1 /*
2 * Copyright (c) 1997, 2021, 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 "jvm_io.h"
27 #include "asm/macroAssembler.hpp"
28 #include "asm/macroAssembler.inline.hpp"
29 #include "ci/ciReplay.hpp"
30 #include "classfile/javaClasses.hpp"
31 #include "code/exceptionHandlerTable.hpp"
32 #include "code/nmethod.hpp"
33 #include "compiler/compileBroker.hpp"
34 #include "compiler/compileLog.hpp"
35 #include "compiler/disassembler.hpp"
36 #include "compiler/oopMap.hpp"
37 #include "gc/shared/barrierSet.hpp"
38 #include "gc/shared/c2/barrierSetC2.hpp"
39 #include "jfr/jfrEvents.hpp"
40 #include "memory/resourceArea.hpp"
41 #include "opto/addnode.hpp"
42 #include "opto/block.hpp"
43 #include "opto/c2compiler.hpp"
44 #include "opto/callGenerator.hpp"
45 #include "opto/callnode.hpp"
46 #include "opto/castnode.hpp"
47 #include "opto/cfgnode.hpp"
48 #include "opto/chaitin.hpp"
49 #include "opto/compile.hpp"
50 #include "opto/connode.hpp"
51 #include "opto/convertnode.hpp"
52 #include "opto/divnode.hpp"
53 #include "opto/escape.hpp"
54 #include "opto/idealGraphPrinter.hpp"
55 #include "opto/inlinetypenode.hpp"
56 #include "opto/loopnode.hpp"
57 #include "opto/machnode.hpp"
58 #include "opto/macro.hpp"
59 #include "opto/matcher.hpp"
60 #include "opto/mathexactnode.hpp"
61 #include "opto/memnode.hpp"
62 #include "opto/mulnode.hpp"
63 #include "opto/narrowptrnode.hpp"
64 #include "opto/node.hpp"
65 #include "opto/opcodes.hpp"
66 #include "opto/output.hpp"
67 #include "opto/parse.hpp"
68 #include "opto/phaseX.hpp"
69 #include "opto/rootnode.hpp"
70 #include "opto/runtime.hpp"
71 #include "opto/stringopts.hpp"
72 #include "opto/type.hpp"
73 #include "opto/vector.hpp"
74 #include "opto/vectornode.hpp"
75 #include "runtime/globals_extension.hpp"
76 #include "runtime/sharedRuntime.hpp"
77 #include "runtime/signature.hpp"
78 #include "runtime/stubRoutines.hpp"
79 #include "runtime/timer.hpp"
80 #include "utilities/align.hpp"
81 #include "utilities/copy.hpp"
82 #include "utilities/macros.hpp"
83 #include "utilities/resourceHash.hpp"
84
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 PhaseCCP::print_statistics();
254 PhaseRegAlloc::print_statistics();
255 PhaseOutput::print_statistics();
256 PhasePeephole::print_statistics();
257 PhaseIdealLoop::print_statistics();
258 if (xtty != NULL) xtty->tail("statistics");
259 }
260 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
261 // put this under its own <statistics> element.
262 print_intrinsic_statistics();
263 }
264 }
265 #endif //PRODUCT
266
267 void Compile::gvn_replace_by(Node* n, Node* nn) {
268 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
269 Node* use = n->last_out(i);
270 bool is_in_table = initial_gvn()->hash_delete(use);
271 uint uses_found = 0;
272 for (uint j = 0; j < use->len(); j++) {
273 if (use->in(j) == n) {
274 if (j < use->req())
275 use->set_req(j, nn);
276 else
277 use->set_prec(j, nn);
278 uses_found++;
279 }
280 }
281 if (is_in_table) {
282 // reinsert into table
283 initial_gvn()->hash_find_insert(use);
284 }
285 record_for_igvn(use);
286 i -= uses_found; // we deleted 1 or more copies of this edge
287 }
288 }
289
290
291 // Identify all nodes that are reachable from below, useful.
292 // Use breadth-first pass that records state in a Unique_Node_List,
293 // recursive traversal is slower.
294 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
295 int estimated_worklist_size = live_nodes();
296 useful.map( estimated_worklist_size, NULL ); // preallocate space
297
298 // Initialize worklist
299 if (root() != NULL) { useful.push(root()); }
300 // If 'top' is cached, declare it useful to preserve cached node
301 if( cached_top_node() ) { useful.push(cached_top_node()); }
302
303 // Push all useful nodes onto the list, breadthfirst
304 for( uint next = 0; next < useful.size(); ++next ) {
305 assert( next < unique(), "Unique useful nodes < total nodes");
306 Node *n = useful.at(next);
307 uint max = n->len();
308 for( uint i = 0; i < max; ++i ) {
309 Node *m = n->in(i);
310 if (not_a_node(m)) continue;
311 useful.push(m);
312 }
313 }
314 }
315
316 // Update dead_node_list with any missing dead nodes using useful
317 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
318 void Compile::update_dead_node_list(Unique_Node_List &useful) {
319 uint max_idx = unique();
320 VectorSet& useful_node_set = useful.member_set();
321
322 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
323 // If node with index node_idx is not in useful set,
324 // mark it as dead in dead node list.
325 if (!useful_node_set.test(node_idx)) {
326 record_dead_node(node_idx);
327 }
328 }
329 }
330
331 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
332 int shift = 0;
333 for (int i = 0; i < inlines->length(); i++) {
334 CallGenerator* cg = inlines->at(i);
335 if (useful.member(cg->call_node())) {
336 if (shift > 0) {
337 inlines->at_put(i - shift, cg);
338 }
339 } else {
340 shift++; // skip over the dead element
341 }
342 }
343 if (shift > 0) {
344 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
345 }
346 }
347
348 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
349 assert(dead != NULL && dead->is_Call(), "sanity");
350 int found = 0;
351 for (int i = 0; i < inlines->length(); i++) {
352 if (inlines->at(i)->call_node() == dead) {
353 inlines->remove_at(i);
354 found++;
355 NOT_DEBUG( break; ) // elements are unique, so exit early
356 }
357 }
358 assert(found <= 1, "not unique");
359 }
360
361 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) {
362 for (int i = node_list.length() - 1; i >= 0; i--) {
363 Node* n = node_list.at(i);
364 if (!useful.member(n)) {
365 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
366 }
367 }
368 }
369
370 void Compile::remove_useless_node(Node* dead) {
371 remove_modified_node(dead);
372
373 // Constant node that has no out-edges and has only one in-edge from
374 // root is usually dead. However, sometimes reshaping walk makes
375 // it reachable by adding use edges. So, we will NOT count Con nodes
376 // as dead to be conservative about the dead node count at any
377 // given time.
378 if (!dead->is_Con()) {
379 record_dead_node(dead->_idx);
380 }
381 if (dead->is_macro()) {
382 remove_macro_node(dead);
383 }
384 if (dead->is_expensive()) {
385 remove_expensive_node(dead);
386 }
387 if (dead->Opcode() == Op_Opaque4) {
388 remove_skeleton_predicate_opaq(dead);
389 }
390 if (dead->for_post_loop_opts_igvn()) {
391 remove_from_post_loop_opts_igvn(dead);
392 }
393 if (dead->is_InlineTypeBase()) {
394 remove_inline_type(dead);
395 }
396 if (dead->is_Call()) {
397 remove_useless_late_inlines( &_late_inlines, dead);
398 remove_useless_late_inlines( &_string_late_inlines, dead);
399 remove_useless_late_inlines( &_boxing_late_inlines, dead);
400 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
401 }
402 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
403 bs->unregister_potential_barrier_node(dead);
404 }
405
406 // Disconnect all useless nodes by disconnecting those at the boundary.
407 void Compile::disconnect_useless_nodes(Unique_Node_List &useful, Unique_Node_List* worklist) {
408 uint next = 0;
409 while (next < useful.size()) {
410 Node *n = useful.at(next++);
411 if (n->is_SafePoint()) {
412 // We're done with a parsing phase. Replaced nodes are not valid
413 // beyond that point.
414 n->as_SafePoint()->delete_replaced_nodes();
415 }
416 // Use raw traversal of out edges since this code removes out edges
417 int max = n->outcnt();
418 for (int j = 0; j < max; ++j) {
419 Node* child = n->raw_out(j);
420 if (!useful.member(child)) {
421 assert(!child->is_top() || child != top(),
422 "If top is cached in Compile object it is in useful list");
423 // Only need to remove this out-edge to the useless node
424 n->raw_del_out(j);
425 --j;
426 --max;
427 }
428 }
429 if (n->outcnt() == 1 && n->has_special_unique_user()) {
430 worklist->push(n->unique_out());
431 }
432 if (n->outcnt() == 0) {
433 worklist->push(n);
434 }
435 }
436
437 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
438 remove_useless_nodes(_predicate_opaqs, useful); // remove useless predicate opaque nodes
439 remove_useless_nodes(_skeleton_predicate_opaqs, useful);
440 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
441 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
442 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
443 #ifdef ASSERT
444 if (_modified_nodes != NULL) {
445 _modified_nodes->remove_useless_nodes(useful.member_set());
446 }
447 #endif
448 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
449
450 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
451 bs->eliminate_useless_gc_barriers(useful, this);
452 // clean up the late inline lists
453 remove_useless_late_inlines( &_late_inlines, useful);
454 remove_useless_late_inlines( &_string_late_inlines, useful);
455 remove_useless_late_inlines( &_boxing_late_inlines, useful);
456 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
457 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
458 }
459
460 // ============================================================================
461 //------------------------------CompileWrapper---------------------------------
462 class CompileWrapper : public StackObj {
463 Compile *const _compile;
464 public:
465 CompileWrapper(Compile* compile);
466
467 ~CompileWrapper();
468 };
469
470 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
471 // the Compile* pointer is stored in the current ciEnv:
472 ciEnv* env = compile->env();
473 assert(env == ciEnv::current(), "must already be a ciEnv active");
474 assert(env->compiler_data() == NULL, "compile already active?");
475 env->set_compiler_data(compile);
476 assert(compile == Compile::current(), "sanity");
477
478 compile->set_type_dict(NULL);
479 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
480 compile->clone_map().set_clone_idx(0);
481 compile->set_type_last_size(0);
482 compile->set_last_tf(NULL, NULL);
483 compile->set_indexSet_arena(NULL);
484 compile->set_indexSet_free_block_list(NULL);
485 compile->init_type_arena();
486 Type::Initialize(compile);
487 _compile->begin_method();
488 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
489 }
490 CompileWrapper::~CompileWrapper() {
491 // simulate crash during compilation
492 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
493
494 _compile->end_method();
495 _compile->env()->set_compiler_data(NULL);
496 }
497
498
499 //----------------------------print_compile_messages---------------------------
500 void Compile::print_compile_messages() {
501 #ifndef PRODUCT
502 // Check if recompiling
503 if (_subsume_loads == false && PrintOpto) {
504 // Recompiling without allowing machine instructions to subsume loads
505 tty->print_cr("*********************************************************");
506 tty->print_cr("** Bailout: Recompile without subsuming loads **");
507 tty->print_cr("*********************************************************");
508 }
509 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
510 // Recompiling without escape analysis
511 tty->print_cr("*********************************************************");
512 tty->print_cr("** Bailout: Recompile without escape analysis **");
513 tty->print_cr("*********************************************************");
514 }
515 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
516 // Recompiling without boxing elimination
517 tty->print_cr("*********************************************************");
518 tty->print_cr("** Bailout: Recompile without boxing elimination **");
519 tty->print_cr("*********************************************************");
520 }
521 if ((_do_locks_coarsening != EliminateLocks) && PrintOpto) {
522 // Recompiling without locks coarsening
523 tty->print_cr("*********************************************************");
524 tty->print_cr("** Bailout: Recompile without locks coarsening **");
525 tty->print_cr("*********************************************************");
526 }
527 if (env()->break_at_compile()) {
528 // Open the debugger when compiling this method.
529 tty->print("### Breaking when compiling: ");
530 method()->print_short_name();
531 tty->cr();
532 BREAKPOINT;
533 }
534
535 if( PrintOpto ) {
536 if (is_osr_compilation()) {
537 tty->print("[OSR]%3d", _compile_id);
538 } else {
539 tty->print("%3d", _compile_id);
540 }
541 }
542 #endif
543 }
544
545 // ============================================================================
546 //------------------------------Compile standard-------------------------------
547 debug_only( int Compile::_debug_idx = 100000; )
548
549 // Compile a method. entry_bci is -1 for normal compilations and indicates
550 // the continuation bci for on stack replacement.
551
552
553 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
554 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing,
555 bool do_locks_coarsening, bool install_code, DirectiveSet* directive)
556 : Phase(Compiler),
557 _compile_id(ci_env->compile_id()),
558 _subsume_loads(subsume_loads),
559 _do_escape_analysis(do_escape_analysis),
560 _install_code(install_code),
561 _eliminate_boxing(eliminate_boxing),
562 _do_locks_coarsening(do_locks_coarsening),
563 _method(target),
564 _entry_bci(osr_bci),
565 _stub_function(NULL),
566 _stub_name(NULL),
567 _stub_entry_point(NULL),
568 _max_node_limit(MaxNodeLimit),
569 _post_loop_opts_phase(false),
570 _inlining_progress(false),
571 _inlining_incrementally(false),
572 _do_cleanup(false),
573 _has_reserved_stack_access(target->has_reserved_stack_access()),
574 #ifndef PRODUCT
575 _igv_idx(0),
576 _trace_opto_output(directive->TraceOptoOutputOption),
577 _print_ideal(directive->PrintIdealOption),
578 #endif
579 _has_method_handle_invokes(false),
580 _clinit_barrier_on_entry(false),
581 _stress_seed(0),
582 _comp_arena(mtCompiler),
583 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
584 _env(ci_env),
585 _directive(directive),
586 _log(ci_env->log()),
587 _failure_reason(NULL),
588 _intrinsics (comp_arena(), 0, 0, NULL),
589 _macro_nodes (comp_arena(), 8, 0, NULL),
590 _predicate_opaqs (comp_arena(), 8, 0, NULL),
591 _skeleton_predicate_opaqs (comp_arena(), 8, 0, NULL),
592 _expensive_nodes (comp_arena(), 8, 0, NULL),
593 _for_post_loop_igvn(comp_arena(), 8, 0, NULL),
594 _inline_type_nodes (comp_arena(), 8, 0, NULL),
595 _coarsened_locks (comp_arena(), 8, 0, NULL),
596 _congraph(NULL),
597 NOT_PRODUCT(_printer(NULL) COMMA)
598 _dead_node_list(comp_arena()),
599 _dead_node_count(0),
600 _node_arena(mtCompiler),
601 _old_arena(mtCompiler),
602 _mach_constant_base_node(NULL),
603 _Compile_types(mtCompiler),
604 _initial_gvn(NULL),
605 _for_igvn(NULL),
606 _late_inlines(comp_arena(), 2, 0, NULL),
607 _string_late_inlines(comp_arena(), 2, 0, NULL),
608 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
609 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL),
610 _late_inlines_pos(0),
611 _number_of_mh_late_inlines(0),
612 _native_invokers(comp_arena(), 1, 0, NULL),
613 _print_inlining_stream(NULL),
614 _print_inlining_list(NULL),
615 _print_inlining_idx(0),
616 _print_inlining_output(NULL),
617 _replay_inline_data(NULL),
618 _java_calls(0),
619 _inner_loops(0),
620 _interpreter_frame_size(0)
621 #ifndef PRODUCT
622 , _in_dump_cnt(0)
623 #endif
624 {
625 C = this;
626 CompileWrapper cw(this);
627
628 if (CITimeVerbose) {
629 tty->print(" ");
630 target->holder()->name()->print();
631 tty->print(".");
632 target->print_short_name();
633 tty->print(" ");
634 }
635 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
636 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
637
638 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
639 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
640 // We can always print a disassembly, either abstract (hex dump) or
641 // with the help of a suitable hsdis library. Thus, we should not
642 // couple print_assembly and print_opto_assembly controls.
643 // But: always print opto and regular assembly on compile command 'print'.
644 bool print_assembly = directive->PrintAssemblyOption;
645 set_print_assembly(print_opto_assembly || print_assembly);
646 #else
647 set_print_assembly(false); // must initialize.
648 #endif
649
650 #ifndef PRODUCT
651 set_parsed_irreducible_loop(false);
652
653 if (directive->ReplayInlineOption) {
654 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
655 }
656 #endif
657 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
658 set_print_intrinsics(directive->PrintIntrinsicsOption);
659 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
660
661 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
662 // Make sure the method being compiled gets its own MDO,
663 // so we can at least track the decompile_count().
664 // Need MDO to record RTM code generation state.
665 method()->ensure_method_data();
666 }
667
668 Init(::AliasLevel);
669
670
671 print_compile_messages();
672
673 _ilt = InlineTree::build_inline_tree_root();
674
675 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
676 assert(num_alias_types() >= AliasIdxRaw, "");
677
678 #define MINIMUM_NODE_HASH 1023
679 // Node list that Iterative GVN will start with
680 Unique_Node_List for_igvn(comp_arena());
681 set_for_igvn(&for_igvn);
682
683 // GVN that will be run immediately on new nodes
684 uint estimated_size = method()->code_size()*4+64;
685 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
686 PhaseGVN gvn(node_arena(), estimated_size);
687 set_initial_gvn(&gvn);
688
689 print_inlining_init();
690 { // Scope for timing the parser
691 TracePhase tp("parse", &timers[_t_parser]);
692
693 // Put top into the hash table ASAP.
694 initial_gvn()->transform_no_reclaim(top());
695
696 // Set up tf(), start(), and find a CallGenerator.
697 CallGenerator* cg = NULL;
698 if (is_osr_compilation()) {
699 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
700 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
701 initial_gvn()->set_type_bottom(s);
702 init_start(s);
703 cg = CallGenerator::for_osr(method(), entry_bci());
704 } else {
705 // Normal case.
706 init_tf(TypeFunc::make(method()));
707 StartNode* s = new StartNode(root(), tf()->domain_cc());
708 initial_gvn()->set_type_bottom(s);
709 init_start(s);
710 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
711 // With java.lang.ref.reference.get() we must go through the
712 // intrinsic - even when get() is the root
713 // method of the compile - so that, if necessary, the value in
714 // the referent field of the reference object gets recorded by
715 // the pre-barrier code.
716 cg = find_intrinsic(method(), false);
717 }
718 if (cg == NULL) {
719 float past_uses = method()->interpreter_invocation_count();
720 float expected_uses = past_uses;
721 cg = CallGenerator::for_inline(method(), expected_uses);
722 }
723 }
724 if (failing()) return;
725 if (cg == NULL) {
726 record_method_not_compilable("cannot parse method");
727 return;
728 }
729 JVMState* jvms = build_start_state(start(), tf());
730 if ((jvms = cg->generate(jvms)) == NULL) {
731 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
732 record_method_not_compilable("method parse failed");
733 }
734 return;
735 }
736 GraphKit kit(jvms);
737
738 if (!kit.stopped()) {
739 // Accept return values, and transfer control we know not where.
740 // This is done by a special, unique ReturnNode bound to root.
741 return_values(kit.jvms());
742 }
743
744 if (kit.has_exceptions()) {
745 // Any exceptions that escape from this call must be rethrown
746 // to whatever caller is dynamically above us on the stack.
747 // This is done by a special, unique RethrowNode bound to root.
748 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
749 }
750
751 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
752
753 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
754 inline_string_calls(true);
755 }
756
757 if (failing()) return;
758
759 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
760
761 // Remove clutter produced by parsing.
762 if (!failing()) {
763 ResourceMark rm;
764 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
765 }
766 }
767
768 // Note: Large methods are capped off in do_one_bytecode().
769 if (failing()) return;
770
771 // After parsing, node notes are no longer automagic.
772 // They must be propagated by register_new_node_with_optimizer(),
773 // clone(), or the like.
774 set_default_node_notes(NULL);
775
776 #ifndef PRODUCT
777 if (should_print(1)) {
778 _printer->print_inlining();
779 }
780 #endif
781
782 if (failing()) return;
783 NOT_PRODUCT( verify_graph_edges(); )
784
785 // If any phase is randomized for stress testing, seed random number
786 // generation and log the seed for repeatability.
787 if (StressLCM || StressGCM || StressIGVN || StressCCP) {
788 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) {
789 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
790 FLAG_SET_ERGO(StressSeed, _stress_seed);
791 } else {
792 _stress_seed = StressSeed;
793 }
794 if (_log != NULL) {
795 _log->elem("stress_test seed='%u'", _stress_seed);
796 }
797 }
798
799 // Now optimize
800 Optimize();
801 if (failing()) return;
802 NOT_PRODUCT( verify_graph_edges(); )
803
804 #ifndef PRODUCT
805 if (print_ideal()) {
806 ttyLocker ttyl; // keep the following output all in one block
807 // This output goes directly to the tty, not the compiler log.
808 // To enable tools to match it up with the compilation activity,
809 // be sure to tag this tty output with the compile ID.
810 if (xtty != NULL) {
811 xtty->head("ideal compile_id='%d'%s", compile_id(),
812 is_osr_compilation() ? " compile_kind='osr'" :
813 "");
814 }
815 root()->dump(9999);
816 if (xtty != NULL) {
817 xtty->tail("ideal");
818 }
819 }
820 #endif
821
822 #ifdef ASSERT
823 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
824 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
825 #endif
826
827 // Dump compilation data to replay it.
828 if (directive->DumpReplayOption) {
829 env()->dump_replay_data(_compile_id);
830 }
831 if (directive->DumpInlineOption && (ilt() != NULL)) {
832 env()->dump_inline_data(_compile_id);
833 }
834
835 // Now that we know the size of all the monitors we can add a fixed slot
836 // for the original deopt pc.
