1 /*
2 * Copyright (C) 2014 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "inliner.h"
18
19 #include "art_method-inl.h"
20 #include "base/enums.h"
21 #include "base/logging.h"
22 #include "builder.h"
23 #include "class_linker.h"
24 #include "class_root-inl.h"
25 #include "constant_folding.h"
26 #include "data_type-inl.h"
27 #include "dead_code_elimination.h"
28 #include "dex/inline_method_analyser.h"
29 #include "driver/compiler_options.h"
30 #include "driver/dex_compilation_unit.h"
31 #include "instruction_simplifier.h"
32 #include "intrinsics.h"
33 #include "jit/jit.h"
34 #include "jit/jit_code_cache.h"
35 #include "mirror/class_loader.h"
36 #include "mirror/dex_cache.h"
37 #include "mirror/object_array-alloc-inl.h"
38 #include "mirror/object_array-inl.h"
39 #include "nodes.h"
40 #include "reference_type_propagation.h"
41 #include "register_allocator_linear_scan.h"
42 #include "scoped_thread_state_change-inl.h"
43 #include "sharpening.h"
44 #include "ssa_builder.h"
45 #include "ssa_phi_elimination.h"
46 #include "thread.h"
47 #include "verifier/verifier_compiler_binding.h"
48
49 namespace art HIDDEN {
50
51 // Instruction limit to control memory.
52 static constexpr size_t kMaximumNumberOfTotalInstructions = 1024;
53
54 // Maximum number of instructions for considering a method small,
55 // which we will always try to inline if the other non-instruction limits
56 // are not reached.
57 static constexpr size_t kMaximumNumberOfInstructionsForSmallMethod = 3;
58
59 // Limit the number of dex registers that we accumulate while inlining
60 // to avoid creating large amount of nested environments.
61 static constexpr size_t kMaximumNumberOfCumulatedDexRegisters = 32;
62
63 // Limit recursive call inlining, which do not benefit from too
64 // much inlining compared to code locality.
65 static constexpr size_t kMaximumNumberOfRecursiveCalls = 4;
66
67 // Limit recursive polymorphic call inlining to prevent code bloat, since it can quickly get out of
68 // hand in the presence of multiple Wrapper classes. We set this to 0 to disallow polymorphic
69 // recursive calls at all.
70 static constexpr size_t kMaximumNumberOfPolymorphicRecursiveCalls = 0;
71
72 // Controls the use of inline caches in AOT mode.
73 static constexpr bool kUseAOTInlineCaches = true;
74
75 // Controls the use of inlining try catches.
76 static constexpr bool kInlineTryCatches = true;
77
78 // We check for line numbers to make sure the DepthString implementation
79 // aligns the output nicely.
80 #define LOG_INTERNAL(msg) \
81 static_assert(__LINE__ > 10, "Unhandled line number"); \
82 static_assert(__LINE__ < 10000, "Unhandled line number"); \
83 VLOG(compiler) << DepthString(__LINE__) << msg
84
85 #define LOG_TRY() LOG_INTERNAL("Try inlinining call: ")
86 #define LOG_NOTE() LOG_INTERNAL("Note: ")
87 #define LOG_SUCCESS() LOG_INTERNAL("Success: ")
88 #define LOG_FAIL(stats_ptr, stat) MaybeRecordStat(stats_ptr, stat); LOG_INTERNAL("Fail: ")
89 #define LOG_FAIL_NO_STAT() LOG_INTERNAL("Fail: ")
90
DepthString(int line) const91 std::string HInliner::DepthString(int line) const {
92 std::string value;
93 // Indent according to the inlining depth.
94 size_t count = depth_;
95 // Line numbers get printed in the log, so add a space if the log's line is less
96 // than 1000, and two if less than 100. 10 cannot be reached as it's the copyright.
97 if (!kIsTargetBuild) {
98 if (line < 100) {
99 value += " ";
100 }
101 if (line < 1000) {
102 value += " ";
103 }
104 // Safeguard if this file reaches more than 10000 lines.
105 DCHECK_LT(line, 10000);
106 }
107 for (size_t i = 0; i < count; ++i) {
108 value += " ";
109 }
110 return value;
111 }
112
CountNumberOfInstructions(HGraph * graph)113 static size_t CountNumberOfInstructions(HGraph* graph) {
114 size_t number_of_instructions = 0;
115 for (HBasicBlock* block : graph->GetReversePostOrderSkipEntryBlock()) {
116 for (HInstructionIterator instr_it(block->GetInstructions());
117 !instr_it.Done();
118 instr_it.Advance()) {
119 ++number_of_instructions;
120 }
121 }
122 return number_of_instructions;
123 }
124
UpdateInliningBudget()125 void HInliner::UpdateInliningBudget() {
126 if (total_number_of_instructions_ >= kMaximumNumberOfTotalInstructions) {
127 // Always try to inline small methods.
128 inlining_budget_ = kMaximumNumberOfInstructionsForSmallMethod;
129 } else {
130 inlining_budget_ = std::max(
131 kMaximumNumberOfInstructionsForSmallMethod,
132 kMaximumNumberOfTotalInstructions - total_number_of_instructions_);
133 }
134 }
135
Run()136 bool HInliner::Run() {
137 if (codegen_->GetCompilerOptions().GetInlineMaxCodeUnits() == 0) {
138 // Inlining effectively disabled.
139 return false;
140 } else if (graph_->IsDebuggable()) {
141 // For simplicity, we currently never inline when the graph is debuggable. This avoids
142 // doing some logic in the runtime to discover if a method could have been inlined.
143 return false;
144 }
145
146 bool did_inline = false;
147 // The inliner is the only phase that sets invokes as `always throwing`, and since we only run the
148 // inliner once per graph this value should always be false at the beginning of the inlining
149 // phase. This is important since we use `HasAlwaysThrowingInvokes` to know whether the inliner
150 // phase performed a relevant change in the graph.
151 DCHECK(!graph_->HasAlwaysThrowingInvokes());
152
153 // Initialize the number of instructions for the method being compiled. Recursive calls
154 // to HInliner::Run have already updated the instruction count.
155 if (outermost_graph_ == graph_) {
156 total_number_of_instructions_ = CountNumberOfInstructions(graph_);
157 }
158
159 UpdateInliningBudget();
160 DCHECK_NE(total_number_of_instructions_, 0u);
161 DCHECK_NE(inlining_budget_, 0u);
162
163 // If we're compiling tests, honor inlining directives in method names:
164 // - if a method's name contains the substring "$noinline$", do not
165 // inline that method;
166 // - if a method's name contains the substring "$inline$", ensure
167 // that this method is actually inlined.
168 // We limit the latter to AOT compilation, as the JIT may or may not inline
169 // depending on the state of classes at runtime.
170 const bool honor_noinline_directives = codegen_->GetCompilerOptions().CompileArtTest();
171 const bool honor_inline_directives =
172 honor_noinline_directives && Runtime::Current()->IsAotCompiler();
173
174 // Keep a copy of all blocks when starting the visit.
175 ArenaVector<HBasicBlock*> blocks = graph_->GetReversePostOrder();
176 DCHECK(!blocks.empty());
177 // Because we are changing the graph when inlining,
178 // we just iterate over the blocks of the outer method.
179 // This avoids doing the inlining work again on the inlined blocks.
180 for (HBasicBlock* block : blocks) {
181 for (HInstruction* instruction = block->GetFirstInstruction(); instruction != nullptr;) {
182 HInstruction* next = instruction->GetNext();
183 HInvoke* call = instruction->AsInvoke();
184 // As long as the call is not intrinsified, it is worth trying to inline.
185 if (call != nullptr && !codegen_->IsImplementedIntrinsic(call)) {
186 if (honor_noinline_directives) {
187 // Debugging case: directives in method names control or assert on inlining.
188 std::string callee_name =
189 call->GetMethodReference().PrettyMethod(/* with_signature= */ false);
190 // Tests prevent inlining by having $noinline$ in their method names.
191 if (callee_name.find("$noinline$") == std::string::npos) {
192 if (TryInline(call)) {
193 did_inline = true;
194 } else if (honor_inline_directives) {
195 bool should_have_inlined = (callee_name.find("$inline$") != std::string::npos);
196 CHECK(!should_have_inlined) << "Could not inline " << callee_name;
197 }
198 }
199 } else {
200 DCHECK(!honor_inline_directives);
201 // Normal case: try to inline.
202 if (TryInline(call)) {
203 did_inline = true;
204 }
205 }
206 }
207 instruction = next;
208 }
209 }
210
211 // We return true if we either inlined at least one method, or we marked one of our methods as
212 // always throwing.
213 return did_inline || graph_->HasAlwaysThrowingInvokes();
214 }
215
IsMethodOrDeclaringClassFinal(ArtMethod * method)216 static bool IsMethodOrDeclaringClassFinal(ArtMethod* method)
217 REQUIRES_SHARED(Locks::mutator_lock_) {
218 return method->IsFinal() || method->GetDeclaringClass()->IsFinal();
219 }
220
221 /**
222 * Given the `resolved_method` looked up in the dex cache, try to find
223 * the actual runtime target of an interface or virtual call.
224 * Return nullptr if the runtime target cannot be proven.
225 */
FindVirtualOrInterfaceTarget(HInvoke * invoke)226 static ArtMethod* FindVirtualOrInterfaceTarget(HInvoke* invoke)
227 REQUIRES_SHARED(Locks::mutator_lock_) {
228 ArtMethod* resolved_method = invoke->GetResolvedMethod();
229 if (IsMethodOrDeclaringClassFinal(resolved_method)) {
230 // No need to lookup further, the resolved method will be the target.
231 return resolved_method;
232 }
233
234 HInstruction* receiver = invoke->InputAt(0);
235 if (receiver->IsNullCheck()) {
236 // Due to multiple levels of inlining within the same pass, it might be that
237 // null check does not have the reference type of the actual receiver.
238 receiver = receiver->InputAt(0);
239 }
240 ReferenceTypeInfo info = receiver->GetReferenceTypeInfo();
241 DCHECK(info.IsValid()) << "Invalid RTI for " << receiver->DebugName();
242 if (!info.IsExact()) {
243 // We currently only support inlining with known receivers.
244 // TODO: Remove this check, we should be able to inline final methods
245 // on unknown receivers.
246 return nullptr;
247 } else if (info.GetTypeHandle()->IsInterface()) {
248 // Statically knowing that the receiver has an interface type cannot
249 // help us find what is the target method.
250 return nullptr;
251 } else if (!resolved_method->GetDeclaringClass()->IsAssignableFrom(info.GetTypeHandle().Get())) {
252 // The method that we're trying to call is not in the receiver's class or super classes.
253 return nullptr;
254 } else if (info.GetTypeHandle()->IsErroneous()) {
255 // If the type is erroneous, do not go further, as we are going to query the vtable or
256 // imt table, that we can only safely do on non-erroneous classes.
257 return nullptr;
258 }
259
260 ClassLinker* cl = Runtime::Current()->GetClassLinker();
261 PointerSize pointer_size = cl->GetImagePointerSize();
262 if (invoke->IsInvokeInterface()) {
263 resolved_method = info.GetTypeHandle()->FindVirtualMethodForInterface(
264 resolved_method, pointer_size);
265 } else {
266 DCHECK(invoke->IsInvokeVirtual());
267 resolved_method = info.GetTypeHandle()->FindVirtualMethodForVirtual(
268 resolved_method, pointer_size);
269 }
270
271 if (resolved_method == nullptr) {
272 // The information we had on the receiver was not enough to find
273 // the target method. Since we check above the exact type of the receiver,
274 // the only reason this can happen is an IncompatibleClassChangeError.
275 return nullptr;
276 } else if (!resolved_method->IsInvokable()) {
277 // The information we had on the receiver was not enough to find
278 // the target method. Since we check above the exact type of the receiver,
279 // the only reason this can happen is an IncompatibleClassChangeError.
280 return nullptr;
281 } else if (IsMethodOrDeclaringClassFinal(resolved_method)) {
282 // A final method has to be the target method.
283 return resolved_method;
284 } else if (info.IsExact()) {
285 // If we found a method and the receiver's concrete type is statically
286 // known, we know for sure the target.
287 return resolved_method;
288 } else {
289 // Even if we did find a method, the receiver type was not enough to
290 // statically find the runtime target.
291 return nullptr;
292 }
293 }
294
FindMethodIndexIn(ArtMethod * method,const DexFile & dex_file,uint32_t name_and_signature_index)295 static uint32_t FindMethodIndexIn(ArtMethod* method,
296 const DexFile& dex_file,
297 uint32_t name_and_signature_index)
298 REQUIRES_SHARED(Locks::mutator_lock_) {
299 if (IsSameDexFile(*method->GetDexFile(), dex_file)) {
300 return method->GetDexMethodIndex();
301 } else {
302 return method->FindDexMethodIndexInOtherDexFile(dex_file, name_and_signature_index);
303 }
304 }
305
FindClassIndexIn(ObjPtr<mirror::Class> cls,const DexCompilationUnit & compilation_unit)306 static dex::TypeIndex FindClassIndexIn(ObjPtr<mirror::Class> cls,
307 const DexCompilationUnit& compilation_unit)
308 REQUIRES_SHARED(Locks::mutator_lock_) {
309 const DexFile& dex_file = *compilation_unit.GetDexFile();
310 dex::TypeIndex index;
311 if (cls->GetDexCache() == nullptr) {
312 DCHECK(cls->IsArrayClass()) << cls->PrettyClass();
313 index = cls->FindTypeIndexInOtherDexFile(dex_file);
314 } else if (!cls->GetDexTypeIndex().IsValid()) {
315 DCHECK(cls->IsProxyClass()) << cls->PrettyClass();
316 // TODO: deal with proxy classes.
317 } else if (IsSameDexFile(cls->GetDexFile(), dex_file)) {
318 DCHECK_EQ(cls->GetDexCache(), compilation_unit.GetDexCache().Get());
319 index = cls->GetDexTypeIndex();
320 } else {
321 index = cls->FindTypeIndexInOtherDexFile(dex_file);
322 // We cannot guarantee the entry will resolve to the same class,
323 // as there may be different class loaders. So only return the index if it's
324 // the right class already resolved with the class loader.
