Mon, 26 Sep 2011 10:24:05 -0700
7081933: Use zeroing elimination optimization for large array
Summary: Don't zero new typeArray during runtime call if the allocation is followed by arraycopy into it.
Reviewed-by: twisti
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/ciCallSite.hpp"
28 #include "ci/ciCPCache.hpp"
29 #include "ci/ciMethodHandle.hpp"
30 #include "classfile/javaClasses.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/callGenerator.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/connode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "opto/subnode.hpp"
42 CallGenerator::CallGenerator(ciMethod* method) {
43 _method = method;
44 }
46 // Utility function.
47 const TypeFunc* CallGenerator::tf() const {
48 return TypeFunc::make(method());
49 }
51 //-----------------------------ParseGenerator---------------------------------
52 // Internal class which handles all direct bytecode traversal.
53 class ParseGenerator : public InlineCallGenerator {
54 private:
55 bool _is_osr;
56 float _expected_uses;
58 public:
59 ParseGenerator(ciMethod* method, float expected_uses, bool is_osr = false)
60 : InlineCallGenerator(method)
61 {
62 _is_osr = is_osr;
63 _expected_uses = expected_uses;
64 assert(InlineTree::check_can_parse(method) == NULL, "parse must be possible");
65 }
67 virtual bool is_parse() const { return true; }
68 virtual JVMState* generate(JVMState* jvms);
69 int is_osr() { return _is_osr; }
71 };
73 JVMState* ParseGenerator::generate(JVMState* jvms) {
74 Compile* C = Compile::current();
76 if (is_osr()) {
77 // The JVMS for a OSR has a single argument (see its TypeFunc).
78 assert(jvms->depth() == 1, "no inline OSR");
79 }
81 if (C->failing()) {
82 return NULL; // bailing out of the compile; do not try to parse
83 }
85 Parse parser(jvms, method(), _expected_uses);
86 // Grab signature for matching/allocation
87 #ifdef ASSERT
88 if (parser.tf() != (parser.depth() == 1 ? C->tf() : tf())) {
89 MutexLockerEx ml(Compile_lock, Mutex::_no_safepoint_check_flag);
90 assert(C->env()->system_dictionary_modification_counter_changed(),
91 "Must invalidate if TypeFuncs differ");
92 }
93 #endif
95 GraphKit& exits = parser.exits();
97 if (C->failing()) {
98 while (exits.pop_exception_state() != NULL) ;
99 return NULL;
100 }
102 assert(exits.jvms()->same_calls_as(jvms), "sanity");
104 // Simply return the exit state of the parser,
105 // augmented by any exceptional states.
106 return exits.transfer_exceptions_into_jvms();
107 }
109 //---------------------------DirectCallGenerator------------------------------
110 // Internal class which handles all out-of-line calls w/o receiver type checks.
111 class DirectCallGenerator : public CallGenerator {
112 private:
113 CallStaticJavaNode* _call_node;
114 // Force separate memory and I/O projections for the exceptional
115 // paths to facilitate late inlinig.
116 bool _separate_io_proj;
118 public:
119 DirectCallGenerator(ciMethod* method, bool separate_io_proj)
120 : CallGenerator(method),
121 _separate_io_proj(separate_io_proj)
122 {
123 }
124 virtual JVMState* generate(JVMState* jvms);
126 CallStaticJavaNode* call_node() const { return _call_node; }
127 };
129 JVMState* DirectCallGenerator::generate(JVMState* jvms) {
130 GraphKit kit(jvms);
131 bool is_static = method()->is_static();
132 address target = is_static ? SharedRuntime::get_resolve_static_call_stub()
133 : SharedRuntime::get_resolve_opt_virtual_call_stub();
135 if (kit.C->log() != NULL) {
136 kit.C->log()->elem("direct_call bci='%d'", jvms->bci());
137 }
139 CallStaticJavaNode *call = new (kit.C, tf()->domain()->cnt()) CallStaticJavaNode(tf(), target, method(), kit.bci());
140 if (!is_static) {
141 // Make an explicit receiver null_check as part of this call.
142 // Since we share a map with the caller, his JVMS gets adjusted.
143 kit.null_check_receiver(method());
144 if (kit.stopped()) {
145 // And dump it back to the caller, decorated with any exceptions:
146 return kit.transfer_exceptions_into_jvms();
147 }
148 // Mark the call node as virtual, sort of:
149 call->set_optimized_virtual(true);
150 if (method()->is_method_handle_invoke()) {
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 _call_node = call; // Save the call node in case we need it later
159 return kit.transfer_exceptions_into_jvms();
160 }
162 //---------------------------DynamicCallGenerator-----------------------------
163 // Internal class which handles all out-of-line invokedynamic calls.
164 class DynamicCallGenerator : public CallGenerator {
165 public:
166 DynamicCallGenerator(ciMethod* method)
167 : CallGenerator(method)
168 {
169 }
170 virtual JVMState* generate(JVMState* jvms);
171 };
173 JVMState* DynamicCallGenerator::generate(JVMState* jvms) {
174 GraphKit kit(jvms);
176 if (kit.C->log() != NULL) {
177 kit.C->log()->elem("dynamic_call bci='%d'", jvms->bci());
178 }
180 // Get the constant pool cache from the caller class.
181 ciMethod* caller_method = jvms->method();
182 ciBytecodeStream str(caller_method);
183 str.force_bci(jvms->bci()); // Set the stream to the invokedynamic bci.
184 assert(str.cur_bc() == Bytecodes::_invokedynamic, "wrong place to issue a dynamic call!");
185 ciCPCache* cpcache = str.get_cpcache();
187 // Get the offset of the CallSite from the constant pool cache
188 // pointer.