837 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
838 if (needs_stack_repair()) {
839 // One extra slot for the special stack increment value
840 next_slot += 2;
841 }
842 set_fixed_slots(next_slot);
843
844 // Compute when to use implicit null checks. Used by matching trap based
845 // nodes and NullCheck optimization.
846 set_allowed_deopt_reasons();
847
848 // Now generate code
849 Code_Gen();
850 }
851
852 //------------------------------Compile----------------------------------------
853 // Compile a runtime stub
854 Compile::Compile( ciEnv* ci_env,
855 TypeFunc_generator generator,
856 address stub_function,
857 const char *stub_name,
858 int is_fancy_jump,
859 bool pass_tls,
860 bool return_pc,
861 DirectiveSet* directive)
862 : Phase(Compiler),
863 _compile_id(0),
864 _subsume_loads(true),
865 _do_escape_analysis(false),
866 _install_code(true),
867 _eliminate_boxing(false),
868 _do_locks_coarsening(false),
869 _method(NULL),
870 _entry_bci(InvocationEntryBci),
871 _stub_function(stub_function),
872 _stub_name(stub_name),
873 _stub_entry_point(NULL),
874 _max_node_limit(MaxNodeLimit),
875 _post_loop_opts_phase(false),
876 _inlining_progress(false),
877 _inlining_incrementally(false),
878 _has_reserved_stack_access(false),
879 #ifndef PRODUCT
880 _igv_idx(0),
881 _trace_opto_output(directive->TraceOptoOutputOption),
882 _print_ideal(directive->PrintIdealOption),
883 #endif
884 _has_method_handle_invokes(false),
885 _clinit_barrier_on_entry(false),
886 _stress_seed(0),
887 _comp_arena(mtCompiler),
888 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
889 _env(ci_env),
890 _directive(directive),
891 _log(ci_env->log()),
892 _failure_reason(NULL),
893 _congraph(NULL),
894 NOT_PRODUCT(_printer(NULL) COMMA)
895 _dead_node_list(comp_arena()),
896 _dead_node_count(0),
897 _node_arena(mtCompiler),
898 _old_arena(mtCompiler),
899 _mach_constant_base_node(NULL),
900 _Compile_types(mtCompiler),
901 _initial_gvn(NULL),
902 _for_igvn(NULL),
903 _number_of_mh_late_inlines(0),
904 _native_invokers(),
905 _print_inlining_stream(NULL),
906 _print_inlining_list(NULL),
907 _print_inlining_idx(0),
908 _print_inlining_output(NULL),
909 _replay_inline_data(NULL),
910 _java_calls(0),
911 _inner_loops(0),
912 _interpreter_frame_size(0),
913 #ifndef PRODUCT
914 _in_dump_cnt(0),
915 #endif
916 _allowed_reasons(0) {
917 C = this;
918
919 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
920 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
921
922 #ifndef PRODUCT
923 set_print_assembly(PrintFrameConverterAssembly);
924 set_parsed_irreducible_loop(false);
925 #else
926 set_print_assembly(false); // Must initialize.
927 #endif
928 set_has_irreducible_loop(false); // no loops
929
930 CompileWrapper cw(this);
931 Init(/*AliasLevel=*/ 0);
932 init_tf((*generator)());
933
934 {
935 // The following is a dummy for the sake of GraphKit::gen_stub
936 Unique_Node_List for_igvn(comp_arena());
937 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
938 PhaseGVN gvn(Thread::current()->resource_area(),255);
939 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
940 gvn.transform_no_reclaim(top());
941
942 GraphKit kit;
943 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
944 }
945
946 NOT_PRODUCT( verify_graph_edges(); )
947
948 Code_Gen();
949 }
950
951 //------------------------------Init-------------------------------------------
952 // Prepare for a single compilation
953 void Compile::Init(int aliaslevel) {
954 _unique = 0;
955 _regalloc = NULL;
956
957 _tf = NULL; // filled in later
958 _top = NULL; // cached later
959 _matcher = NULL; // filled in later
960 _cfg = NULL; // filled in later
961
962 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
963
964 _node_note_array = NULL;
965 _default_node_notes = NULL;
966 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
967
968 _immutable_memory = NULL; // filled in at first inquiry
969
970 // Globally visible Nodes
971 // First set TOP to NULL to give safe behavior during creation of RootNode
972 set_cached_top_node(NULL);
973 set_root(new RootNode());
974 // Now that you have a Root to point to, create the real TOP
975 set_cached_top_node( new ConNode(Type::TOP) );
976 set_recent_alloc(NULL, NULL);
977
978 // Create Debug Information Recorder to record scopes, oopmaps, etc.
979 env()->set_oop_recorder(new OopRecorder(env()->arena()));
980 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
981 env()->set_dependencies(new Dependencies(env()));
982
983 _fixed_slots = 0;
984 set_has_split_ifs(false);
985 set_has_loops(false); // first approximation
986 set_has_stringbuilder(false);
987 set_has_boxed_value(false);
988 _trap_can_recompile = false; // no traps emitted yet
989 _major_progress = true; // start out assuming good things will happen
990 set_has_unsafe_access(false);
991 set_max_vector_size(0);
992 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
993 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
994 set_decompile_count(0);
995
996 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
997 _loop_opts_cnt = LoopOptsCount;
998 _has_flattened_accesses = false;
999 _flattened_accesses_share_alias = true;
1000 _scalarize_in_safepoints = false;
1001
1002 set_do_inlining(Inline);
1003 set_max_inline_size(MaxInlineSize);
1004 set_freq_inline_size(FreqInlineSize);
1005 set_do_scheduling(OptoScheduling);
1006
1007 set_do_vector_loop(false);
1008
1009 if (AllowVectorizeOnDemand) {
1010 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1011 set_do_vector_loop(true);
1012 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());})
1013 } else if (has_method() && method()->name() != 0 &&
1014 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1015 set_do_vector_loop(true);
1016 }
1017 }
1018 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1019 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());})
1020
1021 set_age_code(has_method() && method()->profile_aging());
1022 set_rtm_state(NoRTM); // No RTM lock eliding by default
1023 _max_node_limit = _directive->MaxNodeLimitOption;
1024
1025 #if INCLUDE_RTM_OPT
1026 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1027 int rtm_state = method()->method_data()->rtm_state();
1028 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1029 // Don't generate RTM lock eliding code.
1030 set_rtm_state(NoRTM);
1031 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1032 // Generate RTM lock eliding code without abort ratio calculation code.
1033 set_rtm_state(UseRTM);
1034 } else if (UseRTMDeopt) {
1035 // Generate RTM lock eliding code and include abort ratio calculation
1036 // code if UseRTMDeopt is on.
1037 set_rtm_state(ProfileRTM);
1038 }
1039 }
1040 #endif
1041 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1042 set_clinit_barrier_on_entry(true);
1043 }
1044 if (debug_info()->recording_non_safepoints()) {
1045 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1046 (comp_arena(), 8, 0, NULL));
1047 set_default_node_notes(Node_Notes::make(this));
1048 }
1049
1050 // // -- Initialize types before each compile --
1051 // // Update cached type information
1052 // if( _method && _method->constants() )
1053 // Type::update_loaded_types(_method, _method->constants());
1054
1055 // Init alias_type map.
1056 if (!_do_escape_analysis && aliaslevel == 3)
1057 aliaslevel = 2; // No unique types without escape analysis
1058 _AliasLevel = aliaslevel;
1059 const int grow_ats = 16;
1060 _max_alias_types = grow_ats;
1061 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1062 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1063 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1064 {
1065 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1066 }
1067 // Initialize the first few types.
1068 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1069 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1070 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1071 _num_alias_types = AliasIdxRaw+1;
1072 // Zero out the alias type cache.
1073 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1074 // A NULL adr_type hits in the cache right away. Preload the right answer.
1075 probe_alias_cache(NULL)->_index = AliasIdxTop;
1076
1077 #ifdef ASSERT
1078 _type_verify_symmetry = true;
1079 _phase_optimize_finished = false;
1080 _exception_backedge = false;
1081 #endif
1082 }
1083
1084 //---------------------------init_start----------------------------------------
1085 // Install the StartNode on this compile object.
1086 void Compile::init_start(StartNode* s) {
1087 if (failing())
1088 return; // already failing
1089 assert(s == start(), "");
1090 }
1091
1092 /**
1093 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1094 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1095 * the ideal graph.
1096 */
1097 StartNode* Compile::start() const {
1098 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1099 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1100 Node* start = root()->fast_out(i);
1101 if (start->is_Start()) {
1102 return start->as_Start();
1103 }
1104 }
1105 fatal("Did not find Start node!");
1106 return NULL;
1107 }
1108
1109 //-------------------------------immutable_memory-------------------------------------
1110 // Access immutable memory
1111 Node* Compile::immutable_memory() {
1112 if (_immutable_memory != NULL) {
1113 return _immutable_memory;
1114 }
1115 StartNode* s = start();
1116 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1117 Node *p = s->fast_out(i);
1118 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1119 _immutable_memory = p;
1120 return _immutable_memory;
1121 }
1122 }
1123 ShouldNotReachHere();
1124 return NULL;
1125 }
1126
1127 //----------------------set_cached_top_node------------------------------------
1128 // Install the cached top node, and make sure Node::is_top works correctly.
1129 void Compile::set_cached_top_node(Node* tn) {
1130 if (tn != NULL) verify_top(tn);
1131 Node* old_top = _top;
1132 _top = tn;
1133 // Calling Node::setup_is_top allows the nodes the chance to adjust
1134 // their _out arrays.
1135 if (_top != NULL) _top->setup_is_top();
1136 if (old_top != NULL) old_top->setup_is_top();
1137 assert(_top == NULL || top()->is_top(), "");
1138 }
1139
1140 #ifdef ASSERT
1141 uint Compile::count_live_nodes_by_graph_walk() {
1142 Unique_Node_List useful(comp_arena());
1143 // Get useful node list by walking the graph.
1144 identify_useful_nodes(useful);
1145 return useful.size();
1146 }
1147
1148 void Compile::print_missing_nodes() {
1149
1150 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1151 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1152 return;
1153 }
1154
1155 // This is an expensive function. It is executed only when the user
1156 // specifies VerifyIdealNodeCount option or otherwise knows the
1157 // additional work that needs to be done to identify reachable nodes
1158 // by walking the flow graph and find the missing ones using
1159 // _dead_node_list.
1160
1161 Unique_Node_List useful(comp_arena());
1162 // Get useful node list by walking the graph.
1163 identify_useful_nodes(useful);
1164
1165 uint l_nodes = C->live_nodes();
1166 uint l_nodes_by_walk = useful.size();
1167
1168 if (l_nodes != l_nodes_by_walk) {
1169 if (_log != NULL) {
1170 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1171 _log->stamp();
1172 _log->end_head();
1173 }
1174 VectorSet& useful_member_set = useful.member_set();
1175 int last_idx = l_nodes_by_walk;
1176 for (int i = 0; i < last_idx; i++) {
1177 if (useful_member_set.test(i)) {
1178 if (_dead_node_list.test(i)) {
1179 if (_log != NULL) {
1180 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1181 }
1182 if (PrintIdealNodeCount) {
1183 // Print the log message to tty
1184 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1185 useful.at(i)->dump();
1186 }
1187 }
1188 }
1189 else if (! _dead_node_list.test(i)) {
1190 if (_log != NULL) {
1191 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1192 }
1193 if (PrintIdealNodeCount) {
1194 // Print the log message to tty
1195 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1196 }
1197 }
1198 }
1199 if (_log != NULL) {
1200 _log->tail("mismatched_nodes");
1201 }
1202 }
1203 }
1204 void Compile::record_modified_node(Node* n) {
1205 if (_modified_nodes != NULL && !_inlining_incrementally && !n->is_Con()) {
1206 _modified_nodes->push(n);
1207 }
1208 }
1209
1210 void Compile::remove_modified_node(Node* n) {
1211 if (_modified_nodes != NULL) {
1212 _modified_nodes->remove(n);
1213 }
1214 }
1215 #endif
1216
1217 #ifndef PRODUCT
1218 void Compile::verify_top(Node* tn) const {
1219 if (tn != NULL) {
1220 assert(tn->is_Con(), "top node must be a constant");
1221 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1222 assert(tn->in(0) != NULL, "must have live top node");
1223 }
1224 }
1225 #endif
1226
1227
1228 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1229
1230 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1231 guarantee(arr != NULL, "");
1232 int num_blocks = arr->length();
1233 if (grow_by < num_blocks) grow_by = num_blocks;
1234 int num_notes = grow_by * _node_notes_block_size;
1235 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1236 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1237 while (num_notes > 0) {
1238 arr->append(notes);
1239 notes += _node_notes_block_size;
1240 num_notes -= _node_notes_block_size;
1241 }
1242 assert(num_notes == 0, "exact multiple, please");
1243 }
1244
1245 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1246 if (source == NULL || dest == NULL) return false;
1247
1248 if (dest->is_Con())
1249 return false; // Do not push debug info onto constants.
1250
1251 #ifdef ASSERT
1252 // Leave a bread crumb trail pointing to the original node:
1253 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1254 dest->set_debug_orig(source);
1255 }
1256 #endif
1257
1258 if (node_note_array() == NULL)
1259 return false; // Not collecting any notes now.
1260
1261 // This is a copy onto a pre-existing node, which may already have notes.
1262 // If both nodes have notes, do not overwrite any pre-existing notes.
1263 Node_Notes* source_notes = node_notes_at(source->_idx);
1264 if (source_notes == NULL || source_notes->is_clear()) return false;
1265 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1266 if (dest_notes == NULL || dest_notes->is_clear()) {
1267 return set_node_notes_at(dest->_idx, source_notes);
1268 }
1269
1270 Node_Notes merged_notes = (*source_notes);
1271 // The order of operations here ensures that dest notes will win...
1272 merged_notes.update_from(dest_notes);
1273 return set_node_notes_at(dest->_idx, &merged_notes);
1274 }
1275
1276
1277 //--------------------------allow_range_check_smearing-------------------------
1278 // Gating condition for coalescing similar range checks.
1279 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1280 // single covering check that is at least as strong as any of them.
1281 // If the optimization succeeds, the simplified (strengthened) range check
1282 // will always succeed. If it fails, we will deopt, and then give up
1283 // on the optimization.
1284 bool Compile::allow_range_check_smearing() const {
1285 // If this method has already thrown a range-check,
1286 // assume it was because we already tried range smearing
1287 // and it failed.
1288 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1289 return !already_trapped;
1290 }
1291
1292
1293 //------------------------------flatten_alias_type-----------------------------
1294 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1295 int offset = tj->offset();
1296 TypePtr::PTR ptr = tj->ptr();
1297
1298 // Known instance (scalarizable allocation) alias only with itself.
1299 bool is_known_inst = tj->isa_oopptr() != NULL &&
1300 tj->is_oopptr()->is_known_instance();
1301
1302 // Process weird unsafe references.
1303 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1304 bool default_value_load = EnableValhalla && tj->is_instptr()->klass() == ciEnv::current()->Class_klass();
1305 assert(InlineUnsafeOps || default_value_load, "indeterminate pointers come only from unsafe ops");
1306 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1307 tj = TypeOopPtr::BOTTOM;
1308 ptr = tj->ptr();
1309 offset = tj->offset();
1310 }
1311
1312 // Array pointers need some flattening
1313 const TypeAryPtr *ta = tj->isa_aryptr();
1314 if (ta && ta->is_stable()) {
1315 // Erase stability property for alias analysis.
1316 tj = ta = ta->cast_to_stable(false);
1317 }
1318 if (ta && ta->is_not_flat()) {
1319 // Erase not flat property for alias analysis.
1320 tj = ta = ta->cast_to_not_flat(false);
1321 }
1322 if (ta && ta->is_not_null_free()) {
1323 // Erase not null free property for alias analysis.
1324 tj = ta = ta->cast_to_not_null_free(false);
1325 }
1326
1327 if( ta && is_known_inst ) {
1328 if ( offset != Type::OffsetBot &&
1329 offset > arrayOopDesc::length_offset_in_bytes() ) {
1330 offset = Type::OffsetBot; // Flatten constant access into array body only
1331 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id());
1332 }
1333 } else if( ta && _AliasLevel >= 2 ) {
1334 // For arrays indexed by constant indices, we flatten the alias
1335 // space to include all of the array body. Only the header, klass
1336 // and array length can be accessed un-aliased.
1337 // For flattened inline type array, each field has its own slice so
1338 // we must include the field offset.
1339 if( offset != Type::OffsetBot ) {
1340 if( ta->const_oop() ) { // MethodData* or Method*
1341 offset = Type::OffsetBot; // Flatten constant access into array body
1342 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1343 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1344 // range is OK as-is.
1345 tj = ta = TypeAryPtr::RANGE;
1346 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1347 tj = TypeInstPtr::KLASS; // all klass loads look alike
1348 ta = TypeAryPtr::RANGE; // generic ignored junk
1349 ptr = TypePtr::BotPTR;
1350 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1351 tj = TypeInstPtr::MARK;
1352 ta = TypeAryPtr::RANGE; // generic ignored junk
1353 ptr = TypePtr::BotPTR;
1354 } else { // Random constant offset into array body
1355 offset = Type::OffsetBot; // Flatten constant access into array body
1356 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1357 }
1358 }
1359 // Arrays of fixed size alias with arrays of unknown size.
1360 if (ta->size() != TypeInt::POS) {
1361 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1362 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset());
1363 }
1364 // Arrays of known objects become arrays of unknown objects.
1365 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1366 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1367 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1368 }
1369 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1370 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1371 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1372 }
1373 // Initially all flattened array accesses share a single slice
1374 if (ta->is_flat() && ta->elem() != TypeInlineType::BOTTOM && _flattened_accesses_share_alias) {
1375 const TypeAry *tary = TypeAry::make(TypeInlineType::BOTTOM, ta->size());
1376 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1377 }
1378 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1379 // cannot be distinguished by bytecode alone.
1380 if (ta->elem() == TypeInt::BOOL) {
1381 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1382 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1383 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1384 }
1385 // During the 2nd round of IterGVN, NotNull castings are removed.
1386 // Make sure the Bottom and NotNull variants alias the same.
1387 // Also, make sure exact and non-exact variants alias the same.
1388 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1389 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1390 }
1391 }
1392
1393 // Oop pointers need some flattening
1394 const TypeInstPtr *to = tj->isa_instptr();
1395 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1396 ciInstanceKlass *k = to->klass()->as_instance_klass();
1397 if( ptr == TypePtr::Constant ) {
1398 if (to->klass() != ciEnv::current()->Class_klass() ||
1399 offset < k->layout_helper_size_in_bytes()) {
1400 // No constant oop pointers (such as Strings); they alias with
1401 // unknown strings.
1402 assert(!is_known_inst, "not scalarizable allocation");
1403 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1404 }
1405 } else if( is_known_inst ) {
1406 tj = to; // Keep NotNull and klass_is_exact for instance type
1407 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1408 // During the 2nd round of IterGVN, NotNull castings are removed.
1409 // Make sure the Bottom and NotNull variants alias the same.
1410 // Also, make sure exact and non-exact variants alias the same.
1411 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1412 }
1413 if (to->speculative() != NULL) {
1414 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->klass()->flatten_array(), to->instance_id());
1415 }
1416 // Canonicalize the holder of this field
1417 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1418 // First handle header references such as a LoadKlassNode, even if the
1419 // object's klass is unloaded at compile time (4965979).
1420 if (!is_known_inst) { // Do it only for non-instance types
1421 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset));
1422 }
1423 } else if (offset < 0 || offset >= k->layout_helper_size_in_bytes()) {
1424 // Static fields are in the space above the normal instance
1425 // fields in the java.lang.Class instance.
1426 if (to->klass() != ciEnv::current()->Class_klass()) {
1427 to = NULL;
1428 tj = TypeOopPtr::BOTTOM;
1429 offset = tj->offset();
1430 }
1431 } else {
1432 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1433 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1434 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1435 if( is_known_inst ) {
1436 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), canonical_holder->flatten_array(), to->instance_id());
1437 } else {
1438 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset));
1439 }
1440 }
1441 }
1442 }
1443
1444 // Klass pointers to object array klasses need some flattening
1445 const TypeKlassPtr *tk = tj->isa_klassptr();
1446 if( tk ) {
1447 // If we are referencing a field within a Klass, we need
1448 // to assume the worst case of an Object. Both exact and
1449 // inexact types must flatten to the same alias class so
1450 // use NotNull as the PTR.
1451 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1452
1453 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1454 TypeInstKlassPtr::OBJECT->klass(),
1455 Type::Offset(offset));
1456 }
1457
1458 ciKlass* klass = tk->klass();
1459 if (klass != NULL && klass->is_obj_array_klass()) {
1460 ciKlass* k = TypeAryPtr::OOPS->klass();
1461 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1462 k = TypeInstPtr::BOTTOM->klass();
1463 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset));
1464 }
1465
1466 // Check for precise loads from the primary supertype array and force them
1467 // to the supertype cache alias index. Check for generic array loads from
1468 // the primary supertype array and also force them to the supertype cache
1469 // alias index. Since the same load can reach both, we need to merge
1470 // these 2 disparate memories into the same alias class. Since the
1471 // primary supertype array is read-only, there's no chance of confusion
1472 // where we bypass an array load and an array store.
1473 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1474 if (offset == Type::OffsetBot ||
1475 (offset >= primary_supers_offset &&
1476 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1477 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1478 offset = in_bytes(Klass::secondary_super_cache_offset());
1479 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset));
1480 }
1481 }
1482
1483 // Flatten all Raw pointers together.
