Op_ModI:
4084 if (UseDivMod) {
4085 // Check if a%b and a/b both exist
4086 Node* d = n->find_similar(Op_DivI);
4087 if (d) {
4088 // Replace them with a fused divmod if supported
4089 if (Matcher::has_match_rule(Op_DivModI)) {
4090 DivModINode* divmod = DivModINode::make(n);
4091 d->subsume_by(divmod->div_proj(), this);
4092 n->subsume_by(divmod->mod_proj(), this);
4093 } else {
4094 // replace a%b with a-((a/b)*b)
4095 Node* mult = new MulINode(d, d->in(2));
4096 Node* sub = new SubINode(d->in(1), mult);
4097 n->subsume_by(sub, this);
4098 }
4099 }
4100 }
4101 break;
4102
4103 case Op_ModL:
4104 if (UseDivMod) {
4105 // Check if a%b and a/b both exist
4106 Node* d = n->find_similar(Op_DivL);
4107 if (d) {
4108 // Replace them with a fused divmod if supported
4109 if (Matcher::has_match_rule(Op_DivModL)) {
4110 DivModLNode* divmod = DivModLNode::make(n);
4111 d->subsume_by(divmod->div_proj(), this);
4112 n->subsume_by(divmod->mod_proj(), this);
4113 } else {
4114 // replace a%b with a-((a/b)*b)
4115 Node* mult = new MulLNode(d, d->in(2));
4116 Node* sub = new SubLNode(d->in(1), mult);
4117 n->subsume_by(sub, this);
4118 }
4119 }
4120 }
4121 break;
4122
4123 case Op_UModI:
4124 if (UseDivMod) {
4125 // Check if a%b and a/b both exist
4126 Node* d = n->find_similar(Op_UDivI);
4127 if (d) {
4128 // Replace them with a fused unsigned divmod if supported
4129 if (Matcher::has_match_rule(Op_UDivModI)) {
4130 UDivModINode* divmod = UDivModINode::make(n);
4131 d->subsume_by(divmod->div_proj(), this);
4132 n->subsume_by(divmod->mod_proj(), this);
4133 } else {
4134 // replace a%b with a-((a/b)*b)
4135 Node* mult = new MulINode(d, d->in(2));
4136 Node* sub = new SubINode(d->in(1), mult);
4137 n->subsume_by(sub, this);
4138 }
4139 }
4140 }
4141 break;
4142
4143 case Op_UModL:
4144 if (UseDivMod) {
4145 // Check if a%b and a/b both exist
4146 Node* d = n->find_similar(Op_UDivL);
4147 if (d) {
4148 // Replace them with a fused unsigned divmod if supported
4149 if (Matcher::has_match_rule(Op_UDivModL)) {
4150 UDivModLNode* divmod = UDivModLNode::make(n);
4151 d->subsume_by(divmod->div_proj(), this);
4152 n->subsume_by(divmod->mod_proj(), this);
4153 } else {
4154 // replace a%b with a-((a/b)*b)
4155 Node* mult = new MulLNode(d, d->in(2));
4156 Node* sub = new SubLNode(d->in(1), mult);
4157 n->subsume_by(sub, this);
4158 }
4159 }
4160 }
4161 break;
4162
4163 case Op_LoadVector:
4164 case Op_StoreVector:
4165 case Op_LoadVectorGather:
4166 case Op_StoreVectorScatter:
4167 case Op_LoadVectorGatherMasked:
4168 case Op_StoreVectorScatterMasked:
4169 case Op_VectorCmpMasked:
4170 case Op_VectorMaskGen:
4171 case Op_LoadVectorMasked:
4172 case Op_StoreVectorMasked:
4173 break;
4174
4175 case Op_AddReductionVI:
4176 case Op_AddReductionVL:
4177 case Op_AddReductionVF:
4178 case Op_AddReductionVD:
4179 case Op_MulReductionVI:
4180 case Op_MulReductionVL:
4181 case Op_MulReductionVF:
4182 case Op_MulReductionVD:
4183 case Op_MinReductionV:
4184 case Op_MaxReductionV:
4185 case Op_AndReductionV:
4186 case Op_OrReductionV:
4187 case Op_XorReductionV:
4188 break;
4189
4190 case Op_PackB:
4191 case Op_PackS:
4192 case Op_PackI:
4193 case Op_PackF:
4194 case Op_PackL:
4195 case Op_PackD:
4196 if (n->req()-1 > 2) {
4197 // Replace many operand PackNodes with a binary tree for matching
4198 PackNode* p = (PackNode*) n;
4199 Node* btp = p->binary_tree_pack(1, n->req());
4200 n->subsume_by(btp, this);
4201 }
4202 break;
4203 case Op_Loop:
4204 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4205 case Op_CountedLoop:
4206 case Op_LongCountedLoop:
4207 case Op_OuterStripMinedLoop:
4208 if (n->as_Loop()->is_inner_loop()) {
4209 frc.inc_inner_loop_count();
4210 }
4211 n->as_Loop()->verify_strip_mined(0);
4212 break;
4213 case Op_LShiftI:
4214 case Op_RShiftI:
4215 case Op_URShiftI:
4216 case Op_LShiftL:
4217 case Op_RShiftL:
4218 case Op_URShiftL:
4219 if (Matcher::need_masked_shift_count) {
4220 // The cpu's shift instructions don't restrict the count to the
4221 // lower 5/6 bits. We need to do the masking ourselves.
4222 Node* in2 = n->in(2);
4223 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4224 const TypeInt* t = in2->find_int_type();
4225 if (t != NULL && t->is_con()) {
4226 juint shift = t->get_con();
4227 if (shift > mask) { // Unsigned cmp
4228 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4229 }
4230 } else {
4231 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
4232 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4233 n->set_req(2, shift);
4234 }
4235 }
4236 if (in2->outcnt() == 0) { // Remove dead node
4237 in2->disconnect_inputs(this);
4238 }
4239 }
4240 break;
4241 case Op_MemBarStoreStore:
4242 case Op_MemBarRelease:
4243 // Break the link with AllocateNode: it is no longer useful and
4244 // confuses register allocation.
4245 if (n->req() > MemBarNode::Precedent) {
4246 n->set_req(MemBarNode::Precedent, top());
4247 }
4248 break;
4249 case Op_MemBarAcquire: {
4250 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4251 // At parse time, the trailing MemBarAcquire for a volatile load
4252 // is created with an edge to the load. After optimizations,
4253 // that input may be a chain of Phis. If those phis have no
4254 // other use, then the MemBarAcquire keeps them alive and
4255 // register allocation can be confused.
4256 dead_nodes.push(n->in(MemBarNode::Precedent));
4257 n->set_req(MemBarNode::Precedent, top());
4258 }
4259 break;
4260 }
4261 case Op_Blackhole:
4262 break;
4263 case Op_RangeCheck: {
4264 RangeCheckNode* rc = n->as_RangeCheck();
4265 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4266 n->subsume_by(iff, this);
4267 frc._tests.push(iff);
4268 break;
4269 }
4270 case Op_ConvI2L: {
4271 if (!Matcher::convi2l_type_required) {
4272 // Code generation on some platforms doesn't need accurate
4273 // ConvI2L types. Widening the type can help remove redundant
4274 // address computations.
