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/logging.h"
21 #include "base/pointer_size.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 "profiling_info_builder.h"
41 #include "reference_type_propagation.h"
42 #include "register_allocator_linear_scan.h"
43 #include "scoped_thread_state_change-inl.h"
44 #include "sharpening.h"
45 #include "ssa_builder.h"
46 #include "ssa_phi_elimination.h"
47 #include "thread.h"
48 #include "verifier/verifier_compiler_binding.h"
49
50 namespace art HIDDEN {
51
52 // Instruction limit to control memory.
53 static constexpr size_t kMaximumNumberOfTotalInstructions = 1024;
54
55 // Maximum number of instructions for considering a method small,
56 // which we will always try to inline if the other non-instruction limits
57 // are not reached.
58 static constexpr size_t kMaximumNumberOfInstructionsForSmallMethod = 3;
59
60 // Limit the number of dex registers that we accumulate while inlining
61 // to avoid creating large amount of nested environments.
62 static constexpr size_t kMaximumNumberOfCumulatedDexRegisters = 32;
63
64 // Limit recursive call inlining, which do not benefit from too
65 // much inlining compared to code locality.
66 static constexpr size_t kMaximumNumberOfRecursiveCalls = 4;
67
68 // Limit recursive polymorphic call inlining to prevent code bloat, since it can quickly get out of
69 // hand in the presence of multiple Wrapper classes. We set this to 0 to disallow polymorphic
70 // recursive calls at all.
71 static constexpr size_t kMaximumNumberOfPolymorphicRecursiveCalls = 0;
72
73 // Controls the use of inline caches in AOT mode.
74 static constexpr bool kUseAOTInlineCaches = true;
75
76 // Controls the use of inlining try catches.
77 static constexpr bool kInlineTryCatches = true;
78
79 // We check for line numbers to make sure the DepthString implementation
80 // aligns the output nicely.
81 #define LOG_INTERNAL(msg) \
82 static_assert(__LINE__ > 10, "Unhandled line number"); \
83 static_assert(__LINE__ < 10000, "Unhandled line number"); \
84 VLOG(compiler) << DepthString(__LINE__) << msg
85
86 #define LOG_TRY() LOG_INTERNAL("Try inlinining call: ")
87 #define LOG_NOTE() LOG_INTERNAL("Note: ")
88 #define LOG_SUCCESS() LOG_INTERNAL("Success: ")
89 #define LOG_FAIL(stats_ptr, stat) MaybeRecordStat(stats_ptr, stat); LOG_INTERNAL("Fail: ")
90 #define LOG_FAIL_NO_STAT() LOG_INTERNAL("Fail: ")
91
DepthString(int line) const92 std::string HInliner::DepthString(int line) const {
93 std::string value;
94 // Indent according to the inlining depth.
95 size_t count = depth_;
96 // Line numbers get printed in the log, so add a space if the log's line is less
97 // than 1000, and two if less than 100. 10 cannot be reached as it's the copyright.
98 if (!kIsTargetBuild) {
99 if (line < 100) {
100 value += " ";
101 }
102 if (line < 1000) {
103 value += " ";
104 }
105 // Safeguard if this file reaches more than 10000 lines.
106 DCHECK_LT(line, 10000);
107 }
108 for (size_t i = 0; i < count; ++i) {
109 value += " ";
110 }
111 return value;
112 }
113
CountNumberOfInstructions(HGraph * graph)114 static size_t CountNumberOfInstructions(HGraph* graph) {
115 size_t number_of_instructions = 0;
116 for (HBasicBlock* block : graph->GetReversePostOrderSkipEntryBlock()) {
117 for (HInstructionIterator instr_it(block->GetInstructions());
118 !instr_it.Done();
119 instr_it.Advance()) {
120 ++number_of_instructions;
121 }
122 }
123 return number_of_instructions;
124 }
125
UpdateInliningBudget()126 void HInliner::UpdateInliningBudget() {
127 if (total_number_of_instructions_ >= kMaximumNumberOfTotalInstructions) {
128 // Always try to inline small methods.
129 inlining_budget_ = kMaximumNumberOfInstructionsForSmallMethod;
130 } else {
131 inlining_budget_ = std::max(
132 kMaximumNumberOfInstructionsForSmallMethod,
133 kMaximumNumberOfTotalInstructions - total_number_of_instructions_);
134 }
135 }
136
Run()137 bool HInliner::Run() {
138 if (codegen_->GetCompilerOptions().GetInlineMaxCodeUnits() == 0) {
139 // Inlining effectively disabled.
140 return false;
141 } else if (graph_->IsDebuggable()) {
142 // For simplicity, we currently never inline when the graph is debuggable. This avoids
143 // doing some logic in the runtime to discover if a method could have been inlined.
144 return false;
145 }
146
147 bool did_inline = false;
148
149 // Initialize the number of instructions for the method being compiled. Recursive calls
150 // to HInliner::Run have already updated the instruction count.
151 if (outermost_graph_ == graph_) {
152 total_number_of_instructions_ = CountNumberOfInstructions(graph_);
153 }
154
155 UpdateInliningBudget();
156 DCHECK_NE(total_number_of_instructions_, 0u);
157 DCHECK_NE(inlining_budget_, 0u);
158
159 // If we're compiling tests, honor inlining directives in method names:
160 // - if a method's name contains the substring "$noinline$", do not
161 // inline that method;
162 // - if a method's name contains the substring "$inline$", ensure
163 // that this method is actually inlined.
164 // We limit the latter to AOT compilation, as the JIT may or may not inline
165 // depending on the state of classes at runtime.
166 const bool honor_noinline_directives = codegen_->GetCompilerOptions().CompileArtTest();
167 const bool honor_inline_directives =
168 honor_noinline_directives &&
169 Runtime::Current()->IsAotCompiler() &&
170 !graph_->IsCompilingBaseline();
171
172 // Keep a copy of all blocks when starting the visit.
173 ArenaVector<HBasicBlock*> blocks = graph_->GetReversePostOrder();
174 DCHECK(!blocks.empty());
175 // Because we are changing the graph when inlining,
176 // we just iterate over the blocks of the outer method.
177 // This avoids doing the inlining work again on the inlined blocks.
178 for (HBasicBlock* block : blocks) {
179 for (HInstruction* instruction = block->GetFirstInstruction(); instruction != nullptr;) {
180 HInstruction* next = instruction->GetNext();
181 HInvoke* call = instruction->AsInvokeOrNull();
182 // As long as the call is not intrinsified, it is worth trying to inline.
183 if (call != nullptr && !codegen_->IsImplementedIntrinsic(call)) {
184 if (honor_noinline_directives) {
185 // Debugging case: directives in method names control or assert on inlining.
186 std::string callee_name =
187 call->GetMethodReference().PrettyMethod(/* with_signature= */ false);
188 // Tests prevent inlining by having $noinline$ in their method names.
189 if (callee_name.find("$noinline$") == std::string::npos) {
190 if (TryInline(call)) {
191 did_inline = true;
192 } else if (honor_inline_directives) {
193 bool should_have_inlined = (callee_name.find("$inline$") != std::string::npos);
194 CHECK(!should_have_inlined) << "Could not inline " << callee_name;
195 }
196 }
197 } else {
198 DCHECK(!honor_inline_directives);
199 // Normal case: try to inline.
200 if (TryInline(call)) {
201 did_inline = true;
202 }
203 }
204 }
205 instruction = next;
206 }
207 }
208
209 if (run_extra_type_propagation_) {
210 ReferenceTypePropagation rtp_fixup(graph_,
211 outer_compilation_unit_.GetDexCache(),
212 /* is_first_run= */ false);
213 rtp_fixup.Run();
214 }
215
216 // We return true if we either inlined at least one method, or we marked one of our methods as
217 // always throwing.
218 // To check if we added an always throwing method we can either:
219 // 1) Pass a boolean throughout the pipeline and get an accurate result, or
220 // 2) Just check that the `HasAlwaysThrowingInvokes()` flag is true now. This is not 100%
221 // accurate but the only other part where we set `HasAlwaysThrowingInvokes` is constant
222 // folding the DivideUnsigned intrinsics for when the divisor is known to be 0. This case is
223 // rare enough that changing the pipeline for this is not worth it. In the case of the false
224 // positive (i.e. A) we didn't inline at all, B) the graph already had an always throwing
225 // invoke, and C) we didn't set any new always throwing invokes), we will be running constant
226 // folding, instruction simplifier, and dead code elimination one more time even though it
227 // shouldn't change things. There's no false negative case.
228 return did_inline || graph_->HasAlwaysThrowingInvokes();
229 }
230
IsMethodOrDeclaringClassFinal(ArtMethod * method)231 static bool IsMethodOrDeclaringClassFinal(ArtMethod* method)
232 REQUIRES_SHARED(Locks::mutator_lock_) {
233 return method->IsFinal() || method->GetDeclaringClass()->IsFinal();
234 }
235
236 /**
237 * Given the `resolved_method` looked up in the dex cache, try to find
238 * the actual runtime target of an interface or virtual call.
239 * Return nullptr if the runtime target cannot be proven.
240 */
FindVirtualOrInterfaceTarget(HInvoke * invoke,ReferenceTypeInfo info)241 static ArtMethod* FindVirtualOrInterfaceTarget(HInvoke* invoke, ReferenceTypeInfo info)
242 REQUIRES_SHARED(Locks::mutator_lock_) {
243 ArtMethod* resolved_method = invoke->GetResolvedMethod();
244 if (IsMethodOrDeclaringClassFinal(resolved_method)) {
245 // No need to lookup further, the resolved method will be the target.
246 return resolved_method;
247 }
248
249 if (info.GetTypeHandle()->IsInterface()) {
250 // Statically knowing that the receiver has an interface type cannot
251 // help us find what is the target method.
252 return nullptr;
253 } else if (!resolved_method->GetDeclaringClass()->IsAssignableFrom(info.GetTypeHandle().Get())) {
254 // The method that we're trying to call is not in the receiver's class or super classes.
255 return nullptr;
256 } else if (info.GetTypeHandle()->IsErroneous()) {
257 // If the type is erroneous, do not go further, as we are going to query the vtable or
258 // imt table, that we can only safely do on non-erroneous classes.
259 return nullptr;
260 }
261
262 ClassLinker* cl = Runtime::Current()->GetClassLinker();
263 PointerSize pointer_size = cl->GetImagePointerSize();
264 if (invoke->IsInvokeInterface()) {
265 resolved_method = info.GetTypeHandle()->FindVirtualMethodForInterface(
266 resolved_method, pointer_size);
267 } else {
268 DCHECK(invoke->IsInvokeVirtual());
269 resolved_method = info.GetTypeHandle()->FindVirtualMethodForVirtual(
270 resolved_method, pointer_size);
271 }
272
273 if (resolved_method == nullptr) {
274 // The information we had on the receiver was not enough to find
275 // the target method. Since we check above the exact type of the receiver,
276 // the only reason this can happen is an IncompatibleClassChangeError.
277 return nullptr;
278 } else if (!resolved_method->IsInvokable()) {
279 // The information we had on the receiver was not enough to find
280 // the target method. Since we check above the exact type of the receiver,
281 // the only reason this can happen is an IncompatibleClassChangeError.
282 return nullptr;
283 } else if (IsMethodOrDeclaringClassFinal(resolved_method)) {
284 // A final method has to be the target method.
285 return resolved_method;
286 } else if (info.IsExact()) {
287 // If we found a method and the receiver's concrete type is statically
288 // known, we know for sure the target.
289 return resolved_method;
290 } else {
291 // Even if we did find a method, the receiver type was not enough to
292 // statically find the runtime target.
293 return nullptr;
294 }
295 }
296
FindMethodIndexIn(ArtMethod * method,const DexFile & dex_file,uint32_t name_and_signature_index)297 static uint32_t FindMethodIndexIn(ArtMethod* method,
298 const DexFile& dex_file,
299 uint32_t name_and_signature_index)
300 REQUIRES_SHARED(Locks::mutator_lock_) {
301 if (IsSameDexFile(*method->GetDexFile(), dex_file)) {
302 return method->GetDexMethodIndex();
303 } else {
304 return method->FindDexMethodIndexInOtherDexFile(dex_file, name_and_signature_index);
305 }
306 }
307
FindClassIndexIn(ObjPtr<mirror::Class> cls,const DexCompilationUnit & compilation_unit)308 static dex::TypeIndex FindClassIndexIn(ObjPtr<mirror::Class> cls,
309 const DexCompilationUnit& compilation_unit)
310 REQUIRES_SHARED(Locks::mutator_lock_) {
311 const DexFile& dex_file = *compilation_unit.GetDexFile();
312 dex::TypeIndex index;
313 if (cls->GetDexCache() == nullptr) {
314 DCHECK(cls->IsArrayClass()) << cls->PrettyClass();
315 index = cls->FindTypeIndexInOtherDexFile(dex_file);
316 } else if (!cls->GetDexTypeIndex().IsValid()) {
317 DCHECK(cls->IsProxyClass()) << cls->PrettyClass();
318 // TODO: deal with proxy classes.
319 } else if (IsSameDexFile(cls->GetDexFile(), dex_file)) {
320 DCHECK_EQ(cls->GetDexCache(), compilation_unit.GetDexCache().Get());
321 index = cls->GetDexTypeIndex();
322 } else {
323 index = cls->FindTypeIndexInOtherDexFile(dex_file);
324 // We cannot guarantee the entry will resolve to the same class,
325 // as there may be different class loaders. So only return the index if it's
326 // the right class already resolved with the class loader.
327 if (index.IsValid()) {
328 ObjPtr<mirror::Class> resolved = compilation_unit.GetClassLinker()->LookupResolvedType(
329 index, compilation_unit.GetDexCache().Get(), compilation_unit.GetClassLoader().Get());
330 if (resolved != cls) {
331 index = dex::TypeIndex::Invalid();
332 }
333 }
334 }
335
336 return index;
337 }
338
GetInlineCacheType(const StackHandleScope<InlineCache::kIndividualCacheSize> & classes)339 HInliner::InlineCacheType HInliner::GetInlineCacheType(
340 const StackHandleScope<InlineCache::kIndividualCacheSize>& classes) {
341 DCHECK_EQ(classes.Capacity(), InlineCache::kIndividualCacheSize);
342 uint8_t number_of_types = classes.Size();
343 if (number_of_types == 0) {
344 return kInlineCacheUninitialized;
345 } else if (number_of_types == 1) {
346 return kInlineCacheMonomorphic;
347 } else if (number_of_types == InlineCache::kIndividualCacheSize) {
348 return kInlineCacheMegamorphic;
349 } else {
350 return kInlineCachePolymorphic;
351 }
352 }
353
GetMonomorphicType(const StackHandleScope<InlineCache::kIndividualCacheSize> & classes)354 static inline ObjPtr<mirror::Class> GetMonomorphicType(
355 const StackHandleScope<InlineCache::kIndividualCacheSize>& classes)
356 REQUIRES_SHARED(Locks::mutator_lock_) {
357 DCHECK(classes.GetReference(0) != nullptr);
358 return classes.GetReference(0)->AsClass();
359 }
360
FindMethodFromCHA(ArtMethod * resolved_method)361 ArtMethod* HInliner::FindMethodFromCHA(ArtMethod* resolved_method) {
362 if (!resolved_method->HasSingleImplementation()) {
363 return nullptr;
364 }
365 if (Runtime::Current()->IsAotCompiler()) {
366 // No CHA-based devirtulization for AOT compiler (yet).