189 int index = str.get_method_index();
190 size_t call_site_offset = cpcache->get_f1_offset(index);
192 // Load the CallSite object from the constant pool cache.
193 const TypeOopPtr* cpcache_ptr = TypeOopPtr::make_from_constant(cpcache);
194 Node* cpcache_adr = kit.makecon(cpcache_ptr);
195 Node* call_site_adr = kit.basic_plus_adr(cpcache_adr, cpcache_adr, call_site_offset);
196 Node* call_site = kit.make_load(kit.control(), call_site_adr, TypeInstPtr::BOTTOM, T_OBJECT, Compile::AliasIdxRaw);
198 // Load the target MethodHandle from the CallSite object.
199 Node* target_mh_adr = kit.basic_plus_adr(call_site, call_site, java_lang_invoke_CallSite::target_offset_in_bytes());
200 Node* target_mh = kit.make_load(kit.control(), target_mh_adr, TypeInstPtr::BOTTOM, T_OBJECT);
202 address resolve_stub = SharedRuntime::get_resolve_opt_virtual_call_stub();
204 CallStaticJavaNode *call = new (kit.C, tf()->domain()->cnt()) CallStaticJavaNode(tf(), resolve_stub, method(), kit.bci());
205 // invokedynamic is treated as an optimized invokevirtual.
206 call->set_optimized_virtual(true);
207 // Take extra care (in the presence of argument motion) not to trash the SP:
208 call->set_method_handle_invoke(true);
210 // Pass the target MethodHandle as first argument and shift the
211 // other arguments.
212 call->init_req(0 + TypeFunc::Parms, target_mh);
213 uint nargs = call->method()->arg_size();
214 for (uint i = 1; i < nargs; i++) {
215 Node* arg = kit.argument(i - 1);
216 call->init_req(i + TypeFunc::Parms, arg);
217 }
219 kit.set_edges_for_java_call(call);
220 Node* ret = kit.set_results_for_java_call(call);
221 kit.push_node(method()->return_type()->basic_type(), ret);
222 return kit.transfer_exceptions_into_jvms();
223 }
225 //--------------------------VirtualCallGenerator------------------------------
226 // Internal class which handles all out-of-line calls checking receiver type.
227 class VirtualCallGenerator : public CallGenerator {
228 private:
229 int _vtable_index;
230 public:
231 VirtualCallGenerator(ciMethod* method, int vtable_index)
232 : CallGenerator(method), _vtable_index(vtable_index)
233 {
234 assert(vtable_index == methodOopDesc::invalid_vtable_index ||
235 vtable_index >= 0, "either invalid or usable");
236 }
237 virtual bool is_virtual() const { return true; }
238 virtual JVMState* generate(JVMState* jvms);
239 };
241 JVMState* VirtualCallGenerator::generate(JVMState* jvms) {
242 GraphKit kit(jvms);
243 Node* receiver = kit.argument(0);
245 if (kit.C->log() != NULL) {
246 kit.C->log()->elem("virtual_call bci='%d'", jvms->bci());
247 }
249 // If the receiver is a constant null, do not torture the system
250 // by attempting to call through it. The compile will proceed
251 // correctly, but may bail out in final_graph_reshaping, because
252 // the call instruction will have a seemingly deficient out-count.
253 // (The bailout says something misleading about an "infinite loop".)
254 if (kit.gvn().type(receiver)->higher_equal(TypePtr::NULL_PTR)) {
255 kit.inc_sp(method()->arg_size()); // restore arguments
256 kit.uncommon_trap(Deoptimization::Reason_null_check,
257 Deoptimization::Action_none,
258 NULL, "null receiver");
259 return kit.transfer_exceptions_into_jvms();
260 }
262 // Ideally we would unconditionally do a null check here and let it
263 // be converted to an implicit check based on profile information.
264 // However currently the conversion to implicit null checks in
265 // Block::implicit_null_check() only looks for loads and stores, not calls.
266 ciMethod *caller = kit.method();
267 ciMethodData *caller_md = (caller == NULL) ? NULL : caller->method_data();
268 if (!UseInlineCaches || !ImplicitNullChecks ||
269 ((ImplicitNullCheckThreshold > 0) && caller_md &&
270 (caller_md->trap_count(Deoptimization::Reason_null_check)
271 >= (uint)ImplicitNullCheckThreshold))) {
272 // Make an explicit receiver null_check as part of this call.
273 // Since we share a map with the caller, his JVMS gets adjusted.
274 receiver = kit.null_check_receiver(method());
275 if (kit.stopped()) {
276 // And dump it back to the caller, decorated with any exceptions:
277 return kit.transfer_exceptions_into_jvms();
278 }
279 }
281 assert(!method()->is_static(), "virtual call must not be to static");
282 assert(!method()->is_final(), "virtual call should not be to final");
283 assert(!method()->is_private(), "virtual call should not be to private");
284 assert(_vtable_index == methodOopDesc::invalid_vtable_index || !UseInlineCaches,
285 "no vtable calls if +UseInlineCaches ");
286 address target = SharedRuntime::get_resolve_virtual_call_stub();
287 // Normal inline cache used for call
288 CallDynamicJavaNode *call = new (kit.C, tf()->domain()->cnt()) CallDynamicJavaNode(tf(), target, method(), _vtable_index, kit.bci());
289 kit.set_arguments_for_java_call(call);
290 kit.set_edges_for_java_call(call);
291 Node* ret = kit.set_results_for_java_call(call);
292 kit.push_node(method()->return_type()->basic_type(), ret);
294 // Represent the effect of an implicit receiver null_check
295 // as part of this call. Since we share a map with the caller,
296 // his JVMS gets adjusted.