4275 n->as_Type()->set_type(TypeLong::INT);
4276 ResourceMark rm;
4277 Unique_Node_List wq;
4278 wq.push(n);
4279 for (uint next = 0; next < wq.size(); next++) {
4280 Node *m = wq.at(next);
4281
4282 for(;;) {
4283 // Loop over all nodes with identical inputs edges as m
4284 Node* k = m->find_similar(m->Opcode());
4285 if (k == NULL) {
4286 break;
4287 }
4288 // Push their uses so we get a chance to remove node made
4289 // redundant
4290 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4291 Node* u = k->fast_out(i);
4292 if (u->Opcode() == Op_LShiftL ||
4293 u->Opcode() == Op_AddL ||
4294 u->Opcode() == Op_SubL ||
4295 u->Opcode() == Op_AddP) {
4296 wq.push(u);
4297 }
4298 }
4299 // Replace all nodes with identical edges as m with m
4300 k->subsume_by(m, this);
4301 }
4302 }
4303 }
4304 break;
4305 }
4306 case Op_CmpUL: {
4307 if (!Matcher::has_match_rule(Op_CmpUL)) {
4308 // No support for unsigned long comparisons
4309 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4310 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4311 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4312 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4313 Node* andl = new AndLNode(orl, remove_sign_mask);
4314 Node* cmp = new CmpLNode(andl, n->in(2));
4315 n->subsume_by(cmp, this);
4316 }
4317 break;
4318 }
4319 #ifdef ASSERT
4320 case Op_InlineType: {
4321 n->dump(-1);
4322 assert(false, "inline type node was not removed");
4323 break;
4324 }
4325 #endif
4326 default:
4327 assert(!n->is_Call(), "");
4328 assert(!n->is_Mem(), "");
4329 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4330 break;
4331 }
4332 }
4333
4334 //------------------------------final_graph_reshaping_walk---------------------
4335 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4336 // requires that the walk visits a node's inputs before visiting the node.
4337 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4338 Unique_Node_List sfpt;
4339
4340 frc._visited.set(root->_idx); // first, mark node as visited
4341 uint cnt = root->req();
4342 Node *n = root;
4343 uint i = 0;
4344 while (true) {
4345 if (i < cnt) {
4346 // Place all non-visited non-null inputs onto stack
4347 Node* m = n->in(i);
4348 ++i;
4349 if (m != NULL && !frc._visited.test_set(m->_idx)) {
4350 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
4351 // compute worst case interpreter size in case of a deoptimization
4352 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4353
4354 sfpt.push(m);
4355 }
4356 cnt = m->req();
4357 nstack.push(n, i); // put on stack parent and next input's index
4358 n = m;
4359 i = 0;
4360 }
4361 } else {
4362 // Now do post-visit work
4363 final_graph_reshaping_impl(n, frc, dead_nodes);
4364 if (nstack.is_empty())
4365 break; // finished
4366 n = nstack.node(); // Get node from stack
4367 cnt = n->req();
4368 i = nstack.index();
4369 nstack.pop(); // Shift to the next node on stack
4370 }
4371 }
4372
4373 // Skip next transformation if compressed oops are not used.
4374 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4375 (!UseCompressedOops && !UseCompressedClassPointers))
4376 return;
4377
4378 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4379 // It could be done for an uncommon traps or any safepoints/calls
4380 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4381 while (sfpt.size() > 0) {
4382 n = sfpt.pop();
4383 JVMState *jvms = n->as_SafePoint()->jvms();
4384 assert(jvms != NULL, "sanity");
4385 int start = jvms->debug_start();
4386 int end = n->req();
4387 bool is_uncommon = (n->is_CallStaticJava() &&
4388 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4389 for (int j = start; j < end; j++) {
4390 Node* in = n->in(j);
4391 if (in->is_DecodeNarrowPtr()) {
4392 bool safe_to_skip = true;
4393 if (!is_uncommon ) {
4394 // Is it safe to skip?
4395 for (uint i = 0; i < in->outcnt(); i++) {
4396 Node* u = in->raw_out(i);
4397 if (!u->is_SafePoint() ||
4398 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4399 safe_to_skip = false;
4400 }
4401 }
4402 }
4403 if (safe_to_skip) {
4404 n->set_req(j, in->in(1));
4405 }
4406 if (in->outcnt() == 0) {
4407 in->disconnect_inputs(this);
4408 }
4409 }
4410 }
4411 }
4412 }
4413
4414 //------------------------------final_graph_reshaping--------------------------
4415 // Final Graph Reshaping.
4416 //
4417 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4418 // and not commoned up and forced early. Must come after regular
4419 // optimizations to avoid GVN undoing the cloning. Clone constant
4420 // inputs to Loop Phis; these will be split by the allocator anyways.
4421 // Remove Opaque nodes.
4422 // (2) Move last-uses by commutative operations to the left input to encourage
4423 // Intel update-in-place two-address operations and better register usage
4424 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4425 // calls canonicalizing them back.
4426 // (3) Count the number of double-precision FP ops, single-precision FP ops
4427 // and call sites. On Intel, we can get correct rounding either by
4428 // forcing singles to memory (requires extra stores and loads after each
4429 // FP bytecode) or we can set a rounding mode bit (requires setting and
4430 // clearing the mode bit around call sites). The mode bit is only used
4431 // if the relative frequency of single FP ops to calls is low enough.
4432 // This is a key transform for SPEC mpeg_audio.
4433 // (4) Detect infinite loops; blobs of code reachable from above but not
4434 // below. Several of the Code_Gen algorithms fail on such code shapes,
4435 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4436 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4437 // Detection is by looking for IfNodes where only 1 projection is
4438 // reachable from below or CatchNodes missing some targets.
4439 // (5) Assert for insane oop offsets in debug mode.
4440
4441 bool Compile::final_graph_reshaping() {
4442 // an infinite loop may have been eliminated by the optimizer,
4443 // in which case the graph will be empty.
4444 if (root()->req() == 1) {
4445 record_method_not_compilable("trivial infinite loop");
4446 return true;
4447 }
4448
4449 // Expensive nodes have their control input set to prevent the GVN
4450 // from freely commoning them. There's no GVN beyond this point so
4451 // no need to keep the control input. We want the expensive nodes to
4452 // be freely moved to the least frequent code path by gcm.
4453 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4454 for (int i = 0; i < expensive_count(); i++) {
4455 _expensive_nodes.at(i)->set_req(0, NULL);
4456 }
4457
4458 Final_Reshape_Counts frc;
4459
4460 // Visit everybody reachable!
4461 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4462 Node_Stack nstack(live_nodes() >> 1);
4463 Unique_Node_List dead_nodes;
4464 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4465
4466 // Check for unreachable (from below) code (i.e., infinite loops).
4467 for( uint i = 0; i < frc._tests.size(); i++ ) {
4468 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4469 // Get number of CFG targets.