325 if (index.IsValid()) {
326 ObjPtr<mirror::Class> resolved = compilation_unit.GetClassLinker()->LookupResolvedType(
327 index, compilation_unit.GetDexCache().Get(), compilation_unit.GetClassLoader().Get());
328 if (resolved != cls) {
329 index = dex::TypeIndex::Invalid();
330 }
331 }
332 }
333
334 return index;
335 }
336
GetInlineCacheType(const StackHandleScope<InlineCache::kIndividualCacheSize> & classes)337 HInliner::InlineCacheType HInliner::GetInlineCacheType(
338 const StackHandleScope<InlineCache::kIndividualCacheSize>& classes) {
339 DCHECK_EQ(classes.NumberOfReferences(), InlineCache::kIndividualCacheSize);
340 uint8_t number_of_types = InlineCache::kIndividualCacheSize - classes.RemainingSlots();
341 if (number_of_types == 0) {
342 return kInlineCacheUninitialized;
343 } else if (number_of_types == 1) {
344 return kInlineCacheMonomorphic;
345 } else if (number_of_types == InlineCache::kIndividualCacheSize) {
346 return kInlineCacheMegamorphic;
347 } else {
348 return kInlineCachePolymorphic;
349 }
350 }
351
GetMonomorphicType(const StackHandleScope<InlineCache::kIndividualCacheSize> & classes)352 static inline ObjPtr<mirror::Class> GetMonomorphicType(
353 const StackHandleScope<InlineCache::kIndividualCacheSize>& classes)
354 REQUIRES_SHARED(Locks::mutator_lock_) {
355 DCHECK(classes.GetReference(0) != nullptr);
356 return classes.GetReference(0)->AsClass();
357 }
358
FindMethodFromCHA(ArtMethod * resolved_method)359 ArtMethod* HInliner::FindMethodFromCHA(ArtMethod* resolved_method) {
360 if (!resolved_method->HasSingleImplementation()) {
361 return nullptr;
362 }
363 if (Runtime::Current()->IsAotCompiler()) {
364 // No CHA-based devirtulization for AOT compiler (yet).
365 return nullptr;
366 }
367 if (Runtime::Current()->IsZygote()) {
368 // No CHA-based devirtulization for Zygote, as it compiles with
369 // offline information.
370 return nullptr;
371 }
372 if (outermost_graph_->IsCompilingOsr()) {
373 // We do not support HDeoptimize in OSR methods.
374 return nullptr;
375 }
376 PointerSize pointer_size = caller_compilation_unit_.GetClassLinker()->GetImagePointerSize();
377 ArtMethod* single_impl = resolved_method->GetSingleImplementation(pointer_size);
378 if (single_impl == nullptr) {
379 return nullptr;
380 }
381 if (single_impl->IsProxyMethod()) {
382 // Proxy method is a generic invoker that's not worth
383 // devirtualizing/inlining. It also causes issues when the proxy
384 // method is in another dex file if we try to rewrite invoke-interface to
385 // invoke-virtual because a proxy method doesn't have a real dex file.
386 return nullptr;
387 }
388 if (!single_impl->GetDeclaringClass()->IsResolved()) {
389 // There's a race with the class loading, which updates the CHA info
390 // before setting the class to resolved. So we just bail for this
391 // rare occurence.
392 return nullptr;
393 }
394 return single_impl;
395 }
396
IsMethodVerified(ArtMethod * method)397 static bool IsMethodVerified(ArtMethod* method)
398 REQUIRES_SHARED(Locks::mutator_lock_) {
399 if (method->GetDeclaringClass()->IsVerified()) {
400 return true;
401 }
402 // For AOT, we check if the class has a verification status that allows us to
403 // inline / analyze.
404 // At runtime, we know this is cold code if the class is not verified, so don't
405 // bother analyzing.
406 if (Runtime::Current()->IsAotCompiler()) {
407 if (method->GetDeclaringClass()->IsVerifiedNeedsAccessChecks() ||
408 method->GetDeclaringClass()->ShouldVerifyAtRuntime()) {
409 return true;
410 }
411 }
412 return false;
413 }
414
AlwaysThrows(ArtMethod * method)415 static bool AlwaysThrows(ArtMethod* method)
416 REQUIRES_SHARED(Locks::mutator_lock_) {
417 DCHECK(method != nullptr);
418 // Skip non-compilable and unverified methods.
419 if (!method->IsCompilable() || !IsMethodVerified(method)) {
420 return false;
421 }
422 // Skip native methods, methods with try blocks, and methods that are too large.
423 CodeItemDataAccessor accessor(method->DexInstructionData());
424 if (!accessor.HasCodeItem() ||
425 accessor.TriesSize() != 0 ||
426 accessor.InsnsSizeInCodeUnits() > kMaximumNumberOfTotalInstructions) {
427 return false;
428 }
429 // Scan for exits.
430 bool throw_seen = false;
431 for (const DexInstructionPcPair& pair : accessor) {
432 switch (pair.Inst().Opcode()) {
433 case Instruction::RETURN:
434 case Instruction::RETURN_VOID:
435 case Instruction::RETURN_WIDE:
436 case Instruction::RETURN_OBJECT:
437 return false; // found regular control flow back
438 case Instruction::THROW:
439 throw_seen = true;
440 break;
441 default:
442 break;
443 }
444 }
445 return throw_seen;
446 }
447
TryInline(HInvoke * invoke_instruction)448 bool HInliner::TryInline(HInvoke* invoke_instruction) {
449 MaybeRecordStat(stats_, MethodCompilationStat::kTryInline);
450
451 // Don't bother to move further if we know the method is unresolved or the invocation is
452 // polymorphic (invoke-{polymorphic,custom}).
453 if (invoke_instruction->IsInvokeUnresolved()) {
454 MaybeRecordStat(stats_, MethodCompilationStat::kNotInlinedUnresolved);
455 return false;
456 } else if (invoke_instruction->IsInvokePolymorphic()) {
457 MaybeRecordStat(stats_, MethodCompilationStat::kNotInlinedPolymorphic);
458 return false;
459 } else if (invoke_instruction->IsInvokeCustom()) {
460 MaybeRecordStat(stats_, MethodCompilationStat::kNotInlinedCustom);
461 return false;
462 }
463
464 ScopedObjectAccess soa(Thread::Current());
465 LOG_TRY() << invoke_instruction->GetMethodReference().PrettyMethod();
466
467 ArtMethod* resolved_method = invoke_instruction->GetResolvedMethod();
468 if (resolved_method == nullptr) {
469 DCHECK(invoke_instruction->IsInvokeStaticOrDirect());
470 DCHECK(invoke_instruction->AsInvokeStaticOrDirect()->IsStringInit());
471 LOG_FAIL_NO_STAT() << "Not inlining a String.<init> method";
472 return false;
473 }
474
475 ArtMethod* actual_method = invoke_instruction->IsInvokeStaticOrDirect()
476 ? invoke_instruction->GetResolvedMethod()
477 : FindVirtualOrInterfaceTarget(invoke_instruction);
478
479 if (actual_method != nullptr) {
480 // Single target.
481 bool result = TryInlineAndReplace(invoke_instruction,
482 actual_method,
483 ReferenceTypeInfo::CreateInvalid(),
484 /* do_rtp= */ true,
485 /* is_speculative= */ false);
486 if (result) {
487 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedInvokeVirtualOrInterface);
488 if (outermost_graph_ == graph_) {
489 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedLastInvokeVirtualOrInterface);
490 }
491 } else {
492 HInvoke* invoke_to_analyze = nullptr;
493 if (TryDevirtualize(invoke_instruction, actual_method, &invoke_to_analyze)) {
494 // Consider devirtualization as inlining.
495 result = true;
496 MaybeRecordStat(stats_, MethodCompilationStat::kDevirtualized);
497 } else {
498 invoke_to_analyze = invoke_instruction;
499 }
500 // Set always throws property for non-inlined method call with single target.
501 if (invoke_instruction->AlwaysThrows() || AlwaysThrows(actual_method)) {
502 invoke_to_analyze->SetAlwaysThrows(/* always_throws= */ true);
503 graph_->SetHasAlwaysThrowingInvokes(/* value= */ true);
504 }
505 }
506 return result;
507 }
508
509 DCHECK(!invoke_instruction->IsInvokeStaticOrDirect());
510
511 // No try catch inlining allowed here, or recursively. For try catch inlining we are banking on
512 // the fact that we have a unique dex pc list. We cannot guarantee that for some TryInline methods
513 // e.g. `TryInlinePolymorphicCall`.
514 // TODO(solanes): Setting `try_catch_inlining_allowed_` to false here covers all cases from
515 // `TryInlineFromCHA` and from `TryInlineFromInlineCache` as well (e.g.
516 // `TryInlinePolymorphicCall`). Reassess to see if we can inline inline catch blocks in
517 // `TryInlineFromCHA`, `TryInlineMonomorphicCall` and `TryInlinePolymorphicCallToSameTarget`.
518
519 // We store the value to restore it since we will use the same HInliner instance for other inlinee
520 // candidates.
521 const bool previous_value = try_catch_inlining_allowed_;
522 try_catch_inlining_allowed_ = false;
523
524 if (TryInlineFromCHA(invoke_instruction)) {
525 try_catch_inlining_allowed_ = previous_value;
526 return true;
527 }
528
529 const bool result = TryInlineFromInlineCache(invoke_instruction);
530 try_catch_inlining_allowed_ = previous_value;
531 return result;
532 }
533
TryInlineFromCHA(HInvoke * invoke_instruction)534 bool HInliner::TryInlineFromCHA(HInvoke* invoke_instruction) {
535 ArtMethod* method = FindMethodFromCHA(invoke_instruction->GetResolvedMethod());
536 if (method == nullptr) {
537 return false;
538 }
539 LOG_NOTE() << "Try CHA-based inlining of " << method->PrettyMethod();
540
541 uint32_t dex_pc = invoke_instruction->GetDexPc();
542 HInstruction* cursor = invoke_instruction->GetPrevious();
543 HBasicBlock* bb_cursor = invoke_instruction->GetBlock();
544 if (!TryInlineAndReplace(invoke_instruction,
545 method,
546 ReferenceTypeInfo::CreateInvalid(),
547 /* do_rtp= */ true,
548 /* is_speculative= */ true)) {
549 return false;
550 }
551 AddCHAGuard(invoke_instruction, dex_pc, cursor, bb_cursor);
552 // Add dependency due to devirtualization: we are assuming the resolved method
553 // has a single implementation.
554 outermost_graph_->AddCHASingleImplementationDependency(invoke_instruction->GetResolvedMethod());
555 MaybeRecordStat(stats_, MethodCompilationStat::kCHAInline);
556 return true;
557 }
558
UseOnlyPolymorphicInliningWithNoDeopt()559 bool HInliner::UseOnlyPolymorphicInliningWithNoDeopt() {
560 // If we are compiling AOT or OSR, pretend the call using inline caches is polymorphic and
561 // do not generate a deopt.
562 //
563 // For AOT:
564 // Generating a deopt does not ensure that we will actually capture the new types;
565 // and the danger is that we could be stuck in a loop with "forever" deoptimizations.
566 // Take for example the following scenario:
567 // - we capture the inline cache in one run
568 // - the next run, we deoptimize because we miss a type check, but the method
569 // never becomes hot again
570 // In this case, the inline cache will not be updated in the profile and the AOT code
571 // will keep deoptimizing.
572 // Another scenario is if we use profile compilation for a process which is not allowed
573 // to JIT (e.g. system server). If we deoptimize we will run interpreted code for the
574 // rest of the lifetime.
575 // TODO(calin):
576 // This is a compromise because we will most likely never update the inline cache
577 // in the profile (unless there's another reason to deopt). So we might be stuck with
578 // a sub-optimal inline cache.
579 // We could be smarter when capturing inline caches to mitigate this.
580 // (e.g. by having different thresholds for new and old methods).
581 //
582 // For OSR:
583 // We may come from the interpreter and it may have seen different receiver types.
584 return Runtime::Current()->IsAotCompiler() || outermost_graph_->IsCompilingOsr();
585 }
TryInlineFromInlineCache(HInvoke * invoke_instruction)586 bool HInliner::TryInlineFromInlineCache(HInvoke* invoke_instruction)
587 REQUIRES_SHARED(Locks::mutator_lock_) {
588 if (Runtime::Current()->IsAotCompiler() && !kUseAOTInlineCaches) {
589 return false;
590 }
591
592 StackHandleScope<InlineCache::kIndividualCacheSize> classes(Thread::Current());
593 // The Zygote JIT compiles based on a profile, so we shouldn't use runtime inline caches
594 // for it.
595 InlineCacheType inline_cache_type =
596 (Runtime::Current()->IsAotCompiler() || Runtime::Current()->IsZygote())
597 ? GetInlineCacheAOT(invoke_instruction, &classes)
598 : GetInlineCacheJIT(invoke_instruction, &classes);
599
600 switch (inline_cache_type) {
601 case kInlineCacheNoData: {
602 LOG_FAIL_NO_STAT()
603 << "No inline cache information for call to "
604 << invoke_instruction->GetMethodReference().PrettyMethod();
605 return false;
606 }
607
608 case kInlineCacheUninitialized: {
609 LOG_FAIL_NO_STAT()
610 << "Interface or virtual call to "
611 << invoke_instruction->GetMethodReference().PrettyMethod()
612 << " is not hit and not inlined";
613 return false;
614 }
615
616 case kInlineCacheMonomorphic: {
617 MaybeRecordStat(stats_, MethodCompilationStat::kMonomorphicCall);
618 if (UseOnlyPolymorphicInliningWithNoDeopt()) {
619 return TryInlinePolymorphicCall(invoke_instruction, classes);
620 } else {
621 return TryInlineMonomorphicCall(invoke_instruction, classes);
622 }
623 }
624
625 case kInlineCachePolymorphic: {
626 MaybeRecordStat(stats_, MethodCompilationStat::kPolymorphicCall);
627 return TryInlinePolymorphicCall(invoke_instruction, classes);
628 }
629
630 case kInlineCacheMegamorphic: {
631 LOG_FAIL_NO_STAT()
632 << "Interface or virtual call to "
633 << invoke_instruction->GetMethodReference().PrettyMethod()
634 << " is megamorphic and not inlined";
635 MaybeRecordStat(stats_, MethodCompilationStat::kMegamorphicCall);
636 return false;
637 }
638
639 case kInlineCacheMissingTypes: {
640 LOG_FAIL_NO_STAT()
641 << "Interface or virtual call to "
642 << invoke_instruction->GetMethodReference().PrettyMethod()
643 << " is missing types and not inlined";
644 return false;
645 }
646 }
647 UNREACHABLE();
648 }
649
GetInlineCacheJIT(HInvoke * invoke_instruction,StackHandleScope<InlineCache::kIndividualCacheSize> * classes)650 HInliner::InlineCacheType HInliner::GetInlineCacheJIT(
651 HInvoke* invoke_instruction,
652 /*out*/StackHandleScope<InlineCache::kIndividualCacheSize>* classes) {
653 DCHECK(codegen_->GetCompilerOptions().IsJitCompiler());
654
655 ArtMethod* caller = graph_->GetArtMethod();
656 // Under JIT, we should always know the caller.
