1 //===-- X86ShuffleDecode.cpp - X86 shuffle decode logic -------------------===// 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 into a 11 // generic vector mask. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "X86ShuffleDecode.h" 16 #include "llvm/IR/Constants.h" 17 #include "llvm/CodeGen/MachineValueType.h" 18 19 //===----------------------------------------------------------------------===// 20 // Vector Mask Decoding 21 //===----------------------------------------------------------------------===// 22 23 namespace llvm { 24 25 void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) { 26 // Defaults the copying the dest value. 27 ShuffleMask.push_back(0); 28 ShuffleMask.push_back(1); 29 ShuffleMask.push_back(2); 30 ShuffleMask.push_back(3); 31 32 // Decode the immediate. 33 unsigned ZMask = Imm & 15; 34 unsigned CountD = (Imm >> 4) & 3; 35 unsigned CountS = (Imm >> 6) & 3; 36 37 // CountS selects which input element to use. 38 unsigned InVal = 4+CountS; 39 // CountD specifies which element of destination to update. 40 ShuffleMask[CountD] = InVal; 41 // ZMask zaps values, potentially overriding the CountD elt. 42 if (ZMask & 1) ShuffleMask[0] = SM_SentinelZero; 43 if (ZMask & 2) ShuffleMask[1] = SM_SentinelZero; 44 if (ZMask & 4) ShuffleMask[2] = SM_SentinelZero; 45 if (ZMask & 8) ShuffleMask[3] = SM_SentinelZero; 46 } 47 48 // <3,1> or <6,7,2,3> 49 void DecodeMOVHLPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) { 50 for (unsigned i = NElts/2; i != NElts; ++i) 51 ShuffleMask.push_back(NElts+i); 52 53 for (unsigned i = NElts/2; i != NElts; ++i) 54 ShuffleMask.push_back(i); 55 } 56 57 // <0,2> or <0,1,4,5> 58 void DecodeMOVLHPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) { 59 for (unsigned i = 0; i != NElts/2; ++i) 60 ShuffleMask.push_back(i); 61 62 for (unsigned i = 0; i != NElts/2; ++i) 63 ShuffleMask.push_back(NElts+i); 64 } 65 66 void DecodeMOVSLDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) { 67 unsigned NumElts = VT.getVectorNumElements(); 68 for (int i = 0, e = NumElts / 2; i < e; ++i) { 69 ShuffleMask.push_back(2 * i); 70 ShuffleMask.push_back(2 * i); 71 } 72 } 73 74 void DecodeMOVSHDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) { 75 unsigned NumElts = VT.getVectorNumElements(); 76 for (int i = 0, e = NumElts / 2; i < e; ++i) { 77 ShuffleMask.push_back(2 * i + 1); 78 ShuffleMask.push_back(2 * i + 1); 79 } 80 } 81 82 void DecodeMOVDDUPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) { 83 unsigned VectorSizeInBits = VT.getSizeInBits(); 84 unsigned ScalarSizeInBits = VT.getScalarSizeInBits(); 85 unsigned NumElts = VT.getVectorNumElements(); 86 unsigned NumLanes = VectorSizeInBits / 128; 87 unsigned NumLaneElts = NumElts / NumLanes; 88 unsigned NumLaneSubElts = 64 / ScalarSizeInBits; 89 90 for (unsigned l = 0; l < NumElts; l += NumLaneElts) 91 for (unsigned i = 0; i < NumLaneElts; i += NumLaneSubElts) 92 for (unsigned s = 0; s != NumLaneSubElts; s++) 93 ShuffleMask.push_back(l + s); 94 } 95 96 void DecodePSLLDQMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) { 97 unsigned VectorSizeInBits = VT.getSizeInBits(); 98 unsigned NumElts = VectorSizeInBits / 8; 99 unsigned NumLanes = VectorSizeInBits / 128; 100 unsigned NumLaneElts = NumElts / NumLanes; 101 102 for (unsigned l = 0; l < NumElts; l += NumLaneElts) 103 for (unsigned i = 0; i < NumLaneElts; ++i) { 104 int M = SM_SentinelZero; 105 if (i >= Imm) M = i - Imm + l; 106 ShuffleMask.push_back(M); 107 } 108 } 109 110 void DecodePSRLDQMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) { 111 unsigned VectorSizeInBits = VT.getSizeInBits(); 112 unsigned NumElts = VectorSizeInBits / 8; 113 unsigned NumLanes = VectorSizeInBits / 128; 114 unsigned NumLaneElts = NumElts / NumLanes; 115 116 for (unsigned l = 0; l < NumElts; l += NumLaneElts) 117 for (unsigned i = 0; i < NumLaneElts; ++i) { 118 unsigned Base = i + Imm; 119 int M = Base + l; 120 if (Base >= NumLaneElts) M = SM_SentinelZero; 121 ShuffleMask.push_back(M); 122 } 123 } 124 125 void DecodePALIGNRMask(MVT VT, unsigned Imm, 126 SmallVectorImpl<int> &ShuffleMask) { 127 unsigned NumElts = VT.getVectorNumElements(); 128 unsigned Offset = Imm * (VT.getVectorElementType().getSizeInBits() / 8); 129 130 unsigned NumLanes = VT.getSizeInBits() / 128; 131 unsigned NumLaneElts = NumElts / NumLanes; 132 133 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 134 for (unsigned i = 0; i != NumLaneElts; ++i) { 135 unsigned Base = i + Offset; 136 // if i+offset is out of this lane then we actually need the other source 137 if (Base >= NumLaneElts) Base += NumElts - NumLaneElts; 138 ShuffleMask.push_back(Base + l); 139 } 140 } 141 } 142 143 /// DecodePSHUFMask - This decodes the shuffle masks for pshufd, and vpermilp*. 144 /// VT indicates the type of the vector allowing it to handle different 145 /// datatypes and vector widths. 146 void DecodePSHUFMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) { 147 unsigned NumElts = VT.getVectorNumElements(); 148 149 unsigned NumLanes = VT.getSizeInBits() / 128; 150 unsigned NumLaneElts = NumElts / NumLanes; 151 152 unsigned NewImm = Imm; 153 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 154 for (unsigned i = 0; i != NumLaneElts; ++i) { 155 ShuffleMask.push_back(NewImm % NumLaneElts + l); 156 NewImm /= NumLaneElts; 157 } 158 if (NumLaneElts == 4) NewImm = Imm; // reload imm 159 } 160 } 161 162 void DecodePSHUFHWMask(MVT VT, unsigned Imm, 163 SmallVectorImpl<int> &ShuffleMask) { 164 unsigned NumElts = VT.getVectorNumElements(); 165 166 for (unsigned l = 0; l != NumElts; l += 8) { 167 unsigned NewImm = Imm; 168 for (unsigned i = 0, e = 4; i != e; ++i) { 169 ShuffleMask.push_back(l + i); 170 } 171 for (unsigned i = 4, e = 8; i != e; ++i) { 172 ShuffleMask.push_back(l + 4 + (NewImm & 3)); 173 NewImm >>= 2; 174 } 175 } 176 } 177 178 void DecodePSHUFLWMask(MVT VT, unsigned Imm, 179 SmallVectorImpl<int> &ShuffleMask) { 180 unsigned NumElts = VT.getVectorNumElements(); 181 182 for (unsigned l = 0; l != NumElts; l += 8) { 183 unsigned NewImm = Imm; 184 for (unsigned i = 0, e = 4; i != e; ++i) { 185 ShuffleMask.push_back(l + (NewImm & 3)); 186 NewImm >>= 2; 187 } 188 for (unsigned i = 4, e = 8; i != e; ++i) { 189 ShuffleMask.push_back(l + i); 190 } 191 } 192 } 193 194 /// DecodeSHUFPMask - This decodes the shuffle masks for shufp*. VT indicates 195 /// the type of the vector allowing it to handle different datatypes and vector 196 /// widths. 197 void DecodeSHUFPMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) { 198 unsigned NumElts = VT.getVectorNumElements(); 199 200 unsigned NumLanes = VT.getSizeInBits() / 128; 201 unsigned NumLaneElts = NumElts / NumLanes; 202 203 unsigned NewImm = Imm; 204 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 205 // each half of a lane comes from different source 206 for (unsigned s = 0; s != NumElts*2; s += NumElts) { 207 for (unsigned i = 0; i != NumLaneElts/2; ++i) { 208 ShuffleMask.push_back(NewImm % NumLaneElts + s + l); 209 NewImm /= NumLaneElts; 210 } 211 } 212 if (NumLaneElts == 4) NewImm = Imm; // reload imm 213 } 214 } 215 216 /// DecodeUNPCKHMask - This decodes the shuffle masks for unpckhps/unpckhpd 217 /// and punpckh*. VT indicates the type of the vector allowing it to handle 218 /// different datatypes and vector widths. 219 void DecodeUNPCKHMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) { 220 unsigned NumElts = VT.getVectorNumElements(); 221 222 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate 223 // independently on 128-bit lanes. 224 unsigned NumLanes = VT.getSizeInBits() / 128; 225 if (NumLanes == 0 ) NumLanes = 1; // Handle MMX 226 unsigned NumLaneElts = NumElts / NumLanes; 227 228 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 229 for (unsigned i = l + NumLaneElts/2, e = l + NumLaneElts; i != e; ++i) { 230 ShuffleMask.push_back(i); // Reads from dest/src1 231 ShuffleMask.push_back(i+NumElts); // Reads from src/src2 232 } 233 } 234 } 235 236 /// DecodeUNPCKLMask - This decodes the shuffle masks for unpcklps/unpcklpd 237 /// and punpckl*. VT indicates the type of the vector allowing it to handle 238 /// different datatypes and vector widths. 