Basic SIMD reduction support.
Rationale:
Enables vectorization of x += .... for very basic (simple, same-type)
constructs. Paves the way for more complex (narrower and/or mixed-type)
constructs, which will be handled by the next CL.
This is a revert of Icb5d6c805516db0a1d911c3ede9a246ccef89a22
and thus a revert^2 of I2454778dd0ef1da915c178c7274e1cf33e271d0f
and thus a revert^3 of I1c1c87b6323e01442e8fbd94869ddc9e760ea1fc
and thus a revert^4 of I7880c135aee3ed0a39da9ae5b468cbf80e613766
PS1-2 shows what needed to change
Test: test-art-host test-art-target
Bug: 64091002
Change-Id: I647889e0da0959ca405b70081b79c7d3c9bcb2e9
diff --git a/compiler/optimizing/code_generator_vector_arm64.cc b/compiler/optimizing/code_generator_vector_arm64.cc
index 9095ecd..18a55c8 100644
--- a/compiler/optimizing/code_generator_vector_arm64.cc
+++ b/compiler/optimizing/code_generator_vector_arm64.cc
@@ -27,12 +27,13 @@
using helpers::ARM64EncodableConstantOrRegister;
using helpers::Arm64CanEncodeConstantAsImmediate;
using helpers::DRegisterFrom;
-using helpers::VRegisterFrom;
using helpers::HeapOperand;
using helpers::InputRegisterAt;
using helpers::Int64ConstantFrom;
-using helpers::XRegisterFrom;
+using helpers::OutputRegister;
+using helpers::VRegisterFrom;
using helpers::WRegisterFrom;
+using helpers::XRegisterFrom;
#define __ GetVIXLAssembler()->
@@ -127,20 +128,51 @@
}
}
-void LocationsBuilderARM64::VisitVecSetScalars(HVecSetScalars* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+void LocationsBuilderARM64::VisitVecExtractScalar(HVecExtractScalar* instruction) {
+ LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction);
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimBoolean:
+ case Primitive::kPrimByte:
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ case Primitive::kPrimInt:
+ case Primitive::kPrimLong:
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ locations->SetOut(Location::RequiresRegister());
+ break;
+ case Primitive::kPrimFloat:
+ case Primitive::kPrimDouble:
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ locations->SetOut(Location::SameAsFirstInput());
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
-void InstructionCodeGeneratorARM64::VisitVecSetScalars(HVecSetScalars* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
-}
-
-void LocationsBuilderARM64::VisitVecSumReduce(HVecSumReduce* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
-}
-
-void InstructionCodeGeneratorARM64::VisitVecSumReduce(HVecSumReduce* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+void InstructionCodeGeneratorARM64::VisitVecExtractScalar(HVecExtractScalar* instruction) {
+ LocationSummary* locations = instruction->GetLocations();
+ VRegister src = VRegisterFrom(locations->InAt(0));
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ Umov(OutputRegister(instruction), src.V4S(), 0);
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ Umov(OutputRegister(instruction), src.V2D(), 0);
+ break;
+ case Primitive::kPrimFloat:
+ case Primitive::kPrimDouble:
+ DCHECK_LE(2u, instruction->GetVectorLength());
+ DCHECK_LE(instruction->GetVectorLength(), 4u);
+ DCHECK(locations->InAt(0).Equals(locations->Out())); // no code required
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
// Helper to set up locations for vector unary operations.
@@ -169,6 +201,46 @@
}
}
+void LocationsBuilderARM64::VisitVecReduce(HVecReduce* instruction) {
+ CreateVecUnOpLocations(GetGraph()->GetArena(), instruction);
+}
+
+void InstructionCodeGeneratorARM64::VisitVecReduce(HVecReduce* instruction) {
+ LocationSummary* locations = instruction->GetLocations();
+ VRegister src = VRegisterFrom(locations->InAt(0));
+ VRegister dst = DRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ switch (instruction->GetKind()) {
+ case HVecReduce::kSum:
+ __ Addv(dst.S(), src.V4S());
+ break;
+ case HVecReduce::kMin:
+ __ Sminv(dst.S(), src.V4S());
+ break;
+ case HVecReduce::kMax:
+ __ Smaxv(dst.S(), src.V4S());
+ break;
+ }
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ switch (instruction->GetKind()) {
+ case HVecReduce::kSum:
+ __ Addp(dst.D(), src.V2D());
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD min/max";
+ UNREACHABLE();
+ }
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
+}
+
void LocationsBuilderARM64::VisitVecCnv(HVecCnv* instruction) {
CreateVecUnOpLocations(GetGraph()->GetArena(), instruction);
}
@@ -263,6 +335,7 @@
break;
default:
LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
}
}
@@ -805,6 +878,77 @@
}
}
+void LocationsBuilderARM64::VisitVecSetScalars(HVecSetScalars* instruction) {
+ LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction);
+
+ DCHECK_EQ(1u, instruction->InputCount()); // only one input currently implemented
+
+ HInstruction* input = instruction->InputAt(0);
+ bool is_zero = IsZeroBitPattern(input);
+
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimBoolean:
+ case Primitive::kPrimByte:
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ case Primitive::kPrimInt:
+ case Primitive::kPrimLong:
+ locations->SetInAt(0, is_zero ? Location::ConstantLocation(input->AsConstant())
+ : Location::RequiresRegister());
+ locations->SetOut(Location::RequiresFpuRegister());
+ break;
+ case Primitive::kPrimFloat:
+ case Primitive::kPrimDouble:
+ locations->SetInAt(0, is_zero ? Location::ConstantLocation(input->AsConstant())
+ : Location::RequiresFpuRegister());
+ locations->SetOut(Location::RequiresFpuRegister());
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
+}
+
+void InstructionCodeGeneratorARM64::VisitVecSetScalars(HVecSetScalars* instruction) {
+ LocationSummary* locations = instruction->GetLocations();
+ VRegister dst = VRegisterFrom(locations->Out());
+
+ DCHECK_EQ(1u, instruction->InputCount()); // only one input currently implemented
+
+ // Zero out all other elements first.
+ __ Movi(dst.V16B(), 0);
+
+ // Shorthand for any type of zero.
+ if (IsZeroBitPattern(instruction->InputAt(0))) {
+ return;
+ }
+
+ // Set required elements.
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimBoolean:
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ Mov(dst.V16B(), 0, InputRegisterAt(instruction, 0));
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ Mov(dst.V8H(), 0, InputRegisterAt(instruction, 0));
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ Mov(dst.V4S(), 0, InputRegisterAt(instruction, 0));
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ Mov(dst.V2D(), 0, InputRegisterAt(instruction, 0));
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
+}
+
void LocationsBuilderARM64::VisitVecMultiplyAccumulate(HVecMultiplyAccumulate* instr) {
LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instr);
switch (instr->GetPackedType()) {