1 //===-- X86ShuffleDecodeConstantPool.cpp - X86 shuffle decode -------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // Define several functions to decode x86 specific shuffle semantics using 11 // constants from the constant pool. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "X86ShuffleDecodeConstantPool.h" 16 #include "Utils/X86ShuffleDecode.h" 17 #include "llvm/CodeGen/MachineValueType.h" 18 #include "llvm/IR/Constants.h" 19 20 //===----------------------------------------------------------------------===// 21 // Vector Mask Decoding 22 //===----------------------------------------------------------------------===// 23 24 namespace llvm { 25 26 void DecodePSHUFBMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) { 27 Type *MaskTy = C->getType(); 28 // It is not an error for the PSHUFB mask to not be a vector of i8 because the 29 // constant pool uniques constants by their bit representation. 30 // e.g. the following take up the same space in the constant pool: 31 // i128 -170141183420855150465331762880109871104 32 // 33 // <2 x i64> <i64 -9223372034707292160, i64 -9223372034707292160> 34 // 35 // <4 x i32> <i32 -2147483648, i32 -2147483648, 36 // i32 -2147483648, i32 -2147483648> 37 38 #ifndef NDEBUG 39 unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits(); 40 assert(MaskTySize == 128 || MaskTySize == 256 || MaskTySize == 512); 41 #endif 42 43 if (!MaskTy->isVectorTy()) 44 return; 45 int NumElts = MaskTy->getVectorNumElements(); 46 47 Type *EltTy = MaskTy->getVectorElementType(); 48 if (!EltTy->isIntegerTy()) 49 return; 50 51 // The shuffle mask requires a byte vector - decode cases with 52 // wider elements as well. 53 unsigned BitWidth = cast<IntegerType>(EltTy)->getBitWidth(); 54 if ((BitWidth % 8) != 0) 55 return; 56 57 int Scale = BitWidth / 8; 58 int NumBytes = NumElts * Scale; 59 ShuffleMask.reserve(NumBytes); 60 61 for (int i = 0; i != NumElts; ++i) { 62 Constant *COp = C->getAggregateElement(i); 63 if (!COp) { 64 ShuffleMask.clear(); 65 return; 66 } else if (isa<UndefValue>(COp)) { 67 ShuffleMask.append(Scale, SM_SentinelUndef); 68 continue; 69 } 70 71 APInt APElt = cast<ConstantInt>(COp)->getValue(); 72 for (int j = 0; j != Scale; ++j) { 73 // For AVX vectors with 32 bytes the base of the shuffle is the 16-byte 74 // lane of the vector we're inside. 75 int Base = ((i * Scale) + j) & ~0xf; 76 77 uint64_t Element = APElt.getLoBits(8).getZExtValue(); 78 APElt = APElt.lshr(8); 79 80 // If the high bit (7) of the byte is set, the element is zeroed. 81 if (Element & (1 << 7)) 82 ShuffleMask.push_back(SM_SentinelZero); 83 else { 84 // Only the least significant 4 bits of the byte are used. 85 int Index = Base + (Element & 0xf); 86 ShuffleMask.push_back(Index); 87 } 88 } 89 } 90 91 assert(NumBytes == (int)ShuffleMask.size() && "Unexpected shuffle mask size"); 92 } 93 94 void DecodeVPERMILPMask(const Constant *C, unsigned ElSize, 95 SmallVectorImpl<int> &ShuffleMask) { 96 Type *MaskTy = C->getType(); 97 // It is not an error for the PSHUFB mask to not be a vector of i8 because the 98 // constant pool uniques constants by their bit representation. 99 // e.g. the following take up the same space in the constant pool: 100 // i128 -170141183420855150465331762880109871104 101 // 102 // <2 x i64> <i64 -9223372034707292160, i64 -9223372034707292160> 103 // 104 // <4 x i32> <i32 -2147483648, i32 -2147483648, 105 // i32 -2147483648, i32 -2147483648> 106 107 unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits(); 108 109 if (MaskTySize != 128 && MaskTySize != 256) // FIXME: Add support for AVX-512. 110 return; 111 112 // Only support vector types. 