657 DCHECK(caller != nullptr);
658 ProfilingInfo* profiling_info = graph_->GetProfilingInfo();
659 if (profiling_info == nullptr) {
660 return kInlineCacheNoData;
661 }
662
663 Runtime::Current()->GetJit()->GetCodeCache()->CopyInlineCacheInto(
664 *profiling_info->GetInlineCache(invoke_instruction->GetDexPc()),
665 classes);
666 return GetInlineCacheType(*classes);
667 }
668
GetInlineCacheAOT(HInvoke * invoke_instruction,StackHandleScope<InlineCache::kIndividualCacheSize> * classes)669 HInliner::InlineCacheType HInliner::GetInlineCacheAOT(
670 HInvoke* invoke_instruction,
671 /*out*/StackHandleScope<InlineCache::kIndividualCacheSize>* classes) {
672 DCHECK_EQ(classes->NumberOfReferences(), InlineCache::kIndividualCacheSize);
673 DCHECK_EQ(classes->RemainingSlots(), InlineCache::kIndividualCacheSize);
674
675 const ProfileCompilationInfo* pci = codegen_->GetCompilerOptions().GetProfileCompilationInfo();
676 if (pci == nullptr) {
677 return kInlineCacheNoData;
678 }
679
680 ProfileCompilationInfo::MethodHotness hotness = pci->GetMethodHotness(MethodReference(
681 caller_compilation_unit_.GetDexFile(), caller_compilation_unit_.GetDexMethodIndex()));
682 if (!hotness.IsHot()) {
683 return kInlineCacheNoData; // no profile information for this invocation.
684 }
685
686 const ProfileCompilationInfo::InlineCacheMap* inline_caches = hotness.GetInlineCacheMap();
687 DCHECK(inline_caches != nullptr);
688 const auto it = inline_caches->find(invoke_instruction->GetDexPc());
689 if (it == inline_caches->end()) {
690 return kInlineCacheUninitialized;
691 }
692
693 const ProfileCompilationInfo::DexPcData& dex_pc_data = it->second;
694 if (dex_pc_data.is_missing_types) {
695 return kInlineCacheMissingTypes;
696 }
697 if (dex_pc_data.is_megamorphic) {
698 return kInlineCacheMegamorphic;
699 }
700 DCHECK_LE(dex_pc_data.classes.size(), InlineCache::kIndividualCacheSize);
701
702 // Walk over the class descriptors and look up the actual classes.
703 // If we cannot find a type we return kInlineCacheMissingTypes.
704 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
705 for (const dex::TypeIndex& type_index : dex_pc_data.classes) {
706 const DexFile* dex_file = caller_compilation_unit_.GetDexFile();
707 const char* descriptor = pci->GetTypeDescriptor(dex_file, type_index);
708 ObjPtr<mirror::ClassLoader> class_loader = caller_compilation_unit_.GetClassLoader().Get();
709 ObjPtr<mirror::Class> clazz = class_linker->LookupResolvedType(descriptor, class_loader);
710 if (clazz == nullptr) {
711 VLOG(compiler) << "Could not find class from inline cache in AOT mode "
712 << invoke_instruction->GetMethodReference().PrettyMethod()
713 << " : "
714 << descriptor;
715 return kInlineCacheMissingTypes;
716 }
717 DCHECK_NE(classes->RemainingSlots(), 0u);
718 classes->NewHandle(clazz);
719 }
720
721 return GetInlineCacheType(*classes);
722 }
723
BuildGetReceiverClass(ClassLinker * class_linker,HInstruction * receiver,uint32_t dex_pc) const724 HInstanceFieldGet* HInliner::BuildGetReceiverClass(ClassLinker* class_linker,
725 HInstruction* receiver,
726 uint32_t dex_pc) const {
727 ArtField* field = GetClassRoot<mirror::Object>(class_linker)->GetInstanceField(0);
728 DCHECK_EQ(std::string(field->GetName()), "shadow$_klass_");
729 HInstanceFieldGet* result = new (graph_->GetAllocator()) HInstanceFieldGet(
730 receiver,
731 field,
732 DataType::Type::kReference,
733 field->GetOffset(),
734 field->IsVolatile(),
735 field->GetDexFieldIndex(),
736 field->GetDeclaringClass()->GetDexClassDefIndex(),
737 *field->GetDexFile(),
738 dex_pc);
739 // The class of a field is effectively final, and does not have any memory dependencies.
740 result->SetSideEffects(SideEffects::None());
741 return result;
742 }
743
ResolveMethodFromInlineCache(Handle<mirror::Class> klass,HInvoke * invoke_instruction,PointerSize pointer_size)744 static ArtMethod* ResolveMethodFromInlineCache(Handle<mirror::Class> klass,
745 HInvoke* invoke_instruction,
746 PointerSize pointer_size)
747 REQUIRES_SHARED(Locks::mutator_lock_) {
748 ArtMethod* resolved_method = invoke_instruction->GetResolvedMethod();
749 if (Runtime::Current()->IsAotCompiler()) {
750 // We can get unrelated types when working with profiles (corruption,
751 // systme updates, or anyone can write to it). So first check if the class
752 // actually implements the declaring class of the method that is being
753 // called in bytecode.
754 // Note: the lookup methods used below require to have assignable types.
755 if (!resolved_method->GetDeclaringClass()->IsAssignableFrom(klass.Get())) {
756 return nullptr;
757 }
758
759 // Also check whether the type in the inline cache is an interface or an
760 // abstract class. We only expect concrete classes in inline caches, so this
761 // means the class was changed.
762 if (klass->IsAbstract() || klass->IsInterface()) {
763 return nullptr;
764 }
765 }
766
767 if (invoke_instruction->IsInvokeInterface()) {
768 resolved_method = klass->FindVirtualMethodForInterface(resolved_method, pointer_size);
769 } else {
770 DCHECK(invoke_instruction->IsInvokeVirtual());
771 resolved_method = klass->FindVirtualMethodForVirtual(resolved_method, pointer_size);
772 }
773 // Even if the class exists we can still not have the function the
774 // inline-cache targets if the profile is from far enough in the past/future.
775 // We need to allow this since we don't update boot-profiles very often. This
776 // can occur in boot-profiles with inline-caches.
777 DCHECK(Runtime::Current()->IsAotCompiler() || resolved_method != nullptr);
778 return resolved_method;
779 }
780
TryInlineMonomorphicCall(HInvoke * invoke_instruction,const StackHandleScope<InlineCache::kIndividualCacheSize> & classes)781 bool HInliner::TryInlineMonomorphicCall(
782 HInvoke* invoke_instruction,
783 const StackHandleScope<InlineCache::kIndividualCacheSize>& classes) {
784 DCHECK(invoke_instruction->IsInvokeVirtual() || invoke_instruction->IsInvokeInterface())
785 << invoke_instruction->DebugName();
786
787 dex::TypeIndex class_index = FindClassIndexIn(
788 GetMonomorphicType(classes), caller_compilation_unit_);
789 if (!class_index.IsValid()) {
790 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedDexCacheInaccessibleToCaller)
791 << "Call to " << ArtMethod::PrettyMethod(invoke_instruction->GetResolvedMethod())
792 << " from inline cache is not inlined because its class is not"
793 << " accessible to the caller";
794 return false;
795 }
796
797 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
798 PointerSize pointer_size = class_linker->GetImagePointerSize();
799 Handle<mirror::Class> monomorphic_type =
800 graph_->GetHandleCache()->NewHandle(GetMonomorphicType(classes));
801 ArtMethod* resolved_method = ResolveMethodFromInlineCache(
802 monomorphic_type, invoke_instruction, pointer_size);
803 if (resolved_method == nullptr) {
804 // Bogus AOT profile, bail.
805 DCHECK(Runtime::Current()->IsAotCompiler());
806 return false;
807 }
808
809 LOG_NOTE() << "Try inline monomorphic call to " << resolved_method->PrettyMethod();
810 HInstruction* receiver = invoke_instruction->InputAt(0);
811 HInstruction* cursor = invoke_instruction->GetPrevious();
812 HBasicBlock* bb_cursor = invoke_instruction->GetBlock();
813 if (!TryInlineAndReplace(invoke_instruction,
814 resolved_method,
815 ReferenceTypeInfo::Create(monomorphic_type, /* is_exact= */ true),
816 /* do_rtp= */ false,
817 /* is_speculative= */ true)) {
818 return false;
819 }
820
821 // We successfully inlined, now add a guard.
822 AddTypeGuard(receiver,
823 cursor,
824 bb_cursor,
825 class_index,
826 monomorphic_type,
827 invoke_instruction,
828 /* with_deoptimization= */ true);
829
830 // Run type propagation to get the guard typed, and eventually propagate the
831 // type of the receiver.
832 ReferenceTypePropagation rtp_fixup(graph_,
833 outer_compilation_unit_.GetDexCache(),
834 /* is_first_run= */ false);
835 rtp_fixup.Run();
836
837 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedMonomorphicCall);
838 return true;
839 }
840
AddCHAGuard(HInstruction * invoke_instruction,uint32_t dex_pc,HInstruction * cursor,HBasicBlock * bb_cursor)841 void HInliner::AddCHAGuard(HInstruction* invoke_instruction,
842 uint32_t dex_pc,
843 HInstruction* cursor,
844 HBasicBlock* bb_cursor) {
845 HShouldDeoptimizeFlag* deopt_flag = new (graph_->GetAllocator())
846 HShouldDeoptimizeFlag(graph_->GetAllocator(), dex_pc);
847 // ShouldDeoptimizeFlag is used to perform a deoptimization because of a CHA
848 // invalidation or for debugging reasons. It is OK to just check for non-zero
849 // value here instead of the specific CHA value. When a debugging deopt is
850 // requested we deoptimize before we execute any code and hence we shouldn't
851 // see that case here.
852 HInstruction* compare = new (graph_->GetAllocator()) HNotEqual(
853 deopt_flag, graph_->GetIntConstant(0, dex_pc));
854 HInstruction* deopt = new (graph_->GetAllocator()) HDeoptimize(
855 graph_->GetAllocator(), compare, DeoptimizationKind::kCHA, dex_pc);
856
857 if (cursor != nullptr) {
858 bb_cursor->InsertInstructionAfter(deopt_flag, cursor);
859 } else {
860 bb_cursor->InsertInstructionBefore(deopt_flag, bb_cursor->GetFirstInstruction());
861 }
862 bb_cursor->InsertInstructionAfter(compare, deopt_flag);
863 bb_cursor->InsertInstructionAfter(deopt, compare);
864
865 // Add receiver as input to aid CHA guard optimization later.
866 deopt_flag->AddInput(invoke_instruction->InputAt(0));
867 DCHECK_EQ(deopt_flag->InputCount(), 1u);
868 deopt->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
869 outermost_graph_->IncrementNumberOfCHAGuards();
870 }
871
AddTypeGuard(HInstruction * receiver,HInstruction * cursor,HBasicBlock * bb_cursor,dex::TypeIndex class_index,Handle<mirror::Class> klass,HInstruction * invoke_instruction,bool with_deoptimization)872 HInstruction* HInliner::AddTypeGuard(HInstruction* receiver,
873 HInstruction* cursor,
874 HBasicBlock* bb_cursor,
875 dex::TypeIndex class_index,
876 Handle<mirror::Class> klass,
877 HInstruction* invoke_instruction,
878 bool with_deoptimization) {
879 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
880 HInstanceFieldGet* receiver_class = BuildGetReceiverClass(
881 class_linker, receiver, invoke_instruction->GetDexPc());
882 if (cursor != nullptr) {
883 bb_cursor->InsertInstructionAfter(receiver_class, cursor);
884 } else {
885 bb_cursor->InsertInstructionBefore(receiver_class, bb_cursor->GetFirstInstruction());
886 }
887
888 const DexFile& caller_dex_file = *caller_compilation_unit_.GetDexFile();
889 bool is_referrer;
890 ArtMethod* outermost_art_method = outermost_graph_->GetArtMethod();
891 if (outermost_art_method == nullptr) {
892 DCHECK(Runtime::Current()->IsAotCompiler());
893 // We are in AOT mode and we don't have an ART method to determine
894 // if the inlined method belongs to the referrer. Assume it doesn't.
895 is_referrer = false;
896 } else {
897 is_referrer = klass.Get() == outermost_art_method->GetDeclaringClass();
898 }
899
900 // Note that we will just compare the classes, so we don't need Java semantics access checks.
901 // Note that the type index and the dex file are relative to the method this type guard is
902 // inlined into.
903 HLoadClass* load_class = new (graph_->GetAllocator()) HLoadClass(graph_->GetCurrentMethod(),
904 class_index,
905 caller_dex_file,
906 klass,
907 is_referrer,
908 invoke_instruction->GetDexPc(),
909 /* needs_access_check= */ false);
910 HLoadClass::LoadKind kind = HSharpening::ComputeLoadClassKind(
911 load_class, codegen_, caller_compilation_unit_);
912 DCHECK(kind != HLoadClass::LoadKind::kInvalid)
913 << "We should always be able to reference a class for inline caches";
914 // Load kind must be set before inserting the instruction into the graph.
915 load_class->SetLoadKind(kind);
916 bb_cursor->InsertInstructionAfter(load_class, receiver_class);
917 // In AOT mode, we will most likely load the class from BSS, which will involve a call
918 // to the runtime. In this case, the load instruction will need an environment so copy
919 // it from the invoke instruction.
