1 /*
2 * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSet.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "jfr/jfrEvents.hpp"
39 #include "memory/resourceArea.hpp"
40 #include "opto/addnode.hpp"
41 #include "opto/block.hpp"
42 #include "opto/c2compiler.hpp"
43 #include "opto/callGenerator.hpp"
44 #include "opto/callnode.hpp"
45 #include "opto/castnode.hpp"
46 #include "opto/cfgnode.hpp"
47 #include "opto/chaitin.hpp"
48 #include "opto/compile.hpp"
49 #include "opto/connode.hpp"
50 #include "opto/convertnode.hpp"
51 #include "opto/divnode.hpp"
52 #include "opto/escape.hpp"
53 #include "opto/idealGraphPrinter.hpp"
54 #include "opto/loopnode.hpp"
55 #include "opto/machnode.hpp"
56 #include "opto/macro.hpp"
57 #include "opto/matcher.hpp"
58 #include "opto/mathexactnode.hpp"
59 #include "opto/memnode.hpp"
60 #include "opto/mulnode.hpp"
61 #include "opto/narrowptrnode.hpp"
62 #include "opto/node.hpp"
63 #include "opto/opcodes.hpp"
64 #include "opto/output.hpp"
65 #include "opto/parse.hpp"
66 #include "opto/phaseX.hpp"
67 #include "opto/rootnode.hpp"
68 #include "opto/runtime.hpp"
69 #include "opto/stringopts.hpp"
70 #include "opto/type.hpp"
71 #include "opto/vectornode.hpp"
72 #include "runtime/arguments.hpp"
73 #include "runtime/sharedRuntime.hpp"
74 #include "runtime/signature.hpp"
75 #include "runtime/stubRoutines.hpp"
76 #include "runtime/timer.hpp"
77 #include "utilities/align.hpp"
78 #include "utilities/copy.hpp"
79 #include "utilities/macros.hpp"
80 #include "utilities/resourceHash.hpp"
81
82
83 // -------------------- Compile::mach_constant_base_node -----------------------
84 // Constant table base node singleton.
mach_constant_base_node()85 MachConstantBaseNode* Compile::mach_constant_base_node() {
86 if (_mach_constant_base_node == NULL) {
87 _mach_constant_base_node = new MachConstantBaseNode();
88 _mach_constant_base_node->add_req(C->root());
89 }
90 return _mach_constant_base_node;
91 }
92
93
94 /// Support for intrinsics.
95
96 // Return the index at which m must be inserted (or already exists).
97 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
98 class IntrinsicDescPair {
99 private:
100 ciMethod* _m;
101 bool _is_virtual;
102 public:
IntrinsicDescPair(ciMethod * m,bool is_virtual)103 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
compare(IntrinsicDescPair * const & key,CallGenerator * const & elt)104 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
105 ciMethod* m= elt->method();
106 ciMethod* key_m = key->_m;
107 if (key_m < m) return -1;
108 else if (key_m > m) return 1;
109 else {
110 bool is_virtual = elt->is_virtual();
111 bool key_virtual = key->_is_virtual;
112 if (key_virtual < is_virtual) return -1;
113 else if (key_virtual > is_virtual) return 1;
114 else return 0;
115 }
116 }
117 };
intrinsic_insertion_index(ciMethod * m,bool is_virtual,bool & found)118 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
119 #ifdef ASSERT
120 for (int i = 1; i < _intrinsics->length(); i++) {
121 CallGenerator* cg1 = _intrinsics->at(i-1);
122 CallGenerator* cg2 = _intrinsics->at(i);
123 assert(cg1->method() != cg2->method()
124 ? cg1->method() < cg2->method()
125 : cg1->is_virtual() < cg2->is_virtual(),
126 "compiler intrinsics list must stay sorted");
127 }
128 #endif
129 IntrinsicDescPair pair(m, is_virtual);
130 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
131 }
132
register_intrinsic(CallGenerator * cg)133 void Compile::register_intrinsic(CallGenerator* cg) {
134 if (_intrinsics == NULL) {
135 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
136 }
137 int len = _intrinsics->length();
138 bool found = false;
139 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
140 assert(!found, "registering twice");
141 _intrinsics->insert_before(index, cg);
142 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
143 }
144
find_intrinsic(ciMethod * m,bool is_virtual)145 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
146 assert(m->is_loaded(), "don't try this on unloaded methods");
147 if (_intrinsics != NULL) {
148 bool found = false;
149 int index = intrinsic_insertion_index(m, is_virtual, found);
150 if (found) {
151 return _intrinsics->at(index);
152 }
153 }
154 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
155 if (m->intrinsic_id() != vmIntrinsics::_none &&
156 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
157 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
158 if (cg != NULL) {
159 // Save it for next time:
160 register_intrinsic(cg);
161 return cg;
162 } else {
163 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
164 }
165 }
166 return NULL;
167 }
168
169 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
170 // in library_call.cpp.
171
172
173 #ifndef PRODUCT
174 // statistics gathering...
175
176 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
177 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
178
gather_intrinsic_statistics(vmIntrinsics::ID id,bool is_virtual,int flags)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[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[id] += 1);
189 if (count == 1) {
190 changed = true; // first time
191 }
192 // increment the overall count also:
193 _intrinsic_hist_count[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[id] = (jubyte) (oflags | flags);
210 }
211 // update the overall flags also:
212 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
213 return changed;
214 }
215
format_flags(int flags,char * buf)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
print_intrinsic_statistics()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[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 (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
238 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
239 int flags = _intrinsic_hist_flags[id];
240 juint count = _intrinsic_hist_count[id];
241 if ((flags | count) != 0) {
242 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
243 }
244 }
245 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
246 if (xtty != NULL) xtty->tail("statistics");
247 }
248
print_statistics()249 void Compile::print_statistics() {
250 { ttyLocker ttyl;
251 if (xtty != NULL) xtty->head("statistics type='opto'");
252 Parse::print_statistics();
253 PhaseCCP::print_statistics();
254 PhaseRegAlloc::print_statistics();
255 PhaseOutput::print_statistics();
256 PhasePeephole::print_statistics();
257 PhaseIdealLoop::print_statistics();
258 if (xtty != NULL) xtty->tail("statistics");
259 }
260 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
261 // put this under its own <statistics> element.
262 print_intrinsic_statistics();
263 }
264 }
265 #endif //PRODUCT
266
gvn_replace_by(Node * n,Node * nn)267 void Compile::gvn_replace_by(Node* n, Node* nn) {
268 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
269 Node* use = n->last_out(i);
270 bool is_in_table = initial_gvn()->hash_delete(use);
271 uint uses_found = 0;
272 for (uint j = 0; j < use->len(); j++) {
273 if (use->in(j) == n) {
274 if (j < use->req())
275 use->set_req(j, nn);
276 else
277 use->set_prec(j, nn);
278 uses_found++;
279 }
280 }
281 if (is_in_table) {
282 // reinsert into table
283 initial_gvn()->hash_find_insert(use);
284 }
285 record_for_igvn(use);
286 i -= uses_found; // we deleted 1 or more copies of this edge
287 }
288 }
289
290
not_a_node(const Node * n)291 static inline bool not_a_node(const Node* n) {
292 if (n == NULL) return true;
293 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
294 if (*(address*)n == badAddress) return true; // kill by Node::destruct
295 return false;
296 }
297
298 // Identify all nodes that are reachable from below, useful.
299 // Use breadth-first pass that records state in a Unique_Node_List,
300 // recursive traversal is slower.
identify_useful_nodes(Unique_Node_List & useful)301 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
302 int estimated_worklist_size = live_nodes();
303 useful.map( estimated_worklist_size, NULL ); // preallocate space
304
305 // Initialize worklist
306 if (root() != NULL) { useful.push(root()); }
307 // If 'top' is cached, declare it useful to preserve cached node
308 if( cached_top_node() ) { useful.push(cached_top_node()); }
309
310 // Push all useful nodes onto the list, breadthfirst
311 for( uint next = 0; next < useful.size(); ++next ) {
312 assert( next < unique(), "Unique useful nodes < total nodes");
313 Node *n = useful.at(next);
314 uint max = n->len();
315 for( uint i = 0; i < max; ++i ) {
316 Node *m = n->in(i);
317 if (not_a_node(m)) continue;
318 useful.push(m);
319 }
320 }
321 }
322
323 // Update dead_node_list with any missing dead nodes using useful
324 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
update_dead_node_list(Unique_Node_List & useful)325 void Compile::update_dead_node_list(Unique_Node_List &useful) {
326 uint max_idx = unique();
327 VectorSet& useful_node_set = useful.member_set();
328
329 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
330 // If node with index node_idx is not in useful set,
331 // mark it as dead in dead node list.
332 if (!useful_node_set.test(node_idx)) {
333 record_dead_node(node_idx);
334 }
335 }
336 }
337
remove_useless_late_inlines(GrowableArray<CallGenerator * > * inlines,Unique_Node_List & useful)338 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
339 int shift = 0;
340 for (int i = 0; i < inlines->length(); i++) {
341 CallGenerator* cg = inlines->at(i);
342 CallNode* call = cg->call_node();
343 if (shift > 0) {
344 inlines->at_put(i-shift, cg);
345 }
346 if (!useful.member(call)) {
347 shift++;
348 }
349 }
350 inlines->trunc_to(inlines->length()-shift);
351 }
352
353 // Disconnect all useless nodes by disconnecting those at the boundary.
remove_useless_nodes(Unique_Node_List & useful)354 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
355 uint next = 0;
356 while (next < useful.size()) {
357 Node *n = useful.at(next++);
358 if (n->is_SafePoint()) {
359 // We're done with a parsing phase. Replaced nodes are not valid
360 // beyond that point.
361 n->as_SafePoint()->delete_replaced_nodes();
362 }
363 // Use raw traversal of out edges since this code removes out edges
364 int max = n->outcnt();
365 for (int j = 0; j < max; ++j) {
366 Node* child = n->raw_out(j);
367 if (! useful.member(child)) {
368 assert(!child->is_top() || child != top(),
369 "If top is cached in Compile object it is in useful list");
370 // Only need to remove this out-edge to the useless node
371 n->raw_del_out(j);
372 --j;
373 --max;
374 }
375 }
376 if (n->outcnt() == 1 && n->has_special_unique_user()) {
377 record_for_igvn(n->unique_out());
378 }
379 }
380 // Remove useless macro and predicate opaq nodes
381 for (int i = C->macro_count()-1; i >= 0; i--) {
382 Node* n = C->macro_node(i);
383 if (!useful.member(n)) {
384 remove_macro_node(n);
385 }
386 }
387 // Remove useless CastII nodes with range check dependency
388 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
389 Node* cast = range_check_cast_node(i);
390 if (!useful.member(cast)) {
391 remove_range_check_cast(cast);
392 }
393 }
394 // Remove useless expensive nodes
395 for (int i = C->expensive_count()-1; i >= 0; i--) {
396 Node* n = C->expensive_node(i);
397 if (!useful.member(n)) {
398 remove_expensive_node(n);
399 }
400 }
401 // Remove useless Opaque4 nodes
402 for (int i = opaque4_count() - 1; i >= 0; i--) {
403 Node* opaq = opaque4_node(i);
404 if (!useful.member(opaq)) {
405 remove_opaque4_node(opaq);
406 }
407 }
408 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
409 bs->eliminate_useless_gc_barriers(useful, this);
410 // clean up the late inline lists
411 remove_useless_late_inlines(&_string_late_inlines, useful);
412 remove_useless_late_inlines(&_boxing_late_inlines, useful);
413 remove_useless_late_inlines(&_late_inlines, useful);
414 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
415 }
416
417 // ============================================================================
418 //------------------------------CompileWrapper---------------------------------
419 class CompileWrapper : public StackObj {
420 Compile *const _compile;
421 public:
422 CompileWrapper(Compile* compile);
423
424 ~CompileWrapper();
425 };
426
CompileWrapper(Compile * compile)427 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
428 // the Compile* pointer is stored in the current ciEnv:
429 ciEnv* env = compile->env();
430 assert(env == ciEnv::current(), "must already be a ciEnv active");
431 assert(env->compiler_data() == NULL, "compile already active?");
432 env->set_compiler_data(compile);
433 assert(compile == Compile::current(), "sanity");
434
435 compile->set_type_dict(NULL);
436 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
437 compile->clone_map().set_clone_idx(0);
438 compile->set_type_last_size(0);
439 compile->set_last_tf(NULL, NULL);
440 compile->set_indexSet_arena(NULL);
441 compile->set_indexSet_free_block_list(NULL);
442 compile->init_type_arena();
443 Type::Initialize(compile);
444 _compile->begin_method();
445 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
446 }
~CompileWrapper()447 CompileWrapper::~CompileWrapper() {
448 _compile->end_method();
449 _compile->env()->set_compiler_data(NULL);
450 }
451
452
453 //----------------------------print_compile_messages---------------------------
print_compile_messages()454 void Compile::print_compile_messages() {
455 #ifndef PRODUCT
456 // Check if recompiling
457 if (_subsume_loads == false && PrintOpto) {
458 // Recompiling without allowing machine instructions to subsume loads
459 tty->print_cr("*********************************************************");
460 tty->print_cr("** Bailout: Recompile without subsuming loads **");
461 tty->print_cr("*********************************************************");
462 }
463 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
464 // Recompiling without escape analysis
465 tty->print_cr("*********************************************************");
466 tty->print_cr("** Bailout: Recompile without escape analysis **");
467 tty->print_cr("*********************************************************");
468 }
469 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
470 // Recompiling without boxing elimination
471 tty->print_cr("*********************************************************");
472 tty->print_cr("** Bailout: Recompile without boxing elimination **");
473 tty->print_cr("*********************************************************");
474 }
475 if (C->directive()->BreakAtCompileOption) {
476 // Open the debugger when compiling this method.
477 tty->print("### Breaking when compiling: ");
478 method()->print_short_name();
479 tty->cr();
480 BREAKPOINT;
481 }
482
483 if( PrintOpto ) {
484 if (is_osr_compilation()) {
485 tty->print("[OSR]%3d", _compile_id);
486 } else {
487 tty->print("%3d", _compile_id);
488 }
489 }
490 #endif
491 }
492
493 // ============================================================================
494 //------------------------------Compile standard-------------------------------
debug_only(int Compile::_debug_idx=100000;)495 debug_only( int Compile::_debug_idx = 100000; )
496
497 // Compile a method. entry_bci is -1 for normal compilations and indicates
498 // the continuation bci for on stack replacement.
499
500
501 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
502 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
503 : Phase(Compiler),
504 _compile_id(ci_env->compile_id()),
505 _save_argument_registers(false),
506 _subsume_loads(subsume_loads),
507 _do_escape_analysis(do_escape_analysis),
508 _eliminate_boxing(eliminate_boxing),
509 _method(target),
510 _entry_bci(osr_bci),
511 _stub_function(NULL),
512 _stub_name(NULL),
513 _stub_entry_point(NULL),
514 _max_node_limit(MaxNodeLimit),
515 _inlining_progress(false),
516 _inlining_incrementally(false),
517 _do_cleanup(false),
518 _has_reserved_stack_access(target->has_reserved_stack_access()),
519 #ifndef PRODUCT
520 _trace_opto_output(directive->TraceOptoOutputOption),
521 _print_ideal(directive->PrintIdealOption),
522 #endif
523 _has_method_handle_invokes(false),
524 _clinit_barrier_on_entry(false),
525 _comp_arena(mtCompiler),
526 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
527 _env(ci_env),
528 _directive(directive),
529 _log(ci_env->log()),
530 _failure_reason(NULL),
531 _congraph(NULL),
532 #ifndef PRODUCT
533 _printer(IdealGraphPrinter::printer()),
534 #endif
535 _dead_node_list(comp_arena()),
536 _dead_node_count(0),
537 _node_arena(mtCompiler),
538 _old_arena(mtCompiler),
539 _mach_constant_base_node(NULL),
540 _Compile_types(mtCompiler),
541 _initial_gvn(NULL),
542 _for_igvn(NULL),
543 _warm_calls(NULL),
544 _late_inlines(comp_arena(), 2, 0, NULL),
545 _string_late_inlines(comp_arena(), 2, 0, NULL),
546 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
547 _late_inlines_pos(0),
548 _number_of_mh_late_inlines(0),
549 _print_inlining_stream(NULL),
550 _print_inlining_list(NULL),
551 _print_inlining_idx(0),
552 _print_inlining_output(NULL),
553 _replay_inline_data(NULL),
554 _java_calls(0),
555 _inner_loops(0),
556 _interpreter_frame_size(0)
557 #ifndef PRODUCT
558 , _in_dump_cnt(0)
559 #endif
560 {
561 C = this;
562 #ifndef PRODUCT
563 if (_printer != NULL) {
564 _printer->set_compile(this);
565 }
566 #endif
567 CompileWrapper cw(this);
568
569 if (CITimeVerbose) {
570 tty->print(" ");
571 target->holder()->name()->print();
572 tty->print(".");
573 target->print_short_name();
574 tty->print(" ");
575 }
576 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
577 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
578
579 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
580 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
581 // We can always print a disassembly, either abstract (hex dump) or
582 // with the help of a suitable hsdis library. Thus, we should not
583 // couple print_assembly and print_opto_assembly controls.
584 // But: always print opto and regular assembly on compile command 'print'.
585 bool print_assembly = directive->PrintAssemblyOption;
586 set_print_assembly(print_opto_assembly || print_assembly);
587 #else
588 set_print_assembly(false); // must initialize.
589 #endif
590
591 #ifndef PRODUCT
592 set_parsed_irreducible_loop(false);
593
594 if (directive->ReplayInlineOption) {
595 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
596 }
597 #endif
598 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
599 set_print_intrinsics(directive->PrintIntrinsicsOption);
600 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
601
602 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
603 // Make sure the method being compiled gets its own MDO,
604 // so we can at least track the decompile_count().
605 // Need MDO to record RTM code generation state.
606 method()->ensure_method_data();
607 }
608
609 Init(::AliasLevel);
610
611
612 print_compile_messages();
613
614 _ilt = InlineTree::build_inline_tree_root();
615
616 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
617 assert(num_alias_types() >= AliasIdxRaw, "");
618
619 #define MINIMUM_NODE_HASH 1023
620 // Node list that Iterative GVN will start with
621 Unique_Node_List for_igvn(comp_arena());
622 set_for_igvn(&for_igvn);
623
624 // GVN that will be run immediately on new nodes
625 uint estimated_size = method()->code_size()*4+64;
626 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
627 PhaseGVN gvn(node_arena(), estimated_size);
628 set_initial_gvn(&gvn);
629
630 print_inlining_init();
631 { // Scope for timing the parser
632 TracePhase tp("parse", &timers[_t_parser]);
633
634 // Put top into the hash table ASAP.
635 initial_gvn()->transform_no_reclaim(top());
636
637 // Set up tf(), start(), and find a CallGenerator.
638 CallGenerator* cg = NULL;
639 if (is_osr_compilation()) {
640 const TypeTuple *domain = StartOSRNode::osr_domain();
641 const TypeTuple *range = TypeTuple::make_range(method()->signature());
642 init_tf(TypeFunc::make(domain, range));
643 StartNode* s = new StartOSRNode(root(), domain);
644 initial_gvn()->set_type_bottom(s);
645 init_start(s);
646 cg = CallGenerator::for_osr(method(), entry_bci());
647 } else {
648 // Normal case.
