MIPS32: ART Vectorizer
MIPS32 implementation which uses MSA extension.
Note: Testing is done with checker parts of tests 640, 645, 646 and
651, locally changed to cover MIPS32 cases. These changes can't
be included in this patch since MSA is not a default option.
Test: ./testrunner.py --target --optimizing -j1 in QEMU (mips32r6)
Change-Id: Ieba28f94c48c943d5444017bede9a5d409149762
diff --git a/compiler/optimizing/code_generator_mips.h b/compiler/optimizing/code_generator_mips.h
index 1afa1b9..52ee852 100644
--- a/compiler/optimizing/code_generator_mips.h
+++ b/compiler/optimizing/code_generator_mips.h
@@ -346,6 +346,10 @@
uint32_t num_entries,
HBasicBlock* switch_block,
HBasicBlock* default_block);
+
+ int32_t VecAddress(LocationSummary* locations,
+ size_t size,
+ /* out */ Register* adjusted_base);
void GenConditionalMoveR2(HSelect* select);
void GenConditionalMoveR6(HSelect* select);
diff --git a/compiler/optimizing/code_generator_vector_mips.cc b/compiler/optimizing/code_generator_vector_mips.cc
index c4a3225..ea36e90 100644
--- a/compiler/optimizing/code_generator_vector_mips.cc
+++ b/compiler/optimizing/code_generator_vector_mips.cc
@@ -15,6 +15,7 @@
*/
#include "code_generator_mips.h"
+#include "mirror/array-inl.h"
namespace art {
namespace mips {
@@ -23,11 +24,68 @@
#define __ down_cast<MipsAssembler*>(GetAssembler())-> // NOLINT
void LocationsBuilderMIPS::VisitVecReplicateScalar(HVecReplicateScalar* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ 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::RequiresRegister());
+ locations->SetOut(Location::RequiresFpuRegister());
+ break;
+ case Primitive::kPrimFloat:
+ case Primitive::kPrimDouble:
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void InstructionCodeGeneratorMIPS::VisitVecReplicateScalar(HVecReplicateScalar* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimBoolean:
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ FillB(dst, locations->InAt(0).AsRegister<Register>());
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ FillH(dst, locations->InAt(0).AsRegister<Register>());
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ FillW(dst, locations->InAt(0).AsRegister<Register>());
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ Mtc1(locations->InAt(0).AsRegisterPairLow<Register>(), FTMP);
+ __ MoveToFpuHigh(locations->InAt(0).AsRegisterPairHigh<Register>(), FTMP);
+ __ ReplicateFPToVectorRegister(dst, FTMP, /* is_double */ true);
+ break;
+ case Primitive::kPrimFloat:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ ReplicateFPToVectorRegister(dst,
+ locations->InAt(0).AsFpuRegister<FRegister>(),
+ /* is_double */ false);
+ break;
+ case Primitive::kPrimDouble:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ ReplicateFPToVectorRegister(dst,
+ locations->InAt(0).AsFpuRegister<FRegister>(),
+ /* is_double */ true);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecSetScalars(HVecSetScalars* instruction) {
@@ -51,13 +109,23 @@
LocationSummary* locations = new (arena) LocationSummary(instruction);
switch (instruction->GetPackedType()) {
case Primitive::kPrimBoolean:
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ locations->SetOut(Location::RequiresFpuRegister(),
+ instruction->IsVecNot() ? Location::kOutputOverlap
+ : Location::kNoOutputOverlap);
+ break;
case Primitive::kPrimByte:
case Primitive::kPrimChar:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
+ case Primitive::kPrimLong:
case Primitive::kPrimFloat:
case Primitive::kPrimDouble:
- DCHECK(locations);
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ locations->SetOut(Location::RequiresFpuRegister(),