920 if (load_class->NeedsEnvironment()) {
921 DCHECK(Runtime::Current()->IsAotCompiler());
922 load_class->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
923 }
924
925 HNotEqual* compare = new (graph_->GetAllocator()) HNotEqual(load_class, receiver_class);
926 bb_cursor->InsertInstructionAfter(compare, load_class);
927 if (with_deoptimization) {
928 HDeoptimize* deoptimize = new (graph_->GetAllocator()) HDeoptimize(
929 graph_->GetAllocator(),
930 compare,
931 receiver,
932 Runtime::Current()->IsAotCompiler()
933 ? DeoptimizationKind::kAotInlineCache
934 : DeoptimizationKind::kJitInlineCache,
935 invoke_instruction->GetDexPc());
936 bb_cursor->InsertInstructionAfter(deoptimize, compare);
937 deoptimize->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
938 DCHECK_EQ(invoke_instruction->InputAt(0), receiver);
939 receiver->ReplaceUsesDominatedBy(deoptimize, deoptimize);
940 deoptimize->SetReferenceTypeInfo(receiver->GetReferenceTypeInfo());
941 }
942 return compare;
943 }
944
MaybeReplaceAndRemove(HInstruction * new_instruction,HInstruction * old_instruction)945 static void MaybeReplaceAndRemove(HInstruction* new_instruction, HInstruction* old_instruction) {
946 DCHECK(new_instruction != old_instruction);
947 if (new_instruction != nullptr) {
948 old_instruction->ReplaceWith(new_instruction);
949 }
950 old_instruction->GetBlock()->RemoveInstruction(old_instruction);
951 }
952
TryInlinePolymorphicCall(HInvoke * invoke_instruction,const StackHandleScope<InlineCache::kIndividualCacheSize> & classes)953 bool HInliner::TryInlinePolymorphicCall(
954 HInvoke* invoke_instruction,
955 const StackHandleScope<InlineCache::kIndividualCacheSize>& classes) {
956 DCHECK(invoke_instruction->IsInvokeVirtual() || invoke_instruction->IsInvokeInterface())
957 << invoke_instruction->DebugName();
958
959 if (TryInlinePolymorphicCallToSameTarget(invoke_instruction, classes)) {
960 return true;
961 }
962
963 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
964 PointerSize pointer_size = class_linker->GetImagePointerSize();
965
966 bool all_targets_inlined = true;
967 bool one_target_inlined = false;
968 DCHECK_EQ(classes.NumberOfReferences(), InlineCache::kIndividualCacheSize);
969 uint8_t number_of_types = InlineCache::kIndividualCacheSize - classes.RemainingSlots();
970 for (size_t i = 0; i != number_of_types; ++i) {
971 DCHECK(classes.GetReference(i) != nullptr);
972 Handle<mirror::Class> handle =
973 graph_->GetHandleCache()->NewHandle(classes.GetReference(i)->AsClass());
974 ArtMethod* method = ResolveMethodFromInlineCache(handle, invoke_instruction, pointer_size);
975 if (method == nullptr) {
976 DCHECK(Runtime::Current()->IsAotCompiler());
977 // AOT profile is bogus. This loop expects to iterate over all entries,
978 // so just just continue.
979 all_targets_inlined = false;
980 continue;
981 }
982
983 HInstruction* receiver = invoke_instruction->InputAt(0);
984 HInstruction* cursor = invoke_instruction->GetPrevious();
985 HBasicBlock* bb_cursor = invoke_instruction->GetBlock();
986
987 dex::TypeIndex class_index = FindClassIndexIn(handle.Get(), caller_compilation_unit_);
988 HInstruction* return_replacement = nullptr;
989
990 // In monomorphic cases when UseOnlyPolymorphicInliningWithNoDeopt() is true, we call
991 // `TryInlinePolymorphicCall` even though we are monomorphic.
992 const bool actually_monomorphic = number_of_types == 1;
993 DCHECK_IMPLIES(actually_monomorphic, UseOnlyPolymorphicInliningWithNoDeopt());
994
995 // We only want to limit recursive polymorphic cases, not monomorphic ones.
996 const bool too_many_polymorphic_recursive_calls =
997 !actually_monomorphic &&
998 CountRecursiveCallsOf(method) > kMaximumNumberOfPolymorphicRecursiveCalls;
999 if (too_many_polymorphic_recursive_calls) {
1000 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedPolymorphicRecursiveBudget)
1001 << "Method " << method->PrettyMethod()
1002 << " is not inlined because it has reached its polymorphic recursive call budget.";
1003 } else if (class_index.IsValid()) {
1004 LOG_NOTE() << "Try inline polymorphic call to " << method->PrettyMethod();
1005 }
1006
1007 if (too_many_polymorphic_recursive_calls ||
1008 !class_index.IsValid() ||
1009 !TryBuildAndInline(invoke_instruction,
1010 method,
1011 ReferenceTypeInfo::Create(handle, /* is_exact= */ true),
1012 &return_replacement,
1013 /* is_speculative= */ true)) {
1014 all_targets_inlined = false;
1015 } else {
1016 one_target_inlined = true;
1017
1018 LOG_SUCCESS() << "Polymorphic call to "
1019 << invoke_instruction->GetMethodReference().PrettyMethod()
1020 << " has inlined " << ArtMethod::PrettyMethod(method);
1021
1022 // If we have inlined all targets before, and this receiver is the last seen,
1023 // we deoptimize instead of keeping the original invoke instruction.
1024 bool deoptimize = !UseOnlyPolymorphicInliningWithNoDeopt() &&
1025 all_targets_inlined &&
1026 (i + 1 == number_of_types);
1027
1028 HInstruction* compare = AddTypeGuard(receiver,
1029 cursor,
1030 bb_cursor,
1031 class_index,
1032 handle,
1033 invoke_instruction,
1034 deoptimize);
1035 if (deoptimize) {
1036 MaybeReplaceAndRemove(return_replacement, invoke_instruction);
1037 } else {
1038 CreateDiamondPatternForPolymorphicInline(compare, return_replacement, invoke_instruction);
1039 }
1040 }
1041 }
1042
1043 if (!one_target_inlined) {
1044 LOG_FAIL_NO_STAT()
1045 << "Call to " << invoke_instruction->GetMethodReference().PrettyMethod()
1046 << " from inline cache is not inlined because none"
1047 << " of its targets could be inlined";
1048 return false;
1049 }
1050
1051 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedPolymorphicCall);
1052
1053 // Run type propagation to get the guards typed.
1054 ReferenceTypePropagation rtp_fixup(graph_,
1055 outer_compilation_unit_.GetDexCache(),
1056 /* is_first_run= */ false);
1057 rtp_fixup.Run();
1058 return true;
1059 }
1060
CreateDiamondPatternForPolymorphicInline(HInstruction * compare,HInstruction * return_replacement,HInstruction * invoke_instruction)1061 void HInliner::CreateDiamondPatternForPolymorphicInline(HInstruction* compare,
1062 HInstruction* return_replacement,
1063 HInstruction* invoke_instruction) {
1064 uint32_t dex_pc = invoke_instruction->GetDexPc();
1065 HBasicBlock* cursor_block = compare->GetBlock();
1066 HBasicBlock* original_invoke_block = invoke_instruction->GetBlock();
1067 ArenaAllocator* allocator = graph_->GetAllocator();
1068
1069 // Spit the block after the compare: `cursor_block` will now be the start of the diamond,
1070 // and the returned block is the start of the then branch (that could contain multiple blocks).
1071 HBasicBlock* then = cursor_block->SplitAfterForInlining(compare);
1072
1073 // Split the block containing the invoke before and after the invoke. The returned block
1074 // of the split before will contain the invoke and will be the otherwise branch of
1075 // the diamond. The returned block of the split after will be the merge block
1076 // of the diamond.
1077 HBasicBlock* end_then = invoke_instruction->GetBlock();
1078 HBasicBlock* otherwise = end_then->SplitBeforeForInlining(invoke_instruction);
1079 HBasicBlock* merge = otherwise->SplitAfterForInlining(invoke_instruction);
1080
1081 // If the methods we are inlining return a value, we create a phi in the merge block
1082 // that will have the `invoke_instruction and the `return_replacement` as inputs.
1083 if (return_replacement != nullptr) {
1084 HPhi* phi = new (allocator) HPhi(
1085 allocator, kNoRegNumber, 0, HPhi::ToPhiType(invoke_instruction->GetType()), dex_pc);
1086 merge->AddPhi(phi);
1087 invoke_instruction->ReplaceWith(phi);
1088 phi->AddInput(return_replacement);
1089 phi->AddInput(invoke_instruction);
1090 }
1091
1092 // Add the control flow instructions.
1093 otherwise->AddInstruction(new (allocator) HGoto(dex_pc));
1094 end_then->AddInstruction(new (allocator) HGoto(dex_pc));
1095 cursor_block->AddInstruction(new (allocator) HIf(compare, dex_pc));
1096
1097 // Add the newly created blocks to the graph.
1098 graph_->AddBlock(then);
1099 graph_->AddBlock(otherwise);
1100 graph_->AddBlock(merge);
1101
1102 // Set up successor (and implictly predecessor) relations.
1103 cursor_block->AddSuccessor(otherwise);
1104 cursor_block->AddSuccessor(then);
1105 end_then->AddSuccessor(merge);
1106 otherwise->AddSuccessor(merge);
1107
1108 // Set up dominance information.
1109 then->SetDominator(cursor_block);
1110 cursor_block->AddDominatedBlock(then);
1111 otherwise->SetDominator(cursor_block);
1112 cursor_block->AddDominatedBlock(otherwise);
1113 merge->SetDominator(cursor_block);
1114 cursor_block->AddDominatedBlock(merge);
1115
1116 // Update the revert post order.
1117 size_t index = IndexOfElement(graph_->reverse_post_order_, cursor_block);
1118 MakeRoomFor(&graph_->reverse_post_order_, 1, index);
1119 graph_->reverse_post_order_[++index] = then;
1120 index = IndexOfElement(graph_->reverse_post_order_, end_then);
1121 MakeRoomFor(&graph_->reverse_post_order_, 2, index);
1122 graph_->reverse_post_order_[++index] = otherwise;
1123 graph_->reverse_post_order_[++index] = merge;
1124
1125
1126 graph_->UpdateLoopAndTryInformationOfNewBlock(
1127 then, original_invoke_block, /* replace_if_back_edge= */ false);
1128 graph_->UpdateLoopAndTryInformationOfNewBlock(
1129 otherwise, original_invoke_block, /* replace_if_back_edge= */ false);
1130
1131 // In case the original invoke location was a back edge, we need to update
1132 // the loop to now have the merge block as a back edge.
1133 graph_->UpdateLoopAndTryInformationOfNewBlock(
1134 merge, original_invoke_block, /* replace_if_back_edge= */ true);
1135 }
1136
TryInlinePolymorphicCallToSameTarget(HInvoke * invoke_instruction,const StackHandleScope<InlineCache::kIndividualCacheSize> & classes)1137 bool HInliner::TryInlinePolymorphicCallToSameTarget(
1138 HInvoke* invoke_instruction,
1139 const StackHandleScope<InlineCache::kIndividualCacheSize>& classes) {
1140 // This optimization only works under JIT for now.
1141 if (!codegen_->GetCompilerOptions().IsJitCompiler()) {
1142 return false;
1143 }
1144
1145 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
1146 PointerSize pointer_size = class_linker->GetImagePointerSize();
1147
1148 ArtMethod* actual_method = nullptr;
1149 size_t method_index = invoke_instruction->IsInvokeVirtual()
1150 ? invoke_instruction->AsInvokeVirtual()->GetVTableIndex()
1151 : invoke_instruction->AsInvokeInterface()->GetImtIndex();
1152
1153 // Check whether we are actually calling the same method among
1154 // the different types seen.
1155 DCHECK_EQ(classes.NumberOfReferences(), InlineCache::kIndividualCacheSize);
1156 uint8_t number_of_types = InlineCache::kIndividualCacheSize - classes.RemainingSlots();
1157 for (size_t i = 0; i != number_of_types; ++i) {
1158 DCHECK(classes.GetReference(i) != nullptr);
1159 ArtMethod* new_method = nullptr;
1160 if (invoke_instruction->IsInvokeInterface()) {
1161 new_method = classes.GetReference(i)->AsClass()->GetImt(pointer_size)->Get(
1162 method_index, pointer_size);
1163 if (new_method->IsRuntimeMethod()) {
1164 // Bail out as soon as we see a conflict trampoline in one of the target's
1165 // interface table.
1166 return false;
1167 }
1168 } else {
1169 DCHECK(invoke_instruction->IsInvokeVirtual());
1170 new_method =
1171 classes.GetReference(i)->AsClass()->GetEmbeddedVTableEntry(method_index, pointer_size);
1172 }
1173 DCHECK(new_method != nullptr);
1174 if (actual_method == nullptr) {
1175 actual_method = new_method;
1176 } else if (actual_method != new_method) {
1177 // Different methods, bailout.
1178 return false;
1179 }
1180 }
1181
1182 HInstruction* receiver = invoke_instruction->InputAt(0);
1183 HInstruction* cursor = invoke_instruction->GetPrevious();
1184 HBasicBlock* bb_cursor = invoke_instruction->GetBlock();
1185
1186 HInstruction* return_replacement = nullptr;
1187 if (!TryBuildAndInline(invoke_instruction,
1188 actual_method,
1189 ReferenceTypeInfo::CreateInvalid(),
1190 &return_replacement,
1191 /* is_speculative= */ true)) {
1192 return false;
1193 }
1194
1195 // We successfully inlined, now add a guard.