649 init_tf(TypeFunc::make(method()));
650 StartNode* s = new StartNode(root(), tf()->domain());
651 initial_gvn()->set_type_bottom(s);
652 init_start(s);
653 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
654 // With java.lang.ref.reference.get() we must go through the
655 // intrinsic - even when get() is the root
656 // method of the compile - so that, if necessary, the value in
657 // the referent field of the reference object gets recorded by
658 // the pre-barrier code.
659 cg = find_intrinsic(method(), false);
660 }
661 if (cg == NULL) {
662 float past_uses = method()->interpreter_invocation_count();
663 float expected_uses = past_uses;
664 cg = CallGenerator::for_inline(method(), expected_uses);
665 }
666 }
667 if (failing()) return;
668 if (cg == NULL) {
669 record_method_not_compilable("cannot parse method");
670 return;
671 }
672 JVMState* jvms = build_start_state(start(), tf());
673 if ((jvms = cg->generate(jvms)) == NULL) {
674 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
675 record_method_not_compilable("method parse failed");
676 }
677 return;
678 }
679 GraphKit kit(jvms);
680
681 if (!kit.stopped()) {
682 // Accept return values, and transfer control we know not where.
683 // This is done by a special, unique ReturnNode bound to root.
684 return_values(kit.jvms());
685 }
686
687 if (kit.has_exceptions()) {
688 // Any exceptions that escape from this call must be rethrown
689 // to whatever caller is dynamically above us on the stack.
690 // This is done by a special, unique RethrowNode bound to root.
691 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
692 }
693
694 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
695
696 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
697 inline_string_calls(true);
698 }
699
700 if (failing()) return;
701
702 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
703
704 // Remove clutter produced by parsing.
705 if (!failing()) {
706 ResourceMark rm;
707 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
708 }
709 }
710
711 // Note: Large methods are capped off in do_one_bytecode().
712 if (failing()) return;
713
714 // After parsing, node notes are no longer automagic.
715 // They must be propagated by register_new_node_with_optimizer(),
716 // clone(), or the like.
717 set_default_node_notes(NULL);
718
719 for (;;) {
720 int successes = Inline_Warm();
721 if (failing()) return;
722 if (successes == 0) break;
723 }
724
725 // Drain the list.
726 Finish_Warm();
727 #ifndef PRODUCT
728 if (_printer && _printer->should_print(1)) {
729 _printer->print_inlining();
730 }
731 #endif
732
733 if (failing()) return;
734 NOT_PRODUCT( verify_graph_edges(); )
735
736 // Now optimize
737 Optimize();
738 if (failing()) return;
739 NOT_PRODUCT( verify_graph_edges(); )
740
741 #ifndef PRODUCT
742 if (print_ideal()) {
743 ttyLocker ttyl; // keep the following output all in one block
744 // This output goes directly to the tty, not the compiler log.
745 // To enable tools to match it up with the compilation activity,
746 // be sure to tag this tty output with the compile ID.
747 if (xtty != NULL) {
748 xtty->head("ideal compile_id='%d'%s", compile_id(),
749 is_osr_compilation() ? " compile_kind='osr'" :
750 "");
751 }
752 root()->dump(9999);
753 if (xtty != NULL) {
754 xtty->tail("ideal");
755 }
756 }
757 #endif
758
759 #ifdef ASSERT
760 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
761 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
762 #endif
763
764 // Dump compilation data to replay it.
765 if (directive->DumpReplayOption) {
766 env()->dump_replay_data(_compile_id);
767 }
768 if (directive->DumpInlineOption && (ilt() != NULL)) {
769 env()->dump_inline_data(_compile_id);
770 }
771
772 // Now that we know the size of all the monitors we can add a fixed slot
773 // for the original deopt pc.
774 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
775 set_fixed_slots(next_slot);
776
777 // Compute when to use implicit null checks. Used by matching trap based
778 // nodes and NullCheck optimization.
779 set_allowed_deopt_reasons();
780
781 // Now generate code
782 Code_Gen();
783 }
784
785 //------------------------------Compile----------------------------------------
786 // Compile a runtime stub
Compile(ciEnv * ci_env,TypeFunc_generator generator,address stub_function,const char * stub_name,int is_fancy_jump,bool pass_tls,bool save_arg_registers,bool return_pc,DirectiveSet * directive)787 Compile::Compile( ciEnv* ci_env,
788 TypeFunc_generator generator,
789 address stub_function,
790 const char *stub_name,
791 int is_fancy_jump,
792 bool pass_tls,
793 bool save_arg_registers,
794 bool return_pc,
795 DirectiveSet* directive)
796 : Phase(Compiler),
797 _compile_id(0),
798 _save_argument_registers(save_arg_registers),
799 _subsume_loads(true),
800 _do_escape_analysis(false),
801 _eliminate_boxing(false),
802 _method(NULL),
803 _entry_bci(InvocationEntryBci),
804 _stub_function(stub_function),
805 _stub_name(stub_name),
806 _stub_entry_point(NULL),
807 _max_node_limit(MaxNodeLimit),
808 _inlining_progress(false),
809 _inlining_incrementally(false),
810 _has_reserved_stack_access(false),
811 #ifndef PRODUCT
812 _trace_opto_output(directive->TraceOptoOutputOption),
813 _print_ideal(directive->PrintIdealOption),
814 #endif
815 _has_method_handle_invokes(false),
816 _clinit_barrier_on_entry(false),
817 _comp_arena(mtCompiler),
818 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
819 _env(ci_env),
820 _directive(directive),
821 _log(ci_env->log()),
822 _failure_reason(NULL),
823 _congraph(NULL),
824 #ifndef PRODUCT
825 _printer(NULL),
826 #endif
827 _dead_node_list(comp_arena()),
828 _dead_node_count(0),
829 _node_arena(mtCompiler),
830 _old_arena(mtCompiler),
831 _mach_constant_base_node(NULL),
832 _Compile_types(mtCompiler),
833 _initial_gvn(NULL),
834 _for_igvn(NULL),
835 _warm_calls(NULL),
836 _number_of_mh_late_inlines(0),
837 _print_inlining_stream(NULL),
838 _print_inlining_list(NULL),
839 _print_inlining_idx(0),
840 _print_inlining_output(NULL),
841 _replay_inline_data(NULL),
842 _java_calls(0),
843 _inner_loops(0),
844 _interpreter_frame_size(0),
845 #ifndef PRODUCT
846 _in_dump_cnt(0),
847 #endif
848 _allowed_reasons(0) {
849 C = this;
850
851 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
852 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
853
854 #ifndef PRODUCT
855 set_print_assembly(PrintFrameConverterAssembly);
856 set_parsed_irreducible_loop(false);
857 #else
858 set_print_assembly(false); // Must initialize.
859 #endif
860 set_has_irreducible_loop(false); // no loops
861
862 CompileWrapper cw(this);
863 Init(/*AliasLevel=*/ 0);
864 init_tf((*generator)());
865
866 {
867 // The following is a dummy for the sake of GraphKit::gen_stub
868 Unique_Node_List for_igvn(comp_arena());
869 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
870 PhaseGVN gvn(Thread::current()->resource_area(),255);
871 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
872 gvn.transform_no_reclaim(top());
873
874 GraphKit kit;
875 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
876 }
877
878 NOT_PRODUCT( verify_graph_edges(); )
879
880 Code_Gen();
881 }
882
883 //------------------------------Init-------------------------------------------
884 // Prepare for a single compilation
Init(int aliaslevel)885 void Compile::Init(int aliaslevel) {
886 _unique = 0;
887 _regalloc = NULL;
888
889 _tf = NULL; // filled in later
890 _top = NULL; // cached later
891 _matcher = NULL; // filled in later
892 _cfg = NULL; // filled in later
893
894 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
895
896 _node_note_array = NULL;
897 _default_node_notes = NULL;
898 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
899
900 _immutable_memory = NULL; // filled in at first inquiry
901
902 // Globally visible Nodes
903 // First set TOP to NULL to give safe behavior during creation of RootNode
904 set_cached_top_node(NULL);
905 set_root(new RootNode());
906 // Now that you have a Root to point to, create the real TOP
907 set_cached_top_node( new ConNode(Type::TOP) );
908 set_recent_alloc(NULL, NULL);
909
910 // Create Debug Information Recorder to record scopes, oopmaps, etc.
911 env()->set_oop_recorder(new OopRecorder(env()->arena()));
912 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
913 env()->set_dependencies(new Dependencies(env()));
914
915 _fixed_slots = 0;
916 set_has_split_ifs(false);
917 set_has_loops(has_method() && method()->has_loops()); // first approximation
918 set_has_stringbuilder(false);
919 set_has_boxed_value(false);
920 _trap_can_recompile = false; // no traps emitted yet
921 _major_progress = true; // start out assuming good things will happen
922 set_has_unsafe_access(false);
923 set_max_vector_size(0);
924 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
925 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
926 set_decompile_count(0);
927
928 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
929 _loop_opts_cnt = LoopOptsCount;
930 set_do_inlining(Inline);
931 set_max_inline_size(MaxInlineSize);
932 set_freq_inline_size(FreqInlineSize);
933 set_do_scheduling(OptoScheduling);
934 set_do_count_invocations(false);
935 set_do_method_data_update(false);
936
937 set_do_vector_loop(false);
938
939 if (AllowVectorizeOnDemand) {
940 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
941 set_do_vector_loop(true);
942 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());})
943 } else if (has_method() && method()->name() != 0 &&
944 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
945 set_do_vector_loop(true);
946 }
947 }
948 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
949 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());})
950
951 set_age_code(has_method() && method()->profile_aging());
952 set_rtm_state(NoRTM); // No RTM lock eliding by default
953 _max_node_limit = _directive->MaxNodeLimitOption;
954
955 #if INCLUDE_RTM_OPT
956 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
957 int rtm_state = method()->method_data()->rtm_state();
958 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
959 // Don't generate RTM lock eliding code.
960 set_rtm_state(NoRTM);
961 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
962 // Generate RTM lock eliding code without abort ratio calculation code.
963 set_rtm_state(UseRTM);
964 } else if (UseRTMDeopt) {
965 // Generate RTM lock eliding code and include abort ratio calculation
966 // code if UseRTMDeopt is on.
967 set_rtm_state(ProfileRTM);
968 }
969 }
970 #endif
971 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
972 set_clinit_barrier_on_entry(true);
973 }
974 if (debug_info()->recording_non_safepoints()) {
975 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
976 (comp_arena(), 8, 0, NULL));
977 set_default_node_notes(Node_Notes::make(this));
978 }
979
980 // // -- Initialize types before each compile --
981 // // Update cached type information
982 // if( _method && _method->constants() )
983 // Type::update_loaded_types(_method, _method->constants());
984
985 // Init alias_type map.
986 if (!_do_escape_analysis && aliaslevel == 3)
987 aliaslevel = 2; // No unique types without escape analysis
988 _AliasLevel = aliaslevel;
989 const int grow_ats = 16;
990 _max_alias_types = grow_ats;
991 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
992 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
993 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
994 {
995 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
996 }
997 // Initialize the first few types.
998 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
999 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1000 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1001 _num_alias_types = AliasIdxRaw+1;
1002 // Zero out the alias type cache.
1003 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1004 // A NULL adr_type hits in the cache right away. Preload the right answer.
1005 probe_alias_cache(NULL)->_index = AliasIdxTop;
1006
1007 _intrinsics = NULL;
1008 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1009 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1010 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1011 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1012 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1013 register_library_intrinsics();
1014 #ifdef ASSERT
1015 _type_verify_symmetry = true;
1016 _phase_optimize_finished = false;
1017 #endif
1018 }
1019
1020 //---------------------------init_start----------------------------------------
1021 // Install the StartNode on this compile object.
init_start(StartNode * s)1022 void Compile::init_start(StartNode* s) {
1023 if (failing())
1024 return; // already failing
1025 assert(s == start(), "");
1026 }
1027
1028 /**
1029 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1030 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1031 * the ideal graph.
1032 */
start() const1033 StartNode* Compile::start() const {
1034 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1035 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1036 Node* start = root()->fast_out(i);
1037 if (start->is_Start()) {
1038 return start->as_Start();
1039 }
1040 }
1041 fatal("Did not find Start node!");
1042 return NULL;
1043 }
1044
1045 //-------------------------------immutable_memory-------------------------------------
1046 // Access immutable memory
immutable_memory()1047 Node* Compile::immutable_memory() {
1048 if (_immutable_memory != NULL) {
1049 return _immutable_memory;
1050 }
1051 StartNode* s = start();
1052 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1053 Node *p = s->fast_out(i);
1054 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1055 _immutable_memory = p;
1056 return _immutable_memory;
1057 }
1058 }
1059 ShouldNotReachHere();
1060 return NULL;
1061 }
1062
1063 //----------------------set_cached_top_node------------------------------------
1064 // Install the cached top node, and make sure Node::is_top works correctly.
set_cached_top_node(Node * tn)1065 void Compile::set_cached_top_node(Node* tn) {
1066 if (tn != NULL) verify_top(tn);
1067 Node* old_top = _top;
1068 _top = tn;
1069 // Calling Node::setup_is_top allows the nodes the chance to adjust
1070 // their _out arrays.
1071 if (_top != NULL) _top->setup_is_top();
1072 if (old_top != NULL) old_top->setup_is_top();
1073 assert(_top == NULL || top()->is_top(), "");
1074 }
1075
1076 #ifdef ASSERT
count_live_nodes_by_graph_walk()1077 uint Compile::count_live_nodes_by_graph_walk() {
1078 Unique_Node_List useful(comp_arena());
1079 // Get useful node list by walking the graph.
1080 identify_useful_nodes(useful);
1081 return useful.size();
1082 }
1083
print_missing_nodes()1084 void Compile::print_missing_nodes() {
1085
1086 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1087 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1088 return;
1089 }
1090
1091 // This is an expensive function. It is executed only when the user
1092 // specifies VerifyIdealNodeCount option or otherwise knows the
1093 // additional work that needs to be done to identify reachable nodes
1094 // by walking the flow graph and find the missing ones using
1095 // _dead_node_list.
1096
1097 Unique_Node_List useful(comp_arena());
1098 // Get useful node list by walking the graph.
1099 identify_useful_nodes(useful);
1100
1101 uint l_nodes = C->live_nodes();
1102 uint l_nodes_by_walk = useful.size();
1103
1104 if (l_nodes != l_nodes_by_walk) {
1105 if (_log != NULL) {
1106 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1107 _log->stamp();
1108 _log->end_head();
1109 }
1110 VectorSet& useful_member_set = useful.member_set();
1111 int last_idx = l_nodes_by_walk;
1112 for (int i = 0; i < last_idx; i++) {
1113 if (useful_member_set.test(i)) {
1114 if (_dead_node_list.test(i)) {
1115 if (_log != NULL) {
1116 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1117 }
1118 if (PrintIdealNodeCount) {
1119 // Print the log message to tty
1120 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1121 useful.at(i)->dump();
1122 }
1123 }
1124 }
1125 else if (! _dead_node_list.test(i)) {
1126 if (_log != NULL) {
1127 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1128 }
1129 if (PrintIdealNodeCount) {
1130 // Print the log message to tty
1131 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1132 }
1133 }
1134 }
1135 if (_log != NULL) {
1136 _log->tail("mismatched_nodes");
1137 }
1138 }
1139 }
record_modified_node(Node * n)1140 void Compile::record_modified_node(Node* n) {
1141 if (_modified_nodes != NULL && !_inlining_incrementally &&
1142 n->outcnt() != 0 && !n->is_Con()) {
1143 _modified_nodes->push(n);
1144 }
1145 }
1146
remove_modified_node(Node * n)1147 void Compile::remove_modified_node(Node* n) {
1148 if (_modified_nodes != NULL) {
1149 _modified_nodes->remove(n);
1150 }
1151 }
1152 #endif
1153
1154 #ifndef PRODUCT
verify_top(Node * tn) const1155 void Compile::verify_top(Node* tn) const {
1156 if (tn != NULL) {
1157 assert(tn->is_Con(), "top node must be a constant");
1158 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1159 assert(tn->in(0) != NULL, "must have live top node");
1160 }
1161 }
1162 #endif
1163
1164
1165 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1166
grow_node_notes(GrowableArray<Node_Notes * > * arr,int grow_by)1167 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1168 guarantee(arr != NULL, "");
1169 int num_blocks = arr->length();
1170 if (grow_by < num_blocks) grow_by = num_blocks;
1171 int num_notes = grow_by * _node_notes_block_size;
1172 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1173 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1174 while (num_notes > 0) {
1175 arr->append(notes);
1176 notes += _node_notes_block_size;
1177 num_notes -= _node_notes_block_size;
1178 }
1179 assert(num_notes == 0, "exact multiple, please");
1180 }
1181
copy_node_notes_to(Node * dest,Node * source)1182 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1183 if (source == NULL || dest == NULL) return false;
1184
1185 if (dest->is_Con())
1186 return false; // Do not push debug info onto constants.
1187
1188 #ifdef ASSERT
1189 // Leave a bread crumb trail pointing to the original node:
1190 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1191 dest->set_debug_orig(source);
1192 }
1193 #endif
1194
1195 if (node_note_array() == NULL)
1196 return false; // Not collecting any notes now.
1197
1198 // This is a copy onto a pre-existing node, which may already have notes.
1199 // If both nodes have notes, do not overwrite any pre-existing notes.
1200 Node_Notes* source_notes = node_notes_at(source->_idx);
1201 if (source_notes == NULL || source_notes->is_clear()) return false;
1202 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1203 if (dest_notes == NULL || dest_notes->is_clear()) {
1204 return set_node_notes_at(dest->_idx, source_notes);
1205 }
1206
1207 Node_Notes merged_notes = (*source_notes);
1208 // The order of operations here ensures that dest notes will win...
1209 merged_notes.update_from(dest_notes);
1210 return set_node_notes_at(dest->_idx, &merged_notes);
1211 }
1212
1213
1214 //--------------------------allow_range_check_smearing-------------------------
1215 // Gating condition for coalescing similar range checks.
1216 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1217 // single covering check that is at least as strong as any of them.
1218 // If the optimization succeeds, the simplified (strengthened) range check
1219 // will always succeed. If it fails, we will deopt, and then give up
1220 // on the optimization.
allow_range_check_smearing() const1221 bool Compile::allow_range_check_smearing() const {
1222 // If this method has already thrown a range-check,
1223 // assume it was because we already tried range smearing
1224 // and it failed.
1225 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1226 return !already_trapped;
1227 }
1228
1229
1230 //------------------------------flatten_alias_type-----------------------------
flatten_alias_type(const TypePtr * tj) const1231 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1232 int offset = tj->offset();
1233 TypePtr::PTR ptr = tj->ptr();
1234
1235 // Known instance (scalarizable allocation) alias only with itself.
1236 bool is_known_inst = tj->isa_oopptr() != NULL &&
1237 tj->is_oopptr()->is_known_instance();
1238
1239 // Process weird unsafe references.
1240 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1241 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1242 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1243 tj = TypeOopPtr::BOTTOM;
1244 ptr = tj->ptr();
1245 offset = tj->offset();
1246 }
1247
1248 // Array pointers need some flattening
1249 const TypeAryPtr *ta = tj->isa_aryptr();
1250 if (ta && ta->is_stable()) {
1251 // Erase stability property for alias analysis.
1252 tj = ta = ta->cast_to_stable(false);
1253 }
1254 if( ta && is_known_inst ) {
1255 if ( offset != Type::OffsetBot &&
1256 offset > arrayOopDesc::length_offset_in_bytes() ) {
1257 offset = Type::OffsetBot; // Flatten constant access into array body only
1258 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1259 }
1260 } else if( ta && _AliasLevel >= 2 ) {
1261 // For arrays indexed by constant indices, we flatten the alias
1262 // space to include all of the array body. Only the header, klass
1263 // and array length can be accessed un-aliased.
