Tue, 18 Dec 2012 14:55:25 +0100
8005031: Some cleanup in c2 to prepare for incremental inlining support
Summary: collection of small changes to prepare for incremental inlining.
Reviewed-by: twisti, kvn
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 CallGenerator* _inline_cg;
267 public:
268 LateInlineCallGenerator(ciMethod* method, CallGenerator* inline_cg) :
269 DirectCallGenerator(method, true), _inline_cg(inline_cg) {}
271 virtual bool is_late_inline() const { return true; }
273 // Convert the CallStaticJava into an inline
274 virtual void do_late_inline();
276 virtual JVMState* generate(JVMState* jvms) {
277 Compile *C = Compile::current();
278 C->print_inlining_skip(this);
280 // Record that this call site should be revisited once the main
281 // parse is finished.
282 Compile::current()->add_late_inline(this);
284 // Emit the CallStaticJava and request separate projections so
285 // that the late inlining logic can distinguish between fall
286 // through and exceptional uses of the memory and io projections
287 // as is done for allocations and macro expansion.
288 return DirectCallGenerator::generate(jvms);
289 }
290 };
293 void LateInlineCallGenerator::do_late_inline() {
294 // Can't inline it
295 if (call_node() == NULL || call_node()->outcnt() == 0 ||
296 call_node()->in(0) == NULL || call_node()->in(0)->is_top())
297 return;
299 CallStaticJavaNode* call = call_node();
301 // Make a clone of the JVMState that appropriate to use for driving a parse
302 Compile* C = Compile::current();
303 JVMState* jvms = call->jvms()->clone_shallow(C);
304 uint size = call->req();
305 SafePointNode* map = new (C) SafePointNode(size, jvms);
306 for (uint i1 = 0; i1 < size; i1++) {
307 map->init_req(i1, call->in(i1));
308 }
310 // Make sure the state is a MergeMem for parsing.
311 if (!map->in(TypeFunc::Memory)->is_MergeMem()) {
312 Node* mem = MergeMemNode::make(C, map->in(TypeFunc::Memory));
313 C->initial_gvn()->set_type_bottom(mem);
314 map->set_req(TypeFunc::Memory, mem);
315 }
317 // Make enough space for the expression stack and transfer the incoming arguments
318 int nargs = method()->arg_size();
319 jvms->set_map(map);
320 map->ensure_stack(jvms, jvms->method()->max_stack());
321 if (nargs > 0) {
322 for (int i1 = 0; i1 < nargs; i1++) {
323 map->set_req(i1 + jvms->argoff(), call->in(TypeFunc::Parms + i1));
324 }
325 }
327 C->print_inlining_insert(this);
329 CompileLog* log = C->log();
330 if (log != NULL) {
331 log->head("late_inline method='%d'", log->identify(method()));
332 JVMState* p = jvms;
333 while (p != NULL) {
334 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
335 p = p->caller();
336 }
337 log->tail("late_inline");
338 }
340 // Setup default node notes to be picked up by the inlining
341 Node_Notes* old_nn = C->default_node_notes();
342 if (old_nn != NULL) {
343 Node_Notes* entry_nn = old_nn->clone(C);
344 entry_nn->set_jvms(jvms);
345 C->set_default_node_notes(entry_nn);
346 }
348 // Now perform the inling using the synthesized JVMState
349 JVMState* new_jvms = _inline_cg->generate(jvms);
350 if (new_jvms == NULL) return; // no change
351 if (C->failing()) return;
353 // Capture any exceptional control flow
354 GraphKit kit(new_jvms);
356 // Find the result object
357 Node* result = C->top();
358 int result_size = method()->return_type()->size();
359 if (result_size != 0 && !kit.stopped()) {
360 result = (result_size == 1) ? kit.pop() : kit.pop_pair();
361 }
363 kit.replace_call(call, result);
364 }
367 CallGenerator* CallGenerator::for_late_inline(ciMethod* method, CallGenerator* inline_cg) {
368 return new LateInlineCallGenerator(method, inline_cg);
369 }
372 //---------------------------WarmCallGenerator--------------------------------
373 // Internal class which handles initial deferral of inlining decisions.