+ (instruction->IsVecNeg() || instruction->IsVecAbs())
+ ? Location::kOutputOverlap
+ : Location::kNoOutputOverlap);
break;
default:
LOG(FATAL) << "Unsupported SIMD type";
@@ -70,7 +138,17 @@
}
void InstructionCodeGeneratorMIPS::VisitVecCnv(HVecCnv* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister src = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ Primitive::Type from = instruction->GetInputType();
+ Primitive::Type to = instruction->GetResultType();
+ if (from == Primitive::kPrimInt && to == Primitive::kPrimFloat) {
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ Ffint_sW(dst, src);
+ } else {
+ LOG(FATAL) << "Unsupported SIMD type";
+ }
}
void LocationsBuilderMIPS::VisitVecNeg(HVecNeg* instruction) {
@@ -78,7 +156,45 @@
}
void InstructionCodeGeneratorMIPS::VisitVecNeg(HVecNeg* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister src = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ FillB(dst, ZERO);
+ __ SubvB(dst, dst, src);
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ FillH(dst, ZERO);
+ __ SubvH(dst, dst, src);
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ FillW(dst, ZERO);
+ __ SubvW(dst, dst, src);
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ FillW(dst, ZERO);
+ __ SubvD(dst, dst, src);
+ break;
+ case Primitive::kPrimFloat:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ FillW(dst, ZERO);
+ __ FsubW(dst, dst, src);
+ break;
+ case Primitive::kPrimDouble:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ FillW(dst, ZERO);
+ __ FsubD(dst, dst, src);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecAbs(HVecAbs* instruction) {
@@ -86,7 +202,47 @@
}
void InstructionCodeGeneratorMIPS::VisitVecAbs(HVecAbs* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister src = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ FillB(dst, ZERO); // all zeroes
+ __ Add_aB(dst, dst, src); // dst = abs(0) + abs(src)
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ FillH(dst, ZERO); // all zeroes
+ __ Add_aH(dst, dst, src); // dst = abs(0) + abs(src)
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ FillW(dst, ZERO); // all zeroes
+ __ Add_aW(dst, dst, src); // dst = abs(0) + abs(src)
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ FillW(dst, ZERO); // all zeroes
+ __ Add_aD(dst, dst, src); // dst = abs(0) + abs(src)
+ break;
+ case Primitive::kPrimFloat:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ LdiW(dst, -1); // all ones
+ __ SrliW(dst, dst, 1);
+ __ AndV(dst, dst, src);
+ break;
+ case Primitive::kPrimDouble:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ LdiD(dst, -1); // all ones
+ __ SrliD(dst, dst, 1);
+ __ AndV(dst, dst, src);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecNot(HVecNot* instruction) {
@@ -94,7 +250,30 @@
}
void InstructionCodeGeneratorMIPS::VisitVecNot(HVecNot* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister src = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimBoolean: // special case boolean-not
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ LdiB(dst, 1);
+ __ XorV(dst, dst, src);
+ break;
+ case Primitive::kPrimByte:
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ case Primitive::kPrimInt:
+ case Primitive::kPrimLong:
+ case Primitive::kPrimFloat:
+ case Primitive::kPrimDouble:
+ DCHECK_LE(2u, instruction->GetVectorLength());
+ DCHECK_LE(instruction->GetVectorLength(), 16u);
+ __ NorV(dst, src, src); // lanes do not matter
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
// Helper to set up locations for vector binary operations.