1264 if( offset != Type::OffsetBot ) {
1265 if( ta->const_oop() ) { // MethodData* or Method*
1266 offset = Type::OffsetBot; // Flatten constant access into array body
1267 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1268 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1269 // range is OK as-is.
1270 tj = ta = TypeAryPtr::RANGE;
1271 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1272 tj = TypeInstPtr::KLASS; // all klass loads look alike
1273 ta = TypeAryPtr::RANGE; // generic ignored junk
1274 ptr = TypePtr::BotPTR;
1275 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1276 tj = TypeInstPtr::MARK;
1277 ta = TypeAryPtr::RANGE; // generic ignored junk
1278 ptr = TypePtr::BotPTR;
1279 } else { // Random constant offset into array body
1280 offset = Type::OffsetBot; // Flatten constant access into array body
1281 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1282 }
1283 }
1284 // Arrays of fixed size alias with arrays of unknown size.
1285 if (ta->size() != TypeInt::POS) {
1286 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1287 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1288 }
1289 // Arrays of known objects become arrays of unknown objects.
1290 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1291 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1292 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1293 }
1294 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1295 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1296 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1297 }
1298 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1299 // cannot be distinguished by bytecode alone.
1300 if (ta->elem() == TypeInt::BOOL) {
1301 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1302 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1303 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1304 }
1305 // During the 2nd round of IterGVN, NotNull castings are removed.
1306 // Make sure the Bottom and NotNull variants alias the same.
1307 // Also, make sure exact and non-exact variants alias the same.
1308 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1309 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1310 }
1311 }
1312
1313 // Oop pointers need some flattening
1314 const TypeInstPtr *to = tj->isa_instptr();
1315 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1316 ciInstanceKlass *k = to->klass()->as_instance_klass();
1317 if( ptr == TypePtr::Constant ) {
1318 if (to->klass() != ciEnv::current()->Class_klass() ||
1319 offset < k->size_helper() * wordSize) {
1320 // No constant oop pointers (such as Strings); they alias with
1321 // unknown strings.
1322 assert(!is_known_inst, "not scalarizable allocation");
1323 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1324 }
1325 } else if( is_known_inst ) {
1326 tj = to; // Keep NotNull and klass_is_exact for instance type
1327 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1328 // During the 2nd round of IterGVN, NotNull castings are removed.
1329 // Make sure the Bottom and NotNull variants alias the same.
1330 // Also, make sure exact and non-exact variants alias the same.
1331 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1332 }
1333 if (to->speculative() != NULL) {
1334 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1335 }
1336 // Canonicalize the holder of this field
1337 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1338 // First handle header references such as a LoadKlassNode, even if the
1339 // object's klass is unloaded at compile time (4965979).
1340 if (!is_known_inst) { // Do it only for non-instance types
1341 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1342 }
1343 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1344 // Static fields are in the space above the normal instance
1345 // fields in the java.lang.Class instance.
1346 if (to->klass() != ciEnv::current()->Class_klass()) {
1347 to = NULL;
1348 tj = TypeOopPtr::BOTTOM;
1349 offset = tj->offset();
1350 }
1351 } else {
1352 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1353 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1354 if( is_known_inst ) {
1355 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1356 } else {
1357 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1358 }
1359 }
1360 }
1361 }
1362
1363 // Klass pointers to object array klasses need some flattening
1364 const TypeKlassPtr *tk = tj->isa_klassptr();
1365 if( tk ) {
1366 // If we are referencing a field within a Klass, we need
1367 // to assume the worst case of an Object. Both exact and
1368 // inexact types must flatten to the same alias class so
1369 // use NotNull as the PTR.
1370 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1371
1372 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1373 TypeKlassPtr::OBJECT->klass(),
1374 offset);
1375 }
1376
1377 ciKlass* klass = tk->klass();
1378 if( klass->is_obj_array_klass() ) {
1379 ciKlass* k = TypeAryPtr::OOPS->klass();
1380 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1381 k = TypeInstPtr::BOTTOM->klass();
1382 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1383 }
1384
1385 // Check for precise loads from the primary supertype array and force them
1386 // to the supertype cache alias index. Check for generic array loads from
1387 // the primary supertype array and also force them to the supertype cache
1388 // alias index. Since the same load can reach both, we need to merge
1389 // these 2 disparate memories into the same alias class. Since the
1390 // primary supertype array is read-only, there's no chance of confusion
1391 // where we bypass an array load and an array store.
1392 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1393 if (offset == Type::OffsetBot ||
1394 (offset >= primary_supers_offset &&
1395 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1396 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1397 offset = in_bytes(Klass::secondary_super_cache_offset());
1398 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1399 }
1400 }
1401
1402 // Flatten all Raw pointers together.
1403 if (tj->base() == Type::RawPtr)
1404 tj = TypeRawPtr::BOTTOM;
1405
1406 if (tj->base() == Type::AnyPtr)
1407 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1408
1409 // Flatten all to bottom for now
1410 switch( _AliasLevel ) {
1411 case 0:
1412 tj = TypePtr::BOTTOM;
1413 break;
1414 case 1: // Flatten to: oop, static, field or array
1415 switch (tj->base()) {
1416 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1417 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1418 case Type::AryPtr: // do not distinguish arrays at all
1419 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1420 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1421 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1422 default: ShouldNotReachHere();
1423 }
1424 break;
1425 case 2: // No collapsing at level 2; keep all splits
1426 case 3: // No collapsing at level 3; keep all splits
1427 break;
1428 default:
1429 Unimplemented();
1430 }
1431
1432 offset = tj->offset();
1433 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1434
1435 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1436 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1437 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1438 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1439 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1440 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1441 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1442 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1443 assert( tj->ptr() != TypePtr::TopPTR &&
1444 tj->ptr() != TypePtr::AnyNull &&
1445 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1446 // assert( tj->ptr() != TypePtr::Constant ||
1447 // tj->base() == Type::RawPtr ||
1448 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1449
1450 return tj;
1451 }
1452
Init(int i,const TypePtr * at)1453 void Compile::AliasType::Init(int i, const TypePtr* at) {
1454 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1455 _index = i;
1456 _adr_type = at;
1457 _field = NULL;
1458 _element = NULL;
1459 _is_rewritable = true; // default
1460 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1461 if (atoop != NULL && atoop->is_known_instance()) {
1462 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1463 _general_index = Compile::current()->get_alias_index(gt);
1464 } else {
1465 _general_index = 0;
1466 }
1467 }
1468
basic_type() const1469 BasicType Compile::AliasType::basic_type() const {
1470 if (element() != NULL) {
1471 const Type* element = adr_type()->is_aryptr()->elem();
1472 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1473 } if (field() != NULL) {
1474 return field()->layout_type();
1475 } else {
1476 return T_ILLEGAL; // unknown
1477 }
1478 }
1479
1480 //---------------------------------print_on------------------------------------
1481 #ifndef PRODUCT
print_on(outputStream * st)1482 void Compile::AliasType::print_on(outputStream* st) {
1483 if (index() < 10)
1484 st->print("@ <%d> ", index());
1485 else st->print("@ <%d>", index());
1486 st->print(is_rewritable() ? " " : " RO");
1487 int offset = adr_type()->offset();
1488 if (offset == Type::OffsetBot)
1489 st->print(" +any");
1490 else st->print(" +%-3d", offset);
1491 st->print(" in ");
1492 adr_type()->dump_on(st);
1493 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1494 if (field() != NULL && tjp) {
1495 if (tjp->klass() != field()->holder() ||
1496 tjp->offset() != field()->offset_in_bytes()) {
1497 st->print(" != ");
1498 field()->print();
1499 st->print(" ***");
1500 }
1501 }
1502 }
1503
print_alias_types()1504 void print_alias_types() {
1505 Compile* C = Compile::current();
1506 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1507 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1508 C->alias_type(idx)->print_on(tty);
1509 tty->cr();
1510 }
1511 }
1512 #endif
1513
1514
1515 //----------------------------probe_alias_cache--------------------------------
probe_alias_cache(const TypePtr * adr_type)1516 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1517 intptr_t key = (intptr_t) adr_type;
1518 key ^= key >> logAliasCacheSize;
1519 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1520 }
1521
1522
1523 //-----------------------------grow_alias_types--------------------------------
grow_alias_types()1524 void Compile::grow_alias_types() {
1525 const int old_ats = _max_alias_types; // how many before?
1526 const int new_ats = old_ats; // how many more?
1527 const int grow_ats = old_ats+new_ats; // how many now?
1528 _max_alias_types = grow_ats;
1529 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1530 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1531 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1532 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1533 }
1534
1535
1536 //--------------------------------find_alias_type------------------------------
find_alias_type(const TypePtr * adr_type,bool no_create,ciField * original_field)1537 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1538 if (_AliasLevel == 0)
1539 return alias_type(AliasIdxBot);
1540
1541 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1542 if (ace->_adr_type == adr_type) {
1543 return alias_type(ace->_index);
1544 }
1545
1546 // Handle special cases.
1547 if (adr_type == NULL) return alias_type(AliasIdxTop);
1548 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1549
1550 // Do it the slow way.
1551 const TypePtr* flat = flatten_alias_type(adr_type);
1552
1553 #ifdef ASSERT
1554 {
1555 ResourceMark rm;
1556 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1557 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1558 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1559 Type::str(adr_type));
1560 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1561 const TypeOopPtr* foop = flat->is_oopptr();
1562 // Scalarizable allocations have exact klass always.
1563 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1564 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1565 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1566 Type::str(foop), Type::str(xoop));
1567 }
1568 }
1569 #endif
1570
1571 int idx = AliasIdxTop;
1572 for (int i = 0; i < num_alias_types(); i++) {
1573 if (alias_type(i)->adr_type() == flat) {
1574 idx = i;
1575 break;
1576 }
1577 }
1578
1579 if (idx == AliasIdxTop) {
1580 if (no_create) return NULL;
1581 // Grow the array if necessary.
1582 if (_num_alias_types == _max_alias_types) grow_alias_types();
1583 // Add a new alias type.
1584 idx = _num_alias_types++;
1585 _alias_types[idx]->Init(idx, flat);
1586 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1587 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1588 if (flat->isa_instptr()) {
1589 if (flat->offset() == java_lang_Class::klass_offset()
1590 && flat->is_instptr()->klass() == env()->Class_klass())
1591 alias_type(idx)->set_rewritable(false);
1592 }
1593 if (flat->isa_aryptr()) {
1594 #ifdef ASSERT
1595 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1596 // (T_BYTE has the weakest alignment and size restrictions...)
1597 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1598 #endif
1599 if (flat->offset() == TypePtr::OffsetBot) {
1600 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1601 }
1602 }
1603 if (flat->isa_klassptr()) {
1604 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1605 alias_type(idx)->set_rewritable(false);
1606 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1607 alias_type(idx)->set_rewritable(false);
1608 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1609 alias_type(idx)->set_rewritable(false);
1610 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1611 alias_type(idx)->set_rewritable(false);
1612 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1613 alias_type(idx)->set_rewritable(false);
1614 }
1615 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1616 // but the base pointer type is not distinctive enough to identify
1617 // references into JavaThread.)
1618
1619 // Check for final fields.
1620 const TypeInstPtr* tinst = flat->isa_instptr();
1621 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1622 ciField* field;
1623 if (tinst->const_oop() != NULL &&
1624 tinst->klass() == ciEnv::current()->Class_klass() &&
1625 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1626 // static field
1627 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1628 field = k->get_field_by_offset(tinst->offset(), true);
1629 } else {
1630 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1631 field = k->get_field_by_offset(tinst->offset(), false);
1632 }
1633 assert(field == NULL ||
1634 original_field == NULL ||
1635 (field->holder() == original_field->holder() &&
1636 field->offset() == original_field->offset() &&
1637 field->is_static() == original_field->is_static()), "wrong field?");
1638 // Set field() and is_rewritable() attributes.
1639 if (field != NULL) alias_type(idx)->set_field(field);
1640 }
1641 }
1642
1643 // Fill the cache for next time.
1644 ace->_adr_type = adr_type;
1645 ace->_index = idx;
1646 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1647
1648 // Might as well try to fill the cache for the flattened version, too.
1649 AliasCacheEntry* face = probe_alias_cache(flat);
1650 if (face->_adr_type == NULL) {
1651 face->_adr_type = flat;
1652 face->_index = idx;
1653 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1654 }
1655
1656 return alias_type(idx);
1657 }
1658
1659
alias_type(ciField * field)1660 Compile::AliasType* Compile::alias_type(ciField* field) {
1661 const TypeOopPtr* t;
1662 if (field->is_static())
1663 t = TypeInstPtr::make(field->holder()->java_mirror());
1664 else
1665 t = TypeOopPtr::make_from_klass_raw(field->holder());
1666 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1667 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1668 return atp;
1669 }
1670
1671
1672 //------------------------------have_alias_type--------------------------------
have_alias_type(const TypePtr * adr_type)1673 bool Compile::have_alias_type(const TypePtr* adr_type) {
1674 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1675 if (ace->_adr_type == adr_type) {
1676 return true;
1677 }
1678
1679 // Handle special cases.
1680 if (adr_type == NULL) return true;
1681 if (adr_type == TypePtr::BOTTOM) return true;
1682
1683 return find_alias_type(adr_type, true, NULL) != NULL;
1684 }
1685
1686 //-----------------------------must_alias--------------------------------------
1687 // True if all values of the given address type are in the given alias category.
must_alias(const TypePtr * adr_type,int alias_idx)1688 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1689 if (alias_idx == AliasIdxBot) return true; // the universal category
1690 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1691 if (alias_idx == AliasIdxTop) return false; // the empty category
1692 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1693
1694 // the only remaining possible overlap is identity
1695 int adr_idx = get_alias_index(adr_type);
1696 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1697 assert(adr_idx == alias_idx ||
1698 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1699 && adr_type != TypeOopPtr::BOTTOM),
1700 "should not be testing for overlap with an unsafe pointer");
1701 return adr_idx == alias_idx;
1702 }
1703
1704 //------------------------------can_alias--------------------------------------
1705 // True if any values of the given address type are in the given alias category.
can_alias(const TypePtr * adr_type,int alias_idx)1706 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1707 if (alias_idx == AliasIdxTop) return false; // the empty category
1708 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1709 // Known instance doesn't alias with bottom memory
1710 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1711 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1712
1713 // the only remaining possible overlap is identity
1714 int adr_idx = get_alias_index(adr_type);
1715 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1716 return adr_idx == alias_idx;
1717 }
1718
1719
1720
1721 //---------------------------pop_warm_call-------------------------------------
pop_warm_call()1722 WarmCallInfo* Compile::pop_warm_call() {
1723 WarmCallInfo* wci = _warm_calls;
1724 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1725 return wci;
1726 }
1727
1728 //----------------------------Inline_Warm--------------------------------------
Inline_Warm()1729 int Compile::Inline_Warm() {
1730 // If there is room, try to inline some more warm call sites.
1731 // %%% Do a graph index compaction pass when we think we're out of space?
1732 if (!InlineWarmCalls) return 0;
1733
1734 int calls_made_hot = 0;
1735 int room_to_grow = NodeCountInliningCutoff - unique();
1736 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1737 int amount_grown = 0;
1738 WarmCallInfo* call;
1739 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1740 int est_size = (int)call->size();
1741 if (est_size > (room_to_grow - amount_grown)) {
1742 // This one won't fit anyway. Get rid of it.
1743 call->make_cold();
1744 continue;
1745 }
1746 call->make_hot();
1747 calls_made_hot++;
1748 amount_grown += est_size;
1749 amount_to_grow -= est_size;
1750 }
1751
1752 if (calls_made_hot > 0) set_major_progress();
1753 return calls_made_hot;
1754 }
1755
1756
1757 //----------------------------Finish_Warm--------------------------------------
Finish_Warm()1758 void Compile::Finish_Warm() {
1759 if (!InlineWarmCalls) return;
1760 if (failing()) return;
1761 if (warm_calls() == NULL) return;
1762
1763 // Clean up loose ends, if we are out of space for inlining.
1764 WarmCallInfo* call;
1765 while ((call = pop_warm_call()) != NULL) {
1766 call->make_cold();
1767 }
1768 }
1769
1770 //---------------------cleanup_loop_predicates-----------------------
1771 // Remove the opaque nodes that protect the predicates so that all unused
1772 // checks and uncommon_traps will be eliminated from the ideal graph
cleanup_loop_predicates(PhaseIterGVN & igvn)1773 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1774 if (predicate_count()==0) return;
1775 for (int i = predicate_count(); i > 0; i--) {
1776 Node * n = predicate_opaque1_node(i-1);
1777 assert(n->Opcode() == Op_Opaque1, "must be");
1778 igvn.replace_node(n, n->in(1));
1779 }
1780 assert(predicate_count()==0, "should be clean!");
1781 }
1782
add_range_check_cast(Node * n)1783 void Compile::add_range_check_cast(Node* n) {
1784 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1785 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1786 _range_check_casts->append(n);
1787 }
1788
1789 // Remove all range check dependent CastIINodes.
remove_range_check_casts(PhaseIterGVN & igvn)1790 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1791 for (int i = range_check_cast_count(); i > 0; i--) {
1792 Node* cast = range_check_cast_node(i-1);
1793 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1794 igvn.replace_node(cast, cast->in(1));
1795 }
1796 assert(range_check_cast_count() == 0, "should be empty");
1797 }
1798
add_opaque4_node(Node * n)1799 void Compile::add_opaque4_node(Node* n) {
1800 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
1801 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
1802 _opaque4_nodes->append(n);
1803 }
1804
1805 // Remove all Opaque4 nodes.
remove_opaque4_nodes(PhaseIterGVN & igvn)1806 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
1807 for (int i = opaque4_count(); i > 0; i--) {
1808 Node* opaq = opaque4_node(i-1);
1809 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
1810 // With Opaque4 nodes, the expectation is that the test of input 1
1811 // is always equal to the constant value of input 2. So we can
1812 // remove the Opaque4 and replace it by input 2. In debug builds,
1813 // leave the non constant test in instead to sanity check that it
1814 // never fails (if it does, that subgraph was constructed so, at
1815 // runtime, a Halt node is executed).