1484 if (tj->base() == Type::RawPtr)
1485 tj = TypeRawPtr::BOTTOM;
1486
1487 if (tj->base() == Type::AnyPtr)
1488 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1489
1490 // Flatten all to bottom for now
1491 switch( _AliasLevel ) {
1492 case 0:
1493 tj = TypePtr::BOTTOM;
1494 break;
1495 case 1: // Flatten to: oop, static, field or array
1496 switch (tj->base()) {
1497 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1498 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1499 case Type::AryPtr: // do not distinguish arrays at all
1500 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1501 case Type::KlassPtr:
1502 case Type::AryKlassPtr:
1503 case Type::InstKlassPtr: tj = TypeInstKlassPtr::OBJECT; break;
1504 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1505 default: ShouldNotReachHere();
1506 }
1507 break;
1508 case 2: // No collapsing at level 2; keep all splits
1509 case 3: // No collapsing at level 3; keep all splits
1510 break;
1511 default:
1512 Unimplemented();
1513 }
1514
1515 offset = tj->offset();
1516 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1517
1518 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1519 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1520 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1521 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1522 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1523 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1524 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1525 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1526 assert( tj->ptr() != TypePtr::TopPTR &&
1527 tj->ptr() != TypePtr::AnyNull &&
1528 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1529 // assert( tj->ptr() != TypePtr::Constant ||
1530 // tj->base() == Type::RawPtr ||
1531 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1532
1533 return tj;
1534 }
1535
1536 void Compile::AliasType::Init(int i, const TypePtr* at) {
1537 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1538 _index = i;
1539 _adr_type = at;
1540 _field = NULL;
1541 _element = NULL;
1542 _is_rewritable = true; // default
1543 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1544 if (atoop != NULL && atoop->is_known_instance()) {
1545 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1546 _general_index = Compile::current()->get_alias_index(gt);
1547 } else {
1548 _general_index = 0;
1549 }
1550 }
1551
1552 BasicType Compile::AliasType::basic_type() const {
1553 if (element() != NULL) {
1554 const Type* element = adr_type()->is_aryptr()->elem();
1555 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1556 } if (field() != NULL) {
1557 return field()->layout_type();
1558 } else {
1559 return T_ILLEGAL; // unknown
1560 }
1561 }
1562
1563 //---------------------------------print_on------------------------------------
1564 #ifndef PRODUCT
1565 void Compile::AliasType::print_on(outputStream* st) {
1566 if (index() < 10)
1567 st->print("@ <%d> ", index());
1568 else st->print("@ <%d>", index());
1569 st->print(is_rewritable() ? " " : " RO");
1570 int offset = adr_type()->offset();
1571 if (offset == Type::OffsetBot)
1572 st->print(" +any");
1573 else st->print(" +%-3d", offset);
1574 st->print(" in ");
1575 adr_type()->dump_on(st);
1576 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1577 if (field() != NULL && tjp) {
1578 if (tjp->klass() != field()->holder() ||
1579 tjp->offset() != field()->offset_in_bytes()) {
1580 st->print(" != ");
1581 field()->print();
1582 st->print(" ***");
1583 }
1584 }
1585 }
1586
1587 void print_alias_types() {
1588 Compile* C = Compile::current();
1589 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1590 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1591 C->alias_type(idx)->print_on(tty);
1592 tty->cr();
1593 }
1594 }
1595 #endif
1596
1597
1598 //----------------------------probe_alias_cache--------------------------------
1599 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1600 intptr_t key = (intptr_t) adr_type;
1601 key ^= key >> logAliasCacheSize;
1602 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1603 }
1604
1605
1606 //-----------------------------grow_alias_types--------------------------------
1607 void Compile::grow_alias_types() {
1608 const int old_ats = _max_alias_types; // how many before?
1609 const int new_ats = old_ats; // how many more?
1610 const int grow_ats = old_ats+new_ats; // how many now?
1611 _max_alias_types = grow_ats;
1612 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1613 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1614 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1615 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1616 }
1617
1618
1619 //--------------------------------find_alias_type------------------------------
1620 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1621 if (_AliasLevel == 0)
1622 return alias_type(AliasIdxBot);
1623
1624 AliasCacheEntry* ace = NULL;
1625 if (!uncached) {
1626 ace = probe_alias_cache(adr_type);
1627 if (ace->_adr_type == adr_type) {
1628 return alias_type(ace->_index);
1629 }
1630 }
1631
1632 // Handle special cases.
1633 if (adr_type == NULL) return alias_type(AliasIdxTop);
1634 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1635
1636 // Do it the slow way.
1637 const TypePtr* flat = flatten_alias_type(adr_type);
1638
1639 #ifdef ASSERT
1640 {
1641 ResourceMark rm;
1642 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1643 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1644 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1645 Type::str(adr_type));
1646 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1647 const TypeOopPtr* foop = flat->is_oopptr();
1648 // Scalarizable allocations have exact klass always.
1649 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1650 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1651 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1652 Type::str(foop), Type::str(xoop));
1653 }
1654 }
1655 #endif
1656
1657 int idx = AliasIdxTop;
1658 for (int i = 0; i < num_alias_types(); i++) {
1659 if (alias_type(i)->adr_type() == flat) {
1660 idx = i;
1661 break;
1662 }
1663 }
1664
1665 if (idx == AliasIdxTop) {
1666 if (no_create) return NULL;
1667 // Grow the array if necessary.
1668 if (_num_alias_types == _max_alias_types) grow_alias_types();
1669 // Add a new alias type.
1670 idx = _num_alias_types++;
1671 _alias_types[idx]->Init(idx, flat);
1672 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1673 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1674 if (flat->isa_instptr()) {
1675 if (flat->offset() == java_lang_Class::klass_offset()
1676 && flat->is_instptr()->klass() == env()->Class_klass())
1677 alias_type(idx)->set_rewritable(false);
1678 }
1679 ciField* field = NULL;
1680 if (flat->isa_aryptr()) {
1681 #ifdef ASSERT
1682 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1683 // (T_BYTE has the weakest alignment and size restrictions...)
1684 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1685 #endif
1686 const Type* elemtype = flat->is_aryptr()->elem();
1687 if (flat->offset() == TypePtr::OffsetBot) {
1688 alias_type(idx)->set_element(elemtype);
1689 }
1690 int field_offset = flat->is_aryptr()->field_offset().get();
1691 if (elemtype->isa_inlinetype() &&
1692 field_offset != Type::OffsetBot) {
1693 ciInlineKlass* vk = elemtype->inline_klass();
1694 field_offset += vk->first_field_offset();
1695 field = vk->get_field_by_offset(field_offset, false);
1696 }
1697 }
1698 if (flat->isa_klassptr()) {
1699 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1700 alias_type(idx)->set_rewritable(false);
1701 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1702 alias_type(idx)->set_rewritable(false);
1703 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1704 alias_type(idx)->set_rewritable(false);
1705 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1706 alias_type(idx)->set_rewritable(false);
1707 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1708 alias_type(idx)->set_rewritable(false);
1709 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1710 alias_type(idx)->set_rewritable(false);
1711 }
1712 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1713 // but the base pointer type is not distinctive enough to identify
1714 // references into JavaThread.)
1715
1716 // Check for final fields.
1717 const TypeInstPtr* tinst = flat->isa_instptr();
1718 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1719 if (tinst->const_oop() != NULL &&
1720 tinst->klass() == ciEnv::current()->Class_klass() &&
1721 tinst->offset() >= (tinst->klass()->as_instance_klass()->layout_helper_size_in_bytes())) {
1722 // static field
1723 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1724 field = k->get_field_by_offset(tinst->offset(), true);
1725 } else if (tinst->klass()->is_inlinetype()) {
1726 // Inline type field
1727 ciInlineKlass* vk = tinst->inline_klass();
1728 field = vk->get_field_by_offset(tinst->offset(), false);
1729 } else {
1730 ciInstanceKlass* k = tinst->klass()->as_instance_klass();
1731 field = k->get_field_by_offset(tinst->offset(), false);
1732 }
1733 }
1734 assert(field == NULL ||
1735 original_field == NULL ||
1736 (field->holder() == original_field->holder() &&
1737 field->offset() == original_field->offset() &&
1738 field->is_static() == original_field->is_static()), "wrong field?");
1739 // Set field() and is_rewritable() attributes.
1740 if (field != NULL) {
1741 alias_type(idx)->set_field(field);
1742 if (flat->isa_aryptr()) {
1743 // Fields of flat arrays are rewritable although they are declared final
1744 assert(flat->is_aryptr()->is_flat(), "must be a flat array");
1745 alias_type(idx)->set_rewritable(true);
1746 }
1747 }
1748 }
1749
1750 // Fill the cache for next time.
1751 if (!uncached) {
1752 ace->_adr_type = adr_type;
1753 ace->_index = idx;
1754 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1755
1756 // Might as well try to fill the cache for the flattened version, too.
1757 AliasCacheEntry* face = probe_alias_cache(flat);
1758 if (face->_adr_type == NULL) {
1759 face->_adr_type = flat;
1760 face->_index = idx;
1761 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1762 }
1763 }
1764
1765 return alias_type(idx);
1766 }
1767
1768
1769 Compile::AliasType* Compile::alias_type(ciField* field) {
1770 const TypeOopPtr* t;
1771 if (field->is_static())
1772 t = TypeInstPtr::make(field->holder()->java_mirror());
1773 else
1774 t = TypeOopPtr::make_from_klass_raw(field->holder());
1775 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1776 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1777 return atp;
1778 }
1779
1780
1781 //------------------------------have_alias_type--------------------------------
1782 bool Compile::have_alias_type(const TypePtr* adr_type) {
1783 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1784 if (ace->_adr_type == adr_type) {
1785 return true;
1786 }
1787
1788 // Handle special cases.
1789 if (adr_type == NULL) return true;
1790 if (adr_type == TypePtr::BOTTOM) return true;
1791
1792 return find_alias_type(adr_type, true, NULL) != NULL;
1793 }
1794
1795 //-----------------------------must_alias--------------------------------------
1796 // True if all values of the given address type are in the given alias category.
1797 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1798 if (alias_idx == AliasIdxBot) return true; // the universal category
1799 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1800 if (alias_idx == AliasIdxTop) return false; // the empty category
1801 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1802
1803 // the only remaining possible overlap is identity
1804 int adr_idx = get_alias_index(adr_type);
1805 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1806 assert(adr_idx == alias_idx ||
1807 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1808 && adr_type != TypeOopPtr::BOTTOM),
1809 "should not be testing for overlap with an unsafe pointer");
1810 return adr_idx == alias_idx;
1811 }
1812
1813 //------------------------------can_alias--------------------------------------
1814 // True if any values of the given address type are in the given alias category.
1815 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1816 if (alias_idx == AliasIdxTop) return false; // the empty category
1817 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1818 // Known instance doesn't alias with bottom memory
1819 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1820 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1821
1822 // the only remaining possible overlap is identity
1823 int adr_idx = get_alias_index(adr_type);
1824 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1825 return adr_idx == alias_idx;
1826 }
1827
1828 //---------------------cleanup_loop_predicates-----------------------
1829 // Remove the opaque nodes that protect the predicates so that all unused
1830 // checks and uncommon_traps will be eliminated from the ideal graph
1831 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1832 if (predicate_count()==0) return;
1833 for (int i = predicate_count(); i > 0; i--) {
1834 Node * n = predicate_opaque1_node(i-1);
1835 assert(n->Opcode() == Op_Opaque1, "must be");
1836 igvn.replace_node(n, n->in(1));
1837 }
1838 assert(predicate_count()==0, "should be clean!");
1839 }
1840
1841 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1842 if (!n->for_post_loop_opts_igvn()) {
1843 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1844 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1845 _for_post_loop_igvn.append(n);
1846 }
1847 }
1848
1849 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1850 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1851 _for_post_loop_igvn.remove(n);
1852 }
1853
1854 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1855 // Verify that all previous optimizations produced a valid graph
1856 // at least to this point, even if no loop optimizations were done.
1857 PhaseIdealLoop::verify(igvn);
1858
1859 C->set_post_loop_opts_phase(); // no more loop opts allowed
1860
1861 assert(!C->major_progress(), "not cleared");
1862
1863 if (_for_post_loop_igvn.length() > 0) {
1864 while (_for_post_loop_igvn.length() > 0) {
1865 Node* n = _for_post_loop_igvn.pop();
1866 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1867 igvn._worklist.push(n);
1868 }
1869 igvn.optimize();
1870 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1871
1872 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1873 if (C->major_progress()) {
1874 C->clear_major_progress(); // ensure that major progress is now clear
1875 }
1876 }
1877 }
1878
1879 void Compile::add_inline_type(Node* n) {
1880 assert(n->is_InlineTypeBase(), "unexpected node");
1881 _inline_type_nodes.push(n);
1882 }
1883
1884 void Compile::remove_inline_type(Node* n) {
1885 assert(n->is_InlineTypeBase(), "unexpected node");
1886 if (_inline_type_nodes.contains(n)) {
1887 _inline_type_nodes.remove(n);
1888 }
1889 }
1890
1891 // Does the return value keep otherwise useless inline type allocations alive?
1892 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1893 ResourceMark rm;
1894 Unique_Node_List wq;
1895 wq.push(ret_val);
1896 bool some_allocations = false;
1897 for (uint i = 0; i < wq.size(); i++) {
1898 Node* n = wq.at(i);
1899 assert(!n->is_InlineType(), "chain of inline type nodes");
1900 if (n->outcnt() > 1) {
1901 // Some other use for the allocation
1902 return false;
1903 } else if (n->is_InlineTypePtr()) {
1904 wq.push(n->in(1));
1905 } else if (n->is_Phi()) {
1906 for (uint j = 1; j < n->req(); j++) {
1907 wq.push(n->in(j));
1908 }
1909 } else if (n->is_CheckCastPP() &&
1910 n->in(1)->is_Proj() &&
1911 n->in(1)->in(0)->is_Allocate()) {
1912 some_allocations = true;
1913 } else if (n->is_CheckCastPP()) {
1914 wq.push(n->in(1));
1915 }
1916 }
1917 return some_allocations;
1918 }
1919
1920 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1921 // Make sure that the return value does not keep an otherwise unused allocation alive
1922 if (tf()->returns_inline_type_as_fields()) {
1923 Node* ret = NULL;
1924 for (uint i = 1; i < root()->req(); i++) {
1925 Node* in = root()->in(i);
1926 if (in->Opcode() == Op_Return) {
1927 assert(ret == NULL, "only one return");
1928 ret = in;
1929 }
1930 }
1931 if (ret != NULL) {
1932 Node* ret_val = ret->in(TypeFunc::Parms);
1933 if (igvn.type(ret_val)->isa_oopptr() &&
1934 return_val_keeps_allocations_alive(ret_val)) {
1935 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
1936 assert(ret_val->outcnt() == 0, "should be dead now");
1937 igvn.remove_dead_node(ret_val);
1938 }
1939 }
1940 }
1941 if (_inline_type_nodes.length() == 0) {
1942 return;
1943 }
1944 // Scalarize inline types in safepoint debug info.
1945 // Delay this until all inlining is over to avoid getting inconsistent debug info.
1946 set_scalarize_in_safepoints(true);
1947 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
1948 _inline_type_nodes.at(i)->as_InlineTypeBase()->make_scalar_in_safepoints(&igvn);
1949 }
1950 if (remove) {
1951 // Remove inline type nodes
1952 while (_inline_type_nodes.length() > 0) {
1953 InlineTypeBaseNode* vt = _inline_type_nodes.pop()->as_InlineTypeBase();
1954 if (vt->outcnt() == 0) {
1955 igvn.remove_dead_node(vt);
1956 } else if (vt->is_InlineTypePtr()) {
1957 igvn.replace_node(vt, vt->get_oop());
1958 } else {
1959 // Check if any users are blackholes. If so, rewrite them to use either the
1960 // allocated buffer, or individual components, instead of the inline type node
1961 // that goes away.
1962 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
1963 if (vt->out(i)->is_Blackhole()) {
1964 BlackholeNode* bh = vt->out(i)->as_Blackhole();
1965
1966 // Unlink the old input
1967 int idx = bh->find_edge(vt);
1968 assert(idx != -1, "The edge should be there");
1969 bh->del_req(idx);
1970 --i;
1971
1972 if (vt->is_allocated(&igvn)) {
1973 // Already has the allocated instance, blackhole that
1974 bh->add_req(vt->get_oop());
1975 } else {
1976 // Not allocated yet, blackhole the components
1977 for (uint c = 0; c < vt->field_count(); c++) {
1978 bh->add_req(vt->field_value(c));
1979 }
1980 }
1981
1982 // Node modified, record for IGVN
1983 igvn.record_for_igvn(bh);
1984 }
1985 }
1986
1987 #ifdef ASSERT
1988 for (DUIterator_Fast imax, i = vt->fast_outs(imax); i < imax; i++) {
1989 assert(vt->fast_out(i)->is_InlineTypeBase(), "Unexpected inline type user");
1990 }
1991 #endif
1992 igvn.replace_node(vt, igvn.C->top());
1993 }
1994 }
1995 }
1996 igvn.optimize();
1997 }
1998
1999 void Compile::adjust_flattened_array_access_aliases(PhaseIterGVN& igvn) {
2000 if (!_has_flattened_accesses) {
2001 return;
2002 }
2003 // Initially, all flattened array accesses share the same slice to
2004 // keep dependencies with Object[] array accesses (that could be
2005 // to a flattened array) correct. We're done with parsing so we
2006 // now know all flattened array accesses in this compile
2007 // unit. Let's move flattened array accesses to their own slice,
2008 // one per element field. This should help memory access
2009 // optimizations.
2010 ResourceMark rm;
2011 Unique_Node_List wq;
2012 wq.push(root());
2013
2014 Node_List mergememnodes;
2015 Node_List memnodes;
2016
2017 // Alias index currently shared by all flattened memory accesses
2018 int index = get_alias_index(TypeAryPtr::INLINES);
2019
2020 // Find MergeMem nodes and flattened array accesses
2021 for (uint i = 0; i < wq.size(); i++) {
2022 Node* n = wq.at(i);
2023 if (n->is_Mem()) {
2024 const TypePtr* adr_type = NULL;
2025 if (n->Opcode() == Op_StoreCM) {
2026 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type()));
2027 } else {
2028 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2029 }
2030 if (adr_type == TypeAryPtr::INLINES) {
2031 memnodes.push(n);
2032 }
2033 } else if (n->is_MergeMem()) {
2034 MergeMemNode* mm = n->as_MergeMem();
2035 if (mm->memory_at(index) != mm->base_memory()) {
2036 mergememnodes.push(n);
2037 }
2038 }
2039 for (uint j = 0; j < n->req(); j++) {
2040 Node* m = n->in(j);
2041 if (m != NULL) {
2042 wq.push(m);
2043 }
2044 }
2045 }
2046
2047 if (memnodes.size() > 0) {
2048 _flattened_accesses_share_alias = false;
2049
2050 // We are going to change the slice for the flattened array
2051 // accesses so we need to clear the cache entries that refer to
2052 // them.
2053 for (uint i = 0; i < AliasCacheSize; i++) {
2054 AliasCacheEntry* ace = &_alias_cache[i];
2055 if (ace->_adr_type != NULL &&
2056 ace->_adr_type->isa_aryptr() &&
2057 ace->_adr_type->is_aryptr()->is_flat()) {
2058 ace->_adr_type = NULL;
2059 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the NULL adr_type resolves to AliasIdxTop
2060 }
2061 }
2062
2063 // Find what aliases we are going to add
2064 int start_alias = num_alias_types()-1;
2065 int stop_alias = 0;
2066
2067 for (uint i = 0; i < memnodes.size(); i++) {
2068 Node* m = memnodes.at(i);
2069 const TypePtr* adr_type = NULL;
2070 if (m->Opcode() == Op_StoreCM) {
2071 adr_type = m->in(MemNode::OopStore)->adr_type();
2072 if (adr_type != TypeAryPtr::INLINES) {
2073 // store was optimized out and we lost track of the adr_type
2074 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address),
2075 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore),
2076 get_alias_index(adr_type));
2077 igvn.register_new_node_with_optimizer(clone);
2078 igvn.replace_node(m, clone);
2079 }
2080 } else {
2081 adr_type = m->adr_type();
2082 #ifdef ASSERT
2083 m->as_Mem()->set_adr_type(adr_type);
2084 #endif
2085 }
2086 int idx = get_alias_index(adr_type);
2087 start_alias = MIN2(start_alias, idx);
2088 stop_alias = MAX2(stop_alias, idx);
2089 }
2090
2091 assert(stop_alias >= start_alias, "should have expanded aliases");
2092
2093 Node_Stack stack(0);
2094 #ifdef ASSERT
2095 VectorSet seen(Thread::current()->resource_area());
2096 #endif
2097 // Now let's fix the memory graph so each flattened array access
2098 // is moved to the right slice. Start from the MergeMem nodes.
2099 uint last = unique();
2100 for (uint i = 0; i < mergememnodes.size(); i++) {
2101 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2102 Node* n = current->memory_at(index);
2103 MergeMemNode* mm = NULL;
2104 do {
2105 // Follow memory edges through memory accesses, phis and
2106 // narrow membars and push nodes on the stack. Once we hit
2107 // bottom memory, we pop element off the stack one at a
2108 // time, in reverse order, and move them to the right slice
2109 // by changing their memory edges.
2110 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2111 assert(!seen.test_set(n->_idx), "");
2112 // Uses (a load for instance) will need to be moved to the
2113 // right slice as well and will get a new memory state
2114 // that we don't know yet. The use could also be the
2115 // backedge of a loop. We put a place holder node between
2116 // the memory node and its uses. We replace that place
2117 // holder with the correct memory state once we know it,
2118 // i.e. when nodes are popped off the stack. Using the
2119 // place holder make the logic work in the presence of
2120 // loops.