@@ -106,9 +285,12 @@
case Primitive::kPrimChar:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
+ case Primitive::kPrimLong:
case Primitive::kPrimFloat:
case Primitive::kPrimDouble:
- DCHECK(locations);
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ locations->SetInAt(1, Location::RequiresFpuRegister());
+ locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
break;
default:
LOG(FATAL) << "Unsupported SIMD type";
@@ -121,7 +303,40 @@
}
void InstructionCodeGeneratorMIPS::VisitVecAdd(HVecAdd* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister rhs = VectorRegisterFrom(locations->InAt(1));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ AddvB(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ AddvH(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ AddvW(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ AddvD(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimFloat:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ FaddW(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimDouble:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ FaddD(dst, lhs, rhs);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecHalvingAdd(HVecHalvingAdd* instruction) {
@@ -129,7 +344,40 @@
}
void InstructionCodeGeneratorMIPS::VisitVecHalvingAdd(HVecHalvingAdd* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister rhs = VectorRegisterFrom(locations->InAt(1));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ if (instruction->IsUnsigned()) {
+ instruction->IsRounded()
+ ? __ Aver_uB(dst, lhs, rhs)
+ : __ Ave_uB(dst, lhs, rhs);
+ } else {
+ instruction->IsRounded()
+ ? __ Aver_sB(dst, lhs, rhs)
+ : __ Ave_sB(dst, lhs, rhs);
+ }
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ if (instruction->IsUnsigned()) {
+ instruction->IsRounded()
+ ? __ Aver_uH(dst, lhs, rhs)
+ : __ Ave_uH(dst, lhs, rhs);
+ } else {
+ instruction->IsRounded()
+ ? __ Aver_sH(dst, lhs, rhs)
+ : __ Ave_sH(dst, lhs, rhs);
+ }
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecSub(HVecSub* instruction) {
@@ -137,7 +385,40 @@
}
void InstructionCodeGeneratorMIPS::VisitVecSub(HVecSub* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister rhs = VectorRegisterFrom(locations->InAt(1));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ SubvB(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ SubvH(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ SubvW(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ SubvD(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimFloat:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ FsubW(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimDouble:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ FsubD(dst, lhs, rhs);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecMul(HVecMul* instruction) {
@@ -145,7 +426,40 @@
}
void InstructionCodeGeneratorMIPS::VisitVecMul(HVecMul* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister rhs = VectorRegisterFrom(locations->InAt(1));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ MulvB(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ MulvH(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ MulvW(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ MulvD(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimFloat:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ FmulW(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimDouble:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ FmulD(dst, lhs, rhs);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecDiv(HVecDiv* instruction) {
@@ -153,7 +467,23 @@
}
void InstructionCodeGeneratorMIPS::VisitVecDiv(HVecDiv* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister rhs = VectorRegisterFrom(locations->InAt(1));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimFloat:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ FdivW(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimDouble:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ FdivD(dst, lhs, rhs);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecMin(HVecMin* instruction) {
@@ -161,7 +491,60 @@
}
void InstructionCodeGeneratorMIPS::VisitVecMin(HVecMin* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister rhs = VectorRegisterFrom(locations->InAt(1));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ if (instruction->IsUnsigned()) {
+ __ Min_uB(dst, lhs, rhs);
+ } else {
+ __ Min_sB(dst, lhs, rhs);
+ }
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ if (instruction->IsUnsigned()) {
+ __ Min_uH(dst, lhs, rhs);
+ } else {
+ __ Min_sH(dst, lhs, rhs);
+ }
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ if (instruction->IsUnsigned()) {
+ __ Min_uW(dst, lhs, rhs);
+ } else {
+ __ Min_sW(dst, lhs, rhs);
+ }
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ if (instruction->IsUnsigned()) {
+ __ Min_uD(dst, lhs, rhs);
+ } else {
+ __ Min_sD(dst, lhs, rhs);
+ }
+ break;
+ // When one of arguments is NaN, fmin.df returns other argument, but Java expects a NaN value.
+ // TODO: Fix min(x, NaN) cases for float and double.
+ case Primitive::kPrimFloat:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ DCHECK(!instruction->IsUnsigned());
+ __ FminW(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimDouble:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ DCHECK(!instruction->IsUnsigned());
+ __ FminD(dst, lhs, rhs);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecMax(HVecMax* instruction) {
@@ -169,7 +552,60 @@
}
void InstructionCodeGeneratorMIPS::VisitVecMax(HVecMax* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister rhs = VectorRegisterFrom(locations->InAt(1));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ if (instruction->IsUnsigned()) {
+ __ Max_uB(dst, lhs, rhs);
+ } else {
+ __ Max_sB(dst, lhs, rhs);
+ }
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ if (instruction->IsUnsigned()) {
+ __ Max_uH(dst, lhs, rhs);
+ } else {
+ __ Max_sH(dst, lhs, rhs);
+ }
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ if (instruction->IsUnsigned()) {
+ __ Max_uW(dst, lhs, rhs);
+ } else {
+ __ Max_sW(dst, lhs, rhs);
+ }
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ if (instruction->IsUnsigned()) {
+ __ Max_uD(dst, lhs, rhs);
+ } else {
+ __ Max_sD(dst, lhs, rhs);
+ }
+ break;
+ // When one of arguments is NaN, fmax.df returns other argument, but Java expects a NaN value.
