Mon, 21 Mar 2011 11:28:14 -0700
7008866: Missing loop predicate for loop with multiple entries
Summary: Add predicates when loop head bytecode is parsed instead of when back branch bytecode is parsed.
Reviewed-by: never
1 /*
2 * Copyright (c) 2000, 2011, 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 */
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "ci/ciCPCache.hpp"
28 #include "ci/ciMethodHandle.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/connode.hpp"
36 #include "opto/parse.hpp"
37 #include "opto/rootnode.hpp"
38 #include "opto/runtime.hpp"
39 #include "opto/subnode.hpp"
41 CallGenerator::CallGenerator(ciMethod* method) {
42 _method = method;
43 }
45 // Utility function.
46 const TypeFunc* CallGenerator::tf() const {
47 return TypeFunc::make(method());
48 }
50 //-----------------------------ParseGenerator---------------------------------
51 // Internal class which handles all direct bytecode traversal.
52 class ParseGenerator : public InlineCallGenerator {
53 private:
54 bool _is_osr;
55 float _expected_uses;
57 public:
58 ParseGenerator(ciMethod* method, float expected_uses, bool is_osr = false)
59 : InlineCallGenerator(method)
60 {
61 _is_osr = is_osr;
62 _expected_uses = expected_uses;
63 assert(can_parse(method, is_osr), "parse must be possible");
64 }
66 // Can we build either an OSR or a regular parser for this method?
67 static bool can_parse(ciMethod* method, int is_osr = false);
69 virtual bool is_parse() const { return true; }
70 virtual JVMState* generate(JVMState* jvms);
71 int is_osr() { return _is_osr; }
73 };
75 JVMState* ParseGenerator::generate(JVMState* jvms) {
76 Compile* C = Compile::current();
78 if (is_osr()) {
79 // The JVMS for a OSR has a single argument (see its TypeFunc).
80 assert(jvms->depth() == 1, "no inline OSR");
81 }
83 if (C->failing()) {
84 return NULL; // bailing out of the compile; do not try to parse
85 }
87 Parse parser(jvms, method(), _expected_uses);
88 // Grab signature for matching/allocation
89 #ifdef ASSERT
90 if (parser.tf() != (parser.depth() == 1 ? C->tf() : tf())) {
91 MutexLockerEx ml(Compile_lock, Mutex::_no_safepoint_check_flag);
92 assert(C->env()->system_dictionary_modification_counter_changed(),
93 "Must invalidate if TypeFuncs differ");
94 }
95 #endif
97 GraphKit& exits = parser.exits();
99 if (C->failing()) {
100 while (exits.pop_exception_state() != NULL) ;
101 return NULL;
102 }
104 assert(exits.jvms()->same_calls_as(jvms), "sanity");
106 // Simply return the exit state of the parser,
107 // augmented by any exceptional states.
108 return exits.transfer_exceptions_into_jvms();
109 }
111 //---------------------------DirectCallGenerator------------------------------
112 // Internal class which handles all out-of-line calls w/o receiver type checks.
113 class DirectCallGenerator : public CallGenerator {
114 private:
115 CallStaticJavaNode* _call_node;
116 // Force separate memory and I/O projections for the exceptional
117 // paths to facilitate late inlinig.
118 bool _separate_io_proj;
120 public:
121 DirectCallGenerator(ciMethod* method, bool separate_io_proj)
122 : CallGenerator(method),
123 _separate_io_proj(separate_io_proj)
124 {
125 }
126 virtual JVMState* generate(JVMState* jvms);
128 CallStaticJavaNode* call_node() const { return _call_node; }
129 };
131 JVMState* DirectCallGenerator::generate(JVMState* jvms) {
132 GraphKit kit(jvms);
133 bool is_static = method()->is_static();
134 address target = is_static ? SharedRuntime::get_resolve_static_call_stub()
135 : SharedRuntime::get_resolve_opt_virtual_call_stub();
137 if (kit.C->log() != NULL) {
138 kit.C->log()->elem("direct_call bci='%d'", jvms->bci());
139 }
141 CallStaticJavaNode *call = new (kit.C, tf()->domain()->cnt()) CallStaticJavaNode(tf(), target, method(), kit.bci());
142 if (!is_static) {
143 // Make an explicit receiver null_check as part of this call.
144 // Since we share a map with the caller, his JVMS gets adjusted.
145 kit.null_check_receiver(method());
146 if (kit.stopped()) {
147 // And dump it back to the caller, decorated with any exceptions:
148 return kit.transfer_exceptions_into_jvms();
149 }
150 // Mark the call node as virtual, sort of:
151 call->set_optimized_virtual(true);
152 if (method()->is_method_handle_invoke()) {
153 call->set_method_handle_invoke(true);
154 kit.C->set_has_method_handle_invokes(true);
155 }
156 }
157 kit.set_arguments_for_java_call(call);
158 kit.set_edges_for_java_call(call, false, _separate_io_proj);
159 Node* ret = kit.set_results_for_java_call(call, _separate_io_proj);
160 kit.push_node(method()->return_type()->basic_type(), ret);
161 _call_node = call; // Save the call node in case we need it later
162 return kit.transfer_exceptions_into_jvms();
163 }
165 //---------------------------DynamicCallGenerator-----------------------------
166 // Internal class which handles all out-of-line invokedynamic calls.
