Thu, 03 Jan 2013 15:09:55 -0800
8005522: use fast-string instructions on x86 for zeroing
Summary: use 'rep stosb' instead of 'rep stosq' when fast-string operations are available.
Reviewed-by: twisti, roland
1 /*
2 * Copyright (c) 2000, 2012, 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/ciCallSite.hpp"
28 #include "ci/ciObjArray.hpp"
29 #include "ci/ciMemberName.hpp"
30 #include "ci/ciMethodHandle.hpp"
31 #include "classfile/javaClasses.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/callGenerator.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/connode.hpp"
38 #include "opto/parse.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/runtime.hpp"
41 #include "opto/subnode.hpp"
44 // Utility function.
45 const TypeFunc* CallGenerator::tf() const {
46 return TypeFunc::make(method());
47 }
49 //-----------------------------ParseGenerator---------------------------------
50 // Internal class which handles all direct bytecode traversal.
51 class ParseGenerator : public InlineCallGenerator {
52 private:
53 bool _is_osr;
54 float _expected_uses;
56 public:
57 ParseGenerator(ciMethod* method, float expected_uses, bool is_osr = false)
58 : InlineCallGenerator(method)
59 {
60 _is_osr = is_osr;
61 _expected_uses = expected_uses;
62 assert(InlineTree::check_can_parse(method) == NULL, "parse must be possible");
63 }
65 virtual bool is_parse() const { return true; }
66 virtual JVMState* generate(JVMState* jvms);
67 int is_osr() { return _is_osr; }
69 };
71 JVMState* ParseGenerator::generate(JVMState* jvms) {
72 Compile* C = Compile::current();
74 if (is_osr()) {
75 // The JVMS for a OSR has a single argument (see its TypeFunc).
76 assert(jvms->depth() == 1, "no inline OSR");
77 }
79 if (C->failing()) {
80 return NULL; // bailing out of the compile; do not try to parse
81 }
83 Parse parser(jvms, method(), _expected_uses);
84 // Grab signature for matching/allocation
85 #ifdef ASSERT
86 if (parser.tf() != (parser.depth() == 1 ? C->tf() : tf())) {
87 MutexLockerEx ml(Compile_lock, Mutex::_no_safepoint_check_flag);
88 assert(C->env()->system_dictionary_modification_counter_changed(),
89 "Must invalidate if TypeFuncs differ");
90 }
91 #endif
93 GraphKit& exits = parser.exits();
95 if (C->failing()) {
96 while (exits.pop_exception_state() != NULL) ;
97 return NULL;
98 }
100 assert(exits.jvms()->same_calls_as(jvms), "sanity");
102 // Simply return the exit state of the parser,
103 // augmented by any exceptional states.
104 return exits.transfer_exceptions_into_jvms();
105 }
107 //---------------------------DirectCallGenerator------------------------------
108 // Internal class which handles all out-of-line calls w/o receiver type checks.
109 class DirectCallGenerator : public CallGenerator {
110 private:
111 CallStaticJavaNode* _call_node;
112 // Force separate memory and I/O projections for the exceptional
113 // paths to facilitate late inlinig.
114 bool _separate_io_proj;
116 public:
117 DirectCallGenerator(ciMethod* method, bool separate_io_proj)
118 : CallGenerator(method),
119 _separate_io_proj(separate_io_proj)
120 {
121 }
122 virtual JVMState* generate(JVMState* jvms);
124 CallStaticJavaNode* call_node() const { return _call_node; }
125 };
127 JVMState* DirectCallGenerator::generate(JVMState* jvms) {
128 GraphKit kit(jvms);
129 bool is_static = method()->is_static();
130 address target = is_static ? SharedRuntime::get_resolve_static_call_stub()
131 : SharedRuntime::get_resolve_opt_virtual_call_stub();
133 if (kit.C->log() != NULL) {
134 kit.C->log()->elem("direct_call bci='%d'", jvms->bci());
135 }
137 CallStaticJavaNode *call = new (kit.C) CallStaticJavaNode(tf(), target, method(), kit.bci());
138 _call_node = call; // Save the call node in case we need it later
139 if (!is_static) {
140 // Make an explicit receiver null_check as part of this call.
141 // Since we share a map with the caller, his JVMS gets adjusted.
142 kit.null_check_receiver_before_call(method());
143 if (kit.stopped()) {
144 // And dump it back to the caller, decorated with any exceptions:
145 return kit.transfer_exceptions_into_jvms();
146 }
147 // Mark the call node as virtual, sort of:
148 call->set_optimized_virtual(true);
149 if (method()->is_method_handle_intrinsic() ||
150 method()->is_compiled_lambda_form()) {
151 call->set_method_handle_invoke(true);
152 }
153 }
154 kit.set_arguments_for_java_call(call);
155 kit.set_edges_for_java_call(call, false, _separate_io_proj);
156 Node* ret = kit.set_results_for_java_call(call, _separate_io_proj);
157 kit.push_node(method()->return_type()->basic_type(), ret);
158 return kit.transfer_exceptions_into_jvms();
159 }
161 //--------------------------VirtualCallGenerator------------------------------
162 // Internal class which handles all out-of-line calls checking receiver type.