113 if (!MaskTy->isVectorTy()) 114 return; 115 116 // Make sure its an integer type. 117 Type *VecEltTy = MaskTy->getVectorElementType(); 118 if (!VecEltTy->isIntegerTy()) 119 return; 120 121 // Support any element type from byte up to element size. 122 // This is necessary primarily because 64-bit elements get split to 32-bit 123 // in the constant pool on 32-bit target. 124 unsigned EltTySize = VecEltTy->getIntegerBitWidth(); 125 if (EltTySize < 8 || EltTySize > ElSize) 126 return; 127 128 unsigned NumElements = MaskTySize / ElSize; 129 assert((NumElements == 2 || NumElements == 4 || NumElements == 8) && 130 "Unexpected number of vector elements."); 131 ShuffleMask.reserve(NumElements); 132 unsigned NumElementsPerLane = 128 / ElSize; 133 unsigned Factor = ElSize / EltTySize; 134 135 for (unsigned i = 0; i < NumElements; ++i) { 136 Constant *COp = C->getAggregateElement(i * Factor); 137 if (!COp) { 138 ShuffleMask.clear(); 139 return; 140 } else if (isa<UndefValue>(COp)) { 141 ShuffleMask.push_back(SM_SentinelUndef); 142 continue; 143 } 144 int Index = i & ~(NumElementsPerLane - 1); 145 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue(); 146 if (ElSize == 64) 147 Index += (Element >> 1) & 0x1; 148 else 149 Index += Element & 0x3; 150 ShuffleMask.push_back(Index); 151 } 152 153 // TODO: Handle funny-looking vectors too. 154 } 155 156 void DecodeVPERMIL2PMask(const Constant *C, unsigned M2Z, unsigned ElSize, 157 SmallVectorImpl<int> &ShuffleMask) { 158 Type *MaskTy = C->getType(); 159 160 unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits(); 161 if (MaskTySize != 128 && MaskTySize != 256) 162 return; 163 164 // Only support vector types. 165 if (!MaskTy->isVectorTy()) 166 return; 167 168 // Make sure its an integer type. 169 Type *VecEltTy = MaskTy->getVectorElementType(); 170 if (!VecEltTy->isIntegerTy()) 171 return; 172 173 // Support any element type from byte up to element size. 174 // This is necessary primarily because 64-bit elements get split to 32-bit 175 // in the constant pool on 32-bit target. 176 unsigned EltTySize = VecEltTy->getIntegerBitWidth(); 177 if (EltTySize < 8 || EltTySize > ElSize) 178 return; 179 180 unsigned NumElements = MaskTySize / ElSize; 181 assert((NumElements == 2 || NumElements == 4 || NumElements == 8) && 182 "Unexpected number of vector elements."); 183 ShuffleMask.reserve(NumElements); 184 unsigned NumElementsPerLane = 128 / ElSize; 185 unsigned Factor = ElSize / EltTySize; 186 187 for (unsigned i = 0; i < NumElements; ++i) { 188 Constant *COp = C->getAggregateElement(i * Factor); 189 if (!COp) { 190 ShuffleMask.clear(); 191 return; 192 } else if (isa<UndefValue>(COp)) { 193 ShuffleMask.push_back(SM_SentinelUndef); 194 continue; 195 } 196 197 // VPERMIL2 Operation. 198 // Bits[3] - Match Bit. 199 // Bits[2:1] - (Per Lane) PD Shuffle Mask. 200 // Bits[2:0] - (Per Lane) PS Shuffle Mask. 201 uint64_t Selector = cast<ConstantInt>(COp)->getZExtValue(); 202 int MatchBit = (Selector >> 3) & 0x1; 203 204 // M2Z[0:1] MatchBit 205 // 0Xb X Source selected by Selector index. 206 // 10b 0 Source selected by Selector index. 207 // 10b 1 Zero. 208 // 11b 0 Zero. 209 // 11b 1 Source selected by Selector index. 210 if ((M2Z & 0x2) != 0u && MatchBit != (M2Z & 0x1)) { 211 ShuffleMask.push_back(SM_SentinelZero); 212 continue; 213 } 214 215 int Index = i & ~(NumElementsPerLane - 1); 216 if (ElSize == 64) 217 Index += (Selector >> 1) & 0x1; 218 else 219 Index += Selector & 0x3; 220 221 int Src = (Selector >> 2) & 0x1; 222 Index += Src * NumElements; 223 ShuffleMask.push_back(Index); 224 } 225 226 // TODO: Handle funny-looking vectors too. 227 } 228 229 void DecodeVPPERMMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) { 230 Type *MaskTy = C->getType(); 231 assert(MaskTy->getPrimitiveSizeInBits() == 128); 232 233 // Only support vector types. 