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/loopnode.hpp"
56 #include "opto/machnode.hpp"
57 #include "opto/macro.hpp"
58 #include "opto/matcher.hpp"
59 #include "opto/mathexactnode.hpp"
60 #include "opto/memnode.hpp"
61 #include "opto/mulnode.hpp"
62 #include "opto/narrowptrnode.hpp"
63 #include "opto/node.hpp"
64 #include "opto/opcodes.hpp"
65 #include "opto/output.hpp"
66 #include "opto/parse.hpp"
67 #include "opto/phaseX.hpp"
68 #include "opto/rootnode.hpp"
69 #include "opto/runtime.hpp"
70 #include "opto/stringopts.hpp"
71 #include "opto/type.hpp"
72 #include "opto/vector.hpp"
73 #include "opto/vectornode.hpp"
74 #include "runtime/globals_extension.hpp"
75 #include "runtime/sharedRuntime.hpp"
76 #include "runtime/signature.hpp"
77 #include "runtime/stubRoutines.hpp"
78 #include "runtime/timer.hpp"
79 #include "utilities/align.hpp"
80 #include "utilities/copy.hpp"
81 #include "utilities/macros.hpp"
82 #include "utilities/resourceHash.hpp"
83
84
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == NULL) {
89 _mach_constant_base_node = new MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
94
95
96 /// Support for intrinsics.
97
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 class IntrinsicDescPair {
101 private:
102 ciMethod* _m;
103 bool _is_virtual;
104 public:
105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
107 ciMethod* m= elt->method();
108 ciMethod* key_m = key->_m;
109 if (key_m < m) return -1;
110 else if (key_m > m) return 1;
111 else {
112 bool is_virtual = elt->is_virtual();
113 bool key_virtual = key->_is_virtual;
114 if (key_virtual < is_virtual) return -1;
115 else if (key_virtual > is_virtual) return 1;
116 else return 0;
117 }
118 }
119 };
120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
121 #ifdef ASSERT
122 for (int i = 1; i < _intrinsics.length(); i++) {
123 CallGenerator* cg1 = _intrinsics.at(i-1);
124 CallGenerator* cg2 = _intrinsics.at(i);
125 assert(cg1->method() != cg2->method()
126 ? cg1->method() < cg2->method()
127 : cg1->is_virtual() < cg2->is_virtual(),
128 "compiler intrinsics list must stay sorted");
129 }
130 #endif
131 IntrinsicDescPair pair(m, is_virtual);
132 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
133 }
134
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 bool found = false;
137 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
138 assert(!found, "registering twice");
139 _intrinsics.insert_before(index, cg);
140 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
141 }
142
143 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
144 assert(m->is_loaded(), "don't try this on unloaded methods");
145 if (_intrinsics.length() > 0) {
146 bool found = false;
147 int index = intrinsic_insertion_index(m, is_virtual, found);
148 if (found) {
149 return _intrinsics.at(index);
150 }
151 }
152 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
153 if (m->intrinsic_id() != vmIntrinsics::_none &&
154 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
155 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
156 if (cg != NULL) {
157 // Save it for next time:
158 register_intrinsic(cg);
159 return cg;
160 } else {
161 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
162 }
163 }
164 return NULL;
165 }
166
167 // Compile::make_vm_intrinsic is defined in library_call.cpp.
168
169 #ifndef PRODUCT
170 // statistics gathering...
171
172 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
174
175 inline int as_int(vmIntrinsics::ID id) {
176 return vmIntrinsics::as_int(id);
177 }
178
179 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
180 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
181 int oflags = _intrinsic_hist_flags[as_int(id)];
182 assert(flags != 0, "what happened?");
183 if (is_virtual) {
184 flags |= _intrinsic_virtual;
185 }
186 bool changed = (flags != oflags);
187 if ((flags & _intrinsic_worked) != 0) {
188 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
189 if (count == 1) {
190 changed = true; // first time
191 }
192 // increment the overall count also:
193 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
194 }
195 if (changed) {
196 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
197 // Something changed about the intrinsic's virtuality.
198 if ((flags & _intrinsic_virtual) != 0) {
199 // This is the first use of this intrinsic as a virtual call.
200 if (oflags != 0) {
201 // We already saw it as a non-virtual, so note both cases.
202 flags |= _intrinsic_both;
203 }
204 } else if ((oflags & _intrinsic_both) == 0) {
205 // This is the first use of this intrinsic as a non-virtual
206 flags |= _intrinsic_both;
207 }
208 }
209 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
210 }
211 // update the overall flags also:
212 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
213 return changed;
214 }
215
216 static char* format_flags(int flags, char* buf) {
217 buf[0] = 0;
218 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
219 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
220 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
221 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
222 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
223 if (buf[0] == 0) strcat(buf, ",");
224 assert(buf[0] == ',', "must be");
225 return &buf[1];
226 }
227
228 void Compile::print_intrinsic_statistics() {
229 char flagsbuf[100];
230 ttyLocker ttyl;
231 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
232 tty->print_cr("Compiler intrinsic usage:");
233 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
234 if (total == 0) total = 1; // avoid div0 in case of no successes
235 #define PRINT_STAT_LINE(name, c, f) \
236 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
237 for (auto id : EnumRange<vmIntrinsicID>{}) {
238 int flags = _intrinsic_hist_flags[as_int(id)];
239 juint count = _intrinsic_hist_count[as_int(id)];
240 if ((flags | count) != 0) {
241 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
242 }
243 }
244 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
245 if (xtty != NULL) xtty->tail("statistics");
246 }
247
248 void Compile::print_statistics() {
249 { ttyLocker ttyl;
250 if (xtty != NULL) xtty->head("statistics type='opto'");
251 Parse::print_statistics();
252 PhaseCCP::print_statistics();
253 PhaseRegAlloc::print_statistics();
254 PhaseOutput::print_statistics();
255 PhasePeephole::print_statistics();
256 PhaseIdealLoop::print_statistics();
257 if (xtty != NULL) xtty->tail("statistics");
258 }
259 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
260 // put this under its own <statistics> element.
261 print_intrinsic_statistics();
262 }
263 }
264 #endif //PRODUCT
265
266 void Compile::gvn_replace_by(Node* n, Node* nn) {
267 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
268 Node* use = n->last_out(i);
269 bool is_in_table = initial_gvn()->hash_delete(use);
270 uint uses_found = 0;
271 for (uint j = 0; j < use->len(); j++) {
272 if (use->in(j) == n) {
273 if (j < use->req())
274 use->set_req(j, nn);
275 else
276 use->set_prec(j, nn);
277 uses_found++;
278 }
279 }
280 if (is_in_table) {
281 // reinsert into table
282 initial_gvn()->hash_find_insert(use);
283 }
284 record_for_igvn(use);
285 i -= uses_found; // we deleted 1 or more copies of this edge
286 }
287 }
288
289
290 // Identify all nodes that are reachable from below, useful.
291 // Use breadth-first pass that records state in a Unique_Node_List,
292 // recursive traversal is slower.
293 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
294 int estimated_worklist_size = live_nodes();
295 useful.map( estimated_worklist_size, NULL ); // preallocate space
296
297 // Initialize worklist
298 if (root() != NULL) { useful.push(root()); }
299 // If 'top' is cached, declare it useful to preserve cached node
300 if( cached_top_node() ) { useful.push(cached_top_node()); }
301
302 // Push all useful nodes onto the list, breadthfirst
303 for( uint next = 0; next < useful.size(); ++next ) {
304 assert( next < unique(), "Unique useful nodes < total nodes");
305 Node *n = useful.at(next);
306 uint max = n->len();
307 for( uint i = 0; i < max; ++i ) {
308 Node *m = n->in(i);
309 if (not_a_node(m)) continue;
310 useful.push(m);
311 }
312 }
313 }
314
315 // Update dead_node_list with any missing dead nodes using useful
316 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
317 void Compile::update_dead_node_list(Unique_Node_List &useful) {
318 uint max_idx = unique();
319 VectorSet& useful_node_set = useful.member_set();
320
321 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
322 // If node with index node_idx is not in useful set,
323 // mark it as dead in dead node list.
324 if (!useful_node_set.test(node_idx)) {
325 record_dead_node(node_idx);
326 }
327 }
328 }
329
330 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
331 int shift = 0;
332 for (int i = 0; i < inlines->length(); i++) {
333 CallGenerator* cg = inlines->at(i);
334 if (useful.member(cg->call_node())) {
335 if (shift > 0) {
336 inlines->at_put(i - shift, cg);
337 }
338 } else {
339 shift++; // skip over the dead element
340 }
341 }
342 if (shift > 0) {
343 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
344 }
345 }
346
347 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
348 assert(dead != NULL && dead->is_Call(), "sanity");
349 int found = 0;
350 for (int i = 0; i < inlines->length(); i++) {
351 if (inlines->at(i)->call_node() == dead) {
352 inlines->remove_at(i);
353 found++;
354 NOT_DEBUG( break; ) // elements are unique, so exit early
355 }
356 }
357 assert(found <= 1, "not unique");
358 }
359
360 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) {
361 for (int i = node_list.length() - 1; i >= 0; i--) {
362 Node* n = node_list.at(i);
363 if (!useful.member(n)) {
364 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
365 }
366 }
367 }
368
369 void Compile::remove_useless_node(Node* dead) {
370 remove_modified_node(dead);
371
372 // Constant node that has no out-edges and has only one in-edge from
373 // root is usually dead. However, sometimes reshaping walk makes
374 // it reachable by adding use edges. So, we will NOT count Con nodes
375 // as dead to be conservative about the dead node count at any
376 // given time.
377 if (!dead->is_Con()) {
378 record_dead_node(dead->_idx);
379 }
380 if (dead->is_macro()) {
381 remove_macro_node(dead);
382 }
383 if (dead->is_expensive()) {
384 remove_expensive_node(dead);
385 }
386 if (dead->Opcode() == Op_Opaque4) {
387 remove_skeleton_predicate_opaq(dead);
388 }
389 if (dead->for_post_loop_opts_igvn()) {
390 remove_from_post_loop_opts_igvn(dead);
391 }
392 if (dead->is_Call()) {
393 remove_useless_late_inlines( &_late_inlines, dead);
394 remove_useless_late_inlines( &_string_late_inlines, dead);
395 remove_useless_late_inlines( &_boxing_late_inlines, dead);
396 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
397 }
398 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
399 bs->unregister_potential_barrier_node(dead);
400 }
401
402 // Disconnect all useless nodes by disconnecting those at the boundary.
403 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
404 uint next = 0;
405 while (next < useful.size()) {
406 Node *n = useful.at(next++);
407 if (n->is_SafePoint()) {
408 // We're done with a parsing phase. Replaced nodes are not valid
409 // beyond that point.
410 n->as_SafePoint()->delete_replaced_nodes();
411 }
412 // Use raw traversal of out edges since this code removes out edges
413 int max = n->outcnt();
414 for (int j = 0; j < max; ++j) {
415 Node* child = n->raw_out(j);
416 if (!useful.member(child)) {
417 assert(!child->is_top() || child != top(),
418 "If top is cached in Compile object it is in useful list");
419 // Only need to remove this out-edge to the useless node
420 n->raw_del_out(j);
421 --j;
422 --max;
423 }
424 }
425 if (n->outcnt() == 1 && n->has_special_unique_user()) {
426 record_for_igvn(n->unique_out());
427 }
428 }
429
430 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
431 remove_useless_nodes(_predicate_opaqs, useful); // remove useless predicate opaque nodes
432 remove_useless_nodes(_skeleton_predicate_opaqs, useful);
433 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
434 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
435 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
436
437 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
438 bs->eliminate_useless_gc_barriers(useful, this);
439 // clean up the late inline lists
440 remove_useless_late_inlines( &_late_inlines, useful);
441 remove_useless_late_inlines( &_string_late_inlines, useful);
442 remove_useless_late_inlines( &_boxing_late_inlines, useful);
443 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
444 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
445 }
446
447 // ============================================================================
448 //------------------------------CompileWrapper---------------------------------
449 class CompileWrapper : public StackObj {
450 Compile *const _compile;
451 public:
452 CompileWrapper(Compile* compile);
453
454 ~CompileWrapper();
455 };
456
457 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
458 // the Compile* pointer is stored in the current ciEnv:
459 ciEnv* env = compile->env();
460 assert(env == ciEnv::current(), "must already be a ciEnv active");
461 assert(env->compiler_data() == NULL, "compile already active?");
462 env->set_compiler_data(compile);
463 assert(compile == Compile::current(), "sanity");
464
465 compile->set_type_dict(NULL);
466 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
467 compile->clone_map().set_clone_idx(0);
468 compile->set_type_last_size(0);
469 compile->set_last_tf(NULL, NULL);
470 compile->set_indexSet_arena(NULL);
471 compile->set_indexSet_free_block_list(NULL);
472 compile->init_type_arena();
473 Type::Initialize(compile);
474 _compile->begin_method();
475 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
476 }
477 CompileWrapper::~CompileWrapper() {
478 // simulate crash during compilation
479 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
480
481 _compile->end_method();
482 _compile->env()->set_compiler_data(NULL);
483 }
484
485
486 //----------------------------print_compile_messages---------------------------
487 void Compile::print_compile_messages() {
488 #ifndef PRODUCT
489 // Check if recompiling
490 if (_subsume_loads == false && PrintOpto) {
491 // Recompiling without allowing machine instructions to subsume loads
492 tty->print_cr("*********************************************************");
493 tty->print_cr("** Bailout: Recompile without subsuming loads **");
494 tty->print_cr("*********************************************************");
495 }
496 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
497 // Recompiling without escape analysis
498 tty->print_cr("*********************************************************");
499 tty->print_cr("** Bailout: Recompile without escape analysis **");
500 tty->print_cr("*********************************************************");
501 }
502 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
503 // Recompiling without boxing elimination
504 tty->print_cr("*********************************************************");
505 tty->print_cr("** Bailout: Recompile without boxing elimination **");
506 tty->print_cr("*********************************************************");
507 }
508 if ((_do_locks_coarsening != EliminateLocks) && PrintOpto) {
509 // Recompiling without locks coarsening
510 tty->print_cr("*********************************************************");
511 tty->print_cr("** Bailout: Recompile without locks coarsening **");
512 tty->print_cr("*********************************************************");
513 }
514 if (env()->break_at_compile()) {
515 // Open the debugger when compiling this method.
516 tty->print("### Breaking when compiling: ");
517 method()->print_short_name();
518 tty->cr();
519 BREAKPOINT;
520 }
521
522 if( PrintOpto ) {
523 if (is_osr_compilation()) {
524 tty->print("[OSR]%3d", _compile_id);
525 } else {
526 tty->print("%3d", _compile_id);
527 }
528 }
529 #endif
530 }
531
532 // ============================================================================
533 //------------------------------Compile standard-------------------------------
534 debug_only( int Compile::_debug_idx = 100000; )
535
536 // Compile a method. entry_bci is -1 for normal compilations and indicates
537 // the continuation bci for on stack replacement.
538
539
540 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
541 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing,
542 bool do_locks_coarsening, bool install_code, DirectiveSet* directive)
543 : Phase(Compiler),
544 _compile_id(ci_env->compile_id()),
545 _subsume_loads(subsume_loads),
546 _do_escape_analysis(do_escape_analysis),
547 _install_code(install_code),
548 _eliminate_boxing(eliminate_boxing),
549 _do_locks_coarsening(do_locks_coarsening),
550 _method(target),
551 _entry_bci(osr_bci),
552 _stub_function(NULL),
553 _stub_name(NULL),
554 _stub_entry_point(NULL),
555 _max_node_limit(MaxNodeLimit),
556 _post_loop_opts_phase(false),
557 _inlining_progress(false),
558 _inlining_incrementally(false),
559 _do_cleanup(false),
560 _has_reserved_stack_access(target->has_reserved_stack_access()),
561 #ifndef PRODUCT
562 _igv_idx(0),
563 _trace_opto_output(directive->TraceOptoOutputOption),
564 _print_ideal(directive->PrintIdealOption),
565 #endif
566 _has_method_handle_invokes(false),
567 _clinit_barrier_on_entry(false),
568 _stress_seed(0),
569 _comp_arena(mtCompiler),
570 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
571 _env(ci_env),
572 _directive(directive),
573 _log(ci_env->log()),
574 _failure_reason(NULL),
575 _intrinsics (comp_arena(), 0, 0, NULL),
576 _macro_nodes (comp_arena(), 8, 0, NULL),
577 _predicate_opaqs (comp_arena(), 8, 0, NULL),
578 _skeleton_predicate_opaqs (comp_arena(), 8, 0, NULL),
579 _expensive_nodes (comp_arena(), 8, 0, NULL),
580 _for_post_loop_igvn(comp_arena(), 8, 0, NULL),
581 _coarsened_locks (comp_arena(), 8, 0, NULL),
582 _congraph(NULL),
583 NOT_PRODUCT(_printer(NULL) COMMA)
584 _dead_node_list(comp_arena()),
585 _dead_node_count(0),
586 _node_arena(mtCompiler),
587 _old_arena(mtCompiler),
588 _mach_constant_base_node(NULL),
589 _Compile_types(mtCompiler),
590 _initial_gvn(NULL),
591 _for_igvn(NULL),
592 _late_inlines(comp_arena(), 2, 0, NULL),
593 _string_late_inlines(comp_arena(), 2, 0, NULL),
594 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
595 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL),
596 _late_inlines_pos(0),
597 _number_of_mh_late_inlines(0),
598 _native_invokers(comp_arena(), 1, 0, NULL),
599 _print_inlining_stream(NULL),
600 _print_inlining_list(NULL),
601 _print_inlining_idx(0),
602 _print_inlining_output(NULL),
603 _replay_inline_data(NULL),
604 _java_calls(0),
605 _inner_loops(0),
606 _interpreter_frame_size(0)
607 #ifndef PRODUCT
608 , _in_dump_cnt(0)
609 #endif
610 {
611 C = this;
612 CompileWrapper cw(this);
613
614 if (CITimeVerbose) {
615 tty->print(" ");
616 target->holder()->name()->print();
617 tty->print(".");
618 target->print_short_name();
619 tty->print(" ");
620 }
621 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
622 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
623
624 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
625 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
626 // We can always print a disassembly, either abstract (hex dump) or
627 // with the help of a suitable hsdis library. Thus, we should not
628 // couple print_assembly and print_opto_assembly controls.
629 // But: always print opto and regular assembly on compile command 'print'.
630 bool print_assembly = directive->PrintAssemblyOption;
631 set_print_assembly(print_opto_assembly || print_assembly);
632 #else
633 set_print_assembly(false); // must initialize.
634 #endif
635
636 #ifndef PRODUCT
637 set_parsed_irreducible_loop(false);
638
639 if (directive->ReplayInlineOption) {
640 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
641 }
642 #endif
643 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
644 set_print_intrinsics(directive->PrintIntrinsicsOption);
645 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
646
647 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
648 // Make sure the method being compiled gets its own MDO,
649 // so we can at least track the decompile_count().
650 // Need MDO to record RTM code generation state.
651 method()->ensure_method_data();
652 }
653
654 Init(::AliasLevel);
655
656
657 print_compile_messages();
658
659 _ilt = InlineTree::build_inline_tree_root();
660
661 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
662 assert(num_alias_types() >= AliasIdxRaw, "");
663
664 #define MINIMUM_NODE_HASH 1023
665 // Node list that Iterative GVN will start with
666 Unique_Node_List for_igvn(comp_arena());
667 set_for_igvn(&for_igvn);
668
669 // GVN that will be run immediately on new nodes
670 uint estimated_size = method()->code_size()*4+64;
671 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
672 PhaseGVN gvn(node_arena(), estimated_size);
673 set_initial_gvn(&gvn);
674
675 print_inlining_init();
676 { // Scope for timing the parser
677 TracePhase tp("parse", &timers[_t_parser]);
678
679 // Put top into the hash table ASAP.