163 class VirtualCallGenerator : public CallGenerator {
164 private:
165 int _vtable_index;
166 public:
167 VirtualCallGenerator(ciMethod* method, int vtable_index)
168 : CallGenerator(method), _vtable_index(vtable_index)
169 {
170 assert(vtable_index == Method::invalid_vtable_index ||
171 vtable_index >= 0, "either invalid or usable");
172 }
173 virtual bool is_virtual() const { return true; }
174 virtual JVMState* generate(JVMState* jvms);
175 };
177 JVMState* VirtualCallGenerator::generate(JVMState* jvms) {
178 GraphKit kit(jvms);
179 Node* receiver = kit.argument(0);
181 if (kit.C->log() != NULL) {
182 kit.C->log()->elem("virtual_call bci='%d'", jvms->bci());
183 }
185 // If the receiver is a constant null, do not torture the system
186 // by attempting to call through it. The compile will proceed
187 // correctly, but may bail out in final_graph_reshaping, because
188 // the call instruction will have a seemingly deficient out-count.
189 // (The bailout says something misleading about an "infinite loop".)
190 if (kit.gvn().type(receiver)->higher_equal(TypePtr::NULL_PTR)) {
191 kit.inc_sp(method()->arg_size()); // restore arguments
192 kit.uncommon_trap(Deoptimization::Reason_null_check,
193 Deoptimization::Action_none,
194 NULL, "null receiver");
195 return kit.transfer_exceptions_into_jvms();
196 }
198 // Ideally we would unconditionally do a null check here and let it
199 // be converted to an implicit check based on profile information.
200 // However currently the conversion to implicit null checks in
201 // Block::implicit_null_check() only looks for loads and stores, not calls.
202 ciMethod *caller = kit.method();
203 ciMethodData *caller_md = (caller == NULL) ? NULL : caller->method_data();
204 if (!UseInlineCaches || !ImplicitNullChecks ||
205 ((ImplicitNullCheckThreshold > 0) && caller_md &&
206 (caller_md->trap_count(Deoptimization::Reason_null_check)
207 >= (uint)ImplicitNullCheckThreshold))) {
208 // Make an explicit receiver null_check as part of this call.
209 // Since we share a map with the caller, his JVMS gets adjusted.
210 receiver = kit.null_check_receiver_before_call(method());
211 if (kit.stopped()) {
212 // And dump it back to the caller, decorated with any exceptions:
213 return kit.transfer_exceptions_into_jvms();
214 }
215 }
217 assert(!method()->is_static(), "virtual call must not be to static");
218 assert(!method()->is_final(), "virtual call should not be to final");
219 assert(!method()->is_private(), "virtual call should not be to private");
220 assert(_vtable_index == Method::invalid_vtable_index || !UseInlineCaches,
221 "no vtable calls if +UseInlineCaches ");
222 address target = SharedRuntime::get_resolve_virtual_call_stub();
223 // Normal inline cache used for call
224 CallDynamicJavaNode *call = new (kit.C) CallDynamicJavaNode(tf(), target, method(), _vtable_index, kit.bci());
225 kit.set_arguments_for_java_call(call);
226 kit.set_edges_for_java_call(call);
227 Node* ret = kit.set_results_for_java_call(call);
228 kit.push_node(method()->return_type()->basic_type(), ret);
230 // Represent the effect of an implicit receiver null_check
231 // as part of this call. Since we share a map with the caller,
232 // his JVMS gets adjusted.
233 kit.cast_not_null(receiver);
234 return kit.transfer_exceptions_into_jvms();
235 }
237 CallGenerator* CallGenerator::for_inline(ciMethod* m, float expected_uses) {
238 if (InlineTree::check_can_parse(m) != NULL) return NULL;
239 return new ParseGenerator(m, expected_uses);
240 }
242 // As a special case, the JVMS passed to this CallGenerator is
243 // for the method execution already in progress, not just the JVMS
244 // of the caller. Thus, this CallGenerator cannot be mixed with others!
245 CallGenerator* CallGenerator::for_osr(ciMethod* m, int osr_bci) {
246 if (InlineTree::check_can_parse(m) != NULL) return NULL;
247 float past_uses = m->interpreter_invocation_count();
248 float expected_uses = past_uses;
249 return new ParseGenerator(m, expected_uses, true);
250 }
252 CallGenerator* CallGenerator::for_direct_call(ciMethod* m, bool separate_io_proj) {
253 assert(!m->is_abstract(), "for_direct_call mismatch");
254 return new DirectCallGenerator(m, separate_io_proj);
255 }
257 CallGenerator* CallGenerator::for_virtual_call(ciMethod* m, int vtable_index) {
258 assert(!m->is_static(), "for_virtual_call mismatch");
259 assert(!m->is_method_handle_intrinsic(), "should be a direct call");
260 return new VirtualCallGenerator(m, vtable_index);
261 }
263 // Allow inlining decisions to be delayed
264 class LateInlineCallGenerator : public DirectCallGenerator {
265 protected:
266 CallGenerator* _inline_cg;
268 virtual bool do_late_inline_check(JVMState* jvms) { return true; }
270 public:
271 LateInlineCallGenerator(ciMethod* method, CallGenerator* inline_cg) :
272 DirectCallGenerator(method, true), _inline_cg(inline_cg) {}
274 virtual bool is_late_inline() const { return true; }
276 // Convert the CallStaticJava into an inline
277 virtual void do_late_inline();
279 virtual JVMState* generate(JVMState* jvms) {
280 Compile *C = Compile::current();
281 C->print_inlining_skip(this);
283 // Record that this call site should be revisited once the main
284 // parse is finished.
285 if (!is_mh_late_inline()) {
286 C->add_late_inline(this);
287 }
289 // Emit the CallStaticJava and request separate projections so
290 // that the late inlining logic can distinguish between fall
291 // through and exceptional uses of the memory and io projections
292 // as is done for allocations and macro expansion.
