Skip to content

Commit

Permalink
[X86] Move shuffle decoding for constant pool into the X86CodeGen lib…
Browse files Browse the repository at this point in the history
…rary to remove a layering violation in the Util library.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@256680 91177308-0d34-0410-b5e6-96231b3b80d8
  • Loading branch information
topperc committed Dec 31, 2015
1 parent 2e3f468 commit 8cb3fe2
Show file tree
Hide file tree
Showing 7 changed files with 238 additions and 181 deletions.
1 change: 1 addition & 0 deletions lib/Target/X86/CMakeLists.txt
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,7 @@ set(sources
X86PadShortFunction.cpp
X86RegisterInfo.cpp
X86SelectionDAGInfo.cpp
X86ShuffleDecodeConstantPool.cpp
X86Subtarget.cpp
X86TargetMachine.cpp
X86TargetObjectFile.cpp
Expand Down
165 changes: 0 additions & 165 deletions lib/Target/X86/Utils/X86ShuffleDecode.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,6 @@
//===----------------------------------------------------------------------===//

#include "X86ShuffleDecode.h"
#include "llvm/IR/Constants.h"
#include "llvm/CodeGen/MachineValueType.h"

//===----------------------------------------------------------------------===//
Expand Down Expand Up @@ -296,54 +295,6 @@ void DecodeVPERM2X128Mask(MVT VT, unsigned Imm,
}
}

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.
// 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>

#ifndef NDEBUG
unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits();
assert(MaskTySize == 128 || MaskTySize == 256 || MaskTySize == 512);
#endif

// 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 & ~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,
SmallVectorImpl<int> &ShuffleMask) {
for (int i = 0, e = RawMask.size(); i < e; ++i) {
Expand Down Expand Up @@ -388,68 +339,6 @@ void DecodeVPERMMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
}
}

void DecodeVPERMILPMask(const Constant *C, unsigned ElSize,
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.
// 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);
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) {
unsigned NumDstElts = DstVT.getVectorNumElements();
unsigned SrcScalarBits = SrcVT.getScalarSizeInBits();
Expand Down Expand Up @@ -572,58 +461,4 @@ void DecodeVPERMV3Mask(ArrayRef<uint64_t> RawMask,
}
}

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
16 changes: 0 additions & 16 deletions lib/Target/X86/Utils/X86ShuffleDecode.h
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,6 @@
//===----------------------------------------------------------------------===//

namespace llvm {
class Constant;
class MVT;

enum { SM_SentinelUndef = -1, SM_SentinelZero = -2 };
Expand Down Expand Up @@ -72,9 +71,6 @@ void DecodeUNPCKHMask(MVT VT, SmallVectorImpl<int> &ShuffleMask);
/// different datatypes and vector widths.
void DecodeUNPCKLMask(MVT VT, SmallVectorImpl<int> &ShuffleMask);

/// \brief Decode a PSHUFB mask from an IR-level vector constant.
void DecodePSHUFBMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask);

/// \brief Decode a PSHUFB mask from a raw array of constants such as from
/// BUILD_VECTOR.
void DecodePSHUFBMask(ArrayRef<uint64_t> RawMask,
Expand All @@ -95,10 +91,6 @@ void decodeVSHUF64x2FamilyMask(MVT VT, unsigned Imm,
/// No VT provided since it only works on 256-bit, 4 element vectors.
void DecodeVPERMMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask);

/// \brief Decode a VPERMILP variable mask from an IR-level vector constant.
void DecodeVPERMILPMask(const Constant *C, unsigned ElSize,
SmallVectorImpl<int> &ShuffleMask);

/// \brief Decode a zero extension instruction as a shuffle mask.
void DecodeZeroExtendMask(MVT SrcVT, MVT DstVT,
SmallVectorImpl<int> &ShuffleMask);
Expand All @@ -118,18 +110,10 @@ void DecodeEXTRQIMask(int Len, int Idx,
void DecodeINSERTQIMask(int Len, int Idx,
SmallVectorImpl<int> &ShuffleMask);

/// \brief Decode a VPERM W/D/Q/PS/PD mask from an IR-level vector constant.
void DecodeVPERMVMask(const Constant *C, MVT VT,
SmallVectorImpl<int> &ShuffleMask);

/// \brief Decode a VPERM W/D/Q/PS/PD mask from a raw array of constants.
void DecodeVPERMVMask(ArrayRef<uint64_t> RawMask,
SmallVectorImpl<int> &ShuffleMask);

/// \brief Decode a VPERMT2 W/D/Q/PS/PD mask from an IR-level vector constant.
void DecodeVPERMV3Mask(const Constant *C, MVT VT,
SmallVectorImpl<int> &ShuffleMask);

/// \brief Decode a VPERMT2 W/D/Q/PS/PD mask from a raw array of constants.
void DecodeVPERMV3Mask(ArrayRef<uint64_t> RawMask,
SmallVectorImpl<int> &ShuffleMask);
Expand Down
1 change: 1 addition & 0 deletions lib/Target/X86/X86ISelLowering.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,7 @@
#include "X86FrameLowering.h"
#include "X86InstrBuilder.h"
#include "X86MachineFunctionInfo.h"
#include "X86ShuffleDecodeConstantPool.h"
#include "X86TargetMachine.h"
#include "X86TargetObjectFile.h"
#include "llvm/ADT/SmallBitVector.h"
Expand Down
1 change: 1 addition & 0 deletions lib/Target/X86/X86MCInstLower.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,7 @@

#include "X86AsmPrinter.h"
#include "X86RegisterInfo.h"
#include "X86ShuffleDecodeConstantPool.h"
#include "InstPrinter/X86ATTInstPrinter.h"
#include "MCTargetDesc/X86BaseInfo.h"
#include "Utils/X86ShuffleDecode.h"
Expand Down
Loading

0 comments on commit 8cb3fe2

Please sign in to comment.