680 initial_gvn()->transform_no_reclaim(top());
681
682 // Set up tf(), start(), and find a CallGenerator.
683 CallGenerator* cg = NULL;
684 if (is_osr_compilation()) {
685 const TypeTuple *domain = StartOSRNode::osr_domain();
686 const TypeTuple *range = TypeTuple::make_range(method()->signature());
687 init_tf(TypeFunc::make(domain, range));
688 StartNode* s = new StartOSRNode(root(), domain);
689 initial_gvn()->set_type_bottom(s);
690 init_start(s);
691 cg = CallGenerator::for_osr(method(), entry_bci());
692 } else {
693 // Normal case.
694 init_tf(TypeFunc::make(method()));
695 StartNode* s = new StartNode(root(), tf()->domain());
696 initial_gvn()->set_type_bottom(s);
697 init_start(s);
698 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
699 // With java.lang.ref.reference.get() we must go through the
700 // intrinsic - even when get() is the root
701 // method of the compile - so that, if necessary, the value in
702 // the referent field of the reference object gets recorded by
703 // the pre-barrier code.
704 cg = find_intrinsic(method(), false);
705 }
706 if (cg == NULL) {
707 float past_uses = method()->interpreter_invocation_count();
708 float expected_uses = past_uses;
709 cg = CallGenerator::for_inline(method(), expected_uses);
710 }
711 }
712 if (failing()) return;
713 if (cg == NULL) {
714 record_method_not_compilable("cannot parse method");
715 return;
716 }
717 JVMState* jvms = build_start_state(start(), tf());
718 if ((jvms = cg->generate(jvms)) == NULL) {
719 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
720 record_method_not_compilable("method parse failed");
721 }
722 return;
723 }
724 GraphKit kit(jvms);
725
726 if (!kit.stopped()) {
727 // Accept return values, and transfer control we know not where.
728 // This is done by a special, unique ReturnNode bound to root.
729 return_values(kit.jvms());
730 }
731
732 if (kit.has_exceptions()) {
733 // Any exceptions that escape from this call must be rethrown
734 // to whatever caller is dynamically above us on the stack.
735 // This is done by a special, unique RethrowNode bound to root.
736 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
737 }
738
739 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
740
741 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
742 inline_string_calls(true);
743 }
744
745 if (failing()) return;
746
747 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
748
749 // Remove clutter produced by parsing.
750 if (!failing()) {
751 ResourceMark rm;
752 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
753 }
754 }
755
756 // Note: Large methods are capped off in do_one_bytecode().
757 if (failing()) return;
758
759 // After parsing, node notes are no longer automagic.
760 // They must be propagated by register_new_node_with_optimizer(),
761 // clone(), or the like.
762 set_default_node_notes(NULL);
763
764 #ifndef PRODUCT
765 if (should_print(1)) {
766 _printer->print_inlining();
767 }
768 #endif
769
770 if (failing()) return;
771 NOT_PRODUCT( verify_graph_edges(); )
772
773 // If any phase is randomized for stress testing, seed random number
774 // generation and log the seed for repeatability.
775 if (StressLCM || StressGCM || StressIGVN || StressCCP) {
776 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) {
777 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
778 FLAG_SET_ERGO(StressSeed, _stress_seed);
779 } else {
780 _stress_seed = StressSeed;
781 }
782 if (_log != NULL) {
783 _log->elem("stress_test seed='%u'", _stress_seed);
784 }
785 }
786
787 // Now optimize
788 Optimize();
789 if (failing()) return;
790 NOT_PRODUCT( verify_graph_edges(); )
791
792 #ifndef PRODUCT
793 if (print_ideal()) {
794 ttyLocker ttyl; // keep the following output all in one block
795 // This output goes directly to the tty, not the compiler log.
796 // To enable tools to match it up with the compilation activity,
797 // be sure to tag this tty output with the compile ID.
798 if (xtty != NULL) {
799 xtty->head("ideal compile_id='%d'%s", compile_id(),
800 is_osr_compilation() ? " compile_kind='osr'" :
801 "");
802 }
803 root()->dump(9999);
804 if (xtty != NULL) {
805 xtty->tail("ideal");
806 }
807 }
808 #endif
809
810 #ifdef ASSERT
811 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
812 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
813 #endif
814
815 // Dump compilation data to replay it.
816 if (directive->DumpReplayOption) {
817 env()->dump_replay_data(_compile_id);
818 }
819 if (directive->DumpInlineOption && (ilt() != NULL)) {
820 env()->dump_inline_data(_compile_id);
821 }
822
823 // Now that we know the size of all the monitors we can add a fixed slot
824 // for the original deopt pc.
825 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
826 set_fixed_slots(next_slot);
827
828 // Compute when to use implicit null checks. Used by matching trap based
829 // nodes and NullCheck optimization.
830 set_allowed_deopt_reasons();
831
832 // Now generate code
833 Code_Gen();
834 }
835
836 //------------------------------Compile----------------------------------------
837 // Compile a runtime stub
838 Compile::Compile( ciEnv* ci_env,
839 TypeFunc_generator generator,
840 address stub_function,
841 const char *stub_name,
842 int is_fancy_jump,
843 bool pass_tls,
844 bool return_pc,
845 DirectiveSet* directive)
846 : Phase(Compiler),
847 _compile_id(0),
848 _subsume_loads(true),
849 _do_escape_analysis(false),
850 _install_code(true),
851 _eliminate_boxing(false),
852 _do_locks_coarsening(false),
853 _method(NULL),
854 _entry_bci(InvocationEntryBci),
855 _stub_function(stub_function),
856 _stub_name(stub_name),
857 _stub_entry_point(NULL),
858 _max_node_limit(MaxNodeLimit),
859 _post_loop_opts_phase(false),
860 _inlining_progress(false),
861 _inlining_incrementally(false),
862 _has_reserved_stack_access(false),
863 #ifndef PRODUCT
864 _igv_idx(0),
865 _trace_opto_output(directive->TraceOptoOutputOption),
866 _print_ideal(directive->PrintIdealOption),
867 #endif
868 _has_method_handle_invokes(false),
869 _clinit_barrier_on_entry(false),
870 _stress_seed(0),
871 _comp_arena(mtCompiler),
872 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
873 _env(ci_env),
874 _directive(directive),
875 _log(ci_env->log()),
876 _failure_reason(NULL),
877 _congraph(NULL),
878 NOT_PRODUCT(_printer(NULL) COMMA)
879 _dead_node_list(comp_arena()),
880 _dead_node_count(0),
881 _node_arena(mtCompiler),
882 _old_arena(mtCompiler),
883 _mach_constant_base_node(NULL),
884 _Compile_types(mtCompiler),
885 _initial_gvn(NULL),
886 _for_igvn(NULL),
887 _number_of_mh_late_inlines(0),
888 _native_invokers(),
889 _print_inlining_stream(NULL),
890 _print_inlining_list(NULL),
891 _print_inlining_idx(0),
892 _print_inlining_output(NULL),
893 _replay_inline_data(NULL),
894 _java_calls(0),
895 _inner_loops(0),
896 _interpreter_frame_size(0),
897 #ifndef PRODUCT
898 _in_dump_cnt(0),
899 #endif
900 _allowed_reasons(0) {
901 C = this;
902
903 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
904 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
905
906 #ifndef PRODUCT
907 set_print_assembly(PrintFrameConverterAssembly);
908 set_parsed_irreducible_loop(false);
909 #else
910 set_print_assembly(false); // Must initialize.
911 #endif
912 set_has_irreducible_loop(false); // no loops
913
914 CompileWrapper cw(this);
915 Init(/*AliasLevel=*/ 0);
916 init_tf((*generator)());
917
918 {
919 // The following is a dummy for the sake of GraphKit::gen_stub
920 Unique_Node_List for_igvn(comp_arena());
921 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
922 PhaseGVN gvn(Thread::current()->resource_area(),255);
923 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
924 gvn.transform_no_reclaim(top());
925
926 GraphKit kit;
927 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
928 }
929
930 NOT_PRODUCT( verify_graph_edges(); )
931
932 Code_Gen();
933 }
934
935 //------------------------------Init-------------------------------------------
936 // Prepare for a single compilation
937 void Compile::Init(int aliaslevel) {
938 _unique = 0;
939 _regalloc = NULL;
940
941 _tf = NULL; // filled in later
942 _top = NULL; // cached later
943 _matcher = NULL; // filled in later
944 _cfg = NULL; // filled in later
945
946 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
947
948 _node_note_array = NULL;
949 _default_node_notes = NULL;
950 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
951
952 _immutable_memory = NULL; // filled in at first inquiry
953
954 // Globally visible Nodes
955 // First set TOP to NULL to give safe behavior during creation of RootNode
956 set_cached_top_node(NULL);
957 set_root(new RootNode());
958 // Now that you have a Root to point to, create the real TOP
959 set_cached_top_node( new ConNode(Type::TOP) );
960 set_recent_alloc(NULL, NULL);
961
962 // Create Debug Information Recorder to record scopes, oopmaps, etc.
963 env()->set_oop_recorder(new OopRecorder(env()->arena()));
964 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
965 env()->set_dependencies(new Dependencies(env()));
966
967 _fixed_slots = 0;
968 set_has_split_ifs(false);
969 set_has_loops(false); // first approximation
970 set_has_stringbuilder(false);
971 set_has_boxed_value(false);
972 _trap_can_recompile = false; // no traps emitted yet
973 _major_progress = true; // start out assuming good things will happen
974 set_has_unsafe_access(false);
975 set_max_vector_size(0);
976 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
977 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
978 set_decompile_count(0);
979
980 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
981 _loop_opts_cnt = LoopOptsCount;
982 set_do_inlining(Inline);
983 set_max_inline_size(MaxInlineSize);
984 set_freq_inline_size(FreqInlineSize);
985 set_do_scheduling(OptoScheduling);
986
987 set_do_vector_loop(false);
988
989 if (AllowVectorizeOnDemand) {
990 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
991 set_do_vector_loop(true);
992 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());})
993 } else if (has_method() && method()->name() != 0 &&
994 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
995 set_do_vector_loop(true);
996 }
997 }
998 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
999 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());})
1000
1001 set_age_code(has_method() && method()->profile_aging());
1002 set_rtm_state(NoRTM); // No RTM lock eliding by default
1003 _max_node_limit = _directive->MaxNodeLimitOption;
1004
1005 #if INCLUDE_RTM_OPT
1006 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1007 int rtm_state = method()->method_data()->rtm_state();
1008 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1009 // Don't generate RTM lock eliding code.
1010 set_rtm_state(NoRTM);
1011 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1012 // Generate RTM lock eliding code without abort ratio calculation code.
1013 set_rtm_state(UseRTM);
1014 } else if (UseRTMDeopt) {
1015 // Generate RTM lock eliding code and include abort ratio calculation
1016 // code if UseRTMDeopt is on.
1017 set_rtm_state(ProfileRTM);
1018 }
1019 }
1020 #endif
1021 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1022 set_clinit_barrier_on_entry(true);
1023 }
1024 if (debug_info()->recording_non_safepoints()) {
1025 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1026 (comp_arena(), 8, 0, NULL));
1027 set_default_node_notes(Node_Notes::make(this));
1028 }
1029
1030 // // -- Initialize types before each compile --
1031 // // Update cached type information
1032 // if( _method && _method->constants() )
1033 // Type::update_loaded_types(_method, _method->constants());
1034
1035 // Init alias_type map.
1036 if (!_do_escape_analysis && aliaslevel == 3)
1037 aliaslevel = 2; // No unique types without escape analysis
1038 _AliasLevel = aliaslevel;
1039 const int grow_ats = 16;
1040 _max_alias_types = grow_ats;
1041 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1042 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1043 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1044 {
1045 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1046 }
1047 // Initialize the first few types.
1048 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1049 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1050 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1051 _num_alias_types = AliasIdxRaw+1;
1052 // Zero out the alias type cache.
1053 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1054 // A NULL adr_type hits in the cache right away. Preload the right answer.
1055 probe_alias_cache(NULL)->_index = AliasIdxTop;
1056
1057 #ifdef ASSERT
1058 _type_verify_symmetry = true;
1059 _phase_optimize_finished = false;
1060 _exception_backedge = false;
1061 #endif
1062 }
1063
1064 //---------------------------init_start----------------------------------------
1065 // Install the StartNode on this compile object.
1066 void Compile::init_start(StartNode* s) {
1067 if (failing())
1068 return; // already failing
1069 assert(s == start(), "");
1070 }
1071
1072 /**
1073 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1074 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1075 * the ideal graph.
1076 */
1077 StartNode* Compile::start() const {
1078 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1079 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1080 Node* start = root()->fast_out(i);
1081 if (start->is_Start()) {
1082 return start->as_Start();
1083 }
1084 }
1085 fatal("Did not find Start node!");
1086 return NULL;
1087 }
1088
1089 //-------------------------------immutable_memory-------------------------------------
1090 // Access immutable memory
1091 Node* Compile::immutable_memory() {
1092 if (_immutable_memory != NULL) {
1093 return _immutable_memory;
1094 }
1095 StartNode* s = start();
1096 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1097 Node *p = s->fast_out(i);
1098 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1099 _immutable_memory = p;
1100 return _immutable_memory;
1101 }
1102 }
1103 ShouldNotReachHere();
1104 return NULL;
1105 }
1106
1107 //----------------------set_cached_top_node------------------------------------
1108 // Install the cached top node, and make sure Node::is_top works correctly.
1109 void Compile::set_cached_top_node(Node* tn) {
1110 if (tn != NULL) verify_top(tn);
1111 Node* old_top = _top;
1112 _top = tn;
1113 // Calling Node::setup_is_top allows the nodes the chance to adjust
1114 // their _out arrays.
1115 if (_top != NULL) _top->setup_is_top();
1116 if (old_top != NULL) old_top->setup_is_top();
1117 assert(_top == NULL || top()->is_top(), "");
1118 }
1119
1120 #ifdef ASSERT
1121 uint Compile::count_live_nodes_by_graph_walk() {
1122 Unique_Node_List useful(comp_arena());
1123 // Get useful node list by walking the graph.
1124 identify_useful_nodes(useful);
1125 return useful.size();
1126 }
1127
1128 void Compile::print_missing_nodes() {
1129
1130 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1131 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1132 return;
1133 }
1134
1135 // This is an expensive function. It is executed only when the user
1136 // specifies VerifyIdealNodeCount option or otherwise knows the
1137 // additional work that needs to be done to identify reachable nodes
1138 // by walking the flow graph and find the missing ones using
1139 // _dead_node_list.
1140
1141 Unique_Node_List useful(comp_arena());
1142 // Get useful node list by walking the graph.
1143 identify_useful_nodes(useful);
1144
1145 uint l_nodes = C->live_nodes();
1146 uint l_nodes_by_walk = useful.size();
1147
1148 if (l_nodes != l_nodes_by_walk) {
1149 if (_log != NULL) {
1150 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1151 _log->stamp();
1152 _log->end_head();
1153 }
1154 VectorSet& useful_member_set = useful.member_set();
1155 int last_idx = l_nodes_by_walk;
1156 for (int i = 0; i < last_idx; i++) {
1157 if (useful_member_set.test(i)) {
1158 if (_dead_node_list.test(i)) {
1159 if (_log != NULL) {
1160 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1161 }
1162 if (PrintIdealNodeCount) {
1163 // Print the log message to tty
1164 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1165 useful.at(i)->dump();
1166 }
1167 }
1168 }
1169 else if (! _dead_node_list.test(i)) {
1170 if (_log != NULL) {
1171 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1172 }
1173 if (PrintIdealNodeCount) {
1174 // Print the log message to tty
1175 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1176 }
1177 }
1178 }
1179 if (_log != NULL) {
1180 _log->tail("mismatched_nodes");
1181 }
1182 }
1183 }
1184 void Compile::record_modified_node(Node* n) {
1185 if (_modified_nodes != NULL && !_inlining_incrementally && !n->is_Con()) {
1186 _modified_nodes->push(n);
1187 }
1188 }
1189
1190 void Compile::remove_modified_node(Node* n) {
1191 if (_modified_nodes != NULL) {
1192 _modified_nodes->remove(n);
1193 }
1194 }
1195 #endif
1196
1197 #ifndef PRODUCT
1198 void Compile::verify_top(Node* tn) const {
1199 if (tn != NULL) {
1200 assert(tn->is_Con(), "top node must be a constant");
1201 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1202 assert(tn->in(0) != NULL, "must have live top node");
1203 }
1204 }
1205 #endif
1206
1207
1208 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1209
1210 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1211 guarantee(arr != NULL, "");
1212 int num_blocks = arr->length();
1213 if (grow_by < num_blocks) grow_by = num_blocks;
1214 int num_notes = grow_by * _node_notes_block_size;
1215 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1216 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1217 while (num_notes > 0) {
1218 arr->append(notes);
1219 notes += _node_notes_block_size;
1220 num_notes -= _node_notes_block_size;
1221 }
1222 assert(num_notes == 0, "exact multiple, please");
1223 }
1224
1225 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1226 if (source == NULL || dest == NULL) return false;
1227
1228 if (dest->is_Con())
1229 return false; // Do not push debug info onto constants.
1230
1231 #ifdef ASSERT
1232 // Leave a bread crumb trail pointing to the original node:
1233 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1234 dest->set_debug_orig(source);
1235 }
1236 #endif
1237
1238 if (node_note_array() == NULL)
1239 return false; // Not collecting any notes now.
1240
1241 // This is a copy onto a pre-existing node, which may already have notes.
1242 // If both nodes have notes, do not overwrite any pre-existing notes.
1243 Node_Notes* source_notes = node_notes_at(source->_idx);
1244 if (source_notes == NULL || source_notes->is_clear()) return false;
1245 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1246 if (dest_notes == NULL || dest_notes->is_clear()) {
1247 return set_node_notes_at(dest->_idx, source_notes);
1248 }
1249
1250 Node_Notes merged_notes = (*source_notes);
1251 // The order of operations here ensures that dest notes will win...
1252 merged_notes.update_from(dest_notes);
1253 return set_node_notes_at(dest->_idx, &merged_notes);
1254 }
1255
1256
1257 //--------------------------allow_range_check_smearing-------------------------
1258 // Gating condition for coalescing similar range checks.
1259 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1260 // single covering check that is at least as strong as any of them.
1261 // If the optimization succeeds, the simplified (strengthened) range check
1262 // will always succeed. If it fails, we will deopt, and then give up
1263 // on the optimization.
1264 bool Compile::allow_range_check_smearing() const {
1265 // If this method has already thrown a range-check,
1266 // assume it was because we already tried range smearing
1267 // and it failed.
1268 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1269 return !already_trapped;
1270 }
1271
1272
1273 //------------------------------flatten_alias_type-----------------------------
1274 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1275 int offset = tj->offset();
1276 TypePtr::PTR ptr = tj->ptr();
1277
1278 // Known instance (scalarizable allocation) alias only with itself.
1279 bool is_known_inst = tj->isa_oopptr() != NULL &&
1280 tj->is_oopptr()->is_known_instance();
1281
1282 // Process weird unsafe references.
1283 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1284 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1285 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1286 tj = TypeOopPtr::BOTTOM;
1287 ptr = tj->ptr();
1288 offset = tj->offset();
1289 }
1290
1291 // Array pointers need some flattening
1292 const TypeAryPtr *ta = tj->isa_aryptr();
1293 if (ta && ta->is_stable()) {
1294 // Erase stability property for alias analysis.