293 return DirectCallGenerator::generate(jvms);
294 }
296 virtual void print_inlining_late(const char* msg) {
297 CallNode* call = call_node();
298 Compile* C = Compile::current();
299 C->print_inlining_insert(this);
300 C->print_inlining(method(), call->jvms()->depth()-1, call->jvms()->bci(), msg);
301 }
303 };
305 void LateInlineCallGenerator::do_late_inline() {
306 // Can't inline it
307 if (call_node() == NULL || call_node()->outcnt() == 0 ||
308 call_node()->in(0) == NULL || call_node()->in(0)->is_top())
309 return;
311 for (int i1 = 0; i1 < method()->arg_size(); i1++) {
312 if (call_node()->in(TypeFunc::Parms + i1)->is_top()) {
313 assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing");
314 return;
315 }
316 }
318 if (call_node()->in(TypeFunc::Memory)->is_top()) {
319 assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing");
320 return;
321 }
323 CallStaticJavaNode* call = call_node();
325 // Make a clone of the JVMState that appropriate to use for driving a parse
326 Compile* C = Compile::current();
327 JVMState* jvms = call->jvms()->clone_shallow(C);
328 uint size = call->req();
329 SafePointNode* map = new (C) SafePointNode(size, jvms);
330 for (uint i1 = 0; i1 < size; i1++) {
331 map->init_req(i1, call->in(i1));
332 }
334 // Make sure the state is a MergeMem for parsing.
335 if (!map->in(TypeFunc::Memory)->is_MergeMem()) {
336 Node* mem = MergeMemNode::make(C, map->in(TypeFunc::Memory));
337 C->initial_gvn()->set_type_bottom(mem);
338 map->set_req(TypeFunc::Memory, mem);
339 }
341 // Make enough space for the expression stack and transfer the incoming arguments
342 int nargs = method()->arg_size();
343 jvms->set_map(map);
344 map->ensure_stack(jvms, jvms->method()->max_stack());
345 if (nargs > 0) {
346 for (int i1 = 0; i1 < nargs; i1++) {
347 map->set_req(i1 + jvms->argoff(), call->in(TypeFunc::Parms + i1));
348 }
349 }
351 if (!do_late_inline_check(jvms)) {
352 map->disconnect_inputs(NULL, C);
353 return;
354 }
356 C->print_inlining_insert(this);
358 CompileLog* log = C->log();
359 if (log != NULL) {
360 log->head("late_inline method='%d'", log->identify(method()));
361 JVMState* p = jvms;
362 while (p != NULL) {
363 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
364 p = p->caller();
365 }
366 log->tail("late_inline");
367 }
369 // Setup default node notes to be picked up by the inlining
370 Node_Notes* old_nn = C->default_node_notes();
371 if (old_nn != NULL) {
372 Node_Notes* entry_nn = old_nn->clone(C);
373 entry_nn->set_jvms(jvms);
374 C->set_default_node_notes(entry_nn);
375 }
377 // Now perform the inling using the synthesized JVMState
378 JVMState* new_jvms = _inline_cg->generate(jvms);
379 if (new_jvms == NULL) return; // no change
380 if (C->failing()) return;
382 // Capture any exceptional control flow
383 GraphKit kit(new_jvms);
385 // Find the result object
386 Node* result = C->top();
387 int result_size = method()->return_type()->size();
388 if (result_size != 0 && !kit.stopped()) {
389 result = (result_size == 1) ? kit.pop() : kit.pop_pair();
390 }
392 C->set_has_loops(C->has_loops() || _inline_cg->method()->has_loops());
393 C->env()->notice_inlined_method(_inline_cg->method());
394 C->set_inlining_progress(true);
396 kit.replace_call(call, result);
397 }
400 CallGenerator* CallGenerator::for_late_inline(ciMethod* method, CallGenerator* inline_cg) {
401 return new LateInlineCallGenerator(method, inline_cg);
402 }
404 class LateInlineMHCallGenerator : public LateInlineCallGenerator {
405 ciMethod* _caller;
406 int _attempt;
407 bool _input_not_const;
409 virtual bool do_late_inline_check(JVMState* jvms);
410 virtual bool already_attempted() const { return _attempt > 0; }
412 public:
413 LateInlineMHCallGenerator(ciMethod* caller, ciMethod* callee, bool input_not_const) :
414 LateInlineCallGenerator(callee, NULL), _caller(caller), _attempt(0), _input_not_const(input_not_const) {}
416 virtual bool is_mh_late_inline() const { return true; }
418 virtual JVMState* generate(JVMState* jvms) {
419 JVMState* new_jvms = LateInlineCallGenerator::generate(jvms);
420 if (_input_not_const) {
421 // inlining won't be possible so no need to enqueue right now.
