1 /* 2 * Copyright (c) 2000, 2019, 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/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/castnode.hpp" 37 #include "opto/cfgnode.hpp" 38 #include "opto/parse.hpp" 39 #include "opto/rootnode.hpp" 40 #include "opto/runtime.hpp" 41 #include "opto/subnode.hpp" 42 #include "opto/valuetypenode.hpp" 43 #include "runtime/sharedRuntime.hpp" 44 45 // Utility function. 46 const TypeFunc* CallGenerator::tf() const { 47 return TypeFunc::make(method()); 48 } 49 50 bool CallGenerator::is_inlined_method_handle_intrinsic(JVMState* jvms, ciMethod* m) { 51 return is_inlined_method_handle_intrinsic(jvms->method(), jvms->bci(), m); 52 } 53 54 bool CallGenerator::is_inlined_method_handle_intrinsic(ciMethod* caller, int bci, ciMethod* m) { 55 ciMethod* symbolic_info = caller->get_method_at_bci(bci); 56 return is_inlined_method_handle_intrinsic(symbolic_info, m); 57 } 58 59 bool CallGenerator::is_inlined_method_handle_intrinsic(ciMethod* symbolic_info, ciMethod* m) { 60 return symbolic_info->is_method_handle_intrinsic() && !m->is_method_handle_intrinsic(); 61 } 62 63 //-----------------------------ParseGenerator--------------------------------- 64 // Internal class which handles all direct bytecode traversal. 65 class ParseGenerator : public InlineCallGenerator { 66 private: 67 bool _is_osr; 68 float _expected_uses; 69 70 public: 71 ParseGenerator(ciMethod* method, float expected_uses, bool is_osr = false) 72 : InlineCallGenerator(method) 73 { 74 _is_osr = is_osr; 75 _expected_uses = expected_uses; 76 assert(InlineTree::check_can_parse(method) == NULL, "parse must be possible"); 77 } 78 79 virtual bool is_parse() const { return true; } 80 virtual JVMState* generate(JVMState* jvms); 81 int is_osr() { return _is_osr; } 82 83 }; 84 85 JVMState* ParseGenerator::generate(JVMState* jvms) { 86 Compile* C = Compile::current(); 87 C->print_inlining_update(this); 88 89 if (is_osr()) { 90 // The JVMS for a OSR has a single argument (see its TypeFunc). 91 assert(jvms->depth() == 1, "no inline OSR"); 92 } 93 94 if (C->failing()) { 95 return NULL; // bailing out of the compile; do not try to parse 96 } 97 98 Parse parser(jvms, method(), _expected_uses); 99 // Grab signature for matching/allocation 100 GraphKit& exits = parser.exits(); 101 102 if (C->failing()) { 103 while (exits.pop_exception_state() != NULL) ; 104 return NULL; 105 } 106 107 assert(exits.jvms()->same_calls_as(jvms), "sanity"); 108 109 // Simply return the exit state of the parser, 110 // augmented by any exceptional states. 111 return exits.transfer_exceptions_into_jvms(); 112 } 113 114 //---------------------------DirectCallGenerator------------------------------ 115 // Internal class which handles all out-of-line calls w/o receiver type checks. 116 class DirectCallGenerator : public CallGenerator { 117 private: 118 CallStaticJavaNode* _call_node; 119 // Force separate memory and I/O projections for the exceptional 120 // paths to facilitate late inlining. 121 bool _separate_io_proj; 122 123 public: 124 DirectCallGenerator(ciMethod* method, bool separate_io_proj) 125 : CallGenerator(method), 126 _call_node(NULL), 127 _separate_io_proj(separate_io_proj) 128 { 129 if (ValueTypeReturnedAsFields && method->is_method_handle_intrinsic()) { 130 // If that call has not been optimized by the time optimizations are over, 131 // we'll need to add a call to create a value type instance from the klass 132 // returned by the call (see PhaseMacroExpand::expand_mh_intrinsic_return). 133 // Separating memory and I/O projections for exceptions is required to 134 // perform that graph transformation. 135 _separate_io_proj = true; 136 } 137 } 138 virtual JVMState* generate(JVMState* jvms); 139 140 CallStaticJavaNode* call_node() const { return _call_node; } 141 }; 142 143 JVMState* DirectCallGenerator::generate(JVMState* jvms) { 144 GraphKit kit(jvms); 145 kit.C->print_inlining_update(this); 146 PhaseGVN& gvn = kit.gvn(); 147 bool is_static = method()->is_static(); 148 address target = is_static ? SharedRuntime::get_resolve_static_call_stub() 149 : SharedRuntime::get_resolve_opt_virtual_call_stub(); 150 151 if (kit.C->log() != NULL) { 152 kit.C->log()->elem("direct_call bci='%d'", jvms->bci()); 153 } 154 155 CallStaticJavaNode *call = new CallStaticJavaNode(kit.C, tf(), target, method(), kit.bci()); 156 if (is_inlined_method_handle_intrinsic(jvms, method())) { 157 // To be able to issue a direct call and skip a call to MH.linkTo*/invokeBasic adapter, 158 // additional information about the method being invoked should be attached 159 // to the call site to make resolution logic work 160 // (see SharedRuntime::resolve_static_call_C). 161 call->set_override_symbolic_info(true); 162 } 163 _call_node = call; // Save the call node in case we need it later 164 if (!is_static) { 165 // Make an explicit receiver null_check as part of this call. 166 // Since we share a map with the caller, his JVMS gets adjusted. 167 kit.null_check_receiver_before_call(method()); 168 if (kit.stopped()) { 169 // And dump it back to the caller, decorated with any exceptions: 170 return kit.transfer_exceptions_into_jvms(); 171 } 172 // Mark the call node as virtual, sort of: 173 call->set_optimized_virtual(true); 174 if (method()->is_method_handle_intrinsic() || 175 method()->is_compiled_lambda_form()) { 176 call->set_method_handle_invoke(true); 177 } 178 } 179 kit.set_arguments_for_java_call(call, is_late_inline()); 180 if (kit.stopped()) { 181 return kit.transfer_exceptions_into_jvms(); 182 } 183 kit.set_edges_for_java_call(call, false, _separate_io_proj); 184 Node* ret = kit.set_results_for_java_call(call, _separate_io_proj); 185 kit.push_node(method()->return_type()->basic_type(), ret); 186 return kit.transfer_exceptions_into_jvms(); 187 } 188 189 //--------------------------VirtualCallGenerator------------------------------ 190 // Internal class which handles all out-of-line calls checking receiver type. 191 class VirtualCallGenerator : public CallGenerator { 192 private: 193 int _vtable_index; 194 public: 195 VirtualCallGenerator(ciMethod* method, int vtable_index) 196 : CallGenerator(method), _vtable_index(vtable_index) 197 { 198 assert(vtable_index == Method::invalid_vtable_index || 199 vtable_index >= 0, "either invalid or usable"); 200 } 201 virtual bool is_virtual() const { return true; } 202 virtual JVMState* generate(JVMState* jvms); 203 }; 204 205 JVMState* VirtualCallGenerator::generate(JVMState* jvms) { 206 GraphKit kit(jvms); 207 Node* receiver = kit.argument(0); 208 kit.C->print_inlining_update(this); 209 210 if (kit.C->log() != NULL) { 211 kit.C->log()->elem("virtual_call bci='%d'", jvms->bci()); 212 } 213 214 // If the receiver is a constant null, do not torture the system 215 // by attempting to call through it. The compile will proceed 216 // correctly, but may bail out in final_graph_reshaping, because 217 // the call instruction will have a seemingly deficient out-count. 