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 */ 24 25 #include "precompiled.hpp" 26 #include "ci/bcEscapeAnalyzer.hpp" 27 #include "ci/ciCallSite.hpp" 28 #include "ci/ciCPCache.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" 42 43 44 // Utility function. 45 const TypeFunc* CallGenerator::tf() const { 46 return TypeFunc::make(method()); 47 } 48 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; 55 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 } 64 65 virtual bool is_parse() const { return true; } 66 virtual JVMState* generate(JVMState* jvms); 67 int is_osr() { return _is_osr; } 68 69 }; 70 71 JVMState* ParseGenerator::generate(JVMState* jvms) { 72 Compile* C = Compile::current(); 73 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 } 78 79 if (C->failing()) { 80 return NULL; // bailing out of the compile; do not try to parse 81 } 82 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 92 93 GraphKit& exits = parser.exits(); 94 95 if (C->failing()) { 96 while (exits.pop_exception_state() != NULL) ; 97 return NULL; 98 } 99 100 assert(exits.jvms()->same_calls_as(jvms), "sanity"); 101 102 // Simply return the exit state of the parser, 103 // augmented by any exceptional states. 104 return exits.transfer_exceptions_into_jvms(); 105 } 106 107 //---------------------------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; 115 116 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); 123 124 CallStaticJavaNode* call_node() const { return _call_node; } 125 }; 126 127 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(); 132 133 if (kit.C->log() != NULL) { 134 kit.C->log()->elem("direct_call bci='%d'", jvms->bci()); 135 } 136 137 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 } 160 161 //--------------------------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 == methodOopDesc::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 }; 176 177 JVMState* VirtualCallGenerator::generate(JVMState* jvms) { 178 GraphKit kit(jvms); 179 Node* receiver = kit.argument(0); 180 181 if (kit.C->log() != NULL) { 182 kit.C->log()->elem("virtual_call bci='%d'", jvms->bci()); 183 } 184 185 // 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 } 197 198 // 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 || !os::zero_page_read_protected() || 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 } 216 217 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 == methodOopDesc::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); 229 230 // 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 } 236 237 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 } 241 242 // 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 } 251 252 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 } 256 257 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 } 262 263 // Allow inlining decisions to be delayed 264 class LateInlineCallGenerator : public DirectCallGenerator { 265 protected: 266 CallGenerator* _inline_cg; 267 268 virtual bool do_late_inline_check(JVMState* jvms) { return true; } 269 270 public: 271 LateInlineCallGenerator(ciMethod* method, CallGenerator* inline_cg) : 272 DirectCallGenerator(method, true), _inline_cg(inline_cg) {} 273 274 virtual bool is_late_inline() const { return true; } 275 276 // Convert the CallStaticJava into an inline 277 virtual void do_late_inline(); 278 279 virtual JVMState* generate(JVMState* jvms) { 280 Compile *C = Compile::current(); 281 C->print_inlining_skip(this); 282 283 // 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 } 288 289 // 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 } 295 296 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 } 302 303 }; 304 305 void LateInlineCallGenerator::do_late_inline() { 306 // Can't inline it 307 if (call_node() == NULL || call_node()->outcnt() == 0 || 308 call_node()->in(0) == NULL || call_node()->in(0)->is_top()) { 309 return; 310 } 311 312 const TypeTuple *r = call_node()->tf()->domain(); 313 for (int i1 = 0; i1 < method()->arg_size(); i1++) { 314 if (call_node()->in(TypeFunc::Parms + i1)->is_top() && r->field_at(TypeFunc::Parms + i1) != Type::HALF) { 315 assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing"); 316 return; 317 } 318 } 319 320 if (call_node()->in(TypeFunc::Memory)->is_top()) { 321 assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing"); 322 return; 323 } 324 325 CallStaticJavaNode* call = call_node(); 326 327 // Make a clone of the JVMState that appropriate to use for driving a parse 328 Compile* C = Compile::current(); 329 JVMState* jvms = call->jvms()->clone_shallow(C); 330 uint size = call->req(); 331 SafePointNode* map = new (C) SafePointNode(size, jvms); 332 for (uint i1 = 0; i1 < size; i1++) { 333 map->init_req(i1, call->in(i1)); 334 } 335 336 // Make sure the state is a MergeMem for parsing. 