2121 if (n->outcnt() > 1) {
2122 Node* place_holder = NULL;
2123 assert(!n->has_out_with(Op_Node), "");
2124 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2125 Node* u = n->out(k);
2126 if (u != current && u->_idx < last) {
2127 bool success = false;
2128 for (uint l = 0; l < u->req(); l++) {
2129 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2130 continue;
2131 }
2132 Node* in = u->in(l);
2133 if (in == n) {
2134 if (place_holder == NULL) {
2135 place_holder = new Node(1);
2136 place_holder->init_req(0, n);
2137 }
2138 igvn.replace_input_of(u, l, place_holder);
2139 success = true;
2140 }
2141 }
2142 if (success) {
2143 --k;
2144 }
2145 }
2146 }
2147 }
2148 if (n->is_Phi()) {
2149 stack.push(n, 1);
2150 n = n->in(1);
2151 } else if (n->is_Mem()) {
2152 stack.push(n, n->req());
2153 n = n->in(MemNode::Memory);
2154 } else {
2155 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2156 stack.push(n, n->req());
2157 n = n->in(0)->in(TypeFunc::Memory);
2158 }
2159 } else {
2160 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2161 // Build a new MergeMem node to carry the new memory state
2162 // as we build it. IGVN should fold extraneous MergeMem
2163 // nodes.
2164 mm = MergeMemNode::make(n);
2165 igvn.register_new_node_with_optimizer(mm);
2166 while (stack.size() > 0) {
2167 Node* m = stack.node();
2168 uint idx = stack.index();
2169 if (m->is_Mem()) {
2170 // Move memory node to its new slice
2171 const TypePtr* adr_type = m->adr_type();
2172 int alias = get_alias_index(adr_type);
2173 Node* prev = mm->memory_at(alias);
2174 igvn.replace_input_of(m, MemNode::Memory, prev);
2175 mm->set_memory_at(alias, m);
2176 } else if (m->is_Phi()) {
2177 // We need as many new phis as there are new aliases
2178 igvn.replace_input_of(m, idx, mm);
2179 if (idx == m->req()-1) {
2180 Node* r = m->in(0);
2181 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2182 const Type* adr_type = get_adr_type(j);
2183 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->is_flat() || j == (uint)index) {
2184 continue;
2185 }
2186 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2187 igvn.register_new_node_with_optimizer(phi);
2188 for (uint k = 1; k < m->req(); k++) {
2189 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2190 }
2191 mm->set_memory_at(j, phi);
2192 }
2193 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2194 igvn.register_new_node_with_optimizer(base_phi);
2195 for (uint k = 1; k < m->req(); k++) {
2196 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2197 }
2198 mm->set_base_memory(base_phi);
2199 }
2200 } else {
2201 // This is a MemBarCPUOrder node from
2202 // Parse::array_load()/Parse::array_store(), in the
2203 // branch that handles flattened arrays hidden under
2204 // an Object[] array. We also need one new membar per
2205 // new alias to keep the unknown access that the
2206 // membars protect properly ordered with accesses to
2207 // known flattened array.
2208 assert(m->is_Proj(), "projection expected");
2209 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2210 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2211 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2212 const Type* adr_type = get_adr_type(j);
2213 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->is_flat() || j == (uint)index) {
2214 continue;
2215 }
2216 MemBarNode* mb = new MemBarCPUOrderNode(this, j, NULL);
2217 igvn.register_new_node_with_optimizer(mb);
2218 Node* mem = mm->memory_at(j);
2219 mb->init_req(TypeFunc::Control, ctrl);
2220 mb->init_req(TypeFunc::Memory, mem);
2221 ctrl = new ProjNode(mb, TypeFunc::Control);
2222 igvn.register_new_node_with_optimizer(ctrl);
2223 mem = new ProjNode(mb, TypeFunc::Memory);
2224 igvn.register_new_node_with_optimizer(mem);
2225 mm->set_memory_at(j, mem);
2226 }
2227 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2228 }
2229 if (idx < m->req()-1) {
2230 idx += 1;
2231 stack.set_index(idx);
2232 n = m->in(idx);
2233 break;
2234 }
2235 // Take care of place holder nodes
2236 if (m->has_out_with(Op_Node)) {
2237 Node* place_holder = m->find_out_with(Op_Node);
2238 if (place_holder != NULL) {
2239 Node* mm_clone = mm->clone();
2240 igvn.register_new_node_with_optimizer(mm_clone);
2241 Node* hook = new Node(1);
2242 hook->init_req(0, mm);
2243 igvn.replace_node(place_holder, mm_clone);
2244 hook->destruct(&igvn);
2245 }
2246 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2247 }
2248 stack.pop();
2249 }
2250 }
2251 } while(stack.size() > 0);
2252 // Fix the memory state at the MergeMem we started from
2253 igvn.rehash_node_delayed(current);
2254 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2255 const Type* adr_type = get_adr_type(j);
2256 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->is_flat()) {
2257 continue;
2258 }
2259 current->set_memory_at(j, mm);
2260 }
2261 current->set_memory_at(index, current->base_memory());
2262 }
2263 igvn.optimize();
2264 }
2265 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2266 #ifdef ASSERT
2267 if (!_flattened_accesses_share_alias) {
2268 wq.clear();
2269 wq.push(root());
2270 for (uint i = 0; i < wq.size(); i++) {
2271 Node* n = wq.at(i);
2272 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2273 for (uint j = 0; j < n->req(); j++) {
2274 Node* m = n->in(j);
2275 if (m != NULL) {
2276 wq.push(m);
2277 }
2278 }
2279 }
2280 }
2281 #endif
2282 }
2283
2284
2285 // StringOpts and late inlining of string methods
2286 void Compile::inline_string_calls(bool parse_time) {
2287 {
2288 // remove useless nodes to make the usage analysis simpler
2289 ResourceMark rm;
2290 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2291 }
2292
2293 {
2294 ResourceMark rm;
2295 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2296 PhaseStringOpts pso(initial_gvn(), for_igvn());
2297 print_method(PHASE_AFTER_STRINGOPTS, 3);
2298 }
2299
2300 // now inline anything that we skipped the first time around
2301 if (!parse_time) {
2302 _late_inlines_pos = _late_inlines.length();
2303 }
2304
2305 while (_string_late_inlines.length() > 0) {
2306 CallGenerator* cg = _string_late_inlines.pop();
2307 cg->do_late_inline();
2308 if (failing()) return;
2309 }
2310 _string_late_inlines.trunc_to(0);
2311 }
2312
2313 // Late inlining of boxing methods
2314 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2315 if (_boxing_late_inlines.length() > 0) {
2316 assert(has_boxed_value(), "inconsistent");
2317
2318 PhaseGVN* gvn = initial_gvn();
2319 set_inlining_incrementally(true);
2320
2321 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2322 for_igvn()->clear();
2323 gvn->replace_with(&igvn);
2324
2325 _late_inlines_pos = _late_inlines.length();
2326
2327 while (_boxing_late_inlines.length() > 0) {
2328 CallGenerator* cg = _boxing_late_inlines.pop();
2329 cg->do_late_inline();
2330 if (failing()) return;
2331 }
2332 _boxing_late_inlines.trunc_to(0);
2333
2334 inline_incrementally_cleanup(igvn);
2335
2336 set_inlining_incrementally(false);
2337 }
2338 }
2339
2340 bool Compile::inline_incrementally_one() {
2341 assert(IncrementalInline, "incremental inlining should be on");
2342
2343 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2344
2345 set_inlining_progress(false);
2346 set_do_cleanup(false);
2347
2348 for (int i = 0; i < _late_inlines.length(); i++) {
2349 _late_inlines_pos = i+1;
2350 CallGenerator* cg = _late_inlines.at(i);
2351 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2352 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2353 cg->do_late_inline();
2354 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2355 if (failing()) {
2356 return false;
2357 } else if (inlining_progress()) {
2358 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2359 print_method(PHASE_INCREMENTAL_INLINE_STEP, cg->call_node(), 3);
2360 break; // process one call site at a time
2361 }
2362 } else {
2363 // Ignore late inline direct calls when inlining is not allowed.
2364 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2365 }
2366 }
2367 // Remove processed elements.
2368 _late_inlines.remove_till(_late_inlines_pos);
2369 _late_inlines_pos = 0;
2370
2371 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2372
2373 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2374
2375 set_inlining_progress(false);
2376 set_do_cleanup(false);
2377
2378 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2379 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2380 }
2381
2382 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2383 {
2384 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2385 ResourceMark rm;
2386 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2387 }
2388 {
2389 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2390 igvn = PhaseIterGVN(initial_gvn());
2391 igvn.optimize();
2392 }
2393 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2394 }
2395
2396 // Perform incremental inlining until bound on number of live nodes is reached
2397 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2398 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2399
2400 set_inlining_incrementally(true);
2401 uint low_live_nodes = 0;
2402
2403 while (_late_inlines.length() > 0) {
2404 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2405 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2406 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2407 // PhaseIdealLoop is expensive so we only try it once we are
2408 // out of live nodes and we only try it again if the previous
2409 // helped got the number of nodes down significantly
2410 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2411 if (failing()) return;
2412 low_live_nodes = live_nodes();
2413 _major_progress = true;
2414 }
2415
2416 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2417 bool do_print_inlining = print_inlining() || print_intrinsics();
2418 if (do_print_inlining || log() != NULL) {
2419 // Print inlining message for candidates that we couldn't inline for lack of space.
2420 for (int i = 0; i < _late_inlines.length(); i++) {
2421 CallGenerator* cg = _late_inlines.at(i);
2422 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2423 if (do_print_inlining) {
2424 cg->print_inlining_late(msg);
2425 }
2426 log_late_inline_failure(cg, msg);
2427 }
2428 }
2429 break; // finish
2430 }
2431 }
2432
2433 for_igvn()->clear();
2434 initial_gvn()->replace_with(&igvn);
2435
2436 while (inline_incrementally_one()) {
2437 assert(!failing(), "inconsistent");
2438 }
2439 if (failing()) return;
2440
2441 inline_incrementally_cleanup(igvn);
2442
2443 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2444
2445 if (failing()) return;
2446
2447 if (_late_inlines.length() == 0) {
2448 break; // no more progress
2449 }
2450 }
2451 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2452
2453 if (_string_late_inlines.length() > 0) {
2454 assert(has_stringbuilder(), "inconsistent");
2455 for_igvn()->clear();
2456 initial_gvn()->replace_with(&igvn);
2457
2458 inline_string_calls(false);
2459
2460 if (failing()) return;
2461
2462 inline_incrementally_cleanup(igvn);
2463 }
2464
2465 set_inlining_incrementally(false);
2466 }
2467
2468 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2469 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2470 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2471 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == NULL"
2472 // as if "inlining_incrementally() == true" were set.
2473 assert(inlining_incrementally() == false, "not allowed");
2474 #ifdef ASSERT
2475 Unique_Node_List* modified_nodes = _modified_nodes;
2476 _modified_nodes = NULL;
2477 #endif
2478 assert(_late_inlines.length() > 0, "sanity");
2479
2480 while (_late_inlines.length() > 0) {
2481 for_igvn()->clear();
2482 initial_gvn()->replace_with(&igvn);
2483
2484 while (inline_incrementally_one()) {
2485 assert(!failing(), "inconsistent");
2486 }
2487 if (failing()) return;
2488
2489 inline_incrementally_cleanup(igvn);
2490 }
2491 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2492 }
2493
2494 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2495 if (_loop_opts_cnt > 0) {
2496 debug_only( int cnt = 0; );
2497 while (major_progress() && (_loop_opts_cnt > 0)) {
2498 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2499 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2500 PhaseIdealLoop::optimize(igvn, mode);
2501 _loop_opts_cnt--;
2502 if (failing()) return false;
2503 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2504 }
2505 }
2506 return true;
2507 }
2508
2509 // Remove edges from "root" to each SafePoint at a backward branch.
2510 // They were inserted during parsing (see add_safepoint()) to make
2511 // infinite loops without calls or exceptions visible to root, i.e.,
2512 // useful.
2513 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2514 Node *r = root();
2515 if (r != NULL) {
2516 for (uint i = r->req(); i < r->len(); ++i) {
2517 Node *n = r->in(i);
2518 if (n != NULL && n->is_SafePoint()) {
2519 r->rm_prec(i);
2520 if (n->outcnt() == 0) {
2521 igvn.remove_dead_node(n);
2522 }
2523 --i;
2524 }
2525 }
2526 // Parsing may have added top inputs to the root node (Path
2527 // leading to the Halt node proven dead). Make sure we get a
2528 // chance to clean them up.
2529 igvn._worklist.push(r);
2530 igvn.optimize();
2531 }
2532 }
2533
2534 //------------------------------Optimize---------------------------------------
2535 // Given a graph, optimize it.
2536 void Compile::Optimize() {
2537 TracePhase tp("optimizer", &timers[_t_optimizer]);
2538
2539 #ifndef PRODUCT
2540 if (env()->break_at_compile()) {
2541 BREAKPOINT;
2542 }
2543
2544 #endif
2545
2546 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2547 #ifdef ASSERT
2548 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2549 #endif
2550
2551 ResourceMark rm;
2552
2553 print_inlining_reinit();
2554
2555 NOT_PRODUCT( verify_graph_edges(); )
2556
2557 print_method(PHASE_AFTER_PARSING);
2558
2559 {
2560 // Iterative Global Value Numbering, including ideal transforms
2561 // Initialize IterGVN with types and values from parse-time GVN
2562 PhaseIterGVN igvn(initial_gvn());
2563 #ifdef ASSERT
2564 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2565 #endif
2566 {
2567 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2568 igvn.optimize();
2569 }
2570
2571 if (failing()) return;
2572
2573 print_method(PHASE_ITER_GVN1, 2);
2574
2575 inline_incrementally(igvn);
2576
2577 print_method(PHASE_INCREMENTAL_INLINE, 2);
2578
2579 if (failing()) return;
2580
2581 if (eliminate_boxing()) {
2582 // Inline valueOf() methods now.
2583 inline_boxing_calls(igvn);
2584
2585 if (AlwaysIncrementalInline) {
2586 inline_incrementally(igvn);
2587 }
2588
2589 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2590
2591 if (failing()) return;
2592 }
2593
2594 // Remove the speculative part of types and clean up the graph from
2595 // the extra CastPP nodes whose only purpose is to carry them. Do
2596 // that early so that optimizations are not disrupted by the extra
2597 // CastPP nodes.
2598 remove_speculative_types(igvn);
2599
2600 // No more new expensive nodes will be added to the list from here
2601 // so keep only the actual candidates for optimizations.
2602 cleanup_expensive_nodes(igvn);
2603
2604 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2605 if (EnableVectorSupport && has_vbox_nodes()) {
2606 TracePhase tp("", &timers[_t_vector]);
2607 PhaseVector pv(igvn);
2608 pv.optimize_vector_boxes();
2609
2610 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2611 }
2612 assert(!has_vbox_nodes(), "sanity");
2613
2614 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2615 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2616 initial_gvn()->replace_with(&igvn);
2617 for_igvn()->clear();
2618 Unique_Node_List new_worklist(C->comp_arena());
2619 {
2620 ResourceMark rm;
2621 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2622 }
2623 Unique_Node_List* save_for_igvn = for_igvn();
2624 set_for_igvn(&new_worklist);
2625 igvn = PhaseIterGVN(initial_gvn());
2626 igvn.optimize();
2627 set_for_igvn(save_for_igvn);
2628 }
2629
2630 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2631 // safepoints
2632 remove_root_to_sfpts_edges(igvn);
2633
2634 // Process inline type nodes now that all inlining is over
2635 process_inline_types(igvn);
2636
2637 adjust_flattened_array_access_aliases(igvn);
2638
2639 // Perform escape analysis
2640 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2641 if (has_loops()) {
2642 // Cleanup graph (remove dead nodes).
2643 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2644 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2645 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2646 if (failing()) return;
2647 }
2648 ConnectionGraph::do_analysis(this, &igvn);
2649
2650 if (failing()) return;
2651
2652 // Optimize out fields loads from scalar replaceable allocations.
2653 igvn.optimize();
2654 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2655
2656 if (failing()) return;
2657
2658 if (congraph() != NULL && macro_count() > 0) {
2659 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2660 PhaseMacroExpand mexp(igvn);
2661 mexp.eliminate_macro_nodes();
2662 igvn.set_delay_transform(false);
2663
2664 igvn.optimize();
2665 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2666
2667 if (failing()) return;
2668 }
2669 }
2670
2671 // Loop transforms on the ideal graph. Range Check Elimination,
2672 // peeling, unrolling, etc.
2673
2674 // Set loop opts counter
2675 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2676 {
2677 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2678 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2679 _loop_opts_cnt--;
2680 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2681 if (failing()) return;
2682 }
2683 // Loop opts pass if partial peeling occurred in previous pass
2684 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2685 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2686 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2687 _loop_opts_cnt--;
2688 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2689 if (failing()) return;
2690 }
2691 // Loop opts pass for loop-unrolling before CCP
2692 if(major_progress() && (_loop_opts_cnt > 0)) {
2693 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2694 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2695 _loop_opts_cnt--;
2696 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2697 }
2698 if (!failing()) {
2699 // Verify that last round of loop opts produced a valid graph
2700 PhaseIdealLoop::verify(igvn);
2701 }
2702 }
2703 if (failing()) return;
2704
2705 // Conditional Constant Propagation;
2706 PhaseCCP ccp( &igvn );
2707 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2708 {
2709 TracePhase tp("ccp", &timers[_t_ccp]);
2710 ccp.do_transform();
2711 }
2712 print_method(PHASE_CCP1, 2);
2713
2714 assert( true, "Break here to ccp.dump_old2new_map()");
2715
2716 // Iterative Global Value Numbering, including ideal transforms
2717 {
2718 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2719 igvn = ccp;
2720 igvn.optimize();
2721 }
2722 print_method(PHASE_ITER_GVN2, 2);
2723
2724 if (failing()) return;
2725
2726 // Loop transforms on the ideal graph. Range Check Elimination,
2727 // peeling, unrolling, etc.
2728 if (!optimize_loops(igvn, LoopOptsDefault)) {
2729 return;
2730 }
2731
2732 if (failing()) return;
2733
2734 C->clear_major_progress(); // ensure that major progress is now clear
2735
2736 process_for_post_loop_opts_igvn(igvn);
2737
2738 #ifdef ASSERT
2739 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2740 #endif
2741
2742 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2743
2744 if (_late_inlines.length() > 0) {
2745 // More opportunities to optimize virtual and MH calls.
2746 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2747 process_late_inline_calls_no_inline(igvn);
2748 }
2749
2750 {
2751 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2752 PhaseMacroExpand mex(igvn);
2753 if (mex.expand_macro_nodes()) {
2754 assert(failing(), "must bail out w/ explicit message");
2755 return;
2756 }
2757 print_method(PHASE_MACRO_EXPANSION, 2);
2758 }
2759
2760 // Process inline type nodes again and remove them. From here
2761 // on we don't need to keep track of field values anymore.
2762 process_inline_types(igvn, /* remove= */ true);
2763
2764 {
2765 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2766 if (bs->expand_barriers(this, igvn)) {
2767 assert(failing(), "must bail out w/ explicit message");
2768 return;
2769 }
2770 print_method(PHASE_BARRIER_EXPANSION, 2);
2771 }
2772
2773 if (C->max_vector_size() > 0) {
2774 C->optimize_logic_cones(igvn);
2775 igvn.optimize();
2776 }
2777
2778 DEBUG_ONLY( _modified_nodes = NULL; )
2779
2780 assert(igvn._worklist.size() == 0, "not empty");
2781 assert(_late_inlines.length() == 0, "missed optimization opportunity");
2782 } // (End scope of igvn; run destructor if necessary for asserts.)