167 class DynamicCallGenerator : public CallGenerator {
168 public:
169 DynamicCallGenerator(ciMethod* method)
170 : CallGenerator(method)
171 {
172 }
173 virtual JVMState* generate(JVMState* jvms);
174 };
176 JVMState* DynamicCallGenerator::generate(JVMState* jvms) {
177 GraphKit kit(jvms);
179 if (kit.C->log() != NULL) {
180 kit.C->log()->elem("dynamic_call bci='%d'", jvms->bci());
181 }
183 // Get the constant pool cache from the caller class.
184 ciMethod* caller_method = jvms->method();
185 ciBytecodeStream str(caller_method);
186 str.force_bci(jvms->bci()); // Set the stream to the invokedynamic bci.
187 assert(str.cur_bc() == Bytecodes::_invokedynamic, "wrong place to issue a dynamic call!");
188 ciCPCache* cpcache = str.get_cpcache();
190 // Get the offset of the CallSite from the constant pool cache
191 // pointer.
192 int index = str.get_method_index();
193 size_t call_site_offset = cpcache->get_f1_offset(index);
195 // Load the CallSite object from the constant pool cache.
196 const TypeOopPtr* cpcache_ptr = TypeOopPtr::make_from_constant(cpcache);
197 Node* cpcache_adr = kit.makecon(cpcache_ptr);
198 Node* call_site_adr = kit.basic_plus_adr(cpcache_adr, cpcache_adr, call_site_offset);
199 Node* call_site = kit.make_load(kit.control(), call_site_adr, TypeInstPtr::BOTTOM, T_OBJECT, Compile::AliasIdxRaw);
201 // Load the target MethodHandle from the CallSite object.
202 Node* target_mh_adr = kit.basic_plus_adr(call_site, call_site, java_lang_invoke_CallSite::target_offset_in_bytes());
203 Node* target_mh = kit.make_load(kit.control(), target_mh_adr, TypeInstPtr::BOTTOM, T_OBJECT);
205 address resolve_stub = SharedRuntime::get_resolve_opt_virtual_call_stub();
207 CallStaticJavaNode *call = new (kit.C, tf()->domain()->cnt()) CallStaticJavaNode(tf(), resolve_stub, method(), kit.bci());
208 // invokedynamic is treated as an optimized invokevirtual.
209 call->set_optimized_virtual(true);
210 // Take extra care (in the presence of argument motion) not to trash the SP:
211 call->set_method_handle_invoke(true);
212 kit.C->set_has_method_handle_invokes(true);
214 // Pass the target MethodHandle as first argument and shift the
215 // other arguments.
216 call->init_req(0 + TypeFunc::Parms, target_mh);
217 uint nargs = call->method()->arg_size();
218 for (uint i = 1; i < nargs; i++) {
219 Node* arg = kit.argument(i - 1);
220 call->init_req(i + TypeFunc::Parms, arg);
221 }
223 kit.set_edges_for_java_call(call);
224 Node* ret = kit.set_results_for_java_call(call);
225 kit.push_node(method()->return_type()->basic_type(), ret);
226 return kit.transfer_exceptions_into_jvms();
227 }
229 //--------------------------VirtualCallGenerator------------------------------
230 // Internal class which handles all out-of-line calls checking receiver type.
231 class VirtualCallGenerator : public CallGenerator {
232 private:
233 int _vtable_index;
234 public:
235 VirtualCallGenerator(ciMethod* method, int vtable_index)
236 : CallGenerator(method), _vtable_index(vtable_index)
237 {
238 assert(vtable_index == methodOopDesc::invalid_vtable_index ||
239 vtable_index >= 0, "either invalid or usable");
240 }
241 virtual bool is_virtual() const { return true; }
242 virtual JVMState* generate(JVMState* jvms);
243 };
245 JVMState* VirtualCallGenerator::generate(JVMState* jvms) {
246 GraphKit kit(jvms);
247 Node* receiver = kit.argument(0);
249 if (kit.C->log() != NULL) {
250 kit.C->log()->elem("virtual_call bci='%d'", jvms->bci());
251 }
253 // If the receiver is a constant null, do not torture the system
254 // by attempting to call through it. The compile will proceed
255 // correctly, but may bail out in final_graph_reshaping, because
256 // the call instruction will have a seemingly deficient out-count.
257 // (The bailout says something misleading about an "infinite loop".)
258 if (kit.gvn().type(receiver)->higher_equal(TypePtr::NULL_PTR)) {
259 kit.inc_sp(method()->arg_size()); // restore arguments
260 kit.uncommon_trap(Deoptimization::Reason_null_check,
261 Deoptimization::Action_none,
262 NULL, "null receiver");
263 return kit.transfer_exceptions_into_jvms();
264 }
266 // Ideally we would unconditionally do a null check here and let it
267 // be converted to an implicit check based on profile information.
268 // However currently the conversion to implicit null checks in
269 // Block::implicit_null_check() only looks for loads and stores, not calls.
270 ciMethod *caller = kit.method();
271 ciMethodData *caller_md = (caller == NULL) ? NULL : caller->method_data();
272 if (!UseInlineCaches || !ImplicitNullChecks ||
273 ((ImplicitNullCheckThreshold > 0) && caller_md &&
274 (caller_md->trap_count(Deoptimization::Reason_null_check)
275 >= (uint)ImplicitNullCheckThreshold))) {
276 // Make an explicit receiver null_check as part of this call.