218 // (The bailout says something misleading about an "infinite loop".) 219 if (!receiver->is_ValueType() && kit.gvn().type(receiver)->higher_equal(TypePtr::NULL_PTR)) { 220 assert(Bytecodes::is_invoke(kit.java_bc()), "%d: %s", kit.java_bc(), Bytecodes::name(kit.java_bc())); 221 ciMethod* declared_method = kit.method()->get_method_at_bci(kit.bci()); 222 int arg_size = declared_method->signature()->arg_size_for_bc(kit.java_bc()); 223 kit.inc_sp(arg_size); // restore arguments 224 kit.uncommon_trap(Deoptimization::Reason_null_check, 225 Deoptimization::Action_none, 226 NULL, "null receiver"); 227 return kit.transfer_exceptions_into_jvms(); 228 } 229 230 // Ideally we would unconditionally do a null check here and let it 231 // be converted to an implicit check based on profile information. 232 // However currently the conversion to implicit null checks in 233 // Block::implicit_null_check() only looks for loads and stores, not calls. 234 ciMethod *caller = kit.method(); 235 ciMethodData *caller_md = (caller == NULL) ? NULL : caller->method_data(); 236 if (!UseInlineCaches || !ImplicitNullChecks || !os::zero_page_read_protected() || 237 ((ImplicitNullCheckThreshold > 0) && caller_md && 238 (caller_md->trap_count(Deoptimization::Reason_null_check) 239 >= (uint)ImplicitNullCheckThreshold))) { 240 // Make an explicit receiver null_check as part of this call. 241 // Since we share a map with the caller, his JVMS gets adjusted. 242 receiver = kit.null_check_receiver_before_call(method()); 243 if (kit.stopped()) { 244 // And dump it back to the caller, decorated with any exceptions: 245 return kit.transfer_exceptions_into_jvms(); 246 } 247 } 248 249 assert(!method()->is_static(), "virtual call must not be to static"); 250 assert(!method()->is_final(), "virtual call should not be to final"); 251 assert(!method()->is_private(), "virtual call should not be to private"); 252 assert(_vtable_index == Method::invalid_vtable_index || !UseInlineCaches, 253 "no vtable calls if +UseInlineCaches "); 254 address target = SharedRuntime::get_resolve_virtual_call_stub(); 255 // Normal inline cache used for call 256 CallDynamicJavaNode *call = new CallDynamicJavaNode(tf(), target, method(), _vtable_index, kit.bci()); 257 if (is_inlined_method_handle_intrinsic(jvms, method())) { 258 // To be able to issue a direct call (optimized virtual or virtual) 259 // and skip a call to MH.linkTo*/invokeBasic adapter, additional information 260 // about the method being invoked should be attached to the call site to 261 // make resolution logic work (see SharedRuntime::resolve_{virtual,opt_virtual}_call_C). 262 call->set_override_symbolic_info(true); 263 } 264 kit.set_arguments_for_java_call(call); 265 if (kit.stopped()) { 266 return kit.transfer_exceptions_into_jvms(); 267 } 268 kit.set_edges_for_java_call(call); 269 Node* ret = kit.set_results_for_java_call(call); 270 kit.push_node(method()->return_type()->basic_type(), ret); 271 272 // Represent the effect of an implicit receiver null_check 273 // as part of this call. Since we share a map with the caller, 274 // his JVMS gets adjusted. 275 kit.cast_not_null(receiver); 276 return kit.transfer_exceptions_into_jvms(); 277 } 278 279 CallGenerator* CallGenerator::for_inline(ciMethod* m, float expected_uses) { 280 if (InlineTree::check_can_parse(m) != NULL) return NULL; 281 return new ParseGenerator(m, expected_uses); 282 } 283 284 // As a special case, the JVMS passed to this CallGenerator is 285 // for the method execution already in progress, not just the JVMS 286 // of the caller. Thus, this CallGenerator cannot be mixed with others! 287 CallGenerator* CallGenerator::for_osr(ciMethod* m, int osr_bci) { 288 if (InlineTree::check_can_parse(m) != NULL) return NULL; 289 float past_uses = m->interpreter_invocation_count(); 290 float expected_uses = past_uses; 291 return new ParseGenerator(m, expected_uses, true); 292 } 293 294 CallGenerator* CallGenerator::for_direct_call(ciMethod* m, bool separate_io_proj) { 295 assert(!m->is_abstract(), "for_direct_call mismatch"); 296 return new DirectCallGenerator(m, separate_io_proj); 297 } 298 299 CallGenerator* CallGenerator::for_virtual_call(ciMethod* m, int vtable_index) { 300 assert(!m->is_static(), "for_virtual_call mismatch"); 301 assert(!m->is_method_handle_intrinsic(), "should be a direct call"); 302 return new VirtualCallGenerator(m, vtable_index); 303 } 304 305 // Allow inlining decisions to be delayed 306 class LateInlineCallGenerator : public DirectCallGenerator { 307 private: 308 jlong _unique_id; // unique id for log compilation 309 bool _is_pure_call; // a hint that the call doesn't have important side effects to care about 310 311 protected: 312 CallGenerator* _inline_cg; 313 virtual bool do_late_inline_check(JVMState* jvms) { return true; } 314 315 public: 316 LateInlineCallGenerator(ciMethod* method, CallGenerator* inline_cg, bool is_pure_call = false) : 317 DirectCallGenerator(method, true), _unique_id(0), _is_pure_call(is_pure_call), _inline_cg(inline_cg) {} 318 319 virtual bool is_late_inline() const { return true; } 320 321 // Convert the CallStaticJava into an inline 322 virtual void do_late_inline(); 323 324 virtual JVMState* generate(JVMState* jvms) { 325 Compile *C = Compile::current(); 326 327 C->log_inline_id(this); 328 329 // Record that this call site should be revisited once the main 330 // parse is finished. 331 if (!is_mh_late_inline()) { 332 C->add_late_inline(this); 333 } 334 335 // Emit the CallStaticJava and request separate projections so 336 // that the late inlining logic can distinguish between fall 337 // through and exceptional uses of the memory and io projections 338 // as is done for allocations and macro expansion. 