422 call_node()->set_generator(this);
423 } else {
424 Compile::current()->add_late_inline(this);
425 }
426 return new_jvms;
427 }
429 virtual void print_inlining_late(const char* msg) {
430 if (!_input_not_const) return;
431 LateInlineCallGenerator::print_inlining_late(msg);
432 }
433 };
435 bool LateInlineMHCallGenerator::do_late_inline_check(JVMState* jvms) {
437 CallGenerator* cg = for_method_handle_inline(jvms, _caller, method(), _input_not_const);
439 if (!_input_not_const) {
440 _attempt++;
441 }
443 if (cg != NULL) {
444 assert(!cg->is_late_inline() && cg->is_inline(), "we're doing late inlining");
445 _inline_cg = cg;
446 Compile::current()->dec_number_of_mh_late_inlines();
447 return true;
448 }
450 call_node()->set_generator(this);
451 return false;
452 }
454 CallGenerator* CallGenerator::for_mh_late_inline(ciMethod* caller, ciMethod* callee, bool input_not_const) {
455 Compile::current()->inc_number_of_mh_late_inlines();
456 CallGenerator* cg = new LateInlineMHCallGenerator(caller, callee, input_not_const);
457 return cg;
458 }
460 class LateInlineStringCallGenerator : public LateInlineCallGenerator {
462 public:
463 LateInlineStringCallGenerator(ciMethod* method, CallGenerator* inline_cg) :
464 LateInlineCallGenerator(method, inline_cg) {}
466 virtual JVMState* generate(JVMState* jvms) {
467 Compile *C = Compile::current();
468 C->print_inlining_skip(this);
470 C->add_string_late_inline(this);
472 JVMState* new_jvms = DirectCallGenerator::generate(jvms);
473 return new_jvms;
474 }
475 };
477 CallGenerator* CallGenerator::for_string_late_inline(ciMethod* method, CallGenerator* inline_cg) {
478 return new LateInlineStringCallGenerator(method, inline_cg);
479 }
482 //---------------------------WarmCallGenerator--------------------------------
483 // Internal class which handles initial deferral of inlining decisions.
484 class WarmCallGenerator : public CallGenerator {
485 WarmCallInfo* _call_info;
486 CallGenerator* _if_cold;
487 CallGenerator* _if_hot;
488 bool _is_virtual; // caches virtuality of if_cold
489 bool _is_inline; // caches inline-ness of if_hot
491 public:
492 WarmCallGenerator(WarmCallInfo* ci,
493 CallGenerator* if_cold,
494 CallGenerator* if_hot)
495 : CallGenerator(if_cold->method())
496 {
497 assert(method() == if_hot->method(), "consistent choices");
498 _call_info = ci;
499 _if_cold = if_cold;
500 _if_hot = if_hot;
501 _is_virtual = if_cold->is_virtual();
502 _is_inline = if_hot->is_inline();
503 }
505 virtual bool is_inline() const { return _is_inline; }
506 virtual bool is_virtual() const { return _is_virtual; }
507 virtual bool is_deferred() const { return true; }
509 virtual JVMState* generate(JVMState* jvms);
510 };
513 CallGenerator* CallGenerator::for_warm_call(WarmCallInfo* ci,
514 CallGenerator* if_cold,
515 CallGenerator* if_hot) {
516 return new WarmCallGenerator(ci, if_cold, if_hot);
517 }
519 JVMState* WarmCallGenerator::generate(JVMState* jvms) {
520 Compile* C = Compile::current();
521 if (C->log() != NULL) {
522 C->log()->elem("warm_call bci='%d'", jvms->bci());
523 }
524 jvms = _if_cold->generate(jvms);
525 if (jvms != NULL) {
526 Node* m = jvms->map()->control();
527 if (m->is_CatchProj()) m = m->in(0); else m = C->top();
528 if (m->is_Catch()) m = m->in(0); else m = C->top();
529 if (m->is_Proj()) m = m->in(0); else m = C->top();
530 if (m->is_CallJava()) {
531 _call_info->set_call(m->as_Call());
532 _call_info->set_hot_cg(_if_hot);
533 #ifndef PRODUCT
534 if (PrintOpto || PrintOptoInlining) {
535 tty->print_cr("Queueing for warm inlining at bci %d:", jvms->bci());
536 tty->print("WCI: ");
537 _call_info->print();
538 }
539 #endif
540 _call_info->set_heat(_call_info->compute_heat());
541 C->set_warm_calls(_call_info->insert_into(C->warm_calls()));
542 }
543 }
544 return jvms;
545 }
547 void WarmCallInfo::make_hot() {
548 Unimplemented();
549 }
551 void WarmCallInfo::make_cold() {
552 // No action: Just dequeue.
553 }
556 //------------------------PredictedCallGenerator------------------------------
557 // Internal class which handles all out-of-line calls checking receiver type.
558 class PredictedCallGenerator : public CallGenerator {
559 ciKlass* _predicted_receiver;
560 CallGenerator* _if_missed;
561 CallGenerator* _if_hit;
562 float _hit_prob;
564 public:
565 PredictedCallGenerator(ciKlass* predicted_receiver,
566 CallGenerator* if_missed,
567 CallGenerator* if_hit, float hit_prob)
568 : CallGenerator(if_missed->method())
569 {
570 // The call profile data may predict the hit_prob as extreme as 0 or 1.
571 // Remove the extremes values from the range.
572 if (hit_prob > PROB_MAX) hit_prob = PROB_MAX;
573 if (hit_prob < PROB_MIN) hit_prob = PROB_MIN;
575 _predicted_receiver = predicted_receiver;
576 _if_missed = if_missed;
577 _if_hit = if_hit;
578 _hit_prob = hit_prob;
579 }
581 virtual bool is_virtual() const { return true; }
582 virtual bool is_inline() const { return _if_hit->is_inline(); }
583 virtual bool is_deferred() const { return _if_hit->is_deferred(); }
585 virtual JVMState* generate(JVMState* jvms);
586 };
589 CallGenerator* CallGenerator::for_predicted_call(ciKlass* predicted_receiver,
590 CallGenerator* if_missed,
591 CallGenerator* if_hit,
592 float hit_prob) {
593 return new PredictedCallGenerator(predicted_receiver, if_missed, if_hit, hit_prob);
594 }
597 JVMState* PredictedCallGenerator::generate(JVMState* jvms) {
598 GraphKit kit(jvms);
599 PhaseGVN& gvn = kit.gvn();
600 // We need an explicit receiver null_check before checking its type.
601 // We share a map with the caller, so his JVMS gets adjusted.
