Wed, 18 Sep 2013 14:34:56 -0700
8024342: PPC64 (part 111): Support for C calling conventions that require 64-bit ints.
Summary: Some platforms, as ppc and s390x/zArch require that 32-bit ints are passed as 64-bit values to C functions. This change adds support to adapt the signature and to issue proper casts to c2-compiled stubs. The functions are used in generate_native_wrapper(). Adapt signature used by the compiler as in PhaseIdealLoop::intrinsify_fill().
Reviewed-by: 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(kit.C, 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 CallStaticJavaNode* call = call_node();
308 if (call == NULL || call->outcnt() == 0 ||
309 call->in(0) == NULL || call->in(0)->is_top()) {
310 return;
311 }
313 const TypeTuple *r = call->tf()->domain();
314 for (int i1 = 0; i1 < method()->arg_size(); i1++) {
315 if (call->in(TypeFunc::Parms + i1)->is_top() && r->field_at(TypeFunc::Parms + i1) != Type::HALF) {
316 assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing");
317 return;
318 }
319 }
321 if (call->in(TypeFunc::Memory)->is_top()) {
322 assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing");
323 return;
324 }
326 Compile* C = Compile::current();
327 // Remove inlined methods from Compiler's lists.
328 if (call->is_macro()) {
329 C->remove_macro_node(call);
330 }
332 // Make a clone of the JVMState that appropriate to use for driving a parse
333 JVMState* old_jvms = call->jvms();
334 JVMState* jvms = old_jvms->clone_shallow(C);
335 uint size = call->req();
336 SafePointNode* map = new (C) SafePointNode(size, jvms);
337 for (uint i1 = 0; i1 < size; i1++) {
338 map->init_req(i1, call->in(i1));
339 }
341 // Make sure the state is a MergeMem for parsing.
342 if (!map->in(TypeFunc::Memory)->is_MergeMem()) {
343 Node* mem = MergeMemNode::make(C, map->in(TypeFunc::Memory));
344 C->initial_gvn()->set_type_bottom(mem);
345 map->set_req(TypeFunc::Memory, mem);
346 }
348 uint nargs = method()->arg_size();
349 // blow away old call arguments
350 Node* top = C->top();
351 for (uint i1 = 0; i1 < nargs; i1++) {
352 map->set_req(TypeFunc::Parms + i1, top);
353 }
354 jvms->set_map(map);
356 // Make enough space in the expression stack to transfer
357 // the incoming arguments and return value.
358 map->ensure_stack(jvms, jvms->method()->max_stack());
359 for (uint i1 = 0; i1 < nargs; i1++) {
360 map->set_argument(jvms, i1, call->in(TypeFunc::Parms + i1));
361 }
363 // This check is done here because for_method_handle_inline() method
364 // needs jvms for inlined state.
365 if (!do_late_inline_check(jvms)) {
366 map->disconnect_inputs(NULL, C);
367 return;
368 }
370 C->print_inlining_insert(this);
372 CompileLog* log = C->log();
373 if (log != NULL) {
374 log->head("late_inline method='%d'", log->identify(method()));
375 JVMState* p = jvms;
376 while (p != NULL) {
377 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
378 p = p->caller();
379 }
380 log->tail("late_inline");
381 }
383 // Setup default node notes to be picked up by the inlining
384 Node_Notes* old_nn = C->default_node_notes();
385 if (old_nn != NULL) {
386 Node_Notes* entry_nn = old_nn->clone(C);
387 entry_nn->set_jvms(jvms);
388 C->set_default_node_notes(entry_nn);
389 }
391 // Now perform the inling using the synthesized JVMState
392 JVMState* new_jvms = _inline_cg->generate(jvms);
393 if (new_jvms == NULL) return; // no change
394 if (C->failing()) return;
396 // Capture any exceptional control flow
397 GraphKit kit(new_jvms);
399 // Find the result object
400 Node* result = C->top();
401 int result_size = method()->return_type()->size();
402 if (result_size != 0 && !kit.stopped()) {
403 result = (result_size == 1) ? kit.pop() : kit.pop_pair();
404 }
406 C->set_has_loops(C->has_loops() || _inline_cg->method()->has_loops());
407 C->env()->notice_inlined_method(_inline_cg->method());
408 C->set_inlining_progress(true);
410 kit.replace_call(call, result);
411 }
414 CallGenerator* CallGenerator::for_late_inline(ciMethod* method, CallGenerator* inline_cg) {
415 return new LateInlineCallGenerator(method, inline_cg);
416 }
418 class LateInlineMHCallGenerator : public LateInlineCallGenerator {
419 ciMethod* _caller;
420 int _attempt;
421 bool _input_not_const;
423 virtual bool do_late_inline_check(JVMState* jvms);
424 virtual bool already_attempted() const { return _attempt > 0; }
426 public:
427 LateInlineMHCallGenerator(ciMethod* caller, ciMethod* callee, bool input_not_const) :
428 LateInlineCallGenerator(callee, NULL), _caller(caller), _attempt(0), _input_not_const(input_not_const) {}
430 virtual bool is_mh_late_inline() const { return true; }
432 virtual JVMState* generate(JVMState* jvms) {
433 JVMState* new_jvms = LateInlineCallGenerator::generate(jvms);
434 if (_input_not_const) {
435 // inlining won't be possible so no need to enqueue right now.