234 if (!MaskTy->isVectorTy()) 235 return; 236 237 // Make sure its an integer type. 238 Type *VecEltTy = MaskTy->getVectorElementType(); 239 if (!VecEltTy->isIntegerTy()) 240 return; 241 242 // The shuffle mask requires a byte vector - decode cases with 243 // wider elements as well. 244 unsigned BitWidth = cast<IntegerType>(VecEltTy)->getBitWidth(); 245 if ((BitWidth % 8) != 0) 246 return; 247 248 int NumElts = MaskTy->getVectorNumElements(); 249 int Scale = BitWidth / 8; 250 int NumBytes = NumElts * Scale; 251 ShuffleMask.reserve(NumBytes); 252 253 for (int i = 0; i != NumElts; ++i) { 254 Constant *COp = C->getAggregateElement(i); 255 if (!COp) { 256 ShuffleMask.clear(); 257 return; 258 } else if (isa<UndefValue>(COp)) { 259 ShuffleMask.append(Scale, SM_SentinelUndef); 260 continue; 261 } 262 263 // VPPERM Operation 264 // Bits[4:0] - Byte Index (0 - 31) 265 // Bits[7:5] - Permute Operation 266 // 267 // Permute Operation: 268 // 0 - Source byte (no logical operation). 269 // 1 - Invert source byte. 270 // 2 - Bit reverse of source byte. 271 // 3 - Bit reverse of inverted source byte. 272 // 4 - 00h (zero - fill). 273 // 5 - FFh (ones - fill). 274 // 6 - Most significant bit of source byte replicated in all bit positions. 275 // 7 - Invert most significant bit of source byte and replicate in all bit positions. 276 APInt MaskElt = cast<ConstantInt>(COp)->getValue(); 277 for (int j = 0; j != Scale; ++j) { 278 APInt Index = MaskElt.getLoBits(5); 279 APInt PermuteOp = MaskElt.lshr(5).getLoBits(3); 280 MaskElt = MaskElt.lshr(8); 281 282 if (PermuteOp == 4) { 283 ShuffleMask.push_back(SM_SentinelZero); 284 continue; 285 } 286 if (PermuteOp != 0) { 287 ShuffleMask.clear(); 288 return; 289 } 290 ShuffleMask.push_back((int)Index.getZExtValue()); 291 } 292 } 293 294 assert(NumBytes == (int)ShuffleMask.size() && "Unexpected shuffle mask size"); 295 } 296 297 void DecodeVPERMVMask(const Constant *C, MVT VT, 298 SmallVectorImpl<int> &ShuffleMask) { 299 Type *MaskTy = C->getType(); 300 if (MaskTy->isVectorTy()) { 301 unsigned NumElements = MaskTy->getVectorNumElements(); 302 if (NumElements == VT.getVectorNumElements()) { 303 for (unsigned i = 0; i < NumElements; ++i) { 304 Constant *COp = C->getAggregateElement(i); 305 if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp))) { 306 ShuffleMask.clear(); 307 return; 308 } 309 if (isa<UndefValue>(COp)) 310 ShuffleMask.push_back(SM_SentinelUndef); 311 else { 312 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue(); 313 Element &= (1 << NumElements) - 1; 314 ShuffleMask.push_back(Element); 315 } 316 } 317 } 318 return; 319 } 320 // Scalar value; just broadcast it 321 if (!isa<ConstantInt>(C)) 322 return; 323 uint64_t Element = cast<ConstantInt>(C)->getZExtValue(); 324 int NumElements = VT.getVectorNumElements(); 325 Element &= (1 << NumElements) - 1; 326 for (int i = 0; i < NumElements; ++i) 327 ShuffleMask.push_back(Element); 328 } 329 330 void DecodeVPERMV3Mask(const Constant *C, MVT VT, 331 SmallVectorImpl<int> &ShuffleMask) { 332 Type *MaskTy = C->getType(); 333 unsigned NumElements = MaskTy->getVectorNumElements(); 334 if (NumElements == VT.getVectorNumElements()) { 335 unsigned EltMaskSize = Log2_64(NumElements * 2); 336 for (unsigned i = 0; i < NumElements; ++i) { 337 Constant *COp = C->getAggregateElement(i); 338 if (!COp) { 339 ShuffleMask.clear(); 340 return; 341 } 342 if (isa<UndefValue>(COp)) 343 ShuffleMask.push_back(SM_SentinelUndef); 344 else { 345 APInt Element = cast<ConstantInt>(COp)->getValue(); 346 Element = Element.getLoBits(EltMaskSize); 347 ShuffleMask.push_back(Element.getZExtValue()); 348 } 349 } 350 } 351 } 352 } // llvm namespace 353