4470 // Note that PCTables include exception targets after calls.
4471 uint required_outcnt = n->required_outcnt();
4472 if (n->outcnt() != required_outcnt) {
4473 // Check for a few special cases. Rethrow Nodes never take the
4474 // 'fall-thru' path, so expected kids is 1 less.
4475 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4476 if (n->in(0)->in(0)->is_Call()) {
4477 CallNode* call = n->in(0)->in(0)->as_Call();
4478 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4479 required_outcnt--; // Rethrow always has 1 less kid
4480 } else if (call->req() > TypeFunc::Parms &&
4481 call->is_CallDynamicJava()) {
4482 // Check for null receiver. In such case, the optimizer has
4483 // detected that the virtual call will always result in a null
4484 // pointer exception. The fall-through projection of this CatchNode
4485 // will not be populated.
4486 Node* arg0 = call->in(TypeFunc::Parms);
4487 if (arg0->is_Type() &&
4488 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4489 required_outcnt--;
4490 }
4491 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4492 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4493 // Check for illegal array length. In such case, the optimizer has
4494 // detected that the allocation attempt will always result in an
4495 // exception. There is no fall-through projection of this CatchNode .
4496 assert(call->is_CallStaticJava(), "static call expected");
4497 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4498 uint valid_length_test_input = call->req() - 1;
4499 Node* valid_length_test = call->in(valid_length_test_input);
4500 call->del_req(valid_length_test_input);
4501 if (valid_length_test->find_int_con(1) == 0) {
4502 required_outcnt--;
4503 }
4504 dead_nodes.push(valid_length_test);
4505 assert(n->outcnt() == required_outcnt, "malformed control flow");
4506 continue;
4507 }
4508 }
4509 }
4510 // Recheck with a better notion of 'required_outcnt'
4511 if (n->outcnt() != required_outcnt) {
4512 record_method_not_compilable("malformed control flow");
4513 return true; // Not all targets reachable!
4514 }
4515 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4516 CallNode* call = n->in(0)->in(0)->as_Call();
4517 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4518 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4519 assert(call->is_CallStaticJava(), "static call expected");
4520 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4521 uint valid_length_test_input = call->req() - 1;
4522 dead_nodes.push(call->in(valid_length_test_input));
4523 call->del_req(valid_length_test_input); // valid length test useless now
4524 }
4525 }
4526 // Check that I actually visited all kids. Unreached kids
4527 // must be infinite loops.
4528 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4529 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4530 record_method_not_compilable("infinite loop");
4531 return true; // Found unvisited kid; must be unreach
4532 }
4533
4534 // Here so verification code in final_graph_reshaping_walk()
4535 // always see an OuterStripMinedLoopEnd
4536 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4537 IfNode* init_iff = n->as_If();
4538 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4539 n->subsume_by(iff, this);
4540 }
4541 }
4542
4543 while (dead_nodes.size() > 0) {
4544 Node* m = dead_nodes.pop();
4545 if (m->outcnt() == 0 && m != top()) {
4546 for (uint j = 0; j < m->req(); j++) {
4547 Node* in = m->in(j);
4548 if (in != NULL) {
4549 dead_nodes.push(in);
4550 }
4551 }
4552 m->disconnect_inputs(this);
4553 }
4554 }
4555
4556 #ifdef IA32
4557 // If original bytecodes contained a mixture of floats and doubles
4558 // check if the optimizer has made it homogeneous, item (3).
4559 if (UseSSE == 0 &&
4560 frc.get_float_count() > 32 &&
4561 frc.get_double_count() == 0 &&
4562 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4563 set_24_bit_selection_and_mode(false, true);
4564 }
4565 #endif // IA32
4566
4567 set_java_calls(frc.get_java_call_count());
4568 set_inner_loops(frc.get_inner_loop_count());
4569
4570 // No infinite loops, no reason to bail out.
4571 return false;
4572 }
4573
4574 //-----------------------------too_many_traps----------------------------------
4575 // Report if there are too many traps at the current method and bci.
4576 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4577 bool Compile::too_many_traps(ciMethod* method,
4578 int bci,
4579 Deoptimization::DeoptReason reason) {
4580 ciMethodData* md = method->method_data();
4581 if (md->is_empty()) {
4582 // Assume the trap has not occurred, or that it occurred only
4583 // because of a transient condition during start-up in the interpreter.
4584 return false;
4585 }
4586 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4587 if (md->has_trap_at(bci, m, reason) != 0) {
4588 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4589 // Also, if there are multiple reasons, or if there is no per-BCI record,
4590 // assume the worst.
4591 if (log())
4592 log()->elem("observe trap='%s' count='%d'",
4593 Deoptimization::trap_reason_name(reason),
4594 md->trap_count(reason));
4595 return true;
4596 } else {
4597 // Ignore method/bci and see if there have been too many globally.
4598 return too_many_traps(reason, md);
4599 }
4600 }
4601
4602 // Less-accurate variant which does not require a method and bci.
4603 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4604 ciMethodData* logmd) {
4605 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4606 // Too many traps globally.
4607 // Note that we use cumulative trap_count, not just md->trap_count.
4608 if (log()) {
4609 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4610 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4611 Deoptimization::trap_reason_name(reason),
4612 mcount, trap_count(reason));
4613 }
4614 return true;
4615 } else {
4616 // The coast is clear.
4617 return false;
4618 }
4619 }
4620
4621 //--------------------------too_many_recompiles--------------------------------
4622 // Report if there are too many recompiles at the current method and bci.
4623 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4624 // Is not eager to return true, since this will cause the compiler to use
4625 // Action_none for a trap point, to avoid too many recompilations.
4626 bool Compile::too_many_recompiles(ciMethod* method,
4627 int bci,
4628 Deoptimization::DeoptReason reason) {
4629 ciMethodData* md = method->method_data();
4630 if (md->is_empty()) {
4631 // Assume the trap has not occurred, or that it occurred only
4632 // because of a transient condition during start-up in the interpreter.
4633 return false;
4634 }
4635 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4636 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4637 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4638 Deoptimization::DeoptReason per_bc_reason
4639 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4640 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4641 if ((per_bc_reason == Deoptimization::Reason_none
4642 || md->has_trap_at(bci, m, reason) != 0)
4643 // The trap frequency measure we care about is the recompile count:
4644 && md->trap_recompiled_at(bci, m)
4645 && md->overflow_recompile_count() >= bc_cutoff) {
4646 // Do not emit a trap here if it has already caused recompilations.
4647 // Also, if there are multiple reasons, or if there is no per-BCI record,
4648 // assume the worst.
4649 if (log())
4650 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4651 Deoptimization::trap_reason_name(reason),
4652 md->trap_count(reason),
4653 md->overflow_recompile_count());
4654 return true;
4655 } else if (trap_count(reason) != 0
4656 && decompile_count() >= m_cutoff) {
4657 // Too many recompiles globally, and we have seen this sort of trap.
4658 // Use cumulative decompile_count, not just md->decompile_count.