436 call_node()->set_generator(this);
437 } else {
438 Compile::current()->add_late_inline(this);
439 }
440 return new_jvms;
441 }
443 virtual void print_inlining_late(const char* msg) {
444 if (!_input_not_const) return;
445 LateInlineCallGenerator::print_inlining_late(msg);
446 }
447 };
449 bool LateInlineMHCallGenerator::do_late_inline_check(JVMState* jvms) {
451 CallGenerator* cg = for_method_handle_inline(jvms, _caller, method(), _input_not_const);
453 if (!_input_not_const) {
454 _attempt++;
455 }
457 if (cg != NULL) {
458 assert(!cg->is_late_inline() && cg->is_inline(), "we're doing late inlining");
459 _inline_cg = cg;
460 Compile::current()->dec_number_of_mh_late_inlines();
461 return true;
462 }
464 call_node()->set_generator(this);
465 return false;
466 }
468 CallGenerator* CallGenerator::for_mh_late_inline(ciMethod* caller, ciMethod* callee, bool input_not_const) {
469 Compile::current()->inc_number_of_mh_late_inlines();
470 CallGenerator* cg = new LateInlineMHCallGenerator(caller, callee, input_not_const);
471 return cg;
472 }
474 class LateInlineStringCallGenerator : public LateInlineCallGenerator {
476 public:
477 LateInlineStringCallGenerator(ciMethod* method, CallGenerator* inline_cg) :
478 LateInlineCallGenerator(method, inline_cg) {}
480 virtual JVMState* generate(JVMState* jvms) {
481 Compile *C = Compile::current();
482 C->print_inlining_skip(this);
484 C->add_string_late_inline(this);
486 JVMState* new_jvms = DirectCallGenerator::generate(jvms);
487 return new_jvms;
488 }
489 };
491 CallGenerator* CallGenerator::for_string_late_inline(ciMethod* method, CallGenerator* inline_cg) {
492 return new LateInlineStringCallGenerator(method, inline_cg);
493 }
495 class LateInlineBoxingCallGenerator : public LateInlineCallGenerator {
497 public:
498 LateInlineBoxingCallGenerator(ciMethod* method, CallGenerator* inline_cg) :
499 LateInlineCallGenerator(method, inline_cg) {}
501 virtual JVMState* generate(JVMState* jvms) {
502 Compile *C = Compile::current();
503 C->print_inlining_skip(this);
505 C->add_boxing_late_inline(this);
507 JVMState* new_jvms = DirectCallGenerator::generate(jvms);
508 return new_jvms;
509 }
510 };
512 CallGenerator* CallGenerator::for_boxing_late_inline(ciMethod* method, CallGenerator* inline_cg) {
513 return new LateInlineBoxingCallGenerator(method, inline_cg);
514 }
516 //---------------------------WarmCallGenerator--------------------------------
517 // Internal class which handles initial deferral of inlining decisions.