1196 HInstanceFieldGet* receiver_class = BuildGetReceiverClass(
1197 class_linker, receiver, invoke_instruction->GetDexPc());
1198
1199 DataType::Type type = Is64BitInstructionSet(graph_->GetInstructionSet())
1200 ? DataType::Type::kInt64
1201 : DataType::Type::kInt32;
1202 HClassTableGet* class_table_get = new (graph_->GetAllocator()) HClassTableGet(
1203 receiver_class,
1204 type,
1205 invoke_instruction->IsInvokeVirtual() ? HClassTableGet::TableKind::kVTable
1206 : HClassTableGet::TableKind::kIMTable,
1207 method_index,
1208 invoke_instruction->GetDexPc());
1209
1210 HConstant* constant;
1211 if (type == DataType::Type::kInt64) {
1212 constant = graph_->GetLongConstant(
1213 reinterpret_cast<intptr_t>(actual_method), invoke_instruction->GetDexPc());
1214 } else {
1215 constant = graph_->GetIntConstant(
1216 reinterpret_cast<intptr_t>(actual_method), invoke_instruction->GetDexPc());
1217 }
1218
1219 HNotEqual* compare = new (graph_->GetAllocator()) HNotEqual(class_table_get, constant);
1220 if (cursor != nullptr) {
1221 bb_cursor->InsertInstructionAfter(receiver_class, cursor);
1222 } else {
1223 bb_cursor->InsertInstructionBefore(receiver_class, bb_cursor->GetFirstInstruction());
1224 }
1225 bb_cursor->InsertInstructionAfter(class_table_get, receiver_class);
1226 bb_cursor->InsertInstructionAfter(compare, class_table_get);
1227
1228 if (outermost_graph_->IsCompilingOsr()) {
1229 CreateDiamondPatternForPolymorphicInline(compare, return_replacement, invoke_instruction);
1230 } else {
1231 HDeoptimize* deoptimize = new (graph_->GetAllocator()) HDeoptimize(
1232 graph_->GetAllocator(),
1233 compare,
1234 receiver,
1235 DeoptimizationKind::kJitSameTarget,
1236 invoke_instruction->GetDexPc());
1237 bb_cursor->InsertInstructionAfter(deoptimize, compare);
1238 deoptimize->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
1239 MaybeReplaceAndRemove(return_replacement, invoke_instruction);
1240 receiver->ReplaceUsesDominatedBy(deoptimize, deoptimize);
1241 deoptimize->SetReferenceTypeInfo(receiver->GetReferenceTypeInfo());
1242 }
1243
1244 // Run type propagation to get the guard typed.
1245 ReferenceTypePropagation rtp_fixup(graph_,
1246 outer_compilation_unit_.GetDexCache(),
1247 /* is_first_run= */ false);
1248 rtp_fixup.Run();
1249
1250 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedPolymorphicCall);
1251
1252 LOG_SUCCESS() << "Inlined same polymorphic target " << actual_method->PrettyMethod();
1253 return true;
1254 }
1255
MaybeRunReferenceTypePropagation(HInstruction * replacement,HInvoke * invoke_instruction)1256 void HInliner::MaybeRunReferenceTypePropagation(HInstruction* replacement,
1257 HInvoke* invoke_instruction) {
1258 if (ReturnTypeMoreSpecific(replacement, invoke_instruction)) {
1259 // Actual return value has a more specific type than the method's declared
1260 // return type. Run RTP again on the outer graph to propagate it.
1261 ReferenceTypePropagation(graph_,
1262 outer_compilation_unit_.GetDexCache(),
1263 /* is_first_run= */ false).Run();
1264 }
1265 }
1266
TryDevirtualize(HInvoke * invoke_instruction,ArtMethod * method,HInvoke ** replacement)1267 bool HInliner::TryDevirtualize(HInvoke* invoke_instruction,
1268 ArtMethod* method,
1269 HInvoke** replacement) {
1270 DCHECK(invoke_instruction != *replacement);
1271 if (!invoke_instruction->IsInvokeInterface() && !invoke_instruction->IsInvokeVirtual()) {
1272 return false;
1273 }
1274
1275 // Don't try to devirtualize intrinsics as it breaks pattern matching from later phases.
1276 // TODO(solanes): This `if` could be removed if we update optimizations like
1277 // TryReplaceStringBuilderAppend.
1278 if (invoke_instruction->IsIntrinsic()) {
1279 return false;
1280 }
1281
1282 // Don't bother trying to call directly a default conflict method. It
1283 // doesn't have a proper MethodReference, but also `GetCanonicalMethod`
1284 // will return an actual default implementation.
1285 if (method->IsDefaultConflicting()) {
1286 return false;
1287 }
1288 DCHECK(!method->IsProxyMethod());
1289 ClassLinker* cl = Runtime::Current()->GetClassLinker();
1290 PointerSize pointer_size = cl->GetImagePointerSize();
1291 // The sharpening logic assumes the caller isn't passing a copied method.
1292 method = method->GetCanonicalMethod(pointer_size);
1293 uint32_t dex_method_index = FindMethodIndexIn(
1294 method,
1295 *invoke_instruction->GetMethodReference().dex_file,
1296 invoke_instruction->GetMethodReference().index);
1297 if (dex_method_index == dex::kDexNoIndex) {
1298 return false;
1299 }
1300 HInvokeStaticOrDirect::DispatchInfo dispatch_info =
1301 HSharpening::SharpenLoadMethod(method,
1302 /* has_method_id= */ true,
1303 /* for_interface_call= */ false,
1304 codegen_);
1305 DCHECK_NE(dispatch_info.code_ptr_location, CodePtrLocation::kCallCriticalNative);
1306 if (dispatch_info.method_load_kind == MethodLoadKind::kRuntimeCall) {
1307 // If sharpening returns that we need to load the method at runtime, keep
1308 // the virtual/interface call which will be faster.
1309 // Also, the entrypoints for runtime calls do not handle devirtualized
1310 // calls.
1311 return false;
1312 }
1313
1314 HInvokeStaticOrDirect* new_invoke = new (graph_->GetAllocator()) HInvokeStaticOrDirect(
1315 graph_->GetAllocator(),
1316 invoke_instruction->GetNumberOfArguments(),
1317 invoke_instruction->GetType(),
1318 invoke_instruction->GetDexPc(),
1319 MethodReference(invoke_instruction->GetMethodReference().dex_file, dex_method_index),
1320 method,
1321 dispatch_info,
1322 kDirect,
1323 MethodReference(method->GetDexFile(), method->GetDexMethodIndex()),
1324 HInvokeStaticOrDirect::ClinitCheckRequirement::kNone,
1325 !graph_->IsDebuggable());
1326 HInputsRef inputs = invoke_instruction->GetInputs();
1327 DCHECK_EQ(inputs.size(), invoke_instruction->GetNumberOfArguments());
1328 for (size_t index = 0; index != inputs.size(); ++index) {
1329 new_invoke->SetArgumentAt(index, inputs[index]);
1330 }
1331 if (HInvokeStaticOrDirect::NeedsCurrentMethodInput(dispatch_info)) {
1332 new_invoke->SetRawInputAt(new_invoke->GetCurrentMethodIndexUnchecked(),
1333 graph_->GetCurrentMethod());
1334 }
1335 invoke_instruction->GetBlock()->InsertInstructionBefore(new_invoke, invoke_instruction);
1336 new_invoke->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
1337 if (invoke_instruction->GetType() == DataType::Type::kReference) {
1338 new_invoke->SetReferenceTypeInfoIfValid(invoke_instruction->GetReferenceTypeInfo());
1339 }
1340 *replacement = new_invoke;
1341
1342 MaybeReplaceAndRemove(*replacement, invoke_instruction);
1343 // No need to call MaybeRunReferenceTypePropagation, as we know the return type
1344 // cannot be more specific.
1345 DCHECK(!ReturnTypeMoreSpecific(*replacement, invoke_instruction));
1346 return true;
1347 }
1348
1349
TryInlineAndReplace(HInvoke * invoke_instruction,ArtMethod * method,ReferenceTypeInfo receiver_type,bool do_rtp,bool is_speculative)1350 bool HInliner::TryInlineAndReplace(HInvoke* invoke_instruction,
1351 ArtMethod* method,
1352 ReferenceTypeInfo receiver_type,
1353 bool do_rtp,
1354 bool is_speculative) {
1355 DCHECK(!codegen_->IsImplementedIntrinsic(invoke_instruction));
1356 HInstruction* return_replacement = nullptr;
1357
1358 if (!TryBuildAndInline(
1359 invoke_instruction, method, receiver_type, &return_replacement, is_speculative)) {
1360 return false;
1361 }
1362
1363 MaybeReplaceAndRemove(return_replacement, invoke_instruction);
1364 FixUpReturnReferenceType(method, return_replacement);
1365 if (do_rtp) {
1366 MaybeRunReferenceTypePropagation(return_replacement, invoke_instruction);
1367 }
1368 return true;
1369 }
1370
CountRecursiveCallsOf(ArtMethod * method) const1371 size_t HInliner::CountRecursiveCallsOf(ArtMethod* method) const {
1372 const HInliner* current = this;
1373 size_t count = 0;
1374 do {
1375 if (current->graph_->GetArtMethod() == method) {
1376 ++count;
1377 }
1378 current = current->parent_;
1379 } while (current != nullptr);
1380 return count;
1381 }
1382
MayInline(const CompilerOptions & compiler_options,const DexFile & inlined_from,const DexFile & inlined_into)1383 static inline bool MayInline(const CompilerOptions& compiler_options,
1384 const DexFile& inlined_from,
1385 const DexFile& inlined_into) {
1386 // We're not allowed to inline across dex files if we're the no-inline-from dex file.
1387 if (!IsSameDexFile(inlined_from, inlined_into) &&
1388 ContainsElement(compiler_options.GetNoInlineFromDexFile(), &inlined_from)) {
1389 return false;
1390 }
1391
1392 return true;
1393 }
1394
1395 // Returns whether inlining is allowed based on ART semantics.
IsInliningAllowed(ArtMethod * method,const CodeItemDataAccessor & accessor) const1396 bool HInliner::IsInliningAllowed(ArtMethod* method, const CodeItemDataAccessor& accessor) const {
1397 if (!accessor.HasCodeItem()) {
1398 LOG_FAIL_NO_STAT()
1399 << "Method " << method->PrettyMethod() << " is not inlined because it is native";
1400 return false;
1401 }
1402
1403 if (!method->IsCompilable()) {
1404 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedNotCompilable)
1405 << "Method " << method->PrettyMethod()
1406 << " has soft failures un-handled by the compiler, so it cannot be inlined";
1407 return false;
1408 }
1409
1410 if (!IsMethodVerified(method)) {
1411 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedNotVerified)
1412 << "Method " << method->PrettyMethod()
1413 << " couldn't be verified, so it cannot be inlined";
1414 return false;
1415 }
1416
1417 if (annotations::MethodIsNeverInline(*method->GetDexFile(),
1418 method->GetClassDef(),
1419 method->GetDexMethodIndex())) {
1420 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedNeverInlineAnnotation)
1421 << "Method " << method->PrettyMethod()
1422 << " has the @NeverInline annotation so it won't be inlined";
1423 return false;
1424 }
1425
1426 return true;
1427 }
1428
1429 // Returns whether ART supports inlining this method.
1430 //
1431 // Some methods are not supported because they have features for which inlining
1432 // is not implemented. For example, we do not currently support inlining throw
1433 // instructions into a try block.
IsInliningSupported(const HInvoke * invoke_instruction,ArtMethod * method,const CodeItemDataAccessor & accessor) const1434 bool HInliner::IsInliningSupported(const HInvoke* invoke_instruction,
1435 ArtMethod* method,
1436 const CodeItemDataAccessor& accessor) const {
1437 if (method->IsProxyMethod()) {
1438 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedProxy)
1439 << "Method " << method->PrettyMethod()
1440 << " is not inlined because of unimplemented inline support for proxy methods.";
1441 return false;
1442 }
1443
1444 if (accessor.TriesSize() != 0) {
1445 if (!kInlineTryCatches) {
1446 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedTryCatchDisabled)
1447 << "Method " << method->PrettyMethod()
1448 << " is not inlined because inlining try catches is disabled globally";
1449 return false;
1450 }
1451 const bool disallowed_try_catch_inlining =
1452 // Direct parent is a try block.
1453 invoke_instruction->GetBlock()->IsTryBlock() ||
1454 // Indirect parent disallows try catch inlining.
1455 !try_catch_inlining_allowed_;
1456 if (disallowed_try_catch_inlining) {
1457 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedTryCatchCallee)
1458 << "Method " << method->PrettyMethod()
1459 << " is not inlined because it has a try catch and we are not supporting it for this"
1460 << " particular call. This is could be because e.g. it would be inlined inside another"
1461 << " try block, we arrived here from TryInlinePolymorphicCall, etc.";
1462 return false;
1463 }
1464 }
1465
1466 if (invoke_instruction->IsInvokeStaticOrDirect() &&
1467 invoke_instruction->AsInvokeStaticOrDirect()->IsStaticWithImplicitClinitCheck()) {
1468 // Case of a static method that cannot be inlined because it implicitly
1469 // requires an initialization check of its declaring class.
1470 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedDexCacheClinitCheck)
1471 << "Method " << method->PrettyMethod()
1472 << " is not inlined because it is static and requires a clinit"
1473 << " check that cannot be emitted due to Dex cache limitations";
1474 return false;
1475 }
1476
1477 return true;
1478 }
1479
IsInliningEncouraged(const HInvoke * invoke_instruction,ArtMethod * method,const CodeItemDataAccessor & accessor) const1480 bool HInliner::IsInliningEncouraged(const HInvoke* invoke_instruction,
1481 ArtMethod* method,
1482 const CodeItemDataAccessor& accessor) const {
1483 if (CountRecursiveCallsOf(method) > kMaximumNumberOfRecursiveCalls) {
1484 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedRecursiveBudget)
1485 << "Method "
1486 << method->PrettyMethod()
1487 << " is not inlined because it has reached its recursive call budget.";
1488 return false;
1489 }
1490
1491 size_t inline_max_code_units = codegen_->GetCompilerOptions().GetInlineMaxCodeUnits();
1492 if (accessor.InsnsSizeInCodeUnits() > inline_max_code_units) {
1493 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedCodeItem)
1494 << "Method " << method->PrettyMethod()
1495 << " is not inlined because its code item is too big: "
1496 << accessor.InsnsSizeInCodeUnits()
1497 << " > "
1498 << inline_max_code_units;
1499 return false;
1500 }
1501
1502 if (invoke_instruction->GetBlock()->GetLastInstruction()->IsThrow()) {
1503 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedEndsWithThrow)
1504 << "Method " << method->PrettyMethod()
1505 << " is not inlined because its block ends with a throw";
1506 return false;
1507 }
1508
1509 return true;
1510 }
1511
TryBuildAndInline(HInvoke * invoke_instruction,ArtMethod * method,ReferenceTypeInfo receiver_type,HInstruction ** return_replacement,bool is_speculative)1512 bool HInliner::TryBuildAndInline(HInvoke* invoke_instruction,
1513 ArtMethod* method,
1514 ReferenceTypeInfo receiver_type,
1515 HInstruction** return_replacement,
1516 bool is_speculative) {
1517 // If invoke_instruction is devirtualized to a different method, give intrinsics
1518 // another chance before we try to inline it.