4659 if (log())
4660 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4661 Deoptimization::trap_reason_name(reason),
4662 md->trap_count(reason), trap_count(reason),
4663 md->decompile_count(), decompile_count());
4664 return true;
4665 } else {
4666 // The coast is clear.
4667 return false;
4668 }
4669 }
4670
4671 // Compute when not to trap. Used by matching trap based nodes and
4672 // NullCheck optimization.
4673 void Compile::set_allowed_deopt_reasons() {
4674 _allowed_reasons = 0;
4675 if (is_method_compilation()) {
4676 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4677 assert(rs < BitsPerInt, "recode bit map");
4678 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4679 _allowed_reasons |= nth_bit(rs);
4680 }
4681 }
4682 }
4683 }
4684
4685 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4686 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4687 }
4688
4689 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4690 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4691 }
4692
4693 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4694 if (holder->is_initialized()) {
4695 return false;
4696 }
4697 if (holder->is_being_initialized()) {
4698 if (accessing_method->holder() == holder) {
4699 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4700 // <init>, or a static method. In all those cases, there was an initialization
4701 // barrier on the holder klass passed.
4702 if (accessing_method->is_class_initializer() ||
4703 accessing_method->is_object_constructor() ||
4704 accessing_method->is_static()) {
4705 return false;
4706 }
4707 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4708 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4709 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4710 // child class can become fully initialized while its parent class is still being initialized.
4711 if (accessing_method->is_class_initializer()) {
4712 return false;
4713 }
4714 }
4715 ciMethod* root = method(); // the root method of compilation
4716 if (root != accessing_method) {
4717 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4718 }
4719 }
4720 return true;
4721 }
4722
4723 #ifndef PRODUCT
4724 //------------------------------verify_bidirectional_edges---------------------
4725 // For each input edge to a node (ie - for each Use-Def edge), verify that
4726 // there is a corresponding Def-Use edge.
4727 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4728 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4729 uint stack_size = live_nodes() >> 4;
4730 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4731 nstack.push(_root);
4732
4733 while (nstack.size() > 0) {
4734 Node* n = nstack.pop();
4735 if (visited.member(n)) {
4736 continue;
4737 }
4738 visited.push(n);
4739
4740 // Walk over all input edges, checking for correspondence
4741 uint length = n->len();
4742 for (uint i = 0; i < length; i++) {
4743 Node* in = n->in(i);
4744 if (in != NULL && !visited.member(in)) {
4745 nstack.push(in); // Put it on stack
4746 }
4747 if (in != NULL && !in->is_top()) {
4748 // Count instances of `next`
4749 int cnt = 0;
4750 for (uint idx = 0; idx < in->_outcnt; idx++) {
4751 if (in->_out[idx] == n) {
4752 cnt++;
4753 }
4754 }
4755 assert(cnt > 0, "Failed to find Def-Use edge.");
4756 // Check for duplicate edges
4757 // walk the input array downcounting the input edges to n
4758 for (uint j = 0; j < length; j++) {
4759 if (n->in(j) == in) {
4760 cnt--;
4761 }
4762 }
4763 assert(cnt == 0, "Mismatched edge count.");
4764 } else if (in == NULL) {
4765 assert(i == 0 || i >= n->req() ||
4766 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4767 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
4768 (n->is_Unlock() && i == (n->req() - 1)) ||
4769 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4770 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
4771 } else {
4772 assert(in->is_top(), "sanity");
4773 // Nothing to check.
4774 }
4775 }
4776 }
4777 }
4778
4779 //------------------------------verify_graph_edges---------------------------
4780 // Walk the Graph and verify that there is a one-to-one correspondence
4781 // between Use-Def edges and Def-Use edges in the graph.
4782 void Compile::verify_graph_edges(bool no_dead_code) {
4783 if (VerifyGraphEdges) {
4784 Unique_Node_List visited;
4785
4786 // Call graph walk to check edges
4787 verify_bidirectional_edges(visited);
4788 if (no_dead_code) {
4789 // Now make sure that no visited node is used by an unvisited node.
4790 bool dead_nodes = false;
4791 Unique_Node_List checked;
4792 while (visited.size() > 0) {
4793 Node* n = visited.pop();
4794 checked.push(n);
4795 for (uint i = 0; i < n->outcnt(); i++) {
4796 Node* use = n->raw_out(i);
4797 if (checked.member(use)) continue; // already checked
4798 if (visited.member(use)) continue; // already in the graph
4799 if (use->is_Con()) continue; // a dead ConNode is OK
4800 // At this point, we have found a dead node which is DU-reachable.
4801 if (!dead_nodes) {
4802 tty->print_cr("*** Dead nodes reachable via DU edges:");
4803 dead_nodes = true;
4804 }
4805 use->dump(2);
4806 tty->print_cr("---");
4807 checked.push(use); // No repeats; pretend it is now checked.
4808 }
4809 }
4810 assert(!dead_nodes, "using nodes must be reachable from root");
4811 }
4812 }
4813 }
4814 #endif
4815
4816 // The Compile object keeps track of failure reasons separately from the ciEnv.
4817 // This is required because there is not quite a 1-1 relation between the
4818 // ciEnv and its compilation task and the Compile object. Note that one
4819 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4820 // to backtrack and retry without subsuming loads. Other than this backtracking
4821 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4822 // by the logic in C2Compiler.
4823 void Compile::record_failure(const char* reason) {
4824 if (log() != NULL) {
4825 log()->elem("failure reason='%s' phase='compile'", reason);
4826 }
4827 if (_failure_reason == NULL) {
4828 // Record the first failure reason.
4829 _failure_reason = reason;
4830 }
4831
4832 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4833 C->print_method(PHASE_FAILURE, 1);
4834 }
4835 _root = NULL; // flush the graph, too
4836 }
4837
4838 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4839 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4840 _phase_name(name), _dolog(CITimeVerbose)
4841 {
4842 if (_dolog) {
4843 C = Compile::current();
4844 _log = C->log();
4845 } else {
4846 C = NULL;
4847 _log = NULL;
4848 }
4849 if (_log != NULL) {
4850 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4851 _log->stamp();
4852 _log->end_head();
4853 }
4854 }
4855
4856 Compile::TracePhase::~TracePhase() {
4857
4858 C = Compile::current();
4859 if (_dolog) {
4860 _log = C->log();
4861 } else {
4862 _log = NULL;
4863 }
4864
4865 #ifdef ASSERT
4866 if (PrintIdealNodeCount) {
4867 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4868 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4869 }
4870
4871 if (VerifyIdealNodeCount) {
4872 Compile::current()->print_missing_nodes();
4873 }
4874 #endif
4875
4876 if (_log != NULL) {
4877 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4878 }
4879 }
4880
4881 //----------------------------static_subtype_check-----------------------------
4882 // Shortcut important common cases when superklass is exact:
4883 // (0) superklass is java.lang.Object (can occur in reflective code)
4884 // (1) subklass is already limited to a subtype of superklass => always ok
4885 // (2) subklass does not overlap with superklass => always fail
4886 // (3) superklass has NO subtypes and we can check with a simple compare.
4887 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4888 if (skip) {
4889 return SSC_full_test; // Let caller generate the general case.