1816 #ifdef ASSERT
1817 igvn.replace_node(opaq, opaq->in(1));
1818 #else
1819 igvn.replace_node(opaq, opaq->in(2));
1820 #endif
1821 }
1822 assert(opaque4_count() == 0, "should be empty");
1823 }
1824
1825 // StringOpts and late inlining of string methods
inline_string_calls(bool parse_time)1826 void Compile::inline_string_calls(bool parse_time) {
1827 {
1828 // remove useless nodes to make the usage analysis simpler
1829 ResourceMark rm;
1830 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1831 }
1832
1833 {
1834 ResourceMark rm;
1835 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1836 PhaseStringOpts pso(initial_gvn(), for_igvn());
1837 print_method(PHASE_AFTER_STRINGOPTS, 3);
1838 }
1839
1840 // now inline anything that we skipped the first time around
1841 if (!parse_time) {
1842 _late_inlines_pos = _late_inlines.length();
1843 }
1844
1845 while (_string_late_inlines.length() > 0) {
1846 CallGenerator* cg = _string_late_inlines.pop();
1847 cg->do_late_inline();
1848 if (failing()) return;
1849 }
1850 _string_late_inlines.trunc_to(0);
1851 }
1852
1853 // Late inlining of boxing methods
inline_boxing_calls(PhaseIterGVN & igvn)1854 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1855 if (_boxing_late_inlines.length() > 0) {
1856 assert(has_boxed_value(), "inconsistent");
1857
1858 PhaseGVN* gvn = initial_gvn();
1859 set_inlining_incrementally(true);
1860
1861 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1862 for_igvn()->clear();
1863 gvn->replace_with(&igvn);
1864
1865 _late_inlines_pos = _late_inlines.length();
1866
1867 while (_boxing_late_inlines.length() > 0) {
1868 CallGenerator* cg = _boxing_late_inlines.pop();
1869 cg->do_late_inline();
1870 if (failing()) return;
1871 }
1872 _boxing_late_inlines.trunc_to(0);
1873
1874 inline_incrementally_cleanup(igvn);
1875
1876 set_inlining_incrementally(false);
1877 }
1878 }
1879
inline_incrementally_one()1880 bool Compile::inline_incrementally_one() {
1881 assert(IncrementalInline, "incremental inlining should be on");
1882
1883 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
1884 set_inlining_progress(false);
1885 set_do_cleanup(false);
1886 int i = 0;
1887 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1888 CallGenerator* cg = _late_inlines.at(i);
1889 _late_inlines_pos = i+1;
1890 cg->do_late_inline();
1891 if (failing()) return false;
1892 }
1893 int j = 0;
1894 for (; i < _late_inlines.length(); i++, j++) {
1895 _late_inlines.at_put(j, _late_inlines.at(i));
1896 }
1897 _late_inlines.trunc_to(j);
1898 assert(inlining_progress() || _late_inlines.length() == 0, "");
1899
1900 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
1901
1902 set_inlining_progress(false);
1903 set_do_cleanup(false);
1904 return (_late_inlines.length() > 0) && !needs_cleanup;
1905 }
1906
inline_incrementally_cleanup(PhaseIterGVN & igvn)1907 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
1908 {
1909 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
1910 ResourceMark rm;
1911 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1912 }
1913 {
1914 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
1915 igvn = PhaseIterGVN(initial_gvn());
1916 igvn.optimize();
1917 }
1918 }
1919
1920 // Perform incremental inlining until bound on number of live nodes is reached
inline_incrementally(PhaseIterGVN & igvn)1921 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1922 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
1923
1924 set_inlining_incrementally(true);
1925 uint low_live_nodes = 0;
1926
1927 while (_late_inlines.length() > 0) {
1928 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1929 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1930 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
1931 // PhaseIdealLoop is expensive so we only try it once we are
1932 // out of live nodes and we only try it again if the previous
1933 // helped got the number of nodes down significantly
1934 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
1935 if (failing()) return;
1936 low_live_nodes = live_nodes();
1937 _major_progress = true;
1938 }
1939
1940 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1941 break; // finish
1942 }
1943 }
1944
1945 for_igvn()->clear();
1946 initial_gvn()->replace_with(&igvn);
1947
1948 while (inline_incrementally_one()) {
1949 assert(!failing(), "inconsistent");
1950 }
1951
1952 if (failing()) return;
1953
1954 inline_incrementally_cleanup(igvn);
1955
1956 if (failing()) return;
1957 }
1958 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1959
1960 if (_string_late_inlines.length() > 0) {
1961 assert(has_stringbuilder(), "inconsistent");
1962 for_igvn()->clear();
1963 initial_gvn()->replace_with(&igvn);
1964
1965 inline_string_calls(false);
1966
1967 if (failing()) return;
1968
1969 inline_incrementally_cleanup(igvn);
1970 }
1971
1972 set_inlining_incrementally(false);
1973 }
1974
1975
optimize_loops(PhaseIterGVN & igvn,LoopOptsMode mode)1976 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
1977 if(_loop_opts_cnt > 0) {
1978 debug_only( int cnt = 0; );
1979 while(major_progress() && (_loop_opts_cnt > 0)) {
1980 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
1981 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1982 PhaseIdealLoop::optimize(igvn, mode);
1983 _loop_opts_cnt--;
1984 if (failing()) return false;
1985 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
1986 }
1987 }
1988 return true;
1989 }
1990
1991 // Remove edges from "root" to each SafePoint at a backward branch.
1992 // They were inserted during parsing (see add_safepoint()) to make
1993 // infinite loops without calls or exceptions visible to root, i.e.,
1994 // useful.
remove_root_to_sfpts_edges(PhaseIterGVN & igvn)1995 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
1996 Node *r = root();
1997 if (r != NULL) {
1998 for (uint i = r->req(); i < r->len(); ++i) {
1999 Node *n = r->in(i);
2000 if (n != NULL && n->is_SafePoint()) {
2001 r->rm_prec(i);
2002 if (n->outcnt() == 0) {
2003 igvn.remove_dead_node(n);
2004 }
2005 --i;
2006 }
2007 }
2008 // Parsing may have added top inputs to the root node (Path
2009 // leading to the Halt node proven dead). Make sure we get a
2010 // chance to clean them up.
2011 igvn._worklist.push(r);
2012 igvn.optimize();
2013 }
2014 }
2015
2016 //------------------------------Optimize---------------------------------------
2017 // Given a graph, optimize it.
Optimize()2018 void Compile::Optimize() {
2019 TracePhase tp("optimizer", &timers[_t_optimizer]);
2020
2021 #ifndef PRODUCT
2022 if (_directive->BreakAtCompileOption) {
2023 BREAKPOINT;
2024 }
2025
2026 #endif
2027
2028 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2029 #ifdef ASSERT
2030 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2031 #endif
2032
2033 ResourceMark rm;
2034
2035 print_inlining_reinit();
2036
2037 NOT_PRODUCT( verify_graph_edges(); )
2038
2039 print_method(PHASE_AFTER_PARSING);
2040
2041 {
2042 // Iterative Global Value Numbering, including ideal transforms
2043 // Initialize IterGVN with types and values from parse-time GVN
2044 PhaseIterGVN igvn(initial_gvn());
2045 #ifdef ASSERT
2046 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2047 #endif
2048 {
2049 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2050 igvn.optimize();
2051 }
2052
2053 if (failing()) return;
2054
2055 print_method(PHASE_ITER_GVN1, 2);
2056
2057 inline_incrementally(igvn);
2058
2059 print_method(PHASE_INCREMENTAL_INLINE, 2);
2060
2061 if (failing()) return;
2062
2063 if (eliminate_boxing()) {
2064 // Inline valueOf() methods now.
2065 inline_boxing_calls(igvn);
2066
2067 if (AlwaysIncrementalInline) {
2068 inline_incrementally(igvn);
2069 }
2070
2071 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2072
2073 if (failing()) return;
2074 }
2075
2076 // Now that all inlining is over, cut edge from root to loop
2077 // safepoints
2078 remove_root_to_sfpts_edges(igvn);
2079
2080 // Remove the speculative part of types and clean up the graph from
2081 // the extra CastPP nodes whose only purpose is to carry them. Do
2082 // that early so that optimizations are not disrupted by the extra
2083 // CastPP nodes.
2084 remove_speculative_types(igvn);
2085
2086 // No more new expensive nodes will be added to the list from here
2087 // so keep only the actual candidates for optimizations.
2088 cleanup_expensive_nodes(igvn);
2089
2090 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2091 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2092 initial_gvn()->replace_with(&igvn);
2093 for_igvn()->clear();
2094 Unique_Node_List new_worklist(C->comp_arena());
2095 {
2096 ResourceMark rm;
2097 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2098 }
2099 set_for_igvn(&new_worklist);
2100 igvn = PhaseIterGVN(initial_gvn());
2101 igvn.optimize();
2102 }
2103
2104 // Perform escape analysis
2105 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2106 if (has_loops()) {
2107 // Cleanup graph (remove dead nodes).
2108 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2109 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2110 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2111 if (failing()) return;
2112 }
2113 ConnectionGraph::do_analysis(this, &igvn);
2114
2115 if (failing()) return;
2116
2117 // Optimize out fields loads from scalar replaceable allocations.
2118 igvn.optimize();
2119 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2120
2121 if (failing()) return;
2122
2123 if (congraph() != NULL && macro_count() > 0) {
2124 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2125 PhaseMacroExpand mexp(igvn);
2126 mexp.eliminate_macro_nodes();
2127 igvn.set_delay_transform(false);
2128
2129 igvn.optimize();
2130 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2131
2132 if (failing()) return;
2133 }
2134 }
2135
2136 // Loop transforms on the ideal graph. Range Check Elimination,
2137 // peeling, unrolling, etc.
2138
2139 // Set loop opts counter
2140 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2141 {
2142 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2143 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2144 _loop_opts_cnt--;
2145 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2146 if (failing()) return;
2147 }
2148 // Loop opts pass if partial peeling occurred in previous pass
2149 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2150 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2151 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2152 _loop_opts_cnt--;
2153 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2154 if (failing()) return;
2155 }
2156 // Loop opts pass for loop-unrolling before CCP
2157 if(major_progress() && (_loop_opts_cnt > 0)) {
2158 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2159 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2160 _loop_opts_cnt--;
2161 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2162 }
2163 if (!failing()) {
2164 // Verify that last round of loop opts produced a valid graph
2165 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2166 PhaseIdealLoop::verify(igvn);
2167 }
2168 }
2169 if (failing()) return;
2170
2171 // Conditional Constant Propagation;
2172 PhaseCCP ccp( &igvn );
2173 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2174 {
2175 TracePhase tp("ccp", &timers[_t_ccp]);
2176 ccp.do_transform();
2177 }
2178 print_method(PHASE_CPP1, 2);
2179
2180 assert( true, "Break here to ccp.dump_old2new_map()");
2181
2182 // Iterative Global Value Numbering, including ideal transforms
2183 {
2184 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2185 igvn = ccp;
2186 igvn.optimize();
2187 }
2188 print_method(PHASE_ITER_GVN2, 2);
2189
2190 if (failing()) return;
2191
2192 // Loop transforms on the ideal graph. Range Check Elimination,
2193 // peeling, unrolling, etc.
2194 if (!optimize_loops(igvn, LoopOptsDefault)) {
2195 return;
2196 }
2197
2198 if (failing()) return;
2199
2200 // Ensure that major progress is now clear
2201 C->clear_major_progress();
2202
2203 {
2204 // Verify that all previous optimizations produced a valid graph
2205 // at least to this point, even if no loop optimizations were done.
2206 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2207 PhaseIdealLoop::verify(igvn);
2208 }
2209
2210 if (range_check_cast_count() > 0) {
2211 // No more loop optimizations. Remove all range check dependent CastIINodes.
2212 C->remove_range_check_casts(igvn);
2213 igvn.optimize();
2214 }
2215
2216 #ifdef ASSERT
2217 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2218 #endif
2219
2220 {
2221 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2222 PhaseMacroExpand mex(igvn);
2223 if (mex.expand_macro_nodes()) {
2224 assert(failing(), "must bail out w/ explicit message");
2225 return;
2226 }
2227 print_method(PHASE_MACRO_EXPANSION, 2);
2228 }
2229
2230 {
2231 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2232 if (bs->expand_barriers(this, igvn)) {
2233 assert(failing(), "must bail out w/ explicit message");
2234 return;
2235 }
2236 print_method(PHASE_BARRIER_EXPANSION, 2);
2237 }
2238
2239 if (opaque4_count() > 0) {
2240 C->remove_opaque4_nodes(igvn);
2241 igvn.optimize();
2242 }
2243
2244 if (C->max_vector_size() > 0) {
2245 C->optimize_logic_cones(igvn);
2246 igvn.optimize();
2247 }
2248
2249 DEBUG_ONLY( _modified_nodes = NULL; )
2250 } // (End scope of igvn; run destructor if necessary for asserts.)
2251
2252 process_print_inlining();
2253 // A method with only infinite loops has no edges entering loops from root
2254 {
2255 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2256 if (final_graph_reshaping()) {
2257 assert(failing(), "must bail out w/ explicit message");
2258 return;
2259 }
2260 }
2261
2262 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2263 DEBUG_ONLY(set_phase_optimize_finished();)
2264 }
2265
2266 //---------------------------- Bitwise operation packing optimization ---------------------------
2267
is_vector_unary_bitwise_op(Node * n)2268 static bool is_vector_unary_bitwise_op(Node* n) {
2269 return n->Opcode() == Op_XorV &&
2270 VectorNode::is_vector_bitwise_not_pattern(n);
2271 }
2272
is_vector_binary_bitwise_op(Node * n)2273 static bool is_vector_binary_bitwise_op(Node* n) {
2274 switch (n->Opcode()) {
2275 case Op_AndV:
2276 case Op_OrV:
2277 return true;
2278
2279 case Op_XorV:
2280 return !is_vector_unary_bitwise_op(n);
2281
2282 default:
2283 return false;
2284 }
2285 }
2286
is_vector_ternary_bitwise_op(Node * n)2287 static bool is_vector_ternary_bitwise_op(Node* n) {
2288 return n->Opcode() == Op_MacroLogicV;
2289 }
2290
is_vector_bitwise_op(Node * n)2291 static bool is_vector_bitwise_op(Node* n) {
2292 return is_vector_unary_bitwise_op(n) ||
2293 is_vector_binary_bitwise_op(n) ||
2294 is_vector_ternary_bitwise_op(n);
2295 }
2296
is_vector_bitwise_cone_root(Node * n)2297 static bool is_vector_bitwise_cone_root(Node* n) {
2298 if (!is_vector_bitwise_op(n)) {
2299 return false;
2300 }
2301 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2302 if (is_vector_bitwise_op(n->fast_out(i))) {
2303 return false;
2304 }
2305 }
2306 return true;
2307 }
2308
collect_unique_inputs(Node * n,Unique_Node_List & partition,Unique_Node_List & inputs)2309 static uint collect_unique_inputs(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2310 uint cnt = 0;
2311 if (is_vector_bitwise_op(n)) {
2312 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2313 for (uint i = 1; i < n->req(); i++) {
2314 Node* in = n->in(i);
2315 bool skip = VectorNode::is_all_ones_vector(in);
2316 if (!skip && !inputs.member(in)) {
2317 inputs.push(in);
2318 cnt++;
2319 }
2320 }
2321 assert(cnt <= 1, "not unary");
2322 } else {
2323 uint last_req = n->req();
2324 if (is_vector_ternary_bitwise_op(n)) {
2325 last_req = n->req() - 1; // skip last input
2326 }
2327 for (uint i = 1; i < last_req; i++) {
2328 Node* def = n->in(i);
2329 if (!inputs.member(def)) {
2330 inputs.push(def);
2331 cnt++;
2332 }
2333 }
2334 }
2335 partition.push(n);
2336 } else { // not a bitwise operations
2337 if (!inputs.member(n)) {
2338 inputs.push(n);
2339 cnt++;
2340 }
2341 }
2342 return cnt;
2343 }
2344
collect_logic_cone_roots(Unique_Node_List & list)2345 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2346 Unique_Node_List useful_nodes;
2347 C->identify_useful_nodes(useful_nodes);
2348
2349 for (uint i = 0; i < useful_nodes.size(); i++) {
2350 Node* n = useful_nodes.at(i);
2351 if (is_vector_bitwise_cone_root(n)) {
2352 list.push(n);
2353 }
2354 }
2355 }
2356
xform_to_MacroLogicV(PhaseIterGVN & igvn,const TypeVect * vt,Unique_Node_List & partition,Unique_Node_List & inputs)2357 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2358 const TypeVect* vt,
2359 Unique_Node_List& partition,
2360 Unique_Node_List& inputs) {
2361 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2362 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2363 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2364
2365 Node* in1 = inputs.at(0);
2366 Node* in2 = inputs.at(1);
2367 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2368
2369 uint func = compute_truth_table(partition, inputs);
2370 return igvn.transform(MacroLogicVNode::make(igvn, in3, in2, in1, func, vt));
2371 }
2372
extract_bit(uint func,uint pos)2373 static uint extract_bit(uint func, uint pos) {
2374 return (func & (1 << pos)) >> pos;
2375 }
2376
2377 //
2378 // A macro logic node represents a truth table. It has 4 inputs,
2379 // First three inputs corresponds to 3 columns of a truth table
2380 // and fourth input captures the logic function.
2381 //
2382 // eg. fn = (in1 AND in2) OR in3;
2383 //
2384 // MacroNode(in1,in2,in3,fn)
2385 //
2386 // -----------------
2387 // in1 in2 in3 fn
2388 // -----------------
2389 // 0 0 0 0
2390 // 0 0 1 1
2391 // 0 1 0 0
2392 // 0 1 1 1
2393 // 1 0 0 0
2394 // 1 0 1 1
2395 // 1 1 0 1
2396 // 1 1 1 1
2397 //
2398
eval_macro_logic_op(uint func,uint in1,uint in2,uint in3)2399 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2400 int res = 0;
2401 for (int i = 0; i < 8; i++) {
2402 int bit1 = extract_bit(in1, i);
2403 int bit2 = extract_bit(in2, i);
2404 int bit3 = extract_bit(in3, i);
2405
2406 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2407 int func_bit = extract_bit(func, func_bit_pos);
2408
2409 res |= func_bit << i;
2410 }
2411 return res;
2412 }
2413
eval_operand(Node * n,ResourceHashtable<Node *,uint> & eval_map)2414 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2415 assert(n != NULL, "");
2416 assert(eval_map.contains(n), "absent");
2417 return *(eval_map.get(n));
2418 }
2419
eval_operands(Node * n,uint & func1,uint & func2,uint & func3,ResourceHashtable<Node *,uint> & eval_map)2420 static void eval_operands(Node* n,
2421 uint& func1, uint& func2, uint& func3,
2422 ResourceHashtable<Node*,uint>& eval_map) {
2423 assert(is_vector_bitwise_op(n), "");
2424 func1 = eval_operand(n->in(1), eval_map);
2425
2426 if (is_vector_binary_bitwise_op(n)) {
2427 func2 = eval_operand(n->in(2), eval_map);
2428 } else if (is_vector_ternary_bitwise_op(n)) {
2429 func2 = eval_operand(n->in(2), eval_map);
2430 func3 = eval_operand(n->in(3), eval_map);
2431 } else {
2432 assert(is_vector_unary_bitwise_op(n), "not unary");
2433 }
2434 }
2435
compute_truth_table(Unique_Node_List & partition,Unique_Node_List & inputs)2436 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2437 assert(inputs.size() <= 3, "sanity");
2438 ResourceMark rm;
2439 uint res = 0;
2440 ResourceHashtable<Node*,uint> eval_map;
2441
2442 // Populate precomputed functions for inputs.
2443 // Each input corresponds to one column of 3 input truth-table.
2444 uint input_funcs[] = { 0xAA, // (_, _, a) -> a
2445 0xCC, // (_, b, _) -> b
2446 0xF0 }; // (c, _, _) -> c
2447 for (uint i = 0; i < inputs.size(); i++) {
2448 eval_map.put(inputs.at(i), input_funcs[i]);
2449 }
2450
2451 for (uint i = 0; i < partition.size(); i++) {
2452 Node* n = partition.at(i);
2453
2454 uint func1 = 0, func2 = 0, func3 = 0;
2455 eval_operands(n, func1, func2, func3, eval_map);
2456
2457 switch (n->Opcode()) {
2458 case Op_OrV:
2459 assert(func3 == 0, "not binary");
2460 res = func1 | func2;
2461 break;
2462 case Op_AndV:
2463 assert(func3 == 0, "not binary");
2464 res = func1 & func2;
2465 break;
2466 case Op_XorV:
2467 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2468 assert(func2 == 0 && func3 == 0, "not unary");
2469 res = (~func1) & 0xFF;
2470 } else {
2471 assert(func3 == 0, "not binary");
2472 res = func1 ^ func2;
2473 }
2474 break;
2475 case Op_MacroLogicV:
2476 // Ordering of inputs may change during evaluation of sub-tree
2477 // containing MacroLogic node as a child node, thus a re-evaluation
2478 // makes sure that function is evaluated in context of current
2479 // inputs.
