//===----------------------------------------------------------------------===//
#include "X86ShuffleDecode.h"
-#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/IR/Constants.h"
-#include "llvm/IR/DataLayout.h"
-#include "llvm/IR/InstrTypes.h"
#include "llvm/CodeGen/MachineValueType.h"
//===----------------------------------------------------------------------===//
unsigned CountS = (Imm >> 6) & 3;
// CountS selects which input element to use.
- unsigned InVal = 4+CountS;
+ unsigned InVal = 4 + CountS;
// CountD specifies which element of destination to update.
ShuffleMask[CountD] = InVal;
// ZMask zaps values, potentially overriding the CountD elt.
// <3,1> or <6,7,2,3>
void DecodeMOVHLPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
- for (unsigned i = NElts/2; i != NElts; ++i)
- ShuffleMask.push_back(NElts+i);
+ for (unsigned i = NElts / 2; i != NElts; ++i)
+ ShuffleMask.push_back(NElts + i);
- for (unsigned i = NElts/2; i != NElts; ++i)
+ for (unsigned i = NElts / 2; i != NElts; ++i)
ShuffleMask.push_back(i);
}
// <0,2> or <0,1,4,5>
void DecodeMOVLHPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
- for (unsigned i = 0; i != NElts/2; ++i)
+ for (unsigned i = 0; i != NElts / 2; ++i)
ShuffleMask.push_back(i);
- for (unsigned i = 0; i != NElts/2; ++i)
- ShuffleMask.push_back(NElts+i);
+ for (unsigned i = 0; i != NElts / 2; ++i)
+ ShuffleMask.push_back(NElts + i);
}
void DecodeMOVSLDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
}
}
+void DecodeMOVDDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
+ unsigned VectorSizeInBits = VT.getSizeInBits();
+ unsigned ScalarSizeInBits = VT.getScalarSizeInBits();
+ unsigned NumElts = VT.getVectorNumElements();
+ unsigned NumLanes = VectorSizeInBits / 128;
+ unsigned NumLaneElts = NumElts / NumLanes;
+ unsigned NumLaneSubElts = 64 / ScalarSizeInBits;
+
+ for (unsigned l = 0; l < NumElts; l += NumLaneElts)
+ for (unsigned i = 0; i < NumLaneElts; i += NumLaneSubElts)
+ for (unsigned s = 0; s != NumLaneSubElts; s++)
+ ShuffleMask.push_back(l + s);
+}
+
void DecodePSLLDQMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
unsigned VectorSizeInBits = VT.getSizeInBits();
unsigned NumElts = VectorSizeInBits / 8;
}
}
-/// DecodePSHUFMask - This decodes the shuffle masks for pshufd, and vpermilp*.
+/// DecodePSHUFMask - This decodes the shuffle masks for pshufw, pshufd, and vpermilp*.
/// VT indicates the type of the vector allowing it to handle different
/// datatypes and vector widths.
void DecodePSHUFMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
unsigned NumElts = VT.getVectorNumElements();
unsigned NumLanes = VT.getSizeInBits() / 128;
+ if (NumLanes == 0) NumLanes = 1; // Handle MMX
unsigned NumLaneElts = NumElts / NumLanes;
unsigned NewImm = Imm;
}
}
+void DecodePSWAPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
+ unsigned NumElts = VT.getVectorNumElements();
+ unsigned NumHalfElts = NumElts / 2;
+
+ for (unsigned l = 0; l != NumHalfElts; ++l)
+ ShuffleMask.push_back(l + NumHalfElts);
+ for (unsigned h = 0; h != NumHalfElts; ++h)
+ ShuffleMask.push_back(h);
+}
+
/// DecodeSHUFPMask - This decodes the shuffle masks for shufp*. VT indicates
/// the type of the vector allowing it to handle different datatypes and vector
/// widths.
unsigned NewImm = Imm;
for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
// each half of a lane comes from different source
- for (unsigned s = 0; s != NumElts*2; s += NumElts) {
- for (unsigned i = 0; i != NumLaneElts/2; ++i) {
+ for (unsigned s = 0; s != NumElts * 2; s += NumElts) {
+ for (unsigned i = 0; i != NumLaneElts / 2; ++i) {
ShuffleMask.push_back(NewImm % NumLaneElts + s + l);
NewImm /= NumLaneElts;
}
// Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
// independently on 128-bit lanes.