1519 if (invoke_instruction->GetResolvedMethod() != method && method->IsIntrinsic()) {
1520 MaybeRecordStat(stats_, MethodCompilationStat::kIntrinsicRecognized);
1521 // For simplicity, always create a new instruction to replace the existing
1522 // invoke.
1523 HInvokeVirtual* new_invoke = new (graph_->GetAllocator()) HInvokeVirtual(
1524 graph_->GetAllocator(),
1525 invoke_instruction->GetNumberOfArguments(),
1526 invoke_instruction->GetType(),
1527 invoke_instruction->GetDexPc(),
1528 invoke_instruction->GetMethodReference(), // Use existing invoke's method's reference.
1529 method,
1530 MethodReference(method->GetDexFile(), method->GetDexMethodIndex()),
1531 method->GetMethodIndex(),
1532 !graph_->IsDebuggable());
1533 DCHECK_NE(new_invoke->GetIntrinsic(), Intrinsics::kNone);
1534 HInputsRef inputs = invoke_instruction->GetInputs();
1535 for (size_t index = 0; index != inputs.size(); ++index) {
1536 new_invoke->SetArgumentAt(index, inputs[index]);
1537 }
1538 invoke_instruction->GetBlock()->InsertInstructionBefore(new_invoke, invoke_instruction);
1539 new_invoke->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
1540 if (invoke_instruction->GetType() == DataType::Type::kReference) {
1541 new_invoke->SetReferenceTypeInfoIfValid(invoke_instruction->GetReferenceTypeInfo());
1542 }
1543 *return_replacement = new_invoke;
1544 return true;
1545 }
1546
1547 // Check whether we're allowed to inline. The outermost compilation unit is the relevant
1548 // dex file here (though the transitivity of an inline chain would allow checking the caller).
1549 if (!MayInline(codegen_->GetCompilerOptions(),
1550 *method->GetDexFile(),
1551 *outer_compilation_unit_.GetDexFile())) {
1552 if (TryPatternSubstitution(invoke_instruction, method, return_replacement)) {
1553 LOG_SUCCESS() << "Successfully replaced pattern of invoke "
1554 << method->PrettyMethod();
1555 MaybeRecordStat(stats_, MethodCompilationStat::kReplacedInvokeWithSimplePattern);
1556 return true;
1557 }
1558 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedWont)
1559 << "Won't inline " << method->PrettyMethod() << " in "
1560 << outer_compilation_unit_.GetDexFile()->GetLocation() << " ("
1561 << caller_compilation_unit_.GetDexFile()->GetLocation() << ") from "
1562 << method->GetDexFile()->GetLocation();
1563 return false;
1564 }
1565
1566 CodeItemDataAccessor accessor(method->DexInstructionData());
1567
1568 if (!IsInliningAllowed(method, accessor)) {
1569 return false;
1570 }
1571
1572 if (!IsInliningSupported(invoke_instruction, method, accessor)) {
1573 return false;
1574 }
1575
1576 if (!IsInliningEncouraged(invoke_instruction, method, accessor)) {
1577 return false;
1578 }
1579
1580 if (!TryBuildAndInlineHelper(
1581 invoke_instruction, method, receiver_type, return_replacement, is_speculative)) {
1582 return false;
1583 }
1584
1585 LOG_SUCCESS() << method->PrettyMethod();
1586 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedInvoke);
1587 if (outermost_graph_ == graph_) {
1588 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedLastInvoke);
1589 }
1590 return true;
1591 }
1592
GetInvokeInputForArgVRegIndex(HInvoke * invoke_instruction,size_t arg_vreg_index)1593 static HInstruction* GetInvokeInputForArgVRegIndex(HInvoke* invoke_instruction,
1594 size_t arg_vreg_index)
1595 REQUIRES_SHARED(Locks::mutator_lock_) {
1596 size_t input_index = 0;
1597 for (size_t i = 0; i < arg_vreg_index; ++i, ++input_index) {
1598 DCHECK_LT(input_index, invoke_instruction->GetNumberOfArguments());
1599 if (DataType::Is64BitType(invoke_instruction->InputAt(input_index)->GetType())) {
1600 ++i;
1601 DCHECK_NE(i, arg_vreg_index);
1602 }
1603 }
1604 DCHECK_LT(input_index, invoke_instruction->GetNumberOfArguments());
1605 return invoke_instruction->InputAt(input_index);
1606 }
1607
1608 // Try to recognize known simple patterns and replace invoke call with appropriate instructions.
TryPatternSubstitution(HInvoke * invoke_instruction,ArtMethod * method,HInstruction ** return_replacement)1609 bool HInliner::TryPatternSubstitution(HInvoke* invoke_instruction,
1610 ArtMethod* method,
1611 HInstruction** return_replacement) {
1612 InlineMethod inline_method;
1613 if (!InlineMethodAnalyser::AnalyseMethodCode(method, &inline_method)) {
1614 return false;
1615 }
1616
1617 switch (inline_method.opcode) {
1618 case kInlineOpNop:
1619 DCHECK_EQ(invoke_instruction->GetType(), DataType::Type::kVoid);
1620 *return_replacement = nullptr;
1621 break;
1622 case kInlineOpReturnArg:
1623 *return_replacement = GetInvokeInputForArgVRegIndex(invoke_instruction,
1624 inline_method.d.return_data.arg);
1625 break;
1626 case kInlineOpNonWideConst: {
1627 char shorty0 = method->GetShorty()[0];
1628 if (shorty0 == 'L') {
1629 DCHECK_EQ(inline_method.d.data, 0u);
1630 *return_replacement = graph_->GetNullConstant();
1631 } else if (shorty0 == 'F') {
1632 *return_replacement = graph_->GetFloatConstant(
1633 bit_cast<float, int32_t>(static_cast<int32_t>(inline_method.d.data)));
1634 } else {
1635 *return_replacement = graph_->GetIntConstant(static_cast<int32_t>(inline_method.d.data));
1636 }
1637 break;
1638 }
1639 case kInlineOpIGet: {
1640 const InlineIGetIPutData& data = inline_method.d.ifield_data;
1641 if (data.method_is_static || data.object_arg != 0u) {
1642 // TODO: Needs null check.
1643 return false;
1644 }
1645 HInstruction* obj = GetInvokeInputForArgVRegIndex(invoke_instruction, data.object_arg);
1646 HInstanceFieldGet* iget = CreateInstanceFieldGet(data.field_idx, method, obj);
1647 DCHECK_EQ(iget->GetFieldOffset().Uint32Value(), data.field_offset);
1648 DCHECK_EQ(iget->IsVolatile() ? 1u : 0u, data.is_volatile);
1649 invoke_instruction->GetBlock()->InsertInstructionBefore(iget, invoke_instruction);
1650 *return_replacement = iget;
1651 break;
1652 }
1653 case kInlineOpIPut: {
1654 const InlineIGetIPutData& data = inline_method.d.ifield_data;
1655 if (data.method_is_static || data.object_arg != 0u) {
1656 // TODO: Needs null check.
1657 return false;
1658 }
1659 HInstruction* obj = GetInvokeInputForArgVRegIndex(invoke_instruction, data.object_arg);
1660 HInstruction* value = GetInvokeInputForArgVRegIndex(invoke_instruction, data.src_arg);
1661 HInstanceFieldSet* iput = CreateInstanceFieldSet(data.field_idx, method, obj, value);
1662 DCHECK_EQ(iput->GetFieldOffset().Uint32Value(), data.field_offset);
1663 DCHECK_EQ(iput->IsVolatile() ? 1u : 0u, data.is_volatile);
1664 invoke_instruction->GetBlock()->InsertInstructionBefore(iput, invoke_instruction);
1665 if (data.return_arg_plus1 != 0u) {
1666 size_t return_arg = data.return_arg_plus1 - 1u;
1667 *return_replacement = GetInvokeInputForArgVRegIndex(invoke_instruction, return_arg);
1668 }
1669 break;
1670 }
1671 case kInlineOpConstructor: {
1672 const InlineConstructorData& data = inline_method.d.constructor_data;
1673 // Get the indexes to arrays for easier processing.
1674 uint16_t iput_field_indexes[] = {
1675 data.iput0_field_index, data.iput1_field_index, data.iput2_field_index
1676 };
1677 uint16_t iput_args[] = { data.iput0_arg, data.iput1_arg, data.iput2_arg };
1678 static_assert(arraysize(iput_args) == arraysize(iput_field_indexes), "Size mismatch");
1679 // Count valid field indexes.
1680 size_t number_of_iputs = 0u;
1681 while (number_of_iputs != arraysize(iput_field_indexes) &&
1682 iput_field_indexes[number_of_iputs] != DexFile::kDexNoIndex16) {
1683 // Check that there are no duplicate valid field indexes.
1684 DCHECK_EQ(0, std::count(iput_field_indexes + number_of_iputs + 1,
1685 iput_field_indexes + arraysize(iput_field_indexes),
1686 iput_field_indexes[number_of_iputs]));
1687 ++number_of_iputs;
1688 }
1689 // Check that there are no valid field indexes in the rest of the array.
1690 DCHECK_EQ(0, std::count_if(iput_field_indexes + number_of_iputs,
1691 iput_field_indexes + arraysize(iput_field_indexes),
1692 [](uint16_t index) { return index != DexFile::kDexNoIndex16; }));
1693
1694 // Create HInstanceFieldSet for each IPUT that stores non-zero data.
1695 HInstruction* obj = GetInvokeInputForArgVRegIndex(invoke_instruction,
1696 /* arg_vreg_index= */ 0u);
1697 bool needs_constructor_barrier = false;
1698 for (size_t i = 0; i != number_of_iputs; ++i) {
1699 HInstruction* value = GetInvokeInputForArgVRegIndex(invoke_instruction, iput_args[i]);
1700 if (!IsZeroBitPattern(value)) {
1701 uint16_t field_index = iput_field_indexes[i];
1702 bool is_final;
1703 HInstanceFieldSet* iput =
1704 CreateInstanceFieldSet(field_index, method, obj, value, &is_final);
1705 invoke_instruction->GetBlock()->InsertInstructionBefore(iput, invoke_instruction);
1706
1707 // Check whether the field is final. If it is, we need to add a barrier.
1708 if (is_final) {
1709 needs_constructor_barrier = true;
1710 }
1711 }
1712 }
1713 if (needs_constructor_barrier) {
1714 // See DexCompilationUnit::RequiresConstructorBarrier for more details.
1715 DCHECK(obj != nullptr) << "only non-static methods can have a constructor fence";
1716
1717 HConstructorFence* constructor_fence =
1718 new (graph_->GetAllocator()) HConstructorFence(obj, kNoDexPc, graph_->GetAllocator());
1719 invoke_instruction->GetBlock()->InsertInstructionBefore(constructor_fence,
1720 invoke_instruction);
1721 }
1722 *return_replacement = nullptr;
1723 break;
1724 }
1725 default:
1726 LOG(FATAL) << "UNREACHABLE";
1727 UNREACHABLE();
1728 }
1729 return true;
1730 }
1731
CreateInstanceFieldGet(uint32_t field_index,ArtMethod * referrer,HInstruction * obj)1732 HInstanceFieldGet* HInliner::CreateInstanceFieldGet(uint32_t field_index,
1733 ArtMethod* referrer,
1734 HInstruction* obj)
1735 REQUIRES_SHARED(Locks::mutator_lock_) {
1736 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
1737 ArtField* resolved_field =
1738 class_linker->LookupResolvedField(field_index, referrer, /* is_static= */ false);
1739 DCHECK(resolved_field != nullptr);
1740 HInstanceFieldGet* iget = new (graph_->GetAllocator()) HInstanceFieldGet(
1741 obj,
1742 resolved_field,
1743 DataType::FromShorty(resolved_field->GetTypeDescriptor()[0]),
1744 resolved_field->GetOffset(),
1745 resolved_field->IsVolatile(),
1746 field_index,
1747 resolved_field->GetDeclaringClass()->GetDexClassDefIndex(),
1748 *referrer->GetDexFile(),
1749 // Read barrier generates a runtime call in slow path and we need a valid
1750 // dex pc for the associated stack map. 0 is bogus but valid. Bug: 26854537.
1751 /* dex_pc= */ 0);
1752 if (iget->GetType() == DataType::Type::kReference) {
1753 // Use the same dex_cache that we used for field lookup as the hint_dex_cache.
1754 Handle<mirror::DexCache> dex_cache =
1755 graph_->GetHandleCache()->NewHandle(referrer->GetDexCache());
1756 ReferenceTypePropagation rtp(graph_,
1757 dex_cache,
1758 /* is_first_run= */ false);
1759 rtp.Visit(iget);
1760 }
1761 return iget;
1762 }
1763
CreateInstanceFieldSet(uint32_t field_index,ArtMethod * referrer,HInstruction * obj,HInstruction * value,bool * is_final)1764 HInstanceFieldSet* HInliner::CreateInstanceFieldSet(uint32_t field_index,
1765 ArtMethod* referrer,
1766 HInstruction* obj,
1767 HInstruction* value,
1768 bool* is_final)
1769 REQUIRES_SHARED(Locks::mutator_lock_) {
1770 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
1771 ArtField* resolved_field =
1772 class_linker->LookupResolvedField(field_index, referrer, /* is_static= */ false);
1773 DCHECK(resolved_field != nullptr);
1774 if (is_final != nullptr) {
1775 // This information is needed only for constructors.
1776 DCHECK(referrer->IsConstructor());
1777 *is_final = resolved_field->IsFinal();
1778 }
1779 HInstanceFieldSet* iput = new (graph_->GetAllocator()) HInstanceFieldSet(
1780 obj,
1781 value,
1782 resolved_field,
1783 DataType::FromShorty(resolved_field->GetTypeDescriptor()[0]),
1784 resolved_field->GetOffset(),
1785 resolved_field->IsVolatile(),
1786 field_index,
1787 resolved_field->GetDeclaringClass()->GetDexClassDefIndex(),
1788 *referrer->GetDexFile(),
1789 // Read barrier generates a runtime call in slow path and we need a valid
1790 // dex pc for the associated stack map. 0 is bogus but valid. Bug: 26854537.