339 return DirectCallGenerator::generate(jvms); 340 } 341 342 virtual void print_inlining_late(const char* msg) { 343 CallNode* call = call_node(); 344 Compile* C = Compile::current(); 345 C->print_inlining_assert_ready(); 346 C->print_inlining(method(), call->jvms()->depth()-1, call->jvms()->bci(), msg); 347 C->print_inlining_move_to(this); 348 C->print_inlining_update_delayed(this); 349 } 350 351 virtual void set_unique_id(jlong id) { 352 _unique_id = id; 353 } 354 355 virtual jlong unique_id() const { 356 return _unique_id; 357 } 358 }; 359 360 void LateInlineCallGenerator::do_late_inline() { 361 // Can't inline it 362 CallStaticJavaNode* call = call_node(); 363 if (call == NULL || call->outcnt() == 0 || 364 call->in(0) == NULL || call->in(0)->is_top()) { 365 return; 366 } 367 368 const GrowableArray<SigEntry>* sig_cc = method()->get_sig_cc(); 369 const TypeTuple* r = call->tf()->domain_cc(); 370 for (uint i1 = TypeFunc::Parms, i2 = 0; i1 < r->cnt(); i1++) { 371 if (sig_cc != NULL) { 372 // Skip reserved entries 373 while (!SigEntry::skip_value_delimiters(sig_cc, i2)) { 374 i2++; 375 } 376 if (SigEntry::is_reserved_entry(sig_cc, i2++)) { 377 assert(call->in(i1)->is_top(), "should be top"); 378 continue; 379 } 380 } 381 if (call->in(i1)->is_top() && r->field_at(i1) != Type::HALF) { 382 assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing"); 383 return; 384 } 385 } 386 387 if (call->in(TypeFunc::Memory)->is_top()) { 388 assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing"); 389 return; 390 } 391 392 // check for unreachable loop 393 CallProjections* callprojs = call->extract_projections(true); 394 if (callprojs->fallthrough_catchproj == call->in(0) || 395 callprojs->catchall_catchproj == call->in(0) || 396 callprojs->fallthrough_memproj == call->in(TypeFunc::Memory) || 397 callprojs->catchall_memproj == call->in(TypeFunc::Memory) || 398 callprojs->fallthrough_ioproj == call->in(TypeFunc::I_O) || 399 callprojs->catchall_ioproj == call->in(TypeFunc::I_O) || 400 (callprojs->exobj != NULL && call->find_edge(callprojs->exobj) != -1)) { 401 return; 402 } 403 bool result_not_used = true; 404 for (uint i = 0; i < callprojs->nb_resproj; i++) { 405 if (callprojs->resproj[i] != NULL) { 406 if (callprojs->resproj[i]->outcnt() != 0) { 407 result_not_used = false; 408 } 409 if (call->find_edge(callprojs->resproj[i]) != -1) { 410 return; 411 } 412 } 413 } 414 415 Compile* C = Compile::current(); 416 // Remove inlined methods from Compiler's lists. 417 if (call->is_macro()) { 418 C->remove_macro_node(call); 419 } 420 421 if (_is_pure_call && result_not_used) { 422 // The call is marked as pure (no important side effects), but result isn't used. 423 // It's safe to remove the call. 424 GraphKit kit(call->jvms()); 425 kit.replace_call(call, C->top(), true); 426 } else { 427 // Make a clone of the JVMState that appropriate to use for driving a parse 428 JVMState* old_jvms = call->jvms(); 429 JVMState* jvms = old_jvms->clone_shallow(C); 430 uint size = call->req(); 431 SafePointNode* map = new SafePointNode(size, jvms); 432 for (uint i1 = 0; i1 < size; i1++) { 433 map->init_req(i1, call->in(i1)); 434 } 435 436 PhaseGVN& gvn = *C->initial_gvn(); 437 // Make sure the state is a MergeMem for parsing. 438 if (!map->in(TypeFunc::Memory)->is_MergeMem()) { 439 Node* mem = MergeMemNode::make(map->in(TypeFunc::Memory)); 440 gvn.set_type_bottom(mem); 441 map->set_req(TypeFunc::Memory, mem); 442 } 443 444 // blow away old call arguments 445 for (uint i1 = TypeFunc::Parms; i1 < r->cnt(); i1++) { 446 map->set_req(i1, C->top()); 447 } 448 jvms->set_map(map); 449 450 // Make enough space in the expression stack to transfer 451 // the incoming arguments and return value. 452 map->ensure_stack(jvms, jvms->method()->max_stack()); 453 const TypeTuple *domain_sig = call->_tf->domain_sig(); 454 ExtendedSignature sig_cc = ExtendedSignature(method()->get_sig_cc(), SigEntryFilter()); 455 uint nargs = method()->arg_size(); 456 assert(domain_sig->cnt() - TypeFunc::Parms == nargs, "inconsistent signature"); 457 458 uint j = TypeFunc::Parms; 459 for (uint i1 = 0; i1 < nargs; i1++) { 460 const Type* t = domain_sig->field_at(TypeFunc::Parms + i1); 461 if (method()->has_scalarized_args() && t->is_valuetypeptr() && !t->maybe_null()) { 462 // Value type arguments are not passed by reference: we get an argument per 463 // field of the value type. Build ValueTypeNodes from the value type arguments. 464 GraphKit arg_kit(jvms, &gvn); 465 ValueTypeNode* vt = ValueTypeNode::make_from_multi(&arg_kit, call, sig_cc, t->value_klass(), j, true); 466 map->set_control(arg_kit.control()); 467 map->set_argument(jvms, i1, vt); 468 } else { 469 map->set_argument(jvms, i1, call->in(j++)); 470 BasicType bt = t->basic_type(); 471 while (SigEntry::next_is_reserved(sig_cc, bt, true)) { 472 j += type2size[bt]; // Skip reserved arguments 473 } 474 } 475 } 476 477 C->print_inlining_assert_ready(); 478 479 C->print_inlining_move_to(this); 480 481 C->log_late_inline(this); 482 483 // This check is done here because for_method_handle_inline() method 484 // needs jvms for inlined state. 485 if (!do_late_inline_check(jvms)) { 486 map->disconnect_inputs(NULL, C); 487 return; 488 } 489 490 // Allocate a buffer for the returned ValueTypeNode because the caller expects an oop return. 491 // Do this before the method handle call in case the buffer allocation triggers deoptimization. 492 Node* buffer_oop = NULL; 493 if (is_mh_late_inline() && _inline_cg->method()->return_type()->is_valuetype()) { 494 GraphKit arg_kit(jvms, &gvn); 495 { 496 PreserveReexecuteState preexecs(&arg_kit); 497 arg_kit.jvms()->set_should_reexecute(true); 498 arg_kit.inc_sp(nargs); 499 Node* klass_node = arg_kit.makecon(TypeKlassPtr::make(_inline_cg->method()->return_type()->as_value_klass())); 500 buffer_oop = arg_kit.new_instance(klass_node, NULL, NULL, /* deoptimize_on_exception */ true); 501 } 502 jvms = arg_kit.transfer_exceptions_into_jvms(); 503 } 504 505 // Setup default node notes to be picked up by the inlining 506 Node_Notes* old_nn = C->node_notes_at(call->_idx); 507 if (old_nn != NULL) { 508 Node_Notes* entry_nn = old_nn->clone(C); 509 entry_nn->set_jvms(jvms); 510 C->set_default_node_notes(entry_nn); 511 } 512 513 // Now perform the inlining using the synthesized JVMState 514 JVMState* new_jvms = _inline_cg->generate(jvms); 515 if (new_jvms == NULL) return; // no change 516 if (C->failing()) return; 517 518 // Capture any exceptional control flow 519 GraphKit kit(new_jvms); 520 521 // Find the result object 522 Node* result = C->top(); 523 int result_size = method()->return_type()->size(); 524 if (result_size != 0 && !kit.stopped()) { 525 result = (result_size == 1) ? kit.pop() : kit.pop_pair(); 526 } 527 528 C->set_has_loops(C->has_loops() || _inline_cg->method()->has_loops()); 529 C->env()->notice_inlined_method(_inline_cg->method()); 530 C->set_inlining_progress(true); 531 C->set_do_cleanup(kit.stopped()); // path is dead; needs cleanup 532 533 // Handle value type returns 534 bool returned_as_fields = call->tf()->returns_value_type_as_fields(); 535 if (result->is_ValueType()) { 536 // Only possible if is_mh_late_inline() when the callee does not "know" that the caller expects an oop 537 assert(is_mh_late_inline() && !returned_as_fields, "sanity"); 538 assert(buffer_oop != NULL, "should have allocated a buffer"); 539 ValueTypeNode* vt = result->as_ValueType(); 540 vt->store(&kit, buffer_oop, buffer_oop, vt->type()->value_klass(), 0); 541 // Do not let stores that initialize this buffer be reordered with a subsequent 542 // store that would make this buffer accessible by other threads. 