367 return nullptr;
368 }
369 if (Runtime::Current()->IsZygote()) {
370 // No CHA-based devirtulization for Zygote, as it compiles with
371 // offline information.
372 return nullptr;
373 }
374 if (outermost_graph_->IsCompilingOsr()) {
375 // We do not support HDeoptimize in OSR methods.
376 return nullptr;
377 }
378 PointerSize pointer_size = caller_compilation_unit_.GetClassLinker()->GetImagePointerSize();
379 ArtMethod* single_impl = resolved_method->GetSingleImplementation(pointer_size);
380 if (single_impl == nullptr) {
381 return nullptr;
382 }
383 if (single_impl->IsProxyMethod()) {
384 // Proxy method is a generic invoker that's not worth
385 // devirtualizing/inlining. It also causes issues when the proxy
386 // method is in another dex file if we try to rewrite invoke-interface to
387 // invoke-virtual because a proxy method doesn't have a real dex file.
388 return nullptr;
389 }
390 if (!single_impl->GetDeclaringClass()->IsResolved()) {
391 // There's a race with the class loading, which updates the CHA info
392 // before setting the class to resolved. So we just bail for this
393 // rare occurence.
394 return nullptr;
395 }
396 return single_impl;
397 }
398
IsMethodVerified(ArtMethod * method)399 static bool IsMethodVerified(ArtMethod* method)
400 REQUIRES_SHARED(Locks::mutator_lock_) {
401 if (method->GetDeclaringClass()->IsVerified()) {
402 return true;
403 }
404 // For AOT, we check if the class has a verification status that allows us to
405 // inline / analyze.
406 // At runtime, we know this is cold code if the class is not verified, so don't
407 // bother analyzing.
408 if (Runtime::Current()->IsAotCompiler()) {
409 if (method->GetDeclaringClass()->IsVerifiedNeedsAccessChecks() ||
410 method->GetDeclaringClass()->ShouldVerifyAtRuntime()) {
411 return true;
412 }
413 }
414 return false;
415 }
416
AlwaysThrows(ArtMethod * method)417 static bool AlwaysThrows(ArtMethod* method)
418 REQUIRES_SHARED(Locks::mutator_lock_) {
419 DCHECK(method != nullptr);
420 // Skip non-compilable and unverified methods.
421 if (!method->IsCompilable() || !IsMethodVerified(method)) {
422 return false;
423 }
424 // Skip native methods, methods with try blocks, and methods that are too large.
425 CodeItemDataAccessor accessor(method->DexInstructionData());
426 if (!accessor.HasCodeItem() ||
427 accessor.TriesSize() != 0 ||
428 accessor.InsnsSizeInCodeUnits() > kMaximumNumberOfTotalInstructions) {
429 return false;
430 }
431 // Scan for exits.
432 bool throw_seen = false;
433 for (const DexInstructionPcPair& pair : accessor) {
434 switch (pair.Inst().Opcode()) {
435 case Instruction::RETURN:
436 case Instruction::RETURN_VOID:
437 case Instruction::RETURN_WIDE:
438 case Instruction::RETURN_OBJECT:
439 return false; // found regular control flow back
440 case Instruction::THROW:
441 throw_seen = true;
442 break;
443 default:
444 break;
445 }
446 }
447 return throw_seen;
448 }
449
TryInline(HInvoke * invoke_instruction)450 bool HInliner::TryInline(HInvoke* invoke_instruction) {
451 MaybeRecordStat(stats_, MethodCompilationStat::kTryInline);
452
453 // Don't bother to move further if we know the method is unresolved or the invocation is
454 // polymorphic (invoke-{polymorphic,custom}).
455 if (invoke_instruction->IsInvokeUnresolved()) {
456 MaybeRecordStat(stats_, MethodCompilationStat::kNotInlinedUnresolved);
457 return false;
458 } else if (invoke_instruction->IsInvokePolymorphic()) {
459 MaybeRecordStat(stats_, MethodCompilationStat::kNotInlinedPolymorphic);
460 return false;
461 } else if (invoke_instruction->IsInvokeCustom()) {
462 MaybeRecordStat(stats_, MethodCompilationStat::kNotInlinedCustom);
463 return false;
464 }
465
466 ScopedObjectAccess soa(Thread::Current());
467 LOG_TRY() << invoke_instruction->GetMethodReference().PrettyMethod();
468
469 ArtMethod* resolved_method = invoke_instruction->GetResolvedMethod();
470 if (resolved_method == nullptr) {
471 DCHECK(invoke_instruction->IsInvokeStaticOrDirect());
472 DCHECK(invoke_instruction->AsInvokeStaticOrDirect()->IsStringInit());
473 LOG_FAIL_NO_STAT() << "Not inlining a String.<init> method";
474 return false;
475 }
476
477 ArtMethod* actual_method = nullptr;
478 ReferenceTypeInfo receiver_info = ReferenceTypeInfo::CreateInvalid();
479 if (invoke_instruction->GetInvokeType() == kStatic) {
480 actual_method = invoke_instruction->GetResolvedMethod();
481 } else {
482 HInstruction* receiver = invoke_instruction->InputAt(0);
483 while (receiver->IsNullCheck()) {
484 // Due to multiple levels of inlining within the same pass, it might be that
485 // null check does not have the reference type of the actual receiver.
486 receiver = receiver->InputAt(0);
487 }
488 receiver_info = receiver->GetReferenceTypeInfo();
489 if (!receiver_info.IsValid()) {
490 // We have to run the extra type propagation now as we are requiring the RTI.
491 DCHECK(run_extra_type_propagation_);
492 run_extra_type_propagation_ = false;
493 ReferenceTypePropagation rtp_fixup(graph_,
494 outer_compilation_unit_.GetDexCache(),
495 /* is_first_run= */ false);
496 rtp_fixup.Run();
497 receiver_info = receiver->GetReferenceTypeInfo();
498 }
499
500 DCHECK(receiver_info.IsValid()) << "Invalid RTI for " << receiver->DebugName();
501 if (invoke_instruction->IsInvokeStaticOrDirect()) {
502 actual_method = invoke_instruction->GetResolvedMethod();
503 } else {
504 actual_method = FindVirtualOrInterfaceTarget(invoke_instruction, receiver_info);
505 }
506 }
507
508 if (actual_method != nullptr) {
509 // Single target.
510 bool result = TryInlineAndReplace(invoke_instruction,
511 actual_method,
512 receiver_info,
513 /* do_rtp= */ true,
514 /* is_speculative= */ false);
515 if (result) {
516 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedInvokeVirtualOrInterface);
517 if (outermost_graph_ == graph_) {
518 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedLastInvokeVirtualOrInterface);
519 }
520 } else {
521 HInvoke* invoke_to_analyze = nullptr;
522 if (TryDevirtualize(invoke_instruction, actual_method, &invoke_to_analyze)) {
523 // Consider devirtualization as inlining.
524 result = true;
525 MaybeRecordStat(stats_, MethodCompilationStat::kDevirtualized);
526 } else {
527 invoke_to_analyze = invoke_instruction;
528 }
529 // Set always throws property for non-inlined method call with single target.
530 if (invoke_instruction->AlwaysThrows() || AlwaysThrows(actual_method)) {
531 invoke_to_analyze->SetAlwaysThrows(/* always_throws= */ true);
532 graph_->SetHasAlwaysThrowingInvokes(/* value= */ true);
533 }
534 }
535 return result;
536 }
537
538 if (graph_->IsCompilingBaseline()) {
539 LOG_FAIL_NO_STAT() << "Call to " << invoke_instruction->GetMethodReference().PrettyMethod()
540 << " not inlined because we are compiling baseline and we could not"
541 << " statically resolve the target";
542 // For baseline compilation, we will collect inline caches, so we should not
543 // try to inline using them.
544 outermost_graph_->SetUsefulOptimizing();
545 return false;
546 }
547
548 DCHECK(!invoke_instruction->IsInvokeStaticOrDirect());
549
550 // No try catch inlining allowed here, or recursively. For try catch inlining we are banking on
551 // the fact that we have a unique dex pc list. We cannot guarantee that for some TryInline methods
552 // e.g. `TryInlinePolymorphicCall`.
553 // TODO(solanes): Setting `try_catch_inlining_allowed_` to false here covers all cases from
554 // `TryInlineFromCHA` and from `TryInlineFromInlineCache` as well (e.g.
555 // `TryInlinePolymorphicCall`). Reassess to see if we can inline inline catch blocks in
556 // `TryInlineFromCHA`, `TryInlineMonomorphicCall` and `TryInlinePolymorphicCallToSameTarget`.
557
558 // We store the value to restore it since we will use the same HInliner instance for other inlinee
559 // candidates.
560 const bool previous_value = try_catch_inlining_allowed_;
561 try_catch_inlining_allowed_ = false;
562
563 if (TryInlineFromCHA(invoke_instruction)) {
564 try_catch_inlining_allowed_ = previous_value;
565 return true;
566 }
567
568 const bool result = TryInlineFromInlineCache(invoke_instruction);
569 try_catch_inlining_allowed_ = previous_value;
570 return result;
571 }
572
TryInlineFromCHA(HInvoke * invoke_instruction)573 bool HInliner::TryInlineFromCHA(HInvoke* invoke_instruction) {
574 ArtMethod* method = FindMethodFromCHA(invoke_instruction->GetResolvedMethod());
575 if (method == nullptr) {
576 return false;
577 }
578 LOG_NOTE() << "Try CHA-based inlining of " << method->PrettyMethod();
579
580 uint32_t dex_pc = invoke_instruction->GetDexPc();
581 HInstruction* cursor = invoke_instruction->GetPrevious();
582 HBasicBlock* bb_cursor = invoke_instruction->GetBlock();
583 Handle<mirror::Class> cls = graph_->GetHandleCache()->NewHandle(method->GetDeclaringClass());
584 if (!TryInlineAndReplace(invoke_instruction,
585 method,
586 ReferenceTypeInfo::Create(cls),
587 /* do_rtp= */ true,
588 /* is_speculative= */ true)) {
589 return false;
590 }
591 AddCHAGuard(invoke_instruction, dex_pc, cursor, bb_cursor);
592 // Add dependency due to devirtualization: we are assuming the resolved method
593 // has a single implementation.
594 outermost_graph_->AddCHASingleImplementationDependency(invoke_instruction->GetResolvedMethod());
595 MaybeRecordStat(stats_, MethodCompilationStat::kCHAInline);
596 return true;
597 }
598
UseOnlyPolymorphicInliningWithNoDeopt()599 bool HInliner::UseOnlyPolymorphicInliningWithNoDeopt() {
600 // If we are compiling AOT or OSR, pretend the call using inline caches is polymorphic and
601 // do not generate a deopt.
602 //
603 // For AOT:
604 // Generating a deopt does not ensure that we will actually capture the new types;
605 // and the danger is that we could be stuck in a loop with "forever" deoptimizations.
606 // Take for example the following scenario:
607 // - we capture the inline cache in one run
608 // - the next run, we deoptimize because we miss a type check, but the method
609 // never becomes hot again
610 // In this case, the inline cache will not be updated in the profile and the AOT code
611 // will keep deoptimizing.
612 // Another scenario is if we use profile compilation for a process which is not allowed
613 // to JIT (e.g. system server). If we deoptimize we will run interpreted code for the
614 // rest of the lifetime.
615 // TODO(calin):
616 // This is a compromise because we will most likely never update the inline cache
617 // in the profile (unless there's another reason to deopt). So we might be stuck with
618 // a sub-optimal inline cache.
619 // We could be smarter when capturing inline caches to mitigate this.
620 // (e.g. by having different thresholds for new and old methods).
621 //
622 // For OSR:
623 // We may come from the interpreter and it may have seen different receiver types.
624 return Runtime::Current()->IsAotCompiler() || outermost_graph_->IsCompilingOsr();
625 }
TryInlineFromInlineCache(HInvoke * invoke_instruction)626 bool HInliner::TryInlineFromInlineCache(HInvoke* invoke_instruction)
627 REQUIRES_SHARED(Locks::mutator_lock_) {
628 if (Runtime::Current()->IsAotCompiler() && !kUseAOTInlineCaches) {
629 return false;
630 }
631
632 StackHandleScope<InlineCache::kIndividualCacheSize> classes(Thread::Current());
633 // The Zygote JIT compiles based on a profile, so we shouldn't use runtime inline caches
634 // for it.
635 InlineCacheType inline_cache_type =
636 (Runtime::Current()->IsAotCompiler() || Runtime::Current()->IsZygote())
637 ? GetInlineCacheAOT(invoke_instruction, &classes)
638 : GetInlineCacheJIT(invoke_instruction, &classes);
639
640 switch (inline_cache_type) {
641 case kInlineCacheNoData: {
642 LOG_FAIL_NO_STAT()
643 << "No inline cache information for call to "
644 << invoke_instruction->GetMethodReference().PrettyMethod();
645 return false;
646 }
647
648 case kInlineCacheUninitialized: {
649 LOG_FAIL_NO_STAT()
650 << "Interface or virtual call to "
651 << invoke_instruction->GetMethodReference().PrettyMethod()
652 << " is not hit and not inlined";
653 return false;
654 }
655
656 case kInlineCacheMonomorphic: {
657 MaybeRecordStat(stats_, MethodCompilationStat::kMonomorphicCall);
658 if (UseOnlyPolymorphicInliningWithNoDeopt()) {
659 return TryInlinePolymorphicCall(invoke_instruction, classes);
660 } else {
661 return TryInlineMonomorphicCall(invoke_instruction, classes);
662 }
663 }
664
665 case kInlineCachePolymorphic: {
666 MaybeRecordStat(stats_, MethodCompilationStat::kPolymorphicCall);
667 return TryInlinePolymorphicCall(invoke_instruction, classes);
668 }
669
670 case kInlineCacheMegamorphic: {
671 LOG_FAIL_NO_STAT()
672 << "Interface or virtual call to "
673 << invoke_instruction->GetMethodReference().PrettyMethod()
674 << " is megamorphic and not inlined";
675 MaybeRecordStat(stats_, MethodCompilationStat::kMegamorphicCall);
676 return false;
677 }
678
679 case kInlineCacheMissingTypes: {
680 LOG_FAIL_NO_STAT()
681 << "Interface or virtual call to "
682 << invoke_instruction->GetMethodReference().PrettyMethod()
683 << " is missing types and not inlined";
684 return false;
685 }
686 }
687 }
688
GetInlineCacheJIT(HInvoke * invoke_instruction,StackHandleScope<InlineCache::kIndividualCacheSize> * classes)689 HInliner::InlineCacheType HInliner::GetInlineCacheJIT(
690 HInvoke* invoke_instruction,
691 /*out*/StackHandleScope<InlineCache::kIndividualCacheSize>* classes) {
692 DCHECK(codegen_->GetCompilerOptions().IsJitCompiler());
693
694 ArtMethod* caller = graph_->GetArtMethod();
695 // Under JIT, we should always know the caller.