518 class WarmCallGenerator : public CallGenerator {
519 WarmCallInfo* _call_info;
520 CallGenerator* _if_cold;
521 CallGenerator* _if_hot;
522 bool _is_virtual; // caches virtuality of if_cold
523 bool _is_inline; // caches inline-ness of if_hot
525 public:
526 WarmCallGenerator(WarmCallInfo* ci,
527 CallGenerator* if_cold,
528 CallGenerator* if_hot)
529 : CallGenerator(if_cold->method())
530 {
531 assert(method() == if_hot->method(), "consistent choices");
532 _call_info = ci;
533 _if_cold = if_cold;
534 _if_hot = if_hot;
535 _is_virtual = if_cold->is_virtual();
536 _is_inline = if_hot->is_inline();
537 }
539 virtual bool is_inline() const { return _is_inline; }
540 virtual bool is_virtual() const { return _is_virtual; }
541 virtual bool is_deferred() const { return true; }
543 virtual JVMState* generate(JVMState* jvms);
544 };
547 CallGenerator* CallGenerator::for_warm_call(WarmCallInfo* ci,
548 CallGenerator* if_cold,
549 CallGenerator* if_hot) {
550 return new WarmCallGenerator(ci, if_cold, if_hot);
551 }
553 JVMState* WarmCallGenerator::generate(JVMState* jvms) {
554 Compile* C = Compile::current();
555 if (C->log() != NULL) {
556 C->log()->elem("warm_call bci='%d'", jvms->bci());
557 }
558 jvms = _if_cold->generate(jvms);
559 if (jvms != NULL) {
560 Node* m = jvms->map()->control();
561 if (m->is_CatchProj()) m = m->in(0); else m = C->top();
562 if (m->is_Catch()) m = m->in(0); else m = C->top();
563 if (m->is_Proj()) m = m->in(0); else m = C->top();
564 if (m->is_CallJava()) {
565 _call_info->set_call(m->as_Call());
566 _call_info->set_hot_cg(_if_hot);
567 #ifndef PRODUCT
568 if (PrintOpto || PrintOptoInlining) {
569 tty->print_cr("Queueing for warm inlining at bci %d:", jvms->bci());
570 tty->print("WCI: ");
571 _call_info->print();
572 }
573 #endif
574 _call_info->set_heat(_call_info->compute_heat());
575 C->set_warm_calls(_call_info->insert_into(C->warm_calls()));
576 }
577 }
578 return jvms;
579 }
581 void WarmCallInfo::make_hot() {
582 Unimplemented();
583 }
585 void WarmCallInfo::make_cold() {
586 // No action: Just dequeue.
587 }
590 //------------------------PredictedCallGenerator------------------------------
591 // Internal class which handles all out-of-line calls checking receiver type.
592 class PredictedCallGenerator : public CallGenerator {
593 ciKlass* _predicted_receiver;
594 CallGenerator* _if_missed;
595 CallGenerator* _if_hit;
596 float _hit_prob;
598 public:
599 PredictedCallGenerator(ciKlass* predicted_receiver,
600 CallGenerator* if_missed,
601 CallGenerator* if_hit, float hit_prob)
602 : CallGenerator(if_missed->method())
603 {
604 // The call profile data may predict the hit_prob as extreme as 0 or 1.
605 // Remove the extremes values from the range.
606 if (hit_prob > PROB_MAX) hit_prob = PROB_MAX;
607 if (hit_prob < PROB_MIN) hit_prob = PROB_MIN;
609 _predicted_receiver = predicted_receiver;
610 _if_missed = if_missed;
611 _if_hit = if_hit;
612 _hit_prob = hit_prob;
613 }
615 virtual bool is_virtual() const { return true; }
616 virtual bool is_inline() const { return _if_hit->is_inline(); }
617 virtual bool is_deferred() const { return _if_hit->is_deferred(); }
619 virtual JVMState* generate(JVMState* jvms);
620 };
623 CallGenerator* CallGenerator::for_predicted_call(ciKlass* predicted_receiver,
624 CallGenerator* if_missed,
625 CallGenerator* if_hit,
626 float hit_prob) {
627 return new PredictedCallGenerator(predicted_receiver, if_missed, if_hit, hit_prob);
628 }
631 JVMState* PredictedCallGenerator::generate(JVMState* jvms) {
632 GraphKit kit(jvms);
633 PhaseGVN& gvn = kit.gvn();
634 // We need an explicit receiver null_check before checking its type.
635 // We share a map with the caller, so his JVMS gets adjusted.
636 Node* receiver = kit.argument(0);
638 CompileLog* log = kit.C->log();
639 if (log != NULL) {
640 log->elem("predicted_call bci='%d' klass='%d'",
641 jvms->bci(), log->identify(_predicted_receiver));
642 }
644 receiver = kit.null_check_receiver_before_call(method());
645 if (kit.stopped()) {
646 return kit.transfer_exceptions_into_jvms();
647 }
649 Node* exact_receiver = receiver; // will get updated in place...
650 Node* slow_ctl = kit.type_check_receiver(receiver,
651 _predicted_receiver, _hit_prob,
652 &exact_receiver);
654 SafePointNode* slow_map = NULL;
655 JVMState* slow_jvms;
656 { PreserveJVMState pjvms(&kit);
657 kit.set_control(slow_ctl);
658 if (!kit.stopped()) {
659 slow_jvms = _if_missed->generate(kit.sync_jvms());
660 if (kit.failing())
661 return NULL; // might happen because of NodeCountInliningCutoff
662 assert(slow_jvms != NULL, "must be");
663 kit.add_exception_states_from(slow_jvms);
664 kit.set_map(slow_jvms->map());
665 if (!kit.stopped())
666 slow_map = kit.stop();
667 }
668 }
670 if (kit.stopped()) {
671 // Instance exactly does not matches the desired type.
