Change interface conflict stub to take the interface method.
To avoid doing dex cache lookup, pass the interface method instead. This
costs a few hundred KBs on speed compiled APKs (< 0.5% code size), but
improves performance when hitting a conflict (as seen on dogfood data).
For nterp, we currently pass the conflict method instead of the
interface method. We need to handle default methods before optimizing
it.
This removes our last use of dex cache in compiled code. A follow-up CL
will remove the NeedsDexCacheOfDeclaringClass from HInvokeInterface.
Test: test.py
Change-Id: I3cdd4543ad7d904b3e81950af46a48a48af6991a
diff --git a/compiler/optimizing/code_generator_arm_vixl.cc b/compiler/optimizing/code_generator_arm_vixl.cc
index 0940f42..85337ed 100644
--- a/compiler/optimizing/code_generator_arm_vixl.cc
+++ b/compiler/optimizing/code_generator_arm_vixl.cc
@@ -3432,7 +3432,10 @@
void LocationsBuilderARMVIXL::VisitInvokeInterface(HInvokeInterface* invoke) {
HandleInvoke(invoke);
// Add the hidden argument.
- invoke->GetLocations()->AddTemp(LocationFrom(r12));
+ if (invoke->GetHiddenArgumentLoadKind() == MethodLoadKind::kRecursive) {
+ // We cannot request r12 as it's blocked by the register allocator.
+ invoke->GetLocations()->SetInAt(invoke->GetNumberOfArguments() - 1, Location::Any());
+ }
}
void CodeGeneratorARMVIXL::MaybeGenerateInlineCacheCheck(HInstruction* instruction,
@@ -3468,7 +3471,6 @@
// TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError.
LocationSummary* locations = invoke->GetLocations();
vixl32::Register temp = RegisterFrom(locations->GetTemp(0));
- vixl32::Register hidden_reg = RegisterFrom(locations->GetTemp(1));
Location receiver = locations->InAt(0);
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
@@ -3509,11 +3511,10 @@
// LR = temp->GetEntryPoint();
GetAssembler()->LoadFromOffset(kLoadWord, lr, temp, entry_point);
- // Set the hidden (in r12) argument. It is done here, right before a BLX to prevent other
- // instruction from clobbering it as they might use r12 as a scratch register.
- DCHECK(hidden_reg.Is(r12));
-
{
+ // Set the hidden (in r12) argument. It is done here, right before a BLX to prevent other
+ // instruction from clobbering it as they might use r12 as a scratch register.
+ Location hidden_reg = Location::RegisterLocation(r12.GetCode());
// The VIXL macro assembler may clobber any of the scratch registers that are available to it,
// so it checks if the application is using them (by passing them to the macro assembler
// methods). The following application of UseScratchRegisterScope corrects VIXL's notion of
@@ -3523,8 +3524,18 @@
// (to materialize the constant), since the destination register becomes available for such use
// internally for the duration of the macro instruction.
UseScratchRegisterScope temps(GetVIXLAssembler());
- temps.Exclude(hidden_reg);
- __ Mov(hidden_reg, invoke->GetMethodReference().index);
+ temps.Exclude(RegisterFrom(hidden_reg));
+ if (invoke->GetHiddenArgumentLoadKind() == MethodLoadKind::kRecursive) {
+ Location current_method = locations->InAt(invoke->GetNumberOfArguments() - 1);
+ if (current_method.IsStackSlot()) {
+ GetAssembler()->LoadFromOffset(
+ kLoadWord, RegisterFrom(hidden_reg), sp, current_method.GetStackIndex());
+ } else {
+ __ Mov(RegisterFrom(hidden_reg), RegisterFrom(current_method));
+ }
+ } else {
+ codegen_->LoadMethod(invoke->GetHiddenArgumentLoadKind(), hidden_reg, invoke);
+ }
}
{
// Ensure the pc position is recorded immediately after the `blx` instruction.
@@ -9069,20 +9080,9 @@
return desired_dispatch_info;
}
-void CodeGeneratorARMVIXL::GenerateStaticOrDirectCall(
- HInvokeStaticOrDirect* invoke, Location temp, SlowPathCode* slow_path) {
- Location callee_method = temp; // For all kinds except kRecursive, callee will be in temp.
- switch (invoke->GetMethodLoadKind()) {
- case MethodLoadKind::kStringInit: {
- uint32_t offset =
- GetThreadOffset<kArmPointerSize>(invoke->GetStringInitEntryPoint()).Int32Value();
- // temp = thread->string_init_entrypoint
- GetAssembler()->LoadFromOffset(kLoadWord, RegisterFrom(temp), tr, offset);
- break;
- }
- case MethodLoadKind::kRecursive:
- callee_method = invoke->GetLocations()->InAt(invoke->GetCurrentMethodIndex());
- break;
+
+void CodeGeneratorARMVIXL::LoadMethod(MethodLoadKind load_kind, Location temp, HInvoke* invoke) {
+ switch (load_kind) {
case MethodLoadKind::kBootImageLinkTimePcRelative: {
DCHECK(GetCompilerOptions().IsBootImage() || GetCompilerOptions().IsBootImageExtension());
PcRelativePatchInfo* labels = NewBootImageMethodPatch(invoke->GetResolvedMethodReference());
@@ -9106,13 +9106,44 @@
GetAssembler()->LoadFromOffset(kLoadWord, temp_reg, temp_reg, /* offset*/ 0);
break;
}
- case MethodLoadKind::kJitDirectAddress:
- __ Mov(RegisterFrom(temp), Operand::From(invoke->GetMethodAddress()));
+ case MethodLoadKind::kJitDirectAddress: {
+ __ Mov(RegisterFrom(temp), Operand::From(invoke->GetResolvedMethod()));
break;
+ }
+ case MethodLoadKind::kRuntimeCall: {
+ // Test situation, don't do anything.
+ break;
+ }
+ default: {
+ LOG(FATAL) << "Load kind should have already been handled " << load_kind;
+ UNREACHABLE();
+ }
+ }
+}
+
+void CodeGeneratorARMVIXL::GenerateStaticOrDirectCall(
+ HInvokeStaticOrDirect* invoke, Location temp, SlowPathCode* slow_path) {
+ Location callee_method = temp; // For all kinds except kRecursive, callee will be in temp.
+ switch (invoke->GetMethodLoadKind()) {
+ case MethodLoadKind::kStringInit: {
+ uint32_t offset =
+ GetThreadOffset<kArmPointerSize>(invoke->GetStringInitEntryPoint()).Int32Value();
+ // temp = thread->string_init_entrypoint
+ GetAssembler()->LoadFromOffset(kLoadWord, RegisterFrom(temp), tr, offset);
+ break;
+ }
+ case MethodLoadKind::kRecursive: {
+ callee_method = invoke->GetLocations()->InAt(invoke->GetCurrentMethodIndex());
+ break;
+ }
case MethodLoadKind::kRuntimeCall: {
GenerateInvokeStaticOrDirectRuntimeCall(invoke, temp, slow_path);
return; // No code pointer retrieval; the runtime performs the call directly.
}
+ default: {
+ LoadMethod(invoke->GetMethodLoadKind(), temp, invoke);
+ break;
+ }
}
auto call_code_pointer_member = [&](MemberOffset offset) {