1295 tj = ta = ta->cast_to_stable(false);
1296 }
1297 if( ta && is_known_inst ) {
1298 if ( offset != Type::OffsetBot &&
1299 offset > arrayOopDesc::length_offset_in_bytes() ) {
1300 offset = Type::OffsetBot; // Flatten constant access into array body only
1301 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1302 }
1303 } else if( ta && _AliasLevel >= 2 ) {
1304 // For arrays indexed by constant indices, we flatten the alias
1305 // space to include all of the array body. Only the header, klass
1306 // and array length can be accessed un-aliased.
1307 if( offset != Type::OffsetBot ) {
1308 if( ta->const_oop() ) { // MethodData* or Method*
1309 offset = Type::OffsetBot; // Flatten constant access into array body
1310 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1311 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1312 // range is OK as-is.
1313 tj = ta = TypeAryPtr::RANGE;
1314 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1315 tj = TypeInstPtr::KLASS; // all klass loads look alike
1316 ta = TypeAryPtr::RANGE; // generic ignored junk
1317 ptr = TypePtr::BotPTR;
1318 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1319 tj = TypeInstPtr::MARK;
1320 ta = TypeAryPtr::RANGE; // generic ignored junk
1321 ptr = TypePtr::BotPTR;
1322 } else { // Random constant offset into array body
1323 offset = Type::OffsetBot; // Flatten constant access into array body
1324 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1325 }
1326 }
1327 // Arrays of fixed size alias with arrays of unknown size.
1328 if (ta->size() != TypeInt::POS) {
1329 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1330 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1331 }
1332 // Arrays of known objects become arrays of unknown objects.
1333 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1334 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1335 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1336 }
1337 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1338 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1339 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1340 }
1341 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1342 // cannot be distinguished by bytecode alone.
1343 if (ta->elem() == TypeInt::BOOL) {
1344 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1345 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1346 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1347 }
1348 // During the 2nd round of IterGVN, NotNull castings are removed.
1349 // Make sure the Bottom and NotNull variants alias the same.
1350 // Also, make sure exact and non-exact variants alias the same.
1351 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1352 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1353 }
1354 }
1355
1356 // Oop pointers need some flattening
1357 const TypeInstPtr *to = tj->isa_instptr();
1358 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1359 ciInstanceKlass *k = to->klass()->as_instance_klass();
1360 if( ptr == TypePtr::Constant ) {
1361 if (to->klass() != ciEnv::current()->Class_klass() ||
1362 offset < k->layout_helper_size_in_bytes()) {
1363 // No constant oop pointers (such as Strings); they alias with
1364 // unknown strings.
1365 assert(!is_known_inst, "not scalarizable allocation");
1366 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1367 }
1368 } else if( is_known_inst ) {
1369 tj = to; // Keep NotNull and klass_is_exact for instance type
1370 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1371 // During the 2nd round of IterGVN, NotNull castings are removed.
1372 // Make sure the Bottom and NotNull variants alias the same.
1373 // Also, make sure exact and non-exact variants alias the same.
1374 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1375 }
1376 if (to->speculative() != NULL) {
1377 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1378 }
1379 // Canonicalize the holder of this field
1380 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1381 // First handle header references such as a LoadKlassNode, even if the
1382 // object's klass is unloaded at compile time (4965979).
1383 if (!is_known_inst) { // Do it only for non-instance types
1384 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1385 }
1386 } else if (offset < 0 || offset >= k->layout_helper_size_in_bytes()) {
1387 // Static fields are in the space above the normal instance
1388 // fields in the java.lang.Class instance.
1389 if (to->klass() != ciEnv::current()->Class_klass()) {
1390 to = NULL;
1391 tj = TypeOopPtr::BOTTOM;
1392 offset = tj->offset();
1393 }
1394 } else {
1395 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1396 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1397 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1398 if( is_known_inst ) {
1399 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1400 } else {
1401 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1402 }
1403 }
1404 }
1405 }
1406
1407 // Klass pointers to object array klasses need some flattening
1408 const TypeKlassPtr *tk = tj->isa_klassptr();
1409 if( tk ) {
1410 // If we are referencing a field within a Klass, we need
1411 // to assume the worst case of an Object. Both exact and
1412 // inexact types must flatten to the same alias class so
1413 // use NotNull as the PTR.
1414 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1415
1416 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1417 TypeInstKlassPtr::OBJECT->klass(),
1418 offset);
1419 }
1420
1421 ciKlass* klass = tk->klass();
1422 if( klass->is_obj_array_klass() ) {
1423 ciKlass* k = TypeAryPtr::OOPS->klass();
1424 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1425 k = TypeInstPtr::BOTTOM->klass();
1426 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1427 }
1428
1429 // Check for precise loads from the primary supertype array and force them
1430 // to the supertype cache alias index. Check for generic array loads from
1431 // the primary supertype array and also force them to the supertype cache
1432 // alias index. Since the same load can reach both, we need to merge
1433 // these 2 disparate memories into the same alias class. Since the
1434 // primary supertype array is read-only, there's no chance of confusion
1435 // where we bypass an array load and an array store.
1436 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1437 if (offset == Type::OffsetBot ||
1438 (offset >= primary_supers_offset &&
1439 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1440 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1441 offset = in_bytes(Klass::secondary_super_cache_offset());
1442 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1443 }
1444 }
1445
1446 // Flatten all Raw pointers together.
1447 if (tj->base() == Type::RawPtr)
1448 tj = TypeRawPtr::BOTTOM;
1449
1450 if (tj->base() == Type::AnyPtr)
1451 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1452
1453 // Flatten all to bottom for now
1454 switch( _AliasLevel ) {
1455 case 0:
1456 tj = TypePtr::BOTTOM;
1457 break;
1458 case 1: // Flatten to: oop, static, field or array
1459 switch (tj->base()) {
1460 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1461 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1462 case Type::AryPtr: // do not distinguish arrays at all
1463 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1464 case Type::KlassPtr:
1465 case Type::AryKlassPtr:
1466 case Type::InstKlassPtr: tj = TypeInstKlassPtr::OBJECT; break;
1467 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1468 default: ShouldNotReachHere();
1469 }
1470 break;
1471 case 2: // No collapsing at level 2; keep all splits
1472 case 3: // No collapsing at level 3; keep all splits
1473 break;
1474 default:
1475 Unimplemented();
1476 }
1477
1478 offset = tj->offset();
1479 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1480
1481 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1482 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1483 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1484 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1485 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1486 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1487 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1488 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1489 assert( tj->ptr() != TypePtr::TopPTR &&
1490 tj->ptr() != TypePtr::AnyNull &&
1491 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1492 // assert( tj->ptr() != TypePtr::Constant ||
1493 // tj->base() == Type::RawPtr ||
1494 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1495
1496 return tj;
1497 }
1498
1499 void Compile::AliasType::Init(int i, const TypePtr* at) {
1500 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1501 _index = i;
1502 _adr_type = at;
1503 _field = NULL;
1504 _element = NULL;
1505 _is_rewritable = true; // default
1506 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1507 if (atoop != NULL && atoop->is_known_instance()) {
1508 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1509 _general_index = Compile::current()->get_alias_index(gt);
1510 } else {
1511 _general_index = 0;
1512 }
1513 }
1514
1515 BasicType Compile::AliasType::basic_type() const {
1516 if (element() != NULL) {
1517 const Type* element = adr_type()->is_aryptr()->elem();
1518 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1519 } if (field() != NULL) {
1520 return field()->layout_type();
1521 } else {
1522 return T_ILLEGAL; // unknown
1523 }
1524 }
1525
1526 //---------------------------------print_on------------------------------------
1527 #ifndef PRODUCT
1528 void Compile::AliasType::print_on(outputStream* st) {
1529 if (index() < 10)
1530 st->print("@ <%d> ", index());
1531 else st->print("@ <%d>", index());
1532 st->print(is_rewritable() ? " " : " RO");
1533 int offset = adr_type()->offset();
1534 if (offset == Type::OffsetBot)
1535 st->print(" +any");
1536 else st->print(" +%-3d", offset);
1537 st->print(" in ");
1538 adr_type()->dump_on(st);
1539 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1540 if (field() != NULL && tjp) {
1541 if (tjp->klass() != field()->holder() ||
1542 tjp->offset() != field()->offset_in_bytes()) {
1543 st->print(" != ");
1544 field()->print();
1545 st->print(" ***");
1546 }
1547 }
1548 }
1549
1550 void print_alias_types() {
1551 Compile* C = Compile::current();
1552 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1553 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1554 C->alias_type(idx)->print_on(tty);
1555 tty->cr();
1556 }
1557 }
1558 #endif
1559
1560
1561 //----------------------------probe_alias_cache--------------------------------
1562 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1563 intptr_t key = (intptr_t) adr_type;
1564 key ^= key >> logAliasCacheSize;
1565 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1566 }
1567
1568
1569 //-----------------------------grow_alias_types--------------------------------
1570 void Compile::grow_alias_types() {
1571 const int old_ats = _max_alias_types; // how many before?
1572 const int new_ats = old_ats; // how many more?
1573 const int grow_ats = old_ats+new_ats; // how many now?
1574 _max_alias_types = grow_ats;
1575 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1576 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1577 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1578 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1579 }
1580
1581
1582 //--------------------------------find_alias_type------------------------------
1583 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1584 if (_AliasLevel == 0)
1585 return alias_type(AliasIdxBot);
1586
1587 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1588 if (ace->_adr_type == adr_type) {
1589 return alias_type(ace->_index);
1590 }
1591
1592 // Handle special cases.
1593 if (adr_type == NULL) return alias_type(AliasIdxTop);
1594 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1595
1596 // Do it the slow way.
1597 const TypePtr* flat = flatten_alias_type(adr_type);
1598
1599 #ifdef ASSERT
1600 {
1601 ResourceMark rm;
1602 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1603 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1604 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1605 Type::str(adr_type));
1606 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1607 const TypeOopPtr* foop = flat->is_oopptr();
1608 // Scalarizable allocations have exact klass always.
1609 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1610 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1611 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1612 Type::str(foop), Type::str(xoop));
1613 }
1614 }
1615 #endif
1616
1617 int idx = AliasIdxTop;
1618 for (int i = 0; i < num_alias_types(); i++) {
1619 if (alias_type(i)->adr_type() == flat) {
1620 idx = i;
1621 break;
1622 }
1623 }
1624
1625 if (idx == AliasIdxTop) {
1626 if (no_create) return NULL;
1627 // Grow the array if necessary.
1628 if (_num_alias_types == _max_alias_types) grow_alias_types();
1629 // Add a new alias type.
1630 idx = _num_alias_types++;
1631 _alias_types[idx]->Init(idx, flat);
1632 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1633 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1634 if (flat->isa_instptr()) {
1635 if (flat->offset() == java_lang_Class::klass_offset()
1636 && flat->is_instptr()->klass() == env()->Class_klass())
1637 alias_type(idx)->set_rewritable(false);
1638 }
1639 if (flat->isa_aryptr()) {
1640 #ifdef ASSERT
1641 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1642 // (T_BYTE has the weakest alignment and size restrictions...)
1643 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1644 #endif
1645 if (flat->offset() == TypePtr::OffsetBot) {
1646 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1647 }
1648 }
1649 if (flat->isa_klassptr()) {
1650 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1651 alias_type(idx)->set_rewritable(false);
1652 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1653 alias_type(idx)->set_rewritable(false);
1654 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1655 alias_type(idx)->set_rewritable(false);
1656 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1657 alias_type(idx)->set_rewritable(false);
1658 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1659 alias_type(idx)->set_rewritable(false);
1660 }
1661 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1662 // but the base pointer type is not distinctive enough to identify
1663 // references into JavaThread.)
1664
1665 // Check for final fields.
1666 const TypeInstPtr* tinst = flat->isa_instptr();
1667 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1668 ciField* field;
1669 if (tinst->const_oop() != NULL &&
1670 tinst->klass() == ciEnv::current()->Class_klass() &&
1671 tinst->offset() >= (tinst->klass()->as_instance_klass()->layout_helper_size_in_bytes())) {
1672 // static field
1673 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1674 field = k->get_field_by_offset(tinst->offset(), true);
1675 } else {
1676 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1677 field = k->get_field_by_offset(tinst->offset(), false);
1678 }
1679 assert(field == NULL ||
1680 original_field == NULL ||
1681 (field->holder() == original_field->holder() &&
1682 field->offset() == original_field->offset() &&
1683 field->is_static() == original_field->is_static()), "wrong field?");
1684 // Set field() and is_rewritable() attributes.
1685 if (field != NULL) alias_type(idx)->set_field(field);
1686 }
1687 }
1688
1689 // Fill the cache for next time.
1690 ace->_adr_type = adr_type;
1691 ace->_index = idx;
1692 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1693
1694 // Might as well try to fill the cache for the flattened version, too.
1695 AliasCacheEntry* face = probe_alias_cache(flat);
1696 if (face->_adr_type == NULL) {
1697 face->_adr_type = flat;
1698 face->_index = idx;
1699 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1700 }
1701
1702 return alias_type(idx);
1703 }
1704
1705
1706 Compile::AliasType* Compile::alias_type(ciField* field) {
1707 const TypeOopPtr* t;
1708 if (field->is_static())
1709 t = TypeInstPtr::make(field->holder()->java_mirror());
1710 else
1711 t = TypeOopPtr::make_from_klass_raw(field->holder());
1712 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1713 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1714 return atp;
1715 }
1716
1717
1718 //------------------------------have_alias_type--------------------------------
1719 bool Compile::have_alias_type(const TypePtr* adr_type) {
1720 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1721 if (ace->_adr_type == adr_type) {
1722 return true;
1723 }
1724
1725 // Handle special cases.
1726 if (adr_type == NULL) return true;
1727 if (adr_type == TypePtr::BOTTOM) return true;
1728
1729 return find_alias_type(adr_type, true, NULL) != NULL;
1730 }
1731
1732 //-----------------------------must_alias--------------------------------------
1733 // True if all values of the given address type are in the given alias category.
1734 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1735 if (alias_idx == AliasIdxBot) return true; // the universal category
1736 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1737 if (alias_idx == AliasIdxTop) return false; // the empty category
1738 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1739
1740 // the only remaining possible overlap is identity
1741 int adr_idx = get_alias_index(adr_type);
1742 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1743 assert(adr_idx == alias_idx ||
1744 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1745 && adr_type != TypeOopPtr::BOTTOM),
1746 "should not be testing for overlap with an unsafe pointer");
1747 return adr_idx == alias_idx;
1748 }
1749
1750 //------------------------------can_alias--------------------------------------
1751 // True if any values of the given address type are in the given alias category.
1752 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1753 if (alias_idx == AliasIdxTop) return false; // the empty category
1754 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1755 // Known instance doesn't alias with bottom memory
1756 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1757 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1758
1759 // the only remaining possible overlap is identity
1760 int adr_idx = get_alias_index(adr_type);
1761 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1762 return adr_idx == alias_idx;
1763 }
1764
1765 //---------------------cleanup_loop_predicates-----------------------
1766 // Remove the opaque nodes that protect the predicates so that all unused
1767 // checks and uncommon_traps will be eliminated from the ideal graph
1768 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1769 if (predicate_count()==0) return;
1770 for (int i = predicate_count(); i > 0; i--) {
1771 Node * n = predicate_opaque1_node(i-1);
1772 assert(n->Opcode() == Op_Opaque1, "must be");
1773 igvn.replace_node(n, n->in(1));
1774 }
1775 assert(predicate_count()==0, "should be clean!");
1776 }
1777
1778 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1779 if (!n->for_post_loop_opts_igvn()) {
1780 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1781 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1782 _for_post_loop_igvn.append(n);
1783 }
1784 }
1785
1786 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1787 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1788 _for_post_loop_igvn.remove(n);
1789 }
1790
1791 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1792 // Verify that all previous optimizations produced a valid graph
1793 // at least to this point, even if no loop optimizations were done.
1794 PhaseIdealLoop::verify(igvn);
1795
1796 C->set_post_loop_opts_phase(); // no more loop opts allowed
1797
1798 assert(!C->major_progress(), "not cleared");
1799
1800 if (_for_post_loop_igvn.length() > 0) {
1801 while (_for_post_loop_igvn.length() > 0) {
1802 Node* n = _for_post_loop_igvn.pop();
1803 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1804 igvn._worklist.push(n);
1805 }
1806 igvn.optimize();
1807 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1808
1809 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1810 if (C->major_progress()) {
1811 C->clear_major_progress(); // ensure that major progress is now clear
1812 }
1813 }
1814 }
1815
1816 // StringOpts and late inlining of string methods
1817 void Compile::inline_string_calls(bool parse_time) {
1818 {
1819 // remove useless nodes to make the usage analysis simpler
1820 ResourceMark rm;
1821 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1822 }
1823
1824 {
1825 ResourceMark rm;
1826 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1827 PhaseStringOpts pso(initial_gvn(), for_igvn());
1828 print_method(PHASE_AFTER_STRINGOPTS, 3);
1829 }
1830
1831 // now inline anything that we skipped the first time around
1832 if (!parse_time) {
1833 _late_inlines_pos = _late_inlines.length();
1834 }
1835
1836 while (_string_late_inlines.length() > 0) {
1837 CallGenerator* cg = _string_late_inlines.pop();
1838 cg->do_late_inline();
1839 if (failing()) return;
1840 }
1841 _string_late_inlines.trunc_to(0);
1842 }
1843
1844 // Late inlining of boxing methods
1845 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1846 if (_boxing_late_inlines.length() > 0) {
1847 assert(has_boxed_value(), "inconsistent");
1848
1849 PhaseGVN* gvn = initial_gvn();
1850 set_inlining_incrementally(true);
1851
1852 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1853 for_igvn()->clear();
1854 gvn->replace_with(&igvn);
1855
1856 _late_inlines_pos = _late_inlines.length();
1857
1858 while (_boxing_late_inlines.length() > 0) {
1859 CallGenerator* cg = _boxing_late_inlines.pop();
1860 cg->do_late_inline();
1861 if (failing()) return;
1862 }
1863 _boxing_late_inlines.trunc_to(0);
1864
1865 inline_incrementally_cleanup(igvn);
1866
1867 set_inlining_incrementally(false);
1868 }
1869 }
1870
1871 bool Compile::inline_incrementally_one() {
1872 assert(IncrementalInline, "incremental inlining should be on");
1873
1874 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
1875
1876 set_inlining_progress(false);
1877 set_do_cleanup(false);
1878
1879 for (int i = 0; i < _late_inlines.length(); i++) {
1880 _late_inlines_pos = i+1;
1881 CallGenerator* cg = _late_inlines.at(i);
1882 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
1883 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
1884 cg->do_late_inline();
1885 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
1886 if (failing()) {
1887 return false;
1888 } else if (inlining_progress()) {
1889 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
1890 print_method(PHASE_INCREMENTAL_INLINE_STEP, cg->call_node(), 3);
1891 break; // process one call site at a time
1892 }
1893 } else {
1894 // Ignore late inline direct calls when inlining is not allowed.
1895 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
1896 }
1897 }
1898 // Remove processed elements.