672 kit.set_jvms(slow_jvms);
673 return kit.transfer_exceptions_into_jvms();
674 }
676 // fall through if the instance exactly matches the desired type
677 kit.replace_in_map(receiver, exact_receiver);
679 // Make the hot call:
680 JVMState* new_jvms = _if_hit->generate(kit.sync_jvms());
681 if (new_jvms == NULL) {
682 // Inline failed, so make a direct call.
683 assert(_if_hit->is_inline(), "must have been a failed inline");
684 CallGenerator* cg = CallGenerator::for_direct_call(_if_hit->method());
685 new_jvms = cg->generate(kit.sync_jvms());
686 }
687 kit.add_exception_states_from(new_jvms);
688 kit.set_jvms(new_jvms);
690 // Need to merge slow and fast?
691 if (slow_map == NULL) {
692 // The fast path is the only path remaining.
693 return kit.transfer_exceptions_into_jvms();
694 }
696 if (kit.stopped()) {
697 // Inlined method threw an exception, so it's just the slow path after all.
698 kit.set_jvms(slow_jvms);
699 return kit.transfer_exceptions_into_jvms();
700 }
702 // Finish the diamond.
703 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization
704 RegionNode* region = new (kit.C) RegionNode(3);
705 region->init_req(1, kit.control());
706 region->init_req(2, slow_map->control());
707 kit.set_control(gvn.transform(region));
708 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO);
709 iophi->set_req(2, slow_map->i_o());
710 kit.set_i_o(gvn.transform(iophi));
711 kit.merge_memory(slow_map->merged_memory(), region, 2);
712 uint tos = kit.jvms()->stkoff() + kit.sp();
713 uint limit = slow_map->req();
714 for (uint i = TypeFunc::Parms; i < limit; i++) {
715 // Skip unused stack slots; fast forward to monoff();
716 if (i == tos) {
717 i = kit.jvms()->monoff();
718 if( i >= limit ) break;
719 }
720 Node* m = kit.map()->in(i);
721 Node* n = slow_map->in(i);
722 if (m != n) {
723 const Type* t = gvn.type(m)->meet(gvn.type(n));
724 Node* phi = PhiNode::make(region, m, t);
725 phi->set_req(2, n);
726 kit.map()->set_req(i, gvn.transform(phi));
727 }
728 }
729 return kit.transfer_exceptions_into_jvms();
730 }
733 CallGenerator* CallGenerator::for_method_handle_call(JVMState* jvms, ciMethod* caller, ciMethod* callee, bool delayed_forbidden) {
734 assert(callee->is_method_handle_intrinsic() ||
735 callee->is_compiled_lambda_form(), "for_method_handle_call mismatch");
736 bool input_not_const;
737 CallGenerator* cg = CallGenerator::for_method_handle_inline(jvms, caller, callee, input_not_const);
738 Compile* C = Compile::current();
739 if (cg != NULL) {
740 if (!delayed_forbidden && AlwaysIncrementalInline) {
741 return CallGenerator::for_late_inline(callee, cg);
742 } else {
743 return cg;
744 }
745 }
746 int bci = jvms->bci();
747 ciCallProfile profile = caller->call_profile_at_bci(bci);
748 int call_site_count = caller->scale_count(profile.count());
750 if (IncrementalInline && call_site_count > 0 &&
751 (input_not_const || !C->inlining_incrementally() || C->over_inlining_cutoff())) {
752 return CallGenerator::for_mh_late_inline(caller, callee, input_not_const);
753 } else {
754 // Out-of-line call.