277 // Since we share a map with the caller, his JVMS gets adjusted.
278 receiver = kit.null_check_receiver(method());
279 if (kit.stopped()) {
280 // And dump it back to the caller, decorated with any exceptions:
281 return kit.transfer_exceptions_into_jvms();
282 }
283 }
285 assert(!method()->is_static(), "virtual call must not be to static");
286 assert(!method()->is_final(), "virtual call should not be to final");
287 assert(!method()->is_private(), "virtual call should not be to private");
288 assert(_vtable_index == methodOopDesc::invalid_vtable_index || !UseInlineCaches,
289 "no vtable calls if +UseInlineCaches ");
290 address target = SharedRuntime::get_resolve_virtual_call_stub();
291 // Normal inline cache used for call
292 CallDynamicJavaNode *call = new (kit.C, tf()->domain()->cnt()) CallDynamicJavaNode(tf(), target, method(), _vtable_index, kit.bci());
293 kit.set_arguments_for_java_call(call);
294 kit.set_edges_for_java_call(call);
295 Node* ret = kit.set_results_for_java_call(call);
296 kit.push_node(method()->return_type()->basic_type(), ret);
298 // Represent the effect of an implicit receiver null_check
299 // as part of this call. Since we share a map with the caller,
300 // his JVMS gets adjusted.
301 kit.cast_not_null(receiver);
302 return kit.transfer_exceptions_into_jvms();
303 }
305 bool ParseGenerator::can_parse(ciMethod* m, int entry_bci) {
306 // Certain methods cannot be parsed at all:
307 if (!m->can_be_compiled()) return false;
308 if (!m->has_balanced_monitors()) return false;
309 if (m->get_flow_analysis()->failing()) return false;
311 // (Methods may bail out for other reasons, after the parser is run.
312 // We try to avoid this, but if forced, we must return (Node*)NULL.
313 // The user of the CallGenerator must check for this condition.)
314 return true;
315 }
317 CallGenerator* CallGenerator::for_inline(ciMethod* m, float expected_uses) {
318 if (!ParseGenerator::can_parse(m)) return NULL;
319 return new ParseGenerator(m, expected_uses);
320 }
322 // As a special case, the JVMS passed to this CallGenerator is
323 // for the method execution already in progress, not just the JVMS
324 // of the caller. Thus, this CallGenerator cannot be mixed with others!
325 CallGenerator* CallGenerator::for_osr(ciMethod* m, int osr_bci) {
326 if (!ParseGenerator::can_parse(m, true)) return NULL;
327 float past_uses = m->interpreter_invocation_count();
328 float expected_uses = past_uses;
329 return new ParseGenerator(m, expected_uses, true);
330 }
332 CallGenerator* CallGenerator::for_direct_call(ciMethod* m, bool separate_io_proj) {
333 assert(!m->is_abstract(), "for_direct_call mismatch");
334 return new DirectCallGenerator(m, separate_io_proj);
335 }
337 CallGenerator* CallGenerator::for_dynamic_call(ciMethod* m) {
338 assert(m->is_method_handle_invoke(), "for_dynamic_call mismatch");
339 return new DynamicCallGenerator(m);
340 }
342 CallGenerator* CallGenerator::for_virtual_call(ciMethod* m, int vtable_index) {
343 assert(!m->is_static(), "for_virtual_call mismatch");
344 assert(!m->is_method_handle_invoke(), "should be a direct call");
345 return new VirtualCallGenerator(m, vtable_index);
346 }
348 // Allow inlining decisions to be delayed
349 class LateInlineCallGenerator : public DirectCallGenerator {
350 CallGenerator* _inline_cg;
352 public:
353 LateInlineCallGenerator(ciMethod* method, CallGenerator* inline_cg) :
354 DirectCallGenerator(method, true), _inline_cg(inline_cg) {}
356 virtual bool is_late_inline() const { return true; }
358 // Convert the CallStaticJava into an inline
359 virtual void do_late_inline();
361 JVMState* generate(JVMState* jvms) {
362 // Record that this call site should be revisited once the main
363 // parse is finished.
364 Compile::current()->add_late_inline(this);
366 // Emit the CallStaticJava and request separate projections so
367 // that the late inlining logic can distinguish between fall
368 // through and exceptional uses of the memory and io projections
369 // as is done for allocations and macro expansion.
370 return DirectCallGenerator::generate(jvms);
371 }
373 };
376 void LateInlineCallGenerator::do_late_inline() {
377 // Can't inline it
378 if (call_node() == NULL || call_node()->outcnt() == 0 ||
379 call_node()->in(0) == NULL || call_node()->in(0)->is_top())
380 return;
382 CallStaticJavaNode* call = call_node();
384 // Make a clone of the JVMState that appropriate to use for driving a parse
385 Compile* C = Compile::current();
386 JVMState* jvms = call->jvms()->clone_shallow(C);
387 uint size = call->req();
388 SafePointNode* map = new (C, size) SafePointNode(size, jvms);
389 for (uint i1 = 0; i1 < size; i1++) {
390 map->init_req(i1, call->in(i1));
391 }
393 // Make sure the state is a MergeMem for parsing.