2480 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2481 break;
2482
2483 default: assert(false, "not supported: %s", n->Name());
2484 }
2485 assert(res <= 0xFF, "invalid");
2486 eval_map.put(n, res);
2487 }
2488 return res;
2489 }
2490
compute_logic_cone(Node * n,Unique_Node_List & partition,Unique_Node_List & inputs)2491 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2492 assert(partition.size() == 0, "not empty");
2493 assert(inputs.size() == 0, "not empty");
2494 if (is_vector_ternary_bitwise_op(n)) {
2495 return false;
2496 }
2497
2498 bool is_unary_op = is_vector_unary_bitwise_op(n);
2499 if (is_unary_op) {
2500 assert(collect_unique_inputs(n, partition, inputs) == 1, "not unary");
2501 return false; // too few inputs
2502 }
2503
2504 assert(is_vector_binary_bitwise_op(n), "not binary");
2505 Node* in1 = n->in(1);
2506 Node* in2 = n->in(2);
2507
2508 int in1_unique_inputs_cnt = collect_unique_inputs(in1, partition, inputs);
2509 int in2_unique_inputs_cnt = collect_unique_inputs(in2, partition, inputs);
2510 partition.push(n);
2511
2512 // Too many inputs?
2513 if (inputs.size() > 3) {
2514 partition.clear();
2515 inputs.clear();
2516 { // Recompute in2 inputs
2517 Unique_Node_List not_used;
2518 in2_unique_inputs_cnt = collect_unique_inputs(in2, not_used, not_used);
2519 }
2520 // Pick the node with minimum number of inputs.
2521 if (in1_unique_inputs_cnt >= 3 && in2_unique_inputs_cnt >= 3) {
2522 return false; // still too many inputs
2523 }
2524 // Recompute partition & inputs.
2525 Node* child = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in1 : in2);
2526 collect_unique_inputs(child, partition, inputs);
2527
2528 Node* other_input = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in2 : in1);
2529 inputs.push(other_input);
2530
2531 partition.push(n);
2532 }
2533
2534 return (partition.size() == 2 || partition.size() == 3) &&
2535 (inputs.size() == 2 || inputs.size() == 3);
2536 }
2537
2538
process_logic_cone_root(PhaseIterGVN & igvn,Node * n,VectorSet & visited)2539 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2540 assert(is_vector_bitwise_op(n), "not a root");
2541
2542 visited.set(n->_idx);
2543
2544 // 1) Do a DFS walk over the logic cone.
2545 for (uint i = 1; i < n->req(); i++) {
2546 Node* in = n->in(i);
2547 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2548 process_logic_cone_root(igvn, in, visited);
2549 }
2550 }
2551
2552 // 2) Bottom up traversal: Merge node[s] with
2553 // the parent to form macro logic node.
2554 Unique_Node_List partition;
2555 Unique_Node_List inputs;
2556 if (compute_logic_cone(n, partition, inputs)) {
2557 const TypeVect* vt = n->bottom_type()->is_vect();
2558 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2559 igvn.replace_node(n, macro_logic);
2560 }
2561 }
2562
optimize_logic_cones(PhaseIterGVN & igvn)2563 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2564 ResourceMark rm;
2565 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2566 Unique_Node_List list;
2567 collect_logic_cone_roots(list);
2568
2569 while (list.size() > 0) {
2570 Node* n = list.pop();
2571 const TypeVect* vt = n->bottom_type()->is_vect();
2572 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2573 if (supported) {
2574 VectorSet visited(comp_arena());
2575 process_logic_cone_root(igvn, n, visited);
2576 }
2577 }
2578 }
2579 }
2580
2581 //------------------------------Code_Gen---------------------------------------
2582 // Given a graph, generate code for it
Code_Gen()2583 void Compile::Code_Gen() {
2584 if (failing()) {
2585 return;
2586 }
2587
2588 // Perform instruction selection. You might think we could reclaim Matcher
2589 // memory PDQ, but actually the Matcher is used in generating spill code.
2590 // Internals of the Matcher (including some VectorSets) must remain live
2591 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2592 // set a bit in reclaimed memory.
2593
2594 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2595 // nodes. Mapping is only valid at the root of each matched subtree.
2596 NOT_PRODUCT( verify_graph_edges(); )
2597
2598 Matcher matcher;
2599 _matcher = &matcher;
2600 {
2601 TracePhase tp("matcher", &timers[_t_matcher]);
2602 matcher.match();
2603 if (failing()) {
2604 return;
2605 }
2606 }
2607
2608 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2609 // nodes. Mapping is only valid at the root of each matched subtree.
2610 NOT_PRODUCT( verify_graph_edges(); )
2611
2612 // If you have too many nodes, or if matching has failed, bail out
2613 check_node_count(0, "out of nodes matching instructions");
2614 if (failing()) {
2615 return;
2616 }
2617
2618 print_method(PHASE_MATCHING, 2);
2619
2620 // Build a proper-looking CFG
2621 PhaseCFG cfg(node_arena(), root(), matcher);
2622 _cfg = &cfg;
2623 {
2624 TracePhase tp("scheduler", &timers[_t_scheduler]);
2625 bool success = cfg.do_global_code_motion();
2626 if (!success) {
2627 return;
2628 }
2629
2630 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2631 NOT_PRODUCT( verify_graph_edges(); )
2632 debug_only( cfg.verify(); )
2633 }
2634
2635 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2636 _regalloc = ®alloc;
2637 {
2638 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2639 // Perform register allocation. After Chaitin, use-def chains are
2640 // no longer accurate (at spill code) and so must be ignored.
2641 // Node->LRG->reg mappings are still accurate.
2642 _regalloc->Register_Allocate();
2643
2644 // Bail out if the allocator builds too many nodes
2645 if (failing()) {
2646 return;
2647 }
2648 }
2649
2650 // Prior to register allocation we kept empty basic blocks in case the
2651 // the allocator needed a place to spill. After register allocation we
2652 // are not adding any new instructions. If any basic block is empty, we
2653 // can now safely remove it.
2654 {
2655 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2656 cfg.remove_empty_blocks();
2657 if (do_freq_based_layout()) {
2658 PhaseBlockLayout layout(cfg);
2659 } else {
2660 cfg.set_loop_alignment();
2661 }
2662 cfg.fixup_flow();
2663 }
2664
2665 // Apply peephole optimizations
2666 if( OptoPeephole ) {
2667 TracePhase tp("peephole", &timers[_t_peephole]);
2668 PhasePeephole peep( _regalloc, cfg);
2669 peep.do_transform();
2670 }
2671
2672 // Do late expand if CPU requires this.
2673 if (Matcher::require_postalloc_expand) {
2674 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2675 cfg.postalloc_expand(_regalloc);
2676 }
2677
2678 // Convert Nodes to instruction bits in a buffer
2679 {
2680 TracePhase tp("output", &timers[_t_output]);
2681 PhaseOutput output;
2682 output.Output();
2683 if (failing()) return;
2684 output.install();
2685 }
2686
2687 print_method(PHASE_FINAL_CODE);
2688
2689 // He's dead, Jim.
2690 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2691 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2692 }
2693
2694 //------------------------------Final_Reshape_Counts---------------------------
2695 // This class defines counters to help identify when a method
2696 // may/must be executed using hardware with only 24-bit precision.
2697 struct Final_Reshape_Counts : public StackObj {
2698 int _call_count; // count non-inlined 'common' calls
2699 int _float_count; // count float ops requiring 24-bit precision
2700 int _double_count; // count double ops requiring more precision
2701 int _java_call_count; // count non-inlined 'java' calls
2702 int _inner_loop_count; // count loops which need alignment
2703 VectorSet _visited; // Visitation flags
2704 Node_List _tests; // Set of IfNodes & PCTableNodes
2705
Final_Reshape_CountsFinal_Reshape_Counts2706 Final_Reshape_Counts() :
2707 _call_count(0), _float_count(0), _double_count(0),
2708 _java_call_count(0), _inner_loop_count(0),
2709 _visited( Thread::current()->resource_area() ) { }
2710
inc_call_countFinal_Reshape_Counts2711 void inc_call_count () { _call_count ++; }
inc_float_countFinal_Reshape_Counts2712 void inc_float_count () { _float_count ++; }
inc_double_countFinal_Reshape_Counts2713 void inc_double_count() { _double_count++; }
inc_java_call_countFinal_Reshape_Counts2714 void inc_java_call_count() { _java_call_count++; }
inc_inner_loop_countFinal_Reshape_Counts2715 void inc_inner_loop_count() { _inner_loop_count++; }
2716
get_call_countFinal_Reshape_Counts2717 int get_call_count () const { return _call_count ; }
get_float_countFinal_Reshape_Counts2718 int get_float_count () const { return _float_count ; }
get_double_countFinal_Reshape_Counts2719 int get_double_count() const { return _double_count; }
get_java_call_countFinal_Reshape_Counts2720 int get_java_call_count() const { return _java_call_count; }
get_inner_loop_countFinal_Reshape_Counts2721 int get_inner_loop_count() const { return _inner_loop_count; }
2722 };
2723
2724 #ifdef ASSERT
oop_offset_is_sane(const TypeInstPtr * tp)2725 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2726 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2727 // Make sure the offset goes inside the instance layout.
2728 return k->contains_field_offset(tp->offset());
2729 // Note that OffsetBot and OffsetTop are very negative.
2730 }
2731 #endif
2732
2733 // Eliminate trivially redundant StoreCMs and accumulate their
2734 // precedence edges.
eliminate_redundant_card_marks(Node * n)2735 void Compile::eliminate_redundant_card_marks(Node* n) {
2736 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2737 if (n->in(MemNode::Address)->outcnt() > 1) {
2738 // There are multiple users of the same address so it might be
2739 // possible to eliminate some of the StoreCMs
2740 Node* mem = n->in(MemNode::Memory);
2741 Node* adr = n->in(MemNode::Address);
2742 Node* val = n->in(MemNode::ValueIn);
2743 Node* prev = n;
2744 bool done = false;
2745 // Walk the chain of StoreCMs eliminating ones that match. As
2746 // long as it's a chain of single users then the optimization is
2747 // safe. Eliminating partially redundant StoreCMs would require
2748 // cloning copies down the other paths.
2749 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2750 if (adr == mem->in(MemNode::Address) &&
2751 val == mem->in(MemNode::ValueIn)) {
2752 // redundant StoreCM
2753 if (mem->req() > MemNode::OopStore) {
2754 // Hasn't been processed by this code yet.
2755 n->add_prec(mem->in(MemNode::OopStore));
2756 } else {
2757 // Already converted to precedence edge
2758 for (uint i = mem->req(); i < mem->len(); i++) {
2759 // Accumulate any precedence edges
2760 if (mem->in(i) != NULL) {
2761 n->add_prec(mem->in(i));
2762 }
2763 }
2764 // Everything above this point has been processed.
2765 done = true;
2766 }
2767 // Eliminate the previous StoreCM
2768 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2769 assert(mem->outcnt() == 0, "should be dead");
2770 mem->disconnect_inputs(NULL, this);
2771 } else {
2772 prev = mem;
2773 }
2774 mem = prev->in(MemNode::Memory);
2775 }
2776 }
2777 }
2778
2779 //------------------------------final_graph_reshaping_impl----------------------
2780 // Implement items 1-5 from final_graph_reshaping below.
final_graph_reshaping_impl(Node * n,Final_Reshape_Counts & frc)2781 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2782
2783 if ( n->outcnt() == 0 ) return; // dead node
2784 uint nop = n->Opcode();
2785
2786 // Check for 2-input instruction with "last use" on right input.
2787 // Swap to left input. Implements item (2).
2788 if( n->req() == 3 && // two-input instruction
2789 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2790 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2791 n->in(2)->outcnt() == 1 &&// right use IS a last use
2792 !n->in(2)->is_Con() ) { // right use is not a constant
2793 // Check for commutative opcode
2794 switch( nop ) {
2795 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2796 case Op_MaxI: case Op_MinI:
2797 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2798 case Op_AndL: case Op_XorL: case Op_OrL:
2799 case Op_AndI: case Op_XorI: case Op_OrI: {
2800 // Move "last use" input to left by swapping inputs
2801 n->swap_edges(1, 2);
2802 break;
2803 }
2804 default:
2805 break;
2806 }
2807 }
2808
2809 #ifdef ASSERT
2810 if( n->is_Mem() ) {
2811 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2812 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2813 // oop will be recorded in oop map if load crosses safepoint
2814 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2815 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2816 "raw memory operations should have control edge");
2817 }
2818 if (n->is_MemBar()) {
2819 MemBarNode* mb = n->as_MemBar();
2820 if (mb->trailing_store() || mb->trailing_load_store()) {
2821 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2822 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
2823 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2824 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2825 } else if (mb->leading()) {
2826 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2827 }
2828 }
2829 #endif
2830 // Count FPU ops and common calls, implements item (3)
2831 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2832 if (!gc_handled) {
2833 final_graph_reshaping_main_switch(n, frc, nop);
2834 }
2835
2836 // Collect CFG split points
2837 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2838 frc._tests.push(n);
2839 }
2840 }
2841
final_graph_reshaping_main_switch(Node * n,Final_Reshape_Counts & frc,uint nop)2842 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
2843 switch( nop ) {
2844 // Count all float operations that may use FPU
2845 case Op_AddF:
2846 case Op_SubF:
2847 case Op_MulF:
2848 case Op_DivF:
2849 case Op_NegF:
2850 case Op_ModF:
2851 case Op_ConvI2F:
2852 case Op_ConF:
2853 case Op_CmpF:
2854 case Op_CmpF3:
2855 // case Op_ConvL2F: // longs are split into 32-bit halves
2856 frc.inc_float_count();
2857 break;
2858
2859 case Op_ConvF2D:
2860 case Op_ConvD2F:
2861 frc.inc_float_count();
2862 frc.inc_double_count();
2863 break;
2864
2865 // Count all double operations that may use FPU
2866 case Op_AddD:
2867 case Op_SubD:
2868 case Op_MulD:
2869 case Op_DivD:
2870 case Op_NegD:
2871 case Op_ModD:
2872 case Op_ConvI2D:
2873 case Op_ConvD2I:
2874 // case Op_ConvL2D: // handled by leaf call
2875 // case Op_ConvD2L: // handled by leaf call
2876 case Op_ConD:
2877 case Op_CmpD:
2878 case Op_CmpD3:
2879 frc.inc_double_count();
2880 break;
2881 case Op_Opaque1: // Remove Opaque Nodes before matching
2882 case Op_Opaque2: // Remove Opaque Nodes before matching
2883 case Op_Opaque3:
2884 n->subsume_by(n->in(1), this);
2885 break;
2886 case Op_CallStaticJava:
2887 case Op_CallJava:
2888 case Op_CallDynamicJava:
2889 frc.inc_java_call_count(); // Count java call site;
2890 case Op_CallRuntime:
2891 case Op_CallLeaf:
2892 case Op_CallLeafNoFP: {
2893 assert (n->is_Call(), "");
2894 CallNode *call = n->as_Call();
2895 // Count call sites where the FP mode bit would have to be flipped.
2896 // Do not count uncommon runtime calls:
2897 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2898 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2899 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2900 frc.inc_call_count(); // Count the call site
2901 } else { // See if uncommon argument is shared
2902 Node *n = call->in(TypeFunc::Parms);
2903 int nop = n->Opcode();
2904 // Clone shared simple arguments to uncommon calls, item (1).
2905 if (n->outcnt() > 1 &&
2906 !n->is_Proj() &&
2907 nop != Op_CreateEx &&
2908 nop != Op_CheckCastPP &&
2909 nop != Op_DecodeN &&
2910 nop != Op_DecodeNKlass &&
2911 !n->is_Mem() &&
2912 !n->is_Phi()) {
2913 Node *x = n->clone();
2914 call->set_req(TypeFunc::Parms, x);
2915 }
2916 }
2917 break;
2918 }
2919
2920 case Op_StoreD:
2921 case Op_LoadD:
2922 case Op_LoadD_unaligned:
2923 frc.inc_double_count();
2924 goto handle_mem;
2925 case Op_StoreF:
2926 case Op_LoadF:
2927 frc.inc_float_count();
2928 goto handle_mem;
2929
2930 case Op_StoreCM:
2931 {
2932 // Convert OopStore dependence into precedence edge
2933 Node* prec = n->in(MemNode::OopStore);
2934 n->del_req(MemNode::OopStore);
2935 n->add_prec(prec);
2936 eliminate_redundant_card_marks(n);
2937 }
2938
2939 // fall through
2940
2941 case Op_StoreB:
2942 case Op_StoreC:
2943 case Op_StorePConditional:
2944 case Op_StoreI:
2945 case Op_StoreL:
2946 case Op_StoreIConditional:
2947 case Op_StoreLConditional:
2948 case Op_CompareAndSwapB:
2949 case Op_CompareAndSwapS:
2950 case Op_CompareAndSwapI:
2951 case Op_CompareAndSwapL:
2952 case Op_CompareAndSwapP:
2953 case Op_CompareAndSwapN:
2954 case Op_WeakCompareAndSwapB:
2955 case Op_WeakCompareAndSwapS:
2956 case Op_WeakCompareAndSwapI:
2957 case Op_WeakCompareAndSwapL:
2958 case Op_WeakCompareAndSwapP:
2959 case Op_WeakCompareAndSwapN:
2960 case Op_CompareAndExchangeB:
2961 case Op_CompareAndExchangeS:
2962 case Op_CompareAndExchangeI:
2963 case Op_CompareAndExchangeL:
2964 case Op_CompareAndExchangeP:
2965 case Op_CompareAndExchangeN:
2966 case Op_GetAndAddS:
2967 case Op_GetAndAddB:
2968 case Op_GetAndAddI:
2969 case Op_GetAndAddL:
2970 case Op_GetAndSetS:
2971 case Op_GetAndSetB:
2972 case Op_GetAndSetI:
2973 case Op_GetAndSetL:
2974 case Op_GetAndSetP:
2975 case Op_GetAndSetN:
2976 case Op_StoreP:
2977 case Op_StoreN:
2978 case Op_StoreNKlass:
2979 case Op_LoadB:
2980 case Op_LoadUB:
2981 case Op_LoadUS:
2982 case Op_LoadI:
2983 case Op_LoadKlass:
2984 case Op_LoadNKlass:
2985 case Op_LoadL:
2986 case Op_LoadL_unaligned:
2987 case Op_LoadPLocked:
2988 case Op_LoadP:
2989 case Op_LoadN:
2990 case Op_LoadRange:
2991 case Op_LoadS: {
2992 handle_mem:
2993 #ifdef ASSERT
2994 if( VerifyOptoOopOffsets ) {
2995 MemNode* mem = n->as_Mem();
2996 // Check to see if address types have grounded out somehow.
2997 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2998 assert( !tp || oop_offset_is_sane(tp), "" );
2999 }
3000 #endif
3001 break;
3002 }
3003
3004 case Op_AddP: { // Assert sane base pointers
3005 Node *addp = n->in(AddPNode::Address);
3006 assert( !addp->is_AddP() ||
3007 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3008 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3009 "Base pointers must match (addp %u)", addp->_idx );
3010 #ifdef _LP64
3011 if ((UseCompressedOops || UseCompressedClassPointers) &&
3012 addp->Opcode() == Op_ConP &&
3013 addp == n->in(AddPNode::Base) &&
3014 n->in(AddPNode::Offset)->is_Con()) {
3015 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3016 // on the platform and on the compressed oops mode.