1899 _late_inlines.remove_till(_late_inlines_pos);
1900 _late_inlines_pos = 0;
1901
1902 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
1903
1904 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
1905
1906 set_inlining_progress(false);
1907 set_do_cleanup(false);
1908
1909 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
1910 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
1911 }
1912
1913 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
1914 {
1915 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
1916 ResourceMark rm;
1917 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1918 }
1919 {
1920 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
1921 igvn = PhaseIterGVN(initial_gvn());
1922 igvn.optimize();
1923 }
1924 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
1925 }
1926
1927 // Perform incremental inlining until bound on number of live nodes is reached
1928 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1929 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
1930
1931 set_inlining_incrementally(true);
1932 uint low_live_nodes = 0;
1933
1934 while (_late_inlines.length() > 0) {
1935 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1936 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1937 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
1938 // PhaseIdealLoop is expensive so we only try it once we are
1939 // out of live nodes and we only try it again if the previous
1940 // helped got the number of nodes down significantly
1941 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
1942 if (failing()) return;
1943 low_live_nodes = live_nodes();
1944 _major_progress = true;
1945 }
1946
1947 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1948 bool do_print_inlining = print_inlining() || print_intrinsics();
1949 if (do_print_inlining || log() != NULL) {
1950 // Print inlining message for candidates that we couldn't inline for lack of space.
1951 for (int i = 0; i < _late_inlines.length(); i++) {
1952 CallGenerator* cg = _late_inlines.at(i);
1953 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
1954 if (do_print_inlining) {
1955 cg->print_inlining_late(msg);
1956 }
1957 log_late_inline_failure(cg, msg);
1958 }
1959 }
1960 break; // finish
1961 }
1962 }
1963
1964 for_igvn()->clear();
1965 initial_gvn()->replace_with(&igvn);
1966
1967 while (inline_incrementally_one()) {
1968 assert(!failing(), "inconsistent");
1969 }
1970 if (failing()) return;
1971
1972 inline_incrementally_cleanup(igvn);
1973
1974 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
1975
1976 if (failing()) return;
1977
1978 if (_late_inlines.length() == 0) {
1979 break; // no more progress
1980 }
1981 }
1982 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1983
1984 if (_string_late_inlines.length() > 0) {
1985 assert(has_stringbuilder(), "inconsistent");
1986 for_igvn()->clear();
1987 initial_gvn()->replace_with(&igvn);
1988
1989 inline_string_calls(false);
1990
1991 if (failing()) return;
1992
1993 inline_incrementally_cleanup(igvn);
1994 }
1995
1996 set_inlining_incrementally(false);
1997 }
1998
1999 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2000 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2001 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2002 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == NULL"
2003 // as if "inlining_incrementally() == true" were set.
2004 assert(inlining_incrementally() == false, "not allowed");
2005 assert(_modified_nodes == NULL, "not allowed");
2006 assert(_late_inlines.length() > 0, "sanity");
2007
2008 while (_late_inlines.length() > 0) {
2009 for_igvn()->clear();
2010 initial_gvn()->replace_with(&igvn);
2011
2012 while (inline_incrementally_one()) {
2013 assert(!failing(), "inconsistent");
2014 }
2015 if (failing()) return;
2016
2017 inline_incrementally_cleanup(igvn);
2018 }
2019 }
2020
2021 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2022 if (_loop_opts_cnt > 0) {
2023 debug_only( int cnt = 0; );
2024 while (major_progress() && (_loop_opts_cnt > 0)) {
2025 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2026 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2027 PhaseIdealLoop::optimize(igvn, mode);
2028 _loop_opts_cnt--;
2029 if (failing()) return false;
2030 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2031 }
2032 }
2033 return true;
2034 }
2035
2036 // Remove edges from "root" to each SafePoint at a backward branch.
2037 // They were inserted during parsing (see add_safepoint()) to make
2038 // infinite loops without calls or exceptions visible to root, i.e.,
2039 // useful.
2040 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2041 Node *r = root();
2042 if (r != NULL) {
2043 for (uint i = r->req(); i < r->len(); ++i) {
2044 Node *n = r->in(i);
2045 if (n != NULL && n->is_SafePoint()) {
2046 r->rm_prec(i);
2047 if (n->outcnt() == 0) {
2048 igvn.remove_dead_node(n);
2049 }
2050 --i;
2051 }
2052 }
2053 // Parsing may have added top inputs to the root node (Path
2054 // leading to the Halt node proven dead). Make sure we get a
2055 // chance to clean them up.
2056 igvn._worklist.push(r);
2057 igvn.optimize();
2058 }
2059 }
2060
2061 //------------------------------Optimize---------------------------------------
2062 // Given a graph, optimize it.
2063 void Compile::Optimize() {
2064 TracePhase tp("optimizer", &timers[_t_optimizer]);
2065
2066 #ifndef PRODUCT
2067 if (env()->break_at_compile()) {
2068 BREAKPOINT;
2069 }
2070
2071 #endif
2072
2073 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2074 #ifdef ASSERT
2075 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2076 #endif
2077
2078 ResourceMark rm;
2079
2080 print_inlining_reinit();
2081
2082 NOT_PRODUCT( verify_graph_edges(); )
2083
2084 print_method(PHASE_AFTER_PARSING);
2085
2086 {
2087 // Iterative Global Value Numbering, including ideal transforms
2088 // Initialize IterGVN with types and values from parse-time GVN
2089 PhaseIterGVN igvn(initial_gvn());
2090 #ifdef ASSERT
2091 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2092 #endif
2093 {
2094 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2095 igvn.optimize();
2096 }
2097
2098 if (failing()) return;
2099
2100 print_method(PHASE_ITER_GVN1, 2);
2101
2102 inline_incrementally(igvn);
2103
2104 print_method(PHASE_INCREMENTAL_INLINE, 2);
2105
2106 if (failing()) return;
2107
2108 if (eliminate_boxing()) {
2109 // Inline valueOf() methods now.
2110 inline_boxing_calls(igvn);
2111
2112 if (AlwaysIncrementalInline) {
2113 inline_incrementally(igvn);
2114 }
2115
2116 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2117
2118 if (failing()) return;
2119 }
2120
2121 // Remove the speculative part of types and clean up the graph from
2122 // the extra CastPP nodes whose only purpose is to carry them. Do
2123 // that early so that optimizations are not disrupted by the extra
2124 // CastPP nodes.
2125 remove_speculative_types(igvn);
2126
2127 // No more new expensive nodes will be added to the list from here
2128 // so keep only the actual candidates for optimizations.
2129 cleanup_expensive_nodes(igvn);
2130
2131 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2132 if (EnableVectorSupport && has_vbox_nodes()) {
2133 TracePhase tp("", &timers[_t_vector]);
2134 PhaseVector pv(igvn);
2135 pv.optimize_vector_boxes();
2136
2137 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2138 }
2139 assert(!has_vbox_nodes(), "sanity");
2140
2141 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2142 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2143 initial_gvn()->replace_with(&igvn);
2144 Unique_Node_List* old_worklist = for_igvn();
2145 old_worklist->clear();
2146 Unique_Node_List new_worklist(C->comp_arena());
2147 {
2148 ResourceMark rm;
2149 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2150 }
2151 Unique_Node_List* save_for_igvn = for_igvn();
2152 set_for_igvn(&new_worklist);
2153 igvn = PhaseIterGVN(initial_gvn());
2154 igvn.optimize();
2155 set_for_igvn(old_worklist); // new_worklist is dead beyond this point
2156 }
2157
2158 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2159 // safepoints
2160 remove_root_to_sfpts_edges(igvn);
2161
2162 // Perform escape analysis
2163 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2164 if (has_loops()) {
2165 // Cleanup graph (remove dead nodes).
2166 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2167 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2168 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2169 if (failing()) return;
2170 }
2171 ConnectionGraph::do_analysis(this, &igvn);
2172
2173 if (failing()) return;
2174
2175 // Optimize out fields loads from scalar replaceable allocations.
2176 igvn.optimize();
2177 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2178
2179 if (failing()) return;
2180
2181 if (congraph() != NULL && macro_count() > 0) {
2182 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2183 PhaseMacroExpand mexp(igvn);
2184 mexp.eliminate_macro_nodes();
2185 igvn.set_delay_transform(false);
2186
2187 igvn.optimize();
2188 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2189
2190 if (failing()) return;
2191 }
2192 }
2193
2194 // Loop transforms on the ideal graph. Range Check Elimination,
2195 // peeling, unrolling, etc.
2196
2197 // Set loop opts counter
2198 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2199 {
2200 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2201 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2202 _loop_opts_cnt--;
2203 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2204 if (failing()) return;
2205 }
2206 // Loop opts pass if partial peeling occurred in previous pass
2207 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2208 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2209 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2210 _loop_opts_cnt--;
2211 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2212 if (failing()) return;
2213 }
2214 // Loop opts pass for loop-unrolling before CCP
2215 if(major_progress() && (_loop_opts_cnt > 0)) {
2216 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2217 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2218 _loop_opts_cnt--;
2219 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2220 }
2221 if (!failing()) {
2222 // Verify that last round of loop opts produced a valid graph
2223 PhaseIdealLoop::verify(igvn);
2224 }
2225 }
2226 if (failing()) return;
2227
2228 // Conditional Constant Propagation;
2229 PhaseCCP ccp( &igvn );
2230 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2231 {
2232 TracePhase tp("ccp", &timers[_t_ccp]);
2233 ccp.do_transform();
2234 }
2235 print_method(PHASE_CCP1, 2);
2236
2237 assert( true, "Break here to ccp.dump_old2new_map()");
2238
2239 // Iterative Global Value Numbering, including ideal transforms
2240 {
2241 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2242 igvn = ccp;
2243 igvn.optimize();
2244 }
2245 print_method(PHASE_ITER_GVN2, 2);
2246
2247 if (failing()) return;
2248
2249 // Loop transforms on the ideal graph. Range Check Elimination,
2250 // peeling, unrolling, etc.
2251 if (!optimize_loops(igvn, LoopOptsDefault)) {
2252 return;
2253 }
2254
2255 if (failing()) return;
2256
2257 C->clear_major_progress(); // ensure that major progress is now clear
2258
2259 process_for_post_loop_opts_igvn(igvn);
2260
2261 #ifdef ASSERT
2262 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2263 #endif
2264
2265 {
2266 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2267 PhaseMacroExpand mex(igvn);
2268 if (mex.expand_macro_nodes()) {
2269 assert(failing(), "must bail out w/ explicit message");
2270 return;
2271 }
2272 print_method(PHASE_MACRO_EXPANSION, 2);
2273 }
2274
2275 {
2276 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2277 if (bs->expand_barriers(this, igvn)) {
2278 assert(failing(), "must bail out w/ explicit message");
2279 return;
2280 }
2281 print_method(PHASE_BARRIER_EXPANSION, 2);
2282 }
2283
2284 if (C->max_vector_size() > 0) {
2285 C->optimize_logic_cones(igvn);
2286 igvn.optimize();
2287 }
2288
2289 DEBUG_ONLY( _modified_nodes = NULL; )
2290
2291 assert(igvn._worklist.size() == 0, "not empty");
2292
2293 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2294
2295 if (_late_inlines.length() > 0) {
2296 // More opportunities to optimize virtual and MH calls.
2297 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2298 process_late_inline_calls_no_inline(igvn);
2299 }
2300 } // (End scope of igvn; run destructor if necessary for asserts.)
2301
2302 check_no_dead_use();
2303
2304 process_print_inlining();
2305
2306 // A method with only infinite loops has no edges entering loops from root
2307 {
2308 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2309 if (final_graph_reshaping()) {
2310 assert(failing(), "must bail out w/ explicit message");
2311 return;
2312 }
2313 }
2314
2315 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2316 DEBUG_ONLY(set_phase_optimize_finished();)
2317 }
2318
2319 #ifdef ASSERT
2320 void Compile::check_no_dead_use() const {
2321 ResourceMark rm;
2322 Unique_Node_List wq;
2323 wq.push(root());
2324 for (uint i = 0; i < wq.size(); ++i) {
2325 Node* n = wq.at(i);
2326 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2327 Node* u = n->fast_out(j);
2328 if (u->outcnt() == 0 && !u->is_Con()) {
2329 u->dump();
2330 fatal("no reachable node should have no use");
2331 }
2332 wq.push(u);
2333 }
2334 }
2335 }
2336 #endif
2337
2338 void Compile::inline_vector_reboxing_calls() {
2339 if (C->_vector_reboxing_late_inlines.length() > 0) {
2340 _late_inlines_pos = C->_late_inlines.length();
2341 while (_vector_reboxing_late_inlines.length() > 0) {
2342 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2343 cg->do_late_inline();
2344 if (failing()) return;
2345 print_method(PHASE_INLINE_VECTOR_REBOX, cg->call_node());
2346 }
2347 _vector_reboxing_late_inlines.trunc_to(0);
2348 }
2349 }
2350
2351 bool Compile::has_vbox_nodes() {
2352 if (C->_vector_reboxing_late_inlines.length() > 0) {
2353 return true;
2354 }
2355 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2356 Node * n = C->macro_node(macro_idx);
2357 assert(n->is_macro(), "only macro nodes expected here");
2358 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2359 return true;
2360 }
2361 }
2362 return false;
2363 }
2364
2365 //---------------------------- Bitwise operation packing optimization ---------------------------
2366
2367 static bool is_vector_unary_bitwise_op(Node* n) {
2368 return n->Opcode() == Op_XorV &&
2369 n->req() == 2 &&
2370 VectorNode::is_vector_bitwise_not_pattern(n);
2371 }
2372
2373 static bool is_vector_binary_bitwise_op(Node* n) {
2374 switch (n->Opcode()) {
2375 case Op_AndV:
2376 case Op_OrV:
2377 return n->req() == 2;
2378
2379 case Op_XorV:
2380 return !is_vector_unary_bitwise_op(n);
2381
2382 default:
2383 return false;
2384 }
2385 }
2386
2387 static bool is_vector_ternary_bitwise_op(Node* n) {
2388 return n->Opcode() == Op_MacroLogicV;
2389 }
2390
2391 static bool is_vector_bitwise_op(Node* n) {
2392 return is_vector_unary_bitwise_op(n) ||
2393 is_vector_binary_bitwise_op(n) ||
2394 is_vector_ternary_bitwise_op(n);
2395 }
2396
2397 static bool is_vector_bitwise_cone_root(Node* n) {
2398 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2399 return false;
2400 }
2401 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2402 if (is_vector_bitwise_op(n->fast_out(i))) {
2403 return false;
2404 }
2405 }
2406 return true;
2407 }
2408
2409 static uint collect_unique_inputs(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2410 uint cnt = 0;
2411 if (is_vector_bitwise_op(n)) {
2412 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2413 for (uint i = 1; i < n->req(); i++) {
2414 Node* in = n->in(i);
2415 bool skip = VectorNode::is_all_ones_vector(in);
2416 if (!skip && !inputs.member(in)) {
2417 inputs.push(in);
2418 cnt++;
2419 }
2420 }
2421 assert(cnt <= 1, "not unary");
2422 } else {
2423 uint last_req = n->req();
2424 if (is_vector_ternary_bitwise_op(n)) {
2425 last_req = n->req() - 1; // skip last input
2426 }
2427 for (uint i = 1; i < last_req; i++) {
2428 Node* def = n->in(i);
2429 if (!inputs.member(def)) {
2430 inputs.push(def);
2431 cnt++;
2432 }
2433 }
2434 }
2435 partition.push(n);
2436 } else { // not a bitwise operations
2437 if (!inputs.member(n)) {
2438 inputs.push(n);
2439 cnt++;
2440 }
2441 }
2442 return cnt;
2443 }
2444
2445 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2446 Unique_Node_List useful_nodes;
2447 C->identify_useful_nodes(useful_nodes);
2448
2449 for (uint i = 0; i < useful_nodes.size(); i++) {
2450 Node* n = useful_nodes.at(i);
2451 if (is_vector_bitwise_cone_root(n)) {
2452 list.push(n);
2453 }
2454 }
2455 }
2456
2457 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2458 const TypeVect* vt,
2459 Unique_Node_List& partition,
2460 Unique_Node_List& inputs) {
2461 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2462 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2463 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2464
2465 Node* in1 = inputs.at(0);
2466 Node* in2 = inputs.at(1);
2467 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2468
2469 uint func = compute_truth_table(partition, inputs);
2470 return igvn.transform(MacroLogicVNode::make(igvn, in3, in2, in1, func, vt));
2471 }
2472
2473 static uint extract_bit(uint func, uint pos) {
2474 return (func & (1 << pos)) >> pos;
2475 }
2476
2477 //
2478 // A macro logic node represents a truth table. It has 4 inputs,
2479 // First three inputs corresponds to 3 columns of a truth table
2480 // and fourth input captures the logic function.
2481 //
2482 // eg. fn = (in1 AND in2) OR in3;
2483 //
2484 // MacroNode(in1,in2,in3,fn)
2485 //
2486 // -----------------
2487 // in1 in2 in3 fn
2488 // -----------------
2489 // 0 0 0 0
2490 // 0 0 1 1
2491 // 0 1 0 0
2492 // 0 1 1 1
2493 // 1 0 0 0
2494 // 1 0 1 1
2495 // 1 1 0 1
2496 // 1 1 1 1
2497 //
2498
2499 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2500 int res = 0;
2501 for (int i = 0; i < 8; i++) {
2502 int bit1 = extract_bit(in1, i);
2503 int bit2 = extract_bit(in2, i);
2504 int bit3 = extract_bit(in3, i);
2505
2506 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2507 int func_bit = extract_bit(func, func_bit_pos);
2508
2509 res |= func_bit << i;
2510 }
2511 return res;
2512 }
2513
2514 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2515 assert(n != NULL, "");
2516 assert(eval_map.contains(n), "absent");
2517 return *(eval_map.get(n));
2518 }
2519
2520 static void eval_operands(Node* n,
2521 uint& func1, uint& func2, uint& func3,
2522 ResourceHashtable<Node*,uint>& eval_map) {
2523 assert(is_vector_bitwise_op(n), "");
2524
2525 if (is_vector_unary_bitwise_op(n)) {
2526 Node* opnd = n->in(1);
2527 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2528 opnd = n->in(2);
2529 }
2530 func1 = eval_operand(opnd, eval_map);
2531 } else if (is_vector_binary_bitwise_op(n)) {
2532 func1 = eval_operand(n->in(1), eval_map);
2533 func2 = eval_operand(n->in(2), eval_map);
2534 } else {
2535 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2536 func1 = eval_operand(n->in(1), eval_map);
2537 func2 = eval_operand(n->in(2), eval_map);
2538 func3 = eval_operand(n->in(3), eval_map);
2539 }
2540 }
2541
2542 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2543 assert(inputs.size() <= 3, "sanity");
2544 ResourceMark rm;
2545 uint res = 0;
2546 ResourceHashtable<Node*,uint> eval_map;
2547
2548 // Populate precomputed functions for inputs.
2549 // Each input corresponds to one column of 3 input truth-table.
2550 uint input_funcs[] = { 0xAA, // (_, _, a) -> a
2551 0xCC, // (_, b, _) -> b
2552 0xF0 }; // (c, _, _) -> c
2553 for (uint i = 0; i < inputs.size(); i++) {
2554 eval_map.put(inputs.at(i), input_funcs[i]);
2555 }
2556
2557 for (uint i = 0; i < partition.size(); i++) {
2558 Node* n = partition.at(i);
2559
2560 uint func1 = 0, func2 = 0, func3 = 0;
2561 eval_operands(n, func1, func2, func3, eval_map);
2562
2563 switch (n->Opcode()) {
2564 case Op_OrV:
2565 assert(func3 == 0, "not binary");
2566 res = func1 | func2;
2567 break;
2568 case Op_AndV:
2569 assert(func3 == 0, "not binary");
2570 res = func1 & func2;
2571 break;
2572 case Op_XorV:
2573 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2574 assert(func2 == 0 && func3 == 0, "not unary");
2575 res = (~func1) & 0xFF;
2576 } else {
2577 assert(func3 == 0, "not binary");
2578 res = func1 ^ func2;
2579 }
2580 break;
2581 case Op_MacroLogicV:
2582 // Ordering of inputs may change during evaluation of sub-tree
2583 // containing MacroLogic node as a child node, thus a re-evaluation
2584 // makes sure that function is evaluated in context of current
2585 // inputs.