696 DCHECK(caller != nullptr);
697
698 InlineCache* cache = nullptr;
699 // Start with the outer graph profiling info.
700 ProfilingInfo* profiling_info = outermost_graph_->GetProfilingInfo();
701 if (profiling_info != nullptr) {
702 if (depth_ == 0) {
703 cache = profiling_info->GetInlineCache(invoke_instruction->GetDexPc());
704 } else {
705 uint32_t dex_pc = ProfilingInfoBuilder::EncodeInlinedDexPc(
706 this, codegen_->GetCompilerOptions(), invoke_instruction);
707 if (dex_pc != kNoDexPc) {
708 cache = profiling_info->GetInlineCache(dex_pc);
709 }
710 }
711 }
712
713 if (cache == nullptr) {
714 // Check the current graph profiling info.
715 profiling_info = graph_->GetProfilingInfo();
716 if (profiling_info == nullptr) {
717 return kInlineCacheNoData;
718 }
719
720 cache = profiling_info->GetInlineCache(invoke_instruction->GetDexPc());
721 }
722
723 if (cache == nullptr) {
724 // Either we never hit this invoke and we never compiled the callee,
725 // or the method wasn't resolved when we performed baseline compilation.
726 // Bail for now.
727 return kInlineCacheNoData;
728 }
729 Runtime::Current()->GetJit()->GetCodeCache()->CopyInlineCacheInto(*cache, classes);
730 return GetInlineCacheType(*classes);
731 }
732
GetInlineCacheAOT(HInvoke * invoke_instruction,StackHandleScope<InlineCache::kIndividualCacheSize> * classes)733 HInliner::InlineCacheType HInliner::GetInlineCacheAOT(
734 HInvoke* invoke_instruction,
735 /*out*/StackHandleScope<InlineCache::kIndividualCacheSize>* classes) {
736 DCHECK_EQ(classes->Capacity(), InlineCache::kIndividualCacheSize);
737 DCHECK_EQ(classes->Size(), 0u);
738
739 const ProfileCompilationInfo* pci = codegen_->GetCompilerOptions().GetProfileCompilationInfo();
740 if (pci == nullptr) {
741 return kInlineCacheNoData;
742 }
743
744 ProfileCompilationInfo::MethodHotness hotness = pci->GetMethodHotness(MethodReference(
745 caller_compilation_unit_.GetDexFile(), caller_compilation_unit_.GetDexMethodIndex()));
746 if (!hotness.IsHot()) {
747 return kInlineCacheNoData; // no profile information for this invocation.
748 }
749
750 const ProfileCompilationInfo::InlineCacheMap* inline_caches = hotness.GetInlineCacheMap();
751 DCHECK(inline_caches != nullptr);
752
753 // Inlined inline caches are not supported in AOT, so we use the dex pc directly, and don't
754 // call `InlineCache::EncodeDexPc`.
755 // To support it, we would need to ensure `inline_max_code_units` remain the
756 // same between dex2oat and runtime, for example by adding it to the boot
757 // image oat header.
758 const auto it = inline_caches->find(invoke_instruction->GetDexPc());
759 if (it == inline_caches->end()) {
760 return kInlineCacheUninitialized;
761 }
762
763 const ProfileCompilationInfo::DexPcData& dex_pc_data = it->second;
764 if (dex_pc_data.is_missing_types) {
765 return kInlineCacheMissingTypes;
766 }
767 if (dex_pc_data.is_megamorphic) {
768 return kInlineCacheMegamorphic;
769 }
770 DCHECK_LE(dex_pc_data.classes.size(), InlineCache::kIndividualCacheSize);
771
772 // Walk over the class descriptors and look up the actual classes.
773 // If we cannot find a type we return kInlineCacheMissingTypes.
774 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
775 Thread* self = Thread::Current();
776 for (const dex::TypeIndex& type_index : dex_pc_data.classes) {
777 const DexFile* dex_file = caller_compilation_unit_.GetDexFile();
778 const char* descriptor = pci->GetTypeDescriptor(dex_file, type_index);
779 ObjPtr<mirror::Class> clazz =
780 class_linker->FindClass(self, descriptor, caller_compilation_unit_.GetClassLoader());
781 if (clazz == nullptr) {
782 self->ClearException(); // Clean up the exception left by type resolution.
783 VLOG(compiler) << "Could not find class from inline cache in AOT mode "
784 << invoke_instruction->GetMethodReference().PrettyMethod()
785 << " : "
786 << descriptor;
787 return kInlineCacheMissingTypes;
788 }
789 DCHECK_LT(classes->Size(), classes->Capacity());
790 classes->NewHandle(clazz);
791 }
792
793 return GetInlineCacheType(*classes);
794 }
795
BuildGetReceiverClass(ClassLinker * class_linker,HInstruction * receiver,uint32_t dex_pc) const796 HInstanceFieldGet* HInliner::BuildGetReceiverClass(ClassLinker* class_linker,
797 HInstruction* receiver,
798 uint32_t dex_pc) const {
799 ArtField* field = GetClassRoot<mirror::Object>(class_linker)->GetInstanceField(0);
800 DCHECK_EQ(std::string(field->GetName()), "shadow$_klass_");
801 HInstanceFieldGet* result = new (graph_->GetAllocator()) HInstanceFieldGet(
802 receiver,
803 field,
804 DataType::Type::kReference,
805 field->GetOffset(),
806 field->IsVolatile(),
807 field->GetDexFieldIndex(),
808 field->GetDeclaringClass()->GetDexClassDefIndex(),
809 *field->GetDexFile(),
810 dex_pc);
811 // The class of a field is effectively final, and does not have any memory dependencies.
812 result->SetSideEffects(SideEffects::None());
813 return result;
814 }
815
ResolveMethodFromInlineCache(Handle<mirror::Class> klass,HInvoke * invoke_instruction,PointerSize pointer_size)816 static ArtMethod* ResolveMethodFromInlineCache(Handle<mirror::Class> klass,
817 HInvoke* invoke_instruction,
818 PointerSize pointer_size)
819 REQUIRES_SHARED(Locks::mutator_lock_) {
820 ArtMethod* resolved_method = invoke_instruction->GetResolvedMethod();
821 if (Runtime::Current()->IsAotCompiler()) {
822 // We can get unrelated types when working with profiles (corruption,
823 // systme updates, or anyone can write to it). So first check if the class
824 // actually implements the declaring class of the method that is being
825 // called in bytecode.
826 // Note: the lookup methods used below require to have assignable types.
827 if (!resolved_method->GetDeclaringClass()->IsAssignableFrom(klass.Get())) {
828 return nullptr;
829 }
830
831 // Also check whether the type in the inline cache is an interface or an
832 // abstract class. We only expect concrete classes in inline caches, so this
833 // means the class was changed.
834 if (klass->IsAbstract() || klass->IsInterface()) {
835 return nullptr;
836 }
837 }
838
839 if (invoke_instruction->IsInvokeInterface()) {
840 resolved_method = klass->FindVirtualMethodForInterface(resolved_method, pointer_size);
841 } else {
842 DCHECK(invoke_instruction->IsInvokeVirtual());
843 resolved_method = klass->FindVirtualMethodForVirtual(resolved_method, pointer_size);
844 }
845 // Even if the class exists we can still not have the function the
846 // inline-cache targets if the profile is from far enough in the past/future.
847 // We need to allow this since we don't update boot-profiles very often. This
848 // can occur in boot-profiles with inline-caches.
849 DCHECK(Runtime::Current()->IsAotCompiler() || resolved_method != nullptr);
850 return resolved_method;
851 }
852
TryInlineMonomorphicCall(HInvoke * invoke_instruction,const StackHandleScope<InlineCache::kIndividualCacheSize> & classes)853 bool HInliner::TryInlineMonomorphicCall(
854 HInvoke* invoke_instruction,
855 const StackHandleScope<InlineCache::kIndividualCacheSize>& classes) {
856 DCHECK(invoke_instruction->IsInvokeVirtual() || invoke_instruction->IsInvokeInterface())
857 << invoke_instruction->DebugName();
858
859 dex::TypeIndex class_index = FindClassIndexIn(
860 GetMonomorphicType(classes), caller_compilation_unit_);
861 if (!class_index.IsValid()) {
862 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedDexCacheInaccessibleToCaller)
863 << "Call to " << ArtMethod::PrettyMethod(invoke_instruction->GetResolvedMethod())
864 << " from inline cache is not inlined because its class is not"
865 << " accessible to the caller";
866 return false;
867 }
868
869 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
870 PointerSize pointer_size = class_linker->GetImagePointerSize();
871 Handle<mirror::Class> monomorphic_type =
872 graph_->GetHandleCache()->NewHandle(GetMonomorphicType(classes));
873 ArtMethod* resolved_method = ResolveMethodFromInlineCache(
874 monomorphic_type, invoke_instruction, pointer_size);
875 if (resolved_method == nullptr) {
876 // Bogus AOT profile, bail.
877 DCHECK(Runtime::Current()->IsAotCompiler());
878 return false;
879 }
880
881 LOG_NOTE() << "Try inline monomorphic call to " << resolved_method->PrettyMethod();
882 HInstruction* receiver = invoke_instruction->InputAt(0);
883 HInstruction* cursor = invoke_instruction->GetPrevious();
884 HBasicBlock* bb_cursor = invoke_instruction->GetBlock();
885 if (!TryInlineAndReplace(invoke_instruction,
886 resolved_method,
887 ReferenceTypeInfo::Create(monomorphic_type, /* is_exact= */ true),
888 /* do_rtp= */ false,
889 /* is_speculative= */ true)) {
890 return false;
891 }
892
893 // We successfully inlined, now add a guard.
894 AddTypeGuard(receiver,
895 cursor,
896 bb_cursor,
897 class_index,
898 monomorphic_type,
899 invoke_instruction,
900 /* with_deoptimization= */ true);
901
902 // Lazily run type propagation to get the guard typed, and eventually propagate the
903 // type of the receiver.
904 run_extra_type_propagation_ = true;
905
906 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedMonomorphicCall);
907 return true;
908 }
909
AddCHAGuard(HInstruction * invoke_instruction,uint32_t dex_pc,HInstruction * cursor,HBasicBlock * bb_cursor)910 void HInliner::AddCHAGuard(HInstruction* invoke_instruction,
911 uint32_t dex_pc,
912 HInstruction* cursor,
913 HBasicBlock* bb_cursor) {
914 HShouldDeoptimizeFlag* deopt_flag = new (graph_->GetAllocator())
915 HShouldDeoptimizeFlag(graph_->GetAllocator(), dex_pc);
916 // ShouldDeoptimizeFlag is used to perform a deoptimization because of a CHA
917 // invalidation or for debugging reasons. It is OK to just check for non-zero
918 // value here instead of the specific CHA value. When a debugging deopt is
919 // requested we deoptimize before we execute any code and hence we shouldn't
920 // see that case here.
921 HInstruction* compare = new (graph_->GetAllocator()) HNotEqual(
922 deopt_flag, graph_->GetIntConstant(0, dex_pc));
923 HInstruction* deopt = new (graph_->GetAllocator()) HDeoptimize(
924 graph_->GetAllocator(), compare, DeoptimizationKind::kCHA, dex_pc);
925
926 if (cursor != nullptr) {
927 bb_cursor->InsertInstructionAfter(deopt_flag, cursor);
928 } else {
929 bb_cursor->InsertInstructionBefore(deopt_flag, bb_cursor->GetFirstInstruction());
930 }
931 bb_cursor->InsertInstructionAfter(compare, deopt_flag);
932 bb_cursor->InsertInstructionAfter(deopt, compare);
933
934 // Add receiver as input to aid CHA guard optimization later.
935 deopt_flag->AddInput(invoke_instruction->InputAt(0));
936 DCHECK_EQ(deopt_flag->InputCount(), 1u);
937 deopt->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
938 outermost_graph_->IncrementNumberOfCHAGuards();
939 }
940
AddTypeGuard(HInstruction * receiver,HInstruction * cursor,HBasicBlock * bb_cursor,dex::TypeIndex class_index,Handle<mirror::Class> klass,HInstruction * invoke_instruction,bool with_deoptimization)941 HInstruction* HInliner::AddTypeGuard(HInstruction* receiver,
942 HInstruction* cursor,
943 HBasicBlock* bb_cursor,
944 dex::TypeIndex class_index,
945 Handle<mirror::Class> klass,
946 HInstruction* invoke_instruction,
947 bool with_deoptimization) {
948 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
949 HInstanceFieldGet* receiver_class = BuildGetReceiverClass(
950 class_linker, receiver, invoke_instruction->GetDexPc());
951 if (cursor != nullptr) {
952 bb_cursor->InsertInstructionAfter(receiver_class, cursor);
953 } else {
954 bb_cursor->InsertInstructionBefore(receiver_class, bb_cursor->GetFirstInstruction());
955 }
956
957 const DexFile& caller_dex_file = *caller_compilation_unit_.GetDexFile();
958 bool is_referrer;
959 ArtMethod* outermost_art_method = outermost_graph_->GetArtMethod();
960 if (outermost_art_method == nullptr) {
961 DCHECK(Runtime::Current()->IsAotCompiler());
962 // We are in AOT mode and we don't have an ART method to determine
963 // if the inlined method belongs to the referrer. Assume it doesn't.
964 is_referrer = false;
965 } else {
966 is_referrer = klass.Get() == outermost_art_method->GetDeclaringClass();
967 }
968
969 // Note that we will just compare the classes, so we don't need Java semantics access checks.
970 // Note that the type index and the dex file are relative to the method this type guard is
971 // inlined into.
972 HLoadClass* load_class = new (graph_->GetAllocator()) HLoadClass(graph_->GetCurrentMethod(),
973 class_index,
974 caller_dex_file,
975 klass,
976 is_referrer,
977 invoke_instruction->GetDexPc(),
978 /* needs_access_check= */ false);
979 HLoadClass::LoadKind kind = HSharpening::ComputeLoadClassKind(
980 load_class, codegen_, caller_compilation_unit_);
981 DCHECK(kind != HLoadClass::LoadKind::kInvalid)
982 << "We should always be able to reference a class for inline caches";
983 // Load kind must be set before inserting the instruction into the graph.