374 class WarmCallGenerator : public CallGenerator {
375 WarmCallInfo* _call_info;
376 CallGenerator* _if_cold;
377 CallGenerator* _if_hot;
378 bool _is_virtual; // caches virtuality of if_cold
379 bool _is_inline; // caches inline-ness of if_hot
381 public:
382 WarmCallGenerator(WarmCallInfo* ci,
383 CallGenerator* if_cold,
384 CallGenerator* if_hot)
385 : CallGenerator(if_cold->method())
386 {
387 assert(method() == if_hot->method(), "consistent choices");
388 _call_info = ci;
389 _if_cold = if_cold;
390 _if_hot = if_hot;
391 _is_virtual = if_cold->is_virtual();
392 _is_inline = if_hot->is_inline();
393 }
395 virtual bool is_inline() const { return _is_inline; }
396 virtual bool is_virtual() const { return _is_virtual; }
397 virtual bool is_deferred() const { return true; }
399 virtual JVMState* generate(JVMState* jvms);
400 };
403 CallGenerator* CallGenerator::for_warm_call(WarmCallInfo* ci,
404 CallGenerator* if_cold,
405 CallGenerator* if_hot) {
406 return new WarmCallGenerator(ci, if_cold, if_hot);
407 }
409 JVMState* WarmCallGenerator::generate(JVMState* jvms) {
410 Compile* C = Compile::current();
411 if (C->log() != NULL) {
412 C->log()->elem("warm_call bci='%d'", jvms->bci());
413 }
414 jvms = _if_cold->generate(jvms);
415 if (jvms != NULL) {
416 Node* m = jvms->map()->control();
417 if (m->is_CatchProj()) m = m->in(0); else m = C->top();
418 if (m->is_Catch()) m = m->in(0); else m = C->top();
419 if (m->is_Proj()) m = m->in(0); else m = C->top();
420 if (m->is_CallJava()) {
421 _call_info->set_call(m->as_Call());
422 _call_info->set_hot_cg(_if_hot);
423 #ifndef PRODUCT
424 if (PrintOpto || PrintOptoInlining) {
425 tty->print_cr("Queueing for warm inlining at bci %d:", jvms->bci());
426 tty->print("WCI: ");
427 _call_info->print();
428 }
429 #endif
430 _call_info->set_heat(_call_info->compute_heat());
431 C->set_warm_calls(_call_info->insert_into(C->warm_calls()));
432 }
433 }
434 return jvms;
435 }
437 void WarmCallInfo::make_hot() {
438 Unimplemented();
439 }
441 void WarmCallInfo::make_cold() {
442 // No action: Just dequeue.
443 }
446 //------------------------PredictedCallGenerator------------------------------
447 // Internal class which handles all out-of-line calls checking receiver type.
448 class PredictedCallGenerator : public CallGenerator {
449 ciKlass* _predicted_receiver;
450 CallGenerator* _if_missed;
451 CallGenerator* _if_hit;
452 float _hit_prob;
454 public:
455 PredictedCallGenerator(ciKlass* predicted_receiver,
456 CallGenerator* if_missed,
457 CallGenerator* if_hit, float hit_prob)
458 : CallGenerator(if_missed->method())
459 {
460 // The call profile data may predict the hit_prob as extreme as 0 or 1.
461 // Remove the extremes values from the range.
462 if (hit_prob > PROB_MAX) hit_prob = PROB_MAX;
463 if (hit_prob < PROB_MIN) hit_prob = PROB_MIN;
465 _predicted_receiver = predicted_receiver;
466 _if_missed = if_missed;
467 _if_hit = if_hit;
468 _hit_prob = hit_prob;
469 }
471 virtual bool is_virtual() const { return true; }
472 virtual bool is_inline() const { return _if_hit->is_inline(); }
473 virtual bool is_deferred() const { return _if_hit->is_deferred(); }
475 virtual JVMState* generate(JVMState* jvms);
476 };
479 CallGenerator* CallGenerator::for_predicted_call(ciKlass* predicted_receiver,
480 CallGenerator* if_missed,
481 CallGenerator* if_hit,
482 float hit_prob) {
483 return new PredictedCallGenerator(predicted_receiver, if_missed, if_hit, hit_prob);
484 }
487 JVMState* PredictedCallGenerator::generate(JVMState* jvms) {
488 GraphKit kit(jvms);
489 PhaseGVN& gvn = kit.gvn();
490 // We need an explicit receiver null_check before checking its type.