602 Node* receiver = kit.argument(0);
604 CompileLog* log = kit.C->log();
605 if (log != NULL) {
606 log->elem("predicted_call bci='%d' klass='%d'",
607 jvms->bci(), log->identify(_predicted_receiver));
608 }
610 receiver = kit.null_check_receiver_before_call(method());
611 if (kit.stopped()) {
612 return kit.transfer_exceptions_into_jvms();
613 }
615 Node* exact_receiver = receiver; // will get updated in place...
616 Node* slow_ctl = kit.type_check_receiver(receiver,
617 _predicted_receiver, _hit_prob,
618 &exact_receiver);
620 SafePointNode* slow_map = NULL;
621 JVMState* slow_jvms;
622 { PreserveJVMState pjvms(&kit);
623 kit.set_control(slow_ctl);
624 if (!kit.stopped()) {
625 slow_jvms = _if_missed->generate(kit.sync_jvms());
626 if (kit.failing())
627 return NULL; // might happen because of NodeCountInliningCutoff
628 assert(slow_jvms != NULL, "must be");
629 kit.add_exception_states_from(slow_jvms);
630 kit.set_map(slow_jvms->map());
631 if (!kit.stopped())
632 slow_map = kit.stop();
633 }
634 }
636 if (kit.stopped()) {
637 // Instance exactly does not matches the desired type.
638 kit.set_jvms(slow_jvms);
639 return kit.transfer_exceptions_into_jvms();
640 }
642 // fall through if the instance exactly matches the desired type
643 kit.replace_in_map(receiver, exact_receiver);
645 // Make the hot call:
646 JVMState* new_jvms = _if_hit->generate(kit.sync_jvms());
647 if (new_jvms == NULL) {
648 // Inline failed, so make a direct call.
649 assert(_if_hit->is_inline(), "must have been a failed inline");
650 CallGenerator* cg = CallGenerator::for_direct_call(_if_hit->method());
651 new_jvms = cg->generate(kit.sync_jvms());
652 }
653 kit.add_exception_states_from(new_jvms);
654 kit.set_jvms(new_jvms);
656 // Need to merge slow and fast?
657 if (slow_map == NULL) {
658 // The fast path is the only path remaining.
659 return kit.transfer_exceptions_into_jvms();
660 }
662 if (kit.stopped()) {
663 // Inlined method threw an exception, so it's just the slow path after all.
664 kit.set_jvms(slow_jvms);
665 return kit.transfer_exceptions_into_jvms();
666 }
668 // Finish the diamond.
669 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization
670 RegionNode* region = new (kit.C) RegionNode(3);
671 region->init_req(1, kit.control());
672 region->init_req(2, slow_map->control());
673 kit.set_control(gvn.transform(region));
674 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO);
675 iophi->set_req(2, slow_map->i_o());
676 kit.set_i_o(gvn.transform(iophi));
677 kit.merge_memory(slow_map->merged_memory(), region, 2);
678 uint tos = kit.jvms()->stkoff() + kit.sp();
679 uint limit = slow_map->req();
680 for (uint i = TypeFunc::Parms; i < limit; i++) {
681 // Skip unused stack slots; fast forward to monoff();
682 if (i == tos) {
683 i = kit.jvms()->monoff();
684 if( i >= limit ) break;
685 }
686 Node* m = kit.map()->in(i);
687 Node* n = slow_map->in(i);
688 if (m != n) {
689 const Type* t = gvn.type(m)->meet(gvn.type(n));
690 Node* phi = PhiNode::make(region, m, t);
691 phi->set_req(2, n);
692 kit.map()->set_req(i, gvn.transform(phi));
693 }
694 }
695 return kit.transfer_exceptions_into_jvms();
696 }
699 CallGenerator* CallGenerator::for_method_handle_call(JVMState* jvms, ciMethod* caller, ciMethod* callee, bool delayed_forbidden) {
700 assert(callee->is_method_handle_intrinsic() ||
701 callee->is_compiled_lambda_form(), "for_method_handle_call mismatch");
702 bool input_not_const;
703 CallGenerator* cg = CallGenerator::for_method_handle_inline(jvms, caller, callee, input_not_const);
704 Compile* C = Compile::current();
705 if (cg != NULL) {
706 if (!delayed_forbidden && AlwaysIncrementalInline) {
707 return CallGenerator::for_late_inline(callee, cg);
708 } else {
709 return cg;
710 }
711 }
712 int bci = jvms->bci();
713 ciCallProfile profile = caller->call_profile_at_bci(bci);
714 int call_site_count = caller->scale_count(profile.count());
716 if (IncrementalInline && call_site_count > 0 &&
717 (input_not_const || !C->inlining_incrementally() || C->over_inlining_cutoff())) {
718 return CallGenerator::for_mh_late_inline(caller, callee, input_not_const);
719 } else {
720 return CallGenerator::for_direct_call(callee);
721 }
722 }
724 CallGenerator* CallGenerator::for_method_handle_inline(JVMState* jvms, ciMethod* caller, ciMethod* callee, bool& input_not_const) {
725 GraphKit kit(jvms);
726 PhaseGVN& gvn = kit.gvn();
727 Compile* C = kit.C;
728 vmIntrinsics::ID iid = callee->intrinsic_id();
729 input_not_const = true;
730 switch (iid) {
731 case vmIntrinsics::_invokeBasic:
732 {
733 // Get MethodHandle receiver:
734 Node* receiver = kit.argument(0);
735 if (receiver->Opcode() == Op_ConP) {
736 input_not_const = false;
737 const TypeOopPtr* oop_ptr = receiver->bottom_type()->is_oopptr();
738 ciMethod* target = oop_ptr->const_oop()->as_method_handle()->get_vmtarget();
739 guarantee(!target->is_method_handle_intrinsic(), "should not happen"); // XXX remove
740 const int vtable_index = Method::invalid_vtable_index;
741 CallGenerator* cg = C->call_generator(target, vtable_index, false, jvms, true, PROB_ALWAYS, true, true);
742 assert (!cg->is_late_inline() || cg->is_mh_late_inline(), "no late inline here");
743 if (cg != NULL && cg->is_inline())
744 return cg;
745 }
746 }
747 break;
749 case vmIntrinsics::_linkToVirtual:
750 case vmIntrinsics::_linkToStatic:
751 case vmIntrinsics::_linkToSpecial:
752 case vmIntrinsics::_linkToInterface:
753 {
754 // Get MemberName argument:
755 Node* member_name = kit.argument(callee->arg_size() - 1);
756 if (member_name->Opcode() == Op_ConP) {
757 input_not_const = false;
758 const TypeOopPtr* oop_ptr = member_name->bottom_type()->is_oopptr();
759 ciMethod* target = oop_ptr->const_oop()->as_member_name()->get_vmtarget();
761 // In lamda forms we erase signature types to avoid resolving issues
762 // involving class loaders. When we optimize a method handle invoke
763 // to a direct call we must cast the receiver and arguments to its
764 // actual types.