2586 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2587 break;
2588
2589 default: assert(false, "not supported: %s", n->Name());
2590 }
2591 assert(res <= 0xFF, "invalid");
2592 eval_map.put(n, res);
2593 }
2594 return res;
2595 }
2596
2597 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2598 assert(partition.size() == 0, "not empty");
2599 assert(inputs.size() == 0, "not empty");
2600 if (is_vector_ternary_bitwise_op(n)) {
2601 return false;
2602 }
2603
2604 bool is_unary_op = is_vector_unary_bitwise_op(n);
2605 if (is_unary_op) {
2606 assert(collect_unique_inputs(n, partition, inputs) == 1, "not unary");
2607 return false; // too few inputs
2608 }
2609
2610 assert(is_vector_binary_bitwise_op(n), "not binary");
2611 Node* in1 = n->in(1);
2612 Node* in2 = n->in(2);
2613
2614 int in1_unique_inputs_cnt = collect_unique_inputs(in1, partition, inputs);
2615 int in2_unique_inputs_cnt = collect_unique_inputs(in2, partition, inputs);
2616 partition.push(n);
2617
2618 // Too many inputs?
2619 if (inputs.size() > 3) {
2620 partition.clear();
2621 inputs.clear();
2622 { // Recompute in2 inputs
2623 Unique_Node_List not_used;
2624 in2_unique_inputs_cnt = collect_unique_inputs(in2, not_used, not_used);
2625 }
2626 // Pick the node with minimum number of inputs.
2627 if (in1_unique_inputs_cnt >= 3 && in2_unique_inputs_cnt >= 3) {
2628 return false; // still too many inputs
2629 }
2630 // Recompute partition & inputs.
2631 Node* child = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in1 : in2);
2632 collect_unique_inputs(child, partition, inputs);
2633
2634 Node* other_input = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in2 : in1);
2635 inputs.push(other_input);
2636
2637 partition.push(n);
2638 }
2639
2640 return (partition.size() == 2 || partition.size() == 3) &&
2641 (inputs.size() == 2 || inputs.size() == 3);
2642 }
2643
2644
2645 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2646 assert(is_vector_bitwise_op(n), "not a root");
2647
2648 visited.set(n->_idx);
2649
2650 // 1) Do a DFS walk over the logic cone.
2651 for (uint i = 1; i < n->req(); i++) {
2652 Node* in = n->in(i);
2653 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2654 process_logic_cone_root(igvn, in, visited);
2655 }
2656 }
2657
2658 // 2) Bottom up traversal: Merge node[s] with
2659 // the parent to form macro logic node.
2660 Unique_Node_List partition;
2661 Unique_Node_List inputs;
2662 if (compute_logic_cone(n, partition, inputs)) {
2663 const TypeVect* vt = n->bottom_type()->is_vect();
2664 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2665 igvn.replace_node(n, macro_logic);
2666 }
2667 }
2668
2669 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2670 ResourceMark rm;
2671 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2672 Unique_Node_List list;
2673 collect_logic_cone_roots(list);
2674
2675 while (list.size() > 0) {
2676 Node* n = list.pop();
2677 const TypeVect* vt = n->bottom_type()->is_vect();
2678 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2679 if (supported) {
2680 VectorSet visited(comp_arena());
2681 process_logic_cone_root(igvn, n, visited);
2682 }
2683 }
2684 }
2685 }
2686
2687 //------------------------------Code_Gen---------------------------------------
2688 // Given a graph, generate code for it
2689 void Compile::Code_Gen() {
2690 if (failing()) {
2691 return;
2692 }
2693
2694 // Perform instruction selection. You might think we could reclaim Matcher
2695 // memory PDQ, but actually the Matcher is used in generating spill code.
2696 // Internals of the Matcher (including some VectorSets) must remain live
2697 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2698 // set a bit in reclaimed memory.
2699
2700 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2701 // nodes. Mapping is only valid at the root of each matched subtree.
2702 NOT_PRODUCT( verify_graph_edges(); )
2703
2704 Matcher matcher;
2705 _matcher = &matcher;
2706 {
2707 TracePhase tp("matcher", &timers[_t_matcher]);
2708 matcher.match();
2709 if (failing()) {
2710 return;
2711 }
2712 }
2713 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2714 // nodes. Mapping is only valid at the root of each matched subtree.
2715 NOT_PRODUCT( verify_graph_edges(); )
2716
2717 // If you have too many nodes, or if matching has failed, bail out
2718 check_node_count(0, "out of nodes matching instructions");
2719 if (failing()) {
2720 return;
2721 }
2722
2723 print_method(PHASE_MATCHING, 2);
2724
2725 // Build a proper-looking CFG
2726 PhaseCFG cfg(node_arena(), root(), matcher);
2727 _cfg = &cfg;
2728 {
2729 TracePhase tp("scheduler", &timers[_t_scheduler]);
2730 bool success = cfg.do_global_code_motion();
2731 if (!success) {
2732 return;
2733 }
2734
2735 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2736 NOT_PRODUCT( verify_graph_edges(); )
2737 cfg.verify();
2738 }
2739
2740 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2741 _regalloc = ®alloc;
2742 {
2743 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2744 // Perform register allocation. After Chaitin, use-def chains are
2745 // no longer accurate (at spill code) and so must be ignored.
2746 // Node->LRG->reg mappings are still accurate.
2747 _regalloc->Register_Allocate();
2748
2749 // Bail out if the allocator builds too many nodes
2750 if (failing()) {
2751 return;
2752 }
2753 }
2754
2755 // Prior to register allocation we kept empty basic blocks in case the
2756 // the allocator needed a place to spill. After register allocation we
2757 // are not adding any new instructions. If any basic block is empty, we
2758 // can now safely remove it.
2759 {
2760 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2761 cfg.remove_empty_blocks();
2762 if (do_freq_based_layout()) {
2763 PhaseBlockLayout layout(cfg);
2764 } else {
2765 cfg.set_loop_alignment();
2766 }
2767 cfg.fixup_flow();
2768 }
2769
2770 // Apply peephole optimizations
2771 if( OptoPeephole ) {
2772 TracePhase tp("peephole", &timers[_t_peephole]);
2773 PhasePeephole peep( _regalloc, cfg);
2774 peep.do_transform();
2775 }
2776
2777 // Do late expand if CPU requires this.
2778 if (Matcher::require_postalloc_expand) {
2779 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2780 cfg.postalloc_expand(_regalloc);
2781 }
2782
2783 // Convert Nodes to instruction bits in a buffer
2784 {
2785 TracePhase tp("output", &timers[_t_output]);
2786 PhaseOutput output;
2787 output.Output();
2788 if (failing()) return;
2789 output.install();
2790 }
2791
2792 print_method(PHASE_FINAL_CODE);
2793
2794 // He's dead, Jim.
2795 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2796 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2797 }
2798
2799 //------------------------------Final_Reshape_Counts---------------------------
2800 // This class defines counters to help identify when a method
2801 // may/must be executed using hardware with only 24-bit precision.
2802 struct Final_Reshape_Counts : public StackObj {
2803 int _call_count; // count non-inlined 'common' calls
2804 int _float_count; // count float ops requiring 24-bit precision
2805 int _double_count; // count double ops requiring more precision
2806 int _java_call_count; // count non-inlined 'java' calls
2807 int _inner_loop_count; // count loops which need alignment
2808 VectorSet _visited; // Visitation flags
2809 Node_List _tests; // Set of IfNodes & PCTableNodes
2810
2811 Final_Reshape_Counts() :
2812 _call_count(0), _float_count(0), _double_count(0),
2813 _java_call_count(0), _inner_loop_count(0) { }
2814
2815 void inc_call_count () { _call_count ++; }
2816 void inc_float_count () { _float_count ++; }
2817 void inc_double_count() { _double_count++; }
2818 void inc_java_call_count() { _java_call_count++; }
2819 void inc_inner_loop_count() { _inner_loop_count++; }
2820
2821 int get_call_count () const { return _call_count ; }
2822 int get_float_count () const { return _float_count ; }
2823 int get_double_count() const { return _double_count; }
2824 int get_java_call_count() const { return _java_call_count; }
2825 int get_inner_loop_count() const { return _inner_loop_count; }
2826 };
2827
2828 // Eliminate trivially redundant StoreCMs and accumulate their
2829 // precedence edges.
2830 void Compile::eliminate_redundant_card_marks(Node* n) {
2831 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2832 if (n->in(MemNode::Address)->outcnt() > 1) {
2833 // There are multiple users of the same address so it might be
2834 // possible to eliminate some of the StoreCMs
2835 Node* mem = n->in(MemNode::Memory);
2836 Node* adr = n->in(MemNode::Address);
2837 Node* val = n->in(MemNode::ValueIn);
2838 Node* prev = n;
2839 bool done = false;
2840 // Walk the chain of StoreCMs eliminating ones that match. As
2841 // long as it's a chain of single users then the optimization is
2842 // safe. Eliminating partially redundant StoreCMs would require
2843 // cloning copies down the other paths.
2844 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2845 if (adr == mem->in(MemNode::Address) &&
2846 val == mem->in(MemNode::ValueIn)) {
2847 // redundant StoreCM
2848 if (mem->req() > MemNode::OopStore) {
2849 // Hasn't been processed by this code yet.
2850 n->add_prec(mem->in(MemNode::OopStore));
2851 } else {
2852 // Already converted to precedence edge
2853 for (uint i = mem->req(); i < mem->len(); i++) {
2854 // Accumulate any precedence edges
2855 if (mem->in(i) != NULL) {
2856 n->add_prec(mem->in(i));
2857 }
2858 }
2859 // Everything above this point has been processed.
2860 done = true;
2861 }
2862 // Eliminate the previous StoreCM
2863 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2864 assert(mem->outcnt() == 0, "should be dead");
2865 mem->disconnect_inputs(this);
2866 } else {
2867 prev = mem;
2868 }
2869 mem = prev->in(MemNode::Memory);
2870 }
2871 }
2872 }
2873
2874 //------------------------------final_graph_reshaping_impl----------------------
2875 // Implement items 1-5 from final_graph_reshaping below.
2876 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2877
2878 if ( n->outcnt() == 0 ) return; // dead node
2879 uint nop = n->Opcode();
2880
2881 // Check for 2-input instruction with "last use" on right input.
2882 // Swap to left input. Implements item (2).
2883 if( n->req() == 3 && // two-input instruction
2884 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2885 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2886 n->in(2)->outcnt() == 1 &&// right use IS a last use
2887 !n->in(2)->is_Con() ) { // right use is not a constant
2888 // Check for commutative opcode
2889 switch( nop ) {
2890 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2891 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
2892 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
2893 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2894 case Op_AndL: case Op_XorL: case Op_OrL:
2895 case Op_AndI: case Op_XorI: case Op_OrI: {
2896 // Move "last use" input to left by swapping inputs
2897 n->swap_edges(1, 2);
2898 break;
2899 }
2900 default:
2901 break;
2902 }
2903 }
2904
2905 #ifdef ASSERT
2906 if( n->is_Mem() ) {
2907 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2908 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2909 // oop will be recorded in oop map if load crosses safepoint
2910 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2911 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2912 "raw memory operations should have control edge");
2913 }
2914 if (n->is_MemBar()) {
2915 MemBarNode* mb = n->as_MemBar();
2916 if (mb->trailing_store() || mb->trailing_load_store()) {
2917 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2918 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
2919 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2920 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2921 } else if (mb->leading()) {
2922 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2923 }
2924 }
2925 #endif
2926 // Count FPU ops and common calls, implements item (3)
2927 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2928 if (!gc_handled) {
2929 final_graph_reshaping_main_switch(n, frc, nop);
2930 }
2931
2932 // Collect CFG split points
2933 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2934 frc._tests.push(n);
2935 }
2936 }
2937
2938 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
2939 switch( nop ) {
2940 // Count all float operations that may use FPU
2941 case Op_AddF:
2942 case Op_SubF:
2943 case Op_MulF:
2944 case Op_DivF:
2945 case Op_NegF:
2946 case Op_ModF:
2947 case Op_ConvI2F:
2948 case Op_ConF:
2949 case Op_CmpF:
2950 case Op_CmpF3:
2951 case Op_StoreF:
2952 case Op_LoadF:
2953 // case Op_ConvL2F: // longs are split into 32-bit halves
2954 frc.inc_float_count();
2955 break;
2956
2957 case Op_ConvF2D:
2958 case Op_ConvD2F:
2959 frc.inc_float_count();
2960 frc.inc_double_count();
2961 break;
2962
2963 // Count all double operations that may use FPU
2964 case Op_AddD:
2965 case Op_SubD:
2966 case Op_MulD:
2967 case Op_DivD:
2968 case Op_NegD:
2969 case Op_ModD:
2970 case Op_ConvI2D:
2971 case Op_ConvD2I:
2972 // case Op_ConvL2D: // handled by leaf call
2973 // case Op_ConvD2L: // handled by leaf call
2974 case Op_ConD:
2975 case Op_CmpD:
2976 case Op_CmpD3:
2977 case Op_StoreD:
2978 case Op_LoadD:
2979 case Op_LoadD_unaligned:
2980 frc.inc_double_count();
2981 break;
2982 case Op_Opaque1: // Remove Opaque Nodes before matching
2983 case Op_Opaque2: // Remove Opaque Nodes before matching
2984 case Op_Opaque3:
2985 n->subsume_by(n->in(1), this);
2986 break;
2987 case Op_CallStaticJava:
2988 case Op_CallJava:
2989 case Op_CallDynamicJava:
2990 frc.inc_java_call_count(); // Count java call site;
2991 case Op_CallRuntime:
2992 case Op_CallLeaf:
2993 case Op_CallLeafVector:
2994 case Op_CallNative:
2995 case Op_CallLeafNoFP: {
2996 assert (n->is_Call(), "");
2997 CallNode *call = n->as_Call();
2998 // Count call sites where the FP mode bit would have to be flipped.
2999 // Do not count uncommon runtime calls:
3000 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3001 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3002 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3003 frc.inc_call_count(); // Count the call site
3004 } else { // See if uncommon argument is shared
3005 Node *n = call->in(TypeFunc::Parms);
3006 int nop = n->Opcode();
3007 // Clone shared simple arguments to uncommon calls, item (1).
3008 if (n->outcnt() > 1 &&
3009 !n->is_Proj() &&
3010 nop != Op_CreateEx &&
3011 nop != Op_CheckCastPP &&
3012 nop != Op_DecodeN &&
3013 nop != Op_DecodeNKlass &&
3014 !n->is_Mem() &&
3015 !n->is_Phi()) {
3016 Node *x = n->clone();
3017 call->set_req(TypeFunc::Parms, x);
3018 }
3019 }
3020 break;
3021 }
3022
3023 case Op_StoreCM:
3024 {
3025 // Convert OopStore dependence into precedence edge
3026 Node* prec = n->in(MemNode::OopStore);
3027 n->del_req(MemNode::OopStore);
3028 n->add_prec(prec);
3029 eliminate_redundant_card_marks(n);
3030 }
3031
3032 // fall through
3033
3034 case Op_StoreB:
3035 case Op_StoreC:
3036 case Op_StorePConditional:
3037 case Op_StoreI:
3038 case Op_StoreL:
3039 case Op_StoreIConditional:
3040 case Op_StoreLConditional:
3041 case Op_CompareAndSwapB:
3042 case Op_CompareAndSwapS:
3043 case Op_CompareAndSwapI:
3044 case Op_CompareAndSwapL:
3045 case Op_CompareAndSwapP:
3046 case Op_CompareAndSwapN:
3047 case Op_WeakCompareAndSwapB:
3048 case Op_WeakCompareAndSwapS:
3049 case Op_WeakCompareAndSwapI:
3050 case Op_WeakCompareAndSwapL:
3051 case Op_WeakCompareAndSwapP:
3052 case Op_WeakCompareAndSwapN:
3053 case Op_CompareAndExchangeB:
3054 case Op_CompareAndExchangeS:
3055 case Op_CompareAndExchangeI:
3056 case Op_CompareAndExchangeL:
3057 case Op_CompareAndExchangeP:
3058 case Op_CompareAndExchangeN:
3059 case Op_GetAndAddS:
3060 case Op_GetAndAddB:
3061 case Op_GetAndAddI:
3062 case Op_GetAndAddL:
3063 case Op_GetAndSetS:
3064 case Op_GetAndSetB:
3065 case Op_GetAndSetI:
3066 case Op_GetAndSetL:
3067 case Op_GetAndSetP:
3068 case Op_GetAndSetN:
3069 case Op_StoreP:
3070 case Op_StoreN:
3071 case Op_StoreNKlass:
3072 case Op_LoadB:
3073 case Op_LoadUB:
3074 case Op_LoadUS:
3075 case Op_LoadI:
3076 case Op_LoadKlass:
3077 case Op_LoadNKlass:
3078 case Op_LoadL:
3079 case Op_LoadL_unaligned:
3080 case Op_LoadPLocked:
3081 case Op_LoadP:
3082 case Op_LoadN:
3083 case Op_LoadRange:
3084 case Op_LoadS:
3085 break;
3086
3087 case Op_AddP: { // Assert sane base pointers
3088 Node *addp = n->in(AddPNode::Address);
3089 assert( !addp->is_AddP() ||
3090 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3091 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3092 "Base pointers must match (addp %u)", addp->_idx );
3093 #ifdef _LP64
3094 if ((UseCompressedOops || UseCompressedClassPointers) &&
3095 addp->Opcode() == Op_ConP &&
3096 addp == n->in(AddPNode::Base) &&
3097 n->in(AddPNode::Offset)->is_Con()) {
3098 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3099 // on the platform and on the compressed oops mode.
3100 // Use addressing with narrow klass to load with offset on x86.
3101 // Some platforms can use the constant pool to load ConP.
3102 // Do this transformation here since IGVN will convert ConN back to ConP.
3103 const Type* t = addp->bottom_type();
3104 bool is_oop = t->isa_oopptr() != NULL;
3105 bool is_klass = t->isa_klassptr() != NULL;
3106
3107 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3108 (is_klass && Matcher::const_klass_prefer_decode())) {
3109 Node* nn = NULL;
3110
3111 int op = is_oop ? Op_ConN : Op_ConNKlass;
3112
3113 // Look for existing ConN node of the same exact type.
3114 Node* r = root();
3115 uint cnt = r->outcnt();
3116 for (uint i = 0; i < cnt; i++) {
3117 Node* m = r->raw_out(i);
3118 if (m!= NULL && m->Opcode() == op &&
3119 m->bottom_type()->make_ptr() == t) {
3120 nn = m;
3121 break;
3122 }
3123 }
3124 if (nn != NULL) {
3125 // Decode a narrow oop to match address
3126 // [R12 + narrow_oop_reg<<3 + offset]
3127 if (is_oop) {
3128 nn = new DecodeNNode(nn, t);
3129 } else {
3130 nn = new DecodeNKlassNode(nn, t);
3131 }
3132 // Check for succeeding AddP which uses the same Base.
3133 // Otherwise we will run into the assertion above when visiting that guy.