394 if (!map->in(TypeFunc::Memory)->is_MergeMem()) {
395 map->set_req(TypeFunc::Memory, MergeMemNode::make(C, map->in(TypeFunc::Memory)));
396 }
398 // Make enough space for the expression stack and transfer the incoming arguments
399 int nargs = method()->arg_size();
400 jvms->set_map(map);
401 map->ensure_stack(jvms, jvms->method()->max_stack());
402 if (nargs > 0) {
403 for (int i1 = 0; i1 < nargs; i1++) {
404 map->set_req(i1 + jvms->argoff(), call->in(TypeFunc::Parms + i1));
405 }
406 }
408 CompileLog* log = C->log();
409 if (log != NULL) {
410 log->head("late_inline method='%d'", log->identify(method()));
411 JVMState* p = jvms;
412 while (p != NULL) {
413 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
414 p = p->caller();
415 }
416 log->tail("late_inline");
417 }
419 // Setup default node notes to be picked up by the inlining
420 Node_Notes* old_nn = C->default_node_notes();
421 if (old_nn != NULL) {
422 Node_Notes* entry_nn = old_nn->clone(C);
423 entry_nn->set_jvms(jvms);
424 C->set_default_node_notes(entry_nn);
425 }
427 // Now perform the inling using the synthesized JVMState
428 JVMState* new_jvms = _inline_cg->generate(jvms);
429 if (new_jvms == NULL) return; // no change
430 if (C->failing()) return;
432 // Capture any exceptional control flow
433 GraphKit kit(new_jvms);
435 // Find the result object
436 Node* result = C->top();
437 int result_size = method()->return_type()->size();
438 if (result_size != 0 && !kit.stopped()) {
439 result = (result_size == 1) ? kit.pop() : kit.pop_pair();
440 }
442 kit.replace_call(call, result);
443 }
446 CallGenerator* CallGenerator::for_late_inline(ciMethod* method, CallGenerator* inline_cg) {
447 return new LateInlineCallGenerator(method, inline_cg);
448 }
451 //---------------------------WarmCallGenerator--------------------------------
452 // Internal class which handles initial deferral of inlining decisions.
453 class WarmCallGenerator : public CallGenerator {
454 WarmCallInfo* _call_info;
455 CallGenerator* _if_cold;
456 CallGenerator* _if_hot;
457 bool _is_virtual; // caches virtuality of if_cold
458 bool _is_inline; // caches inline-ness of if_hot
460 public:
461 WarmCallGenerator(WarmCallInfo* ci,
462 CallGenerator* if_cold,
463 CallGenerator* if_hot)
464 : CallGenerator(if_cold->method())
465 {
466 assert(method() == if_hot->method(), "consistent choices");
467 _call_info = ci;
468 _if_cold = if_cold;
469 _if_hot = if_hot;
470 _is_virtual = if_cold->is_virtual();
471 _is_inline = if_hot->is_inline();
472 }
474 virtual bool is_inline() const { return _is_inline; }
475 virtual bool is_virtual() const { return _is_virtual; }
476 virtual bool is_deferred() const { return true; }
478 virtual JVMState* generate(JVMState* jvms);
479 };
482 CallGenerator* CallGenerator::for_warm_call(WarmCallInfo* ci,
483 CallGenerator* if_cold,
484 CallGenerator* if_hot) {
485 return new WarmCallGenerator(ci, if_cold, if_hot);
486 }
488 JVMState* WarmCallGenerator::generate(JVMState* jvms) {
489 Compile* C = Compile::current();
490 if (C->log() != NULL) {
491 C->log()->elem("warm_call bci='%d'", jvms->bci());
492 }
493 jvms = _if_cold->generate(jvms);
494 if (jvms != NULL) {
495 Node* m = jvms->map()->control();
496 if (m->is_CatchProj()) m = m->in(0); else m = C->top();
497 if (m->is_Catch()) m = m->in(0); else m = C->top();
498 if (m->is_Proj()) m = m->in(0); else m = C->top();
499 if (m->is_CallJava()) {
500 _call_info->set_call(m->as_Call());
501 _call_info->set_hot_cg(_if_hot);
502 #ifndef PRODUCT
503 if (PrintOpto || PrintOptoInlining) {
504 tty->print_cr("Queueing for warm inlining at bci %d:", jvms->bci());
505 tty->print("WCI: ");
506 _call_info->print();
507 }
508 #endif
509 _call_info->set_heat(_call_info->compute_heat());
510 C->set_warm_calls(_call_info->insert_into(C->warm_calls()));
511 }
512 }
513 return jvms;
514 }
516 void WarmCallInfo::make_hot() {
517 Unimplemented();
518 }
520 void WarmCallInfo::make_cold() {
521 // No action: Just dequeue.
522 }
525 //------------------------PredictedCallGenerator------------------------------
526 // Internal class which handles all out-of-line calls checking receiver type.
527 class PredictedCallGenerator : public CallGenerator {
528 ciKlass* _predicted_receiver;
529 CallGenerator* _if_missed;
530 CallGenerator* _if_hit;
531 float _hit_prob;
533 public:
534 PredictedCallGenerator(ciKlass* predicted_receiver,
535 CallGenerator* if_missed,
536 CallGenerator* if_hit, float hit_prob)
537 : CallGenerator(if_missed->method())
538 {
539 // The call profile data may predict the hit_prob as extreme as 0 or 1.