3017 // Use addressing with narrow klass to load with offset on x86.
3018 // Some platforms can use the constant pool to load ConP.
3019 // Do this transformation here since IGVN will convert ConN back to ConP.
3020 const Type* t = addp->bottom_type();
3021 bool is_oop = t->isa_oopptr() != NULL;
3022 bool is_klass = t->isa_klassptr() != NULL;
3023
3024 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3025 (is_klass && Matcher::const_klass_prefer_decode())) {
3026 Node* nn = NULL;
3027
3028 int op = is_oop ? Op_ConN : Op_ConNKlass;
3029
3030 // Look for existing ConN node of the same exact type.
3031 Node* r = root();
3032 uint cnt = r->outcnt();
3033 for (uint i = 0; i < cnt; i++) {
3034 Node* m = r->raw_out(i);
3035 if (m!= NULL && m->Opcode() == op &&
3036 m->bottom_type()->make_ptr() == t) {
3037 nn = m;
3038 break;
3039 }
3040 }
3041 if (nn != NULL) {
3042 // Decode a narrow oop to match address
3043 // [R12 + narrow_oop_reg<<3 + offset]
3044 if (is_oop) {
3045 nn = new DecodeNNode(nn, t);
3046 } else {
3047 nn = new DecodeNKlassNode(nn, t);
3048 }
3049 // Check for succeeding AddP which uses the same Base.
3050 // Otherwise we will run into the assertion above when visiting that guy.
3051 for (uint i = 0; i < n->outcnt(); ++i) {
3052 Node *out_i = n->raw_out(i);
3053 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3054 out_i->set_req(AddPNode::Base, nn);
3055 #ifdef ASSERT
3056 for (uint j = 0; j < out_i->outcnt(); ++j) {
3057 Node *out_j = out_i->raw_out(j);
3058 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3059 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3060 }
3061 #endif
3062 }
3063 }
3064 n->set_req(AddPNode::Base, nn);
3065 n->set_req(AddPNode::Address, nn);
3066 if (addp->outcnt() == 0) {
3067 addp->disconnect_inputs(NULL, this);
3068 }
3069 }
3070 }
3071 }
3072 #endif
3073 // platform dependent reshaping of the address expression
3074 reshape_address(n->as_AddP());
3075 break;
3076 }
3077
3078 case Op_CastPP: {
3079 // Remove CastPP nodes to gain more freedom during scheduling but
3080 // keep the dependency they encode as control or precedence edges
3081 // (if control is set already) on memory operations. Some CastPP
3082 // nodes don't have a control (don't carry a dependency): skip
3083 // those.
3084 if (n->in(0) != NULL) {
3085 ResourceMark rm;
3086 Unique_Node_List wq;
3087 wq.push(n);
3088 for (uint next = 0; next < wq.size(); ++next) {
3089 Node *m = wq.at(next);
3090 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3091 Node* use = m->fast_out(i);
3092 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3093 use->ensure_control_or_add_prec(n->in(0));
3094 } else {
3095 switch(use->Opcode()) {
3096 case Op_AddP:
3097 case Op_DecodeN:
3098 case Op_DecodeNKlass:
3099 case Op_CheckCastPP:
3100 case Op_CastPP:
3101 wq.push(use);
3102 break;
3103 }
3104 }
3105 }
3106 }
3107 }
3108 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3109 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3110 Node* in1 = n->in(1);
3111 const Type* t = n->bottom_type();
3112 Node* new_in1 = in1->clone();
3113 new_in1->as_DecodeN()->set_type(t);
3114
3115 if (!Matcher::narrow_oop_use_complex_address()) {
3116 //
3117 // x86, ARM and friends can handle 2 adds in addressing mode
3118 // and Matcher can fold a DecodeN node into address by using
3119 // a narrow oop directly and do implicit NULL check in address:
3120 //
3121 // [R12 + narrow_oop_reg<<3 + offset]
3122 // NullCheck narrow_oop_reg
3123 //
3124 // On other platforms (Sparc) we have to keep new DecodeN node and
3125 // use it to do implicit NULL check in address:
3126 //
3127 // decode_not_null narrow_oop_reg, base_reg
3128 // [base_reg + offset]
3129 // NullCheck base_reg
3130 //
3131 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3132 // to keep the information to which NULL check the new DecodeN node
3133 // corresponds to use it as value in implicit_null_check().
3134 //
3135 new_in1->set_req(0, n->in(0));
3136 }
3137
3138 n->subsume_by(new_in1, this);
3139 if (in1->outcnt() == 0) {
3140 in1->disconnect_inputs(NULL, this);
3141 }
3142 } else {
3143 n->subsume_by(n->in(1), this);
3144 if (n->outcnt() == 0) {
3145 n->disconnect_inputs(NULL, this);
3146 }
3147 }
3148 break;
3149 }
3150 #ifdef _LP64
3151 case Op_CmpP:
3152 // Do this transformation here to preserve CmpPNode::sub() and
3153 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3154 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3155 Node* in1 = n->in(1);
3156 Node* in2 = n->in(2);
3157 if (!in1->is_DecodeNarrowPtr()) {
3158 in2 = in1;
3159 in1 = n->in(2);
3160 }
3161 assert(in1->is_DecodeNarrowPtr(), "sanity");
3162
3163 Node* new_in2 = NULL;
3164 if (in2->is_DecodeNarrowPtr()) {
3165 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3166 new_in2 = in2->in(1);
3167 } else if (in2->Opcode() == Op_ConP) {
3168 const Type* t = in2->bottom_type();
3169 if (t == TypePtr::NULL_PTR) {
3170 assert(in1->is_DecodeN(), "compare klass to null?");
3171 // Don't convert CmpP null check into CmpN if compressed
3172 // oops implicit null check is not generated.
3173 // This will allow to generate normal oop implicit null check.
3174 if (Matcher::gen_narrow_oop_implicit_null_checks())
3175 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3176 //
3177 // This transformation together with CastPP transformation above
3178 // will generated code for implicit NULL checks for compressed oops.
3179 //
3180 // The original code after Optimize()
3181 //
3182 // LoadN memory, narrow_oop_reg
3183 // decode narrow_oop_reg, base_reg
3184 // CmpP base_reg, NULL
3185 // CastPP base_reg // NotNull
3186 // Load [base_reg + offset], val_reg
3187 //
3188 // after these transformations will be
3189 //
3190 // LoadN memory, narrow_oop_reg
3191 // CmpN narrow_oop_reg, NULL
3192 // decode_not_null narrow_oop_reg, base_reg
3193 // Load [base_reg + offset], val_reg
3194 //
3195 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3196 // since narrow oops can be used in debug info now (see the code in
3197 // final_graph_reshaping_walk()).
3198 //
3199 // At the end the code will be matched to
3200 // on x86:
3201 //
3202 // Load_narrow_oop memory, narrow_oop_reg
3203 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3204 // NullCheck narrow_oop_reg
3205 //
3206 // and on sparc:
3207 //
3208 // Load_narrow_oop memory, narrow_oop_reg
3209 // decode_not_null narrow_oop_reg, base_reg
3210 // Load [base_reg + offset], val_reg
3211 // NullCheck base_reg
3212 //
3213 } else if (t->isa_oopptr()) {
3214 new_in2 = ConNode::make(t->make_narrowoop());
3215 } else if (t->isa_klassptr()) {
3216 new_in2 = ConNode::make(t->make_narrowklass());
3217 }
3218 }
3219 if (new_in2 != NULL) {
3220 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3221 n->subsume_by(cmpN, this);
3222 if (in1->outcnt() == 0) {
3223 in1->disconnect_inputs(NULL, this);
3224 }
3225 if (in2->outcnt() == 0) {
3226 in2->disconnect_inputs(NULL, this);
3227 }
3228 }
3229 }
3230 break;
3231
3232 case Op_DecodeN:
3233 case Op_DecodeNKlass:
3234 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3235 // DecodeN could be pinned when it can't be fold into
3236 // an address expression, see the code for Op_CastPP above.
3237 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3238 break;
3239
3240 case Op_EncodeP:
3241 case Op_EncodePKlass: {
3242 Node* in1 = n->in(1);
3243 if (in1->is_DecodeNarrowPtr()) {
3244 n->subsume_by(in1->in(1), this);
3245 } else if (in1->Opcode() == Op_ConP) {
3246 const Type* t = in1->bottom_type();
3247 if (t == TypePtr::NULL_PTR) {
3248 assert(t->isa_oopptr(), "null klass?");
3249 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3250 } else if (t->isa_oopptr()) {
3251 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3252 } else if (t->isa_klassptr()) {
3253 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3254 }
3255 }
3256 if (in1->outcnt() == 0) {
3257 in1->disconnect_inputs(NULL, this);
3258 }
3259 break;
3260 }
3261
3262 case Op_Proj: {
3263 if (OptimizeStringConcat) {
3264 ProjNode* p = n->as_Proj();
3265 if (p->_is_io_use) {
3266 // Separate projections were used for the exception path which
3267 // are normally removed by a late inline. If it wasn't inlined
3268 // then they will hang around and should just be replaced with
3269 // the original one.
3270 Node* proj = NULL;
3271 // Replace with just one
3272 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3273 Node *use = i.get();
3274 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3275 proj = use;
3276 break;
3277 }
3278 }
3279 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3280 if (proj != NULL) {
3281 p->subsume_by(proj, this);
3282 }
3283 }
3284 }
3285 break;
3286 }
3287
3288 case Op_Phi:
3289 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3290 // The EncodeP optimization may create Phi with the same edges
3291 // for all paths. It is not handled well by Register Allocator.
3292 Node* unique_in = n->in(1);
3293 assert(unique_in != NULL, "");
3294 uint cnt = n->req();
3295 for (uint i = 2; i < cnt; i++) {
3296 Node* m = n->in(i);
3297 assert(m != NULL, "");
3298 if (unique_in != m)
3299 unique_in = NULL;
3300 }
3301 if (unique_in != NULL) {
3302 n->subsume_by(unique_in, this);
3303 }
3304 }
3305 break;
3306
3307 #endif
3308
3309 #ifdef ASSERT
3310 case Op_CastII:
3311 // Verify that all range check dependent CastII nodes were removed.
3312 if (n->isa_CastII()->has_range_check()) {
3313 n->dump(3);
3314 assert(false, "Range check dependent CastII node was not removed");
3315 }
3316 break;
3317 #endif
3318
3319 case Op_ModI:
3320 if (UseDivMod) {
3321 // Check if a%b and a/b both exist
3322 Node* d = n->find_similar(Op_DivI);
3323 if (d) {
3324 // Replace them with a fused divmod if supported
3325 if (Matcher::has_match_rule(Op_DivModI)) {
3326 DivModINode* divmod = DivModINode::make(n);
3327 d->subsume_by(divmod->div_proj(), this);
3328 n->subsume_by(divmod->mod_proj(), this);
3329 } else {
3330 // replace a%b with a-((a/b)*b)
3331 Node* mult = new MulINode(d, d->in(2));
3332 Node* sub = new SubINode(d->in(1), mult);
3333 n->subsume_by(sub, this);
3334 }
3335 }
3336 }
3337 break;
3338
3339 case Op_ModL:
3340 if (UseDivMod) {
3341 // Check if a%b and a/b both exist
3342 Node* d = n->find_similar(Op_DivL);
3343 if (d) {
3344 // Replace them with a fused divmod if supported
3345 if (Matcher::has_match_rule(Op_DivModL)) {
3346 DivModLNode* divmod = DivModLNode::make(n);
3347 d->subsume_by(divmod->div_proj(), this);
3348 n->subsume_by(divmod->mod_proj(), this);
3349 } else {
3350 // replace a%b with a-((a/b)*b)
3351 Node* mult = new MulLNode(d, d->in(2));
3352 Node* sub = new SubLNode(d->in(1), mult);
3353 n->subsume_by(sub, this);
3354 }
3355 }
3356 }
3357 break;
3358
3359 case Op_LoadVector:
3360 case Op_StoreVector:
3361 break;
3362
3363 case Op_AddReductionVI:
3364 case Op_AddReductionVL:
3365 case Op_AddReductionVF:
3366 case Op_AddReductionVD:
3367 case Op_MulReductionVI:
3368 case Op_MulReductionVL:
3369 case Op_MulReductionVF:
3370 case Op_MulReductionVD:
3371 case Op_MinReductionV:
3372 case Op_MaxReductionV:
3373 case Op_AndReductionV:
3374 case Op_OrReductionV:
3375 case Op_XorReductionV:
3376 break;
3377
3378 case Op_PackB:
3379 case Op_PackS:
3380 case Op_PackI:
3381 case Op_PackF:
3382 case Op_PackL:
3383 case Op_PackD:
3384 if (n->req()-1 > 2) {
3385 // Replace many operand PackNodes with a binary tree for matching
3386 PackNode* p = (PackNode*) n;
3387 Node* btp = p->binary_tree_pack(1, n->req());
3388 n->subsume_by(btp, this);
3389 }
3390 break;
3391 case Op_Loop:
3392 case Op_CountedLoop:
3393 case Op_OuterStripMinedLoop:
3394 if (n->as_Loop()->is_inner_loop()) {
3395 frc.inc_inner_loop_count();
3396 }
3397 n->as_Loop()->verify_strip_mined(0);
3398 break;
3399 case Op_LShiftI:
3400 case Op_RShiftI:
3401 case Op_URShiftI:
3402 case Op_LShiftL:
3403 case Op_RShiftL:
3404 case Op_URShiftL:
3405 if (Matcher::need_masked_shift_count) {
3406 // The cpu's shift instructions don't restrict the count to the
3407 // lower 5/6 bits. We need to do the masking ourselves.
3408 Node* in2 = n->in(2);
3409 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3410 const TypeInt* t = in2->find_int_type();
3411 if (t != NULL && t->is_con()) {
3412 juint shift = t->get_con();
3413 if (shift > mask) { // Unsigned cmp
3414 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3415 }
3416 } else {
3417 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3418 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3419 n->set_req(2, shift);
3420 }
3421 }
3422 if (in2->outcnt() == 0) { // Remove dead node
3423 in2->disconnect_inputs(NULL, this);
3424 }
3425 }
3426 break;
3427 case Op_MemBarStoreStore:
3428 case Op_MemBarRelease:
3429 // Break the link with AllocateNode: it is no longer useful and
3430 // confuses register allocation.
3431 if (n->req() > MemBarNode::Precedent) {
3432 n->set_req(MemBarNode::Precedent, top());
3433 }
3434 break;
3435 case Op_MemBarAcquire: {
3436 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3437 // At parse time, the trailing MemBarAcquire for a volatile load
3438 // is created with an edge to the load. After optimizations,
3439 // that input may be a chain of Phis. If those phis have no
3440 // other use, then the MemBarAcquire keeps them alive and
3441 // register allocation can be confused.
3442 ResourceMark rm;
3443 Unique_Node_List wq;
3444 wq.push(n->in(MemBarNode::Precedent));
3445 n->set_req(MemBarNode::Precedent, top());
3446 while (wq.size() > 0) {
3447 Node* m = wq.pop();
3448 if (m->outcnt() == 0) {
3449 for (uint j = 0; j < m->req(); j++) {
3450 Node* in = m->in(j);
3451 if (in != NULL) {
3452 wq.push(in);
3453 }
3454 }
3455 m->disconnect_inputs(NULL, this);
3456 }
3457 }
3458 }
3459 break;
3460 }
3461 case Op_RangeCheck: {
3462 RangeCheckNode* rc = n->as_RangeCheck();
3463 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3464 n->subsume_by(iff, this);
3465 frc._tests.push(iff);
3466 break;
3467 }
3468 case Op_ConvI2L: {
3469 if (!Matcher::convi2l_type_required) {
3470 // Code generation on some platforms doesn't need accurate
3471 // ConvI2L types. Widening the type can help remove redundant
3472 // address computations.
3473 n->as_Type()->set_type(TypeLong::INT);
3474 ResourceMark rm;
3475 Unique_Node_List wq;
3476 wq.push(n);
3477 for (uint next = 0; next < wq.size(); next++) {
3478 Node *m = wq.at(next);
3479
3480 for(;;) {
3481 // Loop over all nodes with identical inputs edges as m
3482 Node* k = m->find_similar(m->Opcode());
3483 if (k == NULL) {
3484 break;
3485 }
3486 // Push their uses so we get a chance to remove node made
3487 // redundant
3488 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3489 Node* u = k->fast_out(i);
3490 if (u->Opcode() == Op_LShiftL ||
3491 u->Opcode() == Op_AddL ||
3492 u->Opcode() == Op_SubL ||
3493 u->Opcode() == Op_AddP) {
3494 wq.push(u);
3495 }
3496 }
3497 // Replace all nodes with identical edges as m with m
3498 k->subsume_by(m, this);
3499 }
3500 }
3501 }
3502 break;
3503 }
3504 case Op_CmpUL: {
3505 if (!Matcher::has_match_rule(Op_CmpUL)) {
3506 // No support for unsigned long comparisons
3507 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3508 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3509 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3510 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3511 Node* andl = new AndLNode(orl, remove_sign_mask);
3512 Node* cmp = new CmpLNode(andl, n->in(2));
3513 n->subsume_by(cmp, this);
3514 }
3515 break;
3516 }
3517 default:
3518 assert(!n->is_Call(), "");
3519 assert(!n->is_Mem(), "");
3520 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3521 break;
3522 }
3523 }
3524
3525 //------------------------------final_graph_reshaping_walk---------------------
3526 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3527 // requires that the walk visits a node's inputs before visiting the node.
final_graph_reshaping_walk(Node_Stack & nstack,Node * root,Final_Reshape_Counts & frc)3528 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3529 ResourceArea *area = Thread::current()->resource_area();
3530 Unique_Node_List sfpt(area);
3531
3532 frc._visited.set(root->_idx); // first, mark node as visited
3533 uint cnt = root->req();
3534 Node *n = root;
3535 uint i = 0;
3536 while (true) {
3537 if (i < cnt) {
3538 // Place all non-visited non-null inputs onto stack
3539 Node* m = n->in(i);
3540 ++i;
3541 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3542 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3543 // compute worst case interpreter size in case of a deoptimization
3544 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3545
3546 sfpt.push(m);
3547 }
3548 cnt = m->req();
3549 nstack.push(n, i); // put on stack parent and next input's index
3550 n = m;
3551 i = 0;
3552 }
3553 } else {
3554 // Now do post-visit work
3555 final_graph_reshaping_impl( n, frc );
3556 if (nstack.is_empty())
3557 break; // finished
3558 n = nstack.node(); // Get node from stack
3559 cnt = n->req();
3560 i = nstack.index();
3561 nstack.pop(); // Shift to the next node on stack
3562 }
3563 }
3564
3565 // Skip next transformation if compressed oops are not used.
3566 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3567 (!UseCompressedOops && !UseCompressedClassPointers))
3568 return;
3569
3570 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3571 // It could be done for an uncommon traps or any safepoints/calls
3572 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3573 while (sfpt.size() > 0) {
3574 n = sfpt.pop();
3575 JVMState *jvms = n->as_SafePoint()->jvms();
3576 assert(jvms != NULL, "sanity");
3577 int start = jvms->debug_start();
3578 int end = n->req();
3579 bool is_uncommon = (n->is_CallStaticJava() &&
3580 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3581 for (int j = start; j < end; j++) {
3582 Node* in = n->in(j);
3583 if (in->is_DecodeNarrowPtr()) {
3584 bool safe_to_skip = true;
3585 if (!is_uncommon ) {
3586 // Is it safe to skip?