337 if (!map->in(TypeFunc::Memory)->is_MergeMem()) { 338 Node* mem = MergeMemNode::make(C, map->in(TypeFunc::Memory)); 339 C->initial_gvn()->set_type_bottom(mem); 340 map->set_req(TypeFunc::Memory, mem); 341 } 342 343 // Make enough space for the expression stack and transfer the incoming arguments 344 int nargs = method()->arg_size(); 345 jvms->set_map(map); 346 map->ensure_stack(jvms, jvms->method()->max_stack()); 347 if (nargs > 0) { 348 for (int i1 = 0; i1 < nargs; i1++) { 349 map->set_req(i1 + jvms->argoff(), call->in(TypeFunc::Parms + i1)); 350 } 351 } 352 353 if (!do_late_inline_check(jvms)) { 354 map->disconnect_inputs(NULL, C); 355 return; 356 } 357 358 C->print_inlining_insert(this); 359 360 CompileLog* log = C->log(); 361 if (log != NULL) { 362 log->head("late_inline method='%d'", log->identify(method())); 363 JVMState* p = jvms; 364 while (p != NULL) { 365 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 366 p = p->caller(); 367 } 368 log->tail("late_inline"); 369 } 370 371 // Setup default node notes to be picked up by the inlining 372 Node_Notes* old_nn = C->default_node_notes(); 373 if (old_nn != NULL) { 374 Node_Notes* entry_nn = old_nn->clone(C); 375 entry_nn->set_jvms(jvms); 376 C->set_default_node_notes(entry_nn); 377 } 378 379 // Now perform the inling using the synthesized JVMState 380 JVMState* new_jvms = _inline_cg->generate(jvms); 381 if (new_jvms == NULL) return; // no change 382 if (C->failing()) return; 383 384 // Capture any exceptional control flow 385 GraphKit kit(new_jvms); 386 387 // Find the result object 388 Node* result = C->top(); 389 int result_size = method()->return_type()->size(); 390 if (result_size != 0 && !kit.stopped()) { 391 result = (result_size == 1) ? kit.pop() : kit.pop_pair(); 392 } 393 394 C->set_has_loops(C->has_loops() || _inline_cg->method()->has_loops()); 395 C->env()->notice_inlined_method(_inline_cg->method()); 396 C->set_inlining_progress(true); 397 398 kit.replace_call(call, result, true); 399 } 400 401 402 CallGenerator* CallGenerator::for_late_inline(ciMethod* method, CallGenerator* inline_cg) { 403 return new LateInlineCallGenerator(method, inline_cg); 404 } 405 406 class LateInlineMHCallGenerator : public LateInlineCallGenerator { 407 ciMethod* _caller; 408 int _attempt; 409 bool _input_not_const; 410 411 virtual bool do_late_inline_check(JVMState* jvms); 412 virtual bool already_attempted() const { return _attempt > 0; } 413 414 public: 415 LateInlineMHCallGenerator(ciMethod* caller, ciMethod* callee, bool input_not_const) : 416 LateInlineCallGenerator(callee, NULL), _caller(caller), _attempt(0), _input_not_const(input_not_const) {} 417 418 virtual bool is_mh_late_inline() const { return true; } 419 420 virtual JVMState* generate(JVMState* jvms) { 421 JVMState* new_jvms = LateInlineCallGenerator::generate(jvms); 422 if (_input_not_const) { 423 // inlining won't be possible so no need to enqueue right now. 424 call_node()->set_generator(this); 425 } else { 426 Compile::current()->add_late_inline(this); 427 } 428 return new_jvms; 429 } 430 431 virtual void print_inlining_late(const char* msg) { 432 if (!_input_not_const) return; 433 LateInlineCallGenerator::print_inlining_late(msg); 434 } 435 }; 436 437 bool LateInlineMHCallGenerator::do_late_inline_check(JVMState* jvms) { 438 439 CallGenerator* cg = for_method_handle_inline(jvms, _caller, method(), _input_not_const); 440 441 if (!