unsigned NumLanes = VT.getSizeInBits() / 128;
- if (NumLanes == 0 ) NumLanes = 1; // Handle MMX
+ if (NumLanes == 0) NumLanes = 1; // Handle MMX
unsigned NumLaneElts = NumElts / NumLanes;
for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
- for (unsigned i = l + NumLaneElts/2, e = l + NumLaneElts; i != e; ++i) {
- ShuffleMask.push_back(i); // Reads from dest/src1
- ShuffleMask.push_back(i+NumElts); // Reads from src/src2
+ for (unsigned i = l + NumLaneElts / 2, e = l + NumLaneElts; i != e; ++i) {
+ ShuffleMask.push_back(i); // Reads from dest/src1
+ ShuffleMask.push_back(i + NumElts); // Reads from src/src2
}
}
}
unsigned NumLaneElts = NumElts / NumLanes;
for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
- for (unsigned i = l, e = l + NumLaneElts/2; i != e; ++i) {
- ShuffleMask.push_back(i); // Reads from dest/src1
- ShuffleMask.push_back(i+NumElts); // Reads from src/src2
+ for (unsigned i = l, e = l + NumLaneElts / 2; i != e; ++i) {
+ ShuffleMask.push_back(i); // Reads from dest/src1
+ ShuffleMask.push_back(i + NumElts); // Reads from src/src2
}
}
}
+/// \brief Decode a shuffle packed values at 128-bit granularity
+/// (SHUFF32x4/SHUFF64x2/SHUFI32x4/SHUFI64x2)
+/// immediate mask into a shuffle mask.
+void decodeVSHUF64x2FamilyMask(MVT VT, unsigned Imm,
+ SmallVectorImpl<int> &ShuffleMask) {
+ unsigned NumLanes = VT.getSizeInBits() / 128;
+ unsigned NumElementsInLane = 128 / VT.getScalarSizeInBits();
+ unsigned ControlBitsMask = NumLanes - 1;
+ unsigned NumControlBits = NumLanes / 2;
+
+ for (unsigned l = 0; l != NumLanes; ++l) {
+ unsigned LaneMask = (Imm >> (l * NumControlBits)) & ControlBitsMask;
+ // We actually need the other source.
+ if (l >= NumLanes / 2)
+ LaneMask += NumLanes;
+ for (unsigned i = 0; i != NumElementsInLane; ++i)
+ ShuffleMask.push_back(LaneMask * NumElementsInLane + i);
+ }
+}
+
void DecodeVPERM2X128Mask(MVT VT, unsigned Imm,
SmallVectorImpl<int> &ShuffleMask) {
- if (Imm & 0x88)
- return; // Not a shuffle
-
- unsigned HalfSize = VT.getVectorNumElements()/2;
+ unsigned HalfSize = VT.getVectorNumElements() / 2;
for (unsigned l = 0; l != 2; ++l) {
- unsigned HalfBegin = ((Imm >> (l*4)) & 0x3) * HalfSize;
- for (unsigned i = HalfBegin, e = HalfBegin+HalfSize; i != e; ++i)
- ShuffleMask.push_back(i);
+ unsigned HalfMask = Imm >> (l * 4);
+ unsigned HalfBegin = (HalfMask & 0x3) * HalfSize;
+ for (unsigned i = HalfBegin, e = HalfBegin + HalfSize; i != e; ++i)
+ ShuffleMask.push_back(HalfMask & 8 ? SM_SentinelZero : (int)i);
}
}
-void DecodePSHUFBMask(const Constant *C, const DataLayout *TD,
- SmallVectorImpl<int> &ShuffleMask) {
+void DecodePSHUFBMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) {
Type *MaskTy = C->getType();
// It is not an error for the PSHUFB mask to not be a vector of i8 because the
// constant pool uniques constants by their bit representation.
//
// <2 x i64> <i64 -9223372034707292160, i64 -9223372034707292160>
//
- // <4 x i32> <i32 -2147483648, i32 -2147483648, i32 -2147483648, i32
- // -2147483648>
+ // <4 x i32> <i32 -2147483648, i32 -2147483648,
+ // i32 -2147483648, i32 -2147483648>
+#ifndef NDEBUG
unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits();
+ assert(MaskTySize == 128 || MaskTySize == 256 || MaskTySize == 512);
+#endif
- VectorType *DestTy = nullptr;
- if (MaskTySize == 128)
- DestTy = VectorType::get(Type::getInt8Ty(MaskTy->getContext()), 16);
- else if (MaskTySize == 256)
- DestTy = VectorType::get(Type::getInt8Ty(MaskTy->getContext()), 32);
-
- // FIXME: Add support for AVX-512.