984 load_class->SetLoadKind(kind);
985 bb_cursor->InsertInstructionAfter(load_class, receiver_class);
986 // In AOT mode, we will most likely load the class from BSS, which will involve a call
987 // to the runtime. In this case, the load instruction will need an environment so copy
988 // it from the invoke instruction.
989 if (load_class->NeedsEnvironment()) {
990 DCHECK(Runtime::Current()->IsAotCompiler());
991 load_class->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
992 }
993
994 HNotEqual* compare = new (graph_->GetAllocator()) HNotEqual(load_class, receiver_class);
995 bb_cursor->InsertInstructionAfter(compare, load_class);
996 if (with_deoptimization) {
997 HDeoptimize* deoptimize = new (graph_->GetAllocator()) HDeoptimize(
998 graph_->GetAllocator(),
999 compare,
1000 receiver,
1001 Runtime::Current()->IsAotCompiler()
1002 ? DeoptimizationKind::kAotInlineCache
1003 : DeoptimizationKind::kJitInlineCache,
1004 invoke_instruction->GetDexPc());
1005 bb_cursor->InsertInstructionAfter(deoptimize, compare);
1006 deoptimize->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
1007 DCHECK_EQ(invoke_instruction->InputAt(0), receiver);
1008 receiver->ReplaceUsesDominatedBy(deoptimize, deoptimize);
1009 deoptimize->SetReferenceTypeInfo(receiver->GetReferenceTypeInfo());
1010 }
1011 return compare;
1012 }
1013
MaybeReplaceAndRemove(HInstruction * new_instruction,HInstruction * old_instruction)1014 static void MaybeReplaceAndRemove(HInstruction* new_instruction, HInstruction* old_instruction) {
1015 DCHECK(new_instruction != old_instruction);
1016 if (new_instruction != nullptr) {
1017 old_instruction->ReplaceWith(new_instruction);
1018 }
1019 old_instruction->GetBlock()->RemoveInstruction(old_instruction);
1020 }
1021
TryInlinePolymorphicCall(HInvoke * invoke_instruction,const StackHandleScope<InlineCache::kIndividualCacheSize> & classes)1022 bool HInliner::TryInlinePolymorphicCall(
1023 HInvoke* invoke_instruction,
1024 const StackHandleScope<InlineCache::kIndividualCacheSize>& classes) {
1025 DCHECK(invoke_instruction->IsInvokeVirtual() || invoke_instruction->IsInvokeInterface())
1026 << invoke_instruction->DebugName();
1027
1028 if (TryInlinePolymorphicCallToSameTarget(invoke_instruction, classes)) {
1029 return true;
1030 }
1031
1032 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
1033 PointerSize pointer_size = class_linker->GetImagePointerSize();
1034
1035 bool all_targets_inlined = true;
1036 bool one_target_inlined = false;
1037 DCHECK_EQ(classes.Capacity(), InlineCache::kIndividualCacheSize);
1038 uint8_t number_of_types = classes.Size();
1039 for (size_t i = 0; i != number_of_types; ++i) {
1040 DCHECK(classes.GetReference(i) != nullptr);
1041 Handle<mirror::Class> handle =
1042 graph_->GetHandleCache()->NewHandle(classes.GetReference(i)->AsClass());
1043 ArtMethod* method = ResolveMethodFromInlineCache(handle, invoke_instruction, pointer_size);
1044 if (method == nullptr) {
1045 DCHECK(Runtime::Current()->IsAotCompiler());
1046 // AOT profile is bogus. This loop expects to iterate over all entries,
1047 // so just just continue.
1048 all_targets_inlined = false;
1049 continue;
1050 }
1051
1052 HInstruction* receiver = invoke_instruction->InputAt(0);
1053 HInstruction* cursor = invoke_instruction->GetPrevious();
1054 HBasicBlock* bb_cursor = invoke_instruction->GetBlock();
1055
1056 dex::TypeIndex class_index = FindClassIndexIn(handle.Get(), caller_compilation_unit_);
1057 HInstruction* return_replacement = nullptr;
1058
1059 // In monomorphic cases when UseOnlyPolymorphicInliningWithNoDeopt() is true, we call
1060 // `TryInlinePolymorphicCall` even though we are monomorphic.
1061 const bool actually_monomorphic = number_of_types == 1;
1062 DCHECK_IMPLIES(actually_monomorphic, UseOnlyPolymorphicInliningWithNoDeopt());
1063
1064 // We only want to limit recursive polymorphic cases, not monomorphic ones.
1065 const bool too_many_polymorphic_recursive_calls =
1066 !actually_monomorphic &&
1067 CountRecursiveCallsOf(method) > kMaximumNumberOfPolymorphicRecursiveCalls;
1068 if (too_many_polymorphic_recursive_calls) {
1069 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedPolymorphicRecursiveBudget)
1070 << "Method " << method->PrettyMethod()
1071 << " is not inlined because it has reached its polymorphic recursive call budget.";
1072 } else if (class_index.IsValid()) {
1073 LOG_NOTE() << "Try inline polymorphic call to " << method->PrettyMethod();
1074 }
1075
1076 if (too_many_polymorphic_recursive_calls ||
1077 !class_index.IsValid() ||
1078 !TryBuildAndInline(invoke_instruction,
1079 method,
1080 ReferenceTypeInfo::Create(handle, /* is_exact= */ true),
1081 &return_replacement,
1082 /* is_speculative= */ true)) {
1083 all_targets_inlined = false;
1084 } else {
1085 one_target_inlined = true;
1086
1087 LOG_SUCCESS() << "Polymorphic call to "
1088 << invoke_instruction->GetMethodReference().PrettyMethod()
1089 << " has inlined " << ArtMethod::PrettyMethod(method);
1090
1091 // If we have inlined all targets before, and this receiver is the last seen,
1092 // we deoptimize instead of keeping the original invoke instruction.
1093 bool deoptimize = !UseOnlyPolymorphicInliningWithNoDeopt() &&
1094 all_targets_inlined &&
1095 (i + 1 == number_of_types);
1096
1097 HInstruction* compare = AddTypeGuard(receiver,
1098 cursor,
1099 bb_cursor,
1100 class_index,
1101 handle,
1102 invoke_instruction,
1103 deoptimize);
1104 if (deoptimize) {
1105 MaybeReplaceAndRemove(return_replacement, invoke_instruction);
1106 } else {
1107 CreateDiamondPatternForPolymorphicInline(compare, return_replacement, invoke_instruction);
1108 }
1109 }
1110 }
1111
1112 if (!one_target_inlined) {
1113 LOG_FAIL_NO_STAT()
1114 << "Call to " << invoke_instruction->GetMethodReference().PrettyMethod()
1115 << " from inline cache is not inlined because none"
1116 << " of its targets could be inlined";
1117 return false;
1118 }
1119
1120 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedPolymorphicCall);
1121
1122 // Lazily run type propagation to get the guards typed.
1123 run_extra_type_propagation_ = true;
1124 return true;
1125 }
1126
CreateDiamondPatternForPolymorphicInline(HInstruction * compare,HInstruction * return_replacement,HInstruction * invoke_instruction)1127 void HInliner::CreateDiamondPatternForPolymorphicInline(HInstruction* compare,
1128 HInstruction* return_replacement,
1129 HInstruction* invoke_instruction) {
1130 uint32_t dex_pc = invoke_instruction->GetDexPc();
1131 HBasicBlock* cursor_block = compare->GetBlock();
1132 HBasicBlock* original_invoke_block = invoke_instruction->GetBlock();
1133 ArenaAllocator* allocator = graph_->GetAllocator();
1134
1135 // Spit the block after the compare: `cursor_block` will now be the start of the diamond,
1136 // and the returned block is the start of the then branch (that could contain multiple blocks).
1137 HBasicBlock* then = cursor_block->SplitAfterForInlining(compare);
1138
1139 // Split the block containing the invoke before and after the invoke. The returned block
1140 // of the split before will contain the invoke and will be the otherwise branch of
1141 // the diamond. The returned block of the split after will be the merge block
1142 // of the diamond.
1143 HBasicBlock* end_then = invoke_instruction->GetBlock();
1144 HBasicBlock* otherwise = end_then->SplitBeforeForInlining(invoke_instruction);
1145 HBasicBlock* merge = otherwise->SplitAfterForInlining(invoke_instruction);
1146
1147 // If the methods we are inlining return a value, we create a phi in the merge block
1148 // that will have the `invoke_instruction and the `return_replacement` as inputs.
1149 if (return_replacement != nullptr) {
1150 HPhi* phi = new (allocator) HPhi(
1151 allocator, kNoRegNumber, 0, HPhi::ToPhiType(invoke_instruction->GetType()), dex_pc);
1152 merge->AddPhi(phi);
1153 invoke_instruction->ReplaceWith(phi);
1154 phi->AddInput(return_replacement);
1155 phi->AddInput(invoke_instruction);
1156 }
1157
1158 // Add the control flow instructions.
1159 otherwise->AddInstruction(new (allocator) HGoto(dex_pc));
1160 end_then->AddInstruction(new (allocator) HGoto(dex_pc));
1161 cursor_block->AddInstruction(new (allocator) HIf(compare, dex_pc));
1162
1163 // Add the newly created blocks to the graph.
1164 graph_->AddBlock(then);
1165 graph_->AddBlock(otherwise);
1166 graph_->AddBlock(merge);
1167
1168 // Set up successor (and implictly predecessor) relations.
1169 cursor_block->AddSuccessor(otherwise);
1170 cursor_block->AddSuccessor(then);
1171 end_then->AddSuccessor(merge);
1172 otherwise->AddSuccessor(merge);
1173
1174 // Set up dominance information.
1175 then->SetDominator(cursor_block);
1176 cursor_block->AddDominatedBlock(then);
1177 otherwise->SetDominator(cursor_block);
1178 cursor_block->AddDominatedBlock(otherwise);
1179 merge->SetDominator(cursor_block);
1180 cursor_block->AddDominatedBlock(merge);
1181
1182 // Update the revert post order.
1183 size_t index = IndexOfElement(graph_->reverse_post_order_, cursor_block);
1184 MakeRoomFor(&graph_->reverse_post_order_, 1, index);
1185 graph_->reverse_post_order_[++index] = then;
1186 index = IndexOfElement(graph_->reverse_post_order_, end_then);
1187 MakeRoomFor(&graph_->reverse_post_order_, 2, index);
1188 graph_->reverse_post_order_[++index] = otherwise;
1189 graph_->reverse_post_order_[++index] = merge;
1190
1191
1192 graph_->UpdateLoopAndTryInformationOfNewBlock(
1193 then, original_invoke_block, /* replace_if_back_edge= */ false);
1194 graph_->UpdateLoopAndTryInformationOfNewBlock(
1195 otherwise, original_invoke_block, /* replace_if_back_edge= */ false);
1196
1197 // In case the original invoke location was a back edge, we need to update
1198 // the loop to now have the merge block as a back edge.
1199 graph_->UpdateLoopAndTryInformationOfNewBlock(
1200 merge, original_invoke_block, /* replace_if_back_edge= */ true);
1201 }
1202
TryInlinePolymorphicCallToSameTarget(HInvoke * invoke_instruction,const StackHandleScope<InlineCache::kIndividualCacheSize> & classes)1203 bool HInliner::TryInlinePolymorphicCallToSameTarget(
1204 HInvoke* invoke_instruction,
1205 const StackHandleScope<InlineCache::kIndividualCacheSize>& classes) {
1206 // This optimization only works under JIT for now.
1207 if (!codegen_->GetCompilerOptions().IsJitCompiler()) {
1208 return false;
1209 }
1210
1211 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
1212 PointerSize pointer_size = class_linker->GetImagePointerSize();
1213
1214 ArtMethod* actual_method = nullptr;
1215 size_t method_index = invoke_instruction->IsInvokeVirtual()
1216 ? invoke_instruction->AsInvokeVirtual()->GetVTableIndex()
1217 : invoke_instruction->AsInvokeInterface()->GetImtIndex();
1218
1219 // Check whether we are actually calling the same method among
1220 // the different types seen.
1221 DCHECK_EQ(classes.Capacity(), InlineCache::kIndividualCacheSize);
1222 uint8_t number_of_types = classes.Size();
1223 for (size_t i = 0; i != number_of_types; ++i) {
1224 DCHECK(classes.GetReference(i) != nullptr);
1225 ArtMethod* new_method = nullptr;
1226 if (invoke_instruction->IsInvokeInterface()) {
1227 new_method = classes.GetReference(i)->AsClass()->GetImt(pointer_size)->Get(
1228 method_index, pointer_size);
1229 if (new_method->IsRuntimeMethod()) {
1230 // Bail out as soon as we see a conflict trampoline in one of the target's
1231 // interface table.
1232 return false;
1233 }
1234 } else {
1235 DCHECK(invoke_instruction->IsInvokeVirtual());
1236 new_method =
1237 classes.GetReference(i)->AsClass()->GetEmbeddedVTableEntry(method_index, pointer_size);
1238 }
1239 DCHECK(new_method != nullptr);
1240 if (actual_method == nullptr) {
1241 actual_method = new_method;
1242 } else if (actual_method != new_method) {
1243 // Different methods, bailout.
1244 return false;
1245 }
1246 }
1247
1248 HInstruction* receiver = invoke_instruction->InputAt(0);
1249 HInstruction* cursor = invoke_instruction->GetPrevious();
1250 HBasicBlock* bb_cursor = invoke_instruction->GetBlock();
1251
1252 HInstruction* return_replacement = nullptr;
1253 Handle<mirror::Class> cls =
1254 graph_->GetHandleCache()->NewHandle(actual_method->GetDeclaringClass());
1255 if (!TryBuildAndInline(invoke_instruction,
1256 actual_method,
1257 ReferenceTypeInfo::Create(cls),
1258 &return_replacement,
1259 /* is_speculative= */ true)) {
1260 return false;
1261 }
1262
1263 // We successfully inlined, now add a guard.