_input_not_const) { 442 _attempt++; 443 } 444 445 if (cg != NULL) { 446 assert(!cg->is_late_inline() && cg->is_inline(), "we're doing late inlining"); 447 _inline_cg = cg; 448 Compile::current()->dec_number_of_mh_late_inlines(); 449 return true; 450 } 451 452 call_node()->set_generator(this); 453 return false; 454 } 455 456 CallGenerator* CallGenerator::for_mh_late_inline(ciMethod* caller, ciMethod* callee, bool input_not_const) { 457 Compile::current()->inc_number_of_mh_late_inlines(); 458 CallGenerator* cg = new LateInlineMHCallGenerator(caller, callee, input_not_const); 459 return cg; 460 } 461 462 class LateInlineStringCallGenerator : public LateInlineCallGenerator { 463 464 public: 465 LateInlineStringCallGenerator(ciMethod* method, CallGenerator* inline_cg) : 466 LateInlineCallGenerator(method, inline_cg) {} 467 468 virtual JVMState* generate(JVMState* jvms) { 469 Compile *C = Compile::current(); 470 C->print_inlining_skip(this); 471 472 C->add_string_late_inline(this); 473 474 JVMState* new_jvms = DirectCallGenerator::generate(jvms); 475 return new_jvms; 476 } 477 }; 478 479 CallGenerator* CallGenerator::for_string_late_inline(ciMethod* method, CallGenerator* inline_cg) { 480 return new LateInlineStringCallGenerator(method, inline_cg); 481 } 482 483 484 //---------------------------WarmCallGenerator-------------------------------- 485 // Internal class which handles initial deferral of inlining decisions. 486 class WarmCallGenerator : public CallGenerator { 487 WarmCallInfo* _call_info; 488 CallGenerator* _if_cold; 489 CallGenerator* _if_hot; 490 bool _is_virtual; // caches virtuality of if_cold 491 bool _is_inline; // caches inline-ness of if_hot 492 493 public: 494 WarmCallGenerator(WarmCallInfo* ci, 495 CallGenerator* if_cold, 496 CallGenerator* if_hot) 497 : CallGenerator(if_cold->method()) 498 { 499 assert(method() == if_hot->method(), "consistent choices"); 500 _call_info = ci; 501 _if_cold = if_cold; 502 _if_hot = if_hot; 503 _is_virtual = if_cold->is_virtual(); 504 _is_inline = if_hot->is_inline(); 505 } 506 507 virtual bool is_inline() const { return _is_inline; } 508 virtual bool is_virtual() const { return _is_virtual; } 509 virtual bool is_deferred() const { return true; } 510 511 virtual JVMState* generate(JVMState* jvms); 512 }; 513 514 515 CallGenerator* CallGenerator::for_warm_call(WarmCallInfo* ci, 516 CallGenerator* if_cold, 517 CallGenerator* if_hot) { 518 return new WarmCallGenerator(ci, if_cold, if_hot); 519 } 520 521 JVMState* WarmCallGenerator::generate(JVMState* jvms) { 522 Compile* C = Compile::current(); 523 if (C->log() != NULL) { 524 C->log()->elem("warm_call bci='%d'", jvms->bci()); 525 } 526 jvms = _if_cold->generate(jvms); 527 if (jvms != NULL) { 528 Node* m = jvms->map()->control(); 529 if (m->is_CatchProj()) m = m->in(0); else m = C->top(); 530 if (m->is_Catch()) m = m->in(0); else m = C->top(); 531 if (m->is_Proj()) m = m->in(0); else m = C->top(); 532 if (m->is_CallJava()) { 533 _call_info->set_call(m->as_Call()); 534 _call_info->set_hot_cg(_if_hot); 535 #ifndef PRODUCT 536 if (PrintOpto || PrintOptoInlining) { 537 tty->print_cr("Queueing for warm inlining at bci %d:", jvms->bci()); 538 tty->print("WCI: "); 539 _call_info->print(); 540 } 541 #endif 542 _call_info->set_heat(_call_info->compute_heat()); 543 C->set_warm_calls(_call_info->insert_into(C->warm_calls())); 544 } 545 } 546 return jvms; 547 } 548 549 void WarmCallInfo::make_hot() { 550 Unimplemented(); 551 } 552 553 void WarmCallInfo::make_cold() { 554 // No action: Just dequeue. 555 } 556 557 558 //------------------------PredictedCallGenerator------------------------------ 559 // Internal class which handles all out-of-line calls checking receiver type. 560 class PredictedCallGenerator : public CallGenerator { 561 ciKlass* _predicted_receiver; 562 CallGenerator* _if_missed; 563 CallGenerator* _if_hit; 564 float _hit_prob; 565 566 public: 567 PredictedCallGenerator(ciKlass* predicted_receiver, 568 CallGenerator* if_missed, 569 CallGenerator* if_hit, float hit_prob) 570 : CallGenerator(if_missed->method()) 571 { 572 // The call profile data may predict the hit_prob as extreme as 0 or 1. 573 // Remove the extremes values from the range. 