1791 /* dex_pc= */ 0);
1792 return iput;
1793 }
1794
1795 template <typename T>
NewHandleIfDifferent(ObjPtr<T> object,Handle<T> hint,HGraph * graph)1796 static inline Handle<T> NewHandleIfDifferent(ObjPtr<T> object, Handle<T> hint, HGraph* graph)
1797 REQUIRES_SHARED(Locks::mutator_lock_) {
1798 return (object != hint.Get()) ? graph->GetHandleCache()->NewHandle(object) : hint;
1799 }
1800
CanEncodeInlinedMethodInStackMap(const DexFile & outer_dex_file,ArtMethod * callee,const CodeGenerator * codegen,bool * out_needs_bss_check)1801 static bool CanEncodeInlinedMethodInStackMap(const DexFile& outer_dex_file,
1802 ArtMethod* callee,
1803 const CodeGenerator* codegen,
1804 bool* out_needs_bss_check)
1805 REQUIRES_SHARED(Locks::mutator_lock_) {
1806 if (!Runtime::Current()->IsAotCompiler()) {
1807 // JIT can always encode methods in stack maps.
1808 return true;
1809 }
1810
1811 const DexFile* dex_file = callee->GetDexFile();
1812 if (IsSameDexFile(outer_dex_file, *dex_file)) {
1813 return true;
1814 }
1815
1816 // Inline across dexfiles if the callee's DexFile is:
1817 // 1) in the bootclasspath, or
1818 if (callee->GetDeclaringClass()->IsBootStrapClassLoaded()) {
1819 // In multi-image, each BCP DexFile has their own OatWriter. Since they don't cooperate with
1820 // each other, we request the BSS check for them.
1821 // TODO(solanes, 154012332): Add .bss support for BCP multi-image.
1822 *out_needs_bss_check = codegen->GetCompilerOptions().IsMultiImage();
1823 return true;
1824 }
1825
1826 // 2) is a non-BCP dexfile with the OatFile we are compiling.
1827 if (codegen->GetCompilerOptions().WithinOatFile(dex_file)) {
1828 return true;
1829 }
1830
1831 // TODO(solanes): Support more AOT cases for inlining:
1832 // - methods in class loader context's DexFiles
1833 return false;
1834 }
1835
1836 // Substitutes parameters in the callee graph with their values from the caller.
SubstituteArguments(HGraph * callee_graph,HInvoke * invoke_instruction,ReferenceTypeInfo receiver_type,const DexCompilationUnit & dex_compilation_unit)1837 void HInliner::SubstituteArguments(HGraph* callee_graph,
1838 HInvoke* invoke_instruction,
1839 ReferenceTypeInfo receiver_type,
1840 const DexCompilationUnit& dex_compilation_unit) {
1841 ArtMethod* const resolved_method = callee_graph->GetArtMethod();
1842 size_t parameter_index = 0;
1843 bool run_rtp = false;
1844 for (HInstructionIterator instructions(callee_graph->GetEntryBlock()->GetInstructions());
1845 !instructions.Done();
1846 instructions.Advance()) {
1847 HInstruction* current = instructions.Current();
1848 if (current->IsParameterValue()) {
1849 HInstruction* argument = invoke_instruction->InputAt(parameter_index);
1850 if (argument->IsNullConstant()) {
1851 current->ReplaceWith(callee_graph->GetNullConstant());
1852 } else if (argument->IsIntConstant()) {
1853 current->ReplaceWith(callee_graph->GetIntConstant(argument->AsIntConstant()->GetValue()));
1854 } else if (argument->IsLongConstant()) {
1855 current->ReplaceWith(callee_graph->GetLongConstant(argument->AsLongConstant()->GetValue()));
1856 } else if (argument->IsFloatConstant()) {
1857 current->ReplaceWith(
1858 callee_graph->GetFloatConstant(argument->AsFloatConstant()->GetValue()));
1859 } else if (argument->IsDoubleConstant()) {
1860 current->ReplaceWith(
1861 callee_graph->GetDoubleConstant(argument->AsDoubleConstant()->GetValue()));
1862 } else if (argument->GetType() == DataType::Type::kReference) {
1863 if (!resolved_method->IsStatic() && parameter_index == 0 && receiver_type.IsValid()) {
1864 run_rtp = true;
1865 current->SetReferenceTypeInfo(receiver_type);
1866 } else {
1867 current->SetReferenceTypeInfoIfValid(argument->GetReferenceTypeInfo());
1868 }
1869 current->AsParameterValue()->SetCanBeNull(argument->CanBeNull());
1870 }
1871 ++parameter_index;
1872 }
1873 }
1874
1875 // We have replaced formal arguments with actual arguments. If actual types
1876 // are more specific than the declared ones, run RTP again on the inner graph.
1877 if (run_rtp || ArgumentTypesMoreSpecific(invoke_instruction, resolved_method)) {
1878 ReferenceTypePropagation(callee_graph,
1879 dex_compilation_unit.GetDexCache(),
1880 /* is_first_run= */ false).Run();
1881 }
1882 }
1883
1884 // Returns whether we can inline the callee_graph into the target_block.
1885 //
1886 // This performs a combination of semantics checks, compiler support checks, and
1887 // resource limit checks.
1888 //
1889 // If this function returns true, it will also set out_number_of_instructions to
1890 // the number of instructions in the inlined body.
CanInlineBody(const HGraph * callee_graph,HInvoke * invoke,size_t * out_number_of_instructions,bool is_speculative) const1891 bool HInliner::CanInlineBody(const HGraph* callee_graph,
1892 HInvoke* invoke,
1893 size_t* out_number_of_instructions,
1894 bool is_speculative) const {
1895 ArtMethod* const resolved_method = callee_graph->GetArtMethod();
1896
1897 HBasicBlock* exit_block = callee_graph->GetExitBlock();
1898 if (exit_block == nullptr) {
1899 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedInfiniteLoop)
1900 << "Method " << resolved_method->PrettyMethod()
1901 << " could not be inlined because it has an infinite loop";
1902 return false;
1903 }
1904
1905 bool has_one_return = false;
1906 for (HBasicBlock* predecessor : exit_block->GetPredecessors()) {
1907 const HInstruction* last_instruction = predecessor->GetLastInstruction();
1908 // On inlinees, we can have Return/ReturnVoid/Throw -> TryBoundary -> Exit. To check for the
1909 // actual last instruction, we have to skip the TryBoundary instruction.
1910 if (last_instruction->IsTryBoundary()) {
1911 predecessor = predecessor->GetSinglePredecessor();
1912 last_instruction = predecessor->GetLastInstruction();
1913
1914 // If the last instruction chain is Return/ReturnVoid -> TryBoundary -> Exit we will have to
1915 // split a critical edge in InlineInto and might recompute loop information, which is
1916 // unsupported for irreducible loops.
1917 if (!last_instruction->IsThrow() && graph_->HasIrreducibleLoops()) {
1918 DCHECK(last_instruction->IsReturn() || last_instruction->IsReturnVoid());
1919 // TODO(ngeoffray): Support re-computing loop information to graphs with
1920 // irreducible loops?
1921 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedIrreducibleLoopCaller)
1922 << "Method " << resolved_method->PrettyMethod()
1923 << " could not be inlined because we will have to recompute the loop information and"
1924 << " the caller has irreducible loops";
1925 return false;
1926 }
1927 }
1928
1929 if (last_instruction->IsThrow()) {
1930 if (graph_->GetExitBlock() == nullptr) {
1931 // TODO(ngeoffray): Support adding HExit in the caller graph.
1932 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedInfiniteLoop)
1933 << "Method " << resolved_method->PrettyMethod()
1934 << " could not be inlined because one branch always throws and"
1935 << " caller does not have an exit block";
1936 return false;
1937 } else if (graph_->HasIrreducibleLoops()) {
1938 // TODO(ngeoffray): Support re-computing loop information to graphs with
1939 // irreducible loops?
1940 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedIrreducibleLoopCaller)
1941 << "Method " << resolved_method->PrettyMethod()
1942 << " could not be inlined because one branch always throws and"
1943 << " the caller has irreducible loops";
1944 return false;
1945 }
1946 } else {
1947 has_one_return = true;
1948 }
1949 }
1950
1951 if (!has_one_return) {
1952 if (!is_speculative) {
1953 // If we know that the method always throws with the particular parameters, set it as such.
1954 // This is better than using the dex instructions as we have more information about this
1955 // particular call. We don't mark speculative inlines (e.g. the ones from the inline cache) as
1956 // always throwing since they might not throw when executed.
1957 invoke->SetAlwaysThrows(/* always_throws= */ true);
1958 graph_->SetHasAlwaysThrowingInvokes(/* value= */ true);
1959 }
1960
1961 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedAlwaysThrows)
1962 << "Method " << resolved_method->PrettyMethod()
1963 << " could not be inlined because it always throws";
1964 return false;
1965 }
1966
1967 const bool too_many_registers =
1968 total_number_of_dex_registers_ > kMaximumNumberOfCumulatedDexRegisters;
1969 bool needs_bss_check = false;
1970 const bool can_encode_in_stack_map = CanEncodeInlinedMethodInStackMap(
1971 *outer_compilation_unit_.GetDexFile(), resolved_method, codegen_, &needs_bss_check);
1972 size_t number_of_instructions = 0;
1973 // Skip the entry block, it does not contain instructions that prevent inlining.
1974 for (HBasicBlock* block : callee_graph->GetReversePostOrderSkipEntryBlock()) {
1975 if (block->IsLoopHeader()) {
1976 if (block->GetLoopInformation()->IsIrreducible()) {
1977 // Don't inline methods with irreducible loops, they could prevent some
1978 // optimizations to run.
1979 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedIrreducibleLoopCallee)
1980 << "Method " << resolved_method->PrettyMethod()
1981 << " could not be inlined because it contains an irreducible loop";
1982 return false;
1983 }
1984 if (!block->GetLoopInformation()->HasExitEdge()) {
1985 // Don't inline methods with loops without exit, since they cause the
1986 // loop information to be computed incorrectly when updating after
1987 // inlining.
1988 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedLoopWithoutExit)
1989 << "Method " << resolved_method->PrettyMethod()
1990 << " could not be inlined because it contains a loop with no exit";
1991 return false;
1992 }
1993 }
1994
1995 for (HInstructionIterator instr_it(block->GetInstructions());
1996 !instr_it.Done();
1997 instr_it.Advance()) {
1998 if (++number_of_instructions > inlining_budget_) {
1999 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedInstructionBudget)
2000 << "Method " << resolved_method->PrettyMethod()
2001 << " is not inlined because the outer method has reached"
2002 << " its instruction budget limit.";
2003 return false;
2004 }
2005 HInstruction* current = instr_it.Current();
2006 if (current->NeedsEnvironment()) {
2007 if (too_many_registers) {
2008 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedEnvironmentBudget)
2009 << "Method " << resolved_method->PrettyMethod()
2010 << " is not inlined because its caller has reached"
2011 << " its environment budget limit.";
2012 return false;
2013 }
2014
2015 if (!can_encode_in_stack_map) {
2016 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedStackMaps)
2017 << "Method " << resolved_method->PrettyMethod() << " could not be inlined because "
2018 << current->DebugName() << " needs an environment, is in a different dex file"
2019 << ", and cannot be encoded in the stack maps.";
2020 return false;
2021 }
2022 }
2023
2024 if (current->IsUnresolvedStaticFieldGet() ||
2025 current->IsUnresolvedInstanceFieldGet() ||
2026 current->IsUnresolvedStaticFieldSet() ||
2027 current->IsUnresolvedInstanceFieldSet() ||
2028 current->IsInvokeUnresolved()) {
2029 // Unresolved invokes / field accesses are expensive at runtime when decoding inlining info,
2030 // so don't inline methods that have them.
2031 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedUnresolvedEntrypoint)
2032 << "Method " << resolved_method->PrettyMethod()
2033 << " could not be inlined because it is using an unresolved"
2034 << " entrypoint";
2035 return false;
2036 }
2037
2038 // We currently don't have support for inlining across dex files if we are:
2039 // 1) In AoT,
2040 // 2) cross-dex inlining,
2041 // 3) the callee is a BCP DexFile,
2042 // 4) we are compiling multi image, and
2043 // 5) have an instruction that needs a bss entry, which will always be
2044 // 5)b) an instruction that needs an environment.
2045 // 1) - 4) are encoded in `needs_bss_check` (see CanEncodeInlinedMethodInStackMap).
2046 if (needs_bss_check && current->NeedsBss()) {
2047 DCHECK(current->NeedsEnvironment());
2048 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedBss)
2049 << "Method " << resolved_method->PrettyMethod()
2050 << " could not be inlined because it needs a BSS check";
2051 return false;
2052 }
2053 }
2054 }
2055
2056 *out_number_of_instructions = number_of_instructions;
2057 return true;
2058 }
2059
TryBuildAndInlineHelper(HInvoke * invoke_instruction,ArtMethod * resolved_method,ReferenceTypeInfo receiver_type,HInstruction ** return_replacement,bool is_speculative)2060 bool HInliner::TryBuildAndInlineHelper(HInvoke* invoke_instruction,
2061 ArtMethod* resolved_method,
2062 ReferenceTypeInfo receiver_type,
2063 HInstruction** return_replacement,
2064 bool is_speculative) {
2065 DCHECK(!(resolved_method->IsStatic() && receiver_type.IsValid()));
2066 const dex::CodeItem* code_item = resolved_method->GetCodeItem();
2067 const DexFile& callee_dex_file = *resolved_method->GetDexFile();
2068 uint32_t method_index = resolved_method->GetDexMethodIndex();
2069 CodeItemDebugInfoAccessor code_item_accessor(resolved_method->DexInstructionDebugInfo());
2070 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
2071 Handle<mirror::DexCache> dex_cache = NewHandleIfDifferent(resolved_method->GetDexCache(),
2072 caller_compilation_unit_.GetDexCache(),
2073 graph_);
2074 Handle<mirror::ClassLoader> class_loader =
2075 NewHandleIfDifferent(resolved_method->GetDeclaringClass()->GetClassLoader(),
2076 caller_compilation_unit_.GetClassLoader(),
2077 graph_);
2078
2079 Handle<mirror::Class> compiling_class =
2080 graph_->GetHandleCache()->NewHandle(resolved_method->GetDeclaringClass());
2081 DexCompilationUnit dex_compilation_unit(
2082 class_loader,
2083 class_linker,
2084 callee_dex_file,
2085 code_item,
2086 resolved_method->GetDeclaringClass()->GetDexClassDefIndex(),
2087 method_index,
2088 resolved_method->GetAccessFlags(),
2089 /* verified_method= */ nullptr,
2090 dex_cache,
2091 compiling_class);
2092
2093 InvokeType invoke_type = invoke_instruction->GetInvokeType();
2094 if (invoke_type == kInterface) {
2095 // We have statically resolved the dispatch. To please the class linker
2096 // at runtime, we change this call as if it was a virtual call.