1264 HInstanceFieldGet* receiver_class = BuildGetReceiverClass(
1265 class_linker, receiver, invoke_instruction->GetDexPc());
1266
1267 DataType::Type type = Is64BitInstructionSet(graph_->GetInstructionSet())
1268 ? DataType::Type::kInt64
1269 : DataType::Type::kInt32;
1270 HClassTableGet* class_table_get = new (graph_->GetAllocator()) HClassTableGet(
1271 receiver_class,
1272 type,
1273 invoke_instruction->IsInvokeVirtual() ? HClassTableGet::TableKind::kVTable
1274 : HClassTableGet::TableKind::kIMTable,
1275 method_index,
1276 invoke_instruction->GetDexPc());
1277
1278 HConstant* constant;
1279 if (type == DataType::Type::kInt64) {
1280 constant = graph_->GetLongConstant(
1281 reinterpret_cast<intptr_t>(actual_method), invoke_instruction->GetDexPc());
1282 } else {
1283 constant = graph_->GetIntConstant(
1284 reinterpret_cast<intptr_t>(actual_method), invoke_instruction->GetDexPc());
1285 }
1286
1287 HNotEqual* compare = new (graph_->GetAllocator()) HNotEqual(class_table_get, constant);
1288 if (cursor != nullptr) {
1289 bb_cursor->InsertInstructionAfter(receiver_class, cursor);
1290 } else {
1291 bb_cursor->InsertInstructionBefore(receiver_class, bb_cursor->GetFirstInstruction());
1292 }
1293 bb_cursor->InsertInstructionAfter(class_table_get, receiver_class);
1294 bb_cursor->InsertInstructionAfter(compare, class_table_get);
1295
1296 if (outermost_graph_->IsCompilingOsr()) {
1297 CreateDiamondPatternForPolymorphicInline(compare, return_replacement, invoke_instruction);
1298 } else {
1299 HDeoptimize* deoptimize = new (graph_->GetAllocator()) HDeoptimize(
1300 graph_->GetAllocator(),
1301 compare,
1302 receiver,
1303 DeoptimizationKind::kJitSameTarget,
1304 invoke_instruction->GetDexPc());
1305 bb_cursor->InsertInstructionAfter(deoptimize, compare);
1306 deoptimize->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
1307 MaybeReplaceAndRemove(return_replacement, invoke_instruction);
1308 receiver->ReplaceUsesDominatedBy(deoptimize, deoptimize);
1309 deoptimize->SetReferenceTypeInfo(receiver->GetReferenceTypeInfo());
1310 }
1311
1312 // Lazily run type propagation to get the guard typed.
1313 run_extra_type_propagation_ = true;
1314 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedPolymorphicCall);
1315
1316 LOG_SUCCESS() << "Inlined same polymorphic target " << actual_method->PrettyMethod();
1317 return true;
1318 }
1319
MaybeRunReferenceTypePropagation(HInstruction * replacement,HInvoke * invoke_instruction)1320 void HInliner::MaybeRunReferenceTypePropagation(HInstruction* replacement,
1321 HInvoke* invoke_instruction) {
1322 if (ReturnTypeMoreSpecific(replacement, invoke_instruction)) {
1323 // Actual return value has a more specific type than the method's declared
1324 // return type. Run RTP again on the outer graph to propagate it.
1325 ReferenceTypePropagation(graph_,
1326 outer_compilation_unit_.GetDexCache(),
1327 /* is_first_run= */ false).Run();
1328 }
1329 }
1330
TryDevirtualize(HInvoke * invoke_instruction,ArtMethod * method,HInvoke ** replacement)1331 bool HInliner::TryDevirtualize(HInvoke* invoke_instruction,
1332 ArtMethod* method,
1333 HInvoke** replacement) {
1334 DCHECK(invoke_instruction != *replacement);
1335 if (!invoke_instruction->IsInvokeInterface() && !invoke_instruction->IsInvokeVirtual()) {
1336 return false;
1337 }
1338
1339 // Don't try to devirtualize intrinsics as it breaks pattern matching from later phases.
1340 // TODO(solanes): This `if` could be removed if we update optimizations like
1341 // TryReplaceStringBuilderAppend.
1342 if (invoke_instruction->IsIntrinsic()) {
1343 return false;
1344 }
1345
1346 // Don't bother trying to call directly a default conflict method. It
1347 // doesn't have a proper MethodReference, but also `GetCanonicalMethod`
1348 // will return an actual default implementation.
1349 if (method->IsDefaultConflicting()) {
1350 return false;
1351 }
1352 DCHECK(!method->IsProxyMethod());
1353 ClassLinker* cl = Runtime::Current()->GetClassLinker();
1354 PointerSize pointer_size = cl->GetImagePointerSize();
1355 // The sharpening logic assumes the caller isn't passing a copied method.
1356 method = method->GetCanonicalMethod(pointer_size);
1357 uint32_t dex_method_index = FindMethodIndexIn(
1358 method,
1359 *invoke_instruction->GetMethodReference().dex_file,
1360 invoke_instruction->GetMethodReference().index);
1361 if (dex_method_index == dex::kDexNoIndex) {
1362 return false;
1363 }
1364 HInvokeStaticOrDirect::DispatchInfo dispatch_info =
1365 HSharpening::SharpenLoadMethod(method,
1366 /* has_method_id= */ true,
1367 /* for_interface_call= */ false,
1368 codegen_);
1369 DCHECK_NE(dispatch_info.code_ptr_location, CodePtrLocation::kCallCriticalNative);
1370 if (dispatch_info.method_load_kind == MethodLoadKind::kRuntimeCall) {
1371 // If sharpening returns that we need to load the method at runtime, keep
1372 // the virtual/interface call which will be faster.
1373 // Also, the entrypoints for runtime calls do not handle devirtualized
1374 // calls.
1375 return false;
1376 }
1377
1378 HInvokeStaticOrDirect* new_invoke = new (graph_->GetAllocator()) HInvokeStaticOrDirect(
1379 graph_->GetAllocator(),
1380 invoke_instruction->GetNumberOfArguments(),
1381 invoke_instruction->GetType(),
1382 invoke_instruction->GetDexPc(),
1383 MethodReference(invoke_instruction->GetMethodReference().dex_file, dex_method_index),
1384 method,
1385 dispatch_info,
1386 kDirect,
1387 MethodReference(method->GetDexFile(), method->GetDexMethodIndex()),
1388 HInvokeStaticOrDirect::ClinitCheckRequirement::kNone,
1389 !graph_->IsDebuggable());
1390 HInputsRef inputs = invoke_instruction->GetInputs();
1391 DCHECK_EQ(inputs.size(), invoke_instruction->GetNumberOfArguments());
1392 for (size_t index = 0; index != inputs.size(); ++index) {
1393 new_invoke->SetArgumentAt(index, inputs[index]);
1394 }
1395 if (HInvokeStaticOrDirect::NeedsCurrentMethodInput(dispatch_info)) {
1396 new_invoke->SetRawInputAt(new_invoke->GetCurrentMethodIndexUnchecked(),
1397 graph_->GetCurrentMethod());
1398 }
1399 invoke_instruction->GetBlock()->InsertInstructionBefore(new_invoke, invoke_instruction);
1400 new_invoke->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
1401 if (invoke_instruction->GetType() == DataType::Type::kReference) {
1402 new_invoke->SetReferenceTypeInfoIfValid(invoke_instruction->GetReferenceTypeInfo());
1403 }
1404 *replacement = new_invoke;
1405
1406 MaybeReplaceAndRemove(*replacement, invoke_instruction);
1407 // No need to call MaybeRunReferenceTypePropagation, as we know the return type
1408 // cannot be more specific.
1409 DCHECK(!ReturnTypeMoreSpecific(*replacement, invoke_instruction));
1410 return true;
1411 }
1412
1413
TryInlineAndReplace(HInvoke * invoke_instruction,ArtMethod * method,ReferenceTypeInfo receiver_type,bool do_rtp,bool is_speculative)1414 bool HInliner::TryInlineAndReplace(HInvoke* invoke_instruction,
1415 ArtMethod* method,
1416 ReferenceTypeInfo receiver_type,
1417 bool do_rtp,
1418 bool is_speculative) {
1419 DCHECK(!codegen_->IsImplementedIntrinsic(invoke_instruction));
1420 HInstruction* return_replacement = nullptr;
1421
1422 if (!TryBuildAndInline(
1423 invoke_instruction, method, receiver_type, &return_replacement, is_speculative)) {
1424 return false;
1425 }
1426
1427 MaybeReplaceAndRemove(return_replacement, invoke_instruction);
1428 FixUpReturnReferenceType(method, return_replacement);
1429 if (do_rtp) {
1430 MaybeRunReferenceTypePropagation(return_replacement, invoke_instruction);
1431 }
1432 return true;
1433 }
1434
CountRecursiveCallsOf(ArtMethod * method) const1435 size_t HInliner::CountRecursiveCallsOf(ArtMethod* method) const {
1436 const HInliner* current = this;
1437 size_t count = 0;
1438 do {
1439 if (current->graph_->GetArtMethod() == method) {
1440 ++count;
1441 }
1442 current = current->parent_;
1443 } while (current != nullptr);
1444 return count;
1445 }
1446
MayInline(const CompilerOptions & compiler_options,const DexFile & inlined_from,const DexFile & inlined_into)1447 static inline bool MayInline(const CompilerOptions& compiler_options,
1448 const DexFile& inlined_from,
1449 const DexFile& inlined_into) {
1450 // We're not allowed to inline across dex files if we're the no-inline-from dex file.
1451 if (!IsSameDexFile(inlined_from, inlined_into) &&
1452 ContainsElement(compiler_options.GetNoInlineFromDexFile(), &inlined_from)) {
1453 return false;
1454 }
1455
1456 return true;
1457 }
1458
1459 // Returns whether inlining is allowed based on ART semantics.
IsInliningAllowed(ArtMethod * method,const CodeItemDataAccessor & accessor) const1460 bool HInliner::IsInliningAllowed(ArtMethod* method, const CodeItemDataAccessor& accessor) const {
1461 if (!accessor.HasCodeItem()) {
1462 LOG_FAIL_NO_STAT()
1463 << "Method " << method->PrettyMethod() << " is not inlined because it is native";
1464 return false;
1465 }
1466
1467 if (!method->IsCompilable()) {
1468 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedNotCompilable)
1469 << "Method " << method->PrettyMethod()
1470 << " has soft failures un-handled by the compiler, so it cannot be inlined";
1471 return false;
1472 }
1473
1474 if (!IsMethodVerified(method)) {
1475 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedNotVerified)
1476 << "Method " << method->PrettyMethod()
1477 << " couldn't be verified, so it cannot be inlined";
1478 return false;
1479 }
1480
1481 if (annotations::MethodIsNeverInline(*method->GetDexFile(),
1482 method->GetClassDef(),
1483 method->GetDexMethodIndex())) {
1484 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedNeverInlineAnnotation)
1485 << "Method " << method->PrettyMethod()
1486 << " has the @NeverInline annotation so it won't be inlined";
1487 return false;
1488 }
1489
1490 return true;
1491 }
1492
1493 // Returns whether ART supports inlining this method.
1494 //
1495 // Some methods are not supported because they have features for which inlining
1496 // is not implemented. For example, we do not currently support inlining throw
1497 // instructions into a try block.
IsInliningSupported(const HInvoke * invoke_instruction,ArtMethod * method,const CodeItemDataAccessor & accessor) const1498 bool HInliner::IsInliningSupported(const HInvoke* invoke_instruction,
1499 ArtMethod* method,
1500 const CodeItemDataAccessor& accessor) const {
1501 if (method->IsProxyMethod()) {
1502 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedProxy)
1503 << "Method " << method->PrettyMethod()
1504 << " is not inlined because of unimplemented inline support for proxy methods.";
1505 return false;
1506 }
1507
1508 if (accessor.TriesSize() != 0) {
1509 if (!kInlineTryCatches) {
1510 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedTryCatchDisabled)
1511 << "Method " << method->PrettyMethod()
1512 << " is not inlined because inlining try catches is disabled globally";
1513 return false;
1514 }
1515 const bool disallowed_try_catch_inlining =
1516 // Direct parent is a try block.
1517 invoke_instruction->GetBlock()->IsTryBlock() ||
1518 // Indirect parent disallows try catch inlining.
1519 !try_catch_inlining_allowed_;
1520 if (disallowed_try_catch_inlining) {
1521 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedTryCatchCallee)
1522 << "Method " << method->PrettyMethod()
1523 << " is not inlined because it has a try catch and we are not supporting it for this"
1524 << " particular call. This is could be because e.g. it would be inlined inside another"
1525 << " try block, we arrived here from TryInlinePolymorphicCall, etc.";
1526 return false;
1527 }
1528 }
1529
1530 if (invoke_instruction->IsInvokeStaticOrDirect() &&
1531 invoke_instruction->AsInvokeStaticOrDirect()->IsStaticWithImplicitClinitCheck()) {
1532 // Case of a static method that cannot be inlined because it implicitly
1533 // requires an initialization check of its declaring class.
1534 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedDexCacheClinitCheck)
1535 << "Method " << method->PrettyMethod()
1536 << " is not inlined because it is static and requires a clinit"
1537 << " check that cannot be emitted due to Dex cache limitations";
1538 return false;
1539 }
1540
1541 return true;
1542 }
1543
IsInliningEncouraged(const HInvoke * invoke_instruction,ArtMethod * method,const CodeItemDataAccessor & accessor) const1544 bool HInliner::IsInliningEncouraged(const HInvoke* invoke_instruction,
1545 ArtMethod* method,
1546 const CodeItemDataAccessor& accessor) const {
1547 if (CountRecursiveCallsOf(method) > kMaximumNumberOfRecursiveCalls) {
1548 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedRecursiveBudget)
1549 << "Method "
1550 << method->PrettyMethod()
1551 << " is not inlined because it has reached its recursive call budget.";
1552 return false;
1553 }
1554
1555 size_t inline_max_code_units = codegen_->GetCompilerOptions().GetInlineMaxCodeUnits();
1556 if (accessor.InsnsSizeInCodeUnits() > inline_max_code_units) {
1557 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedCodeItem)
1558 << "Method " << method->PrettyMethod()
1559 << " is not inlined because its code item is too big: "
1560 << accessor.InsnsSizeInCodeUnits()
1561 << " > "
1562 << inline_max_code_units;
1563 return false;
1564 }
1565
1566 if (graph_->IsCompilingBaseline() &&
1567 accessor.InsnsSizeInCodeUnits() > CompilerOptions::kBaselineInlineMaxCodeUnits) {
1568 LOG_FAIL_NO_STAT() << "Reached baseline maximum code unit for inlining "
1569 << method->PrettyMethod();
1570 outermost_graph_->SetUsefulOptimizing();
1571 return false;
1572 }
1573
1574 if (invoke_instruction->GetBlock()->GetLastInstruction()->IsThrow()) {
1575 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedEndsWithThrow)
1576 << "Method " << method->PrettyMethod()
1577 << " is not inlined because its block ends with a throw";
1578 return false;
1579 }
1580
1581 return true;
1582 }
1583
TryBuildAndInline(HInvoke * invoke_instruction,ArtMethod * method,ReferenceTypeInfo receiver_type,HInstruction ** return_replacement,bool is_speculative)1584 bool HInliner::TryBuildAndInline(HInvoke* invoke_instruction,
1585 ArtMethod* method,
1586 ReferenceTypeInfo receiver_type,
1587 HInstruction** return_replacement,
1588 bool is_speculative) {
1589 // If invoke_instruction is devirtualized to a different method, give intrinsics
1590 // another chance before we try to inline it.