574 if (hit_prob > PROB_MAX) hit_prob = PROB_MAX; 575 if (hit_prob < PROB_MIN) hit_prob = PROB_MIN; 576 577 _predicted_receiver = predicted_receiver; 578 _if_missed = if_missed; 579 _if_hit = if_hit; 580 _hit_prob = hit_prob; 581 } 582 583 virtual bool is_virtual() const { return true; } 584 virtual bool is_inline() const { return _if_hit->is_inline(); } 585 virtual bool is_deferred() const { return _if_hit->is_deferred(); } 586 587 virtual JVMState* generate(JVMState* jvms); 588 }; 589 590 591 CallGenerator* CallGenerator::for_predicted_call(ciKlass* predicted_receiver, 592 CallGenerator* if_missed, 593 CallGenerator* if_hit, 594 float hit_prob) { 595 return new PredictedCallGenerator(predicted_receiver, if_missed, if_hit, hit_prob); 596 } 597 598 599 JVMState* PredictedCallGenerator::generate(JVMState* jvms) { 600 GraphKit kit(jvms); 601 PhaseGVN& gvn = kit.gvn(); 602 // We need an explicit receiver null_check before checking its type. 603 // We share a map with the caller, so his JVMS gets adjusted. 604 Node* receiver = kit.argument(0); 605 606 CompileLog* log = kit.C->log(); 607 if (log != NULL) { 608 log->elem("predicted_call bci='%d' klass='%d'", 609 jvms->bci(), log->identify(_predicted_receiver)); 610 } 611 612 receiver = kit.null_check_receiver_before_call(method()); 613 if (kit.stopped()) { 614 return kit.transfer_exceptions_into_jvms(); 615 } 616 617 // Make a copy of the replaced nodes in case we need to restore them 618 ReplacedNodes replaced_nodes = kit.map()->replaced_nodes(); 619 replaced_nodes.clone(); 620 621 Node* exact_receiver = receiver; // will get updated in place... 622 Node* slow_ctl = kit.type_check_receiver(receiver, 623 _predicted_receiver, _hit_prob, 624 &exact_receiver); 625 626 SafePointNode* slow_map = NULL; 627 JVMState* slow_jvms; 628 { PreserveJVMState pjvms(&kit); 629 kit.set_control(slow_ctl); 630 if (!kit.stopped()) { 631 slow_jvms = _if_missed->generate(kit.sync_jvms()); 632 if (kit.failing()) 633 return NULL; // might happen because of NodeCountInliningCutoff 634 assert(slow_jvms != NULL, "must be"); 635 kit.add_exception_states_from(slow_jvms); 636 kit.set_map(slow_jvms->map()); 637 if (!kit.stopped()) 638 slow_map = kit.stop(); 639 } 640 } 641 642 if (kit.stopped()) { 643 // Instance exactly does not matches the desired type. 644 kit.set_jvms(slow_jvms); 645 return kit.transfer_exceptions_into_jvms(); 646 } 647 648 // fall through if the instance exactly matches the desired type 649 kit.replace_in_map(receiver, exact_receiver); 650 651 // Make the hot call: 652 JVMState* new_jvms = _if_hit->generate(kit.sync_jvms()); 653 if (new_jvms == NULL) { 654 // Inline failed, so make a direct call. 655 assert(_if_hit->is_inline(), "must have been a failed inline"); 656 CallGenerator* cg = CallGenerator::for_direct_call(_if_hit->method()); 657 new_jvms = cg->generate(kit.sync_jvms()); 658 } 659 kit.add_exception_states_from(new_jvms); 660 kit.set_jvms(new_jvms); 661 662 // Need to merge slow and fast? 663 if (slow_map == NULL) { 664 // The fast path is the only path remaining. 665 return kit.transfer_exceptions_into_jvms(); 666 } 667 668 if (kit.stopped()) { 669 // Inlined method threw an exception, so it's just the slow path after all. 670 kit.set_jvms(slow_jvms); 671 return kit.transfer_exceptions_into_jvms(); 672 } 673 674 // There are 2 branches and the replaced nodes are only valid on 675 // one: restore the replaced nodes to what they were before the 676 // branch. 677 kit.map()->set_replaced_nodes(replaced_nodes); 678 679 // Finish the diamond. 680 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization 681 RegionNode* region = new (kit.C) RegionNode(3); 682 region->init_req(1, kit.control()); 683 region->init_req(2, slow_map->control()); 684 kit.set_control(gvn.transform(region)); 685 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO); 686 iophi->set_req(2, slow_map->i_o()); 687 kit.set_i_o(gvn.transform(iophi)); 688 kit.merge_memory(slow_map->merged_memory(), region, 2); 689 uint tos = kit.jvms()->stkoff() + kit.sp(); 690 uint limit = slow_map->req(); 691 for (uint i = TypeFunc::Parms; i < limit; i++) { 692 // Skip unused stack slots; fast forward to monoff(); 693 if (i == tos) { 694 i = kit.jvms()->monoff(); 695 if( i >= limit ) break; 696 } 697 Node* m = kit.map()->in(i); 698 Node* n = slow_map->in(i); 699 if (m != n) { 700 const Type* t = gvn.type(m)->meet(gvn.type(n)); 701 Node* phi = PhiNode::make(region, m, t); 702 phi->set_req(2, n); 703 kit.map()->set_req(i, gvn.transform(phi)); 704 } 705 } 706 return kit.transfer_exceptions_into_jvms(); 707 } 708 709 710 