540 // Remove the extremes values from the range.
541 if (hit_prob > PROB_MAX) hit_prob = PROB_MAX;
542 if (hit_prob < PROB_MIN) hit_prob = PROB_MIN;
544 _predicted_receiver = predicted_receiver;
545 _if_missed = if_missed;
546 _if_hit = if_hit;
547 _hit_prob = hit_prob;
548 }
550 virtual bool is_virtual() const { return true; }
551 virtual bool is_inline() const { return _if_hit->is_inline(); }
552 virtual bool is_deferred() const { return _if_hit->is_deferred(); }
554 virtual JVMState* generate(JVMState* jvms);
555 };
558 CallGenerator* CallGenerator::for_predicted_call(ciKlass* predicted_receiver,
559 CallGenerator* if_missed,
560 CallGenerator* if_hit,
561 float hit_prob) {
562 return new PredictedCallGenerator(predicted_receiver, if_missed, if_hit, hit_prob);
563 }
566 JVMState* PredictedCallGenerator::generate(JVMState* jvms) {
567 GraphKit kit(jvms);
568 PhaseGVN& gvn = kit.gvn();
569 // We need an explicit receiver null_check before checking its type.
570 // We share a map with the caller, so his JVMS gets adjusted.
571 Node* receiver = kit.argument(0);
573 CompileLog* log = kit.C->log();
574 if (log != NULL) {
575 log->elem("predicted_call bci='%d' klass='%d'",
576 jvms->bci(), log->identify(_predicted_receiver));
577 }
579 receiver = kit.null_check_receiver(method());
580 if (kit.stopped()) {
581 return kit.transfer_exceptions_into_jvms();
582 }
584 Node* exact_receiver = receiver; // will get updated in place...
585 Node* slow_ctl = kit.type_check_receiver(receiver,
586 _predicted_receiver, _hit_prob,
587 &exact_receiver);
589 SafePointNode* slow_map = NULL;
590 JVMState* slow_jvms;
591 { PreserveJVMState pjvms(&kit);
592 kit.set_control(slow_ctl);
593 if (!kit.stopped()) {
594 slow_jvms = _if_missed->generate(kit.sync_jvms());
595 assert(slow_jvms != NULL, "miss path must not fail to generate");
596 kit.add_exception_states_from(slow_jvms);
597 kit.set_map(slow_jvms->map());
598 if (!kit.stopped())
599 slow_map = kit.stop();
600 }
601 }
603 if (kit.stopped()) {
604 // Instance exactly does not matches the desired type.
605 kit.set_jvms(slow_jvms);
606 return kit.transfer_exceptions_into_jvms();
607 }
609 // fall through if the instance exactly matches the desired type
610 kit.replace_in_map(receiver, exact_receiver);
612 // Make the hot call:
613 JVMState* new_jvms = _if_hit->generate(kit.sync_jvms());
614 if (new_jvms == NULL) {
615 // Inline failed, so make a direct call.
616 assert(_if_hit->is_inline(), "must have been a failed inline");
617 CallGenerator* cg = CallGenerator::for_direct_call(_if_hit->method());
618 new_jvms = cg->generate(kit.sync_jvms());
619 }
620 kit.add_exception_states_from(new_jvms);
621 kit.set_jvms(new_jvms);
623 // Need to merge slow and fast?
624 if (slow_map == NULL) {
625 // The fast path is the only path remaining.
626 return kit.transfer_exceptions_into_jvms();
627 }
629 if (kit.stopped()) {
630 // Inlined method threw an exception, so it's just the slow path after all.
631 kit.set_jvms(slow_jvms);
632 return kit.transfer_exceptions_into_jvms();
633 }
635 // Finish the diamond.
636 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization
637 RegionNode* region = new (kit.C, 3) RegionNode(3);
638 region->init_req(1, kit.control());
639 region->init_req(2, slow_map->control());
640 kit.set_control(gvn.transform(region));
641 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO);
642 iophi->set_req(2, slow_map->i_o());
643 kit.set_i_o(gvn.transform(iophi));
644 kit.merge_memory(slow_map->merged_memory(), region, 2);
645 uint tos = kit.jvms()->stkoff() + kit.sp();
646 uint limit = slow_map->req();
647 for (uint i = TypeFunc::Parms; i < limit; i++) {
648 // Skip unused stack slots; fast forward to monoff();
649 if (i == tos) {
650 i = kit.jvms()->monoff();
651 if( i >= limit ) break;
652 }
653 Node* m = kit.map()->in(i);
654 Node* n = slow_map->in(i);
655 if (m != n) {
656 const Type* t = gvn.type(m)->meet(gvn.type(n));
657 Node* phi = PhiNode::make(region, m, t);
658 phi->set_req(2, n);
659 kit.map()->set_req(i, gvn.transform(phi));
660 }
661 }
662 return kit.transfer_exceptions_into_jvms();
663 }
666 //------------------------PredictedDynamicCallGenerator-----------------------
667 // Internal class which handles all out-of-line calls checking receiver type.