4890 }
4891
4892 if (subk->is_java_subtype_of(superk)) {
4893 return SSC_always_true; // (0) and (1) this test cannot fail
4894 }
4895
4896 if (!subk->maybe_java_subtype_of(superk)) {
4897 return SSC_always_false; // (2) true path dead; no dynamic test needed
4898 }
4899
4900 const Type* superelem = superk;
4901 if (superk->isa_aryklassptr()) {
4902 int ignored;
4903 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4904
4905 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays.
4906 // [QMyValue is a subtype of [LMyValue but the klass for [QMyValue is not equal to
4907 // the klass for [LMyValue. Perform a full test.
4908 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
4909 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
4910 return SSC_full_test;
4911 }
4912 }
4913
4914 if (superelem->isa_instklassptr()) {
4915 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4916 if (!ik->has_subklass()) {
4917 if (!ik->is_final()) {
4918 // Add a dependency if there is a chance of a later subclass.
4919 dependencies()->assert_leaf_type(ik);
4920 }
4921 if (!superk->maybe_java_subtype_of(subk)) {
4922 return SSC_always_false;
4923 }
4924 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4925 }
4926 } else {
4927 // A primitive array type has no subtypes.
4928 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4929 }
4930
4931 return SSC_full_test;
4932 }
4933
4934 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4935 #ifdef _LP64
4936 // The scaled index operand to AddP must be a clean 64-bit value.
4937 // Java allows a 32-bit int to be incremented to a negative
4938 // value, which appears in a 64-bit register as a large
4939 // positive number. Using that large positive number as an
4940 // operand in pointer arithmetic has bad consequences.
4941 // On the other hand, 32-bit overflow is rare, and the possibility
4942 // can often be excluded, if we annotate the ConvI2L node with
4943 // a type assertion that its value is known to be a small positive
4944 // number. (The prior range check has ensured this.)
4945 // This assertion is used by ConvI2LNode::Ideal.
4946 int index_max = max_jint - 1; // array size is max_jint, index is one less
4947 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4948 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4949 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4950 #endif
4951 return idx;
4952 }
4953
4954 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4955 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4956 if (ctrl != NULL) {
4957 // Express control dependency by a CastII node with a narrow type.
4958 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4959 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4960 // node from floating above the range check during loop optimizations. Otherwise, the
4961 // ConvI2L node may be eliminated independently of the range check, causing the data path
4962 // to become TOP while the control path is still there (although it's unreachable).
4963 value->set_req(0, ctrl);
4964 value = phase->transform(value);
4965 }
4966 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4967 return phase->transform(new ConvI2LNode(value, ltype));
4968 }
4969
4970 // The message about the current inlining is accumulated in
4971 // _print_inlining_stream and transferred into the _print_inlining_list
4972 // once we know whether inlining succeeds or not. For regular
4973 // inlining, messages are appended to the buffer pointed by
4974 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4975 // a new buffer is added after _print_inlining_idx in the list. This
4976 // way we can update the inlining message for late inlining call site
4977 // when the inlining is attempted again.
4978 void Compile::print_inlining_init() {
4979 if (print_inlining() || print_intrinsics()) {
4980 // print_inlining_init is actually called several times.
4981 print_inlining_reset();
4982 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4983 }
4984 }
4985
4986 void Compile::print_inlining_reinit() {
4987 if (print_inlining() || print_intrinsics()) {
4988 print_inlining_reset();
4989 }
4990 }
4991
4992 void Compile::print_inlining_reset() {
4993 _print_inlining_stream->reset();
4994 }
4995
4996 void Compile::print_inlining_commit() {
4997 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4998 // Transfer the message from _print_inlining_stream to the current
4999 // _print_inlining_list buffer and clear _print_inlining_stream.
5000 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
5001 print_inlining_reset();
5002 }
5003
5004 void Compile::print_inlining_push() {
5005 // Add new buffer to the _print_inlining_list at current position
5006 _print_inlining_idx++;
5007 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
5008 }
5009
5010 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
5011 return _print_inlining_list->at(_print_inlining_idx);
5012 }
5013
5014 void Compile::print_inlining_update(CallGenerator* cg) {
5015 if (print_inlining() || print_intrinsics()) {
5016 if (cg->is_late_inline()) {
5017 if (print_inlining_current()->cg() != cg &&
5018 (print_inlining_current()->cg() != NULL ||
5019 print_inlining_current()->ss()->size() != 0)) {
5020 print_inlining_push();
5021 }
5022 print_inlining_commit();
5023 print_inlining_current()->set_cg(cg);
5024 } else {
5025 if (print_inlining_current()->cg() != NULL) {
5026 print_inlining_push();
5027 }
5028 print_inlining_commit();
5029 }
5030 }
5031 }
5032
5033 void Compile::print_inlining_move_to(CallGenerator* cg) {
5034 // We resume inlining at a late inlining call site. Locate the
5035 // corresponding inlining buffer so that we can update it.
5036 if (print_inlining() || print_intrinsics()) {
5037 for (int i = 0; i < _print_inlining_list->length(); i++) {
5038 if (_print_inlining_list->at(i)->cg() == cg) {
5039 _print_inlining_idx = i;
5040 return;
5041 }
5042 }
5043 ShouldNotReachHere();
5044 }
5045 }
5046
5047 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
5048 if (print_inlining() || print_intrinsics()) {
5049 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
5050 assert(print_inlining_current()->cg() == cg, "wrong entry");
5051 // replace message with new message
5052 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
5053 print_inlining_commit();
5054 print_inlining_current()->set_cg(cg);
5055 }
5056 }
5057
5058 void Compile::print_inlining_assert_ready() {
5059 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
5060 }
5061
5062 void Compile::process_print_inlining() {
5063 assert(_late_inlines.length() == 0, "not drained yet");
5064 if (print_inlining() || print_intrinsics()) {
5065 ResourceMark rm;
5066 stringStream ss;
5067 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
5068 for (int i = 0; i < _print_inlining_list->length(); i++) {
5069 PrintInliningBuffer* pib = _print_inlining_list->at(i);
5070 ss.print("%s", pib->ss()->freeze());
5071 delete pib;
5072 DEBUG_ONLY(_print_inlining_list->at_put(i, NULL));
5073 }
5074 // Reset _print_inlining_list, it only contains destructed objects.
5075 // It is on the arena, so it will be freed when the arena is reset.
5076 _print_inlining_list = NULL;
5077 // _print_inlining_stream won't be used anymore, either.
