x86_64: Add instrinsic for MethodHandle::invokeExact...

... which targets invoke-virtual methods.

New entrypoint changes deliverException's offset, hence arm test
change.

Bug: 297147201
Test: ./art/test/testrunner/testrunner.py --host --64 -b --optimizing
Test: ./art/test.py --host -g
Change-Id: I636fc60c088bfdf9b695c92de47f1c539e3956f1

5 files changed, 174 insertions, 4 deletions

diff --git a/compiler/optimizing/code_generator_x86_64.h b/compiler/optimizing/code_generator_x86_64.h
index cbb4b17fe5..ad4a60e091 100644
--- a/compiler/optimizing/code_generator_x86_64.h
+++ b/compiler/optimizing/code_generator_x86_64.h
@@ -93,7 +93,6 @@ static constexpr FloatRegister non_volatile_xmm_regs[] = { XMM12, XMM13, XMM14,
   V(StringBuilderLength)                       \
   V(StringBuilderToString)                     \
   /* 1.8 */                                    \
-  V(MethodHandleInvokeExact)                   \
   V(MethodHandleInvoke)
 
 class InvokeRuntimeCallingConvention : public CallingConvention<Register, FloatRegister> {
diff --git a/compiler/optimizing/instruction_builder.cc b/compiler/optimizing/instruction_builder.cc
index c97c78ca17..d7553dd14f 100644
--- a/compiler/optimizing/instruction_builder.cc
+++ b/compiler/optimizing/instruction_builder.cc
@@ -1390,8 +1390,20 @@ bool HInstructionBuilder::BuildInvokePolymorphic(uint32_t dex_pc,
                                             &is_string_constructor);
 
   MethodReference method_reference(&graph_->GetDexFile(), method_idx);
+
+  bool is_invoke_exact =
+      static_cast<Intrinsics>(resolved_method->GetIntrinsic()) ==
+          Intrinsics::kMethodHandleInvokeExact;
+  bool can_be_virtual = number_of_arguments >= 2 &&
+      DataType::FromShorty(shorty[1]) == DataType::Type::kReference;
+
+  bool can_be_intrinsified = is_invoke_exact && can_be_virtual;
+
+  uint32_t number_of_other_inputs = can_be_intrinsified ? 1u : 0u;
+
   HInvoke* invoke = new (allocator_) HInvokePolymorphic(allocator_,
                                                         number_of_arguments,
+                                                        number_of_other_inputs,
                                                         return_type,
                                                         dex_pc,
                                                         method_reference,
@@ -1402,6 +1414,8 @@ bool HInstructionBuilder::BuildInvokePolymorphic(uint32_t dex_pc,
     return false;
   }
 
+  DCHECK_EQ(invoke->AsInvokePolymorphic()->CanTargetInvokeVirtual(), can_be_intrinsified);
+
   if (invoke->GetIntrinsic() != Intrinsics::kNone &&
       invoke->GetIntrinsic() != Intrinsics::kMethodHandleInvoke &&
       invoke->GetIntrinsic() != Intrinsics::kMethodHandleInvokeExact &&
@@ -1879,6 +1893,24 @@ bool HInstructionBuilder::SetupInvokeArguments(HInstruction* invoke,
                           graph_->GetCurrentMethod());
   }
 
+  if (invoke->IsInvokePolymorphic()) {
+    HInvokePolymorphic* invoke_polymorphic = invoke->AsInvokePolymorphic();
+
+    if (invoke_polymorphic->CanTargetInvokeVirtual()) {
+      HLoadMethodType* load_method_type =
+          new (allocator_) HLoadMethodType(graph_->GetCurrentMethod(),
+                                           invoke_polymorphic->GetProtoIndex(),
+                                           graph_->GetDexFile(),
+                                           invoke_polymorphic->GetDexPc());
+      HSharpening::ProcessLoadMethodType(load_method_type,
+                                         code_generator_,
+                                         *dex_compilation_unit_,
+                                         graph_->GetHandleCache()->GetHandles());
+      invoke->SetRawInputAt(invoke_polymorphic->GetNumberOfArguments(), load_method_type);
+      AppendInstruction(load_method_type);
+    }
+  }
+
   return true;
 }
 