3587 for (uint i = 0; i < in->outcnt(); i++) {
3588 Node* u = in->raw_out(i);
3589 if (!u->is_SafePoint() ||
3590 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3591 safe_to_skip = false;
3592 }
3593 }
3594 }
3595 if (safe_to_skip) {
3596 n->set_req(j, in->in(1));
3597 }
3598 if (in->outcnt() == 0) {
3599 in->disconnect_inputs(NULL, this);
3600 }
3601 }
3602 }
3603 }
3604 }
3605
3606 //------------------------------final_graph_reshaping--------------------------
3607 // Final Graph Reshaping.
3608 //
3609 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3610 // and not commoned up and forced early. Must come after regular
3611 // optimizations to avoid GVN undoing the cloning. Clone constant
3612 // inputs to Loop Phis; these will be split by the allocator anyways.
3613 // Remove Opaque nodes.
3614 // (2) Move last-uses by commutative operations to the left input to encourage
3615 // Intel update-in-place two-address operations and better register usage
3616 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3617 // calls canonicalizing them back.
3618 // (3) Count the number of double-precision FP ops, single-precision FP ops
3619 // and call sites. On Intel, we can get correct rounding either by
3620 // forcing singles to memory (requires extra stores and loads after each
3621 // FP bytecode) or we can set a rounding mode bit (requires setting and
3622 // clearing the mode bit around call sites). The mode bit is only used
3623 // if the relative frequency of single FP ops to calls is low enough.
3624 // This is a key transform for SPEC mpeg_audio.
3625 // (4) Detect infinite loops; blobs of code reachable from above but not
3626 // below. Several of the Code_Gen algorithms fail on such code shapes,
3627 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3628 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3629 // Detection is by looking for IfNodes where only 1 projection is
3630 // reachable from below or CatchNodes missing some targets.
3631 // (5) Assert for insane oop offsets in debug mode.
3632
final_graph_reshaping()3633 bool Compile::final_graph_reshaping() {
3634 // an infinite loop may have been eliminated by the optimizer,
3635 // in which case the graph will be empty.
3636 if (root()->req() == 1) {
3637 record_method_not_compilable("trivial infinite loop");
3638 return true;
3639 }
3640
3641 // Expensive nodes have their control input set to prevent the GVN
3642 // from freely commoning them. There's no GVN beyond this point so
3643 // no need to keep the control input. We want the expensive nodes to
3644 // be freely moved to the least frequent code path by gcm.
3645 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3646 for (int i = 0; i < expensive_count(); i++) {
3647 _expensive_nodes->at(i)->set_req(0, NULL);
3648 }
3649
3650 Final_Reshape_Counts frc;
3651
3652 // Visit everybody reachable!
3653 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3654 Node_Stack nstack(live_nodes() >> 1);
3655 final_graph_reshaping_walk(nstack, root(), frc);
3656
3657 // Check for unreachable (from below) code (i.e., infinite loops).
3658 for( uint i = 0; i < frc._tests.size(); i++ ) {
3659 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3660 // Get number of CFG targets.
3661 // Note that PCTables include exception targets after calls.
3662 uint required_outcnt = n->required_outcnt();
3663 if (n->outcnt() != required_outcnt) {
3664 // Check for a few special cases. Rethrow Nodes never take the
3665 // 'fall-thru' path, so expected kids is 1 less.
3666 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3667 if (n->in(0)->in(0)->is_Call()) {
3668 CallNode *call = n->in(0)->in(0)->as_Call();
3669 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3670 required_outcnt--; // Rethrow always has 1 less kid
3671 } else if (call->req() > TypeFunc::Parms &&
3672 call->is_CallDynamicJava()) {
3673 // Check for null receiver. In such case, the optimizer has
3674 // detected that the virtual call will always result in a null
3675 // pointer exception. The fall-through projection of this CatchNode
3676 // will not be populated.
3677 Node *arg0 = call->in(TypeFunc::Parms);
3678 if (arg0->is_Type() &&
3679 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3680 required_outcnt--;
3681 }
3682 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3683 call->req() > TypeFunc::Parms+1 &&
3684 call->is_CallStaticJava()) {
3685 // Check for negative array length. In such case, the optimizer has
3686 // detected that the allocation attempt will always result in an
3687 // exception. There is no fall-through projection of this CatchNode .
3688 Node *arg1 = call->in(TypeFunc::Parms+1);
3689 if (arg1->is_Type() &&
3690 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3691 required_outcnt--;
3692 }
3693 }
3694 }
3695 }
3696 // Recheck with a better notion of 'required_outcnt'
3697 if (n->outcnt() != required_outcnt) {
3698 record_method_not_compilable("malformed control flow");
3699 return true; // Not all targets reachable!
3700 }
3701 }
3702 // Check that I actually visited all kids. Unreached kids
3703 // must be infinite loops.
3704 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3705 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3706 record_method_not_compilable("infinite loop");
3707 return true; // Found unvisited kid; must be unreach
3708 }
3709
3710 // Here so verification code in final_graph_reshaping_walk()
3711 // always see an OuterStripMinedLoopEnd
3712 if (n->is_OuterStripMinedLoopEnd()) {
3713 IfNode* init_iff = n->as_If();
3714 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3715 n->subsume_by(iff, this);
3716 }
3717 }
3718
3719 #ifdef IA32
3720 // If original bytecodes contained a mixture of floats and doubles
3721 // check if the optimizer has made it homogenous, item (3).
3722 if (UseSSE == 0 &&
3723 frc.get_float_count() > 32 &&
3724 frc.get_double_count() == 0 &&
3725 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3726 set_24_bit_selection_and_mode(false, true);
3727 }
3728 #endif // IA32
3729
3730 set_java_calls(frc.get_java_call_count());
3731 set_inner_loops(frc.get_inner_loop_count());
3732
3733 // No infinite loops, no reason to bail out.
3734 return false;
3735 }
3736
3737 //-----------------------------too_many_traps----------------------------------
3738 // Report if there are too many traps at the current method and bci.
3739 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
too_many_traps(ciMethod * method,int bci,Deoptimization::DeoptReason reason)3740 bool Compile::too_many_traps(ciMethod* method,
3741 int bci,
3742 Deoptimization::DeoptReason reason) {
3743 ciMethodData* md = method->method_data();
3744 if (md->is_empty()) {
3745 // Assume the trap has not occurred, or that it occurred only
3746 // because of a transient condition during start-up in the interpreter.
3747 return false;
3748 }
3749 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3750 if (md->has_trap_at(bci, m, reason) != 0) {
3751 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3752 // Also, if there are multiple reasons, or if there is no per-BCI record,
3753 // assume the worst.
3754 if (log())
3755 log()->elem("observe trap='%s' count='%d'",
3756 Deoptimization::trap_reason_name(reason),
3757 md->trap_count(reason));
3758 return true;
3759 } else {
3760 // Ignore method/bci and see if there have been too many globally.
3761 return too_many_traps(reason, md);
3762 }
3763 }
3764
3765 // Less-accurate variant which does not require a method and bci.
too_many_traps(Deoptimization::DeoptReason reason,ciMethodData * logmd)3766 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3767 ciMethodData* logmd) {
3768 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3769 // Too many traps globally.
3770 // Note that we use cumulative trap_count, not just md->trap_count.
3771 if (log()) {
3772 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3773 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3774 Deoptimization::trap_reason_name(reason),
3775 mcount, trap_count(reason));
3776 }
3777 return true;
3778 } else {
3779 // The coast is clear.
3780 return false;
3781 }
3782 }
3783
3784 //--------------------------too_many_recompiles--------------------------------
3785 // Report if there are too many recompiles at the current method and bci.
3786 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3787 // Is not eager to return true, since this will cause the compiler to use
3788 // Action_none for a trap point, to avoid too many recompilations.
too_many_recompiles(ciMethod * method,int bci,Deoptimization::DeoptReason reason)3789 bool Compile::too_many_recompiles(ciMethod* method,
3790 int bci,
3791 Deoptimization::DeoptReason reason) {
3792 ciMethodData* md = method->method_data();
3793 if (md->is_empty()) {
3794 // Assume the trap has not occurred, or that it occurred only
3795 // because of a transient condition during start-up in the interpreter.
3796 return false;
3797 }
3798 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3799 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3800 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3801 Deoptimization::DeoptReason per_bc_reason
3802 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3803 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3804 if ((per_bc_reason == Deoptimization::Reason_none
3805 || md->has_trap_at(bci, m, reason) != 0)
3806 // The trap frequency measure we care about is the recompile count:
3807 && md->trap_recompiled_at(bci, m)
3808 && md->overflow_recompile_count() >= bc_cutoff) {
3809 // Do not emit a trap here if it has already caused recompilations.
3810 // Also, if there are multiple reasons, or if there is no per-BCI record,
3811 // assume the worst.
3812 if (log())
3813 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3814 Deoptimization::trap_reason_name(reason),
3815 md->trap_count(reason),
3816 md->overflow_recompile_count());
3817 return true;
3818 } else if (trap_count(reason) != 0
3819 && decompile_count() >= m_cutoff) {
3820 // Too many recompiles globally, and we have seen this sort of trap.
3821 // Use cumulative decompile_count, not just md->decompile_count.
3822 if (log())
3823 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3824 Deoptimization::trap_reason_name(reason),
3825 md->trap_count(reason), trap_count(reason),
3826 md->decompile_count(), decompile_count());
3827 return true;
3828 } else {
3829 // The coast is clear.
3830 return false;
3831 }
3832 }
3833
3834 // Compute when not to trap. Used by matching trap based nodes and
3835 // NullCheck optimization.
set_allowed_deopt_reasons()3836 void Compile::set_allowed_deopt_reasons() {
3837 _allowed_reasons = 0;
3838 if (is_method_compilation()) {
3839 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3840 assert(rs < BitsPerInt, "recode bit map");
3841 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3842 _allowed_reasons |= nth_bit(rs);
3843 }
3844 }
3845 }
3846 }
3847
needs_clinit_barrier(ciMethod * method,ciMethod * accessing_method)3848 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
3849 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
3850 }
3851
needs_clinit_barrier(ciField * field,ciMethod * accessing_method)3852 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
3853 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
3854 }
3855
needs_clinit_barrier(ciInstanceKlass * holder,ciMethod * accessing_method)3856 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
3857 if (holder->is_initialized()) {
3858 return false;
3859 }
3860 if (holder->is_being_initialized()) {
3861 if (accessing_method->holder() == holder) {
3862 // Access inside a class. The barrier can be elided when access happens in <clinit>,
3863 // <init>, or a static method. In all those cases, there was an initialization
3864 // barrier on the holder klass passed.
3865 if (accessing_method->is_static_initializer() ||
3866 accessing_method->is_object_initializer() ||
3867 accessing_method->is_static()) {
3868 return false;
3869 }
3870 } else if (accessing_method->holder()->is_subclass_of(holder)) {
3871 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
3872 // In case of <init> or a static method, the barrier is on the subclass is not enough:
3873 // child class can become fully initialized while its parent class is still being initialized.
3874 if (accessing_method->is_static_initializer()) {
3875 return false;
3876 }
3877 }
3878 ciMethod* root = method(); // the root method of compilation
3879 if (root != accessing_method) {
3880 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
3881 }
3882 }
3883 return true;
3884 }
3885
3886 #ifndef PRODUCT
3887 //------------------------------verify_graph_edges---------------------------
3888 // Walk the Graph and verify that there is a one-to-one correspondence
3889 // between Use-Def edges and Def-Use edges in the graph.
verify_graph_edges(bool no_dead_code)3890 void Compile::verify_graph_edges(bool no_dead_code) {
3891 if (VerifyGraphEdges) {
3892 ResourceArea *area = Thread::current()->resource_area();
3893 Unique_Node_List visited(area);
3894 // Call recursive graph walk to check edges
3895 _root->verify_edges(visited);
3896 if (no_dead_code) {
3897 // Now make sure that no visited node is used by an unvisited node.
3898 bool dead_nodes = false;
3899 Unique_Node_List checked(area);
3900 while (visited.size() > 0) {
3901 Node* n = visited.pop();
3902 checked.push(n);
3903 for (uint i = 0; i < n->outcnt(); i++) {
3904 Node* use = n->raw_out(i);
3905 if (checked.member(use)) continue; // already checked
3906 if (visited.member(use)) continue; // already in the graph
3907 if (use->is_Con()) continue; // a dead ConNode is OK
3908 // At this point, we have found a dead node which is DU-reachable.
3909 if (!dead_nodes) {
3910 tty->print_cr("*** Dead nodes reachable via DU edges:");
3911 dead_nodes = true;
3912 }
3913 use->dump(2);
3914 tty->print_cr("---");
3915 checked.push(use); // No repeats; pretend it is now checked.
3916 }
3917 }
3918 assert(!dead_nodes, "using nodes must be reachable from root");
3919 }
3920 }
3921 }
3922 #endif
3923
3924 // The Compile object keeps track of failure reasons separately from the ciEnv.
3925 // This is required because there is not quite a 1-1 relation between the
3926 // ciEnv and its compilation task and the Compile object. Note that one
3927 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3928 // to backtrack and retry without subsuming loads. Other than this backtracking
3929 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3930 // by the logic in C2Compiler.
record_failure(const char * reason)3931 void Compile::record_failure(const char* reason) {
3932 if (log() != NULL) {
3933 log()->elem("failure reason='%s' phase='compile'", reason);
3934 }
3935 if (_failure_reason == NULL) {
3936 // Record the first failure reason.
3937 _failure_reason = reason;
3938 }
3939
3940 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3941 C->print_method(PHASE_FAILURE);
3942 }
3943 _root = NULL; // flush the graph, too
3944 }
3945
TracePhase(const char * name,elapsedTimer * accumulator)3946 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3947 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3948 _phase_name(name), _dolog(CITimeVerbose)
3949 {
3950 if (_dolog) {
3951 C = Compile::current();
3952 _log = C->log();
3953 } else {
3954 C = NULL;
3955 _log = NULL;
3956 }
3957 if (_log != NULL) {
3958 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3959 _log->stamp();
3960 _log->end_head();
3961 }
3962 }
3963
~TracePhase()3964 Compile::TracePhase::~TracePhase() {
3965
3966 C = Compile::current();
3967 if (_dolog) {
3968 _log = C->log();
3969 } else {
3970 _log = NULL;
3971 }
3972
3973 #ifdef ASSERT
3974 if (PrintIdealNodeCount) {
3975 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3976 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3977 }
3978
3979 if (VerifyIdealNodeCount) {
3980 Compile::current()->print_missing_nodes();
3981 }
3982 #endif
3983
3984 if (_log != NULL) {
3985 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3986 }
3987 }
3988
3989 //----------------------------static_subtype_check-----------------------------
3990 // Shortcut important common cases when superklass is exact:
3991 // (0) superklass is java.lang.Object (can occur in reflective code)
3992 // (1) subklass is already limited to a subtype of superklass => always ok
3993 // (2) subklass does not overlap with superklass => always fail
3994 // (3) superklass has NO subtypes and we can check with a simple compare.
static_subtype_check(ciKlass * superk,ciKlass * subk)3995 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
3996 if (StressReflectiveCode) {
3997 return SSC_full_test; // Let caller generate the general case.
3998 }
3999
4000 if (superk == env()->Object_klass()) {
4001 return SSC_always_true; // (0) this test cannot fail
4002 }
4003
4004 ciType* superelem = superk;
4005 if (superelem->is_array_klass())
4006 superelem = superelem->as_array_klass()->base_element_type();
4007
4008 if (!subk->is_interface()) { // cannot trust static interface types yet
4009 if (subk->is_subtype_of(superk)) {
4010 return SSC_always_true; // (1) false path dead; no dynamic test needed
4011 }
4012 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4013 !superk->is_subtype_of(subk)) {
4014 return SSC_always_false;
4015 }
4016 }
4017
4018 // If casting to an instance klass, it must have no subtypes
4019 if (superk->is_interface()) {
4020 // Cannot trust interfaces yet.
4021 // %%% S.B. superk->nof_implementors() == 1
4022 } else if (superelem->is_instance_klass()) {
4023 ciInstanceKlass* ik = superelem->as_instance_klass();
4024 if (!ik->has_subklass() && !ik->is_interface()) {
4025 if (!ik->is_final()) {
4026 // Add a dependency if there is a chance of a later subclass.
4027 dependencies()->assert_leaf_type(ik);
4028 }
4029 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4030 }
4031 } else {
4032 // A primitive array type has no subtypes.
4033 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4034 }
4035
4036 return SSC_full_test;
4037 }
4038
conv_I2X_index(PhaseGVN * phase,Node * idx,const TypeInt * sizetype,Node * ctrl)4039 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4040 #ifdef _LP64
4041 // The scaled index operand to AddP must be a clean 64-bit value.
4042 // Java allows a 32-bit int to be incremented to a negative
4043 // value, which appears in a 64-bit register as a large
4044 // positive number. Using that large positive number as an
4045 // operand in pointer arithmetic has bad consequences.
4046 // On the other hand, 32-bit overflow is rare, and the possibility
4047 // can often be excluded, if we annotate the ConvI2L node with
4048 // a type assertion that its value is known to be a small positive
4049 // number. (The prior range check has ensured this.)
4050 // This assertion is used by ConvI2LNode::Ideal.
4051 int index_max = max_jint - 1; // array size is max_jint, index is one less
4052 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4053 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4054 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4055 #endif
4056 return idx;
4057 }
4058
4059 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
constrained_convI2L(PhaseGVN * phase,Node * value,const TypeInt * itype,Node * ctrl)4060 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4061 if (ctrl != NULL) {
4062 // Express control dependency by a CastII node with a narrow type.
4063 value = new CastIINode(value, itype, false, true /* range check dependency */);
4064 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4065 // node from floating above the range check during loop optimizations. Otherwise, the
4066 // ConvI2L node may be eliminated independently of the range check, causing the data path
4067 // to become TOP while the control path is still there (although it's unreachable).
4068 value->set_req(0, ctrl);
4069 // Save CastII node to remove it after loop optimizations.
4070 phase->C->add_range_check_cast(value);
4071 value = phase->transform(value);
4072 }
4073 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4074 return phase->transform(new ConvI2LNode(value, ltype));
4075 }
4076
print_inlining_stream_free()4077 void Compile::print_inlining_stream_free() {
4078 if (_print_inlining_stream != NULL) {
4079 _print_inlining_stream->~stringStream();
4080 _print_inlining_stream = NULL;
4081 }
4082 }
4083
4084 // The message about the current inlining is accumulated in
4085 // _print_inlining_stream and transfered into the _print_inlining_list
4086 // once we know whether inlining succeeds or not. For regular
4087 // inlining, messages are appended to the buffer pointed by
4088 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4089 // a new buffer is added after _print_inlining_idx in the list. This
4090 // way we can update the inlining message for late inlining call site
4091 // when the inlining is attempted again.
print_inlining_init()4092 void Compile::print_inlining_init() {
4093 if (print_inlining() || print_intrinsics()) {
4094 // print_inlining_init is actually called several times.
4095 print_inlining_stream_free();
4096 _print_inlining_stream = new stringStream();
4097 // Watch out: The memory initialized by the constructor call PrintInliningBuffer()
4098 // will be copied into the only initial element. The default destructor of
4099 // PrintInliningBuffer will be called when leaving the scope here. If it
4100 // would destuct the enclosed stringStream _print_inlining_list[0]->_ss
4101 // would be destructed, too!