543 AllocateNode* alloc = AllocateNode::Ideal_allocation(buffer_oop, &kit.gvn()); 544 assert(alloc != NULL, "must have an allocation node"); 545 kit.insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 546 result = buffer_oop; 547 } else if (result->is_ValueTypePtr() && returned_as_fields) { 548 result->as_ValueTypePtr()->replace_call_results(&kit, call, C); 549 } 550 551 kit.replace_call(call, result, true); 552 } 553 } 554 555 556 CallGenerator* CallGenerator::for_late_inline(ciMethod* method, CallGenerator* inline_cg) { 557 return new LateInlineCallGenerator(method, inline_cg); 558 } 559 560 class LateInlineMHCallGenerator : public LateInlineCallGenerator { 561 ciMethod* _caller; 562 int _attempt; 563 bool _input_not_const; 564 565 virtual bool do_late_inline_check(JVMState* jvms); 566 virtual bool already_attempted() const { return _attempt > 0; } 567 568 public: 569 LateInlineMHCallGenerator(ciMethod* caller, ciMethod* callee, bool input_not_const) : 570 LateInlineCallGenerator(callee, NULL), _caller(caller), _attempt(0), _input_not_const(input_not_const) {} 571 572 virtual bool is_mh_late_inline() const { return true; } 573 574 virtual JVMState* generate(JVMState* jvms) { 575 JVMState* new_jvms = LateInlineCallGenerator::generate(jvms); 576 577 Compile* C = Compile::current(); 578 if (_input_not_const) { 579 // inlining won't be possible so no need to enqueue right now. 580 call_node()->set_generator(this); 581 } else { 582 C->add_late_inline(this); 583 } 584 return new_jvms; 585 } 586 }; 587 588 bool LateInlineMHCallGenerator::do_late_inline_check(JVMState* jvms) { 589 590 CallGenerator* cg = for_method_handle_inline(jvms, _caller, method(), _input_not_const, AlwaysIncrementalInline); 591 592 Compile::current()->print_inlining_update_delayed(this); 593 594 if (!_input_not_const) { 595 _attempt++; 596 } 597 598 if (cg != NULL && (cg->is_inline() || cg->is_inlined_method_handle_intrinsic(jvms, cg->method()))) { 599 assert(!cg->is_late_inline(), "we're doing late inlining"); 600 _inline_cg = cg; 601 Compile::current()->dec_number_of_mh_late_inlines(); 602 return true; 603 } 604 605 call_node()->set_generator(this); 606 return false; 607 } 608 609 CallGenerator* CallGenerator::for_mh_late_inline(ciMethod* caller, ciMethod* callee, bool input_not_const) { 610 Compile::current()->inc_number_of_mh_late_inlines(); 611 CallGenerator* cg = new LateInlineMHCallGenerator(caller, callee, input_not_const); 612 return cg; 613 } 614 615 class LateInlineStringCallGenerator : public LateInlineCallGenerator { 616 617 public: 618 LateInlineStringCallGenerator(ciMethod* method, CallGenerator* inline_cg) : 619 LateInlineCallGenerator(method, inline_cg) {} 620 621 virtual JVMState* generate(JVMState* jvms) { 622 Compile *C = Compile::current(); 623 624 C->log_inline_id(this); 625 626 C->add_string_late_inline(this); 627 628 JVMState* new_jvms = DirectCallGenerator::generate(jvms); 629 return new_jvms; 630 } 631 632 virtual bool is_string_late_inline() const { return true; } 633 }; 634 635 CallGenerator* CallGenerator::for_string_late_inline(ciMethod* method, CallGenerator* inline_cg) { 636 return new LateInlineStringCallGenerator(method, inline_cg); 637 } 638 639 class LateInlineBoxingCallGenerator : public LateInlineCallGenerator { 640 641 public: 642 LateInlineBoxingCallGenerator(ciMethod* method, CallGenerator* inline_cg) : 643 LateInlineCallGenerator(method, inline_cg, /*is_pure=*/true) {} 644 645 virtual JVMState* generate(JVMState* jvms) { 646 Compile *C = Compile::current(); 647 648 C->log_inline_id(this); 649 650 C->add_boxing_late_inline(this); 651 652 JVMState* new_jvms = DirectCallGenerator::generate(jvms); 653 return new_jvms; 654 } 655 }; 656 657 CallGenerator* CallGenerator::for_boxing_late_inline(ciMethod* method, CallGenerator* inline_cg) { 658 return new LateInlineBoxingCallGenerator(method, inline_cg); 659 } 660 661 //---------------------------WarmCallGenerator-------------------------------- 662 // Internal class which handles initial deferral of inlining decisions. 663 class WarmCallGenerator : public CallGenerator { 664 WarmCallInfo* _call_info; 665 CallGenerator* _if_cold; 666 CallGenerator* _if_hot; 667 bool _is_virtual; // caches virtuality of if_cold 668 bool _is_inline; // caches inline-ness of if_hot 669 670 public: 671 WarmCallGenerator(WarmCallInfo* ci, 672 CallGenerator* if_cold, 673 CallGenerator* if_hot) 674 : CallGenerator(if_cold->method()) 675 { 676 assert(method() == if_hot->method(), "consistent choices"); 677 _call_info = ci; 678 _if_cold = if_cold; 679 _if_hot = if_hot; 680 _is_virtual = if_cold->is_virtual(); 681 _is_inline = if_hot->is_inline(); 682 } 683 684 virtual bool is_inline() const { return _is_inline; } 685 virtual bool is_virtual() const { return _is_virtual; } 686 virtual bool is_deferred() const { return true; } 687 688 virtual JVMState* generate(JVMState* jvms); 689 }; 690 691 692 CallGenerator* CallGenerator::for_warm_call(WarmCallInfo* ci, 693 CallGenerator* if_cold, 694 CallGenerator* if_hot) { 695 return new WarmCallGenerator(ci, if_cold, if_hot); 696 } 697 698 JVMState* WarmCallGenerator::generate(JVMState* jvms) { 699 Compile* C = Compile::current(); 700 C->print_inlining_update(this); 701 702 if (C->log() != NULL) { 703 C->log()->elem("warm_call bci='%d'", jvms->bci()); 704 } 705 jvms = _if_cold->generate(jvms); 706 if (jvms != NULL) { 707 Node* m = jvms->map()->control(); 708 if (m->is_CatchProj()) m = m->in(0); else m = C->top(); 709 if (m->is_Catch()) m = m->in(0); else m = C->top(); 710 if (m->is_Proj()) m = m->in(0); else m = C->top(); 711 if (m->is_CallJava()) { 712 _call_info->set_call(m->as_Call()); 713 _call_info->set_hot_cg(_if_hot); 714 #ifndef PRODUCT 715 if (PrintOpto || PrintOptoInlining) { 716 tty->print_cr("Queueing for warm inlining at bci %d:", jvms->bci()); 717 tty->print("WCI: "); 718 _call_info->print(); 719 } 720 #endif 721 _call_info->set_heat(_call_info->compute_heat()); 722 C->set_warm_calls(_call_info->insert_into(C->warm_calls())); 723 } 724 } 725 return jvms; 726 } 727 728 void WarmCallInfo::make_hot() { 729 Unimplemented(); 730 } 731 732 void WarmCallInfo::make_cold() { 733 // No action: Just dequeue. 734 } 735 736 737 //------------------------PredictedCallGenerator------------------------------ 738 // Internal class which handles all out-of-line calls checking receiver type. 739 class PredictedCallGenerator : public CallGenerator { 740 ciKlass* _predicted_receiver; 741 CallGenerator* _if_missed; 742 CallGenerator* _if_hit; 743 float _hit_prob; 744 bool _exact_check; 745 746 public: 747 PredictedCallGenerator(ciKlass* predicted_receiver, 748 CallGenerator* if_missed, 749 CallGenerator* if_hit, bool exact_check, 750 float hit_prob) 751 : CallGenerator(if_missed->method()) 752 { 753 // The call profile data may predict the hit_prob as extreme as 0 or 1. 754 // Remove the extremes values from the range. 