491 // We share a map with the caller, so his JVMS gets adjusted.
492 Node* receiver = kit.argument(0);
494 CompileLog* log = kit.C->log();
495 if (log != NULL) {
496 log->elem("predicted_call bci='%d' klass='%d'",
497 jvms->bci(), log->identify(_predicted_receiver));
498 }
500 receiver = kit.null_check_receiver_before_call(method());
501 if (kit.stopped()) {
502 return kit.transfer_exceptions_into_jvms();
503 }
505 Node* exact_receiver = receiver; // will get updated in place...
506 Node* slow_ctl = kit.type_check_receiver(receiver,
507 _predicted_receiver, _hit_prob,
508 &exact_receiver);
510 SafePointNode* slow_map = NULL;
511 JVMState* slow_jvms;
512 { PreserveJVMState pjvms(&kit);
513 kit.set_control(slow_ctl);
514 if (!kit.stopped()) {
515 slow_jvms = _if_missed->generate(kit.sync_jvms());
516 if (kit.failing())
517 return NULL; // might happen because of NodeCountInliningCutoff
518 assert(slow_jvms != NULL, "must be");
519 kit.add_exception_states_from(slow_jvms);
520 kit.set_map(slow_jvms->map());
521 if (!kit.stopped())
522 slow_map = kit.stop();
523 }
524 }
526 if (kit.stopped()) {
527 // Instance exactly does not matches the desired type.
528 kit.set_jvms(slow_jvms);
529 return kit.transfer_exceptions_into_jvms();
530 }
532 // fall through if the instance exactly matches the desired type
533 kit.replace_in_map(receiver, exact_receiver);
535 // Make the hot call:
536 JVMState* new_jvms = _if_hit->generate(kit.sync_jvms());
537 if (new_jvms == NULL) {
538 // Inline failed, so make a direct call.
539 assert(_if_hit->is_inline(), "must have been a failed inline");
540 CallGenerator* cg = CallGenerator::for_direct_call(_if_hit->method());
541 new_jvms = cg->generate(kit.sync_jvms());
542 }
543 kit.add_exception_states_from(new_jvms);
544 kit.set_jvms(new_jvms);
546 // Need to merge slow and fast?
547 if (slow_map == NULL) {
548 // The fast path is the only path remaining.
549 return kit.transfer_exceptions_into_jvms();
550 }
552 if (kit.stopped()) {
553 // Inlined method threw an exception, so it's just the slow path after all.
554 kit.set_jvms(slow_jvms);
555 return kit.transfer_exceptions_into_jvms();
556 }
558 // Finish the diamond.
559 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization
560 RegionNode* region = new (kit.C) RegionNode(3);
561 region->init_req(1, kit.control());
562 region->init_req(2, slow_map->control());
563 kit.set_control(gvn.transform(region));
564 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO);
565 iophi->set_req(2, slow_map->i_o());
566 kit.set_i_o(gvn.transform(iophi));
567 kit.merge_memory(slow_map->merged_memory(), region, 2);
568 uint tos = kit.jvms()->stkoff() + kit.sp();
569 uint limit = slow_map->req();
570 for (uint i = TypeFunc::Parms; i < limit; i++) {
571 // Skip unused stack slots; fast forward to monoff();
572 if (i == tos) {
573 i = kit.jvms()->monoff();
574 if( i >= limit ) break;
575 }
576 Node* m = kit.map()->in(i);
577 Node* n = slow_map->in(i);
578 if (m != n) {
579 const Type* t = gvn.type(m)->meet(gvn.type(n));
580 Node* phi = PhiNode::make(region, m, t);
581 phi->set_req(2, n);
582 kit.map()->set_req(i, gvn.transform(phi));
583 }
584 }
585 return kit.transfer_exceptions_into_jvms();
586 }
589 CallGenerator* CallGenerator::for_method_handle_call(JVMState* jvms, ciMethod* caller, ciMethod* callee) {
590 assert(callee->is_method_handle_intrinsic() ||
591 callee->is_compiled_lambda_form(), "for_method_handle_call mismatch");
592 CallGenerator* cg = CallGenerator::for_method_handle_inline(jvms, caller, callee);