4102 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4103 }
4104 }
4105
print_inlining_reinit()4106 void Compile::print_inlining_reinit() {
4107 if (print_inlining() || print_intrinsics()) {
4108 print_inlining_stream_free();
4109 // Re allocate buffer when we change ResourceMark
4110 _print_inlining_stream = new stringStream();
4111 }
4112 }
4113
print_inlining_reset()4114 void Compile::print_inlining_reset() {
4115 _print_inlining_stream->reset();
4116 }
4117
print_inlining_commit()4118 void Compile::print_inlining_commit() {
4119 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4120 // Transfer the message from _print_inlining_stream to the current
4121 // _print_inlining_list buffer and clear _print_inlining_stream.
4122 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4123 print_inlining_reset();
4124 }
4125
print_inlining_push()4126 void Compile::print_inlining_push() {
4127 // Add new buffer to the _print_inlining_list at current position
4128 _print_inlining_idx++;
4129 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4130 }
4131
print_inlining_current()4132 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4133 return _print_inlining_list->at(_print_inlining_idx);
4134 }
4135
print_inlining_update(CallGenerator * cg)4136 void Compile::print_inlining_update(CallGenerator* cg) {
4137 if (print_inlining() || print_intrinsics()) {
4138 if (!cg->is_late_inline()) {
4139 if (print_inlining_current().cg() != NULL) {
4140 print_inlining_push();
4141 }
4142 print_inlining_commit();
4143 } else {
4144 if (print_inlining_current().cg() != cg &&
4145 (print_inlining_current().cg() != NULL ||
4146 print_inlining_current().ss()->size() != 0)) {
4147 print_inlining_push();
4148 }
4149 print_inlining_commit();
4150 print_inlining_current().set_cg(cg);
4151 }
4152 }
4153 }
4154
print_inlining_move_to(CallGenerator * cg)4155 void Compile::print_inlining_move_to(CallGenerator* cg) {
4156 // We resume inlining at a late inlining call site. Locate the
4157 // corresponding inlining buffer so that we can update it.
4158 if (print_inlining()) {
4159 for (int i = 0; i < _print_inlining_list->length(); i++) {
4160 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4161 _print_inlining_idx = i;
4162 return;
4163 }
4164 }
4165 ShouldNotReachHere();
4166 }
4167 }
4168
print_inlining_update_delayed(CallGenerator * cg)4169 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4170 if (print_inlining()) {
4171 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4172 assert(print_inlining_current().cg() == cg, "wrong entry");
4173 // replace message with new message
4174 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4175 print_inlining_commit();
4176 print_inlining_current().set_cg(cg);
4177 }
4178 }
4179
print_inlining_assert_ready()4180 void Compile::print_inlining_assert_ready() {
4181 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4182 }
4183
process_print_inlining()4184 void Compile::process_print_inlining() {
4185 bool do_print_inlining = print_inlining() || print_intrinsics();
4186 if (do_print_inlining || log() != NULL) {
4187 // Print inlining message for candidates that we couldn't inline
4188 // for lack of space
4189 for (int i = 0; i < _late_inlines.length(); i++) {
4190 CallGenerator* cg = _late_inlines.at(i);
4191 if (!cg->is_mh_late_inline()) {
4192 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4193 if (do_print_inlining) {
4194 cg->print_inlining_late(msg);
4195 }
4196 log_late_inline_failure(cg, msg);
4197 }
4198 }
4199 }
4200 if (do_print_inlining) {
4201 ResourceMark rm;
4202 stringStream ss;
4203 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
4204 for (int i = 0; i < _print_inlining_list->length(); i++) {
4205 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4206 _print_inlining_list->at(i).freeStream();
4207 }
4208 // Reset _print_inlining_list, it only contains destructed objects.
4209 // It is on the arena, so it will be freed when the arena is reset.
4210 _print_inlining_list = NULL;
4211 // _print_inlining_stream won't be used anymore, either.
4212 print_inlining_stream_free();
4213 size_t end = ss.size();
4214 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4215 strncpy(_print_inlining_output, ss.base(), end+1);
4216 _print_inlining_output[end] = 0;
4217 }
4218 }
4219
dump_print_inlining()4220 void Compile::dump_print_inlining() {
4221 if (_print_inlining_output != NULL) {
4222 tty->print_raw(_print_inlining_output);
4223 }
4224 }
4225
log_late_inline(CallGenerator * cg)4226 void Compile::log_late_inline(CallGenerator* cg) {
4227 if (log() != NULL) {
4228 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4229 cg->unique_id());
4230 JVMState* p = cg->call_node()->jvms();
4231 while (p != NULL) {
4232 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4233 p = p->caller();
4234 }
4235 log()->tail("late_inline");
4236 }
4237 }
4238
log_late_inline_failure(CallGenerator * cg,const char * msg)4239 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4240 log_late_inline(cg);
4241 if (log() != NULL) {
4242 log()->inline_fail(msg);
4243 }
4244 }
4245
log_inline_id(CallGenerator * cg)4246 void Compile::log_inline_id(CallGenerator* cg) {
4247 if (log() != NULL) {
4248 // The LogCompilation tool needs a unique way to identify late
4249 // inline call sites. This id must be unique for this call site in
4250 // this compilation. Try to have it unique across compilations as
4251 // well because it can be convenient when grepping through the log
4252 // file.
4253 // Distinguish OSR compilations from others in case CICountOSR is
4254 // on.
4255 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4256 cg->set_unique_id(id);
4257 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4258 }
4259 }
4260
log_inline_failure(const char * msg)4261 void Compile::log_inline_failure(const char* msg) {
4262 if (C->log() != NULL) {
4263 C->log()->inline_fail(msg);
4264 }
4265 }
4266
4267
4268 // Dump inlining replay data to the stream.
4269 // Don't change thread state and acquire any locks.
dump_inline_data(outputStream * out)4270 void Compile::dump_inline_data(outputStream* out) {
4271 InlineTree* inl_tree = ilt();
4272 if (inl_tree != NULL) {
4273 out->print(" inline %d", inl_tree->count());
4274 inl_tree->dump_replay_data(out);
4275 }
4276 }
4277
cmp_expensive_nodes(Node * n1,Node * n2)4278 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4279 if (n1->Opcode() < n2->Opcode()) return -1;
4280 else if (n1->Opcode() > n2->Opcode()) return 1;
4281
4282 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4283 for (uint i = 1; i < n1->req(); i++) {
4284 if (n1->in(i) < n2->in(i)) return -1;
4285 else if (n1->in(i) > n2->in(i)) return 1;
4286 }
4287
4288 return 0;
4289 }
4290
cmp_expensive_nodes(Node ** n1p,Node ** n2p)4291 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4292 Node* n1 = *n1p;
4293 Node* n2 = *n2p;
4294
4295 return cmp_expensive_nodes(n1, n2);
4296 }
4297
sort_expensive_nodes()4298 void Compile::sort_expensive_nodes() {
4299 if (!expensive_nodes_sorted()) {
4300 _expensive_nodes->sort(cmp_expensive_nodes);
4301 }
4302 }
4303
expensive_nodes_sorted() const4304 bool Compile::expensive_nodes_sorted() const {
4305 for (int i = 1; i < _expensive_nodes->length(); i++) {
4306 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4307 return false;
4308 }
4309 }
4310 return true;
4311 }
4312
should_optimize_expensive_nodes(PhaseIterGVN & igvn)4313 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4314 if (_expensive_nodes->length() == 0) {
4315 return false;
4316 }
4317
4318 assert(OptimizeExpensiveOps, "optimization off?");
4319
4320 // Take this opportunity to remove dead nodes from the list
4321 int j = 0;
4322 for (int i = 0; i < _expensive_nodes->length(); i++) {
4323 Node* n = _expensive_nodes->at(i);
4324 if (!n->is_unreachable(igvn)) {
4325 assert(n->is_expensive(), "should be expensive");
4326 _expensive_nodes->at_put(j, n);
4327 j++;
4328 }
4329 }
4330 _expensive_nodes->trunc_to(j);
4331
4332 // Then sort the list so that similar nodes are next to each other
4333 // and check for at least two nodes of identical kind with same data
4334 // inputs.
4335 sort_expensive_nodes();
4336
4337 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4338 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4339 return true;
4340 }
4341 }
4342
4343 return false;
4344 }
4345
cleanup_expensive_nodes(PhaseIterGVN & igvn)4346 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4347 if (_expensive_nodes->length() == 0) {
4348 return;
4349 }
4350
4351 assert(OptimizeExpensiveOps, "optimization off?");
4352
4353 // Sort to bring similar nodes next to each other and clear the
4354 // control input of nodes for which there's only a single copy.
4355 sort_expensive_nodes();
4356
4357 int j = 0;
4358 int identical = 0;
4359 int i = 0;
4360 bool modified = false;
4361 for (; i < _expensive_nodes->length()-1; i++) {
4362 assert(j <= i, "can't write beyond current index");
4363 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4364 identical++;
4365 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4366 continue;
4367 }
4368 if (identical > 0) {
4369 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4370 identical = 0;
4371 } else {
4372 Node* n = _expensive_nodes->at(i);
4373 igvn.replace_input_of(n, 0, NULL);
4374 igvn.hash_insert(n);
4375 modified = true;
4376 }
4377 }
4378 if (identical > 0) {
4379 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4380 } else if (_expensive_nodes->length() >= 1) {
4381 Node* n = _expensive_nodes->at(i);
4382 igvn.replace_input_of(n, 0, NULL);
4383 igvn.hash_insert(n);
4384 modified = true;
4385 }
4386 _expensive_nodes->trunc_to(j);
4387 if (modified) {
4388 igvn.optimize();
4389 }
4390 }
4391
add_expensive_node(Node * n)4392 void Compile::add_expensive_node(Node * n) {
4393 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4394 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4395 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4396 if (OptimizeExpensiveOps) {
4397 _expensive_nodes->append(n);
4398 } else {
4399 // Clear control input and let IGVN optimize expensive nodes if
4400 // OptimizeExpensiveOps is off.
4401 n->set_req(0, NULL);
4402 }
4403 }
4404
4405 /**
4406 * Remove the speculative part of types and clean up the graph
4407 */
remove_speculative_types(PhaseIterGVN & igvn)4408 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4409 if (UseTypeSpeculation) {
4410 Unique_Node_List worklist;
4411 worklist.push(root());
4412 int modified = 0;
4413 // Go over all type nodes that carry a speculative type, drop the
4414 // speculative part of the type and enqueue the node for an igvn
4415 // which may optimize it out.
4416 for (uint next = 0; next < worklist.size(); ++next) {
4417 Node *n = worklist.at(next);
4418 if (n->is_Type()) {
4419 TypeNode* tn = n->as_Type();
4420 const Type* t = tn->type();
4421 const Type* t_no_spec = t->remove_speculative();
4422 if (t_no_spec != t) {
4423 bool in_hash = igvn.hash_delete(n);
4424 assert(in_hash, "node should be in igvn hash table");
4425 tn->set_type(t_no_spec);
4426 igvn.hash_insert(n);
4427 igvn._worklist.push(n); // give it a chance to go away
4428 modified++;
4429 }
4430 }
4431 uint max = n->len();
4432 for( uint i = 0; i < max; ++i ) {
4433 Node *m = n->in(i);
4434 if (not_a_node(m)) continue;
4435 worklist.push(m);
4436 }
4437 }
4438 // Drop the speculative part of all types in the igvn's type table
4439 igvn.remove_speculative_types();
4440 if (modified > 0) {
4441 igvn.optimize();
4442 }
4443 #ifdef ASSERT
4444 // Verify that after the IGVN is over no speculative type has resurfaced
4445 worklist.clear();
4446 worklist.push(root());
4447 for (uint next = 0; next < worklist.size(); ++next) {
4448 Node *n = worklist.at(next);
4449 const Type* t = igvn.type_or_null(n);
4450 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4451 if (n->is_Type()) {
4452 t = n->as_Type()->type();
4453 assert(t == t->remove_speculative(), "no more speculative types");
4454 }
4455 uint max = n->len();
4456 for( uint i = 0; i < max; ++i ) {
4457 Node *m = n->in(i);
4458 if (not_a_node(m)) continue;
4459 worklist.push(m);
4460 }
4461 }
4462 igvn.check_no_speculative_types();
4463 #endif
4464 }
4465 }
4466
4467 // Auxiliary method to support randomized stressing/fuzzing.
4468 //
4469 // This method can be called the arbitrary number of times, with current count
4470 // as the argument. The logic allows selecting a single candidate from the
4471 // running list of candidates as follows:
4472 // int count = 0;
4473 // Cand* selected = null;
4474 // while(cand = cand->next()) {
4475 // if (randomized_select(++count)) {
4476 // selected = cand;
4477 // }
4478 // }
4479 //
4480 // Including count equalizes the chances any candidate is "selected".
4481 // This is useful when we don't have the complete list of candidates to choose
4482 // from uniformly. In this case, we need to adjust the randomicity of the
4483 // selection, or else we will end up biasing the selection towards the latter
4484 // candidates.
4485 //
4486 // Quick back-envelope calculation shows that for the list of n candidates
4487 // the equal probability for the candidate to persist as "best" can be
4488 // achieved by replacing it with "next" k-th candidate with the probability
4489 // of 1/k. It can be easily shown that by the end of the run, the
4490 // probability for any candidate is converged to 1/n, thus giving the
4491 // uniform distribution among all the candidates.
4492 //
4493 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4494 #define RANDOMIZED_DOMAIN_POW 29
4495 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4496 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
randomized_select(int count)4497 bool Compile::randomized_select(int count) {
4498 assert(count > 0, "only positive");
4499 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4500 }
4501
clone_map()4502 CloneMap& Compile::clone_map() { return _clone_map; }
set_clone_map(Dict * d)4503 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4504
dump() const4505 void NodeCloneInfo::dump() const {
4506 tty->print(" {%d:%d} ", idx(), gen());
4507 }
4508
clone(Node * old,Node * nnn,int gen)4509 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4510 uint64_t val = value(old->_idx);
4511 NodeCloneInfo cio(val);
4512 assert(val != 0, "old node should be in the map");
4513 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4514 insert(nnn->_idx, cin.get());
4515 #ifndef PRODUCT
4516 if (is_debug()) {
4517 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4518 }
4519 #endif
4520 }
4521
verify_insert_and_clone(Node * old,Node * nnn,int gen)4522 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4523 NodeCloneInfo cio(value(old->_idx));
4524 if (cio.get() == 0) {
4525 cio.set(old->_idx, 0);
4526 insert(old->_idx, cio.get());
4527 #ifndef PRODUCT
4528 if (is_debug()) {
4529 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4530 }
4531 #endif
4532 }
4533 clone(old, nnn, gen);
4534 }
4535
max_gen() const4536 int CloneMap::max_gen() const {
4537 int g = 0;
4538 DictI di(_dict);
4539 for(; di.test(); ++di) {
4540 int t = gen(di._key);
4541 if (g < t) {
4542 g = t;
4543 #ifndef PRODUCT
4544 if (is_debug()) {
4545 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4546 }
4547 #endif
4548 }
4549 }
4550 return g;
4551 }
4552
dump(node_idx_t key) const4553 void CloneMap::dump(node_idx_t key) const {
4554 uint64_t val = value(key);
4555 if (val != 0) {
4556 NodeCloneInfo ni(val);
4557 ni.dump();
4558 }
4559 }
4560
4561 // Move Allocate nodes to the start of the list
sort_macro_nodes()4562 void Compile::sort_macro_nodes() {
4563 int count = macro_count();
4564 int allocates = 0;
4565 for (int i = 0; i < count; i++) {
4566 Node* n = macro_node(i);
4567 if (n->is_Allocate()) {
4568 if (i != allocates) {
4569 Node* tmp = macro_node(allocates);
4570 _macro_nodes->at_put(allocates, n);
4571 _macro_nodes->at_put(i, tmp);
4572 }
4573 allocates++;
4574 }
4575 }
4576 }
4577
print_method(CompilerPhaseType cpt,int level,int idx)4578 void Compile::print_method(CompilerPhaseType cpt, int level, int idx) {
4579 EventCompilerPhase event;
4580 if (event.should_commit()) {
4581 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
4582 }
4583
4584 #ifndef PRODUCT
4585 if (should_print(level)) {
4586 char output[1024];
4587 if (idx != 0) {
4588 jio_snprintf(output, sizeof(output), "%s:%d", CompilerPhaseTypeHelper::to_string(cpt), idx);
4589 } else {
4590 jio_snprintf(output, sizeof(output), "%s", CompilerPhaseTypeHelper::to_string(cpt));
4591 }
4592 _printer->print_method(output, level);
4593 }
4594 #endif
4595 C->_latest_stage_start_counter.stamp();
4596 }
4597
end_method(int level)4598 void Compile::end_method(int level) {
4599 EventCompilerPhase event;
4600 if (event.should_commit()) {
4601 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, level);
4602 }
4603
4604 #ifndef PRODUCT
4605 if (_printer && _printer->should_print(level)) {
4606 _printer->end_method();
4607 }
4608 #endif
4609 }
4610
4611
4612 #ifndef PRODUCT
4613 IdealGraphPrinter* Compile::_debug_file_printer = NULL;
4614 IdealGraphPrinter* Compile::_debug_network_printer = NULL;
4615
4616 // Called from debugger. Prints method to the default file with the default phase name.
4617 // This works regardless of any Ideal Graph Visualizer flags set or not.
igv_print()4618 void igv_print() {
4619 Compile::current()->igv_print_method_to_file();
4620 }
4621
4622 // Same as igv_print() above but with a specified phase name.
igv_print(const char * phase_name)4623 void igv_print(const char* phase_name) {
4624 Compile::current()->igv_print_method_to_file(phase_name);
4625 }
4626
4627 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
4628 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
4629 // This works regardless of any Ideal Graph Visualizer flags set or not.
igv_print(bool network)4630 void igv_print(bool network) {
4631 if (network) {
4632 Compile::current()->igv_print_method_to_network();
4633 } else {
4634 Compile::current()->igv_print_method_to_file();
4635 }
4636 }
4637
4638 // Same as igv_print(bool network) above but with a specified phase name.
igv_print(bool network,const char * phase_name)4639 void igv_print(bool network, const char* phase_name) {
4640 if (network) {
4641 Compile::current()->igv_print_method_to_network(phase_name);
4642 } else {
4643 Compile::current()->igv_print_method_to_file(phase_name);
4644 }
4645 }
4646
4647 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
igv_print_default()4648 void igv_print_default() {
4649 Compile::current()->print_method(PHASE_DEBUG, 0, 0);
4650 }
4651
4652 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
4653 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
4654 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
igv_append()4655 void igv_append() {
4656 Compile::current()->igv_print_method_to_file("Debug", true);
4657 }
4658
4659 // Same as igv_append() above but with a specified phase name.
igv_append(const char * phase_name)4660 void igv_append(const char* phase_name) {
4661 Compile::current()->igv_print_method_to_file(phase_name, true);
4662 }
4663
igv_print_method_to_file(const char * phase_name,bool append)4664 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
4665 const char* file_name = "custom_debug.xml";
4666 if (_debug_file_printer == NULL) {
4667 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
4668 } else {
4669 _debug_file_printer->update_compiled_method(C->method());
4670 }
4671 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
4672 _debug_file_printer->print_method(phase_name, 0);
4673 }
4674
igv_print_method_to_network(const char * phase_name)4675 void Compile::igv_print_method_to_network(const char* phase_name) {
4676 if (_debug_network_printer == NULL) {
4677 _debug_network_printer = new IdealGraphPrinter(C);
4678 } else {
4679 _debug_network_printer->update_compiled_method(C->method());
4680 }
4681 tty->print_cr("Method printed over network stream to IGV");
4682 _debug_network_printer->print_method(phase_name, 0);
4683 }
4684 #endif
4685
4686