297 kit.cast_not_null(receiver);
298 return kit.transfer_exceptions_into_jvms();
299 }
301 CallGenerator* CallGenerator::for_inline(ciMethod* m, float expected_uses) {
302 if (InlineTree::check_can_parse(m) != NULL) return NULL;
303 return new ParseGenerator(m, expected_uses);
304 }
306 // As a special case, the JVMS passed to this CallGenerator is
307 // for the method execution already in progress, not just the JVMS
308 // of the caller. Thus, this CallGenerator cannot be mixed with others!
309 CallGenerator* CallGenerator::for_osr(ciMethod* m, int osr_bci) {
310 if (InlineTree::check_can_parse(m) != NULL) return NULL;
311 float past_uses = m->interpreter_invocation_count();
312 float expected_uses = past_uses;
313 return new ParseGenerator(m, expected_uses, true);
314 }
316 CallGenerator* CallGenerator::for_direct_call(ciMethod* m, bool separate_io_proj) {
317 assert(!m->is_abstract(), "for_direct_call mismatch");
318 return new DirectCallGenerator(m, separate_io_proj);
319 }
321 CallGenerator* CallGenerator::for_dynamic_call(ciMethod* m) {
322 assert(m->is_method_handle_invoke() || m->is_method_handle_adapter(), "for_dynamic_call mismatch");
323 return new DynamicCallGenerator(m);
324 }
326 CallGenerator* CallGenerator::for_virtual_call(ciMethod* m, int vtable_index) {
327 assert(!m->is_static(), "for_virtual_call mismatch");
328 assert(!m->is_method_handle_invoke(), "should be a direct call");
329 return new VirtualCallGenerator(m, vtable_index);
330 }
332 // Allow inlining decisions to be delayed
333 class LateInlineCallGenerator : public DirectCallGenerator {
334 CallGenerator* _inline_cg;
336 public:
337 LateInlineCallGenerator(ciMethod* method, CallGenerator* inline_cg) :
338 DirectCallGenerator(method, true), _inline_cg(inline_cg) {}
340 virtual bool is_late_inline() const { return true; }
342 // Convert the CallStaticJava into an inline
343 virtual void do_late_inline();
345 JVMState* generate(JVMState* jvms) {
346 // Record that this call site should be revisited once the main
347 // parse is finished.
348 Compile::current()->add_late_inline(this);
350 // Emit the CallStaticJava and request separate projections so
351 // that the late inlining logic can distinguish between fall
352 // through and exceptional uses of the memory and io projections
353 // as is done for allocations and macro expansion.
354 return DirectCallGenerator::generate(jvms);
355 }
357 };
360 void LateInlineCallGenerator::do_late_inline() {
361 // Can't inline it
362 if (call_node() == NULL || call_node()->outcnt() == 0 ||
363 call_node()->in(0) == NULL || call_node()->in(0)->is_top())
364 return;
366 CallStaticJavaNode* call = call_node();
368 // Make a clone of the JVMState that appropriate to use for driving a parse
369 Compile* C = Compile::current();
370 JVMState* jvms = call->jvms()->clone_shallow(C);
371 uint size = call->req();
372 SafePointNode* map = new (C, size) SafePointNode(size, jvms);
373 for (uint i1 = 0; i1 < size; i1++) {
374 map->init_req(i1, call->in(i1));
375 }
377 // Make sure the state is a MergeMem for parsing.
378 if (!map->in(TypeFunc::Memory)->is_MergeMem()) {
379 map->set_req(TypeFunc::Memory, MergeMemNode::make(C, map->in(TypeFunc::Memory)));
380 }
382 // Make enough space for the expression stack and transfer the incoming arguments
383 int nargs = method()->arg_size();
384 jvms->set_map(map);
385 map->ensure_stack(jvms, jvms->method()->max_stack());
386 if (nargs > 0) {
387 for (int i1 = 0; i1 < nargs; i1++) {
388 map->set_req(i1 + jvms->argoff(), call->in(TypeFunc::Parms + i1));
389 }
390 }
392 CompileLog* log = C->log();
393 if (log != NULL) {
394 log->head("late_inline method='%d'", log->identify(method()));
395 JVMState* p = jvms;
396 while (p != NULL) {
397 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
398 p = p->caller();
399 }
400 log->tail("late_inline");
401 }
403 // Setup default node notes to be picked up by the inlining
404 Node_Notes* old_nn = C->default_node_notes();
405 if (old_nn != NULL) {
406 Node_Notes* entry_nn = old_nn->clone(C);
407 entry_nn->set_jvms(jvms);
408 C->set_default_node_notes(entry_nn);
409 }
411 // Now perform the inling using the synthesized JVMState
412 JVMState* new_jvms = _inline_cg->generate(jvms);
413 if (new_jvms == NULL) return; // no change
414 if (C->failing()) return;
416 // Capture any exceptional control flow
417 GraphKit kit(new_jvms);
419 // Find the result object
420 Node* result = C->top();
421 int result_size = method()->return_type()->size();
422 if (result_size != 0 && !kit.stopped()) {
423 result = (result_size == 1) ? kit.pop() : kit.pop_pair();
424 }
426 kit.replace_call(call, result);
427 }
430 CallGenerator* CallGenerator::for_late_inline(ciMethod* method, CallGenerator* inline_cg) {
431 return new LateInlineCallGenerator(method, inline_cg);
432 }
435 //---------------------------WarmCallGenerator--------------------------------
436 // Internal class which handles initial deferral of inlining decisions.