2783
2784 check_no_dead_use();
2785
2786 process_print_inlining();
2787
2788 // A method with only infinite loops has no edges entering loops from root
2789 {
2790 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2791 if (final_graph_reshaping()) {
2792 assert(failing(), "must bail out w/ explicit message");
2793 return;
2794 }
2795 }
2796
2797 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2798 DEBUG_ONLY(set_phase_optimize_finished();)
2799 }
2800
2801 #ifdef ASSERT
2802 void Compile::check_no_dead_use() const {
2803 ResourceMark rm;
2804 Unique_Node_List wq;
2805 wq.push(root());
2806 for (uint i = 0; i < wq.size(); ++i) {
2807 Node* n = wq.at(i);
2808 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2809 Node* u = n->fast_out(j);
2810 if (u->outcnt() == 0 && !u->is_Con()) {
2811 u->dump();
2812 fatal("no reachable node should have no use");
2813 }
2814 wq.push(u);
2815 }
2816 }
2817 }
2818 #endif
2819
2820 void Compile::inline_vector_reboxing_calls() {
2821 if (C->_vector_reboxing_late_inlines.length() > 0) {
2822 _late_inlines_pos = C->_late_inlines.length();
2823 while (_vector_reboxing_late_inlines.length() > 0) {
2824 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2825 cg->do_late_inline();
2826 if (failing()) return;
2827 print_method(PHASE_INLINE_VECTOR_REBOX, cg->call_node());
2828 }
2829 _vector_reboxing_late_inlines.trunc_to(0);
2830 }
2831 }
2832
2833 bool Compile::has_vbox_nodes() {
2834 if (C->_vector_reboxing_late_inlines.length() > 0) {
2835 return true;
2836 }
2837 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2838 Node * n = C->macro_node(macro_idx);
2839 assert(n->is_macro(), "only macro nodes expected here");
2840 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2841 return true;
2842 }
2843 }
2844 return false;
2845 }
2846
2847 //---------------------------- Bitwise operation packing optimization ---------------------------
2848
2849 static bool is_vector_unary_bitwise_op(Node* n) {
2850 return n->Opcode() == Op_XorV &&
2851 VectorNode::is_vector_bitwise_not_pattern(n);
2852 }
2853
2854 static bool is_vector_binary_bitwise_op(Node* n) {
2855 switch (n->Opcode()) {
2856 case Op_AndV:
2857 case Op_OrV:
2858 return true;
2859
2860 case Op_XorV:
2861 return !is_vector_unary_bitwise_op(n);
2862
2863 default:
2864 return false;
2865 }
2866 }
2867
2868 static bool is_vector_ternary_bitwise_op(Node* n) {
2869 return n->Opcode() == Op_MacroLogicV;
2870 }
2871
2872 static bool is_vector_bitwise_op(Node* n) {
2873 return is_vector_unary_bitwise_op(n) ||
2874 is_vector_binary_bitwise_op(n) ||
2875 is_vector_ternary_bitwise_op(n);
2876 }
2877
2878 static bool is_vector_bitwise_cone_root(Node* n) {
2879 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2880 return false;
2881 }
2882 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2883 if (is_vector_bitwise_op(n->fast_out(i))) {
2884 return false;
2885 }
2886 }
2887 return true;
2888 }
2889
2890 static uint collect_unique_inputs(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2891 uint cnt = 0;
2892 if (is_vector_bitwise_op(n)) {
2893 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2894 for (uint i = 1; i < n->req(); i++) {
2895 Node* in = n->in(i);
2896 bool skip = VectorNode::is_all_ones_vector(in);
2897 if (!skip && !inputs.member(in)) {
2898 inputs.push(in);
2899 cnt++;
2900 }
2901 }
2902 assert(cnt <= 1, "not unary");
2903 } else {
2904 uint last_req = n->req();
2905 if (is_vector_ternary_bitwise_op(n)) {
2906 last_req = n->req() - 1; // skip last input
2907 }
2908 for (uint i = 1; i < last_req; i++) {
2909 Node* def = n->in(i);
2910 if (!inputs.member(def)) {
2911 inputs.push(def);
2912 cnt++;
2913 }
2914 }
2915 }
2916 partition.push(n);
2917 } else { // not a bitwise operations
2918 if (!inputs.member(n)) {
2919 inputs.push(n);
2920 cnt++;
2921 }
2922 }
2923 return cnt;
2924 }
2925
2926 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2927 Unique_Node_List useful_nodes;
2928 C->identify_useful_nodes(useful_nodes);
2929
2930 for (uint i = 0; i < useful_nodes.size(); i++) {
2931 Node* n = useful_nodes.at(i);
2932 if (is_vector_bitwise_cone_root(n)) {
2933 list.push(n);
2934 }
2935 }
2936 }
2937
2938 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2939 const TypeVect* vt,
2940 Unique_Node_List& partition,
2941 Unique_Node_List& inputs) {
2942 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2943 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2944 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2945
2946 Node* in1 = inputs.at(0);
2947 Node* in2 = inputs.at(1);
2948 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2949
2950 uint func = compute_truth_table(partition, inputs);
2951 return igvn.transform(MacroLogicVNode::make(igvn, in3, in2, in1, func, vt));
2952 }
2953
2954 static uint extract_bit(uint func, uint pos) {
2955 return (func & (1 << pos)) >> pos;
2956 }
2957
2958 //
2959 // A macro logic node represents a truth table. It has 4 inputs,
2960 // First three inputs corresponds to 3 columns of a truth table
2961 // and fourth input captures the logic function.
2962 //
2963 // eg. fn = (in1 AND in2) OR in3;
2964 //
2965 // MacroNode(in1,in2,in3,fn)
2966 //
2967 // -----------------
2968 // in1 in2 in3 fn
2969 // -----------------
2970 // 0 0 0 0
2971 // 0 0 1 1
2972 // 0 1 0 0
2973 // 0 1 1 1
2974 // 1 0 0 0
2975 // 1 0 1 1
2976 // 1 1 0 1
2977 // 1 1 1 1
2978 //
2979
2980 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2981 int res = 0;
2982 for (int i = 0; i < 8; i++) {
2983 int bit1 = extract_bit(in1, i);
2984 int bit2 = extract_bit(in2, i);
2985 int bit3 = extract_bit(in3, i);
2986
2987 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2988 int func_bit = extract_bit(func, func_bit_pos);
2989
2990 res |= func_bit << i;
2991 }
2992 return res;
2993 }
2994
2995 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2996 assert(n != NULL, "");
2997 assert(eval_map.contains(n), "absent");
2998 return *(eval_map.get(n));
2999 }
3000
3001 static void eval_operands(Node* n,
3002 uint& func1, uint& func2, uint& func3,
3003 ResourceHashtable<Node*,uint>& eval_map) {
3004 assert(is_vector_bitwise_op(n), "");
3005
3006 if (is_vector_unary_bitwise_op(n)) {
3007 Node* opnd = n->in(1);
3008 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3009 opnd = n->in(2);
3010 }
3011 func1 = eval_operand(opnd, eval_map);
3012 } else if (is_vector_binary_bitwise_op(n)) {
3013 func1 = eval_operand(n->in(1), eval_map);
3014 func2 = eval_operand(n->in(2), eval_map);
3015 } else {
3016 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3017 func1 = eval_operand(n->in(1), eval_map);
3018 func2 = eval_operand(n->in(2), eval_map);
3019 func3 = eval_operand(n->in(3), eval_map);
3020 }
3021 }
3022
3023 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3024 assert(inputs.size() <= 3, "sanity");
3025 ResourceMark rm;
3026 uint res = 0;
3027 ResourceHashtable<Node*,uint> eval_map;
3028
3029 // Populate precomputed functions for inputs.
3030 // Each input corresponds to one column of 3 input truth-table.
3031 uint input_funcs[] = { 0xAA, // (_, _, a) -> a
3032 0xCC, // (_, b, _) -> b
3033 0xF0 }; // (c, _, _) -> c
3034 for (uint i = 0; i < inputs.size(); i++) {
3035 eval_map.put(inputs.at(i), input_funcs[i]);
3036 }
3037
3038 for (uint i = 0; i < partition.size(); i++) {
3039 Node* n = partition.at(i);
3040
3041 uint func1 = 0, func2 = 0, func3 = 0;
3042 eval_operands(n, func1, func2, func3, eval_map);
3043
3044 switch (n->Opcode()) {
3045 case Op_OrV:
3046 assert(func3 == 0, "not binary");
3047 res = func1 | func2;
3048 break;
3049 case Op_AndV:
3050 assert(func3 == 0, "not binary");
3051 res = func1 & func2;
3052 break;
3053 case Op_XorV:
3054 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3055 assert(func2 == 0 && func3 == 0, "not unary");
3056 res = (~func1) & 0xFF;
3057 } else {
3058 assert(func3 == 0, "not binary");
3059 res = func1 ^ func2;
3060 }
3061 break;
3062 case Op_MacroLogicV:
3063 // Ordering of inputs may change during evaluation of sub-tree
3064 // containing MacroLogic node as a child node, thus a re-evaluation
3065 // makes sure that function is evaluated in context of current
3066 // inputs.
3067 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3068 break;
3069
3070 default: assert(false, "not supported: %s", n->Name());
3071 }
3072 assert(res <= 0xFF, "invalid");
3073 eval_map.put(n, res);
3074 }
3075 return res;
3076 }
3077
3078 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3079 assert(partition.size() == 0, "not empty");
3080 assert(inputs.size() == 0, "not empty");
3081 if (is_vector_ternary_bitwise_op(n)) {
3082 return false;
3083 }
3084
3085 bool is_unary_op = is_vector_unary_bitwise_op(n);
3086 if (is_unary_op) {
3087 assert(collect_unique_inputs(n, partition, inputs) == 1, "not unary");
3088 return false; // too few inputs
3089 }
3090
3091 assert(is_vector_binary_bitwise_op(n), "not binary");
3092 Node* in1 = n->in(1);
3093 Node* in2 = n->in(2);
3094
3095 int in1_unique_inputs_cnt = collect_unique_inputs(in1, partition, inputs);
3096 int in2_unique_inputs_cnt = collect_unique_inputs(in2, partition, inputs);
3097 partition.push(n);
3098
3099 // Too many inputs?
3100 if (inputs.size() > 3) {
3101 partition.clear();
3102 inputs.clear();
3103 { // Recompute in2 inputs
3104 Unique_Node_List not_used;
3105 in2_unique_inputs_cnt = collect_unique_inputs(in2, not_used, not_used);
3106 }
3107 // Pick the node with minimum number of inputs.
3108 if (in1_unique_inputs_cnt >= 3 && in2_unique_inputs_cnt >= 3) {
3109 return false; // still too many inputs
3110 }
3111 // Recompute partition & inputs.
3112 Node* child = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in1 : in2);
3113 collect_unique_inputs(child, partition, inputs);
3114
3115 Node* other_input = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in2 : in1);
3116 inputs.push(other_input);
3117
3118 partition.push(n);
3119 }
3120
3121 return (partition.size() == 2 || partition.size() == 3) &&
3122 (inputs.size() == 2 || inputs.size() == 3);
3123 }
3124
3125
3126 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3127 assert(is_vector_bitwise_op(n), "not a root");
3128
3129 visited.set(n->_idx);
3130
3131 // 1) Do a DFS walk over the logic cone.
3132 for (uint i = 1; i < n->req(); i++) {
3133 Node* in = n->in(i);
3134 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3135 process_logic_cone_root(igvn, in, visited);
3136 }
3137 }
3138
3139 // 2) Bottom up traversal: Merge node[s] with
3140 // the parent to form macro logic node.
3141 Unique_Node_List partition;
3142 Unique_Node_List inputs;
3143 if (compute_logic_cone(n, partition, inputs)) {
3144 const TypeVect* vt = n->bottom_type()->is_vect();
3145 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3146 igvn.replace_node(n, macro_logic);
3147 }
3148 }
3149
3150 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3151 ResourceMark rm;
3152 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3153 Unique_Node_List list;
3154 collect_logic_cone_roots(list);
3155
3156 while (list.size() > 0) {
3157 Node* n = list.pop();
3158 const TypeVect* vt = n->bottom_type()->is_vect();
3159 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3160 if (supported) {
3161 VectorSet visited(comp_arena());
3162 process_logic_cone_root(igvn, n, visited);
3163 }
3164 }
3165 }
3166 }
3167
3168 //------------------------------Code_Gen---------------------------------------
3169 // Given a graph, generate code for it
3170 void Compile::Code_Gen() {
3171 if (failing()) {
3172 return;
3173 }
3174
3175 // Perform instruction selection. You might think we could reclaim Matcher
3176 // memory PDQ, but actually the Matcher is used in generating spill code.
3177 // Internals of the Matcher (including some VectorSets) must remain live
3178 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3179 // set a bit in reclaimed memory.
3180
3181 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3182 // nodes. Mapping is only valid at the root of each matched subtree.
3183 NOT_PRODUCT( verify_graph_edges(); )
3184
3185 Matcher matcher;
3186 _matcher = &matcher;
3187 {
3188 TracePhase tp("matcher", &timers[_t_matcher]);
3189 matcher.match();
3190 if (failing()) {
3191 return;
3192 }
3193 }
3194 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3195 // nodes. Mapping is only valid at the root of each matched subtree.
3196 NOT_PRODUCT( verify_graph_edges(); )
3197
3198 // If you have too many nodes, or if matching has failed, bail out
3199 check_node_count(0, "out of nodes matching instructions");
3200 if (failing()) {
3201 return;
3202 }
3203
3204 print_method(PHASE_MATCHING, 2);
3205
3206 // Build a proper-looking CFG
3207 PhaseCFG cfg(node_arena(), root(), matcher);
3208 _cfg = &cfg;
3209 {
3210 TracePhase tp("scheduler", &timers[_t_scheduler]);
3211 bool success = cfg.do_global_code_motion();
3212 if (!success) {
3213 return;
3214 }
3215
3216 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3217 NOT_PRODUCT( verify_graph_edges(); )
3218 cfg.verify();
3219 }
3220
3221 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3222 _regalloc = ®alloc;
3223 {
3224 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
3225 // Perform register allocation. After Chaitin, use-def chains are
3226 // no longer accurate (at spill code) and so must be ignored.
3227 // Node->LRG->reg mappings are still accurate.
3228 _regalloc->Register_Allocate();
3229
3230 // Bail out if the allocator builds too many nodes
3231 if (failing()) {
3232 return;
3233 }
3234 }
3235
3236 // Prior to register allocation we kept empty basic blocks in case the
3237 // the allocator needed a place to spill. After register allocation we
3238 // are not adding any new instructions. If any basic block is empty, we
3239 // can now safely remove it.
3240 {
3241 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
3242 cfg.remove_empty_blocks();
3243 if (do_freq_based_layout()) {
3244 PhaseBlockLayout layout(cfg);
3245 } else {
3246 cfg.set_loop_alignment();
3247 }
3248 cfg.fixup_flow();
3249 }
3250
3251 // Apply peephole optimizations
3252 if( OptoPeephole ) {
3253 TracePhase tp("peephole", &timers[_t_peephole]);
3254 PhasePeephole peep( _regalloc, cfg);
3255 peep.do_transform();
3256 }
3257
3258 // Do late expand if CPU requires this.
3259 if (Matcher::require_postalloc_expand) {
3260 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3261 cfg.postalloc_expand(_regalloc);
3262 }
3263
3264 // Convert Nodes to instruction bits in a buffer
3265 {
3266 TracePhase tp("output", &timers[_t_output]);
3267 PhaseOutput output;
3268 output.Output();
3269 if (failing()) return;
3270 output.install();
3271 }
3272
3273 print_method(PHASE_FINAL_CODE);
3274
3275 // He's dead, Jim.
3276 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3277 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3278 }
3279
3280 //------------------------------Final_Reshape_Counts---------------------------
3281 // This class defines counters to help identify when a method
3282 // may/must be executed using hardware with only 24-bit precision.
3283 struct Final_Reshape_Counts : public StackObj {
3284 int _call_count; // count non-inlined 'common' calls
3285 int _float_count; // count float ops requiring 24-bit precision
3286 int _double_count; // count double ops requiring more precision
3287 int _java_call_count; // count non-inlined 'java' calls
3288 int _inner_loop_count; // count loops which need alignment
3289 VectorSet _visited; // Visitation flags
3290 Node_List _tests; // Set of IfNodes & PCTableNodes
3291
3292 Final_Reshape_Counts() :
3293 _call_count(0), _float_count(0), _double_count(0),
3294 _java_call_count(0), _inner_loop_count(0) { }
3295
3296 void inc_call_count () { _call_count ++; }
3297 void inc_float_count () { _float_count ++; }
3298 void inc_double_count() { _double_count++; }
3299 void inc_java_call_count() { _java_call_count++; }
3300 void inc_inner_loop_count() { _inner_loop_count++; }
3301
3302 int get_call_count () const { return _call_count ; }
3303 int get_float_count () const { return _float_count ; }
3304 int get_double_count() const { return _double_count; }
3305 int get_java_call_count() const { return _java_call_count; }
3306 int get_inner_loop_count() const { return _inner_loop_count; }
3307 };
3308
3309 // Eliminate trivially redundant StoreCMs and accumulate their
3310 // precedence edges.
3311 void Compile::eliminate_redundant_card_marks(Node* n) {
3312 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3313 if (n->in(MemNode::Address)->outcnt() > 1) {
3314 // There are multiple users of the same address so it might be
3315 // possible to eliminate some of the StoreCMs
3316 Node* mem = n->in(MemNode::Memory);
3317 Node* adr = n->in(MemNode::Address);
3318 Node* val = n->in(MemNode::ValueIn);
3319 Node* prev = n;
3320 bool done = false;
3321 // Walk the chain of StoreCMs eliminating ones that match. As
3322 // long as it's a chain of single users then the optimization is
3323 // safe. Eliminating partially redundant StoreCMs would require
3324 // cloning copies down the other paths.
3325 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3326 if (adr == mem->in(MemNode::Address) &&
3327 val == mem->in(MemNode::ValueIn)) {
3328 // redundant StoreCM
3329 if (mem->req() > MemNode::OopStore) {
3330 // Hasn't been processed by this code yet.
3331 n->add_prec(mem->in(MemNode::OopStore));
3332 } else {
3333 // Already converted to precedence edge
3334 for (uint i = mem->req(); i < mem->len(); i++) {
3335 // Accumulate any precedence edges
3336 if (mem->in(i) != NULL) {
3337 n->add_prec(mem->in(i));
3338 }
3339 }
3340 // Everything above this point has been processed.
3341 done = true;
3342 }
3343 // Eliminate the previous StoreCM
3344 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3345 assert(mem->outcnt() == 0, "should be dead");
3346 mem->disconnect_inputs(this);
3347 } else {
3348 prev = mem;
3349 }
3350 mem = prev->in(MemNode::Memory);
3351 }
3352 }
3353 }
3354
3355
3356 //------------------------------final_graph_reshaping_impl----------------------
3357 // Implement items 1-5 from final_graph_reshaping below.
3358 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
3359
3360 if ( n->outcnt() == 0 ) return; // dead node
3361 uint nop = n->Opcode();
3362
3363 // Check for 2-input instruction with "last use" on right input.
3364 // Swap to left input. Implements item (2).
3365 if( n->req() == 3 && // two-input instruction
3366 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3367 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3368 n->in(2)->outcnt() == 1 &&// right use IS a last use
3369 !n->in(2)->is_Con() ) { // right use is not a constant
3370 // Check for commutative opcode
3371 switch( nop ) {
3372 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3373 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3374 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3375 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3376 case Op_AndL: case Op_XorL: case Op_OrL:
3377 case Op_AndI: case Op_XorI: case Op_OrI: {
3378 // Move "last use" input to left by swapping inputs
3379 n->swap_edges(1, 2);
3380 break;
3381 }
3382 default:
3383 break;
3384 }
3385 }
3386
3387 #ifdef ASSERT
3388 if( n->is_Mem() ) {
3389 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3390 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
3391 // oop will be recorded in oop map if load crosses safepoint
3392 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3393 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3394 "raw memory operations should have control edge");
3395 }
3396 if (n->is_MemBar()) {
3397 MemBarNode* mb = n->as_MemBar();
3398 if (mb->trailing_store() || mb->trailing_load_store()) {
3399 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3400 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3401 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3402 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3403 } else if (mb->leading()) {
3404 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3405 }
3406 }
3407 #endif
3408 // Count FPU ops and common calls, implements item (3)
3409 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
3410 if (!gc_handled) {
3411 final_graph_reshaping_main_switch(n, frc, nop);
3412 }
3413
3414 // Collect CFG split points
3415 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3416 frc._tests.push(n);
3417 }
3418 }
3419
3420 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
3421 switch( nop ) {
3422 // Count all float operations that may use FPU
3423 case Op_AddF:
3424 case Op_SubF:
3425 case Op_MulF:
3426 case Op_DivF:
3427 case Op_NegF:
3428 case Op_ModF:
3429 case Op_ConvI2F:
3430 case Op_ConF:
3431 case Op_CmpF:
3432 case Op_CmpF3:
3433 case Op_StoreF:
3434 case Op_LoadF:
3435 // case Op_ConvL2F: // longs are split into 32-bit halves
3436 frc.inc_float_count();
3437 break;
3438
3439 case Op_ConvF2D:
3440 case Op_ConvD2F:
3441 frc.inc_float_count();
3442 frc.inc_double_count();
3443 break;
3444
3445 // Count all double operations that may use FPU
3446 case Op_AddD:
3447 case Op_SubD:
3448 case Op_MulD:
3449 case Op_DivD:
3450 case Op_NegD:
3451 case Op_ModD:
3452 case Op_ConvI2D:
3453 case Op_ConvD2I:
3454 // case Op_ConvL2D: // handled by leaf call
3455 // case Op_ConvD2L: // handled by leaf call
3456 case Op_ConD:
3457 case Op_CmpD:
3458 case Op_CmpD3:
3459 case Op_StoreD:
3460 case Op_LoadD:
3461 case Op_LoadD_unaligned:
3462 frc.inc_double_count();
3463 break;
3464 case Op_Opaque1: // Remove Opaque Nodes before matching
3465 case Op_Opaque2: // Remove Opaque Nodes before matching
3466 case Op_Opaque3:
3467 n->subsume_by(n->in(1), this);
3468 break;
3469 case Op_CallStaticJava:
3470 case Op_CallJava:
3471 case Op_CallDynamicJava:
3472 frc.inc_java_call_count(); // Count java call site;
3473 case Op_CallRuntime:
3474 case Op_CallLeaf:
3475 case Op_CallLeafVector:
3476 case Op_CallNative:
3477 case Op_CallLeafNoFP: {
3478 assert (n->is_Call(), "");
3479 CallNode *call = n->as_Call();
3480 // Count call sites where the FP mode bit would have to be flipped.
3481 // Do not count uncommon runtime calls:
3482 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3483 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3484 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3485 frc.inc_call_count(); // Count the call site
3486 } else { // See if uncommon argument is shared
3487 Node *n = call->in(TypeFunc::Parms);
3488 int nop = n->Opcode();
3489 // Clone shared simple arguments to uncommon calls, item (1).