CallGenerator* CallGenerator::for_method_handle_call(JVMState* jvms, ciMethod* caller, ciMethod* callee, bool delayed_forbidden) { 711 assert(callee->is_method_handle_intrinsic() || 712 callee->is_compiled_lambda_form(), "for_method_handle_call mismatch"); 713 bool input_not_const; 714 CallGenerator* cg = CallGenerator::for_method_handle_inline(jvms, caller, callee, input_not_const); 715 Compile* C = Compile::current(); 716 if (cg != NULL) { 717 if (!delayed_forbidden && AlwaysIncrementalInline) { 718 return CallGenerator::for_late_inline(callee, cg); 719 } else { 720 return cg; 721 } 722 } 723 int bci = jvms->bci(); 724 ciCallProfile profile = caller->call_profile_at_bci(bci); 725 int call_site_count = caller->scale_count(profile.count()); 726 727 if (IncrementalInline && call_site_count > 0 && 728 (input_not_const || !C->inlining_incrementally() || C->over_inlining_cutoff())) { 729 return CallGenerator::for_mh_late_inline(caller, callee, input_not_const); 730 } else { 731 // Out-of-line call. 732 return CallGenerator::for_direct_call(callee); 733 } 734 } 735 736 CallGenerator* CallGenerator::for_method_handle_inline(JVMState* jvms, ciMethod* caller, ciMethod* callee, bool& input_not_const) { 737 GraphKit kit(jvms); 738 PhaseGVN& gvn = kit.gvn(); 739 Compile* C = kit.C; 740 vmIntrinsics::ID iid = callee->intrinsic_id(); 741 input_not_const = true; 742 switch (iid) { 743 case vmIntrinsics::_invokeBasic: 744 { 745 // Get MethodHandle receiver: 746 Node* receiver = kit.argument(0); 747 if (receiver->Opcode() == Op_ConP) { 748 input_not_const = false; 749 const TypeOopPtr* oop_ptr = receiver->bottom_type()->is_oopptr(); 750 ciMethod* target = oop_ptr->const_oop()->as_method_handle()->get_vmtarget(); 751 guarantee(!target->is_method_handle_intrinsic(), "should not happen"); // XXX remove 752 const int vtable_index = methodOopDesc::invalid_vtable_index; 753 CallGenerator* cg = C->call_generator(target, vtable_index, false, jvms, true, PROB_ALWAYS, true, true); 754 assert(cg == NULL || !cg->is_late_inline() || cg->is_mh_late_inline(), "no late inline here"); 755 if (cg != NULL && cg->is_inline()) 756 return cg; 757 } 758 } 759 break; 760 761 case vmIntrinsics::_linkToVirtual: 762 case vmIntrinsics::_linkToStatic: 763 case vmIntrinsics::_linkToSpecial: 764 case vmIntrinsics::_linkToInterface: 765 { 766 // Get MemberName argument: 767 Node* member_name = kit.argument(callee->arg_size() - 1); 768 if (member_name->Opcode() == Op_ConP) { 769 input_not_const = false; 770 const TypeOopPtr* oop_ptr = member_name->bottom_type()->is_oopptr(); 771 ciMethod* target = oop_ptr->const_oop()->as_member_name()->get_vmtarget(); 772 773 // In lamda forms we erase signature types to avoid resolving issues 774 // involving class loaders. When we optimize a method handle invoke 775 // to a direct call we must cast the receiver and arguments to its 776 // actual types. 777 ciSignature* signature = target->signature(); 778 const int receiver_skip = target->is_static() ? 0 : 1; 779 // Cast receiver to its type. 780 if (!target->is_static()) { 781 Node* arg = kit.argument(0); 782 const TypeOopPtr* arg_type = arg->bottom_type()->isa_oopptr(); 783 const Type* sig_type = TypeOopPtr::make_from_klass(signature->accessing_klass()); 784 if (arg_type != NULL && !arg_type->higher_equal(sig_type)) { 785 Node* cast_obj = gvn.transform(new (C) CheckCastPPNode(kit.control(), arg, sig_type)); 786 kit.set_argument(0, cast_obj); 787 } 788 } 789 // Cast reference arguments to its type. 790 for (int i = 0; i < signature->count(); i++) { 791 ciType* t = signature->type_at(i); 792 if (t->is_klass()) { 793 Node* arg = kit.argument(receiver_skip + i); 794 const TypeOopPtr* arg_type = arg->bottom_type()->isa_oopptr(); 795 const Type* sig_type = TypeOopPtr::make_from_klass(t->as_klass()); 796 if (arg_type != NULL && !arg_type->higher_equal(sig_type)) { 797 Node* cast_obj = gvn.transform(new (C) CheckCastPPNode(kit.control(), arg, sig_type)); 798 kit.set_argument(receiver_skip + i, cast_obj); 799 } 800 } 801 } 802 803 // Try to get the most accurate receiver type 804 const bool is_virtual = (iid == vmIntrinsics::_linkToVirtual); 805 const bool is_virtual_or_interface = (is_virtual || iid == vmIntrinsics::_linkToInterface); 806 int vtable_index = methodOopDesc::invalid_vtable_index; 807 bool call_does_dispatch = false; 808 809 if (is_virtual_or_interface) { 810 ciInstanceKlass* klass = target->holder(); 811 Node* receiver_node = kit.argument(0); 812 const TypeOopPtr* receiver_type = gvn.type(receiver_node)->isa_oopptr(); 813 // call_does_dispatch and vtable_index are out-parameters. They might be changed. 