5078 print_inlining_reset();
5079 size_t end = ss.size();
5080 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
5081 strncpy(_print_inlining_output, ss.freeze(), end+1);
5082 _print_inlining_output[end] = 0;
5083 }
5084 }
5085
5086 void Compile::dump_print_inlining() {
5087 if (_print_inlining_output != NULL) {
5088 tty->print_raw(_print_inlining_output);
5089 }
5090 }
5091
5092 void Compile::log_late_inline(CallGenerator* cg) {
5093 if (log() != NULL) {
5094 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5095 cg->unique_id());
5096 JVMState* p = cg->call_node()->jvms();
5097 while (p != NULL) {
5098 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5099 p = p->caller();
5100 }
5101 log()->tail("late_inline");
5102 }
5103 }
5104
5105 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5106 log_late_inline(cg);
5107 if (log() != NULL) {
5108 log()->inline_fail(msg);
5109 }
5110 }
5111
5112 void Compile::log_inline_id(CallGenerator* cg) {
5113 if (log() != NULL) {
5114 // The LogCompilation tool needs a unique way to identify late
5115 // inline call sites. This id must be unique for this call site in
5116 // this compilation. Try to have it unique across compilations as
5117 // well because it can be convenient when grepping through the log
5118 // file.
5119 // Distinguish OSR compilations from others in case CICountOSR is
5120 // on.
5121 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5122 cg->set_unique_id(id);
5123 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5124 }
5125 }
5126
5127 void Compile::log_inline_failure(const char* msg) {
5128 if (C->log() != NULL) {
5129 C->log()->inline_fail(msg);
5130 }
5131 }
5132
5133
5134 // Dump inlining replay data to the stream.
5135 // Don't change thread state and acquire any locks.
5136 void Compile::dump_inline_data(outputStream* out) {
5137 InlineTree* inl_tree = ilt();
5138 if (inl_tree != NULL) {
5139 out->print(" inline %d", inl_tree->count());
5140 inl_tree->dump_replay_data(out);
5141 }
5142 }
5143
5144 void Compile::dump_inline_data_reduced(outputStream* out) {
5145 assert(ReplayReduce, "");
5146
5147 InlineTree* inl_tree = ilt();
5148 if (inl_tree == NULL) {
5149 return;
5150 }
5151 // Enable iterative replay file reduction
5152 // Output "compile" lines for depth 1 subtrees,
5153 // simulating that those trees were compiled
5154 // instead of inlined.
5155 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5156 InlineTree* sub = inl_tree->subtrees().at(i);
5157 if (sub->inline_level() != 1) {
5158 continue;
5159 }
5160
5161 ciMethod* method = sub->method();
5162 int entry_bci = -1;
5163 int comp_level = env()->task()->comp_level();
5164 out->print("compile ");
5165 method->dump_name_as_ascii(out);
5166 out->print(" %d %d", entry_bci, comp_level);
5167 out->print(" inline %d", sub->count());
5168 sub->dump_replay_data(out, -1);
5169 out->cr();
5170 }
5171 }
5172
5173 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5174 if (n1->Opcode() < n2->Opcode()) return -1;
5175 else if (n1->Opcode() > n2->Opcode()) return 1;
5176
5177 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5178 for (uint i = 1; i < n1->req(); i++) {
5179 if (n1->in(i) < n2->in(i)) return -1;
5180 else if (n1->in(i) > n2->in(i)) return 1;
5181 }
5182
5183 return 0;
5184 }
5185
5186 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5187 Node* n1 = *n1p;
5188 Node* n2 = *n2p;
5189
5190 return cmp_expensive_nodes(n1, n2);
5191 }
5192
5193 void Compile::sort_expensive_nodes() {
5194 if (!expensive_nodes_sorted()) {
5195 _expensive_nodes.sort(cmp_expensive_nodes);
5196 }
5197 }
5198
5199 bool Compile::expensive_nodes_sorted() const {
5200 for (int i = 1; i < _expensive_nodes.length(); i++) {
5201 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5202 return false;
5203 }
5204 }
5205 return true;
5206 }
5207
5208 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5209 if (_expensive_nodes.length() == 0) {
5210 return false;
5211 }
5212
5213 assert(OptimizeExpensiveOps, "optimization off?");
5214
5215 // Take this opportunity to remove dead nodes from the list
5216 int j = 0;
5217 for (int i = 0; i < _expensive_nodes.length(); i++) {
5218 Node* n = _expensive_nodes.at(i);
5219 if (!n->is_unreachable(igvn)) {
5220 assert(n->is_expensive(), "should be expensive");
5221 _expensive_nodes.at_put(j, n);
5222 j++;
5223 }
5224 }
5225 _expensive_nodes.trunc_to(j);
5226
5227 // Then sort the list so that similar nodes are next to each other
5228 // and check for at least two nodes of identical kind with same data
5229 // inputs.
5230 sort_expensive_nodes();
5231
5232 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5233 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5234 return true;
5235 }
5236 }
5237
5238 return false;
5239 }
5240
5241 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5242 if (_expensive_nodes.length() == 0) {
5243 return;
5244 }
5245
5246 assert(OptimizeExpensiveOps, "optimization off?");
5247
5248 // Sort to bring similar nodes next to each other and clear the
5249 // control input of nodes for which there's only a single copy.
5250 sort_expensive_nodes();
5251
5252 int j = 0;
5253 int identical = 0;
5254 int i = 0;
5255 bool modified = false;
5256 for (; i < _expensive_nodes.length()-1; i++) {
5257 assert(j <= i, "can't write beyond current index");
5258 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5259 identical++;
5260 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5261 continue;
5262 }
5263 if (identical > 0) {
5264 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5265 identical = 0;
5266 } else {
5267 Node* n = _expensive_nodes.at(i);
5268 igvn.replace_input_of(n, 0, NULL);
5269 igvn.hash_insert(n);
5270 modified = true;
5271 }
5272 }
5273 if (identical > 0) {
5274 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5275 } else if (_expensive_nodes.length() >= 1) {
5276 Node* n = _expensive_nodes.at(i);
5277 igvn.replace_input_of(n, 0, NULL);
5278 igvn.hash_insert(n);
5279 modified = true;
5280 }
5281 _expensive_nodes.trunc_to(j);
5282 if (modified) {
5283 igvn.optimize();
5284 }
5285 }
5286
5287 void Compile::add_expensive_node(Node * n) {
5288 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5289 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5290 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5291 if (OptimizeExpensiveOps) {
5292 _expensive_nodes.append(n);
5293 } else {
5294 // Clear control input and let IGVN optimize expensive nodes if
5295 // OptimizeExpensiveOps is off.
5296 n->set_req(0, NULL);
5297 }
5298 }
5299
5300 /**
5301 * Track coarsened Lock and Unlock nodes.
5302 */
5303
5304 class Lock_List : public Node_List {
5305 uint _origin_cnt;
5306 public:
5307 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5308 uint origin_cnt() const { return _origin_cnt; }
5309 };
5310
5311 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5312 int length = locks.length();
5313 if (length > 0) {
5314 // Have to keep this list until locks elimination during Macro nodes elimination.