diff --git a/compiler/optimizing/intrinsics_x86_64.cc b/compiler/optimizing/intrinsics_x86_64.cc
index 0a60e9c642..c32595c486 100644
--- a/compiler/optimizing/intrinsics_x86_64.cc
+++ b/compiler/optimizing/intrinsics_x86_64.cc
@@ -19,10 +19,13 @@
 #include <limits>
 
 #include "arch/x86_64/instruction_set_features_x86_64.h"
+#include "arch/x86_64/registers_x86_64.h"
 #include "art_method.h"
 #include "base/bit_utils.h"
 #include "code_generator_x86_64.h"
+#include "dex/modifiers.h"
 #include "entrypoints/quick/quick_entrypoints.h"
+#include "entrypoints/quick/quick_entrypoints_enum.h"
 #include "heap_poisoning.h"
 #include "intrinsics.h"
 #include "intrinsic_objects.h"
@@ -32,6 +35,7 @@
 #include "mirror/object_array-inl.h"
 #include "mirror/reference.h"
 #include "mirror/string.h"
+#include "optimizing/code_generator.h"
 #include "scoped_thread_state_change-inl.h"
 #include "thread-current-inl.h"
 #include "utils/x86_64/assembler_x86_64.h"
@@ -141,6 +145,36 @@ class ReadBarrierSystemArrayCopySlowPathX86_64 : public SlowPathCode {
   DISALLOW_COPY_AND_ASSIGN(ReadBarrierSystemArrayCopySlowPathX86_64);
 };
 
+// invoke-polymorphic's slow-path which does not move arguments.
+class InvokePolymorphicSlowPathX86_64 : public SlowPathCode {
+ public:
+  explicit InvokePolymorphicSlowPathX86_64(HInstruction* instruction, CpuRegister method_handle)
+      : SlowPathCode(instruction), method_handle_(method_handle) {
+    DCHECK(instruction->IsInvokePolymorphic());
+  }
+
+  void EmitNativeCode(CodeGenerator* codegen) override {
+    CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
+    X86_64Assembler* assembler = x86_64_codegen->GetAssembler();
+    __ Bind(GetEntryLabel());
+    SaveLiveRegisters(codegen, instruction_->GetLocations());
+
+    __ movq(CpuRegister(RDI), method_handle_);
+    x86_64_codegen->InvokeRuntime(QuickEntrypointEnum::kQuickInvokePolymorphicWithHiddenReceiver,
+                                  instruction_,
+                                  instruction_->GetDexPc());
+
+    RestoreLiveRegisters(codegen, instruction_->GetLocations());
+    __ jmp(GetExitLabel());
+  }
+
+  const char* GetDescription() const override { return "InvokePolymorphicSlowPathX86_64"; }
+
+ private:
+  const CpuRegister method_handle_;
+  DISALLOW_COPY_AND_ASSIGN(InvokePolymorphicSlowPathX86_64);
+};
+
 static void CreateFPToIntLocations(ArenaAllocator* allocator, HInvoke* invoke) {
   LocationSummary* locations =
       new (allocator) LocationSummary(invoke, LocationSummary::kNoCall, kIntrinsified);
@@ -3606,7 +3640,7 @@ void IntrinsicLocationsBuilderX86_64::VisitMathFmaFloat(HInvoke* invoke) {
 
 // Generate subtype check without read barriers.
 static void GenerateSubTypeObjectCheckNoReadBarrier(CodeGeneratorX86_64* codegen,
-                                                    VarHandleSlowPathX86_64* slow_path,
+                                                    SlowPathCode* slow_path,
                                                     CpuRegister object,
                                                     CpuRegister temp,
                                                     Address type_address,
@@ -4062,6 +4096,104 @@ static void GenerateVarHandleGet(HInvoke* invoke,
   }
 }
 