3134 for (uint i = 0; i < n->outcnt(); ++i) {
3135 Node *out_i = n->raw_out(i);
3136 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3137 out_i->set_req(AddPNode::Base, nn);
3138 #ifdef ASSERT
3139 for (uint j = 0; j < out_i->outcnt(); ++j) {
3140 Node *out_j = out_i->raw_out(j);
3141 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3142 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3143 }
3144 #endif
3145 }
3146 }
3147 n->set_req(AddPNode::Base, nn);
3148 n->set_req(AddPNode::Address, nn);
3149 if (addp->outcnt() == 0) {
3150 addp->disconnect_inputs(this);
3151 }
3152 }
3153 }
3154 }
3155 #endif
3156 break;
3157 }
3158
3159 case Op_CastPP: {
3160 // Remove CastPP nodes to gain more freedom during scheduling but
3161 // keep the dependency they encode as control or precedence edges
3162 // (if control is set already) on memory operations. Some CastPP
3163 // nodes don't have a control (don't carry a dependency): skip
3164 // those.
3165 if (n->in(0) != NULL) {
3166 ResourceMark rm;
3167 Unique_Node_List wq;
3168 wq.push(n);
3169 for (uint next = 0; next < wq.size(); ++next) {
3170 Node *m = wq.at(next);
3171 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3172 Node* use = m->fast_out(i);
3173 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3174 use->ensure_control_or_add_prec(n->in(0));
3175 } else {
3176 switch(use->Opcode()) {
3177 case Op_AddP:
3178 case Op_DecodeN:
3179 case Op_DecodeNKlass:
3180 case Op_CheckCastPP:
3181 case Op_CastPP:
3182 wq.push(use);
3183 break;
3184 }
3185 }
3186 }
3187 }
3188 }
3189 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3190 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3191 Node* in1 = n->in(1);
3192 const Type* t = n->bottom_type();
3193 Node* new_in1 = in1->clone();
3194 new_in1->as_DecodeN()->set_type(t);
3195
3196 if (!Matcher::narrow_oop_use_complex_address()) {
3197 //
3198 // x86, ARM and friends can handle 2 adds in addressing mode
3199 // and Matcher can fold a DecodeN node into address by using
3200 // a narrow oop directly and do implicit NULL check in address:
3201 //
3202 // [R12 + narrow_oop_reg<<3 + offset]
3203 // NullCheck narrow_oop_reg
3204 //
3205 // On other platforms (Sparc) we have to keep new DecodeN node and
3206 // use it to do implicit NULL check in address:
3207 //
3208 // decode_not_null narrow_oop_reg, base_reg
3209 // [base_reg + offset]
3210 // NullCheck base_reg
3211 //
3212 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3213 // to keep the information to which NULL check the new DecodeN node
3214 // corresponds to use it as value in implicit_null_check().
3215 //
3216 new_in1->set_req(0, n->in(0));
3217 }
3218
3219 n->subsume_by(new_in1, this);
3220 if (in1->outcnt() == 0) {
3221 in1->disconnect_inputs(this);
3222 }
3223 } else {
3224 n->subsume_by(n->in(1), this);
3225 if (n->outcnt() == 0) {
3226 n->disconnect_inputs(this);
3227 }
3228 }
3229 break;
3230 }
3231 #ifdef _LP64
3232 case Op_CmpP:
3233 // Do this transformation here to preserve CmpPNode::sub() and
3234 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3235 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3236 Node* in1 = n->in(1);
3237 Node* in2 = n->in(2);
3238 if (!in1->is_DecodeNarrowPtr()) {
3239 in2 = in1;
3240 in1 = n->in(2);
3241 }
3242 assert(in1->is_DecodeNarrowPtr(), "sanity");
3243
3244 Node* new_in2 = NULL;
3245 if (in2->is_DecodeNarrowPtr()) {
3246 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3247 new_in2 = in2->in(1);
3248 } else if (in2->Opcode() == Op_ConP) {
3249 const Type* t = in2->bottom_type();
3250 if (t == TypePtr::NULL_PTR) {
3251 assert(in1->is_DecodeN(), "compare klass to null?");
3252 // Don't convert CmpP null check into CmpN if compressed
3253 // oops implicit null check is not generated.
3254 // This will allow to generate normal oop implicit null check.
3255 if (Matcher::gen_narrow_oop_implicit_null_checks())
3256 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3257 //
3258 // This transformation together with CastPP transformation above
3259 // will generated code for implicit NULL checks for compressed oops.
3260 //
3261 // The original code after Optimize()
3262 //
3263 // LoadN memory, narrow_oop_reg
3264 // decode narrow_oop_reg, base_reg
3265 // CmpP base_reg, NULL
3266 // CastPP base_reg // NotNull
3267 // Load [base_reg + offset], val_reg
3268 //
3269 // after these transformations will be
3270 //
3271 // LoadN memory, narrow_oop_reg
3272 // CmpN narrow_oop_reg, NULL
3273 // decode_not_null narrow_oop_reg, base_reg
3274 // Load [base_reg + offset], val_reg
3275 //
3276 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3277 // since narrow oops can be used in debug info now (see the code in
3278 // final_graph_reshaping_walk()).
3279 //
3280 // At the end the code will be matched to
3281 // on x86:
3282 //
3283 // Load_narrow_oop memory, narrow_oop_reg
3284 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3285 // NullCheck narrow_oop_reg
3286 //
3287 // and on sparc:
3288 //
3289 // Load_narrow_oop memory, narrow_oop_reg
3290 // decode_not_null narrow_oop_reg, base_reg
3291 // Load [base_reg + offset], val_reg
3292 // NullCheck base_reg
3293 //
3294 } else if (t->isa_oopptr()) {
3295 new_in2 = ConNode::make(t->make_narrowoop());
3296 } else if (t->isa_klassptr()) {
3297 new_in2 = ConNode::make(t->make_narrowklass());
3298 }
3299 }
3300 if (new_in2 != NULL) {
3301 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3302 n->subsume_by(cmpN, this);
3303 if (in1->outcnt() == 0) {
3304 in1->disconnect_inputs(this);
3305 }
3306 if (in2->outcnt() == 0) {
3307 in2->disconnect_inputs(this);
3308 }
3309 }
3310 }
3311 break;
3312
3313 case Op_DecodeN:
3314 case Op_DecodeNKlass:
3315 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3316 // DecodeN could be pinned when it can't be fold into
3317 // an address expression, see the code for Op_CastPP above.
3318 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3319 break;
3320
3321 case Op_EncodeP:
3322 case Op_EncodePKlass: {
3323 Node* in1 = n->in(1);
3324 if (in1->is_DecodeNarrowPtr()) {
3325 n->subsume_by(in1->in(1), this);
3326 } else if (in1->Opcode() == Op_ConP) {
3327 const Type* t = in1->bottom_type();
3328 if (t == TypePtr::NULL_PTR) {
3329 assert(t->isa_oopptr(), "null klass?");
3330 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3331 } else if (t->isa_oopptr()) {
3332 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3333 } else if (t->isa_klassptr()) {
3334 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3335 }
3336 }
3337 if (in1->outcnt() == 0) {
3338 in1->disconnect_inputs(this);
3339 }
3340 break;
3341 }
3342
3343 case Op_Proj: {
3344 if (OptimizeStringConcat || IncrementalInline) {
3345 ProjNode* proj = n->as_Proj();
3346 if (proj->_is_io_use) {
3347 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3348 // Separate projections were used for the exception path which
3349 // are normally removed by a late inline. If it wasn't inlined
3350 // then they will hang around and should just be replaced with
3351 // the original one. Merge them.
3352 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3353 if (non_io_proj != NULL) {
3354 proj->subsume_by(non_io_proj , this);
3355 }
3356 }
3357 }
3358 break;
3359 }
3360
3361 case Op_Phi:
3362 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3363 // The EncodeP optimization may create Phi with the same edges
3364 // for all paths. It is not handled well by Register Allocator.
3365 Node* unique_in = n->in(1);
3366 assert(unique_in != NULL, "");
3367 uint cnt = n->req();
3368 for (uint i = 2; i < cnt; i++) {
3369 Node* m = n->in(i);
3370 assert(m != NULL, "");
3371 if (unique_in != m)
3372 unique_in = NULL;
3373 }
3374 if (unique_in != NULL) {
3375 n->subsume_by(unique_in, this);
3376 }
3377 }
3378 break;
3379
3380 #endif
3381
3382 #ifdef ASSERT
3383 case Op_CastII:
3384 // Verify that all range check dependent CastII nodes were removed.
3385 if (n->isa_CastII()->has_range_check()) {
3386 n->dump(3);
3387 assert(false, "Range check dependent CastII node was not removed");
3388 }
3389 break;
3390 #endif
3391
3392 case Op_ModI:
3393 if (UseDivMod) {
3394 // Check if a%b and a/b both exist
3395 Node* d = n->find_similar(Op_DivI);
3396 if (d) {
3397 // Replace them with a fused divmod if supported
3398 if (Matcher::has_match_rule(Op_DivModI)) {
3399 DivModINode* divmod = DivModINode::make(n);
3400 d->subsume_by(divmod->div_proj(), this);
3401 n->subsume_by(divmod->mod_proj(), this);
3402 } else {
3403 // replace a%b with a-((a/b)*b)
3404 Node* mult = new MulINode(d, d->in(2));
3405 Node* sub = new SubINode(d->in(1), mult);
3406 n->subsume_by(sub, this);
3407 }
3408 }
3409 }
3410 break;
3411
3412 case Op_ModL:
3413 if (UseDivMod) {
3414 // Check if a%b and a/b both exist
3415 Node* d = n->find_similar(Op_DivL);
3416 if (d) {
3417 // Replace them with a fused divmod if supported
3418 if (Matcher::has_match_rule(Op_DivModL)) {
3419 DivModLNode* divmod = DivModLNode::make(n);
3420 d->subsume_by(divmod->div_proj(), this);
3421 n->subsume_by(divmod->mod_proj(), this);
3422 } else {
3423 // replace a%b with a-((a/b)*b)
3424 Node* mult = new MulLNode(d, d->in(2));
3425 Node* sub = new SubLNode(d->in(1), mult);
3426 n->subsume_by(sub, this);
3427 }
3428 }
3429 }
3430 break;
3431
3432 case Op_LoadVector:
3433 case Op_StoreVector:
3434 case Op_LoadVectorGather:
3435 case Op_StoreVectorScatter:
3436 case Op_LoadVectorGatherMasked:
3437 case Op_StoreVectorScatterMasked:
3438 case Op_VectorCmpMasked:
3439 case Op_VectorMaskGen:
3440 case Op_LoadVectorMasked:
3441 case Op_StoreVectorMasked:
3442 break;
3443
3444 case Op_AddReductionVI:
3445 case Op_AddReductionVL:
3446 case Op_AddReductionVF:
3447 case Op_AddReductionVD:
3448 case Op_MulReductionVI:
3449 case Op_MulReductionVL:
3450 case Op_MulReductionVF:
3451 case Op_MulReductionVD:
3452 case Op_MinReductionV:
3453 case Op_MaxReductionV:
3454 case Op_AndReductionV:
3455 case Op_OrReductionV:
3456 case Op_XorReductionV:
3457 break;
3458
3459 case Op_PackB:
3460 case Op_PackS:
3461 case Op_PackI:
3462 case Op_PackF:
3463 case Op_PackL:
3464 case Op_PackD:
3465 if (n->req()-1 > 2) {
3466 // Replace many operand PackNodes with a binary tree for matching
3467 PackNode* p = (PackNode*) n;
3468 Node* btp = p->binary_tree_pack(1, n->req());
3469 n->subsume_by(btp, this);
3470 }
3471 break;
3472 case Op_Loop:
3473 assert(!n->as_Loop()->is_transformed_long_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3474 case Op_CountedLoop:
3475 case Op_LongCountedLoop:
3476 case Op_OuterStripMinedLoop:
3477 if (n->as_Loop()->is_inner_loop()) {
3478 frc.inc_inner_loop_count();
3479 }
3480 n->as_Loop()->verify_strip_mined(0);
3481 break;
3482 case Op_LShiftI:
3483 case Op_RShiftI:
3484 case Op_URShiftI:
3485 case Op_LShiftL:
3486 case Op_RShiftL:
3487 case Op_URShiftL:
3488 if (Matcher::need_masked_shift_count) {
3489 // The cpu's shift instructions don't restrict the count to the
3490 // lower 5/6 bits. We need to do the masking ourselves.
3491 Node* in2 = n->in(2);
3492 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3493 const TypeInt* t = in2->find_int_type();
3494 if (t != NULL && t->is_con()) {
3495 juint shift = t->get_con();
3496 if (shift > mask) { // Unsigned cmp
3497 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3498 }
3499 } else {
3500 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3501 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3502 n->set_req(2, shift);
3503 }
3504 }
3505 if (in2->outcnt() == 0) { // Remove dead node
3506 in2->disconnect_inputs(this);
3507 }
3508 }
3509 break;
3510 case Op_MemBarStoreStore:
3511 case Op_MemBarRelease:
3512 // Break the link with AllocateNode: it is no longer useful and
3513 // confuses register allocation.
3514 if (n->req() > MemBarNode::Precedent) {
3515 n->set_req(MemBarNode::Precedent, top());
3516 }
3517 break;
3518 case Op_MemBarAcquire: {
3519 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3520 // At parse time, the trailing MemBarAcquire for a volatile load
3521 // is created with an edge to the load. After optimizations,
3522 // that input may be a chain of Phis. If those phis have no
3523 // other use, then the MemBarAcquire keeps them alive and
3524 // register allocation can be confused.
3525 ResourceMark rm;
3526 Unique_Node_List wq;
3527 wq.push(n->in(MemBarNode::Precedent));
3528 n->set_req(MemBarNode::Precedent, top());
3529 while (wq.size() > 0) {
3530 Node* m = wq.pop();
3531 if (m->outcnt() == 0) {
3532 for (uint j = 0; j < m->req(); j++) {
3533 Node* in = m->in(j);
3534 if (in != NULL) {
3535 wq.push(in);
3536 }
3537 }
3538 m->disconnect_inputs(this);
3539 }
3540 }
3541 }
3542 break;
3543 }
3544 case Op_Blackhole:
3545 break;
3546 case Op_RangeCheck: {
3547 RangeCheckNode* rc = n->as_RangeCheck();
3548 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3549 n->subsume_by(iff, this);
3550 frc._tests.push(iff);
3551 break;
3552 }
3553 case Op_ConvI2L: {
3554 if (!Matcher::convi2l_type_required) {
3555 // Code generation on some platforms doesn't need accurate
3556 // ConvI2L types. Widening the type can help remove redundant
3557 // address computations.
3558 n->as_Type()->set_type(TypeLong::INT);
3559 ResourceMark rm;
3560 Unique_Node_List wq;
3561 wq.push(n);
3562 for (uint next = 0; next < wq.size(); next++) {
3563 Node *m = wq.at(next);
3564
3565 for(;;) {
3566 // Loop over all nodes with identical inputs edges as m
3567 Node* k = m->find_similar(m->Opcode());
3568 if (k == NULL) {
3569 break;
3570 }
3571 // Push their uses so we get a chance to remove node made
3572 // redundant
3573 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3574 Node* u = k->fast_out(i);
3575 if (u->Opcode() == Op_LShiftL ||
3576 u->Opcode() == Op_AddL ||
3577 u->Opcode() == Op_SubL ||
3578 u->Opcode() == Op_AddP) {
3579 wq.push(u);
3580 }
3581 }
3582 // Replace all nodes with identical edges as m with m
3583 k->subsume_by(m, this);
3584 }
3585 }
3586 }
3587 break;
3588 }
3589 case Op_CmpUL: {
3590 if (!Matcher::has_match_rule(Op_CmpUL)) {
3591 // No support for unsigned long comparisons
3592 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3593 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3594 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3595 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3596 Node* andl = new AndLNode(orl, remove_sign_mask);
3597 Node* cmp = new CmpLNode(andl, n->in(2));
3598 n->subsume_by(cmp, this);
3599 }
3600 break;
3601 }
3602 default:
3603 assert(!n->is_Call(), "");
3604 assert(!n->is_Mem(), "");
3605 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3606 break;
3607 }
3608 }
3609
3610 //------------------------------final_graph_reshaping_walk---------------------
3611 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3612 // requires that the walk visits a node's inputs before visiting the node.
3613 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3614 Unique_Node_List sfpt;
3615
3616 frc._visited.set(root->_idx); // first, mark node as visited
3617 uint cnt = root->req();
3618 Node *n = root;
3619 uint i = 0;
3620 while (true) {
3621 if (i < cnt) {
3622 // Place all non-visited non-null inputs onto stack
3623 Node* m = n->in(i);
3624 ++i;
3625 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3626 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3627 // compute worst case interpreter size in case of a deoptimization
3628 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3629
3630 sfpt.push(m);
3631 }
3632 cnt = m->req();
3633 nstack.push(n, i); // put on stack parent and next input's index
3634 n = m;
3635 i = 0;
3636 }
3637 } else {
3638 // Now do post-visit work
3639 final_graph_reshaping_impl( n, frc );
3640 if (nstack.is_empty())
3641 break; // finished
3642 n = nstack.node(); // Get node from stack
3643 cnt = n->req();
3644 i = nstack.index();
3645 nstack.pop(); // Shift to the next node on stack
3646 }
3647 }
3648
3649 // Skip next transformation if compressed oops are not used.
3650 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3651 (!UseCompressedOops && !UseCompressedClassPointers))
3652 return;
3653
3654 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3655 // It could be done for an uncommon traps or any safepoints/calls
3656 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3657 while (sfpt.size() > 0) {
3658 n = sfpt.pop();
3659 JVMState *jvms = n->as_SafePoint()->jvms();
3660 assert(jvms != NULL, "sanity");
3661 int start = jvms->debug_start();
3662 int end = n->req();
3663 bool is_uncommon = (n->is_CallStaticJava() &&
3664 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3665 for (int j = start; j < end; j++) {
3666 Node* in = n->in(j);
3667 if (in->is_DecodeNarrowPtr()) {
3668 bool safe_to_skip = true;
3669 if (!is_uncommon ) {
3670 // Is it safe to skip?
3671 for (uint i = 0; i < in->outcnt(); i++) {
3672 Node* u = in->raw_out(i);
3673 if (!u->is_SafePoint() ||
3674 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3675 safe_to_skip = false;
3676 }
3677 }
3678 }
3679 if (safe_to_skip) {
3680 n->set_req(j, in->in(1));
3681 }
3682 if (in->outcnt() == 0) {
3683 in->disconnect_inputs(this);
3684 }
3685 }
3686 }
3687 }
3688 }
3689
3690 //------------------------------final_graph_reshaping--------------------------
3691 // Final Graph Reshaping.
3692 //
3693 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3694 // and not commoned up and forced early. Must come after regular
3695 // optimizations to avoid GVN undoing the cloning. Clone constant
3696 // inputs to Loop Phis; these will be split by the allocator anyways.
3697 // Remove Opaque nodes.
3698 // (2) Move last-uses by commutative operations to the left input to encourage
3699 // Intel update-in-place two-address operations and better register usage
3700 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3701 // calls canonicalizing them back.
3702 // (3) Count the number of double-precision FP ops, single-precision FP ops
3703 // and call sites. On Intel, we can get correct rounding either by
3704 // forcing singles to memory (requires extra stores and loads after each
3705 // FP bytecode) or we can set a rounding mode bit (requires setting and
3706 // clearing the mode bit around call sites). The mode bit is only used
3707 // if the relative frequency of single FP ops to calls is low enough.
3708 // This is a key transform for SPEC mpeg_audio.