239 void DecodeUNPCKLMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) { 240 unsigned NumElts = VT.getVectorNumElements(); 241 242 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate 243 // independently on 128-bit lanes. 244 unsigned NumLanes = VT.getSizeInBits() / 128; 245 if (NumLanes == 0 ) NumLanes = 1; // Handle MMX 246 unsigned NumLaneElts = NumElts / NumLanes; 247 248 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 249 for (unsigned i = l, e = l + NumLaneElts/2; i != e; ++i) { 250 ShuffleMask.push_back(i); // Reads from dest/src1 251 ShuffleMask.push_back(i+NumElts); // Reads from src/src2 252 } 253 } 254 } 255 256 void DecodeVPERM2X128Mask(MVT VT, unsigned Imm, 257 SmallVectorImpl<int> &ShuffleMask) { 258 if (Imm & 0x88) 259 return; // Not a shuffle 260 261 unsigned HalfSize = VT.getVectorNumElements()/2; 262 263 for (unsigned l = 0; l != 2; ++l) { 264 unsigned HalfBegin = ((Imm >> (l*4)) & 0x3) * HalfSize; 265 for (unsigned i = HalfBegin, e = HalfBegin+HalfSize; i != e; ++i) 266 ShuffleMask.push_back(i); 267 } 268 } 269 270 void DecodePSHUFBMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) { 271 Type *MaskTy = C->getType(); 272 // It is not an error for the PSHUFB mask to not be a vector of i8 because the 273 // constant pool uniques constants by their bit representation. 274 // e.g. the following take up the same space in the constant pool: 275 // i128 -170141183420855150465331762880109871104 276 // 277 // <2 x i64> <i64 -9223372034707292160, i64 -9223372034707292160> 278 // 279 // <4 x i32> <i32 -2147483648, i32 -2147483648, 280 // i32 -2147483648, i32 -2147483648> 281 282 unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits(); 283 284 if (MaskTySize != 128 && MaskTySize != 256) // FIXME: Add support for AVX-512. 285 return; 286 287 // This is a straightforward byte vector. 288 if (MaskTy->isVectorTy() && MaskTy->getVectorElementType()->isIntegerTy(8)) { 289 int NumElements = MaskTy->getVectorNumElements(); 290 ShuffleMask.reserve(NumElements); 291 292 for (int i = 0; i < NumElements; ++i) { 293 // For AVX vectors with 32 bytes the base of the shuffle is the 16-byte 294 // lane of the vector we're inside. 295 int Base = i < 16 ? 0 : 16; 296 Constant *COp = C->getAggregateElement(i); 297 if (!COp) { 298 ShuffleMask.clear(); 299 return; 300 } else if (isa<UndefValue>(COp)) { 301 ShuffleMask.push_back(SM_SentinelUndef); 302 continue; 303 } 304 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue(); 305 // If the high bit (7) of the byte is set, the element is zeroed. 306 if (Element & (1 << 7)) 307 ShuffleMask.push_back(SM_SentinelZero); 308 else { 309 // Only the least significant 4 bits of the byte are used. 310 int Index = Base + (Element & 0xf); 311 ShuffleMask.push_back(Index); 312 } 313 } 314 } 315 // TODO: Handle funny-looking vectors too. 316 } 317 318 void DecodePSHUFBMask(ArrayRef<uint64_t> RawMask, 319 SmallVectorImpl<int> &ShuffleMask) { 320 for (int i = 0, e = RawMask.size(); i < e; ++i) { 321 uint64_t M = RawMask[i]; 322 if (M == (uint64_t)SM_SentinelUndef) { 323 ShuffleMask.push_back(M); 324 continue; 325 } 326 // For AVX vectors with 32 bytes the base of the shuffle is the half of 327 // the vector we're inside. 328 int Base = i < 16 ? 0 : 16; 329 // If the high bit (7) of the byte is set, the element is zeroed. 330 if (M & (1 << 7)) 331 ShuffleMask.push_back(SM_SentinelZero); 332 else { 333 // Only the least significant 4 bits of the byte are used. 334 int Index = Base + (M & 0xf); 335 ShuffleMask.push_back(Index); 336 } 337 } 338 } 339 340 void DecodeBLENDMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) { 341 int ElementBits = VT.getScalarSizeInBits(); 342 int NumElements = VT.getVectorNumElements(); 343 for (int i = 0; i < NumElements; ++i) { 344 // If there are more than 8 elements in the vector, then any immediate blend 345 // mask applies to each 128-bit lane. There can never be more than 346 // 8 elements in a 128-bit lane with an immediate blend. 