+void IntrinsicLocationsBuilderX86_64::VisitMethodHandleInvokeExact(HInvoke* invoke) {
+  // Don't emit intrinsic code for MethodHandle.invokeExact when it certainly does not target
+  // invoke-virtual: if invokeExact is called w/o arguments or if the first argument in that
+  // call is not a reference.
+  if (!invoke->AsInvokePolymorphic()->CanTargetInvokeVirtual()) {
+    return;
+  }
+  ArenaAllocator* allocator = invoke->GetBlock()->GetGraph()->GetAllocator();
+  LocationSummary* locations = new (allocator)
+      LocationSummary(invoke, LocationSummary::kCallOnMainAndSlowPath, kIntrinsified);
+
+  InvokeDexCallingConventionVisitorX86_64 calling_convention;
+  locations->SetOut(calling_convention.GetReturnLocation(invoke->GetType()));
+
+  locations->SetInAt(0, Location::RequiresRegister());
+
+  // Accomodating LocationSummary for underlying invoke-* call.
+  uint32_t number_of_args = invoke->GetNumberOfArguments();
+  for (uint32_t i = 1; i < number_of_args; ++i) {
+    locations->SetInAt(i, calling_convention.GetNextLocation(invoke->InputAt(i)->GetType()));
+  }
+
+  // The last input is MethodType object corresponding to the call-site.
+  locations->SetInAt(number_of_args, Location::RequiresRegister());
+
+  // We use a fixed-register temporary to pass the target method.
+  locations->AddTemp(calling_convention.GetMethodLocation());
+  locations->AddTemp(Location::RequiresRegister());
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitMethodHandleInvokeExact(HInvoke* invoke) {
+  DCHECK(invoke->AsInvokePolymorphic()->CanTargetInvokeVirtual());
+  LocationSummary* locations = invoke->GetLocations();
+
+  CpuRegister method_handle = locations->InAt(0).AsRegister<CpuRegister>();
+
+  SlowPathCode* slow_path =
+      new (codegen_->GetScopedAllocator()) InvokePolymorphicSlowPathX86_64(invoke, method_handle);
+  codegen_->AddSlowPath(slow_path);
+  X86_64Assembler* assembler = codegen_->GetAssembler();
+
+  Address method_handle_kind = Address(method_handle, mirror::MethodHandle::HandleKindOffset());
+
+  // If it is not InvokeVirtual then go to slow path.
+  // Even if MethodHandle's kind is kInvokeVirtual underlying method still can be an interface or
+  // direct method (that's what current `MethodHandles$Lookup.findVirtual` is doing). We don't check
+  // whether `method` is an interface method explicitly: in that case the subtype check will fail.
+  // TODO(b/297147201): check whether it can be more precise and what d8/r8 can produce.
+  __ cmpl(method_handle_kind, Immediate(mirror::MethodHandle::Kind::kInvokeVirtual));
+  __ j(kNotEqual, slow_path->GetEntryLabel());
+
+  CpuRegister call_site_type =
+      locations->InAt(invoke->GetNumberOfArguments()).AsRegister<CpuRegister>();
+
+  // Call site should match with MethodHandle's type.
+  __ cmpl(call_site_type, Address(method_handle, mirror::MethodHandle::MethodTypeOffset()));
+  __ j(kNotEqual, slow_path->GetEntryLabel());
+
+  CpuRegister method = locations->GetTemp(0).AsRegister<CpuRegister>();
+
+  // Find method to call.
+  __ movq(method, Address(method_handle, mirror::MethodHandle::ArtFieldOrMethodOffset()));
+
+  CpuRegister receiver = locations->InAt(1).AsRegister<CpuRegister>();
+
+  // Using vtable_index register as temporary in subtype check. It will be overridden later.
+  // If `method` is an interface method this check will fail.
+  CpuRegister vtable_index = locations->GetTemp(1).AsRegister<CpuRegister>();
+  GenerateSubTypeObjectCheckNoReadBarrier(codegen_,
+                                          slow_path,
+                                          receiver,
+                                          vtable_index,
+                                          Address(method, ArtMethod::DeclaringClassOffset()));
+
+  NearLabel execute_target_method;
+  // Skip virtual dispatch if `method` is private.
+  __ testl(Address(method, ArtMethod::AccessFlagsOffset()), Immediate(kAccPrivate));
+  __ j(kNotZero, &execute_target_method);
+
+  // MethodIndex is uint16_t.
+  __ movzxw(vtable_index, Address(method, ArtMethod::MethodIndexOffset()));
+
+  constexpr uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
+  // Re-using method register for receiver class.
+  __ movl(method, Address(receiver, class_offset));
+
+  constexpr uint32_t vtable_offset =
+      mirror::Class::EmbeddedVTableOffset(art::PointerSize::k64).Int32Value();
+  __ movq(method, Address(method, vtable_index, TIMES_8, vtable_offset));
+
+  __ Bind(&execute_target_method);
+  __ call(Address(
+      method,
+      ArtMethod::EntryPointFromQuickCompiledCodeOffset(art::PointerSize::k64).SizeValue()));
+  codegen_->RecordPcInfo(invoke, invoke->GetDexPc(), slow_path);
+  __ Bind(slow_path->GetExitLabel());
+}
+
 void IntrinsicLocationsBuilderX86_64::VisitVarHandleGet(HInvoke* invoke) {
   CreateVarHandleGetLocations(invoke, codegen_);
 }
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index 23cc3704e9..ffddd25843 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -4912,6 +4912,7 @@ class HInvokePolymorphic final : public HInvoke {
  public:
   HInvokePolymorphic(ArenaAllocator* allocator,
                      uint32_t number_of_arguments,
+                     uint32_t number_of_other_inputs,
                      DataType::Type return_type,
                      uint32_t dex_pc,
                      MethodReference method_reference,
@@ -4924,7 +4925,7 @@ class HInvokePolymorphic final : public HInvoke {
       : HInvoke(kInvokePolymorphic,
                 allocator,
                 number_of_arguments,
-                /* number_of_other_inputs= */ 0u,
+                number_of_other_inputs,
                 return_type,
                 dex_pc,
                 method_reference,
@@ -4938,6 +4939,12 @@ class HInvokePolymorphic final : public HInvoke {
 