+ // TODO: Fix max(x, NaN) cases for float and double.
+ case Primitive::kPrimFloat:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ DCHECK(!instruction->IsUnsigned());
+ __ FmaxW(dst, lhs, rhs);
+ break;
+ case Primitive::kPrimDouble:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ DCHECK(!instruction->IsUnsigned());
+ __ FmaxD(dst, lhs, rhs);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecAnd(HVecAnd* instruction) {
@@ -177,7 +613,27 @@
}
void InstructionCodeGeneratorMIPS::VisitVecAnd(HVecAnd* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister rhs = VectorRegisterFrom(locations->InAt(1));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimBoolean:
+ case Primitive::kPrimByte:
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ case Primitive::kPrimInt:
+ case Primitive::kPrimLong:
+ case Primitive::kPrimFloat:
+ case Primitive::kPrimDouble:
+ DCHECK_LE(2u, instruction->GetVectorLength());
+ DCHECK_LE(instruction->GetVectorLength(), 16u);
+ __ AndV(dst, lhs, rhs); // lanes do not matter
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecAndNot(HVecAndNot* instruction) {
@@ -193,7 +649,27 @@
}
void InstructionCodeGeneratorMIPS::VisitVecOr(HVecOr* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister rhs = VectorRegisterFrom(locations->InAt(1));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimBoolean:
+ case Primitive::kPrimByte:
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ case Primitive::kPrimInt:
+ case Primitive::kPrimLong:
+ case Primitive::kPrimFloat:
+ case Primitive::kPrimDouble:
+ DCHECK_LE(2u, instruction->GetVectorLength());
+ DCHECK_LE(instruction->GetVectorLength(), 16u);
+ __ OrV(dst, lhs, rhs); // lanes do not matter
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecXor(HVecXor* instruction) {
@@ -201,7 +677,27 @@
}
void InstructionCodeGeneratorMIPS::VisitVecXor(HVecXor* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister rhs = VectorRegisterFrom(locations->InAt(1));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimBoolean:
+ case Primitive::kPrimByte:
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ case Primitive::kPrimInt:
+ case Primitive::kPrimLong:
+ case Primitive::kPrimFloat:
+ case Primitive::kPrimDouble:
+ DCHECK_LE(2u, instruction->GetVectorLength());
+ DCHECK_LE(instruction->GetVectorLength(), 16u);
+ __ XorV(dst, lhs, rhs); // lanes do not matter
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
// Helper to set up locations for vector shift operations.
@@ -213,7 +709,9 @@
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimLong:
- DCHECK(locations);
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ locations->SetInAt(1, Location::ConstantLocation(instruction->InputAt(1)->AsConstant()));
+ locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
break;
default:
LOG(FATAL) << "Unsupported SIMD type";
@@ -226,7 +724,32 @@
}
void InstructionCodeGeneratorMIPS::VisitVecShl(HVecShl* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ int32_t value = locations->InAt(1).GetConstant()->AsIntConstant()->GetValue();
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ SlliB(dst, lhs, value);
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ SlliH(dst, lhs, value);
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ SlliW(dst, lhs, value);
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ SlliD(dst, lhs, value);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecShr(HVecShr* instruction) {
@@ -234,7 +757,32 @@
}
void InstructionCodeGeneratorMIPS::VisitVecShr(HVecShr* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ int32_t value = locations->InAt(1).GetConstant()->AsIntConstant()->GetValue();
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ SraiB(dst, lhs, value);
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ SraiH(dst, lhs, value);
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ SraiW(dst, lhs, value);
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ SraiD(dst, lhs, value);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecUShr(HVecUShr* instruction) {
@@ -242,7 +790,32 @@
}
void InstructionCodeGeneratorMIPS::VisitVecUShr(HVecUShr* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ VectorRegister lhs = VectorRegisterFrom(locations->InAt(0));
+ VectorRegister dst = VectorRegisterFrom(locations->Out());
+ int32_t value = locations->InAt(1).GetConstant()->AsIntConstant()->GetValue();
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ SrliB(dst, lhs, value);
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ SrliH(dst, lhs, value);
+ break;
+ case Primitive::kPrimInt:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ SrliW(dst, lhs, value);
+ break;
+ case Primitive::kPrimLong:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ SrliD(dst, lhs, value);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecMultiplyAccumulate(HVecMultiplyAccumulate* instr) {
@@ -253,20 +826,143 @@
LOG(FATAL) << "No SIMD for " << instr->GetId();
}
+// Helper to set up locations for vector memory operations.