765 ciSignature* signature = target->signature();
766 const int receiver_skip = target->is_static() ? 0 : 1;
767 // Cast receiver to its type.
768 if (!target->is_static()) {
769 Node* arg = kit.argument(0);
770 const TypeOopPtr* arg_type = arg->bottom_type()->isa_oopptr();
771 const Type* sig_type = TypeOopPtr::make_from_klass(signature->accessing_klass());
772 if (arg_type != NULL && !arg_type->higher_equal(sig_type)) {
773 Node* cast_obj = gvn.transform(new (C) CheckCastPPNode(kit.control(), arg, sig_type));
774 kit.set_argument(0, cast_obj);
775 }
776 }
777 // Cast reference arguments to its type.
778 for (int i = 0; i < signature->count(); i++) {
779 ciType* t = signature->type_at(i);
780 if (t->is_klass()) {
781 Node* arg = kit.argument(receiver_skip + i);
782 const TypeOopPtr* arg_type = arg->bottom_type()->isa_oopptr();
783 const Type* sig_type = TypeOopPtr::make_from_klass(t->as_klass());
784 if (arg_type != NULL && !arg_type->higher_equal(sig_type)) {
785 Node* cast_obj = gvn.transform(new (C) CheckCastPPNode(kit.control(), arg, sig_type));
786 kit.set_argument(receiver_skip + i, cast_obj);
787 }
788 }
789 }
790 const int vtable_index = Method::invalid_vtable_index;
791 const bool call_is_virtual = target->is_abstract(); // FIXME workaround
792 CallGenerator* cg = C->call_generator(target, vtable_index, call_is_virtual, jvms, true, PROB_ALWAYS, true, true);
793 assert (!cg->is_late_inline() || cg->is_mh_late_inline(), "no late inline here");
794 if (cg != NULL && cg->is_inline())
795 return cg;
796 }
797 }
798 break;
800 default:
801 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
802 break;
803 }
804 return NULL;
805 }
808 //------------------------PredictedIntrinsicGenerator------------------------------
809 // Internal class which handles all predicted Intrinsic calls.
810 class PredictedIntrinsicGenerator : public CallGenerator {
811 CallGenerator* _intrinsic;
812 CallGenerator* _cg;
814 public:
815 PredictedIntrinsicGenerator(CallGenerator* intrinsic,
816 CallGenerator* cg)
817 : CallGenerator(cg->method())
818 {
819 _intrinsic = intrinsic;
820 _cg = cg;
821 }
823 virtual bool is_virtual() const { return true; }
824 virtual bool is_inlined() const { return true; }
825 virtual bool is_intrinsic() const { return true; }
827 virtual JVMState* generate(JVMState* jvms);
828 };
831 CallGenerator* CallGenerator::for_predicted_intrinsic(CallGenerator* intrinsic,
832 CallGenerator* cg) {
833 return new PredictedIntrinsicGenerator(intrinsic, cg);
834 }
837 JVMState* PredictedIntrinsicGenerator::generate(JVMState* jvms) {
838 GraphKit kit(jvms);
839 PhaseGVN& gvn = kit.gvn();
841 CompileLog* log = kit.C->log();
842 if (log != NULL) {
843 log->elem("predicted_intrinsic bci='%d' method='%d'",
844 jvms->bci(), log->identify(method()));
845 }
847 Node* slow_ctl = _intrinsic->generate_predicate(kit.sync_jvms());
848 if (kit.failing())
849 return NULL; // might happen because of NodeCountInliningCutoff
851 SafePointNode* slow_map = NULL;
852 JVMState* slow_jvms;
853 if (slow_ctl != NULL) {
854 PreserveJVMState pjvms(&kit);
855 kit.set_control(slow_ctl);
856 if (!kit.stopped()) {
857 slow_jvms = _cg->generate(kit.sync_jvms());
858 if (kit.failing())
859 return NULL; // might happen because of NodeCountInliningCutoff
860 assert(slow_jvms != NULL, "must be");
861 kit.add_exception_states_from(slow_jvms);
862 kit.set_map(slow_jvms->map());
863 if (!kit.stopped())
864 slow_map = kit.stop();
865 }
866 }
868 if (kit.stopped()) {
869 // Predicate is always false.