437 class WarmCallGenerator : public CallGenerator {
438 WarmCallInfo* _call_info;
439 CallGenerator* _if_cold;
440 CallGenerator* _if_hot;
441 bool _is_virtual; // caches virtuality of if_cold
442 bool _is_inline; // caches inline-ness of if_hot
444 public:
445 WarmCallGenerator(WarmCallInfo* ci,
446 CallGenerator* if_cold,
447 CallGenerator* if_hot)
448 : CallGenerator(if_cold->method())
449 {
450 assert(method() == if_hot->method(), "consistent choices");
451 _call_info = ci;
452 _if_cold = if_cold;
453 _if_hot = if_hot;
454 _is_virtual = if_cold->is_virtual();
455 _is_inline = if_hot->is_inline();
456 }
458 virtual bool is_inline() const { return _is_inline; }
459 virtual bool is_virtual() const { return _is_virtual; }
460 virtual bool is_deferred() const { return true; }
462 virtual JVMState* generate(JVMState* jvms);
463 };
466 CallGenerator* CallGenerator::for_warm_call(WarmCallInfo* ci,
467 CallGenerator* if_cold,
468 CallGenerator* if_hot) {
469 return new WarmCallGenerator(ci, if_cold, if_hot);
470 }
472 JVMState* WarmCallGenerator::generate(JVMState* jvms) {
473 Compile* C = Compile::current();
474 if (C->log() != NULL) {
475 C->log()->elem("warm_call bci='%d'", jvms->bci());
476 }
477 jvms = _if_cold->generate(jvms);
478 if (jvms != NULL) {
479 Node* m = jvms->map()->control();
480 if (m->is_CatchProj()) m = m->in(0); else m = C->top();
481 if (m->is_Catch()) m = m->in(0); else m = C->top();
482 if (m->is_Proj()) m = m->in(0); else m = C->top();
483 if (m->is_CallJava()) {
484 _call_info->set_call(m->as_Call());
485 _call_info->set_hot_cg(_if_hot);
486 #ifndef PRODUCT
487 if (PrintOpto || PrintOptoInlining) {
488 tty->print_cr("Queueing for warm inlining at bci %d:", jvms->bci());
489 tty->print("WCI: ");
490 _call_info->print();
491 }
492 #endif
493 _call_info->set_heat(_call_info->compute_heat());
494 C->set_warm_calls(_call_info->insert_into(C->warm_calls()));
495 }
496 }
497 return jvms;
498 }
500 void WarmCallInfo::make_hot() {
501 Unimplemented();
502 }
504 void WarmCallInfo::make_cold() {
505 // No action: Just dequeue.
506 }
509 //------------------------PredictedCallGenerator------------------------------
510 // Internal class which handles all out-of-line calls checking receiver type.
511 class PredictedCallGenerator : public CallGenerator {
512 ciKlass* _predicted_receiver;
513 CallGenerator* _if_missed;
514 CallGenerator* _if_hit;
515 float _hit_prob;
517 public:
518 PredictedCallGenerator(ciKlass* predicted_receiver,
519 CallGenerator* if_missed,
520 CallGenerator* if_hit, float hit_prob)
521 : CallGenerator(if_missed->method())
522 {
523 // The call profile data may predict the hit_prob as extreme as 0 or 1.
524 // Remove the extremes values from the range.
525 if (hit_prob > PROB_MAX) hit_prob = PROB_MAX;
526 if (hit_prob < PROB_MIN) hit_prob = PROB_MIN;
528 _predicted_receiver = predicted_receiver;
529 _if_missed = if_missed;
530 _if_hit = if_hit;
531 _hit_prob = hit_prob;
532 }
534 virtual bool is_virtual() const { return true; }
535 virtual bool is_inline() const { return _if_hit->is_inline(); }
536 virtual bool is_deferred() const { return _if_hit->is_deferred(); }
538 virtual JVMState* generate(JVMState* jvms);
539 };
542 CallGenerator* CallGenerator::for_predicted_call(ciKlass* predicted_receiver,
543 CallGenerator* if_missed,
544 CallGenerator* if_hit,
545 float hit_prob) {
546 return new PredictedCallGenerator(predicted_receiver, if_missed, if_hit, hit_prob);
547 }
550 JVMState* PredictedCallGenerator::generate(JVMState* jvms) {
551 GraphKit kit(jvms);
552 PhaseGVN& gvn = kit.gvn();
553 // We need an explicit receiver null_check before checking its type.
554 // We share a map with the caller, so his JVMS gets adjusted.
555 Node* receiver = kit.argument(0);
557 CompileLog* log = kit.C->log();
558 if (log != NULL) {
559 log->elem("predicted_call bci='%d' klass='%d'",
560 jvms->bci(), log->identify(_predicted_receiver));
561 }
563 receiver = kit.null_check_receiver(method());
564 if (kit.stopped()) {
565 return kit.transfer_exceptions_into_jvms();
566 }
568 Node* exact_receiver = receiver; // will get updated in place...
569 Node* slow_ctl = kit.type_check_receiver(receiver,
570 _predicted_receiver, _hit_prob,
571 &exact_receiver);
573 SafePointNode* slow_map = NULL;
574 JVMState* slow_jvms;
575 { PreserveJVMState pjvms(&kit);
576 kit.set_control(slow_ctl);
577 if (!kit.stopped()) {
578 slow_jvms = _if_missed->generate(kit.sync_jvms());
579 assert(slow_jvms != NULL, "miss path must not fail to generate");
580 kit.add_exception_states_from(slow_jvms);
581 kit.set_map(slow_jvms->map());
582 if (!kit.stopped())
583 slow_map = kit.stop();
584 }
585 }
587 if (kit.stopped()) {
588 // Instance exactly does not matches the desired type.