347 int Bit = NumElements > 8 ? i % (128 / ElementBits) : i; 348 assert(Bit < 8 && 349 "Immediate blends only operate over 8 elements at a time!"); 350 ShuffleMask.push_back(((Imm >> Bit) & 1) ? NumElements + i : i); 351 } 352 } 353 354 /// DecodeVPERMMask - this decodes the shuffle masks for VPERMQ/VPERMPD. 355 /// No VT provided since it only works on 256-bit, 4 element vectors. 356 void DecodeVPERMMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) { 357 for (unsigned i = 0; i != 4; ++i) { 358 ShuffleMask.push_back((Imm >> (2*i)) & 3); 359 } 360 } 361 362 void DecodeVPERMILPMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) { 363 Type *MaskTy = C->getType(); 364 assert(MaskTy->isVectorTy() && "Expected a vector constant mask!"); 365 assert(MaskTy->getVectorElementType()->isIntegerTy() && 366 "Expected integer constant mask elements!"); 367 int ElementBits = MaskTy->getScalarSizeInBits(); 368 int NumElements = MaskTy->getVectorNumElements(); 369 assert((NumElements == 2 || NumElements == 4 || NumElements == 8) && 370 "Unexpected number of vector elements."); 371 ShuffleMask.reserve(NumElements); 372 if (auto *CDS = dyn_cast<ConstantDataSequential>(C)) { 373 assert((unsigned)NumElements == CDS->getNumElements() && 374 "Constant mask has a different number of elements!"); 375 376 for (int i = 0; i < NumElements; ++i) { 377 int Base = (i * ElementBits / 128) * (128 / ElementBits); 378 uint64_t Element = CDS->getElementAsInteger(i); 379 // Only the least significant 2 bits of the integer are used. 380 int Index = Base + (Element & 0x3); 381 ShuffleMask.push_back(Index); 382 } 383 } else if (auto *CV = dyn_cast<ConstantVector>(C)) { 384 assert((unsigned)NumElements == C->getNumOperands() && 385 "Constant mask has a different number of elements!"); 386 387 for (int i = 0; i < NumElements; ++i) { 388 int Base = (i * ElementBits / 128) * (128 / ElementBits); 389 Constant *COp = CV->getOperand(i); 390 if (isa<UndefValue>(COp)) { 391 ShuffleMask.push_back(SM_SentinelUndef); 392 continue; 393 } 394 uint64_t Element = cast<ConstantInt>(COp)->getZExtValue(); 395 // Only the least significant 2 bits of the integer are used. 396 int Index = Base + (Element & 0x3); 397 ShuffleMask.push_back(Index); 398 } 399 } 400 } 401 402 void DecodeZeroExtendMask(MVT SrcVT, MVT DstVT, SmallVectorImpl<int> &Mask) { 403 unsigned NumDstElts = DstVT.getVectorNumElements(); 404 unsigned SrcScalarBits = SrcVT.getScalarSizeInBits(); 405 unsigned DstScalarBits = DstVT.getScalarSizeInBits(); 406 unsigned Scale = DstScalarBits / SrcScalarBits; 407 assert(SrcScalarBits < DstScalarBits && 408 "Expected zero extension mask to increase scalar size"); 409 assert(SrcVT.getVectorNumElements() >= NumDstElts && 410 "Too many zero extension lanes"); 411 412 for (unsigned i = 0; i != NumDstElts; i++) { 413 Mask.push_back(i); 414 for (unsigned j = 1; j != Scale; j++) 415 Mask.push_back(SM_SentinelZero); 416 } 417 } 418 419 void DecodeZeroMoveLowMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) { 420 unsigned NumElts = VT.getVectorNumElements(); 421 ShuffleMask.push_back(0); 422 for (unsigned i = 1; i < NumElts; i++) 423 ShuffleMask.push_back(SM_SentinelZero); 424 } 425 426 void DecodeScalarMoveMask(MVT VT, bool IsLoad, SmallVectorImpl<int> &Mask) { 427 // First element comes from the first element of second source. 428 // Remaining elements: Load zero extends / Move copies from first source. 429 unsigned NumElts = VT.getVectorNumElements(); 430 Mask.push_back(NumElts); 431 for (unsigned i = 1; i < NumElts; i++) 432 Mask.push_back(IsLoad ? static_cast<int>(SM_SentinelZero) : i); 433 } 434 } // llvm namespace 435