- if (!DestTy)
- return;
-
- if (DestTy != MaskTy) {
- if (!CastInst::castIsValid(Instruction::BitCast, const_cast<Constant *>(C),
- DestTy))
- return;
-
- C = ConstantFoldInstOperands(Instruction::BitCast, DestTy,
- const_cast<Constant *>(C), TD);
- MaskTy = DestTy;
- }
-
- int NumElements = MaskTy->getVectorNumElements();
- ShuffleMask.reserve(NumElements);
+ // This is a straightforward byte vector.
+ if (MaskTy->isVectorTy() && MaskTy->getVectorElementType()->isIntegerTy(8)) {
+ int NumElements = MaskTy->getVectorNumElements();
+ ShuffleMask.reserve(NumElements);
- for (int i = 0; i < NumElements; ++i) {
- // For AVX vectors with 32 bytes the base of the shuffle is the 16-byte
- // lane of the vector we're inside.
- int Base = i < 16 ? 0 : 16;
- Constant *COp = C->getAggregateElement(i);
- if (!COp) {
- ShuffleMask.clear();
- return;
- } else if (isa<UndefValue>(COp)) {
- ShuffleMask.push_back(SM_SentinelUndef);
- continue;
- }
- uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
- // If the high bit (7) of the byte is set, the element is zeroed.
- if (Element & (1 << 7))
- ShuffleMask.push_back(SM_SentinelZero);
- else {
- // Only the least significant 4 bits of the byte are used.
- int Index = Base + (Element & 0xf);
- ShuffleMask.push_back(Index);
+ for (int i = 0; i < NumElements; ++i) {
+ // For AVX vectors with 32 bytes the base of the shuffle is the 16-byte
+ // lane of the vector we're inside.
+ int Base = i & ~0xf;
+ Constant *COp = C->getAggregateElement(i);
+ if (!COp) {
+ ShuffleMask.clear();
+ return;
+ } else if (isa<UndefValue>(COp)) {
+ ShuffleMask.push_back(SM_SentinelUndef);
+ continue;
+ }
+ uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
+ // If the high bit (7) of the byte is set, the element is zeroed.
+ if (Element & (1 << 7))
+ ShuffleMask.push_back(SM_SentinelZero);
+ else {
+ // Only the least significant 4 bits of the byte are used.
+ int Index = Base + (Element & 0xf);
+ ShuffleMask.push_back(Index);
+ }
}
}
+ // TODO: Handle funny-looking vectors too.
}
void DecodePSHUFBMask(ArrayRef<uint64_t> RawMask,
/// No VT provided since it only works on 256-bit, 4 element vectors.
void DecodeVPERMMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
for (unsigned i = 0; i != 4; ++i) {
- ShuffleMask.push_back((Imm >> (2*i)) & 3);
+ ShuffleMask.push_back((Imm >> (2 * i)) & 3);
}
}
-void DecodeVPERMILPMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) {
+void DecodeVPERMILPMask(const Constant *C, unsigned ElSize,
+ SmallVectorImpl<int> &ShuffleMask) {
Type *MaskTy = C->getType();
- assert(MaskTy->isVectorTy() && "Expected a vector constant mask!");
- assert(MaskTy->getVectorElementType()->isIntegerTy() &&
- "Expected integer constant mask elements!");
- int ElementBits = MaskTy->getScalarSizeInBits();
- int NumElements = MaskTy->getVectorNumElements();
+ // It is not an error for the PSHUFB mask to not be a vector of i8 because the
+ // constant pool uniques constants by their bit representation.
+ // e.g. the following take up the same space in the constant pool:
+ // i128 -170141183420855150465331762880109871104
+ //
+ // <2 x i64> <i64 -9223372034707292160, i64 -9223372034707292160>
+ //
+ // <4 x i32> <i32 -2147483648, i32 -2147483648,
+ // i32 -2147483648, i32 -2147483648>
+
+ unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits();
+
+ if (MaskTySize != 128 && MaskTySize != 256) // FIXME: Add support for AVX-512.
+ return;
+
+ // Only support vector types.
+ if (!MaskTy->isVectorTy())
+ return;
+
+ // Make sure its an integer type.
+ Type *VecEltTy = MaskTy->getVectorElementType();
+ if (!VecEltTy->isIntegerTy())
+ return;
+
+ // Support any element type from byte up to element size.
+ // This is necesary primarily because 64-bit elements get split to 32-bit
+ // in the constant pool on 32-bit target.