1591 if (invoke_instruction->GetResolvedMethod() != method &&
1592 method->IsIntrinsic() &&
1593 IsValidIntrinsicAfterBuilder(static_cast<Intrinsics>(method->GetIntrinsic()))) {
1594 MaybeRecordStat(stats_, MethodCompilationStat::kIntrinsicRecognized);
1595 // For simplicity, always create a new instruction to replace the existing
1596 // invoke.
1597 HInvokeVirtual* new_invoke = new (graph_->GetAllocator()) HInvokeVirtual(
1598 graph_->GetAllocator(),
1599 invoke_instruction->GetNumberOfArguments(),
1600 invoke_instruction->GetType(),
1601 invoke_instruction->GetDexPc(),
1602 invoke_instruction->GetMethodReference(), // Use existing invoke's method's reference.
1603 method,
1604 MethodReference(method->GetDexFile(), method->GetDexMethodIndex()),
1605 method->GetMethodIndex(),
1606 !graph_->IsDebuggable());
1607 DCHECK_NE(new_invoke->GetIntrinsic(), Intrinsics::kNone);
1608 HInputsRef inputs = invoke_instruction->GetInputs();
1609 for (size_t index = 0; index != inputs.size(); ++index) {
1610 new_invoke->SetArgumentAt(index, inputs[index]);
1611 }
1612 invoke_instruction->GetBlock()->InsertInstructionBefore(new_invoke, invoke_instruction);
1613 new_invoke->CopyEnvironmentFrom(invoke_instruction->GetEnvironment());
1614 if (invoke_instruction->GetType() == DataType::Type::kReference) {
1615 new_invoke->SetReferenceTypeInfoIfValid(invoke_instruction->GetReferenceTypeInfo());
1616 }
1617 *return_replacement = new_invoke;
1618 return true;
1619 }
1620
1621 CodeItemDataAccessor accessor(method->DexInstructionData());
1622
1623 if (!IsInliningAllowed(method, accessor)) {
1624 return false;
1625 }
1626
1627 // We have checked above that inlining is "allowed" to make sure that the method has bytecode
1628 // (is not native), is compilable and verified and to enforce the @NeverInline annotation.
1629 // However, the pattern substitution is always preferable, so we do it before the check if
1630 // inlining is "encouraged". It also has an exception to the `MayInline()` restriction.
1631 if (TryPatternSubstitution(invoke_instruction, method, accessor, return_replacement)) {
1632 LOG_SUCCESS() << "Successfully replaced pattern of invoke "
1633 << method->PrettyMethod();
1634 MaybeRecordStat(stats_, MethodCompilationStat::kReplacedInvokeWithSimplePattern);
1635 return true;
1636 }
1637
1638 // Check whether we're allowed to inline. The outermost compilation unit is the relevant
1639 // dex file here (though the transitivity of an inline chain would allow checking the caller).
1640 if (!MayInline(codegen_->GetCompilerOptions(),
1641 *method->GetDexFile(),
1642 *outer_compilation_unit_.GetDexFile())) {
1643 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedWont)
1644 << "Won't inline " << method->PrettyMethod() << " in "
1645 << outer_compilation_unit_.GetDexFile()->GetLocation() << " ("
1646 << caller_compilation_unit_.GetDexFile()->GetLocation() << ") from "
1647 << method->GetDexFile()->GetLocation();
1648 return false;
1649 }
1650
1651 if (!IsInliningSupported(invoke_instruction, method, accessor)) {
1652 return false;
1653 }
1654
1655 if (!IsInliningEncouraged(invoke_instruction, method, accessor)) {
1656 return false;
1657 }
1658
1659 if (!TryBuildAndInlineHelper(
1660 invoke_instruction, method, receiver_type, return_replacement, is_speculative)) {
1661 return false;
1662 }
1663
1664 LOG_SUCCESS() << method->PrettyMethod();
1665 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedInvoke);
1666 if (outermost_graph_ == graph_) {
1667 MaybeRecordStat(stats_, MethodCompilationStat::kInlinedLastInvoke);
1668 }
1669 return true;
1670 }
1671
GetInvokeInputForArgVRegIndex(HInvoke * invoke_instruction,size_t arg_vreg_index)1672 static HInstruction* GetInvokeInputForArgVRegIndex(HInvoke* invoke_instruction,
1673 size_t arg_vreg_index)
1674 REQUIRES_SHARED(Locks::mutator_lock_) {
1675 size_t input_index = 0;
1676 for (size_t i = 0; i < arg_vreg_index; ++i, ++input_index) {
1677 DCHECK_LT(input_index, invoke_instruction->GetNumberOfArguments());
1678 if (DataType::Is64BitType(invoke_instruction->InputAt(input_index)->GetType())) {
1679 ++i;
1680 DCHECK_NE(i, arg_vreg_index);
1681 }
1682 }
1683 DCHECK_LT(input_index, invoke_instruction->GetNumberOfArguments());
1684 return invoke_instruction->InputAt(input_index);
1685 }
1686
1687 // Try to recognize known simple patterns and replace invoke call with appropriate instructions.
TryPatternSubstitution(HInvoke * invoke_instruction,ArtMethod * method,const CodeItemDataAccessor & accessor,HInstruction ** return_replacement)1688 bool HInliner::TryPatternSubstitution(HInvoke* invoke_instruction,
1689 ArtMethod* method,
1690 const CodeItemDataAccessor& accessor,
1691 HInstruction** return_replacement) {
1692 InlineMethod inline_method;
1693 if (!InlineMethodAnalyser::AnalyseMethodCode(method, &accessor, &inline_method)) {
1694 return false;
1695 }
1696
1697 size_t number_of_instructions = 0u; // Note: We do not count constants.
1698 switch (inline_method.opcode) {
1699 case kInlineOpNop:
1700 DCHECK_EQ(invoke_instruction->GetType(), DataType::Type::kVoid);
1701 *return_replacement = nullptr;
1702 break;
1703 case kInlineOpReturnArg:
1704 *return_replacement = GetInvokeInputForArgVRegIndex(invoke_instruction,
1705 inline_method.d.return_data.arg);
1706 break;
1707 case kInlineOpNonWideConst: {
1708 char shorty0 = method->GetShorty()[0];
1709 if (shorty0 == 'L') {
1710 DCHECK_EQ(inline_method.d.data, 0u);
1711 *return_replacement = graph_->GetNullConstant();
1712 } else if (shorty0 == 'F') {
1713 *return_replacement = graph_->GetFloatConstant(
1714 bit_cast<float, int32_t>(static_cast<int32_t>(inline_method.d.data)));
1715 } else {
1716 *return_replacement = graph_->GetIntConstant(static_cast<int32_t>(inline_method.d.data));
1717 }
1718 break;
1719 }
1720 case kInlineOpIGet: {
1721 const InlineIGetIPutData& data = inline_method.d.ifield_data;
1722 if (data.method_is_static || data.object_arg != 0u) {
1723 // TODO: Needs null check.
1724 return false;
1725 }
1726 HInstruction* obj = GetInvokeInputForArgVRegIndex(invoke_instruction, data.object_arg);
1727 HInstanceFieldGet* iget = CreateInstanceFieldGet(data.field_idx, method, obj);
1728 DCHECK_EQ(iget->GetFieldOffset().Uint32Value(), data.field_offset);
1729 DCHECK_EQ(iget->IsVolatile() ? 1u : 0u, data.is_volatile);
1730 invoke_instruction->GetBlock()->InsertInstructionBefore(iget, invoke_instruction);
1731 *return_replacement = iget;
1732 number_of_instructions = 1u;
1733 break;
1734 }
1735 case kInlineOpIPut: {
1736 const InlineIGetIPutData& data = inline_method.d.ifield_data;
1737 if (data.method_is_static || data.object_arg != 0u) {
1738 // TODO: Needs null check.
1739 return false;
1740 }
1741 HInstruction* obj = GetInvokeInputForArgVRegIndex(invoke_instruction, data.object_arg);
1742 HInstruction* value = GetInvokeInputForArgVRegIndex(invoke_instruction, data.src_arg);
1743 HInstanceFieldSet* iput = CreateInstanceFieldSet(data.field_idx, method, obj, value);
1744 DCHECK_EQ(iput->GetFieldOffset().Uint32Value(), data.field_offset);
1745 DCHECK_EQ(iput->IsVolatile() ? 1u : 0u, data.is_volatile);
1746 invoke_instruction->GetBlock()->InsertInstructionBefore(iput, invoke_instruction);
1747 if (data.return_arg_plus1 != 0u) {
1748 size_t return_arg = data.return_arg_plus1 - 1u;
1749 *return_replacement = GetInvokeInputForArgVRegIndex(invoke_instruction, return_arg);
1750 }
1751 number_of_instructions = 1u;
1752 break;
1753 }
1754 case kInlineOpConstructor: {
1755 const InlineConstructorData& data = inline_method.d.constructor_data;
1756 // Get the indexes to arrays for easier processing.
1757 uint16_t iput_field_indexes[] = {
1758 data.iput0_field_index, data.iput1_field_index, data.iput2_field_index
1759 };
1760 uint16_t iput_args[] = { data.iput0_arg, data.iput1_arg, data.iput2_arg };
1761 static_assert(arraysize(iput_args) == arraysize(iput_field_indexes), "Size mismatch");
1762 // Count valid field indexes.
1763 size_t number_of_iputs = 0u;
1764 while (number_of_iputs != arraysize(iput_field_indexes) &&
1765 iput_field_indexes[number_of_iputs] != DexFile::kDexNoIndex16) {
1766 // Check that there are no duplicate valid field indexes.
1767 DCHECK_EQ(0, std::count(iput_field_indexes + number_of_iputs + 1,
1768 iput_field_indexes + arraysize(iput_field_indexes),
1769 iput_field_indexes[number_of_iputs]));
1770 ++number_of_iputs;
1771 }
1772 // Check that there are no valid field indexes in the rest of the array.
1773 DCHECK_EQ(0, std::count_if(iput_field_indexes + number_of_iputs,
1774 iput_field_indexes + arraysize(iput_field_indexes),
1775 [](uint16_t index) { return index != DexFile::kDexNoIndex16; }));
1776
1777 // Create HInstanceFieldSet for each IPUT that stores non-zero data.
1778 HInstruction* obj = GetInvokeInputForArgVRegIndex(invoke_instruction,
1779 /* arg_vreg_index= */ 0u);
1780 bool needs_constructor_barrier = false;
1781 for (size_t i = 0; i != number_of_iputs; ++i) {
1782 HInstruction* value = GetInvokeInputForArgVRegIndex(invoke_instruction, iput_args[i]);
1783 if (!IsZeroBitPattern(value)) {
1784 uint16_t field_index = iput_field_indexes[i];
1785 bool is_final;
1786 HInstanceFieldSet* iput =
1787 CreateInstanceFieldSet(field_index, method, obj, value, &is_final);
1788 invoke_instruction->GetBlock()->InsertInstructionBefore(iput, invoke_instruction);
1789
1790 // Check whether the field is final. If it is, we need to add a barrier.
1791 if (is_final) {
1792 needs_constructor_barrier = true;
1793 }
1794 }
1795 }
1796 if (needs_constructor_barrier) {
1797 // See DexCompilationUnit::RequiresConstructorBarrier for more details.
1798 DCHECK(obj != nullptr) << "only non-static methods can have a constructor fence";
1799
1800 HConstructorFence* constructor_fence =
1801 new (graph_->GetAllocator()) HConstructorFence(obj, kNoDexPc, graph_->GetAllocator());
1802 invoke_instruction->GetBlock()->InsertInstructionBefore(constructor_fence,
1803 invoke_instruction);
1804 }
1805 *return_replacement = nullptr;
1806 number_of_instructions = number_of_iputs + (needs_constructor_barrier ? 1u : 0u);
1807 break;
1808 }
1809 }
1810 if (number_of_instructions != 0u) {
1811 total_number_of_instructions_ += number_of_instructions;
1812 UpdateInliningBudget();
1813 }
1814 return true;
1815 }
1816
CreateInstanceFieldGet(uint32_t field_index,ArtMethod * referrer,HInstruction * obj)1817 HInstanceFieldGet* HInliner::CreateInstanceFieldGet(uint32_t field_index,
1818 ArtMethod* referrer,
1819 HInstruction* obj)
1820 REQUIRES_SHARED(Locks::mutator_lock_) {
1821 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
1822 ArtField* resolved_field =
1823 class_linker->LookupResolvedField(field_index, referrer, /* is_static= */ false);
1824 DCHECK(resolved_field != nullptr);
1825 HInstanceFieldGet* iget = new (graph_->GetAllocator()) HInstanceFieldGet(
1826 obj,
1827 resolved_field,
1828 DataType::FromShorty(resolved_field->GetTypeDescriptor()[0]),
1829 resolved_field->GetOffset(),
1830 resolved_field->IsVolatile(),
1831 field_index,
1832 resolved_field->GetDeclaringClass()->GetDexClassDefIndex(),
1833 *referrer->GetDexFile(),
1834 // Read barrier generates a runtime call in slow path and we need a valid
1835 // dex pc for the associated stack map. 0 is bogus but valid. Bug: 26854537.
1836 /* dex_pc= */ 0);
1837 if (iget->GetType() == DataType::Type::kReference) {
1838 // Use the same dex_cache that we used for field lookup as the hint_dex_cache.
1839 Handle<mirror::DexCache> dex_cache =
1840 graph_->GetHandleCache()->NewHandle(referrer->GetDexCache());
1841 ReferenceTypePropagation rtp(graph_,
1842 dex_cache,
1843 /* is_first_run= */ false);
1844 rtp.Visit(iget);
1845 }
1846 return iget;
1847 }
1848
CreateInstanceFieldSet(uint32_t field_index,ArtMethod * referrer,HInstruction * obj,HInstruction * value,bool * is_final)1849 HInstanceFieldSet* HInliner::CreateInstanceFieldSet(uint32_t field_index,
1850 ArtMethod* referrer,
1851 HInstruction* obj,
1852 HInstruction* value,
1853 bool* is_final)
1854 REQUIRES_SHARED(Locks::mutator_lock_) {
1855 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
1856 ArtField* resolved_field =
1857 class_linker->LookupResolvedField(field_index, referrer, /* is_static= */ false);
1858 DCHECK(resolved_field != nullptr);
1859 if (is_final != nullptr) {
1860 // This information is needed only for constructors.
1861 DCHECK(referrer->IsConstructor());
1862 *is_final = resolved_field->IsFinal();
1863 }
1864 HInstanceFieldSet* iput = new (graph_->GetAllocator()) HInstanceFieldSet(
1865 obj,
1866 value,
1867 resolved_field,
1868 DataType::FromShorty(resolved_field->GetTypeDescriptor()[0]),
1869 resolved_field->GetOffset(),
1870 resolved_field->IsVolatile(),
1871 field_index,
1872 resolved_field->GetDeclaringClass()->GetDexClassDefIndex(),
1873 *referrer->GetDexFile(),
1874 // Read barrier generates a runtime call in slow path and we need a valid
1875 // dex pc for the associated stack map. 0 is bogus but valid. Bug: 26854537.