589 kit.set_jvms(slow_jvms);
590 return kit.transfer_exceptions_into_jvms();
591 }
593 // fall through if the instance exactly matches the desired type
594 kit.replace_in_map(receiver, exact_receiver);
596 // Make the hot call:
597 JVMState* new_jvms = _if_hit->generate(kit.sync_jvms());
598 if (new_jvms == NULL) {
599 // Inline failed, so make a direct call.
600 assert(_if_hit->is_inline(), "must have been a failed inline");
601 CallGenerator* cg = CallGenerator::for_direct_call(_if_hit->method());
602 new_jvms = cg->generate(kit.sync_jvms());
603 }
604 kit.add_exception_states_from(new_jvms);
605 kit.set_jvms(new_jvms);
607 // Need to merge slow and fast?
608 if (slow_map == NULL) {
609 // The fast path is the only path remaining.
610 return kit.transfer_exceptions_into_jvms();
611 }
613 if (kit.stopped()) {
614 // Inlined method threw an exception, so it's just the slow path after all.
615 kit.set_jvms(slow_jvms);
616 return kit.transfer_exceptions_into_jvms();
617 }
619 // Finish the diamond.
620 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization
621 RegionNode* region = new (kit.C, 3) RegionNode(3);
622 region->init_req(1, kit.control());
623 region->init_req(2, slow_map->control());
624 kit.set_control(gvn.transform(region));
625 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO);
626 iophi->set_req(2, slow_map->i_o());
627 kit.set_i_o(gvn.transform(iophi));
628 kit.merge_memory(slow_map->merged_memory(), region, 2);
629 uint tos = kit.jvms()->stkoff() + kit.sp();
630 uint limit = slow_map->req();
631 for (uint i = TypeFunc::Parms; i < limit; i++) {
632 // Skip unused stack slots; fast forward to monoff();
633 if (i == tos) {
634 i = kit.jvms()->monoff();
635 if( i >= limit ) break;
636 }
637 Node* m = kit.map()->in(i);
638 Node* n = slow_map->in(i);
639 if (m != n) {
640 const Type* t = gvn.type(m)->meet(gvn.type(n));
641 Node* phi = PhiNode::make(region, m, t);
642 phi->set_req(2, n);
643 kit.map()->set_req(i, gvn.transform(phi));
644 }
645 }
646 return kit.transfer_exceptions_into_jvms();
647 }
650 //------------------------PredictedDynamicCallGenerator-----------------------
651 // Internal class which handles all out-of-line calls checking receiver type.
652 class PredictedDynamicCallGenerator : public CallGenerator {
653 ciMethodHandle* _predicted_method_handle;
654 CallGenerator* _if_missed;
655 CallGenerator* _if_hit;
656 float _hit_prob;
658 public:
659 PredictedDynamicCallGenerator(ciMethodHandle* predicted_method_handle,
660 CallGenerator* if_missed,
661 CallGenerator* if_hit,
662 float hit_prob)
663 : CallGenerator(if_missed->method()),
664 _predicted_method_handle(predicted_method_handle),
665 _if_missed(if_missed),
666 _if_hit(if_hit),
667 _hit_prob(hit_prob)
668 {}
670 virtual bool is_inline() const { return _if_hit->is_inline(); }
671 virtual bool is_deferred() const { return _if_hit->is_deferred(); }
673 virtual JVMState* generate(JVMState* jvms);
674 };
677 CallGenerator* CallGenerator::for_predicted_dynamic_call(ciMethodHandle* predicted_method_handle,
678 CallGenerator* if_missed,
679 CallGenerator* if_hit,
680 float hit_prob) {
681 return new PredictedDynamicCallGenerator(predicted_method_handle, if_missed, if_hit, hit_prob);
682 }
685 CallGenerator* CallGenerator::for_method_handle_inline(Node* method_handle, JVMState* jvms,
686 ciMethod* caller, ciMethod* callee, ciCallProfile profile) {
687 if (method_handle->Opcode() == Op_ConP) {
688 const TypeOopPtr* oop_ptr = method_handle->bottom_type()->is_oopptr();
689 ciObject* const_oop = oop_ptr->const_oop();
690 ciMethodHandle* method_handle = const_oop->as_method_handle();
692 // Set the callee to have access to the class and signature in
693 // the MethodHandleCompiler.
694 method_handle->set_callee(callee);
695 method_handle->set_caller(caller);
696 method_handle->set_call_profile(profile);
698 // Get an adapter for the MethodHandle.
699 ciMethod* target_method = method_handle->get_method_handle_adapter();
700 if (target_method != NULL) {
701 CallGenerator* cg = Compile::current()->call_generator(target_method, -1, false, jvms, true, PROB_ALWAYS);
702 if (cg != NULL && cg->is_inline())
703 return cg;
704 }
705 } else if (method_handle->Opcode() == Op_Phi && method_handle->req() == 3 &&
706 method_handle->in(1)->Opcode() == Op_ConP && method_handle->in(2)->Opcode() == Op_ConP) {
707 float prob = PROB_FAIR;
708 Node* meth_region = method_handle->in(0);
709 if (meth_region->is_Region() &&
710 meth_region->in(1)->is_Proj() && meth_region->in(2)->is_Proj() &&
711 meth_region->in(1)->in(0) == meth_region->in(2)->in(0) &&
712 meth_region->in(1)->in(0)->is_If()) {
713 // If diamond, so grab the probability of the test to drive the inlining below
714 prob = meth_region->in(1)->in(0)->as_If()->_prob;
715 if (meth_region->in(1)->is_IfTrue()) {
716 prob = 1 - prob;
717 }
718 }
720 // selectAlternative idiom merging two constant MethodHandles.