2097 invoke_type = kVirtual;
2098 }
2099
2100 bool caller_dead_reference_safe = graph_->IsDeadReferenceSafe();
2101 const dex::ClassDef& callee_class = resolved_method->GetClassDef();
2102 // MethodContainsRSensitiveAccess is currently slow, but HasDeadReferenceSafeAnnotation()
2103 // is currently rarely true.
2104 bool callee_dead_reference_safe =
2105 annotations::HasDeadReferenceSafeAnnotation(callee_dex_file, callee_class)
2106 && !annotations::MethodContainsRSensitiveAccess(callee_dex_file, callee_class, method_index);
2107
2108 const int32_t caller_instruction_counter = graph_->GetCurrentInstructionId();
2109 HGraph* callee_graph = new (graph_->GetAllocator()) HGraph(
2110 graph_->GetAllocator(),
2111 graph_->GetArenaStack(),
2112 graph_->GetHandleCache()->GetHandles(),
2113 callee_dex_file,
2114 method_index,
2115 codegen_->GetCompilerOptions().GetInstructionSet(),
2116 invoke_type,
2117 callee_dead_reference_safe,
2118 graph_->IsDebuggable(),
2119 graph_->GetCompilationKind(),
2120 /* start_instruction_id= */ caller_instruction_counter);
2121 callee_graph->SetArtMethod(resolved_method);
2122
2123 ScopedProfilingInfoUse spiu(Runtime::Current()->GetJit(), resolved_method, Thread::Current());
2124 if (Runtime::Current()->GetJit() != nullptr) {
2125 callee_graph->SetProfilingInfo(spiu.GetProfilingInfo());
2126 }
2127
2128 // When they are needed, allocate `inline_stats_` on the Arena instead
2129 // of on the stack, as Clang might produce a stack frame too large
2130 // for this function, that would not fit the requirements of the
2131 // `-Wframe-larger-than` option.
2132 if (stats_ != nullptr) {
2133 // Reuse one object for all inline attempts from this caller to keep Arena memory usage low.
2134 if (inline_stats_ == nullptr) {
2135 void* storage = graph_->GetAllocator()->Alloc<OptimizingCompilerStats>(kArenaAllocMisc);
2136 inline_stats_ = new (storage) OptimizingCompilerStats;
2137 } else {
2138 inline_stats_->Reset();
2139 }
2140 }
2141 HGraphBuilder builder(callee_graph,
2142 code_item_accessor,
2143 &dex_compilation_unit,
2144 &outer_compilation_unit_,
2145 codegen_,
2146 inline_stats_);
2147
2148 if (builder.BuildGraph() != kAnalysisSuccess) {
2149 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedCannotBuild)
2150 << "Method " << callee_dex_file.PrettyMethod(method_index)
2151 << " could not be built, so cannot be inlined";
2152 return false;
2153 }
2154
2155 SubstituteArguments(callee_graph, invoke_instruction, receiver_type, dex_compilation_unit);
2156
2157 const bool try_catch_inlining_allowed_for_recursive_inline =
2158 // It was allowed previously.
2159 try_catch_inlining_allowed_ &&
2160 // The current invoke is not a try block.
2161 !invoke_instruction->GetBlock()->IsTryBlock();
2162 RunOptimizations(callee_graph,
2163 code_item,
2164 dex_compilation_unit,
2165 try_catch_inlining_allowed_for_recursive_inline);
2166
2167 size_t number_of_instructions = 0;
2168 if (!CanInlineBody(callee_graph, invoke_instruction, &number_of_instructions, is_speculative)) {
2169 return false;
2170 }
2171
2172 DCHECK_EQ(caller_instruction_counter, graph_->GetCurrentInstructionId())
2173 << "No instructions can be added to the outer graph while inner graph is being built";
2174
2175 // Inline the callee graph inside the caller graph.
2176 const int32_t callee_instruction_counter = callee_graph->GetCurrentInstructionId();
2177 graph_->SetCurrentInstructionId(callee_instruction_counter);
2178 *return_replacement = callee_graph->InlineInto(graph_, invoke_instruction);
2179 // Update our budget for other inlining attempts in `caller_graph`.
2180 total_number_of_instructions_ += number_of_instructions;
2181 UpdateInliningBudget();
2182
2183 DCHECK_EQ(callee_instruction_counter, callee_graph->GetCurrentInstructionId())
2184 << "No instructions can be added to the inner graph during inlining into the outer graph";
2185
2186 if (stats_ != nullptr) {
2187 DCHECK(inline_stats_ != nullptr);
2188 inline_stats_->AddTo(stats_);
2189 }
2190
2191 if (caller_dead_reference_safe && !callee_dead_reference_safe) {
2192 // Caller was dead reference safe, but is not anymore, since we inlined dead
2193 // reference unsafe code. Prior transformations remain valid, since they did not
2194 // affect the inlined code.
2195 graph_->MarkDeadReferenceUnsafe();
2196 }
2197
2198 return true;
2199 }
2200
RunOptimizations(HGraph * callee_graph,const dex::CodeItem * code_item,const DexCompilationUnit & dex_compilation_unit,bool try_catch_inlining_allowed_for_recursive_inline)2201 void HInliner::RunOptimizations(HGraph* callee_graph,
2202 const dex::CodeItem* code_item,
2203 const DexCompilationUnit& dex_compilation_unit,
2204 bool try_catch_inlining_allowed_for_recursive_inline) {
2205 // Note: if the outermost_graph_ is being compiled OSR, we should not run any
2206 // optimization that could lead to a HDeoptimize. The following optimizations do not.
2207 HDeadCodeElimination dce(callee_graph, inline_stats_, "dead_code_elimination$inliner");
2208 HConstantFolding fold(callee_graph, inline_stats_, "constant_folding$inliner");
2209 InstructionSimplifier simplify(callee_graph, codegen_, inline_stats_);
2210
2211 HOptimization* optimizations[] = {
2212 &fold,
2213 &simplify,
2214 &dce,
2215 };
2216
2217 for (size_t i = 0; i < arraysize(optimizations); ++i) {
2218 HOptimization* optimization = optimizations[i];
2219 optimization->Run();
2220 }
2221
2222 // Bail early for pathological cases on the environment (for example recursive calls,
2223 // or too large environment).
2224 if (total_number_of_dex_registers_ > kMaximumNumberOfCumulatedDexRegisters) {
2225 LOG_NOTE() << "Calls in " << callee_graph->GetArtMethod()->PrettyMethod()
2226 << " will not be inlined because the outer method has reached"
2227 << " its environment budget limit.";
2228 return;
2229 }
2230
2231 // Bail early if we know we already are over the limit.
2232 size_t number_of_instructions = CountNumberOfInstructions(callee_graph);
2233 if (number_of_instructions > inlining_budget_) {
2234 LOG_NOTE() << "Calls in " << callee_graph->GetArtMethod()->PrettyMethod()
2235 << " will not be inlined because the outer method has reached"
2236 << " its instruction budget limit. " << number_of_instructions;
2237 return;
2238 }
2239
2240 CodeItemDataAccessor accessor(callee_graph->GetDexFile(), code_item);
2241 HInliner inliner(callee_graph,
2242 outermost_graph_,
2243 codegen_,
2244 outer_compilation_unit_,
2245 dex_compilation_unit,
2246 inline_stats_,
2247 total_number_of_dex_registers_ + accessor.RegistersSize(),
2248 total_number_of_instructions_ + number_of_instructions,
2249 this,
2250 depth_ + 1,
2251 try_catch_inlining_allowed_for_recursive_inline);
2252 inliner.Run();
2253 }
2254
IsReferenceTypeRefinement(ObjPtr<mirror::Class> declared_class,bool declared_is_exact,bool declared_can_be_null,HInstruction * actual_obj)2255 static bool IsReferenceTypeRefinement(ObjPtr<mirror::Class> declared_class,
2256 bool declared_is_exact,
2257 bool declared_can_be_null,
2258 HInstruction* actual_obj)
2259 REQUIRES_SHARED(Locks::mutator_lock_) {
2260 if (declared_can_be_null && !actual_obj->CanBeNull()) {
2261 return true;
2262 }
2263
2264 ReferenceTypeInfo actual_rti = actual_obj->GetReferenceTypeInfo();
2265 if (!actual_rti.IsValid()) {
2266 return false;
2267 }
2268
2269 ObjPtr<mirror::Class> actual_class = actual_rti.GetTypeHandle().Get();
2270 return (actual_rti.IsExact() && !declared_is_exact) ||
2271 (declared_class != actual_class && declared_class->IsAssignableFrom(actual_class));
2272 }
2273
IsReferenceTypeRefinement(ObjPtr<mirror::Class> declared_class,bool declared_can_be_null,HInstruction * actual_obj)2274 static bool IsReferenceTypeRefinement(ObjPtr<mirror::Class> declared_class,
2275 bool declared_can_be_null,
2276 HInstruction* actual_obj)
2277 REQUIRES_SHARED(Locks::mutator_lock_) {
2278 bool admissible = ReferenceTypePropagation::IsAdmissible(declared_class);
2279 return IsReferenceTypeRefinement(
2280 admissible ? declared_class : GetClassRoot<mirror::Class>(),
2281 /*declared_is_exact=*/ admissible && declared_class->CannotBeAssignedFromOtherTypes(),
2282 declared_can_be_null,
2283 actual_obj);
2284 }
2285
ArgumentTypesMoreSpecific(HInvoke * invoke_instruction,ArtMethod * resolved_method)2286 bool HInliner::ArgumentTypesMoreSpecific(HInvoke* invoke_instruction, ArtMethod* resolved_method) {
2287 // If this is an instance call, test whether the type of the `this` argument
2288 // is more specific than the class which declares the method.
2289 if (!resolved_method->IsStatic()) {
2290 if (IsReferenceTypeRefinement(resolved_method->GetDeclaringClass(),
2291 /*declared_can_be_null=*/ false,
2292 invoke_instruction->InputAt(0u))) {
2293 return true;
2294 }
2295 }
2296
2297 // Iterate over the list of parameter types and test whether any of the
2298 // actual inputs has a more specific reference type than the type declared in
2299 // the signature.
2300 const dex::TypeList* param_list = resolved_method->GetParameterTypeList();
2301 for (size_t param_idx = 0,
2302 input_idx = resolved_method->IsStatic() ? 0 : 1,
2303 e = (param_list == nullptr ? 0 : param_list->Size());
2304 param_idx < e;
2305 ++param_idx, ++input_idx) {
2306 HInstruction* input = invoke_instruction->InputAt(input_idx);
2307 if (input->GetType() == DataType::Type::kReference) {
2308 ObjPtr<mirror::Class> param_cls = resolved_method->LookupResolvedClassFromTypeIndex(
2309 param_list->GetTypeItem(param_idx).type_idx_);
2310 if (IsReferenceTypeRefinement(param_cls, /*declared_can_be_null=*/ true, input)) {
2311 return true;
2312 }
2313 }
2314 }
2315
2316 return false;
2317 }
2318
ReturnTypeMoreSpecific(HInstruction * return_replacement,HInvoke * invoke_instruction)2319 bool HInliner::ReturnTypeMoreSpecific(HInstruction* return_replacement,
2320 HInvoke* invoke_instruction) {
2321 // Check the integrity of reference types and run another type propagation if needed.
2322 if (return_replacement != nullptr) {
2323 if (return_replacement->GetType() == DataType::Type::kReference) {
2324 // Test if the return type is a refinement of the declared return type.
2325 ReferenceTypeInfo invoke_rti = invoke_instruction->GetReferenceTypeInfo();
2326 if (IsReferenceTypeRefinement(invoke_rti.GetTypeHandle().Get(),
2327 invoke_rti.IsExact(),
2328 /*declared_can_be_null=*/ true,
2329 return_replacement)) {
2330 return true;
2331 } else if (return_replacement->IsInstanceFieldGet()) {
2332 HInstanceFieldGet* field_get = return_replacement->AsInstanceFieldGet();
2333 if (field_get->GetFieldInfo().GetField() ==
2334 GetClassRoot<mirror::Object>()->GetInstanceField(0)) {
2335 return true;
2336 }
2337 }
2338 } else if (return_replacement->IsInstanceOf()) {
2339 // Inlining InstanceOf into an If may put a tighter bound on reference types.
2340 return true;
2341 }
2342 }
2343
2344 return false;
2345 }
2346
FixUpReturnReferenceType(ArtMethod * resolved_method,HInstruction * return_replacement)2347 void HInliner::FixUpReturnReferenceType(ArtMethod* resolved_method,
2348 HInstruction* return_replacement) {
2349 if (return_replacement != nullptr) {
2350 if (return_replacement->GetType() == DataType::Type::kReference) {
2351 if (!return_replacement->GetReferenceTypeInfo().IsValid()) {
2352 // Make sure that we have a valid type for the return. We may get an invalid one when
2353 // we inline invokes with multiple branches and create a Phi for the result.
2354 // TODO: we could be more precise by merging the phi inputs but that requires
2355 // some functionality from the reference type propagation.
2356 DCHECK(return_replacement->IsPhi());
2357 ObjPtr<mirror::Class> cls = resolved_method->LookupResolvedReturnType();
2358 ReferenceTypeInfo rti = ReferenceTypePropagation::IsAdmissible(cls)
2359 ? ReferenceTypeInfo::Create(graph_->GetHandleCache()->NewHandle(cls))
2360 : graph_->GetInexactObjectRti();
2361 return_replacement->SetReferenceTypeInfo(rti);
2362 }
2363 }
2364 }
2365 }
2366
2367 } // namespace art
2368