}
}
-/// 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.
// 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) {
}
}
+/// \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;
for (unsigned l = 0; l != 2; ++l) {
- unsigned HalfBegin = ((Imm >> (l * 4)) & 0x3) * HalfSize;
+ unsigned HalfMask = Imm >> (l * 4);
+ unsigned HalfBegin = (HalfMask & 0x3) * HalfSize;
for (unsigned i = HalfBegin, e = HalfBegin + HalfSize; i != e; ++i)
- ShuffleMask.push_back(i);
+ ShuffleMask.push_back(HalfMask & 8 ? SM_SentinelZero : (int)i);
}
}
// <4 x i32> <i32 -2147483648, i32 -2147483648,
// i32 -2147483648, i32 -2147483648>
+#ifndef NDEBUG
unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits();
-
- if (MaskTySize != 128 && MaskTySize != 256) // FIXME: Add support for AVX-512.
- return;
+ assert(MaskTySize == 128 || MaskTySize == 256 || MaskTySize == 512);
+#endif
// This is a straightforward byte vector.
if (MaskTy->isVectorTy() && MaskTy->getVectorElementType()->isIntegerTy(8)) {
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;
+ int Base = i & ~0xf;
Constant *COp = C->getAggregateElement(i);
if (!COp) {
ShuffleMask.clear();
}
}
-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!");
-
- 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);
- }
- } else if (auto *CV = dyn_cast<ConstantVector>(C)) {
- assert((unsigned)NumElements == C->getNumOperands() &&
- "Constant mask has a different number of elements!");
-
- 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;
- }
- 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);
+ unsigned NumElementsPerLane = 128 / ElSize;
+ unsigned Factor = ElSize / EltTySize;
+
+ 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;
}
+ 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);
}
+
+ // TODO: Handle funny-looking vectors too.
}
void DecodeZeroExtendMask(MVT SrcVT, MVT DstVT, SmallVectorImpl<int> &Mask) {
for (unsigned i = 1; i < NumElts; i++)
Mask.push_back(IsLoad ? static_cast<int>(SM_SentinelZero) : i);
}
-} // namespace llvm
+
+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);
+ }
+ }
+ }
+ 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