721 // Generate a guard so that each can be inlined. We might want to
722 // do more inputs at later point but this gets the most common
723 // case.
724 CallGenerator* cg1 = for_method_handle_inline(method_handle->in(1), jvms, caller, callee, profile.rescale(1.0 - prob));
725 CallGenerator* cg2 = for_method_handle_inline(method_handle->in(2), jvms, caller, callee, profile.rescale(prob));
726 if (cg1 != NULL && cg2 != NULL) {
727 const TypeOopPtr* oop_ptr = method_handle->in(1)->bottom_type()->is_oopptr();
728 ciObject* const_oop = oop_ptr->const_oop();
729 ciMethodHandle* mh = const_oop->as_method_handle();
730 return new PredictedDynamicCallGenerator(mh, cg2, cg1, prob);
731 }
732 }
733 return NULL;
734 }
737 CallGenerator* CallGenerator::for_invokedynamic_inline(ciCallSite* call_site, JVMState* jvms,
738 ciMethod* caller, ciMethod* callee, ciCallProfile profile) {
739 ciMethodHandle* method_handle = call_site->get_target();
741 // Set the callee to have access to the class and signature in the
742 // MethodHandleCompiler.
743 method_handle->set_callee(callee);
744 method_handle->set_caller(caller);
745 method_handle->set_call_profile(profile);
747 // Get an adapter for the MethodHandle.
748 ciMethod* target_method = method_handle->get_invokedynamic_adapter();
749 if (target_method != NULL) {
750 Compile *C = Compile::current();
751 CallGenerator* cg = C->call_generator(target_method, -1, false, jvms, true, PROB_ALWAYS);
752 if (cg != NULL && cg->is_inline()) {
753 // Add a dependence for invalidation of the optimization.
754 if (!call_site->is_constant_call_site()) {
755 C->dependencies()->assert_call_site_target_value(call_site, method_handle);
756 }
757 return cg;
758 }
759 }
760 return NULL;
761 }
764 JVMState* PredictedDynamicCallGenerator::generate(JVMState* jvms) {
765 GraphKit kit(jvms);
766 PhaseGVN& gvn = kit.gvn();
768 CompileLog* log = kit.C->log();
769 if (log != NULL) {
770 log->elem("predicted_dynamic_call bci='%d'", jvms->bci());
771 }
773 const TypeOopPtr* predicted_mh_ptr = TypeOopPtr::make_from_constant(_predicted_method_handle, true);
774 Node* predicted_mh = kit.makecon(predicted_mh_ptr);
776 Node* bol = NULL;
777 int bc = jvms->method()->java_code_at_bci(jvms->bci());
778 if (bc == Bytecodes::_invokespecial) {
779 // This is the selectAlternative idiom for guardWithTest
780 Node* receiver = kit.argument(0);
782 // Check if the MethodHandle is the expected one
783 Node* cmp = gvn.transform(new(kit.C, 3) CmpPNode(receiver, predicted_mh));
784 bol = gvn.transform(new(kit.C, 2) BoolNode(cmp, BoolTest::eq) );
785 } else {
786 assert(bc == Bytecodes::_invokedynamic, "must be");
787 // Get the constant pool cache from the caller class.
788 ciMethod* caller_method = jvms->method();
789 ciBytecodeStream str(caller_method);
790 str.force_bci(jvms->bci()); // Set the stream to the invokedynamic bci.
791 ciCPCache* cpcache = str.get_cpcache();
793 // Get the offset of the CallSite from the constant pool cache
794 // pointer.
795 int index = str.get_method_index();
796 size_t call_site_offset = cpcache->get_f1_offset(index);
798 // Load the CallSite object from the constant pool cache.
799 const TypeOopPtr* cpcache_ptr = TypeOopPtr::make_from_constant(cpcache);
800 Node* cpcache_adr = kit.makecon(cpcache_ptr);
801 Node* call_site_adr = kit.basic_plus_adr(cpcache_adr, cpcache_adr, call_site_offset);
802 Node* call_site = kit.make_load(kit.control(), call_site_adr, TypeInstPtr::BOTTOM, T_OBJECT, Compile::AliasIdxRaw);
804 // Load the target MethodHandle from the CallSite object.
805 Node* target_adr = kit.basic_plus_adr(call_site, call_site, java_lang_invoke_CallSite::target_offset_in_bytes());
806 Node* target_mh = kit.make_load(kit.control(), target_adr, TypeInstPtr::BOTTOM, T_OBJECT);
808 // Check if the MethodHandle is still the same.
809 Node* cmp = gvn.transform(new(kit.C, 3) CmpPNode(target_mh, predicted_mh));
810 bol = gvn.transform(new(kit.C, 2) BoolNode(cmp, BoolTest::eq) );
811 }
812 IfNode* iff = kit.create_and_xform_if(kit.control(), bol, _hit_prob, COUNT_UNKNOWN);
813 kit.set_control( gvn.transform(new(kit.C, 1) IfTrueNode (iff)));
814 Node* slow_ctl = gvn.transform(new(kit.C, 1) IfFalseNode(iff));
816 SafePointNode* slow_map = NULL;
817 JVMState* slow_jvms;
818 { PreserveJVMState pjvms(&kit);
819 kit.set_control(slow_ctl);
820 if (!kit.stopped()) {
821 slow_jvms = _if_missed->generate(kit.sync_jvms());
822 assert(slow_jvms != NULL, "miss path must not fail to generate");
823 kit.add_exception_states_from(slow_jvms);
824 kit.set_map(slow_jvms->map());
825 if (!kit.stopped())
826 slow_map = kit.stop();
827 }
828 }
830 if (kit.stopped()) {
831 // Instance exactly does not matches the desired type.