+ unsigned EltTySize = VecEltTy->getIntegerBitWidth();
+ if (EltTySize < 8 || EltTySize > ElSize)
+ return;
+
+ unsigned NumElements = MaskTySize / ElSize;
assert((NumElements == 2 || NumElements == 4 || NumElements == 8) &&
"Unexpected number of vector elements.");
ShuffleMask.reserve(NumElements);
- if (auto *CDS = dyn_cast<ConstantDataSequential>(C)) {
- assert((unsigned)NumElements == CDS->getNumElements() &&
- "Constant mask has a different number of elements!");
+ unsigned NumElementsPerLane = 128 / ElSize;
+ unsigned Factor = ElSize / EltTySize;
- for (int i = 0; i < NumElements; ++i) {
- int Base = (i * ElementBits / 128) * (128 / ElementBits);
- uint64_t Element = CDS->getElementAsInteger(i);
- // Only the least significant 2 bits of the integer are used.
- int Index = Base + (Element & 0x3);
- ShuffleMask.push_back(Index);
+ for (unsigned i = 0; i < NumElements; ++i) {
+ Constant *COp = C->getAggregateElement(i * Factor);
+ if (!COp) {
+ ShuffleMask.clear();
+ return;
+ } else if (isa<UndefValue>(COp)) {
+ ShuffleMask.push_back(SM_SentinelUndef);
+ continue;
}
- } else if (auto *CV = dyn_cast<ConstantVector>(C)) {
- assert((unsigned)NumElements == C->getNumOperands() &&
- "Constant mask has a different number of elements!");
+ int Index = i & ~(NumElementsPerLane - 1);
+ uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
+ if (ElSize == 64)
+ Index += (Element >> 1) & 0x1;
+ else
+ Index += Element & 0x3;
+ ShuffleMask.push_back(Index);
+ }
- for (int i = 0; i < NumElements; ++i) {
- int Base = (i * ElementBits / 128) * (128 / ElementBits);
- Constant *COp = CV->getOperand(i);
- if (isa<UndefValue>(COp)) {
- ShuffleMask.push_back(SM_SentinelUndef);
- continue;
+ // TODO: Handle funny-looking vectors too.
+}
+
+void DecodeZeroExtendMask(MVT SrcVT, MVT DstVT, SmallVectorImpl<int> &Mask) {
+ unsigned NumDstElts = DstVT.getVectorNumElements();
+ unsigned SrcScalarBits = SrcVT.getScalarSizeInBits();
+ unsigned DstScalarBits = DstVT.getScalarSizeInBits();
+ unsigned Scale = DstScalarBits / SrcScalarBits;
+ assert(SrcScalarBits < DstScalarBits &&
+ "Expected zero extension mask to increase scalar size");
+ assert(SrcVT.getVectorNumElements() >= NumDstElts &&
+ "Too many zero extension lanes");
+
+ for (unsigned i = 0; i != NumDstElts; i++) {
+ Mask.push_back(i);
+ for (unsigned j = 1; j != Scale; j++)
+ Mask.push_back(SM_SentinelZero);
+ }
+}
+
+void DecodeZeroMoveLowMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
+ unsigned NumElts = VT.getVectorNumElements();
+ ShuffleMask.push_back(0);
+ for (unsigned i = 1; i < NumElts; i++)
+ ShuffleMask.push_back(SM_SentinelZero);
+}
+
+void DecodeScalarMoveMask(MVT VT, bool IsLoad, SmallVectorImpl<int> &Mask) {
+ // First element comes from the first element of second source.
+ // Remaining elements: Load zero extends / Move copies from first source.
+ unsigned NumElts = VT.getVectorNumElements();
+ Mask.push_back(NumElts);
+ for (unsigned i = 1; i < NumElts; i++)
+ Mask.push_back(IsLoad ? static_cast<int>(SM_SentinelZero) : i);
+}
+
+void DecodeEXTRQIMask(int Len, int Idx,
+ SmallVectorImpl<int> &ShuffleMask) {
+ // Only the bottom 6 bits are valid for each immediate.
+ Len &= 0x3F;
+ Idx &= 0x3F;
+
+ // We can only decode this bit extraction instruction as a shuffle if both the
+ // length and index work with whole bytes.
+ if (0 != (Len % 8) || 0 != (Idx % 8))
+ return;
+
+ // A length of zero is equivalent to a bit length of 64.
+ if (Len == 0)
+ Len = 64;
+
+ // If the length + index exceeds the bottom 64 bits the result is undefined.
+ if ((Len + Idx) > 64) {
+ ShuffleMask.append(16, SM_SentinelUndef);
+ return;
+ }
+
+ // Convert index and index to work with bytes.
+ Len /= 8;
+ Idx /= 8;
+
+ // EXTRQ: Extract Len bytes starting from Idx. Zero pad the remaining bytes
+ // of the lower 64-bits. The upper 64-bits are undefined.