1876 /* dex_pc= */ 0);
1877 return iput;
1878 }
1879
1880 template <typename T>
NewHandleIfDifferent(ObjPtr<T> object,Handle<T> hint,HGraph * graph)1881 static inline Handle<T> NewHandleIfDifferent(ObjPtr<T> object, Handle<T> hint, HGraph* graph)
1882 REQUIRES_SHARED(Locks::mutator_lock_) {
1883 return (object != hint.Get()) ? graph->GetHandleCache()->NewHandle(object) : hint;
1884 }
1885
CanEncodeInlinedMethodInStackMap(const DexFile & outer_dex_file,ArtMethod * callee,const CodeGenerator * codegen,bool * out_needs_bss_check)1886 static bool CanEncodeInlinedMethodInStackMap(const DexFile& outer_dex_file,
1887 ArtMethod* callee,
1888 const CodeGenerator* codegen,
1889 bool* out_needs_bss_check)
1890 REQUIRES_SHARED(Locks::mutator_lock_) {
1891 if (!Runtime::Current()->IsAotCompiler()) {
1892 // JIT can always encode methods in stack maps.
1893 return true;
1894 }
1895
1896 const DexFile* dex_file = callee->GetDexFile();
1897 if (IsSameDexFile(outer_dex_file, *dex_file)) {
1898 return true;
1899 }
1900
1901 // Inline across dexfiles if the callee's DexFile is:
1902 // 1) in the bootclasspath, or
1903 if (callee->GetDeclaringClass()->IsBootStrapClassLoaded()) {
1904 // In multi-image, each BCP DexFile has their own OatWriter. Since they don't cooperate with
1905 // each other, we request the BSS check for them.
1906 // TODO(solanes, 154012332): Add .bss support for BCP multi-image.
1907 *out_needs_bss_check = codegen->GetCompilerOptions().IsMultiImage();
1908 return true;
1909 }
1910
1911 // 2) is a non-BCP dexfile with the OatFile we are compiling.
1912 if (codegen->GetCompilerOptions().WithinOatFile(dex_file)) {
1913 return true;
1914 }
1915
1916 // TODO(solanes): Support more AOT cases for inlining:
1917 // - methods in class loader context's DexFiles
1918 return false;
1919 }
1920
1921 // 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)1922 void HInliner::SubstituteArguments(HGraph* callee_graph,
1923 HInvoke* invoke_instruction,
1924 ReferenceTypeInfo receiver_type,
1925 const DexCompilationUnit& dex_compilation_unit) {
1926 ArtMethod* const resolved_method = callee_graph->GetArtMethod();
1927 size_t parameter_index = 0;
1928 bool run_rtp = false;
1929 for (HInstructionIterator instructions(callee_graph->GetEntryBlock()->GetInstructions());
1930 !instructions.Done();
1931 instructions.Advance()) {
1932 HInstruction* current = instructions.Current();
1933 if (current->IsParameterValue()) {
1934 HInstruction* argument = invoke_instruction->InputAt(parameter_index);
1935 if (argument->IsNullConstant()) {
1936 current->ReplaceWith(callee_graph->GetNullConstant());
1937 } else if (argument->IsIntConstant()) {
1938 current->ReplaceWith(callee_graph->GetIntConstant(argument->AsIntConstant()->GetValue()));
1939 } else if (argument->IsLongConstant()) {
1940 current->ReplaceWith(callee_graph->GetLongConstant(argument->AsLongConstant()->GetValue()));
1941 } else if (argument->IsFloatConstant()) {
1942 current->ReplaceWith(
1943 callee_graph->GetFloatConstant(argument->AsFloatConstant()->GetValue()));
1944 } else if (argument->IsDoubleConstant()) {
1945 current->ReplaceWith(
1946 callee_graph->GetDoubleConstant(argument->AsDoubleConstant()->GetValue()));
1947 } else if (argument->GetType() == DataType::Type::kReference) {
1948 if (!resolved_method->IsStatic() && parameter_index == 0 && receiver_type.IsValid()) {
1949 run_rtp = true;
1950 current->SetReferenceTypeInfo(receiver_type);
1951 } else {
1952 current->SetReferenceTypeInfoIfValid(argument->GetReferenceTypeInfo());
1953 }
1954 current->AsParameterValue()->SetCanBeNull(argument->CanBeNull());
1955 }
1956 ++parameter_index;
1957 }
1958 }
1959
1960 // We have replaced formal arguments with actual arguments. If actual types
1961 // are more specific than the declared ones, run RTP again on the inner graph.
1962 if (run_rtp || ArgumentTypesMoreSpecific(invoke_instruction, resolved_method)) {
1963 ReferenceTypePropagation(callee_graph,
1964 dex_compilation_unit.GetDexCache(),
1965 /* is_first_run= */ false).Run();
1966 }
1967 }
1968
1969 // Returns whether we can inline the callee_graph into the target_block.
1970 //
1971 // This performs a combination of semantics checks, compiler support checks, and
1972 // resource limit checks.
1973 //
1974 // If this function returns true, it will also set out_number_of_instructions to
1975 // the number of instructions in the inlined body.
CanInlineBody(const HGraph * callee_graph,HInvoke * invoke,size_t * out_number_of_instructions,bool is_speculative) const1976 bool HInliner::CanInlineBody(const HGraph* callee_graph,
1977 HInvoke* invoke,
1978 size_t* out_number_of_instructions,
1979 bool is_speculative) const {
1980 ArtMethod* const resolved_method = callee_graph->GetArtMethod();
1981
1982 HBasicBlock* exit_block = callee_graph->GetExitBlock();
1983 if (exit_block == nullptr) {
1984 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedInfiniteLoop)
1985 << "Method " << resolved_method->PrettyMethod()
1986 << " could not be inlined because it has an infinite loop";
1987 return false;
1988 }
1989
1990 bool has_one_return = false;
1991 for (HBasicBlock* predecessor : exit_block->GetPredecessors()) {
1992 const HInstruction* last_instruction = predecessor->GetLastInstruction();
1993 // On inlinees, we can have Return/ReturnVoid/Throw -> TryBoundary -> Exit. To check for the
1994 // actual last instruction, we have to skip the TryBoundary instruction.
1995 if (last_instruction->IsTryBoundary()) {
1996 predecessor = predecessor->GetSinglePredecessor();
1997 last_instruction = predecessor->GetLastInstruction();
1998
1999 // If the last instruction chain is Return/ReturnVoid -> TryBoundary -> Exit we will have to
2000 // split a critical edge in InlineInto and might recompute loop information, which is
2001 // unsupported for irreducible loops.
2002 if (!last_instruction->IsThrow() && graph_->HasIrreducibleLoops()) {
2003 DCHECK(last_instruction->IsReturn() || last_instruction->IsReturnVoid());
2004 // TODO(ngeoffray): Support re-computing loop information to graphs with
2005 // irreducible loops?
2006 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedIrreducibleLoopCaller)
2007 << "Method " << resolved_method->PrettyMethod()
2008 << " could not be inlined because we will have to recompute the loop information and"
2009 << " the caller has irreducible loops";
2010 return false;
2011 }
2012 }
2013
2014 if (last_instruction->IsThrow()) {
2015 if (graph_->GetExitBlock() == nullptr) {
2016 // TODO(ngeoffray): Support adding HExit in the caller graph.
2017 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedInfiniteLoop)
2018 << "Method " << resolved_method->PrettyMethod()
2019 << " could not be inlined because one branch always throws and"
2020 << " caller does not have an exit block";
2021 return false;
2022 } else if (graph_->HasIrreducibleLoops()) {
2023 // TODO(ngeoffray): Support re-computing loop information to graphs with
2024 // irreducible loops?
2025 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedIrreducibleLoopCaller)
2026 << "Method " << resolved_method->PrettyMethod()
2027 << " could not be inlined because one branch always throws and"
2028 << " the caller has irreducible loops";
2029 return false;
2030 }
2031 } else {
2032 has_one_return = true;
2033 }
2034 }
2035
2036 if (!has_one_return) {
2037 if (!is_speculative) {
2038 // If we know that the method always throws with the particular parameters, set it as such.
2039 // This is better than using the dex instructions as we have more information about this
2040 // particular call. We don't mark speculative inlines (e.g. the ones from the inline cache) as
2041 // always throwing since they might not throw when executed.
2042 invoke->SetAlwaysThrows(/* always_throws= */ true);
2043 graph_->SetHasAlwaysThrowingInvokes(/* value= */ true);
2044 }
2045
2046 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedAlwaysThrows)
2047 << "Method " << resolved_method->PrettyMethod()
2048 << " could not be inlined because it always throws";
2049 return false;
2050 }
2051
2052 const bool too_many_registers =
2053 total_number_of_dex_registers_ > kMaximumNumberOfCumulatedDexRegisters;
2054 bool needs_bss_check = false;
2055 const bool can_encode_in_stack_map = CanEncodeInlinedMethodInStackMap(
2056 *outer_compilation_unit_.GetDexFile(), resolved_method, codegen_, &needs_bss_check);
2057 size_t number_of_instructions = 0;
2058 // Skip the entry block, it does not contain instructions that prevent inlining.
2059 for (HBasicBlock* block : callee_graph->GetReversePostOrderSkipEntryBlock()) {
2060 if (block->IsLoopHeader()) {
2061 if (block->GetLoopInformation()->IsIrreducible()) {
2062 // Don't inline methods with irreducible loops, they could prevent some
2063 // optimizations to run.
2064 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedIrreducibleLoopCallee)
2065 << "Method " << resolved_method->PrettyMethod()
2066 << " could not be inlined because it contains an irreducible loop";
2067 return false;
2068 }
2069 if (!block->GetLoopInformation()->HasExitEdge()) {
2070 // Don't inline methods with loops without exit, since they cause the
2071 // loop information to be computed incorrectly when updating after
2072 // inlining.
2073 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedLoopWithoutExit)
2074 << "Method " << resolved_method->PrettyMethod()
2075 << " could not be inlined because it contains a loop with no exit";
2076 return false;
2077 }
2078 }
2079
2080 for (HInstructionIterator instr_it(block->GetInstructions());
2081 !instr_it.Done();
2082 instr_it.Advance()) {
2083 if (++number_of_instructions > inlining_budget_) {
2084 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedInstructionBudget)
2085 << "Method " << resolved_method->PrettyMethod()
2086 << " is not inlined because the outer method has reached"
2087 << " its instruction budget limit.";
2088 return false;
2089 }
2090 HInstruction* current = instr_it.Current();
2091 if (current->NeedsEnvironment()) {
2092 if (too_many_registers) {
2093 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedEnvironmentBudget)
2094 << "Method " << resolved_method->PrettyMethod()
2095 << " is not inlined because its caller has reached"
2096 << " its environment budget limit.";
2097 return false;
2098 }
2099
2100 if (!can_encode_in_stack_map) {
2101 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedStackMaps)
2102 << "Method " << resolved_method->PrettyMethod() << " could not be inlined because "
2103 << current->DebugName() << " needs an environment, is in a different dex file"
2104 << ", and cannot be encoded in the stack maps.";
2105 return false;
2106 }
2107 }
2108
2109 if (current->IsUnresolvedStaticFieldGet() ||
2110 current->IsUnresolvedInstanceFieldGet() ||
2111 current->IsUnresolvedStaticFieldSet() ||
2112 current->IsUnresolvedInstanceFieldSet() ||
2113 current->IsInvokeUnresolved()) {
2114 // Unresolved invokes / field accesses are expensive at runtime when decoding inlining info,
2115 // so don't inline methods that have them.
2116 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedUnresolvedEntrypoint)
2117 << "Method " << resolved_method->PrettyMethod()
2118 << " could not be inlined because it is using an unresolved"
2119 << " entrypoint";
2120 return false;
2121 }
2122
2123 // We currently don't have support for inlining across dex files if we are:
2124 // 1) In AoT,
2125 // 2) cross-dex inlining,
2126 // 3) the callee is a BCP DexFile,
2127 // 4) we are compiling multi image, and
2128 // 5) have an instruction that needs a bss entry, which will always be
2129 // 5)b) an instruction that needs an environment.
2130 // 1) - 4) are encoded in `needs_bss_check` (see CanEncodeInlinedMethodInStackMap).
2131 if (needs_bss_check && current->NeedsBss()) {
2132 DCHECK(current->NeedsEnvironment());
2133 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedBss)
2134 << "Method " << resolved_method->PrettyMethod()
2135 << " could not be inlined because it needs a BSS check";
2136 return false;
2137 }
2138
2139 if (outermost_graph_->IsCompilingBaseline() &&
2140 (current->IsInvokeVirtual() || current->IsInvokeInterface()) &&
2141 ProfilingInfoBuilder::IsInlineCacheUseful(current->AsInvoke(), codegen_)) {
2142 uint32_t maximum_inlining_depth_for_baseline =
2143 InlineCache::MaxDexPcEncodingDepth(
2144 outermost_graph_->GetArtMethod(),
2145 codegen_->GetCompilerOptions().GetInlineMaxCodeUnits());
2146 if (depth_ + 1 > maximum_inlining_depth_for_baseline) {
2147 LOG_FAIL_NO_STAT() << "Reached maximum depth for inlining in baseline compilation: "
2148 << depth_ << " for " << callee_graph->GetArtMethod()->PrettyMethod();
2149 outermost_graph_->SetUsefulOptimizing();
2150 return false;
2151 }
2152 }
2153 }
2154 }
2155
2156 *out_number_of_instructions = number_of_instructions;
2157 return true;
2158 }
2159
TryBuildAndInlineHelper(HInvoke * invoke_instruction,ArtMethod * resolved_method,ReferenceTypeInfo receiver_type,HInstruction ** return_replacement,bool is_speculative)2160 bool HInliner::TryBuildAndInlineHelper(HInvoke* invoke_instruction,
2161 ArtMethod* resolved_method,
2162 ReferenceTypeInfo receiver_type,
2163 HInstruction** return_replacement,
2164 bool is_speculative) {
2165 DCHECK_IMPLIES(resolved_method->IsStatic(), !receiver_type.IsValid());
2166 DCHECK_IMPLIES(!resolved_method->IsStatic(), receiver_type.IsValid());
2167 const dex::CodeItem* code_item = resolved_method->GetCodeItem();
2168 const DexFile& callee_dex_file = *resolved_method->GetDexFile();
2169 uint32_t method_index = resolved_method->GetDexMethodIndex();
2170 CodeItemDebugInfoAccessor code_item_accessor(resolved_method->DexInstructionDebugInfo());
2171 ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
2172 Handle<mirror::DexCache> dex_cache = NewHandleIfDifferent(resolved_method->GetDexCache(),
2173 caller_compilation_unit_.GetDexCache(),
2174 graph_);
2175 Handle<mirror::ClassLoader> class_loader =
2176 NewHandleIfDifferent(resolved_method->GetDeclaringClass()->GetClassLoader(),
2177 caller_compilation_unit_.GetClassLoader(),
2178 graph_);
2179
2180 Handle<mirror::Class> compiling_class =
2181 graph_->GetHandleCache()->NewHandle(resolved_method->GetDeclaringClass());
2182 DexCompilationUnit dex_compilation_unit(
2183 class_loader,
2184 class_linker,
2185 callee_dex_file,
2186 code_item,
2187 resolved_method->GetDeclaringClass()->GetDexClassDefIndex(),
2188 method_index,
2189 resolved_method->GetAccessFlags(),
2190 /* verified_method= */ nullptr,
2191 dex_cache,
2192 compiling_class);
2193
2194 InvokeType invoke_type = invoke_instruction->GetInvokeType();
2195 if (invoke_type == kInterface) {
2196 // We have statically resolved the dispatch. To please the class linker
2197 // at runtime, we change this call as if it was a virtual call.