5315 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5316 for (int i = 0; i < length; i++) {
5317 AbstractLockNode* lock = locks.at(i);
5318 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5319 locks_list->push(lock);
5320 }
5321 _coarsened_locks.append(locks_list);
5322 }
5323 }
5324
5325 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5326 int count = coarsened_count();
5327 for (int i = 0; i < count; i++) {
5328 Node_List* locks_list = _coarsened_locks.at(i);
5329 for (uint j = 0; j < locks_list->size(); j++) {
5330 Node* lock = locks_list->at(j);
5331 assert(lock->is_AbstractLock(), "sanity");
5332 if (!useful.member(lock)) {
5333 locks_list->yank(lock);
5334 }
5335 }
5336 }
5337 }
5338
5339 void Compile::remove_coarsened_lock(Node* n) {
5340 if (n->is_AbstractLock()) {
5341 int count = coarsened_count();
5342 for (int i = 0; i < count; i++) {
5343 Node_List* locks_list = _coarsened_locks.at(i);
5344 locks_list->yank(n);
5345 }
5346 }
5347 }
5348
5349 bool Compile::coarsened_locks_consistent() {
5350 int count = coarsened_count();
5351 for (int i = 0; i < count; i++) {
5352 bool unbalanced = false;
5353 bool modified = false; // track locks kind modifications
5354 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5355 uint size = locks_list->size();
5356 if (size == 0) {
5357 unbalanced = false; // All locks were eliminated - good
5358 } else if (size != locks_list->origin_cnt()) {
5359 unbalanced = true; // Some locks were removed from list
5360 } else {
5361 for (uint j = 0; j < size; j++) {
5362 Node* lock = locks_list->at(j);
5363 // All nodes in group should have the same state (modified or not)
5364 if (!lock->as_AbstractLock()->is_coarsened()) {
5365 if (j == 0) {
5366 // first on list was modified, the rest should be too for consistency
5367 modified = true;
5368 } else if (!modified) {
5369 // this lock was modified but previous locks on the list were not
5370 unbalanced = true;
5371 break;
5372 }
5373 } else if (modified) {
5374 // previous locks on list were modified but not this lock
5375 unbalanced = true;
5376 break;
5377 }
5378 }
5379 }
5380 if (unbalanced) {
5381 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5382 #ifdef ASSERT
5383 if (PrintEliminateLocks) {
5384 tty->print_cr("=== unbalanced coarsened locks ===");
5385 for (uint l = 0; l < size; l++) {
5386 locks_list->at(l)->dump();
5387 }
5388 }
5389 #endif
5390 record_failure(C2Compiler::retry_no_locks_coarsening());
5391 return false;
5392 }
5393 }
5394 return true;
5395 }
5396
5397 /**
5398 * Remove the speculative part of types and clean up the graph
5399 */
5400 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5401 if (UseTypeSpeculation) {
5402 Unique_Node_List worklist;
5403 worklist.push(root());
5404 int modified = 0;
5405 // Go over all type nodes that carry a speculative type, drop the
5406 // speculative part of the type and enqueue the node for an igvn
5407 // which may optimize it out.
5408 for (uint next = 0; next < worklist.size(); ++next) {
5409 Node *n = worklist.at(next);
5410 if (n->is_Type()) {
5411 TypeNode* tn = n->as_Type();
5412 const Type* t = tn->type();
5413 const Type* t_no_spec = t->remove_speculative();
5414 if (t_no_spec != t) {
5415 bool in_hash = igvn.hash_delete(n);
5416 assert(in_hash, "node should be in igvn hash table");
5417 tn->set_type(t_no_spec);
5418 igvn.hash_insert(n);
5419 igvn._worklist.push(n); // give it a chance to go away
5420 modified++;
5421 }
5422 }
5423 // Iterate over outs - endless loops is unreachable from below
5424 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5425 Node *m = n->fast_out(i);
5426 if (not_a_node(m)) {
5427 continue;
5428 }
5429 worklist.push(m);
5430 }
5431 }
5432 // Drop the speculative part of all types in the igvn's type table
5433 igvn.remove_speculative_types();
5434 if (modified > 0) {
5435 igvn.optimize();
5436 }
5437 #ifdef ASSERT
5438 // Verify that after the IGVN is over no speculative type has resurfaced
5439 worklist.clear();
5440 worklist.push(root());
5441 for (uint next = 0; next < worklist.size(); ++next) {
5442 Node *n = worklist.at(next);
5443 const Type* t = igvn.type_or_null(n);
5444 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
5445 if (n->is_Type()) {
5446 t = n->as_Type()->type();
5447 assert(t == t->remove_speculative(), "no more speculative types");
5448 }
5449 // Iterate over outs - endless loops is unreachable from below
5450 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5451 Node *m = n->fast_out(i);
5452 if (not_a_node(m)) {
5453 continue;
5454 }
5455 worklist.push(m);
5456 }
5457 }
5458 igvn.check_no_speculative_types();
5459 #endif
5460 }
5461 }
5462
5463 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5464 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5465 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5466 if (!EnableValhalla || ta == NULL || tb == NULL ||
5467 ta->is_zero_type() || tb->is_zero_type() ||
5468 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5469 // Use old acmp if one operand is null or not an inline type
5470 return new CmpPNode(a, b);
5471 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5472 // We know that one operand is an inline type. Therefore,
5473 // new acmp will only return true if both operands are NULL.
5474 // Check if both operands are null by or'ing the oops.
5475 a = phase->transform(new CastP2XNode(NULL, a));
5476 b = phase->transform(new CastP2XNode(NULL, b));
5477 a = phase->transform(new OrXNode(a, b));
5478 return new CmpXNode(a, phase->MakeConX(0));
5479 }
5480 // Use new acmp
5481 return NULL;
5482 }
5483
5484 // Auxiliary methods to support randomized stressing/fuzzing.
5485
5486 int Compile::random() {
5487 _stress_seed = os::next_random(_stress_seed);
5488 return static_cast<int>(_stress_seed);
5489 }
5490
5491 // This method can be called the arbitrary number of times, with current count
5492 // as the argument. The logic allows selecting a single candidate from the
5493 // running list of candidates as follows:
5494 // int count = 0;
5495 // Cand* selected = null;
5496 // while(cand = cand->next()) {
5497 // if (randomized_select(++count)) {
5498 // selected = cand;
5499 // }
5500 // }
5501 //
5502 // Including count equalizes the chances any candidate is "selected".
5503 // This is useful when we don't have the complete list of candidates to choose
5504 // from uniformly. In this case, we need to adjust the randomicity of the
5505 // selection, or else we will end up biasing the selection towards the latter
5506 // candidates.
5507 //
5508 // Quick back-envelope calculation shows that for the list of n candidates
5509 // the equal probability for the candidate to persist as "best" can be
5510 // achieved by replacing it with "next" k-th candidate with the probability
5511 // of 1/k. It can be easily shown that by the end of the run, the
5512 // probability for any candidate is converged to 1/n, thus giving the
5513 // uniform distribution among all the candidates.