755 return CallGenerator::for_direct_call(callee);
756 }
757 }
759 CallGenerator* CallGenerator::for_method_handle_inline(JVMState* jvms, ciMethod* caller, ciMethod* callee, bool& input_not_const) {
760 GraphKit kit(jvms);
761 PhaseGVN& gvn = kit.gvn();
762 Compile* C = kit.C;
763 vmIntrinsics::ID iid = callee->intrinsic_id();
764 input_not_const = true;
765 switch (iid) {
766 case vmIntrinsics::_invokeBasic:
767 {
768 // Get MethodHandle receiver:
769 Node* receiver = kit.argument(0);
770 if (receiver->Opcode() == Op_ConP) {
771 input_not_const = false;
772 const TypeOopPtr* oop_ptr = receiver->bottom_type()->is_oopptr();
773 ciMethod* target = oop_ptr->const_oop()->as_method_handle()->get_vmtarget();
774 guarantee(!target->is_method_handle_intrinsic(), "should not happen"); // XXX remove
775 const int vtable_index = Method::invalid_vtable_index;
776 CallGenerator* cg = C->call_generator(target, vtable_index, false, jvms, true, PROB_ALWAYS, true, true);
777 assert(!cg->is_late_inline() || cg->is_mh_late_inline(), "no late inline here");
778 if (cg != NULL && cg->is_inline())
779 return cg;
780 }
781 }
782 break;
784 case vmIntrinsics::_linkToVirtual:
785 case vmIntrinsics::_linkToStatic:
786 case vmIntrinsics::_linkToSpecial:
787 case vmIntrinsics::_linkToInterface:
788 {
789 // Get MemberName argument:
790 Node* member_name = kit.argument(callee->arg_size() - 1);
791 if (member_name->Opcode() == Op_ConP) {
792 input_not_const = false;
793 const TypeOopPtr* oop_ptr = member_name->bottom_type()->is_oopptr();
794 ciMethod* target = oop_ptr->const_oop()->as_member_name()->get_vmtarget();
796 // In lamda forms we erase signature types to avoid resolving issues
797 // involving class loaders. When we optimize a method handle invoke
798 // to a direct call we must cast the receiver and arguments to its
799 // actual types.
800 ciSignature* signature = target->signature();
801 const int receiver_skip = target->is_static() ? 0 : 1;
802 // Cast receiver to its type.
803 if (!target->is_static()) {
804 Node* arg = kit.argument(0);
805 const TypeOopPtr* arg_type = arg->bottom_type()->isa_oopptr();
806 const Type* sig_type = TypeOopPtr::make_from_klass(signature->accessing_klass());
807 if (arg_type != NULL && !arg_type->higher_equal(sig_type)) {
808 Node* cast_obj = gvn.transform(new (C) CheckCastPPNode(kit.control(), arg, sig_type));
809 kit.set_argument(0, cast_obj);
810 }
811 }
812 // Cast reference arguments to its type.
813 for (int i = 0; i < signature->count(); i++) {
814 ciType* t = signature->type_at(i);
815 if (t->is_klass()) {
816 Node* arg = kit.argument(receiver_skip + i);
817 const TypeOopPtr* arg_type = arg->bottom_type()->isa_oopptr();
818 const Type* sig_type = TypeOopPtr::make_from_klass(t->as_klass());
819 if (arg_type != NULL && !arg_type->higher_equal(sig_type)) {
820 Node* cast_obj = gvn.transform(new (C) CheckCastPPNode(kit.control(), arg, sig_type));
821 kit.set_argument(receiver_skip + i, cast_obj);
822 }
823 }
824 }
826 // Try to get the most accurate receiver type
827 const bool is_virtual = (iid == vmIntrinsics::_linkToVirtual);
828 const bool is_virtual_or_interface = (is_virtual || iid == vmIntrinsics::_linkToInterface);
829 int vtable_index = Method::invalid_vtable_index;
830 bool call_does_dispatch = false;
832 if (is_virtual_or_interface) {
833 ciInstanceKlass* klass = target->holder();
834 Node* receiver_node = kit.argument(0);
835 const TypeOopPtr* receiver_type = gvn.type(receiver_node)->isa_oopptr();
836 // call_does_dispatch and vtable_index are out-parameters. They might be changed.
837 target = C->optimize_virtual_call(caller, jvms->bci(), klass, target, receiver_type,
838 is_virtual,
839 call_does_dispatch, vtable_index); // out-parameters
840 }
842 CallGenerator* cg = C->call_generator(target, vtable_index, call_does_dispatch, jvms, true, PROB_ALWAYS, true, true);
843 assert(!cg->is_late_inline() || cg->is_mh_late_inline(), "no late inline here");
844 if (cg != NULL && cg->is_inline())
845 return cg;
846 }
847 }
848 break;
850 default:
851 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
852 break;
853 }
854 return NULL;
855 }
858 //------------------------PredictedIntrinsicGenerator------------------------------
859 // Internal class which handles all predicted Intrinsic calls.