   dex::ProtoIndex GetProtoIndex() { return proto_idx_; }
 
+  bool CanTargetInvokeVirtual() const {
+    return GetIntrinsic() == Intrinsics::kMethodHandleInvokeExact &&
+        GetNumberOfArguments() >= 2 &&
+        InputAt(1)->GetType() == DataType::Type::kReference;
+  }
+
   DECLARE_INSTRUCTION(InvokePolymorphic);
 
  protected:
diff --git a/compiler/utils/assembler_thumb_test_expected.cc.inc b/compiler/utils/assembler_thumb_test_expected.cc.inc
index f96a9bc154..5184b2c897 100644
--- a/compiler/utils/assembler_thumb_test_expected.cc.inc
+++ b/compiler/utils/assembler_thumb_test_expected.cc.inc
@@ -154,7 +154,7 @@ const char* const VixlJniHelpersResults = {
   "     210: f8d9 8020     ldr.w r8, [r9, #32]\n"
   "     214: 4770          bx lr\n"
   "     216: f8d9 0094     ldr.w r0, [r9, #148]\n"
-  "     21a: f8d9 e2bc     ldr.w lr, [r9, #700]\n"
+  "     21a: f8d9 e2c0     ldr.w lr, [r9, #704]\n"
   "     21e: 47f0          blx lr\n"
 };