+static void CreateVecMemLocations(ArenaAllocator* arena,
+ HVecMemoryOperation* instruction,
+ bool is_load) {
+ LocationSummary* locations = new (arena) LocationSummary(instruction);
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimBoolean:
+ case Primitive::kPrimByte:
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ case Primitive::kPrimInt:
+ case Primitive::kPrimLong:
+ case Primitive::kPrimFloat:
+ case Primitive::kPrimDouble:
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1)));
+ if (is_load) {
+ locations->SetOut(Location::RequiresFpuRegister());
+ } else {
+ locations->SetInAt(2, Location::RequiresFpuRegister());
+ }
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
+}
+
+// Helper to prepare register and offset for vector memory operations. Returns the offset and sets
+// the output parameter adjusted_base to the original base or to a reserved temporary register (AT).
+int32_t InstructionCodeGeneratorMIPS::VecAddress(LocationSummary* locations,
+ size_t size,
+ /* out */ Register* adjusted_base) {
+ Register base = locations->InAt(0).AsRegister<Register>();
+ Location index = locations->InAt(1);
+ int scale = TIMES_1;
+ switch (size) {
+ case 2: scale = TIMES_2; break;
+ case 4: scale = TIMES_4; break;
+ case 8: scale = TIMES_8; break;
+ default: break;
+ }
+ int32_t offset = mirror::Array::DataOffset(size).Int32Value();
+
+ if (index.IsConstant()) {
+ offset += index.GetConstant()->AsIntConstant()->GetValue() << scale;
+ __ AdjustBaseOffsetAndElementSizeShift(base, offset, scale);
+ *adjusted_base = base;
+ } else {
+ Register index_reg = index.AsRegister<Register>();
+ if (scale != TIMES_1) {
+ __ Lsa(AT, index_reg, base, scale);
+ } else {
+ __ Addu(AT, base, index_reg);
+ }
+ *adjusted_base = AT;
+ }
+ return offset;
+}
+
void LocationsBuilderMIPS::VisitVecLoad(HVecLoad* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ CreateVecMemLocations(GetGraph()->GetArena(), instruction, /* is_load */ true);
}
void InstructionCodeGeneratorMIPS::VisitVecLoad(HVecLoad* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ size_t size = Primitive::ComponentSize(instruction->GetPackedType());
+ VectorRegister reg = VectorRegisterFrom(locations->Out());
+ Register base;
+ int32_t offset = VecAddress(locations, size, &base);
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimBoolean:
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ LdB(reg, base, offset);
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ // Loading 8-bytes (needed if dealing with compressed strings in StringCharAt) from unaligned
+ // memory address may cause a trap to the kernel if the CPU doesn't directly support unaligned
+ // loads and stores.
+ // TODO: Implement support for StringCharAt.