593 if (cg != NULL)
594 return cg;
595 return CallGenerator::for_direct_call(callee);
596 }
598 CallGenerator* CallGenerator::for_method_handle_inline(JVMState* jvms, ciMethod* caller, ciMethod* callee) {
599 GraphKit kit(jvms);
600 PhaseGVN& gvn = kit.gvn();
601 Compile* C = kit.C;
602 vmIntrinsics::ID iid = callee->intrinsic_id();
603 switch (iid) {
604 case vmIntrinsics::_invokeBasic:
605 {
606 // Get MethodHandle receiver:
607 Node* receiver = kit.argument(0);
608 if (receiver->Opcode() == Op_ConP) {
609 const TypeOopPtr* oop_ptr = receiver->bottom_type()->is_oopptr();
610 ciMethod* target = oop_ptr->const_oop()->as_method_handle()->get_vmtarget();
611 guarantee(!target->is_method_handle_intrinsic(), "should not happen"); // XXX remove
612 const int vtable_index = Method::invalid_vtable_index;
613 CallGenerator* cg = C->call_generator(target, vtable_index, false, jvms, true, PROB_ALWAYS);
614 if (cg != NULL && cg->is_inline())
615 return cg;
616 } else {
617 if (PrintInlining) C->print_inlining(callee, jvms->depth() - 1, jvms->bci(), "receiver not constant");
618 }
619 }
620 break;
622 case vmIntrinsics::_linkToVirtual:
623 case vmIntrinsics::_linkToStatic:
624 case vmIntrinsics::_linkToSpecial:
625 case vmIntrinsics::_linkToInterface:
626 {
627 // Get MemberName argument:
628 Node* member_name = kit.argument(callee->arg_size() - 1);
629 if (member_name->Opcode() == Op_ConP) {
630 const TypeOopPtr* oop_ptr = member_name->bottom_type()->is_oopptr();
631 ciMethod* target = oop_ptr->const_oop()->as_member_name()->get_vmtarget();
633 // In lamda forms we erase signature types to avoid resolving issues
634 // involving class loaders. When we optimize a method handle invoke
635 // to a direct call we must cast the receiver and arguments to its
636 // actual types.
637 ciSignature* signature = target->signature();
638 const int receiver_skip = target->is_static() ? 0 : 1;
639 // Cast receiver to its type.
640 if (!target->is_static()) {
641 Node* arg = kit.argument(0);
642 const TypeOopPtr* arg_type = arg->bottom_type()->isa_oopptr();
643 const Type* sig_type = TypeOopPtr::make_from_klass(signature->accessing_klass());
644 if (arg_type != NULL && !arg_type->higher_equal(sig_type)) {
645 Node* cast_obj = gvn.transform(new (C) CheckCastPPNode(kit.control(), arg, sig_type));
646 kit.set_argument(0, cast_obj);
647 }
648 }
649 // Cast reference arguments to its type.
650 for (int i = 0; i < signature->count(); i++) {
651 ciType* t = signature->type_at(i);
652 if (t->is_klass()) {
653 Node* arg = kit.argument(receiver_skip + i);
654 const TypeOopPtr* arg_type = arg->bottom_type()->isa_oopptr();
655 const Type* sig_type = TypeOopPtr::make_from_klass(t->as_klass());
656 if (arg_type != NULL && !arg_type->higher_equal(sig_type)) {
657 Node* cast_obj = gvn.transform(new (C) CheckCastPPNode(kit.control(), arg, sig_type));
658 kit.set_argument(receiver_skip + i, cast_obj);
659 }
660 }
661 }
662 const int vtable_index = Method::invalid_vtable_index;
663 const bool call_is_virtual = target->is_abstract(); // FIXME workaround
664 CallGenerator* cg = C->call_generator(target, vtable_index, call_is_virtual, jvms, true, PROB_ALWAYS);
665 if (cg != NULL && cg->is_inline())
666 return cg;
667 }
668 }
669 break;
671 default:
672 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
673 break;
674 }
675 return NULL;
676 }
679 //------------------------PredictedIntrinsicGenerator------------------------------
680 // Internal class which handles all predicted Intrinsic calls.