755 if (hit_prob > PROB_MAX) hit_prob = PROB_MAX; 756 if (hit_prob < PROB_MIN) hit_prob = PROB_MIN; 757 758 _predicted_receiver = predicted_receiver; 759 _if_missed = if_missed; 760 _if_hit = if_hit; 761 _hit_prob = hit_prob; 762 _exact_check = exact_check; 763 } 764 765 virtual bool is_virtual() const { return true; } 766 virtual bool is_inline() const { return _if_hit->is_inline(); } 767 virtual bool is_deferred() const { return _if_hit->is_deferred(); } 768 769 virtual JVMState* generate(JVMState* jvms); 770 }; 771 772 773 CallGenerator* CallGenerator::for_predicted_call(ciKlass* predicted_receiver, 774 CallGenerator* if_missed, 775 CallGenerator* if_hit, 776 float hit_prob) { 777 return new PredictedCallGenerator(predicted_receiver, if_missed, if_hit, 778 /*exact_check=*/true, hit_prob); 779 } 780 781 CallGenerator* CallGenerator::for_guarded_call(ciKlass* guarded_receiver, 782 CallGenerator* if_missed, 783 CallGenerator* if_hit) { 784 return new PredictedCallGenerator(guarded_receiver, if_missed, if_hit, 785 /*exact_check=*/false, PROB_ALWAYS); 786 } 787 788 JVMState* PredictedCallGenerator::generate(JVMState* jvms) { 789 GraphKit kit(jvms); 790 kit.C->print_inlining_update(this); 791 PhaseGVN& gvn = kit.gvn(); 792 // We need an explicit receiver null_check before checking its type. 793 // We share a map with the caller, so his JVMS gets adjusted. 794 Node* receiver = kit.argument(0); 795 CompileLog* log = kit.C->log(); 796 if (log != NULL) { 797 log->elem("predicted_call bci='%d' exact='%d' klass='%d'", 798 jvms->bci(), (_exact_check ? 1 : 0), log->identify(_predicted_receiver)); 799 } 800 801 receiver = kit.null_check_receiver_before_call(method()); 802 if (kit.stopped()) { 803 return kit.transfer_exceptions_into_jvms(); 804 } 805 806 // Make a copy of the replaced nodes in case we need to restore them 807 ReplacedNodes replaced_nodes = kit.map()->replaced_nodes(); 808 replaced_nodes.clone(); 809 810 Node* casted_receiver = receiver; // will get updated in place... 811 Node* slow_ctl = NULL; 812 if (_exact_check) { 813 slow_ctl = kit.type_check_receiver(receiver, _predicted_receiver, _hit_prob, 814 &casted_receiver); 815 } else { 816 slow_ctl = kit.subtype_check_receiver(receiver, _predicted_receiver, 817 &casted_receiver); 818 } 819 820 SafePointNode* slow_map = NULL; 821 JVMState* slow_jvms = NULL; 822 { PreserveJVMState pjvms(&kit); 823 kit.set_control(slow_ctl); 824 if (!kit.stopped()) { 825 slow_jvms = _if_missed->generate(kit.sync_jvms()); 826 if (kit.failing()) 827 return NULL; // might happen because of NodeCountInliningCutoff 828 assert(slow_jvms != NULL, "must be"); 829 kit.add_exception_states_from(slow_jvms); 830 kit.set_map(slow_jvms->map()); 831 if (!kit.stopped()) 832 slow_map = kit.stop(); 833 } 834 } 835 836 if (kit.stopped()) { 837 // Instance exactly does not matches the desired type. 838 kit.set_jvms(slow_jvms); 839 return kit.transfer_exceptions_into_jvms(); 840 } 841 842 // fall through if the instance exactly matches the desired type 843 kit.replace_in_map(receiver, casted_receiver); 844 845 // Make the hot call: 846 JVMState* new_jvms = _if_hit->generate(kit.sync_jvms()); 847 if (new_jvms == NULL) { 848 // Inline failed, so make a direct call. 849 assert(_if_hit->is_inline(), "must have been a failed inline"); 850 CallGenerator* cg = CallGenerator::for_direct_call(_if_hit->method()); 851 new_jvms = cg->generate(kit.sync_jvms()); 852 } 853 kit.add_exception_states_from(new_jvms); 854 kit.set_jvms(new_jvms); 855 856 // Need to merge slow and fast? 857 if (slow_map == NULL) { 858 // The fast path is the only path remaining. 859 return kit.transfer_exceptions_into_jvms(); 860 } 861 862 if (kit.stopped()) { 863 // Inlined method threw an exception, so it's just the slow path after all. 864 kit.set_jvms(slow_jvms); 865 return kit.transfer_exceptions_into_jvms(); 866 } 867 868 // Allocate value types if they are merged with objects (similar to Parse::merge_common()) 869 uint tos = kit.jvms()->stkoff() + kit.sp(); 870 uint limit = slow_map->req(); 871 for (uint i = TypeFunc::Parms; i < limit; i++) { 872 Node* m = kit.map()->in(i); 873 Node* n = slow_map->in(i); 874 const Type* t = gvn.type(m)->meet_speculative(gvn.type(n)); 875 if (m->is_ValueType() && !t->isa_valuetype()) { 876 // Allocate value type in fast path 877 m = ValueTypePtrNode::make_from_value_type(&kit, m->as_ValueType()); 878 kit.map()->set_req(i, m); 879 } 880 if (n->is_ValueType() && !t->isa_valuetype()) { 881 // Allocate value type in slow path 882 PreserveJVMState pjvms(&kit); 883 kit.set_map(slow_map); 884 n = ValueTypePtrNode::make_from_value_type(&kit, n->as_ValueType()); 885 kit.map()->set_req(i, n); 886 slow_map = kit.stop(); 887 } 888 } 889 890 // There are 2 branches and the replaced nodes are only valid on 891 // one: restore the replaced nodes to what they were before the 892 // branch. 893 kit.map()->set_replaced_nodes(replaced_nodes); 894 895 // Finish the diamond. 896 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization 897 RegionNode* region = new RegionNode(3); 898 region->init_req(1, kit.control()); 899 region->init_req(2, slow_map->control()); 900 kit.set_control(gvn.transform(region)); 901 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO); 902 iophi->set_req(2, slow_map->i_o()); 903 kit.set_i_o(gvn.transform(iophi)); 904 // Merge memory 905 kit.merge_memory(slow_map->merged_memory(), region, 2); 906 // Transform new memory Phis. 907 for (MergeMemStream mms(kit.merged_memory()); mms.next_non_empty();) { 908 Node* phi = mms.memory(); 909 if (phi->is_Phi() && phi->in(0) == region) { 910 mms.set_memory(gvn.transform(phi)); 911 } 912 } 913 for (uint i = TypeFunc::Parms; i < limit; i++) { 914 // Skip unused stack slots; fast forward to monoff(); 915 if (i == tos) { 916 i = kit.jvms()->monoff(); 917 if( i >= limit ) break; 918 } 919 Node* m = kit.map()->in(i); 920 Node* n = slow_map->in(i); 921 if (m != n) { 922 const Type* t = gvn.type(m)->meet_speculative(gvn.type(n)); 923 Node* phi = PhiNode::make(region, m, t); 924 phi->set_req(2, n); 925 kit.map()->set_req(i, gvn.transform(phi)); 926 } 927 } 928 return kit.transfer_exceptions_into_jvms(); 929 } 930 931 932 CallGenerator* CallGenerator::for_method_handle_call(JVMState* jvms, ciMethod* caller, ciMethod* callee, bool delayed_forbidden) { 933 assert(callee->is_method_handle_intrinsic(), "for_method_handle_call mismatch"); 934 bool input_not_const; 935 CallGenerator* cg = CallGenerator::for_method_handle_inline(jvms, caller, callee, input_not_const, false); 936 Compile* C = Compile::current(); 937 if (cg != NULL) { 938 if (!delayed_forbidden && AlwaysIncrementalInline) { 939 return CallGenerator::for_late_inline(callee, cg); 940 } else { 941 return cg; 942 } 943 } 944 int bci = jvms->bci(); 945 ciCallProfile profile = caller->call_profile_at_bci(bci); 946 int call_site_count = caller->scale_count(profile.count()); 947 