3709 // (4) Detect infinite loops; blobs of code reachable from above but not
3710 // below. Several of the Code_Gen algorithms fail on such code shapes,
3711 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3712 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3713 // Detection is by looking for IfNodes where only 1 projection is
3714 // reachable from below or CatchNodes missing some targets.
3715 // (5) Assert for insane oop offsets in debug mode.
3716
3717 bool Compile::final_graph_reshaping() {
3718 // an infinite loop may have been eliminated by the optimizer,
3719 // in which case the graph will be empty.
3720 if (root()->req() == 1) {
3721 record_method_not_compilable("trivial infinite loop");
3722 return true;
3723 }
3724
3725 // Expensive nodes have their control input set to prevent the GVN
3726 // from freely commoning them. There's no GVN beyond this point so
3727 // no need to keep the control input. We want the expensive nodes to
3728 // be freely moved to the least frequent code path by gcm.
3729 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3730 for (int i = 0; i < expensive_count(); i++) {
3731 _expensive_nodes.at(i)->set_req(0, NULL);
3732 }
3733
3734 Final_Reshape_Counts frc;
3735
3736 // Visit everybody reachable!
3737 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3738 Node_Stack nstack(live_nodes() >> 1);
3739 final_graph_reshaping_walk(nstack, root(), frc);
3740
3741 // Check for unreachable (from below) code (i.e., infinite loops).
3742 for( uint i = 0; i < frc._tests.size(); i++ ) {
3743 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3744 // Get number of CFG targets.
3745 // Note that PCTables include exception targets after calls.
3746 uint required_outcnt = n->required_outcnt();
3747 if (n->outcnt() != required_outcnt) {
3748 // Check for a few special cases. Rethrow Nodes never take the
3749 // 'fall-thru' path, so expected kids is 1 less.
3750 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3751 if (n->in(0)->in(0)->is_Call()) {
3752 CallNode *call = n->in(0)->in(0)->as_Call();
3753 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3754 required_outcnt--; // Rethrow always has 1 less kid
3755 } else if (call->req() > TypeFunc::Parms &&
3756 call->is_CallDynamicJava()) {
3757 // Check for null receiver. In such case, the optimizer has
3758 // detected that the virtual call will always result in a null
3759 // pointer exception. The fall-through projection of this CatchNode
3760 // will not be populated.
3761 Node *arg0 = call->in(TypeFunc::Parms);
3762 if (arg0->is_Type() &&
3763 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3764 required_outcnt--;
3765 }
3766 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3767 call->req() > TypeFunc::Parms+1 &&
3768 call->is_CallStaticJava()) {
3769 // Check for negative array length. In such case, the optimizer has
3770 // detected that the allocation attempt will always result in an
3771 // exception. There is no fall-through projection of this CatchNode .
3772 Node *arg1 = call->in(TypeFunc::Parms+1);
3773 if (arg1->is_Type() &&
3774 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3775 required_outcnt--;
3776 }
3777 }
3778 }
3779 }
3780 // Recheck with a better notion of 'required_outcnt'
3781 if (n->outcnt() != required_outcnt) {
3782 record_method_not_compilable("malformed control flow");
3783 return true; // Not all targets reachable!
3784 }
3785 }
3786 // Check that I actually visited all kids. Unreached kids
3787 // must be infinite loops.
3788 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3789 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3790 record_method_not_compilable("infinite loop");
3791 return true; // Found unvisited kid; must be unreach
3792 }
3793
3794 // Here so verification code in final_graph_reshaping_walk()
3795 // always see an OuterStripMinedLoopEnd
3796 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
3797 IfNode* init_iff = n->as_If();
3798 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3799 n->subsume_by(iff, this);
3800 }
3801 }
3802
3803 #ifdef IA32
3804 // If original bytecodes contained a mixture of floats and doubles
3805 // check if the optimizer has made it homogenous, item (3).
3806 if (UseSSE == 0 &&
3807 frc.get_float_count() > 32 &&
3808 frc.get_double_count() == 0 &&
3809 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3810 set_24_bit_selection_and_mode(false, true);
3811 }
3812 #endif // IA32
3813
3814 set_java_calls(frc.get_java_call_count());
3815 set_inner_loops(frc.get_inner_loop_count());
3816
3817 // No infinite loops, no reason to bail out.
3818 return false;
3819 }
3820
3821 //-----------------------------too_many_traps----------------------------------
3822 // Report if there are too many traps at the current method and bci.
3823 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3824 bool Compile::too_many_traps(ciMethod* method,
3825 int bci,
3826 Deoptimization::DeoptReason reason) {
3827 ciMethodData* md = method->method_data();
3828 if (md->is_empty()) {
3829 // Assume the trap has not occurred, or that it occurred only
3830 // because of a transient condition during start-up in the interpreter.
3831 return false;
3832 }
3833 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3834 if (md->has_trap_at(bci, m, reason) != 0) {
3835 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3836 // Also, if there are multiple reasons, or if there is no per-BCI record,
3837 // assume the worst.
3838 if (log())
3839 log()->elem("observe trap='%s' count='%d'",
3840 Deoptimization::trap_reason_name(reason),
3841 md->trap_count(reason));
3842 return true;
3843 } else {
3844 // Ignore method/bci and see if there have been too many globally.
3845 return too_many_traps(reason, md);
3846 }
3847 }
3848
3849 // Less-accurate variant which does not require a method and bci.
3850 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3851 ciMethodData* logmd) {
3852 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3853 // Too many traps globally.
3854 // Note that we use cumulative trap_count, not just md->trap_count.
3855 if (log()) {
3856 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3857 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3858 Deoptimization::trap_reason_name(reason),
3859 mcount, trap_count(reason));
3860 }
3861 return true;
3862 } else {
3863 // The coast is clear.
3864 return false;
3865 }
3866 }
3867
3868 //--------------------------too_many_recompiles--------------------------------
3869 // Report if there are too many recompiles at the current method and bci.
3870 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3871 // Is not eager to return true, since this will cause the compiler to use
3872 // Action_none for a trap point, to avoid too many recompilations.
3873 bool Compile::too_many_recompiles(ciMethod* method,
3874 int bci,
3875 Deoptimization::DeoptReason reason) {
3876 ciMethodData* md = method->method_data();
3877 if (md->is_empty()) {
3878 // Assume the trap has not occurred, or that it occurred only
3879 // because of a transient condition during start-up in the interpreter.
3880 return false;
3881 }
3882 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3883 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3884 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3885 Deoptimization::DeoptReason per_bc_reason
3886 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3887 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3888 if ((per_bc_reason == Deoptimization::Reason_none
3889 || md->has_trap_at(bci, m, reason) != 0)
3890 // The trap frequency measure we care about is the recompile count:
3891 && md->trap_recompiled_at(bci, m)
3892 && md->overflow_recompile_count() >= bc_cutoff) {
3893 // Do not emit a trap here if it has already caused recompilations.
3894 // Also, if there are multiple reasons, or if there is no per-BCI record,
3895 // assume the worst.
3896 if (log())
3897 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3898 Deoptimization::trap_reason_name(reason),
3899 md->trap_count(reason),
3900 md->overflow_recompile_count());
3901 return true;
3902 } else if (trap_count(reason) != 0
3903 && decompile_count() >= m_cutoff) {
3904 // Too many recompiles globally, and we have seen this sort of trap.
3905 // Use cumulative decompile_count, not just md->decompile_count.
3906 if (log())
3907 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3908 Deoptimization::trap_reason_name(reason),
3909 md->trap_count(reason), trap_count(reason),
3910 md->decompile_count(), decompile_count());
3911 return true;
3912 } else {
3913 // The coast is clear.
3914 return false;
3915 }
3916 }
3917
3918 // Compute when not to trap. Used by matching trap based nodes and
3919 // NullCheck optimization.
3920 void Compile::set_allowed_deopt_reasons() {
3921 _allowed_reasons = 0;
3922 if (is_method_compilation()) {
3923 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3924 assert(rs < BitsPerInt, "recode bit map");
3925 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3926 _allowed_reasons |= nth_bit(rs);
3927 }
3928 }
3929 }
3930 }
3931
3932 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
3933 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
3934 }
3935
3936 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
3937 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
3938 }
3939
3940 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
3941 if (holder->is_initialized()) {
3942 return false;
3943 }
3944 if (holder->is_being_initialized()) {
3945 if (accessing_method->holder() == holder) {
3946 // Access inside a class. The barrier can be elided when access happens in <clinit>,
3947 // <init>, or a static method. In all those cases, there was an initialization
3948 // barrier on the holder klass passed.
3949 if (accessing_method->is_static_initializer() ||
3950 accessing_method->is_object_initializer() ||
3951 accessing_method->is_static()) {
3952 return false;
3953 }
3954 } else if (accessing_method->holder()->is_subclass_of(holder)) {
3955 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
3956 // In case of <init> or a static method, the barrier is on the subclass is not enough:
3957 // child class can become fully initialized while its parent class is still being initialized.
3958 if (accessing_method->is_static_initializer()) {
3959 return false;
3960 }
3961 }
3962 ciMethod* root = method(); // the root method of compilation
3963 if (root != accessing_method) {
3964 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
3965 }
3966 }
3967 return true;
3968 }
3969
3970 #ifndef PRODUCT
3971 //------------------------------verify_graph_edges---------------------------
3972 // Walk the Graph and verify that there is a one-to-one correspondence
3973 // between Use-Def edges and Def-Use edges in the graph.
3974 void Compile::verify_graph_edges(bool no_dead_code) {
3975 if (VerifyGraphEdges) {
3976 Unique_Node_List visited;
3977 // Call recursive graph walk to check edges
3978 _root->verify_edges(visited);
3979 if (no_dead_code) {
3980 // Now make sure that no visited node is used by an unvisited node.
3981 bool dead_nodes = false;
3982 Unique_Node_List checked;
3983 while (visited.size() > 0) {
3984 Node* n = visited.pop();
3985 checked.push(n);
3986 for (uint i = 0; i < n->outcnt(); i++) {
3987 Node* use = n->raw_out(i);
3988 if (checked.member(use)) continue; // already checked
3989 if (visited.member(use)) continue; // already in the graph
3990 if (use->is_Con()) continue; // a dead ConNode is OK
3991 // At this point, we have found a dead node which is DU-reachable.
3992 if (!dead_nodes) {
3993 tty->print_cr("*** Dead nodes reachable via DU edges:");
3994 dead_nodes = true;
3995 }
3996 use->dump(2);
3997 tty->print_cr("---");
3998 checked.push(use); // No repeats; pretend it is now checked.
3999 }
4000 }
4001 assert(!dead_nodes, "using nodes must be reachable from root");
4002 }
4003 }
4004 }
4005 #endif
4006
4007 // The Compile object keeps track of failure reasons separately from the ciEnv.
4008 // This is required because there is not quite a 1-1 relation between the
4009 // ciEnv and its compilation task and the Compile object. Note that one
4010 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4011 // to backtrack and retry without subsuming loads. Other than this backtracking
4012 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4013 // by the logic in C2Compiler.
4014 void Compile::record_failure(const char* reason) {
4015 if (log() != NULL) {
4016 log()->elem("failure reason='%s' phase='compile'", reason);
4017 }
4018 if (_failure_reason == NULL) {
4019 // Record the first failure reason.
4020 _failure_reason = reason;
4021 }
4022
4023 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4024 C->print_method(PHASE_FAILURE);
4025 }
4026 _root = NULL; // flush the graph, too
4027 }
4028
4029 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4030 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4031 _phase_name(name), _dolog(CITimeVerbose)
4032 {
4033 if (_dolog) {
4034 C = Compile::current();
4035 _log = C->log();
4036 } else {
4037 C = NULL;
4038 _log = NULL;
4039 }
4040 if (_log != NULL) {
4041 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4042 _log->stamp();
4043 _log->end_head();
4044 }
4045 }
4046
4047 Compile::TracePhase::~TracePhase() {
4048
4049 C = Compile::current();
4050 if (_dolog) {
4051 _log = C->log();
4052 } else {
4053 _log = NULL;
4054 }
4055
4056 #ifdef ASSERT
4057 if (PrintIdealNodeCount) {
4058 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4059 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4060 }
4061
4062 if (VerifyIdealNodeCount) {
4063 Compile::current()->print_missing_nodes();
4064 }
4065 #endif
4066
4067 if (_log != NULL) {
4068 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4069 }
4070 }
4071
4072 //----------------------------static_subtype_check-----------------------------
4073 // Shortcut important common cases when superklass is exact:
4074 // (0) superklass is java.lang.Object (can occur in reflective code)
4075 // (1) subklass is already limited to a subtype of superklass => always ok
4076 // (2) subklass does not overlap with superklass => always fail
4077 // (3) superklass has NO subtypes and we can check with a simple compare.
4078 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4079 if (StressReflectiveCode) {
4080 return SSC_full_test; // Let caller generate the general case.
4081 }
4082
4083 if (superk == env()->Object_klass()) {
4084 return SSC_always_true; // (0) this test cannot fail
4085 }
4086
4087 ciType* superelem = superk;
4088 ciType* subelem = subk;
4089 if (superelem->is_array_klass()) {
4090 superelem = superelem->as_array_klass()->base_element_type();
4091 }
4092 if (subelem->is_array_klass()) {
4093 subelem = subelem->as_array_klass()->base_element_type();
4094 }
4095
4096 if (!subk->is_interface()) { // cannot trust static interface types yet
4097 if (subk->is_subtype_of(superk)) {
4098 return SSC_always_true; // (1) false path dead; no dynamic test needed
4099 }
4100 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4101 !(subelem->is_klass() && subelem->as_klass()->is_interface()) &&
4102 !superk->is_subtype_of(subk)) {
4103 return SSC_always_false; // (2) true path dead; no dynamic test needed
4104 }
4105 }
4106
4107 // If casting to an instance klass, it must have no subtypes
4108 if (superk->is_interface()) {
4109 // Cannot trust interfaces yet.
4110 // %%% S.B. superk->nof_implementors() == 1
4111 } else if (superelem->is_instance_klass()) {
4112 ciInstanceKlass* ik = superelem->as_instance_klass();
4113 if (!ik->has_subklass() && !ik->is_interface()) {
4114 if (!ik->is_final()) {
4115 // Add a dependency if there is a chance of a later subclass.
4116 dependencies()->assert_leaf_type(ik);
4117 }
4118 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4119 }
4120 } else {
4121 // A primitive array type has no subtypes.
4122 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4123 }
4124
4125 return SSC_full_test;
4126 }
4127
4128 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4129 #ifdef _LP64
4130 // The scaled index operand to AddP must be a clean 64-bit value.
4131 // Java allows a 32-bit int to be incremented to a negative
4132 // value, which appears in a 64-bit register as a large
4133 // positive number. Using that large positive number as an
4134 // operand in pointer arithmetic has bad consequences.
4135 // On the other hand, 32-bit overflow is rare, and the possibility
4136 // can often be excluded, if we annotate the ConvI2L node with
4137 // a type assertion that its value is known to be a small positive
4138 // number. (The prior range check has ensured this.)
4139 // This assertion is used by ConvI2LNode::Ideal.
4140 int index_max = max_jint - 1; // array size is max_jint, index is one less
4141 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4142 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4143 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4144 #endif
4145 return idx;
4146 }
4147
4148 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4149 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4150 if (ctrl != NULL) {
4151 // Express control dependency by a CastII node with a narrow type.
4152 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4153 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4154 // node from floating above the range check during loop optimizations. Otherwise, the
4155 // ConvI2L node may be eliminated independently of the range check, causing the data path
4156 // to become TOP while the control path is still there (although it's unreachable).
4157 value->set_req(0, ctrl);
4158 value = phase->transform(value);
4159 }
4160 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4161 return phase->transform(new ConvI2LNode(value, ltype));
4162 }
4163
4164 void Compile::print_inlining_stream_free() {
4165 if (_print_inlining_stream != NULL) {
4166 _print_inlining_stream->~stringStream();
4167 _print_inlining_stream = NULL;
4168 }
4169 }
4170
4171 // The message about the current inlining is accumulated in
4172 // _print_inlining_stream and transfered into the _print_inlining_list
4173 // once we know whether inlining succeeds or not. For regular
4174 // inlining, messages are appended to the buffer pointed by
4175 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4176 // a new buffer is added after _print_inlining_idx in the list. This
4177 // way we can update the inlining message for late inlining call site
4178 // when the inlining is attempted again.
4179 void Compile::print_inlining_init() {
4180 if (print_inlining() || print_intrinsics()) {
4181 // print_inlining_init is actually called several times.
4182 print_inlining_stream_free();
4183 _print_inlining_stream = new stringStream();
4184 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4185 }
4186 }
4187
4188 void Compile::print_inlining_reinit() {
4189 if (print_inlining() || print_intrinsics()) {
4190 print_inlining_stream_free();
4191 // Re allocate buffer when we change ResourceMark
4192 _print_inlining_stream = new stringStream();
4193 }
4194 }
4195
4196 void Compile::print_inlining_reset() {
4197 _print_inlining_stream->reset();
4198 }
4199
4200 void Compile::print_inlining_commit() {
4201 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4202 // Transfer the message from _print_inlining_stream to the current
4203 // _print_inlining_list buffer and clear _print_inlining_stream.
4204 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4205 print_inlining_reset();
4206 }
4207
4208 void Compile::print_inlining_push() {
4209 // Add new buffer to the _print_inlining_list at current position
4210 _print_inlining_idx++;
4211 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4212 }
4213
4214 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4215 return _print_inlining_list->at(_print_inlining_idx);
4216 }
4217
4218 void Compile::print_inlining_update(CallGenerator* cg) {
4219 if (print_inlining() || print_intrinsics()) {
4220 if (cg->is_late_inline()) {
4221 if (print_inlining_current()->cg() != cg &&
4222 (print_inlining_current()->cg() != NULL ||
4223 print_inlining_current()->ss()->size() != 0)) {
4224 print_inlining_push();
4225 }
4226 print_inlining_commit();
4227 print_inlining_current()->set_cg(cg);
4228 } else {
4229 if (print_inlining_current()->cg() != NULL) {
4230 print_inlining_push();
4231 }
4232 print_inlining_commit();
4233 }
4234 }
4235 }
4236
4237 void Compile::print_inlining_move_to(CallGenerator* cg) {
4238 // We resume inlining at a late inlining call site. Locate the
4239 // corresponding inlining buffer so that we can update it.
4240 if (print_inlining() || print_intrinsics()) {
4241 for (int i = 0; i < _print_inlining_list->length(); i++) {
4242 if (_print_inlining_list->at(i)->cg() == cg) {
4243 _print_inlining_idx = i;
4244 return;
4245 }
4246 }
4247 ShouldNotReachHere();
4248 }
4249 }
4250
4251 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4252 if (print_inlining() || print_intrinsics()) {
4253 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4254 assert(print_inlining_current()->cg() == cg, "wrong entry");
4255 // replace message with new message
4256 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4257 print_inlining_commit();
4258 print_inlining_current()->set_cg(cg);
4259 }
4260 }
4261
4262 void Compile::print_inlining_assert_ready() {
4263 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4264 }
4265
4266 void Compile::process_print_inlining() {
4267 assert(_late_inlines.length() == 0, "not drained yet");
4268 if (print_inlining() || print_intrinsics()) {
4269 ResourceMark rm;
4270 stringStream ss;
4271 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
4272 for (int i = 0; i < _print_inlining_list->length(); i++) {
4273 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4274 ss.print("%s", pib->ss()->as_string());
4275 delete pib;
4276 DEBUG_ONLY(_print_inlining_list->at_put(i, NULL));
4277 }
4278 // Reset _print_inlining_list, it only contains destructed objects.
4279 // It is on the arena, so it will be freed when the arena is reset.