860 class PredictedIntrinsicGenerator : public CallGenerator {
861 CallGenerator* _intrinsic;
862 CallGenerator* _cg;
864 public:
865 PredictedIntrinsicGenerator(CallGenerator* intrinsic,
866 CallGenerator* cg)
867 : CallGenerator(cg->method())
868 {
869 _intrinsic = intrinsic;
870 _cg = cg;
871 }
873 virtual bool is_virtual() const { return true; }
874 virtual bool is_inlined() const { return true; }
875 virtual bool is_intrinsic() const { return true; }
877 virtual JVMState* generate(JVMState* jvms);
878 };
881 CallGenerator* CallGenerator::for_predicted_intrinsic(CallGenerator* intrinsic,
882 CallGenerator* cg) {
883 return new PredictedIntrinsicGenerator(intrinsic, cg);
884 }
887 JVMState* PredictedIntrinsicGenerator::generate(JVMState* jvms) {
888 GraphKit kit(jvms);
889 PhaseGVN& gvn = kit.gvn();
891 CompileLog* log = kit.C->log();
892 if (log != NULL) {
893 log->elem("predicted_intrinsic bci='%d' method='%d'",
894 jvms->bci(), log->identify(method()));
895 }
897 Node* slow_ctl = _intrinsic->generate_predicate(kit.sync_jvms());
898 if (kit.failing())
899 return NULL; // might happen because of NodeCountInliningCutoff
901 SafePointNode* slow_map = NULL;
902 JVMState* slow_jvms;
903 if (slow_ctl != NULL) {
904 PreserveJVMState pjvms(&kit);
905 kit.set_control(slow_ctl);
906 if (!kit.stopped()) {
907 slow_jvms = _cg->generate(kit.sync_jvms());
908 if (kit.failing())
909 return NULL; // might happen because of NodeCountInliningCutoff
910 assert(slow_jvms != NULL, "must be");
911 kit.add_exception_states_from(slow_jvms);
912 kit.set_map(slow_jvms->map());
913 if (!kit.stopped())
914 slow_map = kit.stop();
915 }
916 }
918 if (kit.stopped()) {
919 // Predicate is always false.
920 kit.set_jvms(slow_jvms);
921 return kit.transfer_exceptions_into_jvms();
922 }
924 // Generate intrinsic code:
925 JVMState* new_jvms = _intrinsic->generate(kit.sync_jvms());
926 if (new_jvms == NULL) {
927 // Intrinsic failed, so use slow code or make a direct call.
928 if (slow_map == NULL) {
929 CallGenerator* cg = CallGenerator::for_direct_call(method());
930 new_jvms = cg->generate(kit.sync_jvms());
931 } else {
932 kit.set_jvms(slow_jvms);
933 return kit.transfer_exceptions_into_jvms();
934 }
935 }
936 kit.add_exception_states_from(new_jvms);
937 kit.set_jvms(new_jvms);
939 // Need to merge slow and fast?
940 if (slow_map == NULL) {
941 // The fast path is the only path remaining.
942 return kit.transfer_exceptions_into_jvms();
943 }
945 if (kit.stopped()) {
946 // Intrinsic method threw an exception, so it's just the slow path after all.
947 kit.set_jvms(slow_jvms);
948 return kit.transfer_exceptions_into_jvms();
949 }
951 // Finish the diamond.
952 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization
953 RegionNode* region = new (kit.C) RegionNode(3);
954 region->init_req(1, kit.control());
955 region->init_req(2, slow_map->control());
956 kit.set_control(gvn.transform(region));
957 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO);
958 iophi->set_req(2, slow_map->i_o());
959 kit.set_i_o(gvn.transform(iophi));
960 kit.merge_memory(slow_map->merged_memory(), region, 2);
961 uint tos = kit.jvms()->stkoff() + kit.sp();
962 uint limit = slow_map->req();
963 for (uint i = TypeFunc::Parms; i < limit; i++) {
964 // Skip unused stack slots; fast forward to monoff();
965 if (i == tos) {
966 i = kit.jvms()->monoff();
967 if( i >= limit ) break;
968 }
969 Node* m = kit.map()->in(i);
970 Node* n = slow_map->in(i);
971 if (m != n) {
972 const Type* t = gvn.type(m)->meet(gvn.type(n));
973 Node* phi = PhiNode::make(region, m, t);
974 phi->set_req(2, n);
975 kit.map()->set_req(i, gvn.transform(phi));
976 }
977 }
978 return kit.transfer_exceptions_into_jvms();
979 }
981 //-------------------------UncommonTrapCallGenerator-----------------------------
982 // Internal class which handles all out-of-line calls checking receiver type.