948 if (IncrementalInline && (AlwaysIncrementalInline || 949 (call_site_count > 0 && (input_not_const || !C->inlining_incrementally() || C->over_inlining_cutoff())))) { 950 return CallGenerator::for_mh_late_inline(caller, callee, input_not_const); 951 } else { 952 // Out-of-line call. 953 return CallGenerator::for_direct_call(callee); 954 } 955 } 956 957 static void cast_argument(int nargs, int arg_nb, ciType* t, GraphKit& kit) { 958 PhaseGVN& gvn = kit.gvn(); 959 Node* arg = kit.argument(arg_nb); 960 const Type* arg_type = arg->bottom_type(); 961 const Type* sig_type = TypeOopPtr::make_from_klass(t->as_klass()); 962 if (arg_type->isa_oopptr() && !arg_type->higher_equal(sig_type)) { 963 const Type* narrowed_arg_type = arg_type->join_speculative(sig_type); // keep speculative part 964 arg = gvn.transform(new CheckCastPPNode(kit.control(), arg, narrowed_arg_type)); 965 kit.set_argument(arg_nb, arg); 966 } 967 if (sig_type->is_valuetypeptr() && !arg->is_ValueType() && 968 !kit.gvn().type(arg)->maybe_null() && t->as_value_klass()->is_scalarizable()) { 969 arg = ValueTypeNode::make_from_oop(&kit, arg, t->as_value_klass()); 970 kit.set_argument(arg_nb, arg); 971 } 972 } 973 974 CallGenerator* CallGenerator::for_method_handle_inline(JVMState* jvms, ciMethod* caller, ciMethod* callee, bool& input_not_const, bool delayed_forbidden) { 975 GraphKit kit(jvms); 976 PhaseGVN& gvn = kit.gvn(); 977 Compile* C = kit.C; 978 vmIntrinsics::ID iid = callee->intrinsic_id(); 979 input_not_const = true; 980 switch (iid) { 981 case vmIntrinsics::_invokeBasic: 982 { 983 // Get MethodHandle receiver: 984 Node* receiver = kit.argument(0); 985 if (receiver->Opcode() == Op_ConP) { 986 input_not_const = false; 987 const TypeOopPtr* oop_ptr = receiver->bottom_type()->is_oopptr(); 988 ciMethod* target = oop_ptr->const_oop()->as_method_handle()->get_vmtarget(); 989 const int vtable_index = Method::invalid_vtable_index; 990 991 if (!ciMethod::is_consistent_info(callee, target)) { 992 print_inlining_failure(C, callee, jvms->depth() - 1, jvms->bci(), 993 "signatures mismatch"); 994 return NULL; 995 } 996 997 CallGenerator* cg = C->call_generator(target, vtable_index, 998 false /* call_does_dispatch */, 999 jvms, 1000 true /* allow_inline */, 1001 PROB_ALWAYS, 1002 NULL, 1003 true, 1004 delayed_forbidden); 1005 return cg; 1006 } else { 1007 print_inlining_failure(C, callee, jvms->depth() - 1, jvms->bci(), 1008 "receiver not constant"); 1009 } 1010 } 1011 break; 1012 1013 case vmIntrinsics::_linkToVirtual: 1014 case vmIntrinsics::_linkToStatic: 1015 case vmIntrinsics::_linkToSpecial: 1016 case vmIntrinsics::_linkToInterface: 1017 { 1018 int nargs = callee->arg_size(); 1019 // Get MemberName argument: 1020 Node* member_name = kit.argument(nargs - 1); 1021 if (member_name->Opcode() == Op_ConP) { 1022 input_not_const = false; 1023 const TypeOopPtr* oop_ptr = member_name->bottom_type()->is_oopptr(); 1024 ciMethod* target = oop_ptr->const_oop()->as_member_name()->get_vmtarget(); 1025 1026 if (!ciMethod::is_consistent_info(callee, target)) { 1027 print_inlining_failure(C, callee, jvms->depth() - 1, jvms->bci(), 1028 "signatures mismatch"); 1029 return NULL; 1030 } 1031 1032 // In lambda forms we erase signature types to avoid resolving issues 1033 // involving class loaders. When we optimize a method handle invoke 1034 // to a direct call we must cast the receiver and arguments to its 1035 // actual types. 1036 ciSignature* signature = target->signature(); 1037 const int receiver_skip = target->is_static() ? 0 : 1; 1038 // Cast receiver to its type. 1039 if (!target->is_static()) { 1040 cast_argument(nargs, 0, signature->accessing_klass(), kit); 1041 } 1042 // Cast reference arguments to its type. 1043 for (int i = 0, j = 0; i < signature->count(); i++) { 1044 ciType* t = signature->type_at(i); 1045 if (t->is_klass()) { 1046 cast_argument(nargs, receiver_skip + j, t, kit); 1047 } 1048 j += t->size(); // long and double take two slots 1049 } 1050 1051 // Try to get the most accurate receiver type 1052 const bool is_virtual = (iid == vmIntrinsics::_linkToVirtual); 1053 const bool is_virtual_or_interface = (is_virtual || iid == vmIntrinsics::_linkToInterface); 1054 int vtable_index = Method::invalid_vtable_index; 1055 bool call_does_dispatch = false; 1056 1057 ciKlass* speculative_receiver_type = NULL; 1058 if (is_virtual_or_interface) { 1059 ciInstanceKlass* klass = target->holder(); 1060 Node* receiver_node = kit.argument(0); 1061 const TypeOopPtr* receiver_type = gvn.type(receiver_node)->isa_oopptr(); 1062 // call_does_dispatch and vtable_index are out-parameters. They might be changed. 1063 // optimize_virtual_call() takes 2 different holder 1064 // arguments for a corner case that doesn't apply here (see 1065 // Parse::do_call()) 1066 target = C->optimize_virtual_call(caller, jvms->bci(), klass, klass, 1067 target, receiver_type, is_virtual, 1068 call_does_dispatch, vtable_index, // out-parameters 1069 false /* check_access */); 1070 // We lack profiling at this call but type speculation may 1071 // provide us with a type 1072 speculative_receiver_type = (receiver_type != NULL) ? receiver_type->speculative_type() : NULL; 1073 } 1074 CallGenerator* cg = C->call_generator(target, vtable_index, call_does_dispatch, jvms, 1075 !StressMethodHandleLinkerInlining /* allow_inline */, 1076 PROB_ALWAYS, 1077 speculative_receiver_type, 1078 true, 1079 delayed_forbidden); 1080 return cg; 1081 } else { 1082 print_inlining_failure(C, callee, jvms->depth() - 1, jvms->bci(), 1083 "member_name not constant"); 1084 } 1085 } 1086 break; 1087 1088 default: 1089 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)); 1090 break; 1091 } 1092 return NULL; 1093 } 1094 1095 1096 //------------------------PredicatedIntrinsicGenerator------------------------------ 1097 // Internal class which handles all predicated Intrinsic calls. 1098 class PredicatedIntrinsicGenerator : public CallGenerator { 1099 CallGenerator* _intrinsic; 1100 CallGenerator* _cg; 1101 1102 public: 1103 PredicatedIntrinsicGenerator(CallGenerator* intrinsic, 1104 CallGenerator* cg) 1105 : CallGenerator(cg->method()) 1106 { 1107 _intrinsic = intrinsic; 1108 _cg = cg; 1109 } 1110 1111 virtual bool is_virtual() const { return true; } 1112 virtual bool is_inlined() const { return true; } 1113 virtual bool is_intrinsic() const { return true; } 1114 1115 virtual JVMState* generate(JVMState* jvms); 1116 }; 1117 1118 1119 CallGenerator* CallGenerator::for_predicated_intrinsic(CallGenerator* intrinsic, 1120 CallGenerator* cg) { 1121 return new PredicatedIntrinsicGenerator(intrinsic, cg); 1122 } 1123 1124 1125 JVMState* PredicatedIntrinsicGenerator::generate(JVMState* jvms) { 1126 // The code we want to generate here is: 1127 // if (receiver == NULL) 1128 // uncommon_Trap 1129 // if (predicate(0)) 1130 // do_intrinsic(0) 1131 // else 1132 // if (predicate(1)) 1133 // do_intrinsic(1) 1134 // ... 