681 class PredictedIntrinsicGenerator : public CallGenerator {
682 CallGenerator* _intrinsic;
683 CallGenerator* _cg;
685 public:
686 PredictedIntrinsicGenerator(CallGenerator* intrinsic,
687 CallGenerator* cg)
688 : CallGenerator(cg->method())
689 {
690 _intrinsic = intrinsic;
691 _cg = cg;
692 }
694 virtual bool is_virtual() const { return true; }
695 virtual bool is_inlined() const { return true; }
696 virtual bool is_intrinsic() const { return true; }
698 virtual JVMState* generate(JVMState* jvms);
699 };
702 CallGenerator* CallGenerator::for_predicted_intrinsic(CallGenerator* intrinsic,
703 CallGenerator* cg) {
704 return new PredictedIntrinsicGenerator(intrinsic, cg);
705 }
708 JVMState* PredictedIntrinsicGenerator::generate(JVMState* jvms) {
709 GraphKit kit(jvms);
710 PhaseGVN& gvn = kit.gvn();
712 CompileLog* log = kit.C->log();
713 if (log != NULL) {
714 log->elem("predicted_intrinsic bci='%d' method='%d'",
715 jvms->bci(), log->identify(method()));
716 }
718 Node* slow_ctl = _intrinsic->generate_predicate(kit.sync_jvms());
719 if (kit.failing())
720 return NULL; // might happen because of NodeCountInliningCutoff
722 SafePointNode* slow_map = NULL;
723 JVMState* slow_jvms;
724 if (slow_ctl != NULL) {
725 PreserveJVMState pjvms(&kit);
726 kit.set_control(slow_ctl);
727 if (!kit.stopped()) {
728 slow_jvms = _cg->generate(kit.sync_jvms());
729 if (kit.failing())
730 return NULL; // might happen because of NodeCountInliningCutoff
731 assert(slow_jvms != NULL, "must be");
732 kit.add_exception_states_from(slow_jvms);
733 kit.set_map(slow_jvms->map());
734 if (!kit.stopped())
735 slow_map = kit.stop();
736 }
737 }
739 if (kit.stopped()) {
740 // Predicate is always false.
741 kit.set_jvms(slow_jvms);
742 return kit.transfer_exceptions_into_jvms();
743 }
745 // Generate intrinsic code:
746 JVMState* new_jvms = _intrinsic->generate(kit.sync_jvms());
747 if (new_jvms == NULL) {
748 // Intrinsic failed, so use slow code or make a direct call.
749 if (slow_map == NULL) {
750 CallGenerator* cg = CallGenerator::for_direct_call(method());
751 new_jvms = cg->generate(kit.sync_jvms());
752 } else {
753 kit.set_jvms(slow_jvms);
754 return kit.transfer_exceptions_into_jvms();
755 }
756 }
757 kit.add_exception_states_from(new_jvms);
758 kit.set_jvms(new_jvms);
760 // Need to merge slow and fast?
761 if (slow_map == NULL) {
762 // The fast path is the only path remaining.
763 return kit.transfer_exceptions_into_jvms();
764 }
766 if (kit.stopped()) {
767 // Intrinsic method threw an exception, so it's just the slow path after all.
768 kit.set_jvms(slow_jvms);
769 return kit.transfer_exceptions_into_jvms();
770 }
772 // Finish the diamond.
773 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization
774 RegionNode* region = new (kit.C) RegionNode(3);
775 region->init_req(1, kit.control());
776 region->init_req(2, slow_map->control());
777 kit.set_control(gvn.transform(region));
778 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO);
779 iophi->set_req(2, slow_map->i_o());
780 kit.set_i_o(gvn.transform(iophi));
781 kit.merge_memory(slow_map->merged_memory(), region, 2);
782 uint tos = kit.jvms()->stkoff() + kit.sp();
783 uint limit = slow_map->req();
784 for (uint i = TypeFunc::Parms; i < limit; i++) {
785 // Skip unused stack slots; fast forward to monoff();
786 if (i == tos) {
787 i = kit.jvms()->monoff();
788 if( i >= limit ) break;
789 }
790 Node* m = kit.map()->in(i);
791 Node* n = slow_map->in(i);
792 if (m != n) {
793 const Type* t = gvn.type(m)->meet(gvn.type(n));
794 Node* phi = PhiNode::make(region, m, t);
795 phi->set_req(2, n);
796 kit.map()->set_req(i, gvn.transform(phi));
797 }
798 }
799 return kit.transfer_exceptions_into_jvms();
800 }
802 //-------------------------UncommonTrapCallGenerator-----------------------------
803 // Internal class which handles all out-of-line calls checking receiver type.