832 kit.set_jvms(slow_jvms);
833 return kit.transfer_exceptions_into_jvms();
834 }
836 // Make the hot call:
837 JVMState* new_jvms = _if_hit->generate(kit.sync_jvms());
838 if (new_jvms == NULL) {
839 // Inline failed, so make a direct call.
840 assert(_if_hit->is_inline(), "must have been a failed inline");
841 CallGenerator* cg = CallGenerator::for_direct_call(_if_hit->method());
842 new_jvms = cg->generate(kit.sync_jvms());
843 }
844 kit.add_exception_states_from(new_jvms);
845 kit.set_jvms(new_jvms);
847 // Need to merge slow and fast?
848 if (slow_map == NULL) {
849 // The fast path is the only path remaining.
850 return kit.transfer_exceptions_into_jvms();
851 }
853 if (kit.stopped()) {
854 // Inlined method threw an exception, so it's just the slow path after all.
855 kit.set_jvms(slow_jvms);
856 return kit.transfer_exceptions_into_jvms();
857 }
859 // Finish the diamond.
860 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization
861 RegionNode* region = new (kit.C, 3) RegionNode(3);
862 region->init_req(1, kit.control());
863 region->init_req(2, slow_map->control());
864 kit.set_control(gvn.transform(region));
865 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO);
866 iophi->set_req(2, slow_map->i_o());
867 kit.set_i_o(gvn.transform(iophi));
868 kit.merge_memory(slow_map->merged_memory(), region, 2);
869 uint tos = kit.jvms()->stkoff() + kit.sp();
870 uint limit = slow_map->req();
871 for (uint i = TypeFunc::Parms; i < limit; i++) {
872 // Skip unused stack slots; fast forward to monoff();
873 if (i == tos) {
874 i = kit.jvms()->monoff();
875 if( i >= limit ) break;
876 }
877 Node* m = kit.map()->in(i);
878 Node* n = slow_map->in(i);
879 if (m != n) {
880 const Type* t = gvn.type(m)->meet(gvn.type(n));
881 Node* phi = PhiNode::make(region, m, t);
882 phi->set_req(2, n);
883 kit.map()->set_req(i, gvn.transform(phi));
884 }
885 }
886 return kit.transfer_exceptions_into_jvms();
887 }
890 //-------------------------UncommonTrapCallGenerator-----------------------------
891 // Internal class which handles all out-of-line calls checking receiver type.
892 class UncommonTrapCallGenerator : public CallGenerator {
893 Deoptimization::DeoptReason _reason;
894 Deoptimization::DeoptAction _action;
896 public:
897 UncommonTrapCallGenerator(ciMethod* m,
898 Deoptimization::DeoptReason reason,
899 Deoptimization::DeoptAction action)
900 : CallGenerator(m)
901 {
902 _reason = reason;
903 _action = action;
904 }
906 virtual bool is_virtual() const { ShouldNotReachHere(); return false; }
907 virtual bool is_trap() const { return true; }
909 virtual JVMState* generate(JVMState* jvms);
910 };
913 CallGenerator*
914 CallGenerator::for_uncommon_trap(ciMethod* m,
915 Deoptimization::DeoptReason reason,
916 Deoptimization::DeoptAction action) {
917 return new UncommonTrapCallGenerator(m, reason, action);
918 }
921 JVMState* UncommonTrapCallGenerator::generate(JVMState* jvms) {
922 GraphKit kit(jvms);
923 // Take the trap with arguments pushed on the stack. (Cf. null_check_receiver).
924 int nargs = method()->arg_size();
925 kit.inc_sp(nargs);
926 assert(nargs <= kit.sp() && kit.sp() <= jvms->stk_size(), "sane sp w/ args pushed");
927 if (_reason == Deoptimization::Reason_class_check &&
928 _action == Deoptimization::Action_maybe_recompile) {
929 // Temp fix for 6529811
930 // Don't allow uncommon_trap to override our decision to recompile in the event
931 // of a class cast failure for a monomorphic call as it will never let us convert
932 // the call to either bi-morphic or megamorphic and can lead to unc-trap loops
933 bool keep_exact_action = true;
934 kit.uncommon_trap(_reason, _action, NULL, "monomorphic vcall checkcast", false, keep_exact_action);
935 } else {
936 kit.uncommon_trap(_reason, _action);
937 }
938 return kit.transfer_exceptions_into_jvms();
939 }
941 // (Note: Moved hook_up_call to GraphKit::set_edges_for_java_call.)
943 // (Node: Merged hook_up_exits into ParseGenerator::generate.)
945 #define NODES_OVERHEAD_PER_METHOD (30.0)
946 #define NODES_PER_BYTECODE (9.5)
948 void WarmCallInfo::init(JVMState* call_site, ciMethod* call_method, ciCallProfile& profile, float prof_factor) {
949 int call_count = profile.count();
950 int code_size = call_method->code_size();
952 // Expected execution count is based on the historical count:
953 _count = call_count < 0 ? 1 : call_site->method()->scale_count(call_count, prof_factor);
955 // Expected profit from inlining, in units of simple call-overheads.
956 _profit = 1.0;
958 // Expected work performed by the call in units of call-overheads.
959 // %%% need an empirical curve fit for "work" (time in call)
960 float bytecodes_per_call = 3;
961 _work = 1.0 + code_size / bytecodes_per_call;
963 // Expected size of compilation graph:
964 // -XX:+PrintParseStatistics once reported:
965 // Methods seen: 9184 Methods parsed: 9184 Nodes created: 1582391
966 // Histogram of 144298 parsed bytecodes:
967 // %%% Need an better predictor for graph size.