2198 invoke_type = kVirtual;
2199 }
2200
2201 bool caller_dead_reference_safe = graph_->IsDeadReferenceSafe();
2202 const dex::ClassDef& callee_class = resolved_method->GetClassDef();
2203 // MethodContainsRSensitiveAccess is currently slow, but HasDeadReferenceSafeAnnotation()
2204 // is currently rarely true.
2205 bool callee_dead_reference_safe =
2206 annotations::HasDeadReferenceSafeAnnotation(callee_dex_file, callee_class)
2207 && !annotations::MethodContainsRSensitiveAccess(callee_dex_file, callee_class, method_index);
2208
2209 const int32_t caller_instruction_counter = graph_->GetCurrentInstructionId();
2210 HGraph* callee_graph = new (graph_->GetAllocator()) HGraph(
2211 graph_->GetAllocator(),
2212 graph_->GetArenaStack(),
2213 graph_->GetHandleCache()->GetHandles(),
2214 callee_dex_file,
2215 method_index,
2216 codegen_->GetCompilerOptions().GetInstructionSet(),
2217 invoke_type,
2218 callee_dead_reference_safe,
2219 graph_->IsDebuggable(),
2220 graph_->GetCompilationKind(),
2221 /* start_instruction_id= */ caller_instruction_counter);
2222 callee_graph->SetArtMethod(resolved_method);
2223
2224 ScopedProfilingInfoUse spiu(Runtime::Current()->GetJit(), resolved_method, Thread::Current());
2225 if (Runtime::Current()->GetJit() != nullptr) {
2226 callee_graph->SetProfilingInfo(spiu.GetProfilingInfo());
2227 }
2228
2229 // When they are needed, allocate `inline_stats_` on the Arena instead
2230 // of on the stack, as Clang might produce a stack frame too large
2231 // for this function, that would not fit the requirements of the
2232 // `-Wframe-larger-than` option.
2233 if (stats_ != nullptr) {
2234 // Reuse one object for all inline attempts from this caller to keep Arena memory usage low.
2235 if (inline_stats_ == nullptr) {
2236 void* storage = graph_->GetAllocator()->Alloc<OptimizingCompilerStats>(kArenaAllocMisc);
2237 inline_stats_ = new (storage) OptimizingCompilerStats;
2238 } else {
2239 inline_stats_->Reset();
2240 }
2241 }
2242 HGraphBuilder builder(callee_graph,
2243 code_item_accessor,
2244 &dex_compilation_unit,
2245 &outer_compilation_unit_,
2246 codegen_,
2247 inline_stats_);
2248
2249 if (builder.BuildGraph() != kAnalysisSuccess) {
2250 LOG_FAIL(stats_, MethodCompilationStat::kNotInlinedCannotBuild)
2251 << "Method " << callee_dex_file.PrettyMethod(method_index)
2252 << " could not be built, so cannot be inlined";
2253 return false;
2254 }
2255
2256 SubstituteArguments(callee_graph, invoke_instruction, receiver_type, dex_compilation_unit);
2257
2258 const bool try_catch_inlining_allowed_for_recursive_inline =
2259 // It was allowed previously.
2260 try_catch_inlining_allowed_ &&
2261 // The current invoke is not a try block.
2262 !invoke_instruction->GetBlock()->IsTryBlock();
2263 RunOptimizations(callee_graph,
2264 invoke_instruction->GetEnvironment(),
2265 code_item,
2266 dex_compilation_unit,
2267 try_catch_inlining_allowed_for_recursive_inline);
2268
2269 size_t number_of_instructions = 0;
2270 if (!CanInlineBody(callee_graph, invoke_instruction, &number_of_instructions, is_speculative)) {
2271 return false;
2272 }
2273
2274 DCHECK_EQ(caller_instruction_counter, graph_->GetCurrentInstructionId())
2275 << "No instructions can be added to the outer graph while inner graph is being built";
2276
2277 // Inline the callee graph inside the caller graph.
2278 const int32_t callee_instruction_counter = callee_graph->GetCurrentInstructionId();
2279 graph_->SetCurrentInstructionId(callee_instruction_counter);
2280 *return_replacement = callee_graph->InlineInto(graph_, invoke_instruction);
2281 // Update our budget for other inlining attempts in `caller_graph`.
2282 total_number_of_instructions_ += number_of_instructions;
2283 UpdateInliningBudget();
2284
2285 DCHECK_EQ(callee_instruction_counter, callee_graph->GetCurrentInstructionId())
2286 << "No instructions can be added to the inner graph during inlining into the outer graph";
2287
2288 if (stats_ != nullptr) {
2289 DCHECK(inline_stats_ != nullptr);
2290 inline_stats_->AddTo(stats_);
2291 }
2292
2293 if (caller_dead_reference_safe && !callee_dead_reference_safe) {
2294 // Caller was dead reference safe, but is not anymore, since we inlined dead
2295 // reference unsafe code. Prior transformations remain valid, since they did not
2296 // affect the inlined code.
2297 graph_->MarkDeadReferenceUnsafe();
2298 }
2299
2300 return true;
2301 }
2302
RunOptimizations(HGraph * callee_graph,HEnvironment * caller_environment,const dex::CodeItem * code_item,const DexCompilationUnit & dex_compilation_unit,bool try_catch_inlining_allowed_for_recursive_inline)2303 void HInliner::RunOptimizations(HGraph* callee_graph,
2304 HEnvironment* caller_environment,
2305 const dex::CodeItem* code_item,
2306 const DexCompilationUnit& dex_compilation_unit,
2307 bool try_catch_inlining_allowed_for_recursive_inline) {
2308 // Note: if the outermost_graph_ is being compiled OSR, we should not run any
2309 // optimization that could lead to a HDeoptimize. The following optimizations do not.
2310 HDeadCodeElimination dce(callee_graph, inline_stats_, "dead_code_elimination$inliner");
2311 HConstantFolding fold(callee_graph, inline_stats_, "constant_folding$inliner");
2312 InstructionSimplifier simplify(callee_graph, codegen_, inline_stats_);
2313
2314 HOptimization* optimizations[] = {
2315 &fold,
2316 &simplify,
2317 &dce,
2318 };
2319
2320 for (size_t i = 0; i < arraysize(optimizations); ++i) {
2321 HOptimization* optimization = optimizations[i];
2322 optimization->Run();
2323 }
2324
2325 // Bail early for pathological cases on the environment (for example recursive calls,
2326 // or too large environment).
2327 if (total_number_of_dex_registers_ > kMaximumNumberOfCumulatedDexRegisters) {
2328 LOG_NOTE() << "Calls in " << callee_graph->GetArtMethod()->PrettyMethod()
2329 << " will not be inlined because the outer method has reached"
2330 << " its environment budget limit.";
2331 return;
2332 }
2333
2334 // Bail early if we know we already are over the limit.
2335 size_t number_of_instructions = CountNumberOfInstructions(callee_graph);
2336 if (number_of_instructions > inlining_budget_) {
2337 LOG_NOTE() << "Calls in " << callee_graph->GetArtMethod()->PrettyMethod()
2338 << " will not be inlined because the outer method has reached"
2339 << " its instruction budget limit. " << number_of_instructions;
2340 return;
2341 }
2342
2343 CodeItemDataAccessor accessor(callee_graph->GetDexFile(), code_item);
2344 HInliner inliner(callee_graph,
2345 outermost_graph_,
2346 codegen_,
2347 outer_compilation_unit_,
2348 dex_compilation_unit,
2349 inline_stats_,
2350 total_number_of_dex_registers_ + accessor.RegistersSize(),
2351 total_number_of_instructions_ + number_of_instructions,
2352 this,
2353 caller_environment,
2354 depth_ + 1,
2355 try_catch_inlining_allowed_for_recursive_inline);
2356 inliner.Run();
2357 }
2358
IsReferenceTypeRefinement(ObjPtr<mirror::Class> declared_class,bool declared_is_exact,bool declared_can_be_null,HInstruction * actual_obj)2359 static bool IsReferenceTypeRefinement(ObjPtr<mirror::Class> declared_class,
2360 bool declared_is_exact,
2361 bool declared_can_be_null,
2362 HInstruction* actual_obj)
2363 REQUIRES_SHARED(Locks::mutator_lock_) {
2364 if (declared_can_be_null && !actual_obj->CanBeNull()) {
2365 return true;
2366 }
2367
2368 ReferenceTypeInfo actual_rti = actual_obj->GetReferenceTypeInfo();
2369 if (!actual_rti.IsValid()) {
2370 return false;
2371 }
2372
2373 ObjPtr<mirror::Class> actual_class = actual_rti.GetTypeHandle().Get();
2374 return (actual_rti.IsExact() && !declared_is_exact) ||
2375 (declared_class != actual_class && declared_class->IsAssignableFrom(actual_class));
2376 }
2377
IsReferenceTypeRefinement(ObjPtr<mirror::Class> declared_class,bool declared_can_be_null,HInstruction * actual_obj)2378 static bool IsReferenceTypeRefinement(ObjPtr<mirror::Class> declared_class,
2379 bool declared_can_be_null,
2380 HInstruction* actual_obj)
2381 REQUIRES_SHARED(Locks::mutator_lock_) {
2382 bool admissible = ReferenceTypePropagation::IsAdmissible(declared_class);
2383 return IsReferenceTypeRefinement(
2384 admissible ? declared_class : GetClassRoot<mirror::Class>(),
2385 /*declared_is_exact=*/ admissible && declared_class->CannotBeAssignedFromOtherTypes(),
2386 declared_can_be_null,
2387 actual_obj);
2388 }
2389
ArgumentTypesMoreSpecific(HInvoke * invoke_instruction,ArtMethod * resolved_method)2390 bool HInliner::ArgumentTypesMoreSpecific(HInvoke* invoke_instruction, ArtMethod* resolved_method) {
2391 // If this is an instance call, test whether the type of the `this` argument
2392 // is more specific than the class which declares the method.
2393 if (!resolved_method->IsStatic()) {
2394 if (IsReferenceTypeRefinement(resolved_method->GetDeclaringClass(),
2395 /*declared_can_be_null=*/ false,
2396 invoke_instruction->InputAt(0u))) {
2397 return true;
2398 }
2399 }
2400
2401 // Iterate over the list of parameter types and test whether any of the
2402 // actual inputs has a more specific reference type than the type declared in
2403 // the signature.
2404 const dex::TypeList* param_list = resolved_method->GetParameterTypeList();
2405 for (size_t param_idx = 0,
2406 input_idx = resolved_method->IsStatic() ? 0 : 1,
2407 e = (param_list == nullptr ? 0 : param_list->Size());
2408 param_idx < e;
2409 ++param_idx, ++input_idx) {
2410 HInstruction* input = invoke_instruction->InputAt(input_idx);
2411 if (input->GetType() == DataType::Type::kReference) {
2412 ObjPtr<mirror::Class> param_cls = resolved_method->LookupResolvedClassFromTypeIndex(
2413 param_list->GetTypeItem(param_idx).type_idx_);
2414 if (IsReferenceTypeRefinement(param_cls, /*declared_can_be_null=*/ true, input)) {
2415 return true;
2416 }
2417 }
2418 }
2419
2420 return false;
2421 }
2422
ReturnTypeMoreSpecific(HInstruction * return_replacement,HInvoke * invoke_instruction)2423 bool HInliner::ReturnTypeMoreSpecific(HInstruction* return_replacement,
2424 HInvoke* invoke_instruction) {
2425 // Check the integrity of reference types and run another type propagation if needed.
2426 if (return_replacement != nullptr) {
2427 if (return_replacement->GetType() == DataType::Type::kReference) {
2428 // Test if the return type is a refinement of the declared return type.
2429 ReferenceTypeInfo invoke_rti = invoke_instruction->GetReferenceTypeInfo();
2430 if (IsReferenceTypeRefinement(invoke_rti.GetTypeHandle().Get(),
2431 invoke_rti.IsExact(),
2432 /*declared_can_be_null=*/ true,
2433 return_replacement)) {
2434 return true;
2435 } else if (return_replacement->IsInstanceFieldGet()) {
2436 HInstanceFieldGet* field_get = return_replacement->AsInstanceFieldGet();
2437 if (field_get->GetFieldInfo().GetField() ==
2438 GetClassRoot<mirror::Object>()->GetInstanceField(0)) {
2439 return true;
2440 }
2441 }
2442 } else if (return_replacement->IsInstanceOf()) {
2443 // Inlining InstanceOf into an If may put a tighter bound on reference types.
2444 return true;
2445 }
2446 }
2447
2448 return false;
2449 }
2450
FixUpReturnReferenceType(ArtMethod * resolved_method,HInstruction * return_replacement)2451 void HInliner::FixUpReturnReferenceType(ArtMethod* resolved_method,
2452 HInstruction* return_replacement) {
2453 if (return_replacement != nullptr) {
2454 if (return_replacement->GetType() == DataType::Type::kReference) {
2455 if (!return_replacement->GetReferenceTypeInfo().IsValid()) {
2456 // Make sure that we have a valid type for the return. We may get an invalid one when
2457 // we inline invokes with multiple branches and create a Phi for the result.
2458 // TODO: we could be more precise by merging the phi inputs but that requires
2459 // some functionality from the reference type propagation.
2460 DCHECK(return_replacement->IsPhi());
2461 ObjPtr<mirror::Class> cls = resolved_method->LookupResolvedReturnType();
2462 ReferenceTypeInfo rti = ReferenceTypePropagation::IsAdmissible(cls)
2463 ? ReferenceTypeInfo::Create(graph_->GetHandleCache()->NewHandle(cls))
2464 : graph_->GetInexactObjectRti();
2465 return_replacement->SetReferenceTypeInfo(rti);
2466 }
2467 }
2468 }
2469 }
2470
2471 } // namespace art
2472