814 target = C->optimize_virtual_call(caller, jvms->bci(), klass, target, receiver_type, 815 is_virtual, 816 call_does_dispatch, vtable_index); // out-parameters 817 } 818 819 CallGenerator* cg = C->call_generator(target, vtable_index, call_does_dispatch, jvms, true, PROB_ALWAYS, true, true); 820 assert(cg == NULL || !cg->is_late_inline() || cg->is_mh_late_inline(), "no late inline here"); 821 if (cg != NULL && cg->is_inline()) 822 return cg; 823 } 824 } 825 break; 826 827 default: 828 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 829 break; 830 } 831 return NULL; 832 } 833 834 835 //------------------------PredictedIntrinsicGenerator------------------------------ 836 // Internal class which handles all predicted Intrinsic calls. 837 class PredictedIntrinsicGenerator : public CallGenerator { 838 CallGenerator* _intrinsic; 839 CallGenerator* _cg; 840 841 public: 842 PredictedIntrinsicGenerator(CallGenerator* intrinsic, 843 CallGenerator* cg) 844 : CallGenerator(cg->method()) 845 { 846 _intrinsic = intrinsic; 847 _cg = cg; 848 } 849 850 virtual bool is_virtual() const { return true; } 851 virtual bool is_inlined() const { return true; } 852 virtual bool is_intrinsic() const { return true; } 853 854 virtual JVMState* generate(JVMState* jvms); 855 }; 856 857 858 CallGenerator* CallGenerator::for_predicted_intrinsic(CallGenerator* intrinsic, 859 CallGenerator* cg) { 860 return new PredictedIntrinsicGenerator(intrinsic, cg); 861 } 862 863 864 JVMState* PredictedIntrinsicGenerator::generate(JVMState* jvms) { 865 GraphKit kit(jvms); 866 PhaseGVN& gvn = kit.gvn(); 867 868 CompileLog* log = kit.C->log(); 869 if (log != NULL) { 870 log->elem("predicted_intrinsic bci='%d' method='%d'", 871 jvms->bci(), log->identify(method())); 872 } 873 874 Node* slow_ctl = _intrinsic->generate_predicate(kit.sync_jvms()); 875 if (kit.failing()) 876 return NULL; // might happen because of NodeCountInliningCutoff 877 878 SafePointNode* slow_map = NULL; 879 JVMState* slow_jvms; 880 if (slow_ctl != NULL) { 881 PreserveJVMState pjvms(&kit); 882 kit.set_control(slow_ctl); 883 if (!kit.stopped()) { 884 slow_jvms = _cg->generate(kit.sync_jvms()); 885 if (kit.failing()) 886 return NULL; // might happen because of NodeCountInliningCutoff 887 assert(slow_jvms != NULL, "must be"); 888 kit.add_exception_states_from(slow_jvms); 889 kit.set_map(slow_jvms->map()); 890 if (!kit.stopped()) 891 slow_map = kit.stop(); 892 } 893 } 894 895 if (kit.stopped()) { 896 // Predicate is always false. 897 kit.set_jvms(slow_jvms); 898 return kit.transfer_exceptions_into_jvms(); 899 } 900 901 // Generate intrinsic code: 902 JVMState* new_jvms = _intrinsic->generate(kit.sync_jvms()); 903 if (new_jvms == NULL) { 904 // Intrinsic failed, so use slow code or make a direct call. 905 if (slow_map == NULL) { 906 CallGenerator* cg = CallGenerator::for_direct_call(method()); 907 new_jvms = cg->generate(kit.sync_jvms()); 908 } else { 909 kit.set_jvms(slow_jvms); 910 return kit.transfer_exceptions_into_jvms(); 911 } 912 } 913 kit.add_exception_states_from(new_jvms); 914 kit.set_jvms(new_jvms); 915 916 // Need to merge slow and fast? 917 if (slow_map == NULL) { 918 // The fast path is the only path remaining. 919 return kit.transfer_exceptions_into_jvms(); 920 } 921 922 if (kit.stopped()) { 923 // Intrinsic method threw an exception, so it's just the slow path after all. 924 kit.set_jvms(slow_jvms); 925 return kit.transfer_exceptions_into_jvms(); 926 } 927 928 // Finish the diamond. 