983 class UncommonTrapCallGenerator : public CallGenerator {
984 Deoptimization::DeoptReason _reason;
985 Deoptimization::DeoptAction _action;
987 public:
988 UncommonTrapCallGenerator(ciMethod* m,
989 Deoptimization::DeoptReason reason,
990 Deoptimization::DeoptAction action)
991 : CallGenerator(m)
992 {
993 _reason = reason;
994 _action = action;
995 }
997 virtual bool is_virtual() const { ShouldNotReachHere(); return false; }
998 virtual bool is_trap() const { return true; }
1000 virtual JVMState* generate(JVMState* jvms);
1001 };
1004 CallGenerator*
1005 CallGenerator::for_uncommon_trap(ciMethod* m,
1006 Deoptimization::DeoptReason reason,
1007 Deoptimization::DeoptAction action) {
1008 return new UncommonTrapCallGenerator(m, reason, action);
1009 }
1012 JVMState* UncommonTrapCallGenerator::generate(JVMState* jvms) {
1013 GraphKit kit(jvms);
1014 // Take the trap with arguments pushed on the stack. (Cf. null_check_receiver).
1015 int nargs = method()->arg_size();
1016 kit.inc_sp(nargs);
1017 assert(nargs <= kit.sp() && kit.sp() <= jvms->stk_size(), "sane sp w/ args pushed");
1018 if (_reason == Deoptimization::Reason_class_check &&
1019 _action == Deoptimization::Action_maybe_recompile) {
1020 // Temp fix for 6529811
1021 // Don't allow uncommon_trap to override our decision to recompile in the event
1022 // of a class cast failure for a monomorphic call as it will never let us convert
1023 // the call to either bi-morphic or megamorphic and can lead to unc-trap loops
1024 bool keep_exact_action = true;
1025 kit.uncommon_trap(_reason, _action, NULL, "monomorphic vcall checkcast", false, keep_exact_action);
1026 } else {
1027 kit.uncommon_trap(_reason, _action);
1028 }
1029 return kit.transfer_exceptions_into_jvms();
1030 }
1032 // (Note: Moved hook_up_call to GraphKit::set_edges_for_java_call.)
1034 // (Node: Merged hook_up_exits into ParseGenerator::generate.)
1036 #define NODES_OVERHEAD_PER_METHOD (30.0)
1037 #define NODES_PER_BYTECODE (9.5)
1039 void WarmCallInfo::init(JVMState* call_site, ciMethod* call_method, ciCallProfile& profile, float prof_factor) {
1040 int call_count = profile.count();
1041 int code_size = call_method->code_size();
1043 // Expected execution count is based on the historical count:
1044 _count = call_count < 0 ? 1 : call_site->method()->scale_count(call_count, prof_factor);
1046 // Expected profit from inlining, in units of simple call-overheads.
1047 _profit = 1.0;
1049 // Expected work performed by the call in units of call-overheads.
1050 // %%% need an empirical curve fit for "work" (time in call)
1051 float bytecodes_per_call = 3;
1052 _work = 1.0 + code_size / bytecodes_per_call;
1054 // Expected size of compilation graph:
1055 // -XX:+PrintParseStatistics once reported:
1056 // Methods seen: 9184 Methods parsed: 9184 Nodes created: 1582391
1057 // Histogram of 144298 parsed bytecodes:
1058 // %%% Need an better predictor for graph size.
1059 _size = NODES_OVERHEAD_PER_METHOD + (NODES_PER_BYTECODE * code_size);
1060 }
1062 // is_cold: Return true if the node should never be inlined.
1063 // This is true if any of the key metrics are extreme.
1064 bool WarmCallInfo::is_cold() const {
1065 if (count() < WarmCallMinCount) return true;
1066 if (profit() < WarmCallMinProfit) return true;
1067 if (work() > WarmCallMaxWork) return true;
1068 if (size() > WarmCallMaxSize) return true;
1069 return false;
1070 }
1072 // is_hot: Return true if the node should be inlined immediately.
1073 // This is true if any of the key metrics are extreme.