668 class PredictedDynamicCallGenerator : public CallGenerator {
669 ciMethodHandle* _predicted_method_handle;
670 CallGenerator* _if_missed;
671 CallGenerator* _if_hit;
672 float _hit_prob;
674 public:
675 PredictedDynamicCallGenerator(ciMethodHandle* predicted_method_handle,
676 CallGenerator* if_missed,
677 CallGenerator* if_hit,
678 float hit_prob)
679 : CallGenerator(if_missed->method()),
680 _predicted_method_handle(predicted_method_handle),
681 _if_missed(if_missed),
682 _if_hit(if_hit),
683 _hit_prob(hit_prob)
684 {}
686 virtual bool is_inline() const { return _if_hit->is_inline(); }
687 virtual bool is_deferred() const { return _if_hit->is_deferred(); }
689 virtual JVMState* generate(JVMState* jvms);
690 };
693 CallGenerator* CallGenerator::for_predicted_dynamic_call(ciMethodHandle* predicted_method_handle,
694 CallGenerator* if_missed,
695 CallGenerator* if_hit,
696 float hit_prob) {
697 return new PredictedDynamicCallGenerator(predicted_method_handle, if_missed, if_hit, hit_prob);
698 }
701 JVMState* PredictedDynamicCallGenerator::generate(JVMState* jvms) {
702 GraphKit kit(jvms);
703 PhaseGVN& gvn = kit.gvn();
705 CompileLog* log = kit.C->log();
706 if (log != NULL) {
707 log->elem("predicted_dynamic_call bci='%d'", jvms->bci());
708 }
710 // Get the constant pool cache from the caller class.
711 ciMethod* caller_method = jvms->method();
712 ciBytecodeStream str(caller_method);
713 str.force_bci(jvms->bci()); // Set the stream to the invokedynamic bci.
714 ciCPCache* cpcache = str.get_cpcache();
716 // Get the offset of the CallSite from the constant pool cache
717 // pointer.
718 int index = str.get_method_index();
719 size_t call_site_offset = cpcache->get_f1_offset(index);
721 // Load the CallSite object from the constant pool cache.
722 const TypeOopPtr* cpcache_ptr = TypeOopPtr::make_from_constant(cpcache);
723 Node* cpcache_adr = kit.makecon(cpcache_ptr);
724 Node* call_site_adr = kit.basic_plus_adr(cpcache_adr, cpcache_adr, call_site_offset);
725 Node* call_site = kit.make_load(kit.control(), call_site_adr, TypeInstPtr::BOTTOM, T_OBJECT, Compile::AliasIdxRaw);
727 // Load the target MethodHandle from the CallSite object.
728 Node* target_adr = kit.basic_plus_adr(call_site, call_site, java_lang_invoke_CallSite::target_offset_in_bytes());
729 Node* target_mh = kit.make_load(kit.control(), target_adr, TypeInstPtr::BOTTOM, T_OBJECT);
731 // Check if the MethodHandle is still the same.
732 const TypeOopPtr* predicted_mh_ptr = TypeOopPtr::make_from_constant(_predicted_method_handle, true);
733 Node* predicted_mh = kit.makecon(predicted_mh_ptr);
735 Node* cmp = gvn.transform(new(kit.C, 3) CmpPNode(target_mh, predicted_mh));
736 Node* bol = gvn.transform(new(kit.C, 2) BoolNode(cmp, BoolTest::eq) );
737 IfNode* iff = kit.create_and_xform_if(kit.control(), bol, _hit_prob, COUNT_UNKNOWN);
738 kit.set_control( gvn.transform(new(kit.C, 1) IfTrueNode (iff)));
739 Node* slow_ctl = gvn.transform(new(kit.C, 1) IfFalseNode(iff));
741 SafePointNode* slow_map = NULL;
742 JVMState* slow_jvms;
743 { PreserveJVMState pjvms(&kit);
744 kit.set_control(slow_ctl);
745 if (!kit.stopped()) {
746 slow_jvms = _if_missed->generate(kit.sync_jvms());
747 assert(slow_jvms != NULL, "miss path must not fail to generate");
748 kit.add_exception_states_from(slow_jvms);
749 kit.set_map(slow_jvms->map());
750 if (!kit.stopped())
751 slow_map = kit.stop();
752 }
753 }
755 if (kit.stopped()) {
756 // Instance exactly does not matches the desired type.
757 kit.set_jvms(slow_jvms);
758 return kit.transfer_exceptions_into_jvms();
759 }
761 // Make the hot call:
762 JVMState* new_jvms = _if_hit->generate(kit.sync_jvms());
763 if (new_jvms == NULL) {
764 // Inline failed, so make a direct call.
765 assert(_if_hit->is_inline(), "must have been a failed inline");
766 CallGenerator* cg = CallGenerator::for_direct_call(_if_hit->method());
767 new_jvms = cg->generate(kit.sync_jvms());
768 }
769 kit.add_exception_states_from(new_jvms);
770 kit.set_jvms(new_jvms);
772 // Need to merge slow and fast?
773 if (slow_map == NULL) {
774 // The fast path is the only path remaining.
775 return kit.transfer_exceptions_into_jvms();
776 }
778 if (kit.stopped()) {
779 // Inlined method threw an exception, so it's just the slow path after all.
780 kit.set_jvms(slow_jvms);
781 return kit.transfer_exceptions_into_jvms();
782 }
784 // Finish the diamond.