3490 if (n->outcnt() > 1 &&
3491 !n->is_Proj() &&
3492 nop != Op_CreateEx &&
3493 nop != Op_CheckCastPP &&
3494 nop != Op_DecodeN &&
3495 nop != Op_DecodeNKlass &&
3496 !n->is_Mem() &&
3497 !n->is_Phi()) {
3498 Node *x = n->clone();
3499 call->set_req(TypeFunc::Parms, x);
3500 }
3501 }
3502 break;
3503 }
3504
3505 case Op_StoreCM:
3506 {
3507 // Convert OopStore dependence into precedence edge
3508 Node* prec = n->in(MemNode::OopStore);
3509 n->del_req(MemNode::OopStore);
3510 if (prec->is_MergeMem()) {
3511 MergeMemNode* mm = prec->as_MergeMem();
3512 Node* base = mm->base_memory();
3513 for (int i = AliasIdxRaw + 1; i < num_alias_types(); i++) {
3514 const Type* adr_type = get_adr_type(i);
3515 if (adr_type->isa_aryptr() && adr_type->is_aryptr()->is_flat()) {
3516 Node* m = mm->memory_at(i);
3517 n->add_prec(m);
3518 }
3519 }
3520 if (mm->outcnt() == 0) {
3521 mm->disconnect_inputs(this);
3522 }
3523 } else {
3524 n->add_prec(prec);
3525 }
3526 eliminate_redundant_card_marks(n);
3527 }
3528
3529 // fall through
3530
3531 case Op_StoreB:
3532 case Op_StoreC:
3533 case Op_StorePConditional:
3534 case Op_StoreI:
3535 case Op_StoreL:
3536 case Op_StoreIConditional:
3537 case Op_StoreLConditional:
3538 case Op_CompareAndSwapB:
3539 case Op_CompareAndSwapS:
3540 case Op_CompareAndSwapI:
3541 case Op_CompareAndSwapL:
3542 case Op_CompareAndSwapP:
3543 case Op_CompareAndSwapN:
3544 case Op_WeakCompareAndSwapB:
3545 case Op_WeakCompareAndSwapS:
3546 case Op_WeakCompareAndSwapI:
3547 case Op_WeakCompareAndSwapL:
3548 case Op_WeakCompareAndSwapP:
3549 case Op_WeakCompareAndSwapN:
3550 case Op_CompareAndExchangeB:
3551 case Op_CompareAndExchangeS:
3552 case Op_CompareAndExchangeI:
3553 case Op_CompareAndExchangeL:
3554 case Op_CompareAndExchangeP:
3555 case Op_CompareAndExchangeN:
3556 case Op_GetAndAddS:
3557 case Op_GetAndAddB:
3558 case Op_GetAndAddI:
3559 case Op_GetAndAddL:
3560 case Op_GetAndSetS:
3561 case Op_GetAndSetB:
3562 case Op_GetAndSetI:
3563 case Op_GetAndSetL:
3564 case Op_GetAndSetP:
3565 case Op_GetAndSetN:
3566 case Op_StoreP:
3567 case Op_StoreN:
3568 case Op_StoreNKlass:
3569 case Op_LoadB:
3570 case Op_LoadUB:
3571 case Op_LoadUS:
3572 case Op_LoadI:
3573 case Op_LoadKlass:
3574 case Op_LoadNKlass:
3575 case Op_LoadL:
3576 case Op_LoadL_unaligned:
3577 case Op_LoadPLocked:
3578 case Op_LoadP:
3579 case Op_LoadN:
3580 case Op_LoadRange:
3581 case Op_LoadS:
3582 break;
3583
3584 case Op_AddP: { // Assert sane base pointers
3585 Node *addp = n->in(AddPNode::Address);
3586 assert( !addp->is_AddP() ||
3587 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3588 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3589 "Base pointers must match (addp %u)", addp->_idx );
3590 #ifdef _LP64
3591 if ((UseCompressedOops || UseCompressedClassPointers) &&
3592 addp->Opcode() == Op_ConP &&
3593 addp == n->in(AddPNode::Base) &&
3594 n->in(AddPNode::Offset)->is_Con()) {
3595 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3596 // on the platform and on the compressed oops mode.
3597 // Use addressing with narrow klass to load with offset on x86.
3598 // Some platforms can use the constant pool to load ConP.
3599 // Do this transformation here since IGVN will convert ConN back to ConP.
3600 const Type* t = addp->bottom_type();
3601 bool is_oop = t->isa_oopptr() != NULL;
3602 bool is_klass = t->isa_klassptr() != NULL;
3603
3604 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3605 (is_klass && Matcher::const_klass_prefer_decode())) {
3606 Node* nn = NULL;
3607
3608 int op = is_oop ? Op_ConN : Op_ConNKlass;
3609
3610 // Look for existing ConN node of the same exact type.
3611 Node* r = root();
3612 uint cnt = r->outcnt();
3613 for (uint i = 0; i < cnt; i++) {
3614 Node* m = r->raw_out(i);
3615 if (m!= NULL && m->Opcode() == op &&
3616 m->bottom_type()->make_ptr() == t) {
3617 nn = m;
3618 break;
3619 }
3620 }
3621 if (nn != NULL) {
3622 // Decode a narrow oop to match address
3623 // [R12 + narrow_oop_reg<<3 + offset]
3624 if (is_oop) {
3625 nn = new DecodeNNode(nn, t);
3626 } else {
3627 nn = new DecodeNKlassNode(nn, t);
3628 }
3629 // Check for succeeding AddP which uses the same Base.
3630 // Otherwise we will run into the assertion above when visiting that guy.
3631 for (uint i = 0; i < n->outcnt(); ++i) {
3632 Node *out_i = n->raw_out(i);
3633 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3634 out_i->set_req(AddPNode::Base, nn);
3635 #ifdef ASSERT
3636 for (uint j = 0; j < out_i->outcnt(); ++j) {
3637 Node *out_j = out_i->raw_out(j);
3638 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3639 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3640 }
3641 #endif
3642 }
3643 }
3644 n->set_req(AddPNode::Base, nn);
3645 n->set_req(AddPNode::Address, nn);
3646 if (addp->outcnt() == 0) {
3647 addp->disconnect_inputs(this);
3648 }
3649 }
3650 }
3651 }
3652 #endif
3653 break;
3654 }
3655
3656 case Op_CastPP: {
3657 // Remove CastPP nodes to gain more freedom during scheduling but
3658 // keep the dependency they encode as control or precedence edges
3659 // (if control is set already) on memory operations. Some CastPP
3660 // nodes don't have a control (don't carry a dependency): skip
3661 // those.
3662 if (n->in(0) != NULL) {
3663 ResourceMark rm;
3664 Unique_Node_List wq;
3665 wq.push(n);
3666 for (uint next = 0; next < wq.size(); ++next) {
3667 Node *m = wq.at(next);
3668 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3669 Node* use = m->fast_out(i);
3670 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3671 use->ensure_control_or_add_prec(n->in(0));
3672 } else {
3673 switch(use->Opcode()) {
3674 case Op_AddP:
3675 case Op_DecodeN:
3676 case Op_DecodeNKlass:
3677 case Op_CheckCastPP:
3678 case Op_CastPP:
3679 wq.push(use);
3680 break;
3681 }
3682 }
3683 }
3684 }
3685 }
3686 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3687 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3688 Node* in1 = n->in(1);
3689 const Type* t = n->bottom_type();
3690 Node* new_in1 = in1->clone();
3691 new_in1->as_DecodeN()->set_type(t);
3692
3693 if (!Matcher::narrow_oop_use_complex_address()) {
3694 //
3695 // x86, ARM and friends can handle 2 adds in addressing mode
3696 // and Matcher can fold a DecodeN node into address by using
3697 // a narrow oop directly and do implicit NULL check in address:
3698 //
3699 // [R12 + narrow_oop_reg<<3 + offset]
3700 // NullCheck narrow_oop_reg
3701 //
3702 // On other platforms (Sparc) we have to keep new DecodeN node and
3703 // use it to do implicit NULL check in address:
3704 //
3705 // decode_not_null narrow_oop_reg, base_reg
3706 // [base_reg + offset]
3707 // NullCheck base_reg
3708 //
3709 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3710 // to keep the information to which NULL check the new DecodeN node
3711 // corresponds to use it as value in implicit_null_check().
3712 //
3713 new_in1->set_req(0, n->in(0));
3714 }
3715
3716 n->subsume_by(new_in1, this);
3717 if (in1->outcnt() == 0) {
3718 in1->disconnect_inputs(this);
3719 }
3720 } else {
3721 n->subsume_by(n->in(1), this);
3722 if (n->outcnt() == 0) {
3723 n->disconnect_inputs(this);
3724 }
3725 }
3726 break;
3727 }
3728 #ifdef _LP64
3729 case Op_CmpP:
3730 // Do this transformation here to preserve CmpPNode::sub() and
3731 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3732 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3733 Node* in1 = n->in(1);
3734 Node* in2 = n->in(2);
3735 if (!in1->is_DecodeNarrowPtr()) {
3736 in2 = in1;
3737 in1 = n->in(2);
3738 }
3739 assert(in1->is_DecodeNarrowPtr(), "sanity");
3740
3741 Node* new_in2 = NULL;
3742 if (in2->is_DecodeNarrowPtr()) {
3743 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3744 new_in2 = in2->in(1);
3745 } else if (in2->Opcode() == Op_ConP) {
3746 const Type* t = in2->bottom_type();
3747 if (t == TypePtr::NULL_PTR) {
3748 assert(in1->is_DecodeN(), "compare klass to null?");
3749 // Don't convert CmpP null check into CmpN if compressed
3750 // oops implicit null check is not generated.
3751 // This will allow to generate normal oop implicit null check.
3752 if (Matcher::gen_narrow_oop_implicit_null_checks())
3753 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3754 //
3755 // This transformation together with CastPP transformation above
3756 // will generated code for implicit NULL checks for compressed oops.
3757 //
3758 // The original code after Optimize()
3759 //
3760 // LoadN memory, narrow_oop_reg
3761 // decode narrow_oop_reg, base_reg
3762 // CmpP base_reg, NULL
3763 // CastPP base_reg // NotNull
3764 // Load [base_reg + offset], val_reg
3765 //
3766 // after these transformations will be
3767 //
3768 // LoadN memory, narrow_oop_reg
3769 // CmpN narrow_oop_reg, NULL
3770 // decode_not_null narrow_oop_reg, base_reg
3771 // Load [base_reg + offset], val_reg
3772 //
3773 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3774 // since narrow oops can be used in debug info now (see the code in
3775 // final_graph_reshaping_walk()).
3776 //
3777 // At the end the code will be matched to
3778 // on x86:
3779 //
3780 // Load_narrow_oop memory, narrow_oop_reg
3781 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3782 // NullCheck narrow_oop_reg
3783 //
3784 // and on sparc:
3785 //
3786 // Load_narrow_oop memory, narrow_oop_reg
3787 // decode_not_null narrow_oop_reg, base_reg
3788 // Load [base_reg + offset], val_reg
3789 // NullCheck base_reg
3790 //
3791 } else if (t->isa_oopptr()) {
3792 new_in2 = ConNode::make(t->make_narrowoop());
3793 } else if (t->isa_klassptr()) {
3794 new_in2 = ConNode::make(t->make_narrowklass());
3795 }
3796 }
3797 if (new_in2 != NULL) {
3798 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3799 n->subsume_by(cmpN, this);
3800 if (in1->outcnt() == 0) {
3801 in1->disconnect_inputs(this);
3802 }
3803 if (in2->outcnt() == 0) {
3804 in2->disconnect_inputs(this);
3805 }
3806 }
3807 }
3808 break;
3809
3810 case Op_DecodeN:
3811 case Op_DecodeNKlass:
3812 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3813 // DecodeN could be pinned when it can't be fold into
3814 // an address expression, see the code for Op_CastPP above.
3815 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3816 break;
3817
3818 case Op_EncodeP:
3819 case Op_EncodePKlass: {
3820 Node* in1 = n->in(1);
3821 if (in1->is_DecodeNarrowPtr()) {
3822 n->subsume_by(in1->in(1), this);
3823 } else if (in1->Opcode() == Op_ConP) {
3824 const Type* t = in1->bottom_type();
3825 if (t == TypePtr::NULL_PTR) {
3826 assert(t->isa_oopptr(), "null klass?");
3827 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3828 } else if (t->isa_oopptr()) {
3829 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3830 } else if (t->isa_klassptr()) {
3831 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3832 }
3833 }
3834 if (in1->outcnt() == 0) {
3835 in1->disconnect_inputs(this);
3836 }
3837 break;
3838 }
3839
3840 case Op_Proj: {
3841 if (OptimizeStringConcat || IncrementalInline) {
3842 ProjNode* proj = n->as_Proj();
3843 if (proj->_is_io_use) {
3844 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3845 // Separate projections were used for the exception path which
3846 // are normally removed by a late inline. If it wasn't inlined
3847 // then they will hang around and should just be replaced with
3848 // the original one. Merge them.
3849 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3850 if (non_io_proj != NULL) {
3851 proj->subsume_by(non_io_proj , this);
3852 }
3853 }
3854 }
3855 break;
3856 }
3857
3858 case Op_Phi:
3859 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3860 // The EncodeP optimization may create Phi with the same edges
3861 // for all paths. It is not handled well by Register Allocator.
3862 Node* unique_in = n->in(1);
3863 assert(unique_in != NULL, "");
3864 uint cnt = n->req();
3865 for (uint i = 2; i < cnt; i++) {
3866 Node* m = n->in(i);
3867 assert(m != NULL, "");
3868 if (unique_in != m)
3869 unique_in = NULL;
3870 }
3871 if (unique_in != NULL) {
3872 n->subsume_by(unique_in, this);
3873 }
3874 }
3875 break;
3876
3877 #endif
3878
3879 #ifdef ASSERT
3880 case Op_CastII:
3881 // Verify that all range check dependent CastII nodes were removed.
3882 if (n->isa_CastII()->has_range_check()) {
3883 n->dump(3);
3884 assert(false, "Range check dependent CastII node was not removed");
3885 }
3886 break;
3887 #endif
3888
3889 case Op_ModI:
3890 if (UseDivMod) {
3891 // Check if a%b and a/b both exist
3892 Node* d = n->find_similar(Op_DivI);
3893 if (d) {
3894 // Replace them with a fused divmod if supported
3895 if (Matcher::has_match_rule(Op_DivModI)) {
3896 DivModINode* divmod = DivModINode::make(n);
3897 d->subsume_by(divmod->div_proj(), this);
3898 n->subsume_by(divmod->mod_proj(), this);
3899 } else {
3900 // replace a%b with a-((a/b)*b)
3901 Node* mult = new MulINode(d, d->in(2));
3902 Node* sub = new SubINode(d->in(1), mult);
3903 n->subsume_by(sub, this);
3904 }
3905 }
3906 }
3907 break;
3908
3909 case Op_ModL:
3910 if (UseDivMod) {
3911 // Check if a%b and a/b both exist
3912 Node* d = n->find_similar(Op_DivL);
3913 if (d) {
3914 // Replace them with a fused divmod if supported
3915 if (Matcher::has_match_rule(Op_DivModL)) {
3916 DivModLNode* divmod = DivModLNode::make(n);
3917 d->subsume_by(divmod->div_proj(), this);
3918 n->subsume_by(divmod->mod_proj(), this);
3919 } else {
3920 // replace a%b with a-((a/b)*b)
3921 Node* mult = new MulLNode(d, d->in(2));
3922 Node* sub = new SubLNode(d->in(1), mult);
3923 n->subsume_by(sub, this);
3924 }
3925 }
3926 }
3927 break;
3928
3929 case Op_LoadVector:
3930 case Op_StoreVector:
3931 case Op_LoadVectorGather:
3932 case Op_StoreVectorScatter:
3933 case Op_VectorCmpMasked:
3934 case Op_VectorMaskGen:
3935 case Op_LoadVectorMasked:
3936 case Op_StoreVectorMasked:
3937 break;
3938
3939 case Op_AddReductionVI:
3940 case Op_AddReductionVL:
3941 case Op_AddReductionVF:
3942 case Op_AddReductionVD:
3943 case Op_MulReductionVI:
3944 case Op_MulReductionVL:
3945 case Op_MulReductionVF:
3946 case Op_MulReductionVD:
3947 case Op_MinReductionV:
3948 case Op_MaxReductionV:
3949 case Op_AndReductionV:
3950 case Op_OrReductionV:
3951 case Op_XorReductionV:
3952 break;
3953
3954 case Op_PackB:
3955 case Op_PackS:
3956 case Op_PackI:
3957 case Op_PackF:
3958 case Op_PackL:
3959 case Op_PackD:
3960 if (n->req()-1 > 2) {
3961 // Replace many operand PackNodes with a binary tree for matching
3962 PackNode* p = (PackNode*) n;
3963 Node* btp = p->binary_tree_pack(1, n->req());
3964 n->subsume_by(btp, this);
3965 }
3966 break;
3967 case Op_Loop:
3968 assert(!n->as_Loop()->is_transformed_long_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3969 case Op_CountedLoop:
3970 case Op_LongCountedLoop:
3971 case Op_OuterStripMinedLoop:
3972 if (n->as_Loop()->is_inner_loop()) {
3973 frc.inc_inner_loop_count();
3974 }
3975 n->as_Loop()->verify_strip_mined(0);
3976 break;
3977 case Op_LShiftI:
3978 case Op_RShiftI:
3979 case Op_URShiftI:
3980 case Op_LShiftL:
3981 case Op_RShiftL:
3982 case Op_URShiftL:
3983 if (Matcher::need_masked_shift_count) {
3984 // The cpu's shift instructions don't restrict the count to the
3985 // lower 5/6 bits. We need to do the masking ourselves.
3986 Node* in2 = n->in(2);
3987 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3988 const TypeInt* t = in2->find_int_type();
3989 if (t != NULL && t->is_con()) {
3990 juint shift = t->get_con();
3991 if (shift > mask) { // Unsigned cmp
3992 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3993 }
3994 } else {
3995 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3996 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3997 n->set_req(2, shift);
3998 }
3999 }
4000 if (in2->outcnt() == 0) { // Remove dead node
4001 in2->disconnect_inputs(this);
4002 }
4003 }
4004 break;
4005 case Op_MemBarStoreStore:
4006 case Op_MemBarRelease:
4007 // Break the link with AllocateNode: it is no longer useful and
4008 // confuses register allocation.
4009 if (n->req() > MemBarNode::Precedent) {
4010 n->set_req(MemBarNode::Precedent, top());
4011 }
4012 break;
4013 case Op_MemBarAcquire: {
4014 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4015 // At parse time, the trailing MemBarAcquire for a volatile load
4016 // is created with an edge to the load. After optimizations,
4017 // that input may be a chain of Phis. If those phis have no
4018 // other use, then the MemBarAcquire keeps them alive and
4019 // register allocation can be confused.
4020 ResourceMark rm;
4021 Unique_Node_List wq;
4022 wq.push(n->in(MemBarNode::Precedent));
4023 n->set_req(MemBarNode::Precedent, top());
4024 while (wq.size() > 0) {
4025 Node* m = wq.pop();
4026 if (m->outcnt() == 0) {
4027 for (uint j = 0; j < m->req(); j++) {
4028 Node* in = m->in(j);
4029 if (in != NULL) {
4030 wq.push(in);
4031 }
4032 }
4033 m->disconnect_inputs(this);
4034 }
4035 }
4036 }
4037 break;
4038 }
4039 case Op_Blackhole:
4040 break;
4041 case Op_RangeCheck: {
4042 RangeCheckNode* rc = n->as_RangeCheck();
4043 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4044 n->subsume_by(iff, this);
4045 frc._tests.push(iff);
4046 break;
4047 }
4048 case Op_ConvI2L: {
4049 if (!Matcher::convi2l_type_required) {
4050 // Code generation on some platforms doesn't need accurate
4051 // ConvI2L types. Widening the type can help remove redundant
4052 // address computations.
4053 n->as_Type()->set_type(TypeLong::INT);
4054 ResourceMark rm;
4055 Unique_Node_List wq;
4056 wq.push(n);
4057 for (uint next = 0; next < wq.size(); next++) {
4058 Node *m = wq.at(next);
4059
4060 for(;;) {
4061 // Loop over all nodes with identical inputs edges as m
4062 Node* k = m->find_similar(m->Opcode());
4063 if (k == NULL) {
4064 break;
4065 }
4066 // Push their uses so we get a chance to remove node made
4067 // redundant
4068 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4069 Node* u = k->fast_out(i);
4070 if (u->Opcode() == Op_LShiftL ||
4071 u->Opcode() == Op_AddL ||
4072 u->Opcode() == Op_SubL ||
4073 u->Opcode() == Op_AddP) {
4074 wq.push(u);
4075 }
4076 }
4077 // Replace all nodes with identical edges as m with m
4078 k->subsume_by(m, this);
4079 }
4080 }
4081 }
4082 break;
4083 }
4084 case Op_CmpUL: {
4085 if (!Matcher::has_match_rule(Op_CmpUL)) {
4086 // No support for unsigned long comparisons
4087 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4088 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4089 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4090 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4091 Node* andl = new AndLNode(orl, remove_sign_mask);
4092 Node* cmp = new CmpLNode(andl, n->in(2));
4093 n->subsume_by(cmp, this);
4094 }
4095 break;
4096 }
4097 #ifdef ASSERT
4098 case Op_InlineTypePtr:
4099 case Op_InlineType: {
4100 n->dump(-1);
4101 assert(false, "inline type node was not removed");
4102 break;
4103 }
4104 #endif
4105 default:
4106 assert(!n->is_Call(), "");
4107 assert(!n->is_Mem(), "");
4108 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4109 break;
4110 }
4111 }
4112
4113 //------------------------------final_graph_reshaping_walk---------------------
4114 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4115 // requires that the walk visits a node's inputs before visiting the node.