4280 _print_inlining_list = NULL;
4281 // _print_inlining_stream won't be used anymore, either.
4282 print_inlining_stream_free();
4283 size_t end = ss.size();
4284 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4285 strncpy(_print_inlining_output, ss.base(), end+1);
4286 _print_inlining_output[end] = 0;
4287 }
4288 }
4289
4290 void Compile::dump_print_inlining() {
4291 if (_print_inlining_output != NULL) {
4292 tty->print_raw(_print_inlining_output);
4293 }
4294 }
4295
4296 void Compile::log_late_inline(CallGenerator* cg) {
4297 if (log() != NULL) {
4298 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4299 cg->unique_id());
4300 JVMState* p = cg->call_node()->jvms();
4301 while (p != NULL) {
4302 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4303 p = p->caller();
4304 }
4305 log()->tail("late_inline");
4306 }
4307 }
4308
4309 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4310 log_late_inline(cg);
4311 if (log() != NULL) {
4312 log()->inline_fail(msg);
4313 }
4314 }
4315
4316 void Compile::log_inline_id(CallGenerator* cg) {
4317 if (log() != NULL) {
4318 // The LogCompilation tool needs a unique way to identify late
4319 // inline call sites. This id must be unique for this call site in
4320 // this compilation. Try to have it unique across compilations as
4321 // well because it can be convenient when grepping through the log
4322 // file.
4323 // Distinguish OSR compilations from others in case CICountOSR is
4324 // on.
4325 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4326 cg->set_unique_id(id);
4327 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4328 }
4329 }
4330
4331 void Compile::log_inline_failure(const char* msg) {
4332 if (C->log() != NULL) {
4333 C->log()->inline_fail(msg);
4334 }
4335 }
4336
4337
4338 // Dump inlining replay data to the stream.
4339 // Don't change thread state and acquire any locks.
4340 void Compile::dump_inline_data(outputStream* out) {
4341 InlineTree* inl_tree = ilt();
4342 if (inl_tree != NULL) {
4343 out->print(" inline %d", inl_tree->count());
4344 inl_tree->dump_replay_data(out);
4345 }
4346 }
4347
4348 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4349 if (n1->Opcode() < n2->Opcode()) return -1;
4350 else if (n1->Opcode() > n2->Opcode()) return 1;
4351
4352 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4353 for (uint i = 1; i < n1->req(); i++) {
4354 if (n1->in(i) < n2->in(i)) return -1;
4355 else if (n1->in(i) > n2->in(i)) return 1;
4356 }
4357
4358 return 0;
4359 }
4360
4361 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4362 Node* n1 = *n1p;
4363 Node* n2 = *n2p;
4364
4365 return cmp_expensive_nodes(n1, n2);
4366 }
4367
4368 void Compile::sort_expensive_nodes() {
4369 if (!expensive_nodes_sorted()) {
4370 _expensive_nodes.sort(cmp_expensive_nodes);
4371 }
4372 }
4373
4374 bool Compile::expensive_nodes_sorted() const {
4375 for (int i = 1; i < _expensive_nodes.length(); i++) {
4376 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4377 return false;
4378 }
4379 }
4380 return true;
4381 }
4382
4383 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4384 if (_expensive_nodes.length() == 0) {
4385 return false;
4386 }
4387
4388 assert(OptimizeExpensiveOps, "optimization off?");
4389
4390 // Take this opportunity to remove dead nodes from the list
4391 int j = 0;
4392 for (int i = 0; i < _expensive_nodes.length(); i++) {
4393 Node* n = _expensive_nodes.at(i);
4394 if (!n->is_unreachable(igvn)) {
4395 assert(n->is_expensive(), "should be expensive");
4396 _expensive_nodes.at_put(j, n);
4397 j++;
4398 }
4399 }
4400 _expensive_nodes.trunc_to(j);
4401
4402 // Then sort the list so that similar nodes are next to each other
4403 // and check for at least two nodes of identical kind with same data
4404 // inputs.
4405 sort_expensive_nodes();
4406
4407 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4408 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4409 return true;
4410 }
4411 }
4412
4413 return false;
4414 }
4415
4416 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4417 if (_expensive_nodes.length() == 0) {
4418 return;
4419 }
4420
4421 assert(OptimizeExpensiveOps, "optimization off?");
4422
4423 // Sort to bring similar nodes next to each other and clear the
4424 // control input of nodes for which there's only a single copy.
4425 sort_expensive_nodes();
4426
4427 int j = 0;
4428 int identical = 0;
4429 int i = 0;
4430 bool modified = false;
4431 for (; i < _expensive_nodes.length()-1; i++) {
4432 assert(j <= i, "can't write beyond current index");
4433 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4434 identical++;
4435 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4436 continue;
4437 }
4438 if (identical > 0) {
4439 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4440 identical = 0;
4441 } else {
4442 Node* n = _expensive_nodes.at(i);
4443 igvn.replace_input_of(n, 0, NULL);
4444 igvn.hash_insert(n);
4445 modified = true;
4446 }
4447 }
4448 if (identical > 0) {
4449 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4450 } else if (_expensive_nodes.length() >= 1) {
4451 Node* n = _expensive_nodes.at(i);
4452 igvn.replace_input_of(n, 0, NULL);
4453 igvn.hash_insert(n);
4454 modified = true;
4455 }
4456 _expensive_nodes.trunc_to(j);
4457 if (modified) {
4458 igvn.optimize();
4459 }
4460 }
4461
4462 void Compile::add_expensive_node(Node * n) {
4463 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4464 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4465 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4466 if (OptimizeExpensiveOps) {
4467 _expensive_nodes.append(n);
4468 } else {
4469 // Clear control input and let IGVN optimize expensive nodes if
4470 // OptimizeExpensiveOps is off.
4471 n->set_req(0, NULL);
4472 }
4473 }
4474
4475 /**
4476 * Track coarsened Lock and Unlock nodes.
4477 */
4478
4479 class Lock_List : public Node_List {
4480 uint _origin_cnt;
4481 public:
4482 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4483 uint origin_cnt() const { return _origin_cnt; }
4484 };
4485
4486 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4487 int length = locks.length();
4488 if (length > 0) {
4489 // Have to keep this list until locks elimination during Macro nodes elimination.
4490 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4491 for (int i = 0; i < length; i++) {
4492 AbstractLockNode* lock = locks.at(i);
4493 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4494 locks_list->push(lock);
4495 }
4496 _coarsened_locks.append(locks_list);
4497 }
4498 }
4499
4500 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4501 int count = coarsened_count();
4502 for (int i = 0; i < count; i++) {
4503 Node_List* locks_list = _coarsened_locks.at(i);
4504 for (uint j = 0; j < locks_list->size(); j++) {
4505 Node* lock = locks_list->at(j);
4506 assert(lock->is_AbstractLock(), "sanity");
4507 if (!useful.member(lock)) {
4508 locks_list->yank(lock);
4509 }
4510 }
4511 }
4512 }
4513
4514 void Compile::remove_coarsened_lock(Node* n) {
4515 if (n->is_AbstractLock()) {
4516 int count = coarsened_count();
4517 for (int i = 0; i < count; i++) {
4518 Node_List* locks_list = _coarsened_locks.at(i);
4519 locks_list->yank(n);
4520 }
4521 }
4522 }
4523
4524 bool Compile::coarsened_locks_consistent() {
4525 int count = coarsened_count();
4526 for (int i = 0; i < count; i++) {
4527 bool unbalanced = false;
4528 bool modified = false; // track locks kind modifications
4529 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4530 uint size = locks_list->size();
4531 if (size == 0) {
4532 unbalanced = false; // All locks were eliminated - good
4533 } else if (size != locks_list->origin_cnt()) {
4534 unbalanced = true; // Some locks were removed from list
4535 } else {
4536 for (uint j = 0; j < size; j++) {
4537 Node* lock = locks_list->at(j);
4538 // All nodes in group should have the same state (modified or not)
4539 if (!lock->as_AbstractLock()->is_coarsened()) {
4540 if (j == 0) {
4541 // first on list was modified, the rest should be too for consistency
4542 modified = true;
4543 } else if (!modified) {
4544 // this lock was modified but previous locks on the list were not
4545 unbalanced = true;
4546 break;
4547 }
4548 } else if (modified) {
4549 // previous locks on list were modified but not this lock
4550 unbalanced = true;
4551 break;
4552 }
4553 }
4554 }
4555 if (unbalanced) {
4556 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4557 #ifdef ASSERT
4558 if (PrintEliminateLocks) {
4559 tty->print_cr("=== unbalanced coarsened locks ===");
4560 for (uint l = 0; l < size; l++) {
4561 locks_list->at(l)->dump();
4562 }
4563 }
4564 #endif
4565 record_failure(C2Compiler::retry_no_locks_coarsening());
4566 return false;
4567 }
4568 }
4569 return true;
4570 }
4571
4572 /**
4573 * Remove the speculative part of types and clean up the graph
4574 */
4575 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4576 if (UseTypeSpeculation) {
4577 Unique_Node_List worklist;
4578 worklist.push(root());
4579 int modified = 0;
4580 // Go over all type nodes that carry a speculative type, drop the
4581 // speculative part of the type and enqueue the node for an igvn
4582 // which may optimize it out.
4583 for (uint next = 0; next < worklist.size(); ++next) {
4584 Node *n = worklist.at(next);
4585 if (n->is_Type()) {
4586 TypeNode* tn = n->as_Type();
4587 const Type* t = tn->type();
4588 const Type* t_no_spec = t->remove_speculative();
4589 if (t_no_spec != t) {
4590 bool in_hash = igvn.hash_delete(n);
4591 assert(in_hash, "node should be in igvn hash table");
4592 tn->set_type(t_no_spec);
4593 igvn.hash_insert(n);
4594 igvn._worklist.push(n); // give it a chance to go away
4595 modified++;
4596 }
4597 }
4598 // Iterate over outs - endless loops is unreachable from below
4599 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4600 Node *m = n->fast_out(i);
4601 if (not_a_node(m)) {
4602 continue;
4603 }
4604 worklist.push(m);
4605 }
4606 }
4607 // Drop the speculative part of all types in the igvn's type table
4608 igvn.remove_speculative_types();
4609 if (modified > 0) {
4610 igvn.optimize();
4611 }
4612 #ifdef ASSERT
4613 // Verify that after the IGVN is over no speculative type has resurfaced
4614 worklist.clear();
4615 worklist.push(root());
4616 for (uint next = 0; next < worklist.size(); ++next) {
4617 Node *n = worklist.at(next);
4618 const Type* t = igvn.type_or_null(n);
4619 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4620 if (n->is_Type()) {
4621 t = n->as_Type()->type();
4622 assert(t == t->remove_speculative(), "no more speculative types");
4623 }
4624 // Iterate over outs - endless loops is unreachable from below
4625 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4626 Node *m = n->fast_out(i);
4627 if (not_a_node(m)) {
4628 continue;
4629 }
4630 worklist.push(m);
4631 }
4632 }
4633 igvn.check_no_speculative_types();
4634 #endif
4635 }
4636 }
4637
4638 // Auxiliary methods to support randomized stressing/fuzzing.
4639
4640 int Compile::random() {
4641 _stress_seed = os::next_random(_stress_seed);
4642 return static_cast<int>(_stress_seed);
4643 }
4644
4645 // This method can be called the arbitrary number of times, with current count
4646 // as the argument. The logic allows selecting a single candidate from the
4647 // running list of candidates as follows:
4648 // int count = 0;
4649 // Cand* selected = null;
4650 // while(cand = cand->next()) {
4651 // if (randomized_select(++count)) {
4652 // selected = cand;
4653 // }
4654 // }
4655 //
4656 // Including count equalizes the chances any candidate is "selected".
4657 // This is useful when we don't have the complete list of candidates to choose
4658 // from uniformly. In this case, we need to adjust the randomicity of the
4659 // selection, or else we will end up biasing the selection towards the latter
4660 // candidates.
4661 //
4662 // Quick back-envelope calculation shows that for the list of n candidates
4663 // the equal probability for the candidate to persist as "best" can be
4664 // achieved by replacing it with "next" k-th candidate with the probability
4665 // of 1/k. It can be easily shown that by the end of the run, the
4666 // probability for any candidate is converged to 1/n, thus giving the
4667 // uniform distribution among all the candidates.
4668 //
4669 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4670 #define RANDOMIZED_DOMAIN_POW 29
4671 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4672 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4673 bool Compile::randomized_select(int count) {
4674 assert(count > 0, "only positive");
4675 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4676 }
4677
4678 CloneMap& Compile::clone_map() { return _clone_map; }
4679 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4680
4681 void NodeCloneInfo::dump() const {
4682 tty->print(" {%d:%d} ", idx(), gen());
4683 }
4684
4685 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4686 uint64_t val = value(old->_idx);
4687 NodeCloneInfo cio(val);
4688 assert(val != 0, "old node should be in the map");
4689 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4690 insert(nnn->_idx, cin.get());
4691 #ifndef PRODUCT
4692 if (is_debug()) {
4693 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4694 }
4695 #endif
4696 }
4697
4698 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4699 NodeCloneInfo cio(value(old->_idx));
4700 if (cio.get() == 0) {
4701 cio.set(old->_idx, 0);
4702 insert(old->_idx, cio.get());
4703 #ifndef PRODUCT
4704 if (is_debug()) {
4705 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4706 }
4707 #endif
4708 }
4709 clone(old, nnn, gen);
4710 }
4711
4712 int CloneMap::max_gen() const {
4713 int g = 0;
4714 DictI di(_dict);
4715 for(; di.test(); ++di) {
4716 int t = gen(di._key);
4717 if (g < t) {
4718 g = t;
4719 #ifndef PRODUCT
4720 if (is_debug()) {
4721 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4722 }
4723 #endif
4724 }
4725 }
4726 return g;
4727 }
4728
4729 void CloneMap::dump(node_idx_t key) const {
4730 uint64_t val = value(key);
4731 if (val != 0) {
4732 NodeCloneInfo ni(val);
4733 ni.dump();
4734 }
4735 }
4736
4737 // Move Allocate nodes to the start of the list
4738 void Compile::sort_macro_nodes() {
4739 int count = macro_count();
4740 int allocates = 0;
4741 for (int i = 0; i < count; i++) {
4742 Node* n = macro_node(i);
4743 if (n->is_Allocate()) {
4744 if (i != allocates) {
4745 Node* tmp = macro_node(allocates);
4746 _macro_nodes.at_put(allocates, n);
4747 _macro_nodes.at_put(i, tmp);
4748 }
4749 allocates++;
4750 }
4751 }
4752 }
4753
4754 void Compile::print_method(CompilerPhaseType cpt, const char *name, int level) {
4755 EventCompilerPhase event;
4756 if (event.should_commit()) {
4757 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
4758 }
4759 #ifndef PRODUCT
4760 if (should_print(level)) {
4761 _printer->print_method(name, level);
4762 }
4763 #endif
4764 C->_latest_stage_start_counter.stamp();
4765 }
4766
4767 void Compile::print_method(CompilerPhaseType cpt, int level, int idx) {
4768 char output[1024];
4769 #ifndef PRODUCT
4770 if (idx != 0) {
4771 jio_snprintf(output, sizeof(output), "%s:%d", CompilerPhaseTypeHelper::to_string(cpt), idx);
4772 } else {
4773 jio_snprintf(output, sizeof(output), "%s", CompilerPhaseTypeHelper::to_string(cpt));
4774 }
4775 #endif
4776 print_method(cpt, output, level);
4777 }
4778
4779 void Compile::print_method(CompilerPhaseType cpt, Node* n, int level) {
4780 ResourceMark rm;
4781 stringStream ss;
4782 ss.print_raw(CompilerPhaseTypeHelper::to_string(cpt));
4783 if (n != NULL) {
4784 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
4785 } else {
4786 ss.print_raw(": NULL");
4787 }
4788 C->print_method(cpt, ss.as_string(), level);
4789 }
4790
4791 void Compile::end_method(int level) {
4792 EventCompilerPhase event;
4793 if (event.should_commit()) {
4794 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, level);
4795 }
4796
4797 #ifndef PRODUCT
4798 if (_method != NULL && should_print(level)) {
4799 _printer->end_method();
4800 }
4801 #endif
4802 }
4803
4804
4805 #ifndef PRODUCT
4806 IdealGraphPrinter* Compile::_debug_file_printer = NULL;
4807 IdealGraphPrinter* Compile::_debug_network_printer = NULL;
4808
4809 // Called from debugger. Prints method to the default file with the default phase name.
4810 // This works regardless of any Ideal Graph Visualizer flags set or not.
4811 void igv_print() {
4812 Compile::current()->igv_print_method_to_file();
4813 }
4814
4815 // Same as igv_print() above but with a specified phase name.
4816 void igv_print(const char* phase_name) {
4817 Compile::current()->igv_print_method_to_file(phase_name);
4818 }
4819
4820 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
4821 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
4822 // This works regardless of any Ideal Graph Visualizer flags set or not.
4823 void igv_print(bool network) {
4824 if (network) {
4825 Compile::current()->igv_print_method_to_network();
4826 } else {
4827 Compile::current()->igv_print_method_to_file();
4828 }
4829 }
4830
4831 // Same as igv_print(bool network) above but with a specified phase name.
4832 void igv_print(bool network, const char* phase_name) {
4833 if (network) {
4834 Compile::current()->igv_print_method_to_network(phase_name);
4835 } else {
4836 Compile::current()->igv_print_method_to_file(phase_name);
4837 }
4838 }
4839
4840 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
4841 void igv_print_default() {
4842 Compile::current()->print_method(PHASE_DEBUG, 0);
4843 }
4844
4845 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
4846 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
4847 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
4848 void igv_append() {
4849 Compile::current()->igv_print_method_to_file("Debug", true);
4850 }
4851
4852 // Same as igv_append() above but with a specified phase name.
4853 void igv_append(const char* phase_name) {
4854 Compile::current()->igv_print_method_to_file(phase_name, true);
4855 }
4856
4857 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
4858 const char* file_name = "custom_debug.xml";
4859 if (_debug_file_printer == NULL) {
4860 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
4861 } else {
4862 _debug_file_printer->update_compiled_method(C->method());
4863 }
4864 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
4865 _debug_file_printer->print(phase_name, (Node*)C->root());
4866 }
4867
4868 void Compile::igv_print_method_to_network(const char* phase_name) {
4869 if (_debug_network_printer == NULL) {
4870 _debug_network_printer = new IdealGraphPrinter(C);
4871 } else {
4872 _debug_network_printer->update_compiled_method(C->method());
4873 }
4874 tty->print_cr("Method printed over network stream to IGV");
4875 _debug_network_printer->print(phase_name, (Node*)C->root());
4876 }
4877 #endif
4878
4879 void Compile::add_native_invoker(RuntimeStub* stub) {
4880 _native_invokers.append(stub);
4881 }
4882
4883 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
4884 if (type != NULL && phase->type(value)->higher_equal(type)) {
4885 return value;
4886 }
4887 Node* result = NULL;
4888 if (bt == T_BYTE) {
4889 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
4890 result = new RShiftINode(result, phase->intcon(24));
4891 } else if (bt == T_BOOLEAN) {
4892 result = new AndINode(value, phase->intcon(0xFF));
4893 } else if (bt == T_CHAR) {
4894 result = new AndINode(value,phase->intcon(0xFFFF));
4895 } else {
4896 assert(bt == T_SHORT, "unexpected narrow type");
4897 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
4898 result = new RShiftINode(result, phase->intcon(16));
4899 }
4900 if (transform_res) {
4901 result = phase->transform(result);
4902 }
4903 return result;
4904 }
4905