929 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization 930 RegionNode* region = new (kit.C) RegionNode(3); 931 region->init_req(1, kit.control()); 932 region->init_req(2, slow_map->control()); 933 kit.set_control(gvn.transform(region)); 934 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO); 935 iophi->set_req(2, slow_map->i_o()); 936 kit.set_i_o(gvn.transform(iophi)); 937 kit.merge_memory(slow_map->merged_memory(), region, 2); 938 uint tos = kit.jvms()->stkoff() + kit.sp(); 939 uint limit = slow_map->req(); 940 for (uint i = TypeFunc::Parms; i < limit; i++) { 941 // Skip unused stack slots; fast forward to monoff(); 942 if (i == tos) { 943 i = kit.jvms()->monoff(); 944 if( i >= limit ) break; 945 } 946 Node* m = kit.map()->in(i); 947 Node* n = slow_map->in(i); 948 if (m != n) { 949 const Type* t = gvn.type(m)->meet(gvn.type(n)); 950 Node* phi = PhiNode::make(region, m, t); 951 phi->set_req(2, n); 952 kit.map()->set_req(i, gvn.transform(phi)); 953 } 954 } 955 return kit.transfer_exceptions_into_jvms(); 956 } 957 958 //-------------------------UncommonTrapCallGenerator----------------------------- 959 // Internal class which handles all out-of-line calls checking receiver type. 960 class UncommonTrapCallGenerator : public CallGenerator { 961 Deoptimization::DeoptReason _reason; 962 Deoptimization::DeoptAction _action; 963 964 public: 965 UncommonTrapCallGenerator(ciMethod* m, 966 Deoptimization::DeoptReason reason, 967 Deoptimization::DeoptAction action) 968 : CallGenerator(m) 969 { 970 _reason = reason; 971 _action = action; 972 } 973 974 virtual bool is_virtual() const { ShouldNotReachHere(); return false; } 975 virtual bool is_trap() const { return true; } 976 977 virtual JVMState* generate(JVMState* jvms); 978 }; 979 980 981 CallGenerator* 982 CallGenerator::for_uncommon_trap(ciMethod* m, 983 Deoptimization::DeoptReason reason, 984 Deoptimization::DeoptAction action) { 985 return new UncommonTrapCallGenerator(m, reason, action); 986 } 987 988 989 JVMState* UncommonTrapCallGenerator::generate(JVMState* jvms) { 990 GraphKit kit(jvms); 991 // Take the trap with arguments pushed on the stack. (Cf. null_check_receiver). 992 int nargs = method()->arg_size(); 993 kit.inc_sp(nargs); 994 assert(nargs <= kit.sp() && kit.sp() <= jvms->stk_size(), "sane sp w/ args pushed"); 995 if (_reason == Deoptimization::Reason_class_check && 996 _action == Deoptimization::Action_maybe_recompile) { 997 // Temp fix for 6529811 998 // Don't allow uncommon_trap to override our decision to recompile in the event 999 // of a class cast failure for a monomorphic call as it will never let us convert 1000 // the call to either bi-morphic or megamorphic and can lead to unc-trap loops 1001 bool keep_exact_action = true; 1002 kit.uncommon_trap(_reason, _action, NULL, "monomorphic vcall checkcast", false, keep_exact_action); 1003 } else { 1004 kit.uncommon_trap(_reason, _action); 1005 } 1006 return kit.transfer_exceptions_into_jvms(); 1007 } 1008 1009 // (Note: Moved hook_up_call to GraphKit::set_edges_for_java_call.) 1010 1011 // (Node: Merged hook_up_exits into ParseGenerator::generate.) 1012 1013 #define NODES_OVERHEAD_PER_METHOD (30.0) 1014 #define NODES_PER_BYTECODE (9.5) 1015 1016 void WarmCallInfo::init(JVMState* call_site, ciMethod* call_method, ciCallProfile& profile, float prof_factor) { 1017 int call_count = profile.count(); 1018 int code_size = call_method->code_size(); 1019 1020 // Expected execution count is based on the historical count: 1021 _count = call_count < 0 ? 1 : call_site->method()->scale_count(call_count, prof_factor); 1022 1023 // Expected profit from inlining, in units of simple call-overheads. 1024 _profit = 1.0; 1025 1026 // Expected work performed by the call in units of call-overheads. 1027 // %%% need an empirical curve fit for "work" (time in call) 1028 float bytecodes_per_call = 3; 1029 _work = 1.0 + code_size / bytecodes_per_call; 1030 1031 // Expected size of compilation graph: 1032 // -XX:+PrintParseStatistics once reported: 1033 // Methods seen: 9184 Methods parsed: 9184 Nodes created: 1582391 1034 // Histogram of 144298 parsed bytecodes: 1035 // %%% Need an better predictor for graph size. 1036 _size = NODES_OVERHEAD_PER_METHOD + (NODES_PER_BYTECODE * code_size); 1037 } 1038 1039 // is_cold: Return true if the node should never be inlined. 1040 // This is true if any of the key metrics are extreme. 1041 bool WarmCallInfo::is_cold() const { 1042 if (count() < WarmCallMinCount) return true; 1043 if (profit() < WarmCallMinProfit) return true; 1044 if (work() > WarmCallMaxWork) return true; 1045 if (size() > WarmCallMaxSize) return true; 1046 return false; 1047 } 1048 1049 // is_hot: Return true if the node should be inlined immediately. 