870 kit.set_jvms(slow_jvms);
871 return kit.transfer_exceptions_into_jvms();
872 }
874 // Generate intrinsic code:
875 JVMState* new_jvms = _intrinsic->generate(kit.sync_jvms());
876 if (new_jvms == NULL) {
877 // Intrinsic failed, so use slow code or make a direct call.
878 if (slow_map == NULL) {
879 CallGenerator* cg = CallGenerator::for_direct_call(method());
880 new_jvms = cg->generate(kit.sync_jvms());
881 } else {
882 kit.set_jvms(slow_jvms);
883 return kit.transfer_exceptions_into_jvms();
884 }
885 }
886 kit.add_exception_states_from(new_jvms);
887 kit.set_jvms(new_jvms);
889 // Need to merge slow and fast?
890 if (slow_map == NULL) {
891 // The fast path is the only path remaining.
892 return kit.transfer_exceptions_into_jvms();
893 }
895 if (kit.stopped()) {
896 // Intrinsic method threw an exception, so it's just the slow path after all.
897 kit.set_jvms(slow_jvms);
898 return kit.transfer_exceptions_into_jvms();
899 }
901 // Finish the diamond.
902 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization
903 RegionNode* region = new (kit.C) RegionNode(3);
904 region->init_req(1, kit.control());
905 region->init_req(2, slow_map->control());
906 kit.set_control(gvn.transform(region));
907 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO);
908 iophi->set_req(2, slow_map->i_o());
909 kit.set_i_o(gvn.transform(iophi));
910 kit.merge_memory(slow_map->merged_memory(), region, 2);
911 uint tos = kit.jvms()->stkoff() + kit.sp();
912 uint limit = slow_map->req();
913 for (uint i = TypeFunc::Parms; i < limit; i++) {
914 // Skip unused stack slots; fast forward to monoff();
915 if (i == tos) {
916 i = kit.jvms()->monoff();
917 if( i >= limit ) break;
918 }
919 Node* m = kit.map()->in(i);
920 Node* n = slow_map->in(i);
921 if (m != n) {
922 const Type* t = gvn.type(m)->meet(gvn.type(n));
923 Node* phi = PhiNode::make(region, m, t);
924 phi->set_req(2, n);
925 kit.map()->set_req(i, gvn.transform(phi));
926 }
927 }
928 return kit.transfer_exceptions_into_jvms();
929 }
931 //-------------------------UncommonTrapCallGenerator-----------------------------
932 // Internal class which handles all out-of-line calls checking receiver type.
933 class UncommonTrapCallGenerator : public CallGenerator {
934 Deoptimization::DeoptReason _reason;
935 Deoptimization::DeoptAction _action;
937 public:
938 UncommonTrapCallGenerator(ciMethod* m,
939 Deoptimization::DeoptReason reason,
940 Deoptimization::DeoptAction action)
941 : CallGenerator(m)
942 {
943 _reason = reason;
944 _action = action;
945 }
947 virtual bool is_virtual() const { ShouldNotReachHere(); return false; }
948 virtual bool is_trap() const { return true; }
950 virtual JVMState* generate(JVMState* jvms);
951 };
954 CallGenerator*
955 CallGenerator::for_uncommon_trap(ciMethod* m,
956 Deoptimization::DeoptReason reason,
957 Deoptimization::DeoptAction action) {
958 return new UncommonTrapCallGenerator(m, reason, action);
959 }
962 JVMState* UncommonTrapCallGenerator::generate(JVMState* jvms) {
963 GraphKit kit(jvms);
964 // Take the trap with arguments pushed on the stack. (Cf. null_check_receiver).
965 int nargs = method()->arg_size();
966 kit.inc_sp(nargs);
967 assert(nargs <= kit.sp() && kit.sp() <= jvms->stk_size(), "sane sp w/ args pushed");
968 if (_reason == Deoptimization::Reason_class_check &&
969 _action == Deoptimization::Action_maybe_recompile) {
970 // Temp fix for 6529811
971 // Don't allow uncommon_trap to override our decision to recompile in the event
972 // of a class cast failure for a monomorphic call as it will never let us convert
973 // the call to either bi-morphic or megamorphic and can lead to unc-trap loops
974 bool keep_exact_action = true;
975 kit.uncommon_trap(_reason, _action, NULL, "monomorphic vcall checkcast", false, keep_exact_action);
976 } else {
977 kit.uncommon_trap(_reason, _action);
978 }
979 return kit.transfer_exceptions_into_jvms();
980 }
982 // (Note: Moved hook_up_call to GraphKit::set_edges_for_java_call.)
984 // (Node: Merged hook_up_exits into ParseGenerator::generate.)
986 #define NODES_OVERHEAD_PER_METHOD (30.0)
987 #define NODES_PER_BYTECODE (9.5)
989 void WarmCallInfo::init(JVMState* call_site, ciMethod* call_method, ciCallProfile& profile, float prof_factor) {
990 int call_count = profile.count();
991 int code_size = call_method->code_size();
993 // Expected execution count is based on the historical count:
994 _count = call_count < 0 ? 1 : call_site->method()->scale_count(call_count, prof_factor);
996 // Expected profit from inlining, in units of simple call-overheads.
997 _profit = 1.0;
999 // Expected work performed by the call in units of call-overheads.
1000 // %%% need an empirical curve fit for "work" (time in call)
1001 float bytecodes_per_call = 3;
1002 _work = 1.0 + code_size / bytecodes_per_call;
1004 // Expected size of compilation graph:
1005 // -XX:+PrintParseStatistics once reported:
1006 // Methods seen: 9184 Methods parsed: 9184 Nodes created: 1582391
1007 // Histogram of 144298 parsed bytecodes:
1008 // %%% Need an better predictor for graph size.