1135 // else 1136 // do_java_comp 1137 1138 GraphKit kit(jvms); 1139 PhaseGVN& gvn = kit.gvn(); 1140 1141 CompileLog* log = kit.C->log(); 1142 if (log != NULL) { 1143 log->elem("predicated_intrinsic bci='%d' method='%d'", 1144 jvms->bci(), log->identify(method())); 1145 } 1146 1147 if (!method()->is_static()) { 1148 // We need an explicit receiver null_check before checking its type in predicate. 1149 // We share a map with the caller, so his JVMS gets adjusted. 1150 kit.null_check_receiver_before_call(method()); 1151 if (kit.stopped()) { 1152 return kit.transfer_exceptions_into_jvms(); 1153 } 1154 } 1155 1156 int n_predicates = _intrinsic->predicates_count(); 1157 assert(n_predicates > 0, "sanity"); 1158 1159 JVMState** result_jvms = NEW_RESOURCE_ARRAY(JVMState*, (n_predicates+1)); 1160 1161 // Region for normal compilation code if intrinsic failed. 1162 Node* slow_region = new RegionNode(1); 1163 1164 int results = 0; 1165 for (int predicate = 0; (predicate < n_predicates) && !kit.stopped(); predicate++) { 1166 #ifdef ASSERT 1167 JVMState* old_jvms = kit.jvms(); 1168 SafePointNode* old_map = kit.map(); 1169 Node* old_io = old_map->i_o(); 1170 Node* old_mem = old_map->memory(); 1171 Node* old_exc = old_map->next_exception(); 1172 #endif 1173 Node* else_ctrl = _intrinsic->generate_predicate(kit.sync_jvms(), predicate); 1174 #ifdef ASSERT 1175 // Assert(no_new_memory && no_new_io && no_new_exceptions) after generate_predicate. 1176 assert(old_jvms == kit.jvms(), "generate_predicate should not change jvm state"); 1177 SafePointNode* new_map = kit.map(); 1178 assert(old_io == new_map->i_o(), "generate_predicate should not change i_o"); 1179 assert(old_mem == new_map->memory(), "generate_predicate should not change memory"); 1180 assert(old_exc == new_map->next_exception(), "generate_predicate should not add exceptions"); 1181 #endif 1182 if (!kit.stopped()) { 1183 PreserveJVMState pjvms(&kit); 1184 // Generate intrinsic code: 1185 JVMState* new_jvms = _intrinsic->generate(kit.sync_jvms()); 1186 if (new_jvms == NULL) { 1187 // Intrinsic failed, use normal compilation path for this predicate. 1188 slow_region->add_req(kit.control()); 1189 } else { 1190 kit.add_exception_states_from(new_jvms); 1191 kit.set_jvms(new_jvms); 1192 if (!kit.stopped()) { 1193 result_jvms[results++] = kit.jvms(); 1194 } 1195 } 1196 } 1197 if (else_ctrl == NULL) { 1198 else_ctrl = kit.C->top(); 1199 } 1200 kit.set_control(else_ctrl); 1201 } 1202 if (!kit.stopped()) { 1203 // Final 'else' after predicates. 1204 slow_region->add_req(kit.control()); 1205 } 1206 if (slow_region->req() > 1) { 1207 PreserveJVMState pjvms(&kit); 1208 // Generate normal compilation code: 1209 kit.set_control(gvn.transform(slow_region)); 1210 JVMState* new_jvms = _cg->generate(kit.sync_jvms()); 1211 if (kit.failing()) 1212 return NULL; // might happen because of NodeCountInliningCutoff 1213 assert(new_jvms != NULL, "must be"); 1214 kit.add_exception_states_from(new_jvms); 1215 kit.set_jvms(new_jvms); 1216 if (!kit.stopped()) { 1217 result_jvms[results++] = kit.jvms(); 1218 } 1219 } 1220 1221 if (results == 0) { 1222 // All paths ended in uncommon traps. 1223 (void) kit.stop(); 1224 return kit.transfer_exceptions_into_jvms(); 1225 } 1226 1227 if (results == 1) { // Only one path 1228 kit.set_jvms(result_jvms[0]); 1229 return kit.transfer_exceptions_into_jvms(); 1230 } 1231 1232 // Merge all paths. 1233 kit.C->set_has_split_ifs(true); // Has chance for split-if optimization 1234 RegionNode* region = new RegionNode(results + 1); 1235 Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO); 1236 for (int i = 0; i < results; i++) { 1237 JVMState* jvms = result_jvms[i]; 1238 int path = i + 1; 1239 SafePointNode* map = jvms->map(); 1240 region->init_req(path, map->control()); 1241 iophi->set_req(path, map->i_o()); 1242 if (i == 0) { 1243 kit.set_jvms(jvms); 1244 } else { 1245 kit.merge_memory(map->merged_memory(), region, path); 1246 } 1247 } 1248 kit.set_control(gvn.transform(region)); 1249 kit.set_i_o(gvn.transform(iophi)); 1250 // Transform new memory Phis. 1251 for (MergeMemStream mms(kit.merged_memory()); mms.next_non_empty();) { 1252 Node* phi = mms.memory(); 1253 if (phi->is_Phi() && phi->in(0) == region) { 1254 mms.set_memory(gvn.transform(phi)); 1255 } 1256 } 1257 1258 // Merge debug info. 1259 Node** ins = NEW_RESOURCE_ARRAY(Node*, results); 1260 uint tos = kit.jvms()->stkoff() + kit.sp(); 1261 Node* map = kit.map(); 1262 uint limit = map->req(); 1263 for (uint i = TypeFunc::Parms; i < limit; i++) { 1264 // Skip unused stack slots; fast forward to monoff(); 1265 if (i == tos) { 1266 i = kit.jvms()->monoff(); 1267 if( i >= limit ) break; 1268 } 1269 Node* n = map->in(i); 1270 ins[0] = n; 1271 const Type* t = gvn.type(n); 1272 bool needs_phi = false; 1273 for (int j = 1; j < results; j++) { 1274 JVMState* jvms = result_jvms[j]; 1275 Node* jmap = jvms->map(); 1276 Node* m = NULL; 1277 if (jmap->req() > i) { 1278 m = jmap->in(i); 1279 if (m != n) { 1280 needs_phi = true; 1281 t = t->meet_speculative(gvn.type(m)); 1282 } 1283 } 1284 ins[j] = m; 1285 } 1286 if (needs_phi) { 1287 Node* phi = PhiNode::make(region, n, t); 1288 for (int j = 1; j < results; j++) { 1289 phi->set_req(j + 1, ins[j]); 1290 } 1291 map->set_req(i, gvn.transform(phi)); 1292 } 1293 } 1294 1295 return kit.transfer_exceptions_into_jvms(); 1296 } 1297 1298 //-------------------------UncommonTrapCallGenerator----------------------------- 1299 // Internal class which handles all out-of-line calls checking receiver type. 1300 class UncommonTrapCallGenerator : public CallGenerator { 1301 Deoptimization::DeoptReason _reason; 1302 Deoptimization::DeoptAction _action; 1303 1304 public: 1305 UncommonTrapCallGenerator(ciMethod* m, 1306 Deoptimization::DeoptReason reason, 1307 Deoptimization::DeoptAction action) 1308 : CallGenerator(m) 1309 { 1310 _reason = reason; 1311 _action = action; 1312 } 1313 1314 virtual bool is_virtual() const { ShouldNotReachHere(); return false; } 1315 virtual bool is_trap() const { return true; } 1316 1317 virtual JVMState* generate(JVMState* jvms); 1318 }; 1319 1320 1321 CallGenerator* 1322 CallGenerator::for_uncommon_trap(ciMethod* m, 1323 Deoptimization::DeoptReason reason, 1324 Deoptimization::DeoptAction action) { 1325 return new UncommonTrapCallGenerator(m, reason, action); 1326 } 1327 1328 1329 JVMState* UncommonTrapCallGenerator::generate(JVMState* jvms) { 1330 GraphKit kit(jvms); 1331 kit.C->print_inlining_update(this); 1332 // Take the trap with arguments pushed on the stack. (Cf. null_check_receiver). 1333 // Callsite signature can be different from actual method being called (i.e _linkTo* sites). 1334 // Use callsite signature always. 