4116 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
4117 Unique_Node_List sfpt;
4118
4119 frc._visited.set(root->_idx); // first, mark node as visited
4120 uint cnt = root->req();
4121 Node *n = root;
4122 uint i = 0;
4123 while (true) {
4124 if (i < cnt) {
4125 // Place all non-visited non-null inputs onto stack
4126 Node* m = n->in(i);
4127 ++i;
4128 if (m != NULL && !frc._visited.test_set(m->_idx)) {
4129 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
4130 // compute worst case interpreter size in case of a deoptimization
4131 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4132
4133 sfpt.push(m);
4134 }
4135 cnt = m->req();
4136 nstack.push(n, i); // put on stack parent and next input's index
4137 n = m;
4138 i = 0;
4139 }
4140 } else {
4141 // Now do post-visit work
4142 final_graph_reshaping_impl( n, frc );
4143 if (nstack.is_empty())
4144 break; // finished
4145 n = nstack.node(); // Get node from stack
4146 cnt = n->req();
4147 i = nstack.index();
4148 nstack.pop(); // Shift to the next node on stack
4149 }
4150 }
4151
4152 // Skip next transformation if compressed oops are not used.
4153 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4154 (!UseCompressedOops && !UseCompressedClassPointers))
4155 return;
4156
4157 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4158 // It could be done for an uncommon traps or any safepoints/calls
4159 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4160 while (sfpt.size() > 0) {
4161 n = sfpt.pop();
4162 JVMState *jvms = n->as_SafePoint()->jvms();
4163 assert(jvms != NULL, "sanity");
4164 int start = jvms->debug_start();
4165 int end = n->req();
4166 bool is_uncommon = (n->is_CallStaticJava() &&
4167 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4168 for (int j = start; j < end; j++) {
4169 Node* in = n->in(j);
4170 if (in->is_DecodeNarrowPtr()) {
4171 bool safe_to_skip = true;
4172 if (!is_uncommon ) {
4173 // Is it safe to skip?
4174 for (uint i = 0; i < in->outcnt(); i++) {
4175 Node* u = in->raw_out(i);
4176 if (!u->is_SafePoint() ||
4177 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4178 safe_to_skip = false;
4179 }
4180 }
4181 }
4182 if (safe_to_skip) {
4183 n->set_req(j, in->in(1));
4184 }
4185 if (in->outcnt() == 0) {
4186 in->disconnect_inputs(this);
4187 }
4188 }
4189 }
4190 }
4191 }
4192
4193 //------------------------------final_graph_reshaping--------------------------
4194 // Final Graph Reshaping.
4195 //
4196 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4197 // and not commoned up and forced early. Must come after regular
4198 // optimizations to avoid GVN undoing the cloning. Clone constant
4199 // inputs to Loop Phis; these will be split by the allocator anyways.
4200 // Remove Opaque nodes.
4201 // (2) Move last-uses by commutative operations to the left input to encourage
4202 // Intel update-in-place two-address operations and better register usage
4203 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4204 // calls canonicalizing them back.
4205 // (3) Count the number of double-precision FP ops, single-precision FP ops
4206 // and call sites. On Intel, we can get correct rounding either by
4207 // forcing singles to memory (requires extra stores and loads after each
4208 // FP bytecode) or we can set a rounding mode bit (requires setting and
4209 // clearing the mode bit around call sites). The mode bit is only used
4210 // if the relative frequency of single FP ops to calls is low enough.
4211 // This is a key transform for SPEC mpeg_audio.
4212 // (4) Detect infinite loops; blobs of code reachable from above but not
4213 // below. Several of the Code_Gen algorithms fail on such code shapes,
4214 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4215 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4216 // Detection is by looking for IfNodes where only 1 projection is
4217 // reachable from below or CatchNodes missing some targets.
4218 // (5) Assert for insane oop offsets in debug mode.
4219
4220 bool Compile::final_graph_reshaping() {
4221 // an infinite loop may have been eliminated by the optimizer,
4222 // in which case the graph will be empty.
4223 if (root()->req() == 1) {
4224 record_method_not_compilable("trivial infinite loop");
4225 return true;
4226 }
4227
4228 // Expensive nodes have their control input set to prevent the GVN
4229 // from freely commoning them. There's no GVN beyond this point so
4230 // no need to keep the control input. We want the expensive nodes to
4231 // be freely moved to the least frequent code path by gcm.
4232 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4233 for (int i = 0; i < expensive_count(); i++) {
4234 _expensive_nodes.at(i)->set_req(0, NULL);
4235 }
4236
4237 Final_Reshape_Counts frc;
4238
4239 // Visit everybody reachable!
4240 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4241 Node_Stack nstack(live_nodes() >> 1);
4242 final_graph_reshaping_walk(nstack, root(), frc);
4243
4244 // Check for unreachable (from below) code (i.e., infinite loops).
4245 for( uint i = 0; i < frc._tests.size(); i++ ) {
4246 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4247 // Get number of CFG targets.
4248 // Note that PCTables include exception targets after calls.
4249 uint required_outcnt = n->required_outcnt();
4250 if (n->outcnt() != required_outcnt) {
4251 // Check for a few special cases. Rethrow Nodes never take the
4252 // 'fall-thru' path, so expected kids is 1 less.
4253 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4254 if (n->in(0)->in(0)->is_Call()) {
4255 CallNode *call = n->in(0)->in(0)->as_Call();
4256 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4257 required_outcnt--; // Rethrow always has 1 less kid
4258 } else if (call->req() > TypeFunc::Parms &&
4259 call->is_CallDynamicJava()) {
4260 // Check for null receiver. In such case, the optimizer has
4261 // detected that the virtual call will always result in a null
4262 // pointer exception. The fall-through projection of this CatchNode
4263 // will not be populated.
4264 Node *arg0 = call->in(TypeFunc::Parms);
4265 if (arg0->is_Type() &&
4266 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4267 required_outcnt--;
4268 }
4269 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
4270 call->req() > TypeFunc::Parms+1 &&
4271 call->is_CallStaticJava()) {
4272 // Check for negative array length. In such case, the optimizer has
4273 // detected that the allocation attempt will always result in an
4274 // exception. There is no fall-through projection of this CatchNode .
4275 Node *arg1 = call->in(TypeFunc::Parms+1);
4276 if (arg1->is_Type() &&
4277 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
4278 required_outcnt--;
4279 }
4280 }
4281 }
4282 }
4283 // Recheck with a better notion of 'required_outcnt'
4284 if (n->outcnt() != required_outcnt) {
4285 record_method_not_compilable("malformed control flow");
4286 return true; // Not all targets reachable!
4287 }
4288 }
4289 // Check that I actually visited all kids. Unreached kids
4290 // must be infinite loops.
4291 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4292 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4293 record_method_not_compilable("infinite loop");
4294 return true; // Found unvisited kid; must be unreach
4295 }
4296
4297 // Here so verification code in final_graph_reshaping_walk()
4298 // always see an OuterStripMinedLoopEnd
4299 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4300 IfNode* init_iff = n->as_If();
4301 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4302 n->subsume_by(iff, this);
4303 }
4304 }
4305
4306 #ifdef IA32
4307 // If original bytecodes contained a mixture of floats and doubles
4308 // check if the optimizer has made it homogenous, item (3).
4309 if (UseSSE == 0 &&
4310 frc.get_float_count() > 32 &&
4311 frc.get_double_count() == 0 &&
4312 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4313 set_24_bit_selection_and_mode(false, true);
4314 }
4315 #endif // IA32
4316
4317 set_java_calls(frc.get_java_call_count());
4318 set_inner_loops(frc.get_inner_loop_count());
4319
4320 // No infinite loops, no reason to bail out.
4321 return false;
4322 }
4323
4324 //-----------------------------too_many_traps----------------------------------
4325 // Report if there are too many traps at the current method and bci.
4326 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4327 bool Compile::too_many_traps(ciMethod* method,
4328 int bci,
4329 Deoptimization::DeoptReason reason) {
4330 ciMethodData* md = method->method_data();
4331 if (md->is_empty()) {
4332 // Assume the trap has not occurred, or that it occurred only
4333 // because of a transient condition during start-up in the interpreter.
4334 return false;
4335 }
4336 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4337 if (md->has_trap_at(bci, m, reason) != 0) {
4338 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4339 // Also, if there are multiple reasons, or if there is no per-BCI record,
4340 // assume the worst.
4341 if (log())
4342 log()->elem("observe trap='%s' count='%d'",
4343 Deoptimization::trap_reason_name(reason),
4344 md->trap_count(reason));
4345 return true;
4346 } else {
4347 // Ignore method/bci and see if there have been too many globally.
4348 return too_many_traps(reason, md);
4349 }
4350 }
4351
4352 // Less-accurate variant which does not require a method and bci.
4353 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4354 ciMethodData* logmd) {
4355 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4356 // Too many traps globally.
4357 // Note that we use cumulative trap_count, not just md->trap_count.
4358 if (log()) {
4359 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4360 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4361 Deoptimization::trap_reason_name(reason),
4362 mcount, trap_count(reason));
4363 }
4364 return true;
4365 } else {
4366 // The coast is clear.
4367 return false;
4368 }
4369 }
4370
4371 //--------------------------too_many_recompiles--------------------------------
4372 // Report if there are too many recompiles at the current method and bci.
4373 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4374 // Is not eager to return true, since this will cause the compiler to use
4375 // Action_none for a trap point, to avoid too many recompilations.
4376 bool Compile::too_many_recompiles(ciMethod* method,
4377 int bci,
4378 Deoptimization::DeoptReason reason) {
4379 ciMethodData* md = method->method_data();
4380 if (md->is_empty()) {
4381 // Assume the trap has not occurred, or that it occurred only
4382 // because of a transient condition during start-up in the interpreter.
4383 return false;
4384 }
4385 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4386 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4387 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4388 Deoptimization::DeoptReason per_bc_reason
4389 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4390 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4391 if ((per_bc_reason == Deoptimization::Reason_none
4392 || md->has_trap_at(bci, m, reason) != 0)
4393 // The trap frequency measure we care about is the recompile count:
4394 && md->trap_recompiled_at(bci, m)
4395 && md->overflow_recompile_count() >= bc_cutoff) {
4396 // Do not emit a trap here if it has already caused recompilations.
4397 // Also, if there are multiple reasons, or if there is no per-BCI record,
4398 // assume the worst.
4399 if (log())
4400 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4401 Deoptimization::trap_reason_name(reason),
4402 md->trap_count(reason),
4403 md->overflow_recompile_count());
4404 return true;
4405 } else if (trap_count(reason) != 0
4406 && decompile_count() >= m_cutoff) {
4407 // Too many recompiles globally, and we have seen this sort of trap.
4408 // Use cumulative decompile_count, not just md->decompile_count.
4409 if (log())
4410 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4411 Deoptimization::trap_reason_name(reason),
4412 md->trap_count(reason), trap_count(reason),
4413 md->decompile_count(), decompile_count());
4414 return true;
4415 } else {
4416 // The coast is clear.
4417 return false;
4418 }
4419 }
4420
4421 // Compute when not to trap. Used by matching trap based nodes and
4422 // NullCheck optimization.
4423 void Compile::set_allowed_deopt_reasons() {
4424 _allowed_reasons = 0;
4425 if (is_method_compilation()) {
4426 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4427 assert(rs < BitsPerInt, "recode bit map");
4428 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4429 _allowed_reasons |= nth_bit(rs);
4430 }
4431 }
4432 }
4433 }
4434
4435 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4436 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4437 }
4438
4439 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4440 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4441 }
4442
4443 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4444 if (holder->is_initialized()) {
4445 return false;
4446 }
4447 if (holder->is_being_initialized()) {
4448 if (accessing_method->holder() == holder) {
4449 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4450 // <init>, or a static method. In all those cases, there was an initialization
4451 // barrier on the holder klass passed.
4452 if (accessing_method->is_class_initializer() ||
4453 accessing_method->is_object_constructor() ||
4454 accessing_method->is_static()) {
4455 return false;
4456 }
4457 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4458 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4459 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4460 // child class can become fully initialized while its parent class is still being initialized.
4461 if (accessing_method->is_class_initializer()) {
4462 return false;
4463 }
4464 }
4465 ciMethod* root = method(); // the root method of compilation
4466 if (root != accessing_method) {
4467 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4468 }
4469 }
4470 return true;
4471 }
4472
4473 #ifndef PRODUCT
4474 //------------------------------verify_graph_edges---------------------------
4475 // Walk the Graph and verify that there is a one-to-one correspondence
4476 // between Use-Def edges and Def-Use edges in the graph.
4477 void Compile::verify_graph_edges(bool no_dead_code) {
4478 if (VerifyGraphEdges) {
4479 Unique_Node_List visited;
4480 // Call recursive graph walk to check edges
4481 _root->verify_edges(visited);
4482 if (no_dead_code) {
4483 // Now make sure that no visited node is used by an unvisited node.
4484 bool dead_nodes = false;
4485 Unique_Node_List checked;
4486 while (visited.size() > 0) {
4487 Node* n = visited.pop();
4488 checked.push(n);
4489 for (uint i = 0; i < n->outcnt(); i++) {
4490 Node* use = n->raw_out(i);
4491 if (checked.member(use)) continue; // already checked
4492 if (visited.member(use)) continue; // already in the graph
4493 if (use->is_Con()) continue; // a dead ConNode is OK
4494 // At this point, we have found a dead node which is DU-reachable.
4495 if (!dead_nodes) {
4496 tty->print_cr("*** Dead nodes reachable via DU edges:");
4497 dead_nodes = true;
4498 }
4499 use->dump(2);
4500 tty->print_cr("---");
4501 checked.push(use); // No repeats; pretend it is now checked.
4502 }
4503 }
4504 assert(!dead_nodes, "using nodes must be reachable from root");
4505 }
4506 }
4507 }
4508 #endif
4509
4510 // The Compile object keeps track of failure reasons separately from the ciEnv.
4511 // This is required because there is not quite a 1-1 relation between the
4512 // ciEnv and its compilation task and the Compile object. Note that one
4513 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4514 // to backtrack and retry without subsuming loads. Other than this backtracking
4515 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4516 // by the logic in C2Compiler.
4517 void Compile::record_failure(const char* reason) {
4518 if (log() != NULL) {
4519 log()->elem("failure reason='%s' phase='compile'", reason);
4520 }
4521 if (_failure_reason == NULL) {
4522 // Record the first failure reason.
4523 _failure_reason = reason;
4524 }
4525
4526 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4527 C->print_method(PHASE_FAILURE);
4528 }
4529 _root = NULL; // flush the graph, too
4530 }
4531
4532 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4533 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4534 _phase_name(name), _dolog(CITimeVerbose)
4535 {
4536 if (_dolog) {
4537 C = Compile::current();
4538 _log = C->log();
4539 } else {
4540 C = NULL;
4541 _log = NULL;
4542 }
4543 if (_log != NULL) {
4544 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4545 _log->stamp();
4546 _log->end_head();
4547 }
4548 }
4549
4550 Compile::TracePhase::~TracePhase() {
4551
4552 C = Compile::current();
4553 if (_dolog) {
4554 _log = C->log();
4555 } else {
4556 _log = NULL;
4557 }
4558
4559 #ifdef ASSERT
4560 if (PrintIdealNodeCount) {
4561 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4562 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4563 }
4564
4565 if (VerifyIdealNodeCount) {
4566 Compile::current()->print_missing_nodes();
4567 }
4568 #endif
4569
4570 if (_log != NULL) {
4571 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4572 }
4573 }
4574
4575 //----------------------------static_subtype_check-----------------------------
4576 // Shortcut important common cases when superklass is exact:
4577 // (0) superklass is java.lang.Object (can occur in reflective code)
4578 // (1) subklass is already limited to a subtype of superklass => always ok
4579 // (2) subklass does not overlap with superklass => always fail
4580 // (3) superklass has NO subtypes and we can check with a simple compare.
4581 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4582 if (StressReflectiveCode || superk == NULL || subk == NULL) {
4583 return SSC_full_test; // Let caller generate the general case.
4584 }
4585
4586 if (superk == env()->Object_klass()) {
4587 return SSC_always_true; // (0) this test cannot fail
4588 }
4589
4590 ciType* superelem = superk;
4591 ciType* subelem = subk;
4592 if (superelem->is_array_klass()) {
4593 superelem = superelem->as_array_klass()->base_element_type();
4594 }
4595 if (subelem->is_array_klass()) {
4596 subelem = subelem->as_array_klass()->base_element_type();
4597 }
4598
4599 if (!subk->is_interface()) { // cannot trust static interface types yet
4600 if (subk->is_subtype_of(superk)) {
4601 return SSC_always_true; // (1) false path dead; no dynamic test needed
4602 }
4603 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4604 !(subelem->is_klass() && subelem->as_klass()->is_interface()) &&
4605 !superk->is_subtype_of(subk)) {
4606 return SSC_always_false; // (2) true path dead; no dynamic test needed
4607 }
4608 }
4609
4610 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays.
4611 // [QMyValue is a subtype of [LMyValue but the klass for [QMyValue is not equal to
4612 // the klass for [LMyValue. Perform a full test.
4613 if (superk->is_obj_array_klass() && !superk->as_array_klass()->is_elem_null_free() &&
4614 superk->as_array_klass()->element_klass()->is_inlinetype()) {
4615 return SSC_full_test;
4616 }
4617 // If casting to an instance klass, it must have no subtypes
4618 if (superk->is_interface()) {
4619 // Cannot trust interfaces yet.
4620 // %%% S.B. superk->nof_implementors() == 1
4621 } else if (superelem->is_instance_klass()) {
4622 ciInstanceKlass* ik = superelem->as_instance_klass();
4623 if (!ik->has_subklass() && !ik->is_interface()) {
4624 if (!ik->is_final()) {
4625 // Add a dependency if there is a chance of a later subclass.
4626 dependencies()->assert_leaf_type(ik);
4627 }
4628 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4629 }
4630 } else {
4631 // A primitive array type has no subtypes.
4632 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4633 }
4634
4635 return SSC_full_test;
4636 }
4637
4638 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4639 #ifdef _LP64
4640 // The scaled index operand to AddP must be a clean 64-bit value.
4641 // Java allows a 32-bit int to be incremented to a negative
4642 // value, which appears in a 64-bit register as a large
4643 // positive number. Using that large positive number as an
4644 // operand in pointer arithmetic has bad consequences.
4645 // On the other hand, 32-bit overflow is rare, and the possibility
4646 // can often be excluded, if we annotate the ConvI2L node with
4647 // a type assertion that its value is known to be a small positive
4648 // number. (The prior range check has ensured this.)
4649 // This assertion is used by ConvI2LNode::Ideal.
4650 int index_max = max_jint - 1; // array size is max_jint, index is one less
4651 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4652 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4653 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4654 #endif
4655 return idx;
4656 }
4657
4658 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4659 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4660 if (ctrl != NULL) {
4661 // Express control dependency by a CastII node with a narrow type.
4662 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4663 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4664 // node from floating above the range check during loop optimizations. Otherwise, the
4665 // ConvI2L node may be eliminated independently of the range check, causing the data path
4666 // to become TOP while the control path is still there (although it's unreachable).
4667 value->set_req(0, ctrl);
4668 value = phase->transform(value);
4669 }
4670 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4671 return phase->transform(new ConvI2LNode(value, ltype));
4672 }
4673
4674 void Compile::print_inlining_stream_free() {
4675 if (_print_inlining_stream != NULL) {
4676 _print_inlining_stream->~stringStream();
4677 _print_inlining_stream = NULL;
4678 }
4679 }
4680
4681 // The message about the current inlining is accumulated in
4682 // _print_inlining_stream and transfered into the _print_inlining_list
4683 // once we know whether inlining succeeds or not. For regular
4684 // inlining, messages are appended to the buffer pointed by
4685 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4686 // a new buffer is added after _print_inlining_idx in the list. This
4687 // way we can update the inlining message for late inlining call site
4688 // when the inlining is attempted again.
4689 void Compile::print_inlining_init() {
4690 if (print_inlining() || print_intrinsics()) {
4691 // print_inlining_init is actually called several times.
4692 print_inlining_stream_free();
4693 _print_inlining_stream = new stringStream();
4694 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4695 }
4696 }
4697
4698 void Compile::print_inlining_reinit() {
4699 if (print_inlining() || print_intrinsics()) {
4700 print_inlining_stream_free();
4701 // Re allocate buffer when we change ResourceMark
4702 _print_inlining_stream = new stringStream();
4703 }
4704 }
4705
4706 void Compile::print_inlining_reset() {
4707 _print_inlining_stream->reset();
4708 }
4709
4710 void Compile::print_inlining_commit() {
4711 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4712 // Transfer the message from _print_inlining_stream to the current
4713 // _print_inlining_list buffer and clear _print_inlining_stream.
4714 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4715 print_inlining_reset();
4716 }
4717
4718 void Compile::print_inlining_push() {
4719 // Add new buffer to the _print_inlining_list at current position
4720 _print_inlining_idx++;
4721 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4722 }
4723
4724 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4725 return _print_inlining_list->at(_print_inlining_idx);
4726 }
4727
4728 void Compile::print_inlining_update(CallGenerator* cg) {
4729 if (print_inlining() || print_intrinsics()) {
4730 if (cg->is_late_inline()) {
4731 if (print_inlining_current()->cg() != cg &&
4732 (print_inlining_current()->cg() != NULL ||
4733 print_inlining_current()->ss()->size() != 0)) {
4734 print_inlining_push();
4735 }
4736 print_inlining_commit();
4737 print_inlining_current()->set_cg(cg);
4738 } else {
4739 if (print_inlining_current()->cg() != NULL) {
4740 print_inlining_push();
4741 }
4742 print_inlining_commit();
4743 }
4744 }
4745 }
4746
4747 void Compile::print_inlining_move_to(CallGenerator* cg) {
4748 // We resume inlining at a late inlining call site. Locate the
4749 // corresponding inlining buffer so that we can update it.