1009 _size = NODES_OVERHEAD_PER_METHOD + (NODES_PER_BYTECODE * code_size);
1010 }
1012 // is_cold: Return true if the node should never be inlined.
1013 // This is true if any of the key metrics are extreme.
1014 bool WarmCallInfo::is_cold() const {
1015 if (count() < WarmCallMinCount) return true;
1016 if (profit() < WarmCallMinProfit) return true;
1017 if (work() > WarmCallMaxWork) return true;
1018 if (size() > WarmCallMaxSize) return true;
1019 return false;
1020 }
1022 // is_hot: Return true if the node should be inlined immediately.
1023 // This is true if any of the key metrics are extreme.
1024 bool WarmCallInfo::is_hot() const {
1025 assert(!is_cold(), "eliminate is_cold cases before testing is_hot");
1026 if (count() >= HotCallCountThreshold) return true;
1027 if (profit() >= HotCallProfitThreshold) return true;
1028 if (work() <= HotCallTrivialWork) return true;
1029 if (size() <= HotCallTrivialSize) return true;
1030 return false;
1031 }
1033 // compute_heat:
1034 float WarmCallInfo::compute_heat() const {
1035 assert(!is_cold(), "compute heat only on warm nodes");
1036 assert(!is_hot(), "compute heat only on warm nodes");
1037 int min_size = MAX2(0, (int)HotCallTrivialSize);
1038 int max_size = MIN2(500, (int)WarmCallMaxSize);
1039 float method_size = (size() - min_size) / MAX2(1, max_size - min_size);
1040 float size_factor;
1041 if (method_size < 0.05) size_factor = 4; // 2 sigmas better than avg.
1042 else if (method_size < 0.15) size_factor = 2; // 1 sigma better than avg.
1043 else if (method_size < 0.5) size_factor = 1; // better than avg.
1044 else size_factor = 0.5; // worse than avg.
1045 return (count() * profit() * size_factor);
1046 }
1048 bool WarmCallInfo::warmer_than(WarmCallInfo* that) {
1049 assert(this != that, "compare only different WCIs");
1050 assert(this->heat() != 0 && that->heat() != 0, "call compute_heat 1st");
1051 if (this->heat() > that->heat()) return true;
1052 if (this->heat() < that->heat()) return false;
1053 assert(this->heat() == that->heat(), "no NaN heat allowed");
1054 // Equal heat. Break the tie some other way.
1055 if (!this->call() || !that->call()) return (address)this > (address)that;
1056 return this->call()->_idx > that->call()->_idx;
1057 }
1059 //#define UNINIT_NEXT ((WarmCallInfo*)badAddress)
1060 #define UNINIT_NEXT ((WarmCallInfo*)NULL)
1062 WarmCallInfo* WarmCallInfo::insert_into(WarmCallInfo* head) {
1063 assert(next() == UNINIT_NEXT, "not yet on any list");
1064 WarmCallInfo* prev_p = NULL;
1065 WarmCallInfo* next_p = head;
1066 while (next_p != NULL && next_p->warmer_than(this)) {
1067 prev_p = next_p;
1068 next_p = prev_p->next();
1069 }
1070 // Install this between prev_p and next_p.
1071 this->set_next(next_p);
1072 if (prev_p == NULL)
1073 head = this;
1074 else
1075 prev_p->set_next(this);
1076 return head;
1077 }
1079 WarmCallInfo* WarmCallInfo::remove_from(WarmCallInfo* head) {
1080 WarmCallInfo* prev_p = NULL;
1081 WarmCallInfo* next_p = head;
1082 while (next_p != this) {
1083 assert(next_p != NULL, "this must be in the list somewhere");
1084 prev_p = next_p;
1085 next_p = prev_p->next();
1086 }
1087 next_p = this->next();
1088 debug_only(this->set_next(UNINIT_NEXT));
1089 // Remove this from between prev_p and next_p.
1090 if (prev_p == NULL)
1091 head = next_p;
1092 else
1093 prev_p->set_next(next_p);
1094 return head;
1095 }
1097 WarmCallInfo WarmCallInfo::_always_hot(WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE(),
1098 WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE());
1099 WarmCallInfo WarmCallInfo::_always_cold(WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE(),
1100 WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE());
1102 WarmCallInfo* WarmCallInfo::always_hot() {
1103 assert(_always_hot.is_hot(), "must always be hot");
1104 return &_always_hot;
1105 }
1107 WarmCallInfo* WarmCallInfo::always_cold() {
1108 assert(_always_cold.is_cold(), "must always be cold");
1109 return &_always_cold;
1110 }
1113 #ifndef PRODUCT
1115 void WarmCallInfo::print() const {
1116 tty->print("%s : C=%6.1f P=%6.1f W=%6.1f S=%6.1f H=%6.1f -> %p",
1117 is_cold() ? "cold" : is_hot() ? "hot " : "warm",
1118 count(), profit(), work(), size(), compute_heat(), next());
1119 tty->cr();
1120 if (call() != NULL) call()->dump();
1121 }
1123 void print_wci(WarmCallInfo* ci) {
1124 ci->print();
1125 }
1127 void WarmCallInfo::print_all() const {
1128 for (const WarmCallInfo* p = this; p != NULL; p = p->next())
1129 p->print();
1130 }
1132 int WarmCallInfo::count_all() const {
1133 int cnt = 0;
1134 for (const WarmCallInfo* p = this; p != NULL; p = p->next())
1135 cnt++;
1136 return cnt;
1137 }
1139 #endif //PRODUCT