5514 //
5515 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5516 #define RANDOMIZED_DOMAIN_POW 29
5517 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5518 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5519 bool Compile::randomized_select(int count) {
5520 assert(count > 0, "only positive");
5521 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5522 }
5523
5524 CloneMap& Compile::clone_map() { return _clone_map; }
5525 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5526
5527 void NodeCloneInfo::dump() const {
5528 tty->print(" {%d:%d} ", idx(), gen());
5529 }
5530
5531 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5532 uint64_t val = value(old->_idx);
5533 NodeCloneInfo cio(val);
5534 assert(val != 0, "old node should be in the map");
5535 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5536 insert(nnn->_idx, cin.get());
5537 #ifndef PRODUCT
5538 if (is_debug()) {
5539 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5540 }
5541 #endif
5542 }
5543
5544 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5545 NodeCloneInfo cio(value(old->_idx));
5546 if (cio.get() == 0) {
5547 cio.set(old->_idx, 0);
5548 insert(old->_idx, cio.get());
5549 #ifndef PRODUCT
5550 if (is_debug()) {
5551 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5552 }
5553 #endif
5554 }
5555 clone(old, nnn, gen);
5556 }
5557
5558 int CloneMap::max_gen() const {
5559 int g = 0;
5560 DictI di(_dict);
5561 for(; di.test(); ++di) {
5562 int t = gen(di._key);
5563 if (g < t) {
5564 g = t;
5565 #ifndef PRODUCT
5566 if (is_debug()) {
5567 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5568 }
5569 #endif
5570 }
5571 }
5572 return g;
5573 }
5574
5575 void CloneMap::dump(node_idx_t key) const {
5576 uint64_t val = value(key);
5577 if (val != 0) {
5578 NodeCloneInfo ni(val);
5579 ni.dump();
5580 }
5581 }
5582
5583 // Move Allocate nodes to the start of the list
5584 void Compile::sort_macro_nodes() {
5585 int count = macro_count();
5586 int allocates = 0;
5587 for (int i = 0; i < count; i++) {
5588 Node* n = macro_node(i);
5589 if (n->is_Allocate()) {
5590 if (i != allocates) {
5591 Node* tmp = macro_node(allocates);
5592 _macro_nodes.at_put(allocates, n);
5593 _macro_nodes.at_put(i, tmp);
5594 }
5595 allocates++;
5596 }
5597 }
5598 }
5599
5600 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5601 EventCompilerPhase event;
5602 if (event.should_commit()) {
5603 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5604 }
5605 #ifndef PRODUCT
5606 ResourceMark rm;
5607 stringStream ss;
5608 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5609 if (n != nullptr) {
5610 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5611 }
5612
5613 const char* name = ss.as_string();
5614 if (should_print_igv(level)) {
5615 _igv_printer->print_method(name, level);
5616 }
5617 if (should_print_phase(cpt)) {
5618 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5619 }
5620 #endif
5621 C->_latest_stage_start_counter.stamp();
5622 }
5623
5624 // Only used from CompileWrapper
5625 void Compile::begin_method() {
5626 #ifndef PRODUCT
5627 if (_method != NULL && should_print_igv(1)) {
5628 _igv_printer->begin_method();
5629 }
5630 #endif
5631 C->_latest_stage_start_counter.stamp();
5632 }
5633
5634 // Only used from CompileWrapper
5635 void Compile::end_method() {
5636 EventCompilerPhase event;
5637 if (event.should_commit()) {
5638 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5639 }
5640
5641 #ifndef PRODUCT
5642 if (_method != NULL && should_print_igv(1)) {
5643 _igv_printer->end_method();
5644 }
5645 #endif
5646 }
5647
5648 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5649 #ifndef PRODUCT
5650 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) {
5651 return true;
5652 }
5653 #endif
5654 return false;
5655 }
5656
5657 bool Compile::should_print_igv(int level) {
5658 #ifndef PRODUCT
5659 if (PrintIdealGraphLevel < 0) { // disabled by the user
5660 return false;
5661 }
5662
5663 bool need = directive()->IGVPrintLevelOption >= level;
5664 if (need && !_igv_printer) {
5665 _igv_printer = IdealGraphPrinter::printer();
5666 _igv_printer->set_compile(this);
5667 }
5668 return need;
5669 #else
5670 return false;
5671 #endif
5672 }
5673
5674 #ifndef PRODUCT
5675 IdealGraphPrinter* Compile::_debug_file_printer = NULL;
5676 IdealGraphPrinter* Compile::_debug_network_printer = NULL;
5677
5678 // Called from debugger. Prints method to the default file with the default phase name.
5679 // This works regardless of any Ideal Graph Visualizer flags set or not.
5680 void igv_print() {
5681 Compile::current()->igv_print_method_to_file();
5682 }
5683
5684 // Same as igv_print() above but with a specified phase name.
5685 void igv_print(const char* phase_name) {
5686 Compile::current()->igv_print_method_to_file(phase_name);
5687 }
5688
5689 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5690 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5691 // This works regardless of any Ideal Graph Visualizer flags set or not.
5692 void igv_print(bool network) {
5693 if (network) {
5694 Compile::current()->igv_print_method_to_network();
5695 } else {
5696 Compile::current()->igv_print_method_to_file();
5697 }
5698 }
5699
5700 // Same as igv_print(bool network) above but with a specified phase name.
5701 void igv_print(bool network, const char* phase_name) {
5702 if (network) {
5703 Compile::current()->igv_print_method_to_network(phase_name);
5704 } else {
5705 Compile::current()->igv_print_method_to_file(phase_name);
5706 }
5707 }
5708
5709 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5710 void igv_print_default() {
5711 Compile::current()->print_method(PHASE_DEBUG, 0);
5712 }
5713
5714 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5715 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5716 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5717 void igv_append() {
5718 Compile::current()->igv_print_method_to_file("Debug", true);
5719 }
5720
5721 // Same as igv_append() above but with a specified phase name.
5722 void igv_append(const char* phase_name) {
5723 Compile::current()->igv_print_method_to_file(phase_name, true);
5724 }
5725
5726 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5727 const char* file_name = "custom_debug.xml";
5728 if (_debug_file_printer == NULL) {
5729 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5730 } else {
5731 _debug_file_printer->update_compiled_method(C->method());
5732 }
5733 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5734 _debug_file_printer->print(phase_name, (Node*)C->root());
5735 }
5736
5737 void Compile::igv_print_method_to_network(const char* phase_name) {
5738 if (_debug_network_printer == NULL) {
5739 _debug_network_printer = new IdealGraphPrinter(C);
5740 } else {
5741 _debug_network_printer->update_compiled_method(C->method());
5742 }
5743 tty->print_cr("Method printed over network stream to IGV");
5744 _debug_network_printer->print(phase_name, (Node*)C->root());
5745 }
5746 #endif
5747
5748 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5749 if (type != NULL && phase->type(value)->higher_equal(type)) {
5750 return value;
5751 }
5752 Node* result = NULL;
5753 if (bt == T_BYTE) {
5754 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5755 result = new RShiftINode(result, phase->intcon(24));
5756 } else if (bt == T_BOOLEAN) {
5757 result = new AndINode(value, phase->intcon(0xFF));
5758 } else if (bt == T_CHAR) {
5759 result = new AndINode(value,phase->intcon(0xFFFF));
5760 } else {
5761 assert(bt == T_SHORT, "unexpected narrow type");
5762 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5763 result = new RShiftINode(result, phase->intcon(16));
5764 }
5765 if (transform_res) {
5766 result = phase->transform(result);
5767 }
5768 return result;
5769 }
5770