1335 ciMethod* declared_method = kit.method()->get_method_at_bci(kit.bci()); 1336 int nargs = declared_method->arg_size(); 1337 kit.inc_sp(nargs); 1338 assert(nargs <= kit.sp() && kit.sp() <= jvms->stk_size(), "sane sp w/ args pushed"); 1339 if (_reason == Deoptimization::Reason_class_check && 1340 _action == Deoptimization::Action_maybe_recompile) { 1341 // Temp fix for 6529811 1342 // Don't allow uncommon_trap to override our decision to recompile in the event 1343 // of a class cast failure for a monomorphic call as it will never let us convert 1344 // the call to either bi-morphic or megamorphic and can lead to unc-trap loops 1345 bool keep_exact_action = true; 1346 kit.uncommon_trap(_reason, _action, NULL, "monomorphic vcall checkcast", false, keep_exact_action); 1347 } else { 1348 kit.uncommon_trap(_reason, _action); 1349 } 1350 return kit.transfer_exceptions_into_jvms(); 1351 } 1352 1353 // (Note: Moved hook_up_call to GraphKit::set_edges_for_java_call.) 1354 1355 // (Node: Merged hook_up_exits into ParseGenerator::generate.) 1356 1357 #define NODES_OVERHEAD_PER_METHOD (30.0) 1358 #define NODES_PER_BYTECODE (9.5) 1359 1360 void WarmCallInfo::init(JVMState* call_site, ciMethod* call_method, ciCallProfile& profile, float prof_factor) { 1361 int call_count = profile.count(); 1362 int code_size = call_method->code_size(); 1363 1364 // Expected execution count is based on the historical count: 1365 _count = call_count < 0 ? 1 : call_site->method()->scale_count(call_count, prof_factor); 1366 1367 // Expected profit from inlining, in units of simple call-overheads. 1368 _profit = 1.0; 1369 1370 // Expected work performed by the call in units of call-overheads. 1371 // %%% need an empirical curve fit for "work" (time in call) 1372 float bytecodes_per_call = 3; 1373 _work = 1.0 + code_size / bytecodes_per_call; 1374 1375 // Expected size of compilation graph: 1376 // -XX:+PrintParseStatistics once reported: 1377 // Methods seen: 9184 Methods parsed: 9184 Nodes created: 1582391 1378 // Histogram of 144298 parsed bytecodes: 1379 // %%% Need an better predictor for graph size. 1380 _size = NODES_OVERHEAD_PER_METHOD + (NODES_PER_BYTECODE * code_size); 1381 } 1382 1383 // is_cold: Return true if the node should never be inlined. 1384 // This is true if any of the key metrics are extreme. 1385 bool WarmCallInfo::is_cold() const { 1386 if (count() < WarmCallMinCount) return true; 1387 if (profit() < WarmCallMinProfit) return true; 1388 if (work() > WarmCallMaxWork) return true; 1389 if (size() > WarmCallMaxSize) return true; 1390 return false; 1391 } 1392 1393 // is_hot: Return true if the node should be inlined immediately. 1394 // This is true if any of the key metrics are extreme. 1395 bool WarmCallInfo::is_hot() const { 1396 assert(!is_cold(), "eliminate is_cold cases before testing is_hot"); 1397 if (count() >= HotCallCountThreshold) return true; 1398 if (profit() >= HotCallProfitThreshold) return true; 1399 if (work() <= HotCallTrivialWork) return true; 1400 if (size() <= HotCallTrivialSize) return true; 1401 return false; 1402 } 1403 1404 // compute_heat: 1405 float WarmCallInfo::compute_heat() const { 1406 assert(!is_cold(), "compute heat only on warm nodes"); 1407 assert(!is_hot(), "compute heat only on warm nodes"); 1408 int min_size = MAX2(0, (int)HotCallTrivialSize); 1409 int max_size = MIN2(500, (int)WarmCallMaxSize); 1410 float method_size = (size() - min_size) / MAX2(1, max_size - min_size); 1411 float size_factor; 1412 if (method_size < 0.05) size_factor = 4; // 2 sigmas better than avg. 1413 else if (method_size < 0.15) size_factor = 2; // 1 sigma better than avg. 1414 else if (method_size < 0.5) size_factor = 1; // better than avg. 1415 else size_factor = 0.5; // worse than avg. 1416 return (count() * profit() * size_factor); 1417 } 1418 1419 bool WarmCallInfo::warmer_than(WarmCallInfo* that) { 1420 assert(this != that, "compare only different WCIs"); 1421 assert(this->heat() != 0 && that->heat() != 0, "call compute_heat 1st"); 1422 if (this->heat() > that->heat()) return true; 1423 if (this->heat() < that->heat()) return false; 1424 assert(this->heat() == that->heat(), "no NaN heat allowed"); 1425 // Equal heat. Break the tie some other way. 1426 if (!this->call() || !that->call()) return (address)this > (address)that; 1427 return this->call()->_idx > that->call()->_idx; 1428 } 1429 1430 //#define UNINIT_NEXT ((WarmCallInfo*)badAddress) 1431 #define UNINIT_NEXT ((WarmCallInfo*)NULL) 1432 1433 WarmCallInfo* WarmCallInfo::insert_into(WarmCallInfo* head) { 1434 assert(next() == UNINIT_NEXT, "not yet on any list"); 1435 WarmCallInfo* prev_p = NULL; 1436 WarmCallInfo* next_p = head; 1437 while (next_p != NULL && next_p->warmer_than(this)) { 1438 prev_p = next_p; 1439 next_p = prev_p->next(); 1440 } 1441 // Install this between prev_p and next_p. 1442 this->set_next(next_p); 1443 if (prev_p == NULL) 1444 head = this; 1445 else 1446 prev_p->set_next(this); 1447 return head; 1448 } 1449 1450 WarmCallInfo* WarmCallInfo::remove_from(WarmCallInfo* head) { 1451 WarmCallInfo* prev_p = NULL; 1452 WarmCallInfo* next_p = head; 1453 while (next_p != this) { 1454 assert(next_p != NULL, "this must be in the list somewhere"); 1455 prev_p = next_p; 1456 next_p = prev_p->next(); 1457 } 1458 next_p = this->next(); 1459 debug_only(this->set_next(UNINIT_NEXT)); 1460 // Remove this from between prev_p and next_p. 1461 if (prev_p == NULL) 1462 head = next_p; 1463 else 1464 prev_p->set_next(next_p); 1465 return head; 1466 } 1467 1468 WarmCallInfo WarmCallInfo::_always_hot(WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE(), 1469 WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE()); 1470 WarmCallInfo WarmCallInfo::_always_cold(WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE(), 1471 WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE()); 1472 1473 WarmCallInfo* WarmCallInfo::always_hot() { 1474 assert(_always_hot.is_hot(), "must always be hot"); 1475 return &_always_hot; 1476 } 1477 1478 WarmCallInfo* WarmCallInfo::always_cold() { 1479 assert(_always_cold.is_cold(), "must always be cold"); 1480 return &_always_cold; 1481 } 1482 1483 1484 #ifndef PRODUCT 1485 1486 void WarmCallInfo::print() const { 1487 tty->print("%s : C=%6.1f P=%6.1f W=%6.1f S=%6.1f H=%6.1f -> %p", 1488 is_cold() ? "cold" : is_hot() ? "hot " : "warm", 1489 count(), profit(), work(), size(), compute_heat(), next()); 1490 tty->cr(); 1491 if (call() != NULL) call()->dump(); 1492 } 1493 1494 void print_wci(WarmCallInfo* ci) { 1495 ci->print(); 1496 } 1497 1498 void WarmCallInfo::print_all() const { 1499 for (const WarmCallInfo* p = this; p != NULL; p = p->next()) 1500 p->print(); 1501 } 1502 1503 int WarmCallInfo::count_all() const { 1504 int cnt = 0; 1505 for (const WarmCallInfo* p = this; p != NULL; p = p->next()) 1506 cnt++; 1507 return cnt; 1508 } 1509 1510 #endif //PRODUCT