968 _size = NODES_OVERHEAD_PER_METHOD + (NODES_PER_BYTECODE * code_size);
969 }
971 // is_cold: Return true if the node should never be inlined.
972 // This is true if any of the key metrics are extreme.
973 bool WarmCallInfo::is_cold() const {
974 if (count() < WarmCallMinCount) return true;
975 if (profit() < WarmCallMinProfit) return true;
976 if (work() > WarmCallMaxWork) return true;
977 if (size() > WarmCallMaxSize) return true;
978 return false;
979 }
981 // is_hot: Return true if the node should be inlined immediately.
982 // This is true if any of the key metrics are extreme.
983 bool WarmCallInfo::is_hot() const {
984 assert(!is_cold(), "eliminate is_cold cases before testing is_hot");
985 if (count() >= HotCallCountThreshold) return true;
986 if (profit() >= HotCallProfitThreshold) return true;
987 if (work() <= HotCallTrivialWork) return true;
988 if (size() <= HotCallTrivialSize) return true;
989 return false;
990 }
992 // compute_heat:
993 float WarmCallInfo::compute_heat() const {
994 assert(!is_cold(), "compute heat only on warm nodes");
995 assert(!is_hot(), "compute heat only on warm nodes");
996 int min_size = MAX2(0, (int)HotCallTrivialSize);
997 int max_size = MIN2(500, (int)WarmCallMaxSize);
998 float method_size = (size() - min_size) / MAX2(1, max_size - min_size);
999 float size_factor;
1000 if (method_size < 0.05) size_factor = 4; // 2 sigmas better than avg.
1001 else if (method_size < 0.15) size_factor = 2; // 1 sigma better than avg.
1002 else if (method_size < 0.5) size_factor = 1; // better than avg.
1003 else size_factor = 0.5; // worse than avg.
1004 return (count() * profit() * size_factor);
1005 }
1007 bool WarmCallInfo::warmer_than(WarmCallInfo* that) {
1008 assert(this != that, "compare only different WCIs");
1009 assert(this->heat() != 0 && that->heat() != 0, "call compute_heat 1st");
1010 if (this->heat() > that->heat()) return true;
1011 if (this->heat() < that->heat()) return false;
1012 assert(this->heat() == that->heat(), "no NaN heat allowed");
1013 // Equal heat. Break the tie some other way.
1014 if (!this->call() || !that->call()) return (address)this > (address)that;
1015 return this->call()->_idx > that->call()->_idx;
1016 }
1018 //#define UNINIT_NEXT ((WarmCallInfo*)badAddress)
1019 #define UNINIT_NEXT ((WarmCallInfo*)NULL)
1021 WarmCallInfo* WarmCallInfo::insert_into(WarmCallInfo* head) {
1022 assert(next() == UNINIT_NEXT, "not yet on any list");
1023 WarmCallInfo* prev_p = NULL;
1024 WarmCallInfo* next_p = head;
1025 while (next_p != NULL && next_p->warmer_than(this)) {
1026 prev_p = next_p;
1027 next_p = prev_p->next();
1028 }
1029 // Install this between prev_p and next_p.
1030 this->set_next(next_p);
1031 if (prev_p == NULL)
1032 head = this;
1033 else
1034 prev_p->set_next(this);
1035 return head;
1036 }
1038 WarmCallInfo* WarmCallInfo::remove_from(WarmCallInfo* head) {
1039 WarmCallInfo* prev_p = NULL;
1040 WarmCallInfo* next_p = head;
1041 while (next_p != this) {
1042 assert(next_p != NULL, "this must be in the list somewhere");
1043 prev_p = next_p;
1044 next_p = prev_p->next();
1045 }
1046 next_p = this->next();
1047 debug_only(this->set_next(UNINIT_NEXT));
1048 // Remove this from between prev_p and next_p.
1049 if (prev_p == NULL)
1050 head = next_p;
1051 else
1052 prev_p->set_next(next_p);
1053 return head;
1054 }
1056 WarmCallInfo WarmCallInfo::_always_hot(WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE(),
1057 WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE());
1058 WarmCallInfo WarmCallInfo::_always_cold(WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE(),
1059 WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE());
1061 WarmCallInfo* WarmCallInfo::always_hot() {
1062 assert(_always_hot.is_hot(), "must always be hot");
1063 return &_always_hot;
1064 }
1066 WarmCallInfo* WarmCallInfo::always_cold() {
1067 assert(_always_cold.is_cold(), "must always be cold");
1068 return &_always_cold;
1069 }
1072 #ifndef PRODUCT
1074 void WarmCallInfo::print() const {
1075 tty->print("%s : C=%6.1f P=%6.1f W=%6.1f S=%6.1f H=%6.1f -> %p",
1076 is_cold() ? "cold" : is_hot() ? "hot " : "warm",
1077 count(), profit(), work(), size(), compute_heat(), next());
1078 tty->cr();
1079 if (call() != NULL) call()->dump();
1080 }
1082 void print_wci(WarmCallInfo* ci) {
1083 ci->print();
1084 }
1086 void WarmCallInfo::print_all() const {
1087 for (const WarmCallInfo* p = this; p != NULL; p = p->next())
1088 p->print();
1089 }
1091 int WarmCallInfo::count_all() const {
1092 int cnt = 0;
1093 for (const WarmCallInfo* p = this; p != NULL; p = p->next())
1094 cnt++;
1095 return cnt;
1096 }
1098 #endif //PRODUCT