785 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization
786 RegionNode* region = new (kit.C, 3) RegionNode(3);
787 region->init_req(1, kit.control());
788 region->init_req(2, slow_map->control());
789 kit.set_control(gvn.transform(region));
790 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO);
791 iophi->set_req(2, slow_map->i_o());
792 kit.set_i_o(gvn.transform(iophi));
793 kit.merge_memory(slow_map->merged_memory(), region, 2);
794 uint tos = kit.jvms()->stkoff() + kit.sp();
795 uint limit = slow_map->req();
796 for (uint i = TypeFunc::Parms; i < limit; i++) {
797 // Skip unused stack slots; fast forward to monoff();
798 if (i == tos) {
799 i = kit.jvms()->monoff();
800 if( i >= limit ) break;
801 }
802 Node* m = kit.map()->in(i);
803 Node* n = slow_map->in(i);
804 if (m != n) {
805 const Type* t = gvn.type(m)->meet(gvn.type(n));
806 Node* phi = PhiNode::make(region, m, t);
807 phi->set_req(2, n);
808 kit.map()->set_req(i, gvn.transform(phi));
809 }
810 }
811 return kit.transfer_exceptions_into_jvms();
812 }
815 //-------------------------UncommonTrapCallGenerator-----------------------------
816 // Internal class which handles all out-of-line calls checking receiver type.
817 class UncommonTrapCallGenerator : public CallGenerator {
818 Deoptimization::DeoptReason _reason;
819 Deoptimization::DeoptAction _action;
821 public:
822 UncommonTrapCallGenerator(ciMethod* m,
823 Deoptimization::DeoptReason reason,
824 Deoptimization::DeoptAction action)
825 : CallGenerator(m)
826 {
827 _reason = reason;
828 _action = action;
829 }
831 virtual bool is_virtual() const { ShouldNotReachHere(); return false; }
832 virtual bool is_trap() const { return true; }
834 virtual JVMState* generate(JVMState* jvms);
835 };
838 CallGenerator*
839 CallGenerator::for_uncommon_trap(ciMethod* m,
840 Deoptimization::DeoptReason reason,
841 Deoptimization::DeoptAction action) {
842 return new UncommonTrapCallGenerator(m, reason, action);
843 }
846 JVMState* UncommonTrapCallGenerator::generate(JVMState* jvms) {
847 GraphKit kit(jvms);
848 // Take the trap with arguments pushed on the stack. (Cf. null_check_receiver).
849 int nargs = method()->arg_size();
850 kit.inc_sp(nargs);
851 assert(nargs <= kit.sp() && kit.sp() <= jvms->stk_size(), "sane sp w/ args pushed");
852 if (_reason == Deoptimization::Reason_class_check &&
853 _action == Deoptimization::Action_maybe_recompile) {
854 // Temp fix for 6529811
855 // Don't allow uncommon_trap to override our decision to recompile in the event
856 // of a class cast failure for a monomorphic call as it will never let us convert
857 // the call to either bi-morphic or megamorphic and can lead to unc-trap loops
858 bool keep_exact_action = true;
859 kit.uncommon_trap(_reason, _action, NULL, "monomorphic vcall checkcast", false, keep_exact_action);
860 } else {
861 kit.uncommon_trap(_reason, _action);
862 }
863 return kit.transfer_exceptions_into_jvms();
864 }
866 // (Note: Moved hook_up_call to GraphKit::set_edges_for_java_call.)
868 // (Node: Merged hook_up_exits into ParseGenerator::generate.)
870 #define NODES_OVERHEAD_PER_METHOD (30.0)
871 #define NODES_PER_BYTECODE (9.5)
873 void WarmCallInfo::init(JVMState* call_site, ciMethod* call_method, ciCallProfile& profile, float prof_factor) {
874 int call_count = profile.count();
875 int code_size = call_method->code_size();
877 // Expected execution count is based on the historical count:
878 _count = call_count < 0 ? 1 : call_site->method()->scale_count(call_count, prof_factor);
880 // Expected profit from inlining, in units of simple call-overheads.
881 _profit = 1.0;
883 // Expected work performed by the call in units of call-overheads.
884 // %%% need an empirical curve fit for "work" (time in call)
885 float bytecodes_per_call = 3;
886 _work = 1.0 + code_size / bytecodes_per_call;
888 // Expected size of compilation graph:
889 // -XX:+PrintParseStatistics once reported:
890 // Methods seen: 9184 Methods parsed: 9184 Nodes created: 1582391
891 // Histogram of 144298 parsed bytecodes:
892 // %%% Need an better predictor for graph size.
893 _size = NODES_OVERHEAD_PER_METHOD + (NODES_PER_BYTECODE * code_size);
894 }
896 // is_cold: Return true if the node should never be inlined.
897 // This is true if any of the key metrics are extreme.
898 bool WarmCallInfo::is_cold() const {
899 if (count() < WarmCallMinCount) return true;
900 if (profit() < WarmCallMinProfit) return true;
901 if (work() > WarmCallMaxWork) return true;
902 if (size() > WarmCallMaxSize) return true;
903 return false;
904 }
906 // is_hot: Return true if the node should be inlined immediately.
907 // This is true if any of the key metrics are extreme.