1074 bool WarmCallInfo::is_hot() const {
1075 assert(!is_cold(), "eliminate is_cold cases before testing is_hot");
1076 if (count() >= HotCallCountThreshold) return true;
1077 if (profit() >= HotCallProfitThreshold) return true;
1078 if (work() <= HotCallTrivialWork) return true;
1079 if (size() <= HotCallTrivialSize) return true;
1080 return false;
1081 }
1083 // compute_heat:
1084 float WarmCallInfo::compute_heat() const {
1085 assert(!is_cold(), "compute heat only on warm nodes");
1086 assert(!is_hot(), "compute heat only on warm nodes");
1087 int min_size = MAX2(0, (int)HotCallTrivialSize);
1088 int max_size = MIN2(500, (int)WarmCallMaxSize);
1089 float method_size = (size() - min_size) / MAX2(1, max_size - min_size);
1090 float size_factor;
1091 if (method_size < 0.05) size_factor = 4; // 2 sigmas better than avg.
1092 else if (method_size < 0.15) size_factor = 2; // 1 sigma better than avg.
1093 else if (method_size < 0.5) size_factor = 1; // better than avg.
1094 else size_factor = 0.5; // worse than avg.
1095 return (count() * profit() * size_factor);
1096 }
1098 bool WarmCallInfo::warmer_than(WarmCallInfo* that) {
1099 assert(this != that, "compare only different WCIs");
1100 assert(this->heat() != 0 && that->heat() != 0, "call compute_heat 1st");
1101 if (this->heat() > that->heat()) return true;
1102 if (this->heat() < that->heat()) return false;
1103 assert(this->heat() == that->heat(), "no NaN heat allowed");
1104 // Equal heat. Break the tie some other way.
1105 if (!this->call() || !that->call()) return (address)this > (address)that;
1106 return this->call()->_idx > that->call()->_idx;
1107 }
1109 //#define UNINIT_NEXT ((WarmCallInfo*)badAddress)
1110 #define UNINIT_NEXT ((WarmCallInfo*)NULL)
1112 WarmCallInfo* WarmCallInfo::insert_into(WarmCallInfo* head) {
1113 assert(next() == UNINIT_NEXT, "not yet on any list");
1114 WarmCallInfo* prev_p = NULL;
1115 WarmCallInfo* next_p = head;
1116 while (next_p != NULL && next_p->warmer_than(this)) {
1117 prev_p = next_p;
1118 next_p = prev_p->next();
1119 }
1120 // Install this between prev_p and next_p.
1121 this->set_next(next_p);
1122 if (prev_p == NULL)
1123 head = this;
1124 else
1125 prev_p->set_next(this);
1126 return head;
1127 }
1129 WarmCallInfo* WarmCallInfo::remove_from(WarmCallInfo* head) {
1130 WarmCallInfo* prev_p = NULL;
1131 WarmCallInfo* next_p = head;
1132 while (next_p != this) {
1133 assert(next_p != NULL, "this must be in the list somewhere");
1134 prev_p = next_p;
1135 next_p = prev_p->next();
1136 }
1137 next_p = this->next();
1138 debug_only(this->set_next(UNINIT_NEXT));
1139 // Remove this from between prev_p and next_p.
1140 if (prev_p == NULL)
1141 head = next_p;
1142 else
1143 prev_p->set_next(next_p);
1144 return head;
1145 }
1147 WarmCallInfo WarmCallInfo::_always_hot(WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE(),
1148 WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE());
1149 WarmCallInfo WarmCallInfo::_always_cold(WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE(),
1150 WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE());
1152 WarmCallInfo* WarmCallInfo::always_hot() {
1153 assert(_always_hot.is_hot(), "must always be hot");
1154 return &_always_hot;
1155 }
1157 WarmCallInfo* WarmCallInfo::always_cold() {
1158 assert(_always_cold.is_cold(), "must always be cold");
1159 return &_always_cold;
1160 }
1163 #ifndef PRODUCT
1165 void WarmCallInfo::print() const {
1166 tty->print("%s : C=%6.1f P=%6.1f W=%6.1f S=%6.1f H=%6.1f -> %p",
1167 is_cold() ? "cold" : is_hot() ? "hot " : "warm",
1168 count(), profit(), work(), size(), compute_heat(), next());
1169 tty->cr();
1170 if (call() != NULL) call()->dump();
1171 }
1173 void print_wci(WarmCallInfo* ci) {
1174 ci->print();
1175 }
1177 void WarmCallInfo::print_all() const {
1178 for (const WarmCallInfo* p = this; p != NULL; p = p->next())
1179 p->print();
1180 }
1182 int WarmCallInfo::count_all() const {
1183 int cnt = 0;
1184 for (const WarmCallInfo* p = this; p != NULL; p = p->next())
1185 cnt++;
1186 return cnt;
1187 }
1189 #endif //PRODUCT