+ DCHECK(!instruction->IsStringCharAt());
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ LdH(reg, base, offset);
+ break;
+ case Primitive::kPrimInt:
+ case Primitive::kPrimFloat:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ LdW(reg, base, offset);
+ break;
+ case Primitive::kPrimLong:
+ case Primitive::kPrimDouble:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ LdD(reg, base, offset);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
void LocationsBuilderMIPS::VisitVecStore(HVecStore* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ CreateVecMemLocations(GetGraph()->GetArena(), instruction, /* is_load */ false);
}
void InstructionCodeGeneratorMIPS::VisitVecStore(HVecStore* instruction) {
- LOG(FATAL) << "No SIMD for " << instruction->GetId();
+ LocationSummary* locations = instruction->GetLocations();
+ size_t size = Primitive::ComponentSize(instruction->GetPackedType());
+ VectorRegister reg = VectorRegisterFrom(locations->InAt(2));
+ Register base;
+ int32_t offset = VecAddress(locations, size, &base);
+ switch (instruction->GetPackedType()) {
+ case Primitive::kPrimBoolean:
+ case Primitive::kPrimByte:
+ DCHECK_EQ(16u, instruction->GetVectorLength());
+ __ StB(reg, base, offset);
+ break;
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ DCHECK_EQ(8u, instruction->GetVectorLength());
+ __ StH(reg, base, offset);
+ break;
+ case Primitive::kPrimInt:
+ case Primitive::kPrimFloat:
+ DCHECK_EQ(4u, instruction->GetVectorLength());
+ __ StW(reg, base, offset);
+ break;
+ case Primitive::kPrimLong:
+ case Primitive::kPrimDouble:
+ DCHECK_EQ(2u, instruction->GetVectorLength());
+ __ StD(reg, base, offset);
+ break;
+ default:
+ LOG(FATAL) << "Unsupported SIMD type";
+ UNREACHABLE();
+ }
}
#undef __
diff --git a/compiler/optimizing/loop_optimization.cc b/compiler/optimizing/loop_optimization.cc
index 83f31c7..422e58d 100644
--- a/compiler/optimizing/loop_optimization.cc
+++ b/compiler/optimizing/loop_optimization.cc
@@ -1171,7 +1171,32 @@
}
return false;
case kMips:
- // TODO: implement MIPS SIMD.
+ if (features->AsMipsInstructionSetFeatures()->HasMsa()) {
+ switch (type) {
+ case Primitive::kPrimBoolean:
+ case Primitive::kPrimByte:
+ *restrictions |= kNoDiv;
+ return TrySetVectorLength(16);
+ case Primitive::kPrimChar:
+ case Primitive::kPrimShort:
+ *restrictions |= kNoDiv | kNoStringCharAt;
+ return TrySetVectorLength(8);
+ case Primitive::kPrimInt:
+ *restrictions |= kNoDiv;
+ return TrySetVectorLength(4);
+ case Primitive::kPrimLong:
+ *restrictions |= kNoDiv;
+ return TrySetVectorLength(2);
+ case Primitive::kPrimFloat:
+ *restrictions |= kNoMinMax; // min/max(x, NaN)
+ return TrySetVectorLength(4);
+ case Primitive::kPrimDouble:
+ *restrictions |= kNoMinMax; // min/max(x, NaN)
+ return TrySetVectorLength(2);
+ default:
+ break;
+ } // switch type
+ }
return false;
case kMips64:
if (features->AsMips64InstructionSetFeatures()->HasMsa()) {
diff --git a/compiler/utils/mips/assembler_mips.cc b/compiler/utils/mips/assembler_mips.cc
index 44b9bb4..c581f1c 100644
--- a/compiler/utils/mips/assembler_mips.cc
+++ b/compiler/utils/mips/assembler_mips.cc
@@ -2904,6 +2904,17 @@
static_cast<FRegister>(wt));
}
+void MipsAssembler::ReplicateFPToVectorRegister(VectorRegister dst,
+ FRegister src,
+ bool is_double) {
+ // Float or double in FPU register Fx can be considered as 0th element in vector register Wx.
+ if (is_double) {
+ SplatiD(dst, static_cast<VectorRegister>(src), 0);
+ } else {
+ SplatiW(dst, static_cast<VectorRegister>(src), 0);
+ }
+}
+
void MipsAssembler::LoadConst32(Register rd, int32_t value) {
if (IsUint<16>(value)) {
// Use OR with (unsigned) immediate to encode 16b unsigned int.
diff --git a/compiler/utils/mips/assembler_mips.h b/compiler/utils/mips/assembler_mips.h
index a229882..33803bb 100644
--- a/compiler/utils/mips/assembler_mips.h
+++ b/compiler/utils/mips/assembler_mips.h
@@ -612,6 +612,9 @@
void IlvrW(VectorRegister wd, VectorRegister ws, VectorRegister wt);
void IlvrD(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ // Helper for replicating floating point value in all destination elements.
+ void ReplicateFPToVectorRegister(VectorRegister dst, FRegister src, bool is_double);
+
// Higher level composite instructions.
void LoadConst32(Register rd, int32_t value);
void LoadConst64(Register reg_hi, Register reg_lo, int64_t value);