908 bool WarmCallInfo::is_hot() const {
909 assert(!is_cold(), "eliminate is_cold cases before testing is_hot");
910 if (count() >= HotCallCountThreshold) return true;
911 if (profit() >= HotCallProfitThreshold) return true;
912 if (work() <= HotCallTrivialWork) return true;
913 if (size() <= HotCallTrivialSize) return true;
914 return false;
915 }
917 // compute_heat:
918 float WarmCallInfo::compute_heat() const {
919 assert(!is_cold(), "compute heat only on warm nodes");
920 assert(!is_hot(), "compute heat only on warm nodes");
921 int min_size = MAX2(0, (int)HotCallTrivialSize);
922 int max_size = MIN2(500, (int)WarmCallMaxSize);
923 float method_size = (size() - min_size) / MAX2(1, max_size - min_size);
924 float size_factor;
925 if (method_size < 0.05) size_factor = 4; // 2 sigmas better than avg.
926 else if (method_size < 0.15) size_factor = 2; // 1 sigma better than avg.
927 else if (method_size < 0.5) size_factor = 1; // better than avg.
928 else size_factor = 0.5; // worse than avg.
929 return (count() * profit() * size_factor);
930 }
932 bool WarmCallInfo::warmer_than(WarmCallInfo* that) {
933 assert(this != that, "compare only different WCIs");
934 assert(this->heat() != 0 && that->heat() != 0, "call compute_heat 1st");
935 if (this->heat() > that->heat()) return true;
936 if (this->heat() < that->heat()) return false;
937 assert(this->heat() == that->heat(), "no NaN heat allowed");
938 // Equal heat. Break the tie some other way.
939 if (!this->call() || !that->call()) return (address)this > (address)that;
940 return this->call()->_idx > that->call()->_idx;
941 }
943 //#define UNINIT_NEXT ((WarmCallInfo*)badAddress)
944 #define UNINIT_NEXT ((WarmCallInfo*)NULL)
946 WarmCallInfo* WarmCallInfo::insert_into(WarmCallInfo* head) {
947 assert(next() == UNINIT_NEXT, "not yet on any list");
948 WarmCallInfo* prev_p = NULL;
949 WarmCallInfo* next_p = head;
950 while (next_p != NULL && next_p->warmer_than(this)) {
951 prev_p = next_p;
952 next_p = prev_p->next();
953 }
954 // Install this between prev_p and next_p.
955 this->set_next(next_p);
956 if (prev_p == NULL)
957 head = this;
958 else
959 prev_p->set_next(this);
960 return head;
961 }
963 WarmCallInfo* WarmCallInfo::remove_from(WarmCallInfo* head) {
964 WarmCallInfo* prev_p = NULL;
965 WarmCallInfo* next_p = head;
966 while (next_p != this) {
967 assert(next_p != NULL, "this must be in the list somewhere");
968 prev_p = next_p;
969 next_p = prev_p->next();
970 }
971 next_p = this->next();
972 debug_only(this->set_next(UNINIT_NEXT));
973 // Remove this from between prev_p and next_p.
974 if (prev_p == NULL)
975 head = next_p;
976 else
977 prev_p->set_next(next_p);
978 return head;
979 }
981 WarmCallInfo* WarmCallInfo::_always_hot = NULL;
982 WarmCallInfo* WarmCallInfo::_always_cold = NULL;
984 WarmCallInfo* WarmCallInfo::always_hot() {
985 if (_always_hot == NULL) {
986 static double bits[sizeof(WarmCallInfo) / sizeof(double) + 1] = {0};
987 WarmCallInfo* ci = (WarmCallInfo*) bits;
988 ci->_profit = ci->_count = MAX_VALUE();
989 ci->_work = ci->_size = MIN_VALUE();
990 _always_hot = ci;
991 }
992 assert(_always_hot->is_hot(), "must always be hot");
993 return _always_hot;
994 }
996 WarmCallInfo* WarmCallInfo::always_cold() {
997 if (_always_cold == NULL) {
998 static double bits[sizeof(WarmCallInfo) / sizeof(double) + 1] = {0};
999 WarmCallInfo* ci = (WarmCallInfo*) bits;
1000 ci->_profit = ci->_count = MIN_VALUE();
1001 ci->_work = ci->_size = MAX_VALUE();
1002 _always_cold = ci;
1003 }
1004 assert(_always_cold->is_cold(), "must always be cold");
1005 return _always_cold;
1006 }
1009 #ifndef PRODUCT
1011 void WarmCallInfo::print() const {
1012 tty->print("%s : C=%6.1f P=%6.1f W=%6.1f S=%6.1f H=%6.1f -> %p",
1013 is_cold() ? "cold" : is_hot() ? "hot " : "warm",
1014 count(), profit(), work(), size(), compute_heat(), next());
1015 tty->cr();
1016 if (call() != NULL) call()->dump();
1017 }
1019 void print_wci(WarmCallInfo* ci) {
1020 ci->print();
1021 }
1023 void WarmCallInfo::print_all() const {
1024 for (const WarmCallInfo* p = this; p != NULL; p = p->next())
1025 p->print();
1026 }
1028 int WarmCallInfo::count_all() const {
1029 int cnt = 0;
1030 for (const WarmCallInfo* p = this; p != NULL; p = p->next())
1031 cnt++;
1032 return cnt;
1033 }
1035 #endif //PRODUCT