4750 if (print_inlining() || print_intrinsics()) {
4751 for (int i = 0; i < _print_inlining_list->length(); i++) {
4752 if (_print_inlining_list->at(i)->cg() == cg) {
4753 _print_inlining_idx = i;
4754 return;
4755 }
4756 }
4757 ShouldNotReachHere();
4758 }
4759 }
4760
4761 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4762 if (print_inlining() || print_intrinsics()) {
4763 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4764 assert(print_inlining_current()->cg() == cg, "wrong entry");
4765 // replace message with new message
4766 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4767 print_inlining_commit();
4768 print_inlining_current()->set_cg(cg);
4769 }
4770 }
4771
4772 void Compile::print_inlining_assert_ready() {
4773 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4774 }
4775
4776 void Compile::process_print_inlining() {
4777 assert(_late_inlines.length() == 0, "not drained yet");
4778 if (print_inlining() || print_intrinsics()) {
4779 ResourceMark rm;
4780 stringStream ss;
4781 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
4782 for (int i = 0; i < _print_inlining_list->length(); i++) {
4783 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4784 ss.print("%s", pib->ss()->as_string());
4785 delete pib;
4786 DEBUG_ONLY(_print_inlining_list->at_put(i, NULL));
4787 }
4788 // Reset _print_inlining_list, it only contains destructed objects.
4789 // It is on the arena, so it will be freed when the arena is reset.
4790 _print_inlining_list = NULL;
4791 // _print_inlining_stream won't be used anymore, either.
4792 print_inlining_stream_free();
4793 size_t end = ss.size();
4794 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4795 strncpy(_print_inlining_output, ss.base(), end+1);
4796 _print_inlining_output[end] = 0;
4797 }
4798 }
4799
4800 void Compile::dump_print_inlining() {
4801 if (_print_inlining_output != NULL) {
4802 tty->print_raw(_print_inlining_output);
4803 }
4804 }
4805
4806 void Compile::log_late_inline(CallGenerator* cg) {
4807 if (log() != NULL) {
4808 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4809 cg->unique_id());
4810 JVMState* p = cg->call_node()->jvms();
4811 while (p != NULL) {
4812 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4813 p = p->caller();
4814 }
4815 log()->tail("late_inline");
4816 }
4817 }
4818
4819 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4820 log_late_inline(cg);
4821 if (log() != NULL) {
4822 log()->inline_fail(msg);
4823 }
4824 }
4825
4826 void Compile::log_inline_id(CallGenerator* cg) {
4827 if (log() != NULL) {
4828 // The LogCompilation tool needs a unique way to identify late
4829 // inline call sites. This id must be unique for this call site in
4830 // this compilation. Try to have it unique across compilations as
4831 // well because it can be convenient when grepping through the log
4832 // file.
4833 // Distinguish OSR compilations from others in case CICountOSR is
4834 // on.
4835 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4836 cg->set_unique_id(id);
4837 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4838 }
4839 }
4840
4841 void Compile::log_inline_failure(const char* msg) {
4842 if (C->log() != NULL) {
4843 C->log()->inline_fail(msg);
4844 }
4845 }
4846
4847
4848 // Dump inlining replay data to the stream.
4849 // Don't change thread state and acquire any locks.
4850 void Compile::dump_inline_data(outputStream* out) {
4851 InlineTree* inl_tree = ilt();
4852 if (inl_tree != NULL) {
4853 out->print(" inline %d", inl_tree->count());
4854 inl_tree->dump_replay_data(out);
4855 }
4856 }
4857
4858 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4859 if (n1->Opcode() < n2->Opcode()) return -1;
4860 else if (n1->Opcode() > n2->Opcode()) return 1;
4861
4862 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4863 for (uint i = 1; i < n1->req(); i++) {
4864 if (n1->in(i) < n2->in(i)) return -1;
4865 else if (n1->in(i) > n2->in(i)) return 1;
4866 }
4867
4868 return 0;
4869 }
4870
4871 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4872 Node* n1 = *n1p;
4873 Node* n2 = *n2p;
4874
4875 return cmp_expensive_nodes(n1, n2);
4876 }
4877
4878 void Compile::sort_expensive_nodes() {
4879 if (!expensive_nodes_sorted()) {
4880 _expensive_nodes.sort(cmp_expensive_nodes);
4881 }
4882 }
4883
4884 bool Compile::expensive_nodes_sorted() const {
4885 for (int i = 1; i < _expensive_nodes.length(); i++) {
4886 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4887 return false;
4888 }
4889 }
4890 return true;
4891 }
4892
4893 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4894 if (_expensive_nodes.length() == 0) {
4895 return false;
4896 }
4897
4898 assert(OptimizeExpensiveOps, "optimization off?");
4899
4900 // Take this opportunity to remove dead nodes from the list
4901 int j = 0;
4902 for (int i = 0; i < _expensive_nodes.length(); i++) {
4903 Node* n = _expensive_nodes.at(i);
4904 if (!n->is_unreachable(igvn)) {
4905 assert(n->is_expensive(), "should be expensive");
4906 _expensive_nodes.at_put(j, n);
4907 j++;
4908 }
4909 }
4910 _expensive_nodes.trunc_to(j);
4911
4912 // Then sort the list so that similar nodes are next to each other
4913 // and check for at least two nodes of identical kind with same data
4914 // inputs.
4915 sort_expensive_nodes();
4916
4917 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4918 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4919 return true;
4920 }
4921 }
4922
4923 return false;
4924 }
4925
4926 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4927 if (_expensive_nodes.length() == 0) {
4928 return;
4929 }
4930
4931 assert(OptimizeExpensiveOps, "optimization off?");
4932
4933 // Sort to bring similar nodes next to each other and clear the
4934 // control input of nodes for which there's only a single copy.
4935 sort_expensive_nodes();
4936
4937 int j = 0;
4938 int identical = 0;
4939 int i = 0;
4940 bool modified = false;
4941 for (; i < _expensive_nodes.length()-1; i++) {
4942 assert(j <= i, "can't write beyond current index");
4943 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4944 identical++;
4945 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4946 continue;
4947 }
4948 if (identical > 0) {
4949 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4950 identical = 0;
4951 } else {
4952 Node* n = _expensive_nodes.at(i);
4953 igvn.replace_input_of(n, 0, NULL);
4954 igvn.hash_insert(n);
4955 modified = true;
4956 }
4957 }
4958 if (identical > 0) {
4959 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4960 } else if (_expensive_nodes.length() >= 1) {
4961 Node* n = _expensive_nodes.at(i);
4962 igvn.replace_input_of(n, 0, NULL);
4963 igvn.hash_insert(n);
4964 modified = true;
4965 }
4966 _expensive_nodes.trunc_to(j);
4967 if (modified) {
4968 igvn.optimize();
4969 }
4970 }
4971
4972 void Compile::add_expensive_node(Node * n) {
4973 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4974 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4975 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4976 if (OptimizeExpensiveOps) {
4977 _expensive_nodes.append(n);
4978 } else {
4979 // Clear control input and let IGVN optimize expensive nodes if
4980 // OptimizeExpensiveOps is off.
4981 n->set_req(0, NULL);
4982 }
4983 }
4984
4985 /**
4986 * Track coarsened Lock and Unlock nodes.
4987 */
4988
4989 class Lock_List : public Node_List {
4990 uint _origin_cnt;
4991 public:
4992 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4993 uint origin_cnt() const { return _origin_cnt; }
4994 };
4995
4996 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4997 int length = locks.length();
4998 if (length > 0) {
4999 // Have to keep this list until locks elimination during Macro nodes elimination.
5000 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5001 for (int i = 0; i < length; i++) {
5002 AbstractLockNode* lock = locks.at(i);
5003 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5004 locks_list->push(lock);
5005 }
5006 _coarsened_locks.append(locks_list);
5007 }
5008 }
5009
5010 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5011 int count = coarsened_count();
5012 for (int i = 0; i < count; i++) {
5013 Node_List* locks_list = _coarsened_locks.at(i);
5014 for (uint j = 0; j < locks_list->size(); j++) {
5015 Node* lock = locks_list->at(j);
5016 assert(lock->is_AbstractLock(), "sanity");
5017 if (!useful.member(lock)) {
5018 locks_list->yank(lock);
5019 }
5020 }
5021 }
5022 }
5023
5024 void Compile::remove_coarsened_lock(Node* n) {
5025 if (n->is_AbstractLock()) {
5026 int count = coarsened_count();
5027 for (int i = 0; i < count; i++) {
5028 Node_List* locks_list = _coarsened_locks.at(i);
5029 locks_list->yank(n);
5030 }
5031 }
5032 }
5033
5034 bool Compile::coarsened_locks_consistent() {
5035 int count = coarsened_count();
5036 for (int i = 0; i < count; i++) {
5037 bool unbalanced = false;
5038 bool modified = false; // track locks kind modifications
5039 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5040 uint size = locks_list->size();
5041 if (size == 0) {
5042 unbalanced = false; // All locks were eliminated - good
5043 } else if (size != locks_list->origin_cnt()) {
5044 unbalanced = true; // Some locks were removed from list
5045 } else {
5046 for (uint j = 0; j < size; j++) {
5047 Node* lock = locks_list->at(j);
5048 // All nodes in group should have the same state (modified or not)
5049 if (!lock->as_AbstractLock()->is_coarsened()) {
5050 if (j == 0) {
5051 // first on list was modified, the rest should be too for consistency
5052 modified = true;
5053 } else if (!modified) {
5054 // this lock was modified but previous locks on the list were not
5055 unbalanced = true;
5056 break;
5057 }
5058 } else if (modified) {
5059 // previous locks on list were modified but not this lock
5060 unbalanced = true;
5061 break;
5062 }
5063 }
5064 }
5065 if (unbalanced) {
5066 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5067 #ifdef ASSERT
5068 if (PrintEliminateLocks) {
5069 tty->print_cr("=== unbalanced coarsened locks ===");
5070 for (uint l = 0; l < size; l++) {
5071 locks_list->at(l)->dump();
5072 }
5073 }
5074 #endif
5075 record_failure(C2Compiler::retry_no_locks_coarsening());
5076 return false;
5077 }
5078 }
5079 return true;
5080 }
5081
5082 /**
5083 * Remove the speculative part of types and clean up the graph
5084 */
5085 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5086 if (UseTypeSpeculation) {
5087 Unique_Node_List worklist;
5088 worklist.push(root());
5089 int modified = 0;
5090 // Go over all type nodes that carry a speculative type, drop the
5091 // speculative part of the type and enqueue the node for an igvn
5092 // which may optimize it out.
5093 for (uint next = 0; next < worklist.size(); ++next) {
5094 Node *n = worklist.at(next);
5095 if (n->is_Type()) {
5096 TypeNode* tn = n->as_Type();
5097 const Type* t = tn->type();
5098 const Type* t_no_spec = t->remove_speculative();
5099 if (t_no_spec != t) {
5100 bool in_hash = igvn.hash_delete(n);
5101 assert(in_hash, "node should be in igvn hash table");
5102 tn->set_type(t_no_spec);
5103 igvn.hash_insert(n);
5104 igvn._worklist.push(n); // give it a chance to go away
5105 modified++;
5106 }
5107 }
5108 // Iterate over outs - endless loops is unreachable from below
5109 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5110 Node *m = n->fast_out(i);
5111 if (not_a_node(m)) {
5112 continue;
5113 }
5114 worklist.push(m);
5115 }
5116 }
5117 // Drop the speculative part of all types in the igvn's type table
5118 igvn.remove_speculative_types();
5119 if (modified > 0) {
5120 igvn.optimize();
5121 }
5122 #ifdef ASSERT
5123 // Verify that after the IGVN is over no speculative type has resurfaced
5124 worklist.clear();
5125 worklist.push(root());
5126 for (uint next = 0; next < worklist.size(); ++next) {
5127 Node *n = worklist.at(next);
5128 const Type* t = igvn.type_or_null(n);
5129 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
5130 if (n->is_Type()) {
5131 t = n->as_Type()->type();
5132 assert(t == t->remove_speculative(), "no more speculative types");
5133 }
5134 // Iterate over outs - endless loops is unreachable from below
5135 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5136 Node *m = n->fast_out(i);
5137 if (not_a_node(m)) {
5138 continue;
5139 }
5140 worklist.push(m);
5141 }
5142 }
5143 igvn.check_no_speculative_types();
5144 #endif
5145 }
5146 }
5147
5148 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5149 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5150 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5151 if (!EnableValhalla || ta == NULL || tb == NULL ||
5152 ta->is_zero_type() || tb->is_zero_type() ||
5153 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5154 // Use old acmp if one operand is null or not an inline type
5155 return new CmpPNode(a, b);
5156 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5157 // We know that one operand is an inline type. Therefore,
5158 // new acmp will only return true if both operands are NULL.
5159 // Check if both operands are null by or'ing the oops.
5160 a = phase->transform(new CastP2XNode(NULL, a));
5161 b = phase->transform(new CastP2XNode(NULL, b));
5162 a = phase->transform(new OrXNode(a, b));
5163 return new CmpXNode(a, phase->MakeConX(0));
5164 }
5165 // Use new acmp
5166 return NULL;
5167 }
5168
5169 // Auxiliary methods to support randomized stressing/fuzzing.
5170
5171 int Compile::random() {
5172 _stress_seed = os::next_random(_stress_seed);
5173 return static_cast<int>(_stress_seed);
5174 }
5175
5176 // This method can be called the arbitrary number of times, with current count
5177 // as the argument. The logic allows selecting a single candidate from the
5178 // running list of candidates as follows:
5179 // int count = 0;
5180 // Cand* selected = null;
5181 // while(cand = cand->next()) {
5182 // if (randomized_select(++count)) {
5183 // selected = cand;
5184 // }
5185 // }
5186 //
5187 // Including count equalizes the chances any candidate is "selected".
5188 // This is useful when we don't have the complete list of candidates to choose
5189 // from uniformly. In this case, we need to adjust the randomicity of the
5190 // selection, or else we will end up biasing the selection towards the latter
5191 // candidates.
5192 //
5193 // Quick back-envelope calculation shows that for the list of n candidates
5194 // the equal probability for the candidate to persist as "best" can be
5195 // achieved by replacing it with "next" k-th candidate with the probability
5196 // of 1/k. It can be easily shown that by the end of the run, the
5197 // probability for any candidate is converged to 1/n, thus giving the
5198 // uniform distribution among all the candidates.
5199 //
5200 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5201 #define RANDOMIZED_DOMAIN_POW 29
5202 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5203 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5204 bool Compile::randomized_select(int count) {
5205 assert(count > 0, "only positive");
5206 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5207 }
5208
5209 CloneMap& Compile::clone_map() { return _clone_map; }
5210 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5211
5212 void NodeCloneInfo::dump() const {
5213 tty->print(" {%d:%d} ", idx(), gen());
5214 }
5215
5216 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5217 uint64_t val = value(old->_idx);
5218 NodeCloneInfo cio(val);
5219 assert(val != 0, "old node should be in the map");
5220 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5221 insert(nnn->_idx, cin.get());
5222 #ifndef PRODUCT
5223 if (is_debug()) {
5224 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5225 }
5226 #endif
5227 }
5228
5229 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5230 NodeCloneInfo cio(value(old->_idx));
5231 if (cio.get() == 0) {
5232 cio.set(old->_idx, 0);
5233 insert(old->_idx, cio.get());
5234 #ifndef PRODUCT
5235 if (is_debug()) {
5236 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5237 }
5238 #endif
5239 }
5240 clone(old, nnn, gen);
5241 }
5242
5243 int CloneMap::max_gen() const {
5244 int g = 0;
5245 DictI di(_dict);
5246 for(; di.test(); ++di) {
5247 int t = gen(di._key);
5248 if (g < t) {
5249 g = t;
5250 #ifndef PRODUCT
5251 if (is_debug()) {
5252 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5253 }
5254 #endif
5255 }
5256 }
5257 return g;
5258 }
5259
5260 void CloneMap::dump(node_idx_t key) const {
5261 uint64_t val = value(key);
5262 if (val != 0) {
5263 NodeCloneInfo ni(val);
5264 ni.dump();
5265 }
5266 }
5267
5268 // Move Allocate nodes to the start of the list
5269 void Compile::sort_macro_nodes() {
5270 int count = macro_count();
5271 int allocates = 0;
5272 for (int i = 0; i < count; i++) {
5273 Node* n = macro_node(i);
5274 if (n->is_Allocate()) {
5275 if (i != allocates) {
5276 Node* tmp = macro_node(allocates);
5277 _macro_nodes.at_put(allocates, n);
5278 _macro_nodes.at_put(i, tmp);
5279 }
5280 allocates++;
5281 }
5282 }
5283 }
5284
5285 void Compile::print_method(CompilerPhaseType cpt, const char *name, int level) {
5286 EventCompilerPhase event;
5287 if (event.should_commit()) {
5288 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5289 }
5290 #ifndef PRODUCT
5291 if (should_print(level)) {
5292 _printer->print_method(name, level);
5293 }
5294 #endif
5295 C->_latest_stage_start_counter.stamp();
5296 }
5297
5298 void Compile::print_method(CompilerPhaseType cpt, int level, int idx) {
5299 char output[1024];
5300 #ifndef PRODUCT
5301 if (idx != 0) {
5302 jio_snprintf(output, sizeof(output), "%s:%d", CompilerPhaseTypeHelper::to_string(cpt), idx);
5303 } else {
5304 jio_snprintf(output, sizeof(output), "%s", CompilerPhaseTypeHelper::to_string(cpt));
5305 }
5306 #endif
5307 print_method(cpt, output, level);
5308 }
5309
5310 void Compile::print_method(CompilerPhaseType cpt, Node* n, int level) {
5311 ResourceMark rm;
5312 stringStream ss;
5313 ss.print_raw(CompilerPhaseTypeHelper::to_string(cpt));
5314 if (n != NULL) {
5315 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5316 } else {
5317 ss.print_raw(": NULL");
5318 }
5319 C->print_method(cpt, ss.as_string(), level);
5320 }
5321
5322 void Compile::end_method(int level) {
5323 EventCompilerPhase event;
5324 if (event.should_commit()) {
5325 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, level);
5326 }
5327
5328 #ifndef PRODUCT
5329 if (_method != NULL && should_print(level)) {
5330 _printer->end_method();
5331 }
5332 #endif
5333 }
5334
5335
5336 #ifndef PRODUCT
5337 IdealGraphPrinter* Compile::_debug_file_printer = NULL;
5338 IdealGraphPrinter* Compile::_debug_network_printer = NULL;
5339
5340 // Called from debugger. Prints method to the default file with the default phase name.
5341 // This works regardless of any Ideal Graph Visualizer flags set or not.
5342 void igv_print() {
5343 Compile::current()->igv_print_method_to_file();
5344 }
5345
5346 // Same as igv_print() above but with a specified phase name.
5347 void igv_print(const char* phase_name) {
5348 Compile::current()->igv_print_method_to_file(phase_name);
5349 }
5350
5351 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5352 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5353 // This works regardless of any Ideal Graph Visualizer flags set or not.
5354 void igv_print(bool network) {
5355 if (network) {
5356 Compile::current()->igv_print_method_to_network();
5357 } else {
5358 Compile::current()->igv_print_method_to_file();
5359 }
5360 }
5361
5362 // Same as igv_print(bool network) above but with a specified phase name.
5363 void igv_print(bool network, const char* phase_name) {
5364 if (network) {
5365 Compile::current()->igv_print_method_to_network(phase_name);
5366 } else {
5367 Compile::current()->igv_print_method_to_file(phase_name);
5368 }
5369 }
5370
5371 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5372 void igv_print_default() {
5373 Compile::current()->print_method(PHASE_DEBUG, 0);
5374 }
5375
5376 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5377 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5378 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5379 void igv_append() {
5380 Compile::current()->igv_print_method_to_file("Debug", true);
5381 }
5382
5383 // Same as igv_append() above but with a specified phase name.
5384 void igv_append(const char* phase_name) {
5385 Compile::current()->igv_print_method_to_file(phase_name, true);
5386 }
5387
5388 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5389 const char* file_name = "custom_debug.xml";
5390 if (_debug_file_printer == NULL) {
5391 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5392 } else {
5393 _debug_file_printer->update_compiled_method(C->method());
5394 }
5395 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5396 _debug_file_printer->print(phase_name, (Node*)C->root());
5397 }
5398
5399 void Compile::igv_print_method_to_network(const char* phase_name) {
5400 if (_debug_network_printer == NULL) {
5401 _debug_network_printer = new IdealGraphPrinter(C);
5402 } else {
5403 _debug_network_printer->update_compiled_method(C->method());
5404 }
5405 tty->print_cr("Method printed over network stream to IGV");
5406 _debug_network_printer->print(phase_name, (Node*)C->root());
5407 }
5408 #endif
5409
5410 void Compile::add_native_invoker(RuntimeStub* stub) {
5411 _native_invokers.append(stub);
5412 }
5413
5414 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5415 if (type != NULL && phase->type(value)->higher_equal(type)) {
5416 return value;
5417 }
5418 Node* result = NULL;
5419 if (bt == T_BYTE) {
5420 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5421 result = new RShiftINode(result, phase->intcon(24));
5422 } else if (bt == T_BOOLEAN) {
5423 result = new AndINode(value, phase->intcon(0xFF));
5424 } else if (bt == T_CHAR) {
5425 result = new AndINode(value,phase->intcon(0xFFFF));
5426 } else {
5427 assert(bt == T_SHORT, "unexpected narrow type");
5428 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5429 result = new RShiftINode(result, phase->intcon(16));
5430 }
5431 if (transform_res) {
5432 result = phase->transform(result);
5433 }
5434 return result;
5435 }
5436