804 class UncommonTrapCallGenerator : public CallGenerator {
805 Deoptimization::DeoptReason _reason;
806 Deoptimization::DeoptAction _action;
808 public:
809 UncommonTrapCallGenerator(ciMethod* m,
810 Deoptimization::DeoptReason reason,
811 Deoptimization::DeoptAction action)
812 : CallGenerator(m)
813 {
814 _reason = reason;
815 _action = action;
816 }
818 virtual bool is_virtual() const { ShouldNotReachHere(); return false; }
819 virtual bool is_trap() const { return true; }
821 virtual JVMState* generate(JVMState* jvms);
822 };
825 CallGenerator*
826 CallGenerator::for_uncommon_trap(ciMethod* m,
827 Deoptimization::DeoptReason reason,
828 Deoptimization::DeoptAction action) {
829 return new UncommonTrapCallGenerator(m, reason, action);
830 }
833 JVMState* UncommonTrapCallGenerator::generate(JVMState* jvms) {
834 GraphKit kit(jvms);
835 // Take the trap with arguments pushed on the stack. (Cf. null_check_receiver).
836 int nargs = method()->arg_size();
837 kit.inc_sp(nargs);
838 assert(nargs <= kit.sp() && kit.sp() <= jvms->stk_size(), "sane sp w/ args pushed");
839 if (_reason == Deoptimization::Reason_class_check &&
840 _action == Deoptimization::Action_maybe_recompile) {
841 // Temp fix for 6529811
842 // Don't allow uncommon_trap to override our decision to recompile in the event
843 // of a class cast failure for a monomorphic call as it will never let us convert
844 // the call to either bi-morphic or megamorphic and can lead to unc-trap loops
845 bool keep_exact_action = true;
846 kit.uncommon_trap(_reason, _action, NULL, "monomorphic vcall checkcast", false, keep_exact_action);
847 } else {
848 kit.uncommon_trap(_reason, _action);
849 }
850 return kit.transfer_exceptions_into_jvms();
851 }
853 // (Note: Moved hook_up_call to GraphKit::set_edges_for_java_call.)
855 // (Node: Merged hook_up_exits into ParseGenerator::generate.)
857 #define NODES_OVERHEAD_PER_METHOD (30.0)
858 #define NODES_PER_BYTECODE (9.5)
860 void WarmCallInfo::init(JVMState* call_site, ciMethod* call_method, ciCallProfile& profile, float prof_factor) {
861 int call_count = profile.count();
862 int code_size = call_method->code_size();
864 // Expected execution count is based on the historical count:
865 _count = call_count < 0 ? 1 : call_site->method()->scale_count(call_count, prof_factor);
867 // Expected profit from inlining, in units of simple call-overheads.
868 _profit = 1.0;
870 // Expected work performed by the call in units of call-overheads.
871 // %%% need an empirical curve fit for "work" (time in call)
872 float bytecodes_per_call = 3;
873 _work = 1.0 + code_size / bytecodes_per_call;
875 // Expected size of compilation graph:
876 // -XX:+PrintParseStatistics once reported:
877 // Methods seen: 9184 Methods parsed: 9184 Nodes created: 1582391
878 // Histogram of 144298 parsed bytecodes:
879 // %%% Need an better predictor for graph size.
880 _size = NODES_OVERHEAD_PER_METHOD + (NODES_PER_BYTECODE * code_size);
881 }
883 // is_cold: Return true if the node should never be inlined.
884 // This is true if any of the key metrics are extreme.
885 bool WarmCallInfo::is_cold() const {
886 if (count() < WarmCallMinCount) return true;
887 if (profit() < WarmCallMinProfit) return true;
888 if (work() > WarmCallMaxWork) return true;
889 if (size() > WarmCallMaxSize) return true;
890 return false;
891 }
893 // is_hot: Return true if the node should be inlined immediately.
894 // This is true if any of the key metrics are extreme.