1050 // This is true if any of the key metrics are extreme. 1051 bool WarmCallInfo::is_hot() const { 1052 assert(!is_cold(), "eliminate is_cold cases before testing is_hot"); 1053 if (count() >= HotCallCountThreshold) return true; 1054 if (profit() >= HotCallProfitThreshold) return true; 1055 if (work() <= HotCallTrivialWork) return true; 1056 if (size() <= HotCallTrivialSize) return true; 1057 return false; 1058 } 1059 1060 // compute_heat: 1061 float WarmCallInfo::compute_heat() const { 1062 assert(!is_cold(), "compute heat only on warm nodes"); 1063 assert(!is_hot(), "compute heat only on warm nodes"); 1064 int min_size = MAX2(0, (int)HotCallTrivialSize); 1065 int max_size = MIN2(500, (int)WarmCallMaxSize); 1066 float method_size = (size() - min_size) / MAX2(1, max_size - min_size); 1067 float size_factor; 1068 if (method_size < 0.05) size_factor = 4; // 2 sigmas better than avg. 1069 else if (method_size < 0.15) size_factor = 2; // 1 sigma better than avg. 1070 else if (method_size < 0.5) size_factor = 1; // better than avg. 1071 else size_factor = 0.5; // worse than avg. 1072 return (count() * profit() * size_factor); 1073 } 1074 1075 bool WarmCallInfo::warmer_than(WarmCallInfo* that) { 1076 assert(this != that, "compare only different WCIs"); 1077 assert(this->heat() != 0 && that->heat() != 0, "call compute_heat 1st"); 1078 if (this->heat() > that->heat()) return true; 1079 if (this->heat() < that->heat()) return false; 1080 assert(this->heat() == that->heat(), "no NaN heat allowed"); 1081 // Equal heat. Break the tie some other way. 1082 if (!this->call() || !that->call()) return (address)this > (address)that; 1083 return this->call()->_idx > that->call()->_idx; 1084 } 1085 1086 //#define UNINIT_NEXT ((WarmCallInfo*)badAddress) 1087 #define UNINIT_NEXT ((WarmCallInfo*)NULL) 1088 1089 WarmCallInfo* WarmCallInfo::insert_into(WarmCallInfo* head) { 1090 assert(next() == UNINIT_NEXT, "not yet on any list"); 1091 WarmCallInfo* prev_p = NULL; 1092 WarmCallInfo* next_p = head; 1093 while (next_p != NULL && next_p->warmer_than(this)) { 1094 prev_p = next_p; 1095 next_p = prev_p->next(); 1096 } 1097 // Install this between prev_p and next_p. 1098 this->set_next(next_p); 1099 if (prev_p == NULL) 1100 head = this; 1101 else 1102 prev_p->set_next(this); 1103 return head; 1104 } 1105 1106 WarmCallInfo* WarmCallInfo::remove_from(WarmCallInfo* head) { 1107 WarmCallInfo* prev_p = NULL; 1108 WarmCallInfo* next_p = head; 1109 while (next_p != this) { 1110 assert(next_p != NULL, "this must be in the list somewhere"); 1111 prev_p = next_p; 1112 next_p = prev_p->next(); 1113 } 1114 next_p = this->next(); 1115 debug_only(this->set_next(UNINIT_NEXT)); 1116 // Remove this from between prev_p and next_p. 1117 if (prev_p == NULL) 1118 head = next_p; 1119 else 1120 prev_p->set_next(next_p); 1121 return head; 1122 } 1123 1124 WarmCallInfo WarmCallInfo::_always_hot(WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE(), 1125 WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE()); 1126 WarmCallInfo WarmCallInfo::_always_cold(WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE(), 1127 WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE()); 1128 1129 WarmCallInfo* WarmCallInfo::always_hot() { 1130 assert(_always_hot.is_hot(), "must always be hot"); 1131 return &_always_hot; 1132 } 1133 1134 WarmCallInfo* WarmCallInfo::always_cold() { 1135 assert(_always_cold.is_cold(), "must always be cold"); 1136 return &_always_cold; 1137 } 1138 1139 1140 #ifndef PRODUCT 1141 1142 void WarmCallInfo::print() const { 1143 tty->print("%s : C=%6.1f P=%6.1f W=%6.1f S=%6.1f H=%6.1f -> %p", 1144 is_cold() ? "cold" : is_hot() ? "hot " : "warm", 1145 count(), profit(), work(), size(), compute_heat(), next()); 1146 tty->cr(); 1147 if (call() != NULL) call()->dump(); 1148 } 1149 1150 void print_wci(WarmCallInfo* ci) { 1151 ci->print(); 1152 } 1153 1154 void WarmCallInfo::print_all() const { 1155 for (const WarmCallInfo* p = this; p != NULL; p = p->next()) 1156 p->print(); 1157 } 1158 1159 int WarmCallInfo::count_all() const { 1160 int cnt = 0; 1161 for (const WarmCallInfo* p = this; p != NULL; p = p->next()) 1162 cnt++; 1163 return cnt; 1164 } 1165 1166 #endif //PRODUCT