+ for (int i = 0; i != Len; ++i)
+ ShuffleMask.push_back(i + Idx);
+ for (int i = Len; i != 8; ++i)
+ ShuffleMask.push_back(SM_SentinelZero);
+ for (int i = 8; i != 16; ++i)
+ ShuffleMask.push_back(SM_SentinelUndef);
+}
+
+void DecodeINSERTQIMask(int Len, int Idx,
+ SmallVectorImpl<int> &ShuffleMask) {
+ // Only the bottom 6 bits are valid for each immediate.
+ Len &= 0x3F;
+ Idx &= 0x3F;
+
+ // We can only decode this bit insertion instruction as a shuffle if both the
+ // length and index work with whole bytes.
+ if (0 != (Len % 8) || 0 != (Idx % 8))
+ return;
+
+ // A length of zero is equivalent to a bit length of 64.
+ if (Len == 0)
+ Len = 64;
+
+ // If the length + index exceeds the bottom 64 bits the result is undefined.
+ if ((Len + Idx) > 64) {
+ ShuffleMask.append(16, SM_SentinelUndef);
+ return;
+ }
+
+ // Convert index and index to work with bytes.
+ Len /= 8;
+ Idx /= 8;
+
+ // INSERTQ: Extract lowest Len bytes from lower half of second source and
+ // insert over first source starting at Idx byte. The upper 64-bits are
+ // undefined.
+ for (int i = 0; i != Idx; ++i)
+ ShuffleMask.push_back(i);
+ for (int i = 0; i != Len; ++i)
+ ShuffleMask.push_back(i + 16);
+ for (int i = Idx + Len; i != 8; ++i)
+ ShuffleMask.push_back(i);
+ for (int i = 8; i != 16; ++i)
+ ShuffleMask.push_back(SM_SentinelUndef);
+}
+
+void DecodeVPERMVMask(ArrayRef<uint64_t> RawMask,
+ SmallVectorImpl<int> &ShuffleMask) {
+ for (int i = 0, e = RawMask.size(); i < e; ++i) {
+ uint64_t M = RawMask[i];
+ ShuffleMask.push_back((int)M);
+ }
+}
+
+void DecodeVPERMV3Mask(ArrayRef<uint64_t> RawMask,
+ SmallVectorImpl<int> &ShuffleMask) {
+ for (int i = 0, e = RawMask.size(); i < e; ++i) {
+ uint64_t M = RawMask[i];
+ ShuffleMask.push_back((int)M);
+ }
+}
+
+void DecodeVPERMVMask(const Constant *C, MVT VT,
+ SmallVectorImpl<int> &ShuffleMask) {
+ Type *MaskTy = C->getType();
+ if (MaskTy->isVectorTy()) {
+ unsigned NumElements = MaskTy->getVectorNumElements();
+ if (NumElements == VT.getVectorNumElements()) {
+ for (unsigned i = 0; i < NumElements; ++i) {
+ Constant *COp = C->getAggregateElement(i);
+ if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp))) {
+ ShuffleMask.clear();
+ return;
+ }
+ if (isa<UndefValue>(COp))
+ ShuffleMask.push_back(SM_SentinelUndef);
+ else {
+ uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
+ Element &= (1 << NumElements) - 1;
+ ShuffleMask.push_back(Element);
+ }
}
- uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
- // Only the least significant 2 bits of the integer are used.
- int Index = Base + (Element & 0x3);
- ShuffleMask.push_back(Index);
}
+ return;
}
+ // Scalar value; just broadcast it
+ if (!isa<ConstantInt>(C))
+ return;
+ uint64_t Element = cast<ConstantInt>(C)->getZExtValue();
+ int NumElements = VT.getVectorNumElements();
+ Element &= (1 << NumElements) - 1;
+ for (int i = 0; i < NumElements; ++i)
+ ShuffleMask.push_back(Element);
}
+void DecodeVPERMV3Mask(const Constant *C, MVT VT,
+ SmallVectorImpl<int> &ShuffleMask) {
+ Type *MaskTy = C->getType();
+ unsigned NumElements = MaskTy->getVectorNumElements();
+ if (NumElements == VT.getVectorNumElements()) {
+ for (unsigned i = 0; i < NumElements; ++i) {
+ Constant *COp = C->getAggregateElement(i);
+ if (!COp) {
+ ShuffleMask.clear();
+ return;
+ }
+ if (isa<UndefValue>(COp))
+ ShuffleMask.push_back(SM_SentinelUndef);
+ else {
+ uint64_t Element = cast<ConstantInt>(COp)->getZExtValue();
+ Element &= (1 << NumElements*2) - 1;
+ ShuffleMask.push_back(Element);
+ }
+ }
+ }
+}
} // llvm namespace
projects / oota-llvm.git / blobdiff