895 bool WarmCallInfo::is_hot() const {
896 assert(!is_cold(), "eliminate is_cold cases before testing is_hot");
897 if (count() >= HotCallCountThreshold) return true;
898 if (profit() >= HotCallProfitThreshold) return true;
899 if (work() <= HotCallTrivialWork) return true;
900 if (size() <= HotCallTrivialSize) return true;
901 return false;
902 }
904 // compute_heat:
905 float WarmCallInfo::compute_heat() const {
906 assert(!is_cold(), "compute heat only on warm nodes");
907 assert(!is_hot(), "compute heat only on warm nodes");
908 int min_size = MAX2(0, (int)HotCallTrivialSize);
909 int max_size = MIN2(500, (int)WarmCallMaxSize);
910 float method_size = (size() - min_size) / MAX2(1, max_size - min_size);
911 float size_factor;
912 if (method_size < 0.05) size_factor = 4; // 2 sigmas better than avg.
913 else if (method_size < 0.15) size_factor = 2; // 1 sigma better than avg.
914 else if (method_size < 0.5) size_factor = 1; // better than avg.
915 else size_factor = 0.5; // worse than avg.
916 return (count() * profit() * size_factor);
917 }
919 bool WarmCallInfo::warmer_than(WarmCallInfo* that) {
920 assert(this != that, "compare only different WCIs");
921 assert(this->heat() != 0 && that->heat() != 0, "call compute_heat 1st");
922 if (this->heat() > that->heat()) return true;
923 if (this->heat() < that->heat()) return false;
924 assert(this->heat() == that->heat(), "no NaN heat allowed");
925 // Equal heat. Break the tie some other way.
926 if (!this->call() || !that->call()) return (address)this > (address)that;
927 return this->call()->_idx > that->call()->_idx;
928 }
930 //#define UNINIT_NEXT ((WarmCallInfo*)badAddress)
931 #define UNINIT_NEXT ((WarmCallInfo*)NULL)
933 WarmCallInfo* WarmCallInfo::insert_into(WarmCallInfo* head) {
934 assert(next() == UNINIT_NEXT, "not yet on any list");
935 WarmCallInfo* prev_p = NULL;
936 WarmCallInfo* next_p = head;
937 while (next_p != NULL && next_p->warmer_than(this)) {
938 prev_p = next_p;
939 next_p = prev_p->next();
940 }
941 // Install this between prev_p and next_p.
942 this->set_next(next_p);
943 if (prev_p == NULL)
944 head = this;
945 else
946 prev_p->set_next(this);
947 return head;
948 }
950 WarmCallInfo* WarmCallInfo::remove_from(WarmCallInfo* head) {
951 WarmCallInfo* prev_p = NULL;
952 WarmCallInfo* next_p = head;
953 while (next_p != this) {
954 assert(next_p != NULL, "this must be in the list somewhere");
955 prev_p = next_p;
956 next_p = prev_p->next();
957 }
958 next_p = this->next();
959 debug_only(this->set_next(UNINIT_NEXT));
960 // Remove this from between prev_p and next_p.
961 if (prev_p == NULL)
962 head = next_p;
963 else
964 prev_p->set_next(next_p);
965 return head;
966 }
968 WarmCallInfo WarmCallInfo::_always_hot(WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE(),
969 WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE());
970 WarmCallInfo WarmCallInfo::_always_cold(WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE(),
971 WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE());
973 WarmCallInfo* WarmCallInfo::always_hot() {
974 assert(_always_hot.is_hot(), "must always be hot");
975 return &_always_hot;
976 }
978 WarmCallInfo* WarmCallInfo::always_cold() {
979 assert(_always_cold.is_cold(), "must always be cold");
980 return &_always_cold;
981 }
984 #ifndef PRODUCT
986 void WarmCallInfo::print() const {
987 tty->print("%s : C=%6.1f P=%6.1f W=%6.1f S=%6.1f H=%6.1f -> %p",
988 is_cold() ? "cold" : is_hot() ? "hot " : "warm",
989 count(), profit(), work(), size(), compute_heat(), next());
990 tty->cr();
991 if (call() != NULL) call()->dump();
992 }
994 void print_wci(WarmCallInfo* ci) {
995 ci->print();
996 }
998 void WarmCallInfo::print_all() const {
999 for (const WarmCallInfo* p = this; p != NULL; p = p->next())
1000 p->print();
1001 }
1003 int WarmCallInfo::count_all() const {
1004 int cnt = 0;
1005 for (const WarmCallInfo* p = this; p != NULL; p = p->next())
1006 cnt++;
1007 return cnt;
1008 }
1010 #endif //PRODUCT